Vilfan I. - Lecture Notes in Statistical Mechanics (2002) PDF

Lecture Notes
in
Statistical Mechanics
Igor Vilfan
The Abdus Salam ICTP, Trieste, Italy

and
The J. Stefan Institute, Ljubljana, Slovenia
Trieste, Autumn 2002

I. Vilfan Statistical Mechanics
ii
Preface
The present Lecture Notes in Statistical Mechanics were written for the students
of the ICTP Diploma Course in Condensed Matter Physics at the Abdus Salam
ICTP in Trieste, Italy.
The lectures cover classical and quantum statistical mechanics with some em-
phasis on classical spin systems. I give also an introduction to Bose condensation
and superfluidity but I do not discuss phenomena specific to Fermi particles, being
covered by other lecturers.
I.V.
iii
I. Vilfan Statistical Mechanics
iv
Contents
Preface iii
Tables ix
1 Foundations 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Ensembles in Statistical Mechanics . . . . . . . . . . . . . . . . . 4
1.2.1 Microcanonical Ensemble and the Entropy
(Isolated Systems) . . . . . . . . . . . . . . . . . . . . . 4
1.2.2 Canonical Ensemble and the Free Energy
(Systems at Fixed Temperature) . . . . . . . . . . . . . . 6
1.2.3 Grand Canonical Ensemble
(Open systems) . . . . . . . . . . . . . . . . . . . . . . . 10
1.3 Analogy Between Fluids and Magnetic Systems . . . . . . . . . . 12
1.4 Fluctuations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.5 Phase Transitions . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.5.1 Order Parameter . . . . . . . . . . . . . . . . . . . . . . 18
1.5.2 Critical Point . . . . . . . . . . . . . . . . . . . . . . . . 19
1.5.3 Critical Exponents . . . . . . . . . . . . . . . . . . . . . 21
2 Classical Models 25
2.1 Real Gases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2 Ising Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2.1 Lattice Gas . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2.2 Binary Alloys . . . . . . . . . . . . . . . . . . . . . . . . 32
2.2.3 Ising model in d = 1 . . . . . . . . . . . . . . . . . . . . 33
2.2.4 Ising Model in d = 2 . . . . . . . . . . . . . . . . . . . . 38
2.3 Potts and Related Models . . . . . . . . . . . . . . . . . . . . . . 40
v
I. Vilfan Statistical Mechanics Contents
2.4 Two-dimensional xy Model . . . . . . . . . . . . . . . . . . . . . 42

2.5 Spin Glasses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.5.1 Some Physical Properties of Spin Glasses . . . . . . . . . 49
3 Quantum Models 53
3.1 Quantum Statistics . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.1.1 The Density Matrix . . . . . . . . . . . . . . . . . . . . . 54
3.1.2 Ensembles in Quantum Statistical Mechanics . . . . . . . 55
3.2 Bose Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.2.1 Ideal Bose Gas . . . . . . . . . . . . . . . . . . . . . . . 58
3.2.2 Bose Einstein Condensation . . . . . . . . . . . . . . . . 61
3.3 Liquid 4 He . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.3.1 Superfluidity . . . . . . . . . . . . . . . . . . . . . . . . 68
3.3.2 Vortex Excitations . . . . . . . . . . . . . . . . . . . . . 69
3.3.3 Order Parameter . . . . . . . . . . . . . . . . . . . . . . 70
4 Methods of Statistical Mechanics 73

4.1 Mean-Field Theories . . . . . . . . . . . . . . . . . . . . . . . . 73
4.1.1 Weiss Molecular Field of an Ising System . . . . . . . . . 74
4.1.2 Bragg-Williams Approximation . . . . . . . . . . . . . . 75
4.1.3 Landau Theory of Continuous Phase Transitions . . . . . 79
4.1.4 Critical Exponents in the Landau Theory . . . . . . . . . 85
4.1.5 Short-Range Order and Fluctuations . . . . . . . . . . . . 86
4.1.6 Validity of The Mean-field Theory . . . . . . . . . . . . . 88
4.2 Monte Carlo Simulations . . . . . . . . . . . . . . . . . . . . . . 90
4.2.1 Importance Sampling . . . . . . . . . . . . . . . . . . . . 90
4.2.2 Metropolis Algorithm . . . . . . . . . . . . . . . . . . . 91
4.2.3 Practical Details and Data Analysis . . . . . . . . . . . . 91
4.3 The Renormalization Group Methods . . . . . . . . . . . . . . . 96
4.3.1 A Trivial Example: The d = 1 Ising Model . . . . . . . . 101
4.3.2 RG Transformations on the d = 2 Ising Model . . . . . . 104
4.3.3 General Properties of the RG Transformations . . . . . . . 107
4.4 Momentum-Space Renormalization
Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
4.4.1 The Gaussian Model . . . . . . . . . . . . . . . . . . . . 118
4.4.2 The Landau-Wilson Model . . . . . . . . . . . . . . . . . 121
4.4.3 Critical Exponents . . . . . . . . . . . . . . . . . . . . . 123
vi
Further Reading 129
Appendices 131
A Distribution Functions 131
B Maxwell Relations 133
C Basic Thermodynamic Relations of Magnetic Systems Revisited 135
vii
viii
Tables
Table 1: Ensembles in statistical mechanics
Ensemble Partition function Thermodynamic potential
Microcanonical ∆Ω(E, V, N ) S(E, V, N ) = kB ln ∆Ω(E, V, N )
Canonical Z(T, V, N ) F (T, V, N ) = −kB T ln Z(T, V, N )

= i e−Ei (V,N )/kB T
P
= ∆Ω(E, V, N )e−E(V,N )/kB T dE

R
Y (T, p, N ) G(T, p, N ) = −kB T ln Y (T, p, µ)
Grand canonical Ξ(T, V, µ) J(T, V, µ) = F − µN = −pV

= N eµN/kB T Z(T, V, N ) = −kB T ln Ξ(T, V, µ)
P
ix
I. Vilfan Statistical Mechanics Tables
Table 2: Thermodynamic functions
Thermodynamic functions Natural Total differential

(Definition) variables
Entropy
S hEi, V, N dS = dhEi/T + pdV /T − µdN/T
Internal energy
hEi S, V, N dhEi = T dS − pdV + µdN
Enthalpy
H = hEi + pV S, p, N dH = T dS + V dp + µdN
Helmholtz free energy

F = hEi − T S T, V, N dF = −SdT − pdV + µdN
Gibbs free energy

G = F + pV = N µ T, p, N dG = −SdT + V dp + µdN
Grand potential
J = F − G = −pV T, V, µ dJ = −SdT − pdV − N dµ
x
Table 3: Relations of different thermodynamic quantities to the partition function
Canonical partition function

(constant number of particles)
Z(T, V, N ) = ∆Ω(E, V, N )e−βE dE

R
Discrete energy levels:

Z(T, V, N ) = i e−βEi
P
Free energy
F (T, V, N ) = −kB T ln Z(T, V, N ) G(T, P, N ) = F (T, V, N ) + P V
@ @
@ @
@ @
R
@ R
@
Pressure Entropy Entropy Volume
∂F
P = −( ∂V )T S = −( ∂F )
∂T V
S = −( ∂G )
∂T P
V = ( ∂G )
∂P T
? ? ? ?
Inverse Heat capacity Heat capacity Isothermal

compressibility (constant V ) (constant P ) compressibility
1 ∂P ∂S ∂S
κT
= −V ( ∂V )T CV = T ( ∂T )V CP = T ( ∂T )P κT = − V1 ( ∂V )
∂P T
2
∂ F 2 2 2
= V ( ∂V 2 )T = −T ( ∂∂TF2 )V = −T ( ∂∂TG2 )P = − V1 ( ∂∂PG2 )T
xi
xii
Chapter 1
Foundations
1.1 Introduction
A macroscopic system is, typically, composed of the order of 1023 particles. It
is impossible to know, investigate, or describe the exact microscopic behaviour
of each individual particle in such a system. We must limit our knowledge to the
average properties, thermodynamic quantities like the temperature or pressure and
correlation functions.
Statistical mechanics is the bridge between the microscopic and macroscopic
world, it provides methods of calculating the macroscopic properties, like the spe-
cific heat, from the microscopic information, like the interaction energy between
the particles. The essential ingredient of statistical mechanics is the probability
distribution, i.e., the collection of occupancies of different configurations (micro-
scopic states).
Phase Space. Consider as an example the ideal gas of N particles in a three–
dimensional space. The Hamiltonian is
3N
X p2i X
H= + Vij ; Vij → 0. (1.1)
i=1 2m i,j6=i
(pi is the component of the momentum, m is the particle mass, and Vij the in-
teraction between particles. The interaction term is necessary for the system of
particles to reach thermal equilibrium, but it can be neglected compared to the
kinetic energy.). The 3N coordinates and 3N momenta form the 6N dimensional
phase space. Each point in the phase space represents a microscopic state of the
1
I. Vilfan Statistical Mechanics Chapter 1
system. The microscopic state changes in time according to the canonical equa-
tions of motion:
∂H ∂H
ṗj = − and q̇j = . (1.2)
∂qj ∂pj
and traces a trajectory in the phase space.
Ensembles. The system is completely described if we know the coordinates
and momenta of all particles (6N variables - a terrible task). To get an average of
a physical quantity, say A, one should make the time average over a long segment
of the trajectory (a collection of consecutive configurations) of such a system in
the phase space,
1 Z t0 +∆t
hAi = lim A[{qi (t), pi (t)}]dt. (1.3)
∆t→∞ ∆t t0
However, usually we know neither the exact microscopic state, i.e., the loca-
tion of the system in the phase space, nor the trajectory. What we can know, is
the macroscopic state (temperature, pressure, volume, ...). Instead of studying a
particular microscopic state and the trajectory over the phase space, we investigate
an ensemble of systems, i.e., a collection of all possible microscopic systems, all
belonging to the same macroscopic state.
Ergodicity. To calculate the average of a physical quantity, we substitute the
time average over one system by an average over an ensemble of equivalent sys-
tems at fixed time:
Z
hAi = A({qi , pi })p({qi , pi })d3N qd3N p. (1.4)
W
The integral is over the whole phase space W and p is the probability density
in the phase space, it is the probability that a unit volume of the phase space
is occupied. So, instead of the time average, we make an ensemble average at
fixed time. When the time and the ensemble averages are equal, we say that the
system is ergodic. In an alternative but equivalent definition we say that a system
is ergodic if it evolves in such a way that it visits with equal probability all points
of the phase space which are accessible from the initial configuration subject to
the constraint of energy conservation. In other words, the trajectory in the phase
space of an ergodic system will spend equal time intervals in all regions of the
constant energy surface. Most of the physical systems of interest in statistical
mechanics are ergodic, so we will in most of the cases replace time averages by
the corresponding ensemble averages.
2
Γ(Ε) ∆Ω
states
Figure 1.1: Constant-energy sur- Figure 1.2: Constant-energy shell

face in the phase space. The area of in the 6N -dimensional phase space.
the surface in the 6N -dimensional The number of states in the shell is
phase space is Γ(E). ∆Ω(E).
For the above example of an isolated ideal gas, the total energy of the system
is constant, H = E, the trajectory lies on the constant-energy surface Γ(E) in the
phase space and the ensemble average of A is:
1 Z
hAi = A({qi , pi })d3N q d3N p. (1.5)
Γ(E) Γ(E)
(The integral is over the constant energy surface Γ(E).) For quantum systems as
well as for classical systems with discrete energy levels, the number of configura-
tions is an extremely irregular function of the energy E. Therefore we introduce
a narrow energy interval of the width ∆E and the ergodic hypothesis is general-
ized so that all the configurations, satisfying the condition that their energy is in
the interval between E and E + ∆E, are equally likely to occur. The number
of configurations ∆Ω with the energy in the interval between E and E + ∆E
is proportional to the volume of the corresponding shell in the phase space. For
indistinguishable particles, ∆Ω is:
1 X
∆Ω(E) = 1. (1.6)
h3N N ! E≤H≤E+∆E
In principle, all systems in Nature obey quantum statistical mechanics, the

energy levels are discrete, identical particles are indistinguishable and their wave-
functions are either symmetric (Bosons) or antisymmetric (Fermions) upon ex-
change of two particles. Classical statistical mechanics is valid only as a special,
limiting case when the average occupation of any single–particle quantum state is
<< 1.
3
Although we will discuss only statistical mechanics of systems in equilibrium,

i.e., equilibrium statistical mechanics, there is a growing interest in nonequilib-
rium statistical mechanics which investigates steady, stationary states of nonequi-
librium systems.
1.2 Ensembles in Statistical Mechanics

An ensemble is a collection of many configurations of a system, all describing the
same thermodynamic state. We start with a discussion of isolated systems.
1.2.1 Microcanonical Ensemble and the Entropy

(Isolated Systems)
A microcanonical ensemble represents a collection of configurations of isolated
systems that have reached thermal equilibrium. A system is isolated from its
environment if it exchanges neither particles nor energy with its surrounding. The
volume, internal energy and number of particles of such a system are constant
and are the same for all configurations that are part of the same microcanonical
ensemble. Instead of working with constant energy surface Γ(E), we consider all
the states with the total energy in a narrow energy interval between E and E+∆E.
All such configurations are lying in a thin shell in the phase space and the total
number of configurations is ∆Ω.
Closely related to the ergodic hypothesis is the postulate of equal weights
which says that in equilibrium of an isolated system, all the configurations with
equal energy are equally probable. For a system, belonging to a microcanonical
ensemble, therefore, the probability for being in a configuration qi , pi with the
energy H({qi , pi }) is:
(
1/∆Ω(E) E ≤ H({qi , pi }) ≤ E + ∆E
p({qi , pi }) = (1.7)
0 otherwise.
∆Ω(E) is the number of configurations having the energy in the interval E ≤

H({qi , pi }) ≤ E + ∆E. All configurations {qi , pi } which fall into this energy
interval are occupied with equal probability p({qi , pi }).
4
Entropy and Probability

Probably the most fundamental thermodynamic quantity in statistical physics is
the entropy. Entropy is a concept, defined only for macroscopic systems. A sys-
tem of only a few particles does not have an entropy. The entropy is a measure of
disorder in the system. and is an additive function of state.
Let us consider an isolated system composed of
two subsystems, which are in contact with each other
1 2 so that they can exchange energy and particles. As
a result of the contact between the two subsystems,
an additional uncertainty has been introduced into the
system (at the beginning we knew the energy and the number of particles of each
individual subsystem, at the end we know only the total energy and number of
particles of the whole system).
We postulate that the entropy is The entropy is an additive quantity, therefore:
S = S1 + S2 (1.8)
We postulate that the entropy is
S = kB ln ∆Ω. (1.9)
This relation was discovered by Ludwig Boltzmann and is therefore called the
Boltzmann equation. The proportionality constant kB is the Boltzmann constant
and is determined by relating the statistical definition of the entropy to the ther-
modynamic definition. The Boltzmann equation provides the bridge between ther-
modynamics (entropy) and statistical mechanics (number of configurations).
The entropy plays the role of thermodynamic potential for microcanonical
ensembles, (−S) is minimal for an isolated system in equilibrium (V , E and N
are kept constant). From thermodynamics we know that the total differential of
the entropy is:
T dS(E, V, N ) = dE + P dV − µ dN. (1.10)
(This is the first law of thermodynamics.) This equation provides a statistical
definition of the temperature T , pressure P and chemical potential µ:
!
1 ∂S
= , (1.11)
T ∂E N,V
5
!
∂S
P =T , (1.12)
∂V N,E
!
∂S
µ = −T . (1.13)
∂N E,V
T , p and µ are intensive quantities, they are independent of N whereas the entropy
and the energy E are extensive quantities, they are proportional to the number of
particles N .
1.2.2 Canonical Ensemble and the Free Energy

(Systems at Fixed Temperature)
A canonical ensemble is a collection of closed systems at constant temperature.
Closed systems have fixed number of particles and volume but can exchange en-
ergy with their surroundings. The energy of a closed system is therefore not con-
stant. As an example we can imagine a gas of particles (air molecules, e.g.) in a
closed container at constant temperature.
Consider now an isolated system composed of
a small closed (sub)system s with N particles in
a volume V and of a “heat bath” (b) at tempera-
ture T . Let the two subsystems be in contact so
s
that they can exchange energy but no particles. In
bath
equilibrium, the system s has equal temperature
as the heat bath. Now, the energy of the small
system, Es , is no longer fixed and the system s
Figure 1.3: A small closed will be in any energy state Es , in each one with
system (s) in contact with a a different probability ps . The heat bath must be
large system, called “bath.” large (Nb Ns ), so that the change in the energy
s and bath can exchange en- of the bath, Eb , during energy exchange between
ergy. the small system and the bath, is negligible. Ev-
idently, since the compound system (= system +
bath) is isolated, the total energy Et = Eb + Es is constant and the compound sys-
tem belongs to a microcanonical ensemble for which the number of configurations
in the energy interval between Et and Et + ∆E is:
∆Ωt (Et ) = ∆Ωs (Es )∆Ωb (Et − Es ). (1.14)
6
We first express ∆Ωb with the entropy via the Boltzmann equation (1.9) and then
expand ln ∆Ωb to first order in Es :
!
∂Sb Es
kB ln ∆Ωb (Et − Es ) = Sb (Et − Es ) ≈ Sb (Et ) − Es = Sb (Et ) − .
∂E T
(1.15)
Thus,
∆Ωb (Et − Es ) = ∆Ωb (Et ) e−βEs = C e−βEs . (1.16)
Here, the constant C is independent of Es and β is the inverse temperature, β =
1/kB T .
The (normalized) probability for the system s having the energy between Es
and Es + dEs is:
∆Ωs (Es )∆Ωb (Et − Es )dEs

p(Es )dEs = R . (1.17)
dEs ∆Ωs (Es )∆Ωb (Et − Es )
After expressing ∆Ωb (Et − Es ) with (1.16) we find the probability density:
1
p(Es ) = ρ(Es )e−βEs , (1.18)
Z
where ρ(Es ) = ∆Ω(Es )/∆E is the density of states (number of states per unit en-
ergy interval). Z is called the partition function and is the normalization constant,
depending on the temperature, volume and number of particles of the system s:
Z
Z(T, V, N ) = ρ(E)e−βE dE. (1.19)
(Here we omit the index s.) Z(T, V, N ) is a Laplace transform of the microcanoni-
cal partition function ∆Ω(E) = ρ(E)∆E. For systems with discrete energy levels
we replace the integral either by a sum over all the energy levels l:
gl e−βEl
X
Z(T, V, N ) = (1.20)
l
(gl is the degeneracy of the energy level l), or by a sum over all configurations i:
e−βEi .
X
Z(T, V, N ) = (1.21)
i
7
The role of Z is much more important than just being the normalization con-
stant. Z is obtained by integration over the whole phase space. In this way infor-
mation on microscopic states of the system (coordinates and momenta of individ-
ual particles, e.g.) becomes lost but all thermodynamic information on the system
(with prescribed temperature, number of particles and volume) is retained and can
be obtained from Z, as we shall see below.
The distribution in which the particles are in equilibrium with a heat bath at
temperature T is called a canonical distribution similarly as the ensemble defined
by this distribution is called a canonical ensemble. For a system in equilibrium at
a fixed temperature, the system can (at least in principle) asses any energy. The
probability density p(E) is the same for all states with the same energy but it
decreases exponentially with increasing energy.
Free Energy
In this Section we will see the relation between the average of a physical quantity
of a canonical ensemble and the partition function Z.
The average energy hEi and average generalized force hXi in a canonical
distribution are given by
1Z ∂
hEi = E ρ(E)e−βE dE = − ln Z, (1.22)
Z ∂β
1Z −βE
Z
∂E −βE ∂ 1
hXi = Xρ(E) e dE = e dE = − ln Z. (1.23)
Z ∂x ∂x β
x is the generalized coordinate, conjugate to the generalized force. Examples of
conjugate forces and coordinates are: pressure vs. volume in gases, magnetic field
vs. magnetization in magnetic systems; the generalized force is intensive while
the coordinate is an extensive quantity.
We introduce the (Helmholtz) free energy as:
F (T, V, N ) = −kB T ln Z(T, V, N ) . (1.24)
Like the Boltzmann equation (1.9) in case of the microcanonical ensemble, also
this equation relates thermodynamics (free energy) and statistical mechanics (par-
tition function).
Relation between F and S: We take the well-known relation from thermody-
namics:
F (T, V, N ) = hE(S, V, N )i − T S, (1.25)
8
F (T, V, N ) is a Legendre transform of hE(S, V, N )i, by which the independent

variables are changed from (S, V, N ) to (T, V, N ).
The total differential of F is:
dF = −SdT − P dV + µdN. (1.26)
Another Legendre transform relates the Helmholtz and the Gibbs free energies
F (T, V, N ) and G(T, P, N ):
G(T, P, N ) = F (T, V, N ) + P V (1.27)
For a list of thermodynamic functions and their derivatives, see Tables 1 to 3 in

the Appendix B. Exercise: Construct these tables for magnetic systems!
Stability and convexity of the free energies (See Fig. 1.4). The Helmholtz free
energy F (T, V ) (for simplicity, we will omit N ) is a concave (i.e, the second
derivative is negative) function of T and convex function of V and the Gibbs free
energy G(T, P ) is always a concave function of both T and P .
F G
V p
Figure 1.4: Schematic behaviour of the Helmholtz and Gibbs free energies.
These properties follow from the fact that the system must be thermally and
mechanically stable, therefore the specific heat and the compressibility must be
positive quantities. We begin by noting that
∂G ∂F
S = −( ) = −( ) . (1.28)
∂T p ∂T V
It follows that
∂2G ∂S 1
( 2
) = −( ) ≡ − Cp ≤ 0 (1.29)
∂T p ∂T p T
9
and
∂2F ∂S 1
( 2
) = −( ) ≡ − CV ≤ 0. (1.30)
∂T V ∂T V T
Hence, G(T, p) and F (T, V ) are concave functions of T .
Together with equation (1.26) we see that for a system in equilibrium at con-
stant T, V, N the distribution of a system over possible configurations is such that
F is minimal. Therefore F is also called canonical or Gibbs potential.
The next two properties are obtained by differentiating G(T, P ) and F (T, V )
with respect to pressure and volume, respectively:
∂2G ∂V
( 2
) =( ) ≡ −V κT ≤ 0 (1.31)
∂P T ∂P T
∂2F ∂P 1
( 2
) = −( ) ≡ ≥ 0. (1.32)
∂V T ∂V T V κT
Hence, G(T, P ) is a concave function of pressure P and F (T, V ) is a convex
function of volume V .
The above properties are valid also for systems that undergo phase transitions.
1.2.3 Grand Canonical Ensemble

(Open systems)
Consider now a more general case with the system
of volume Vs being not only in contact with a heat
bath at temperature T , but also exchanging parti-
cles with the bath. As an example we can imagine
s
a room at constant temperature with an open win-
bath
dow. In this case, neither the energy nor the num-
ber of particles in the system (air molecules in the
room, e.g.) are fixed. Such a system is described
by the grand canonical distribution.
We proceed like in case of the canonical ensemble, we consider the total sys-
tem being part of a microcanonical ensemble, therefore the number of configura-
tions of the total system is:
∆Ω(Et , Nt ) = ∆Ωs (Ei , Ns ) ∆Ωb (Et − Es , Nt − Ns ). (1.33)
10
We express ∆Ωb with the entropy via the Boltzmann equation and then expand
ln ∆Ωb to first order in Es and Ns :
kB ln ∆Ωb (Et − Es , Nt − Ns ) = Sb (Et − Es , Nt − Ns )
! !
∂Sb ∂Sb
≈ Sb (Et , Nt ) − Es − Ns
∂E ∂N
Es µNs
= Sb (Et , Nt ) − + .
T T
Thus,
∆Ωb (Et − Es , Nt − Ns ) = C e−β(Es −µNs ) . (1.34)
Eventually we find for the probability density of a system having the energy E
and N particles:
1
p(E, N ) = ρ(E, N )e−β(E−µN ) , (1.35)
Ξ
where the normalization constant Ξ is called the grand (canonical) partition func-
tion:
XZ
Ξ(T, V, µ) = ρ(E, N ) e−β(E−µN ) dE. (1.36)
N
(Again, we omit the index s.) The grand partition function Ξ is related to the
partition function Z:
Z(T, V, N ) eβµN .
X
Ξ(T, V, µ) = (1.37)
N
Thermodynamic quantities in the grand canonical ensemble

In analogy to the free energy in the case of a canonical ensemble, we introduce
the grand (canonical) potential J which will be used for open systems:
J(T, V, µ) = −kB T ln Ξ (1.38)
The total differential of J is:
dJ = −SdT − pdV − N dµ. (1.39)
The entropy can now be expressed as
! !
∂J ∂ ln Ξ
S=− = kB ln Ξ + kB T , (1.40)
∂T V,µ
∂T V,µ
11
the pressure is: ! !

