Statistical Physics

Gianni Blatter

2020
ii

Warning: This document is originally based on a handwritten script

in German. Errors may appear in translation and we appreciate much your
help in their identification. Please send corrections and suggestions to

blatterj@[Link], Subject: Comment to Statphys-Skript

future generation of students will be grateful for your help.

Thanks: I wish to thank Pascal Steger for his help in writing this elec-
tronic form of the script and numerous students for their valuable comments.

G. Blatter

$Id: [Link] 1907 2007-03-19 13:25:21Z blatter$

Contents

0 Literature 1

1 Thermodynamics 3
1.1 Basic concepts and definitions . . . . . . . . . . . . . . . . . . 4
1.1.1 State variables, equation of state . . . . . . . . . . . . 4
1.1.2 Functions of many variables, differentials, integrability 5
1.1.3 Process types . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.4 Units and coefficients . . . . . . . . . . . . . . . . . . 9
1.2 Three basic Laws of Thermodynamics . . . . . . . . . . . . . 11
1.2.1 The First Law of Thermodynamics, energy . . . . . . 11
1.2.2 The Second Law of Thermodynamics, entropy . . . . . 14
1.2.3 The third law of thermodynamics, zero temperature . 17
1.3 Multicomponent systems . . . . . . . . . . . . . . . . . . . . . 20
1.4 Thermodynamic potentials . . . . . . . . . . . . . . . . . . . 21
1.4.1 Thermodynamic potentials with extensive variables . . 21
1.4.2 Constraining parameters . . . . . . . . . . . . . . . . . 22
1.4.3 Legendre transformation . . . . . . . . . . . . . . . . . 24
1.4.4 Thermodynamic potentials with intensive variables . . 26
1.4.5 Equation of Gibbs-Duhem . . . . . . . . . . . . . . . . 28
1.5 Phenomenology of phase transitions . . . . . . . . . . . . . . 30
1.5.1 Gibbs’ phase rule . . . . . . . . . . . . . . . . . . . . . 30
1.5.2 Gibbs surfaces . . . . . . . . . . . . . . . . . . . . . . 32
1.5.3 Ehrenfest classification of phase transitions . . . . . . 33
1.5.4 Phase diagrams . . . . . . . . . . . . . . . . . . . . . . 34

iii
iv CONTENTS

1.5.5 Free energies . . . . . . . . . . . . . . . . . . . . . . . 36

1.5.6 Constraining parameters in phase transitions . . . . . 37
1.5.7 Entropy surfaces . . . . . . . . . . . . . . . . . . . . . 39
1.5.8 Nucleation in first-order transitions . . . . . . . . . . . 40

2 Statistical Description 45

3 Classical statistical mechanics 49

3.1 Microcanonical ensemble . . . . . . . . . . . . . . . . . . . . . 50
3.2 Canonical ensemble . . . . . . . . . . . . . . . . . . . . . . . . 56
3.3 Grand canonical ensemble . . . . . . . . . . . . . . . . . . . . 58
3.4 Fluctuations* . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.4.1 Energy fluctuations in the canonical ensemble . . . . . 61
3.4.2 Density fluctuations in the grand canonical ensemble . 64

4 Quantum statistical mechanics 67

4.1 Expectation values in quantum statistics . . . . . . . . . . . . 67
4.1.1 Density matrix . . . . . . . . . . . . . . . . . . . . . . 69
4.2 Ensembles in quantum statistics . . . . . . . . . . . . . . . . 70
4.3 The 3rd law . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

5 Applications and techniques 73

5.1 One particle statistical mechanics . . . . . . . . . . . . . . . . 73
5.1.1 Classical free particle . . . . . . . . . . . . . . . . . . . 73
5.1.2 Quantum mechanical free particle . . . . . . . . . . . 73
5.1.3 Classical particle in a potential . . . . . . . . . . . . . 74
5.1.4 Interacting classical particles . . . . . . . . . . . . . . 75
5.1.5 Quantum particle in a potential . . . . . . . . . . . . . 75
5.2 Gases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.2.1 The classical ideal gas . . . . . . . . . . . . . . . . . . 79
5.2.2 Ideal quantum gases . . . . . . . . . . . . . . . . . . . 81
5.2.3 Classical limit of ideal quantum gases . . . . . . . . . 84
5.2.4 Quantum corrections to the interacting classical gas . 86
5.2.5 Interacting particles at T → 0 . . . . . . . . . . . . . . 86
CONTENTS v

5.3 Path integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

5.4 Variational principle . . . . . . . . . . . . . . . . . . . . . . . 96
5.5 Cumulant expansion . . . . . . . . . . . . . . . . . . . . . . . 97

6 Fermions 99
6.0.1 Dilute Fermi gas, δ = nλ3  1 . . . . . . . . . . . . . 101
6.0.2 Degenerate Fermi gas, δ = nλ3  1 . . . . . . . . . . . 101
6.1 Electrons in a magnetic field . . . . . . . . . . . . . . . . . . . 106
6.1.1 Landau diamagnetism . . . . . . . . . . . . . . . . . . 108
6.1.2 Pauli spin paramagnetism . . . . . . . . . . . . . . . . 113

7 Bosons 117
7.1 Photons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.2 Phonons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
7.3 Ideal Bose-gas:
Bose-Einstein-condensation . . . . . . . . . . . . . . . . . . . 125

8 Magnetic order 133

8.1 Generic models of local magnetism . . . . . . . . . . . . . . . 134
8.1.1 Heisenberg model . . . . . . . . . . . . . . . . . . . . . 134
8.1.2 XY model . . . . . . . . . . . . . . . . . . . . . . . . . 136
8.1.3 Ising model . . . . . . . . . . . . . . . . . . . . . . . . 136
8.2 Magnetization and Susceptibility . . . . . . . . . . . . . . . . 136
8.3 Partition function for a spin system . . . . . . . . . . . . . . 137
8.4 Broken symmetry and order . . . . . . . . . . . . . . . . . . . 139
8.4.1 Susceptibility and correlations . . . . . . . . . . . . . 141
8.5 Mean-field theory . . . . . . . . . . . . . . . . . . . . . . . . . 142
8.5.1 Ising model . . . . . . . . . . . . . . . . . . . . . . . . 142
8.5.2 The O(n) model . . . . . . . . . . . . . . . . . . . . . 147
8.6 Domain walls in the 1D-Ising model . . . . . . . . . . . . . . 149
8.7 Magnetic order in two dimensions . . . . . . . . . . . . . . . . 152
8.7.1 2D Ising model . . . . . . . . . . . . . . . . . . . . . . 152
8.7.2 XY model in 2D . . . . . . . . . . . . . . . . . . . . . 154
8.7.3 Heisenberg model [O(n) model] . . . . . . . . . . . . . 163
vi CONTENTS

9 Gas–Liquid Transition 165

9.1 The van der Waals Gas . . . . . . . . . . . . . . . . . . . . . 165
9.1.1 Universal form . . . . . . . . . . . . . . . . . . . . . . 166
9.1.2 The Maxwell construction . . . . . . . . . . . . . . . . 167
9.1.3 Microscopic analysis . . . . . . . . . . . . . . . . . . . 169
9.1.4 Critical region of the gas–liquid transition . . . . . . . 171

10 Lee-Yang Theorem∗ 177

11 Landau Theory 181

11.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
11.2 Landau free energy functional . . . . . . . . . . . . . . . . . . 183
11.3 Second-order phase transitions . . . . . . . . . . . . . . . . . 184
11.3.1 Non-local susceptibility and correlation length . . . . 187
11.4 First-order phase transitions . . . . . . . . . . . . . . . . . . . 191
11.4.1 Multicritical points . . . . . . . . . . . . . . . . . . . . 195

12 Field Theory: Gauss and Ginzburg 203

12.1 Lattice field theory and continuum limit . . . . . . . . . . . . 205
12.1.1 ~k-space . . . . . . . . . . . . . . . . . . . . . . . . . . 207
12.2 Gaussian model . . . . . . . . . . . . . . . . . . . . . . . . . . 208
12.2.1 Free energy and specific heat . . . . . . . . . . . . . . 208
12.2.2 Susceptibility and correlator . . . . . . . . . . . . . . . 210
12.2.3 Gaussian approximation of a Hamiltonian . . . . . . . 212
12.3 Self-consistent field, Hartree approximation . . . . . . . . . . 215
12.4 The Ginzburg criterion . . . . . . . . . . . . . . . . . . . . . . 220

13 Critical Fluctuations and Scaling 225

13.1 Scaling laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
13.1.1 Widom scaling . . . . . . . . . . . . . . . . . . . . . . 231

14 Renormalization Group Theory 235

14.1 Kadanoff’s block spin transformation . . . . . . . . . . . . . . 235
14.2 Renormalization group equations . . . . . . . . . . . . . . . . 239
CONTENTS vii

14.3 Fixed points and exponents . . . . . . . . . . . . . . . . . . . 241

15 Real Space RG, Spin Models 247

15.1 1D Ising model . . . . . . . . . . . . . . . . . . . . . . . . . . 247
15.2 2D Ising model . . . . . . . . . . . . . . . . . . . . . . . . . . 252
15.2.1 Fixed points and phase diagram . . . . . . . . . . . . 256

16 k-space RG 259
16.1 Gauss model . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
16.2 Φ4 theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
16.3 Dangerously irrelevant variables . . . . . . . . . . . . . . . . . 274
16.4 β-function and ‘charge’ . . . . . . . . . . . . . . . . . . . . . . 275
16.5 Two-loop expansion for G and finite η . . . . . . . . . . . . . 277

17 BKT transition 281

viii CONTENTS
Chapter 0

Literature

K. Huang, Statistical Mechanics (John Wiley & Sons, New York, 1987),
good survey.
A. Sommerfeld, Thermodynamik und Statistik (Vorlesungen über theo-
retische Physik. Band V, Harri Deutsch, 1977), good survey, old style.
L.E. Reichl, A Modern Course in Statistical Physics (E. Arnold (Publish-
ers), 1980, 1987), good survey.
D. Chandler, Introduction to Modern Statistical Physics (Oxford Univer-
sity Press, New York, 1987), good survey.
H. Smith, H. Jensen, Transport Phenomena, (Claredon Press, Oxford,
1989), Transportphenomena.
N. Straumann, Thermodynamik (Lecture Notes in Physics, Springer, Berlin,
1986), Focus on Thermodynamics.
S.K. Ma, Statistical Mechanics (World Scientific, Singapore, 1985), Focus
on Statistical Mechanics.
R. P. Feynman, Statistical Mechanics, a set of lectures (Frontiers in Physics,
Benjamin/Cummings, Reading, Massachusetts, 1982), Path Integrals.
N. Goldenfeld, Lectures on Phase Transitions and the Renormalization
Group (Frontiers in Physics, Addison-Wesley, Reading, Massachusetts,
1994), Phase Transitions, Renormalization Group Theory.
M. Plischke, B. Bergersen, Equilibrium Statistical Mechanics (Prentice-
Hall, 1989), Phase Transitions.
S.K. Ma, Modern Theory of Critical Phenomena (Frontiers in Physics,
Benjamin/Cummings, Reading, Massachusetts, 1976), Phase Transi-
tions, Renormalization Group Theory.

1
2 CHAPTER 0. LITERATURE

P. Chaikin, T. Lubensky, Principles of Condensed Matter Physics (Cam-

bridge University Press, 1995), Phase Transitions, Renormalization
Group Theory, another view on condensed matter with a strong sta-
tistical mechanics touch.

P.W. Anderson, Basic Notions of Condensed Matter Physics (Frontiers

in Physics, Benjamin/Cummings, Reading, Massachusetts, 1984), Phase
Transitions, Broken Symmetries.

N.G. van Kampen, Stochastic Processes in Physics and Chemistry (North-

Holland, Amsterdam, 1992), Stochastic Processes.

U. Weiss, Quantum Dissipative Systems (World Scientific, Singapore, 1993),

Dissipative Systems.
Chapter 1

Thermodynamics

Thermodynamics (TD) deals with macroscopic properties of a many-body

system in thermodynamic (thdyn) equilibrium. The latter is reached under
free conditions (leaving the system by itself in the absence of any forces,
probes, etc.) in the long-time limit.
Thermodynamics has been developed in the context of studies on steam
engines and is a phenomenological theory resting on three Laws of Thermo-
dynamics, dealing with energy, entropy, and the approach to zero absolute
temperature. Starting with the three Laws of Thermodynamics and using
the mathematical theory of functions of many variables, Thermodynamics
serves to obtain consistent relations between the various thermodynamic
quantities or variables.
Statistical Mechanics is a microscopic theory that has been developed to
derive the Thermodynamic Laws from basic microscopic principles (the sys-
tem Hamiltonian). Over the decades, the goals have expanded in numerous
directions, particularly the understanding of phase transitions, the physics
of defects and of disordered systems, the complex behavior of glasses, to
name a few.
In this chapter, we briefly summarize the basic elements of Thermody-
namics, assuming that students are already familiar with the topic. This
short summary then is meant to provide an initial frame for the discussion
of Statistical Mechanics which is the main focus of this lecture.
So far the lecture deals with Equilibrium Physics. However, the reader
should be aware that Nonequilibrium Phenomena is a topic of great impor-
tance as well. They appear in two facets: linear response (transport, polar-
ization, fluctuations, etc.) studies the behavior of systems under small ap-
plied forces (or no force when discussing static/dynamic fluctuations around
equilibrium) where the system remains close to equilibrium and hence the
response can be calculated perturbatively from the equilibrium state. Hence,

3
4 CHAPTER 1. THERMODYNAMICS

knowing everything about the equilibrium state (including correlators, fluc-

tuations) allows to calculate the linear response. This behavior is often
encountered in traditional condensed matter systems as these equilibrate
quickly and forces are often small (to avoid damage). The other facet then
is to understand systems far from equilibrium, e.g., the evolution of a system
quenched to a (non-equilibrium) initial state, systems under strong drives
generating non-linear response, driven-dissipative steady states, etc. The
latter topic has gained in importance with the advent of quantum engineered
systems, cold atoms and molecules, photonic cavity systems, etc. These
systems are easily driven far away from equilibrium and relax only slowly
(few collisions in atomic clouds, continuous pumping of photons), if at all.
Approaches to non-equilibrium physics include the Boltzmann Transport
Equation, Hydrodynamics, stochastic- or Langevin equations, the Fokker-
Planck equation, classical (rate) and quantum (Lindblad) master equations,
etc. This lecture will not cover these topics.

1.1 Basic concepts and definitions

1.1.1 State variables, equation of state

A thermodynamic system is described by (thermodynamic) state variables,

typical examples being the volume V , pressure p, temperature T , particle
number N , chemical potential µ, magnetic field H, magnetization M , etc.,
but also less trivial ‘derived’ quantities such as the inner energy U , the en-
tropy S and the other thermodynamic potentials F (free energy), G (Gibbs
free energy), H (enthalpy), and Ω (thermodynamic or grand potential). We
distinguish intensive (p, T , µ, H, independent on the system’s size) from ex-
tensive variables (V , S, N , M , proportional to the system’s size). The pairs
p and V , T and S, µ and N , H and M are dual and their product defines
an energy, pV = work, T S = heat, µN = chemical energy, HM = magnetic
energy. In thermal equilibrium, the intensive variables are homogeneous
throughout the system (this is a characteristic of the equilibrium state), i.e.,
they do not depend on position. A gradient ∇T ~ , ∇µ~ is characteristic of
a non-equilibrium state where heat and particles flow through the system;
the gradient in pressure ∇p ~ generates ‘mechanical’ transport. Hence, an
equilibrium state is described by a minimal amount of information.
These thermodynamic variables are interdependent and thus related
through thermodynamic equations, equations of state
f (p, V, T, . . .) = 0 (1.1)
as well as caloric equations
U = U (T, V, . . .), (1.2)
1.1. BASIC CONCEPTS AND DEFINITIONS 5

famous examples being those of the ideal gas, its equation of state pV = nRT
(n = moles of substance, R = 8.314 J/mol K the gas constant) and its caloric
equation U = (3/2)nRT . Note that in statistical mechanics these laws
read pV = N kB T and U = (3/2)N kB T with the Boltzmann constant kB =
1.381 10−23 J/K and n = N/NA , R = kB NA with NA = 6.022 1023 /mol the
Avogadro number. Furthermore, note the R is a phenomenological constant
while kB is a microscopic quantity.

1.1.2 Functions of many variables, differentials, integrability

The basic mathematical tool in the description of thermodynamic relations

are functions of many variables, their differentials, and the integrability of
differential forms. Consider two states A and B in state space, see Fig. 1.1,
e.g., characterised by the state variables pA,B und TA,B . A state function
Z is defined through its characteristic property that it depends only on the
state (A or B) itself but not on the way the state has been reached. Hence,
we can go from A to B using any path γ connecting A and B,
Z Z
ZB = ZA + dZ = ZA + dZ,
γ1 γ2

and correspondingly, we find that the loop (γ = γ1 − γ2 ) integral of the

differential dZ of a state function Z vanishes,
Z
dZ = 0. (1.3)
loop γ
In the thermodynamic context, we encounter both state functions and quan-
tities that do not describe the state. E.g., as already mentioned, U , S and
T , V and p, N and µ are state functions, but W = pV (work) and Q = T S
(heat) are not. It is thermodynamics that says that U and S are state func-
tions (and how to build them via the First and Second Law) but that only the
combination of heat and work define a state function, U = T S − pV (+µN ).
Of central importance when dealing with thermodynamics is the under-
standing of functions of many variables and their differentials. Let Z(x, y)
be a state function, x and y the chosen independent state variables. Then
dZ = Xdx + Y dy,
∂Z ∂Z
X = , Y = ,
∂x y ∂y x
∂X ∂Y
= . (1.4)
∂y ∂x
The last relation is necessary and sufficient for the integrability of the com-
plete or exact differential Xdx + Y dy. 1 Equivalently, we can impose the
1
Given a differential Xdx + Y dy with ∂y X = ∂x Y , i.e., a complete/exact differential,
there exists a primitive (function or integral) Z(x, y), such that dZ = Xdx + Y dy.
6 CHAPTER 1. THERMODYNAMICS

B
A f ( p,V,T ) = 0

γ1 T
B
A γ2
p

Fig. 1.1: States A and B in state space described by the thermodynamic vari-
ables p, V and T . Thermodynamic equilibrium states are defined through
the state function f (p, V, T ) = 0; the condition f (p, V, T ) = 0 defines the
surface of allowed equilibrium states. The state B can be reached from the
state A via many different paths. A state function Z has the characteristic
property that its value ZA in A can be obtained from the value ZB in B via
arbitrary paths γ connecting A and B.

integral condition on Z
I
dZ = 0 (1.5)

that guarantees its integrability. When dealing with two independent vari-
ables x and y, any 1-form Xdx + Y dy can be made exact with the help of an
integrating factor. An example in thermodynamics is the relation between
entropy and heat; the heat δQ(V, T ) is not an exact differential (hence we
write δ) but the product T −1 δQ = dS is. As a result, S is a state func-
tion but Q is not. The statement that dS is an exact differential is the
Second Law of Thermodynamics in mathematical form. In the same vain,
we have that δW and δQ are not exact differentials but the combination
dU = δQ − δW is—this this just the statement given by the First Law of
Thermodynamics in mathematical form.
Integrability becomes more cumbersome when dealing with more than
two variables.
Pn The integrability of a 1-form with n variables (or in n di-
mensions) k Xk dxk requires, that the n(n − 1)/2 conditions ∂Xi /∂xk =
∂Xk /∂xi are satisfied. This can be written in compact form as the condition
~ x ) = X,
for the vanishing of the rotation of the n-dimensional vector field Z(~ ~
~ ~ ~ ~
∇ ∧ Z = 0. For n = 3 we can write ∇ ∧ Z as a vector; for n = 4 we ob-
tain an anti-symmetric tensor. Integrability of a 1-form in three dimensions
1.1. BASIC CONCEPTS AND DEFINITIONS 7

P3 ~ Z~ 6= 0 can be achieved with the help of an integrating

k=1 Xk dxk with ∇×
factor if the condition
Z~ ·∇
~ ×Z ~ =0 (1.6)
~ ⊥∇
~ × Z.
~ The complete or exact differential P n
is fulfilled, i.e., Z k Xk dxk is
also known as a Pfaffian differential.
When dealing with functions of many variables it is helpful to remind
about a few rules of calculus: Consider the quantities/variables x, y, z,
which satisfy the condition f (x, y, z) = 0. Defining the function x(y, z),
we obtain the differential dx = ∂y x|z dy + ∂z x|y dz and setting dx = 0 or
x = const one obtains the chain rule
∂x ∂y ∂z
= −1, (1.7)
∂y z ∂z x ∂x y

where we have used the inversion

∂z 1
= , (1.8)
∂x y
∂x
∂z y

that follows from z(x, y = const) via dz = ∂x z|y dx, 1 = ∂x z|y ∂z x|y . Let
us include the additional function w(x, y) in our considerations. We can
construct the pair of functions x(y, w) und y(z, w) and their differentials

∂x ∂x ∂y ∂y
dx = dy + dw, dy = dz + dw.
∂y w ∂w y ∂z w ∂w z

Fixing w and inserting one in the other, we obtain dx = ∂y x|w ∂z y|w dz and
hence the chain rule at fixed w,

∂x ∂y ∂x
= . (1.9)
∂y w ∂z w ∂z w

1.1.3 Process types

Changing the state of a system is an important action in the context of

thermodynamics (e.g., in operating an engine). We distinguish between
reversible and irreversible processes: A reversible process is characterized by
its slow change of control variables, such that the process can be reversed
when inverting the direction in the evolution of the control parameters.
During such a reversible or quasi-static process the system always remains
in equilibrium. Note that here quasi-static means on a time scale so slow that
equilibrium can establish itself at any moment—since this usually appears on
a microscopic time scale, such a process can be apparently quite fast in real
time! A reversible process can be described by only a few number of variables
8 CHAPTER 1. THERMODYNAMICS

as the equilibrium is always guaranteed. When dealing with an irreversible

process the system leaves the manifold of equilibrium states and has to
return the final equilibrium state via relaxation from out-of-equilibrium, see
Fig. 1.2; it cannot be reversed. Usually, we cannot handle such processes
(as they involve a huge set of parameters in their description which have to
be managed by numerical approaches). If we want to determine the change
of state during an irreversible process we have to invent/find a substitute
reversible process that we can handle.

irreversible
T ? ?

A f ( p,V,T ) = 0
reversible

B
A
p

Fig. 1.2: While a reversible process proceeds within the equilibrium manifold
defined through f (p, V, T ) = 0, an irreversible one leaves the surface and has
to return to it via relaxation. While the reversible process can be described
by a minimal number/evolution of thermodynamic variables, the irreversible
process cannot be easily handled.

Typical examples of reversible and irreversible processes are the rever-

sible- and the free-isothermal expansions of a gas, see Fig. 1.3, with both
assuming the same final state.
Other types of processes encountered in thermodynamics involve restric-
tions through fixing specific variables, e.g., fixing T = const defines an
isothermal process, p = const is isobaric, V = const is isochoric, and the
absence of heat flow to/away from the system defines an adjabatic process2
The trajectories assumed during these processes are called isotherms, iso-
bars, isochors, and adjabatic curves. Running these types of processes
necessitates reservoirs providing heat or work or particles. These reservoirs
then introduce (additional) thermodynamic (intensive) variables such as the
2
Note the difference to the other use of an adjabatic process as describing a slow
process, e.g., in quantum mechanics.
1.1. BASIC CONCEPTS AND DEFINITIONS 9

0 VA VB 0 VA VB
T T

Fig. 1.3: Isothermal expansion of a gas: in the reversible process on the left,
the cylinder picks up heat from the reservoir and can convert it into work
extracted from the displacement of the piston, hence quite a useful process.
The irreversible process on the right where the gas just flows over in the
second compartment misses this opportunity—there is no heat transferred
from the reservoir to the gas and there is no work done by its expansion.
When attempting to calculate the entropy change of the right process, we
substitute it by the left one and obtain the result ∆S = nR ln(VB /VA ).

temperature T characterizing an energy/heat reservoir (changing S or Q),

the pressure p for a ‘work’ reservoir (changing V ), the chemical potential µ
for a particle reservoir (changing N ).

1.1.4 Units and coefficients

We measure

temperature T

in o C (degrees Celsius) oder K (Kelvin),

evaporation
∆T melting
(H2 O) = 100 ◦ C oder = 100 K (1.10)
T0 = T melting
(H2 O) = 0 ◦ C oder = 273.15 K;

the

pressure p

in Torr, atm, bar, Pa,

1 atm(atmosphere) = 760 Torr (760 mm Hg) (1.11)

4 2
1 at(techn. atmosphere) = 9.81 · 10 N/m (≈ 10 m H2 O) = 0.981 bar,
1 at = 0.968 atm, (1.12)
5 2 5
1 bar = 10 N/m = 10 Pa (Pascal);

the
10 CHAPTER 1. THERMODYNAMICS

heat δQ

in cal (calories), kcal, J, erg, Ws,

15.5 C

1 kcal = δQ ∆T 14.5 ◦ C
H2 O at normal pressure = 4187 J,
1 J = 1 m2 kg/s = 1 Ws = 107 erg;
2
(1.13)

the

amount n of a substance

in mol, g,
X
1 mol substance ≈ atomic weight · g, [gram-mol].
atoms in the substance
(1.14)
E.g., one mole of water

ˆ 2 · 1(H) + 1 · 16(O) = 18 g
1 mol H2 O = (1.15)
1 mol ˆ 6.022 ·
= 1023 molecules

where the normalization is defined via the carbon,

1 mol 12
C = 6.022 · 1023 · 12 · 1.66044 · 10−24 g = 12 g, (1.16)

with 1.66044 · 10−24 g corresponding to 1/12 of the mass of a single 12 C-

Atom.

Material coefficients

Finally, it is useful to remind about some of the coefficients describing the

thermodynamic properties of a substance under a change in pressure p and
temperature T .
1 ∂V
expansion coefficient α= ,
V ∂T p
1 ∂p
tension coefficient β= ,
p ∂T V
1 ∂V
(isothermal) compressibility κT = − ,
V ∂p T
1 ∂V
(adiabatic) compressibility κS = − . (1.17)
V ∂p S

These coefficients are easily measured and of much use in thermodynamic

relations. Via Maxwell’s relations, these coefficients can be transformed to
1.2. THREE BASIC LAWS OF THERMODYNAMICS 11

derivatives of non-trivial quantities. For the ideal gas we have α = β = 1/T

and κT = 1/p; indeed, a strongly pumped tire is not much compressible and
κ is small. Of special relevance is the compressibility as it defines a criterion
for the stability of a phases (similar is true for the specific heat).

1.2 Three basic Laws of Thermodynamics

The phenomenological theory of thermodynamics rests on three basic (phe-

nomenological) observations and associated abstractions. As all processes in
nature so far adhere to this consistent set of observations and since their ab-
straction entails a multitude of further physical and observable consequences
that all have turned out correct, these observations/abstractions have been
elevated to the status of ‘Thermodynamic Laws’.

1.2.1 The First Law of Thermodynamics, energy

Consider an arbitrary change of thermodynamic state during which the sys-

tem accepts the heat δQ, while producing an amount of work δW . Such a
process modifies the system’s internal energy U by

dU = δQ − δW, (1.18)

where dU is an exact differential. This statement implies that there exists

a state function U assuming values that depend on the system’s state but
not on the way the state has been reached. This state function then can be
obtained from a starting value UA , where A labels a state, via integration
along any path γAB connecting A with the final state B,
Z
UB = UA + dU. (1.19)
γAB

Alternatively, the first law states that any loop integral (or cyclic process)
over the combination δQ − δW vanishes,
I
δQ − δW = 0, (1.20)

and hence δQ − δW = dU is an exact differential.

As a little exercise, we have a look at Gay-Lussac’s experiment, where
a gas in the volume V1 at temperature T1 flows over into a volume V2 to
assume the larger total volume V1 + V2 . The experiment then tells, that
the temperature remains unchanged, T2 = T1 , hence the system has not
picked up any energy and δQ = 0. Also the system did not produce any
work as the gas merely flowed over into the empty volume without driving
12 CHAPTER 1. THERMODYNAMICS

a piston, hence δW = 0. Thus, we conclude that U = U (T ) depends only

on temperature, ∂V U |T = 0, and hence dU = CV (T )dT or
Z T
U (T ) = dT 0 CV (T 0 ). (1.21)
0

This finding is accurate for an ideal gas; for a real gas (e.g., a van der Waals
gas, see later) the internal energy turns out to depend on both T and V .
This is clear if we think in microscopic terms (which we should do in stat
mech but not in thdyn, but we do it anyway), as the atoms/molecules in
the Van der Waals gas attract, and hence expanding the gas enlarges the
distance between the particles and kinetic energy is transformed to potential
energy, thus cooling the gas internally. For a monoatomic ideal gas, CV is a
constant, hence U = CV T . This is a typical application of thermodynamic
principles—do an experiment (phenomenology) and use the thermodynamic
laws to obtain new insights, here, a simplification of a general state function
U (V, T ) to a simpler one U (T ).
In extending the above exercise, we compare a process that changes the
gas’ state at fixed volume V versus one where the pressure p is fixed,

V = const dUv = δQ = CV dT, CV = ∂T U |V , (1.22)

p = const dUp = δQ − δW = Cp dT − p dV, Cp = ∂T (U + pV )|p . (1.23)

Making use of the equation of state p dV = nR dT in (1.23) (where we have

used that p = const in this second process) we find that


dUp = Cp − nR dT = dUV = CV dT, (1.24)

where we have assumed the same change in temperature dT for the two
processes. Hence, the First Law gives us a new insight on the relation
between specific heats,

Cp − CV = nR > 0, (1.25)

thus relating experiments that probe different caloric properties of the sys-
tem. Again, thinking about the result, it is clear that Cp > CV as in the
second experiment the system produces work (on a piston that holds p)
while changing its temperature by the same amount, so the system has to
pick up more heat in the second process.
Finally, we discuss the adjabatic reversible expansion of an ideal gas
which is characterized by δQ = 0 and δW = p dV , see Fig. 1.4. Making use
of the First Law and the caloric equation dU = CV dT , we find that

CV dT = dU = δQ − δW = −p dV, (1.26)
1.2. THREE BASIC LAWS OF THERMODYNAMICS 13

cylinder
Fig. 1.4: Geometry of the
p cylinder-piston configuration in
T
the adiabatic-reversible expansion.
V piston

δQ = 0

and accounting for the equation of state dT = (p dV + V dp)/nR (both, p

and V change at fixed n), we obtain


CV + nR p dV + CV V dp = 0. (1.27)

Using Eq. (1.25) this simplifyes to

dp cp dV
+ = 0,
p cV V
ln p + ln V cp /cV = const,
p V cp /cV = const . (1.28)

Fig. 1.5: Adjabatic versus

p adiabatic
isothermal processes. Both are
relevant in the Carnot machine
or process when building a cyclic
machine converting between
isothermal heat and work.

Rewriting this relation in different pairs of variables p, V ; p, T ; V, T , we

obtain the equations for adjabatic tranformations in the form (with Cp /CV =
γ, for a monoatomic gas we have CV = 3/2 and γ = 5/3)

p V γ = const, p V 5/3 = const, (1.29)

1−γ
p γ T = const, p−2/3 T = const,
V γ−1 T = const, V 2/3 T = const.

When comparing the adjabatic process to an isothermal one for which p V =

const one immediately observes that the adjabatic curve drops more steeply,
see Fig. 1.5.
14 CHAPTER 1. THERMODYNAMICS

1.2.2 The Second Law of Thermodynamics, entropy

The Second Law of Thermodynamics rests on the observations that

Clausius: heat can never pass from a colder to a warmer body without
some other change, connected therewith and occurring at the same time.

Kelvin: it is impossible to devise a cyclically operating device, the sole

effect of which is to absorb energy in the form of heat from a single thermal
reservoir and to deliver an equivalent amount of work.

thdyn engine heat pump

T1 T1
Q1 Q1

W W

Q2 Q2
T2 T2

Fig. 1.6: Carnot’s cyclic processes: the engine absorbs heat Q1 from the hot
reservoir and transforms part of it into work W = Q1 − Q2 , while the rest
flows as waste heat Q2 into the cold reservoir. The pump or refrigerator
consumes work W = Q1 − Q2 to transfer the heat Q2 to the hot reservoir
(all quantities W , Q1 , and Q2 are positive in this accounting).

The two statements are equivalent. A cyclic machine delivering work

thus produces waste heat that has to be fed to a second reservoir at lower
temperature. This insight has been formalized by Sadi Carnot with his heat
engine or heat pump (refrigerator) shown in Fig. 1.6. They both run a cyclic
trajectory in the p-V diagram of Fig. 1.7 comprising isothermal (reservoirs
attached) and adjabatic (reservoirs detached) processes.
Carnot’s considerations for an ideal maschine which operates reversibly
and hence produces no additional losses leads us to a relation between the
transferred amounts of heat Q1 and Q2 and the temperatures T1 and T2 of
the reservoirs,

Q1 T1
= . (1.30)
Q2 T2

In doing so, we have introduced the concept of absolute temperature T ,

which can be related one to one to the gas temperature Tgas defined in
1.2. THREE BASIC LAWS OF THERMODYNAMICS 15

Fig. 1.7: Cyclic process 1 → 2 →

p 1
3 → 4 → 1 within the p, V -
diagram, using a gas as a working
2 medium.
W T1

4 3
T2
V

(1.10) if we use an ideal gas as the working medium, T = Tgas . Combining

the above result with the First Law in the form W = Q1 − Q2 , we find the
optimal efficiencies for the Carnot cycle,
W Q2 T2
ηengine = =1− =1− (1.31)
Q1 Q1 T1
for the thermodynamic engine and
Q1 1 1
ηpump = = = (1.32)
W 1 − Q2 /Q1 1 − T2 /T1
for the heat pump.

p Fig. 1.8: Subdivison of an arbitrary

reversible and cyclic process into a
sequence of smaller → infinitesimal
Carnot-processes.
δW

Extending these considerations to arbitrary cyclic processes, see Fig. 1.8,

in which amounts of heat δQrev are exchanged with a reservoir at tempera-
ture T in a reversible way one obtains from (1.30) that
I
δQrev
= 0. (1.33)
T
Hence, the quantity δQrev /T is an exact differential and there exists a state
function S called entropy,
Z
δQrev
SB = SA + , (1.34)
γAB T
16 CHAPTER 1. THERMODYNAMICS

of which δQrev /T is the differential, dS = δQrev /T . In the end, we can

give an alternative formulation of the Second Law, similar to that of the
First Law: Consider an arbitrary change of thermodynamic state during
which the system reversibliy exchanges the heat δQrev with a reservoir at
temperature T . Such a process modifies the system’s entropy S by
δQrev
dS = . (1.35)
T
Finally, for a general process that is not necessarily reversible, we have
instead of (1.34) the inequality
Z B
δQ
→ SB > SA + (1.36)
A T

which leads us to yet another statement of the Second Law: In an isolated

system (i.e., δQ = 0) the entropy can only increase,
SB > SA . (1.37)
This equation breaks the time reversal symmetry and defines an arrow of
time.
As a first application, we determine the entropy of the ideal gas. This is
easily done by noting that 1/T is an integration factor on δQ = dU + p dV ,
hence,
δQ dU + p dV
dS = = (1.38)
T T
dT dV
= CV + nR , (1.39)
T V
where we have made use of the thermal and caloric equations of the ideal
gas, U = CV T and p = nRT /V . Proper integration provides the result
Z T,V
T V
S(T, V ) − S0 = dS = CV ln + nR ln , (1.40)
T0 ,V0 T0 V0
h  T 3/2  V i
= nR ln + ln .
T0 V0
Note, that we find S only up to a constant S0 —the problem of unknown S0
will be the subject of the Third Law of Thermodynamics.
Next, we briefly establish the relation between the caloric and thermo-
dynamic equations of state. We start from the First and Second Laws of
Thermodynamics in the form
1 p
dS = dU + dV
T T
∂U ∂U
dU = dT + dV
∂T V ∂V T
1.2. THREE BASIC LAWS OF THERMODYNAMICS 17

and insert the second equation into the first,

∂U  ∂U 
T dS = dT + + p dV. (1.41)
∂T V ∂V T

The differential of S is
∂S ∂S
T dS = T dT + T dV
∂T V ∂V T

and comparing the two expressions for T dS, we obtain the following relations
between partial derivatives,
∂S 1 ∂U ∂S 1  ∂U 
= , = +p .
∂T V T ∂T V ∂V T T ∂V T

Since S is a state function (or, equivalently, dS is exact), we can relate these

partial derivatives via
∂2S ∂2S
= ,
∂T ∂V ∂V ∂T
from which we obtain
∂ h 1 ∂U i ∂ h 1  ∂U i
= +p ,
∂V T ∂T V ∂T T ∂V T
1 ∂ U2 1  ∂U  1 ∂2U 1 ∂p
= − 2 +p + + ,
T ∂V ∂T T ∂V T T ∂T ∂V T ∂T V
and finally
∂U ∂p
=T − p. (1.42)
∂V T ∂T V
This equation relates the caloric quantity U with the thermal equation of
state p(V, T ). It is easy to check this relation for the ideal gas—it immedi-
ately predicts the outcome of the Gay-Lussac experiment, ∂V U = 0.

1.2.3 The third law of thermodynamics, zero temperature

The Third Law of Thermodynamics (due to Planck and Nernst) tells, that

Planck-Nernst: the entropy tends to a constant value S0 , independent

on pressure, volume, state of aggregation, when the temperature is lowered
towards its absolute zero.
In general, we can put this constant equal to zero, S0 = 0. Since, as we
will see, the entropy quantifies the number of microstates that the system
can assume at given thermodynamic conditions (e.g., fixed N , V , and en-
ergy E), the value S0 = 0 at T = 0 is compatible with the statement that
the system goes into its quantum-mechanical groundstate as T approaches
18 CHAPTER 1. THERMODYNAMICS

zero (provided that the groundstate is not degenerate). Complications obvi-

ously arise when the excitation spectrum has no gap (a mild complication,
see the discussion of BEC in the ideal bosonic quantum gas) or when the
phase space develops a complex structure at low energies, as is the case in a
glass (ergdocity breaking, phase space breaks up into well separated ergodic
components, hierarchical symmetry breaking, multi-valley energy landscape,
ultrametricity, still a challenging problem of condensed matter physics). The
full content of the Third Law can only be appreciated in a quantum mechan-
ical setting. In quantum mechanics, we deal with quantized states that we
can properly count, at least in a well defined (normalizable) setting. Hence
‘counting’ of microstates (→ W ) makes sense and S ∝ ln W can be properly
evaluated. In classical physics, phase space is continuous and we do not
know how to ‘count’. Choosing a volume 2π~ → S ∝ ln Γ/2π~ for a degree
of freedom seems fine, but if we choose another numerical α 6= 1 in the basic
phase space volume, i.e., 2πα~, we get a shift in S, S → S − kB ln α, thus
S0 is ill defined in classical physics.
The third law has a number of specific consequences, e.g., the thermody-
namic coefficients α, β, CV , Cp all have to vanish upon approaching T = 0.
Indeed, using Maxwell’s relations for G and F (see next section on thermo-
dynamic potentials) one finds that
1 ∂V 1 ∂S
α= =− →0 for T → 0 (1.43)
V ∂T p V ∂p T

as S does no longer depend on pressure as T → 0, or

1 ∂p 1 ∂S
β= = →0 for T → 0. (1.44)
p ∂T V p ∂V T

Similarly, for the specific heat,

∂S ∂S
CV = T , Cp = T , (1.45)
∂T V ∂T p
we can integrate
T
CV (V, T 0 )
Z
S(V, T ) = dT 0 + S(V, 0), (1.46)
0 T0

T
Cp (p, T 0 )
Z
S(p, T ) = dT 0 + S(p, 0), (1.47)
0 T0
and set S(V, 0) = 0, S(p, 0) = 0, since S(T → 0) becomes independent on V
and p. In order for the integrals not to diverge as T → 0, the specific heats
Cp , CV (T → 0) → 0 must vanish. From condensed matter courses, we know
that indeed CV ∝ T 3 for an elastic system (phonons) and CV ∝ T in the
electron system of a metal.
1.2. THREE BASIC LAWS OF THERMODYNAMICS 19

Checking matters for the ideal gas, we run into a problem—the entropy
S(V, T ) in Eq. (1.40) is ill defined as T → 0 and the specific heat CV does
not approach 0. Again, quantum mechanics comes to our rescue, telling us
that the ideal classical gas is a high temperature approximation of physical
ideal gases which are made from bosons or fermions. Ideal fermions then
become degenerate as T is lowered and CV ∝ T , while bosons condense and
CV ∝ T 3/2 as T → 0.
An important consequence of the Third Law is the statement that the ab-
solute zero in temperature cannot be reached. We show this for a mechanical
process—the translation to other processes, e.g., adjabatic demagnetization,
is straightforward. Using (1.43) and (1.47) we can write

T
Cp (p, T 0 ) T
dT 0
Z Z
∂S ∂ ∂Cp
Vα=− =− dT 0 =− .
∂p T ∂p 0 T0 0 ∂p T0 T0

We expand the specific heat Cp = A(p)T x + . . . and find that

Z T
∂A x−1 ∂p A
Vα=− dT T = −T x ∝ Cp
0 ∂p x

and hence the ratio

Vα
⇒ → const 6= 0 für T → 0 (1.48)
Cp

assumes a finite constant value at low temperatures. Expressing the change

in entropy through material properties,3

T dS = Cp dT − T V α dp,

we find that an adiabatic expansion with a reduction dp in pressure that

cools the system produces ever smaller steps dT in temperature reduction
as T → 0,
V α
dT = T dp. (1.50)
Cp
Hence, such an iterative cooling procedure will never reach the absolute zero
in temperature.
3
Starting from S(p, T ), we can find, in analogy to (1.41), that
 ∂U ∂V  ∂V
T dS = +p dT − T dp. (1.49)
∂T p ∂T p ∂T p

Then use that the prefactor of dT is Cp and the prefactor of dp is T V α, with α the thermal
expansion coefficient.
20 CHAPTER 1. THERMODYNAMICS

1.3 Multicomponent systems

When dealing with more complex systems we have to account for many
components (say r, components i ∈ {1, . . . , r}) that can occur in various
phases or aggregates (say ν; possible phases α ∈ {1, . . . , ν} may be solid,
liquid, or gas). A component i can show up in any of these phases α, hence
the total amount of component i is given by
ν
(α)
X
ni = ni , (1.51)
α=1

(α)
where we count the amount of substance in moles and ni quantifies the
amount of substance i in the phase α ∈ {1, . . . , ν}. Similarly, the total
internal energy, entropy, and volume of the overall system results from the
sum over its phases
X X X
U= U (α) , S = S (α) , V = V (α) . (1.52)
α α α

(α)
We define the chemical potential of the i-th component in phase α, µi , as
(α)
the change in energy under a reversible change in the amount of moles ni ,

(α) ∂U
µi = (α)
. (1.53)
(β) (α)
∂ni S,V,nj without ni

(α)
The chemical potentials µi are equilibrium parameters, the same way as
the other intensive variables p and T . Indeed, a system in thermal equi-
librium has a uniform distribution of temperature, pressure, and chemical
potentials throughout its phases,4
(µ) (ν)
T (µ) = T (ν) = T, p(µ) = p(ν) = p, µi = µi = µi . (1.54)

A gradient (or difference) in the chemical potential between two phases will
lead to a phase transformation between the phases of that component such
as to equalize the chemical potentials.
Finally, the differentials of the entropy and the internal energy pick up
additional terms accounting for a change in mole (n) or particle (N ) number,
r
X
T dS = dU + pdV − µi dni , (1.55)
i=1
r
X
dU = T dS − p dV + µi dni .
i=1
4
The intensive variables are also homogeneous throughtout the individual phases, see
Sec. 1.4.2
1.4. THERMODYNAMIC POTENTIALS 21

1.4 Thermodynamic potentials

Thermodynamic potentials are state functions of specific variables that gen-

erate all thermodynamic equations and the caloric one by taking derivatives—
once a thermodynamic potential is known, the system’s thermodynamics is
completely determined. When going to a statistical mechanics description
of a system, one of the primary goals will be the calculation of the partition
function from which we will obtain an appropriate thermodynamic potential
that is associated with the parameters imposed on the system, e.g., fixed
volume V , energy U , particle number N , or pressure p, chemical potential
µ, or temperature T .

1.4.1 Thermodynamic potentials with extensive variables

The primary thermodynamic potentials we will consider are the entropy

S and internal energy U as a function of the extensive variables U, V and
S, V , respectively; choosing these extensive variables makes sure that the
functions S(U, V ) and U (S, V ) are concave and convex, respectively. Note
that U (S, V ) and S(U, V ) are potentials but the functions U (V, T ) or S(V, T )
are not; rather, these latter are usual state functions. The first (of many)
potentials is the entropy S(U, A1 , . . . , An ), where A1 , . . . , An is the set of
extensive variables generalizing the volume. Indeed, starting from the dif-
ferential (1.38) andP generalizing the expression for the work done on the
system by δW = − k ak dAk , we obtain the generalized differential
n
1 1X
dS = dU − ak dAk . (1.56)
T T
k

Here, ai are the intensive variables dual to Ai , obvious pairs being ai and
Ai = −p and V , µ and N , σ = surface tension and A = area, φ = electric po-
tential and Q = charge, E = electric field and P = electric polarization, H =
magnetic field and M = magnetic moment, etc. The set U, A1 , . . . , An , with
U the caloric variable, is assumed to provide a complete characterization of
the system’s state through extensive variables. The equation (1.56) is known
as the Gibbs fundamental equation. Its knowledge allows to determine the
caloric and all state equations of the system, that is why S(U, A1 , . . . , An )
is called a thermodynamic potential.
In order to find the equations of state, we exploit the integrability of
dS, guaranteeing the existence of the state function S(U, A1 , . . . , An ). Its
expansion
n
∂S X ∂S
dS = dU + dAk (1.57)
∂U A1 ,...,An ∂Ak U,A1 ,...,An without Ak
k
22 CHAPTER 1. THERMODYNAMICS

and comparison with (1.56) provides the relations

1 ∂S
= → T = T (U, A1 , . . . , An ) → U = U (T, A1 , . . . , An ),
T ∂U A1 ,...,An
∂S
ak = −T → ak = ak (T, A1 , . . . , An ).
∂Ak U,A1 ,...,An no Ak

These are nothing but the caloric and thermal equations of state. Hence,
knowing a thermodynamic potential of a system allows for the immediate
derivation of its equations of state and thus its full thermodynamic charac-
terization.
The second thermodynamic potential follows immediately from the en-
tropy. Indeed, the inner energy U is a state function as well and expressing it
through the extensive variables S, A1 , . . . , An , we obtain the thermodynamic
potential U (S, A1 , . . . , An ). Its differential reads
n
X
dU = T dS + ak dAk (1.58)
k

and the equations of state follow from

∂U ∂U
T = , ak = . (1.59)
∂S A1 ,...,An ∂Ak S,A1 ,...,An without Ak

1.4.2 Constraining parameters

Before proceeding with the other thermodynamic potentials obtained through

a suitable Legendre transformation, we briefly discuss the meaning and func-
tionality of constraining parameters. E.g., consider the piston system in Fig.
1.9. The differential for the entropy now involves an additional constraining
parameter V1 ,
T dS = dU + p2 dV + (p1 − p2 )dV1 , (1.60)
where we have used that dV2 = −dV1 . Fixing U and V and varying V1 we
obtain the change in entropy
1
δS|U,V = (p1 − p2 )δV1 .
T
The Second Law tells that S is maximal in an isolated system, δS = 0, and
hence we find that
p1 = p2 in equilibrium. (1.61)
This emphasizes the role of p as an equilibrium parameter. Obviously, push-
ing V1 to a position with p1 6= p2 takes the system out of equilibrium and
we need more than the minimal number of parameters for its description.
1.4. THERMODYNAMIC POTENTIALS 23

P P’

p1, V1 p2 , V2

0 x

Fig. 1.9: Piston system with a constraint. The system volume V is parti-
tioned into two parts V1 and V2 such that V = V1 + V2 . The inner piston
P’ can be moved in order to change the volumes V1 and V2 , the constraint
demanding that ∆V1 = −∆V2 .

Letting P’ free, the piston will move such as to reach equilibrium and po-
tentially produce (useful) work. Thereby, the entropy will increase (see Fig.
1.10)
δS|U,V > 0. (1.62)

Fig. 1.10: The entropy increases

S
when removing the constraint
and approaches the equilibrium
state with maximal entropy.

0 V1

Similarly, we can, on a Gedankenlevel, introduce a constraint in U and

determine
∂S1 ∂S2 1 1
δS|U,V = δU1 + δU2 = − δU1 .
∂U1 U,V ∂U2 U,V T1 T2

Again, δS = 0 at fixed U and V and hence we find that

T1 = T2 in equilibrium. (1.63)

Furthermore, we can consider the curvature (or second order variation)

around the equilibrium state where δ 2 S|U,V < 0 (since S is maximal); this
condition guarantees the stability of the equilibrium state with positive co-
efficients CV und κT

CV > 0, κT > 0 (stability). (1.64)

24 CHAPTER 1. THERMODYNAMICS

The derivation is a bit tedious but nevertheless a nice exercise in differenti-

ating: starting with the constraint entropy

Scon (2U, 2V ; δU, δV ) = S(U + δU, V + δV ) + S(U − δU, V − δV ).

and requiring that δ 2 S|U,V < 0 as well as the extensivity S(2U, 2V ) =

2S(U, V ) we obtain

0 > [S(U + δU, V + δV ) + S(U − δU, V − δV ) − S(2U, 2V )]

∂2S 2 ∂2S ∂2S
= (δU ) + 2 δU δV + (δV )2 = 2 δ 2 S,
∂U 2 ∂U ∂V ∂V 2
where δ 2 S denotes the variation of the entropy S(U, V ) (note, that the linear
terms vanish in equilibrium). Using ∂S/∂U = 1/T , ∂S/∂V = p/T , we find
that
1 ∂2S ∂2S p ∂2S ∂2S
δ = δU + δV, δ = δU + δV
T ∂U 2 ∂V ∂U T ∂U ∂V ∂V 2
1 p
⇒ 2 δ2S = δ δU + δ δV
T T
1  ∂U ∂U  1
= − 2 δT + δV δT − 2 p δV δT
T ∂T V ∂V T T
1  ∂p ∂p 
+ δT + δV δV
T ∂T V ∂V T
CV 1 ∂U
h ∂p i 1
= − 2 (δT )2 − 2 +p−T δV δT − (δV )2 < 0
T T ∂V T ∂T V T V κT
and hence CV , κT > 0 (the second term vanishes due to (1.42)).
Equivalent conditions apply to the internal energy, see Fig. 1.11

dU ≤ 0, U only decreases at fixed S, V,

δU |S,V = 0, in equilibrium,
2
δ U |S,V > 0, stability, (1.65)

which derive from the relation

1
dS ≥ δQ/T = (dU + p dV ) (1.66)
T
replacing the statement of the Second Law (that the entropy can only in-
crease in a closed system, dS ≥ 0) when attaching the system to a reservoir
at temperature T and reversibly exchanging the heat −δQ.

1.4.3 Legendre transformation

The potentials S and U are functions of extensive variables only. When

trying to find thermodynamic potentials at fixed intensive variables we have
1.4. THERMODYNAMIC POTENTIALS 25

S, V fixed
Fig. 1.11: The behavior of
U the internal energy U under a
(virtual) constraint Y .

0 Y

yx
f (x) Fig. 1.12: Legendre trans-
formation Lf (y) of the func-
y tion f (x).

x
L f ( y)
to introduce an additional operation, the Legendre transformation. The
Legendre transformation is a tool that allows to substitute variables without
loss of information.
We start from the fact that S is concave in U and V , i.e.,

S(tX1 + (1 − t)X2 ) ≥ tS(X1 ) + (1 − t)S(X2 ), (1.67)

X = U, V.

Similarly, U is convex in S and V . We then exploit the mathematical theo-

rem that a Legendre transform Lf of a convex function f (x),

Lf (y) = sup[xy − f (x)] (1.68)

x

is again convex; furthermore for f strictly convex and differentiable the

operation L is involutory,
L(Lf ) = f, (1.69)
telling that no information is lost under the action of L. A geometrical
description of the Legendre transformation is shown in Fig. 1.12, while Fig.
1.13 illustrates the action of L in the situation of a non-differentiable f with
kinks. The latter is of physical relevance in the context of phase transitions.

It is well known from mechanics that the Legendre transformation with

the conjugate variable exchanges the role of the two variables. Here, the Leg-
endre transform of a potential will generate a new thermodynamic potential
for the dual variable.
26 CHAPTER 1. THERMODYNAMICS

f Lf

C
A
B BC
B
x y
AB

Fig. 1.13: Legendre transform of a partly convex function with kinks.

Straight pieces transform to kinks and vice versa.

1.4.4 Thermodynamic potentials with intensive variables

We start with the thermodynamic potential U (S, V ) and wish to substitute S

by T , as the temperature is a more suitable variable to fix in an experiment.

Free energy F (T, V, N )

We replace the variable S in U through its conjugate variable ∂S U |V,N = T

and define F as the Legendre transforms of U in S,
h ∂U i
F (T, V, N ) = U − S (T, V, N ) = [U − T S](T, V, N ). (1.70)
∂S V,N

Going to the differential of F , we realize that indeed, the variables T , V ,

and N are the natural ones,

dF = dU − T dS − SdT = T dS − pdV + µdN − T dS − SdT

⇒ dF = −pdV − SdT + µdN. (1.71)

The differentials provide the equations of state,

∂F ∂F ∂F
= −p, = −S, = µ, (1.72)
∂V T,N ∂T V,N ∂N T,V

and the condition for the integrability results in one of the famous Maxwell
relations
∂p ∂S
= . (1.73)
∂T V ∂V T
Replacing the entropy dS by dS = δQrev /T , we find the relation

∂p 1 δQ
= , (1.74)
∂T V T dV T
1.4. THERMODYNAMIC POTENTIALS 27

that will appear in a similar form as the Clapeyron equation in the context
of phase transitions.
By construction, F is convex in the extensive variables V aund N und
concave (note the different signs in (1.68) und (1.70)) in the intensive vari-
able T and the following minimum principle applies,

dF ≤ 0, F can only diminish at fixed T, V, and N,

δF |T,V,N = 0, in equilibrium.

The first equation follows from (1.66),

T dS ≥ dU + pdV − µdN → (dU − T dS)|T,V,N ≤ 0

dF |T,V,N = d(U − T S)|T,V = (dU − T dS)|T,V,N → dF |T,V,N ≤ 0.

Enthalpy H

The enthalpy H(S, p, N ) is obtained from the internal energy U (S, V, N ) by

replacing the volume V by the pressure ∂V U |S,N = −p,

H(S, p, N ) = [U + pV ](S, p, N ). (1.75)

H is concave in p and convex in S and N ; in equilibrium, H is minimal in

the constraints S and N . The differential has the form

dH = T dS + V dp + µ dN,

and the equations of state derive from

∂H ∂H ∂H
= T, = V, = µ. (1.76)
∂S p,V ∂p S,N ∂N S,p

Gibbs potential G

The Gibbs potential G(T, p, N ) derives from the internal energy U (S, V, N )
by replacing S and V ,

G(T, p, N ) = [U − T S + pV ](T, p, N ) = [H − T S](T, p, N ) (1.77)

= [F + pV ](T, p, N ).

G is concave in T and p and proportional to N ; minimization of G is trivial

for a one-component system, see (1.84). The differential assumes the form

dG = −S dT + V dp + µ dN,

and the equations of state derive from

∂G ∂G ∂G
= −S, = V, = µ. (1.78)
∂T p,N ∂p T,N ∂n T,p
28 CHAPTER 1. THERMODYNAMICS

Grand potential Ω

The grand potential Ω(T, V, µ) derives from U (S, V, N ) by replacing S and

N,
Ω(T, V, µ) = [U − T S − µN ](T, V, N ). (1.79)
Ω is concave in T and µ and proportional to V . The differential has the
form
dΩ = −S dT − p dV − n dµ,
and the equations of state derive from
∂Ω ∂Ω ∂Ω
= −S, = −p, = −N. (1.80)
∂T V,µ ∂V T,µ ∂µ T,V
When going to the statistical description of a system, the first goal then
will be to determine the thermodynamic potential associated with the given
parameters.

Maxwell-relations

Consider the potential X in its variables y1 , . . . , yn . The conditions of inte-

grability
∂2X ∂2X
= (1.81)
∂yi ∂yk ∂yk ∂yi
provide us with the Maxwell-relations. Such a Maxwell-relation, ∂ 2 S/∂U ∂V =
∂ 2 S/∂V ∂U has been used to relate the thermodynamic equations of state
in (1.42).

1.4.5 Equation of Gibbs-Duhem

A function f (x1 , . . . , xn ) is homogeneous of order k, if

f (λx1 , . . . , λxn ) = λk f (x1 , . . . , xn ). (1.82)
Let k = 1, then f is homogeneous and of order one and Euler’s theorem
applies (to be proven by taking the derivative of (1.82) with respect to λ),
n
X ∂f
f (x1 , . . . , xn ) = xi (1.83)
∂xi xj 6=xi
i=1
The potential U (S, V, N ) is homogeneous and of first order, hence
U = T S − pV + µN,
F = −pV + µN,
H = T S + µN, (1.84)
G = µN,
Ω = −pV.
1.4. THERMODYNAMIC POTENTIALS 29

Taking the total differential of the first equation in (1.84) we obtain

dU = T dS + S dT − p dV + V dp + µ dN + N dµ,

and exploiting the usual form of the differential dU in its natural variables
S, V , and N ,
dU = T dS − p dV + µ dN, (1.85)

we arrive at the equation of Gibbs-Duhem,

0 = S dT − V dp + N dµ, (1.86)

stating, that the set of intensive variables is not independent. Indeed, the
function p(S, V, Nk ) is homogeneous of order zero and we have

p(S, V, N1 , . . . , Nr ) = p(λS, λV, λN1 , . . . , λNr ). (1.87)

Defining λ = 1/(N1 + N2 + . . . + Nr ) = 1/N , N = total number of particles,

we define
r
Ni X
xi = , xi = 1
N
1

and obtain that

p = p(S/N, V /N, x1 , . . . , xr )
= p(S/N, V /N, x1 , . . . , xr−1 , 1 − x1 − x2 − . . . − xr−1 ). (1.88)

While the equilibrium values of extensive variables of a r-component system

is determined by 2 + r extensive variables, only 1 + r intensive variables
are sufficient to determine the other intensive variables (intensive variables
do not depend on the system size). Correspondingly, the Gibbs-Duhem
equation (1.86) tells that changes in the intensive variables of a system are
interdependent.
An alternative derivation of (1.86) can be done by repeated Legendre
transforms,

U − TS = F
F + pV = G
G − µn = 0
⇒ 0 = SdT − V dp + ndµ

in the variables T, p, µ.
30 CHAPTER 1. THERMODYNAMICS

1.5 Phenomenology of phase transitions

The physics of phase transitions is one of the leading topics in statistical

physics. Here, we focus on their phenomenological aspects and thus quite
naturally will mainly be focusing on first-order transitions. As a compen-
sation, second-order (continuous) transitions will assume a more prominent
role in the statistical mechanics chapters.
In order to introduce the topic, let us consider a simple one-component
gas. In the ideal limit of noninteracting atoms (the ideal gas), we expect
to find only one state or aggregate, the gas phase. However, if the atoms
interact, actually attractive on large distances and repulsive on short ones,
see Fig. 1.14, one may expect that particles go into a state of optimal distance
r0 when the temperature drops below the energy scale V0 , which is a high
density phase. This state may be a liquid (with no order and no rigidity)
or a solid (with long-range translational and orientational order, as well as
rigidity und shear).

Fig. 1.14: Typical interaction be-

V
tween neutral atoms, attractive on
long distances due to the Van der
Waals interaction and repulsive on
r0 r short distances due to Pauli repul-
sion.
−V0

Indeed, all these phases, gas, liquid, and solid do exist under appropri-
ate conditions and transform into one another in so-called phase transitions.
These phases and transitions then are conveniently represented in so-called
phase diagrams. While we discuss here the phenomenological aspects of
phase transitions, their ‘microscopic’ properties and understanding will be
one of the central topics of statistical mechanics. To fix ideas, we will con-
sider a system that can be characterized by the parameters p and V , T and
S, and µ and N (or mole number n)—other systems may be characterized
by alternative pairs, e.g., a magnetic system is described by H and M rather
than p and V (although a piece of magnetic material requires both pairs for
its complete description, e.g., in magnetostriction phenomena the volume V
is non-trivially related to M ).

1.5.1 Gibbs’ phase rule

A thermodynamic system may manifest itself in different states or phases

which appear in isolation or coexist—going to a multicomponent system,
matters become even more complicated. The Gibbs’ phase rule brings some
1.5. PHENOMENOLOGY OF PHASE TRANSITIONS 31

order into this zoo of possibilities by outlining the basic structure a system’s
phase diagram can assume. Here, we discuss a specific situation where
the pressure p, temperature T , and amount of substances (in moles) ni ,
i ∈ {1, . . . , r} are prescribed—the situation where other ‘external’ parame-
ters are implosed can be handled in a similar way. The appropriate thermo-
(α)
dynamic potential for these variables is G and the chemical potentials µi
of the i-th substance in the α-th phase, α ∈ {1, . . . , ν},
(α) (α) (α)

µi p, T, x1 , . . . , xr−1 , (1.89)

are the associated equilibrium parameters; following the Gibbs-Duhem law,

they depend themselves only on the 2 + r − 1 intensive variables p, T , and
(α) (α) (α)
the relative amount xi of substance i in the phase α, xi = ni /n(α) with
Pr (α)
i=1 xi = 1. Equilibrium then imposes the r(ν − 1) conditions

Fig. 1.15: Hypothetical phase dia-

p critical point gram with several phases α, β, and
γ and transitions. A critical point
terminates one of the critical lines
β and tripple points mark locations of
α γ three-phase coexistence. (Question:
is this geomtry compatible with the
triple point concavity of S?)
T

(α) (α) (α)
(β) (β) (β)

µi p, T, x1 , . . . , xr−1 = µi p, T, x1 , . . . , xr−1 (1.90)

on the total of 2 + ν(r − 1) intensive variables (2 from p, T , and ν sets of

r − 1 relative mole numbers). Hence, there remain f = 2 + ν(r − 1) − r(ν − 1)
degrees of freedom, defining the Gibbs’ phase rule in the form

f = 2 + r − ν. (1.91)

How do we make use of this rule? Consider a one-component system with

r = 1 that is present in a unique phase α, i.e., ν = 1, e.g., a gas of atoms
at high temperature and low pressure, hence α = gas. The Gibbs’ rule then
tells that f = 2, meaning that this phase can be realized at all values of p
and T , i.e., in a two-dimensional region, an area, of the p–T phase diagram.
Asking for a coexistence of this gas phase with another, say liquid, phase β,
we have ν = 2 and such coexistence can be realized only along a line in the
p–T diagram, where

µ(α) (p, T ) = µ(β) (p, T ) → pαβ (T ). (1.92)

32 CHAPTER 1. THERMODYNAMICS

Finally, requiring coexistence of three phases α, β, γ, we can find only a point

in the p–T phase diagram where this is possible. Indeed, the three lines

µ(α) (p, T ) = µ(β) (p, T ) → pαβ (T ),

µ(β) (p, T ) = µ(γ) (p, T ) → pβγ (T ),
µ(γ) (p, T ) = µ(α) (p, T ) → pγα (T ), (1.93)

cross in a single point, the triple point pαβγ , Tαβγ . According to Gibbs’
phase rule there is no possibility for the coexistence of four phases in the
one-component system with only two intensive variable p and T . A phase
diagram may have a form as sketched in Fig. 1.15. The critical endpoint of
the α-β phase line can be understood when studying the condition (1.92) in
more detail.