∂J ∂ ln Ξ
p=− = kB T , (1.41)
∂V T,µ
∂V T,µ
and the number of particles is
! !
∂J ∂ ln Ξ
N =− = kB T . (1.42)
∂µ T,V
∂µ T,V
For a system in equilibrium at constant T , V , and µ, the distribution of a system

over possible configurations is such that the potential J is minimal.
1.3 Analogy Between Fluids and Magnetic Systems

Above, we have mainly discussed fluids (liquids, gases). When we want to con-
sider, e.g., magnetic systems, the above relations between thermodynamic quanti-
ties are easily written down for them as well. The potential energy of a magnetic
dipole µ in a magnetic field H is
W = −µH (1.43)
and the work required to increase the magnetization of a system by dM is:
−Hdµ. (1.44)
Hence, the first law of thermodynamics for magnets reads:
T dS = dE − HdM (1.45)
Comparing this equation with 1.10 we see that we only have to replace the ”gen-
eralized force” - pressure P by the ”force”, acting on the magnetic moment, i.e.,
by the negative magnetic field:
p → −H. (1.46)
Similarly, the ”generalized coordinate” V is replaced by the ”generalized coordi-

nate” of magnetic systems, M :
V → M. (1.47)
Notice that P and H are intensive quantities whereas V and M are extensive.
12
1.4 Fluctuations
The microcanonical ensemble describes a collection of isolated systems where
neither energy nor the number of particles can fluctuate.
Energy fluctuations in the canonical ensemble: The canonical ensemble de-
scribes a system in contact with a heat bath. For such a system, the energy is
not exactly fixed, it can fluctuate around hEi. We will now evaluate the square of
the energy fluctuations,
(∆E)2 = h(E − hEi)2 i = hE 2 i − hEi2 . (1.48)
In the canonical ensemble, the thermal average of the energy is

1 ∂Z
hEi = − (1.49)
Z ∂β
and the thermal average of the energy squared is:
2 1Z 2 −βE 1 ∂2Z
hE i = E ρ(E)e dE = . (1.50)
Z Z ∂β 2
So, the square of the energy fluctuation is
1 ∂2Z 1 ∂Z 2 ∂ 2 ln Z
(∆E)2 = − ( ) = . (1.51)
Z ∂β 2 Z ∂β ∂β 2
After expressing this with hEi (see paragraph 2.3.3) we get:
∂hEi ∂hEi
(∆E)2 = − = kB T 2 . (1.52)
∂β ∂T
(∆E)2 = kB T 2 CV . (1.53)
Here, CV is the heat capacity, it is an extensive quantity and is proportional to N .
The relative fluctuation of the energy is thus:
s
∆E kB c V T
= , (1.54)
hEi N ε
where ε is the energy per particle and cV is the specific heat
√ at constant volume,
cV = CV /N . The relative energy fluctuation is ∝ 1/ N and it vanishes as
N → ∞, see Fig. 1.5. For macroscopic systems (i.e., in the limit N → ∞), the
13
p(E)
p(E)
Width Width
0 <E> 0 <E>
E E
(a) (b)
Figure 1.5: Fluctuation in the energy of a small (a) and of a large (b) system.
distribution of energies in a canonical ensemble is so sharply peaked around the

average energy that the canonical ensemble is equivalent to the microcanonical
ensemble. For later use we write Eq. (1.54) in the form:
β 2
cv = hE i − hEi2 (1.55)
NT
Fluctuations in the grand canonical ensemble: Now the number of particles is
not fixed, so we shall calculate ∆N (and ∆E). The calculation is analogous to
the above calculation for the canonical ensemble. We start with the grand partition
function XZ
Ξ(T, V, µ) = ρ(E)e−β(Ei −µN ) dE (1.56)
N
and differentiate it twice with respect to µ by keeping β and V constant,
1 ∂Ξ
hN i = , (1.57)
βΞ ∂(µ)
1 ∂2Ξ
hN 2 i = . (1.58)
β 2 Ξ ∂µ2
The square of the particle number fluctuation is:
∂ 2 ln Ξ
!
2 2 2
(∆N ) = hN i − hN i = . (1.59)
β 2 ∂µ2 T,V
For fluids at constant T and V , we obtain from the thermodynamic relation (See
Table 1) pV = kB T ln Ξ:
14
∂ 2 ln Ξ V ∂2p
= . (1.60)
∂µ2 kB T ∂µ2
We assume that the (Helmholtz) free energy can be written as F (T, V, N )
= N f (T, v), where f (T, v) is the free energy per particle and depends on T and
v ≡ V /N . Then,
µ(T, p) = f (T, v) + pv (1.61)
and
∂2p ∂ 1 1 ∂v

2
= =− , (1.62)
∂µ ∂µ v v 2 ∂µ
where
∂µ ∂µ ∂p ∂p 1
= =v ≡− . (1.63)
∂v ∂p ∂v ∂v κT
κT is the isothermal compressibility. We get finally that
hN ikB T κT
hN 2 i − hN i2 = . (1.64)
v
The square of the fluctuation in the number of particles is proportional
√ to the
(isothermal) compressibility! This equation shows also that ∆N ∝ N and that
∆N/hN i vanishes as N → ∞. We conclude that the grand canonical ensemble is
equivalent to the canonical ensemble when hN i → ∞. The decision about which
ensemble is to be used to describe the investigated system does not depend on
whether the system is closed or open, but it is a matter of convenience. For later
use we rewrite Eq. 1.64 in a different form:
vβ 2
κT = hN i − hN i2 (1.65)
hN i
The limit N → ∞, in which the relative fluctuations become negligible, is
called the thermodynamic limit. So far we assumed cV or κT to be constant.
At a continuous phase transition however (see Section 1.5), these quantities and
consequently also the corresponding fluctuations diverge on all lengthscales. This
makes life of experimentalists and of those who make simulations on computers
(both can deal with finite systems only) very difficult.
The fluctuations of N around hN i or E around hEi are - of course - time-
dependent processes! Therefore we have to assume that the systems are ergodic
in order to identify the time average with an ensemble average which we actually
calculated.
15
Fluctuation-Dissipation Theorem
Above, we have shown that (∆E)2 is proportional to the specific heat cV and that
(∆N )2 is proportional to the compressibility. In the problem 3 we also see that
(∆M )2 is proportional to the susceptibility. All these examples indicate that there
is a general relation between the fluctuations of the generalized coordinate and the
response of this coordinate to the perturbation of the conjugate generalized force.
This is the message of the fluctuation-dissipation theorem which in a most
general way relates the fluctuations of a physical quantity of a system in equilib-
rium to a dissipation process which is realized when the system is subject to an
external force driving it away from equilibrium. The theorem makes it possible
to find the non-equilibrium properties of a system from the analysis of thermal
fluctuations of the system in equilibrium. It is a bridge between non-equilibrium
and equilibrium statistical mechanics.
Now let us have a look at magnetic systems. The magnetic susceptibility, i.e.,
the response of the magnetization M to the extermal magnetic field H is:
β 2
χ= hM i − hM i2 . (1.66)
N
P
If we write M = i Si , the susceptibility becomes:
β X
χ= [hSi Sj i − hSi ihSj i] . (1.67)
N i,j
The summand on the right-hand side is the spin-spin correlation function,
Γij = [hSi Sj i − hSi ihSj i] . (1.68)
which describes the correlations in the spatial fluctuations of the spin around its
thermal average. Because of the interaction between the nearest-neighbouring
spins, one expects that the spins are correlated on short distances and that their
fluctuations are less and less correlated as the distance increases. The correlation
function thus decays with the distance. Often Γ will be assumed to have the form
Γ(r) ∝ r−p e−r/ξ , (1.69)
where the decay length ξ is called the correlation length. As we shall see later, ξ is
a temperature-dependent quantity which plays a crucial role in phase transitions,
it diverges at a continuous phase transition.
16
µ µ
µ2 µ2 CP
µ1 µ1
T T
(a) (b)
Figure 1.6: Chemical potentials of two phases in contact. When µ1 = µ2 , the

two phases coexist. In figure (a) the chemical potentials cross and the transition is
first order. Dashed portions represent unstable states. In figure (b) the transition
is continuous and the two phases meet in the critical point CP.
1.5 Phase Transitions

So far we have considered only homogeneous systems, i.e., single phases. In
Nature there are, however, many systems that exist in different phases. Examples
are solid-liquid-gas, magnetic, ferroelectric, · · ·, systems. In this Section we will
discuss transitions between such phases.
When two phases, say 1 and 2 are in contact with each other, there is an inter-
face separating the phases. In equilibrium, the chemical potentials of both phases
in contact are equal,
µ1 (T, p1 , · · ·) = µ2 (T, p2 , · · ·) (1.70)
and the two phases coexist (See Figure 1.6).
Classification of Phase Transitions

As one of the parameters (e.g., the temperature) is varied in such a way that the
coexistence line is crossed, the system undergoes a phase transition between the
two phases. At the phase transition a derivative of the free energy (or some other
17
thermodynamic potential) is discontinuous. The free energy (thermodynamic po-

tential) becomes non-analytic in this point. The transition can be either first order
or continuous.
(a) First-order phase transitions. The conditions for first-order phase transi-
tions are that the chemical potentials are equal,
∆µ = µ1 (T, p, · · ·) − µ2 (T, p, · · ·) = 0, (1.71)
and that their derivatives are different,
∂∆µ ∂∆µ
6= 0 and 6= 0, (1.72)
∂T ∂p
so that the chemical potentials or the free energies cross at the transition. The
phase with the lower chemical potential is stable and the other phase is metastable.
An example of a first-order phase transition is the liquid-gas transition.
(b) Continuous phase transitions. A transition is continuous when the chem-
ical potentials of both phases are equal and when their first derivatives are also
equal,
∆µ = µ1 (T, p, · · ·) − µ2 (T, p, · · ·) = 0, (1.73)
∂∆µ ∂∆µ
=0 and = 0, (1.74)
∂T ∂p
so that the chemical potentials or the free energies have a common tangent at the
transition. There is no latent heat in this type of transition.
Instead of continuous transitions, some authors write about second, third, and
higher order transitions, depending on which derivative of the chemical poten-
tial diverges or becomes discontinuous. This classification is somewhat arbitrary,
therefore we will distinguish only between the continuous and first-order transi-
tions.
1.5.1 Order Parameter

As an example, let us consider a simple ferromagnetic material. At high temper-
atures, there is no magnetization and the system is rotationally invariant. At low
18
temperatures, when spontaneous magnetization occurs, the magnetization direc-

tion defines a preferred direction in space, destroying the rotational invariance.
We call this spontaneous symmetry breaking and it takes place at the critical tem-
perature (for ferromagnets it is the Curie point). Since a symmetry is either absent
or present, the two phases must be described by different functions of thermody-
namic variables, which cannot be continued analytically across the critical point.
Because of the reduction in symmetry, an additional parameter is needed to
describe the thermodynamics of the low-temperature phase. This parameter is
called the order parameter. It is usually an extensive thermodynamic variable. In
the case of ferromagnets, the order parameter is the thermal average of the (total)
magnetization, hM i. For the gas-liquid transition, the order parameter is either the
volume or the density difference between the two phases. Notice that the density
is not an extensive quantity! Notice also that there is no obvious symmetry change
in the gas-liquid transition.
The basic idea of phase transitions is that near the critical point, the order
parameter is the only relevant thermodynamic quantity.
Some examples of the order parameters and conjugate fields are given in the
following table:
Systems Order Parameter Conjugate Field
Gas-liquid VG − VL p − pc
Ferromagnets M H
Antiferromagnets Mstaggered Hstaggered
Superfluids hψi Not physical
(Condensate wave fn.)
Superconductors ∆ Not physical
(Gap parameter)
1.5.2 Critical Point

1. Liquid-gas transitions. In Figure 1.7(a) a typical phase diagram of a solid-
liquid-gas system in the P −T plane at constant V and N is shown. Here we would
like to concentrate on the liquid-gas transition. The liquid and the gas phases
coexist along a line where the transition between these two phases is first order.
19
P P
T>Tc T=Tc
pc Solid Liq. CP
CP
T<Tc
Gas
TP
Coexistence region Liquid
Gas
Tc T ρ
Figure 1.7: Phase diagram of a fluid. CP is the critical point; T P is the triple
point.
H T > Tc
H
T=Tc
M>0 T > Tc
0 M
M < 0 Tc T
Figure 1.8: Phase diagram of a ferromagnet.
20
If T is increased, the coexistence line terminates in the critical point, where the
transition between the two phases is continuous. Beyond the critical point, there
is no transition between the two phases, there is no singularity in any physical
quantity upon going from one phase to the other. It is also instructive to consider
the isotherms in the P − ρ plane at constant T and N , see Fig. 1.7(b). At low
temperature, there is a large difference between the gas and liquid densities, ρG
and ρL , but as the critical temperature is approached, this difference tends to zero.
The existence of a quantity that is non-zero below the critical temperature and zero
above it, is a common feature of critical points in a variety of different physical
systems. ρL − ρG is the order parameter for the liquid-gas critical point.
2. Ferromagnets. A typical phase diagram of a ferromagnet is shown in Fig.
1.8(a). The order parameter is the magnetization M , i.e., the total magnetic mo-
ment. At low temperature, M is ordered either ”up” or ”down”, depending on the
external field H. When H changes sign, the system undergoes a first-order phase
transition (bold line) in which the magnetization changes its direction discontin-
uously [Fig. 1.8(b)]. As T is increased, the critical point is reached, at which
M changes continuously upon variation of H but its derivative with respect to H,
i.e., the susceptibility diverges. At still higher T , there is no phase transition, the
spins are simply disordered, this is the paramagnetic region.
1.5.3 Critical Exponents

In this Section the behaviour of the thermodynamic quantities near the critical
point is discussed. We suppose that very near the critical point, any thermody-
namic quantity can be decomposed into a ”regular” part, which remains finite (not
necessarily continuous) and a ”singular part” that can be divergent or have diver-
gent derivatives. The singular part is assumed to be proportional to some power
of |T − TC | (or p − pc ).
Definition of the critical exponents. For clarity, we will define the critical ex-
ponents on a ferromagnet. However, their definition will be general and will be
valid for arbitrary system that undergoes a continuous phase transition. The most
important critical exponents are:
β = Order parameter exponent. When H = 0, the magnetization M in a
ferromagnet is a decreasing function of T and vanishes at TC (See Fig. 1.9).
For T very close to (but below) TC , M can be written as
M ∼ (TC − T )β . (1.75)
21
0.5
00 0.5
0.5 H
1
T 1.5
2 0
2.5
Figure 1.9: Order parameter vs. temperature and field.
β is the order parameter exponent which describes the behaviour of M near TC .

Here and later, ∼ means ”the singular part is proportional to”.
δ = Exponent of the critical isotherm. When T = TC , M vanishes for H = 0
but increases very rapidly if H increases. For small H, the singular part of M is
proportional to:
M ∼ H 1/δ , (1.76)
where δ > 1.
γ = Susceptibility exponent. As TC is approached either from above or from
below, the susceptibility diverges. The divergence is described with the exponent
γ:
χ ∼ |T − TC |−γ . (1.77)
Notice that the exponent γ is the same for T > TC and T < TC whereas the
proportionality constants are different.
α = Specific heat exponent. The specific heat at constant volume (for ferro-
magnetic systems one has to take H = 0) has a singularity at TC . This singularity
is described by the exponent α:
CV ∼ |T − TC |−α . (1.78)
As usually, the exponent α is the same for T > TC and T < TC whereas the
proportionality constants are different.
ν = Correlation length exponent. As already mentioned, the correlation length
ξ is a temperature- dependent quantity. As we shall see later, ξ diverges at the
22
critical point with the exponent ν:
ξ ∼ |T − TC |−ν . (1.79)
η = Correlation function exponent at the critical isotherm. At TC , ξ diverges

and the correlation function Γ decays as a power law of r,
Γ ∼ rd−2+η , (1.80)
where d is the dimensionality of the system.

Scaling. As the name suggests, the scaling has something to do with the
change of various quantities under a change of length scale. Important scaling
laws concerning the thermodynamic functions can be derived from the simple
(but strong) assumption that, near the critical point, the correlation length ξ is the
only characteristic length of the system, in terms of which all other lengths must
be measured. This is the ”scaling hypothesis.”
With the aid of the scaling hypothesis we are able to get relations between the
critical exponents. The most common scaling relations are:
γ = β(δ − 1)
γ = ν(2 − η)
(1.81)
α + 2β + γ = 2
νd = 2 − α
so that only two of the above exponents are independent. The last four equations
are called the scaling laws. We shall prove some of the scaling laws later.
Universality. Some quantities depend on the microscopic details of the system.
Examples of such quantities are:
- The critical temperature depends on the strength of interaction.
- The proportionality constants (e.g., for the temperature dependence of the
magnetization) also depend on the microscopic properties of the system.
On the other hand, the critical exponents and the functional form of the equa-
tion of state, of the correlation functions, etc. are independent of many details of
the system. They depend only on global features, such as:
- The spatial dimensionality,
- symmetry (isotropic, uniaxial, planar,...), ... .
Systems that have the same critical exponents and the same functional form
of the equation of state, etc, belong to the same universality class. For example,
23
the Ising model in d = 2 forms one universality class, whereas in d = 3 it forms

another universality class, the Ising models in d = 2 and d = 3 have different
critical exponents because they have different spatial dimensionality. However, on
a first glance very different physical systems often belong to the same universality
class. Example: binary alloys and some magnetic systems can be described with
mathematically similar models and they belong to the same universality class.
Roughening of some crystal surfaces at high temperature (but below the melting
temperature) and the d = 2 Coulomb gas are members of another universality
class.
24
Chapter 2
Classical Models
In this Chapter we will discuss models that can be described by classical statistical
mechanics. We will concentrate on the classical spin models which are used not
only to study magnetism but are valid also for other systems like binary alloys or
lattice gases.
In principle, all particles obey quantum statistical mechanics. Only when the
temperature is so high and the density is so low that the average separation be-
tween the particles is much larger than the thermal wavelength λ,
1/3 !1/2
V h2

λ, where λ = , (2.1)
N 2πmkB T
the quantum statistical mechanics reduces to classical statistical mechanics and we
can use a classical, Maxwell-Boltzmann distribution function (instead of Fermi–
Dirac or Bose–Einstein).
For magnetic materials, the situation is somewhat different. The Hamiltonian
of a magnetic system is a function of spin operators which can usually not be
directly approximated by classical vectors. Quantum models in d dimensions can
be mapped onto classical models in d0 = (d + 1) dimensions. The classical limit
of the Heisenberg model, however, can be constructed for large eigenvalues of
the spin operator by replacing the spin operators by three-dimensional classical
vectors.
There is another quantum-mechanical effect we must discuss. In quantum me-
chanics identical particles are indistinguishable from each other whereas in classi-
cal mechanics they are distinguishable. Therefore, the partition function of a sys-
tem of non-interacting, non-localized particles (ideal gas!) is not just the product
of single-particle partition functions, as one would expect from classical statistical
25
mechanics, but we must take into account that the particles are indistinguishable.
We must divide the total partition function by the number of permutations between
N identical particles, N !,
1
ZN = (Z1 )N , (2.2)
N!
where Z1 is the one-particle partition function. The factor N ! is a purely quantum
effect and could not be obtained from classical statistical mechanics.
In (insulating) magnetic systems we deal with localized particles (spins) and
permutations among particles are not possible. Therefore we must not divide the
partition function by (N !).
A Hamiltonian of a simple magnetic system is:
1X ~b i S
~b j − H
~
Xb
~ i.
H=− Ji,j S S (2.3)
2 i,j i
The first term is the interaction energy between the spins, Ji,j being the (ex-
change) interaction energy between the spins at the sites i and j. The sum is
over all bonds between the spins. S ~b is the spin operator; in general it is a vec-
tor. The order parameter is equal to the thermal and quantum-mechanical av-
~b m
erage of S, ~
~ = hhSii. The thermal average of the first term in Eq. (2.3) is
E(S, m)~ = hE(S, {S})i. (Here, S is the entropy and {S} is a spin configuration.
The thermal average has to be taken over all spin configurations.) E(S, m) ~ is the
internal energy of a magnetic system.
The second term in (2.3) is the interaction with external magnetic field H.~ The
thermal average of this term corresponds to P V in fluids. We see that hH(S, H)i ~
is a Legendre transformation of E(S, m) ~ and is thus equivalent to the enthalpy of
fluids.
2.1 Real Gases

In the ideal gas we neglected the collisions between the particles. In reality, the
atoms (molecules) have a finite diameter and when they come close together, they
interact. The Hamiltonian of interacting particles in a box is:
p~2i
" #
X X
H= + U (~ri ) + u(ri − rj ), (2.4)
i 2m i>j
26
u(ri − rj ) = u(r) is the interaction energy and U the potential of the box walls.
The particles are polarizable and this leads to an attractive van der Waals interac-
tion between the induced electric dipole moments. Such an interaction energy is
proportional to r−6 , where r is the interatomic distance. At short distances, how-
ever, the electrons repel because of the Pauli exclusion principle. Although the
repulsive potential energy usually decays exponentially with r, we will write it in
the form:
(
∞ r < 2r0
u(r) = (2.5)
−(2r0 /r)6 r > 2r0
(2r0 is the minimal distance to which the particles of ”hard-core” radius r0 can
approach.)
The partition function of N interacting identical particles is
0
r
Figure 2.1: Interatomic potential energy u(r) in real gases.
1 Z Z
3 3
Z Z
Z(T, V, N ) = · · · d p 1 · · · d p N · · · d3 r1 · · · d3 rN ×
h3N N !
e−β [ i (pi /2m+U )+ i>j u(ri,j )] .
P 2 P
(2.6)
After integrating over all momenta, the partition function becomes:
!3N/2
1 2πmkB T
Z(T, V, N ) = ZI (T, V, N ), (2.7)
N! h2
where ZI carries all the interactions between the particles. We calculate the par-
tition function in a mean-field approximation in which all the correlations are ne-
glected. The potential acting on one particle is calculated by assuming a uniform
27
spatial distribution of other particles,

"Z #N
V1
3 −βΦ
ZI = d r1 e (2.8)
0
V1 is the volume, available to one particle, V1 = V − (N/2)(4π/3)(2r0 )3 =

V − N b. Here we subtracted from V the volume occupied by other particles. N/2
is the number of pairs of particles. The mean-field potential Φ is:
Z ∞ 2N a
Φ=− (2r0 /r)6 (N/V )4πr2 dr = − (2.9)
2r0 V
The resulting partition function is
1 2πmkB T 3N/2 2
Z(T, V, N ) = ( 2
) (V − N b)N eβN a/V (2.10)
N! h
(a = (2π/3) (2r0 )3 , b = (2π/3)(2r0 )3 ), the Helmholtz free energy is:
 !3/2 
V − Nb 2πmkB T N 2a
F (T, V, N ) = −kB T ln Z = −N kB T ln  − ,
N h2 V
(2.11)
and the equation of state is
∂F N kB T N 2a
p(T, V, N ) = −( ) = − 2 . (2.12)
∂V T,N V − Nb V
The last equation can be written in a more familiar form:
N 2a
(p + )(V − N b) = N kB T. (2.13)
V2
This is the familiar van der Waals equation. We derived the van der Waals equa-
tion of state in a mean-field approximation, we neglected all the correlations be-
tween the molecules. Nevertheless, this equation qualitatively correctly describes
the liquid-gas transition and it also predicts the existence of the critical point.
However, as we neglected the correlations, which are essential in the vicinity of
the critical point, it does not predict the correct critical behaviour.
28
The critical point is located at the inflexion point of the isotherm, (∂p/∂V )T =
(∂ p/∂V 2 )T = 0, and is located at:
2
VC = 3N b
a
pC =
27b2
8a
TC = (2.14)
27bkB
Below the critical point, the liquid and the gas phases are separated by a first-
order phase transition. The equilibrium transition between the two phases takes
place when their chemical potentials and therefore the Gibbs free energies are
equal.
Some comments are in place here.
• Critical exponents. The van der Waals equation was derived in a mean-field
approximation, therefore all the exponents are mean-field like.
• Liquid → gas transition. The liquid phase is stable until the point A is
reached. In equilibrium, at this point, liquid starts to evaporate, the system
enters the coexistence region (mixture of gas and liquid) until the point B is
reached, which corresponds to the pure gas phase.
If equilibrium is not reached, the liquid phase persists beyond the point A
towards the point D. In this region, the liquid is metastable and can (at
least in principle) persist until the point D is reached. Beyond this point,
the compressibility is negative, the liquid state is unstable (not accessible).
• Gas → liquid transition. The situation is reversed. In equilibrium, the

gas phase is stable until the point B is reached. This is the onset of the
coexistence region which exists until the point A is reached. Metastable gas
phase exists between the points B and E and, again, the region E to D is
unstable.
• Convexity. As we have discussed in Section 2.3.3 already, the Helmholtz

free energy F of a stable system has to be a concave function of V . For
metastable states it is concave locally whereas for unstable states, it is con-
vex.
29
• Spinodals. The transition from a metastable (homogeneous) state toward

the equilibrium (mixed) state is connected with the nucleation of droplets of
the new phase. Droplets first nucleate and then grow until the equilibrium
(coexistence) is reached between the new and the old phase. Nucleation is
a dynamic process.
The ultimate stability limit of the metastable state is called the spinodal line.
Along this line the isothermal compressibility diverges, therefore the spin-
odal line is in a sense a line of critical points. In the van der Waals (mean-
field) systems the separation between the metastable and unstable states is
sharp. Beyond MFA, for systems with short-range correlations, however,
the transition between metastable and unstable states is often gradual.
2.2 Ising Model

The Ising model is the prototype model for all magnetic phase transitions and is
probably the most studied model of statistical physics. In this model, the spin
~b is replaced by a number, which represents the z− component of the
operator S
spin and is usually S = ±1 (”up” or ”down”). The order parameter m = hSi is
thus one-dimensional, it is a two-component scalar.
The Ising Hamiltonian is
1X X
H=− Jij Si Sj − H Si (2.15)
2 i,j i
Jij is the interaction energy between the spins i and j. Usually (not always) the
sum is only over nearest neighbouring (NN) pairs, then Jij = J. H is the external
magnetic field (in energy units). The case J > 0 corresponds to ferromagnetism
and J < 0 to antiferromagnetism. As we shall show later, the Ising model has a
phase transition at finite T if d > 1. In d = 1 the transition is at T = 0. Therefore
we say that d = 1 is the lower critical dimension of the Ising model.
In the low-temperature phase, the symmetry is broken, the spins are ordered,
the order parameter is m 6= 0, whereas in the high-temperature phase the sym-
metry is not broken, m vanishes, the spins are disordered. Therefore this model
is a prototype of all order-disorder transitions. Other examples of order-disorder
transitions, I would like to mention, are lattice gases and binary alloys. All these
systems are equivalent to each other, they all have the same critical behaviour
(which - of course - depends on d), for given d they all belong to the same univer-
sality class.
30
2.2.1 Lattice Gas

Imagine that a fluid (gas or liquid) system is divided into a regular lattice of cells
of volume roughly equal to the particle (atom, molecule) volume. We say that
the cell is occupied if a particle centre falls into this cell. Since the cell volume is
comparable to the particle volume, there can be not more than one particle per cell.
Notice that microscopic details (like positions of the particles) are irrelevant in the
critical region, therefore the lattice gas model is suitable for studying the critical
behaviour of real gases although the space is discretized. (It is also assumed that
the kinetic energy of the gas particles, which is neglected in the lattice gas model,
does not contribute a singular part to the partition function.)
Figure 2.2: A configuration of a lattice gas in d = 2. Solid circles represent atoms

and open circles empty sites. Lines show bonds between nearest–neighbouring
atoms.
A lattice gas is thus a collection of particles whose kinetic energy is neglected

and that are arranged into discrete cells. The cells are either occupied by one atom
[closed circles in Fig. 2.2] or empty [open circles]. The Hamiltonian is:
X X
Hlg = u(ri,j )pi pj − µ pi , (2.16)
hi,ji i
where u is the interaction potential,

( ∞ (ri,j = 0)
u(ri,j ) = −u (ri,j = NN distance) (2.17)
0 (otherwise)
and hi, ji is the sum over pairs. pi is the occupation number, it is 1 if the site is
occupied by an atom, and zero otherwise. The occupation is, like the Ising spin,
31
a double-valued function. Therefore we expect that the two models are related.
Indeed, if we set
1 + Si
pi = , (2.18)
2
the Hamiltonian 2.16 becomes:
uX µ uz X

Hlg = − S i Sj − + Si − H0 . (2.19)
4 hi,ji 2 4 i
Here, z is the number of NN neighbours (4 for square lattice, 6 for simple cubic
lattice), and H0 is a constant, independent of S. Thus, we see that Hlg is identical
to the Ising Hamiltonian if we set u/4 = J and (µ/2 + uz/4) = H. For u > 0,
the atoms attract and they ”condense” at low temperatures whereas for u < 0
the atoms repel. At 1/2 coverage (one half of the lattice sites is occupied) and
u < 0, the empty and occupied sites will alternate at low T , like spins in an
antiferromagnet.
The lattice gas model in d = 2 is used to investigate, e.g., hydrogen adsorbed
on metal surfaces. The substrate provides discrete lattice sites on which hydrogen
can be adsorbed. By varying the pressure, the coverage (amount of adsorbed
hydrogen) is varied. This corresponds to varying the field in magnetic systems.
There is an important difference between the lattice gas and the Ising models
of magnetism. The number of atoms is constant (provided no atoms evaporate
from or condense on the lattice), we say that the order parameter in the lattice gas
models is conserved whereas it is not conserved in magnetic systems. This has
important consequences for dynamics.
2.2.2 Binary Alloys

A binary alloy is composed of two metals, say A and B, arranged on a lattice. We
introduce the occupations:
(
1 if i is occupied by an A atom
pi =
0 if i is occupied by a B atom
(
1 if i is occupied by a B atom
qi =
0 if i is occupied by an A atom
pi + qi = 1. (2.20)
32
Depending on which atoms are nearest neighbours, there are three different NN
interaction energies: uAA , uBB , and uAB = uBA . The Hamiltonian is:
X X X
Hba = −uAA pi pj − uBB qi qj − uAB (pi qj + qi pj ), (2.21)
hi,ji hi,ji hi,ji
where each sum hi, ji runs over all NN pairs. If we write pi = (1 + Si )/2 and
qi = (1 − Si )/2 then Si = 1 if i is occupied by A and Si = −1 for a B site.
P
When there are 50% A atoms and 50% B atoms, i Si = 0, and the Hamiltonian
becomes an Ising Hamiltonian in the absence of a magnetic field:
X
Hba = −J Si Sj + H0 . (2.22)
hi,ji
H0 is a constant, and J = (uAA + uBB − 2uAB )/4. For binary alloys, the order
parameter is also conserved.
2.2.3 Ising model in d = 1

It is not difficult to solve the Ising model in the absence of H in d = 1 exactly.
The Hamiltonian of a linear chain of N spins with NN interactions is
X
H1d = −J Si Si+1 . (2.23)
i
We shall use periodic boundary conditions, that means that the spins will be
N 1 2
3
i+1
i
Figure 2.3: Lattice sites of an Ising model in one dimension with periodic bound-
ary conditions.
33
arranged on a ring, see Fig. 2.3. Then, the N −th and the first spins are NN and
the system is periodic. The partition function is:
P N
e−βH({S}) = eβJ Si Si+1
eβJSi Si+1 .
X X XY
Z= i = (2.24)
{S} {S} {S} i=1
(Don’t forget: {S} is the sum over all spin configurations – there are 2N spin
P
P
configurations – whereas i is the sum over all sites!)
The exponent in (2.24) can be written as:
eβJSi Si+1 = cosh βJSi Si+1 +sinh βJSi Si+1 = cosh βJ +Si Si+1 sinh βJ. (2.25)
The last equality holds because Si Si+1 can only be ±1 and cosh is an even function
whereas sinh is an odd function of the argument. The partition function is:
N
Z = (cosh βJ)N
XY
[1 + KSi Si+1 ] , (2.26)
{S} i=1
where K = tanh βJ. We work out the product and sort terms in powers of K:
Z = (cosh βJ)N
X
(1 + S1 S2 K)(1 + S2 S3 K) · · · (1 + SN S1 K)
{S}
= (cosh βJ)N
X
[1 + K(S1 S2 + S2 S3 + · · · + SN S1 )+
{S}
i
+K 2 (S1 S2 S2 S3 + · · ·) + · · · + K N (S1 S2 S2 · · · SN SN S1 ) . (2.27)
The terms, linear in K contain products of two different (neighbouring) spins, like
P
Si Si+1 . The sum over all spin configurations of this product vanish, {S} Si Si+1 =
0, because there are two configurations with parallel spins (Si Si+1 = +1) and two
with antiparallel spins (Si Si+1 = −1). Thus, the term linear in K vanishes after
summation over all spin configurations. For the same reason also the sum over all
spin configurations, which appear at the term proportional to K 2 , vanish. In order
for a term to be different from zero, all the spins in the product must appear twice
(then, {Si } Si2 = 2). This condition is fulfilled only in the last term, which – af-
P
ter summation over all spin configurations – gives 2N K N . Therefore the partition
function of the Ising model of a linear chain of N spins is:
h i
Z = (2 cosh βJ)N 1 + K N . (2.28)
34
In the thermodynamic limit, N → ∞, K N → 0, and the partition function be-

comes:
Z(H = 0) = (2 cosh βJ)N . (2.29)
The free energy,

0
F/N
Temperature
Figure 2.4: Free energy of the Ising model in d = 1.
F = −N kB T ln(2 cosh βJ), (2.30)

is regular at all finite T , therefore the Ising model in d = 1 has no phase transition
at any finite T .
To investigate the possibility of a phase transition at T = 0, we will calculate
the correlation function Γ and the correlation length ξ. We first replace J by Jj
(at the end we put Jj = J), so that
Z = 2N
Y
cosh βJj . (2.31)
j
The nearest-neighbour spin-spin correlation function is:

1X P
Γ(i, i + 1) = hSi Si+1 i = Si Si+1 eβ j Jj Sj Sj+1
Z {S}
1 ∂Z
= = tanh (βJi ). (2.32)
βZ ∂Ji
The spin correlation function for arbitrary distance k,
Γ(i, i + k) = hSi Si+k i
35
= hSi Si+1 Si+1 Si+2 Si+2 Si+k−1 Si+k−1 Si+k i

1 ∂ ∂ ∂ ∂
= ··· Z, (2.33)
Z ∂(βJi ) ∂(βJi+1 ) ∂(βJi+2 ) ∂(βJi+k−1 )
Γ(k) ≡ Γ(i, i + k) = [tanh (βJ)]k , (2.34)

is just the product of all the nearest-neighbour correlation functions between the
sites i and i + k, and is (if all J’s are equal) independent of the site i. The decay of
the spin correlation function with distance is shown in Fig. 2.5. The correlations
decay exponentially with distance and the correlation length is:
ξ 1
=− , (2.35)
a ln tanh βJ
where a is the separation between the NN spins (”lattice constant”). In the limit
T → 0, the correlation length diverges exponentially as T → 0:
ξ 1
≈ e2βJ . (2.36)
a 2
Equations 2.34 and 2.36 tell us that there is no long range order at any finite T and
0
0 2 4 6 8 10
k
Figure 2.5: Decay of the correlations in the d = 1 Ising model.
that T = 0 is the critical temperature of the d = 1 Ising model. Such a behaviour

is typical for systems in the lower critical dimension:
d > dC : TC 6= 0 (2.37)
d ≤ dC : TC = 0, (2.38)
36
therefore we conclude that d = 1 is the lower critical dimension of the Ising

model.
The zero-field isothermal susceptibility of the d = 1 Ising model is, according
to the fluctuation-dissipation theorem,
β X
χT (H = 0) = hSi Sj i. (2.39)
N i,j
In the thermodynamic limit, χT becomes:
β X N/2
X
χT (H = 0) = hSi Si+k i
N i k=−N/2
 
β X  N/2
X
= 2 hSi Si+k i − hSi Si i . (2.40)
N i k=0
In the thermodynamic limit (N → ∞), χT (H = 0) becomes

" #
2 1 2J/kB T
χT (H = 0) = β −1 = e . (2.41)
1 − tanh βJ kB T
The susceptibility has an essential singularity of the type exp(1/T ) as T → TC =
0 when J > 0. When J < 0, i.e., for antiferromagnetic NN interaction, the spin
correlation function alternates in sign and the susceptibility is finite at all T .
The internal energy is:
E = −∂ ln Z/∂β = −JN tanh βJ. (2.42)
Notice the similarity between the internal energy and the short-range correlation
function Γ(i, i + 1)! This is a general property of the internal energy.
The entropy is S = (E − F )/T ,
NJ
S(H = 0) = N kB ln [2 cosh βJ] − tanh βJ. (2.43)
T
In the limit T → 0, S → 0 and in the limit T → ∞, S → N kB ln 2, see Fig. 2.6.
The heat capacity CH = T (∂S/∂T )H is:
!2
βJ
CH = N kB . (2.44)
cosh βJ
The heat capacity is finite at all temperatures (no singularity at T = 0!), see Fig.
2.7.
37
N k ln(2)
0 1 2 3 4
kT/J
Figure 2.6: Temperature dependence of the entropy for the zero-field d = 1 Ising
model.
CH
0 1 2 3 4 5
kT/J
Figure 2.7: Temperature dependence of the heat capacity for the zero-field d = 1
Ising model.
2.2.4 Ising Model in d = 2

The Ising model for H = 0 in d = 2 dimensions has been solved exactly by
Onsager. The calculation is lengthy, therefore we will first show that the Ising
model in d = 2 has a phase transition at finite T and then quote the Onsager’s
result.
Consider a two-dimensional square lattice with all the spins on the outer bound-
ary ”up”. This simulates an arbitrarily weak external field and breaks the up-down
symmetry of the model. To show that spontaneous ordering exists, it is enough
to show that in the thermodynamic limit at a sufficiently small but finite tempera-
ture there are more spins ”up” than ”down”. Now we will estimate the fraction of
”down” spins from the average size of a spin-down domain.
38
Figure 2.8: A possible domain configuration of the Ising model on a square lattice.
With the above boundary conditions, all the domain walls must be closed
loops, see Fig. 2.8. In a wall the neighbouring spins are antiparallel, this costs
an energy 2J per wall segment and the probability, that a domain with its total
wall length (perimeter) b is thermally excited, is given by the Boltzmann factor
exp[−2βJb]. For large b, there are many possible domain wall configurations
with the same wall length b. For an estimate of the number m of possible wall
configurations, assume that on each site the wall can either be straight or make
a kink to either side. This makes a factor of 3 for each wall segment (except for
the last one where it is bound to close the loop). Besides, a wall can start on each
lattice site, this makes a factor of N . Therefore, m(b) is limited by:
m(b) ≤ 3b−1 N (2.45)
and the average number of different thermally excited domains with perimeter b
is:
≤ 3b−1 N e−2βJb . (2.46)
The number of ”down” spins in a domain with perimeter b is N− ≤ (b/4)2 and

the average concentration of all ”down” spins is:
hN− i X b2 b−1 −2βJb
≤ 3 e
N b=4,6,8,··· 16
1 ∂2
3b e−αb
X
=
48 ∂α2 b=4,6,8,···
39
1 ∂2 X
−2α x

= 9e
48 ∂α2 x=2,3,4,···
1 ∂2 q2
!
=
48 ∂α2 1 − q
q 2 (4 − 3q + q 2 )
= , (2.47)
12(1 − q)3
where α = 2βJ, x = 2b, q = 9 exp(−2α). For large, finite β, q is small and
the above ratio is < 1/2. This means that the order parameter hN+ − N− i/N is
greater than zero at low T and the Ising model in d = 2 (and of course also in
d > 2) has a phase transition at finite T .
The exact critical temperature of the (anisotropic) Ising model on a rectangular
lattice is given by the equation:
2Jx 2Jy

sinh sinh = 1,
kB TC kB TC
where Jx and Jy are the spin interaction energies in x and y directions, respec-
tively.
2.3 Potts and Related Models

In the previous Section we have discussed different classical two-state models
(which are all mapped on the Ising model). A large class of classical physical
systems cannot be described by the Ising model. As an example consider ad-
sorption of noble gases on surfaces, like krypton on graphite. Graphite provides
a two-dimensional hexagonal lattice of adsorption sites. Kr is large and once it
occupies one hexagon, it is unfavourable for another Kr atom to occupy any of its
NN hexagons, see Fig. 2.9. At full coverage, thus, Kr forms a regular lattice with
1/3 of hexagons covered by Kr. Since there are three entirely equivalent positions
(states!) for Kr (denoted by A, B, and C), the system is described by a three-state
model. The adsorbed Kr forms A,B,C domains, separated by domain walls. It is
clear from Fig. 2.9 that the A-B and the B-A walls are not equivalent, they have
different energies in general.
In the Potts model there are q states available to each site. The Potts Hamilto-
nian (with only NN interactions) is:
X
H = −J δσi ,σj . (2.48)
hi,ji
40
A B A
A B
C
Figure 2.9: Kr on graphite. There are three equivalent adsorption sites, but Kr
atoms are too big to occupy all of them – only one third of the sites (i.e., either
”A”, ”B”, or ”C” sites) can be occupied. We talk about ”A,” ”B,” or ”C” domains.
The domains are separated by domain walls. The A-B domain wall has a different
structure (and energy) than the B-A wall.
J is the interaction energy and δσi ,σj is the δ-function which is = 1 if the NN sites
i and j are in the same state and zero otherwise,
(
1 σi = σj with (i, j) ∈ N N
δσi ,σj = (2.49)
0 otherwise
σi = 1, 2, · · · , q is a number which describes the state of the site i. Thus, in the

Potts model, the interaction energy is equal to −J if the two NN sites are in the
same state and vanishes otherwise. For q = 2, the Potts model is equivalent to the
Ising model (S = ±1) because each site has two possible states in both cases. The
two states can be visualized as two points on a line (which is a one-dimensional
object). The mapping from the Potts to the Ising model is accomplished by writ-
ing:
1 + Si Sj
δSi Sj = . (2.50)
2
For q = 3 the three states of the Potts model can be visualized as corners of
a triangle (which is a two-dimensional object) whereas for q = 4, the four states
41
belong to corners of a tetrahedron, which is a 3-dimensional object. We conclude

that the order parameter of the q−state Potts model is a q−component scalar. In
d = 2, the Potts model has a continuous phase transition for q ≤ 4 and a first-order
transition for q > 4.
Order parameter. Let xA , xB , · · ·, xq be the fractions of cells (sites) that are
in the states A, B, · · ·, q, respectively. Of course, xA + xB + · · · + xq = 1. In
the disordered phase, all the fractions are equal whereas in the ordered phase one
state, say A, will have greater probability than the other, xA > (xB , xC , · · · , xq ).
We make a simplified approach, assuming that xB = xC = · · · = xq . The order
parameter m measures the asymmetry in occupancies,
xB + · · · + xq
m = xA −
q−1
1 + (q − 1)m 1−m
xA = , xB = x C = · · · = x q =
q q
It is easy to verify that for mA = 0: xA = xB = xC = · · · = xq and for m = 1:
xA = 1 and xB = xC = · · · = xq = 0. thus, mA is the order parameter which is
6= 0 in the ordered phase and vanishes in the disordered phase.
I would like to mention another generalization of the 2-state model, that is the
q−state (chiral) clock model. In this model, the states are imagined as hands of
a clock (i.e., vectors on a circle) that can take only discrete orientations, Θi =
(2π/q)σi (where σi = 0, 1, 2, · · · , q − 1). The q−state chiral clock Hamiltonian is

~ · ~ri,j
X
H = −J cos Θi − Θj + ∆ (2.51)
hi,ji
∆ is the chirality which breaks the symmetry of the model, it produces an asym-
metry in the interaction between the NN sites. The interaction energies between,
say, σi = 0, σj = 1 and σi = 1, σj = 0 are different. The two-state clock model is
identical to the Ising model and, if ∆ = 0, the 3-state clock model is identical to
the 3-state Potts model. For q > 3, the clock and the Potts model are not identical.
The two-dimensional clock and Potts models are widely used in surface physics.
The above mentioned Kr on graphite can be described with the 3-state chiral clock
model with ∆ 6= 0.
2.4 Two-dimensional xy Model

A very interesting situation appears when the spin is a continuous vector in a plane
(dimensionality of the order parameter is 2) and the spatial dimensionality is also
42
d = 2. As we have seen, the Ising model in d ≥ 2 does have a phase transition

at finite TC . In the xy model, however, spin excitations (spin waves) are easier
to excite thermally than for the Ising model, they are so strong that they even
destroy long-range order at any finite T (Mermin-Wagner theorem). On the other
hand, there has been evidence that some kind of a transition does take place at
a finite TC . What is the ordering and what is the nature of this transition? This
question was answered in the early seventies by Berezinskii and by Kosterlitz and
Thouless.
The Hamiltonian of the xy model is:
~i · S
~j = −J
X X
H = −J S cos(Φi − Φj ), (2.52)
hi,ji hi,ji
where the spins are classical vectors of unit length aligned in the plane. Φi is the
angle between the spin direction and an arbitrarily chosen axis and is a continuous
variable. For simplicity, let us consider the square lattice and let the sum be only
over nearest neighbours. Only slowly varying configurations (Φi − Φj 1) give
considerable contributions to the partition function, therefore we expand H:
J X Ja2 X ~
H − E0 = (Φi − Φj )2 = |∇Φ(i)|2
2 hi,ji 2 i
J Z
~
= d2 r|∇Φ(r)| 2
. (2.53)
2
Here a is the lattice constant and we transformed the last sum over all lattice sites
into an integral over the area. Φ(r) is now a scalar field.
Possible excitations above the (perfectly ordered) ground state are spin waves.
They are responsible for destroying long-range order at any finite T . We will not
consider spin waves here. Another possible excitation above the ground state is
a vortex, shown in Fig. 2.10. Any line integral along a closed path around the
centre of the vortex will give
(
Φ = qφ
I
~ r) · d~r = 2πq
∇Φ(~ ⇒ (2.54)
~
|∇Φ| = q/r
where q is the vorticity (q is integer; q = 1 for the vortex shown in Fig. 2.10) and
φ is the polar angle to ~r. From the equations (2.53) and (2.54) we find the energy
of this vortex: Z R
J q 2 R

E= 2πrdr = πJ ln q2. (2.55)
a 2 r a
43
Figure 2.10: Two examples of a vortex with vorticity q = 1.
R is the size of the system. Notice that the energy of the vortex increases logarith-
mically with the size of the system. In estimating the entropy associated with a
vortex we realize that a vortex can be put on each lattice site therefore the number
of configurations is about (R/a)2 and the entropy of one vortex is
R