1.5.2 Gibbs surfaces

Every phase α defines a surface µ(α) (p, T ) (the Gibbs surface) in µ–p–T
space, see Fig. 1.16. Consider the situation (e.g., at appropriate pressure p

µ µ

β β
C

p p
1.
pαβ(T) pαβ(T)
α α
T T 2.

Fig. 1.16: Left: Gibbs surface for two phases α and β. According to the
thermodynamic minimization principle, it is the phase with a lower energy
that will be realised at given p and T . The phases transform into one
another at the crossing line between the surfaces that defines the transition
line pαβ (T ) when projected into the p–T plane. The sudden change of slope
at the crossing of the minimal surfaces implies jumps in the conjugated
variables of p and T , the volume V = ∂p G and the entropy S = −∂T G.
These jumps are related by Clapeyron’s equation. Right: Gibbs surface
describing the situation with a second-order transition: the line where the
two Gibbs surfaces cut at an angle, thus defining a first-order transition,
terminates in a critical point C describing a second-order phase transition.

and temperature T near a transition line pαβ (T )) where two phases α and
β are relevant, see Fig. 1.16. In equilibrium, the Gibbs energy

G(p, T ) = n(α) µ(α) (p, T ) + n(β) µ(β) (p, T ) (1.94)

1.5. PHENOMENOLOGY OF PHASE TRANSITIONS 33

is always minimal, i.e., for p > pαβ phase β is realized and we have n(α) = 0,
while for p < pαβ the phase α has lower energy and hence n(β) = 0. In
general, the Gibbs surfaces cut under a finite angle such that the extensive
variables V (conjugate to p) und S (conjugate to T ; here, we use molar
quantities s and v) undergo a jump at the phase transformation,

∆v(T ) = v (α) pαβ (T ), T − v (β) pαβ (T ), T

∂µ(α) ∂µ(β)
= − 6= 0, (1.95)
∂p T ∂p T
∆s(T ) = s(α) pαβ (T ), T − s(β) pαβ (T ), T

∂µ(α) ∂µ(β)
= − + 6= 0.
∂T pαβ ∂T pαβ

The derivatives of G are geometrically related, see Fig. 1.16, implying a

corresponding thermodynamic relation between the jumps in the conjugate
variables: Following the cut defined by Eq. (1.92) and using dµ = −sdT +vdp
with the definitions of the jumps (1.95), we obtain the equation of Clausius
and Clapeyron (cf. (1.74))

−s(α) dT + v (α) dp = −s(β) dT + v (β) dp

dpαβ ∆s(T ) ∆S(T )
⇒ = = . (1.96)
dT ∆v(T ) ∆V (T )

During the transition at fixed pressure p and temperature T , we have the

transformation of a low-entropy phase with a small volume, e.g., a liquid,
into a high-entropy phase occupying a large volume, e.g., a gas. During this
phase transformation the latent heat

T ∆s = ` (1.97)

has to be provided. This latent heat T dS modifies both the system’s internal
energy dU as well as its volume dV (as the system provides work under
expansion), T dS = dU + p dV . The CC (Clausius-Clapeyron)-equation
then reads
dpαβ `
T = . (1.98)
dT ∆v
Measuring ` and ∆v at fixed p and T for different values of T allows us to
find the transition line pαβ through simple integration.

1.5.3 Ehrenfest classification of phase transitions

A phase transition corresponds to a non-analyticity in the system’s thermo-

dynamic potential. Such a non-analyticity can only appear in the thermody-
34 CHAPTER 1. THERMODYNAMICS

namic limit (i.e., in a system with infinitely many degrees of freedom5 ) and
usually manifests itself through jumps in derivatives. A system described
with a Gibbs potential that has jumps in the first derivatives at the tran-
sition is undergoing a first-order phase transition, see the above example.
A phase transition is called to be of n-th order if the jumps appear in the
n-th derivative. Other (modern) terminology distinguishes between transi-
tions with a latent heat (first-order, with a jump in the order parameter)
and second-order transitions (smooth development of the order parameter,
with divergent susceptibiliy and correlation length), also sometimes called
continuous transitions. But beware, that there are also infinite-order tran-
sitions (that do not break a symmetry and have no order parameter), the
most well known example being the Kosterlitz-Thouless transition; some-
times, ‘continuous’ is referring to an infinite-order transition. One way to
arrive at a second-order transition is via the termination of a first-order
line where two Gibbs surfaces cut, see Fig. 1.16. The point C is called a
critical point or critical end-point. Note that the typical characteristics of a
second-order transition (e.g., the critical exponents) near a critical end-point
manifest themselves along the continuation of the first-order phase line, in
our example the continuation of pαβ (T ).

1.5.4 Phase diagrams

Often, it is convenient to study other diagrams, different from the p–T phase-
diagram, in order to understand and characterize a system. Alternative
diagrams are those providing the pressure p or temperature T as a function
of the volume V , see Fig. 1.17.

p p
C T
α p
pαβ
pαβ T pαβ(T)
β

T v v (α ) v (β ) v v (α ) v (β ) v

Fig. 1.17: Phase diagrams describing a first-order phase transition. The p–

T phase diagram (left) provides an overview on the phases and transitions.
The jump in the (molar) volume v manifests itself as line segments in the
p–v (middle) and T –v diagrams (right).

With V the conjugate variable to p, the p–V and T –V diagrams exhibit

5
See W. Ketterle, N.J. vanDruten, Bose-Einstein condensation of a finite number of
particles trapped in one or three dimensions, Phys. Rev. A 54, 656 (1996), for finite-size
effects in a BEC.
1.5. PHENOMENOLOGY OF PHASE TRANSITIONS 35

features associated with the jump in volume at the transition. Hence, while
the p–T diagram tells everything about the existing phases and transitions,
replacing p by V to arrive at a T –V diagram adds information on the jump
in volume at fixed p, but no longer provides a survey of phases.

p p
T T
C l C
l s K g
K
g T 3−phase
coexistence
v v

Fig. 1.18: Surfaces of thermodynamic equilibrium states in p-v-T space for

a one-component two- and a three- phase system (gas, liquid, solid). The
ruled surfaces K define locations for phase coexistence. The projections into
the p, T -plane of the ruled surfaces define the phase diagram. Phases cannot
be distinguished beyond the critical points C.

The three diagrams in Fig. 1.17 are nothing but the projections of the
surface of thermodynamic states in p-v-T space describing a one-component
system with two phases α and β (e.g., gas and liquid). A typical example is
shown in Fig. 1.18, together with the corresponding surface for a three-phase
system including a third phase γ (e.g., a solid phase). The ruled surfaces K
are made from straight lines (projecting onto individual points in the p–T
phase diagram) and describe regions of phase coexistence; their projection
into the p, T plane defines the transition lines, e.g., pαβ (T ) in Fig. 1.17. The
phases α and β cannot be distiguished from each other beyond the critical
point C. The phase diagram for the three-phase system is shown in Fig.
1.19 and includes a triple-point T .

Fig. 1.19: Phase diagram for

p the three-phase system with solid
l C (s), liquid (l), and gas (g) phases.
s C is a critical point beyond which
gas and liquid phases cannot be
distinguished. T is the triple-
g
T point where all three phases co-
exist.
T

Having understood the various diagrams, we return to the thermody-

36 CHAPTER 1. THERMODYNAMICS

namic potentials that actually generate the latter. We have started from
the ‘good’ variables p and T for which the Gibbs potential generates the
p–T phase diagram. When going to the variables v and T one may ask what
characteristic form the associated thermodynamic potential f (V, T ) actually
has in the presence of a transition—this is the subject of the next section.

1.5.5 Free energies

Let us consider the free energy (per mole) f (v, T ). The phases α and β define
the free energies f (α) and f (β) . The (convex) isotherm f (v, T = const) must
have the form as shown in Fig. 1.20, with the tangents t(α) = t(β) = t in v (α)
and in v (β) connecting the states A and B following from the relation
∂f
p=− (1.99)
∂v T

and the equilibrium condition

pαβ (T ) = p(α) T, v (α) = p(β) T, v (β) .

(1.100)

Fig. 1.20: Isotherm f (v, T =

f f (α)
T = const const) of the free energy f (v, T ) in
the vicinity of a first-order phase-
α’ transition with a jump in volume v.
A
t β’
( β)
f
B

v (α) v (β) v

The Maxwell-construction defining the joint tangent in A and B is equiv-

alent to the equilibrium condition on the chemical potentials µ in Eq. (1.92),

f (α) T, v (α) − f (β) T, v (β)

= −pαβ (1.101)
v (α) − v (β)
⇒ µ(α) = (f + pv)(α) = (f + pv)(β) = µ(β) .

Along the tangent, the free energy assumes the linear form

v (β) (T ) − v
f (t) (T, v) = f (α) T, v (α) (t)


(1.102)
v (β) (T ) − v (α) (T )

v − v (α) (T )
+f (β) T, v (β) (T ) (β)
v (T ) − v (α) (T )
1.5. PHENOMENOLOGY OF PHASE TRANSITIONS 37

describing the two-phase coexistence of phases α and β. The phase-mixture

in the coexistence region is infinitely compressible as the tangent t has no
curvature,
∂2f
= 0 ⇒ κT = ∞.
∂v 2 T
Applying (virtual) pressure pushes the dilute phase (e.g., the gas) into the
dense phase (e.g., the liquid).

f
T = const Fig. 1.21: Two-phase coexistence
α
α & β, metastable states α0 and β 0 ,
α’ A’ unstable and unstable regime with negative
A
B’ compressibility.
α&β β’
β
B

v (α) v (β) v

The branches α0 und β 0 describe metastable states of the phases α and β,

see Fig. 1.21; these metastable states correspond to the high-energy Gibbs
surfaces in Fig. 1.16. In the example of a gas–liquid system these correspond
to undercooled gas and overheated liquid phases. In the points A0 and B 0
the curvatures
∂2f ∂p 1
2
=− = (1.103)
∂v T ∂v T κT v
change their sign, the compressibility turns negative and the phases are
rendered unstable. The states A0 and B 0 define the so-called spinodals in
the p, v-diagram, see Fig. 1.22.

p T C metastable Fig. 1.22: The stable phases α

β − phase
α
and β in the p-v diagram, become
metastable (α0 and β 0 ) at the phase
β
α+β transition and turn unstable at the
T spinodal lines.
α’ β’
spinodal v

1.5.6 Constraining parameters in phase transitions

It is interesting to return from the free energy f (v, T ) in Fig. 1.20 back
to the Gibbs energy g(p, T ) in the slightly different formulation involving
38 CHAPTER 1. THERMODYNAMICS

a constraint parameter. To fix ideas, we use the language pertinent to

a gas–liquid first-order transition and start with the free energy in Fig.
1.23 including the straight segment describing the gas–liquid coexistence
regime as well as the segment (dashed) describing metastable and unstable
phases. The two phases are distinct from one another through their densities
1/v = ρ, v = V /N = the volume per particle.

T
Fig. 1.23: Free energy at fixed tem-
f perature T . The transition occurs
at the pressure p = plg (T ) and cor-
responds, via ∂v f = −plg (T ), to the
straight segment in f .

0 vli vgas v

In returning to the Gibbs free energy µ = f + pv, we keep the volume (or
density) as a constraint parameter and plot µ(p, T ; v) = f (v, T )+p v at fixed
p and T as a function of v in Fig. 1.24. The minima of the constraint Gibbs

metastable
µ p > plg
gas phase
stable
liquid
phase
p = plg (T )

first−order
p < p lg
phase transition
at p = p lg
constraining
0 −1 parameter
ρ−1
liquid ρ−1
gas ρ=v
for µ
metastable liquid stable
phase gas phase

Fig. 1.24: Gibbs free energy (chemical potential) as a function of the

constraining parameter v = ρ−1 . The three curves describe the evolution of
µ when changing the pressure p (the conjugate variable to the constraining
v. The minimization ∂v µ(p, T ; v) = ∂v f (v, T ) + p = 0 is consistent with the
the equation of state p = −∂v f deriving from the thermodynamic potential
f (v, T ). Hence minima in µ(p, T ; v) in the constraining parameter v define
equilibrium states. Other points along the lines do not describe equilibrium
states.
1.5. PHENOMENOLOGY OF PHASE TRANSITIONS 39

energy µ(p, T ; v) as a function of v then provide thermodynamic equilibrium

states, as
∂µ ∂f
= + p = −p + p = 0, (1.104)
∂v ∂v
where we have used that p = −∂v f , see Eq. (1.72). Hence, the two relations
p = −∂v f and ∂v µ = 0 are equivalent. At the phase transition p = plg (T ),
the two chemical potentials should be the same, hence the curve µ(p, T ; v)
plottet versus v shows two degenerate minima, see Fig. 1.24. Increasing p >
plg (T ) the liquid minimum is the stable one and the gas becomes metastable,
and vice versa for p < plg (T ). Increasing p to eliminate the metastable
minimum describing the gas, we reach the spinodal where the gas becomes
unstable (and vice versa when decreasing p to eliminate the metastable liquid
minimum).

1.5.7 Entropy surfaces

Above, we have started with describing a system’s phase transition in its

intensive variables p and T and then replaced one variable, the pressure p,
by its dual, the volume v. Obviously, we can replace the other variable
and repeat the above discussion, or replace both (all) variables by extensive
ones—it is this situation which we are going to discuss now. The resulting
potential in the extensive variables dual to T and p is u(s, v) or s(u, v).
We focus here on the latter, the entropy potential s(u, v) as a function of
its extensive variables. The latter all jump in a first-order phase transition
and the resulting concave entropy surface s(u, v) consists of extremal points
describing pure phases and simplices describing phase mixtures. Thereby,
the two-phase mixtures α, β define ruled surfaces involving straight segments
α-β, while the three-phase mixtures α, β, γ is described by a three-simplex
α-β-γ within the entropy surface, see Fig. 1.25. These simplices transform
under the Legendre transformation

s→u→f →µ

(for the transform u → f via s → T and for f → µ via v → p) to become

the phase lines pαβ , pβγ , pγα and the triple point pαβγ , Tαβγ , see Fig. 1.26.
Note, that in transforming s → µ, we lose the information about the distri-
bution of phases during the transition6 , but all thermodynamic information
is preserved.
6
The relative amount of the phases in the mixture α : β follows from the point within
the simplex; all points in a simplex collaps to a line pαβ (T ), T or a point pαβγ , Tαβγ ;
the information about the relative abundance of phases in the mixture is lost under the
transformation s → µ.
40 CHAPTER 1. THERMODYNAMICS

l C
s
µ
g
T s
l
C

s C
T g

l C
u T g p s l
s g
T
v T

Fig. 1.25: Entropy surface Fig. 1.26: Gibbs energy surface

s(u, v) with extremal points µ(p, T ), twice Legendre trans-
(pure phases), ruled surfaces formed with respect to s(u, v).
α-β (transition lines), and The simplices transform to lines
three-simplex α-β-γ (triple along which the surface µ(p, T )
point) describing pure and co- changes slope abruptly (kinks).
existing phases. The intensive The extensive variables result
variables are given by the deriva- from the derivatives ∂T µ|p = −s,
tives ∂v s|u = p/T , ∂u s|v = 1/T . ∂p µ|T = v; the latter jump at the
phase boundaries.

1.5.8 Nucleation in first-order transitions

The constraint Gibbs free energy can be taken as an inspiration to describe

the process of the transition. Tuning the system, e.g., from the gas phase at
p < plg (T ), to the transition point p = plg (T ) and even a little bit beyond,
the system will stay in the gas phase as it has to overcome a barrier in
order to transform to the liquid. The gas then transforms into a liquid
via a nucleation process, where a liquid droplet is formed as a fluctuation.
This droplet will either grow and take the gas into the liquid, or evaporate
again and a new fluctuation is needed. Our goal then is to find the optimal
droplet (with the smallest barrier) that is needed to successfully trigger the
transition. The discussion below is kept in the language of the gas–liquid
transition but is straightforwardly generalized to other situations.
Let an amount δngas (in moles) of gas condense into an isotropic liquid
droplet of radius r (note that ρ = N/V denotes the usual particle density),

δngas = −δnli = (4π/3NA )r3 ρli ,

where NA denotes Avogadro’s number and ρ = N/V . At p > plg (T ) this

process is favored by a gain δngas (µli −µgas ) < 0 in free energy and hampered
by the energy cost 4πr2 σlg to produce a gas–liquid interface, where σlg is
1.5. PHENOMENOLOGY OF PHASE TRANSITIONS 41

the surface tension (note that such an interface has an intermediate density
ρgas < ρint < ρli and hence costs more energy, see Fig. 1.24). Hence, the
droplet growth involves the competition between a volume gain and a surface
cost, totalling to a Gibbs free energy
4π 3
G(r) = r ρli (µli − µgas ) + 4πr2 σlg . (1.105)
3NA
Depending on the pressure p, the function G(r) changes its behavior, see
Fig. 1.27: At small radii r, the positive quadratic term always dominates.

p < plg Fig. 1.27: Gibbs free energy of a

G
p = p lg
liquid droplet of radius r. The cu-
bic term changes sign at p = plg ,
leading to a maximum energy at a
p > p lg finite critical radius rc . Overcom-
ing the barrier G(rc ) at rc , the sys-
r tems runs away (the droplet grows)
0 rc
to transform into a bulk liquid.

When increasing the pressure across plg (T ), the cubic term changes sign and
produces a maximum at the critical radius rc as obtained from the condition
δG = 0, 7

δG = δ(nli µli + ngas µgas + 4πr2 σlg )

 2σlg NA 
= δnli µli − µgas + = 0, (1.106)
r ρli
resulting in the critical radius (for µli < µgas )
2σlg NA 1
rc = . (1.107)
ρli µgas − µli

In order to obtain the critical radius as a function of pressure p > plg (T ),

we take the derivative of
2σlg NA
µgas − µli =
ρli rc
with respect to p and use ∂p µ = vmol = NA /ρ. As a result, we find that
1 1  1 ∂rc 1 ∂ρli 
− = 2σlg − −
ρgas ρli ρli rc2 ∂p rc ρ2li ∂p
7
We make use of the relations
4π 3 ρli ∂nli ρli
δngas = −δnli , nli = r → = 4πr2 .
3 NA ∂r NA
42 CHAPTER 1. THERMODYNAMICS

and using the liquid’s properties of small volume vli  vgas = RT /p and
poor compressibility ∂p ρli ≈ 0, we arrive at the differential equation

∂rc kB T ρli rc2 1 kB T ρli

=− −→ ∂ = ∂ ln p. (1.108)
∂p 2σ p rc 2σ
(1.109)

Making use of the boundary condition p(rc = ∞) = plg , we obtain the final
results (see Fig. 1.28)
 2σ v 1  2σ vli
li
p = plg exp and rc (p, T ) = . (1.110)
RT rc RT ln(p/plg )

rc Fig. 1.28: Size of critical nucleus

versus pressure p. As p approaches
the phase boundary plg from above
the droplet radius diverges.

p (T ) p
lg

In order to trigger the phase transition at a pressure p > plg , a thermal

fluctuation must generate a liquid droplet of radius r > rc —according to Fig.
1.27 the Gibbs free energy then decreases with growing radius, ∂r G|r>rc < 0
and the droplet expands to transform all the gas into the liquid. At given
conditions p, T , such fluctuations will be realized with a rate

Γ ≈ ωmikro e−G(rc )/kB T . (1.111)

The microscopic attempt rate ωmikro is difficult to estimate; a rough number

is given by the collision time between gas particles.
The above discussion can be generalized to numerous systems, e.g., the
flipping of magnetic domains when changing the externally applied field
H or the motion of elastic membranes in periodic or random potentials.
Furthermore, the ideas can be extended to quantum phase transitions at T =
0, where the Gibbs free energy G/kB T has to be replaced by the (euclidean)
action SE T /~. A famous example in this regard is the nucleation of the
Helium B-phase in the A-phase (Lifshitz, Kagan) sketched in Fig. 1.29.
1.5. PHENOMENOLOGY OF PHASE TRANSITIONS 43

3
He −A Fig. 1.29: Nucleation of a 3 He-B
droplet in 3 He-A as an example of
a quantum phase transition. Simi-
3
lar ideas of ‘false vacuum’ have been
He −B formulated about our universe (see
S. Coleman)
44 CHAPTER 1. THERMODYNAMICS
Chapter 2

Statistical Description

We will now proceed with the microscopic derivation and extension of the
results from thermodynamics. To this end, we will consider a many-body
system which can be described by the coordinates

(p, q) ≡ {p1 , . . . , p3N , q1 , . . . , q3N } ∈ Γ (2.1)

with states (p, q), elements of the phase-space Γ, dim Γ = 6N , N denot-

ing the particle number. The dynamics of the system is described by the
Hamiltonian H(p, q) through

∂H ∂H
q̇i = , ṗi = − .
∂pi ∂qi

In an experiment, the system is investigated by measuring an observable

M(p, q) for a time T and one is interested in the time-averaged value
Z T
T 1
M = lim dt M[p(t), q(t)], (2.2)
T →∞ T 0

where [p(t), q(t)] describes a trajectory in Γ. For a system in thermodynamic

equilibrium, (2.2) will give a finite result. In theory, the time-averaged
T
observable M is an inaccessible quantity. On the other hand, within a
computational-physics approach, modern computers can time evolve sys-
tems of ∼ 109 particles for a few million time steps in molecular dynamics
simulations. With typical1 scattering times around 10−12 s, we obtain a
time interval of 10−6 s — this is quite an respectable result and, quite of-
ten, much can be learned from such molecular dynamics (MD) simulations.
1
The collision time in classical gases is typically 10−10 seconds, the same for electron-
electron scattering in a room temperature metal, while the electron-impurity scattering
time in a metal is typically 10−14 seconds. In a semiconductor, typical collision times of
electrons and holes are in the range of 10−13 seconds to 10−12 seconds.

45
46 CHAPTER 2. STATISTICAL DESCRIPTION

While, these computations still fall short of describing systems with 1023
particles which can be experimentally observed over milliseconds to days,
such huge simulations are often not required and smaller systems/times will
do. Another important aspect, however, is that such computations produce
vast quantities of data that do not really interest us in their great detail,
but from which we will only extract certain specific characteristics of the
system.
In the statistical mechanics approach, the time-average (2.2) is replaced
by the ensemble-average
R 3N 3N
d p d q M(p, q) ρ(p, q)
hM i = R (2.3)
d3N p d3N q ρ(p, q)
as the quantity of interest. Here the density-function
ρ(p, q)d3N p d3N q (2.4)
describes the probability that the system is in the volume d3N p d3N q around
the point (p, q) in phase-space Γ.
While the phase-space density ρ(p, q, t) is a time-dependent function in
general, its total derivative dρ/dt vanishes by Liouville’s theorem: A change
of density ∂ρ/∂t in the volume ω ⊂ Γ can only be due to a flow of density
through the boundary ∂ω,
Z Z
∂ρ
dω = − d~σ · (ṗ, q̇)ρ
ω⊂Γ ∂t
Z∂ω
Gauss
= − dω (∇p , ∇q ) · [(ṗ, q̇)ρ], ∀ω
ω
∂ρ
→ = −(∇p , ∇q ) · [(ṗ, q̇)ρ] (2.5)
∂t
3N
X ∂(ṗi ρ) ∂(q̇i ρ)
= − −
∂pi ∂qi
i=1
3N
X ∂ρ ∂H ∂ρ ∂H
= − = {H, ρ}, (2.6)
∂pi ∂qi ∂qi ∂pi
i=1

where we have used (above, {f, g} denotes the Poisson bracket of f and g)
∂ ṗi ∂ q̇i
= 0, = 0,
∂pi ∂qi
∂H ∂H
= −ṗi , = q̇i .
∂qi ∂pi
According to Eq. (2.5), the material derivative of ρ
dρ
[∂t + (ṗ, q̇) · (∇p , ∇q )]ρ = =0 (2.7)
dt
 47

vanishes; this is Liouville’s theorem. However, note that the partial deriva-
tive ∂t ρ 6= 0 does not vanish in general2 .
In statistical mechanics, we are not interested in an arbitrary state of
the system but specifically in the thermodynamic equilibrium state. For an
isolated system in equilibrium with the volume V , the particle number N ,
and the energy E fixed, the principle of equal a-priori probabilities provides
us with the density-function ρ(p, q) in the form
(
const, E < H(p, q) < E + ∆,
ρ(p, q) = (2.8)
0, otherwise,

i.e., ρ is uniform over all states (p, q) which are compatible with the con-
straints. As ρ(p, q) = ρ(H(p, q)), we immediately have ∂t ρ = 0 from (2.6),
i.e., ρ is time-independent in thermodynamic equilibrium.
A set of systems prepared with identical macroscopic parameters (for
example V , N , E fixed) is called an ensemble. Such a (thermodynamically
equilibrated) ensemble is associated with a corresponding density function
ρ(p, q); the specific ensemble characterized by fixed V , N , E and described
by the density ρ(p, q) as given in (2.8) is called the microcanonical ensemble.
The fundamental idea of statistical mechanics then is that the time-
average and ensemble-average of an observable M coincide,
T
M = hMi, (2.9)
time-average = ensemble-average.

It is easy to convince ourselves that (2.9) is true if the individual system

prepared in the lab explores all of phase space Γ compatible with the con-
straints N , V , E without preference, i.e., the trajectories [p(t), q(t)] cover
ΓN,V,E homogeneously. This ergodic hypothesis is nontrivial. Establishing
which systems (do not) fulfil this criterion is an area of active research. E.g.,
ergodicity is broken in the physics of glasses3 .
The expression (2.9) is linked to the concept of self-averaging: We could
convincingly argue about the validity of (2.9) if instead of one system we
would measure a large number of equally (i.e., at fixed E, V, N ) prepared
systems,
T 1 X T
M = M ⇒ hMi. (2.10)
Number of systems many systems
2
One can then interpret the density function ρ as the density of an incompressible fluid
in Γ-space.
3
When the system freezes into the glass state infinite barriers appear for T → 0.
These divide the phase space Γ into subspaces from which the system can only escape on
hierarchically larger (infinite for T → 0) timescales, i.e., the system becomes trapped in a
subspace of Γ.
48 CHAPTER 2. STATISTICAL DESCRIPTION

The concept of self-averaging then tells us that we may consider a system of

1023 particles
P in V as a set of statistically independent systems in subvolumes
δVi , V = i δVi . Similarly, we may split the time interval T into smaller
intervals δTi and thus create a large number of statistically independent
experiments. The average over time measurements of these subsystems in
space and time then mimick an ensemble average which is equal to the time
average of the original system. It is important to assure that the volumes
δV and intervals δT are larger than the system’s correlations in time and
space, δT  τ , δV  ξ 3 , τ = correlation time, ξ = correlation length.
Starting from the above concepts, we can develop statistical physics
along several directions:

1. We derive thermodynamics from the a-priori Ansatz for ρ = ρmicro in

the microcanonical ensemble and then go to new ensembles by coupling
the system to reservoirs (T -reservoirs allowing the system’s energy
E to fluctuate; µ-reservoirs allowing for a fluctuating N ). The new
ensembles are the canonical and grand canonical ones with ρcan ∼
exp[−H(p, q)/T ] and ρgcan ∼ exp[−(H − µN )/T ], respectively.

2. We generalise the classical ideas to quantum mechanical systems

X
ρ(p, q) → density operator ρ = wi |Ψi ihΨi |, (2.11)
i

where wi are the probabilities to find the system in the quantum state
|Ψi i. The average then is the combination of a classical average with
probabilities wi and a quantum average over the wave functions or
amplitudes Ψi .

3. We simplify the statistical description and analyse the physics away

from equilibrium. Typical non-equilibrium phenomena involve trans-
port (e.g., Boltzmann transport equation), relaxation, response (e.g.,
Kubo formula), stochastic processes (e.g., Langevin equations for Brow-
nian motion), both in classical and quantum frameworks.

In this lecture, we will discuss the aspects 1 and 2 in detail, while the
non-equilibrium aspects described in 3 will be left for another advanced
course.
Chapter 3

Classical statistical
mechanics

We will now derive thermodynamics from the various ensembles. We recall

the concept of the 6N -dimensional phase space Γ whose points (p, q) =
(p1 , . . . , p3N , q1 , . . . , q3N ) describe an instantaneous state of the system. We
also recall the density function ρ(p, q) defined on Γ that provides the weight
in the ensemble-average of an observable M,

d3N p d3N q M(p, q)ρ(p, q)

R
hMi = R , (3.1)
d3N p d3N q ρ(p, q)

which under the assumption of ergodicity is equal to the measured time-

T
average M . The density function ρ depends on the ensemble. We will
now discuss the three relevant ensembles for statistical mechanics (E is the
energy, V the volume, and N denotes the particle number):

E, V, N fixed, ‘isolated’ system, microcanonical ensemble

V, N fixed, ‘closed’ system, system in contact with a thermal reservoir at

temperature T , canonical ensemble.

V fixed, ‘open’ system, system in contact with a thermal reservoir at tem-

perature T and a particle reservoir with a chemical potential µ, grand
canonical ensemble.

49
50 CHAPTER 3. CLASSICAL STATISTICAL MECHANICS

3.1 Microcanonical ensemble

We base statistical mechanics on the postulate of equal a-priori probabilities

ρ(p, q) for every state in Γ constraint to the fixed parameters E, V, N ,
(
1/h3N N !, E < H(p, q) < E + ∆,
ρ(p, q) = (3.2)
0, otherwise.

The constant h has the dimension of an action and vanishes from the ex-
pectation hMi in Eq. (3.1). The factor N ! resolves the Gibbs paradox aris-
ing in the context of the mixing of indistinguishable particles. The factor
N ! makes sure that all permutations of the state (pπ(1) , . . . , pπ(N ) , qπ(1) , . . . ,
qπ(N ) ), π ∈ S N , are counted only once, resulting in the correct ‘Boltzmann-
counting’. That microscopic particles are truly indistinguishable is a postu-
late of quantum statistics and carries over (in its principle) to the classical
description.
We define the (dimensionless) volume
Z
Γ(E) ≡ d3N p d3N q ρ(p, q)

d3N p d3N q
Z
= 3N
(3.3)
E<H(p,q)<E+∆ N ! h

in phase space Γ appertaining to the microcanonical ensemble. With the

definitions
d3N p d3N q
Z
Σ(E) ≡ 3N
,
H(p,q)<E N ! h
dΣ(E)
ω(E) ≡ , (density of states) (3.4)
dE
we may write Γ as

Γ(E) = Σ(E + ∆) − Σ(E) ≈ ω(E)∆. (3.5)

Making use of the volume Γ defined in Eq. (3.3), we introduce the expression
S(E, V, N ),
S(E, V, N ) = kB log Γ(E), (3.6)
where kB = 1.380·10−16 erg/K = (1/11605) eV/K and 1 eV = 1.602·10−19 J.
The expression S is (see the proof below)

− extensive in E, V, N
− maximal in an isolated system, i.e., (3.7)
S(E, V, N ) > S(E, V, N ; constrained).
3.1. MICROCANONICAL ENSEMBLE 51

Given the statements in (3.7), we can identify S(E, V, N ) with the system’s
entropy. This provides us with a microscopic interpretation of entropy: it
counts the number of available states in Γ-space and its maximization directs
the isolated system towards a homogeneous distribution of states compatible
with the fixed variables E, V, N .

S is extensive: We divide the system E, V, N into two parts E1 , V1 , N1

and E2 , V2 , N2 (the interaction between the systems scales as V 2/3 and is
thus neglected in the following considerations). The isolated systems 1 and
2 with energies (E1 , E1 + ∆) and (E2 , E2 + ∆) have entropies

S(E1 , V1 , N1 ) = kB log Γ(E1 ),

S(E2 , V2 , N2 ) = kB log Γ(E2 ), (3.8)

where Γ(Ei ) is the volume in Γ occupied by 1, 2 see Fig. 3.1.

E E2
E p2
p1 E 1+ Δ
E1

0
Δ << E
E 2 +Δ 0
q1 q2

Fig. 3.1: Splitting of the volume Γ in phase space.

The constrained (in E, V , and N ) system occupies the phase space

volume
Γ(E1 )Γ(E2 ). (3.9)
Next, we relax the constraint on E: The total energy E can be distributed
arbitrarily over the subsystems, with the energies E, E1 , E2 related through

E < E1 + E2 < E + 2∆. (3.10)

The phase-space volume Γ(E) is the sum of all possible distributions of the
energy E1 ≡ E1i = i∆ and E2 = E − E1 , i = 0, . . . , E/∆,

E/∆
X
Γ(E) = Γ(E1i )Γ(E − E1i ) and thus
i
E/∆
X
S(E, V, N ) = kB log Γ(E1i )Γ(E − E1i ). (3.11)
i
52 CHAPTER 3. CLASSICAL STATISTICAL MECHANICS

Consider the largest contribution to the sum in (3.11) and let it be

Γ(Ē1 )Γ(Ē2 ) with Ē1 + Ē2 = E.

We then have
E
Γ(Ē1 )Γ(Ē2 ) ≤ Γ(E) ≤ Γ(Ē1 )Γ(Ē2 ). (3.12)
∆
As log Γ(Ēi ) ∝ Ni , log E/∆ ∼ log N , we immediately find that

S(E, V, N ) = S(Ē1 , V1 , N1 ) + S(Ē2 , V2 , N2 ) + O(log N ). (3.13)

Let us next investigate how we split the system when choosing the divi-
sion E = Ē1 + Ē2 . As the energies Ē1 and Ē2 are the main contributors to
the entropy, the variation

δ[Γ(E1 )Γ(E2 )] =0 (3.14)

Ē1 ,Ē2

vanishes and we find an equilibrium parameter (we write Γi = Γ(Ei ))

δΓ1 δΓ2
= − , and using δE1 = −δE2
Γ1 Γ2
∂ ∂
log Γ1 = log Γ2
∂E1 ∂E2
1 ∂S1 ∂S2 1
→ ≡ = ≡ . (3.15)
T1 ∂E1 Ē1 ∂E2 Ē2 T2

Hence, the derivative of S with respect to E defines an equilibrium parameter

that we identify with the (inverse) temperature T ,

1 ∂S
= . (3.16)
T ∂E V,N

Combining the results (3.13) and (3.15), we find that separating our
system into two subsystems while keeping them in equilibrium, our function
S is additive and hence extensive, up to small log corrections. Note that the
result in (3.13) also implies that the most likely (equlibrium) distribution of
energies is E = Ē1 + Ē2 : nearly all configurations have the optimal energy
distribution, i.e., the one belonging to the largest phase space volume.

S is maximal: By removing a constraint, the phase-space volume in-

creases and therefore S(E, V, N ) > S(E, V, N ; constrained). Note that re-
moving a constraint dividing E, V, N into two parts that are out of equilib-
rium will lead to an equilibration that homogenises the temperature (and
3.1. MICROCANONICAL ENSEMBLE 53

other equilibrium parameters) and that will entail a rise in entropy. Hence,
a division of the system into two parts that remain in equilibrum implies
extensiveness, S = S1 + S2 , while division into two parts out of equilibrium
implies a rise in entropy, S > S1 + S2 . 2
Repeating the above steps and relaxing the constraints in V and N leads
us to the additional equilibrium parameters p (pressure) and µ (chemical
potential),
∂S ∂S
p=T , µ = −T . (3.17)
∂V E,N ∂N E,V
With the obvious identification of U ≡ E and using the definition of entropy
through Eq. (3.6), we can identify S expressed as a function of the extensive
variables E, V, N as a thermodynamic potential
S(U ≡ E, V, N ) = kB log Γ(E). (3.18)

Reexpressing the function S(U, V, N ) in the form U (S, V, N ) provides us

with the internal energy U as a function of the variables S, V , N and the
thermal (p = . . . , µ = . . . ) and the caloric (eliminate S in U (S, V, N ) using
T = . . . ) equations of state are obtained from the derivatives with respect
to the conjugate variables,
∂U
p = − ,
∂V S,N
∂U
µ = ,
∂N S,V
∂U
T = .
∂S V,N

Hence, the calculation of the phase-space volume Γ(E) (or Σ(E), ω(E))
provides us with the entire thermodynamics of an isolated system that is
characterised by the extensive variables E, V, N . Note that log Γ ∼ log Σ ∼
log ω. We can go over to open systems by using the Legendre transformations
F = U − T S, G = U − T S + pV .

Example: The ideal gas

With the Hamiltonian for a free gas of particles given by

N
1 X 2
H= pi = E, N, V = const, (3.19)
2m
i=1
we have to calculate the phase-space integral
Z Z
1 3 3
Σ(E) = d q1 . . . d qN d3 p1 . . . d3 pN . (3.20)
N ! h3N
| {z } H<E
VN
54 CHAPTER 3. CLASSICAL STATISTICAL MECHANICS
qP √
The integral over the 3N -dimensional sphere of radius 2 2mE ≡
i pi =
R yields the volume1
π n/2
S 3N (R) = Rn (3.21)
Γ(n/2 + 1) n=3N

and we obtain the result

1 VN π 3N/2
Σ(E) = 3N
(2mE)3N/2 . (3.22)
N ! h Γ(3N/2 + 1)
Making use of Stirling’s formula
√
Γ(1 + z) = z z e−z 2πz(1 + r(z)), 0 < r(z) < 1/12z + . . . , (3.23)

we find the entropy of an ideal gas (Sackur-Tetrode),

S(E, V, N ) = kB log Σ
h V  4πmE 3/2 5 i
= N kB log + log + . (3.24)
N 3N h2 2
Here, the prefactor is extensive ∝ N and the terms in parenthesis are inten-
sive ∝ V /N (from N !) and ∝ E/N (from Γ(3N/2 + 1)). The internal energy

3h2  N 2/3 h2 S 5i
U (S, V, N ) = N exp − (3.25)
4πm V 3 N kB 3
is obtained by inversion and the definition of the temperature T via the
derivative of U , providing us with the caloric equation of state,
∂U 2 U 3
T = = , U = N kB T = E. (3.26)
∂S V 3 N kB 2
Making use of the specific heat
∂U 3
CV = = N kB , (3.27)
∂T 2
we recover the thermodynamic expression for S(V, T )
V
S = N kB log + CV log T + const. (3.28)
N
1 z=n
The integer and half-integer values of the Γ-function derive from Γ(z + 1) = zΓ(z) =
√
n! and Γ(1/2) = π and thus
√ √
π π ππ 4π 3
S1 = √ R = 2R, S 2 = R2 , S 3 = √ R3 = R .
π/2 1 (3/4) π 3
Exercise: Calculate the formula for the surface and the volume of an n-dimensional sphere
by using the integration of a Gaussian in Cartesian and polar coordinates. Show that the
volume of a high-dimensional R ∞sphere concentrates itself on the surface. Calculate the
Stirling formula from N ! = 0 dx eS(x) , S(x) = N log x − x and evaluate the integral by
the method of steepest descent.
3.1. MICROCANONICAL ENSEMBLE 55

in an improved form with V /V0 → V /N under the logarithm. Taking the

derivative of U with respect to V , we obtain the pressure p and the thermo-
dynamic equation of state2 follows from
∂U 2U N kB T
p=− = = . (3.29)
∂V S 3V V
This completes the microscopic derivation of the thermodynamic equations
of state for the ideal gas.

Equipartition theorem

As our second application of statistical mechanics formulated in the micro-

canonical ensemble, we will derive the principle of equipartition; it tells us
that the expectation value of the product xi ∂H/∂xj is given by
D ∂H E
xi = δij kB T. (3.30)
∂xj
P 2
For the ideal gas with H = pi /2m and using xi = xj = pi we obtain
D p2 E k T
i B
= , (3.31)
2m 2
i.e., in equilibrium every motional degree of freedom carries the energy
kB T /2. Thus the internal energy of the ideal gas becomes
3 3
hHi = U = N kB T, CV = kB N, (3.32)
2 2
as expected from (3.26) and (3.27).
For a system of harmonic oscillators H = i ai p2i + bi qi2 , we obtain
P

XD ∂H ∂H E ai ,bi 6=0 X
pi + qi = h2Hi = 2 kB T. (3.33)
∂pi ∂qi
i i

As a result, every kinetic and potential degree of freedom carries the energy
kB T /2.

Proof:

Using the simple manipulation

∂(H − E)
Z Z
∂H
dp dq xi = dp dq xi
H<E ∂xj ∂xj
ZH<E Z
∂
= dp dq (xi (H − E)) −δij dp dq (H − E),
H<E ∂xj H<E
| {z }
=0, boundary term where H−E=0
2
for a relativistic gas p = U/3V .
56 CHAPTER 3. CLASSICAL STATISTICAL MECHANICS

we find
D ∂H E Z Z
1 ∂H ∆ ∂ ∂H
xi = dp dq xi = dp dq xi
∂xj Γ E<H<E+∆ ∂xj Γ ∂E H<E ∂xj
Z Z
δij ∂ δij δij
= dp dq (E − H) = dp dq = Σ
ω ∂E H<E ω H<E ω
1 kB
= δij = δij = δij kB T.
∂E log Σ ∂E S

Using Hamilton’s equation ṗi = −∂qi H, we immediately obtain the virial

theorem,3

3N D 3N
X ∂H E DX E
− qi = qi ṗi = −3N kB T. (3.34)
∂qi
i=1
| i=1{z }
virial from classical mechanics

A few words of caution concerning the principle of equipartition: One may

be tempted to argue that CV for harmonic system is always trivially CV =
(f /2)kB N , where f is the number of degrees of freedom, both kinetic and
potential, per particle. This is only true for classical systems where every
degree of freedom may be excited by arbitrarily small energies. In quantum
mechanics a finite energy δE is often required to excite a degree of freedom
(for example δE = ~ω for the harmonic oscillator). Only when kB T > δE
does the degree of freedom contribute fully with kB T /2 to the specific heat.
The quantum mechanical weight with which a degree of freedom with finite
energy ~ω > 0 contributes, e.g., to the specific heat or the total energy, is
T →0 k T >~ω
∝ ~ω coth(~ω/2kB T ) → ~ω and B → 2kB T .

3.2 Canonical ensemble

We consider a system (E1 , V1 , N1 ) in contact with a large system (E2 , V2 ,

N2 )  (E1 , V1 , N1 ) which we will interpret as a reservoir. We consider the
combined system to be isolated with energy E < (E1 + E2 ) < E + 2∆. With
the phase space volume belonging to the reservoir denoted by Γ(E2 ), the
probability ρ(p1 , q1 ) that system 1 is found in the state (p1 , q1 ) is given by

ρ(p1 , q1 ) ∝ Γ(E − E1 ). (3.35)

3
The virial theorem from mechanics tells us that the average value of the total kinetic
energy hT P
i in a bound system is linked to the total potential energy hV i, more precisely,
2hT i = − N ~ ri i, with F
~i the force acting on the particles.
i hFi · ~
3.2. CANONICAL ENSEMBLE 57

From the previous discussion of the extensiveness of S, see (3.13), it follows

that only a single energy E1 = Ē1  E is relevant and thus we can expand

Γ(E − E1 ) = elog Γ(E−E1 ) = eS(E−E1 )/kB

−(E1 /kB ) ∂E S
≈ eS(E)/kB e E≈E2

= const e−E1 /kB T2 , (3.36)

where T2 is the temperature of the reservoir. With E1 = H(p1 , q1 ), we can

define the probability density of the canonical ensemble as
1
ρ(p, q) = e−H(p,q)/kB T . (3.37)
h3N N !
We define the canonical partition function ZN as
Z
ZN (V, T ) = d3N p d3N q ρ(p, q)
Z 3N 3N  
d pd q H(p, q)
= exp − , (3.38)
N ! h3N kB T

and the quantity4

F (T, V, N ) = −kB T log ZN (V, T ). (3.39)

The quantity F is extensive5 and relates to the internal energy U and the
entropy S via (proof follows)

F = U − T S. (3.40)

Thus, we interpret the expression F = −kB T log ZN as the free energy of

the system.
The result (3.40) follows from the normalization (we define β = 1/kB T )
Z 3N 3N
1 d p d q −βH
1 = e ,
ZN N ! h3N
Z 3N 3N
−βF d p d q −β H
ZN = e = e ,
N ! h3N
and differentiation of this condition with respect to β,
Z 3N 3N
1 d p d q −β H  
3N
e F − H + β∂β F = 0
ZN h N!
∂F ∂F
→ hHi = F + β and thus U = F − T . (3.41)
∂β ∂T
4
Note that we dropped the constraint E1 < E in (3.38) because large energies H(p, q)
give vanishing contributions to ZN .
5
ZN1 +N2 (V1 + V2 ; T ) ≈ ZN1 (V1 , T ) ZN2 (V2 , T )
58 CHAPTER 3. CLASSICAL STATISTICAL MECHANICS

The quantity F (T, V, N ) thus obeys the differential equation (3.41) which
is the one defining the energy potential U (S, V, N ) from the free energy
potential F (T, V, N ) via the Legendre transformation; using −S = ∂T F , we
have U = F + T S = F − T ∂T F .
For a canonical ensemble, we then can recover all of the system’s ther-
modynamic behaviour by calculating its partition function

e−β H(p,q)
Z
ZN (T, V ) = d3N p d3N q
N ! h3N

and the corresponding free energy

F (T, V, N ) = −kB T log ZN (T, V ), (3.42)

(3.43)

with the derivatives providing the equations of state via

∂F
S = − ,
∂T V,N
U = F + T S, caloric equation,
∂F
p = − , thermodynamic equation,
∂V T,N
∂F
µ = , thermodynamic equation.
∂N T,V

3.3 Grand canonical ensemble

We consider a system V1 , N1 coupled to a particle reservoir (V2 , N2 ) 

(V1 , N1 ) (with the combined system coupled to a thermal reservoir at tem-
perature T ). The partition function

d3N p d3N q −βH(p,q,N )

Z
ZN (V, T ) = e (3.44)
N ! h3N
3.3. GRAND CANONICAL ENSEMBLE 59

may be written as a product over the subsystems where N1 = N − N2

particles are present in the considered volume V1 (this is our system)6
N Z
1 X N!
ZN = d3N1 p1 d3N1 q1 e−βH(p1 ,q1 ,N1 )
N ! h3N N1 ! N2 !
N1 =0
Z
× d3N2 p2 d3N2 q2 e−βH(p2 ,q2 ,N2 )
N Z
d3N1 p1 d3N1 q1 −βH(p1 ,q1 ,N1 ) d3N2 p2 d3N2 q2 −βH(p2 ,q2 ,N2 )
X Z
= e e
N1 ! h3N1 N2 ! h3N2
N1 =0
N Z
X
≡ d3N1 p1 d3N1 q1 ρ(p1 , q1 , N1 )ZN , (3.45)
N1 =0

with the density function

e−βH(p1 ,q1 ,N1 ) d3N2 p2 d3N2 q2 −βH(p2 ,q2 ,N2 )
Z
1
ρ(p1 , q1 , N1 ) = e
ZN (T, V ) N1 ! h3N1 N2 ! h3N2
| {z }
ZN2 (T,V2 )

ZN2 (T, V2 ) e−βH(p1 ,q1 ,N1 )

= (3.46)
ZN (T, V ) N1 ! h3N1
and the normalization
N Z
X
d3N1 p1 d3N1 q1 ρ(p1 , q1 , N1 ) = 1. (3.47)
N1 =0

Using the equations of state for the pressure ∂V F = −p and the chemical
potential ∂N F = µ that characterize the reservoir T, V2 , N2 , it follows from
(3.42) that
ZN2 (T, V2 )
= e−β[F (T,V −V1 ,N −N1 )−F (T,V,N )]
ZN (T, V )
≈ e−β[pV1 −µN1 ] . (3.48)

We introduce the fugacity

z = eβµ (3.49)
as our new independent variable for µ and rewrite Eq. (3.46) as
z N1
ρ(p1 , q1 , N1 ) = e−βpV1 −β H(p1 ,q1 )
N1 ! h3N1
e−β [H−µN1 +pV1 ]
= . (3.50)
N1 ! h3N1
6
The factor N !/N1 !N2 ! counts the number of ways the particles can distribute between
the volumes V1 and V2 .
60 CHAPTER 3. CLASSICAL STATISTICAL MECHANICS

We now assume that N, V → ∞ (hence, N1  N can assume any value)

and we drop the index 1 in Eq. (3.50) (hence N1 → N and V1 → V ) to
arrive at the density function of the grand canonical ensemble,
e−β [H−µN +pV ]
ρ(p, q, N ) = . (3.51)
N ! h3N
Next, we define the grand canonical partition function (again in the limit of
N, V → ∞ and V1 , N1 replaced by V, N )
∞
X
Z(T, V, z) = z N ZN (T, V ) (3.52)
N =0
∞ Z
X d3N p d3N q −β [H(p,q)−µN ]
= e
N ! h3N
N =0

and using the density function (3.51) and its normalization (3.47), we find
the grand potential 7
−Ω = pV = kB T log Z(T, V, z). (3.53)
The grand canonical partition function directly gives us the thermal equation
of state p = p(T, V, µ). We find the caloric equation of state by calculating
the average energy8
XZ
U = hHi = d3N p d3N q H(p, q)ρ(p, q, N )
N
d3N p d3N q
X Z
= e|−βpV zN H(p, q)e−βH(p,q) ,
{z } N ! h3N
1/Z N | {z }
−∂β ZN (T,V )
∂
U = − log Z(β, V, z) . (3.54)
∂β V,z

Similarly, we find the particle number

N z N ZN
P
∂
N = hN i = PN N =z log Z(β, V, z) (= −∂µ Ω). (3.55)
N z ZN ∂z β,V

Differentiating with respect to V trivially provides us with the thermal equa-

tion of state
∂
p = kB T log Z(β, V, z) (= −∂V Ω). (3.56)
∂V β,z
7
From the normalization, we have that
XZ X NZ d3N p d3N q −β H
1= dp dq ρ = e−pV /kB T z 3N
e = e−pV /kB T Z.
N N
N ! h

8
Note that here, we view z as an independent variable for µ and not as a function of
β and µ, i.e., ∂β z = 0.
3.4. FLUCTUATIONS* 61

All potentials can be reconstructed from (3.54–3.56), for example

F = U − TS from U according to (3.54)

Z T
CV ∂U
and S = dT , CV = .
0 T ∂T V

3.4 Fluctuations*

In the canonical ensemble the energy E is not fixed, and similar is the case for
N in the grand canonical ensemble. Let us then determine the fluctuations in
energy (E) in the canonical ensemble and of the particle number (N ) in the
grand canonical ensemble. We show that √ the fluctuations in the extensive
quantities E and N are small, of order N ,

hH 2 i − hHi2 = kB T 2 CV , hHi ∼ N, CV ∼ N,
N2 1 ∂V
hN 2 i − hN i2 = kB T κT , κT = − . (3.57)
V V ∂p T

3.4.1 Energy fluctuations in the canonical ensemble

Starting with the mean energy

d3N p d3N q
Z
1
U = hHi = He−βH
ZN N ! h3N

we can rewrite this equation with the help of ZN = e−βF

d3N p d3N q
Z
[U − H]e−β(H−F ) = 0.
N ! h3N

Taking the derivative with respect to β provides us with the expression

d3N p d3N q
Z
∂U h i
+ [U − H] F + β∂β F −H e−β(H−F ) = 0,
∂β N ! h3N | {z }
=hHi=U

from which we obtain the desired result for the energy fluctuations

∂U
h(U − H)2 i = − . (3.58)
∂β

Rewriting the derivative

∂U ∂U
= −kB T 2 = −kB T 2 CV
∂β ∂T
62 CHAPTER 3. CLASSICAL STATISTICAL MECHANICS

immediately gives us the result (3.57a). Using the extensivities√ hHi ∼ N

and CV ∼ N , we find that the relative fluctuations vanish as 1/ N for large
particle numbers N ,
∆H (hH 2 i − hHi2 )1/2 1 N →∞
≡ ∼√ → 0. (3.59)
H hHi N

We consider an alternative derivation of this result which provides us

with additional insights on the roles of energy versus entropy. Let us group
together all states with energy E which contribute to the canonical partition
function with the weight of the microcanonical partition sum ω(E). We thus
rewrite the canonical partition function in the form
Z 3N 3N Z ∞
d p d q −β H
ZN = e = dE ω(E)e−βE
N ! h3N 0
Z ∞
= dE e−βE+log ω(E)
0
F (E)
Z ∞  z }| {
= dE exp −β(E − T S(E)) , (3.60)
0
where the Hamiltonian H assumes the value E in the corresponding en-
ergy shell. When expressed through the free (F ) and spectral free (F(E))
energies, this equation reads
Z ∞
−βF
e = dE e−βF (E) .
0

The integral in (3.60) is dominated by the minimum in the exponent F(E) =

E − T S. We can thus expand around the minimum9

1 ∂2F
F(E) ' F(Ē) + (E − Ē)2 ,
2 ∂E 2 Ē
∂F ∂S 1
= 1−T = 0 → ∂E S|E=Ē = ,
∂E Ē ∂E Ē T
we write ∂E S = T̃ (E)−1 , T̃ (Ē) = T,
∂2F ∂2S
= −T (3.61)
∂E 2
Ē ∂E 2 Ē
∂ ∂S 1 ∂ T̃ 1
= −T = = .
∂E ∂E T ∂E T CV
Ē=U 1
→ F(E) ' [U − T S(U )] + (E − U )2 . (3.62)
2T CV
9
For the ideal gas, we can write F(E) = N [e − (3/2)kB T log e] + rest with e = E/N , see
Eq. (3.24). This spectral free energy has a minimum at ē = (3/2)kB T as trivially obtained
from ∂E F(E) = [1 − (3/2)kB T /e]ē = 0.
3.4. FLUCTUATIONS* 63

The partition function ZN can thus be expressed as a Gaussian integral

 
Z ∞ h (E − U )2 i
ZN ≈ exp −β(U − T S(U )) dE exp − (3.63)
| {z } 0 2kB T 2 CV
F̄ √
| {z }
2πkB T 2 CV
p
≈ 2πkB T 2 CV e−β(U −T S) = e−βF .

Hence, we find that the distributions of internal

√ energies E in the canonical
ensemble is given by a Gaussian of width kB T 2 CV around U , see Fig. 3.2.
When comparing the entropy in the canonical and microcanonical ensembles,

Fig. 3.2: Fluctuations of the in-

e−β
F
ternal energy E around U . The
width
√ of the (Gaussian) distribution
is kB T 2 CV .

U E

we find that the results agree up to an (irrelevant) logarithmic correction,

1
F = U − T Smicrocan − kB T log CV ,
2
1
→ Scan = Smicrocan + log CV . (3.64)
2 | {z }
O(log N )

Thus both ensembles generate the same thermodynamics. This calculation

also highlights the statement that among all the energies E1 = E − Eres
only one value Ē1 = U = hHi is relevant—(3.59) and (3.63) show that the
fluctuations around this energy are small. This result can be traced back to
the behaviour of the integrand in ZN , which is a product of the density of
states
elog ω(E) ∼ eαE

and the Boltzmann factor

e−βH ∼ e−βE .

Indeed, when E ∝ N → ∞, one obtains a sharply peaked function as

illustrated in Fig. 3.63.
64 CHAPTER 3. CLASSICAL STATISTICAL MECHANICS

e−β H ω (E ) Fig. 3.3: Sharply peaked inte-

grand resulting from a rapid growth
in the phase space volume ω(E)
and a rapid decay of the Boltzmann
weight.

0 U E

3.4.2 Density fluctuations in the grand canonical ensemble

The particle number fluctuations ∆N appear from (3.52) (also see (3.55))
and are given by
∂ ∂
hN 2 i − hN i2 = z z log Z(T, V, z)
∂z ∂z | {z }
pV /kT
∂ 1 ∂
∂z
= βz ∂µ V 1 ∂2p ∂2p
= = V k B T . (3.65)
kB T β 2 ∂µ2 ∂µ2
We rewrite ∂µ2 p by using the homogeneity of F

F (T, V, N ) = N f (T, v = V /N ). (3.66)

With
∂F ∂f
p = − =− ,
∂V ∂v
∂F ∂f
µ = =f −v ,
∂N ∂v
we find that ∂µ2 p is proportional to the compressibility κT ,10

∂2p ∂ ∂p
=
∂µ2 ∂µ ∂µ
1 ∂v 1 1 1 κT
= − 2 = 3 2 =− 3 = 2
v ∂µ v ∂v f v ∂v p v
and we obtain the result (3.57b)

(∆N )2 = hN 2 i − hN i2 = nkB T κT N. (3.67)

10
We use the relations
∂v p = −∂v2 f, ∂µ v = (∂v f − ∂v f − v∂v2 f )−1 = −1/v∂v2 f,
∂µ p = ∂v p∂µ v = 1/v.
3.4. FLUCTUATIONS* 65

Note that both fluctuations ∆E and ∆N are given by their respective linear
response coefficients11 CV and κT ; this property is related to the fluctuation–
dissipation theorem, see Landau-Lifschitz.
Since the fluctuations in N are small, we can identify N = hN i = N̄ in
(3.52) as the dominant term and obtain

Z ≈ z N̄ ZN̄ = e−β(F −µN̄ ) (3.68)

with the canonical free energy F (T, V, N̄ ) in the exponent. The result (3.68)
provides us with a useful relation between the grand canonical partition
function Z and the free energy12
F (T, V, N̄ ) = kB T [N̄ log z − log Z(T, V, z)], (3.69)
where z is obtained from N̄ = z∂z log Z(T, V, z). Also, we can find the
optimal particle number N̄ by repeating the analysis leading to the optimal
energy Ē in the previous discussion of energy fluctuations in the canonical
ensemble,
canonical Z ∞ Z ∞
ZN = dE e−β[E−T S(E)] = dE e−βF (E)
0 0
∂S
→ Ē = U from min[E − T S] → 1 − T = 0,
∂E Ē
grand canonical
X X
ZN = z N ZN = e−β[F (T,V,N )−µN ]
N N
∂F
→ N̄ from min[F − µN ] → − µ = 0. (3.70)
∂N N̄

The chemical potential in the grand canonical ensemble plays the role of
a Lagrange multiplier H → H − µN that guarantees the conservation
of particle numbers. If particles may spontaneously be created then the
difference between particle and antiparticle numbers is usually conserved
NT − NAT = const. Correspondingly the expression µN is replaced by
µ(NT − NAT ) in the formula for the grand canonical partition function
X
Z= ZNT ZNAT eβµ(NT −NAT ) . (3.71)
NT ,NAT

The definition (3.71) is relevant in the context of baryon-antibaryon- (QCD),

electron-positron- (QED), electron-hole- (semiconductor), vortex-antivortex-
(superconductors, XY -spin systems in 2D), etc. generation of particles/exci-
tations through thermal fluctuations.
11
linear response: δU = CV δT , δN = −(V /v 2 )δv = N κT δp = (N 2 /V )κT δµ.
12
This corresponds to the Legendre transformation, −kB T log Z + kB T N̄ log z = −pV +
kB T N̄ µβ = µN̄ − pV = G − pV = F .
66 CHAPTER 3. CLASSICAL STATISTICAL MECHANICS

Finally, we note that according to (3.57) the fluctuations in energy and

density become large when CV , κT → ∞. In particular, at the first-order
gas-liquid phase transition we have ∂v p = 0 and κT = ∞, see chapter 2.
This divergence in κT shows that there are large fluctuations in density at
the gas-liquid phase transition. One can show that a large range of particle
numbers becomes relevant in the grand canonical partition function Z: With
vg = Vgas /N , vli = Vli /N , the weights W (N ) = z N ZN in the grand partition
function Z exhibit a behaviour as sketched in Fig. 3.4.