S ≈ 2kB ln . (2.56)
a
The free energy, F = E − T S, is positive and large at low temperatures. That
means that it is not very easy to (thermally) excite single vortices. As temperature
increases, the entropy term causes F to decrease. F vanishes (changes sign) at a
critical temperature:
πJ
kB TC = . (2.57)
2
Above TC a large number of vortices is thermally excited.
However, this does not mean that there are no vortices at low T . The easiest
way to see this is to draw an analogy to a system of electrical charges in two
dimensions. If ~k is a unit vector perpendicular to the plane of the system, then
E~ = ∇(qφ)
~ × ~k is equal to the electric field produced by a charge of strength
2π0 q positioned at the centre of the vortex. The energy of an electric field is
0 Z 2 ~ 2 0 Z 2 ~
Eel = d r|E| = d r|∇(qφ) × ~k|2 , (2.58)
2 2
whereas the energy of a system of vortices is
JZ 2 ~ J
Evort = d r|∇Φ × ~k|2 = Eel . (2.59)
2 0
44
The energy of a system of vortices of strength qi is equivalent to the energy of a

system of electric charges! There is almost a complete analogy between a system
of vortices and the two-dimensional Coulomb gas. We have seen that, like the
electrostatic energy of a single charge in two dimensions, also the energy of a
vortex diverges logarithmically with the size of the system. However, a pair of
opposite charges, i.e., a dipole does have a finite electrostatic energy because at
large distances the fields of two opposite charges almost cancel. In analogy to the
energy of two charges ±q at ~r1 and ~r2 we write the total energy of a system of two
opposite vortices: !
2 |~r1 − ~r2 |
E2 vort = 2πJq ln . (2.60)
a
This self-energy of a vortex pair (of a dipole) is very small when the vortices are
close together. Thus, it is difficult to excite a single vortex but it is very easy
to excite a pair of opposite vortices. Whereas a single vortex influences the spin
configuration on the whole lattice, the effect of a pair of opposite vortices cancels
out on large distances from the pair so that the effect is limited to an area of the
order |~r1 − ~r2 |2 .
Since the size of the system does not appear in (2.60), but it does apper in the
expression for the entropy, we will have a finite density of bound vortex pairs. The
equilibrium density of vortex pairs will depend on the pair-pair (dipole-dipole in
the Coulomb gas case) interaction which we neglected here.
Now we have the following picture. At low but > 0 temperature (T < TC )
there will be a finite density of low-energy vortex pairs (”dipoles”) of zero total
vorticity. At high temperatures (T > TC ), on the other hand, there will be a large
concentration of single vortices. The critical temperature, we estimated above, is
thus associated with unbinding (dissociation) of vortex pairs.
Very interesting is the behaviour of the spin correlation function hS ~i · S
~j i.
We already mentioned that spin-wave excitations break the long-range order. If
we neglect the spin wave excitations (which have nothing to do with the phase
transition, they do not behave in a critical way at any finite temperature), then:
(
~i · S
~j i = > 0 for T < TC
lim hS (2.61)
|~
ri −~
rj |→∞ 0 for T > TC .
This follows from the fact that bound vortices don’t influence the spin orientations
at large distances whereas they do if they are free.
In a more realistic approach, one must take into account interaction between
the vortex pairs, which was neglected in (2.60). Surrounding vortices (dipoles!)
will screen the (electric) field and will thus reduce the self-energy of a vortex.
45
A sophisticated real-space renormalization group treatment of the two-dimensional

xy model [Kosterlitz 1974] shows that the correlation length ξ diverges above TC
as:
ξ ∝ exp (b/t1/2 ) b ≈ 1.5 (2.62)
where t = (T − TC )/TC . Below TC , however, the correlation length is always
infinite and the correlations decay algebraically (like in a critical point)! Therefore
we can say that we have a line of critical points extending from zero to TC . The
phase transition is characterized by a change from algebraic (power-law) decay of
correlations below TC to an exponential decay above TC .
Analogous to vortex unbinding transition in the xy model is the unbinding
of opposite charges of a dipole in the two-dimensional Coulomb gas. At low
temperatures, the Coulomb gas is composed of neutral dipoles, bound pairs, and
is not conducting. At high temperature it is composed of free + and − charges
which are conducting.
Another analogy to the xy model are the dislocation lines in a (three-dimensional)
crystal. At low temperatures, pairs of dislocation lines with opposite Burgers vec-
tors can exist. At TC , the dislocations dissociate. This is the basis of the dislo-
cation theory of melting. In this latter example, stiffness can play the role of the
order parameter, it is 6= 0 below the melting temperature and vanishes in the liquid
phase.
Let me stress again that the described behaviour of Coulomb gases or xy sys-
tems is specific only for two-dimensions!
2.5 Spin Glasses

The (vitreous) glasses are amorphous materials, their lattice is not crystalline but is
strongly disordered. Topological defects destroy the long-range order in the crys-
talline structure whereas the short-range order is still preserved to a high degree.
Example: topological defects on a square lattice would be triangles or pentagons.
Of course, the defects cause also a deformation of the squares.
Spin glasses are the magnetic (spin) analogue of the (vitreous) glasses. In
magnetic systems, the exchange interaction is often an oscillating function of dis-
tance between the spins. If such a system is diluted with non-magnetic ions, the
magnetic atoms are randomly positioned on the lattice and thus experience ran-
dom exchange interactions with other, neighbouring atoms. Example: Au1−x Fex ,
where x is small.
46
Another system, which behaves in a similar way, is a random mixture of a

ferromagnetic and an antiferromagnetic materials. The exchange interaction be-
tween nearest neighbours i and j is random, it depends on the type of the magnetic
atoms i and j.
A Hamiltonian which reflects the essential features of a spin glass is:
~i · S
~j − ~ ·S
~i .
X X
H=− J(Ri,j )S H (2.63)
i,j i
J(Ri,j ) is a random variable and the first sum includes also more distant pairs than
just nearest neighbours. Random exchange interactions cause that – even at low
temperatures – the spins are not ordered, the system has no long-range order in
the the ground state. The “usual” long-range order parameter
1 X
m= hSi i (2.64)
N i
vanishes for a spin glass. Yet, in the experiment, a broad maximum in the specific
heat and a rather sharp maximum in the zero-field susceptibility, connected with
hysteresis and remanescence, are seen. So, there must be a kind of a transition
as the temperature is varied, we shall call this a spin–glass transition and the low-
temperature state the spin-glass phase.
At high temperatures, the spins behave like in a normal paramagnet, they flip
around dynamically so that the thermodynamic average hSi i vanishes. There is no
long-range order in the system and also the local spontaneous magnetization, i.e.,
the (time) average of the spin at a site i vanishes, hSi i = 0 for each site. At low
temperatures, however, the spins freeze in a disordered configuration, the system
is in a state where hSi i =
6 0 although the average magnetization still vanishes,
m = 0. The spin-glass transition is thus a freezing transition.
Something happens to the dynamics of the spins, therefore we must consider
time-dependent spins, S(t). To distinguish between the two phases, we introduce
the (Edwards-Anderson) order parameter:
1 X
q = lim lim hSi (t0 )Si (t0 + t)i. (2.65)
t→∞ N →∞ N
i
q clearly vanishes in the paramagnetic phase and is > 0 in the spin-glass phase,
see Fig. 2.12. It measures the mean square local spontaneous magnetization.
Notice that freezing is a gradual process, which starts at Tf where the first spins
freeze and ends at T = 0 where all the spins are frozen.
47
Figure 2.11: Spins in the low-temperature, spin-glass phase are frozen in a disor-
dered configuration.
Since, in the spin-glass phase the spins are (at least partially) frozen in a dis-
ordered configuration, we can easily imagine that some spins are frustrated and
that the spins do not freeze in a unique way. Freezing is history dependent, the
configuration into which the spins freeze, depends on the conditions (e.g, on the
magnitude of the magnetic filed) close to the freezing temperature. If we, upon
cooling, apply an external field pointing ”up”, there will be more spins frozen in
the ”up” orientation and will stay in this orientation even if we switch the field
off in the low-temperature phase. This means that the systems can fall into a
metastable configuration from which it is not able to escape if the temperature is
low. The free energy, thus, has many, many local minima in the configuration
space, see Fig. 2.13. In the thermodynamic limit (N → ∞), some hills between
the local minima will become infinitely high and the system will be trapped in
such a minimum forever. In the time average, thus, the system will explore only a
part of the total configuration space. This means that the time average of a phys-
ical quantity is no longer equal to the ensemble average (where one explores the
whole configuration space). A system in the spin-glass phase is thus not ergodic
in the thermodynamic limit! Since it is ergodic in the paramagnetic phase, we say
that at the freezing transition the ergodicity breaks down.
48
q 1/2
0 Tf T
Figure 2.12: Temperature dependence of the order parameter q. Upon cooling,

the spins start to freeze at Tf ; below Tf , more and more spins are frozen until,
eventually, at T = 0 all the spins are frozen.
2.5.1 Some Physical Properties of Spin Glasses

The static susceptibility, Fig. 2.14, below the freezing temperature depends strongly
on the way the experiment is performed. χ(T ) is largest and roughly temeprature
independent in a ”field cooling” (FC) cycle, i.e., if the field H is applied above Tf
and then the sample is cooled at constant H below Tf . This susceptibility is re-
versible, it is independent of the history. In contrast, the “zero-field cooled” (ZFC)
susceptibility is obtained if the sample is cooled to T ≈ 0 at H = 0 and then the
sample heated in the field. The difference between χF C and χZF C confirms the
conjecture that the system can be frozen in different spin configurations (which,
of course, also have different susceptibilities).
The conjecture that freezing takes place at Tf is confirmed also by the ac
susceptibility measurements, shown in Fig 2.15 for Cu1−x Mnx (x = 0.9%) at dif-
ferent frequencies. The lower the frequency, the more pronounced (less rounded)
the cusp is. We see also that the position of the cusp shifts to lower temperatures
as the frequency decreases. So, the freezing temperature should be defined as the
maximum of χ(ν) in the limit as ν → 0.
49
Configurations
Configurations
Figure 2.13: Free energy landscape in the configuration space of a spin glass at
low temperatures. There is one global and many local minima.
Figure 2.14: Temperature dependence of the static susceptibility of Cu1−x Mnx for
x = 1.08 and 2.02%. (b) and (d): cooled to low temperature at H = 0 and then
heated at H = 5.9 Oe; (a) and (c): cooled and/or heated in H = 5.9 Oe. [From
Nagata et al., Phys. Rev. B 19, 1633 (1979).]
50
2.6 Hz
10.4 Hz
234 Hz
1.33 KHz
Figure 2.15: Temperature dependence of the frequency-dependent susceptibility

of Cu1−x Mnx (x = 0.9%. [From C.A.M. Mulder et al., Phys. Rev. B 23, 1384
(1982), ibid. 25, 515 (1982).]
51
52
Chapter 3
Quantum Models
In classical statistical mechanics we did not know the exact microscopic state of
the system (location in the phase space). We made use of the ergodic hypothesis
and replaced the time average of a physical quantity with an ensemble average,
i.e., an average over many equivalent systems.
In quantum statistical mechanics the situation is similar.
3.1 Quantum Statistics

In this Chapter we will first introduce the density matrix and then make an intro-
duction to the Bose condensation.
In quantum mechanics, the energy levels are discrete and the state of the sys-
tem is described by a wavefunction. Let us first assume that we know the dynami-
cal state (wavefunction) |ψi of the whole system and the complete set of stationary
states |ni which are the eigenstates of the full Hamiltonian of the compound sys-
tem,
H|ni = En |ni. (3.1)
(We assume that the states |ni are normalized.) n denotes the quantum states (e.g.,
momenta) of all particles in the system. En is the total energy of all particles of
the system in the compound state n. Any dynamical state of the system can be
expressed as a linear combination of these eigenstates,
X
|ψ(t)i = cn (t)|ni, (3.2)
n
53
The coefficients cn (t) define a point in the (Hilbert) space of wavefunctions |ni.
The (time-dependent) wavefunction |ψ(t)i which obeys the Schrödinger equation
d|ψi
ih̄ = H|ψi (3.3)
dt
is thus the quantum-mechanical analogue of the point in phase space of classical
statistical mechanics. The expectation value of a physical quantity A (which does
not necessarily commute with H) is:
c∗n cm hn|A|mi
X
hA(t)i = hψ(t)|A|ψ(t)i =
m,n
(3.4)
The last expression holds for any complete set of states |ni, also if they are not
the eigenstates of the full Hamiltonian.
3.1.1 The Density Matrix

As the number of particles in the system increases, the separation between the
energy levels decreases rapidly, and, in the thermodynamic limit, it becomes ex-
tremely high. On the other hand, the energy levels are never completely sharp.
They are broadened for many reasons, including the uncertainty principle. In a
macroscopic system (with extremely dense energy levels), the levels will thus al-
ways overlap. Consequently, a macroscopic system will never be in a strictly sta-
tionary quantum state, it will always be in a dynamical state, in a time-dependent
mixture of stationary states.
It is neither feasible nor possible to find a complete description of such a sys-
tem. Like in classical statistical mechanics, we first assume that the system is
ergodic, we replace the time average by an ensemble average over many quantum
systems at the same instance. We talk about probabilities of finding the system
in the states |ni. The dynamical state is no longer a unique quantum state but
is a statistical mixture of quantum states. We introduce the density matrix as the
ensemble average of c∗m cn :
ρ = |miρm,n hn|
Ω
1X ∗
ρm,n ≡ hm|ρ|ni = ci ci . (3.5)
Ω i=1 n m
The sum over i runs over Ω configurations of the system, |ψ i i, all being part of
the same ensemble, and cim = hm|ψ i i. We do an ensemble average because our
information on the system is not complete.
54
The density matrix is a statistical operator with the following important prop-
erties:
• The density matrix completely defines an ensemble of quantum systems

and it carries all the information that is available for a quantum statistical
ensemble.
• Its diagonal elements ρn,n tell the probability that the system is in the state
|ni whereas its non-diagonal elements ρm,n tell the probability of a transi-
tion from the state |ni to the state |mi. For stationary states, thus, ρ has to
be diagonal, it commutes with the Hamiltonian.
• The density matrix is normalized,
Tr(ρ) = 1. (3.6)
The trace runs over any complete set of states |ni and is independent of the
choice of the basis set.
• The (ensemble and quantum-mechanical) average value hAi of any physical

quantity A is equal to:
X
hAi = hψ|A|ψi = ρm,n hn|A|mi = Tr(ρA) = Tr(Aρ). (3.7)
n,m
The trace in quantum statistical mechanics plays the role of integration over
the whole available phase space in classical statistical mechanics.
• The dynamics of the system is completely described by:
∂ρ
ih̄ = [H, ρ]. (3.8)
∂t
This is the “Schrödinger equation” of the density matrix.
3.1.2 Ensembles in Quantum Statistical Mechanics

Now, we shall specify the density matrix and its relation to thermodynamic quan-
tities.
55
• Quantum Microcanonical ensemble.

The energy of systems belonging to a microcanonical ensemble is fixed,
therefore it is convenient to use the energy eigenstates as a basis. Then the
density matrix is diagonal. According to the postulate of equal weights,
all the states |ni with the energy En between E and E + ∆E are equally
probable and let there be ∆Ω such states. For a quantum system, ∆Ω is an
integer number. The probability for the system being in one of these states
is p = 1/∆Ω and the entropy is equal to:
S = −kB hln pi. (3.9)
In general, we must use the density matrix instead of simple probabilities p
and the entropy of a quantum system is:
S = −kB hln ρi = −kB Tr(ρ ln ρ) . (3.10)
Again, the average h i means ensemble and quantum average – both aver-
ages cannot be separated.
• Canonical ensemble. In the energy representation ρ is diagonal and we can
use the same arguments as in the case of classical canonical distributions.
The matrix elements are equal to:
1 −βEm
ρm,m = e , (3.11)
Z
where the normalization constant is (again) the partition function
e−βEm .
X
Z= (3.12)
m
In an arbitrary basis, the density matrix can be written more generally as
1 −βH
ρ= e with Z = Tr e−βH (3.13)
Z
(ρ and H are operators!). Thus, for a canonical ensemble, the average (ther-
mal and quantum- mechanical) of an operator A is:
1
hAi = Tr (ρA) = Tr (e−βH A) (3.14)
Z
56
In particular, the internal energy is
1 −βH ∂
E = hHi = Tr He =− ln Tr e−βH
Z ∂β
∂
E=− ln Z. (3.15)
∂β
The free energy is

F = −kB T ln Z. (3.16)
All the thermodynamic relations are the same as before (Tables I - III are
still valid), the only difference is that now we have to calculate the partition
function as the trace of the density matrix.
The entropy in a canonical ensemble, S, is the ensemble average over all
(energy) states of the microcanonical entropy S:
S = −hkB ln ρi = −kB Tr(ρ ln ρ). (3.17)
• Grand canonical ensemble. In the grand canonical ensemble the operator ρ

operates on a generalized Hilbert space which is the direct sum of all Hilbert
spaces with fixed number of particles. The density matrix is:
1 −β(H−µN )
ρ= e ,
Ξ
where the grand partition function is
Ξ(T, V, µ) = Tr e−β(H−µN ) .
Notice that now N is an operator in the generalized Hilbert space. As in the

classical case, Ξ is equal to the canonical partition function Z, weighted by
the fugacity z = exp(βµ) and summed over N . The ensemble average of
an operator A in the grand canonical ensemble is
1
hAi = Tr(A e−β(H−µN ) ).
Ξ
57
3.2 Bose Systems

3.2.1 Ideal Bose Gas
Now we shall consider the properties of an ideal gas of particles with integer spin
and mass m > 0. As usually for ideal gases, we neglect the interaction between
the particles. The Hamiltonian of the ideal Bose gas is:
X p2
H= np~ .
p
~
2m
np~ is the number of particles in the state with momentum p~. Because we are deal-
ing with Bosons, np~ can also be > 1. We want to study condensation where two
phases come into contact, the number of particles in one phase is not fixed, there-
fore we must work with the grand-canonical distribution. The grand-canonical
partition function of ideal Bose gas is:
P
(p2 /2m) np~ +βµN
e−β
X X
Ξ= p
~ , N= np~ .
{np~ } p
~
(Here we will disregard the (2S + 1) factor which comes from the spin degener-
acy.) In the grand partition function, each sum over np~ is over all non-negative
integers and, since the particles do not interact, we can split the partition function
into a product:
∞ ∞ nh in0 h in1 h in2 o
2 2
eβµ e−β(p1 /2m−µ) e−β(p2 /2m−µ)
X X
Ξ= ··· ··· = (3.18)
n0 =0 n1 =0
( )
Y Xh
−β(p2 /2m−µ) n
i Y 1
= e = .
p
~ n p
~
1 − ze−βp2 /2m
Here we have introduced the fugacity z = eβµ . The equation of state of an ideal
Bose gas is (see Table I):
P V = kB T ln Ξ(T, V, µ)
PV X h 2
i
=− ln 1 − e−β(p /2m−µ) (3.19)
kB T p
~
58
The thermal average of N is determined by the chemical potential µ (see Tables):

1 ∂ ln Ξ X 1
hN i = − = β(p2 /2m−µ) . (3.20)
β ∂µ p
~
e −1
We can write hN i =
P
p
~ np~ where
1
np~ = (3.21)
eβ(p2 /2m−µ) −1
is the Bose-Einstein distribution function, which tells the occupation probability
of a given (non-degenerate) state p~. For hN i to be finite and positive, the second
term in the denominator has to be < 1 for any p~, in particular for p = 0. This
means that µ must be < 0 and 0 < z < 1.
For large V we replace the sum by an integral:
X 4πV Z ∞ 2
→ 3 p dp (3.22)
p
~
h 0
and we get:
4πV Z ∞ 2 1
hN i = 3
p dp β(p2 /2m−µ)
h 0 e −1
!3/2
mkB T V
= V g3/2 (z) = g3/2 (z), (3.23)
2πh̄2 λ3
where
∞
1 Z∞ xν−1 X zl
gν (z) = dx x −1 = (3.24)
Γ(ν) 0 e z − 1 l=1 lν
and λ is the thermal wavelength,

!1/2
2πh̄2
λ= .
mkB T
The functions g3/2 (z) and g5/2 (z) are shown in Fig 3.1. For later use: g3/2 (1) ≈
2.612 and g5/2 (1) ≈ 1.342. The pressure is:
m kB T

5/3
P = 2 (kB T ) g5/2 (z) = 3 g5/2 (z). (3.25)
2πh̄ λ
59
2 3/2
g
1
5/2
0 0.5 1
z
Figure 3.1: The functions g3/2 (z) and g5/2 (z).
Now let us start increasing the pressure at fixed (high enough) T . This increases
the chemical potential, the fugacity, and the density of particles, hN i/V . How-
ever, we know that z must be < 1. What happens when z → 1? Can we further
increase the pressure or the density of particles?
To answer this question we must go back to Eq. (3.20). The term √ with p~ = 0
which was not included in (3.23) (The density of states n(E) has a E depen-
dence on E and vanishes as E → 0.) becomes singular (divergent) when µ → 0
and it has to be treated separately. This divergence has very important conse-
quences for a Bose gas, as we shall show now. Instead of Eq. (3.20) we must
write:
X 1 z V z
hN i = + = 3 g3/2 (z) + (3.26)
~6=0
p
eβ(p2 /2m−µ) −1 1−z λ 1−z
and the equation of state is:
PV 4πV Z ∞ 2 h 2
i
=− 3 p dp ln 1 − ze−βp /2m − ln (1 − z)
kB T h 0
P 1 1
= 3 g5/2 (z) − ln (1 − z) , (3.27)
kB T λ V
60
The last terms in Eqs. (3.26) and (3.27) come from the term p = 0 and correspond
to the particles in the lowest energy, in the ground state. In the thermodynamic
limit the last term in (3.27) vanishes whereas it remains finite in (3.26). We will
denote
hN i = hN 0 i + hN0 i, (3.28)
where hN 0 i is the average number of bosons in the continuum (excited, p > 0)

states,
V
hN 0 i = g3/2 (z) (3.29)
λ3
and hN0 i the average number of bosons in the ground state,
z
hN0 i = . (3.30)
1−z
The average density of bosons in the continuum states reaches its maximum value
when g3/2 (z) is maximal, that is for z → 1 (µ → 0),
!3/2
0
hNmax (T )i mkB T
= g3/2 (1) ∝ T 3/2 . (3.31)
V 2πh̄2
The fact that hN 0 i is limited and cannot exceed hNmax 0

i has its origin in the
quantum-mechanical and statistical nature of the system. The number of avail-
able states in a box with volume V is finite and is equal to λV3 . Each state, on the
average, accommodates np~ particles, where np~ is given by the Bose-Einstein dis-
tribution function, Eq. (3.21). If there are more than hN 0 i particles in the system,
they are pushed into the ground state.
3.2.2 Bose Einstein Condensation

We will now investigate the behaviour of ideal Bose gases, described by Eqs.
(3.26, 3.27). Let us consider a system of N bosons in a volume V .
0 0
At high enough temperature, hNmax i is so large that N < hNmax i and the
chemical potential is determined from
!3/2
0 mkB T
N = hN i = V g3/2 (z). (3.32)
2πh̄2
61
This means that z < 1 and µ < 0. The density of particles in the ground state is
hN0 i/V = z/(1 − z)/V → 0 as V → ∞. (3.33)
Almost all the particles are in the |~p| > 0 states and the ground state is macro-
scopically empty. The equation of state simplifies to
P g5/2 (z)
= (3.34)
kB T λ3
At very high T , g3/2 (z) and g5/2 (z) are 1 and we approximate [see the series
for gn , Eq. (3.24)]:
g5/2 (z) ≈ g3/2 ≈ z (3.35)
and the equation of state becomes:

!3/2 !3/2
P mkB T mkB T z
= g5/2 (z) ≈z = . (3.36)
kB T 2πh̄2 2πh̄2 λ3
z is determined by the density of particles:
!3/2 !3/2
N mkB T mkB T z
= g3/2 (z) ≈z = . (3.37)
V 2πh̄2 2πh̄2 λ3
The fugacity z = hN iλ3 /V vanishes as T −3/2 at high T and for fixed density.
This justifies the approximation (3.35). Eliminating z from the last two equations
yields the equation of state of the ideal Bose gas at high T :
P V =< N > kB T (3.38)
which is identical to the equation of state of a classical ideal gas (as it should be
at high T !).
Now we lower the temperature by keeping N and V fixed. Eq. (3.32) tells
us that g3/2 (z) must increase. This means that z and µ also increase. Eventually,
a temperature is reached, where z reaches its maximum value (z → 1, N =
0
hNmax i). This is the transition temperature TC and is determined by the condition
!3/2
N mkB TC
= g3/2 (1) (3.39)
V 2πh̄2
62
from which we find:
2πh̄2
2/3
N