W gas
liquid & gas
liquid

κT = κT <

8
V/ v g V /v li N

Fig. 3.4: Weights W (N ) contributing to the grand canonical partition

function Z in the pure phases and at the liquid-gas transition at fixed volume
V.
Chapter 4

Quantum statistical
mechanics

We consider a system characterized by a Hilbert space H and the Hamilto-

nian H. The basis {Φn } in H is orthonormal and complete,

hΦn |Φm i = δmn (orthonormal),

X
|Φn ihΦn | = 1l (complete). (4.1)
n

Often one chooses the eigenvectors of H to constitute the basis {Φn }

HΦn = En Φn . (4.2)

An arbitrary state vector Ψ of the system can be expanded in the basis,

X X
Ψ= hΦn |ΨiΦn = cn Φn . (4.3)
n n

4.1 Expectation values in quantum statistics

Expectation values of observables M are given by

∗
P
hΨ|M|Ψi n,m cn cm hΦn |M|Φm i
hMi = = P 2
. (4.4)
hΨ|Ψi n |cn |

In statistical physics we divide our system into a subsystem and a reservoir,

X
Ψ= cn Φn (4.5)
n

67
68 CHAPTER 4. QUANTUM STATISTICAL MECHANICS

where cn are the amplitudes of the reservoir1 and Φn are the states of the
subsystem (we choose a time-independent basis {Φn } of the subsystem). As
we do not know the state of the reservoir, we replace the coefficients c∗n cm
by expectation values hcn |cm ires ; similarly |cn |2 becomes hcn |cn ires . The
quantity measured in an experiment
P is the time average of M. With a time-
independent norm hΨ|Ψi = n hcn |cn ires in the isolated system = system +
reservoir, we have

P T
T m,n hcn |cm ires hΦn |M|Φm i
hMi = . (4.6)
hΨ|Ψi

T
As a first step, we consider the non-diagonal elements hcn |cm ires , n 6= m.
The reservoir dynamics generates amplitudes cn with random phases, such
that the time-averages of matrix elements of differing amplitudes vanish

T
hcn |cm ires = 0 if n 6= m. (4.7)

This is the postulate of random phases. Note that inelastic collisions be-
tween particles are sufficient to randomize the phases.2 As a result of phase
T
randomization, only the diagonal weights hcn |cn ires remain in the average
(4.6). In a next step, we have to make a statistical assumption about the
system in equilibrium, as is usually done with the assumption of molecular
chaos in kinetic theory or the equal a priori probabilities in phase-space
Γ for the microcanonical ensemble (at fixed extensive variables E, V , N )
in classical statistical mechanics. The latter is needed to replace the time-
average by the ensemble-average and brings us to the ergodic hypothesis:
The trajectories of particles explore equally often all parts of phase-space
that are compatible with fixed E, V , and N . Thus, in the present case,
we also define the microcanonical ensemble by giving all states Φn that are
compatible with the constraints of fixed energy E, volume V , and particle

1
Here the reservoir acts as a generator of random phases as opposed to one exchanging
energy or particles. The latter types of reservoirs will be added in the context of the
canonical and grand canonical ensembles.
2
The assumption (4.7) breaks down when the reservoir does not properly randomize
phases. This can happen in small systems when the phase is preserved over distances
that are of similar length or longer than the sample. Also, phase information may be
kept in a small reservoir as is often discussed in the context of quantum thermodynamics.
The important element entering the discussion then is the phenomenon of entanglement
between the system and the reservoir, preventing the system from being factorized.
4.1. EXPECTATION VALUES IN QUANTUM STATISTICS 69

number N an equal a priori probability 3

(
T 1, if E < En < E + ∆,
hcn |cn ires = (4.8)
0, otherwise.

Under these assumptions, the experimentally measured time-average is given

by the ensemble average
P
T wn hΦn |M|Φn i
hMi = n P (4.9)
n wn

with weights (
1, E < En < E + ∆,
wn = (4.10)
0, otherwise.

4.1.1 Density matrix

A suitable formalism to describe the quantum behavior of a mixed state is

known from the quantum mechanics course. Indeed, the quantum mechanics
of a system whose state can be written as an incoherent superposition of
states is described by the density matrix
X
ρ= wn |Φn ihΦn |. (4.11)
n

If wn = δn0 the state is called pure; if more than one weight wn isP
non-zero it
is mixed. Usually, the weights wn are normalized according to n wn = 1.
Here, we consider the normalisation explicitly and write

T Tr(ρM) X
hMi = , Tr(A) = Trace(A) = hΦn |A|Φn i, (4.12)
Tr(ρ) n

for the expectation values. The equivalence of (4.9) and (4.12) is trivial.
The time evolution of ρ is given by the quantum mechanical version of
Liouville’s theorem, the Neumann differential equation

i~∂t ρ = [H, ρ]. (4.13)

3
In Rclassical mechanics, we count the states in phase-space by using the
integral d3N p d3N q/h3N where h is an action (plus Gibbs factor 1/N ! for the correct
‘Boltzmann counting’ motivated by the mixing (entropy) of indistinguishable particles);
in quantum statistical mechanics every state obtains a weight 1. If we take Planck’s con-
stants to be the action h = 2π~ and use the Heisenberg uncertainty principle, we obtain a
weight of 1 for every independent state both classically and quantum mechanically. The
indistinguishability of particles is inherent to quantum mechanics and the Gibbs factor
appears straightforwardly.
70 CHAPTER 4. QUANTUM STATISTICAL MECHANICS

With ρ = ρ(H), see (4.9) and (4.10), we have

[H, ρ] = 0 (4.14)

and the density matrix is time-independent. In the density matrix formalism

the (phase-)reservoir causes the collapse of the density matrix onto its diag-
onal and the ergodicity tells that all diagonal elements that are compatible
with the given parameters E, V , and N are given the same weight.
A practical and often used relation is

Tr(AB) = Tr(BA), (4.15)

implying that Tr is independent of the chosen representation,

Tr(U AU −1 ) = Tr(A). (4.16)

Using the concept of density matrices ρ from the beginning, one may
proceed in a more straightforward way. Consider a system that is iso-
lated but still coupled to the rest of the world (a reservoir) with the total
Hilbert space Hsys ⊗ Hres . Note that this reservoir neither exchanges energy
nor particles with our isolated system, hence E, V , N are fixed (alterna-
tively, we look at a subsystem in a larger system). In quantum mechanics,
the measurement of a system observable M provides us with the expec-
tation value hMi = Tr(ρM). Since we do not know anything about the
rest of the world (or reservoir), we trace it out, hMi = Trsys Trres (ρM) =
Trsys [Trres (ρ)M] ≡ Trsys (ρsys M), since M only acts in Hsys . In tracing ρ
over the reservoir, we have eliminated the entanglement with the rest of the
world, what corresponds to the phase randomization above. Requiring ρsys
to describe a time-independent equilibrium state, we have ∂t ρsys = 0 and
As a result, ρsys is diagonal in the eigenstates |Φn i of
hence [Hsys , ρsys ] = 0. P
Hsys and hence ρsys = i wi |Φn ihΦn | with real coefficients wi (since ρ = ρ† ).
Finally, the a priori assumption of equal probabilities wi = 1/Γ(E) provides
the final result for the properly normalized (i.e., Tr(ρsys ) = 1, cf. Eq. (4.18))
density matrix of the microcanonical ensemble, ρsys = 1l/Γ(E).

4.2 Ensembles in quantum statistics

We summarize the three ensembles from classical statistical physics in the

framework of quantum statistics.4 We choose the eigenstates {Φn } of the
4
Note the normalization
Tr(ρ) = 1 (4.17)
(with the trace in Fock-space) for the density matrix ρ of the grand canonical ensemble.
With Tr(ρ) = exp(−pV /kB T )Z, we obtain the expression for the grand potential Ω =
−pV .
4.2. ENSEMBLES IN QUANTUM STATISTICS 71

Hamiltonian HΦn = En Φn as our basis. The states ΦN

n belong to the Fock-
sector N .

microcanonical ensemble, E, V , N constant:

X
density matrix: ρ = wn |Φn ihΦn |, (4.18)
n
(
1, E < En < E + ∆,
weights: wn =
0, otherwise,
volume: Γ(E) = Tr(ρ) = ω(E) · ∆,
X
density of states: ω(E) = δ(E − En ),
n
thdyn potential: S(E, V, N ) = kB log Γ(E).
Using the normalized density matrix P ρsys from above, we can write
S = −kB Trsys [ρsys log ρsys ] = −kB i wi log(wi ) = kB log Γ.
canonical ensemble, T , V , N constant:
X
density matrix: ρ = e−βH = wn |Φn ihΦn |, (4.19)
n
weights: wn = e−βEn ,
inverse temperature: β = 1/kB T,
partition sum: ZN = Tr(e−βH ),
thdyn potential: F (T, V, N ) = −kB T ln ZN .

grand canonical ensemble, T , V , µ constant:

density matrix: ρ = e−pV /kB T e−β(H−µN ) , (4.20)
−β(H−µN )
partition sum: Z = Tr(e )
X∞
= z N ZN (z = eβµ the fugacity)
N =0
X
−β(H−µN ) N
= hΦN
n |e |Φn i
N,n
with the trace in Fock-space.
thdyn potential: Ω = −pV = −kB T log Z(T, V, z)
or
F (T, V, N ) = kB T [N log z − log Z],
N = z∂z log Z.

In modern condensed matter physics one often encounters areas where

traditional quantum statistical mechanics breaks down: For example, cur-
rents in small metalic/semiconducting rings (persistent currents in an iso-
lated ring) are described by the canonical ensemble (N is constant). A
72 CHAPTER 4. QUANTUM STATISTICAL MECHANICS

further example is given by coherence effect in small (mesoscopic) struc-

tures; here there is no reservoir generating random phases that guarantees
the validity of (4.7).

4.3 The 3rd law

With S = kB log Γ the entropy counts the number of states that are avail-
able. According to ρ ∼ e−H/T only the ground state is available when T = 0
and thus S = 0. If the ground state is g ∼ N -times degenerate, we still have
s = S/N ∼ (log N )/N → 0, and the entropy per particle vanishes. A more
relevant problem is the continuity of the spectrum—the argument above
then is still correct but mainly academic as the ground state is only realized
for T < ∆E ∼ ~2 /mL2 ∼ 5 · 10−15 K (m = mnukleon , L = 1 cm). Thus the
density of states ω(E) is relevant in the low temperature phases of matter.
Usual matter transform into one of the following phases for low temperatures
T → 0: crystals (nearly all) or superfluids/superconductors (3 He, 4 He, met-
als). The density of states in these systems then has a gap (e.g., fermionic
superfluids and superconductors) or vanishes as ω(E) ∼ E d−1 for E → 0
(Goldstone modes) and thus S → 0 for T → 0 (see Section 7.2). The notoric
exceptions are frustrated systems, glasses, two-level systems, etc., showing
a residual entropy as T → 0.
Chapter 5

Applications, formalism, and

techniques

5.1 One particle statistical mechanics

It may sound bizzare to study the statistical mechanics of a single particle;

it is, however, a very useful exercise providing quite some insights and ap-
plications. For example, we can imagine a single particle in a thermal bath
of other particles or fields, where the other degrees of freedom are invoked
to bring the considered particle into equilibrium (with the thermal bath).

5.1.1 Classical free particle

The classical partition function for a free particle follows from (3.38)
Z 3 3
d pd q −p2 /2mkB T  2πmk T 3/2
B
Z1 = e = V
h3 h2
Z
where we have used d3 q = V,
Z ∞
2 √
and dx e−x /2σ = 2πσ. (5.1)
−∞

The constant h is not known.

5.1.2 Quantum mechanical free particle

According to (4.19) the canonical partition function is given by

X
Z1 = Tr(e−βH ) = p |e−βH |~
h~ p i, (5.2)
p
~

73
74 CHAPTER 5. APPLICATIONS AND TECHNIQUES

where |~
p i is a basis and using
2
H = −~2 ∇ ~ /2m, (5.3)
~ √
p i = eik·~x / V ,
h~x |~
2π
p~ = ~~k, ~k = ~n, ni ∈ Z,
L
and V = L3 the quantization volume
X Z  L 3 Z
3
= d n= d3 k
2π
~
n
d3 p −p2 /2mkB T
Z  mk T 3/2
−p2 /2mkB T
X B
→ Z1 = e =V e = V . (5.4)
(2π~)3 2π~2
~
n
The quantum mechanical calculation gives the classical result and fixes
the quantity h in phase-space: In accord with the uncertainty principle,
we should divide µ-space (the Γ space for one particle) into boxes of size
d3 p d3 q = (2π~)3 and count one state per box,
→ class. Z1 = q.m. Z1 if h = 2π~,
~ = the reduced Planck’s constant.
With the thermal de Broglie wavelength
s
2π~2
λ= (5.5)
mkB T
the partition function is
V  L 3
Z1 = = . (5.6)
λ3 λ
The thermal de Broglie wavelength is the wavelength of a particle with
energy E = πkB T ,
~2 k 2 ~2 (2π)2  2π~2 1/2
E= = = πkB T → λ = . (5.7)
2m 2mλ2 mkB T

5.1.3 Classical particle in a potential

p, ~q ) = p2 /2m + V (~q ), then

Let H(~
Z 3 3
d p d q −β[p2 /2m+V (~q )]
Z1 = e
h3
Z 3
V d q −βV (~q )
= e , (5.8)
λ3 V
where the contribution from the (massive) particle dynamics is the factor in
front of the integral and the configurational integral, is independent of the
particle dynamics.
5.1. ONE PARTICLE STATISTICAL MECHANICS 75

5.1.4 Interacting classical particles

As a corollary, we obtain for an ideal non-interacting gas1

1 N 1  V N
ZN = Z1 = (5.9)
N! N ! λ3
(see also Eq. (5.20)), while accounting for an interaction V (|~qi −~qj |) between
the gas particles generates a more complex configurational integral,
Z 3N
1  V N d q h X i
ZN = exp −β V (~
q i − ~
q j ) . (5.10)
N ! λ3 VN
i<j

5.1.5 Quantum particle in a potential

We have to calculate the partition function

2
X
Z1 = p |e−β(p /2m+V (~q )) |~
h~ p i. (5.11)
p
~

Our momentum basis is inpractical −βV (~

q)
R 3 for the calculation of the factor e
and we insert the identity 1 = d q |~q ih~q |. For β → 0, T → ∞ the Baker-
Haussdorf expansion provides us with the approximation
β2
e−β(H0 +V ) ≈ e−βH0 e−βV e− 2
[H0 ,V ]

and to lowest order in β, we can easily handle both H0 and V to arrive at

the ‘classical’ result,
XZ 2
→ Z1 ≈ p |e−βp /2m |~q ih~q |e−βV (~q ) |~
d3 q h~ pi
p
~
XZ d3 q −βV (~q )
Z
V
= d3 q e−βεp h~ p i e−βV (~q ) = 3
p |~q ih~q |~ e ,
| {z } λ V
p
~
1/V

valid for small β → 0 or high temperatures T → ∞.

In the next step, we want to investigate quantum corrections on the
partition function. A systematic expansion in β is tricky and we prefer to
expand in ~. To do this, we use the fact that the partially transformed
density matrix h~p |e−βH |~q i obeys the imaginary time Schrödinger equation,

∂ h ~2 i
− p |e−βH |~q i = −
h~ p |e−βH |~q i.
∆ + V h~ (5.12)
∂β 2m
1
For independent particles Z1 → Z1N and the factor 1/N ! takes care about the Gibbs
paradoxon.
76 CHAPTER 5. APPLICATIONS AND TECHNIQUES

In the limit ~ → 0, the leading (highest) derivative is vanishing, motivating

the use of the WKB-Ansatz (Wenzel-Kramers-Brillouin)
√ ∞  
−βH −(i/~)S
X ~ n
h~
p |e |~q i = e / V, S = Sn .
i
n=0
Equation (5.12) then can be written as
∂S ~h 1 ~ 2 i  ~ 2 1
= (∇S) + V (~q ) − ∆S,
∂β i 2m i 2m
and we find the coefficients Sn
∂β S0 = 0 → e−(i/~)S0 = e−(i/~)S0 (β=0) = e−(i/~)S(β=0)
√
p |e−βH |~q i|β=0 = h~
→ h~ p |~q i = e−(i/~)~p·~q/ V
→ S0 = p~ · ~q,

1 ~ p2
∂β S 1 (∇S0 )2 + V =
= +V
2m 2m
→ S1 = βH(~ p, ~q ),
1 ~ ~ 1 ) − 1 ∆S0 = p~ β ∇V ~
∂β S 2 = (∇S0 )(∇S
m 2m m
β2 ~
→ S2 = p~ · ∇V
2m
1 ~ 1~ ~ 1
∂β S 3 = (∇S1 )2 + ∇S 0 · ∇S2 − ∆S1
2m m 2m
β2 ~ 2 β2 ~ )2 V − β ∆V,
= (∇V ) + 2
p·∇
(~
2m 2m 2m
β3 ~ 2 β3 2
~ )2 V − β ∆V,
→ S3 = (∇V ) + (~
p · ∇
6m 6m2 4m
plus further orders Sn≥4 ∝ ~ in the exponent.2 The partition function
3

becomes (note that the factor h~q |~ p i is compensated by S0 )

XZ
Z1 = p |e−βH |~q ih~q |~
d3 q h~ pi
p
~
X Z d3 q h iX ∞  
~ n i
= exp − Sn
V ~ i
p
~ n=1
X Z d3 q h ~  ~ 2 i
= exp −βH(~ p, ~q ) − S2 − S3 + . . .
V i i
p
~
X Z d3 q n ~  ~ 2 h 1 i o
≈ e−βH 1 − S2 − S3 − S22 + . . .
V i i 2
p
~
 ~ D 1 E 
= Z1cl 1 − hS2 icl + ~2 S3 − S22 + . . . .
i 2 cl
2
Note that, indeed, Sn>0 ∝ β and hence S0 (β = 0) = S(β = 0).
5.1. ONE PARTICLE STATISTICAL MECHANICS 77

For the classical expectation values one easily finds (we only point out some
critical steps)
Z
hS2 icl ∼ d3 p p~ · (. . .) = 0, (integral odd in pi )
Z 3 3
1 d p d q −βH β3
hS3 icl = cl 3
e [. . . + pi pj ∂i ∂j V ] → h∆V i,
Z1 ~ R
6m2 |{z}
pi pj e−βH →δij mkB T

d3 p d3 q −βH β 4
Z
1
hS22 icl = e pi pj (∂i V )(∂j V ) → h(∇V )2 i,
Z1cl ~3 R
4m 2 |{z}
pi pj e−βH →δij mkB T

and using
Z Z Z Z
−βH −βH ~ −βH ~
e ~
∆V = − ∇e ~
∇V = β ∇V e ∇V = e−βH β(∇V
~ )2 ,

we can reduce all correction terms to a form ∝ h∆V i. Summarising, we find

Z1 to be
 2
β2 β2 1 β2
 
cl 2 β
Z1 ≈ Z1 1 + ~ h∆V i + h∆V i − h∆V i − h∆V i
6m 6m 4m 2 4m
 β 2 ~2 
= Z1cl 1 − h∆V i (5.13)
24m
and exponentiating,
 β 2 ~2   β 2 ~2  D h β 2 ~2 iE
1− h∆V i ≈ exp − h∆V i ≈ exp − ∆V ,
24m 24m 24m
we can recast (5.13) into
Z 3
V d q  h β~2 i
Z1 ≈ 3 exp −β V (~q ) + ∆V (~q ) (5.14)
λ V 24m

with the quantum correction of order (λ/l)2 V with l ≈ V /|∇V~ | the typical
3
scale of variations in the potential V . Indeed, the condition λ/l  1 of
smooth potential is the standard requirement for the validity of the quasi-
classical (WKB) approximation. Hence, we see that the configurational
integral is modified due to an additional quantum smearing of the potential.
For the free energy F1 = −kB T ln Z1 , we obtain the quantum corrected
version
~2 β 2 ~ 2
F1 = F1cl + h(∇V ) icl + O(~3 ),
24m
where we have again used h∆V i = βh(∇V ~ )2 i. Note that we could perform
2
the expansion in ~ systematically. The expansion in β (for example via
3
We assume that there is only one scale in V (q) such that ∂q2 V /V ∼ (∂q V /V )2 ∼ 1/`2 .
78 CHAPTER 5. APPLICATIONS AND TECHNIQUES

Baker-Haussdorf) is much more tricky, because we can transform pi pj factors

~ V )2 factors in (1/β)∆V such that higher order terms in
in δij m/β and (∇
β can be brought into lower powers of β by using these relations.
Clearly, the further expansion in ~ will become increasingly more diffi-
cult. One therefore preferably starts from the eigenvalue problem

HΦn = En Φn

and constructs the partition function from

X X
Z1 = hΦn |e−βH |Φn i = e−βEn .
n n

Elastic manifolds

A similar analysis can be done for (elastic) manifolds instead of particles.

The Hamiltonian H = p2 /2m + V (~q ) generalises to
Z h1 C ~ 2 i
H= dd q Π2 + (∇u) + V (u(~q )) (5.15)
2ρ 2

with the mass density ρ and the elasticity C. This Hamiltonian derives from
the Lagrangian
Z h ρ  ∂u 2 C i
L = dd ~q ~ 2 − V [u(~q )]
− (∇u) (5.16)
2 ∂t 2
∂L ∂u
Π = =ρ ,
∂u,t ∂t
h = Πu̇ − `

comprising kinetic energy density, elastic energy density, and the potential.4
The displacement field ~u(~q ) can be vectorial in general. For d = 1 at V = 0
(V 6= 0) (5.16) describes a free (bound) string.5
4
The variation produces the equations of motion in Lagrangian or Hamiltonian form,

∂` ∂` ∂`
0 = ∂t + ∂i − ,
∂(∂t u) ∂(∂i u) ∂u
∂h ∂h
−Π̇ = − ∂i ,
∂u ∂(∂i u)
∂h
u̇ = .
∂Π

5
Note the symmetry in the dynamic and elastic
p terms for the “Euclidean” Lagrangian
LE = T + V (compare to (5.60)): t → t0 = t c/ρ gives a d + 1-dimensional isotropic
problem.
5.2. GASES 79

5.2 Gases

The topic of quantum gases has attracted huge interest during recent years.
The ability to laser- and evaporatively cool atoms to amazingly low tem-
peratures (µK and nK) using optical traps and the development of suitable
(magneto-optic) traps allowed Eric Cornell and Carl Wiemann at JILA and
Wolfgang Ketterle, Kendall David, and Marc-Oliver Mewes at MIT to pro-
duce a Bose-Einstein condensate in 1995; in 2001 Cornell, Wiemann, and
Ketterle received the physics Nobel prize for this achievement. At the be-
ginning of the millenium, the cooling of fermionic atoms into a quantum-
degenerate state with kB T  εF followed. Today quantum optics is an ever
blooming field. The focus is on interesting systems described by compli-
cated Hamiltonians that generate non-trivial physics. The quantum gases
are then used as ‘quantum simulators’ for such Hamiltonians. Elements
of such Hamiltonians include non-trivial periodic lattices (→ simulations
of graphene, topological insulators assisted by optical lattices), interactions
(tunable by Feshbach resonances), interactions mediated by gauge fields
(generated by structured laser fields), disorder, Bose-Fermi mixtures, mul-
ticomponent gases, etc.
Note the difference between quantum gases and strongly correlated quan-
tum gases, i.e., gases with strong interactions, and quantum liquids that
commonly interact strongly and have a high density. Typical examples
of quantum liquids are fermionic 3 He and bosonic 4 He. The subject of
fermionic quantum liquids brings one to the Landau theory of Fermi-liquids
(encompassing the electron gas/liquid in a metal), that is renormalized due
to the strong interactions but otherwise behaves phenomenologically like a
fermionic quantum gas.
In what follows, we first discuss the statistics of classical and quantum
gases before we study Fermi and Bose gases more thouroughly in the follow-
ing two chapters, respectively.

5.2.1 The classical ideal gas

For later comparison, we first start with the classical or so called “Boltz-
mann” gas. We calculate6 the partition function in the canonical ensemble

6
We discretize phase-space and work with discrete momenta p
~i .
80 CHAPTER 5. APPLICATIONS AND TECHNIQUES

P
−β np~i εpi
X
ZN = e p
~i
· gc , (5.17)
|{z}
{np~i }N configurations
X
np~i = # particles with momentum p~i , np~i = N,
p
~i

εp i = p2i /2m.

The correct Boltzmann counting accounts for the number of possible ways
to distribute the N particles among the momenta p~0 , . . . , p~i , . . .,7
N!
g̃c = .
np~0 !np~1 ! . . . np~i ! . . .

In addition, we have to take into account a Gibbs factor 1/N ! in order to

avoid the Gibbs’ paradox, hence,

g̃c → gc = g̃c /N !.

We exchange the sum and product by using the multinomial theorem, 8

X Y e−βnp~i εpi 1 X −βεp N

ZN = = e i , (5.18)
np~i ! N!
{np~i }N i i

and then replace the sum over momenta by integrals, counting one state per
phase-space volume h3 → (2π~)3 (alternatively replace d3 n → (L/2π)3 d3 k →
V /(2π~)3 d3 p)

d3 p p2 
Z
X
−βεpi
 V
e =V exp − = 3,
(2π~)3 2mkB T λ
p
~i

where s
2π~2
λ= (5.19)
mkB T
7
In the correct ‘Boltzmann counting’, we first treat count configurations with the num-
ber g̃c > 1 of their appearance and then account for the ‘indistinguishability’ of particles
with the Gibbs factor 1/N !. In quantum mechanics, indistinguishability is accounted for
by the fermionic or bosonic statistics.
8
The multinomial theorem says that
!
X k
(x1 + x2 + · · · + xn )k = xk1 xk2 · · · xknn ,
k1 , . . . , k n 1 2
k1 +···+kn =k

where !
k k!
= .
k1 , . . . , k n k1 ! · · · kn !
5.2. GASES 81

is the thermal de Broglie wavelength. The partition function, the thermo-

dynamic potential, the caloric equation of state, and the thermodynamic
equation of state then become (we use Stirling’s formula for N !)

1  V N
ZN = , (5.20)
N ! λ3
V h
3
i
log ZN = N log + N = N − log(nλ ) + 1 ,
N λ3 h i
F = −kB T log ZN = N kB T log(nλ3 ) − 1 ,
∂F
S = − = −F/T + (3/2)N kB = (−F + U )/T,
∂T V
∂F
p = − = N kB T /V,
∂V T

with the gas parameter δ ≡ nλ3 . For an ideal classical gas δ  1, i.e.,
the gas is dilute (small density n) or hot (large temperature T , small λ).
Using U = (3/2)N kB T , the above results produce the Sackur-Tedrode result
(3.24)
 for the entropywith E replaced by T and h by 2π~, S(T, V, N ) =
N kB − log(nλ3 ) + 5/2 .

5.2.2 Ideal quantum gases

The quantum gases are easiest to describe in the grand canonical ensemble
(this removes the constraint on the particle number N , i.e., all values of N
are admitted)

∞
X
Z(V, T, z) = z N ZN (V, T )
N =0
X∞ X  X 
= z N exp −β np~ εp
N =0 {np~ }N p
~
| {z
}
np
~
Q −βεp
p
~ ze
XX X h
np~
np~ i
= ... ... ze−βεp0 0 . . . ze−βεpk k ...
np~0 np~1 np~k
YhX
n i
= ze−βεp . (5.21)
p
~ n

For bosons, the state p~ can be populated with n = 0, 1, . . . , k, . . . , ∞ parti-

cles. This has to be contrasted with the fermions, where only occupations
n = 0, 1 are allowed (Pauli exclusion principle, every state is occupied at
82 CHAPTER 5. APPLICATIONS AND TECHNIQUES

most once). Thus (5.21) simplifies to the product form

(Q
−βεp ),
~ 1/(1 − ze
p bosons,
Z(V, T, z) = Q (5.22)
−βεp ),
~ (1 + ze
p fermions.
The grand potential Ω = −pV and the particle number N are given by
pV = kB T log Z
( P
− p~ log(1 − ze−βεp ), bosons,
= kB T P (5.23)
−βεp ),
~ log(1 + ze
p fermions
∂
N = z log Z
∂z
(P
βεp /z − 1),
~ 1/(e
p bosons,
= (5.24)
βεp /z + 1),
P
~ 1/(e
p fermions.
From the latter, we obtain the occupation numbers hnp~ i for the state |~
pi
(
1/(eβ(εp −µ) − 1), bosons,
hnp~ i = (5.25)
1/(eβ(εp −µ) + 1), fermions,
P
[by inspection of (5.24) and using N = p~ hnp~ i, or calculate
1 X N X P 1 ∂
hnp~ i = z np~ e−β p~ np~ εp = − log Z
Z β ∂εp
N {np~ }N

and use (5.23).]

The sums in (5.23) and (5.24) can be simplified further. Using
Z ∞ Z ∞
X 4πp2 dp x2 =βp2 /2m V 4
=V 3
= 3
√ dx x2 ,
0 (2π~) λ π 0
p
~

we obtain (
p = (kB T /λ3 )f5/2 (z),
fermions (5.26)
N/V = n = (1/λ3 )f3/2 (z),
where the functions f5/2 and f3/2 derive from [we use the expansion log(1 +
2
s) = s − s2 /2 + s3 /3 − s4 /4 + · · · = ` (−1)`+1 s` /` and s = ze−x ]
P
Z ∞
4 2
f5/2 (z) = √ dx x2 log(1 + ze−x ) (5.27)
π 0
Z ∞ ∞ ∞
4 X (−1)`+1 z ` −`x2 X (−1)`+1 z `
= √ dx x2 e = ,
π 0
`=1
`
`=1
`5/2
∞
X (−1)`+1 z `
∂
f3/2 (z) = z f5/2 (z) = ,
∂z
`=1
`3/2
5.2. GASES 83

replacing (5.23) and (5.24) by simpler expressions. In the second integral,

we have replaced `x2 → r2 (→ additional factor `−3/2 ) and the remaining
√
integral provides the compensating factor π/4.
The analysis for bosons is slightly more complicated as the terms p~ = 0
in (5.23) and (5.24) lead to divergences (Bose-Einstein condensation), see
below. Thus, we first separate these terms from the rest before replacing
the sum by an integral9 and obtain the expressions
(
p = (kB T /λ3 )g5/2 (z) − (kB T /V ) log(1 − z),
bosons (5.28)
N/V = n = (1/λ3 )g3/2 (z) + (1/V ) z/(1 − z),

with

∞ ∞
z`
Z
4 2 −x2
X
g5/2 (z) = − √ dx x log(1 − ze )= , (5.29)
π 0 `=1
`5/2
X z` ∞
∂
g3/2 (z) = z g5/2 (z) = .
∂z `3/2 `=1

For z → 1, the term (1/V ) z/(1 − z) = hn0 i/V will single out a fraction of
the total density as a k-space condensate, i.e., hn0 i ∝ N ; we will discuss
in detail later how the occupation number hn0 i can become macroscopically
occupied. The consequence is the appearance of quantum mechanical effects
on a macroscopic scale: MQP = macroscopic quantum phenomena. The
internal energy U of the quantum gas is obtained from (5.23),

1 X N X X  P
U (V, T, z) = z np~ εp e−β p~ np~ εp
Z
N {np~ }N p
~
(
∂ ∂ h V f5/2 (z) i
= − log Z(V, T, z) = −
∂β ∂β λ3 g5/2 (z)
(
3 V kB T f5/2 (z), fermions,
= (5.30)
2 λ3 g5/2 (z), bosons,
3
= p(V, T, z)V.
2

In the last expression, valid for the pure Bose gas, we have ignored the con-
densation terms for the bosons. Free Fermi and Bose gases have a multitude
of interesting properties and applications which we will come back to later.

9
use again the expansion of the logarithm, log(1 − s) = −(s + s2 /2 + s3 /3 + s4 /4 +
WeP
· · · ) = ` s` /`.
84 CHAPTER 5. APPLICATIONS AND TECHNIQUES

5.2.3 Classical limit of ideal quantum gases

We calculate the canonical partition function of ideal quantum gases and

investigate their classical (high-temperature) limit β → 0, T → ∞. We
have to calculate the quantity
X
ZN = Tr(e−βH0 ) = hΦn |e−βEn |Φn i. (5.31)
n

The energy eigenvalue problem is solved by (note that n describes a 3N

multi-index n → {~
p1 , . . . , p~N })

H0 |Φn i = En |Φn i,
N
X p2i 2π
En = , p~i = ~~ni , ~ni ∈ Z3 ,
2m L
i=1
1 X
h~q1 , ~q2 , . . . ~qN |Φn i = √ (±1)π φp~1 (~qπ1 ) . . . φp~N (~qπN )
N ! π∈S
N

ei~p·~q/~
φp~ (~q ) = √ .
V

Substituting the sum over states |Φn i in (5.31) by an integral

VN
X Z
→ d3N p
n
N ! (2π~)3N

1 = d3N q |~q1 , . . . ih~q1 . . . |, we obtain

R
and inserting the identity

VN
Z
ZN = d3N p d3N q |Φn (~q1 , . . . , ~qN )|2 e−βEn . (5.32)
N !(2π~)3N

The factor 1/N ! takes into account that permutations of the vector p~i in
|Φn i does not generate any new states. We simplify the expression for the
probability |Φn |2 according to

1 X 0
|Φn |2 = (±1)π+π · φp∗~1 (~qπ1 )φp~1 (~qπ10 ) . . . φp∗~N (~qπN )φp~N (~qπN
0 )
N! 0
π,π
1 X
= (±)π ei~p1 ·(~q1 −~qπ1 )/~ . . . ,
VN π

where we have used that every permutation π 0 under the integral d3N p
R

gives the same contribution10 ; this allows us to replace π 0 = 1l and multiply

10
Show this for N = 2, 3.
5.2. GASES 85

by N !. The p~-integrations can be done by completing the square

d3 p −βp2 /2m+i~p·~q/~ d3 p h mq 2 i
Z Z h β  im 2 i
e = exp − p
~ − ~
q exp −
(2π~)3 (2π~)3 2m ~β 2~β
1 h πq 2 i
= 3 exp − 2 , λ2 = 2π~2 /mkB T .
λ λ
Making use of this result in the expression (5.32) for ZN gives us the exact
result
1  V N d3N q X
Z h π
− π2 (~
i
π − λ2 (~ qπ1 )2
q1 −~ qπN )2
qN −~
ZN = (±1) e . . . e λ . (5.33)
N ! λ3 VN N
π∈S

λ2
P
For T → ∞, → 0, the only term in the sum π that survives is the
identity 1lPand thus we obtain the classical result (5.20). For T < ∞ the
terms in π become smaller the more particles are permuted by π. As long
as h(~qi − ~qj )2 i > λ2 (i.e., at low densities), we can expand the sum in the
number of permutations and obtain
X X h 2π i
(±1)π [. . .] ≈ 1 ± exp − 2 (~qi − ~qj )2 + 3 particle − terms.
π
λ
i<j

VS
Fig. 5.1: Statistical inter-
action for bosons (B, attrac-
F tive) and fermions (F, repul-
sive with a hard-core).

0 q /λ
1
B
−k B T log 2

Using the identity x = elog x , we construct the potential

h 2
i
V S (~q ) ≡ −kB T log 1 ± e−2π(q/λ) , (5.34)

allowing us to approximate (5.33) at large temperatures by the expression

Z 3N
1  V N d q h X
S
i
ZN ≈ exp −β V (~
q i − ~
q j ) . (5.35)
N ! λ3 VN
i<j

Comparing (5.35) with the classical result (5.10), we see that the quantum
statistics of the particles expresses itself as an effective potential under a
configurational integral, the so called statistical interaction
2

V S (q) = −kB T log 1 ± e−2π(q/λ) . (5.36)
86 CHAPTER 5. APPLICATIONS AND TECHNIQUES

This interaction is short ranged (λ) and depends on the particle statistics,
attractive for bosons and repulsive for fermions, see Fig. 5.1.

5.2.4 Quantum corrections to the interacting classical gas

In order to find the quantum corrections of particle-particle interactions, we

again perform an expansion in ~. Quantum effects then correct the classical
results with terms due to the statistical interaction as well as the quantum
corrections to V , modifying the configurational integral to read
Z 3N
1  V N d q
ZN (V, T ) ≈ 3
(5.37)
N! λ VN
h X β~2 i
× exp −β [V (~qi − ~qj ) + V S (~qi − ~qj ) + ∆i V (~qi − ~qj )] .
24m
i<j

5.2.5 Interacting particles at T → 0

For T → 0, particles have small kinetic energies and hence ~k → 0; in this

situation, particle-particle interaction can be described by s-wave scattering.
The scattering phases δ`=0 (k → 0) are given by (see the course on quantum
mechanics, scattering theory, partial wave analysis)

1 r0
k cot δ0 ≈ − + k 2 + . . . . (5.38)
a 2
The scattering length is a > 0 for a repulsive interaction and a < 0 for
an attractive interaction. The parameter r0 gives the effective range of the

r Ψ( r )

r Ψ (r )
V

0 a r a 0 r
V

Fig. 5.2: Shape of the wave function Ψ(r) for a repulsive (left) and an
attractive (right) interaction V (r).

potential. At small energies, cot δ0 ≈ −1/ak is a valid approximation. For

a repulsive interaction, the scattering between particles then corresponds to
5.2. GASES 87

that of hard spheres of radius a > 0. The s-wave scattering of hard spheres
can conveniently described by the pseudopotential

a 4πa~2 ∂ 
V`=0 (~r ) = δ(~r ) r (5.39)
m ∂r
i.e., the two Hamiltonians

p2
H = + Vhard sphere, R=a ,
2m
p2 a
H = + V`=0 (~r ), (5.40)
2m

generate the same scattering solutions Ψ`=0,k (~r ) in the exterior space r >
a for ` = 0, k → 0; correspondingly they generate the same eigenvalues
E`=0 (k).
a (~
A further simplification arises when the pseudopotential V`=0 r ) is treated
perturbatively. Within a perturbative analysis, one uses plane waves as the
unperturbed solutions. As these solutions, say ϕ(~r ), are regular at the ori-
gin ~r = 0, we have δ(~r )(∂r rϕ) = δ(~r )ϕ(0) and one may replace11 (5.39) by
a (~
V`=0 r ) = (4πa~2 /m)δ(~r ). To treat repulsively interacting quantum gases
(either fermions or bosons) in the limit of small temperatures T → 0 the
following strategy is suitable:

1. Calculate the spectrum of

X p2 4πa~2 X
i
H= + δ(~qi − ~qj ) (5.41)
2m m
i i<j

in perturbation theory. The result depends on the particle statistics.

Note that divergencies at small scales r → 0 or large momenta k → ∞
have to be treated by reconsidering the full operator δ(~r )(∂r r), e.g.,
in the context of the Bogoliubov spectrum for repulsive condensed
bosons.

2. Calculate the canonical partition function from the spectrum.

The above strategy is succesfull for repulsive particles with

λ, n−1/3 = hdistancei > a,

1/3
a/λ(T ), an (dilute gas)  1. (5.42)
11
One may go beyond perturbation theory if one replaces the operator gδ(~r )∂r r by
ḡ(Λ)δ(~r ) with a renormalized coupling ḡ(Λ) and Λ a short distance cutoff, see W. Zwerger’s
Varenna lectures for details and conditions.
88 CHAPTER 5. APPLICATIONS AND TECHNIQUES

5.3 Path integrals

Here, we restrict ourselves to 1-particle quantum statistical mechanics and

where appropriate make some comments about the many-body problem.
The generalization to bosonic many-body systems is “trivial”, for fermions
Grassman-variables are required, see, e.g., the book of Negele and Orland.
We start with the differential equation for the density matrix (see 5.12)
∂ρ
−~ = Hρ(τ ). (5.43)
∂τ
The correctness of (5.43) can easily be seen from the energy representation
of the density matrix (in a sketchy notation)

hΦi | ρ(β) = e−βH |Φj i,

ρij (β) = e−βEi δij ,
∂ρij
= −Ei ρij ,
∂β
∂ρ
hΦi | = −Hρ |Φj i.
∂β
By defining τ = ~β as imaginary time, we immediately obtain (5.43). For-
mally (5.43) is naturally solved by ρ(τ ) = e−Hτ /~ . In the position represen-
tation, (5.43) is given by
∂ρ(x, x0 ; τ )
−~ = Hx ρ(x, x0 ; τ ) (5.44)
∂τ
(where H acts on the x-coordinate in ρ, therefore Hx ) or, more generally,
Z
∂
−~ hx|ρ|x0 i = dx00 hx|H|x00 ihx00 |ρ|x0 i.
∂τ
The initial condition on (5.44) is

ρ(x, x0 ; 0) = hx|x0 i = δ(x − x0 ). (5.45)

For a free particle with

p2 ~2 2
H0 = =− ∂ (5.46)
2m 2m x
one arrives at the diffusion problem
∂ρ0 ~2 2
−~ = − ∂ ρ0 , (5.47)
∂τ 2m x
ρ0 (x, x0 ; 0) = δ(x − x0 ),
 m 1/2 h m i
→ ρ0 (x, x0 ; τ ) = exp − (x − x0 )2 .
2π~τ 2~τ
5.3. PATH INTEGRALS 89

Hence, the density matrix or imaginary time propagator evolves as a diffusive

particle when starting from the classical limit at τ = 0 or high temperatures
where the particle position is well defined, to a quantum smeared particle
when decreasing the temperature. The smearing is over a distance ∆x ≈
p
λ2 /π, and transforming the wavelength λ to momentum λ = 2π/k = h/p,
we arrive at the Heisenberg uncertainty relation ∆x∆p ∼ ~. For a system
of length L at temperature kB T = 1/β = ~/τ , we find the partition function
Z Z
Z1 = Tr(ρ0 ) = dx hx|ρ0 |xi = dx ρ0 (x, x; β~)
Z  m 1/2  mk T 1/2 L
B
= dx = L = , (5.48)
2π~2 β 2π~2 λ

in agreement with (5.4).

Next, we find a path integral representation of the density matrix for a
free particle. To this end, we write the Boltzmann factor as an n-fold product
e−τ H/~ = [e(−τ /n)H ]n and rewrite the density matrix ρ(τ ) in a corresponding
form

ρ(τ ) = ρ(ε)ρ(ε) . . . ρ(ε), τ = nε,

with ε a small parameter. Inserting unities, we obtain the expression

Z Z
→ ρ(x, x0 ; τ ) = . . . dx1 . . . dxn−1 (5.49)
ρ(x, xn−1 ; ε)ρ(xn−1 , xn−2 ; ε) . . .
. . . ρ(x2 , x1 ; ε)ρ(x1 , x0 ; ε)

for the real space representation of the density matrix. The idea then is that,
while the calculation of ρ(x, x0 ; τ ) is difficult, the determination of ρ(x, x0 ; ε)
is simplified by the presence of the small parameter ε. This will become
apparant when we discuss the particle in a potential V (x) further down—
before going into this, however, we discuss the simplest situation, which is
free particle.
For a free particle, the imaginary time propagator over an interval ε has
the form  m 1/2 h m i
ρ0 (x, x0 ; ε) = exp − (x − x0 )2 , (5.50)
2π~ε 2~ε
and inserting this expression into (5.49), we arrive at the full density matrix
in the form
Z Z
0 1 dx1 dxn−1
ρ0 (x, x ; τ ) = p ... p ... p
2π~ε/m 2π~ε/m 2π~ε/m
h mε  x − x 2 i h mε  x − x0 2 i
n−1 1
exp − . . . exp − .
2~ ε 2~ ε
90 CHAPTER 5. APPLICATIONS AND TECHNIQUES

In the limit of small ε → 0, we can make the substitutions (note that we

have n factors in the product with x0 = x0 and xn = x)

xk − xk−1 dx h mε i h 1Z τ m  dx 2 i
→ , Π exp − 2
(. . .) → exp − dτ 0
ε dτ 2~ ~ 0 2 dτ
| {z }
SE0 [x(τ )]

with SE0 [x(τ )] the Euklidean action (and the index 0 standing for the free
particle). Furthermore, we abbreviate the product of integrals over interme-
diate positions xi as an integral over paths,
Z n−1 Z x(τ )=x
1 Y dxi
p p → D[x(τ )].
2π~ε/m i=1
2π~ε/m x(0)=x0

These replacements provide us with the path integral representation of the

density matrix ρ0 (x, x0 ; τ ) for the free particle,
Z h 1 i
0
ρ0 (x, x ; τ ) = D[x(τ )] exp − SE0 [x(τ )] (5.51)
~
1 h m i
= p exp − (x − x0 )2 ,
2π~τ /m 2~τ

where the second equation can be obtained by an explicit integration over

the internediate positions xi via completing squares and carrying out the
integration over the Gaussians; the final result then agrees with the solution
(5.47) obtained previously by solving the initial value problem in differential
form.
One can arrive to this result in an alternative way: Consider the real time
propagator of the free particle (with S0 [x(t)] the action of a free particle)
Z hi i
0
G0 (x, x ; t) = D[x(t)] exp S0 [x(t)] (5.52)
~
as it is known from quantum mechanics. This propagator solves the Schrödinger
equation

i~∂t G0 = H0 G0 ,

with the initial condition

G0 (x, x0 ; 0) = δ(x − x0 ),

and has the explicit form

1 h im i
G0 (x, x0 ; t) = p exp (x − x0 )2 .
2πi~t/m 2~t
5.3. PATH INTEGRALS 91

Going over to imaginary time,

t → −iτ, (Wick rotation), (5.53)

the action Z
i i m  dx 2
S0 [x(τ )] = dt
~ ~ 2 dt
transforms to the Euclidean action by a Wick rotation
Z Z
i m  dx 2 1 m  dx 2 SE0 [x(τ )]
→ −idτ =− dτ =− , (5.54)
~ 2 −idτ ~ 2 dτ ~
thus reproducing the result (5.51).
In a second step, we focus on a particle in the potential V (x). We can
analyze the imaginary time Schrödinger equation (5.43) by splitting the
Hamilonian into a free part H0 and a potential part V that we treat sepa-
rately, as usually done in a perturbatve approach. 12 Given that

−~∂τ ρ0 = H0 ρ0 (5.55)

has been solved, the task is to find the solution to −~∂τ ρ = Hρ. As usual,
we rewrite the correction eH0 τ /~ ρ = ρD in the Dirac representation. The
dynamical equation for ρD is easily found,
−Hρ
z }| {
H0 τ /~ H0 τ /~
−~∂τ ρD = −e H0 ρ − e ~∂τ ρ = eH0 τ /~ V ρ
= VD ρD , where VD = eH0 τ /~ V e−H0 τ /~ .

Integrating (5.55) with the initial condition

ρD (0) = eH0 τ /~ ρ|τ =0 = 1,

1 τ 0
Z
ρD (τ ) = 1 − dτ VD ρD (τ 0 ),
~ 0
we obtain
1 τ 0
Z
ρ(τ ) = ρ0 (τ ) − dτ ρ0 (τ − τ 0 )V ρ(τ 0 )
~ 0
1 τ 0
Z
1st Order
≈ ρ0 (τ ) − dτ ρ0 (τ − τ 0 )V ρ0 (τ 0 ), (5.56)
~ 0
12
Alternatively, we can use (5.49) and decompose ρ(x, x0 ; ε) = hx| exp[−ε(T +
0 0
R
V )/~]|x i ≈ hx| exp(−εT /~) exp(−εV /~)|x i and insert two complete sets dy|yihy|
exp(−εT /~)|x0 i.
R R
and (dp/2π~)|pihp| to arrive at (dpdy/2π~)hx| exp(−εV /~)|yihy|pihp|
0
R
Evaluating the matrix elements, we arrive at ρ(x, x ; ε) ≈ (dp/2π~) exp[ip(x −
x0 )/~] exp[−ε(T (p) + V (x)/)~]. The integral over p then gives free part ρ0 (x, x0 ; ε) of
the propagator and we obtain ρ(x, x0 ; ε) ≈ ρ0 (x, x0 ; ε) exp(−εV (x)/~), see Eq. (5.59). The
Baker-Campbell-Haussdorff formula tells that corrections are of order ε2 , see also the
Trotter-Suzuki product formula saying that exp(−τ H/~) = [exp(−εT /~) exp(−εV /~)]n
for n = τ / → ∞.
92 CHAPTER 5. APPLICATIONS AND TECHNIQUES

where in the second equation we use the smallness of τ (not of V ). Using this
perturbative result, we can generalize the expression (5.50) for Rρ0 (x, x0 ; ε) to
the situation of a particle in a potential V (x). Inserting 1 = dx0 |x00 ihx00 |
into (5.56), we obtain
ρ(x, x0 ; ε) ≈ ρ0 (x, x0 ; ε) (5.57)
Z ∞ Z ε
1
− dx00 dε0 ρ0 (x, x00 ; ε − ε0 )V (x00 )ρ0 (x00 , x0 ; ε0 ).
~ −∞ 0

For small ε the main contribution to the integral dx00 comes from the
R

region
(ε − ε0 )x0 + ε0 x if x ≈ x0
x̄00 = ≈ x ≈ x0 ,
ε
which can be verified by minimizing the exponent of ρ0 · ρ0 ,
∂  (x − x00 )2 (x00 − x0 )2 
+ = 0.
∂x00 ε − ε0 ε0 x̄00

Then, equation (5.57) simplifies to

Z ∞
1 ε 0
Z
ρ − ρ0 |ε,x−x0 small ≈ − dε V (x) dx00 ρ0 (x, x00 ; ε − ε0 )ρ0 (x00 , x0 ; ε0 )
~ 0 −∞
| {z }
→ρ0 (x,x0 ;ε) (ε small)
ε
≈ − V (x)ρ0 (x, x0 ; ε). (5.58)
~
Inserting
h ε i
ρ(xk , xk−1 ; ε) ≈ 1 − V (xk ) ρ0 (xk , xk−1 ; ε)
~
1 h ε m(xk − xk−1 )2 i
≈ p exp − V (xk ) − (5.59)
2π~ε/m ~ 2~ε
into (5.49) provides us with the path integral for the particle moving in the
potential V (x),
Z  1 
0
ρ(x, x ; τ ) = D[x(τ )] exp − SE [x(τ )] , (5.60)
~
with the Euclidean action
Z τ h m  dx 2 i
SE [x(τ )] = dτ 0 +V (x) .
0 2 dτ
R R
Note that S = dt(T − V ), while SE = dτ (T + V ).
The statistical mechanics of such a particle follows from the trace
Z
Z1 = dx ρ(x, x; τ = ~β),
1
F1 = − ln Z1 . (5.61)
β
5.3. PATH INTEGRALS 93

The generalization to a particle in R3 is trivial.

The final results (5.60 and 5.61) give us a new perspective on the quan-
tum statistics of a single particle: The quantum statistics of a particle is
given by the weighted sum of all periodic imaginary-time world lines. The

hβ
imaginary time

q ( τ)
τ

q2
q
1
0

Fig. 5.3: Imaginary-time world line extending periodically between τ = 0

and τ = ~β. The behavior of the particle lifts the dimensionality of the
problem from D = 2 (a point in a plane) to a periodic world line in 2 + 1-
dimensional space. Deviations from a straight line describe the quantum
fluctuations which the particle undergoes (due to the quantum uncertainty).

stronger a line fluctuates the smaller is its weight in Z1 (∂~q/∂τ = “angle”

away from the R τ -axis = “velocity”). Similarly, the weight in Z1 is suppressed
if the mean ~qdτ /β = h~q i involves a large potential energy. In the limit
β → ∞, T → 0 → τ → ∞ the world line becomes infinitely long: In the
free case it can deviate from the ideal line ~q = const with minimal cost (a
deviation δq over large times τ costs little action); these deviations from the
straight line represent the quantum fluctuations which the particle under-
goes due to the uncertainty principle (this is the phenomenon of ‘quantum
smearing’). In the high-temperature case β → 0, T → ∞ → τ → 0, finite
“velocities” ∂τ ~q cost a lot of energy (even a small deviation δq costs a lot of
action when it involves a small time τ < ~β → 0) and the world line stays
straight with ∂τ ~q ≈ 0; this correspoonds to the classical limit of statistical
mechanics without quantum fluctuations,

Z ~β
SE dτ h m i
= (∂τ ~q )2 + V (~q )
~ 0 ~ 2
SE0
≈ + βV (~q ), (5.62)
~
94 CHAPTER 5. APPLICATIONS AND TECHNIQUES
Z
→ρ ≈ D[~q (τ )]e−SE0 /~ e−V (~q )/kB T = ρ0 e−V (~q )/kB T ,
Z 3 Z 3
d q −V (~q )/kB T cf. (5.8) V d q −V (~q )/kB T
→ Z1 ≈ Z10 e = 3
e .
V λ V

hβ
T

8
top view

τ
hβ
τ
qy
0
q
x

Fig. 5.4: World lines in the limit β → ∞ (left) and the crossover into
classical world lines (right) for β → 0, T → ∞. The top perspective in the
center illustrates the quantum fluctuations (“smearing”) of the particle at
low temperatures.

From Fig. 5.4, we see that classical statistical mechanics in d + 1 di-

mensions is equivalent to quantum statistical mechanics in d-dimensional
systems, for example the classical statistical mechanics of elastic (C) strings
corresponds to the quantum statistical mechanics of point particles with
mass m = C.
When going over to many-body systems, we must take into account the
statistics of the particles. The indistinguishability of the particles appears in
the periodic boundary conditions when considering the trace: when taking
the trace

Z
d3N q ρ(~q1 , . . . , ~qN ; ~q10 , . . . , ~qN0 ; β~)|q~i0 =~qi , (5.63)

we have to include all world lines with permuted endpoints ~qi0 = ~qP i , P ∈
S N , in the partition function. For bosons, we must add all the contributions
5.3. PATH INTEGRALS 95

and obtain the result

Z
1 X
ZN = d3N q ρ(~q1 , . . . , ~qN ; ~qP 1 . . . ~qP N ; β~),
N!
P
Z YN h S [{q (τ )}] i
0 0 E k
ρ(~q1 , . . . , ~qN ; ~q1 , . . . , ~qN ; τ ) = D[~qk (τ )] exp − , (5.64)
~
k
Z τ hX
m 1X i
SE [{~qk (τ )}] = dτ 0 (∂τ 0 ~qi )2 + V (~qi − ~qj ) .
0 2 2
i i6=j

A typical configuration is sketched in Fig. 5.5. In a Path Integral Quantum

disentangled entangled

hβ

Fig. 5.5: Disentangled (left) and entangled (right) world lines. Bosons at
low temperatures entangle easily and thereby attract one another, as sep-
arating entangled bosons stretches their worldlines leading to an increase
of the action; the same result has been obtained earlier on, see Eq. (5.36).
Furthermore, there is an interesting insight into the Bose condensation phe-
nomenon, as Bosons exchange their identity through the periodic boundary
conditions. Bosons entangling in a “ring”within so-called “cooperative ring
exchanges”, assume the “same identity”. Once these permutation “rings”
extend throughout the entire sample, the system undergoes condensation.
Note that there is only one disentangled state but there are N ! states that
involve entanglement.

Monte Carlo simulation one calculates and adds all configurations as those
sketched in Fig. 5.5. For fermions the problem is more complicated: The sum
P
P
p in (5.64) must be weighted by a (−1) prefactor, there is a sign problem
in the PIQMC-technique for simulating fermions. The corresponding path-
integrals involve Grassmann variables.
A cute application is the study of two-boson/two-fermion systems and
the resulting understanding of the statistical interaction in the language of
96 CHAPTER 5. APPLICATIONS AND TECHNIQUES

world lines.

5.4 Variational principle

Consider the problem characterized by the (Euclidean) action S (we suppress

the index E). The partition function is given by the path integral13
Z
−βF
Z=e = D[x(τ )]e−S[x(τ )]/~ . (5.65)

In general, the calculation of Eq. (5.65) is difficult if not impossible. We look

for a simple action S0 , whose partition function can be calculated and which
approximates well the partition function we are looking for. In this endeav-
our the following theorem is often useful: The free energies corresponding
to S and S0 fullfill the inequality
1
F ≤ F0 + hS − S0 i0 . (5.66)
β~
For a classical system we have

F ≤ F0 + hH − H0 i0 . (5.67)

The Bogoliubov variational principle follows directly: The functional

1
F[S0 ] = F0 + hS − S0 i0 ,
β~
F[H0 ] = F0 + hH − H0 i0 , (5.68)

is minimal for the best choice of S0 or H0 . For example, one defines S0 or H0

to be a quadratic action or Hamiltonian and minimizes (5.68) with respect to
the parameter in S0 , H0 ; this procedure is called the self-consistent harmonic
approximation.
In order to prove (5.68) one writes (5.65) in the form

D[x]e−(S−S0 )/~ e−S0 /~ −βF0

R
−βF
e = R e
D[x]e−S0 /~
= e−βF0 he−(S−S0 )/~ i0 . (5.69)

We assume that S and S0 are real and use the convexity of e−x ,

e−βF ≥ e−hS−S0 i0 /~ e−βF0 ,

1
→F ≤ F0 + hS − S0 i0 .
β~
13
We
R simplify
R notation
R and include theR integral over positions into the path integral,
Z = dxρ = dx D[x(t)] exp(−S/~) → D[x(t)] exp(−S/~).
5.5. CUMULANT EXPANSION 97

e −f Fig. 5.6: Convexity of the

function e−f and positions
e−f
of the averages he−f i versus
− f e−hf i .
e

f1 f f2 f

For β → 0, T → ∞, we define L0 = (m/2)(∂τ x)2 + V0 , then

m
L = (∂τ x)2 + V,
2
hS − S0 i0 ≈ ~βhV − V0 i0 ,
and we find the classical result
F ≤ F0 + hH − H0 i0 .

5.5 Cumulant expansion

Consider the situation where we divide up an action S into a simple part S0 ,

for example free or Gaussian, plus rest, S = S0 + S 0 . We want to calculate
F = −kB T ln Z from (see (5.69), ~ = 1)
0
Z = Z0 he−S i0 . (5.70)
If S 0 is small (a perturbation), we can expand hexp(−S 0 )i0 as
0 hS 0 2 i0
he−S i0 ≈ 1 − hS 0 i0 + − ... . (5.71)
2
Using log(1 + x) = x − x2 /2 + O(x3 ), we can rewrite (5.71) as
0
 hS 0 2 i0 
loghe−S i0 ≈ log 1 − hS 0 i0 + − ...
2
1  02
0
≈ −hS i0 + hS i0 − hS 0 i20 − . . .

2
0
h 1 2
i
→ he−S i0 = exp −hS 0 i0 + [hS 0 i0 − hS 0 i20 ] − . . . (5.72)
2
The expression in square brackets in the second line for the expansion of
the logarithm is called a cumulant. The form (5.72) provides us with a new
effective form of the free energy,
kB T n 0 1  02 o
F = −kT ln Z ≈ F0 + hS i0 − hS i0 − hS 0 i20 ] + . . . . (5.73)
~ 2~
98 CHAPTER 5. APPLICATIONS AND TECHNIQUES

The corresponding result for the classical case (S → H, ~ → kB T ) reads

1  02
F = −kT ln Z ≈ F0 + hV 0 i0 − hV i0 − hV 0 i20 + . . . .

(5.74)
2kB T

Higher-order corrections in S 0 or V 0 can be found by going back to the

general definition of theRcumulant expansion: Let x be a random variable
with distribution p(x), dx p(x) = 1. Let V (x) be a function on x. We
define the two generating functions for the moments and cumulants

φV (t) ≡ heitV (x) i and ΨV (t) ≡ ln φV (t).

Then, the n-th derivative of φV and ΨV defines the moment

(n)
Mn ≡ hV n (x)i = (−i)n φV (0)

and the cumulant

(n)
Cn ≡ (−i)n ΨV (0). (5.75)
The moments and cumulants are related via

C0 = 0,
C1 = M1 = hV (x)i,
C2 = M2 − M12 = hV 2 (x)i − hV (x)i2 ,
C3 = M3 − 3M2 M1 − 2M13 ,
...

The cumulants provide us with the expectation value of eV ,

∞
V (x)
hX 1 i
he i = exp Cn . (5.76)
n!
n=1

The generalization to fields x → Φ can be done in a similar way.

Two remarks: Higher-order cumulants introduce new information about
the distribution function, while higher moments ‘repeat’ information. E.g.,
the Gaussian distribution is characterized by two cumulants, but has arbi-
trary many moments. Second, when doing perturbation theory with Feyn-
man graphs, one observes (see, e.g., the chapter on renormalization group
theory) that only a class of graphs, the connected ones, are relevant—these
correspond to cumulants.
Chapter 6

Fermions

We summarize the results for the Fermi gas in the previous chapters:
Partition function:
Y
Z(V, T, z) = (1 + ze−βεp ). (6.1)
p
~

Equation of state:
kB T kB T
p= log Z = 3 f5/2 (z).
V λ
Particle density:
z 1
n= ∂z log Z = 3 f3/2 (z).
V λ
Occupation number:
1
hnp~ i = . (6.2)
eβ(εp −µ) +1

∞ ∞
(−1)`+1 z `
Z
4 2
X
f5/2 (z) = √ dx x2 log(1 + ze−x ) = ,
π 0 `=1
`5/2
∞ ∞
x2 (−1)`+1 z `
Z
4 X
f3/2 (z) = √ dx = .
π 0 1 + ex2 /z `=1
`3/2

We carry out the following program: We first solve the equation relating the
gas parameter nλ3 and the fugacity z = eβµ ,

δ = nλ3 = f3/2 (z), (6.3)

in order to find the chemical potential

µ(n) = kB T log z(n). (6.4)

99
100 CHAPTER 6. FERMIONS

By calculating the internal energy (or caloric equation of state)

X
U= hnp~ iεp , (6.5)
p
~

we can then find the thermal equation of state via (see 8.41)
2U
p= (6.6)
3V
and we are done.
The dimensionless density parameter
δ = nλ3 (6.7)
characterizes the quantumness of the gas, see Fig. 6.1,
δ  1 → classical gas,
δ  1 → quantum gas. (6.8)

Fig. 6.1: Density parameter:

left, a dilute classical gas with
δ  1, right, a dense quantum
gas with δ  1.

We expand f3/2 (z) for large and small z using the series expansion for
z  1,
z2 z3
f3/2 (z) ≈ z − 3/2 + 3/2 − . . . , (6.9)
2 3
and the integral representation for z  1,
Z ∞
4 x2
f3/2 (z) = √ dx (6.10)
π 0 1 + ex2 −βµ
Z ∞ √ Z ∞
2 y 4 y 3/2 ey−βµ
= √ dy = √ dy
π 0 1 + ey−βµ 3 π 0 (1 + ey−βµ )2
where z = eβµ . The function
ey−ν d 1
y−ν 2
=−
(1 + e ) dy 1 + ey−ν
evaluated at fixed ν = log z = βµ is sharply peaked around ν motivating an
expansion of y 3/2 around this point (Sommerfeld expansion),
3 3
y 3/2 ≈ ν 3/2 + ν 1/2 (y − ν) + 1/2 (y − ν)2 + . . . .
2 8ν
 101

We then have to calculate the integral

Z ∞
tn et
In = dt , (6.11)
−∞ (1 + et )2
where we have replaced the lower bound t = y − ν → −ν at y = 0 by −∞
at large values of z  1 or ν. One then finds the coefficients
π2 7π 4 31π 6
I0 = 1, I1 = I2n+1 = 0, I2 =
, I4 = , I6 = , (6.12)
3 15 21
providing us with the Sommerfeld expansion for f3/2 (z),
4 h π2 i
f3/2 (z) ≈ √ (log z)3/2 + (log z)−1/2 + . . . + O(e−ν ). (6.13)
3 π 8

6.0.1 Dilute Fermi gas, δ = nλ3  1

Starting from δ = f3/2 (z), (6.3), and using (6.9), we find (fMB (~
p ) is the
1
classical Maxwell-Boltzmann distribution )
z2 1
δ ≈ z− √ z ≈ δ + √ δ 2 > 0,
→ (6.14)
2 2 2 2
 2π~2 3/2 h p2 i
hnp~ i ≈ ze−βεp = n exp −
mkB T 2mkB T
3
≈ h fMB (~
p ).
The series expansion of f5/2 (z) provides us with the thermal equation of
state
V  z2 
pV ≈ kB T z − 5/2 + . . . (6.15)
λ3 2
 nλ3 
≈ N kB T 1 + √ + . . . , (6.16)
4 2
describing the dilute Fermi Gas as a classical ideal gas with quantum cor-
rections ∝ δ. Note that the chemical potential µ and its dependence on
n appears in the prefactor of the Maxwell-Boltzmann distribution (via the
fugacity z = exp(βµ)).

6.0.2 Degenerate Fermi gas, δ = nλ3  1

Using (6.3) and (6.13), we find the relation

4 h π2 1 i
δ = nλ3 ≈ √ (βµ)3/2 + (6.17)
3 π 8 (βµ)1/2
1
p ) = n(2πmkB T )−3/2 exp(−p2 /2mkB T ), dif-
The classical distribution function fMB (~
fers from the present result by its normalization, hnp~ i = (2π~)3 fMB (~
p ).
102 CHAPTER 6. FERMIONS

between the density n and the chemical potential2 µ, and thus to lowest
order

~2
µ(T = 0) = (6π 2 n)2/3 = εF . (6.18)
2m

Including corrections to lowest order in T , one finds the chemical potential

µ(T ) to decrease with increasing temperature,3

h π 2  kB T 2 i
µ(T ) ≈ εF 1 − . (6.20)
12 εF

The occupation number

1
p ) = hnp i ≈
fFD (~
eβ(εp −µ) +1

is 1 for εp < µ and rapidly decays to 0 in a window of width kB T around µ.

Expressing the Fermi energy through the Fermi momentum, εF = ~2 kF2 /2m,
we see that inside the sphere k < kF the states are occupied at T = 0, while
outside they are unoccupied. The sharp boundary at kF is called the Fermi
surface. Using this insight, we can easily calculate kF and εF ,
Z
X 2π
N = 1= d3 n, where ~k = ~n
k<kF L
k<kF
Z kF
V V 3
= dk 4πk 2 = k ,
(2π)3 0 6π 2 F
→ kF3 = 6π 2 n. (6.21)

For energies εp~ that are g-times degenerate, the corresponding result is

6π 2
kF3 = n.
g

For example, for an electron gas with s = 1/2, g = 2, and one obtains

kF3 = 3π 2 n. (6.22)
2
Note that εF is one of the characterising parameters of a Fermi gas, while µ describes
the thermodynamic chemical potential.
3
First, recast (6.17) into the form
h π 2  kB T i2/3 ~2
µ 1+ ≈ (6π 2 n)2/3 (6.19)
8 µ 2m

and then solve iteratively for small values of kB T /µ ≈ kB T /εF .

 103

In order to find the internal energy, we calculate

Z ∞ Z ∞ 5
X V PI V p 
U = εp hnp i = 2 3 dp p4 hnp i = dp −∂ hn
p p i
p
4π ~ m 0 4π 2 m~3 0 5
√ √
V m 2m ∞ y 5/2 ey−βµ V m 2mµ5/2
Z
= dy ≈ (1 + . . .)
5π 2 ~3 β 5/2 0 (1 + ey−βµ )2 5π 2 ~3
| {z }
(βµ)5/2 +...

3 h 5π 2  kB T 2 i
≈ N εF 1 + + ... . (6.23)
5 12 εF

With the Sommerfeld expansion to order T 2 , we can obtain the specific heat

π 2 kB T T →0
C V = N kB → 0, (6.24)
2 εF
in agreement with the third law.
Finally, we find the thermal equation of state (or the partition function)
to be
2U 2 h 5π 2  kB T 2 i
p= = nεF 1 + + ... . (6.25)
3V 5 12 εF
Note that the pressure, different from the classical ideal gas, does not vanish
at T → 0 but stays large. It corresponds to the pressure of a system at the
degeneracy temperature

~2 kF2 ~2  6π 2 2/3
kB TF = εF = = n . (6.26)
2m 2m g

For a metallic electron system with n ≈ 1022 cm−3 , we find a degeneracy

temperature

1.055 · 10−34 m2 kg/s · 6.58 · 10−16 eV s  6π 2 1022 2/3

kB TF ≈ ·
2 · 0.911 · 10−30 kg 2 10−6 m3
≈ 1.69 eV,
TF ≈ 2 · 104 K. (6.27)

Particle-hole asymmetry

The ∝ T 2 correction in the expression for the chemical potential µ(T ), see
Eq. (6.20), is a consequence of the particle-hole asymmetry: The density
of states ρ(ε) of the free Fermi gas in 3D is (we use ρ(k) = k 2 /π 2 for fermions
with spin and ∂k ε = ~2 k/m)
√ r
dk mk m 2mε ε
ρ(ε) = ρ(k) = 2 2
= 2 3
= ρ(εF ) , (6.28)
dε π ~ π ~ εF
104 CHAPTER 6. FERMIONS

with ρ(εF ) = 3n/2εF the density of states for an electron gas of total density
√
n including both spin directions.4 We see that ρ(ε) ∝ ε, i.e., there are
more states available for the electrons at energies ε > εF than for the holes
at ε < εF . Thus, for T > 0 the chemical potential µ has to decrease such
as to conserve the particle number, see Fig. 6.2. Indeed, in the integral
providing the fixed number of particles N ,
Z ∞
N =V dε fFD (ε)ρ(ε), (6.29)
0

the distribution function fFD (ε) is symmetric with respect to µ, but ρ(ε)
is not. For a free massive particle, we find a constant density of states
ρ(ε) (and hence symmetric around εF ) in dimension d = 2, as ρ(k) ∝ k
and ∂k ε ∝ k. In general, with a dispersion ε(k) ∝ k n and ρ(k) ∝ k d−1
in d dimensions, we have a constant ρ(ε) if d = n. Note that electrons
moving in a periodic potential have complicated dispersions and particle-
hole symmetry may appear in specific regions of the Brioullin zone. In
1D mesoscopic physics, the dispersion around the Fermi energy is often
linearized, εk − εF ≈ vF ~(k − kF ), implying particle-hole symmetry within
this approximation.

Fig. 6.2: Decrease of the chem-

ρ( ε) ical potential µ(T ) in a free elec-
T =0 tron gas. For a free massive
particle in 3D, the particle–hole
asymmetry generates a negative
fFD ( ε) shift in µ(T ) when T increases.

0 µ εF ε

Fermi gases in stars

The most prominent example of a Fermi gas is the electron gas in a metal.
However, other interesting examples are know from astronomy, i.e., the mat-
ter from which stars are made. Depending on the properties of such matter,
gravitational properties (i.e., mass density) versus pressure (i.e., equation of
state), a star may be stable or not.
In white dwarves (made from a 4 He2+ , e− plasma), the pressure is gener-
ated by the Fermi gas of the electrons. At a mass density of ρ ∼ 106 g/cm3 ,
the electron density is of the order of n ∼ 1029 cm−3 (m4 He ≈ 7 · 10−24 g)
4
Sometimes one may encounter the formula ρσ = 3n/4εF for the density of states per
spin σ in an electron gas.
 105

and constitutes a non-relativistic degenerate electron gas with a degeneracy

temperature
TF ∼ 109 K, (6.30)

much larger than the effective temperature T ∼ 107 K in the core of the
star. We may notice a neat division of tasks between the nuclei (the 4 He-
cores) and the electrons: The nucleons generate the attractive gravitational
force while the electrons, through their degeneracy pressure, generate the
stabilizing pressure. In the non-relativistic case, we obtain the polytrope (or
equation of state)
p = Kρ1+2/3 = Kρ5/3 . (6.31)

The exponent with the additional 2/3-contribution from εF coincides with

the adiabate in the classical ideal gas. In the relativistic case, where c pF >
me c2 , one finds the exponent γ = 4/3. Using the hydrostatic equation for a
~ ∇p/ρ)
stable star ∇·( ~ = −4πGρ (G = 6.674 10−11 m3 /kg s2 the gravitational
constant) one can determine the structure of the star. Note that the stability
of the star depends on its polytrope.
Finally, we consider a neutron star, where the reaction p + e− → n + ν
has transformed nearly all protons and electrons into neutrons. Gravity and
pressure are now generated by the same fermionic particle. Typical mass
densities of around ρ ∼ 1015 g/cm3 generate fermion densities n ∼ 1039 cm−3
(non-relativistic case; mN ≈ 1.7 · 10−24 g) corresponding to degeneracy tem-
peratures
TF ∼ 1012 K. (6.32)

The polytropes are p = Kρ5/3 and p = Kρ in the non-relativistic and

relativistic limits, respectively. A typical star life goes through the sequence
of a burning star to a white dwarf when the thermonuclear fuel runs out. If
the mass of the core is too large a collapse takes the object to a neutron-star.
Very complicated variants with supernova-explosions are possible.

Metallic sound

A similar divison of tasks between mass density ρ and pressure p as in the

white dwarf appears in the sound generation in metals. There, the electronic
Fermi-sea generates the pressure p while the ions provide the mass density ρ.
From the dispersion ω = (∂ρ p s/n )1/2 q = c q (s/n is the entropy per particle
in adiabatic sound) and using the compressibility5

5 2
κ−1 = −V ∂V p = n∂n p = p = nεF
3 3
5
At low temperatures, adjabatic and isothermal compressibilities are the same.
106 CHAPTER 6. FERMIONS

in a degenerate Fermi gas, it follows that6

1
2ZnεF 1 mZ p2F Zm 2
c2 = =
= = v , (6.33)
ρκ
nM 3 3 M m 2 3M F
r
Zm
→c = vF , (6.34)
3M
where we have used that ρ = nM/Z with n the electron density, M the mass
of the ions, and Z is the ion’s valency. Using

pF ~(3π 2 n)1/3
vF ∼ ∼ ∼ 108 cm/s,
m m
we obtain a typical sound (or phonon) velocity cmetal ∼ 106 cm/s in a metal.
Note that in these collective excitations both the electrons and the ions are
oscillating. What happens if only the electrons oscillate?

6.1 Electrons in a magnetic field

Electrons in a magnetic field show interesting properties. They define a com-

plex thermodynamic system and their study is relevant to the understanding
of the magnetic properties of metals.
The Hamiltonian of free electrons in a magnetic field has the form (we
~i = ~σi /2 with ~σ the
define e > 0 and include the ~ of the spin in µB , i.e., S
Pauli matrices)

X [~ ~ qi )]2
pi + (e/c)A(~ X
H = − ~
~ i · H,
µ (6.35)
2m
i i
~i ,
~ i = −µB g S
µ
e~
µB = = 0.579 · 10−8 eV/G = 1.165 · 10−19 G cm3 ,
2mc
g = 2[1 + α/2π + . . .] ≈ 2,

with H~ =∇ ~ ∧A ~ the magnetic field, µ ~i the

~ i the magnetic moment, and S
spin of the i-th electron (note that the spin and moment are antiparallel).
Gauge invariance gives us the freedom

A(~ ~ r ) − Φ0 ∇χ(~
~ r ) → A(~ ~ r ),
2π
Φ0
φ(~r ) → φ(~r ) + ∂t χ(~r ),
2πc
Ψ(~r ) → eiχ(~r ) Ψ(~r ), (6.36)
6
√
In 3 He-fluids and in neutron stars c = vF / 3.
6.1. ELECTRONS IN A MAGNETIC FIELD 107

where χ(~r ) is a dimensionless phase and

hc
Φ0 = ≈ 4 · 10−7 G cm2 (6.37)
e
is the quantum of magnetic flux.
The statistical mechanics of this system is given by calculating the par-
tition function ZN or Z. A new quantity that we must define in the present
context is the magnetization density 7 (we choose H~ = (0, 0, H) without loss
of generality) (
kB T ∂ ZN ,
m= log (6.38)
V ∂H Z.
It characterizes the response of the system to an applied magnetic field
H. The external field H induces currents or changes the direction of the
magnetic moments µ ~i associated with the spins S
~ i = −g µB S ~i (= ~σi /2) in
the system. Alternatively, we may define m as the change in energy induced
by an applied magnetic field H,
1 D ∂H E
m = −
V ∂H
1 DX E 1 D ∂H E
orb
= µz,i − (6.39)
V V ∂H
i

involving again spin and orbital contributions. Using H 2 = energy/volume,

it is clear that m has the same dimension as H. The spin contribution causes
an increase in the magnetic field

B = H + 4π mSpin > H,

and thus contributes with a paramagnetic response. The orbital contribution

is diamagnetic,
B = H − 4π|morb | < H,
because the induced currents cost energy. The susceptibility
∂M
χ≡ (6.40)
∂H
is correspondingly positive, χpara > 0, or negative, χdia < 0. In the fol-
lowing, we first study the Landau diamagnetism and then the Pauli spin
paramagnetism of a free electron gas.
7
Alternatively, we define m as the conjugate variable to H through the free energy
density g, m = −∂H g(T, h), where g = G/V and we write g rather then f since H is an
intensive variable.
P In fact, we will see later when considering spin-hamiltonians that the
term −H i µi ↔ −Hm plays the role of a Legendre transform from m to H, see Eq.
(8.1).
108 CHAPTER 6. FERMIONS

6.1.1 Landau diamagnetism

To begin with, we note that the diamagnetism of free electrons is a purely

quantum mechanical phenomenon—there is no classical analog to this effect.
This fact is formulated in the Bohr-VanLeeuwen theorem, which states that
the classical calculation of the magnetisation of free electrons gives zero.
Indeed, the calculation of the classical partition function
Z
d3N p d3N q h X [~ ~ qi )]2
pi − (e/c)A(~ i
ZN = exp −β + . . . (6.41)
N ! h3N 2m
i

can be reduced to the calculation of ZN in the free case by the translation

e~
p~i → p~i0 = p~i − A(~
qi ), d3 pi = d3 p0i .
c
The partition function ZN then is independent of H and thus (6.38) gives
m = 0. Hence, it takes a quantum mechanical analysis to obtain a finite
result.
Calculating the grand canonical partition function (6.1)
Y
Z= (1 + ze−βελ ) (6.42)
λ

requires us to compute the spectrum of (6.35) (we ignore the spin compo-
nent). The single particle problem

[~ ~ q )]2
p + (e/c)A(~
H= , HΨλ = ελ Ψλ (6.43)
2m
can be reduced to the displaced harmonic oscillator. We choose the gauge
~ = (−Hy, 0, 0)
A (6.44)
~ = (0, 0, H) according to H
that generates the correct field H ~ =∇
~ ∧ A.
~ The
eigenvalue problem then has the form

~2 h 2πi 2 i
− ∂x − Hy + ∂y2 + ∂z2 Ψλ = ελ Ψλ . (6.45)
2m Φ0

Separating variables, Ψλ = eikx x eikz z φ(y), we find the differential equation

governing φ(y) to be the Schrödinger equation for a displaced harmonic
oscillator
h ~2 1 i  ~2 kz2 
− ∂y2 + mωc2 (y − y0 )2 φ = ελ − φ, (6.46)
2m 2 2m
h eH Φ0
with ωc = H= and y0 = kx = `2 kx ,
mΦ0 mc 2πH
6.1. ELECTRONS IN A MAGNETIC FIELD 109

with the cyclotron frequency ωc and the magnetic length `. The latter intro-
duces a new length scale in the problem (besides the interparticle distance
or kF ). The eigenvalues and eigenfunctions are those of a harmonic oscillator
and we find
 1  ~2 kz2
ελ = ε(n, kz ) = ~ωc n + + , (6.47)
2 2m
2 2
Ψλ = eikx x eikz z Nn Hn [(y − y0 )/`] e−(y−y0 ) /2` ,

where Hn are the Hermite polynomials. With a quantization box Lx , Ly , Lz ,

Fig. 6.3: Orbit of width ` associated

y
with the lowest Landau level. The or-
bits associated with one Landau level
y0 cover all the plane. Higher Landau
l levels provide a finer resolution of dis-
tances along y with an increasing ex-
tension and number of nodes.
0 x

we have kz = (2π/Lz )nz , kx = (2π/Lx )nx , 0 < y0 = `2 (2π/Lx )nx < Ly , and
hence
Lx Ly HA ΦA
nx ≤ = = , (6.48)
2π`2 Φ0 Φ0
with the flux ΦA through the surface determining the degeneracy of the
Landau levels.

ρ ρ

0 H= 0 ε 0 H>0 ε

Fig. 6.4: Sketch p of the density of states for the free electron gas at H = 0
(left) ∝ (3εF /2n) ε/εF and at a finite field H > 0 (right) with the first
peak at ~ωc /2 followed by further equidistant peaks at (n + 1/2)~ωc .

The change in the density of states of electrons due to the presence of a

magnetic field is a consequence of a dimensional reduction. Indeed, the fully
extended plane-wave eigenfunctions of free electrons in three dimensions
(3D) is reduced to a confined (circular/Larmor-) orbit in 2D (with details
depending on the choice of the gauge) and a remaining plane wave in the
third dimension. As a result, the density of states (DOS) collapses to a
110 CHAPTER 6. FERMIONS

√
discrete spectrum (for the harmonic oscillator) plus a 1D DOS ∝ 1/ ε
DOS for the motion along z, see Fig. 6.4. Going to 2D, one then expects
the quantization into Landau levels to fully manifest itself in a discrete
spectrum, see Fig. 6.5.

ρ(ε ) ΦA Φ states

H= 0

ε
δ−peaks
ρ(ε )
h ωc
H> 0

h ωc / 2 ε

Fig. 6.5: Spectrum describing electrons in a magnetic field. The (constant)

DOS of free electrons moving in a plane collapses to the discrete spectrum
of 2D Landau levels. The dimensional reduction due to the magnetic field
localizes the free extended electrons into Larmor orbits. The weight in
the DOS over segments ~ωc is concentrated in the discrete levels with a
corresponding degeneracy.

With the degeneracy ΦA /Φ0 per Landau level and the inclusion of all
Landau levels, one obtains a complete set of orthogonal states. Indeed, in
2D, the orbits Ψλ=[n,kx (y0 )] with fixed n and varying kx cover the entire plane
with a resolution given by the Hermite function Hn [(y − y0 )/`], see Fig. 6.6.
The n = 0 orbitals have the extent of the magnetic length, hy 2 i1/2 ∼ `.
The size of the orbital at εF , n = εF /~ωc , is given by the Larmor radius
rL = vF /ωc = `(kF `) (note the increasing powers of y in Hn ).
n= 0 n= 1 n= 2
l

y0 y0 y0

Fig. 6.6: Coverage of the plane with orbits of increasing resolution associ-
ated with higher Landau levels.

At the boundaries, the orbitals are deformed by the box potential and
their energy is raised, see Fig. 6.7. Within a semi-classical picture, the orbits
transform from circular-localized in the bulk to extended skipping orbits at
the boundary. These skipping orbits are the simplest semi-classical picture
of the edge-currents in the quantum Hall effect (QHE).
PΦA /Φ0
By superposition i=1 ani Ψn (y −y0i ) of degenerate states within each
6.1. ELECTRONS IN A MAGNETIC FIELD 111

x Fig. 6.7: 2D metallic strip ex-

tending from 0 to Ly along y. The
l
wave functions Ψn [y0 (kx )] per Lan-
dau level n cover all of the plane
when varying 0 < kx < Ly /`2 →
0 y0 Ly y 0 < y0 = `2 kx < Ly . As the or-
bits approach the boundaries at 0
ε and Ly , they are squeezed and their
energy is raised up.

0 y0 y

Landau level, we can generate a new set of eigenfunctions for the n-th level.
For example, with the gauge A~ = (B/2)(y, −x, 0), one obtains the circular
solutions in Fig. 6.8

2∂ m ∂ n −z z̄/2`2
Ψm,n (z = x + iy) ∼ ez z̄/4` e , (6.49)
∂ z̄ m ∂z n
with ∂z = ∂x − i∂y and ∂z̄ = ∂x + i∂y ,
 1
HΨmn = ~ωc n + Ψmn ,
2
n = energy q# and m = angular momentum q#. (6.50)

y y
circular

x x
lattice

Fig. 6.8: Different gauges and superpositions produce other sets of Lan-
dau level functions, e.g., the circular solutions (left) or the lattice solutions
(right).

Using the spectrum (6.47), we can calculate the grand canonical partition
112 CHAPTER 6. FERMIONS

function Z: According to (6.1), we have to calculate

Y
Z = (1 + ze−βελ )
λ
ΦA /Φ0
X XX
log 1 + ze−βε(n,kz )


log Z =
1 n kz
∞ Z
ΦA X dkz
log 1 + ze−βε(n,kz )


= Lz
Φ0 2π
n=0
ΦA X ∞ dkz h β~2 k 2 i
Z h i
z
' Lz z exp − exp −β~ωc (n + 1/2)
Φ0 n −∞ 2π | {z2m } | {z }
→1/λ −β~ωc /2
→e
1−e−β~ωc

zHV exp(−~ωc /2kB T )

= .
Φ0 λ 1 − exp(−~ωc /kB T )

with the three sums over the level degeneracy, the Landau levels, and the
kinetic energy in the z direction. In addition, we have assumed a high
temperature approximation when expanding the logarithm. In the high
temperature limit, the density relates to the fugacity via z = nλ3 , see (6.14).
For small fields, x = ~ωc /2kB T = λ2 /4π`2  1, we can expand and obtain
the following results,

nλ3 V H 1 − x + x2 /2
log Z '
Φ0 λ 1 − (1 − 2x + 4x2 /2 − 8x3 /6)
nλ2 H 1
= V (1 − x + x2 /2) · (1 + x + x2 /3)
Φ0 2x
nλ2 H Φ0 mkB T  x2 
= V 1 −
Φ0 2π~2 H 6
 2
π ~ λ H 2 2   1 λ4 
' nV 1 − = N 1 − , (6.51)
12 mkB T Φ20 96π 2 `4
kB T ∂ π n~2 λ2 H
m = log Z = −
V ∂H 6 m Φ20
nµ2B H nµB λ2
= − =− , (6.52)
3kB T 12π `2
∂m nµ2B
χL = =− , (6.53)
∂H 3kB T

where `2 = Φ0 /2πH, λ2 = 2π~2 /mkB T , and µB = e~/2mc; in the high

temperature and small field limit, nλ3  1 and λ2 /2π`2  1.
The calculation is more difficult for the degenerate Fermi gas. Dimen-
sionally, the result must scale as χ ∼ (n/εF )µ2B . For small H one finds (see
6.1. ELECTRONS IN A MAGNETIC FIELD 113

Fig. 6.9 for an overall sketch of χL (T ))

1 n 2
χL = − ρ(εF )µ2B = − µ . (6.54)
3 2εF B

Fig. 6.9: Temperature de-

n 2
µ
2εF B pendence of the Landau dia-
magnetic susceptibility in a
−χ L Fermi gas.
Curie
~ 1 /T

0 εF T

Note that Landau diamagnetism is a small effect,

|δH| 1 e2
= 4π|χL | = kF 2 ∼ 10−5 (6.55)
H 3π mc
with kF ∼ 1/Å and the classical electron radius e2 /mc2 ∼ 3 · 10−5 Å.
An interesting question is the one about the extent of the regions near the
surface where the currents flow.

6.1.2 Pauli spin paramagnetism

We again consider (6.35), but now ignore the gauge field in the kinetic energy
(we write µ~ i = −gµB S~i ≈ −µB~σi with ~σ the Pauli matrices, g ≈ 2, and the
~ of the spin is included in µB ),
X p2 
i
H= + µ B σi H , σi = ±1,
2m
i

~ = (0, 0, H). The single-particle energies depend on

where we have chosen H
the momentum p~ and spin variable σ,

p2
εp~,σ = + σµB H, (6.56)
2m
with the occupation of these states given by np~,σ ∈ {0, 1}. DefiningP
the total
number of electrons with up (+) and down (−) moments µi , N± = p~ np~,∓1 ,
the total energy of a state {np~,σ } is given by the expression
X p2
E[{np~,σ }] = np~,σ − µB H(N+ − N− ). (6.57)
2m
p
~,σ
114 CHAPTER 6. FERMIONS

We calculate the canonical partition function for N = N+ + N− particles,

X N
X h i
ZN = e−βE[{np~,σ }] = exp βµB H(2N+ − N ) (6.58)
{np~,σ }N N+ =0
X h X p2 i X h X p2 i
× exp −β np~,−1 exp −β np~,x−1 .
2m 2m
{np~,−1 }N+ p
~ {np~,1 }N− p
~

Using the partition function

X h X p2 i
Z0N = e−βF0 (N ) = exp −β np~ (6.59)
2m
{np~ }N p
~

for N spinless fermions of mass m, we can recast (6.58) to

M
X
ZN = e−βµB HN e2βµB HN+ e−βF0 (N+ )−βF0 (N− ) . (6.60)
N+ =0

As usual, it is the largest term in the sum that plays a dominant role: Using
f (N+ ) = 2µB HN+ − F0 (N+ ) − F0 (N − N+ ), we have to find the value
N+ = N̄+ maximizing f ,
∂f
= 0. (6.61)
∂N+ N̄+
With ∂N F0 = µ the chemical potential of the ideal spinless Fermi gas, see
(6.14) and (6.18),

kB T log nλ3 < 0,
µ(n) = 
π 2

k T
2  (6.62)
εF (n) 1 − B
 0,
12 εF

the condition (6.61) tells us that the chemical potentials of the spin ↑ and
spin ↓ systems differ by the magnetic energy 2µB H

µ+ − µ− ≡ µ(N̄+ ) − µ(N − N̄+ ) = 2µB H. (6.63)

Using N̄± = (1/2 ± ε)N and (6.63), we find the relative disbalance

µB HV
ε= . (6.64)
∂n µ N
in the spin population. For the degenerate Fermi gas
∂n 1 3n
= ρ(εF ) = ,
∂µ 2 4εF
6.1. ELECTRONS IN A MAGNETIC FIELD 115

Fig. 6.10: The chemical po-

E tentials for momentum ↑ and
µ+ ↓ electrons in a homgeneous
µ− magnetic field H differ by the
magnetic energy 2µB H.

where the additional factor 1/2 appears since we consider the individual spin
configuration separately. The spin disbalance then becomes

2µB H 3 n 3 µB H
N̄+ − N̄− = 2εN = N= N
n 4 εF 2 εF
and the magnetization density takes on the value

(N̄+ − N̄− )µB 3 µ2B H

m= = n. (6.65)
V 2 εF
For the Pauli spin susceptibility, we find the result
3n 2
χP = µ = ρ(εF )µ2B , (6.66)
2εF B
three times the value of the Landau diamagnetism,

χP = −3χL . (6.67)

For the non-degenerate Fermi gas we have the Curie law behavior

∂n 1 ∂ 1 µ/kB T n µB H
= 3
e = , N̄+ − N̄− = N,
∂µ 2 ∂µ λ 2kB T kB T
µ2 H n 2
m = B n, χP = µ . (6.68)
kB T kB T B
These results are summarized in the sketch of Fig. 6.11.

3 n µ2
Fig. 6.11: Temperature de-
2 εF B pendence of the paramagnetic
Pauli spin-susceptibility in a
χP Curie
Fermi gas.
~ 1/ T

0 εF T
116 CHAPTER 6. FERMIONS

A neat argument is that in a real Fermi gas with repulsive interaction

a ferromagnetic instability8 may arise. The particles with parallel spins
avoid each other due to Pauli’s exclusion principle and thus have to pay less
interaction energy.

8
The finite magnetization density |m(H = 0)| > 0 is a consequence of spontaneous
symmetry breaking where the spins arrange in parallel along some direction in the absence
of an external field.
Chapter 7

Bosons

We study two types of bosonic systems: Photons are the quantum par-
ticles constituting the classical waves of electromagnetic radiation—when
discussing photons in a general context, their particle number is not con-
served and their chemical potential is zero, µ = 0, meaning the cost of
adding a photon is always its energy. Similar is the case for the phonons,
the quantum particle associated with sound waves in a material. Quantum
engineered photonic systems involving cavities and waveguides are an area of
current research, in particular, the non-equilibrium aspect that is associated
with the omnipresent drive and dissipation in this type of systems—steady
states then replace the equilibrium states. The second class of bosons are
massive particles, in particular atoms, that are experimentally studied in
the form of trapped cold gases; below, we will discuss the properties of an
ideal Bose-gas and its Bose-Einstein condensation.

7.1 Photons

The quantum mechanical description of the electromagnetic field is given by

the photon number states

|n~k1 ,λ1 , n~k2 ,λ2 , . . . , n~ki ,λi , . . .i (7.1)

in Fock space, where the ~k, λ-mode is occupied n~k,λ ∈ N0 times. The pa-
rameters ~k and λ describe the wavevector and the polarization of the mode.
In a quantization volume L3 , the wavevectors are given by

~k = 2π ~n, ~n ∈ Z3 . (7.2)
L

117
118 CHAPTER 7. BOSONS

For every ~k there exist two polarizations (either linear or circular), specified
by the polarization vectors

~εx , ~εy ; ~ε+ , ~ε− ; |~ελ | = 1; ~ελ · ~kλ = 0. (7.3)

The corresponding field is transversal 1 . The gauge degree of freedom is a con-

sequence of the masslessness of photons and reduces the (spin-)polarizations
to the values ±~ (Photon = masseless spin-1 boson with Sz 6= 0). A further
consequence of the lack of mass is the linear dispersion
p
E = p2 c2 + m2 c4 |m=0 = pc,
E = ~ω, ωk = ck, p = ~k. (7.4)

Energy, momentum and spin momentum of the photon field are given by
X
E[{n~k,λ }] = ~ωk n~k,λ ,
~k,λ
X
P~ [{n~k,λ }] = ~~k n~k,λ ,
~k,λ
X
~ ~ }] =
S[{n ~k̂ (n~k,+ − n~k,− ). (7.5)
k,λ
~k

The corresponding operators are obtained by substituting the bosonic cre-

ation and annihilation operators

n~k,λ = a~† a~k,λ , [a~k,λ , a~† 0 ] = δ~k~k 0 δλλ0 .

k,λ k ,λ0

We consider a box (cavity), whose walls are at a fixed temperature T

and bring the radiation to equilibrium (alternatively, use a black body which
captures all radiation and radiates a (different) spectral distribution deter-
mined by its temperature). We investigate the spectral distribution of the
radiation in the box by calculating the grandcanonical partition function2
Z Y 1 X
Z= ; log Z = −2 log(1 − e−β~ωk ). (7.6)
1 − e−β~ωk
~k,λ ~k

1~ ~ = 0 in the Coulomb gauge; the physical fields are transverse, with Coulomb’s
∇·A
~ ~
law ∇ · E = 0 and Gauss’ law ∇ ~ ·B ~ = 0.
2
Alternatively consider a (classical) harmonic oscillator and calculate its (canonical)
partition function by a gaussian integral over phase space (p, q). Then go over to the
quantum mechanical oscillator with energies En = ~ωn and calculate the corresponding
result Z = Tr[exp(−βH)] = 1/(1 − exp[−β~ω)] (without quanta/particle interpretation);
for many independent oscillators with frequencies ωk , we find the canonical partition
function Z = Πk Zk , which is exactly the result found in (7.6). The equality Z = Z (or
F = Ω) follows from (4.20) with z = 1, µ = 0. Thus, we can interpret the quanta of
the harmonic oscillator as bosonic particles with chemical potential µ = 0. The linear
spectrum in n of the harmonic oscillator is an essential part of this argumentation.
7.1. PHOTONS 119

The mean occupation numbers are easily found,

1 ∂ log Z 1
hn~k,λ i = − = β~ω ,
β ∂~ωk e k −1
(7.7)

and the internal energy is to be calculated from

∂ X
U = − log Z = ~ωk hn~k,λ i (7.8)
∂β
~k,λ

with the equation of state following immediately from U (use that Z = Z

and p = −∂V F , F = −kB T ln Z)

1 ∂ 1 ∂ U
p= log Z = − log Z = . (7.9)
β ∂V 3V ∂β 3V

Here we used that Z depends on V through x = βωk : With ω~k = 2πc|~n |V −1/3 ,
we find ∂V = (∂V x)∂x and ∂V x = −x/3V , ∂x = ωk−1 ∂β , thus ∂V =
−(β/3V )∂β . Note that (7.9) (m = 0) is different from (8.41) (m > 0)
(for massive particles, we have εp ∝ V −2/3 and ∂V = −(2β/3V )∂β ). To
calculate U one substitutes
Z ∞ Z ∞
d3 k
Z
X V →∞ V 2 V
→ 2V = 2 dk k = 2 3 dω ω 2 (7.10)
(2π)3 π 0 π c 0
~k,λ

and finds the Stefan-Boltzmann law,

Z ∞ Z ∞
V~ ω3 V~ x3
U = dω = dx
π 2 c3 0 eβ~ω − 1 π 2 c3 ~4 β 4 0 ex − 1
| {z }
π 4 /15

π 2 (kB T )4 π2 V
= V = kB T, (7.11)
15 (~c)3 15 λ3

where we have defined the thermal wavelength λ through 2πkB T = ~ck =

hc/λ in the last equation. Comparing this last expression U ∼ (V /λ3 )kB T
with the ideal gas result U ∼ N kB T , we see that the particle number N
corresponds to the ratio V /λ3 , i.e., the ‘particles are created by the temper-
ature’ rather than being there ‘from the beginning’. The specific heat per
unit volume continuously grows with temperature,

4π 2 kB4 T 3 T →∞
cV = → ∞ (7.12)
15 (~c)3

as steadily additional modes get activated when temperature increases.

120 CHAPTER 7. BOSONS

Of particular significance is Wien’s displacement law. We write the (clas-

sical) radiation law in the form (u is the spectral energy density and is given
by the density of states times the thermal energy kB T per mode from the
equipartition theorem)
ω2
u(ω, T ) = kB T f (β~ω), (7.13)
π 2 c3
separating the classical Rayleigh-Jeans law u ∝ ω 2 from the Planck factor
f = β~ω/(eβ~ω − 1) (the modification due to the quantization of energy);
the latter cuts off of the spectral energy density at large frequencies, see Fig.
7.1. Wien’s displacement law, the shift of the maximum in u(ω) with the

u Fig. 7.1: The Rayleigh-Jeans

law valid at low frequencies
is cutoff by the Planck fac-
tor f (β~ω). Their competi-
RJ tion leads to Wien’s displace-
ment law, telling that the maxi-
mum in u(ω) shifts linearly in T .
0 ω
T
temperature T follows from the requirement ∂ω u = 0,
→ [2f + xf 0 ]|xm = 0, β~ωm = xm → ωm ∝ T
since xm is merely a number determined by the shape of the function f (x).
The Stephan-Boltzmann law (7.11) can be derived classically: With the
Maxwellian radiation pressure3 p = u/3 and the internal energy U = V u(T ),
we find the entropy differential
CV 1  ∂U 
dS = dT + + p dV
T T ∂V
V ∂u 4u
→ dS = dT + dV, (7.14)
T ∂T 3T
and using the condition for integrability ∂ 2 S/∂T ∂V = ∂ 2 S/∂V ∂T , we arrive
at the differential equation
1 ∂u 4 ∂ u
=
T ∂T 3 ∂T T
du dT
→ =4 → ln u = 4 ln T + const. (7.15)
u T
→ u = aT 4 . (7.16)
3
Average the pressure (2S/c) cos2 θ, S the Pointing vector, over a half-sphere in order
to describe the reflection (at an angle θ) from a boundary.
7.1. PHOTONS 121

Planck’s derivation of the radiation law

~ω 3 1
u(ω, T ) = 2 3

(7.17)
π c exp ~ω/kB T − 1

makes use of an ingenious trick:4 He considered a charged harmonic oscilla-

tor of frequency ω coupled to a radiation field of temperature T and found
the relation
πc3
U = UHO = u(ω, T ), (7.18)
2ω 2
between the spectral weight u(ω, T ) of the radiation field and the time aver-
aged energy U of the oscillator. Next, he considered the entropy at constant
volume
dU dS 1
dS = → = .
T dU T
He makes use of a more precise Wien displacement law that was based on
improved measurements of Paschen,

u(ω, T ) ∼ ω 3 e−αω/T , (7.19)

and using (7.18) for large frequences ω, he found

dS 1 1 d2 S 1
= =− ln U, 2
=− . (7.20)
dU T αω dU αωU

For small ω, he makes use of the equipartitian theorem U = kB T for the

harmonic oscillator

dS 1 kB d2 S kB
→ = = , 2
= − 2. (7.21)
dU T U dU U

Interpolating, one obtains

d2 S 1
=− (7.22)
dU 2 αωU + U 2 /kB

Using that ∂U2 S = −T −2 ∂U T , one finds a differential equation for U (T ),

which is solved by
αωkB
U = αω/T . (7.23)
e −1
Making use of (7.18) one finally arrives at Planck’s law.
4
One may find more about this sketch in arXiv:quant-ph/0010008, ”...ich dachte mir
nicht viel dabei...”, Planck’s uneven way to the radiation formula.
122 CHAPTER 7. BOSONS

7.2 Phonons

Quite often, the vacuum state of a condensed matter system can be viewed
as an elastic deformable medium. The excitations of this medium, the
phonons, then play the same role as the photons in the electromagnetic
vacuum. The important difference is that condensed matter systems are
composed of atoms: The continuum nature of the vacuum in electromag-
netism is replaced by the discrete atomic lattice of condensed matter. In-
stead of infinitely many modes the condensed matter system has 3N modes
corresponding to the 3N degrees of freedom of the N atoms. The spectrum
in momentum space is reduced to the Brillouin zone (BZ) due to the fact that
large wave numbers k > Kn (with Kn the reciprocal lattice vectors) do not
make sense because such modes resolve the (inexistent) structure between
the lattice sites. Furthermore, the spectrum at the BZ edge is deformed as
the phonons “scatter on the lattice” and are thus reflected ∂ω/∂k|BZ = 0.
Thirdly, there are now “3 polarisations”, two transversal and one longitu-
dinal, for every ~k-vektor, transversal shearing modes and longitudinal com-
pression modes. The latter correspond to the hydrodynamic sound in gas
(longitudinal compression modes)—the shearing modes are a characteristic
of the hard condensed matter system (they appear neither in liquids nor in
gases). Note that the classification of longitudinal (k) and transversal (⊥)
(to ~k ) directions only make sense for special directions of ~k in the BZ.

ω ω
+/−

ck
2x

0 k − π/ a 0 k [1,0,0] π/a

Fig. 7.2: Summary of differences between photons und phonons. Extended

k-space versus finite Brillouin zone, overall linear double-degenerate spec-
trum versus spectral deformation at the BZ edges and three modes, double
degenerate transverse plus longitudinal, velocity of light versus velocity of
sound at small k, abour 104 to 105 times smaller.

For a simple isotropic elastic medium, we can model the phonon disper-
sion by

ωλ = cλ |~k | (7.24)
7.2. PHONONS 123

and the density of states has the form (with d3 k ∼ 4πk 2 dk)
X  L 3
ρ(ω)dω ' 4πk 2 dk
2π
λ
X V 1
2
= ω dω
2π 2 c3λ
λ
3V ω 2
= dω, (7.25)
2π 2 c̄3
3 2 1
= + .
c̄3 c3⊥ c3k

We find the size of the isotropic Brillouin zone through the constraint
Z ωmax 3
V ωmax V 3
ρ(ω)dω = 2 3
= 2 kBZ = 3N
0 2π c̄ 2π
3
→ kBZ = 6π 2 n, cf. (6.22)
ωmax = c̄kBZ . (7.26)

The internal energy follows from Eqs. (7.11) and (7.26),

Z ωmax
3V ~ω
U = dω ω 2 β~ω
2π 2 c̄3 0 e −1
Z β~ωmax
U 3~ 1 t3
→ = dt . (7.27)
N 2π 2 nc̄3 (~β)4 0 et − 1
| {z }
9(kB T )4 /(~ωmax )3

The Debye function D(x) behaves as

Z x (
3 t3 1 − 3x/8 + x2 /20 − . . . , x  1,
D(x) ≡ 3 dt t = (7.28)
x 0 e −1 π 4 /5x3 + O(e−x ), x  1,

for small and large arguments x and we obtain the result


U 1 − 3 Θ + . . . , T  Θ,
8T
≈ 3kB T π4  T 3 (7.29)
N 
5 Θ , T  Θ.

Here, we have introduced the Debye temperature

kB Θ = ~ωmax . (7.30)

At temperatures T above Θ, all modes are excited and we find the classical
result of Dulong-Petit for the specific heat of a solid

CV = 3N kB , T > Θ, (7.31)
124 CHAPTER 7. BOSONS

3Nk B Dulong−Petit Fig. 7.3: Specific heat of

phonons. As the spectrum
is cutoff at the Debye energy
CV kB Θ (at the BZ), the specific
heat saturates at the Dulong-
Petit value.
Debye
0 Θ/4 T

(6 degrees of freedom per bound atom). For T  Θ, excitations are added

continuously starting from zero and we obtain the (SB)-Debye result

12π 4  T 3
CV = N kB , T < Θ. (7.32)
5 Θ
Here is a collection of Debye temperatures

Ag 215 K
Al 428 K
Au 170 K
Cu 343 K
He 25 K
C (diamond) 2250 K
C (graphite) 413 K
Si 645 K
MgO 750 K
SiO2 (quartz) 470 K
NaCl 310 K

Small values of Θ describe “soft”, while large values refer to “hard” materi-
als.
In general, every excitation/degree of freedom contributes to CV ; it is
their spectral density as it follows from the dispersion ω ∼ k n and the
dimensionality of the system (→ ρ ∼ k d−1 ) that determines the temperature
dependence of CV . Simple dimensional considerations, valid for small T such
that the cutoff in the integral is on T and not on ρ, then provide the scaling
behavior

1 ω∼T
Z Z
d ~ω
U ∼ d k β~ω ∼ dk ρ,
e −1 β
Z ω∼T Z
CV ∼ dk ρ ∼ dk k d−1 ∼ k d ∼ ω d/n |ω∼T ∼ T d/n .
7.3. IDEAL BOSE-GAS: BOSE-EINSTEIN-CONDENSATION 125

When all modes are exhausted (with the integral cutoff by ρ) we have U ∝ T
and CV is constant. For phonons in 3D, we have n = 1, d = 3,

CV ∼ T 3 . (7.33)

Massive Bosons in 3D have n = 2 and hence we expect CV ∼ T 3/2 , see Eq.

(7.54) below. Similar estimates can be made for Fermions when properly
accounting for the Pauli blockade.

7.3 Ideal Bose-gas:

Bose-Einstein-condensation

We recapitulate the results from Sec. 5.2.2:

Partition function:
Y
Z(V, T, z) = (1 − ze−βεp )−1 . (7.34)
p
~

Equation of state:
kB T kB T kB T
p= log Z = 3 g5/2 (z) − log(1 − z). (7.35)
V λ V
Particle density:
z ∂ 1 1 z
n= log Z = 3 g3/2 (z) + . (7.36)
V ∂z λ V 1−z
Occupation number:
1
hnp~ i = . (7.37)
eβ(εp −µ) −1
Z ∞ ∞
4 2
X z`
g5/2 (z) = − √ dx x2 log(1 − ze−x ) = 5/2
,
π 0 `
`=1
Z ∞ 2 ∞ `
4 x X z
g3/2 (z) = √ dx x2 = . (7.38)
π 0 e /z − 1 `3/2 `=1

Again, we first solve the fugacity relation

λ3 z
nλ3 = g3/2 (z) + (7.39)
V 1−z
for z. The function g3/2 (z), see Fig. 7.4, can be expanded in the range
0 ≤ z ≤ 1 using the series expansion ` z ` /`3/2 ,
P

( √ √
z + z 2 /2 2 + z 3 /3 3 + . . . , z  1,
g3/2 (z) = (7.40)
ζ(3/2) = 2.612, z = 1.
126 CHAPTER 7. BOSONS

Fig. 7.4: Sketch for g3/2 (z)

2.612 in the interval [0, 1], linear at
small z and saturating at z =
1 with infinite slope. Small
g3/2 densities nλ3 < 2.612 can be
3 brought to the normal phase
n λ < 2.612
without condensation as z <
1.
0 z 1

The slope of g3/2 (z) diverges in z = 1. We see that for nλ3 < 2.612, equation
(7.39) can be satisfied with a value z < 1 deriving from nλ3 = g3/2 (z) with
the second term playing no role in the limit V → ∞. In contrast when
nλ3 > 2.612, we find a solution to (7.39) requires that z = 1 − O(1/V ) and
the second term has to compensate the contribution nλ3 − 2.612. In the
limit V → ∞, we find z → 1 and using (7.39), the ground state occupation
hn0 i = hnp~=0 i

z V
hn0 i = = 3 (nλ3 − nBE λ3 ) ∝ V (7.41)
1−z λ | {z }
=2.612

assumes a macroscopic value hn0 i ∝ N : This is the phenomenon of Bose-

Einstein-condensation (BEC). One often talks about condensation in mo-
mentum space: Instead of condensing in a liquid drop (real space), the con-
densation occurs in a quantum state with definite p~. We will come back to
the peculiarities of the condensate later on. The constraint
 2π~2 3/2
nλ3 = n = 2.612 (7.42)
mkB T
defines the critical temperature for the condensation

~2 n2/3 4π
kB TBE = · 2/3
(7.43)
2m
| {z } |2.612
{z }
Energie 6.63

at constant density n, or, alternatively, the critical density

2.612
nBE = (7.44)
λ3
at constant temperature T ; both criteria signal the onset of condensation.
These strong quantum effects appear when ∼ 2.6 particles can be found in a
de Broglie volume λ3 . For T → 0 the condensation is complete and hn0 i/N =
1. At finite temperatures, the fraction of particles in the condensate is found
7.3. IDEAL BOSE-GAS: BOSE-EINSTEIN-CONDENSATION 127

2 n 0 /N
λ α1 /T BE−condensation
1
2/3
1/ TBE α 1 / n

0 n 0 TBE T

Fig. 7.5: Left: particles Bose condense when sufficiently many particles
occupy the de Broglie volume λ3 . Right: the fraction of particles in the
condensate grows continuously with lowering temperature T .

from (7.41), see Fig. 7.5 for a sketch,

(
hn0 i (1 − nBE /n) = [1 − (T /TBE )3/2 ], T < TBE ,
= (7.45)
N 0, T > TBE .

One may ask why only the p~ = 0 state condenses, as the momentum
p1 = 2π~/L also goes to 0 for V = L3 → ∞. Thus, we would have

N 1 hn0 i hn1 i
= 3 g3/2 (z) + + (7.46)
V λ V V
P
with the last term the next in the sum p~ hnp~ i. We fix T below TBE and

np ~N Fig. 7.6: The condensation with

macroscopic occupation is con-
T 0 fined to only the ground state
p = 0; the occupations hnp>0 i
for finite momenta remains finite,
i.e., do not scale with N .

0 p

determine hn1 i from Eq. (7.37) for a large volume and βε1  1 (we use
hn0 i/V = 1/α > 0),

hn1 i 1 1 z=eβ0 /(1+α/V ) 1/V T V →∞

= βε
≈ ∝ 1/3 → 0,
V V e /z − 1
1 α/V + β(ε1 − ε0 ) V

i.e., in the thermodynamic limit V → ∞ only the ground state condenses;

while p~ = 0 is occupied macroscopically (hn0 i ∝ N ) all hnp>0 i remain finite,
see Fig. 7.6.
128 CHAPTER 7. BOSONS

Knowing the fugacity z as a function of density n and temperature T ,

g3/2 (z) n  T 3/2

BE
= = , nλ3 < 2.612, T > TBE , n < nBE , (7.47)
g3/2 (1) nBE T
z = 1, nλ3 > 2.612, T < TBE , n > nBE ,

we can evaluate the equation of state (see 7.35) of the ideal Bose-gas
(
kB T  nλ3  T > TBE ,
p = g (z) → k T n 1 − √ + . . . , (7.48)
λ3 5/2
B
4 2 n < nBE ,
(
kB T kB T T→0 T < TBE ,
p = 3
g5/2 (1) = 1.342 3 ∝ T 5/2 → 0, (7.49)
λ λ n > nBE .

This result defines the transition line pBE (T ) in the p–T diagram, see Fig.
7.7,
 m 3/2
pBE (T ) = g5/2 (1) (kB T )5/2 .
2π~2

p Fig. 7.7: Transition line for

start the Bose Einstein condensation
condensation in the p–T diagram.
pBE ~ T 5/2

gas

0 T

In a further step, we determine the p–v diagram for the ideal Bose gas.
For a volume v < vBE (T ) (v = 1/n is the volume per particle, n > nBE (T )),
we have p = pBE = const. This is the vapour pressure above the conden-
sate with the gas-phase assuming a fraction nBE (T )/n of the density (the
remaining particles are in the condensate and generate no pressure). At
the transition, nλ3 = λ3 /v = g3/2 (1) and hence TBE (v) ∝ v −2/3 . Using the
expression (7.49) for the pressure at the transition, we find the transition
line in the p–v diagram,

kB TBE (v) 2π~2 g5/2 (1)

pBE (v) = g 5/2 (1) = → pBE v 5/3 = const.
λ3 (TBE (v)) 5/3
m v [g3/2 (1)]5/3

(7.50)
7.3. IDEAL BOSE-GAS: BOSE-EINSTEIN-CONDENSATION 129

p BEC transition line Fig. 7.8: p–v diagram for the

Bose System with a gas phase
gas and a Bose-Einstein condensate.
Curves shown are isothermes
C
G with fixed temperature T .
T 0
p BE ~ v −5/3
gas + condensate
0 vBE ( T ) v = 1/ n

The phase diagrams in the Figs. 7.7 and 7.8 resemble those for the real
gas with the liquid–gas transition. We can make an analogy between Bose
gases and real classical gases with the Bose condensed (C) phase correspond-
ing to the liquid and the Bose gas phase (G) with p > 0 corresponding to
the usual gas phase (G). The condensate replacing the liquid, however, has
different, atypical properties. In particular, it does not sustain a pressure,
and thus cannot oppose any compression. One then may observe the fol-
lowing similarities and differences: First, the transition between G and C in
Fig. 7.8 obeys the Clausius-Clapeyron relation with the volume jump from
vG = vBE (T ) to vC = 0, ∆v = vBE (T ). Correspondingly, one obtains the
latent heat (per particle joining the condensate)

dpBE 5 g5/2 (1)

`= T ∆v = kB T, (7.51)
|dT
{z } 2 g3/2 (1)
(5/2)g5/2 (1) kB /λ3

where we have used vBE /λ3 = 1/g3/2 (1) for the volume jump. On the other
hand, we cannot reach the p > pBE (T ) region of the p-T diagram. If we
want to cool further below TBE , there is no state that can resist a pressure
and thus the system follows the transition line pBE (T ) until it reaches p = 0,
T = 0 (up to the impossibility to reach T = 0 exactly), see Fig. 7.9. The full
condensation can then be realized by i) ‘pressing’ the gas into the condensate
at p und T (reduce the volume), or ii) reducing the temperature at constant
volume (stretching the de Broglie wavelength beyond the system volume);
naturally, variations of these two methods are possible.
Further, we show (see below) that CV (T ) only shows a kink. The inter-
pretation of the BEC as a phase transition of first or second order according
to the Ehrenfest classification is not really applicable, but note that L&L
classify the transition as being second order.5
For the internal energy U , the entropy S, and the specific heat CV , one
5
There is no finite free energy surface that defines a finite slope ∂p g or ∂T g below the
pBE (T ) transition line.
130 CHAPTER 7. BOSONS

Fig. 7.9: The Bose gas

p p BE ( T ) condenses completely by com-
pressing the gas into the con-
forbidden region condensation
densate at TBE and fixed pres-
via n ,v 0

8
sure or by lowering the tem-
perature T → 0 at fixed vol-
condensation ume and following the transi-
via T 0.
tion line pBE (T ).
0 T

finds

U 3  3 kB T3 g5/2 (z),
2 nλ
= pv = (7.52)
N 2  3 kB T g (1),
3 2 nλ 5/2

 5 g5/2 (z)
S 2 nλ3 − log z,
= (7.53)
N kB  5 g5/2 (1) ,
2 nλ3
 g (z)
 15 5/2 9 g3/2 (z)
CV 4 nλ3 − 4 g1/2 (z) ,
= (7.54)
N kB  15 g5/2 (1)
4 nλ3
.
The internal energy appears from the pressure and U = (3/2)pV . The
entropy S follows from
G = N µ = N kB T log z,
∂G
S = − , (7.55)
∂T
∂z 5z g5/2 (z)
= − , (7.56)
∂T 2T g3/2 (z)

where the last equation is derived from ∂ T g5/2 (z) = pλ3 /kB T (using the
 

relation ∂z gn = gn−1 /z). The specific heat is found via

∂U
CV , =
∂T
∂z 3z g3/2
= − , (7.57)
∂T 2T g1/2

and making use of ∂ T g3/2 (z) = nλ3 . The dependence CV (T ), see Fig.
 

7.10, shows a kink at the transition.

In the condensate phase the entropy is (we make use of Eq. (7.45))
S  T 3/2 nBE
= s= s = [1 − hn0 i/N ]s
N TBE n
7.3. IDEAL BOSE-GAS: BOSE-EINSTEIN-CONDENSATION 131

C/Nk B
Fig. 7.10: Specific heat of a
BEC with a cusp at TBE .
3/2

T 3/2
0 TBE T

with the entropy per particle in the gas phase

g5/2 (1) 5
s= kB , (7.58)
g3/2 (1) 2

i.e., the entropy S (and the pressure p) are generated purely by the gas phase.
For T → 0 the gas phase vanishes along with the entropy and the specific
heat, in accordance with the third law of thermodynamics. The latent heat
` = T ∆s = T s originating from (7.58) agrees with (7.51). Note that the
latent heat and volume jump are realized when hitting the transition point
pBE (T ) and either condensing the gas by further decrease of p and T along
the transition line or by squeezing the gas into the condensate at fixed p and
T . Thereafter, there is no new thermodynamic phase which can be further
cooled with a new thermodynamic Gibbs function, i.e., we cannot calculate
an entropy/volume jump as a jump in derivatives of two free energy surfaces
across the transition line as is the case for a usual first-order transition.
132 CHAPTER 7. BOSONS
Chapter 8

Magnetic order

On the way to understanding the general theory of phase transitions, the

study of magnetic systems is particularly useful. We first define the generic
models aapearing in the context of (local moment) magnetism—the Heisen-
berg, XY, and Ising models, restricting ourselves to the classical description.
We discuss their mean field theory (MF theory) and emphasize its (super1 )
universal behaviour. We then study the individual models in more detail
and show that fluctuations destroy various MF transitions: The Ising model
(with the dimension n = 1 of the spin variable/order parameter) shows no
transition in d = 1 space dimensions: the Ising chain is paramagnetic for all
T > 0 and orders only at T = 0. The XY model (with a two-dimensional
spin field n = 2) for the first time shows quasi-order in d = 2 and the
first transition for the Heisenberg model (n = 3) is found in d = 3. This
discussion demonstrates how fluctuations of the order parameter (OP) can
destroy phase transitions in systems with low space dimensions, although
MF theory predicts their existence in every dimension. Hence, fluctuations
are particularly strong in low space dimensions d = 1 and d = 2. The ab-
sence of some of these transitions is the content of the theorem of Hohenberg,
Mermin, Wagner, and Berezinskii. We will get to know several mechanisms
that destroy (quasi-)ordered phases: The generation of topological excita-
tions: domain walls in the d = 1 Ising model and free vortices in d = 2 XY
models. Long wave length excitations (← Goldstone modes) are responsible
for the reduction of order in the d = 2 XY model (order → quasi-order)
and the destruction of order in the d < 2 Heisenberg model. For the d = 2
Heisenberg model, another type of topological excitations, Skyrmions, have
to be considered in addition to spin fluctuations.

1
I.e., the critical exponents do not depend on the dimensionality d and n of space and
of the order parameter.

133
134 CHAPTER 8. MAGNETIC ORDER

8.1 Generic models of local magnetism

8.1.1 Heisenberg model

In Chapter 6, we discussed the (itinerant) magnetism of free electrons,

in particular, the Pauli spin paramgnetism defined by the spin moment
~ = −µB g~σ /2. In this free (non-interacting) system, we did not find any
µ
ordering phenomenon—we did, however, mention the instabilities resulting
from the repulsive Coulomb interaction and the Pauli principle. These are
the basic elements for the Heisenberg model: we consider spins S ~i defining
local moments on a lattice and let them interact through through the energy
~i · S
Jij S ~j . Often, the interaction is limited to the z nearest neighbours and
chosen to be isotropic, Jij = J for hi, ji and 0 otherwise2 . In a finite external
magnetic field ~h one obtains the Hamiltonian3

J X~ ~ X
~i · ~h = HH + Hext ,
H = − Si · Sj − µS (8.1)
2
hi,ji i
J > 0, ferromagnetic coupling, ↑↑: E = −J; ↑↓: E = J,
~
|Si | = 1, classical spins,
µ = magnetic moment.

The magnetic interaction between the spin moments µS ~i is usually small.

The exchange interactions generated by the Coulomb repulsion and the
Pauli exclusion principle are much larger in general. Consider two Hydrogen
atoms at 0 and at R.~ Using the ground state wavefunctions φA (~r ) = φ(~r )
~ ), φ = φ1s = (1/πa3B )1/2 e−r/aB , one generates the
and φB (~r ) = φ(~r − R
variational valence states
1
Ψs,t = p [φA (~r1 )φB (~r2 ) ± φB (~r1 )φA (~r2 )]χs,t (s1 , s2 ) (8.2)
2(1 ± S 2 )

with the overlap integral

Z
~)=
S(R d3 r φA (~r )φB (~r )  1.

The energies Es,t = hΨs,t |HH2 |Ψs,t i of the singlet and triplet states are
separated by

φA (~r2 )φB (~r1 ) 2e2 φB (~r2 )φA (~r1 )

Z
∆E = Es − Et ≈ d3 r1 d3 r2 , (8.3)
|~r1 − ~r2 |
2
hi, ji denotes nearest neighbours. With every pair hi, ji appearing twice in the sum,
the energy per bond is 2J
3
We have to be carefuld about units. Here, µh assumes the unit of energy.
8.1. GENERIC MODELS OF LOCAL MAGNETISM 135

as determined by the Coulomb interaction. Using

(
1/4, triplet,
~s1 · ~s2 =
−3/4, singlet,

we can write this result (up to a common energy shift) as

X
HH2 = −∆E ~si · ~sj ,
ij

i.e., the effective Hamiltonian for the spin problem is of the Heisenberg type.
The exchange interaction J = ∆E/2 (1/2 from the definition |S ~i | = 1) is
purely electronic (non magnetic). Typical values for J can go up to ∼ 103 K,
hence, the temperature scale of electronic magnetism can be rather high.
An important parameter in the study of magnetic systems and phase
transitions in general is the range of the interaction. Restricting the latter
to nearest neighbors, it is the number of nearest neighbours z that plays an
important role. For a cubic lattice, z = 2d where d denotes the dimension-
ality of the crystal.
The interaction J can be negative—in this situation one talks about
anti-ferromagnetic coupling. If the lattice is bipartite, an anti-ferromagnetic
order can appear with opposite feromagnetic orders on the two sub-lattices.
If the lattice is not bipartite the system will be frustrated, for example on a
triangular lattice in d = 2, see Fig. 8.1. In a magnetically doped solid (for

Fig. 8.1: Frustration in the d =

2 triangular lattice: the third
J <0 J<0 spin always costs an exchange en-
ergy.

?
J<0
example Mn in Cu, Ag, Au) with different distances between the magnetic
atoms, the coupling Jij can be +/− randomly4 and the result is a spin glas
with maximal frustration.
The Heisenberg model has O(3)-Symmetry. Generally one talks about
O(n)-symmetric spin models with the continuity of the symmetry playing
~i in the Heisenberg model is
a particularly relevant role—the spin-field S
4
E.g, generated via the RKKY interaction (Ruderman-Kittel-Kasuya-Yosida) and ran-
dom distances between dopants.
136 CHAPTER 8. MAGNETIC ORDER

continuously (O(3)) symmetric, i.e., can be rotated in any direction. If one

reduces n to 2, one obtains the (U(1) continuously symmetric) XY model.
Finally, the Ising model has Z2 symmetry.

8.1.2 XY model

In the XY model, the freedom of the spins is constrained to a plane—one

chooses, without loss of generality, the xy plane. The O(2) symmetry is
isomorphic to the U (1) symmetry of the Bose problem. The complex order
√
parameter hΨi = ρeiφ of the condensate can in fact be identified with
the order parameter hS~ i = S(cos ϑ, sin ϑ) of the XY model; the XY spin
model and the physics of Bose condensation/superfluidity show the same
behaviour, a finding that goes under the notion of universality. An XY-
symmetry is found in magnetic materials due to an anisotropy—for example
an “easy plane” anisotropy where the spin preferentially arranges itself in a
plane.

8.1.3 Ising model

If one limits the degree of freedom of the spin to one axis (an “easy axis”
anisotropy in solids) one obtains the Ising model,

JX z z X
H=− Si Sj − µh Siz = HI + Hext . (8.4)
2
hi,ji i

Its symmetry is given by the discrete group Z2 = {1l, R with R2 = 1l} (Ising
symmetry). The Ising model of magnetism is equivalent to the lattice gas
model where a lattice site can be occupied or not and to the problem of
binary alloys, for example Zn-Cu in a bcc lattice. Furthermore, the Ising
model and the van der Waal gas-liquid transition are in the same universality
class.

8.2 Magnetization and Susceptibility

In correspondence to defining a local particle density

DX E
hn(~r )i = δ(~r − ~ri ) (8.5)
i
8.3. PARTITION FUNCTION FOR A SPIN SYSTEM 137

in a gas or condensate, we can define a local magnetisation density5

DX E
hm(~
~ r )i = µS~i δ(~r − ~ri ) (8.6)
i

~i denotes the
in magnetic systems, see also Eqs. (6.38) and (6.39). Here, S
(classical) spin on site i and µ the magnetic moment. In the homogeneous
case hm(~
~ r )i = hmi
~ and thus
~
M
hmi
~ = (8.7)
V
~ being the (extensive) magnetic mo-
defines the magnetisation density, M
ment. As for the density n = 1/V , the magnetisation density m
~ takes on
the role of a thermodynamic variable that interdepends on quantities such
as entropy or energy.
Another quantity that characterizes the magnetic system and that can be
conveniently measured is the magnetic stiffness or (magnetic) susceptibility

∂m
χ = . (8.8)
∂h
This quantity should be viewed as the magnetic analog of the compressibility
κ = V −1 ∂p V in the mechanical context with variables p and V or the specific
heat CV = T ∂T S in the caloric context with thermal variables T und S. In
the sections 6.1.1 and 6.1.2, we have already seen the Landau- and Pauli-
susceptibilites of the electron gas.