TC = . (3.40)
kB m [g3/2 (1)]2/3 V
If we insert this expression into the equation of state, Eq. (3.34), we obtain the
pressure at which the transition takes place:
!3/2
5/2 mkB 5/2
PC = kB TC g5/2 (1) ∝ TC . (3.41)
2πh̄2
At TC , the ensemble average of the density of particles in the continuum states

reaches its maximum value whereas the ground state is still empty, hN0 i/V = 0.
0
Below TC , hNmax i further decreases and becomes < N (we keep N fixed).
The chemical potential further approaches zero (µ cannot be exactly = 0, because
that would mean that N = ∞) and z → 1 in such a way that the total number of
particles is kept constant. z is very close to 1, we can put gn (z) ≈ gn (1) and the
number of particles in the continuum states is
!3/2
0 mkB T
hN i = V g3/2 (1). (3.42)
2πh̄2
When hN 0 i < N , the ground state starts to fill, hN0 i increases. Below TC , thus,
only a part of the particles can be accommodated in the continuum states. The rest
must go into the ground state! The number of bosons in the ground state is:
!3/2
0 mkB T
hN0 i = N − hN i = N − V g3/2 (1)
2πh̄2
from which we get:
3/2
hN0 i T

=1− . (3.43)
N TC
hN0 i/N is finite, the state with p~ = 0 is occupied with a macroscopic number of
particles, the particles condense in the momentum space into the zero-momentum
63
P Critical line
0
V
Figure 3.2: Isotherms of the ideal Bose gas.
1
<No>/N <N’>/N
<No>/N
<N’>/N
0.5
0
0 0.5 1 1.5
T/Tc
Figure 3.3: Temperature dependence of the relative number of bosons in the

ground (hN0 i/N ) and excited (hN 0 i/N ) states.
state. This is the Bose-Einstein condensation. The fugacity is, from Eq. (3.30):
z = hN0 i/(hN0 i + 1). In the thermodynamic limit, when V and hN0 i → ∞,
z = 1 and µ = 0 below TC . The equation of state is:
!3/2
mkB T g5/2 (1) hN 0 i
P = g5/2 (1) kB T = kB T. (3.44)
2πh̄2 g3/2 (1) V
Only the particles in the continuum states (in the gas phase) contribute to pres-
sure. The particles in the ground state (condensate) are at rest, they cannot exert
any pressure. If the density is increased, the extra particles fall into the ground
state and the pressure does not increase. At fixed T < TC the density of particles
in the gas phase is constant and P is independent of V , see Fig. 3.2. The temper-
0
ature dependence of hN0 i is shown in Fig. 3.3. In the limit T → 0, hNmax i→0
and all the particles are in the ground state, N = hN0 i.
64
Vacuum
Gas phase
0 T
Figure 3.4: Phase diagram of an ideal Bose gas.
The corresponding phase diagram is shown in Fig. 3.4. The normal, gas phase
exists for T > TC , i.e., to the right of the transition line. On the line, given by
(3.41), condensation takes place and since the particles in the ground state don’t
contribute to pressure, the condensate lies on the transition line itself. Notice that
the ”critical points” in Bose systems form a line in the P vs. T or in the P vs. V
planes and are not in a single point like in van der Waals gases.
We now invert Eq. (3.44),
g5/2 (1) hN 0 i
V = kB T. (3.45)
g3/2 (1) P
As T → 0, also hN 0 i → 0 and therefore V → 0 at any finite pressure. The

condensed phase does not occupy any volume. This means that the ideal Bose gas
can be compressed to zero volume without any increase in pressure.
Some comments: The condensate (ground state) contributes neither to E, nor
to CV , P , or V ! At low T , CV vanishes like T 3/2 in contrast to the photon gas,
where CV ∝ T 3 . They differ because the density of states is different in both
cases. There are more excited states available for particles than for photons and
the specific heat is greater.
Of course, this, and infinite compressibility are the artifacts of the ideal gas
model. In reality, because of atomic repulsion, the volume of the condensed phase
does not vanish and the compressibility does not diverge. The equation of state of
a non-ideal Bose gas is shown in Fig. 3.5.
65
P
• Critical point
0
V
Figure 3.5: An isotherm of a non-ideal Bose gas.
Solid
P
He-I
λ− line
Critical point
He-II
Gas
0 T
Figure 3.6: Phase diagram of 4 He. He I is the normal, and He II the superfluid
phase.
ωk
∆
p0 p
Figure 3.7: Dispersion curve of elementary excitations in liquid 4 He. The excita-
tions around the minimum at p0 are called rotons.
66
3.3 Liquid 4He

A candidate for the Bose-Einstein condensation at low T is liquid 4 He. The phase
diagram is shown in Fig. 3.6. He I is the normal liquid and He II exhibits super-
fluid behaviour. From the fact that the line separating the solid and the fluid phases
is horizontal at T → 0, we conclude with the help of the Clausius-Clapeyron
equation,
dp Sfluid − Ssolid
=0= (3.46)
dT Vfluid − Vsolid
that liquid He II at T → 0 has no more entropy than solid helium, which is highly
ordered, of course. We conclude that the particles in the ground state (T → 0)
have already the highest possible degree of order although they are in the fluid
and not in the solid phase. We interpret the He II phase as a mixture of two fluids,
a normal fluid with final viscosity and a super-fluid (formed of particles in the
ground state) with zero viscosity. Characteristic for the superfluid state is also a
very large heat conductivity (as a consequence, He II does not boil, the atoms only
evaporate from the surface). What is the ground state of He II and what are the
excitations?
We shall first discuss possible excitations above the ground state of liquid 4 He
and only then the properties of the ground state. In the liquid, the atoms are
close together, they are strongly interacting (exchanging momenta). A state of
an atom with fixed momentum is not an eigenstate of the Hamiltonian, the atoms
are subject to a collective motion. The elementary excitations in the system are
collective excitations, quasiparticles, phonons (longitudinal because it is a liquid)
and rotons (see Fig. 3.7). At low momenta, the excitations are phonons with a
linear dispersion relation,
p = cp. (3.47)
At high momenta (short wavelengths), the interaction between the atoms causes
a minimum in the dispersion curve. Elementary excitations around this minimum
are (for historical reasons) called rotons. Their energy is written in the form:
(p − p0 )2
p = ∆ + , (3.48)
2m∗
where ∆ is the energy gap, ∆/kB = 8.6K, and m∗ = 1.0 × 10−27 kg is the
effective mass of rotons. p0 = 2.0 × 10−24 kg m/s. At low T , the density of these
quasiparticles is low and they don’t interact.
67
The grand canonical partition function for phonons (µ = 0) is, in analogy

with (3.18):
4πV Z ∞ 2
ln Ξph = − 3 p dp ln(1 − e−βcp )
h 0
4πV Z ∞ x3 dx
=
3(hβc)3 0 ex − 1
4π 5 V
= (kB T )3 (3.49)
45(hc)3
The grand canonical partition function for rotons (also µ = 0) is:
4πV Z ∞ 2 2 ∗
ln Ξrot = − 3 p dp ln(1 − e−β∆−β(p−p0 ) /2m )
h 0
2 ∗
At low T (large β), e−β∆−β(p−p0 ) /2m is small and we approximate ln(1 − x) ≈
−x. (This is equivalent to using the Maxwell-Boltzmann distribution.)
4πV −β∆ Z ∞ 2 −β(p−p0 )2 /2m∗
ln Ξrot = e p dp e
h3 0
4πV −β∆ Z ∞ 2 2 ∗
≈ 3
e p dp e−β(p−p0 ) /2m
h s −∞
4πV 2πm∗ 2 m∗ −β∆
!
= p0 + e (3.50)
h3 β β
From these equations, one can calculate the specific heat (and other thermody-
namic quantities) at low temperatures.
3.3.1 Superfluidity
Consider an isolated system (in which energy and momentum are conserved) com-
posed of a long cylindrical capillary filled with liquid 4 He at T = 0. Initially, let
the capillary tube rotate along the symmetry axis with the tangential velocity v0
and let the fluid be at rest. We would expect that - because of friction - energy
would be transferred from the tube to the fluid and that the fluid will gradually be
excited out of the ground state. Let us check under which conditions the energy
and momentum can be transferred to the fluid. Immediately after the tube started
to rotate, the tube had the kinetic energy M v02 /2 (M is the mass of the tube) and
the fluid was at rest and had the ground-state energy E0 . At a later time, we expect
the tube velocity to decrease by ∆v (because we consider an isolated system) and
68
that in the fluid an (at least one) elementary excitation with energy p and momen-
tum p~ is created. The conservation laws demand that the energy at a later time
is:
1 1
M (v0 − ∆v)2 + p + E0 = M v02 + E0
2 2
(3.51)
and that the momentum is:
M (v0 − ∆v) + pt = M v0 ⇒ M ∆v = pt (3.52)
where pt is the excitation momentum in the tangential direction. For ∆v 6= 0, the
first equation gives the inequality
1
p = M v0 ∆v − M (∆v)2 < M v0 ∆v. (3.53)
2
After eliminating M ∆v, the conservation laws tell us that
~v · p~ > p . (3.54)
Inspection of Fig. 3.7 shows that this is only possible if v is larger than a threshold
velocity vc . At small relative velocity between the tube and the fluid, the rotons
are not excited, the momentum is not transferred to the fluid, the fluid will stay at
rest. This means that there is no friction, no viscosity - this is the superfluid phase.
(As we will see later, roton excitations are not responsible for the excitation out
of the ground state. Vortices have higher energy, but lower /p ratio.)
3.3.2 Vortex Excitations

There are other collective excitations in liquid 4 He that have higher energies than
phonons or rotons, but which nonetheless have much lower energy to momentum
ratio than rotons. The creation of these excitations allows energy and momentum
to be transferred to the fluid at a much lower critical velocity than in the case of
rotons.
Let us now discuss such excitations of the condensate. The wave function of
an excited collective state is
Ψ = ψeiφ(r) , (3.55)
the (mass) current density is
~j = ih̄ Ψ∇Ψ
h i
~ ∗ − Ψ∗ ∇Ψ
~ ~ r).
= h̄ψ 2 ∇φ(~ (3.56)
2
69
On the other hand,

~j = N m~v (3.57)
V
where N/V = ρ = ψ 2 . From the two equations we get for the particle velocity
h̄ ~
~v = ∇φ(~r). (3.58)
m
The wavefunction must be single-valued and the phase φ is determined modulo
2π, therefore the closed-path integral of the momentum is
I I
m ~v · d~l = h̄ ~ ~l = nh,
∇φd (3.59)
where h is the Planck’s constant and n an arbitrary integer.

To understand the nature of these excited states it is best to consider liquid 4 He
within a small rotating bucket. If the bucket is slowly brought into rotation, the
liquid will behave in the following way. At small angular velocities of the bucket,
the liquid will remain at rest. If the angular velocity is increased beyond some
threshold angular velocity, the energy is transferred to the fluid, and the fluid will
make an abrupt transition to a state in which there is a vortex line running from
top to bottom, parallel to the axis of rotation. The vortex line looks like a small
whirlpool, the fluid flows around the line with the velocity that decreases with
distance from the line. The vorticity, defined as the line integral of the momentum
along any path encircling the vortex line, Eq. (3.59), is quantized. As the angular
velocity is increased, more and more quantized vortex lines are created. Once a
vortex line is created, the macroscopically occupied single-particle ground state
is no longer a zero-momentum eigenstate. It is a quantum state with a flow pat-
tern that includes the vortex lines. When there are many vortex lines, they will
distribute themselves in a pattern so that the average fluid velocity (averaged over
many vortices) is the same as in the case of rigid-body rotation.
To summarize, the quantization of vortices is related to the phase factor of the
condensate wave function, defined in Eq. (3.60). In the lowest-energy state, ψ and
φ do not depend on ~r, they are constant. The state in which φ varies with ~r is an
excited state which describes the fluid with one or several vortices.
3.3.3 Order Parameter

A superfluid can be considered as a mixture of two coexisting fluids, of a normal
component and of a superfluid component. At a first glance, one would choose
70
the density of particles in the ground (superfluid) state as the order parameter,
because it is 6= 0 in the low-temperature phase and vanishes at TC . However, in an
interacting system, the operator N0 doesn’t commute with H, the single-particle
momenta are not good quantum numbers. On the other hand we already know
that the ground state is a state with zero momentum p~ = 0 = −ih̄∇ ~ ⇒∇~ = 0.
This means that the wavefunction is a constant in the ground state. Therefore we
choose for the order parameter the statistical average of the ground-state wave
function which commutes with H and which can be written in the form:
1 h −β(H−µN ) i
hΨi = Tr Ψe = ψeiφ (3.60)
Ξ
(ψ is the amplitude and φ the phase of the wavefunction). The density of the
superfluid component is hN0 i/V = ψ 2 = constant. In the superfluid phase, a
macroscopically occupied single-particle quantum-mechanical ground state with
zero momentum extends through the whole volume available to the fluid. Each
atom is not at any particular place, it is simultaneously everywhere in the macro-
scopic volume! This is the meaning of the order parameter defined in (3.60).
Symmetry breaking. There is a strong parallel between hΨi and the sponta-
neous magnetization hM ~ i in ferromagnets,
~i= 1 ~ −βH
hM Tr M e , (3.61)
Z
~ is the (operator for the) total magnetic moment and
where M
~i · S
~j .
X
H = −J S
hi,ji
In the absence of an external magnetic field, thus, H is invariant under rotations,

M~ and −M ~ occur with equal probability and therefore hM~ i is always zero. But
we know that the ferromagnets do have spontaneous magnetization at low T . The
resolution lies in ”spontaneous symmetry breaking” - the ground state of a Hamil-
tonian does not possess the full symmetry of the Hamiltonian. The ground state
~ i points along a definite di-
wavefunction is not rotationally invariant, because hM
rection in space whereas H is rotationally invariant. In the case of ferromagnets,
thus, the transition corresponds to spontaneous breaking of the rotational symme-
try. One must therefore redefine hM ~ i in such a way that +M ~ and −M ~ are not
both included in the partition function with equal weights. This is most easily
done by imposing an infinitesimal external field H:
~ i = lim lim 1 ~ ~
hM TrM e−β(H−M ·H) . (3.62)
H→0 V →∞ Z
71
Returning to Bose systems, the Hamiltonian has a global gauge invariance,

the wave functions are invariant with respect to the phase shift,
Ψ(~r) → eiφ(~r) Ψ(~r) (3.63)
where φ is an arbitrary phase factor (real number). Bose condensation corresponds

to a spontaneous breaking of global gauge invariance. When calculating the order
parameter, we imagine, like in magnetic systems that the system is subjected to an
(infinitesimal) external field η which couples linearly to the order parameter. We
calculate the statistical average of Ψ(~r) in the thermodynamic limit:
1
hΨ(~r)i = lim lim TrΨ(~r)e−βE , (3.64)
η→0 V →∞ Ξ
where
Z
E = H − µN − d3 r[Ψ(~r)η(~r) + Ψ† (~r)η † (~r)]. (3.65)
The only essential difference with the ferromagnetic case is that the external field
here is fictitious, it cannot be realized experimentally.
72
Chapter 4
Methods of Statistical Mechanics
4.1 Mean-Field Theories

Very few models of statistical mechanics have been solved exactly; in most of
the cases one has to rely on approximative methods. Among them, the mean-field
approximation (MFA) is one of the most widely used. The advantage of the mean-
field theory is its simplicity and that it correctly predicts the qualitative features of
a system in most cases.
The essence of the mean-field theory is the assumption of statistical indepen-
dence of the local ordering (spins in the case of magnetic systems). The interac-
tion terms in the Hamiltonian are replaced by an effective, ”mean field” term. In
this way, all the information on correlations in the fluctuations is lost. Therefore,
the mean-field theory is usually inadequate in the critical regime. It usually gives
wrong critical exponents, in particular at low spatial dimensions when the number
of NN is small.
The MFA becomes exact in the limit as the number of interacting neighbours
z → ∞. This is the case when the range of interaction z → ∞ or if the spatial
dimension is high. In both cases, a site ”feels” contributions from many neigh-
bours and therefore the fluctuations average out or become even irrelevant (for
Ising systems, they become irrelevant in d ≥ 4). At low d, the MFA must be used
with caution. Not only that it predicts wrong critical exponents, its predictions are
even qualitatively wrong sometimes. For example, it predicts long-range order
and finite critical temperature for the d = 1 Ising model. In the following, we
shall introduce and formulate the MFA for magnetic systems. In the Appendix A
we summarize basic thermodynamic relations for magnetic systems. The results,
73
however, are valid also for other systems.
Two formulations of the mean-field approximation. We will now formulate

the mean-field approximation in two different ways.
• A: The “Weiss molecular field theory” according to its “inventor” Pierre
Weiss. This method is straightforward and easy to understand. It gives us
an expression for the order parameter but does not tell us anything about the
free energy or partition function.
• B: The second method will be the Bragg-Williams approximation which is
more sophisticated and is based on the free energy minimization.
We will develop the mean-field approximation on the Ising model in an exter-
nal magnetic field.
4.1.1 Weiss Molecular Field of an Ising System

Weiss formulated a theory of ferromagnetism in which he assumed that the effect
of the neighbouring spins on a given spin can be described by a fictitious “molec-
ular field” which is proportional to the average magnetization of the system.
We start with the Ising Hamiltonian in an external magnetic field:
X X
H = −J Si S j − H Si . (4.1)
hiji i
The local field acting on the spin at the site i is:

X
Hi = J Sj + H. (4.2)
j
It depends on the configuration of all the neighbouring spins around the site i.
If the lattice is such that each spin has many neighbours, then it is not a bad
approximation to replace the actual value of the neighbouring spins by the mean
value of all spins,
X
Sj → zm, (4.3)
j
where z is the number of nearest neighbours and m is the mean (average) magne-
tization per site, m = N1 j Sj . With (4.3), the local field becomes
P
Hi = zJm + H, (4.4)
74
1.0
m 0.5
0.0
0.0 0.5 1.0
T/Tc
Figure 4.1: Temperature dependence of the order parameter in the mean-field

approximation.
The actual local field acting on the site i was replaced by its mean value and is
independent of the site, Hi → HM F . In the mean-field approximation the original
Hamiltonian is replaced by a mean-field Hamiltonian
X
HM F = −HM F Si , (4.5)
i
which is equivalent to the Hamiltonian of a non-interacting spin system in an

external magnetic field and can be solved exactly:
m = tanh(βHM F ) = tanh[β(zJm + H)]. (4.6)
This is a self-consistent, transcendental equation for m. It has a non-trivial solu-

tion (m 6= 0) for βzJ > 1, i.e., at low temperatures. The temperature dependence
of the order parameter m for H = 0 is shown in Fig. 4.1. It vanishes at the critical
temperature TC = zJ.
4.1.2 Bragg-Williams Approximation

The Gibbs free energy is a function of T and H,
dG = −SdT − M dH. (4.7)
75
This equation tells us that for fixed T and H, δG = 0. In equilibrium, thus, G

is an extremum of a trial free energy G with respect to any variational parameter.
We will now search the trial free energy G.
We start with the density matrix and calculate the Gibbs free energy of an
Ising system. In the Bragg-Williams (mean-field) approach we approximate the
total density matrix ρ by a direct product of the density matrices for individual
spins,
e−β Hi .
Y Y
ρ≈ ρi ∝ (4.8)
b
i i
Physically this means that the spin fluctuations are considered to be uncorrelated,
the orientation of a spin is statistically independent of the orientation of any other
spin. There are no correlations in the spin orientation except those coming from
the long-range order. Under this assumption, the Ising Hamiltonian is not only
diagonal in the spin space (the state Si = +1 is uncoupled from the state Si =
−1; physically this means that the Ising system has no dynamics), but it is also
independent of all other spins in the system. Thus, the trial density matrices ρi
are diagonal and of the general form (remember, its trace must be = 1):
 1 (1 + m0i ) 0

 

ρi = 
 2 .

(4.9)
1 0 
0 (1 − m )

 
i
 
2
They depend on the variational parameter which is site-independent, m0i = m0 ,

because of translational invariance. With this density matrix the average spin
orientation is:
hSi i = Tr(Si ρi ) = m. (4.10)
Because of statistical independence of the spins in MFA, the short-range spin
correlations are:
hSi Sj i = hSi ihSj i = m2 . (4.11)
We can write the (trial) enthalpy E = hHi [where H is given by the Eq. (4.1)] as:
1 2
E = − zJN m0 − N Hm0 (4.12)
2
and the trial entropy as:
1 + m0 1 + m0 1 − m0 1 − m0
!
S = −kB Tr(ρ ln ρ) = −N kB ln + ln
2 2 2 2
76
-0.3
▲
-0.4 ❍
-0.5
-0.6 ●
G (T,H;m’) ●
-0.7
●
-0.8
-0.9
-1
●
-1.1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
m’
Figure 4.2: Trial Gibbs free energy G(T, H; m0 ) vs. m0 at H = 0.1 zJ. Solid
circles denote stable states, open circle a metastable state, and the solid triangle an
unstable state. The curves are labelled by T /TC . m and G(T, H) are determined
by the minimum of G.
(4.13)
With these expressions, the trial Gibbs free energy becomes:

1 2
G(H, T ; m0 ) = E − T S = E + kB T Tr(ρ ln ρ) = − zJN m0 − N Hm0 +
2
1 + m0 1 + m0 1 − m0 1 − m0
!
N kB T ln + ln (.4.14)
2 2 2 2
It depends on H, T , and on the variational parameter m0 . G for H = 0.1 (in

units of zJ) is shown in Fig. 4.2. To get the true equilibrium Gibbs free energy,
G(H, T ; m) has to be minimized with respect to the variational parameter m0 ,
G(T, H) = min
0
G(H, T ; m0 ). (4.15)
m
The minimization gives:

1 1+m
zJm + H = kB T ln , (4.16)
2 1−m
77
T < Tc Tc T > Tc
m
-1
0
H
Figure 4.3: External-field dependence of the Ising-model magnetization in the

mean-field approximation. TC is the critical temperature. Solid line: stable states,
dashed line: metastable states, dotted line: unstable states.
or:
m = tanh[β(zJm + H)]. (4.17)
This is the familiar “equation of state” of magnetic systems.

The field-dependence of m, Eq. (4.17) is shown in Fig. 4.3. At high tem-
peratures (T > TC ), the order parameter m behaves as in a paramagnet, this is
the paramagnetic phase. The spins are partially ordered only because of H. For
H = 0, the spins are disordered, therefore this phase is also called the disordered
phase. At low temperature (T < TC ), the spins are spontaneously ordered even
in the absence of H. This is the spontaneously ordered phase. The system has
long-range order for J > 0, all the spins are oriented predominantly in the same
direction, this is the ferromagnetic phase.
The Gibbs free energy is:
zN N kB T 1 − m2
G(T, H) = Jm2 + ln
2 2 4
zN 2
= Jm − N kB T ln [2 cosh β(zJm + H)] (4.18)
2
78
where m is related to H via the Eq. (4.17). The equation of state is

!
∂G(T, H)
M =− (M = N m). (4.19)
∂H T
The Helmholtz free energy is F (T, M ) = G(T, H) + HM :
1+m 1+m

FM F (T, M ) = N kB T ln
2 2
(4.20)
1−m 1−m 1

+ ln − N zJm2 .
2 2 2
The field dependence of the equilibrium Gibbs free energy G is shown in Fig.
4.4. The equation of state now reads
!
∂F (T, M )
H= . (4.21)
∂M T
The Helmholtz free energy F (T, M, N ) is shown in Fig. 4.5.

As before, the critical temperature, where the spontaneous ordering vanishes
in the absence of external magnetic field H as T is increased, is determined by the
condition βC zJ = 1,
kB TC = z|J|. (4.22)
In the past, many improvements to the simple MF theory presented here have
been proposed. With these improvements, the MF critical temperature came closer
to the exact critical temperature. None of these theories, however, was able to
bring an improvements to the critical exponents, they kept their mean-field values.
The real break-through brought the renormalization-group theory, which will be
introduced in Section 4.3.
4.1.3 Landau Theory of Continuous Phase Transitions

In the previous Section we have seen that the trial Gibbs free energy G(T, H; M 0 )
was an analytic function which had approximately a parabolic shape as a function
79
0.5
-0.5
0.75
G(T,H)
-1 1.0
T = 1.5 Tc
-1.5
-2
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1
H
Figure 4.4: Equilibrium Gibbs free energy. At low temperatures, when a first-
order transition at H = 0 takes place, part of the curve corresponds to metastable
states and part to unstable states.
of the variational parameter M 0 above the critical temperature and a double min-
imum below TC . Close to the critical point and for H = 0, the minimum of G is
close to m0 = 0 and we can expand the trial mean-field free energy G to a very
good approximation in a power series of m.
1
G(T, H; m) = G0 + rm2 + um4 + · · · (4.23)
2
At T > TC , r must be > 0 and below TC , it must be < 0, thus r changes sign and
in most of the cases it can be approximated by a linear temperature dependence,
r = r0 (T − TC )/TC .
The Landau theory is phenomenological and deals only with macroscopic
quantities. The theory does not deal correctly with the fluctuations and is a mean-
field theory. In fact, the Landau theory of continuous phase transitions is much
more than just simple expansion of G. It is based on the symmetry of the system.
Landau postulated that the free energy can be expanded as a power series in m
where only those terms that are compatible with the symmetry of the system are
allowed.
80
0.5
-0.5 0.75
F(T,m)
1.0
-1
T = 1.5 Tc
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1

m
Figure 4.5: Equilibrium Helmholtz free energy. At low temperatures, where the
transition is first order, F has two minima, corresponding to ±m. The region
between the two minima is the coexistence region when the system can be partially
in the +m, and partially in the −m state (domain structure!).
81
Formulation of the Landau Theory

Critical anomalies take place on long lenthscales. Therefore it is enough and in
order to change from discrete to continuous description:
mi → m(~r)
hi → h(~r).
In formulating the Landau theory let us consider a macroscopic system and

assume that there exists a (generalized; not necessarily magnetic) order parameter
density m(~r) averaged over microscopic, atomic distances a. Because of the aver-
age, m(~r) is now a continuous function of ~r (continuous in m and in ~r)! There can
be no fluctuations in m(~r) on the lengthscale smaller than a (a >∼
lattice spacing).
Correspondingly, m(~r) has no Fourier components with wavevector greater than
k = 2π/a.
In lattice models, the canonical partition function appropriate to the Gibbs free
energy of a spin system is:
e−βH
X
Y (T, H) = (4.24)
{Si }
(we have to sum over all spin configurations). Now, since m(r) is a continuous
function of the coordinate r, we must replace the sum by a functional integral over
all functions m(r). When doing this replacement one must not forget that m(r)
is equal to the average of a microscopic spin Si over a small volume. That means
that there can be several microscopic spin configurations which average into the
same m(r). Let there be Ω(m(r)) spin configurations that lead to the same m(r).
Then, the sum over all spin configurations is equal to the functional integral over
all functions m(r) where each function m(r) is weighted by Ω(m(r)):
X Z
→ Dm(r) Ω(m(r)). (4.25)
{Si }
The partition function is:

Z
Y (T, H) = Dm(r) Ω(m(r)) e−βH(H;m(r)) . (4.26)
Using the Boltzmann formula S = kB ln Ω, we can write Y in the form:

Z
Y (T, H) = Dm(r) e−βH(H;m(r)−T S)
82
Z
= Dm(r) e−βG(T,H;m(r)) , (4.27)
where we have introduced the Landau free energy:
G(T, H; m(r)) = H(H; m(r) − T S(m(r)). (4.28)
Notice the analogy between this equation and the trial Gibbs free energy ob-
tained previously with the Bragg-Williams method! The Landau free energy is in
fact a generalization of the trial Gibbs free energy.
Notice also that the entropy S(m(r)) is not equal to the usual microcanonical
entropy S (related to the number of configurations in an energy interval ∆E)
because we did not make any summation over the energies! S(m(r)) is related
to the number of microscopic configurations leading to the same function m(r).
Only after integration over all the states in a certain energy interval we would
get the usual thermodynamic entropy S. Therefore we can call S(m(r)) a partial
entropy.
We still have to work out the functional integral in order to get the (equilib-
rium) partition function Y (T, H)! From Y (T, H) the Gibbs free energy is:
Z
G(T, H) = −kB T ln Y (T, H) = −kB T ln Dm e−βG(T,H;m) . (4.29)
In the Landau approach we write down the Landau free energy in a most gen-
eral form as an expansion in terms of the order parameter and its derivatives by
taking into account all the symmetry-allowed terms up to a given order:
1 2
Z
G(T, H, m) = dV rm + um4 + · · · − Hm
2
1

+ g|∇m|2 + · · · . (4.30)
2
The highest-order coefficient in the expansion must always be positive, else the
system were unstable. For the Ising model, this form can also be obtained by
expanding G, Eq. (4.14), in powers of m. The last term appears if m is site-
dependent:
1
Si Sj → m(r) m(r + a) = m(r) m(r) + m(r) a∇m(r) + a2 ∇2 m(r) + · · · .
2
(4.31)
83
The first term is already included in the first line of (4.30), the second term van-
ishes in systems with center of inversion (symmetry argument!) and the last term,
after partial integration over the whole volume, gives the term proportional to
|∇m|2 . The prefactor g must also be positive always. A negative g would lead to
completely unphysical wild fluctuations in m on extremely short lengthscales.
Here we have assumed a one-component order parameter. In general, the order
parameter can have several components or be complex. The coefficients are in
principle temperature-dependent and depend also on the symmetry of the system.
For example, the coefficient of the cubic term vanishes if the system is invariant
under the reversal of m(r) when H = 0. Usually we assume that u is temperature
independent whereas the quadratic term has a linear temperature dependence,
r = r0 t, t = (T − TC )/TC (4.32)
In the Landau theory we assume that the fluctuations of the order parameter
are small, that the important values of m lie in a narrow range near its equilibrium
value that minimizes the Landau free energy. In fact, the maximal contribution
to the partition function comes from the term with minimal Landau free energy.
In the thermodynamic limit, the probability of any other configuration will vanish
rapidly. Therefore we make use of the saddle-point approximation, the functional
integral (4.29) is equal to the maximum value of the integrand, multiplied by a
constant. For constant field H, the functional form of m(r) that maximizes the
integrand (minimizes the Landau free energy), is a constant, determined by the
conditions
∂G ∂2G
=0 ≥0 (4.33)
∂m ∂m2
In the saddle-point approximation we immediately obtain G(T, H) from Eq. (4.29):
G(T, H) = G0 + minm G(T, H; m) = (4.34)
In equilibrium, the system is at the bottom of the Landau free energy G. This
determines m and G(T, H). From G(T, H) we get other thermodynamic quanti-
ties, as before.
The advantage of the Landau theory is that it allows one to study a macro-
scopic system without knowing all its microscopic details. It is based on the sym-
metry properties of the macroscopic system alone.
84
4.1.4 Critical Exponents in the Landau Theory

We will now calculate the critical exponents of the Landau model. To be more
specific, let us investigate the Landau model of magnetic phase transitions, when
1 02

4
G = G0 + rm + um0 − m0 H V (4.35)
2
In equilibrium, when ∂G/∂m0 = 0, we get the equation of state:
rm + 4um3 − H = 0 (4.36)
• Order-parameter exponent β ( H = 0). For H = 0, and r > 0, the equation
of state has a single solution, m = 0, there is no spontaneous magnetization
above TC , as expected. For t < 0, the equation of state has three roots when
H = 0. The root at m = 0 belongs to an unstable state, so we are left with
two real roots. The result is:
s
r
( r0 1/2 1
2
m =− ⇒ m=± t β= t<0 (4.37)
4u 4u 2
m=0 t>0
• Critical isotherm exponent δ (T = TC ). Now, r = 0, and we have 4um3 =

H, from which it follows
s
3 H
m= ⇒ δ=3 t=0 (4.38)
4u
• Susceptibility exponent γ. The isothermal susceptibility is defined as
!
∂m
χT = . (4.39)
∂H T
From the equation of state, Eq. (4.36), we get:

! !
∂m 2 ∂m
r + 12um −1=0 (4.40)
∂H T
∂H T
from which we find:

1
∝ t−1 ⇒γ=1 t>0
r0 t
(
χT = (4.41)
1
0
∝ |t|−1 ⇒γ=1 t<0
−2r t
85
• Specific-heat exponent α (H = 0). The specific heat at constant (zero) field

is
∂2G
!
cH = −T (4.42)
∂T 2 H
For t > 0, m = 0, and G = G0 = const. (that means that G0 does not

depend on the reduced temperature t), so cH is zero. For t < 0 (and H = 0),
m2 = −r/4u and
r0 2 2
G = G0 − t (4.43)
16u
It follows:
( 0 ⇒α=0 t>0
cH = 02 (4.44)
r TC
⇒α=0 t<0
8u
Comment 1: We see that the Landau theory gives mean-field (also called ”classi-
cal”) critical exponents.
Comment 2: At H 6= 0, there is no phase transition, all the quantities vary
smoothly, only the exponent δ is defined.
4.1.5 Short-Range Order and Fluctuations

Although the correlations and fluctuations are neglected in the MFA, we can still
get an estimate of the short-range correlations. To obtain the mean-field values of
the exponents ν and η, which describe the behaviour of the correlation function,
we have to include the gradient term in the Landau free energy. In the frame-
work of the mean-field theory, this leads to the Ornstein-Zernike approximation.
To investigate the correlations, we consider the response of m(~r) to a perturbing
field acting, say, at the origin, H(~r) = λδ(~r) via the “response function”, the
susceptibility:
Z
m(~r) = dd r0 χ(~r − ~r0 )H(~r0 ). (4.45)
The Fourier transform of the last convolution is simply:
m(~q) = χ(~q)H(~q). (4.46)
86
To the lowest order in m and ∇m, the Landau free energy is:
1 1
Z
d
G(H; m(r)) = G0 + d r g|∇m(~r)|2 + rm2 (~r)
2 2 i
4
+um (~r) − λm(~r)δ(~r) (4.47)
We first make Fourier transforms to the d−dimensional momentum space,

Z
m(~q) = dd r e−i~q·~r m(~r)
1 Z d i~q·~r
m(~r) = d q e m(~q) (4.48)
(2π)d
The Fourier transform of ∇m is:
1 Z d i~q·~r
∇m(~r) = d q ∇e m(~q)
(2π)d Z
1
= dd q i~qei~q·~r m(~q) (4.49)
(2π)d
The Fourier transform of a δ−function is a constant,
1 Z d i~q·~r
δ(~r) = d qe (4.50)
(2π)d
In the momentum space, the Landau free energy is:
1 2 2 1
Z
G(H; m(~q)) = G0 + dd q gq m (~q) + rm2 (~q)
2 2 i
4
+um (~q) − λm(~q) . (4.51)
Here, we approximated the Fourier transform of m4 (~r) by m4 (~q)!

In the mean-field approximation the modes with different momenta ~q are as-
sumed to be independent of each other. Therefore we can minimize G for each
individual mode separately,
∂G
= 0, (4.52)
∂m(~q)
and the equation of state in momentum space is:
(gq 2 + r)m(~q) + 4um3 (~q) = λ. (4.53)
87
For t > 0, when m vanishes in the absence of the field, we can neglect the cubic
term, and we get:
λ 1
m(~q) = , χ(~q) = . (4.54)
gq 2+r gq 2+r
After a (somewhat tedious) Fourier transform back to real space we find in the
limit of large ~r:
χ(~r) ∝ m(~r) ∝ r2−d exp(−r/ξ) ⇒ η = 0 (4.55)
where
g 1
r
ξ= ⇒ ν= . (4.56)
r0 t 2
Below TC , m 6= 0, we replace m(~r) by m+δm(~r), where m is the spontaneous
magnetization and δm the response to the infinitesimal perturbing field λ. We
expand in powers of δm and retain only linear-order terms. We find:
λ
δm(~q) = (4.57)
gq 2 − 2r
and the same exponents as for t > 0.

In the Ornstein-Zernike approximation (random-phase approximation), the
modes with different momenta are considered to be independent of one another.
The free energy is a sum (integral) of contributions from non-interacting normal
modes.
4.1.6 Validity of The Mean-field Theory

The mean-field theories neglect the fluctuations of the neighbouring spins, each
spin interacts only with the mean value of its neighbours. So, the mean-filed
theories can only be valid when fluctuations are irrelevant. We will now estimate
the relevance of fluctuations close to the critical point. We shall calculate the free
energy of the fluctuations and compare it with the free energy calculated in the
MFA.
To the order of magnitude, only those fluctuations can be thermally excited
whose free energy is of the order kB T (or less). The free energy of a typical
fluctuation will thus be proportional to kB T . Since the size of the fluctuation is of
88
the order of the correlation length ξ (this is how the correlation length is defined),
the free energy of a typical fluctuation per unit volume is proportional to
Gf luct ∝ ξ −d ∝ |t|νd . (4.58)
On the other hand, the singular part of the mean-field free energy is governed by
the specific-heat exponent α:
GM F ∝ |t|2−α (4.59)
(we obtain this expression after two integrations of the specific heat over temper-
ature). For the mean-field theory to be ”exact”, Gf luct GM F in the limit as
t → 0,
Gf luct
lim ∝ lim |t|νd−2+α = 0
|t|→0 GM F |t|→0
2−α
⇒ νd − 2 + α > 0 ⇒ d > . (4.60)
ν
For the Ising model (scalar order parameter) in the mean-field approximation α =
0 and ν = 1/2 and we get the condition
d > 4. (4.61)
The mean-field exponents are correct and the fluctuations are irrelevant only for
d > 4. d = 4 is called the upper critical dimension of the Ising model.
89
4.2 Monte Carlo Simulations

With the modern computers, different computer simulations became extremely
powerful techniques. One sets up in the computer the Hamiltonian describing a
physical system and tries to see how it behaves. However, an efficient and reliable
computer simulation is full of tricks and a lot of experience is needed to produce
reliable results.
Among different computer simulations, the Monte Carlo (MC) simulations are
probably the most widely used in statistical mechanics. MC simulations are used
to study classical as well as quantum systems, continuous spin systems, liquids
and other disordered systems, gauge theories, etc. Here we shall explain the prin-
ciples of the MC methods on the prototype example, the Ising model on a square
lattice.
4.2.1 Importance Sampling

In statistical mechanics we calculate the expectation values of thermodynamic
observables. For a spin system, the expectation value is defined as
Ae−βH
P
{s}
hAi = P −βH
. (4.62)
{s} e
For an Ising system, the sum is over 2N configurations. This number is extremely
large for all but very small systems. Therefore one must find a method which
estimates thermodynamic properties by sampling a small subset of representa-
tive configurations. One possible strategy would be to scan randomly the whole
configuration space (scan randomly over all spin configurations). The expecta-
tion value of A would then be obtained as a sum over the scanned configurations,
weighted by the appropriate Boltzmann factors. However, this random sampling
has a very serious drawback. Because of the Boltzmann factor which brings a
negligible weight to most of the configurations, very few configurations will con-
tribute to the expectation value of A and a very unreliable estimate will result.
This problem occurs because only an extremely restricted part of the configura-
tion space is occupied with a considerable probability in the thermodynamic limit.
Therefore it makes sense to restrict the sampling only to these states. this is the
importance sampling.
Importance sampling is realized if we generate a Markov chain of configura-
tions, i.e., a configuration is generated from the preceding configuration. The dis-
advantage of this Markov chain is that the successive configurations are strongly
90
correlated and therefore very long sequence of configurations is needed to obtain

reliable averages. The expectation value of A is now equal to:
nX
" #
1 max
hAi = lim An (4.63)

nmax →∞ nmax − nmin n=nmin
where An is the average of A over n successive configurations.
4.2.2 Metropolis Algorithm

In generating successive configurations, the transition probability must be normal-
ized, ergodic, and it must obey detailed balance. The Metropolis algorithm fulfills
the above conditions. A new configuration {Sf } is obtained from the previous
configuration {Si } by flipping one or more spins. In the Metropolis algorithm, the
probability that the system changes from {Si } to {Sf } is
(
e−β(Ef −Ei ) if Ef > Ei
P ({Si } → {Sf }) = (4.64)
1 if Ef ≤ Ei
where Ei and Ef are the energies of the previous and final states, respectively.
With this choice of transition probabilities, the system tends asymptotically to a
steady state in which the probability of a given configuration is e−βE({S}) .
The practical implementation of the Metropolis algorithm is very simple. First,
a spin is selected either at random or sequentially. The spin is flipped and the en-
ergies Ei and Ef are calculated. Finally, a pseudo-random number is used to
accept or reject the flip with the probability (4.64). The meaning of this condition
is straightforward. If the systems gains in energy by flipping the spin, the flip
is accepted. It is also accepted if it costs energy to flip the spin, but the energy
difference is small so that the Boltzmann factor is larger than the random number
(analogy to thermal excitations!). The flow chart of a MC program is shown in
Fig. 4.6.
4.2.3 Practical Details and Data Analysis

• Initial Configuration. In principle, the initial configuration should not in-
fluence the final configuration. However, the equilibrium may be reached
faster if a proper initial configuration is chosen. An example, showing the
approach to thermal equilibrium in a d = 3 Ising ferromagnet, is shown in
Fig. 4.7.
91

! " $#
%&
')(*+(
, -.
(
%/- 0/123 547698;:)<>=?#
%&
')(@+" A )B BDCE"
F9GIH;GKJ
n 0?l H n 0?m H
L E7
NM!- OPRQEEP
NM!
o < 1 < k J n n
%/-S?

UT'(VXW>! T
Y[Z
\]
S "^- _<
%/-?

@a`" \

W>! T Y*b
Zd
Y 1 Zd5e'c Zgfih ZgjZgf Y[Z
Figure 4.6: Flow diagram of a Monte-Carlo simulation with Metropolis algorithm.

During the first k1 steps the system still approaches to its equilibrium configura-
tion, therefore these configurations are not included in calculating the average
properties.
92
Figure 4.7: Relaxation of magnetization towards its equilibrium value in a Monte-

Carlo simulation of an Ising system on a 10×10×10 lattice with periodic boundary
conditions. Solid lines are “instantaneous” magnetizations and dashed lines their
equilibrium estimates. (From: K. Binder and Z. Rauch, Z. Phys. 219, 201 (1969).)
93
Figure 4.8: Scaled magnetization vs. scaled temperature of an Ising model on a

square lattice. (From: D.P. Landau, Phys. Rev. B 13, 2997 (1976).)
• Boundary Conditions. Different boundary conditions are used. In periodic

boundary conditions the system is periodically repeated in all directions to
mimic an infinite system. Alternatively, one may describe the effect of the
surrounding spins on an (inevitably) finite lattice by a mean field, acting on
the spins at the border. Open boundary conditions, when the finite lattice is
isolated and the spins at the surface have less neighbours than other spins,
have also been used.
• Finite-size effects. Lattices, considered in MC simulations, must have a
finite size (L) whereas we are usually interested in the macroscopic prop-
erties. The question is, how to take the limit of L to ∞. Usually, one
calculates an average for several different lattice sizes and then extrapolates
the average to infinite L. This extrapolation is simple away from the critical
point where the correlation length is small compared to L. The problem
becomes non-trivial in the critical region because ξ which would normally
diverge, cannot exceed L. As a consequence, singularities associated with
the critical point are rounded. If possible, scaling forms should be used in
extrapolating L → ∞ (see Fig. 4.8).
• Random-number generator. Generation of uncorrelated random numbers is
94
not a trivial problem. Computer languages have built-in ”pseudo-random-

number” generators. Not always these pseudo-random-number generators
are random. In particular, if random sampling is used, then it might happen
that some points on the lattice are visited more often than the other (which
- in the worst case - may never be visited). Pseudo-random-number gen-
erators use algorithms to calculate a sequence of (random) numbers. Care
must be taken to initialize random number generators with different ”seeds”
in order not to obtain the same sequence of random numbers each time.
• Statistical errors. To obtain reliable results, an average must be taken over n

much larger than the number over which the MC states are correlated. This
becomes more difficult near the critical point because of critical slowing
down.
• Free energy. Monte-Carlo simulation samples only a very small part of the
phase space whereas to obtain the partition function or the free energy, we
must sum over the whole configuration space. In MC simulations, there-
fore we calculate the free energy by integrating one of its derivatives, e.g.,
− M dH. This means that we need to calculate M (H) in a range of H.
R
• Conservation laws. Above, the spins were allowed to flip uncorrelated from
one another. This works fine for magnetic systems, where the order param-
eter is not conserved. For binary alloys, e.g., however, the number of atoms
is conserved and so is the order parameter. So, opposite spins have to flip
simultaneously (one ”up”→”down” and another - usually a neighbouring -
one ”down”→”up”). The conservation laws have important consequences
for the kinetics of fluctuations (slowing down).
95
4.3 The Renormalization Group Methods

Some typical results of MC simulations of a d = 2 Ising ferromagnet are shown
in Fig. 4.9. At high T (T = 2TC ), there is only short-range order, the spins
form small clusters. The correlation length (approximately equal to the linear size
of the largest cluster) is small. Close (but above) TC , somewhat larger patches
in which most of the spins are lined up in the same direction begin to develop.
When the system reaches TC , these patches expand to infinite size, but fluctuations
of smaller scale persist! At the critical temperature, spins form clusters at all
lengthscales, including one infinite-size cluster (which cannot be seen on a finite
system, but we know that the correlation length has to diverge at TC ). As a result,
all scales of length must be included in a theoretical description. Another very
important conclusion is reached if we substitute a block of 3 × 3 spins by a single
spin with the orientation prescribed by the majority rule (block spin ”up” if the
majority of the spins in the block is ”up”, and vice versa). In this way the length
scale of the lattice is changed by a factor b = 3 each time. This is a ”real-space
block-spin renormalization-group transformation”.
Examples of such transformations are shown in Figs. 4.10-4.12. Below TC ,
the system has long-range order but, because the system is close to TC , the fluc-
tuations form large clusters of oppositely ordered spins. However, under each
transformation the size of the unit cell increases, the fluctuations become smaller
and smaller and those fluctuations, which are smaller in size than the lattice spac-
ing, disappear. Under successive transformations, thus, the system becomes more
and more ordered, it effectively approaches zero temperature, see Fig. 4.13.
Above TC (Fig. 4.11), there is no long-range order, spins form random ”up”
and ”down” clusters. Under each transformation, the correlation length (expressed
in units of the new unit cell) decreases, the clusters become smaller and smaller
as if the temperature were higher, but they never disappear. At T = TC (Fig
4.12), the clusters are of all length scales, including one infinitely large cluster.
The correlation length is infinite (exceeds the size of the lattice used in the MC
simulation). Under successive transformations, it remains ”infinite”, there is no
change in the distribution of cluster sizes, the system is invariant under scale trans-
formations! The system remains at the critical temperature and the critical point
is a fixed point. Other (trivial) fixed points are T = 0 and T = ∞ because the
system away from TC tends to one of these points under successive scale transfor-
mations. Decrease of the correlation length under successive transformations is an
extremely useful property because it makes the fluctuations uncorrelated and we
can solve the system using an approximate (e.g., mean-field) theory to calculate
96
2TC
1.05TC
TC
Figure 4.9: Results of MC simulations of a d = 2 Ising ferromagnet at different T .