8.3 Partition function for a spin system

The spin system described by the Hamiltonian (8.1) has a fixed number
N of spins or degrees of freedom and thus the canonical ensemble is the
appropriate one for its statistical description. We then have to adapt the
canonical partition sum (3.38)
Z
ZN (V, T ) = d3N p d3N q ρ(p, q)
Z 3N 3N  
d pd q H(p, q)
= exp − ,
h3N N ! kB T
to the spin lattice problem. The continuous variables (p1 , . . . , q3N ) are
replaced by the set (S~1 , . . . , S
~N ) and the integral over phase space goes
5
Be careful about units; here, m(~
~ r ) assumes the unit of moment per volume since it
is a field, while m ~i is a moment. We go from one to the other by multiplying with
~ i = µS
the unit cell volume Vuc .
138 CHAPTER 8. MAGNETIC ORDER

over to the sum over all spin configurations {S~i }N with an integration
1
over angles for n > 1 on each site. The situation is particularly sim-
ple for the Ising model, where the partition sum is over all realizations
of {S1 = ±1, S2 = ±1, . . . , SN = ±1}.
Good thermodynamic variables for the spin system are the temperature
T (and its conjugate entropy S) and the (intensive) external magnetic field
~h (with the magnetization M ~ assuming the role of the conjugate extensive
variable). The variables p and V , pressure and volume, are usually ignored
when discussing generic magnetic phenomena. However, in real compressible
material they may have to be included into the description as well, e.g., in
analyzing the phenomenon of magnetostriction. The partition function of a
magnetic system then is usually written in the form

X  ~i }] 
H[{S
ZN (T, ~h) = exp − (8.9)
kB T
~i }
{S

with angular integrations per site included in the sum when n > 1. The
factor
R h−3N can be dropped as we have replaced the phase-space integral
dpdq by a dimensionless sum. Similarly, the Gibbs factor 1/N ! becomes
irrelevant since the moments/spins are fixed on the lattice sites and there is
no permutation of variables.
The associated thermodynamic potential is a function of the intensive
variables T and ~h, as well as the number N playing the role of the extensive
variable. Hence, the good potential is the Gibbs fee energy,

G(T, ~h, N ) = −kB T ln ZN (T, ~h). (8.10)

P ~ ~
Note that the addition of the term −µ i S i · h in the (microscopic) Hamil-
tonian corresponds to a Legendre transformation (from F (m) ~ to G(~h)) on
a phenomenological level. The derivatives with respect to T and ~h provide
us with the entropy S and the magnetization M ~,

∂T G(T, ~h, N )|~h,N = −S, ∂~h G(T, ~h, N )|T,N = −M

~. (8.11)

The magnetic susceptibility χ follows from the second derivative,

∂mi 1 ∂2G
χij = =− . (8.12)
∂hj V ∂hi ∂hj

It is often useful to go over to conjugate variables. One then replaces

the Gibbs free energy G(T, ~h, N ) = −kB T ln ZN (T, ~h) (with ~h playing the
role of the pressure p) through the Legendre transform

F = G + ~h · M
~ = F (T, M
~ , N ).
8.4. BROKEN SYMMETRY AND ORDER 139

The magnetic information is then obtained from

~h = ∂ ~ F, (8.13)
M
∂2F
χ−1
ij = V . (8.14)
∂Mi ∂Mj

8.4 Broken symmetry and order

In non-interacting systems, an external magnetic field ~h 6= 0 is required to

~ 6= 0 via
generate a finite moment m
Z
H ext
=− ~ r ) · ~h(~r ).
dd r m(~ (8.15)

Rather than applying a homogeneous field ~h, here, we apply an inhomoge-

neous test field ~h(~r ) that will induce a local magnetization m(~
~ r )—this will
allow us later to discuss the phenomena of order and correlations.

Fig. 8.2: Paramagnetic and

ferromagnetic phase in a fer-
romagnet at large (left)/small
(right) temperatures.

A (for example) ferromagnetic interaction tends to align the moments

in the absence of an external field, i.e., the magnetization can appear spon-
taneously. In the free energy F = U − T S, the entropic contribution T S
favors a paramagnetic phase at large temperatures while the internal energy
U tends to order the spins. At small temperatures the energy term wins
and the free energy is minimised by the ferromagnetically ordered phase.

Fig. 8.3: Magnetic phase di-

agram with a first-order transi-
h tion line at h = 0 when crossing
the h = 0 line at fixed temper-
st nd
1 2 ature T < Tc and critical end-
point at Tc where a second-order
0 Tc T transition appears when tuning
T across Tc at h = 0.
critical point
140 CHAPTER 8. MAGNETIC ORDER

In the ferromagnetically ordered phase the symmetry of the Hamiltonian

is spontaneously broken. For example the O(3) symmetry of the ~h = 0
Heisenberg Hamiltonian is broken by the choice of a magnetisation direction
hm̂i in the ferromagnetic phase. As the state {S ~i } belonging to m̂ and the
state {−S~i } belonging to −m̂ have identical energies, we in fact strictly have
hm̂i identically vanishing. Mathematically, we must explicitly break the
symmetry to find a finite magnetisation, for example through the following
limiting procedure,
ext ~
ZN [T, ~h ] = Tr e−β[HH +H (h )] ,


hm(T,
~ ~h, N )i = 1 ∂ ln ZN (T, ~h ),
βV ∂~h
hm(T
~ )i = lim lim hm(T,
~ ~h, N )i. (8.16)
~h→0 N →∞

We then obtain the phase diagram as shown in figure 8.3 (compare to the
p–T phase diagram in Fig. 1.17). The transition at Tc in an h = 0 field
is of second order: The magnetisation grows steadily from 0 when T drops
below Tc . The transition at T = const < Tc and h = 0 is first-order: The
magnetisation jumps from −m ê to m ê with m = |m|. ~ A sketch of the
equation of state hm(T,
~ h)i is shown in Fig. 8.4. For ~h = 0 the spontaneous

m Fig. 8.4: Equation of state

hm(T,
~ h)i. The thick lines trace
the magnetization at h = 0 and
h
0 at T = Tc where mean-field scal-
ing (with exponents β and δ)
shows up near Tc . The jump im
Tc
m at h = 0 and below Tc cor-
T responds to the ruled surface in
the p–V –T diagram in Fig. 1.18
of the gas–liquid transition.

magnetisation hm(T
~ )i plays the role of an order parameter that specifies
how, i.e., in which direction, the symmetry is spontaneously broken.
As in the crystal and the superfluid, the spontaneous symmetry breaking
goes hand in hand with the appearance of long-range ordering (LRO) in
the system.6 The latter is quantified by the spin-spin or magnetisation-
magnetisation correlator
 
C(~r, ~r 0 ) = hm(~ ~ r 0 )i = µ2 hS(~
~ r ) · m(~ ~ r 0 )i .
~ r ) · S(~ (8.17)

6
A special case may appear in 2D, see later.
8.4. BROKEN SYMMETRY AND ORDER 141

Long-range order then implies that the direction of magnetic moments at

far distant points is preserved and hence the correlator does not vanish,

|~ r 0 |→∞
r−~
C(~r, ~r 0 ) → ~ )i2 6= 0
hm(T (LRO). (8.18)

8.4.1 Susceptibility and correlations

Considering space a dependent drive ~h(~r ) and response m(~

~ r ) allows us to
define the non-local magnetic susceptibility

δhmi (~r )i 1 δ 2 ln Z[T, ~h(~r )]

χij (~r, ~r 0 ) = = (8.19)
δhj (~r 0 ) β δhi (~r )δhj (~r 0 )
= β[hmi (~r )mj (~r 0 )i − hmi (~r )ihmj (~r 0 )i]
= βh(mi (~r ) − hmi (~r )i)(mj (~r 0 ) − hmj (~r 0 )i)i
1 1
= h δmi (~r ) δmj (~r 0 ) i = Gij (~r, ~r 0 ), (8.20)
kB T kB T

which is proportional to the correlator of the order parameter fluctuations

according to (8.20). Note that the second derivatives of ln Z generate cumu-
lants (denoted by hh· · · ii) and not moments (we consider a one-component
case),

δ 2 ln Z m(~r 0 )e−βH
R
δ
= (8.21)
∂βh(~r )δβh(~r 0 )
R
δβh(~r ) e−βH
m(~r 0 )m(~r )e−βH m(~r 0 )e−βH m(~r )e−βH
R R R
= R −
e−βH
R
R

e−βH e−βH
= hm(~r )m(~r 0 )i−hm(~r )ihm(~r 0 )i ≡ hhm(~r )m(~r 0 )ii.

The Legendre transformed free energy then reads

Z
F = G + dd r ~h(~r )hm(~
~ r )i = F (T, hm(~r )i),

from which the magnetic information derives from (again, be careful with
units, specially volumes)

δF
hi (~r ) = , (8.22)
δhmi (~r )i
δ2F δhi (~r )
χ−1 r, ~r 0 ) =
ij (~ 0
= . (8.23)
δhmi (~r )iδhmj (~r )i δhmj (~r 0 )i

The last line is found from (where cr denotes the chain rule and fd stands
142 CHAPTER 8. MAGNETIC ORDER

for functional derivative)7

δhmi (~r )i δhj (~r 0 )

Z
δhmi (~r )i cr
= dd r0
δhmk (~r 00 )i δhj (~r 0 ) δhmk (~r 00 )i
k fd k (8.19) k (8.23) (8.25)
Z
δik δ(~r − ~r 00 ) = d3 r0 χij (~r, ~r 0 )χ−1 r 0 , ~r 00 ).
jk (~ (8.26)

The remaining (non-magnetic) thermodynamics follows from the standard

relations. The corresponding variables (µ, N ; p, V ) have been suppressed in
the expressions above.

8.5 Mean-field theory

We begin with the simplest case of the Ising model and then proceed with
the mean-field theory of the O(n) model involving the XY and Heisenberg
spins.

8.5.1 Ising model

We start from the Hamiltonian (8.4),

JX z z X
H=− Si Sj − µh Siz , (8.27)
2
hi,ji i

where each pairing hi, ji appears twice. Every spin is exposed to the external
field h and the effective or “Weiss” field of its neighbours,
X
µh0i = µh + JSjz ≈ µh + zJhSi. (8.28)
j

7
The functional derivative of F [m(~r )] with respect to m(~r 0 ) is defined as

δF F [m(~r ) + εδ(~r − ~r 0 )] − F [m(~r )]

0
= lim , (8.24)
δm(~r ) ε→0 ε

~
e.g., with F [m(~r )] = dd r f [m(~r ), ∇m(~
R
r )], we find that

δF
Z
∂f δm(~r )
h ∂f ~
δ ∇m(~ r )i
= dd r +
δm(~r 0 ) ∂m(~r ) δm(~r )0 ~
∂ ∇m(~ r ) δm(~r 0 )
Z h ∂f ∂f i
= dd r δ(~r − ~r 0 ) − ∇
~ δ(~r − ~r 0 )
∂m(~r ) ~
∂ ∇m(~ r)
∂f ~ ∂f
= −∇ .
∂m(~r 0 ) ~
∂ ∇m(~ r 0)
8.5. MEAN-FIELD THEORY 143

In the last equation, we have implemented a mean-field approximation by

replacing (1/z) j Sjz with the average field hSi produced by the z neigh-
P
bors. This approximation takes us to the simple one-particle Hamiltonian

J X
HMF = zN hSi2 − µh0 Siz . (8.29)
2
i
The following alternative derivation makes the type of approximation, the
dropping of quadratic fluctuation terms, more transparent. We rewrite the
Hamiltonian (8.27)
JX z X
H = − [(Si − hSi) + hSi][(Sjz − hSi) + hSi] − µh Siz
2 | {z }
hi,ji i
| {z }
δSiz δSjz

J X X X
= − N zhSi2 − JzhSi δSiz − µh Siz − J δSiz δSjz
2
i i hi,ji
J X
≈ zN hSi2 − µh0 Siz ,
2
i
where we counted each pairing twice and ignored the quadratic terms in the
fluctuations; this derivation again leads us to the mean-field Hamiltonian
(8.29).
Yet another approach is via Bragg-Williams theory,Pwhere one writes
an Ansatz for a single-particle Hamiltonian HMF = −α i Siz + const and
minimises the Bogoliubov functional (5.67) with respect to α, see later.
The Hamiltonian (8.29) is (up to the constant) a sum of single parti-
cle Hamiltonians H1 = −αS z , α = µh0 and the corresponding statistical
mechanics can be easily calculated, ZN = Z1N with Z1 the single spin parti-
tion function. The average spin hSi is obtained through the self consistency
equation derived from statistical averaging,
∂
hSi = ln Z1 , Z1 = eβα + e−βα = 2 cosh βα,
∂βα
∂  µh + zJhSi 
= ln 2 cosh βα = tanh .
∂βα kB T
Inverting the above equation provides us with a transzendental equation for
hSi,
µh = kB T artanh hSi − zJhSi (8.30)
which can be solved graphically as sketched in Fig. 8.5. The critical tem-
perature Tc separates regimes with positive and negative slopes ∂h/∂hSi at
the origin,
∂µh ∂
' (kB T hSi − zJhSi) = 0
∂hSi ∂hSi
→ kB Tc = zJ. (8.31)
144 CHAPTER 8. MAGNETIC ORDER

Fig. 8.5: Graphical solution

h T > Tc
of the self-consistency equa-
h= 0 tion for the mean field hSi.
S =0 While the magnetization m =
µhSi/Vuc vanishes above Tc , a
−1 0 1 S finite solution appears at tem-
peratures below Tc .
T = Tc T < Tc

A negative slope leads to the appearance of multiple solutions for hSi at

h = 0, i.e., in the absence of an external driving field. This implies that
the symmetry is spontaneously broken with a finite hSi selecting one of two
directions.
The behaviour close to Tc is obtained from the cubic approximation
 hSi3 
µh = kB T hSi + − zJhSi. (8.32)
3
For h = 0 and T . Tc , we find a finite average hSi from
 hSi2 
kB T 1 + = zJ = kB Tc ,
3
√
T r
2 c
 T
hSi = 3 −1 → hSi ≈ ± 3 1 − , (8.33)
T Tc
while for T → 0, hSi saturates according to

hSi ≈ 1 − 2e−2(Tc /T )hSi ≈ 1 − 2e−2Tc /T . (8.34)

These results provide us with an overall temperature dependence for the

magnetisation
µhSi
hmi = (8.35)
Vuc
as shown in Fig. 8.6. The spontaneous magnetisation hmi (or hSi) plays
the role of the order parameter in the magnetic transition as it selects one
of two degenerate directions (Z2 symmetry of the Ising model). Below Tc ,
this order parameter increases from zero with a fractional, i.e., non-analytic,
power-law
hmi ∝ (1 − T /Tc )β = |τ |β with β = 1/2 (8.36)
and T /Tc = 1 + τ : the critical exponent β for the order parameter as a
function of |τ | assumes the value β = 1/2 within mean-field theory. This
statement is universal and applies even across changing the dimensionalities
d and n of space and the order parameter. Within the present discussion,
8.5. MEAN-FIELD THEORY 145

m Fig. 8.6: Spontaneous mag-

h= 0 netisation hmi at h = 0 as a
−2Tc / T
1−2 e function of T . Within mean-
field theory, hmi departs from
1−T / Tc
zero with a square root law,
implying a mean-field critical
0 Tc T exponent β = 1/2.

the dimensionality d of space appears only in the value of the critical tem-
perature kB Tc = 2dJ, indicating that higher dimensions favour the ordered
phase more strongly.
The self-consistency equation (8.32) provides us with further details of
the magnetic transition. At Tc , Eq. (8.32) reduces to µh = kB T hSi3 /3,
telling us that the order parameter hmi = (µ/Vuc )(3µh/kB T )1/3 on the crit-
ical isotherm scales as

hmi ∝ h1/δ with δ = 3, (8.37)

see Fig. 8.4. The critical exponent δ for the order parameter as a function
of h on the critical isotherm T = Tc assumes the value δ = 3 in mean-field
theory. Furthermore, we can calculate the magnetic suceptibility

χ = ∂h hmi. (8.38)

Above Tc , hmi = 0 and we find (from (8.32)) the derivative ∂hSi h = (kB Tc /µ)
(T /Tc − 1), from which we obtain χ = (µ2 /Vuc kB Tc )/τ and hence the scaling
of the susceptibility at h = 0 near Tc is

χ ∝ 1/τ γ with γ = 1. (8.39)

The same exponent γ but with a different prefactor is obtained when evalu-
ating χ below Tc where hmi =6 0. The critical exponent γ for the magnetic
susceptibility χ assumes the value γ = 1 in mean-field theory.
The last exponent characterizing the phase transition is the one for the
specific heat cV . Evaluating this quantity from the free energy f0 (see
below) one finds a jump, telling us that the critical exponent vanishes within
mean-field theory,
cV ∝ τ α with α = 0. (8.40)
The critical exponent α for the specific heat cV assumes the value α = 0
in mean-field theory.
A different formulation of the mean-field theory for the Ising model has
been given by Bragg and Williams. Using the Bogoliubov variational princi-
ple, we construct the free energy F = U −T S by calculating S and estimating
146 CHAPTER 8. MAGNETIC ORDER

U . Consider a state with N↑ up-spins, N↑ + N↓ = N and N↑ − N↓ ≡ N hSi.

The entropy of such a state is (we make use of Sterling’s formula)
   
N N
S = kB ln = kB ln (8.41)
N↑ N (1 + hSi)/2
N kB  
≈ 2 ln 2 − (1 + hSi) ln(1 + hSi) − (1 − hSi) ln(1 − hSi) .
2

This expression vanishes for the ordered phases with hSi = ±1 and amounts
to S = N kB ln 2 for an unpolarized (disordered) state with hSi = 0. For the
internal energy one uses the mean-field approximation

1
Uh = − N zJhSi2 − N µhhSi. (8.42)
2

The results (8.41) and (8.42) add to the free energy (h = 0)

f0 (T, hSi) = (U0 − T S)/N (8.43)

1 kB T
≈ (kB T − Jz)hSi2 + hSi4 − kB T ln 2,
2 12
where in the last equation, we have assumed a small value for the order
parameter hSi. Drawing f0 (T, hSi) for different temperatures T , see Fig.
8.7, we see that the character of the curves change at T = Tc = Jz where
the quadratic term changes sign. For T < Tc , we find a minimal free energy
at a spontaneous magnetisation
√  T 1/2
hSi(T < Tc ) ≈ ± 3 1 − ,
Tc

in agreement with the result in Eq. (8.33).

T >Tc
Fig. 8.7: Free energy
f0 f0 (T, hSi) at h = 0. As
T drops below Tc the
minimal free energy is
Tc T < Tc attained at a finite value
of hSi.

0 S

The equation of state can be alternatively obtained from

∂f0 ∂fh
h= or = 0, (8.44)
∂µhSi ∂µhSi
8.5. MEAN-FIELD THEORY 147

where the latter expression makes use of h as a constraining parameter; both

expressions lead to the same result
kB T 1 + hSi
µh = −zJhSi + ln (8.45)
2 1 − hSi
that agrees with (8.30).
The expansion (8.43) of the free energy around Tc is identical to the
ansatz made within Landau theory. Once the Landau expansion in the
order parameter is known, the thermodynamics around Tc can be completely
derived. We will come back to Landau’s general mean-field theory in section
11.2.

8.5.2 The O(n) model

Starting from the Hamiltonian (8.1) of the Heisenberg model, one again
makes a mean-field ansatz (see 8.28)

µ~h 0 = µ~h + zJhS

~i (8.46)
~ i self consistently,
and calculates hS
~0 ~~
dΩ eβµh ·S S
R
~
hS i = R . (8.47)
dΩ eβµ~h 0 ·S~

Using ~h = hêx , hS
~ i = hS x iêx , one finds

XY model: S x = cos ϕ
0
dϕ eβµhx cos ϕ cos ϕ
R
x
hS i = , (8.48)
dϕ eβµh0x cos ϕ
R

Heisenberg model: S x = cos ϑ

0
dϕ d cos ϑ eβµhx cos ϑ cos ϑ
R
x
hS i = . (8.49)
dϕ d cos ϑ eβµh0x cos ϑ
R

Evaluating the integrals over angles ϑ and ϕ, one arrives at the self-consistency
equations for the XY model,
I1 [βµhx + βzJhS x i]
XY model: hS x i = , (8.50)
I0 [βµhx + βzJhS x i]

For the Heisenberg model, the corresponding equation involves standard

hyperbolic functions,
1
hS x i = coth[βµhx + βzJhS x i] − . (8.51)
βµhx + βzJhS x i
148 CHAPTER 8. MAGNETIC ORDER

I0 Fig. 8.8: Sketch of mod-

ified Besselfunctions of the
first kind I0 and I1 appearing
I1 in the self-consistency equa-
1 tion of the XY model.

0 x

The expansion for small hS x i provides us with the expression for Tc and the
T -dependence of the order parameter around Tc . Using x ≡ βzJhSi we find,

kB T x2
Ising: =1− → kB Tc = zJ,
zJ 3 √ √
hSi = 3 1 − t. (8.52)

kB T 1 x2
XY: = − → kB Tc = zJ/2,
zJ 2 16 √ √
hSi = 2 1 − t. (8.53)

kB T 1 x2
Heisenberg: = − → kB Tc = zJ/3,
zJ 3 45 p √
hSi = 5/3 1 − t. (8.54)

The results can be summarized in the following way: The qualitative

behaviour of the order parameter with mean-field theory is independent
of the dimensionalities d and n of space and the order parameter. The
(cubic) self-consistency equations (8.30), (8.50), (8.51) generate a (quartic)
free energy of the type found in (8.43) around Tc :
r
f (T, hSi) = hSi2 + uhSi4 + const.
2
 T
with r = r0 1 − , r0 < 0,
Tc
u ≈ const. (8.55)

From f we obtain a universal thermodynamic behaviour independent from

n and d.
The differences in the various models only manifest themselves in the nu-
merical values of parameters, e.g., r0 , u, and Tc in (8.55). In particular, for
8.6. DOMAIN WALLS IN THE 1D-ISING MODEL 149

these three models, we have results for the critical temperature that depend
on d and n as
2d
kB Tc = J , (8.56)
n
i.e., larger values of d (more neighbors) increases Tc , while a larger n pro-
viding more degrees of freedom to the order parameter decreases Tc .
The mean field theory provides us with a simple compact and universal
picture of magnetic phase transitions. In fact, mean-field theory always gives
the same structural result: The continuous (second-order) phase transitions
are described by a self-consistency equation which around Tc is given by
T 4u 3
Φ=Φ+ Φ ,
Tc r0
Φ = order parameter, from where we obtain a free energy (by integration)
|r0 |  T 2
f (T, Φ) = − 1− Φ + uΦ4 .
2 Tc
Unfortunately, mean-field theory misses an important point: Its super (i.e.,
independent on d and n) universal character predicts transitions that do not
exist and gives us exponents which are independent of the dimensionalities
d and n of the model. This deficiency will be cured when fluctuations are in-
cluded into the picture. In spite of these drawbacks, its simple and compact
form motivates the use of mean-field theory as a first starting point when
dealing with a phase transition. There are even situations where mean-field
theory is “correct”. This is, e.g., the case when the mean field is supported
by many (or even an infinite number of) neighbours. In this situation, the
averaged field acting on the order parameter is well defined, i.e., the devi-
ations (= fluctuations) from the mean are small. Many neighbours appear
when going to higher dimensions and long-range interactions, and mean-field
theory turns out correct for the magnetic transitions of the Ising-, XY-, and
Heisenberg models in d > 4. Furthermore, mean-field theory is very accurate
for BCS (Bardeen-Cooper-Schrieffer) superconductors with a large correla-
tion or coherence length ξBCS , nξBCS3  1 with n the electron density. In
the following, we study how the mean-field theory fails in magnetic systems
in low dimensions, where z is small and fluctuations in the mean field are
expected to be large.

8.6 Domain walls in the 1D-Ising model

A simple argument demonstrates that the (short range8 ) Ising model cannot
order at finite temperatures in one dimension. Consider the ordered Ising
8
See D.J. Thouless, Phys. Rev. 187, 732 (1969) for a discussion of long-range-order in
one-dimensional Ising systems with long-range (slowly decaying) interactions.
150 CHAPTER 8. MAGNETIC ORDER

chain in Fig. 8.9. The introduction of a domain wall costs an energy 2J.

Fig. 8.9: Ordered Ising chain

(top) and including one domain
domain wall
wall (bottom) that flips half the
spins in the chain.

Such a domain wall can be inserted in any of the N − 1 spacings between

spins, resulting in the domain wall’s free energy

Fdw = 2J − kB T log(N − 1). (8.57)

For N → ∞, we have Fdw < 0 for all T > 0 and thus there is no order present
at any finite temperature. The only place where an ordered ferromagnetic
state can survive is at T = 0 and we can call this a T = 0 critical point,
Tc = 0.
Domain walls are topological objects or defects. The only way to generate
them is in pairs or by dragging them into the system across its boundaries,
see Fig. 8.10.

Fig. 8.10: Generation of

domain walls in pairs (top)
or via entry from a bound-
ary (bottom).

The 1D-Ising model can be solved exactly with the help of transfer
matrices. Consider N Spins Si = ±1 with periodic boundary conditions
S1 = SN +1 . The partition sum ZN (T, h) can be written in the form

N
XX X  X  µh 
ZN (T, h) = ... exp β JSk Sk+1 + (Sk + Sk+1 ) . (8.58)
2
S1 S2 SN k=1

Introducing the transfer matrices (S, S 0 = ±1)

  µh 
hS|T |S 0 i = exp β JSS 0 + (S + S 0 ) ,
2 
e−βJ
 β[J+µh]
e
T = , (8.59)
e−βJ eβ[J−µh]
8.6. DOMAIN WALLS IN THE 1D-ISING MODEL 151

the partition function ZN can be written9 as the trace of T N ,

X X
ZN (T, h) = ... hS1 |T |S2 ihS2 |T |S3 i . . . hSN |T |S1 i
S1 SN

|Sk ihSk | = 1)
X X
(with = hS1 |T N |S1 i
k S1

= λN+ + λN
−. (8.60)

The eigenvalues λ± of the transfermatrix T are easily derived,

h µh  µh 1/2 i
λ± = eJ/kB T cosh ± sinh2 + e−4J/kB T . (8.61)
kB T kB T
In the thermodynamic limit N → ∞, only the larger eigenvalue λ+ > λ− is
relevant and we obtain the free energy per particle in the form
G
g(T, h) = = −kB T log λ+ (8.62)
N
h µh  µh 1/2 i
= −J − kB T log cosh + sinh2 + e−4J/kB T
kB T kB T
2 2
µ h 2J/kB T
≈ −J − kB T e−2J/kB T − e ,
2kB T
where we have assumed that h defines the smallest scale in the expansion,
µh/kB T  e−2J/kB T  1. As expected, the T = 0 groundstate energy per
spin is −J; furthermore, the approach to the groundstate is exponential in
T , hence the excitation spectrum is gapped. Taking the derivative of g(h)
with respect to h, we obtain the magnetization
∂g
hm(T, h)i = −
Vuc ∂h
µ sinh(µh/kB T )
= (8.63)
Vuc sinh2 (µh/kB T ) + e−4J/kB T 1/2


h small µ2 h 2J/kB T h→0

≈ e → 0, (8.64)
Vuc kB T
i.e., there is no spontaneous magnetization upon decreasing the temperature
T . The susceptibility

∂m µ2
χ= = e2J/kB T (8.65)
∂h Vuc kB T
diverges for T → 0, signalling a T = 0 phase transition. Indeed, going to
high temperatures, hm(T, h)i ≈ µ2 h/Vuc kB T and we find the Curie-Weiss
law χ ∝ (T − Tc )−1 with Tc = 0, again indicating a T = 0 phase transition.
9
Here, the bra and ket vectors denote classical spin states.
152 CHAPTER 8. MAGNETIC ORDER

In addition, we can determine correlations between spins, G(i, j) =

G(n) = hSi Sj i with n = |i − j| due to translation invariance. We shift
this discussion to Sec. 15.1.
As not even the Ising model (with the lowest degree of freedom in the
spin) shows a phase transition in 1D, we expect that neither the XY- nor
the Heisenberg model with their continuous spin directions will have a finite
temperature transition. This expectation is confirmed by the Hohenberg-
Mermin-Wagner theorem, see 8.7.2.

8.7 Magnetic order in two dimensions

We first give an argument (due to Peierls) that the Ising model in 2D shows
a phase transition at finite temperatures. We will avoid discussing the exact
solution here (this exact solution is due to Onsager and produces a critical
temperature given by kB Tc ≈ 2.269185 J). Next, we discuss the continuous-
spin XY model and its special quasi–long-range ordered low-temperature
phase, as well as the special type of transition between the short-range or-
dered phase at high temperatures and the quasi–long-range ordered state
(due to Berezinskii and Kosterlitz-Thouless). Finally, we make a few com-
ments on the non-ordering (in 2D) Heisenberg model.

8.7.1 2D Ising model

We show that at h = 0 and positive but sufficiently small T > 0 there exists
a finite magnetisation hSi > 0 in the 2D Ising model.

Peierls’ argument:

Consider the T = 0 ground state of the 2D Ising model with all spins pointing
up, Siz = ↑, denoted by + in Fig. 8.11. As the temperature is raised, spin–
↓ domains (denoted by − in Fig. 8.11) appear. The associated domain
walls separating spin–↑ from spin–↓ regions cost an energy 2JL, where L
is the length of the domain’s circumference. The growth/extension of such
domains costs an energy ∝ L, the same scaling as the entropy associated
with the geometry and location of such a domain. As a result, the energy
dominates over the entropy at sufficiently low but finite temperatures T > 0
and the phase stays ordered.10

10
Note that a domain in 1D costs L0 , in 2D L1 , in d dimensions Ld−1 , hence additional
dimensions “stabilise” the ordered phase.
8.7. MAGNETIC ORDER IN TWO DIMENSIONS 153

++ + + + +++ Fig. 8.11: Domain formation in

the ground state of the 2D Ising
++ − − + +++ model: the domain wall encloses
+− − − − −++ a droplet with inverted magneti-
+− − − − +++ sation.
++ − − + +++
++ + + + +++
++ + + + +++
Idea of proof (Peierls, Griffiths):

Consider an arbitrary spin S ∗ on the lattice; p+ is the probability that

S ∗ = 1, p− + p+ = 1, hS ∗ i = p+ − p− . We choose the edge to be +. If
S ∗ = −1, there are 2n + 1 loops between S ∗ and the edge, otherwise there
are 2n loops, see Fig. 8.12. Let S ∗ = −1. Flipping all spins in the domain
of S ∗ , we lose a loop and S ∗ = 1.

+++++++++++ Fig. 8.12: Switched domains

+−−−−−−+−−+ in the neighbourhood of the
+−+−−−−++++ magnetic phase transition in
+−++−−−++++ the Ising model.
+−++−−+++++
+−−−−++++++
+−−−+++−−−+
+++++++−−−+
+++++++−−++
+++++++++++
Hence, we can go from S ∗ = 1 to S ∗ = −1 by inserting an additional loop
(of length L) around S ∗ , keeping the other loops fixed. Defining g(L) as the
number of such additional loops of length L, we can relate p+ and p− via
(this relation has to be worked out more carefully in a true proof, where one
has to show that ‘∼’ is to be replaced by ‘<’; note that the normalization
enters through p+ )
X 
p− ∼ g(L)e−2JL/kB T p+ ,
L

and using p+ < 1, we obtain the inequality

X
p− < g(L)e−2JL/kB T .
L

The factor g(L) > 1 favors the appearance of an additional loop that flips
the spin, while the Boltzmann factor accounts for the loop’s cost in energy
154 CHAPTER 8. MAGNETIC ORDER

and reduces its probability to show up—obviously, this is the usual interplay
of entropy and energy associated with a loop of size L.
In our estimate of g(L), we start at an arbitrary point in the domain.
For the surface of the domain, we set the maximum value (L/4)2 . In the
first step, we choose one from four directions, in every following step from
three (we ignore crossings). Thus we obtain
 L 2 1
g(L) ∼ · 4 · 3L−1 · . (8.66)
4 2L
The last factor takes into account that we can start at any point of the loop
and can go around it in two directions, clock- and anti-clockwise. We then
find
∞
X L L −2JL/kB T
p− < 3 e
24
L=4,6,...
∞
1 X −d −αL 2J
< e , α= − ln 3
24 dα kB T
L=4,5
1  d e−4α  5 e−4α
= − < . (8.67)
24 dα 1 − e−α 24 1 − e−α
For sufficiently large α or small temperatures T , the average spin hS ∗ i =
1 − 2p− is bounded from below,

5 e−4α
hS ∗ i > 1 − (8.68)
12 1 − e−α
and choosing kB T < J, e−α < 0.41, we find that hS ∗ i > 0.98. Thus below
kB T = J the phase is definitely ordered; in fact, the ordered phase survives
until kB Tc ≈ 2.269185 J (Onsager).
This type of proof for the existence of an ordered phase can be used
for other models. The central object entering the proof is the loop of
length L flipping a domain. One then may argue, that a finite-temperature
phase transition appears if there is a temperature 0 < T < ∞ where
loops start proliferating. The free energy of such a loop is Floop (L) ∼
2LJ − T ln(L3L /24) ∼ L(2J − T ln 3) and we find that loops flood the plane
when T > 2J/ ln 3.

8.7.2 XY model in 2D
~i = (cos ϑi , sin ϑi ),
The XY Hamiltonian can be written in polar coordinates S
X X
H = −J ~i · S
S ~j = −J cos(ϑi − ϑj ) (8.69)
hi,ji hi,ji
8.7. MAGNETIC ORDER IN TWO DIMENSIONS 155

(we work without an external field present, h = 0). At T = 0 the ground

state has all its spins aligned and the symmetry is spontaneously broken
by the selection of a specific angle ϑ0 . Small deviations ϑi − ϑ0 from the
direction ϑ0 generate an “elastic energy” (we use cos ϑ = 1 − ϑ2 /2 + . . . ;
z = 2d)
zN J 1 X
H ≈ − + J (ϑi − ϑj )2
2 2
hi,ji

Ja 2 Z d
d r ~ 2
≈ E0 + (∇ϑ) , (8.70)
2 ad
where E0 is the ground state energy and ϑ(~r ) is the (continuous) deforma-
tion field away from ϑ0 . Note that we kept the underlying space dimension
d arbitrary; we will focus on the planar case d = 2 later. The parameter J
describes the stiffness of the ordered phase; a is the lattice spacing.
We want to check whether the ordered phase as obtained from the mean-
field solution, is stable under fluctuations of the field ϑ(~r ) or whether it gets
destroyed as in the 1D Ising model. To this end, we calculate the expectation
value (we set ϑ0 = 0)
D[ϑ(~r )]e−βHel eiϑ(~r )
R
hSx i = hcos ϑ(~r )i = R (8.71)
D[ϑ(~r )]e−βHel
with the Gaussian Hamiltonian
Ja2
Z d
d r ~ 2
Hel = [∇ϑ]
2 ad
Ja2 X 2
= k ϑ~k ϑ−~k , (8.72)
2
~k

where we have used

dd r
Z
1 X ~ 1 ~
ϑ(~r ) = √ ϑ~k eik·~r , ϑ~k = √ ϑ(~r ) e−ik·~r .
N ~ N ad
k
Next, we use that fluctuations are Gaussian,
2 (~
hSx i = <heiϑ(~r ) i = e−hϑ r )i/2
,
and determine the amplitude of local fluctuations11
1 X kB T  a d Z kB T
2
hϑ (~r )i = 2 2
= dd k 2 2
N Ja k 2π Ja k
~k
S1d kB T 1
(
= (2π) d J d−2 , d > 2,
(8.73)
=∞ for k → 0, d ≤ 2.
11
We make use of hϑ2 (~r )i = (1/N ) ~k hϑ~k ϑ−~k i and the equipartition theorem
P

Ja2 k2 h|ϑ~k |2 i/2 = kB T /2. S1d is the surface of the unit sphere in d dimensions.
156 CHAPTER 8. MAGNETIC ORDER

We thus find that short-range fluctuations (the k-integral lives on large k ∼

1/a) reduce the order parameter hSx i in d > 2 while long-range fluctuations
(with the integral dominated by k → 0) push it to zero for d ≤ 2. This
statement can be refined by calculating the non-local spin-spin correlator

~ r ) · S(0)i
C(~r, 0) = hS(~ ~ = < exp[i(ϑ(~r ) − ϑ(0))]
= exp(−h[ϑ(~r ) − ϑ(0)]2 i/2). (8.74)

Using

2
 a d Z 2kB T (1 − cos ~k · ~r )
h[ϑ(~r ) − ϑ(0)] i = dd k
2π Ja2 k 2
 d
S1 2 kB T
 , d > 2,
 (2π)d d−2 J


∼ kB T R (8.75)
 πJ ln a , d = 2,

 kB T x ,

d = 1,
J a

we find that in
r→∞
3D: C(~r, 0) → const. > 0, LRO, (8.76)
 k T R
~ 0) R→∞ B
2D: C(R, → exp − ln
2πJ a
 a kB T /2πJ R→∞, alg.
= → 0, QLRO,
R
 k T x  x→∞, exp.
x→∞ B
1D: C(x, 0) → exp − → 0, SRO.
2J a
In words: thermally excited, elastic, low-energy fluctuations tend to destroy
the T = 0 order at T > 0. In 3D, C(~r ) stays finite assymptotically and long-
range order survives for 0 < T < Tc . In 1D, C(x) goes to 0 exponentially
for every temperature T > 0 and the order is destroyed. The 2D case is
marginal: C(R) decays to 0, but only slowly, algebraically. An algebraically
decaying correlator is typical for a system at a phase transition Tc , where
no length scale is apparent in the system12 . In 2D, at all temperatures T ,
the system behaves like a critical one at Tc . As the long-range order is
only slowly destroyed one talks about quasi–long-range ordered = QLRO
phase. This result is suspicious—it is hardly possible that QLRO survives
to arbitrary high temperatures. Indeed, we will see below, a very special
phase transition (due to Bereszinskii-Kosterlitz-Thouless) will intervene and
replace the algebraic decay by an exponential one, such that the order is
short-ranged at high temperatures.13
12
Algebraic functions have “no scale”.
13
The mistake is in our use of the Gaussian approximation, which fails at high temper-
atures where another type of fluctuations, vortices, comes into play.
8.7. MAGNETIC ORDER IN TWO DIMENSIONS 157

Hohenberg-Mermin-Wagner Theorem

The above finding is elevated to a theorem due to the work of Hohenberg,

Mermin and Wagner. In systems with continuous symmetry every long-
range ordered phase is destroyed by fluctuations in one or two dimensions.
This statement applies to spin systems (O(n), n = 2: XY model, n = 3:
Heisenberg model), superfluids (U (1) =
ˆ O(2)), crystals (translation symme-
try), etc. Note that 2D crystals show QLRO which is visible in algebraic
(instead of δ-function) Bragg peaks, S(Q) ∼ 1/|Q − G|dG , with G ~ = a lat-
tice vector. The spontaneous breaking of a continuous symmetry leads to
a new, low-energy excitation in the ordered phase: The order parameter
can choose one among a continuum of directions, around which the order
parameter then can fluctuate. The long wave-length modes cost only lit-
tle deformation energy, what leads to a soft mode, ε(k → 0) → 0. These
Goldstone modes are responsible for the destruction of the ordered phases
in d = 1 and 2 dimensions.
On the classical level the phenomenon is trivially universal: With an
elastic energy E ∼ (C/2)(∇Φ )2 for the fluctuating field Φ and in combina-
tion with the equipartition of thermal energy to all the modes, we obtain
the divergence at small k (infrared; in d = 2, we obtain a logarithm)14
Z Z
kB T T d=1,2
hΦ2 i ∼ dd k ∼ T dk k d−3 ∼ k d−2 |0 −→ ∞. (8.79)
Ck 2 0 d − 2
Within a quantum description using quantized modes (Goldstone bosons)
14
Similar (but in frequency rather than real space) can be obtained from the fluctuation-
dissipation theorem: The latter relates the dissipative (i.e., imaginary) component of the
susceptibility α (or Green’s function) with the system’s fluctuations (or noise). Let x = αf ,
f the (fluctuating) “force” and x the displacement. We write the symmetrized fluctuation
correlator (hxω xω0 i + hxω0 xω i)/2 = 2πδ(ω + ω 0 )hx2 iω ; then the amplitude hx2 iω is given
by the fluctuation-dissipation theorem (see, e.g., Landau-Lifshitz V, page 367, or Chaikin-
Lubensky)
~ω
hx2 iω = ~ Im[α] ctgh , (8.77)
2kB T
with ~ωctgh(~ω/2kB T ) the mean energy of the mode ω. As an example, we consider
a particle with the coordinate x(t) in a dissipative bath; the particle’s motion then is
governed by the friction rather than the inertia, ηv = η ẋ = −ηiωx = f , and hence
α = i/ηω [the fluctuating force f (t) is characterized by hf (t)i = 0 and hf (t)f (t0 )i =
2ηkB T δ(t − t0 ), please check]. In this case
Z Z
dω dω ~ ~ω
h[x(t) − x(0)]2 i = [1 − cos(ωt)]hxω x−ω i = [1 − cos(ωt)] ctgh (8.78)
π π ηω 2kB T
Z −1 Z ∞
kB T h t
Z
2k T [1 − cos(ωt)] 2 dω i
≈ B dω ∼ dω t + ∼ (kB T /η) t ∼ Dt.
ηπ ω2 η 0 t−1 ω
2

Thus, we obtain the Einstein relation between the diffusion coefficient D of the Brownian
motion hδx2 i(t) ∼ Dt and the dissipation η or mobility µ = 1/η. As an excercise one may
calculate the T → 0 limit of this “high temperature” result which will replace h[x(t) −
x(0)]2 i ∝ T t by h[x(t) − x(0)]2 i ∝ ~ ln ωUV t, with ωUV a UV cutoff.
158 CHAPTER 8. MAGNETIC ORDER

the analysis looks less universal.15

15
We consider three cases, the reduction of positional (Φ → ~ u ) order in a solid due
to phonons (note that hu2 (~r )i appears in the Debye-Waller factor), the reduction of the
ferromagnetic moment (Φ → m) ~ due to magnons, and the reduction of the condensate
density (Φ → n0 ) in an interacting Bose gas due to the sound waves.

Crystal, phonon = Goldstone mode: Using the Hamiltonian and the ex-
pression for the displacement field u of the phonons (note the appearance of ωk = ck,
c2 = ε/µ, the ratio of elasticity and mass density)
X
H = ~ωk (a~†k a~k + 1/2), ωk = ck,
~
k
 ~ 1/2
u~k = (a~k + a†−~k ) = displacement,
2ωk µL

we obtain the mean-square displacement (with n(ω) = 1/(eβω − 1) for bosons)

X ei(~k+~k 0 )·~r X 1
hu2 (~r )i ∼ √ h(a~k + a†−~k )(a~k 0 + a†−~k 0 )i ∼ (1 + 2n(ωk ))
ωk ωk0 ωk
~
k,~
k0 ~
k
Z Z Z
1 ~ωk T % T
∼ dd k coth ∼ dd k 2 ∼ T dk kd−3 .
ωk 2kB T ωk
In the limit T → 0, we have coth → 1 and in 1D only a quasi–long-range order survives,
as opposed to ferromagnets. Note that E ∼ ~ωk a†k ak ∼ ωk2 µL|uk |2 ∼ c2 µLk2 |uk |2 ∼
(ε/µ)µL|uk |2 → E/L ∼ ε|∇u|2 , which is the usual elastic energy density.

Heisenberg ferromagnet, magnon = Goldstone-mode:

X 1 2
H = εk a~†k a~k , εk = Jk ,
2
~
k
 X n(εk ) 
hmi = µ 1 − ,
N
~
k
Z Z Z
hδmi 1 T% T
∼ dd k βε ∼ dd k ∼T dk kd−3 .
µ e k −1 0 εk 0

In the limit T → 0, all modes k > 0 remain unoccupied and the mode k = 0 leaves the
order unchanged; an ordered ferromagnetic state survives.

BEC, sound wave = Goldstone mode:

X q
H = Ek a~†k a~k , Ek = ε2k + 2µεk ∼ c~k,
~
k

mc2 n0
Z
n0 ∼ n − dd k ha~†k a~k i, hn~k i =
[1 + 2n(Ek )],
Ek 2n
mc2 n0 T
Z Z
T%
∼ n − dd k → ··· − T dk kd−3 ,
nEk2 0

where we have used the Bogoliubov theory of interacting Bose gases to obtain the expres-
sion for hn~k i (with c2 = nV /m the speed of sound and n, n0 the total and condensate
densities). With Ek2 ∝ k2 at small k, we find that the normal density n − n0 diverges
in d = 1, 2 and there is no room for a condensate. Comparing the three cases, we see
that for sound waves (in crystals and in the BEC) the occupation number n(ωk ) and the
8.7. MAGNETIC ORDER IN TWO DIMENSIONS 159

Destruction of quasi–long-range order

According to (8.76) a quasi-ordered phase survives at low temperatures—

this is often called the Berezinskii phase. Would the Gaussian Hamiltonian
(8.70) be valid at all energies, this result would even persist to high energies.
However, (8.70) is not the original Hamiltonian—the latter is given by (8.69)
and involves ϑ as an angular (i.e., compact) variable. As a consequence,
there are not only elastic excitations as described by (8.70), but one has to
take into account the existence of topological excitations as well. The latter
trigger a phase transition into a disordered high-temperature phase that is
characterized by a fast decaying correlator C(R, ~ 0) ∼ exp[−R/ξ(T )]. Using
ϑ(R~ ) = ϑ(R, θ) the field of such an excitations can be simply expressed
through
ϑ(R~ ) = kθ + ϕ, ϕ = const. (8.80)
and a set of typical spin configurations belonging to k = ±1 and ϕ = 0, π/2
is sketched in Fig. 8.13.

Fig. 8.13: Spin configurations with ϑ(R)~ = kθ + ϕ defining the angle of the
XY spin with respect to the x-axis at the position R~ = R(sin θ, sin θ). From
left to right k = 1, ϕ = 0; k = −1, ϕ = 0; k = 1, ϕ = π/2; k = −1, ϕ = π/2.

~ R
Using ∇ϑ( ~ ) = ±R
~ ⊥ /R2 , one finds that
I
d~` · ∇ϑ
~ = ±2π; (8.81)

for k = ±1; in general, one finds the winding number

I
d~` · ∇ϑ
~ = 2πk, k = ±1, ±2, . . . (8.82)

amplitudes (hu2 i ∝ ωk and hn~k i ∝ 1/Ek ) both contribute a factor k, producing the overall
factor k2 in the denominator, while for the ferromagnet, the dispersion is quadratic and
all k2 contribution is from n(εk ). Note that the ideal Bose gas behaves as the ferromagnet
with a quadratic dispersion and spectral factor unity. See HẆatanabe and H. Murayama
for more on this.
160 CHAPTER 8. MAGNETIC ORDER

characterising this type of topological defect called vortex. An isolated vor-

tex costs an energy (k = ±1)
Z Z
J 2 ~ 2 1 R∞
E≈ d R (∇ϑ) = πJ dR R 2 = πJ ln . (8.83)
2 R R0
The limits R0 and R∞ are usually given by the lattice constant R0 ≈ α · a,
with α ∼ 1, and the system size L,
L
EV = πJ ln . (8.84)
αa
Note that the uncertainty of α adds a numerical ∼ J ln α to the vortex
energy—it is the precise shape of the vortex near its core that determines
this numerical and that usually has to be found by a numerical solution of
the problem: the vortex configuration fulfills the equation ∆ϑ = 0 in R2 \{0}
(dotted plane) and thus forms a (local) minimal configuration (there is no
minimal-energy, i.e., force-free solution, for two vortices as these interact at
finite distances).

Fig. 8.14: Vector-fields associ-

j core
ated with vortex configurations
B in 2D and in 3D. This type of
~
vector-fields describe the B-field
B j of a linear current (middle) or the
2D flow-field of a circulating current
3D j in a superfluid or superconductor
(right).

~ corresponds to the one of a B

The gradient-field ∇ϑ ~ field that is gener-
ated by a current carrying conductor or by the vortex current of a rotating
superconductor/superfluid, see Fig. 8.14; in 2D, this is a planar vector-field,
while in 3D the vector-field belongs to a linear topological defect that can
bend.

E VV
Fig. 8.15: Attractive
potential between a vor-
tex and an anti-vortex
ln R /a
at a distance R.

0 1 R/a

A second vortex with opposite winding shields, i.e., compensates, the

field of the first vortex at large distances; for such a vortex–anti-vortex pair
8.7. MAGNETIC ORDER IN TWO DIMENSIONS 161

~ VV̄ (R)| ∝ 1/R2 → 0, since

the far-field vanishes quadratically, |∇ϑ
~ V
∂ ∇ϑ
~ V (R)
|∇ϑ ~ + ∇ϑ
~ V̄ (R
~ −R
~ 0 )| ∼ ~ 0, ∝ 1
·R
~
∂R R2
and we obtain the energy (see Fig. 8.15)
~ ~
EVV̄ (R ~ 2 ) = 2πJ ln |R1 − R2 | .
~ 1, R (8.85)
a
Configurations of vortex–anti-vortex pairs for ϕ = 0 and ϕ = π/2 are
sketched in Fig. 8.16.

−1
1 1

−1

Fig. 8.16: Configurations for vortex–antivortex-pairs, ϕ = 0 left and ϕ =

π/2 right.

Generally, the energy of two vortices of windings k1 , k2 a distance R

away can be written as

Ek1 ,k2 (R) = πJ(k1 + k2 )2 ln(L/a) − 2πJk1 k2 ln(R/a). (8.86)

We conclude that low-lying excitations of the Berezinskii phase are a com-

bination of elastic deformations and a “neutral ” gas of vortices,
Z Z
J ~ r ]2 − πJ d2 R d2 R0 nV (R ~ 0)
~ )nV (R
HB ≈ d2 R [∇ϑ
2
h |R~ −R ~ 0| Li
× ln − ln , (8.87)
X a a
~) =
nV (R ~ −R
ki δ(R ~ i ),
i

~ ) describing the singular part of the vector-

with a vortex density nV (R
~
field ϑ(R ). The term ln(L/a) in the interaction enforces the “charge” =
“winding”- neutrality
Z X
~)=
d2 R nV (R ki = 0. (8.88)
i
162 CHAPTER 8. MAGNETIC ORDER

With Eq. (8.87), we have found a Hamiltonian that can describe the
phase transition that leads to the loss of the quasi-LRO that is present in
the low-temperature phase of the XY -model. The need for the existence of
such a phase transition in 2D became clear from the work of Berezinskii, who
found that the low-temperature phase is characterized by quasi-LRO and
stiffness, and high-temperature expansions, telling that the order is short-
ranged, C(R) ∼ e−R/ξ . It was Kosterlitz and Thouless,16 who in 1973 (in J.
Phys. C 6, 1181(1973)) proposed the following scenario for the transition: In
the low temperature phase the vortices are pairwise bound and thus do not
destroy the quasi-LRO. More pairs appear with increasing temperature—
they screen the interaction between larger pairs which finally leads to the
dissociation of pairs at a temperature TBKT ≈ πJ/2kB . The appearance
of free vortices above TBKT destroys the quasi-LRO. The calculation is a
beautiful example of renormalisation in classical statistical mechanics, see
Sec. 17. They obtained the Nobel prize for this work in 2016 (together with
Haldane, in the more general setting of ‘topology in condensed matter’).

Fig. 8.17: Vortices and correlation

or coherence length ξ(T ) below and
− + above the BKT transition. Vortices
+ + −
− − are bound in pairs at T < TBKT
+ − + and thus do not destroy q-LRO—
+ + the correlation length ξ(T ) is in-
− −
ξ( T )
finite for all T < TBKT , describ-
ing a line of critical points. Above
8

the transition, vortices unbind on

ξ (T ) scale ξ(T ), thus establishing short
range order. The correlation length
T ξ(T ) decreases exponentially with
TBKT
increasing T .

Berezinskii-Kosterlitz-Thouless transition

A simple argument for estimating TBKT is the following: A vortex costs an

energy πJ ln(L/a). The possibility to place the vortex at one of (L/a)2
lattice points provides an the entropy S = kB ln(L/a)2 . The free energy of
a single vortex
F = (πJ − 2kB T ) ln(L/a) (8.89)
becomes negative at the BKT temperature
π
kB TBKT = J (8.90)
2
16
Note that Berezinskii understood the role of vortices in this transition
8.7. MAGNETIC ORDER IN TWO DIMENSIONS 163

and thus it becomes favourable for the system to generate such excitations.
Above TBKT , vortices are bound at short distances R < ξ(T ) and therefore
the quasi-order survives on the scale of ξ. At distances R > ξ, free vortices
appear and C(R) goes to 0 exponentially. Below TBKT , we have ξ(T <
TBKT ) = ∞. Above the transition at TBKT , the coherence length diverges
upon approaching TBKT from above according to
b
ξ(T ) ∼ exp p .
(T − TBKT )/TBKT

8.7.3 Heisenberg model [O(n) model]

Including thermal fluctuations in the Heisenberg model leads to the non-

linear sigma model
Z
Z = D[σ(~r )]D[π(~r )] δ(σ 2 + π 2 − 1)e−H/kB T , (8.91)

Ja2
Z d Z d
d r d r ~
with H = d
(∂i~n ) · (∂i~n ) − µh · ~n,
2 a ad
~n = (~π , σ),

where we chose ~h = hêz and the symmetry broken along ~n0 = (~0, 1) = êz .
The analysis of the nlσ model is non-trivial and we cannot go into detail
here. The renormalisation group analysis delivers a ferromagnetic phase
transition at
d−2
kB Tc ≈ 2πJ , (8.92)
n−2
and hence Tc = 0 for the n = 3, d = 2 Heisenberg model in 2D (note
that in addition to spin-waves, also topological excitations in the form of
Skyrmions are relevant in the characterization of the system, e.g., the cor-
relation length, see Polyakov’s book). In 3D the Heisenberg model shows
a transition at Tc ∼ J. The overall behavior of the critical temperatures
in magnetic transitions as a function of n and d is sketched in Fig. 8.18,
where we include our knowledge gained so far from mean-field theory and
the analysis of fluctuations.
164 CHAPTER 8. MAGNETIC ORDER

MF
MF
MF
QLRO
Tc n BKT
3 Heis
2 XY
LRO
1 Ising

0 1 2 3 4 d

Fig. 8.18: Critical temperatures Tc in the three different magnetic models

with dimensionalities n and d; n = 1—Ising model, n = 2—XY model,
n = 3—Heisenberg model.
Chapter 9

Gas–Liquid Transition

Besides the magnetic phase transitions, the gas–liquid transition in real

gases is another classic example of a first-order phase transition with a crit-
ical endpoint. Although there is no broken symmetry in this example, the
transition can be described by a Landau mean-field theory with the density
jump δn = nli − ngas serving as an ‘order parameter’. The classic model
describing a real gas goes back to the dissertation of Johannes Diderik van
der Waals.

9.1 The van der Waals Gas

In his thesis on ‘Over de Continuiteit van den gas- en vloeistoftoestand’ (Lei-

den, 1873), van der Waals put down an equation of state for real gases that
describes a gas–liquid phase transition1 The Ansatz was pretty ingenious:
van der Waals took the short-range repulsive interaction between atoms into
account by reducing the total volume v (per mole) by the volume b occupied
by the atoms themselves, hence he replace v by v−b in the ideal-gas equation
of state. Second, he accounted for the long-range attraction −C/r6 between
neutral (but polarizable) atoms by reducing the pressure p in the ideal gas
equation with a term2 a/v 2 , p → p + a/v 2 . These corrections results in the
famous van der Waals equation of state (for one mole of substance) in the
form  a
(v − b) p + 2 = RT (9.1)
v
1
To place this achievement into proper perspective, one may note that Boltzmann
called van der Waals ‘the Newton of the real gases’. Van der Waals was awarded the
Nobel prize in 1910, ‘for his work on the equation of state for gases and liquids’.
2
In mean-field theory, where all particles interact with all other, the pairwise attraction
between particles contributes an energy ∝ N 2 /V and the derivative with respect to V
scales as N 2 /V 2 .

165
166 CHAPTER 9. GAS–LIQUID TRANSITION

with the parameters a and b describing the attractive and repulsive interac-
tions. The p-v-diagramm then exhibits the characterisitc isotherms shown
in Fig. 9.1.

Fig. 9.1: Isotherms of the van der

p Waals gas in the p-v diagram. Be-
low a critical temperature Tc , the
T isotherms change from monotoni-
pc cally decreasing to non-monotonic
curves exhibiting minima and max-
ima.

Tc

vc v

9.1.1 Universal form

The isotherms p(v) are monotonically decreasing at high temperatures T >

Tc , while below Tc the cubic equation exhibits three solutions for v at fixed
p < pc and T < Tc . The critical parameters Tc , pc , and vc follow from the
three conditions
RT a
p = pvdW (v, T ) → = p + 2,
v−b v
∂p RT 2a
=0 → = 3,
∂v (v − b)2 v
2
∂ p RT 3a
2
=0 → 3
= 4, (9.2)
∂v (v − b) v
and we obtain the results
1 a 8 a
pc = , vc = 3b, Tc = . (9.3)
27 b2 27 bR
Going over to the dimensionless variables π̄ = p/pc , ν̄ = v/vc , and τ̄ = T /Tc ,
we arrive at the van der Waals equation of corresponding states 3 ,
 3
π̄ + 2 (3ν̄ − 1) = 8τ̄ . (9.4)
ν̄
Combing the equations in (9.3) one finds an interdependence between the
critical paramters,
RTc 8
= . (9.5)
p c vc 3
3
Today, this would probably be called the universal law of real gases.
9.1. THE VAN DER WAALS GAS 167

9.1.2 The Maxwell construction

In correspondence with the discussion in Sec. 1.5.5 where we constructed

the free energy of a two-phase system by drawing the convex envelope of the
two free energy functions, here, we construct the proper isotherms in the p-v
diagram of the van der Waals gas. As the temperature T drops below critical,
i.e., for T < Tc , the isotherms exhibit segments with increasing derivatives
∂V p|T > 0, that correspond to the violation of the stability criterion κT > 0.
This indicates to us the presence of a phase transition appearing below Tc
and we have to correct the van der Waals isotherms. Consider an isotherm
at T < Tc and cut it with an isobar p < pc , see Fig. 9.2.

p Fig. 9.2: In correcting the van der

T < Tc
Waals isotherm, we determine the
pressure p < pc of the isobar such
A that the chemical potentials of the
B two phases A (liquid) and B (gas)
match up, µli (p, T ) = µgas (p, T ).

vA vB v

We determine the pressure p such that the chemical potentials align,

µli (p, T ) = µgas (p, T ), that leads us to a first-order phase transition at
Tgl (p) < Tc . With µ = u − T s + pv we have to satisfy the equation

ugas − uli − T (sgas − sli ) + p(vgas − vli ) = 0. (9.6)

We then need the expressions for the internal energy u and the entropy s:
the internal energy u of a real gase depends on the volume v, see (1.42),

∂u a
∂v u|T = T ∂T p|v − p ⇒ = , (9.7)
∂v T v2
a consequence of the interatomic attraction (i.e., the stored potentiel en-
ergy). We obtain the energy difference ugas − uli by straightforward integra-
tion
Z B
Tli =Tgas a a a
ugas − uli = dv 2 = − + . (9.8)
A v vgas vli

For the entropy we can write

1 1  ∂u ∂u  p
ds = (du + pdv) = dT + dv + dv (9.9)
T T ∂T V ∂v T T
168 CHAPTER 9. GAS–LIQUID TRANSITION

and using (9.7) and (9.1), we obtain

cv R
ds = dT + dv
T v−b
Z B
Tli =Tgas R vgas − b
⇒ sgas − sli = dv = R ln . (9.10)
A v−b vli − b

The condition (9.6) then assumes the form

a a
0=− − RT ln(vgas − b) + + RT ln(vli − b) + p(vgas − vli ) (9.11)
vgas vli
or
a a
p(vgas − vli ) = + RT ln(vgas − b) − − RT ln(vli − b). (9.12)
vgas vli

On the other hand, we can determine the work or area in the p-v diagram
Z B Z B Z B
dv a dv
dv p = RT −
A A v − b li v2
B a B
= RT ln(v − b) + (9.13)
A v A
and find that it agrees with the right hand side of (9.12), hence
Z B
dv p = p(vgas − vli ). (9.14)
A

We find that the equilibrium condition µli = µgas can be reduced to the
condition of equal shaded areas in Fig. 9.2, which is commonly known as
the Maxwell construction for the physical correction of the van der Waals
Isotherms. One easily shows, that this construction corresponds to the con-
vex completion of the free energy surface in section 1.5.5.

p Fig. 9.3: Corrected van der Waals

T < Tc
isotherms describing a first-order
B’ phase transition between liquid (A)
A β’ and gas (B) phases. The points A0
B and B 0 define the spinodals, the end-
α’
A’ points of metastable phases. The
segments α0 and β 0 describe a super-
heated liquid and an undercooled or
v oversaturated gas.

Figure 9.3 provides a sketch of the corrected isotherm describing the

liquid–gas first-order phase transition with points defining the spinodals (A0
9.1. THE VAN DER WAALS GAS 169

and B 0 ) and regions with metastable phases (α0 and β 0 ). Repeating this
construction for all pressures p < pc , we obtain the p-v diagram in Fig.
9.4. At the fog line first liquid droplets appear in the gas phase, while
at the boiling line first gas bubbles show up in the liquid. Going on a
path γ around the critical point we can go continuously from the gas- into
the liquid phase—this is the core statement made in the title of van der
Waals’ thesis. Note that one cannot liquefy the gas through pressurizing
when the temperature is high, T > Tc —gas- and liquid phases cannot be
distinguished at these high temperatures. Interesting further topics are the

p Fig. 9.4: p-v diagram with

γ gas, liquid, and coexistence
regions. The latter starts
C fog line from the boiling line and ends
pc
at the fog line.
Tc
l K g

vc v
boiling line

liquefaction of gases (with a Joule-Kelvin process) and the definition of the

gas temperature, i.e., how to relate the measured gas temperatue θ to the
absolute temperature T . Here, we now focus on the microscopic derivation
of the van der Waals equation and the description of the phase transition
within a mean-field theory.