At high T (T = 2TC ; top), the correlation length is small. Close (but above) TC
(middle), somewhat larger clusters begin to develop. When the system reaches TC
(bottom), these clusters expand to infinite size, but fluctuations on smaller scale
persist! (From: K.G. Wilson, Sci. Am. 241, 140 (1979).))
97
Figure 4.10: Block-spin transformation is applied to a lattice repeatedly at T =

0.99TC . Each time the number of spins is reduced by a factor of 9, elucidating
the behaviour of the system at a larger scale. The correlation length decreases
under successive transformations. Clearly seen is the decrease of the correlation
length and a suppression of the fluctuations. Under each transformation the system
appears more ordered, the system flows under the renormalization transformations
towards zero temperature. (After K.G. Wilson, Sci. Am. 241, 140 (1979))
98
Figure 4.11: As Fig. 5.9 but for T = 1.22TC . Under each transformation, the
correlation length decreases, the system flows towards infinite temperature. (After
K.G. Wilson, Sci. Am. 241, 140 (1979))
99
Figure 4.12: As Fig. 5.9 but for T = TC : The correlation length is infinite
and remains infinite under successive transformations, the system remains at the
critical temperature. (After K.G. Wilson, Sci. Am. 241, 140 (1979))
100
T=0 Tc T =∞
Figure 4.13: Block-spin renormalization at H = 0. Upon successive decimations

of an initial configuration at T < TC the system becomes more and more ordered
as if the temperature were lower and lower. Eventually, the system becomes com-
pletely ordered, this is the T = 0 fixed point. Upon decimations of a state at
T > TC , the system becomes more and more disordered, the correlation length
decreases as if the temperature were higher, until the correlation length becomes
on the order of the lattice spacing. This is the infinite–temperature fixed point.
A system at TC stays, upon successive decimations, at TC . This is another fixed
point, the critical point.
the properties in that region and then transform back to the original lattice. This
is the basic idea of the renormalization-group methods and we shall exploit this
idea on some simple examples.
4.3.1 A Trivial Example: The d = 1 Ising Model

In the renormalization-group methods the temperature changes under successive
transformations, therefore it is convenient to work with the reduced Hamiltonian,
we divide H by kB T . The reduced Hamiltonian of an Ising Hamiltonian with NN
exchange interaction in an external field is
X X J H
H = −K Si Si+1 − h Si , K= h= . (4.65)
i i kB T kB T
Recursion relations. If we want to make several successive renormalization-

group (RG) transformations, the structure of H must not change after each trans-
formation. We are thus seeking for such a transformation from the old Hamil-
tonian H to the new Hamiltonian H0 that keeps the structure of the Hamiltonian
unchanged under successive transformations. In addition, the transformation must
be such that it does not change the partition function. That is necessary because
this ensures that the physical properties of the system are not changed. The parti-
tion function of the one-dimensional Ising model in an external field is written in
101

"!"!""!! "!"!""!! "!"!""!! &%$#&%$#&&%%$$## &%$#&%$#&&%%$$## &%$#&%$#&&%%$$## *)('*)('**))(('' *)('*)('**))(('' *)('*)('**))((''
1 2 3 4 5 6 7 8 9 10
Figure 4.14: Decimation on the d = 1 Ising model.
the form
P P
e−H = eK Si Si+1 +h Si
X X
Z(N, K, h) = i i
{S} {S}
K(S1 S2 +S2 S3 )+hS2 + h (S1 +S3 )
X
= e 2
{S}
h
×eK(S3 S4 +S4 S5 )+hS4 + 2 (S3 +S5 ) × · · · , (4.66)
We divide the lattice into the sublattices × and ◦, denote the spins on the × sub-
lattice by σi (Fig. 4.14),
h h
eσ2 [K(S1 +S3 )+h]+ 2 (S1 +S3 ) eσ4 [K(S3 +S5 )+h]+ 2 (S3 +S5 ) · · · ,
XX
Z=
{S} {σ}
(4.67)
and carry out a partial trace, i.e., we sum over 2N/2 configurations of the {σ} spins
sitting on the × sublattice:
Xh h h
i
Z(N, K, h) = e(K+ 2 )(S1 +S3 )+h + e(−K+ 2 )(S1 +S3 )−h × · · · . (4.68)
{S}
The entire partition function must not change and H must keep its structure, there-
fore we have the condition
N 0
Z(N, K, h) = eN g(K,h) Z( , K 0 , h0 ) = eN g e−H
X
(4.69)
2 {S}
where
H0 = −K 0 Si Si+2 − h0
X X
Si (4.70)
i∈odd i∈odd
and g is a constant. This will ensure that the physical properties will not change
under successive iterations. The two conditions (4.70) and (4.69) imply that
h h
e(K+ 2 )(Si +Si+2 )+h + e(−K+ 2 )(Si +Si+2 )−h
102
0 h0
= eK Si Si+2 + 2 (Si +Si+2 )+2g . (4.71)
The last equation has to hold for any Si , Si+2 = ±1:
0 0
e2K+2h + e−2K = eK +h +2g
0 0
e2K−2h + e−2K = eK −h +2g
0
e+h + e−h = e−K +2g , (4.72)
from which we get:

0 e2K+2h + e−2K e2K−2h + e−2K
e4K =
(eh + e−h )2
0 e2K+2h + e−2K
e2h =
e2K−2h + e−2K
2
e8g = e2K+2h + e−2K e2K−2h + e−2K eh + e−h (4.73)
or:
1 cosh(2K + h) cosh(2K − h)
K0 = ln
4 cosh2 h
1 cosh(2K + h)
h0 = h + ln
2 cosh(2K − h)
1 h i
g = ln 16 cosh(2K + h) cosh(2K − h) cosh2 h . (4.74)
8
Equations (4.73) or (4.74) are the recursion relations and determine the fixed
points and the flow diagram of the system.
In each iteration, the number of degrees of freedom is reduced by one half, the
new Hamiltonian H has only one half of the previous spins,
N 0 = N/b b = 2, (4.75)
and the lattice spacing is increased,
a0 = ba. (4.76)
Other lengths which are measured in units of the lattice spacing are reduced by a
factor b. For example, the new correlation length is
ξ 0 = ξ/b. (4.77)
The remaining spins on the decimated lattice interact with their new nearest neigh-
bours through the renormalized coupling constants K 0 and are subject to renor-
malized fields h0 .
103
Fixed points. At a fixed point the parameters (the coupling constants – in our
case K and h) do not change under successive decimations. This means that the
system at a fixed point stays at this point in the parameter space. With the new
variables
x = e−4K y = e−2h z = e−8g

0 ≤ (x, y, z) ≤ 1 (4.78)
the recursion relations become

(1 + y)2
x0 = x
(x + y)(1 + xy)
x+y
y0 = y
1 + xy
1
z0 = z 2 xy 2 (4.79)
(x + y)(1 + xy)(1 + y)2
The recursion relations for x and y do not depend on z. Physically, this means
that the singular behaviour of the free energy does not depend on a shift in the
energy scale. Therefore, we first investigate the flow diagram and fixed points
in the (x, y) plane, see Fig. 4.15. From the first equation it is easy to see that
x0 = x if x = 1, i.e., x= 1 is a fixed point of this system for arbitrary y . We
denote the fixed points by a ∗. There is a line of fixed points at x∗ = 1 for
0 ≤ y ∗ ≤ 1, corresponding to an infinite-temperature sink. Along this line, the
spins are completely disordered and ξ → 0. If we set x = 0, the second equation
reduces to y 0 = y 2 , which has two fixed points, one at y ∗ = 1 and one at y ∗ = 0.
The fixed point at (x∗ , y ∗ ) = (0, 1) (this means H = 0 and T = 0) corresponds to
the critical point where ξ → ∞. This is correct because we know that the d = 1
Ising model has TC = 0. This fixed point is unstable and any point in its vicinity
will flow under RG transformations away from it towards the fixed points on the
line at x∗ = 1. The critical behaviour of the d = 1 Ising model is governed by
the least stable fixed point at (x∗ , y ∗ ) = (0, 1). The point at (x∗ , y ∗ ) = (0, 0) is
another fixed point which corresponds to T = 0 and H = ∞. At this fixed point,
the spins are completely ordered.
4.3.2 RG Transformations on the d = 2 Ising Model

The recursion relations of the d = 1 Ising model could be formulated exactly.
Usually, however, farther-range interactions are introduced under successive dec-
104
y=e-2h
Figure 4.15: Schematic flow diagram of the d = 1 Ising model. (After J.M.
Yeomans, Statistical Mechanics of Phase Transitions, Oxford 1992.)
imations and approximations have to be made in order to get the closed-form

recursion relations.
As an example, let us apply the decimation on the d = 2 Ising model on
a square lattice in the absence of an external field. We start with NN (nearest
neighbours) exchange interactions and sum the partition function over all possible
configurations of the σ spins (these are the spins that will be decimated, they are
denoted by × in Fig. 4.16).
P P
K Si Sj
eK (Si+x +Si−x +Si+y +Si−y )σi
X XX
Z = e hiji = i
{S} {S} {σ}

XYh i
= eK(Si+x +Si−x +Si+y +Si−y )
+ e−K(Si+x +Si−x +Si+y+Si−y ) · · ·
{S} i
(4.80)
S±x and S±y are the ”right/left” and ”upper/lower” neighbours of σi , respectively.
The conditions are that under decimation Z must not change and that H should
have the same structure as the original Hamiltonian. However, we soon realize
that we cannot fulfill the above conditions with the renormalized form of the initial
105
(& (& (& $" $" $"
)' )' )' %# %# %#
(& (& (& $" $" $"
)' )' )' %# %# %#
(& (& (& $" $" $"
)' )' )' %# %# %#

! !

! !

! !
L K L

,* ,* ,*
-+ -+ -+

,* ,* ,*
-+ -+ -+
K K

,* ,* ,*
-+ -+ -+

L

K L

. .
/ /

. .
/ /

. .
/ /

Figure 4.16: Decimation on the d = 2 square lattice. The spins denoted by ×

are decimated. K is the NN interaction and √ L the NNN interaction. After deci-
mation, the lattice constant increases by b = 2, and previous NNN become NN
(nearest neighbours). Solid lines represent the original lattice and dashed lines the
decimated lattice.
Hamiltonian. After decimation we get a more complicated Hamiltonian:
H0 = − [K 0 (Si+x Si+y + Si+y Si−x + Si−x Si−y + Si−y Si+x )

X
+ L0 (Si+x Si−x + Si+y Si−y ) + Q0 Si+x Si+y Si−x Si−y + C 0 ] ,

(4.81)
where K 0 is the new NN interaction, L0 the new NNN (next nearest neighbours)
interaction, Q0 the four-spin interaction, and C 0 a constant. Only with the Hamil-
tonian of this form, it is possible to satisfy the condition Z 0 = Z for arbitrary
configurations of S±x,y . Although we started with only NN interactions, decima-
tion generated NNN interaction and even four-spin interactions. So, we have to
start from the beginning again with a more complicated Hamiltonian which in-
cludes all these interactions. In the following we shall neglect Q0 and assume that
L is small. In this case we obtain the following approximate recursion relations:
K 0 ≈ L + 2K 2
L0 ≈ K 2 (4.82)
106

K

K

2

K
L+ L

2K
K K
K2
Figure 4.17: The renormalized interaction after a decimation is L + 2K 2 .
In fact, these relations could be guessed immediately, see Fig. 4.17. The new near-
est neighbours are the previous NNN, therefore the term L in the first equation.
Besides, they were connected by two pairs of consecutive bonds, this brings 2K 2 .
In the second equation, the new next nearest neighbours (NNN) are connected by
two consecutive NN bonds on the old lattice - each of them contributing a factor
K.
The recursion relations 4.82 have two trivial fixed points, (K ∗ , L∗ ) = (0, 0)
and (K ∗ , L∗ ) = (∞, ∞) and one non-trivial fixed point at (1/3, 1/9). The first
one is the infinite-temperature, paramagnetic fixed point, the second one is the
zero-temperature, ferromagnetic fixed point. The non-trivial fixed point corre-
sponds to the critical point of the system.
4.3.3 General Properties of the RG Transformations

So far, we have seen that the renormalization-group transformations work by
changing the length scale and by reducing the number of the degrees of freedom.
Only at criticality the system remains unaltered under the change in scale and the
critical behaviour is determined by a fixed point of the transformation.
We start with an effective Hamiltonian, which is characterized by a set of
coupling constants and fields K, L, · · ·. Each set of these constants defines a point
in the ”Hamiltonian space,” the space of coupling constants. The renormalization-
group transformations work in this space,
H0 = R H. (4.83)
The renormalization-group operator R reduces the number of degrees of freedom
107
from N to N 0 . This defines the scale factor b:

N
bd = . (4.84)
N0
In the previous examples, the reduction was made in real space either by block-
spin transformation or by decimation; later we shall do the reduction also by inte-
grating out large momenta (short wavelengths) in momentum space.
The essential condition to be satisfied by any RG transformation is that the
partition function must not change,
ZN 0 (H0 ) = ZN (H). (4.85)
Therefore the total free energy does not change but the free energy per unit cell
(spin) increases as
f (H0 ) = bd f (H). (4.86)
All lengths, measured in units of the new lattice spacing, are reduced by the factor
b. Thus, the correlation length scales as
1
ξ0 = ξ. (4.87)
b
Fixed points. The iteration of R traces a trajectory in the Hamiltonian space,

this is called the ”Hamiltonian flow.” The trajectories end in fixed points H∗ , de-
fined by
H∗ = RH∗ ⇐⇒ H0 = H ≡ H∗ . (4.88)
At the fixed points, the system is invariant under subsequent scale changes and
thus the correlation length does not change
ξ0 = ξ ≡ ξ∗. (4.89)
Let us now take a closer look at the Hamiltonian flow near the fixed points.
We introduce a reduced Hamiltonian which we write in the general form
~
H = ~µ · φ (4.90)
and investigate the flow in the parameter space spanned by the generalized fields
~ which contains product of operators (like spins).
~µ, conjugate to the operator φ
108
• Example: In the case of the d = 2 Ising model on a square lattice, the

components of φ ~ include the terms linear in S (the conjugate filed is h),
quadratic in S (coming from interactions with NN, NNN neighbours), four-
spin terms, etc. The corresponding conjugate fields are the magnetic field h
and the coupling constants K, L, etc. In the beginning, many components
of ~µ are zero, but under successive transformations, more and more complex
interactions emerge and an increasing number of components of ~µ becomes
6= 0.
Under successive RG transformations, the Hamiltonian evolves and the system

moves through the parameter space:
~µ0 = R~µ (4.91)
until it reaches a fixed point when:
~ 0 = ~µ ≡ ~µ∗ .
µ (4.92)
Every particular type of criticality corresponds to a particular fixed point and de-
viations from the critical point are related to deviations from the corresponding
fixed point. To get insight into the critical properties we must investigate the flow
in the Hamiltonian space in the vicinity of the fixed points.
Near a fixed point we linearize the operator R. We write
~ = ~µ∗ + δ~µ
µ
~µ0 = µ
~ ∗ + δ~µ0 (4.93)
where the small deviations from ~µ∗ are related by

∂R
~ 0 = R(~µ∗ ) +
µ δ~µ = ~µ∗ + L(~µ∗ )δ~µ (4.94)
∂~µ µ~ =~µ∗
δ~µ0 = L(~µ∗ )δ~µ. (4.95)
The linear matrix L is evaluated at the fixed point and is thus a constant. The
eigenvalues λi and the eigenvectors ~νi of L,
L~νi = λi~νi , (4.96)
determine the critical properties of the Hamiltonian.
109
The eigenvalues λi are functions of b. Two successive transformations with

scale change b make a total scale change b2 , they are thus equivalent to a single
transformation with the scale change b2 .
λi (b)λi (b) = λi (b2 ) (4.97)
and the eigenvalues must be of the form
λi (b) = bxi . (4.98)
xi are the critical exponents and, as we shall see, they are closely related to the
usual critical exponents α, β, · · ·. So, if we know λi , we can obtain the critical
exponents xi from
ln λi
xi = . (4.99)
ln b
• Example 1: For the d = 1 Ising model, µ1 = x and µ2 = y. After lineariza-
tion of the recursion relations (4.73)around the fixed point at ~µ∗ = (0, 1) we
get:
δx0 = 4δx δy 0 = 2δy. (4.100)
Hence, L is diagonal in this representation. The eigenvalues and corre-

sponding eigenvectors are:
1

λ1 = 4 ν1 =
0
0

λ2 = 2 ν1 = . (4.101)
1
It follows that the two critical exponents are both positive:
x1 = 2 x2 = 1. (4.102)
• Example 2: For the d = 2 Ising model, µ1 = K and µ2 = L, and after

linearization Eqs. (4.82) become
δK 0 = 4K ∗ δK + δL
δL0 = 2K ∗ δK. (4.103)
110
For the non-trivial fixed point at (K ∗ , L∗ ) = ( 13 , 91 ), the RG transformation

matrix is:
4K ∗ 1 4/3 1

L= = (4.104)
2K ∗ 0 2/3 0
The eigenvalues of L are:
√ √
2 + 10 2− 10
λ1 = ≈ 1.72, λ2 = ≈ −0.387 (4.105)
3 3
and the eigenvectors:
√
1+ 10/2
2.58

~ν1 ∝ ≈
1 1
√
1 − 10/2 −0.58

~ν2 ∝ ≈ (4.106)
1 1
(Question: what does negative λ2 mean physically, how does the system
behave under successive RG transformations if λ is negative?) The corre-
sponding critical exponents are:
√
ln[(2 + 10)/3]
x1 = 2 ≈ 1.57
ln 2
√
ln |(2 − 10)/3|
x2 = 2 ≈ −2.74, (4.107)
ln 2
one is positive and the other one is negative.
Flow in the parameter space. We first express δ~µ near the fixed point at ~µ∗ in
terms of the eigenvectors ~νi of L. In case of the d = 1 Ising model this means that
δx0 = λ1 δx = 4δx
δy 0 = λ2 δy = 2δy. (4.108)
In general we write δ~µ as:
X
δ~µ = hi~νi . (4.109)
i
The coefficients hi are the linear scaling fields. They are the components of the
generalized field in the coordinate system of the eigenvectors. Under a renormal-
ization
δ~µ0 = Lδ~µ =
X X
hi L~νi = hi λi~νi , (4.110)
i i
111
where λi = bxi . On the other hand:
δ~µ0 = h0i~νi ,
X
(4.111)
i
therefore
h0i = bxi hi (b > 1). (4.112)
It follows that the flow of the Hamiltonian in the parameter space close to the fixed
point depends crucially on the exponents xi . For positive xi (λi > 1), the scaling
filed hi increases under repeated RG transformations and drives the system away
from the fixed point. In this case, hi is a relevant scaling field. If the exponent
xi is negative, the corresponding scaling field hi decreases under subsequent RG
transformations, thus driving the system towards the fixed point. Such fields are
called irrelevant scaling fields.
Thus the stability of the fixed point and the topology of the flow lines around
it depend on the number of relevant and irrelevant scaling fields associated with
it. To illustrate this, Fig. 4.18 shows the the trajectories in the parameter space
with two irrelevant and one relevant scaling field.
In general, any system which has initially the component of the relevant field
different from zero, will after repeated RG transformations eventually be driven
away from the fixed point (dotted lines in Fig. 4.18). Physically this corresponds
to the system moving away from criticality.
The points in which the relevant field vanishes, define the critical surface.
All the points on the critical surface flow into the fixed point. The points on the
critical surface represent systems with an infinite correlation length. Under RG
transformations, the systems stays on the critical surface (ξ remains infinite), but
it moves towards the fixed point.
• Example: The d = 1 Ising model has h1 = x and h2 = y. For the non-trivial

fixed point at x∗ = 0 and y ∗ = 1 the critical exponents are x1 = 2 > 0 and
x2 = 1 > 0, that means that both, temperature and the external magnetic
field, are relevant.
In the case of the d = 2 Ising model on the square lattice, x1 > 0 and
x2 < 0, as we have seen. It has one relevant scaling field (h1 ) and one irrel-
evant field (h2 ). The relevant scaling field is related to reduced temperature
whereas the irrelevant field is related to L, the coupling with NNN. Notice
that here we considered the Ising model without external field. If the field
112
h1
h2
Fixed
point
h3 Critical
surface
Figure 4.18: A fixed point with one relevant (h1 ) and two irrelevant (h2 and h3 )
scaling fields. All the points on the two-dimensional critical surface flow to the
fixed point (solid trajectories). The points slightly above (or below) the critical
surface first flow towards the fixed point but eventually they are driven away from
it as the relevant scaling field increases (dotted trajectories).
113
ν1
ν2
(a)
ν1
ν2
(b)
ν1
ν2
(c)
Figure 4.19: Topology of the RG flow near fixed points with: (a) two relevant
fields, (b) two irrelevant fields, and (c) one relevant and one irrelevant field.
is included, then we have two relevant scaling fields and one irrelevant (the
magnetic field is always relevant).
Figure 4.19 shows possible scenarios in the case of two scaling fields. Ferro-
magnets have two relevant fields, external field and temperature (or NN exchange
interaction) and the fixed point is of the type shown in Fig. 4.19 (a).
Universality. The critical behaviour is completely determined by the behaviour

of the system close to the fixed point, where the transformation matrix can be
linearized. The transformation matrix depends on the values of the scaling fields
at the critical point and not on the initial values of the irrelevant scaling fields.
Thus, all systems which flow close to the fixed point, independent of where in the
parameter space they start, will exhibit the same critical exponents determined by
the eigenvalues of the linear transformation matrix at that fixed point. The critical
114
exponents are universal and do not depend on the initial point in the parameter
space (like the coupling strength).
Critical exponent ν. To illustrate how the critical exponent ν is obtained from

xi , let us consider ferromagnetic systems. We know that there the critical point is
specified by H and T . These are the relevant fields for magnetic systems. There-
fore we identify h1 by the reduced temperature, h1 = t and h2 by the reduced
magnetic field, h2 = h. Under RG transformations ξ and t scale as
ξ
ξ0 = t0 = bxt t (4.113)
b
The two equations tell us that ξ xt t is constant, it is invariant under RG transfor-
mations. Therefore
ξ(t) ∝ t−1/xt (4.114)
Since, by definition of ν, ξ ∝ t−ν , we get
1
ν= (4.115)
xt
The last equation can be obtained also in a different way. Unless the system is
initially at the critical point, the correlation length will decrease under successive
block-spin RG transformations. The system will eventually reach a point where
the correlation length ξ 0 will become equal to the size of the renormalized unit cell,
so that the nearest-neighbour spins become uncorrelated. This point corresponds
to a certain temperature t0 which cannot depend on the initial temperature. Sup-
pose, the system is initially at the temperature t, and arrives at t0 after n iterations.
Then,
t0 = bnxt t and ξ 0 ≡ 1 = ξ/bn ⇒ ξ = bn (4.116)
From these equations we get Eq. 4.115 again.
• We return again to the d = 2 Ising model. Above, we considered the model
at h = 0. Thus, the exponent xt corresponds to the relevant exponent x1 =
1.57 from Eq. (4.107) and therefore ν = 1/x1 = 0.64. This is better than
the mean-field value of 1/2, but still far from the exact value ν = 1. The
discrepancy has its origin in the approximations we have made to obtain the
RG recursion relations.
115
Other critical exponents. The original and renormalized Hamiltonians H and

H0 correspond to positions in the parameter space, ~µ and ~µ0 . We apply relation
(4.86) to the singular part of the reduced free energy per spin:
gs (~µ0 ) = bd gs (~µ). (4.117)
Near a fixed point ~µ and ~µ0 can be expressed in terms of the scaling fields hi and
h0i which are related by (4.112). Hence eq. (4.117) becomes:
gs (h1 , h2 , ...) = b−d gs (bx1 h1 , bx2 h2 , ...), (4.118)
where gs is a homogeneous function of ~µ. This is the scaling form of the singular
part of the free energy. This equation holds for arbitrary b because we can repeat
transformations many times. In fact, b can be any real number > 1! Again we
identify h1 = t, h2 = h and assume that other scaling fields are irrelevant. Then,
eq. (4.118) becomes
gs (t, h) = b−d gs (bx1 t, bx2 h). (4.119)
Now, to come to the specific-heat critical exponent α, we remind that