9.1.3 Microscopic analysis

We start from the classical canonical partition function for N interacting

particles, see Eq. (5.37),

d3N q
Z
1  V N h X i
ZN (V, T ) = exp −β V (~
q i − ~
q j ) . (9.15)
N ! λ3 VN
i<j

The r−6 interaction between the neutral atoms is long-ranged, i.e., every
atom interacts with many neighbors, motivating a mean-field description
with a potential
1 V0 −r/R0
V (r) = − e (9.16)
8π R03
170 CHAPTER 9. GAS–LIQUID TRANSITION

that decays on the scale R0 → ∞. The prefactor guarantees the regular

limit Z
d3 rV (r) = −V0 . (9.17)

The long-range part of the configurational integralRthen contributes a factor

(we replace the sum over ~qi by an integral (N/V ) d3 q such that each par-
ticle contributes with a term −N V0 /V to the interaction energy; the factor
1/2 corrects for double counting when every particle interacts with all other)
2 /2V d3 rV (r) 2 /2V
R
lr
ZN = e−β(N )
= eβV0 N . (9.18)
The short-range repulsion contributes a factor

d3N q Y
Z
sr
ZN = χσ (~qi − ~qj ) (9.19)
VN
i<j

with χσ (r) the characteristic function on r > σ. This is the partition func-
tion of a gas of hard spheres of radius σ which we approximate by
sr
ZN = (V − N b)N /V N , (9.20)

where b is the excluded volume per particle. Collecting factors, we find the
partition function of the van der Waals gas

1  V N βV0 N 2 /2V  V − N b N
ZN (T, V ) = e (9.21)
N ! λ3 V
as well as its free energy

aN 2
F (T, V, N ) = Fid−gas − − kB T N log(1 − bN/V ) (9.22)
V
with a = V0 /2. The equation of state follows from the partial derivative
p = −∂V F and we obtain the van der Waals equation

kB T a kB T b kB T a
p= − 2+ = − , (9.23)
v v v v−b v − b v2
with v = V /N the volume per particle. The free energy per particle assumes
the form (see Eq. (8.23))
 a
f = kB T log[λ3 /(v − b)] − 1 − ,

(9.24)
v
which coincides with the ideal gas result when a = 0 and b = 0. The entropy
per particle is given by −∂T f ,

s = kB log[(v − b)/λ3 ] + 5/2

 
(9.25)
9.1. THE VAN DER WAALS GAS 171

and the internal energy results from the difference u = f + T s,

u = (3/2)kB T − a/v. (9.26)

Finally, the Gibbs free energy per particle assumes the form
 2a kB T v
g = f + p v = kB T log[λ3 /(v − b)] − 1 −

+ . (9.27)
v v−b
It is interesting to plot
R p this quantity at fixed T by using the fact that
g(p, T ) = g(T, p0 ) + p0 v(p)dp (since v = ∂p g|T ) and use the form of v(p)
sketched in Fig. 9.5(a): we follow this curve from D to E to F to G, see
Fig. 9.5(b), and use that ∂p v is the curvature, then g first rises as a concave
curve, becomes convex at E, and concave again at F. √ Near E (F), we have
p−pE ∝ −(v −vE )2 (p−pF ∝ (v −vF )2 ) and since δv ∝ δp, we find that the
curvature diverges and g has cusps at E and F. The convex branch between
E and F is unstable.
E
p G
(a) g (b) C’’
G
F C C’

vli v gas v
E
C’ C
C’’
D D
F p

Fig. 9.5: p-v diagram (a) and Gibbs energy per particle g(p) at fixed T
(b) for the van der Waals gas with two stable (solid lines), two metastable
(dashed) and one unstable branch (dotted).

9.1.4 Critical region of the gas–liquid transition

We introduce the relative (to the critical point) parameters p/pc = 1 + π,

v/vc = 1 + ν, T /Tc = 1 + τ , and write the van der Waals equation in its
universal form
3 8(1 + τ )
(1 + π) + = . (9.28)
(1 + ν)2 2 + 3ν
In the critical region, π, ν, and τ are all small and we can expand. In order
to select the expansion parameter we identify the volume V as the order
parameter (m in the magnetic system) and p as the external (conjugate) field
(h in the magnetic system), while T plays its usual role as the parameter
separating disordered (T > Tc ), critical (T ≈ Tc ), and ordered (T < Tc )
172 CHAPTER 9. GAS–LIQUID TRANSITION

regions and taking us through the transition at Tc . Hence, it is convenient

to expand π in ν and we obtain the result
3 21 27
π = 4τ − ν 3 + ν 4 − 6τ ν + 9τ ν 2 − τ ν 3 + . . . . (9.29)
2 4 2

We start with the critical isotherm which separates monotonous from

non-monotonous curves at τ = 0,
3 21
π(ν) ≈ − ν 3 + ν 4 + . . . . (9.30)
2 4
The leading term then provides the scaling

ν ∝ π 1/3 (9.31)

for the ‘order parameter’ ν(h) versus its conjugate field π at Tc . This
corresponds to the mean-field exponent δ = 3, ν ∝ π 1/δ . The corresponding
result in the magnetic setup is identical, m ∝ h1/3 . The correction ν 4 tells
that cubic curve is asymmetric, steeper at negative values of ν and flatter
on the positive side. This tells that the gas volume grows faster than the
liquid volume shrinks when the temperature T decreases below Tc , see Fig.
9.6.

π Fig. 9.6: Critical isotherm at

τ = 0 and Maxwell construction
at τ < 0. The volumes νgas ,
νli , and νunst (the unstable solu-
0 ν tion) are found from solving the
τ<0
π gl cubic equation (9.29), while the
τ=0 pressure πgl (τ ) follows from the
Maxwell construction (9.32).
νli νunst νgas

Below the critical isotherm, we have to replace the non-monotonous part

in π(ν) by the Maxwell construction. Defining the constant pressure at
the transition as πgl (τ ), we have to satisfy the condition
Z νgas
dν[π(ν) − πgl ] = 0 (9.32)
νli

which relates πgl to the volumes νgas and νli , see Fig. 9.6,
2
πgl = 4τ − 3τ (νgas + νli ) + 3τ (νgas + νli νgas + νli2 ) + . . . . (9.33)

We then have to solve the cubic problems π(νgas ) = πgl , π(νli ) = πgl , with
π(ν) given by (9.29), for νgas (πgl ) and νli (πgl ) and solve Eq. (9.33) for πgl (τ );
9.1. THE VAN DER WAALS GAS 173

the third solution νunst to the cubic equation defines the unstable volume
with maximal Gibbs energy. Fortunately, to lowest p order we can solve the
problem more easily. Assuming a solution ν = α |τ |, we can rewrite Eq.
(9.29) in the form
3 
π − 4τ = − α3 − 6α |τ |3/2 + O(τ 2 ), (9.34)
2
with the solutions (using α(α2 − 4) = 0)
p p
πgl ≈ 4τ < 0, νgas ≈ 2 |τ |, νunst ≈ 0, νli ≈ −2 |τ | (9.35)

defining the boundaries of the coexistence region. The result νgas/li ∝

p
|τ | for the ‘order parameter’ ν(τ ) as a function of τ defines the mean-
field critical exponent β = 1/2, in agreement with the prediction of Landau’s
mean field theory. The analogous result for the magnetic system is m ∝ |τ |β
with β = 1/2.
p
Pushing the expansion in |τ | further, one can find the next order cor-
rections
p 18 p 18
νgas ≈ 2 |τ | − τ, νli ≈ −2 |τ | − τ, (9.36)
5 5
such that the diameter of the coexistence region is not centered but
increases linearly in |τ |,

(vgas + vli )/2 ≈ vc (1 + 18|τ |/5). (9.37)

The coexistence region together with the diameter is shown in the v–T
diagram of Fig. 9.7; it corresponds to the ferromagnetic region in the m–T
diagram of magnetism, see Fig. 8.6.
In Fig. 9.8, we sketch the analog of the m–T –h diagram in Fig. 8.4 for
the gas–liquid transition; note that Tc in Fig. 9.8(b) is corresponds to the
tip of the ruled surface in Fig. 9.8(a) in rotated coordinates.
When transforming v to p, we obtain the p–T diagram with the first-
order transition line pgl (T ). We can obtain the shape of the transition
line by inserting the results (9.36) into (9.33) and obtain

48 2
πgl ≈ 4τ + τ . (9.38)
5
The insets in Fig. 9.7 show the first-order line in the gas–liquid transition
together with the analogous line in the magnetic transition.
Continuing the transition line πgl within the π–τ diagram beyond τ = 0,
we move on the critical isochore where ν = 0, see (9.29)

πci = 4τ. (9.39)

174 CHAPTER 9. GAS–LIQUID TRANSITION

π h
gas
ν m
liquid 0 τ 0 τ
gas

0 τ 0 τ

liquid

Fig. 9.7: Gas–liquid coexistence region in the v–T diagram. When com-
pared to the ferromagnetic magnetization in the corresponding m–T diagram
of magnetism, we note the absence of symmetry m ↔ −m in the gas–liquid
coexistence region. Insets: Note the similar asymmetry in the p–T dia-
gram versus the h–T diagram with its symmetric first-order transition line
at h = 0.

m
p
T
C h
0
l
Tc
g T

v
Fig. 9.8: Surfaces of thermodynamic equilibrium states for the gas–liquid
transition in p–v–T space as compared to the h–m–T diagram for the ferro-
magnetic system.

The critical isochore in the vdW problem plays the same role as the h = 0,
T > Tc line in the magnetic problem. It is along this line where the critical
scaling behavior develops when approaching the second-order critical point
‘from above’. When the critical point is approached along another line in
the π–τ (or h-τ ) diagram, this scaling behavior is modified.
Next, we calculate the isothermal compressibility κT = −v −1 ∂p v—
its analogue in the magnetic system is the suceptibility χ = ∂h m, see Eq.
(8.38). Above Tc , τ > 0, the derivative of Eq. 9.29 on the critical isochore
ν = 0 provides us with the result
1 ∂ν 1
κT = − ≈ . (9.40)
pc ∂π 6pc τ
Below, we determine the analogous quantity along the coexistence bound-
9.1. THE VAN DER WAALS GAS 175

aries πgas (νgas ) and πli (νli ). Using again Eq. (9.29), we have ∂ν π = −6τ +
18τ ν − (9/2)ν 2 and inserting ν 2 ≈ −4τ > 0, we obtain the compressibility
below Tc ,
1 ∂ν 1
κT = − ≈ , (9.41)
pc (1 + ν) ∂π 12pc |τ |
half the value at temperatures T > Tc , τ > 0. The results Eqs. (9.40) and
(9.41) scale ∝ 1/|τ | as predicted by mean-field theory. The analogous result
for the magnetic system is χ ∝ 1/|τ |γ with γ = 1.
As a last quantity, we calculate the specific heat cv (T ) at constant
volume. On the critical isochore ν = 0, we use the expression Eq. (9.26) for
the internal energy and obtain the standard ideal gas result
3
cv = ∂T u|v = kB (9.42)
2
valid above Tc . Below Tc , we have to take the volume term into account,
since the system is in a mixed gas–liquid state on the critical isochore. The
volume ratios at a given ν are

ν = rνli + (1 − r)νgas , (9.43)

hence, on the critical isochore ν = 0, we have r/(1 − r) = −νgas /νli . The

internal energy is u = r uli + (1 − r) ugas and using again Eq. (9.26), we
obtain the inner energy of the mixture
3 a 1  ν νli 
gas
u = kB T − − . (9.44)
2 vc νgas − νli 1 + νli 1 + νgas

Expanding to order τ 2 , we obtain

3 a 2
u ≈ kB T − (1 − νgas νli + νgas νli (νgas + νli ) + νgas νli2 ) (9.45)
2 vc
3 a 56 
≈ kB T − 1 − 4τ − τ 2 .
2 vc 25
The result for the specific heat below Tc then is
3 9  28 
cv = ∂T u|v = kB + kB 1 + τ , (9.46)
2 2 25
where we have used that a/vc Tc = (9/8)kB . As a result, we find that cv
jumps by (9/2)kB at Tc , see Fig. 9.9. Within mean-field theory, the specific
heat cv ∝ |τ |α is expected to undergo a jump, corresponding to an exponent
α = 0.
Summarizing, we have treated the gas–liquid transition within a mean-
field approximation, assuming a long (infinite) range interaction and found
all the mean-field critical exponents predicted by Landau theory.
176 CHAPTER 9. GAS–LIQUID TRANSITION

cV Fig. 9.9: Specific heat jump at the

6 kB critical temperature τ = 0 on the
critical isochore ν = 0.

3k
2 B

0 τ
Chapter 10

Lee-Yang Theorem∗

The Yang-Lee Theorem provides insights into the appearance of non-analyti-

cities producing a phase transition. Specifically, consider a physical system
with N particles in a volume V ; how can this system show a phase transition
with a singularity in the free energy? In order to find an answer to this
question, one may study the lattice gas model of attractive hard sphere
atoms, see Fig. 10.1. Here, we briefly discuss the main ideas and results of
such an analysis.

V
Fig. 10.1: Shape of the inter-
action potential describing (in-
finitely) hard spheres.
a
r
−V
0

For a hard sphere system, only a finite number of particles M (V ) can fit
in a volume V , ZN >M (V ) (V ) = 0, and the grandcanonical partition function
must be a polynomial of order M in the fugacity z,
Z(T, V, z) = 1 + zZ1 (T, V ) + z 2 Z2 (T, V ) + . . . + z M ZM (T, V ). (10.1)
The equations of state are obtained from
kB T
p = log Z(T, V, z),
V
1 z
= ∂z log Z(T, V, z). (10.2)
v V
As all the coefficients Zk in (10.1) are positive, the function Z(z ∈ C) has no
zeroes on R+ ⊂ C. Therefore, the functions p(z), v(z) and p(v) are analytic

177
178 CHAPTER 10. LEE-YANG THEOREM∗

on R+ and p, −∂p/∂v ≥ 0,
∂p ∂p ∂z kB T 1
= =
∂v ∂z ∂v zv ∂v/∂z
kB T V
= − 3 ≤ 0,
v (hN 2 i − hN i2 )
where hN 2 i − hN i2 = z∂z (z∂z log Z) = z∂z (V /v).
As there are no singularities, no phase transition appears in a finite system.
Only when going to the thermodynamic limit V → ∞, n = N/V = 1/v =
const, there can be singularities in the quantities
1
p = kB T lim log Z(T, V, z),
V →∞ V
1 z
= lim ∂z log Z(T, V, z). (10.3)
v V →∞ V
Lee and Yang have shown that phase transitions are controlled by the dis-
tribution of zeroes in Z(z ∈ C)—phase transitions appear when these zeroes
approach the real axis as sketched in Fig. 10.2, thereby cutting the axis into
separate segments.

Fig. 10.2: Zeroes of Z(z ∈ C)

in the complex plane.

Theorem 1: The limit

1
F∞ (z) ≡ lim log Z(T, V, z) (10.4)
V →∞ V

exists for all z > 0 and is independent of the volume V if the surface
∂V ≤ V 2/3 .
Theorem 2: Let SR be a segment of R+ , U (SR ) ⊂ C a region around SR
with no zeroes of Z. Then (1/V ) log Z converges uniformly in U for
V → ∞ and its limit F∞ (z) is analytic.

Theorem 2 allows us to swap the operations limV →∞ and z∂z in (10.3)

within the region U , and hence
p(z) = kB T F∞ (z),
1 ∂
= z F∞ (z), (10.5)
v(z) ∂z
thus defining a stable phase with p, −∂p/∂v > 0 within the parameter range
SR .
 179

1. Let SR = R+ , see Fig. 10.3. Then there exists only a free phase.

Fig. 10.3: Singularities of log Z

0 SR in the complex plane and phase
diagrams for the situation where
SR = R+ .
p 1/ v p

z z v

2. Let R+ = SR1 + SR2 . Then there exist two phases that transform into
each other via a phase transition.

Fig. 10.4: Singularities of log Z

in the complex plane and phase
1 2
0 SR SR diagrams for the situation where
R+ = SR1 + SR2 . F∞ (z) ∝ p(z) is
piece-wise analytic according to
p 1/ v p
b a theorem 2. A jump in ∂F∞ /∂z
b
a generates a first-order transition
z z vb va v
(middle), whereas a second-order
transition is realized for a con-
p 1/ v p
a tinuous first and a discontinuous
a
second derivative of F∞ (z) (bot-
z z v tom).

The behavior of the function F∞ (z) ∝ p across the two segments SR1
and SR2 determines the type of transition: The transition is first-order if
the derivative ∂z F∞ exhibits a discontinuity or jump. A continuous first
derivative and a discontinuity or jump in the second derivative ∂z2 F∞ signals
a second-order transition, see Fig. 10.4
The Lee-Yang circle theorem tells that for the hard sphere lattice gas
with a potential as sketched in Fig. 10.1 the zeroes of Z(T, V, z) lie on a unit
circle in the complex plane. At any finite V , these zeroes appear in discrete
complex-conjugate pairs away from the real axis. In the thermodynamic
limit, the zeros accumulate and approach the real axis, ‘pinching it off’ at
z = 1, what results in two phases with a transition at z = 1.
180 CHAPTER 10. LEE-YANG THEOREM∗
Chapter 11

Landau Theory

In this chapter, we turn our focus to phenomena that appear in the proximity
of a phase transition. We start with a phenomenological mean-field theory
for the order parameter Φ (including also a gradient term), the Landau
theory of phase transitions, and then discuss the various types of transitions,
second-order, first-order, and multicritical points. We start with a short
overview.

11.1 Overview

The underlying free energy F (T, Φ) is described as a Taylor series in the

~
order parameter and its gradient ∇Φ, taking into account the symmetries
of the problem
r 2 c ~ 2
f (T, Φ) = Φ − wΦ3 + uΦ4 + u0 Φ6 + |∇Φ| , (11.1)
2 Z 2 Z
g(T, h) = f − hΦ, F = dd x f, G = dd x g.

The terms ∝ r, u determine Φ in a phase transition of second order. A cubic

term ∝ w induces a first-order phase transition—its presence depends on the
symmetry of the problem. The term ∝ c allows us to study non-local effects
and spatial correlations at the phase transition. The term ∝ u0 becomes
important when the coefficients r and u vanish or turn negative, then u0
has to stabilise f . Such a situation appears in a multi-dimensional phase
diagram, where next to the temperature T other parameters (for example
the chemical potential µ) modify the coupling constants r, u, u0 in f . One
then finds multicritical points, for example the tricritical point where two
phase transitions of first and second order coincide. Other examples include
bicritical and tetracritical points.

181
182 CHAPTER 11. LANDAU THEORY

The original Landau expansion with powers of Φ has later been extended
to a field-theoretical framework (Landau-Ginzburg-Wilson) including also
gradient terms ∇Φ. The Landau free-energy functional F [Φ(~r )] (or effective
coarse grained Hamiltonian) over the order-parameter field Φ(~r ) can be
understood as the result from taking a partial trace (coarse-graining) over
microscopic degrees of freedom Φµ described by a microscopic Hamilonian
H[Φµ (~r )],
Z Y
Z = D[Φ(~r )]TrΦµ δ[Φ(~r ) − Σ(Φµ (~r ))]e−βH[Φµ (~r )]
~
r
Z
= D[Φ(~r )]e−βF [T,Φ(~r )] , (11.2)
Q
with the product ~r δ[Φ(~r ) − Σ(Φµ (~r ))] filtering out all Φµ configurations
which give the same Φ. Here, Σ denotes a function that takes the micro-
scopic degrees of freedom Φµ (~r ) near ~r as an input to generate a coarse-
grained variable Φ(~r ), e.g., see (14.26). Hence, while the Hamiltonian H
and field Φµ describe the microscopic problem, the free energy F and the
order-parameter field Φ describe an effective coarse-grained problem. The
underlying idea is that F describes the correct long wave length and low-
energy physics that is relevant for the analysis of the phase transition. As an
illustration of such a coarse graining procedure, we may refer to the discrete
microscopic spin degrees of freedom S~i in an Ising model that become con-
tinuous variables φ(~r ) = hS(~r )i under averaging over small local domains;
the resulting effective φ4 theory, see Eq. (8.43), then describes the Ising spin
problem as an effective continuous theory. Note that the coefficients in the
coarse grained free energy or effective Hamiltonian in general depend on T ,
which is due to the coarse-graining over microscopic degrees of freedom in
Eq. (11.2).
Note the differences and similarities between our mean-field approach in
chapters 8 and 9 and the present discussion: in Chaps. 8 and 9, we have
started from a microscopic Hamiltonian and have introduced a Weiss-field
by replacing the neighboring microscopic degrees of freedom with its local
average. This mean-field theory has 1/z as its small parameter, with 1/z
small implying small fluctuations. The original Landau free energy was just
an expansion of a phenomenological free energy f (T, Φ) near Tc in a small
order parameter Φ, so the small parameter here is the order parameter Φ
itself. Taking the saddle-point approximation to the path integral (11.2),
ZMF = e−βFMF (T ) .
then provides us with a mean-field theory of the type constructed in Chaps.
8 and 9, bringing us full circle back.
However, using the path-integral in Eq. (11.2), we can go further and
include fluctuations around the saddle-point. The simplest approximation is
11.2. LANDAU FREE ENERGY FUNCTIONAL 183

to include Gaussian or quadratic fluctuations around the saddle within the

Gauss model. This approximation is fine if fluctuations are small, implying
that corrections to the saddle-point result are small as well. On approach-
ing Tc , however, fluctuations become large and start interacting, usually via
a higher-order term ∝ uΦ4 in f [T, Φ(~r )]. The fluctuations then become
critical or non-Gaussian and produce new numerical values for the critical
exponents that are different from the mean-field results. The division be-
tween small (Gauss) fluctuations and large (critical) fluctuations is given by
the Ginzburg criterion, originally invented for the specific heat but in general
depending on the quantity under investigation. In close vicinity of the phase
transition, the Ginzburg region, there are general scaling laws that relate
different quantities. The relevant exponents appearing in these scaling laws
can be calculated using renormalisation group methods, see Chap. 14. The
mean-field theory in chapters Chaps. 8 and 9 (and further discussed below)
has given us a first shot on these objects.
We then obtain the following hierarchy of descriptions: Weiss-type mean-
field theory gives good results when the interaction extends over many de-
grees of freedom, either in high dimensions or when interactions are long-
range; the small parameter is 1/z. The Landau approach is fine near Tc
when the order parameter is small—in a first order transition, the jump has
to be small (weakly first-order). Small fluctuations can be treated within
a Gauss theory, large fluctuations taking over near Tc require taking their
interaction into account via an RG treatment.

11.2 Landau free energy functional

Within Landau theory, we consider the part F (T, Φi (~r )) of the free energy
that is associated with the phase transition, Ftotal = F + FRest . The field
Φi (~r ), i = 1, . . . , n describes an n components local order parameter. Ac-
cording to Landau, one constructs the free energy F under the following
provisions:

1. Close to the transition, the OP should be small, such that F can be

expanded as a power series in Φ. A modulation ∇Φ ~ 6= 0 of the or-
~
der parameter costs an elastic energy (c/2)|∇Φ| 2 —this contribution is
understood as the first term in a gradient expansion. Higher orders
~ are neglected. In this sense F describes the low-energy/long-
in ∇Φ
wave length physics of the problem/of the order parameter. The cutoff
Λ ∼ 2π/a is defined by the lattice constant a (UV-cutoff). The inter-
esting physics lives at long wavelengths k → 0 (no IR cutoff).

2. The terms appearing in F must be consistent with the symmetries of

the underlying Hamiltonian, such that the principle of spontaneous
184 CHAPTER 11. LANDAU THEORY

symmetry breaking applies. In the n = 1 Ising model this means

that F is symmetric in Φ and −Φ; only terms, which involve Φ2 ,
are allowed in the expansion of F . In the n = 3 Heisenberg model,
the O(3)-symmetry must be taken into account and hence F must be
expanded in X
Φ2 = Φ2i . (11.3)
i

Note that Φ ~ does not describe the microscopic spin field S(~ ~ r ), but
the corresponding order parameter. In particular, in the Ising model
|S z | = 1, but 0 < Φ < 1; according to mean-field theory, we have
Φ = hS z i, Φ(~r ) = hS z (~r )i` , where h. . .i` is a local average over a few
cells.
3. A field ~h conjugate to Φ
~ breaks the symmetry of F and generates
~ is found by
a finite order parameter even above Tc . The value of Φ
minimising the Legendre transformation
Z
G = F − dd r ~h(~r ) · Φ(~
~ r)

~ r ) = − δG .
Φ(~ (11.4)
δ~h(~r )
Inversely, the conjugate field ~h follows from
~h(~r ) = δF δG
→ = 0, (11.5)
~ r)
δ Φ(~ ~ r)
δ Φ(~
~ as a constraining parameter in G.
where we have used Φ
4. The parameters r, u, u0 , w, c, . . . in f , Eq. (11.1), are found by compar-
ing experimental and theoretical results (fitting parameters). When
n > m experimentally determined quantities have fixed the m param-
eters of the theory the latter becomes predictive. In some cases the
parameters can also be calculated from a microscopic theory. This
is the case, e.g., for the derivation of Ginzburg-Landau theory from
BCS-theory (Gorkov).

11.3 Second-order phase transitions

For the Ising, O(n), gas–liquid,1 . . . problems, we use a scalar order param-
eter Φ and expand f in Φ2 ,
r
f (T, Φ) = Φ2 + uΦ4 . (11.6)
2
1
There is no broken symmetry in the gas–liquid transition; nevertheless, we can let the
volume v per particle assume the role of an order parameter
11.3. SECOND-ORDER PHASE TRANSITIONS 185

At high temperatures T > Tc , f should be minimised by Φ = 0, hence

r(T > Tc ) > 0. Conversely, for T < Tc , f should be minimal for a finite Φ,
thus r(T < Tc ) < 0. Note that u > 0 guarantees the stability of f , as well as
the existence of a finite minimum. Using r(T ) = r0 (T − Tc ) and assuming a
constant u, we find the minimum of f at (see Fig. 11.1)

0, T > Tc ,
Φmin = (11.7)
± −r
 1/2
, T < Tc .
4u

T > Tc Tc T < Tc Fig. 11.1: Free energy density

f f (T, Φ), involving only squares
Φ2 of the order parameter Φ,
with a solution Φ = 0 at high
temperature T > Tc and two de-
generate minima with a finite or-
der parameter ±Φ at T < Tc .
Φ
0

In fact (11.7) describes a phase transition of second order, where the

order parameter Φ grows steadily from 0 as T drops below Tc , see Fig.
11.9,
 r 1/2
0
Φ= (Tc − T )β . (11.8)
4u
The critical exponent assumes the value β = 1/2; this value 1/2 for β is called
the mean-field value. If one considers the fact that in the flat “potential”
f ∼ uΦ4 close to Tc the fluctuations of the order parameter become large, one
finds a different value for β which is due to interactions between fluctuations;
a typical value is β ≈ 1/3 in 3D.
Using (11.5), one obtains the relation between the field h and the order
parameter Φ,

h = rΦ + 4uΦ3 .

This provides us with two results, the critical isotherm, see Fig. 11.9,
 h 1/3
Φ= ∼ h1/δ , T = Tc , (11.9)
4u
186 CHAPTER 11. LANDAU THEORY

with the critical exponent δ = 3, and the susceptibility2

(
∂Φ 1/r, T > Tc ,
χ = =
∂h h=0 1/2|r|, T < Tc .

From (11.9), we obtain the critical exponent γ = 1 for the susceptibility,

see Fig. 11.9,
1
χ∼ . (11.10)
|T − Tc |γ
with γ = 1. Again, critical fluctuations around Tc modify this result with γ
typically assuming a value γ ≈ 4/3 in 3D.
Finally, the free energy f assumes the values
(
0, T > Tc ,
f= 2
(11.11)
−r /16u, T < Tc ,

above and below the transition and taking the second derivative in T we
obtain the change in the specific heat associated with the transition,
(
∂2f 0, T > Tc ,
cV = −T = (11.12)
∂T 2 2
T r0 /8u, T < Tc .

We thus find that in the Landau mean-field approximation the specific heat
jumps. Denoting the exponent of the specific heat by α, see Fig. 11.9,
1
cV ∼ (11.13)
|T − Tc |α
this jump is consistent with a value α = 0. Note that the specific heat jump
described by (11.12) describes only the singular contribution associated with
the transition; other degrees of freedom contribute to cV , entropy s, and f
with a regular part.
Examples of second-order transitions are the magnetic transitions of the
Ising, XY, and Heisenberg models. As discussed in detail in Chap. 9, the
mean-field theory of the van der Waals gas exhibits a second-order transition
at the critical point pc , vc , Tc as well, although no symmetry is broken. The
role of the order parameter then is assumed by the volume (or density)
difference Φ ↔ vgas − vli .
Another classic example for a second-order phase transition is the normal-
metal–superconductor (or fermionic superfluid) transition. The microscopic
origin of the superconducting transition is found in a pairing of (fermionic)
electrons into (bosonic) Cooper pairs which then can condense in k-space.
2
The derivative ∂h of (11.9) gives 1 = (r + 12uΦ2 )∂h Φ and thus χ = 1/r at Φ = 0,
χ = 1/(r − 12ur/4u) at Φ2 = −r/4u.
11.3. SECOND-ORDER PHASE TRANSITIONS 187

Fermionic superconductor Bosonic superfluid

sharp jump
cV overlapping cV
Cooper pairs [J/ kg K]
10 ~ln 1/ |T −Tλ |
exp/gapped

linear/metallic

0 Tc T 0 1 2 3 4K T

Fig. 11.2: Left, sharp specific heat jump typical for a mean-field type transi-
tion as it appears in the fermionic normal-metal–superconductor transition.
The mean-field character is a consequence of the overlapping Cooper pairs
(→ many ‘neighbors’ defining a sharp ‘Weiss’ field). This is different for
a bosonic superfluid transition, right, where the coherence length is of the
order of the particle distance.

Its mean-field description is given by the BCS (Bardeen-Cooper-Schrieffer)

theory, a weak coupling theory with a small parameter 1/kF ξ0 , where kF ∼
1/n−1/3 denotes the Fermi wave vector and ξ0 is the coherence length or size
of the Cooper pairs. Typical extensions of Cooper pairs are of the order of
100 Åand the large number of particles in the volume ξ03 , nξ03  1, implies
that many Cooperpairs overlap, see Fig. 11.2. Given the large number of
‘neighbors’ of a pair, the superconducting transition then is well described by
mean-field theory. In particular, the specific heat cV exhibits a sharp jump
at Tc , as predicted by the Landau mean-field theory, see Fig. 11.2. Fur-
thermore, the order parameter (∝ energy gap, experimentally measured in
tunneling experiments) shows a mean-field type behaviour at Tc , ∆ ∼ τ 1/2 .
All this is very different for a bosonic superfluid where Bosons condense with-
out pairing. The number nξ03 ∼ na3 ∼ 1 of ‘neighbors’ then is small (a the
distance between bosons), resulting in a broad, so-called λ transition which
is only very approximately described by mean-field theory. The Ginzburg
criterion discussed in Sec. 12.4 below will give us a more accurate argument
how wide the critical or non–mean-field region of a phase transition is.

11.3.1 Non-local susceptibility and correlation length

Different from the Weiss mean-field theory in Chap. 8, the Landau free
energy functional involves a gradient term (∇Φ)2 and we can define further
quantities (and their critical exponents) associated with the spacial behavior
of Φ. One such quantity is the non-local susceptibility χ(~q ) defined in
188 CHAPTER 11. LANDAU THEORY

(8.23) (see also (8.20))

δ2F
χ−1 (~r, ~r 0 ) = = kB T G−1 (~r, ~r 0 ). (11.14)
δΦ(~r )δΦ(~r 0 )
Applying the usual rules of functional differentiation, we obtain that
χ−1 (~r, ~r 0 ) = (r + 12uΦ2 − c∇2 )δ(~r − ~r 0 ), (11.15)
and an additional Fourier transformation provides the result3
1
χ(~q ) = . (11.16)
r + 12uΦ2 + cq 2
Rewriting the denominator in the dimensionless form 1 + q 2 ξ 2 ,
1 ξ2
χ(q) = (11.17)
c 1 + q2ξ2
we can evaluate the correlation length above or below Tc with the help
of (11.16)
(p
 c 1/2 c/r, T > Tc ,
ξ(T ) = 2
= p
r + 12uΦ −c/2r, T < Tc ,
1
∼ . (11.18)
|T − Tc |ν
The exponent of the correlation length ξ is ν = 1/2 within a mean-field
description; fluctuations typically raise the value to ν ≈ 2/3 in a 3D critical
system. Thep‘microscopic’ correlation length ξ0 follows from the T = 0
value, ξ0 ≈ c/r0 Tc , or from a dimensional analysis, Epot = r0 Tc Φ2 ∼
~ 2 = Eelast , c|∇Φ|
c|∇Φ| ~ 2 ∼ cΦ2 /ξ 2 → ξ 2 ≈ c/r0 Tc .
0
The form (11.17) describing the non-local susceptibility was first intro-
duced by Ornstein and Zernicke in their analysis of light scattering at the
critical point of the gas–liquid transition: the phenomenon of critical opales-
cence is due to the strong scattering of light on bubbles (with different di-
electric constant) of size ξ which become large on approaching Tc ; when the
bubbles match the wave length of the light they produce strong scattering.
Transforming χ(~q ) back from (11.17) to real space we obtain4
Z ∞ d−1
ξ 2 dd q ei~q·~r x=qr 1 eix cos ϑ
Z Z
dΩd x
χ(r) = = dx
c (2π)d 1 + ξ 2 q 2 c (2π)d 0 rd−2 r2 /ξ 2 + x2
1
= d−2 Y (r/ξ) (11.19)
cr (
Z ∞
dΩd eix cos ϑ e−α /4π,
Z
d−1 d = 3,
Y (α) = dx x d 2 2
=
0 (2π) α + x K0 (α)/2π, d = 2.
3
Use Eq. (8.26) to show that χ−1
ij (~
q )χjk (~
q ) = δik , see also (11.21).
√
4
K0 (α) ∼ e−α / α.
11.3. SECOND-ORDER PHASE TRANSITIONS 189

For T 6= Tc , the non-local susceptibility χ(r) or Green function G(r) = T χ(r)

decays exponentially on a scale ξ. At Tc , ξ → ∞ and no scale exists, such
that χ(r) ∼ 1/rd−2 decays algebraically, see Fig. 11.3. The algebraically
decaying correlator in the Berezinskii phase corresponds to a system at the
critical point, i.e., the Berezinskii phase is critical for all T < TBKT . In a
more advanced discussion, see Chap. 13, the exponent d − 2 governing the
decay of χ(r) at Tc picks up an anomalous dimension η,
χ(r) ∝ 1/rd−2+η . (11.20)
E.g., in the Kosterlitz-Thouless transition η = 1/4. Within a mean-field
analysis, the anomalous dimension vanishes, ηMF = 0.

χ Fig. 11.3: The non-local sus-

ceptibility χ(r) decays exponen-
T = Tc tially ∝ e−r/ξ(T ) away from Tc
and algebraically ∝ 1/rd−2+η at
T = Tc Tc . The anomalous exponent η
vanishes in a mean-field descrip-
tion, ηMF = 0.
0 ξ r

The length ξ describes the distance over which a perturbation in the

order parameter will recover, or over which a perturbation stretches (←
correlation length). In particular, ξ gives the extent of a vortex core in
superconductors and superfluids, where one often uses the term coherence
length for ξ instead. The typical size of a domain in the critical phase
near the transition is also described by ξ, see Fig. 11.4. As T approaches
Tc from above, the disordered B phase is suppressed by (critical) fluctua-
tions (ξ % ∞), to allow for bubbles of the new ordered A phase and vice
versa. E.g., the coherence length ξ gives us the typical size of the super-
fluid5 /superconducting fluctuation droplets at T > Tλ , TBCS .

Longitudinal and transverse response

For a (n) vector-valued order parameter Φi , one can define two suceptibilities
χk and χ⊥ , that are directed longitudinally and transverse to the direction
of order, respectively. The relevant susceptibility tensor has been defined
in (8.23) and proceeding as in (11.15) with (11.3), we obtain (see also Eq.
(11.16))
χ−1 q ) = T G−1
ij (~ q ) = (r + 4uΦ2 + cq 2 )δij + 8uΦi Φj .
ij (~ (11.21)
5
The superfluid transition in 4 He is called a λ-transition because of its broad smearing
(in the shape of a λ).
190 CHAPTER 11. LANDAU THEORY

A ξ B

ξ
ξ
K
K

Tc T

Fig. 11.4: Domains of alternative phases near the transition are of typical
size ξ. E.g., a typical critical (K) bubble/droplet in the disordered (B) /
ordered (A) phases above and below Tc measures ξ(T ). The critical domains
K represent the scale invariant phase at the transition. On approaching Tc ,
the correlation length ξ diverges from either side.

Assuming the symmetry to be broken along φ,~ we define the longitudinal

and transverse projectors
k φi φj φi φj
Pij = 2 , Pij⊥ = δij − 2 , (11.22)
φ φ
we can rewrite χ−1
ij as
k
χ−1 q ) = χ−1
ij (~ q )Pij + χ−1
k (~ q )Pij⊥
⊥ (~

where6
1 1
χk (~q ) = , χ⊥ (~q ) = . (11.23)
r + 12uΦ2 + cq 2 r + 4uΦ2 + cq 2
While χk is identical to the scalar susceptibility χ in (11.16), we find a
different expression for the transversal susceptibility
(
1/(r + cq 2 ), T > Tc ,
χ⊥ (~q ) = (11.24)
1/cq 2 , T < Tc .
Thus, the transverse part of the order parameter correlator becomes mass-
less below Tc (above Tc , the parameter r > 0 plays the role of a ‘mass’;
alternatively, it generates a finite ‘length scale’ for the correlations),
T
G⊥ (~q ) = hδΦ⊥ (~q )δΦ⊥ (−~q )i = ,
cq 2
1
G⊥ (~r ) ∼ . (11.25)
rd−2
6
We write χ−1 2 2 2 2
ij = αδij + βφi φj /φ with α = r + 4uΦ + cq and β = 8uΦ . Using
k
δij δij = n, δij φi φj /φ2 = 1, (φi φj /φ2 )(φi φj /φ2 ) = 1 and χ−1
k = Pij χ−1 −1
ij , (n − 1)χ⊥ =
⊥ −1
Pij χij , 11.23 follows immediately.
11.4. FIRST-ORDER PHASE TRANSITIONS 191

mexican hat
f Fig. 11.5: Mexican hat illus-
trating longitudinal modes that
change the order parameter am-
plitude and cost quadratic (in
0 δΦ) energy and transverse modes
φ that cost ‘no’ energy in the limit
k → 0.
φ φ

The softness of transverse fluctuations destroys the long-range order in lower

dimensions d = 1 and d = 2 (Hohenberg-Mermin-Wagner). While the longi-
tudinal susceptibility χk describes the changes in the amplitude of the order
parameter, the transverse part χ⊥ describes the changes in the ‘phase’ or
~ see Fig. 11.5. The separation into
direction of Φ,
OP
longitudinal modes → ↔, δΦk
OP
transverse modes → l, δΦ⊥ (11.26)

only exists when a continuous symmetry has been broken.

11.4 First-order phase transitions

The simplest generic form of the Landau free energy f (T, Φ) for a first order
phase transition is given by
r
f (T, Φ) = Φ2 − wΦ3 + uΦ4 , (11.27)
2
where

r = r0 (T − T ∗ ), w, u > 0, constant.

Note that T ∗ 6= Tc ; the meaning of T ∗ will be made clear shortly. The

behaviour of f for various temperatures is shown in figure 11.6, where the
phase transition at Tc is shown at the left.
At Tc , f is minimised by an additional solution Φ = Φmin > 0 satisfying
the conditions ∂Φ f = 0 with f = 0, hence

(r − 3wΦ + 4uΦ2 )Φ = 0,
(r/2 − wΦ + uΦ2 )Φ2 = 0,

from which we find the first-order jump

w
Φmin =
2u
192 CHAPTER 11. LANDAU THEORY

T =Tc
f f
T >Tc
T = T **
T < Tc

0 st Φ 0 Φ
1
T =T *

Fig. 11.6: Free energy functional for a first-order transition including a

cubic term. Left: the transition appears when the two minima in f turn
degenerate, with the order parameter jumping to a finite (small, to be con-
sistent with the expansion) value. Right: critical temperatures T ∗ and T ∗∗
defining the spinodals where metastable minima vanish.

of the order parameter at

w2
rc = r0 (Tc − T ∗ ) = > 0,
2u
see Fig. 11.9. Decreasing the temperature further, the metastable solution at
Φ = 0 becomes unstable at T ∗ < Tc , where the curvature in 0 turns negative
when r∗ = 0, see Fig. 11.6 right. The ordered phase Φ > 0 becomes unstable
upon increasing T beyond T ∗∗ > Tc , where ∂Φ f = ∂Φ 2 f = 0, and we find

that
9w2
r∗ = 0, r∗∗ = . (11.28)
16u
Note the different separations r∗∗ − rc = w2 /16u and rc − r∗ = 8w2 /16u.
The caloric properties of the transition can be calculated from the free
energy (
0, T > Tc ,
f= 2
(11.29)
(r − rc )Φmin /2, T < Tc ,
from which we obtain the entropy jump ∆S = −∂f /∂T ,
r0 w2
SΦmin − S0 = −
8u2
and the latent heat ` = −Tc ∆S,
r0 w2 Tc
`= . (11.30)
8u2
The jump in S vanishes with the cubic term w → 0. Note that on top of
the entropy associated with the transition, there is a positive regular part
11.4. FIRST-ORDER PHASE TRANSITIONS 193

describing the remaining degrees of freedom in the system, S = SΦ + Sreg .

The jump in entropy ∆S then produces a δ−function in the specific heat
cV , see Fig. 11.9.
As usual, we define the conjugate field h via h = ∂Φ f , and the suscepti-
bility χ = ∂h Φ, that provides us with the result
∂Φ 1
χ= =
∂h r − 6wΦ + 12uΦ2
For T > Tc , we have Φ = 0, χ = 1/r = 1/r0 (T − T ∗ ), hence, different from
the second order transition, χ now diverges at T → T ∗ when coming from
the high temperature phase, however, this divergence is cut off at Tc , see
Fig. 11.7. Inversely, we can calculate χ in the ordered phase and find that
when extrapolating to the domain T > Tc the expression diverges at T ∗∗ .
Indeed, since χ−1 = ∂Φ 2 f , and using the result (11.28), we find that χ−1

vanishes at r∗∗ that corresponds to a divergence χ ∝ 1/(T ∗∗ − T ) at the

temperature T ∗∗ . The temperatures T ∗ and T ∗∗ define the locations where
the metastable phases become unstable, the spinodals that we encountered
in Sec. 1.5.5. The temperatures T ∗ and T ∗∗ set the boundaries within which
hysteretic effects such as overheating of a liquid or undercooling of a gas
show up. In the same way the increase/divergence of the correlation length

Fig. 11.7: The susceptibility ap-

χ parently diverges at the temper-
atures T ∗ and T ∗∗ framing Tc .
The divergencies are cutoff at Tc
where the branches change over
at a finite value of χ.

T* Tc T ** T
ξ(T ) appearing in χ(~q ) is cut off at Tc . The first-order phase transition thus
shows no scaling laws and we cannot define any critical exponents.
Note that our analysis shows a jump in the order parameter from 0 to
Φmin > 0. The expansion of f as a power series in Φ requires that Φ > Φmin
remains small; in this situation, one talks about a weak first-order transition.
Otherwise, higher-order corrections to f become important.
Typical examples for the appearance of a first-order transition are the
nematic to isotropic transition in liquid-crystals (Maier-Saupe model), see
Fig. 11.8 and the liquid to crystal transition. The order parameter of the
I-N transition is not a vector but a director written in a matrix form.7 The
7
For a uniaxial nematic, Qij = S(ni nj −δij /3) with ~n a unit vector, the Frank director.
194 CHAPTER 11. LANDAU THEORY

molecules
Fig. 11.8: The transition be-
tween isotropic (left) and ne-
matic (right) phases of molecules
is first-order.
isotropic

nematic

order parameters in the liquid–crystal transition are the density waves ρG~
defining the density modulation
X ~
ρ = ρ0 + ρG~ eiG~r .
~
G

~ i add up to zero,
The cubic term in f is allowed if the three wavevectors G
~1 + G
G ~2 + G~ 3 = 0, and is given by
X
ρG~ 1 ρG~ 2 ρG~ 3 .
~ 1 ,G
G ~ 2 ,G
~3

A first order transition can also appear in a model without cubic terms, for
example f = (r/2)Φ2 + uΦ4 + u0 Φ6 with u < 0. This model can also describe
a second-order phase transition, namely for r < 0, u > 0. In a system with
many order-destroying variables, for example T and a mixing concentration
x, one can manipulate r and u such that both transitions are present in
the phase diagram and meet at the tricritical point. The appearance of
such multicritical points is the focus of the next section. Before, we briefly
summarize and compare the main results of our discussion of second- and
first-order transitions.

Summary, Landau mean-field theory, second- and first-order

In figure 11.9, we compare the results of the mean-field analysis of second-

and first-order transitions. The diagrams for the order parameter Φ(T, h) are
different in the scaling behavior versus first-order jump. The specific heats
cV (T ) show a jump versus the δ-function due to the latent heat (or entropy
jump). The susceptibilities χ(T ) diverge for the second-order transition,
while the first-order transition is characterized by a kink plus hysteresis.
The correlation lengths ξ(T ) show a similar behavior as the susceptibilities.
11.4. FIRST-ORDER PHASE TRANSITIONS 195

χ
Φ
cV jump
scaling

h
0
scaling Tc Tc T Tc T
T

Φ cV χ

h δ− function
0
jump Tc Tc T T* Tc T ** T
T
hysteresis

Fig. 11.9: Phase diagrams (left), specific heat (middle), and susceptibility
(right) of second-(top) and first-order (bottom) mean-field transitions. The
scaling behavior and divergencies in the second-order transition are replaced
by jumps and a knee in the first-order transition. The correlation length ξ(T )
exhibits the same behavior as the susceptibility χ(T ).

11.4.1 Multicritical points

We start with the functional

1
f (T, ∆, Φ) = r(T, ∆)Φ2 + u(T, ∆)Φ4 + u0 Φ6 . (11.31)
2

The parameter u0 > 0 guarantees the stability of f . For u > 0, the line
r(T, ∆) = 0 gives a line of second-order phase transitions. For u < 0, there
are first-order transitions when f = ∂Φ f = 0 at Φ 6= 0,

0 = (r/2 + uΦ2 + u0 Φ4 )Φ2 ,

0 = (r + 4uΦ2 + 6u0 Φ4 )Φ,
u2 u
→ rc = , Φ2min = − 0 , u < 0. (11.32)
2u0 2u

The resulting u-r phase-diagram with the tricritical point at the origin is
shown in Fig. 11.10.
The transition lines of both the first- and second-order transitions in
the u-r plane are easily transformed to the physical variables T and ∆, see
Fig. 11.11. Again, the two transition lines join smoothly with their slopes
196 CHAPTER 11. LANDAU THEORY

u f
f

nd Φ
Φ 2

tricritical point

r
f f
st
1
Φ
Φ

Fig. 11.10: u-r phase diagram with a second-order transition line at r = 0

for positive u and a first-order transition line at rc = u2 /2u0 for a negative
u. The two transition lines join smoothly at the origin.

properly aligning,8

∂∆ ∂r/∂T rT
= − ≡− ,
∂T r=0 ∂r/∂∆ r∆
∂∆ 0
rT − uuT /u TC: r=0=u rT ∂∆
= − 0
= − = . (11.33)
∂T r=u2 /2u0 r∆ − uu∆ /u r∆ ∂T r=0

Note that the conjugate (to T and ∆) variables S and x jump on the first-

T Fig. 11.11: First- and second-

1st order transition lines joining
smoothly in a tricritical point.
TC
nd
2

∆
order line. As usual, these jumps ∆S and ∆x are related through Clapeyron
equations. For example one finds the jump in s to be

∂f rT |u|
∆s = − ≈− → 0 at TC. (11.34)
∂T 4 u0
A new feature is the coexistence of three phases with Φ = 0, Φ = ±Φc on
the first-order phase boundary.
8
Along r(T, ∆) = 0, we have 0 = dr = rT dT + r∆ d∆.
11.4. FIRST-ORDER PHASE TRANSITIONS 197

Critical behavior at the tricritical point

The critical behaviour at the tricritical point is different from the normal
second-order critical behaviour. We approach the tricritical point TC along
the line u = 0 with r = r0 TTC τ , τ = 1 − T /TTC . Then, the order parameter
Φ increases as
 −r 1/4
Φ=± ∼ τ β with β = 1/4. (11.35)
6u0
With h the field conjugate to Φ, we find the equation of state from h = ∂Φ f ,

h = rΦ + 4uΦ3 + 6u0 Φ5 (11.36)

and the critical isotherm at r = 0, u = 0, i.e., at the tricritical point TC,

 h 1/δ
Φ= , with δ = 5.
6u0
For the susceptibility χ we find
(
∂Φ 1 1/r, T > TTC ,
χ = = =
∂h r + 30u0 Φ4 −1/4r, T < TTC ,
∼ |τ |−γ , with γ = 1. (11.37)

Finally, to find the critical exponent α of the specific heat, we require the
free energy f ,
(
0, T > TTC ,
f= p (11.38)
(r/2) −r/6u , T < TTC → f ∼ |τ |3/2 ,
0

from which we obtain the specific heat divergence

∂2f
cV = −T ∼ |τ |−α with α = 1/2. (11.39)
∂T 2
Note the difference in the critical exponent for the second-order critical and
tricritical transitions

α β γ δ ν η (11.40)
C : 0 1/2 1 3 1/2 0
TC : 1/2 1/4 1 5 1/2 0

By comparing the terms uΦ4 and u0 Φ6 to (r/2)Φ2 , we can identify the regions
where critical and tricritical fluctuations dominate. Within these regions,
the free energies can be approximated by
r
C : f ≈ Φ2 + uΦ4 , u0 Φ6 small,
2
r
TC : f ≈ Φ2 + u0 Φ6 , uΦ4 small. (11.41)
2
198 CHAPTER 11. LANDAU THEORY

u Fig. 11.12: For u > 0, we

tricritical can define the separate regimes
fluctuations where critical and tricritical fluc-
critical
fluctuations tuations dominate.

The crossover between the two regions appears at u0 Φ2 ∼ u, and using Φ2

from balancing (r/2)Φ2 ∼ uΦ4 → Φ2 ∼ r/2u, we obtain the scaling for the
crossover line r ∼ u2 /u0 , see Fig. 11.12.
The most well-known example of a tricritical point is in the mixed 3 He–
4 He system with the tuning parameters temperature T and 3 He-fraction
x, with x the dual variable to ∆ (the chemical potential difference). 4 He
becomes superfluid at Tλ . Mixing in 3 He atoms reduces Tλ . At the tricrit-
ical point Tλ terminates and the mixture phase-separates into an 4 He rich
superfluid and a 3 He rich normal fluid.

Fig. 11.13: Phase separa-

tion in the 3 He–4 He mixture.
T tricritical
point The second-order λ-transition
1.0
Tλ line above TC (mixture with less
0.8 then 67 % 3 He) separates a su-
super− perfluid from normal phase. Be-
0.6 fluid normal
low TC the mixture phase sepa-
0.4 phase rates in a first-order fashion and
separation superfluidity only appears in the
0.2 4 He-rich phase, while the 3 He-

rich phase at large x remains nor-

0 10 3
50% He 100 x mal.

The name “tricritical” becomes clear if we consider a magnetic substitute

model for the 3 He–4 He with a phase diagram shown in Fig. 11.14—the latter
involves the conjugate (to x) variable ∆ = µ3 −µ4 , i.e., the chemical potential
difference between 3 He–4 He and a magnetic field h that is unphysical in the
He-mixture but plays a role in the magnetic substitute. A large negative
∆ produces pure 4 He. The so-called Blume-Emery-Griffiths model involves
spins on a lattice that assume values Si = 0, ±1; sites occupied with a 3 He
atom assume a value Si = 0, while Si = ±1 describe sites with 4 He (all
sites host exactly one atom); hence, hS 2 i ↔ 1 − x and the fraction x of
11.4. FIRST-ORDER PHASE TRANSITIONS 199

3 He acts as an order parameter for phase separation. The ± signs for the
4 He sites allow to define the (magnetic) average hSi i as an order parameter
that describes superfluidity. The spin Hamiltonian driving these magnetic
and phase-separation order parameters looks pretty complicated, see Blume,
Emery, Griffiths, PRA 4 1071 (1971). One finds the T −∆−h-diagram shown
in Fig. 11.14 (Griffiths, PRL 24, 715 (1970)) with a three-phase coexistence
(two magnetic phases that can get split by h and a paramagnetic phase)
on the first-order critical line D and three second-order lines meeting in a
point; the merging of these three second-order lines gives the name to the
tricritical point.

T
λ−line

2 nd

A TC
B’
B
∆
D
h
Fig. 11.14: T -h-∆ phase diagram for the 3 He–4 He mixture. The region (or
surface) A below the λ-line (a line of critical points corresponding to the
λ-line in Fig. 11.13) describes the superfluid phase in the He mixture and a
first-order or coexistence surface with two degenerate magnetic phases (with
OP ±hSi selected by ±|h| → ±0) in the spin lattice model. The surface A
is connected via D (corresponding to the split lines terminating the phase
separated region in Fig. 11.13) to two other first-order (coexistence) surfaces
B and B 0 extending into the finite h regions. In the magnetic analogon, the
two phases coexisting in the surface B involve different values of hS 2 i (and
similar for B 0 ). The three critical lines terminating A (the λ-line), B, and
B 0 meet in the tricritical point TC. Along the line D, three phases coexist.

Multicritical- and Lifshitz points

We close with a brief overview of multicritical points, see Fig. 11.15. The
lesson to learn here is that different phase transition lines connect with
different constraints. E.g., a second-order line can branch off a first-order
line at a finite angle, while it extends smoothly across a tricritical point.
200 CHAPTER 11. LANDAU THEORY

Similarly, two second-order lines emanating from a bicritical point continue

smoothly out of a first-order line. And finally, two second-order lines touch
smoothly in a tetracritical point.

h h λ−line h
tetracritical
critical CP point
point
critical
endpoint
T T T

λ tricritical λ
T T
point TC
x
konj.
to ∆ phase
separation
∆ x
bicritical
h m
point
m phase
separation BP
konj.
to h

T T

Fig. 11.15: Geometry of phase diagrams with a critical, tricritical, bicritical,

and tetracritical point. Note how first-order lines open up when going over
to the conjugate extensive variable.
11.4. FIRST-ORDER PHASE TRANSITIONS 201

Finally, Lifshitz points appear in phase diagrams involving a spatially

modulated phase, see Fig. 11.16. The latter appears when the free energy
involves higher-order derivatives, thus defining a finite k-vecor in the prob-
lem. Specifically, for a free energy of the form
r ck
f = Φ2 + (∇~ k Φ)2 + c⊥ (∇
~ ⊥ Φ)2 + D (∇
~ 2 Φ)2 + uΦ4 ,
2 2 2 2
describing an anisotropic (k, ⊥) system, a modulation of Φ with wavelength
k⊥ = |c⊥ |/2D shows up in the ~r⊥ -direction when the transverse elasticity
c⊥ < 0 turns negative. If the order paramater is a vector Φ, ~ the transi-
tion from the ordered to the modulated phase is second-order. An example
is given by the anisotropic-next-nearest-neighbor (ANNNI) Ising model,9
where the anisotropy defines the ⊥ direction.

r disordered phase r disordered phase

Lifshitz Lifshitz
point point
modulated
0 0
phase
ordered modulated
ordered
phase phase
phase

−c −c
0 0

Fig. 11.16: Phase diagram with a Lifshitz point involving a spatially mod-
~ is vector-valued, the transition
ulated phase (left). If the order parameter Φ
to the modulated phase is second-order (right).

9
with ferromagnetic in-plane coupling, ferromagnetic out-of-plane coupling, and anti-
ferromagnetic next-nearest-neighbor coupling.
202 CHAPTER 11. LANDAU THEORY
Chapter 12

Field Theory: Gauss and

Ginzburg

If we want to go beyond the mean-field approximation, we must account

for the fluctuations in our analysis. We R start from the effective description
through a free energy functional F = dd r f [Φ(~r )] over an order parameter
field Φ(~r ), see the discussion in Sec. 11.1. The functional F over the classi-
cal order parameter field Φ can, in some particular cases, be derived directly
from an underlying classical or quantum mechanical microscopic Hamilto-
nian. It is also common to obtain F phenomenologically, in this case one
must take into account the constraints discussed in Sec. 11.2, e.g., the rel-
evant symmetries of the problem. The underlying idea is that F describes
the correct long-wave-length and low-energy physics that is relevant for the
analysis of the phase transition. The coefficients in F become T dependent
through the partial trace in (11.2) which shows itself in the T dependence
of r(T ) = r0 (T − Tc ). One often writes H instead of F and then calls H
the (effective low-energy) ‘Hamiltonian’ of the problem. Note that such a
‘Hamiltonian’ H provides the energy of a configuration but does not include
its dynamics (in this sense F is a more fitting description). 1
In the following, we focus our discussion on the (classical) critical prob-
lem in the vicinity of a phase transition as described by a generic free-energy
1
The fluctuations in the system we are considering here are driven by T ; including
quantum fluctuations (driven by ~) one has to go over to a Euclidean action SE /~. One
then has to include a dynamical term, typically involving a non-local kernel α(τ ) (with τ
denoting the imaginary time)
Z
T [Φ(~r, τ )] = dτ dτ 0 dd r Φ(~r, τ )α(τ − τ 0 )Φ(~r, τ 0 ), (12.1)
R
SE [Φ(~r, τ )] = T + dτ F [Φ(~r, τ )]. A description of various order-parameter dynamics can
be found in the review article of Hohenberg and Halperin, Theory of dynamic critical
phenomena, Rev. Mod. Phys. 49, 435 (1977).

203
204 CHAPTER 12. FIELD THEORY: GAUSS AND GINZBURG

functional or ‘Hamiltonian’ of the type

Z nr c ~ o
H= dd r Φ2 (~r ) + |∇Φ(~
r )|2 + uΦ4 (~r ) , (12.2)
2 2

the generalisation of (11.6) that includes spatial deformations of the order-

parameter field Φ(~r ), see also (11.1). The Hamiltonian (12.2) allows us to
study the critical phenomena of a generic second-order transition of a scalar
order parameter (the Φ4 theory).2
Setting the coupling u = 0 in (12.2), we obtain the Gauss Modell, that we
can solve exactly.3 In doing so, we go beyond Landau mean-field theory and
include (Gaussian) fluctuations to lowest order, see Fig. 12.1. The full Φ4

H Fig. 12.1: Free energy land-

scape for the order-parameter
field Φ(~r ). The Gauss model
fluctuations
includes small amplitude fluc-
tuations of the order parame-
min ter around the saddle point (or
mean-field) solution.

LMFT Φmin
Φ
− dim Φ −space
8

theory is not exactly solvable. The uΦ4 term introduces interactions between
the fluctuations (mode-mode-coupling). We will make a first attempt to
treat effects of the Φ4 term in Sec. 12.3 on the level of the Hartree- or self-
consistent field approximation; the full treatment via the renormalization
group will be discussed later in Sec. 16.2. Next, we discuss how to deal with
the functional integral over the order parameter field in Eq. (11.2), both on
a lattice in real space and in ~k-space.

2
Note the similarity with the problem of an elastic manifold in a potential V (u) as
described by the configurational integral
Z
dd r {C[∇u(~
~ r )]2 /2+V [u(~
R
Z= D[u(~r )] e−β r )]}
(12.3)

and touched upon in Sec. 5.1.5 (Eq. (12.3) describes the classical limit of (5.15) after a
trivial integration over π). Hence, the order-parameter field Φ(~r ) can be understood as
the deformation field u(~r ) of an elastic manifold trapped in a square/quartic potential
V (u).
3
We will generalize this later by first expanding (12.2) near a homogeneous minimum
and only then setting u to zero.
12.1. LATTICE FIELD THEORY AND CONTINUUM LIMIT 205

12.1 Lattice field theory and continuum limit

We consider a d-dimensional lattice with N sites ν and the volume v per

lattice site, V = N v. The Hamiltonian and partition function4 are then
defined by
Xr X Cνν 0
H[Φν ] = Φ2ν + uΦ4ν + (Φν − Φν 0 )2 ,
2 2
ν ν,ν 0
YZ ∞
ZN = dΦν e−βH[Φν ] . (12.4)
ν −∞

Consider a 1D model in a potential as sketched in Fig. 12.2. We easily

Φmin
x f
Φ Φ

f
x
Φ Φ min Φ

Fig. 12.2: Minimal configuration (solid line) and fluctuations (dashed line)
for the potential shown to the right. Below Tc , the symmetry has been
broken by selecting one minimum.

identify the minimal configuration Φmin and the energetically less favourable
fluctuations δΦ. The continuum limit of (12.4) is obtained from the limit
v → 0 with V = N v = const and the substitutions

ν → ~r,
Z d
X d r
f (Φν ) → f [Φ(~r )],
ν
v
Z d
1X 2 1 d r ~ 2
Cνν (Φν − Φν ) →
0 0 c [∇Φ] (12.5)
2 0 2 v
ν,ν
1X
with c = Cν,0 Rν2 for an isotropic model.
d ν
4
We do not worry about the normalization of the functional integral over Φν , what
accounts for an undetermined constant in the entropy, which is irrelevant in the discussion
of critical phenomena.
206 CHAPTER 12. FIELD THEORY: GAUSS AND GINZBURG

Similarly, one understands the path integral in (12.4) as the limit

Z YZ ∞
D[Φ(~r )] = lim dΦν . (12.6)
v→0 −∞
ν

Hence the discrete model (12.4) transforms to the continuum model5

Z d h
d r r 2 c ~ i
H[Φν ] = Φ (~r ) + uΦ4 (~r ) + |∇Φ(~
r )|2 ,
v 2 2
Z
ZN = D[Φ(~r )] e−βH[Φ(~r )] . (12.7)

Dropping the elastic and quartic terms in the Hamiltonian of Eq. (12.7), we
easily find the partition function
Z N Z ∞
r N
−βH[Φ(~
r )]
Y 2πkB T −(r/2kB T )Φ2ν
ZN = D[Φ(~r )]e = dΦν e = .
ν −∞ r
(12.8)
Its associated free energy scales linearly with N . In order to include the
elastic term as well, we analyze the model in Fourier space.6 The inclusion
5
The expression (12.7) strictly follows from the discrete version (12.4); later in the text,
we will sometimes use v = 1, implying that the coupling constants scale differently with
length. Such different notation is more convenient, e.g., in the discussion of scaling. Also
note that N in the Fourier transforms below goes over to V when we choose v = 1.
6
Alternatively, the discrete quadratic problem (12.4) without quartic term can be solved
directly using the following statements about Gauss integrals: We start from the simplest
expression r
Z ∞
2 2π λ2 /2C
dx e−Cx /2+λx = e (12.9)
−∞ C
that we reduce to a simple Gauss integral through quadratic completion. Its n × n matrix
generalisation reads
n
Z Y r
h 1 i (2π)n h1
−1
i
dxi exp − xi Cij xj + λi xi = exp λi Cij λj , (12.10)
i=1
2 det C 2

where the matrix Cij = hi|C|ji is taken to be real and symmetric. In this case, C can be
diagonalised, hp|C|p0 i = δpp0 Cp , with Cp real and the transformation matrix Mpi = hp|ii
can be chosen to be real and symmetric (with hi|pi the orthonormal eigenfunctions of C).
Using
X X
xi = hi|pixp , λj = hj|piλp ,
p p

| det M | = | det M 2
| 1/2
= | det 1|1/2 = 1,

we find
Z Yn n Z n
s
−xi Cij xj /2+λi xi
Y −Cp x2
p /2+λp xp
Y 2π λ2p /2Cp
dxi e = dxp e = e ,
i=1 p=1 p=1
Cp

in agreement with the result (12.10).

12.1. LATTICE FIELD THEORY AND CONTINUUM LIMIT 207

of the quartic term cannot be done exactly and requires pertubative/RG

techniques.

12.1.1 ~k-space

We consider a volume V = Ld with periodic boundary conditions. The

transformation from Φ(~r ) to ~k-space takes us back to discrete sums,7

dd r −i~k·~r
Z Z d
1 d r i(~k−~k 0 )·~r
Φ~k = e Φ(~r ), e = δ~k,~k 0 ,
v N v
1 X i~k·~r 1 X i~k·(~r−~r 0 )
Φ(~r ) = e Φ~k , e = δ(~r − ~r 0 ).
N V
~k ~k

The limit V → ∞ is then given by

dd k
Z
1 X
→ . (12.11)
V (2π)d
~k

The Fourier transformed expression for our Hamiltonian (12.2) reads

1 X u X
r + c k 2 Φ~k Φ−~k + 3


H= Φ~k1 Φ~k2 Φ~k3 Φ−~k1 −~k2 −~k3 . (12.12)
2N N
~k ~k1 ,~k2 ,~k3

As Φ(~r ) ∈ R is real, we have Φ−~k = Φ~∗ . Writing Φ±~k = <Φ~k ± i=Φ~i , we

k k
can obtain all degrees of freedom in the model by taking half of all ~k-values
(e.g., those with kz > 0) and attribute to every such ~k two real fields <Φ~ k
and =Φ~k . The first (Gaussian) term in (12.12) can then be written in the
form (we combine ~k and −~k into one term)
1 X0
r + c k 2 |Φ~k |2


HG = (12.13)
N
~k

where ~0k refers to the constraint of taking only half of all ~k values. Note
P
that with this counting, N degrees of freedom in real space go over into N
degrees of freedom in ~k-space, namely N/2 ~k-points and 2 real fields <Φ~k
and =Φ~k per ~k-point.
The path integral in ~k-space is written as
Z Z Y Z Y
0 0
D[Φ(~r )] → dΦ~k dΦ~∗k /N → (2/N ) d<Φ~k d=Φ~k , (12.14)
k<Λ k<Λ
7
Note that with this definition, Φ(~r ) and Φ~k have the same units. For v = 1, we have
a change of units and have to replace N by V .
208 CHAPTER 12. FIELD THEORY: GAUSS AND GINZBURG

where the factors N and 2 in the last two equations originate from the Ja-
cobians when going from real- to k-space and changing Φ~k , Φ~∗ to the real
k
fields <Φ~k and =Φ~k (note that the Jacobian for the Fourier transform is
unity only for a unitary, i.e., symmetric definition of the Fourier transform).
The (UV = ultra-violett) cutoff Λ in the product takes into account the un-
derlying lattice, and thus the coarse graining of the Hamiltonian, Λ ∼ 2π/a.
The integrals over real- and imaginary parts <Φ~k and =Φ~k are independent
but with constrained ~k vectors, hence either the product 0 is over half of
Q
~k-space or over all ~k vectors with the result be taken to the half power.
Please note that, here, we have been more careful than usual about a con-
sistent scaling between descriptions in real and k-space in order to provide
a good reference point; however, keep in mind that we have no control on
the absolute normalization in the (classical) functional integral, such that
keeping all factors may appear as an overkill as it goes beyond precision.

12.2 Gaussian model

The Gaussian model with u = 0 in (12.4) and (12.12) can be solved exactly.
The Hamiltonian of the Gaussian model is defined as (see (12.7))
Z d
1 d r ~ r )|2 ]
HG = [rΦ2 (~r ) + c|∇Φ(~ (12.15)
2 v
1 X0 1 X0 ∗ −1 ~
= [r + ck 2 ] |Φ~k |2 ≡ Φ~k G0 (k )Φ~k
N Nβ
~k ~k

with8
kB T
G0 (~k) = . (12.16)
r + ck 2

12.2.1 Free energy and specific heat

The Gauss Hamiltonian (12.15) describes the fluctuations δΦ = Φ on top

of the mean-field solution Φ = 0 at temperatures T > Tc , with each, the
saddle-point (mean-field) and fluctuations contributing to the free energy.
The latter is obtained from the partition function and we find (note that
with our definitions, r is measured in units of energy)
Λ r
−βFG
Y 2πkB T
ZG = e = , 0 < |~k | < Λ. (12.17)
r + ck 2
~k

8
The kernel G−1 0 in the Hamiltonian is the original definition of the Green function of a
free theory; it is then straightforward to show (via a Gauss integration) that the correlator
is the inverse of G−10 , hΦΦi ∼ G0 .
12.2. GAUSSIAN MODEL 209

The result then provides us with the free energy

kB T X 2πkB T
F = FMF − log (12.18)
2 r + ck 2
|~k |<Λ

with the Gaussian part

dd k
Z
FG kB T 2πkB T
= − log , (12.19)
V 2 (2π)d r + ck 2
Sd−1 Λ
Z
Sd−1
≈ kB T dk k d−1 log k ≈ kB T 2 [k d (d log k − 1)]Λ
ξ −1
(2π)d ξ−1 d (2π)d
Sd−1
≈ kB T [Λd log Λ + ξ −d log ξ]
d(2π)d
Sd−1 kB T Sd−1 kB T
≈ d
log Λ + log ξ = Fa + Fsing ,
d a d(2π)d ξ d

where Sd = 2π d /Γ(d/2) is the surface of a sphere in d dimensions, S4 = 2π 2 ,

S3 = 4π, S2 = 2π. On approaching Tc , it is the second, singular part Fsing
of the free energy that attracts our interest.
Indeed, the free energy (12.19) gives us access to the specific heat cV
and its exponent α that will be modified by Gaussian fluctuations. From
the definition of the specific heat
∂ 2 (F/V )
cV = −T
∂T 2
and using (12.19), we obtain

∂ 2 h kB T dd k
Z
2πkB T i
cV = T log
∂T 2 2 (2π)d r(T ) + ck 2
dd k (T ∂T r)2
Z
∼ kB + further less singular terms,
(2π)d (r + ck 2 )2
where the dominant singular contribution is generated by the temperature
dependence in r(T ) = r0 (T − p Tc ) = r0 Tc τ . We write the denominator
r0 Tc τ + ck 2 as c(ξ −2 + k 2 ), ξ = c/r0 (T − Tc ) and switch to dimensionless
variables x = kξ to obtain
Z Λξ
4−d xd−1
cV ∼ kB ξ dx . (12.20)
0 (1 + x2 )2
For τ → 0+ the correlation length ξ → ∞ diverges, while the integral in
(12.20) remains finite for 1 < d < 4: for x → 0 we have xd−1 and for x → ∞
we obtain the scaling xd−5 . Thus, we find that cV diverges at the critical
point,
τ → 0+
cV ∼ const ξ 4−d → ∞ in 1 < d < 4. (12.21)
210 CHAPTER 12. FIELD THEORY: GAUSS AND GINZBURG

The result (12.21) defines a new value for the critical exponent α,
cV ∼ τ −(4−d)/2 ∼ τ −α ,
we find that within the Gaussian approximation
α = 2 − d/2. (12.22)
Tracing back the origin of the divergence in (12.20)
Z Z
d 1 1
cV ∼ d k −2 2 2
∼ dd k 4 ,
(ξ + k ) ξ −1 k

we note that the k-integral diverges at small k as τ → 0, i.e., 1/ξ 2 → 0.

Hence, it is the long-wave length fluctuations that provide the diverging
contribution to cV and we deal with a infrared divergence. Going to d > 4,
the integral in (12.20) diverges as (Λξ)d−4 but the prefactor ξ 4−d goes to
zero, providing a regular result; in d = 4 we have a logarithm. Hence, cV
remains finite in d > 4 and fluctuations do not provide a correction to the
mean-field exponent α.

12.2.2 Susceptibility and correlator

Next, we turn our interest to the susceptibility χ0 = ∂h hΦi and the cor-
relator G0 = hδΦ δΦi. According to (8.20) (which is generally valid for
any Hamiltonian since it is a linear response result), they are related by
χ0 = G0 /kB T . The inverse susceptibility χ−1
0 is directly obtained from the
functional derivative9
v 2 δ 2 HG
χ−1 r, ~r 0 ) =
0 (~ = v(r − c∇2 )δ(~r − ~r 0 )
δΦ(~r )δΦ(~r 0 )
1
↔ χ0 (~q ) = . (12.23)
r + cq 2
Next, let us explicitly calculate the order parameter correlator hδΦ δΦi, see
(8.20); starting from
hδΦ(~r )δΦ(~r 0 )i = hΦ(~r )Φ(~r 0 )i,
we calculate the order-parameter correlator,
1 X i~k·~r i~k 0 ·~r 0
hΦ(~r )Φ(~r 0 )i = 2 e e hΦ~k Φ~k 0 i
N
~k,~k 0
Z Y
1 0
hΦ~k Φ~k 0 i = (dΦq~ dΦ∗q~ /N ) Φ~k Φ~k 0 e−βHG . (12.24)
ZG
q~
9
Note that our measure is dd r/v rather then dd r, that generates the additional factor
v per functional derivative.
12.2. GAUSSIAN MODEL 211

For ~k 6= ~k 0 , the equation (12.24) reduces to hΦ~k ihΦ~k 0 i and hence vanishes.
For ~k = ~k 0 , we have to calculate hΦ~2 i, which will vanish by angular inte-
k
gration of dΦ~k . The only term surviving is the expectation value hΦ~k Φ−~k i,
R∞ R 2 −(β/N )(r+ck2 )|Φ~k |2
0 d|Φ~k | dϕ~k |Φ~k | |Φ~k | e N kB T
hΦ~k Φ−~k i = R ∞ 2 2 = , (12.25)
−(β/N )(r+ck )|Φ~k | r + ck 2
R
0 d|Φ ~k
| dϕ~
k
|Φ ~
k
| e
where we have used dϕ = 2π, d|Φ| |Φ| = dz 2 /2, and a standard inte-
R R R

gration of an exponential. The order-parameter correlator in ~k-space then

is given by
N kB T
hΦ~k Φ~k 0 i = δ~k,−~k 0 = N δ~k,−~k 0 G0 (~k ) = N δ~k,−~k 0 kB T χ0 (~k ), (12.26)
r + ck 2
where we have used the result (12.23) in the last equation (which is in
agreement with (8.20)). Its real space version is obtained by an additional
Fourier transformation,
1 X i~k·(~r−~r 0 ) kB T
G0 (~r − ~r 0 ) = hΦ(~r )Φ(~r 0 )i = e (12.27)
N r + ck 2
~k
1 X i~k·(~r−~r 0 )
= e G0 (~k ).
N
~k

(12.25) is the equipartition theorem: Every ~k-mode with energy E~k = (r +

ck 2 )|Φ~k |2 /N , see (12.13), involves a thermal energy 2 kB T /2 (with a factor 2
from < + =). The relation (12.26) has a flavor of the fluctuation-dissipation
theorem or linear response (Kubo) formula that connects a response χ0
(since χ = ∂h Φ) to a fluctuation correlator G0 (since G0 = hδΦ δΦi).
Using the results Eq. (12.26) for G0 (~k) one trivially confirms the so-called
static susceptibility sum-rule
Z d
d r kB T
G(~r ) = G(~k = 0) = kB T χ(T ) = . (12.28)
v r(T )
Analyzing Eq. (12.26) provides us with the remaining critical exponents
of the Gaussian model. Repeating the steps in Sec. 11.3, we immediately
find the exponents γ = 1, ν = 1/2, and η = 0: The result for χ in Eq.
(12.28) provides us with the exponent γ = 1, while the result Eq. (12.26) for
χ0 (~k) provides us with the expression for the correlation length ξ(T ) and
the exponent ν. Finally, the behavior of G0 (r) at Tc provides us with the
algebraic deacy G0 (r) ∼ 1/rd−2 and the anomalous exponent η = 0. One can
take the point of view that the exponents for ν and η actually belong to the
Gaussian model, depending on whether one wants to define the non-local
χ(~q ) already within mean-field theory or within the Gauss theory, which
more naturally includes the derivative term describing inhomogeneities in
Φ.
212 CHAPTER 12. FIELD THEORY: GAUSS AND GINZBURG

12.2.3 Gaussian approximation of a Hamiltonian

We can extend our discussion to arbitrary Hamiltonians by constructing

their Gaussian approximation. To do so, we expand H[Φ(~r )] according to
the following scheme: We expand H around a stationary point and define
the saddle point Hamiltonian Hsaddle by

δH
= 0,
δΦ(~r ) Φsaddle
Hsaddle = H(Φsaddle ). (12.29)

According to its definition, the “saddle point” can be any extremum of H.

The mean-field approximation of a theory corresponds to a saddle point
approximation where the saddle is a minimum

ZMF = e−H(Φmin )/kB T . (12.30)

In an external field h(~r ), the saddle point10 Φmin (~r ) of H − dd r h(~r )Φ(~r )
R

defines the mean-field equation of state (we choose v = 1)

δH
= h(~r ). (12.31)
δΦ(~r ) Φmin (~
r)

The Gaussian approximation of a theory in the region around the saddle

point is obtained by the second-order term in the expansion

δ2H
Z
1
HG = H(hΦ(~r )i) + dd r dd r0 δΦ(~r ) δΦ(~r 0 ), (12.32)
2 δΦ(~r )δΦ(~r 0 ) hΦ(~r )i

where δΦ = Φ − hΦ(~r )i are the fluctuations around the expectation value

hΦ(~r )i. Note that within a Gauss approximation, we always expand around
the average hΦi rather then the minimum Φmin : The reason is that in general
fluctuations can change the mean value of Φ and take it away from the
minimizing amplitude Φmin . In such a case, we have to calculate hΦi self-
consistently such that the condition hδΦi = hΦ − hΦ(~r )ii = 0 is satisfied.
The second functional derivative of H defines the operator

δ2H
H2 ≡ , (12.33)
δΦ(~r )δΦ(~r 0 ) hΦ(~
r )i

of which we determine the spectrum by solving the eigenvalue problem

H2 δΦλ (~r ) = ελ δΦλ (~r ). (12.34)

10
Φmin (~r ) may inherit an ~r-dependence from h(~r ).
12.2. GAUSSIAN MODEL 213

We then can rewrite (12.32) as

1X
HG = H(hΦ(~r )i) + ελ Cλ2 , (12.35)
2
λ

where we have taken the eigenset {δΦλ } as an orthonormal basis

X
δΦ(~r ) = Cλ δΦλ (~r ). (12.36)
λ

For the Φ4theory this scheme produces the following results:

Saddlepoint:
rΦmin + 4uΦ3min = h → LMFT.
Gaussian fluctuations (note the change Φmin → hΦi):
H2 = (r + 12uhΦi2 − c∇2 )δ(~r − ~r 0 ).
Spectrum: the operator H2 can be diagonalised by a Fourier transformation,
ελ=k = r + 12uhΦi2 + ck 2 ,
Φ4 1 X ∗
→ HG = H(hΦi) + Φ~k (r + 12uhΦi2 + ck 2 ) Φ~k . (12.37)
2V
~k6=0

Susceptibility (we use a general expectation value hΦ(~r )i) accounting for an
external field h(~r )):
χ−1
0 (~r, ~r 0 ) = [r + 12uhΦ(~r )i2 − c∇2 ]δ(~r − ~r 0 ) [= kB T G−1
0 ], (12.38)
−1 ~ 2 2
χ0 (k ) = r + 12uhΦi + ck ,
where the latter expression applies for a homogeneous situation. The ap-
proximation (12.37) to the Φ4 Hamiltonian can be used in both regimes
T > Tc (where hΦi = 0) and T < Tc (where hΦi = 6 0, resulting in a mass
= r + 12uhΦi2 = 2|r|). Close to Tc , we have r → 0, hΦi → 0, and the
fluctuations can no longer be treated in the Gaussian approximation.
Next, we analyse the statistical mechanics of the Φ4 model in the Gaus-
sian approximation for both regions T > Tc and T < Tc .

For large temperatures T > Tc , we use Φmin = hΦi = 0 and can make
straightforward use of the results obtained in our discussion of the Gaussian
model above, e.g., the free energy (12.19), the specific heat (12.21) with
the exponent α = (4 − d)/2, the susceptibility (12.23) and the correlator
(12.26). Analysing (12.23) as done before in Sec. 11.3 provides us with the
exponents γ = 1, ν = 1/2 and η = 0, all three unchanged when compared
to the mean-field result.11
11
One can take the point of view that the exponents for ν and η first appear in the
Gaussian model, depending on whether one wants to define the non-local χ(~ q ) within the
mean-field theory or within the Gauss theory, which naturally includes the derivative term
describing inhomogeneities in Φ.
214 CHAPTER 12. FIELD THEORY: GAUSS AND GINZBURG

For small temperatures T < Tc , we account for fluctuation corrections

around the saddle point using (including an external field h)
p
hΦi ≈ Φmin ≈ |r|/4u + h/2|r|,

that results in a Gaussian Hamiltonian of the form

1 X 
2|r| + 6h u/|r| + ck 2 Φ~∗k Φ~k .
p
H ≈ HG = H(Φmin ) + (12.39)
2V | {z }
~k6=0
kB T G−1 ~
0 (k )

At zero field, the only change then is the substitution r → r+12uhΦ2 i = 2|r|.
~ is vector-valued (with n components), there is an additional contri-
If Φ
bution to the Hamiltonian (12.39)
1 X0 ∗  p 
Φ~k,i 2h u/|r| + ck 2 Φ~k,i ,
V
i,~k

which derives from the n − 1 transverse modes. The Green’s function G0

k
becomes G0 and we have to account for the transverse Green’s function
k T
G⊥ ~
0 (k, h) = p B (12.40)
2h u/|r| + ck 2

in our analysis. This term contributes with

dd k
Z
kB T 2πkB T
− (n − 1) d
log p (12.41)
2 (2π) 2h u/|r| + ck 2

to the free energy density FG /V in (12.19). Differentiating the free energy

to obtain the critical exponent α for the specific heat cV = −T ∂ 2 F/∂T 2 , the
mean-field term gives us the jump in cV , while the fluctuation term provides
the singular contribution leading to the Gaussian exponent α = 2−d/2. The
term (12.41) does not produce an additional singular contribution (check
this). For h → 0, the mass term in G⊥ 0 vanishes and the massless long-wave
length transverse modes destroy the ordered phase in d ≤ 2.
What does the Gaussian theory lead to for the further exponents β and
δ? The answer is subtle: The simplest view is that the Gaussian theory does
not describe the limit as we approach to the critical point Tc . In particular,
approaching
p Tc from below is problematic as the minimum around Φ =
−r/4u vanishes for r → 0. Furthermore, we need the uΦ4 term to take the
theory to and across the critical point in a well defined way. This manifests
itself in the fact that Φ(τ ) = (−r0 Tc τ /4u)1/2 and Φ(h) = (h/4u)1/3 diverge
in the limit where u → 0—it is, however, exactly these relations that give
us β = 1/2 and δ = 3 in mean-field theory. We will come back to this point
later.
12.3. SELF-CONSISTENT FIELD, HARTREE APPROXIMATION 215

Summarising, we find that the inclusion of Gauss fluctuations to the Φ4

Hamiltonian leaves all the mean-field exponents, except for α, unchanged,
provided we simply take over the values for β and δ over from our mean-field
analysis (marked with a ∗ ),

exp α β γ δ ν η
MFA 0 1/2 1 3 1/2 0
∗ ∗
GA 2 − d/2 1/2 1 3 1/2 0

We thus see that a study of cV is best geared towards measuring the validity

divergence in GT
cV Fig. 12.3: Illustration of the var-
ious contributions to the specific
heat cV (T ) near Tc : mean-field
regular jump on top of the regular back-
background ground and the divergence due to
MF−jump Gaussian fluctuations producing
a critical exponent α = 2 − d/2.
Tc T

of the mean-field approximation, a consideration that led Ginzburg to his

criterion, see Sec. 12.4. Before going on with deriving the Ginzburg-criterion,
we extract some more results from the Gauss approximation.

12.3 Self-consistent field, Hartree approximation

We consider the Gauss approximation of the Φ4 Hamiltonian and start from

the free energy

dd q G−1
Z
kB T 0 (q)
F = FMF + log , (12.42)
2 (2π)d 2π

with the inverse Green function G−10 (q), in a real space representation (see
−1
(12.38)) kB T G0 = (r + 12uhΦ(~r )i2 − c∇2 )δ(~r − ~r 0 ). By taking the func-
tional derivative12 of (12.42), we obtain a new expression for G−1 (after
taking derivatives, we assume temperatures T > Tc such that hΦi = 0 and
12
We use (with v = 1)

δ 2 (Φ(x))2 δΦ(x)
δ(x − y) = 2 δ(r − x)δ(x − y) = 2δ(r0 − x)δ(r − x)δ(x − y).
δΦ(r)δΦ(r0 ) δΦ(r0 )
216 CHAPTER 12. FIELD THEORY: GAUSS AND GINZBURG

kB T G−1 2
0 = r + ck ),

δ 2 βF
G−1 (~r, ~r 0 ) =
δhΦ(~r )iδhΦ(~r 0 )i
δ 2 G−1
Z
−1 0 1 0 (~
x, ~y )
= G0 (~r, ~r ) + dd x dd y G0 (~x, ~y )
2 δhΦ(~r )iδhΦ(~r 0 )i
= G−1 r, ~r 0 ) + 12u δ(~r − ~r 0 ) G0 (~r, ~r 0 ),
0 (~ (12.43)

and hence
dd q
Z
−1 ~ −1 ~
kB T G (k ) = kB T G0 (k ) + 12u G0 (~q ), (12.44)
(2π)d
dd q
Z
2 kB T
= r + ck + 12u .
(2π)d r + cq 2

The improved form (12.44) for the correlator G is called a “one-loop”

approximation. One can obtain this result from a perturbative approach by
resumming loop graphs: Using H = H0 + H0 , H0 ∼ uΦ4 , we expand in H0 ,
1
hΦ(~r )Φ(~r 0 )i = Tr[e−βH Φ(~r )Φ(~r 0 )]
Z
1
≈ Tr[e−βH0 (1 − βH0 + . . .)Φ(~r )Φ(~r 0 )].
Z
We rewrite this equation with the help of graphs, see Fig. 12.4, where
hΦ(~r )Φ(~r 0 )i = G is represented by a thick line, hΦ(~r )Φ(~r 0 )i0 = G0 is de-
noted by a thin line, and u corresponds to a thick dot (with four legs indi-
cating the four fields Φ in uΦ4 ). The first line of Fig. 12.4 can be summed
as a geometric series
R or rewritten graphically in the form G = G0 + G0 Σ1 G,
with Σ1 = −12u G0 the “loop diagram”. The result of such a summation
is given by the Dyson equation G = G0 /(1 − G0 Σ1 ) (or G−1 = G−1 0 − Σ1 )
and provides us with the one-loop result of Eq. (12.44).
The correction to G−1
0 in (12.44) is called a self-energy Σ, to lowest order

dd q
Z
Σ1 = 12u G0 (~q ). (12.45)
(2π)d
The next term within a systematic expansion is given by Σ2 , see Fig. 12.4;
with the idea of resummation in the back, we only include diagrams that
cannot be divided into two by cutting a single internal line, that defines the
irreducible self-energy. The second-order term Σ2 plays an important role in
the RG of the Φ4 theory. After resummation, we can write G−1 = G−1 0 −Σ
or G = G0 /(1 − G0 Σ). Within the skeleton approximation, we replace G0
in this expression by G (and vertices by vertex functions).
The self-consistent field approximation (SCFA) picks up this idea by
replacing G0 in the loop integral Σ1 by a self-consistent expression for G. A
12.3. SELF-CONSISTENT FIELD, HARTREE APPROXIMATION 217

G G0
= + + + + ...
u

+ + ... + + ...

~ drop
~ +

1 2

= + + ...

Fig. 12.4: Diagrammatic representation of G (thick line) through the free

Green’s function G0 (thin line) and 4-leg R vertices ∝ u. The loop (or self-
energy Σ1 ) symbolizes the integral 12u G0 in Eq. (12.44). Keeping only
the loop diagrams, these can be resummed (Dyson equation) and provide
the result in (12.44). Note that so called vacuum (or disconnected) graphs
cancel out in the expectation value G = hΦΦi (linked cluster theorem).
Bottom: Diagrammatic expansion of the (irreducible) self-energy Σ up to
second order in u. The irreducible self-energy involves diagrams that cannot
be divided into two by cutting a single internal line. The Dyson equation
for the Green’s function then reads G = G0 + G0 ΣG, see above, from which
follows an expression for G through G0 and Σ, G = G0 /(1 − G0 Σ), or
G−1 = G−10 − Σ. The one-loop order for G involves only the first term Σ1 .

lazy derivation goes like this: We focus on the Φ4 term of our Hamiltonian
H and contract two of the four Φ-factors H,

Φ4 → hΦ2 iΦΦ, (12.46)

in one of the six possible ways (there are 4 · 3/2 = 6 ways to choose 2 of
4 factors Φ to contract). This is equivalent to replacing r by r + 12uhΦ2 i
(note the factor 1/2 in front of r) and thus we obtain the improved result
for the Green function,

kB T G−1 (~k ) = r + ck 2 + 12uhΦ2 i, (12.47)

Note that hΦi = 0 above Tc within the Gauss approximation (12.37), while
the square amplitude hΦ2 i does not vanish above Tc . Since

dd k
Z
hΦ2 i = G(~r − ~r 0 = 0) = G(~k )
(2π)d
dd k
Z
kB T
= , (12.48)
(2π) r + 12uhΦ2 i + ck 2
d
218 CHAPTER 12. FIELD THEORY: GAUSS AND GINZBURG

the Green function in Eq. (12.47) can be rewritten in a form similar to

(12.44)

dd q
Z
kB T G−1
(~k ) = kB T G−1 ~
0 (k ) + 12u G(~q ), (12.49)
(2π)d

i.e., with G0 = kB T /(r + ck 2 ) replaced by G in the loop integral (this cor-

responds to drawing a thick loop line in the diagram of Fig. 12.4). We
then have to solve (12.48) self-consistently for hΦ2 i that amounts to the so-
called self-consistent field (SCFA), Hartree or random phase (RPA) approx-
imation. (Attention: the nomenclature is not quite uniform; e.g., (12.47),
(12.48), and (12.49) are definitely called SCFA, for (12.44) one says “one-
loop” or RPA. Hartree is often used for the simple first-order perturbative
result with only one single loop). Note that the integral in (12.48) is re-
stricted to k < Λ.
Both, the one-loop and the SCFA allow us to calculate the fluctuation-
induced lowering of Tc . To this end, we determine the inverse susceptibility
(with G involving the shifted mass r̃ = r + 12uhΦ2 i; the corresponding
equation in one-loop reads χ−1 (~q = 0) = 0 = r + kB T Σ1 with r set to zero
in the expression for Σ1 )

dd k
Z
1
χ−1 (~q = 0) ≡ r̃ = r + 12ukB T (12.50)
(2π)d r̃ + ck 2

and noting that it vanishes at Tc . In d > 2 the integral dd k/k 2 in the

R

self-energy is divergent at large k and has to be cut at Λ; the low-k cutoff

at r̃ is irrelevant. Using r ≡ r0 (T − Tc0 ) and
Z Λ
dd k S1d k d−1 r̃→0 S1d
Z Z
1
= dk ≈ dk k d−3
(2π)d r̃ + ck 2 (2π)d r̃ + ck 2 c(2π)d 0
S1d Λd−2
= , d > 2, (12.51)
c(2π)d d − 2

we find that χ−1 vanishes at

12ukB Tc S1d Λd−2

r(Tc ) = r0 (Tc − Tc0 ) = − <0 (12.52)
c (2π)d d − 2

producing a reduction in Tc . Dividing by r0 Tc0 , this can be written as

Tc  ξ d−2
0
1− ∼ Gi <0 (12.53)
Tc0 a

with Gi ∼ kB /∆cv ξ0d the Ginzburg number measuring the importance of

fluctuations, see (12.61). Eq. (12.53) provides us with a formula for the
12.3. SELF-CONSISTENT FIELD, HARTREE APPROXIMATION 219

fluctuation-induced reduction of Tc , i.e., Tc < Tc0 , specifically, fluctuations

at large k ∼ Λ reduce Tc .
In d ≤ 2 the long-wave length fluctuations lead to an infrared divergence
in (12.51) and Tc → 0. The first dimension d for which the finite-temperature
transition into an ordered phase is destroyed by fluctuations when reducing
d is called the lower critical dimension dL (Tc goes to 0 for the first time in
dimension d = dL when coming down from high dimensions d; there is no
phase transition at finite temperatures T > 0 in d ≤ dL ). Thus the above
arguments seem to suggest the dL = 2 for the Φ4 theory. However, note that
this result was obtained within the SCFA; in reality, the Φ4 -theory belongs
to the Ising universality class with discrete Z2 symmetry, for which dL = 1.
Hence, the SCFA seems to overestimate the role of fluctuations.13
Calculating the critical exponents γ and ν in a SCFA, we find
2 1
γ= , ν= . (12.54)
d−2 d−2
For d = 4, fluctuations are not relevant and we recover the mean-field values.
Summarising, we have found the following results:

• Large fluctuations (at short scales) near Tc reduce the critical temper-
ature Tc0 (MF) → Tc (SCFA) < Tc0 . For d ≤ dL we have Tc = 0. For
the Φ4 -theory, we found dL = 2 (in SCFA).

• Large fluctuations modify the exponents in the vicinity of Tc ,

4−d
αMF = 0 → αG = ,
2
βMF = 1/2,
2
γMF = 1 → γSCFA = ,
d−2
δMF = 3,
1
νMF = 1/2 → νSCFA = ,
d−2
ηMF = 0. (12.55)

For comparison, we cite some results for critical exponents in 3D (see

S.K. Ma, N. Goldenfeld, Chaikin-Lubensky), for the Ising- or Φ4 -model
13
Alternatively, increasing d, the lower critical dimension dL is the last dimension where
the transition does not occur. According to the HMW-theorem, dL = 2 for any system with
continuous symmetry (since quasi–long-range-order doesn’t count and the BKT transition
is not of the conventional type); in particular, dL = 2 for the XY -model, for the crystal or
the superfluid transition. The lower critical dimension for a system that breaks a discrete
symmetry, e.g., the Ising model or the Φ4 theory, the lower critical dimension is dL = 1;
a finite temperature transition with Tc ∝  exists for d = 1 + .
220 CHAPTER 12. FIELD THEORY: GAUSS AND GINZBURG

(same universality class), α ≈ 0.11, β ≈ 0.325, γ ≈ 1.24, δ ≈ 4.82,

ν ≈ 0.63, η ≈ 0.03. For the XY-model, α ≈ −0.01, β ≈ 0.35, γ ≈ 1.32,
δ ≈ 5, ν ≈ 0.67, η ≈ 0.04. For the 2D Ising model we have the exact
results α = 0, β = 1/8, γ = 7/4, δ = 15, ν = 1, η = 1/4.
Increasing n (from Ising to XY to Heisenberg) generates slightly lin-
creasing exponents (except for α); hence, larger fluctuations generate
larger exponents. Obviously, the large exponents produced by the
SCFA tell that this approximation overestimates the role of fluctua-
tions.
• For d > 4 fluctuations are irrelevant and the mean-field theory gives
the correct results. One calls the dimension in which mean-field theory
becomes reliable the upper critical dimension dU , for the Φ4 -theory we
have dU = 4.
• Close to Tc the mean-field and Gauss approximations are no longer
valid as they either ignore fluctuations altogether (MF) or ignore the
interactions between fluctuations (mode-mode coupling). The latter
changes the exponents which can be more accurately obtained with the
help of a renormalization group treatment or numerical techniques.

12.4 The Ginzburg criterion

A notable feature of mean-field theory is that it leads to a criterion which

determines it validity. Close to the transition at Tc , fluctuations become
important and the mean-field approximation breaks down. The Ginzburg
criterion defines the region around Tc where the fluctuations are important—
this region is called critical. There are a number of ways to estimate the size
of the critical region around Tc . For example Ginzburg compared the mean-
field jump ∆cV at Tc to the divergence in cV due to (Gaussian) fluctuations
to extract the Ginzburg temperature TG ,
r02 Tc (11.12)
= ∆cV ≡ cV (T − Tc )|TG , (12.56)
8u
where cV (T ) is given by (12.20). Here, we use an alternative argument based
on the order parameter Φ and its fluctuations δΦ. The order parameter is
correlated on the scale ξ(T ). We consider mean-square fluctuations of the
order parameter δΦ = Φ − hΦi averaged over the coherence volume Vξ = ξ d ,
Z
1 2
δΦ = dd r δΦ(~r ) → hδΦ i (12.57)
Vξ
and compare it to the square of the mean order parameter hΦi2 . The fluc-
tuations become relevant when
2
hδΦ i > hΦi2 (12.58)
12.4. THE GINZBURG CRITERION 221

and we define the Ginzburg temperature TG by the relation

2
hδΦ i = hΦi2 |TG . Tc . (12.59)

Using (note the relation ξ02 = c/r0 Tc )

Z
2 1
hδΦ i = 2 dd r dd r0 hδΦ(~r )δΦ(~r 0 )i
Vξ
Z Z
1 (11.19) kB T Y (r/ξ)
= dd r G(~r ) = dd r d−2
Vξ cVξ r
2 Z ∞
kB T ξ d kB T Ad kB T ξ02
= S1 dx xY (x) ≡ Ad d−2 =
cVξ 0 cξ 2cξ d |1 − t|
|r| r2 Tc 2Tc |1 − t| 2∆cV Tc 2
and hΦi2 = = 0 = ξ0 |1 − t|, (12.60)
4u 8u c r0 Tc /c c

as well as (1 − t)2 = (ξ0 /ξ)4 /4, we find (according to (12.58) fluctuations are
relevant when (ξ/ξ0 )d−4 < #)
 ξ d−4 Ad kB 1  Ad kB 2/(4−d)
= = Gi or 1 − tG = .(12.61)
ξ0 T =TG ∆cV ξ0d 2 ∆cV ξ0d

The Ginzburg number Gi tells us about the importance of thermal fluctua-

tions at Tc . Typically ∆cV . kB /dof (dof = degree of freedom) and Ad ∼ 1,
in 3D
Z ∞
e−x
A3 = 4π dx x = 1,
0 4π
hence, we obtain the relative extent of the Ginzburg regime (Tc − TG )/Tc
 1 2/(4−d)
1 − tG ∼ . (12.62)
#dof in ξ0d
Equation (12.62) is the famous Ginzburg criterion. Depending on the sys-
tem’s dimension d, fluctuations are irrelevant (if d > 4), relevant (if d < 4),
or marginal (if d = 4). In particular,

For d > 4, we have

 ξ d−4 Ad kB
>
ξ0 ∆cV ξ0d
and fluctuations never become relevant. We conclude that for Φ4 -type theo-
ries, the mean-field theory provides a consistent description of second-order
phase transitions in dimensions d > 4 and we find that the upper critical
dimension dU for the Φ4 theory (above which mean-field theory describes
well the transition) is dU = 4. Other theories have other critical dimensions,
for example, we have dU = 6 for the percolation problem.
222 CHAPTER 12. FIELD THEORY: GAUSS AND GINZBURG

For d < dU (= 4), fluctuations always become relevant close to Tc . The

criterion (12.62) then determines the critical range TG < T < Tc , inside
which the fluctuations dominate. Despite having calculated this range below
Tc , it is equally valid for T > Tc with 1 − tG → |1 − tG |. The question about
the mean-field character of a phase transition thus becomes a quantitative
one.
A particularly relevant example for a mean-field transition in dimen-
sions d < 4 is the transition between the normal-metallic and supercon-
ducting states (or fermionic superfluid–normal-fluid states). In order to
condense, the fermions have to undergo Cooper pairing, resulting in typi-
cal pairing lengths of the order of ξ0 ≈ 102 . . . 104 Å, much larger than the
typical distance 1/kF ∼ n−1/3 between particles. Hence, we typically have
|1 − tG | ∼ 10−16 , an unobservably small fluctuation region (in high temper-
ature superconductors, this number is much larger, of order 10−2 ). Quite
oppositely, the lambda (λ-)transition of bosonic 4 He is broad, ξ03 n ∼ 1 and
|1 − tG | ∼ 1. As a result, the fermionic and bosonic transitions into the
superfluid state are vastly different, sharp for the fermions and broad for
the bosons, see Fig. 12.5

bosons
cV fermions cV λ −transition
BCS

0 TBCS T 0 Tλ T
critical region

Fig. 12.5: Sharpness of the superconducting/superfluid transitions defined

through the behavior of the specific heat cV at Tc . Left: Fermionic systems
show a sharp transition upon k-space condensation (pairing and conden-
sation) as is visible via the sharp jump in cV . Right: The lambda (λ-)
transition in a bosonic fluid is broad, extending over a temperature regime
of order Tc .

In d = 4, the mean-field theory is correct up to logarithmic corrections,

for example
χ ∼ (|1 − t|)−1 [− log |1 − t|]1/3 (12.63)
for the Φ4 theory.

Above, we derived the criterion (12.61) for the width of the critical region
inside Landau mean-field theory for a second-order phase transition. More
12.4. THE GINZBURG CRITERION 223

generally, we compare (see the static susceptibility sum-rule Eq. (12.28))

G(~k = 0)
Z Z
1 d 1
d r hδΦ(~r )δΦ(~r )i ∼ dd r G(~r ) ∼
Vξ Vξ Vξ Vξ ξd
kB T χ(~k = 0)
∼ ∼ |1 − t|dν−γ ,
ξd
hΦi2 ∼ (1 − t)2β ∼ (1 − t)2β , (12.64)
2
to find the upper dimension dU where both quantities hδΦ i and hΦi2 scale
equally in 1 − t,

(1 − t)dU ν−γ ∼ (1 − t)2β (12.65)

and hence
2β + γ
dU = . (12.66)
ν
For the tricritical point, we have β = 1/4, γ = 1, ν = 1/2 and dU = 3: a
tricritical point is already well described by mean-field theory in 3D.
224 CHAPTER 12. FIELD THEORY: GAUSS AND GINZBURG
Chapter 13

Critical Fluctuations and

Scaling Laws

In this chapter, we want to investigate the system’s behavior in the critical

region. We concentrate on the Landau Φ4 theory and show how the uΦ4
interaction term becomes relevant in the critical range (in d < 4) and can-
not be treated perturbatively. Proceeding further with our scaling analysis,
we seemingly are forced to believe that the critical exponents must take on
the mean-field values. We briefly discuss how, nevertheless, we arrive at
the “anomalous” values for the exponents; treating this question properly
requires an extended discussion in the context of the renormalization group
theory. Luckily, we do not have to calculate all the exponents from scratch:
The scaling laws give us multiple conditions such that two independent ex-
ponents determine all the other.
We start with our “Hamiltonian” (12.2) (we choose v = 1),
Z nr c ~ 2 o
H = dd r Φ2 + (∇Φ) + uΦ4 , (13.1)
2 2
and take two steps to go over to dimensionless variables. First, we introduce
the rescaled parameters
p r u
Φ̃ ≡ βcΦ, r̃ ≡ , ũ ≡ 2 (13.2)
c βc
in order to find the dimensionless Hamiltonian
Z n r̃ 1 ~ 2 o
βH = H̃ = dd r Φ̃2 + (∇ Φ̃) + ũΦ̃4 . (13.3)
2 2

The elastic term tells us that the dimension of the field Φ̃, (L = length) is

[Φ̃] = L1−d/2 , (13.4)

225
226 CHAPTER 13. CRITICAL FLUCTUATIONS AND SCALING

and, correspondingly, we find the dimensions of r̃ and ũ to be

[r̃] = 1/L2 , [ũ] = Ld−4 . (13.5)

Therefore, second, we choose ξ 2 ∼ c/r = 1/r̃ as our (squared) length scale

and define
Φ̃ ũ ~r
ϕ ≡ 1−d/2 , r̃ → 1, u0 ≡ d−4 , ~x = . (13.6)
ξ ξ ξ
Thus the dimensionless Hamiltonian expressed as a function of the dimen-
sionless quantities becomes
Z n1 o
H = H0 + Hint = dd x ~ 2 ] + u0 ϕ4 .
[ϕ2 + (∇ϕ) (13.7)
2
As usual, we have to calculate the partition function
Z
Z = D[ϕ]e−H[ϕ] . (13.8)

R u0 6= 0, u0  1, we
For u0 = 0, this is a trivial task (Gaussian model). For
can try perturbation theory in the interaction Hint = dd x u0 ϕ4 ,
Z  1 2 
Z ≈ D[ϕ]e−H0 [ϕ] 1 − Hint + Hint − ... . (13.9)
2!
Analyzing the “small” parameter u0 more accurately, we find that
(
u 4−d 0, d > 4,
u0 = 2
ξ → (13.10)
βc T →Tc ∞, d < 4,

such that for d < dU = 4 perturbation theory in the Φ4 interacting term

breaks down as we approach to Tc . In fact the limiting value u0 ∼ 1 gives
us the Ginzburg criterion, see (12.61),

ukB T 4−d kB Tc2 4−d

u0 = ξ ∼ ξ (13.11)
c2 r02 Tc /u c2 /r02
kB 4−d
 ξ 4−d
T =TG
∼ 4 ξ ∼ Gi ∼ 1.
∆cV ξ0 ξ0

However, this discussion about the divergence/convergence of perturbation

theory is inaccurate: Below dU , the expansion (13.9) diverges term by term,
which doesn’t mean that the series doesn’t converge, see for example e−t =
n
P
n (−t) /n! for t → ∞. On the other hand, there is no guarantee for the
convergence of perturbation theory if d > dU —the series only converges
asymptotically. A hint to the non-analyticity of Z(u0 ) in u0 = 0 (as relevant
in d > 4) derives from the convergence of Z(u0 > 0) while Z(u0 < 0) with
4
its factor ∼ e|u0 |ϕ diverges. We conclude that resummation methods (Pade,
 227

Borel) play an important role in conjuction with a perturbative analysis in

the critical range.
Let us make use of our dimensional analysis (13.1)–(13.7) to learn more
about the critical exponents. Consider the order-parameter correlator

G(~r ) = hΦ̃(~r ) Φ̃(0)i. (13.12)

Using Eq. (13.4), we obtain its dimension

[G] = L2−d , (13.13)

or, in ~k-space (see (12.27)),

1 X i~k·~r ~
G(~r ) = e Ĝ(k ) → [Ĝ(~k )] = Ld L2−d = L2 . (13.14)
V
~k

Changing the length scale L → L0 ≡ εL, Ĝ transforms according to Ĝ0 (~k 0 ) L02 =
Ĝ(~k ) L2 (since increasing the unit leads to a reduction of the number at fixed
physical quantity) and hence
1
Ĝ0 (~k 0 ) = Ĝ(~k ). (13.15)
ε2

Furthermore, k 2 and r̃ transform according to ([k 2 ] = [r̃] = 1/L2 → k 0 L0 −1 =

kL−1 , r̃0 L0 −2 = r̃L−2 )

k 02 = ε2 k 2 and r̃0 = ε2 r̃,

such that indeed

1 1 1 1
Ĝ0 (~k 0 ) = 2 Ĝ(~k ) = 2 2
= .
ε ε r̃ + k r̃ + k̃ 02
0

On the other hand, for T → Tc , we expect that

Ĝ(~k, Tc ) ∼ k −2+η and hence Ĝ0 (~k 0 ) ∼ ε−2+η Ĝ(~k ), (13.16)

such that we finally have to conclude that η = 0 (from comparing (13.15)

with (13.16)). Similarly, dimensional analysis gives us the correlation length

[ξ] = L ∼ 1/[r̃]1/2 → ξ ∼ |1 − t|−1/2 and hence ν = 1/2.

The dimensional analysis with just one length scale in the problem leads
to the conclusion that the mean-field theory provides us with the correct
exponents, contrary to experimental evidence. What went wrong?
The root of the problem lies in our assumption that only one length
scale determines the physics. In fact there is a second, microscopic length
228 CHAPTER 13. CRITICAL FLUCTUATIONS AND SCALING

scale Λ−1 , that influences the fluctuations in the critical region. In fact,
you might have been bothered by the appearance of the cutoff Λ in the var-
ious expressions of the thermodynamic quantities. Technically, the renor-
malization group theory will solve the problem by eliminating high-energy
short-wavelength fluctuations in a way that accounts for their presence via
simultaneously modifying the long-wavelength fluctuations by renormalising
the coupling constants r and u.
On a non-technical level, we may still ask how to set up consistent scaling
laws that fullfill (13.15) as well as (13.16) with η 6= 0? The idea is to consider
the second length scale a = Λ−1 explicitly. For example, replacing (13.16)
by
Ĝ(~k, Tc ) ∼ aη k −2+η , (13.17)
we can consistently modify the k-scaling of Ĝ under a change in length scale
L0 = εL, with the overall scaling Ĝ0 ∝ Ĝ/ε2 preserved,
 a η
Ĝ0 (~k 0 , Tc ) ∼ a0η k 0−2+η ∼ (εk)−2+η
ε
1 1
∼ 2 aη k −2+η ∼ 2 Ĝ(~k, Tc ).
ε ε
By studying Ĝ at Tc and constant ~k as a function of a, we find the anomalous
exponent η from Ĝ ∼ aη .
Similarly, we can find the anomalous dimension of the correlation length
ξ: Using [ξ] = L, [a] = L, [r̃] = 1/L2 , we modify the scaling law for ξ by
introducing a scaling function f (x),
1
ξ = √ f (r̃a2 ), (13.18)
r̃
with a dimensionless argument x. If f (x → 0) is regular, we again obtain
ν = 1/2, the mean-field value. If, however,

f (x → 0) ∼ xθ , (13.19)

equation (13.18) provides us with a modified exponent,

ξ ∼ r̃−1/2 (r̃a2 )θ ∼ r̃−1/2+θ ∼ r̃−ν

→ ν = 1/2 − θ, (13.20)

with the canonical dimension 1/2 and the anomalous dimension θ of ξ.1 In
the same way one says that the order-parameter field

Φ̃ ∼ L1−d/2−η/2 (13.21)
1
Alternatively, we may assume that the correlation length ξ has to be built from c/r
and from a, ξ 2 ∼ (c/r)1−2θ a2θ , from which we find ξ ∼ (c/r)1/2−θ ∼ |τ |−1/2+θ and hence
ν = 1/2 − θ.
13.1. SCALING LAWS 229

picks up an anomalous dimension η/2. Thus, we have understood that the

critical exponents α, β, γ, δ, and η can assume non-trivial values away from
the results of Landau mean-field theory.
The question of the actual value of these exponents remains. The first
good news is that we will discuss the renormalization group theory that will
provide us with numerical values in the next chapter. The second good news
is that only two of the six exponents are independent and the remaining four
follow from the scaling laws.

13.1 Scaling laws

The scaling laws carry the names of their inventors Fisher, Rushbrooke,
Widom, and Josephson and read

(F ) : γ = ν(2 − η),
(R) : α + 2β + γ = 2,
(W ) : γ = β(δ − 1),
(J) : νd = 2 − α. (13.22)

The last relation is called hyperscaling as it involves the dimension d; it

follows from the scaling of the free energy density f ∼ ξ −d . The mean-
field exponents fulfill these scaling laws, except for Josephson’s law which
is consistent with the mean-field exponents only when d = dU —for the Φ4
theory, in dU = 4 (for the tricritical point, the mean field exponents satisfy
the Josephson law in d = dU = 3). For the Gauss model (with α = (4− d)/2,
γ = 1, ν = 1/2) the Fisher and Josephson laws are satisfied. Within the
Gauss model, we have no information about the exponents β and δ—these
exponents require the incorporation of the uΦ4 terms, e.g., within the mean-
field theory (see also section 16.3).
The relations (13.22) follow from the scaling hypothesis that says that the
singular behaviour of physical (observable) quantities near the critical point
depends only on the divergence of the correlation length ξ(T ) ∼ |1 − t|−ν ≡
|τ |−ν ; however, note the relevance of the microscopic scale a that itself
has brought forward the anomalous part of the critical exponent ν in the
divergence of ξ at Tc . The simplest argument that leads to the relations
230 CHAPTER 13. CRITICAL FLUCTUATIONS AND SCALING

(13.22) then is a dimensional one: Using2

[βf ] = L−d , free energy density,

2−d−η
[G] = L , order parameter correlator,
1−d/2−η/2
   
m = Φ = L , order parameter, (13.23)
2−η
[kB T χ] = L , susceptibility,
−1−d/2+η/2
[h] = [βH] = L , conjugate field,

dd r G(~r ) as well as β dd r HΦ = 1 or
R R
(note that we have used kB T χ =
Φ = −∂H f ) and the definitions

cV ∼ |τ |−α , χ ∼ |τ |−γ , Φ ∼ |τ |β , Φ ∼ h1/δ , (13.24)

we find the relations3

1. |τ |−α ∼ cV ∼ T ∂T2 f ∼ |τ |dν−2 → νd = 2 − α, (13.25)

β −ν(2−d−η)/2
2. |τ | ∼ Φ ∼ |τ | → β = −ν(2 − d − η)/2,
−γ −ν(2−η)
3. |τ | ∼ χ ∼ |τ | → γ = ν(2 − η),
4. |τ |βδ ∼ h ∼ |τ |ν(2+d−η)/2 → βδ = ν(2 + d − η)/2.

Equation 1. leads to (J); the equations 2. and 3. give us

γ νd
β=− + , (13.26)
2 2
while combining 3. and 4., we obtain
γ νd
βδ = + .
2 2
Their subtraction provides us with (W),

β(δ − 1) = γ. (13.27)
2
Note that we assume two anomalous scalings where the microscopic scale a has entered,
one for lengths, i.e., ξ (implying ν 6= 1/2 is non-trivial) and one for the field Φ (implying
η 6= 0 is non-trivial). Near the critical point, the microscopic length a, that was needed in
getting nontrivial exponents for distances r and fields Φ, only appears via the anomalous
values of ν and η; it does not appear again in other quantities due to the assumption of ξ
being the only relevant scale. Indeed, since all quantities can be built from lengths/volumes
and fields, assuming a scaling for lengths/volumes L ∼ ξ, V ∼ ξ d and fields Φ ∼ ξ 1−d/2−η/2
does the job. Still, one may ask whether this assumption of two anomalous exponents
(and the appearance of ξ alone as the relevant length near Tc ) is correct. E.g., one could
wonder, whether another quantity Q involves yet one more anomalous scaling exponent,
i.e., follows a scaling Q ∼ ξ y−κ aκ with anomalous exponent κ rather than Q ∼ ξ y . This
question has to be answered by calculation, e.g., using RG. The question is, whether
the integrals appearing in the perturbative analysis are determined by small wavelengths
(k ∼ 1/ξ) or whether large k ∼ 1/a-values play a role.
3
The term −2 in the exponent |τ |dν−2 originates from the derivative ∂T ξ ∝ τ −ν−1 .
13.1. SCALING LAWS 231

while their addition results in (R),

β(δ + 1) = νd = 2 − α → α + 2β + γ = 2.

Finally, equation 3. delivers (F).

13.1.1 Widom scaling

The following consideration gives a deeper insight into the scaling laws and
is due to Widom: To fix our ideas, we assume a magnetic system and analyse
the set of curves m(τ, h) (Φ → m, τ = (T − Tc )/Tc , βH = h conjugate to
m). Widom then suggested combining the two relations
(
0, τ > 0,
m(τ, 0) ∼ β
(13.28)
±|τ | , τ < 0,

and
m(0, h) ∼ ±|h|1/δ , τ = 0, (13.29)
into a single scaling law
 h 
m(τ, h) ∼ |τ |β M± . (13.30)
|τ |∆
In doing so, the exponents β and ∆, the so-called “gap exponent”, are taken
to be universal and M± are two scaling functions, + for τ > 0, − for τ < 0.
The ansatz (13.30) tells us that by choosing the three parameters β, ∆ and
Tc correctly, we can reduce all the curves m(τ, h) to only two curves, one
describing the situation above and one below Tc . This is an elegant ansatz
that allows

• to reduce the experimental data to two curves and thus represent the
data in a consistent manner, and

• to find Tc and the critical exponents as the data reduction to two

curves only works if β, ∆, and Tc are chosen correctly.

A well known example is the data collapse at the magnetic transition in Ni,
see Fig. 13.1
The scaling functions M± (x) fullfill certain conditions, e.g., they are odd
under a sign change in x,

M± (−x) = −M± (x),

(this follows from the symmetry m(τ, −h) = −m(τ, h)) and

M+ (0) = 0, M− (0) 6= 0,
232 CHAPTER 13. CRITICAL FLUCTUATIONS AND SCALING

all curves T < Tc

3 Fig. 13.1: Data collapse at the
24 ·10 magnetic transition in Ni in the
region around Tc = 627.4 K.
20 M− The plot shows (m/|τ |β )2 ver-
16 sus (|h|/|τ |∆ )/(m/|τ |β ). The ex-
tracted exponents are β = 0.378
12 all curves and γ = ∆ − β = 1.34 (J.S. Kou-
T > Tc vel and J.B. Comly, Phys. Rev.
8
M+ Lett. 20, 1237 (1968)).
4

4 8 12 16 20 ·10 5

since m(τ, 0) = |τ |β M− (0) 6= 0 for τ < 0 and m(τ, 0) = τ β M+ (0) = 0 for

τ > 0. By differentiating (13.30) with respect to h, we obtain

∂m 1 ∂m |τ |β
χ(h = 0) = = ∼ ∆ M0± (0) ∼ |τ |−γ
∂H kB T ∂h |τ |
→ ∆ = β + γ. (13.31)

Furthermore, assuming a scaling ansatz

M± (x → ∞) ∼ xλ (13.32)

for h 6= 0, ∆ > 0, and |τ | → 0 such that x → ∞ and using (13.24),

m(τ = 0) ∼ h1/δ , we find
 h 
m(0, h) = |τ |β M±
|τ |β+γ
 h λ
∼ |τ |β ∼ hλ |τ |β(1−λ)−λγ ∼ h1/δ . (13.33)
|τ |β+γ

For m to remain finite at τ = 0 the exponent of |τ | has to vanish,

β(1 − λ) = γλ, (13.34)

and the scaling in h tells us that λ = 1/δ; as a result, we obtain Widom’s

law
β(δ − 1) = γ (W). (13.35)
Setting up this scaling law seems rather simple, it is, however, fairly tricky
and should be treated with the corresponding respect. Here are two more
examples to practise on, the scaling of the free energy density f and that of
the correlator G.
13.1. SCALING LAWS 233

Free energy density:

At h = 0, the free energy density scales as

βf (τ ) ∼ ξ −d ∼ |τ |νd ∼ |τ |2−α ,
from which we directly obtain Josephson’s law (J). We can include finite
values of h with the scaling ansatz
f (τ, h) = |τ |2−α F± (h/|τ |∆ ) (13.36)
and obtain the magnetization by taking the derivative with respect to h,
0 h→0
m ∼ ∂h f ∼ |τ |2−α−∆ F± (h/|τ |∆ ) ∼ |τ |β (13.37)
→ 2 − α − ∆ = β. (13.38)
Taking one more derivative provides us with the susceptibility
∂2f 00 h→0
χ ∼ ∼ |τ |2−α−2∆ F± (h/|τ |∆ ) ∼ |τ |−γ (13.39)
∂h2
→ 2 − α − 2∆ = −γ (13.40)
and combining these scaling relations, we obtain Rushbrooke’s law
α + 2β + γ = 2 (R). (13.41)
(J)
Alternatively, we can write 2 − α − ∆ = β → νd − ∆ = β, and ∆ = β + γ
(from taking the difference in the above expressions, see also (13.31)) →
νd − 2β = γ, see (13.26).
A systematic approach to the analysis of the free energy density f (τ, h, . . . )
as a function of all its variables starts from a scaling ansatz in the form of
a homogeneous function for f in the scale parameter b,
f (τ, h) = b−d f (bλτ τ, bλh h), (13.42)
with λτ and λh the scaling exponents for the reduced temperature τ and
the external field h (plus other exponents for other variables). Choosing
b ∼ |τ |−1/λτ (to set the first argument in f to unity), as well as b ∼ ξ (as ξ is
the diverging scale in the problem), we recover the scaling f ∼ ξ −d ∼ |τ |νd
of Eq. (13.36) at h = 0 if we set
1
λτ =
. (13.43)
ν
For finite h, we obtain the scaling (since the first argument has gone with
the identification b ∼ |τ |−1/λτ ∼ ξ)
f (τ, h) = ξ −d f0 (ξ λh h) = ξ −d f0 (h|τ |−νλh ) = ξ −d f0 (h|τ |−∆ ), (13.44)
in agreement with Eq. (13.36), if we choose the gap exponent ∆ = λh ν =
β + γ, see (13.31), and hence
β+γ
λh = . (13.45)
ν
234 CHAPTER 13. CRITICAL FLUCTUATIONS AND SCALING

Correlator:

So far, we have used only the scale ξ in the problem (but accounted for the
existence of the second, high energy cutoff scale Λ ∼ π/a (or lattice cutoff
a) by assuming an anomalous contribution θ to the canonical dimension
1/2 for the exponent ν, see Eq. (13.20)). When discussing the correlator at
T = Tc , we have to include the microscopic scale explicitly when analyzing
the scaling of the correlator G (since ξ → ∞ and we need the second scale
a to have an anomalous exponent η). Let a be the microscopic length; we
distribute the scaling factor L2 over a and ξ,

Ĝ(~k ) = a2−s ξ s G(kξ).

For τ → 0, ξ → ∞ and assuming a scaling ansatz

G(x → ∞) ∼ xε , (13.46)

we should have

Ĝ(~k ) ∼ a2−s ξ s (kξ)ε ∼ a2−s k ε ξ s+ε ∼ k −2+η

with a proper scaling in k. The latter tells us that ε = −2 + η, while the

finiteness of the expression at τ = 0 requires us to set s + ε = 0. These
conditions then imply that s = 2 − η. On the other hand, for k → 0 we have

Ĝ(0) ∼ ξ s G(0) ∼ χ ∼ |τ |−γ , (13.47)

(the factor a2−s only corrects the length scaling and is otherwise irrelevant
in this context) and we obtain Fisher’s law

(2 − η)ν = γ (F ). (13.48)

Thus, we have rederived all the scaling laws assuming a scaling ansatz for
various quantities. Note, that all the above discussion is based on the scaling
hypothesis that has an empirical root. A firm derivation of scaling is only
provided by the RG treatment introduced in the following chapter.
Chapter 14

Renormalization Group
Theory

In the previous chapter, we have seen that thermodynamic quantities exhibit

scaling behavior in the critical region characterized by critical exponents α,
β, γ, δ, ν, and η. Luckily, only two of these exponents are independent, with
the other four deriving from the four scaling relations in Eq. (13.22). Here,
we first discuss the (heuristic) block spin transformation due to Kadanoff
with the goal to obtain an idea about the origin and consequences of scaling.
Subsequently, we extend these ideas to the general setting leading to the
renormalization group equations, fixed points, and a computational scheme
for the determination of critical exponents.