∂ 2 gs
!
cH = ≡ gtt (h = 0) ∝ |t|−α . (4.120)
∂t2 h=0
Differentiating Eq. (4.119) twice with respect to temperature and putting the scal-
ing field h equal to zero gives:
dτ
gtt (t, h = 0) ∝ b−d gτ τ (τ, 0) = b−d+2x1 gτ τ , (4.121)
dt
where we have used τ = bx1 t. Initially, t is small whereas b > 1 so that the
product bx1 t increases with each iteration. We repeat the iterations until
bx1 |t| = 1. (4.122)
In this way, the arguments of gτ τ and therefore also the function gτ τ are constant,
all the temperature dependence of the specific heat is in the prefactor b−d+2x1 :
gtt (t, h = 0) ∝ |t|(d−2x1 )/x1 gtt (±1, 0) (4.123)
and
d
α=2− . (4.124)
x1
116
Similarly we find
d − x2
β=
x1
2x2 − d
γ=
x1
x2
δ= (4.125)
d − x2
We have expressed five critical exponents in terms of two independent expo-
nents x1 and x2 . That means that there must be relations between the exponents
α · · · ν. One can easily verify the scaling laws:
α + 2β + γ = 2
γ = β(δ − 1)
2 − α = dν. (4.126)
4.4 Momentum-Space Renormalization

Group
RG methods in momentum space are much more powerful and more widely used
than their counterparts in real space. The problem is only that they are much
more complicated than the above real-space prototype examples. We shall define
the RG transformation for the Landau model, that means that we shall now use
continuous instead of discrete spin variables. The main reason is that discrete
spins become essentially continuous after (block-spin) transformations.
We start with the free energy of the form:
1
Z
H(h, m(r)) = dd~r g|∇m(r)|2 − m(r)h
V 2
1 2

4
+ rm (r) + um (r) (4.127)
2
This is also called the Landau-Ginzburg-Wilson (LGW) Hamiltonian. The main
procedure of all the RG transformations is ”coarse graining”(i.e., transformation
to larger unit cells) followed by rescaling (to make the system look like the original
one). In real-space RG this was made either by introducing block spins or by
decimation. Now we shall make coarse graining in momentum space. Instead
117
of expanding the size of the unit cell in real space, we shall shrink the size of
the Brillouin zone in the momentum space. In the following the method of the
momentum-space renormalization group will be demonstrated on the Gaussian
model.
4.4.1 The Gaussian Model

The Hamiltonian of the Gaussian model in the momentum space is:
1 X
H= (r + gq 2 )|m(q)|2 − hm(0). (4.128)
2 |~q|<Λ
We assume a homogeneous magnetic field and we are only interested in the stabil-
ity of the m = 0 solution (r > 0), therefore we take u = 0; this will simplify the
summation over ~q). Λ is the cutoff momentum (Brillouin zone boundary). Taking
u = 0 means that the modes with different momenta become uncoupled.
The corresponding partition function is equal to the functional integral over al
functions m(r). In momentum space, the functional integral becomes:
Z Y Z ∞
Dm(r) → dm(q) (4.129)
0
|q|<Λ
and the partition function:

Y Z ∞
− 21
P
(r+gq 2 )|m(q)|2 −hm(0)
Y (t, h) = dm(q) e |q|<Λ . (4.130)
0
|q|<Λ
The RG transformation consists of ”integrating out” the wavevectors that are

outside the sphere of the radius Λ/b and then rescaling. One RG cycle consists of
three steps:
1. Integration. After ”integrating out” the momenta within a thin shell be-
tween the radii Λ/b and Λ, the partition function (which should not change)
becomes:
Z
0
Y = eΩ dm(q) e−H ,
Y
(4.131)
|q|<Λ/b
where the first term, exp(Ω), is a function of Λ, b, and coupling constants. It

comes from the integration over momenta in the shell between Λ/b and Λ.
118
The second term is the partition function of the states with momenta |q| <
Λ/b. Notice that this separation into two contributions was only possible
because we set u = 0, the partition function is then a product of independent
Gaussian integrals. The new Hamiltonian,
1
H0 (m(q)) = (r + gq 2 )|m(q)|2 − hm(0),
X
(4.132)
|q|<Λ/b
2
depends only on m(q) with |q| < Λ/b and has the same structure as the
original one.
2. Rescaling. The second step restores the old cutoff by blowing up the radius
of integration to the original value Λ in such a way that we change the
variable of integration to
q~0 = b~q. (4.133)
In this way the cutoff momentum is restored back to Λ. This step corre-
sponds to putting a0 = ba in real-space RG. The Hamiltonian now reads:
0 −d
X 1 q02 q0
H (m) = b (r + g 2 )|m( )|2 − hm(0). (4.134)
|q 0 |<Λ
2 b b
3. Normalization. The step 3 restores the standard normalization of the order

parameter. This step corresponds to putting the new spin equal to 1 after
block-spin or decimation in real-space RG. Renormalization is restored with
g 0 = g if we set:
s
1 q0
m0 (q 0 ) = m( ). (4.135)
bd+2 b
The transformed Hamiltonian and the partition function (which must not
change) now look like
X 1 2
H0 (m) = (r0 + gq 0 )|m0 (q 0 )|2 − h0 m0 (0). (4.136)
|q 0 |<Λ
2
and
Y Z 0
Y =e Ω
dm(q 0 ) e−βH . (4.137)
|q 0 |<Λ
119
In this way we integrated out large momenta and transformed H to its orig-
inal form. The recursion relations are:
r 0 = b2 r
h0 = b(d+2)/2 h. (4.138)
Since b ≈ 1 we can write b = 1 + dτ , thus,
bn ≈ 1 + n ln b ≡ 1 + ndτ. (4.139)
We can say that τ counts the number of RG cycles. Then, r and h obey the
following differential equations:
dr
= 2r
dτ
dh d+2
= h (4.140)
dτ 2
with the solution
r = r0 ext τ xt = 2
d+2
h = h0 exh τ xh = . (4.141)
2
Since q is continuous (for macroscopic systems), the RG transformations was

carried out in infinitesimal steps (b ≈ 1) and the problem was reduced to solving
a set of differential equations.
The Gaussian model has two trivial fixed points at h∗ = 0, one at r∗ = ∞
(very high T , m = 0) and one at r∗ = −∞ (very low T , where m → ∞ because
we set u = 0!), and one nontrivial fixed point (the critical point) at h∗ = r∗ = 0.
At the nontrivial fixed point the critical exponents are both positive and r and h are
both relevant scaling fields. From xi , we shall calculate other critical exponents
in the Section 4.4.3.
As we shall see in the next Section, for d < 4 the Gaussian model is unsta-
ble with respect to the parameter u coming from the four-spin interaction in the
Hamiltonian.
120
4.4.2 The Landau-Wilson Model

In the Gaussian model, we have neglected the fourth and higher-order terms in the
Landau Hamiltonian. The next step is to include also the m4 term. This will lead
to non-MFA exponents, as we shall see. Here we shall bring only an outline of
the RG procedure and discuss the results. We start with the LGW Hamiltonian in
real space,
1 1
Z
d
H(m) = d ~r g|∇m(~r)|2 + rm2 (~r) + um4 (~r) (4.142)
2 2
where u > 0. We first integrate out an infinitesimally thin shell of the thickness
δq in momentum space by taking b ≈ 1. Then,
δq/Λ = δΛ/Λ = δ ln Λ = b − 1 ≈ ln b (4.143)
and we write
m(~r) = m̄(~r) + δm(~r) (4.144)
where δm(~r) contains the Fourier components to be integrated out and m̄ the rest.
After some approximations we get the following renormalization-group (differen-
tial) equations for r and u [Huang, 1987]:
dr(τ ) r

= 2r + 12Λd−2 u 1 − 2
dτ Λ
du(τ )
= (4 − d)u − 36Λd−4 u2 . (4.145)
dτ
To analyse these equations, it is convenient to introduce the dimensionless
coupling constants:
r u
x= , y= (4.146)
Λ2 Λ4−d
then, the RG equations become:
dx
= 2x + 12y(1 − x)
dτ
dy
= y − 36y 2 (4.147)
dτ
121
where
≡4−d (4.148)
We shall assume that is small. The system has two fixed points at finite x and y,
fixed by dx/dτ = dy/dτ = 0:
)
x∗ = 0
Gaussian fixed point
y∗ = 0
)
x∗ = −/6
“Nontrivial” fixed point. (4.149)
y ∗ = /36
The Gaussian fixed point is at x = y = 0 whereas the “nontrivial” (also called

the “Wilson–Fisher”) fixed point lies in the upper (y > 0) half of the (x, y) plane
only if > 0, d < 4. In the lower half of the (x, y) plane, u is negative and the
system is unstable. The nontrivial fixed point approaches the Gaussian fixed point
as → 0.
In the neighbourhood of the fixed points we linearize the RG equations
x = x∗ + δx, y = y ∗ + δy (4.150)
and get the following linearized recursion relations:

d
δx = (2 − 12y ∗ )δx + 12(1 − x∗ )δy
dτ
d
δy = ( − 72y ∗ )δy (4.151)
dτ
which are written in the general form as:
! !
d δx δx
=L (4.152)
dτ δy δy
At the Gaussian fixed point the linear matrix L is:
2 12

L = (4.153)
0
The matrix L has the following eigenvalues and eigenvectors:
1

λ1 = 2 ~ν1 =
0
122
1

λ2 = ~ν2 ∝ (4.154)
−(2 − )/12
The corresponding scaling fields are h1 and h2 . Under the (infinitesimal) RG
transformations the scaling fields change as dhi /dτ = λi hi ,
hi (τ ) = hi0 eλi τ = hi0 bλi (4.155)
Thus, the critical exponents are equal to the eigenvalues λi : xt = λ1 = 2 and
x 2 = λ2 =
In the neighbourhood of the nontrivial fixed point the linear transformation
matrix L is:
2(1 − 6y ∗ ) 12(1 − x∗ ) 2(1 − /6) 12(1 + /6)

L= = (4.156)
0 − 72y ∗ 0 −
It has the following eigenvalues and eigenvectors:

1

λ1 = 2 − /3 ν1 =
0
6+
− 1+/3

λ2 = − ν2 = (4.157)
1
Now we have the following picture. For d < 4 ( > 0) the nontrivial fixed
point lies in the upper half of the (x, y) plane whereas the Gaussian fixed point
is (as before) at x = y = 0, see Fig. 4.20. For the nontrivial fixed point, h2 is
irrelevant (it does not affect the critical exponents) and h1 = r ∝ t the relevant
scaling field whereas for the Gaussian fixed point, both fields are relevant. Thus,
the line ν2 is the critical line. For d > 4 ( < 0)the nontrivial fixed point lies in the
lower half of the (x, y) plane, i.e., in the unphysical region (u < 0), see Fig. 4.21,
and the critical behaviour is governed by the Gaussian fixed point. d = 4 is the
special case when both fixed points coincide. The linearized recursion relations
give x2 = 0, h2 is marginal, and x2 cannot tell us the direction of RG flow. In
this case we must include higher–order terms which were neglected when we
linearized the transformation matrix. These terms tell us that the scaling field h2
is irrelevant for u > 0 at d = 4.
4.4.3 Critical Exponents

We have seen that the critical exponents of the nontrivial fixed point are
xt = 2 − /3
123
A BC
y ∝u
x∝ r
Unphysical region
Figure 4.20: Fixed points and flow diagram of the Landau-Ginzburg-Wilson

model for d < 4. The Gaussian fixed point at (x∗ , y ∗ ) = (0, 0) has the mean-
field exponents. The non-trivial fixed point at (x∗ , y ∗ ) = (−/6, /36) is unstable
against the relevant perturbation δx ∼ t and stable against the irrelevant perturba-
tion δy. The system at A has t < 0 and it flows towards the low-temperature fixed
point, r becomes increasingly negative. At the point C the system has t > 0 and
it flows to the high-temperature fixed point, r becomes increasingly positive. The
point B is on the critical line and the system flows towards the non-trivial fixed
point. (After K. Huang, Statistical Mechanics)
124
y∝ u
x∝ r
Unphysical region
Figure 4.21: Fixed points and flow diagram of the Landau-Ginzburg-Wilson

model for d > 4. The nontrivial fixed point is in the unphysical region. To
stabilize the system in the lower half of the (x, y) plane, one needs an m6 term.
(After K. Huang, Statistical Mechanics)
xu = −
xh = 1 + d/2. (4.158)
In fact, we calculated xh only for the Gaussian fixed point in the Eq. (4.138).
However, the term with m4 does not affect the value of xh so that xh of the non-
trivial fixed point is the same. With these values of xi we get from Eqs. (4.115,
4.124, 4.125) to the first order in for d < 4:

α =
6
1
β = −
2 6

γ = 1+
6
δ = 3+
1
ν = +
2 12
η = 0. (4.159)
125
For d ≥ 4 ( ≤ 0) we only have the Gaussian fixed point in the physical region
(u > 0) with the exponents
xt = 2
xu =
xh = 1 + d/2. (4.160)
With Eqs. (4.115, 4.124, 4.125) we find:

d
α = 2−
2
1
(β = − )
2 4
γ = 1
(δ = 3 + )
1
ν =
2
η = 0. (4.161)
The above expressions for β and δ are wrong, because we have neglected the
4
m term in considering the Gaussian fixed point. Below TC , r is negative and for
m to be finite, the term um4 with u > 0 must be included in the Hamiltonian.
Although u is an irrelevant field, it helps to stabilize m at a finite value and in this
way influences the critical behaviour. By treating the um4 term in the Gaussian
approximation (independent modes) we get the usual mean-field values for the
exponents β and η:
1
β =
2
δ = 3 (4.162)
for all d ≥ 4. These results are correct and also agree with the results of the
Landau (MFA) model in Sec. 4.1.4.
The Gaussian model gives the same critical exponents as the mean-field theory
except for the specific-heat exponent α. Therefore the Gaussian fixed point is
identified as the mean-field fixed point. In the Gaussian model the specific heat
126
diverges (α > 0) for d < 4. This divergence is caused by the long-wavelength

fluctuations which were neglected in the calculation of the free energy in the MF
approximation.
In d = 3 ( = 1) we find α ≈ 0.17 and ν ≈ 0.58. The results of this paragraph
are based on an expansion of d around d = 4 and are valid for small > 0 or for
d ≥ 4. Nevertheless, we put = 1 when d = 3. That means that is not small and
higher-order terms in must be included. This is then the −expansion, which
gives good values for the critical exponents.
127
128
Further Reading
[1] K. Binder (ed.), Monte-Carlo methods in statistical physics, Springer-Verlag

(New York, 1986).
[2] J.J. Binney, N.J. Dowrick, A.J. Fisher, and M.E. Newman, The Theory of
Critical Phenomena, Oxford University Press (Oxford, 1992).
[3] J. Cardy, Scaling and Renormalization in Statistical Physics, Cambridge Uni-

versity Press (Cambridge, 1996).
[4] W. Greiner, L. Neise, and H. Stöcker, Thermodynamics and Statistical Me-

chanics, Springer-Verlag (New York, 1995).
[5] K. Huang, Statistical Mechanics, J. Wiley (New York, 1987).
[6] R. Kubo, Statistical Mechanics, North Holland (Amsterdam, 1967).
[7] L.D. Landau and E.M. Lifshitz, Statistical Physics I,II, Pergamon Press (Ox-
ford, 1980).
[8] S.-k. Ma, Modern Theory of Critical Phenomena, Benjamin (Reading, 1976).
[9] B.M. McCoy and T.T. Wu, The Two-Dimensional Ising Model, Harvard (Cam-
bridge, 1973).
[10] M. Plischke and B. Bergersen, Equilibrium Statistical Mechanics, World

Scientific (Singapore, 1994).
[11] F. Schwabl, Statistical Mechanics, Springer Verlag (Berlin, 2002).
[12] H.E. Stanley, Phase Transitions and Critical Phenomena, Oxford University
Press (London, 1971).
129
[13] K.G. Wilson, Scientific American, 241, 140 (1979).
[14] Julia M Yeomans, Statistical Mechanics of Phase Transitions, Clarendon

(Oxford, 1992).
130
Appendix A
Distribution Functions
We have to distinguish three kinds of systems:

(a) identical but distinguishable particles (”classical particles”)
(b) identical indistinguishable particles with half-integer spin (Fermi-ons)
(c) identical indistinguishable particles with integer spins (Bosons)
The occupation probability of a state i with energy Ei is:
(a-Maxwell-Boltzmann):
1
n(Ei ) = βEi (A.1)
e
2
Maxwell Boltzmann distrubution
1: T = 1 000 K
n(E)
2: T = 5 000 K
1
3: T = 10 000 K
3
2
0
0 0.5 1 1.5 2
Energy (eV)
(b-Fermi-Dirac):
1
n(Ei ) = (A.2)
eβ(Ei −εF ) +1
(c-Bose-Einstein):
1
n(Ei ) = . (A.3)
eβEi −1
131
I. Vilfan Statistical Mechanics Appendix A
2
Fermi Dirac distrubution
EF = 1 eV
1: T = 0 K
2: T = 1 000 K
n(E)
2: T = 5 000 K
1
1 3: T = 10 000 K
3 2
4
0
0 0.5 1 1.5 2
Energy (eV)
5
Bose Einstein distrubution
4
3 1: T = 1 000 K
n(E)
2: T = 5 000 K
2 3: T = 10 000 K
3
1 2
1
0
0 0.5 1 1.5 2
Energy (eV)
132
Appendix B
Maxwell Relations
Let df be a total differential of the function f (x, y). Then,
df = u(x, y)dx + v(x, y)dy (B.1)
with ! !
∂f ∂f
u(x, y) = ; v(x, y) = (B.2)
∂x y
∂y x
and
∂2f ∂2f
! ! ! !
∂u ∂v
= = = . (B.3)
∂y x
∂y∂x ∂x∂y ∂x y
This is a very useful relation that - when applied to thermodynamic functions -

gives the Maxwell relations:
! !
∂T ∂p
=− , (B.4)
∂V S
∂S V
! !
∂T ∂V
= , (B.5)
∂p S
∂S p
! !
∂S ∂V
=− , (B.6)
∂p T
∂T p
! !
∂S ∂p
= . (B.7)
∂V T
∂T V
133
I. Vilfan Statistical Mechanics Appendix B
134
Appendix C
Basic Thermodynamic Relations of

Magnetic Systems Revisited
The internal energy is:

X
U (S, M ) = − Ji,j hSi Sj i (C.1)
hiji
where S is the entropy, M is the total magnetic moment (in dimensionless units),
M = N m, m is the magnetic moment per lattice site, m = hSi i. Ji,j is the
exchange interaction between the spins i and j. The (magnetic) enthalpy is (here
we will use E for the enthalpy, to avoid confusion with the field H):
E(S, H) = U − HM ≡ hHi (C.2)
where H is the magnetic field in units of energy. The Helmholtz free energy is:
F (T, M ) = −kB T ln Z(T, M, N ) = U − T S (C.3)
and its total differential is
dF = −S dT + H dM (C.4)
(see Table II of Part I; the number of spins N is constant): The Gibbs free energy
is
G(T, H) = −kB T ln Y (T, H) = E − T S

= F − HM = U − T S − HM (C.5)
135
I. Vilfan Statistical Mechanics Appendix C
and its differential is
dG = −S dT − M dH (C.6)
Here, we introduced also the corresponding partition functions Z and Y . From

Eq. (C.4) we get immediately the equation of state:
!
∂F
H= (C.7)
∂M T,N
From Eq. (C.6) we also get an equation of state:

! !
∂G ∂E
M =− =− (C.8)
∂H T,N
∂H S,N
Of course, the equations (C.7) and (C.8) must lead to the same equation, they are
just two ways of writing the equation of state. When we want to study systems at
constant H (and not at constant M ), the more convenient thermodynamic potential
is the Gibbs free energy G(T, H, N ) instead of the Helmholtz free energy.
Other useful thermodynamic relations are:
∂2G
! !
∂S dU
CH = T = = −T (C.9)
∂T H
dT ∂T 2 H
∂2F
! !
∂S
CM = T = −T (C.10)
∂T M
∂T 2 H
∂2G
! !
∂M
χT = =− (C.11)
∂H T
∂H 2 T
∂2E
! !
∂M
χS = =− (C.12)
∂H S
∂H 2 S
136

Vilfan I. - Lecture Notes in Statistical Mechanics (2002) PDF

Uploaded by

Copyright:

Available Formats

Vilfan I. - Lecture Notes in Statistical Mechanics (2002) PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Vilfan I. - Lecture Notes in Statistical Mechanics (2002) PDF

Uploaded by

Copyright:

Available Formats

Lecture Notes

The Abdus Salam ICTP, Trieste, Italy

Trieste, Autumn 2002

2.4 Two-dimensional xy Model . . . . . . . . . . . . . . . . . . . . . 42

4 Methods of Statistical Mechanics 73

Further Reading 129

A Distribution Functions 131

B Maxwell Relations 133

C Basic Thermodynamic Relations of Magnetic Systems Revisited 135

Table 1: Ensembles in statistical mechanics

Ensemble Partition function Thermodynamic potential

Microcanonical ∆Ω(E, V, N ) S(E, V, N ) = kB ln ∆Ω(E, V, N )

Canonical Z(T, V, N ) F (T, V, N ) = −kB T ln Z(T, V, N )

= ∆Ω(E, V, N )e−E(V,N )/kB T dE

Y (T, p, N ) G(T, p, N ) = −kB T ln Y (T, p, µ)

Grand canonical Ξ(T, V, µ) J(T, V, µ) = F − µN = −pV

Table 2: Thermodynamic functions

Thermodynamic functions Natural Total differential

Helmholtz free energy

Gibbs free energy

Table 3: Relations of different thermodynamic quantities to the partition function

Canonical partition function

Z(T, V, N ) = ∆Ω(E, V, N )e−βE dE

Discrete energy levels:

F (T, V, N ) = −kB T ln Z(T, V, N ) G(T, P, N ) = F (T, V, N ) + P V

Pressure Entropy Entropy Volume

Inverse Heat capacity Heat capacity Isothermal

Figure 1.1: Constant-energy sur- Figure 1.2: Constant-energy shell

In principle, all systems in Nature obey quantum statistical mechanics, the

Although we will discuss only statistical mechanics of systems in equilibrium,

1.2 Ensembles in Statistical Mechanics

1.2.1 Microcanonical Ensemble and the Entropy

∆Ω(E) is the number of configurations having the energy in the interval E ≤

Entropy and Probability

We postulate that the entropy is

1.2.2 Canonical Ensemble and the Free Energy

∆Ωt (Et ) = ∆Ωs (Es )∆Ωb (Et − Es ). (1.14)

∆Ωs (Es )∆Ωb (Et − Es )dEs

F (T, V, N ) = −kB T ln Z(T, V, N ) . (1.24)

F (T, V, N ) is a Legendre transform of hE(S, V, N )i, by which the independent

dF = −SdT − P dV + µdN. (1.26)

G(T, P, N ) = F (T, V, N ) + P V (1.27)

For a list of thermodynamic functions and their derivatives, see Tables 1 to 3 in

1.2.3 Grand Canonical Ensemble

∆Ω(Et , Nt ) = ∆Ωs (Ei , Ns ) ∆Ωb (Et − Es , Nt − Ns ). (1.33)

Thermodynamic quantities in the grand canonical ensemble

the pressure is: ! !

For a system in equilibrium at constant T , V , and µ, the distribution of a system

1.3 Analogy Between Fluids and Magnetic Systems

Similarly, the ”generalized coordinate” V is replaced by the ”generalized coordi-

(∆E)2 = h(E − hEi)2 i = hE 2 i − hEi2 . (1.48)

In the canonical ensemble, the thermal average of the energy is

distribution of energies in a canonical ensemble is so sharply peaked around the

The summand on the right-hand side is the spin-spin correlation function,

Γij = [hSi Sj i − hSi ihSj i] . (1.68)

Γ(r) ∝ r−p e−r/ξ , (1.69)

Figure 1.6: Chemical potentials of two phases in contact. When µ1 = µ2 , the

1.5 Phase Transitions

µ1 (T, p1 , · · ·) = µ2 (T, p2 , · · ·) (1.70)