14.1 Kadanoff ’s block spin transformation

Consider an Ising model, see Eq. (8.4), on a d-dimensional lattice ~ri with
lattice constant a (we set µ = 1 and drop the superscript on Siz ) and N
sites,
X X
H = −J Si Sj − h Si . (14.1)
hi,ji i

Near the critical point, the correlation length ξ diverges and spins are
strongly correlated, allowing us to go over to a coarse grained description
with averaged spins
X
S̄I = b−d Si , (14.2)
i∈BI

with the Block BI containing bd spins, see Fig. 14.1x (in order to keep
the spins normalized, we may replace S̄I → S I = S̄I /|S̄I |, that implies a

235
236 CHAPTER 14. RENORMALIZATION GROUP THEORY

a block spins rescaled

a

2a

Fig. 14.1: Block-spin transformation: The spins on the original lattice ~ri
with lattice constant a are first combined into block spins S̄I . In a second
step, both the spins and the lattice are rescaled, S̄I → s0 and R~ I → ~r 0 .

somewhat different scaling analysis for G, see below). The new lattice with
positions R~ I now has a lattice constant ba, fewer particles b−d N , and a
correlation length that is reduced by b, ξ → ξ/b. Since the new system has a
shorter correlation length it has moved further away from the critical point.
The above transition to Blockspins goes under the idea of ‘integration’
over small-scale or high-energy degrees of freedom. Given the scale invari-
ance of the problem near a critical point, this ‘integration’ shall not change
the form of the theory, the Hamiltonian, or other quantities, such as the
correlator. However, having integrated over small scale degrees of freedom,
all our quantities now live on the scale ba. Instead of keeping track of the
scale, we simply rescale all quantities, which is the second idea or concept
in building the renormalization group.
Before doing that, let us agree on the following notation to denote coor-
dinates and fields after integration and rescaling,

~ I ↔ ~r −→ ~r 0 = ~r/b,
R
S̄I ↔ s(~r ) −→ s0 (~r 0 ) = bλs s(~r ), (14.3)

where we have introduced a field rescaling with an exponent λs . In order

to find the proper rescaling of the field s(~r ), we request that the correlator
shall remain invariant under integration and rescaling. In a first step, we
integrate over microscopic spins SI to arrive at
X X
G(R ~ J ) = b−2d
~I, R ~ J |−(d−2+η) . (14.4)
~I − R
hSi Sj i = hS̄I S̄J i ∼ |R
i∈BI j∈BJ

Using (14.3), this can be brought to the form

G(~r1 , ~r2 ) = hs(~r1 )s(~r2 )i ∼ |~r1 − ~r2 |−(d−2+η) . (14.5)

14.1. KADANOFF’S BLOCK SPIN TRANSFORMATION 237

Under rescaling, this can be rewritten as1

G0 (~r 01 , ~r 02 ) = hs0 (~r 01 )s0 (~r 02 )i = b2λs hs(~r1 )s(~r2 )i (14.6)

2λs −(d−2+η)
∼ b |~r1 − ~r2 | .

Invariance then requires us to set

λs = (d − 2 + η)/2 (14.7)

as it allows us to include the field rescaling b2λs into the decay |~r1 −~r2 |−(d−2+η) ,

G0 (~r 01 , ~r 02 ) ∼ b2λs |~r1 − ~r2 |−(d−2+η) ∼ |(~r1 − ~r2 )/b|−(d−2+η) (14.8)

∼ |~r 01 − ~r 02 |−(d−2+η) .

Note the difference to Eq. (13.4) providing the trivial field scaling [Φ̃] =
L1−d/2 —the Block spin s0 is already assumed to scale with an anomalous
correction η that is different from the original ‘geometric’ scaling of the field
s.2
Next, let us analyze the behavior of the conjugate field h under coarse
graining. Again, we find the scaling of h(~ri ) (assumed to be smooth
P on the
scale ba) from the requirement that the free energy term Hh = − i h(~ri )Si
remains form-invariant under the transformation to block spins,
X X X
−Hh = h(~ri )Si = ~ I )bd S̄I =
h(R h(~r )bd s(~r ) (14.9)
i ~I
R ~
r
!
X X X
= h(~r )bd−λs s0 (~r 0 ) = h(b~r 0 )bd−λs s0 (~r 0 ) = h0 (~r 0 )s0 (~r 0 ),
~
r r0
~ r0
~

resulting in a scaling relation for h in the form

h0 (~r 0 ) = bd−λs h(b~r 0 ) = bd−λs h(~r ). (14.10)

Hence, the scaling exponent for the field h is (see Eq. (13.45))

d+2−η β+γ
λh = d − λs = = , (14.11)
2 ν
in agreement with the gap exponent ∆ = λh ν = β + γ found in Eqs. (13.44)
and (13.36). In (14.11), we have made use of the scaling relations (??),
(d + 2 − η)/2 = βδ/ν, and Widom’s scaling βδ = β + γ.
1
If we want to rescale the normalized field S I , we make the Ansatz that |S I | ∼ b−λs
and G0 (~r 01 , ~r 02 ) ∼ (1/|S I |)2 G(~r1 , ~r2 ).
2
Technically, a finite anomalous exponent η is produced as a result of renormalization
of the couplings in the Hamiltonian when changing the scale parameter b; e.g., in the
Φ4 -theory of chapter 16.2, a nonzero value for η is obtained at order 2 , with  = 4 − d.
238 CHAPTER 14. RENORMALIZATION GROUP THEORY

Equation (14.11) tells us another important fact, namely that a pair of

conjugate variables φ and χ has scaling exponents that relate via

λχ = d − λφ . (14.12)

This relation is helpful in the discussion of the scaling of the reduced Hamil-
tonian KJ = HJ /T . Expanding around the critical point, we write
X
KJ ≈ K c + τ K c s(~r )s(~r + d~ )
r,d~
~
X
= Kc + Kc τ (~r ), (14.13)
~
r

where K = J/T , Kc = J/Tc , and τ = (T − Tc )/Tc . The local energy

X
(~r ) = s(~r )s(~r + d~ )
d~

then assumes the role of the field conjugate to τ , and imposing the scaling
ansatz for  and τ ,

0 (~r 0 ) = bλ (~r ), (14.14)

0 λτ
τ = b τ, (14.15)

we can relate the exponents λ and λτ via

λ = d − λτ . (14.16)

From (13.43), we know that λτ = 1/ν (since τ scales as ξ −1/ν ) and we obtain
the scaling exponent for the energy
dν − 1
λ = . (14.17)
ν
Note that we cannot find the scaling of  via field rescaling since we cannot
disentangle the interactions between spins when going to a block spin. Also,
on a more abstract level, the scaling of energies is related to ν with its
own anomalous contribution θ, rather than field scaling with the anomalous
exponent η.
Finally, we can find the homogeneity relation for the free energy F . De-
termining the latter by calculating the partition sum over all configurations,
we expect to arrive at the same expression independent of the starting point,
the original or the block variables, hence F (τ, h, χ, . . . ) = N f (τ, h, χ, . . . ) =
N 0 f (τ 0 , h0 , χ0 , . . . ). Using N = bd N 0 as well as the scaling properties of the
various fields, we obtain the scaling or homogeneity relation for the free
energy density

f (τ, h, χ, . . . ) = b−d f (bλτ τ, bλh h, bλχ χ, . . . ). (14.18)

14.2. RENORMALIZATION GROUP EQUATIONS 239

This relation is true for any value of b; choosing b = |τ |−ν (i.e., rescaling by
the maximal value ξ/a) we recover the scaling relation (13.44)
f (τ, h, χ, . . . ) = ξ −d f0 (h/|τ |∆h , χ/|τ |∆χ , . . . ) (14.19)
with the gap exponents related to the scaling exponents via ∆ = λν.
This is a very deep result as it tells us about the relevance of terms in
the Hamiltonian: If the scaling exponent λ or gap exponent ∆ of a variable
v is negative, λ, ∆ < 0 (note that ν > 0) then the variable is reduced
under rescaling and vanishes ultimately, v → bλ v → 0 under rescaling or
v/|τ |∆ = v|τ ||∆| → 0 on approaching the critical point τ → 0. Accordingly,
variables are classified as relevant if λ > 0, irrelevant if λ < 0, or marginally
relevant if λ = 0. This is the origin of universality: although Hamiltonians
may look different with additional terms or terms of different forms, once
they approach the same form under rescaling (i.e., differences are reduced
and finally vanish under rescaling) they describe the same physics.

14.2 Renormalization group equations

The above discussion was simple and illustrative but a bit naive. E.g., it is
rather clear that the transition to block spin variables cannot be accurately
done without generation of new terms in the Hamiltonian that differ from
the simple nearest-neighbor Ising form. In order to overcome this deficiency,
we consider a most general Hamiltonian H[{K}] that depends on a complete
set of coupling constants {K}. In fact, to start with, the Hamiltonian may
involve some subset of finite couplings, but new terms may be generated in
the step of renormalization, therefore the set {K} shall be complete. E.g.,
for an Ising spin system such a Hamiltonian has the form
X
H[{K}] = Kα Sα (14.20)
α

with
Y
Sα = Si (14.21)
i∈Iα

and Iα denoting arbitrary sets of site labels. The Ising Hamiltonian (14.1)
then contains only two types of sets, the single sites with K1 = −h and
the set of nearest neighbor pairs with K2 = −J. The general Hamiltonian
H[{K}], however, is supposed to involve all possible sets
X X
H[{K}] = K0 N + K1 Si + K2 Si Sj (14.22)
i hi,jinn
X X
+ K20 Si Sj + K3 Si Sj Sk + . . .
hi,jinnn hi,j,kinn
240 CHAPTER 14. RENORMALIZATION GROUP THEORY

where hi, jinn denotes nearest-neighbor pairs, hi, jinnn next-nearest-neighbor

pairs, and hi, j, kinn nearest-neighbor triplets, etc. In the following, we drop
the set brackets, keeping in mind that {K} → K is a vector in coupling
space.
Next, we define a coarse-graining transformation that averages over (or
‘integrates out’) short wavelength degrees of freedom—this is the general-
ization of the block spin transformation in Kadanoff’s argument. We call
this operation of coarse graining a renormalization group transformation Rb ,
where the index b tells that we have combined bd original degrees of free-
dom into a new one. This change of scale transforms or ‘renormalizes’ the
coupling constants according to
K 0 = Rb [K]. (14.23)
The operator Rb in this recursion relation may be a rather complicated,
in general non-linear equation; furthermore, it may generate entries in the
coupling vector K 0 where there were none before, i.e., a coupling 0 = Kn ∈
{K} may go over into 0 6= Kn0 ∈ {K 0 }. The transformations Rb with b > 1
define a semi-group (there is no inverse to Rb ),
Rb1 Rb2 = Rb1 b2 , (14.24)
thus justifying the term renormalization group.
The ‘bad’ news is that there is no generally valid recipe that tells how to
(technically) determine Rb —rather, each case has to be treated on its own
(and there may even be various possibilities for the same case) and usually
presents a master piece on its own. Using the setting of the generalized
spin Hamiltonian above, the main overarching ideas can be formulated as
follows: Using the dimensionless Hamiltonian K = −H/T , the partition
function ZN [K] = Tr{Si } exp(K[K, {Si }]) defines the free energy per particle
f [K] = N −1 ln ZN [K]. The RG transformation combines the variables Si in
blocks BI into new block variables S̄I ≡ SI0 , see Eq. (14.2), thereby reducing
the number of variables from N to N 0 = N/bd . The new Hamiltonian
K[K 0 , {SI0 }]) governing the new variables SI0 can then be obtained by a
partial trace over all those configurations {Si } that produce a given set {SI0 }.
Defining the corresponding projection operator P ({Si }, {SI0 }), we then can
write
0 0
eK[K ,{SI }] = Tr{Si } P ({Si }, {SI0 })eK[K,{Si }] . (14.25)
The projection operator is positive, P ({Si }, {SI0 }) ≥ 0, and shall be de-
fined such that the new variables SI0 assume the same range as the original
variables Si . E.g., forPIsing variables Si = ±1, a good choice for the new
0
variables is SI = sign i∈BI Si = ±1 and
Y  X 
P ({Si }, {SI0 }) = δ SI0 − sign Si . (14.26)
I i∈BI
14.3. FIXED POINTS AND EXPONENTS 241

The projection operator in (14.26) plays the role of the filter in Eq. (11.2)
when going from microscopic to coarse grained variables. Furthermore, we
require that P ({Si }, {SI0 }) respects the symmetries of the problem in order
0 0
P K[K , {SI }] involves
to make sure that only symmetry-allowed terms. Using,
the property {S 0 } P ({Si }, {SI0 }) = 1 (since a given set {Si } defines the set
I
{SI0 }, try it out with 6 spins and 2 blocks), the partition sum is conserved,
0 0
ZN 0 [K 0 ] = Tr{SI0 } eK[K ,{SI }] (14.27)
= Tr{SI0 } Tr{Si } P ({Si }, {SI0 })eK[K,{Si }]
= Tr{Si } eK[K,{Si }] = ZN [K],

and the free energies per particle relate via

f [K] = b−d f [K 0 ], (14.28)

in agreement with Eq. (14.18). Hence, we have managed to repeat the

heuristic Kadanoff argument in a more general setting. Furthermore, fol-
lowing the above recipe in some manner, the RG equations (14.23) can be
learned at some stage, e.g., after tracing out the original variables Si in
(14.25). Note, however, that the above only provides a formal and less of a
practical recipe to find Rb .

14.3 Fixed points and exponents

The RG equations (14.23) describe a (discrete) flow in the space of coupling

constants K—understanding K as a k-dimensional vector, we deal with a
flow pattern in Rk . Letting b → 1 or defining log b = log(1 + x) ≈ x → 0,
we arrive at a continuous flow. In order to characterize this flow, we find its
fixed points (or rather vectors) defined through

K ∗ = Rb [K ∗ ]. (14.29)

In addition, we will look at the flow around these fixed points which is
determined by the linearized equation around each K ∗ , see below—knowing
the fixed points and the flow in their vicinity will give us a good picture
of the overall flow of couplings and will provide us with the major physical
insights into the system’s behavior.
Let us first consider the scaling of lengths. As explained above, the RG
takes the system further away from a critical point since lengths shrink by
b and thus

ξ 0 = ξ[K 0 ] = ξ[K]/b = ξ/b. (14.30)

242 CHAPTER 14. RENORMALIZATION GROUP THEORY

However, at a fixed point, ξ has to remain invariant, which implies that

ξ[K ∗ ] = ξ[K ∗ ]/b and hence that ξ ∗ = ξ[K ∗ ] is either zero or infinity. A fixed
point with ξ ∗ = ∞ is called critical and its associated set of couplings char-
acterize the critical point of a phase transition. On the other hand, a fixed
point with ξ ∗ = 0 is trivial and describes a (thermodynamic) phase. The
number of fixed points associated with a specific RG problem depends on the
phases and phase transitions described by the given Hamiltonian; usually we
have ordered (low temperature) and disordered (high temperature) phases
that correspond to two trivial fixed points. If these phases transform into
one another via a phase transition a third (critical) fixed point shows up.
As an alternative, we could have a smooth crossover between these phases
without additional fixed point.
An interesting property of a critical fixed point is its basin of attrac-
tion3 —these special K-points are all characterized by an infinite correlation
length, as repeated action with Rb on K shows,

ξ[K] = bξ[Rb [K]] = b2 ξ[Rb2 [K]] (14.31)

n ∗
= b ξ[Rbn [K]] n
→ lim b ξ[K ] = ∞.
n→∞

Hence, all the K points in the basin of attraction are critical; they define
the critical manifold associated with the fixed point. This is the basis for
universality—all the different Hamiltonians in the basin of attraction of a
given critical fixed point exhibit the same critical behavior characterized by
the same exponents.
Next, let us find these exponents. In order to do so, we have to go
slightly away from the critical point and determine the flow in its vicinity.
Hence, we expand K around K ∗ ,

Kn = Kn∗ + δKn , (14.32)

where Kn denotes the n-th component in the coupling vector K. The RG

transformation takes Kn into Kn0 = Kn∗ + δKn0 and a Taylor expansion
provides us with the relation between δKn0 and δKn ,
X ∂K 0 X
δKn0 = n
δKm = Rnm [K ∗ ] δKm , (14.33)
m
∂Km K=K ∗
m

with
∂Rb [K]n
Rnm [K ∗ ] = . (14.34)
∂Km K=K ∗

Equation (14.33) is the linearized RG equation describing the flow near the
fixed point K ∗ . Assuming that Rnm is symmetric (which is often, but not
3
A point K is in the basin of attraction of K ∗ if limn→∞ Rbn [K] → K ∗ .
14.3. FIXED POINTS AND EXPONENTS 243

always, the case, see later about strategies for non-symmetric situations),
we diagonalize the matrix

Rnm [K ∗ ]vm
α
= bλα vm
α
(14.35)

where the algebraic form bλα of the eigenvalue in b is a consequence of the

group property,

Rb1 Rb2 v α = bλ1 α bλ2 α v α = (b1 b2 )λα v α = Rb1 b2 v α . (14.36)

Next, we expand δK in the eigenbasis v α of Rmn [K ∗ ],

X
δKn = cα vnα (14.37)
α

with coefficients
X
cα = vnα δKn = v α · δK, (14.38)
n

where in the last equation, we have to remind that δK and v α are vectors.
Equation (14.37) is not valid for non-symmetric matrices Rmn .
Applying now the RG transformation to δK expressed in the form (14.37),
we immediately see that the coefficients get multiplied by the eigenvalues
bλα ,

[δK ↔ cα ] → [δK 0 ↔ cα0 = bλα cα ]. (14.39)

Hence, depending on the sign of λα , the coefficient cα will either grow

(for λα > 0) or shrink and vanish (for λα < 0). Indeed, assuming triv-
ial eigenvectors, the eigenvalues λα are just the scaling exponents encoun-
tered previously in Eq. (14.18), the homogeneity equation for the free energy
f (τ, h, χ, . . . ) = b−d f (bλτ τ, bλh h, bλχ χ, . . . ), where the parameters τ, h, χ, . . .
stand for specific couplings that here are denoted by Ki , Kj , Kk , . . . . What
we have gained since there is a recipe how to calculate these exponents once
the Hamiltonian and the RG equation for its coupling vector is known.
The linearized RG defines relevant and irrelevant directions v α in cou-
pling space characterized by positive or negative scaling exponents λα or
eigenvalues bλα that are larger or smaller than unity; vanishing scaling ex-
ponents define marginal directions. These directions in coupling space then
are associated with relevant, irrelevant, and marginal operators or variables
appearing in the system Hamiltonian. The eigenvalues bλα and eigenvectors
v α govern the flow around the fixed point K ∗ : relevant eigenvalues or direc-
tions define a flow away from K ∗ , irrelevant directions describe the flow into
K ∗ and hence the basin of attraction or the critical manifold associated
with the fixed point.
244 CHAPTER 14. RENORMALIZATION GROUP THEORY

K3 critical
fixed point

critical manifold

K1
trajectories / flow K2
under RG

Fig. 14.2: Sketch of the critical point with its critical manifold and flow
under the renormalization group in parameter space {K}. Points on the
critical manifold flow into the fixed point and exhibit the same universal
behavior. Moving away from the critical manifold, trajectories first approach
the fixed point and then move away along an unstable direction. This flow
is governed by the same eigenvectors and eigenvalues for all points near
the critical manifold, thus generating the universal behavior with identical
exponents.

Fixed points then can be classified through their flow pattern: A fixed
point associated with a bulk phase is a pure sink, with the couplings K ∗
characterizing the bulk phase. A critical fixed point exhibits both inward
and outgoing flows. The inward flows define the critical manifold, while
the outgoing flows in the relevant directions typically take the system away
from the fixed point and drive it into a stable bulk phase. Hence, all relevant
couplings have to be carefully tuned in order to ultimately reach the desired
endpoint of the trajectory, be it a critical point or a bulk phase. A generic
situation of the flow in coupling space is shown in Fig. 14.2; we will encounter
specific flow patterns in our applications in sections 15–17.
A final remark concerns the notion of universality: All couplings in the
critical manifold will ultimately flow to the fixed point K ∗ . Going slightly
off this manifold will lead to trajectories that first drive the system towards
the fixed point but then will turn away from K ∗ as the relevant directions
14.3. FIXED POINTS AND EXPONENTS 245

start to kick in. This drive to and away from the fixed point is governed
by the same exponents λα characterizing the fixed point K ∗ , which tells us
that the behavior of all these systems is universal, i.e., governed by the same
exponents or fixed point.

Three final remarks: In case of a non-symmetric matrix Rmn we have

to deal with left- and right-eigenvectors (we use a simplified notation)

R[K ∗ ]vRα = bλα vRα vLα R[K ∗ ] = bλα vLα (14.40)

where vLα is a row vector. It is a few steps to show that left- and right-vector
eigenvalues are the same and that the eigenvectors are orthogonal, vLα · vRα =
δα,α0 , assuming proper normalization. The only change then appears in the
expansion of δK which is done in the eigenbasis vRα ,
X
δK = cα vRα (14.41)
α

with coefficients obtained with the help of the left-eigenvectors,

cα = vLα · δK. (14.42)

Second, the RG equations often appear in differential form. Let us

start out at scale ba and go to (1 + ε)ba, i.e., K = Kb and K 0 = K(1+ε)b ,

K(1+ε)b = R1+ε [Kb ] (14.43)

Expanding in ε, we obtain

Kb + bε ∂b Kb = R1 [Kb ] + ∂ε R1+ε [Kb ]|ε=0 ε. (14.44)

Using b∂b = ∂log b and R1 [Kb ] = Kb as well as the definition B[Kb ] ≡

∂ε R1+ε [Kb ]|ε=0 , we obtain the RG equation in differential form

dKb
= B[Kb ] = ∂ε R1+ε [Kb ]|ε=0 . (14.45)
d log b

The fixed-point condition now reads B[K ∗ ] = 0. As an exercise, we consider

the linearized evolution of Kb near K ∗ along v α , δK1 = cα v α with cα (b =
1) = 1. Then Rs [Kb ] ≈ sλα δKb and B[Kb ] = λα δKb , hence ∂log b δKb =
λα δKb , d log δKb = λα d log b, and finally δKb = bλα δK1 or cα (b) = bλα , as
expected. Thus the linearized version of (14.45) directly provides the scaling
exponents λα .
Third, in high energy physics, the issue of renormalizability of a theory
turns up. In contrast to the situation in statistical physics (where we deal
with degrees of freedom on a lattice and hence enjoy a small scale cutoff
246 CHAPTER 14. RENORMALIZATION GROUP THEORY

a ∼ 1/Λ), there is no physical cutoff in continuous space, hence UV physics

at scales Λ → ∞ becomes relevant. Renormalization describes the flow of
couplings as we change the scale. While we study our world on some finite
scale, the question still arises whether the limit Λ → ∞ exists. This is
indeed the case if the theory is renormalizable.
Chapter 15

Real space normalization of

spin models

In this section, we carry out the renormalization procedure in real space as

suggested by Kadanoff with his block spin transformation. We study the
Ising models in one and two dimensions; both can be solved exactly, via the
transfer matrix technique in 1D and via Onsager’s procedure in 2D. This will
allow us to estimate the quality of the RG analysis. The 1D Ising model has
no finite temperature transition and an exponentially diverging correlation
length ξ as T → 0, with Tc = 0 a zero-temperature fixed-point. The 2D
Ising model has a finite-temperature fixed-point and we find estimates for
Tc (on a triangular lattice) and for the scaling exponents λτ and λh .

15.1 1D Ising model

In section 8.6, we have discussed the exact solution for the 1D-Ising model.
This solution provides an ideal example for the illustration of the Kadanoff
block spin transformation as it can be carried out easily (and exactly) with
the help of transfer matrices. Before doing so, we complete the exact solution
with a calculation of the spin correlations in this model. Furthermore, we
will discuss the results of section 8.6 in the light of our expanded knowledge
on critical phenomena.
The Hamiltonian for the Ising model is given in Eq. (14.1). Due to trans-
lation invariance, the spin correlator G(i, j) = G(n) = hSi Sj i only depends
on the distance n = |i − j| between sites. We define the dimensionless cou-
pling K = J/kB T , set h = 0, and use the transfer matrix technique to obtain

247
248 CHAPTER 15. REAL SPACE RG, SPIN MODELS

(choosing i = 1 and j = n + 1)
1 X
G(n) = S1 hS1 |T |S2 ihS2 |T |S3 i . . . Sn+1 hSn+1 |T |Sn+2 i . . . hSN |T |S1 i
ZN
{Si }
1 X
= S1 T n Sn+1 T N −n
ZN
S1 ,Sn+1
1 h i
= (2 cosh K)N (tanh K)n [1 + (tanh K)N −2n ] . (15.1)
ZN
Here, we have used that T (S, S 0 ) = cosh K (1 + SS 0 tanh K) and

T n (S, S 0 ) = 2n−1 (cosh K)n [1 + SS 0 (tanh K)n ]. (15.2)

With ZN = (2 cosh K)N and taking the limit N → ∞ (note that tanh K < 1
at any finite temperature, hence the term ∝ (tanh K)N vanishes), we find
that

G(n) = (tanh K)n ≡ e−n/ξ (15.3)

1 T →0
with ξ = −1/ ln(tanh K) ≈ e2J/kB T . (15.4)
2
Hence, the correlation length ξ diverges as T → 0, signalling the presence
of a critical point Tc = 0.
Following (12.27), we can relate G and χ,
∞
X ∞
X
kB T χ(q = 0) ∝ G(n) = 2 (tanh K)n − 1 (15.5)
n=−∞ n=0
1 + tanh K T →0 2J/kB T
= ≈ e ∝ ξ,
1 − tanh K
and find a susceptibility that diverges as well (see also Eq. (8.65)); below,
we show that the result χ ∝ ξ is consistent with scaling, χ ∝ ξ γ/ν with
γ/ν = 1.
Let us analyze these results in the light of critical fluctuations and scal-
ing. From (8.62), we can extract the singular part (we set µ = 1)
g+J 1 h 1  hξ 2 i 1
=− + = g0 (hξ) (15.6)
kB T ξ 2 kB T ξ
and comparing with the scaling relation (13.36) expressed through ξ,

g(τ, h) = ξ −(2−α)/ν g0 (hξ ∆/ν ), (15.7)

we find that (note that Josephson’s law tells that 2 − α = νd and d = 1

here)
2−α
=1 and ∆/ν = 1. (15.8)
ν
15.1. 1D ISING MODEL 249

The gap exponent ∆ for h relates to the exponents β and γ via ∆ = γ + β

and using Rushbrooke’s law in the form (2−α)/ν = (β +γ)/ν +β/ν, we find
that β/ν = 0 (note, this result does not make much sense in the disordered
phase above Tc = 0, where the exact result (8.64) tells that m(T, h = 0) = 0;
here, we only need the value of β/ν for finding the other relevant exponents.
Alternatively, you may prefer to eliminate β from the equations and thus
avoid its calculation). Inserting this value back into the relation for the gap
exponent, we find that ∆/ν = (γ + β)/ν = γ/ν = 1. This is consistent
with the scaling of χ ∝ |τ |−γ ∝ ξ γ/ν and Eq. (15.5). Next, we use Fisher’s
relation (13.22) γ = ν(2 − η) to find that η = 1. Finally, Widom’s relation
tells that γ/ν = (β/ν)(δ − 1) which tells that δ → ∞ for the T = 0 critical
isotherm (this result is consistent with the limit T → 0 at finite h of (8.63),
which gives m(h) = µ/Vuc ∼ h1/δ when δ → ∞). Overall, we have found
the exponents and relations
γ 2−α
η = 1, δ → ∞, = 1, = 1. (15.9)
ν ν
Furthermore, with ξ going to infinity faster than any power of T as T → Tc =
0, we can set ν → ∞. This provides a value α → −∞, that is consistent with
the gapped form of cV ∼ (2J/T )2 e−2J/T vanishing faster than any power of
T.
We now turn to the RG analysis of the 1D Ising model. We combine b
sites into a Block und write the partition function in the form
0
ZN (K, L, C) = Tr[T N ] = Tr[(T b )N ] = ZN 0 (K 0 , L0 , C 0 ) (15.10)

with
e−K
 K+L 
eC
T (K, L, C) = e , (15.11)
e−K eK−L

C a constant, and L = h/kB T (we have set µ = 1). The renormalized

couplings K 0 , L0 , and C 0 are defined through the decimation operation

T (K 0 , L0 , C 0 ) = T b (K, L, C) (15.12)

and the requirement of form invariance. For vanishing coupling L = 0, we

find the simple recursion

tanh K 0 = (tanh K)b −→ K 0 = tanh−1 [(tanh K)b ], (15.13)

independent of C. However, in general, we need the freedom of an additional

constant C in order to solve all the conditions implied by (15.12); e.g., for
b = 2, the condition (15.12) reads
 K 0 +L0 0 
e−K e−K e−K
 K+L   K+L 
2C e e C0 e
e =e 0 0 0 . (15.14)
e−K eK−L e−K eK−L e−K eK −L
250 CHAPTER 15. REAL SPACE RG, SPIN MODELS

Changing variables to u = e−4K , v = e−2L , and w = e−4C , inserting and

solving for u0 , v 0 , and w0 , we find the recursion relations

u(1 + v)2
u0 = , (15.15)
(u + v)(1 + uv)
v(u + v)
v0 = ,
1 + uv
w2 uv 2
w0 = .
(1 + v)2 (u + v)(1 + uv)

Let us analyze the resulting flow, first for L = 0, see Fig. 15.1. Eq. (15.13)
defines the fixed points

tanh K0∗ = 0, T0∗ = ∞ and ∗

tanh K∞ ∗
= 1, T∞ = 0. (15.16)

All values K < ∞ flow to K0∗ = 0 which thus is a stable or attractive fixed
point—it describes the high-temperature paramagnetic phase of the Ising
model. Only the point K∞∗ = ∞ has tanh K ∗ = 1 and hence maps to itself.
∞
∗ = 0 unstable fixed point describing the ferromagnetic phase
This is the T∞
of the Ising model. There is no finite temperature phase transition in the
model. Combining (15.13) and the definition of ξ in (15.4), we obtain the
scaling

ξ 0 = ξ/b, (15.17)

hence renormalization takes the system to a point with a smaller correlation

∗ = ∞. Under iterative application of
length, away from the critical point K∞
the RG, we arrive at the paramagnetic fixed point with vanishing correlation
length ξ = 0.
At finite h, the system’s flow is described by the recursion relations
(15.15). These have 3 fixed ‘points’ (see Fig. 15.1)

u∗ = v ∗ = 0, u∗ = 1, v ∗ arbitrary, and u∗ = 0, v ∗ = 1. (15.18)

The first with K ∗ = ∞ (zero temperature) and L∗ = ∞ is an isolated

trivial fixed point that describes an ordered phase with spins aligned by an
infinite external field h. The second with K ∗ = 0 (infinite temperature)
and arbitrary h is a paramagnetic line. The third fixed point with K ∗ = ∞
(zero temperature) and L∗ = 0 (zero field) is a critical one describing a
ferromagnetic phase at T = 0 = h. Let us analyse this critical fixed point in
more detail. Linearizing the recursion relations (15.15) around u∗ = 0, v ∗ =
1, we find

δu0 = 4 δu, (15.19)

0
δv = 2 δv.
15.1. 1D ISING MODEL 251

unstable stable
fixed point fixed point
F
1 h =0

h >0
v = e −2L

paramagnetic line
isolated
fixed point
0
0 u = e −2K 1
T 0 T 8
ξ ξ 0
8

Fig. 15.1: Sketch of the RG flow for the 1D Ising model. At h = 0

(dashed box) the flow is away from the unstable T = 0 fixed-point towards
the paramagnetic high-temperature phase. The flow at finite h ends up on
the paramagnetic line with a renormalized (i.e., larger) field h. The isolated
fixed-point at T = 0, h = ∞ describes a frozen state of aligned spins. The 1D
Ising model does not have a finite temperature phase transition. The zero-
temperature (ferromagnetic) transition has a correlation length diverging as
ξ ∝ e2J/kB T .

Transforming to ξ and L, we have u0 = u∗ + δu0 = δu0 = 4δu = 4u (since

u∗ = 0) and δv 0 = −2L0 = 2(−2L) = 2δv (since L∗ = 0). With u = ξ −2 , the
first equation implies ξ 0 −2 = 22 ξ −2 and hence ξ 0 = ξ/2 = ξ/b, as had to be
expected. The second equation says that L0 = 2L = bL. Let us use these
results in the scaling of the free energy g(K, L, C). According to (15.10),
the free energy expressed through the renormalized couplings relates to the
original one via g(K, L, C) = b−1 g(K 0 , L0 , C 0 ). We first extract the singular
part, g = C + gsing (K, L), since C appears only in the prefactor of g. Next,
we express the singular part of g via ξ and L. Choosing b = ξ, we then
obtain (since L0 = bL = ξL and in analogy to (13.44))

1
gsing (ξ, L) = g0 (Lξ), (15.20)
ξ

in agreement with the result (15.7) of the exact solution.

252 CHAPTER 15. REAL SPACE RG, SPIN MODELS

For completeness, we may also study the flow near the other fixed points.
Near u∗ = v ∗ = 0, the small deviations u and v flow as u0 = u/(u + v) and
v 0 = v(u + v), hence a point (u, 0) ends up at u0 = 1, while a point (0, v)
goes to v 0 = v 2 , in agreement with the flow sketched in Fig. 15.1. Near
u∗ = 1, v ∗ = v, we find that a deviation δ < 0 from u∗ = 1 scales as
δ 0 = −δ 2 v/(1 + v)2 while v changes to v 0 = v[1 + δ(1 − v)/(1 + v)], hence
the flow is downwards towards the paramagnetic line.

15.2 2D Ising model

There are many ways to perform an RG analysis on the 2D Ising model.

Here, we will choose a simple version with spins on a triangular lattice, see
Niemeyer and van Leeuwen, Physica 71, 17 (1974). Due to universality,
it doesn’t matter whether we study a square or triangular lattice—the RG
will provide us with the same coarse-grained picture of this system. We will
again follow Kadanoff’s ideas and define an RG step by going over to block
spins,√choosing blocks with three spins corresponding to a scale parameter
b = 3, see Fig. 15.2. We first study the zero-field case h = 0 and then
include finite fields in a second step. The procedure will provide us with
a phase/flow diagram for the couplings K = J/kB T and L = h/kB T and
approximative values of Kc and the critical exponents.

a
1 2 Fig. 15.2: 2D Ising model
3 on a triangular lattice. Blocks
with three spins form a new
3a lattice √
with larger lattice con-
stant 3a. The interaction
1 2 between blocks involves two
bonds 13 and 23.
3

The 2D Ising model has been exactly solved by Onsager and provides
us with the exacts exponents α = 0 and δ = 15—we can use these results
and determine the other exponents, allowing us to check the quality of our
RG analysis. From Josephson’s law 2 − α = νd, see (13.22), we find ν = 1.
Using Widom’s law, we obtain δ = (β + γ)/β. Using Eq. (14.11), β + γ =
(ν/2)(2 + d − η) as well as (13.26) and Fisher’s law to find 2β = ν(d − 2 + η),
15.2. 2D ISING MODEL 253

we obtain the scaling relation between δ and η,

d+2−η
δ= . (15.21)
d−2+η
The value δ = 15 then provides the anomalous exponent η = 1/4. As a
result, see Eqs. (14.17) and (14.11), we obtain the temperature and field
scaling exponents
1 d+2−η 15
λτ = =1 and λh = = . (15.22)
ν 2 8

Starting out with the dimensionless zero-field Hamiltonian,

X
K=K Si Sj , (15.23)
hi,ji

we divide the lattice into blocks BI and define the new block spins SI through
the majority rule,

SI = 1 for (S)I = (111), (111̄), (11̄1), (1̄11), (15.24)

SI = −1 ≡ 1̄ for (S)I = (1̄1̄1̄), (1̄1̄1), (1̄11̄), (11̄1̄),

where the blocked variables (S)I = (SI,1 , SI,2 , SI,3 ) group the original spins
Si into blocks, SI,i denoting the i-th spin in the I-th block. The set {(S)I }
then covers all spin configurations of the original lattice. We can relate
the coarse-grained Hamiltonian for a specific block configuration {SI } by
summing the original Hamiltonian over all configurations {(S)I } compatible
with {SI },
X
eKb [{SI }] = eK[{SI },{(S)I }] . (15.25)
{(S)I }

In order to evaluate the partial sum in Eq. (15.25), we split the Hamiltonian
into a part K0 involving only bonds within blocks and a part K0 accounting
for bonds connecting different blocks,
XX
K0 = K SI,i SI,j , (15.26)
I (S)I
X X
0
K = K SI,i SJ,j . (15.27)
I6=J (S)I ,(S)J

Treating K0 as a perturbation, we make use of the cumulant expansion (5.72)

to rewrite (15.25) in the form
0
X
eKb [{SI }] = heK i0 eK0 [{SI },{(S)I ]} (15.28)
{(S)I }
254 CHAPTER 15. REAL SPACE RG, SPIN MODELS

with
0 0 02i 0 i2 ]/2+...
heK i0 = ehK i0 +[hK 0 −hK 0 (15.29)

and
K0 [{SI },{(S)I ]}
P
{(S)I } A[{SI }, {(S)I ]}]e
hA[{SI }]i0 ≡ P K0 [{SI },{(S)I ]}
. (15.30)
{(S)I } e

The second factor in Eq. (15.28) is easily calculated: the Hamiltonian K0

decouples into independent blocks and the sum over the individual block
configurations (S)I in Block I is independent on the block spin SI , hence
we have
X
eK0 [{SI },{(S)I ]} = [ZI (K)]N/3 , (15.31)
{(S)I }

with
X
ZI (K) = eK(SI,1 SI,2 +SI,2 SI,3 +SI,3 SI,1 ) = 3e−K + e3K (15.32)
(S)I

the partition function of the block I which only depends on K and N/3 is
the number of blocks.
We approximate the first factor in Eq. (15.28) by its lowest-order term
hK0 i0 in the expansion (15.29) which involves just two bonds between adja-
cent blocks, see Fig. 15.2,

hK0 (SI , SJ )i0 = KhSI,1 SJ,3 + SI,2 SJ,3 i0 = 2KhSI,1 SJ,3 i0 . (15.33)

The Hamiltonian K0 does not couple spins in different blocks and hence the
correlator in (15.33) factorizes

hK0 (SI , SJ )i0 = 2KhSI,1 i0 hSJ,3 i0 . (15.34)

It then remains to determine the averages

1 X
hSI,1 i0 = SI,1 eK(SI,1 SI,2 +SI,2 SI,3 +SI,3 SI,1 ) (15.35)
ZI
(S)I

which depend on the block spin SI ,

e−K + e3K
hSI,1 i0 = SI ≡ s(K) SI (15.36)
3e−K + e3K
with the field scaling factor s(K) ≤ 1. Hence, the term hK0 i0 assumes the
simple form

hK0 (SI , SJ )i0 = 2K[s(K)]2 SI SJ (15.37)

15.2. 2D ISING MODEL 255

with rescaled block spin variables and the rescaled Hamiltonian reads
X
Kb [{SI }] = (N/3) ln ZI (K) + K 0 SI SJ + O(K2 ). (15.38)
hI,Ji

Hence, we find that the RG has generated a constant term (trivial energy
shift) and has renormalized the couplings between spins
K 0 = 2[s(K)]2 K. (15.39)

Going further in the cumulant expansion and including the second-order

term ∝ [K0 ]2 two things happen: In addition to corrections of the nearest
neighbor coupling K, new couplings to next-nearest (M ) and third-nearest
(P ) spins are generated. Second, these next- and third-nearest interactions
M and P have to be included in the determination of the lowest-order cor-
rection hK0 i0 as well and we obtain a set of three coupled equations for K,
M , and P replacing the simple equation (15.39) for K. We P will not follow
this avenue but turn our focus on the magnetic term L = L i Si .
The renormalized Hamiltonians without and with magnetic field term
read
X
eKb [{SI }] = eK[{SI },{(S)I }] ,
{(S)I }
 
e K[{SI },{(S)I }]+L[{SI },{(S)I }] .
X
eKb [{SI }]+Lb [{SI }] = (15.40)
{(S)I }

Expanding in L and subtracting the first from the second equation, we

obtain an expression for the renormalized magnetic term
K[{SI },{(S)I }]
P
{(S)I } L[{SI }, {(S)I }]e
Lb [{SI }] = P K[{SI },{(S)I }]
. (15.41)
{(S)I } e

Inserting the definitions for L[{SI }, {(S)I }] and Lb [{SI }],

XX
L[{SI }, {(S)I }] = L SI,i (15.42)
I (S)I
X
0
Lb [{SI }] = L SI ,
I

into Eq. (15.41) and evaluating the right hand side to lowest (i.e., zeroth)
order in K0 , we obtain
X X X
Lb [{SI }] = L hSI,i i0 = 3LhSI,1 i0 . (15.43)
I (S)I I

Making use of Eq. (15.36), we obtain the recursion relation for the magnetic
field
L0 = 3[s(K)] L. (15.44)
256 CHAPTER 15. REAL SPACE RG, SPIN MODELS

15.2.1 Fixed points and phase diagram

The RG equations

K 0 = 2[s(K)]2 K, L0 = 3[s(K)] L, (15.45)

tell us all universal features of the 2D Ising model. We have 5 fixed points,
√
K ∗ = 0, ∞, s−1 (1/ 2) at L∗ = 0 (15.46)
L∗ = ±∞ at K ∗ = 0,

where s−1 is the inverse function of s, i.e., the finite K (and non-trivial or
critical) fixed point is defined by [s(K ∗ )]2 = 1/2. We first focus on this
critical fixed point. Solving for K ∗ ≡ Kc , we obtain the location of critical
fixed point at1
√
Kc = (1/4) log(1 + 2 2) ≈ 0.3356, (15.47)

to be compared with the Onsager2 result Kc = (ln 3)/4 = 0.2745. The

eigenvalue3 (see Eq. (14.35))

∂K K 0 |Kc = 2s(s + 2Ks0 )|Kc ≈ 1.6235 (15.48)

√
has to be compared with the exact value bλτ = [ 3]1 ≈ 1.7321—not too
bad. The eigenvalue for the magnetic field scaling is
3
∂L L0 |Kc = 3s(Kc ) = √ ≈ 2.1213, (15.49)
2
√ 15/8
not too far from the Onsager result bλh = 3 ≈ 2.8009. Both scaling
exponents are positive and hence Kc is unstable in both directions K and
L, see Fig. 15.3.
Let us have a look at the other fixed points. The high temperature
point K ∗ = 0, L∗ = 0 describes the paramagnetic phase, see Fig. 15.3. The
eigenvalues are bλτ = 1/2 and bλh = 3/2, attractive along K and repulsive
along L. The fixed point at K ∗ = ∞, L∗ = 0 has eigenvalues bλτ = 2
and bλh = 3 = bd (note that the RG result λh = d = 2 is again close to
the Onsager result λh = 15/8 = 2 − 1/8). This is seemingly repulsive in
both direction, however, closer inspection tells that the flow in K is towards
∞ (since every K > Kc is pushed to larger K) and hence K ∗ = ∞ is
1
√ √ √
with x ≡ e4K we have to solve (x+1)/(x+3) =√1/ 2 which gives x = (3− √ 2)/( 2−1)
and multiplying numerator and denominator by 2 + 1, we find x = 1 + 2 2.
2
Note that Tc is not universal, for the square lattice we have kB Tc ≈ 2.27 J while for
the triangular lattice kB Tc ≈ 3.64 J. √
3
We combine ∂K K 0 = 2s(s + 2Ks0 ) with s0 = 8e2K /ZI2 and make use of s = 1/ 2 at
the critical fixed point.
15.2. 2D ISING MODEL 257

S bulk phase sink

L critical

h >0
fixed point

paramagnetic
continuity T 0
P F
fixed point T
8

C K ferromagnetic
discontinuity
fixed point
h <0

first−order
transition

S bulk phase sink

Fig. 15.3: Sketch of the flow diagram for the 2D Ising model. The 2D
Ising model has a finite temperature critical point C and two relevant flow
directions. At h = 0 (an invariant manifold), the flow is towards the low-
temperature ferromagnetic critical point when starting above Kc and to-
wards the high-temperature paramagnetic fixed point when starting below
Kc . The ferromagnetic fixed point is a discontinuity fixed point describing
the first-order phase transition at T < Tc , h = 0 of the generic magnetic
phase diagram in Fig. 8.3 (discontinuous order parameter jumping between
T = 0 bulk ferromagnetic phases with spontaneously broken symmetry with
spins/magnetization pointing up or down). The paramagnetic point is a con-
tinuity fixed point; both F and P are unstable towards the sinks S describing
field-aligned spin-ordered phases (without spontaneously broken symmetry).

actually attractive along K. Another speciality of this fixed point is the fact
that the scaling exponent λh = d. This result tells us that this fixed point
describes a first-order transition (Nienhuis-Nauenberg criterion) in h = 0 at
temperatures T < Tc —this is the usual first-order transition in the magnetic
phase diagram of a ferromagnetic system, see Fig. 8.3. Finally, the fixed
points at K ∗ = 0, L∗ = ±∞ describe high-temperature phases with spins
aligned by the strong magnetic field—they are sinks and hence describe bulk
(ferromagnetic) phases. A sketch of the overall flow is given in the phase
diagram of Fig. 15.3.
258 CHAPTER 15. REAL SPACE RG, SPIN MODELS
Chapter 16

k-space renormalization of
field theories

In this chapter, we discuss another type of coarse graining: rather than

building new block spins in real space, we integrate out short wavelength
modes or large-k degrees of freedom in momentum space. This is usually
done in a continuum model and we focus here on the one-component field
theory in d dimensions. The generic Hamiltonian has the form (see Eq.
(12.2), note that we incorporate the factor 1/kB T in the definition of the
couplings c, r, u and h)
Z
1
dd r c(∇Φ(~r ))2 + rΦ2 (~r ) + uΦ4 (~r ) + · · · − hΦ(~r )
 
K =
2
1 X 1 2 u X
ck + r |Φ~k |2 + 3

= Φ~k1 Φ~k2 Φ~k3 Φ−~k1 −~k2 −~k3
V 2 V
~k ~k1 ,~k2 ,~k3

+ · · · − hΦ~k=0 . (16.1)

Alternatively, we replace
dd k
Z Z
1 X
→ ≡ (dd k) (16.2)
V (2π)d
~k

and cut the integral at Λ ∼ 2π/a with a the lattice constant. The central
quantity to be determined is the partition function
Y0 Z
Z(c, r, u, . . . ) = e−F (c,r,u,... ) = N dΦ~k dΦ~∗k e−K[{Φ~k }] , (16.3)
k<Λ

with the prime reminding us that the product is to be taken over half of
~k-space since Φ ~ = Φ~ for a real order parameter field Φ(~r ). The prefactor
−k k
N reminds us that we actually have an undefined normalization in the mea-
sure of the path integral; this normalization drops out in the calculation of

259
260 CHAPTER 16. K-SPACE RG

expectation values but contributes an offset to the entropy. We will usually

drop it in the following.
The basic idea is to integrate out large momenta k ∼ Λ and find a new
renormalized free energy F 0 (c0 , r0 , u0 ). In order to do so, we first fix the
elasticity c = 1 in order to define the scaling of the order parameter—the
other terms then renormalize with respect to the elastic term. The RG
transformation involves three steps:
1) Integration: we define a new Hamiltonian by integrating out the mo-
menta in a shell Λ/b < k < Λ, with b > 1 assuming the same role as the
decimation in the block spin transformation above (where we went from a
lattice constant a separating sites to a new lattice constant ba separating
blocks; in ~k-space, Λ ∼ 2π/a → 2π/ba ∼ Λ/b)
Y0 Z
e−Kb [{Φ~k }] ≡ eC dΦ~k dΦ~∗k e−K[{Φ~k }] , (16.4)
Λ/b<k<Λ

with the additional but physically irrelevant constant C depending on Λ and

all couplings. The new Hamiltonian Kb

1 Λ/b d
Z
Kb [{Φ~k }] = (d k)[c0 k 2 + r] |Φ~k |2 + . . . (16.5)
2
depends on a reduced set of modes with k < Λ/b and in general picks up a
new elasticity c0 6= 1 which we will deal with in step 3).
2) rescaling: we restore the old cutoff Λ by increasing the unit of length
by a factor b (as we did before with the block spin transformation), r →
r0 = r/b and hence

k 0 = bk. (16.6)

The Hamiltonian now reads

Z Λ
1 2
Kb [{Φ~k }] = d (dd k 0 )[(c0 /b2 )k 0 + r] |Φ~k 0 /b |2 + . . . . (16.7)
2b

3) Field rescaling: we restore the unit elasticity (and hence the original
form of the Hamiltonian) by absorbing the factor c0 /bd+2 in the field Φ,

Φ~0k 0 ≡ (c0 /bd+2 )1/2 Φ~k 0 /b=~k . (16.8)

Defining the scaling factor zk for the field Φ~k through Φ~0 0 = zk Φ~k , we find
k
that

zk = 1/b(d+2)/2 (16.9)

Eq. (16.8) corresponds to the introduction of new block spin variables in the
Kadanoff procedure used before. Here, we have used the simpler version of
16.1. GAUSS MODEL 261

field-rescaling valid in the absence of an anomalous dimension η. In general,

η 6= 0 and we have to rescale the fields by (c0 /bd+2−η )1/2 , implying that
zk = 1/b(d+2−η)/2 . Note that the field rescaling for Φ~r is different; defining
Φ~0r 0 = zr Φ~r , we obtain zr = 1/b(2−d−η)/2 , cf. Eq. (14.7) (with the Fourier
transform at the origin of this difference). This guarantees, that we can
calculate correlators hΦq~1 Φq~2 i = G(q = q1 )(2π)d δ(~q1 + ~q2 ) using either the
original or the renormalized Hamitonian and still arrive at the same result
G(q) ∼ A/q 2−η (assuming an isotropic system).
After performing steps 1) to 3), we arrive at the transformed Hamiltonian

1 Λ d 0 02
Z
0
Kb [{Φ~k 0 }] = (d k )[k + r0 ] |Φ~0k 0 |2 + . . . (16.10)
2
with the renormalized ‘mass’
b2
r0 = r. (16.11)
c0
Hence, we have absorbed the volume bd and have shifted the renormalization
of the elasticity to that of the mass by rescaling the field Φ → Φ0 in order to
arrive at a form invariant Hamiltonian. The partition function takes again
the form (16.3) but with a new (physically irrelevant) prefactor. In the
following, we will apply this scheme to two examples, the Gauss model and
the Φ4 theory.

16.1 Gauss model

The renormalization of the Gauss model is trivial but nevertheless instruc-

tive and helpful to get confidence into the procedure. Furthermore, we need
it as a preparation for the discussion of the Φ4 model. The Hamiltonian is
Z
1
K = (dd k)[r + k 2 ]|Φ~k |2 − hΦ~k=0 (16.12)
2
and makes only sense for r > 0. In performing the RG step 1), we split the
modes into slow and fast ones,
~
Φ~< = Φk for k < Λ/b, (16.13)
k
~
Φ~> = Φk for Λ/b < k < Λ,
k

in real space,
Z Λ/b Z Λ
d i~k·~
r ~
Φ(~r ) = (d k) Φ~k e + (dd k) Φ~k eik·~r (16.14)
0 Λ/b
= Φ (~r ) + Φ> (~r ).
<
262 CHAPTER 16. K-SPACE RG

The Hamiltonian then splits into slow and fast modes as well,
Z Z
1 d 2 < 2 1
K = (d k)[r + k ]|Φ~ | + (dd k)[r + k 2 ]|Φ~> |2 − hΦ~k=0
2 k 2 k
< >
= K +K , (16.15)

where we have included the magnetic field term −hΦ~k=0 into K< . The
partition function separates into factors describing slow and fast modes as
well (we abbreviate DΦ~k = dΦ~k dΦ~∗ ),
k
Z Y
0
ZG (r, Λ) = e−FG = DΦ~k e−K (16.16)
k<Λ
Z Y Z Y
0 < 0 >
= DΦ~k e−K DΦ~k e−K = ZG (r, Λ/b)ZG>
k<Λ/b Λ/b<k<Λ

with
h1 X  2π i
ZG> = exp ln 2 (16.17)
2 k
Λ/b<k<Λ

and where we have set r = 0 in the second, regular factor ZG> . Focus-
ing on the critical behavior of the Gauss model, we can ignore this factor.
The partition function Z G(r, Λ/b) can be brought into the form ZG (r, Λ) by
rescaling momenta k 0 = bk and fields Φ~0 0 = (1/bd+2 )1/2 Φ~< ~ 0 . These steps
k k=k /b
2) and 3) of the general RG scheme then provide us with the renormalized
Hamiltonian
1 Λ d
Z
0
Kb [{Φ~k }] = (d k)[k 2 + r0 ] |Φ~0k |2 − h0 Φ0 (0) (16.18)
2
and the renormalized couplings

r0 = b2 r and h0 = b(d+2)/2 h, (16.19)

where we have used the fact that the integration over modes Λ/b < K < Λ
does not renormalise the elasticity, c0 = 1, a consequence of the factorization
of the modes. Choosing b = 1 + ` close to unity, we can expand, r` =
(1 + `)2 r ≈ r(1 + 2`), to obtain the differential scaling laws

∂` r` = 2r` and ∂` h` = (d/2 + 1)h` . (16.20)

Note that ` ≈ log b; the differential equations (16.20) give us direct access
to the scaling exponents λ, in agreement with (14.45).
The recursion relations (16.19) have two fixed-points at r∗ = 0, ∞, of
which the Gaussian one at r∗ = 0, h∗ = 0 is the interesting one. This is
a critical fixed-point describing the critical behavior of the phase transition
16.1. GAUSS MODEL 263

in a Landau-Ginzburg model within the Gaussian approximation. The RG

then predicts the scaling exponents λτ = 2 = 1/ν (since r = r0 Tc (T − Tc ))
and λh = (d + 2)/2 = (d + 2 − η)/2, from which we obtain ν = 1/2 and
η = 0, in agreement with the analysis of the Gauss model in Sec. 12.2. Both
scaling exponents are positive and hence the fields τ and h are both relevant.
Other exponents we can get for the Gauss model are α = 2 − νd = (4 − d)/2
from the free-energy scaling (13.36) and γ = α + 2∆ − 2 = 1, see (13.40)
and use 2∆ = 2νλh . Note, that we cannot go across Tc with the Gauss
model but always have to stay above Tc (or switch to the low-T side of Tc
with a suitable mass describing the fluctuations around the mean-field order
parameter below Tc and replacing τ → |τ |). This is why the exponents for β
and δ differ from the MF values when using the relations (13.38) and (15.21)
in their calculation: β = (d−2)/4 and δ = (d+2)/(d−2). These results agree
with the mean-field values only in d = 4. The failure to provide the correct
answer is related to our dropping of the uΦ4 term which is a dangerously
irrelevant variable in d > 4 (and a relevant one in d < 4). We will return to
this point at the end of section 16.2.
Before turning to the Φ4 theory, we want to have a look at the free
energy FG of the Gauss model, as we can learn another interesting concept.
Consider the partition function (16.16) at h = 0. We keep the (reduced)
temperature τ = 1 − T /Tc fixed and rewrite r(T ) = r0 (Tc /T )τ → m0 as the
theory’s ‘mass’. The expression (16.16) is regular at small k for any finite τ
but divergent at large k, i.e., when letting Λ go to infinity, which is a rather
unpleasant feature, albeit not influencing the critical behavior (which lives
at small k). The way how to get rid of this ultra-violett singularity is by
going over to a running scale for the mass at fixed τ .
We start from Eq. (16.16), which tells us that we can relate the partition
function at scale Λ with that at scale Λ/b

ZG (m0 , Λ) = ZG> ZG (m0 , Λ/b). (16.21)

We ignore the factor ZG> (it could be absorbed into the normalization factor
N of the path integral) and define

ZG (m0 , Λ) = ZG (m0 , Λ/b) (16.22)

Next, we rescale k = k 0 /b, thereby restoring the old cutoff Λ,

ZG (m0 , Λ) = bM (Λ) Z(m0 b2 , Λ), (16.23)

with M (Λ) the number of modes in the integration volume k < Λ. Compar-
ing the right hand sides of the above two equations, we obtain the scaling
relation

ZG (m0 , Λ/b) = bM (Λ) ZG (m0 b2 , Λ) (16.24)

264 CHAPTER 16. K-SPACE RG

or

ZG (m0 , Λ) = b−M (Λ) ZG (m0 /b2 , Λ/b). (16.25)

Defining a new cutoff λ = Λ/b, we can rewrite this in the form

ZG (m0 , Λ) = zs (Λ, λ)ZG (m(λ), λ), (16.26)

i.e., we can describe the physics originating from m0 and Λ (up to an ir-
relevant shift in energy ∝ ln zs ) through the physics described by a new
partition function (or free energy FG ) that lives on the scale λ with a cou-
pling constant

λ2
m(λ) = m0 (16.27)
Λ2

that depends on the scale (running coupling). All we know when looking at
the world with a resolution λ is the associated mass m(λ). From which scale
Λ the ‘microscopic’ mass m0 originated is neither relevant nor accessible.
The ensuing free energy then is perfectly regular, see Eq. (12.19),

F G(m, λ) = − ln ZB (m, λ) (16.28)

Z λ
Sd−1
= dk k d−1 ln(m + k 2 ) + const.,
2(2π)d 0

and we have managed to absorb the imminent divergence at large k into

a theory at given scale λ (upper cutoff in (16.28)) with a corresponding
rescaled coupling constant m(λ).

16.2 Φ4 theory

We have already seen, that the Gauss model does not allow us to go across
the critical point at Tc of the Landau-Ginzburg free-energy functional and
we need to include the regularization property of the Φ4 term in order to do
so. Accounting for this term is non-trivial and quite some work, as modes
now start interacting.
Let us first quickly check the relevance of this term. In order to do
so, we consider the trivial scaling of the various terms in the real space
version of K, see (16.1). Rescaling lengths and fields, r → r0 = r/b and
φ(~r ) → Φ0 (~r 0 ) = zr (b)Φ(~r ), we find that the mass r, elasticity c, interaction
u, and field h energies scale as

r0 = bd zr−2 r, c0 = bd−2 zr−2 c, u0 = bd zr−4 u, h0 = bd zr−1 h. (16.29)

16.2. Φ4 THEORY 265

Assuming that the elastic term does not rescale, we set zr2 = 1/b2−d (note
the difference to the field scaling (16.8) in k-space) and obtain the trivial
scaling relations (due to power counting)

r0 = b2 r, c0 = 1c, u0 = b4−d u, h0 = b(d+2)/2 h, (16.30)

in agreement with (16.19). Thus the coupling u is relevant for d R< 4 and ir-
relevant for d > 4. Higher order contact couplings of the form up dd rΦp (~r )
scale with

u0p = bd zr−p up = b p−d(p/2−1) up (16.31)

and hence powers p > 4 are less dangerous, i.e., they become irrelevant
already at smaller dimensions d, e.g., the Φ6 is irrelevant for d > 3. Hence,
in our physical dimension d = 3, we are well advised to account for the Φ4
term and may be more casual about the Φ6 term in the Landau-Ginzburg
expansion. As a general rule, we learn that every field Φ adds a power
b1−d/2 and every gradient comes with a factor b−1 in the power count, so
more fields Φ and more gradients ∇~ render a term less relevant. Below, we
will consider dimensions
=4−d (16.32)

in our -expansion with  assumed to be small, hence we are safe from the
Φ6 term for  < 1.
Next, we outline the strategy of the calculation. We start from the
Hamiltonian (16.1) at h = 0, written in the form
Z Λ
1 2
(dd k) ck + r |Φ~k |2

K(c, r, u) = (16.33)
2
Z Λ
+ Π4i (dd ki )u(2π)d δ(Σ4i ~ki )Φ~k1 Φ~k2 Φ~k3 Φ~k4 .

We split the modes into slow and fast ones as in (16.14) and divide the
Hamiltonian as in (16.15),

K = K< + K> + V (16.34)

where V accounts for the mode coupling,

Z Λ
V = Π4i (dd ki ) u (2π)d δ(Σ4i ~ki )Φ~k1 Φ~k2 Φ~k3 Φ~k4 . (16.35)

Note that K< and K> are simple Gauss Hamiltonians over the slow and fast
fields.
266 CHAPTER 16. K-SPACE RG

We treat the interaction term V as a perturbation and apply the cumu-

lant expansion of Sec. (5.5). The partition function then can be written in
the form
Z Y
0
Z(r, u, Λ) = DΦ~k e−K (16.36)
k<Λ
Z Y Z Y
0 < 0 >
= DΦ~< e−K DΦ~> e−K e−V
k<Λ/b k Λ/b<k<Λ k
Z Y
0 <
= Z> DΦ~< e−K he−V i>
k<Λ/b k
Z Y
0
 
< > 2 > >2
≈ Z> DΦ~< e−K e− hVi −[hV i −hVi ]/2 ,
k<Λ/b k

with Z > given in (16.17) and the expectation value h. . . i> to be taken with
the Gaussian K> describing fast modes. We drop Z > (as it does not interfere
with the critical behavior) and rescale with k 0 = bk and Φ~0 0 = zk Φ~<0 to
k k /b
obtain (see (16.23))

Z(r, u, Λ) = bM (Λ) Z(r0 , u0 , . . . , Λ), (16.37)

where Z(r0 , u0 , . . . , Λ) involves the renormalized Hamiltonian (we drop the

primes on the renormalized ~k and Φ, cf. (16.10), as they have become irrel-
evant)
Z Λ
0 0 0 1
(dd k) c0 k 2 + r0 |Φ~k |2

Kb (c , r , u ) = (16.38)
2
Z Λ
+ Π4i (dd ki ) u0 (2π)d δ(Σ4i ~ki ) Φ~k1 Φ~k2 Φ~k3 Φ~k4 + . . .

In lowest order (i.e., 1 ), the elasticity does not renormalize (see the more
detailed discussion later) and the field rescales with zk = 1/b(d+2)/2 such as
to keep c0 = c (see Eq. (16.9)). We derive the RG equations for r and u to
order  = 4 − d, slightly away from the point d = 4 where effects from u
become irrelevant,

r0 ≈ b2 r + corr(u) + corr(ur), (16.39)

0  2
u ≈ b u + corr(u ),

where the first terms follow from simple power counting. Including cor-
rections to order u2 or 2 in the equation for u, we will find a non-trivial
fixed-point u∗ ∼ , hence r∗ ∼  as well, and we do not need corrections ∼ u2
in the equation for r. In general, the RG will generate additional terms ∝ Φp
with p > 4 as well as dispersive couplings, e.g., a dispersive elasticity c0 (k)
is obtained in order 2 and will generate a finite η ∝ 2 ; we will not include
such higher-order in  corrections in our lowest-order analysis below.
16.2. Φ4 THEORY 267

Next, we calculate the cumulants hVi> and [hV 2 i> − hVi> 2 ]. We start
out with the linear correction hVi> and insert the decomposition into slow
and fast modes Φ(~r ) = Φ< (~r ) + Φ> (~r ), Eq. (16.14), into the definition of
V. As a result, we obtain terms such as Φ< 4 , Φ< 3 Φ> , Φ< 2 Φ> 2 , Φ< Φ> 3 ,
Φ> 4 . The first term generates the trivial scaling term b−3d z −4 Φ0 4 = b4−d Φ0 4
with z −4 = b2d+4 and the factor b−3d deriving from the k-integrations (b−4d )
combined with the δ-function (bd ). The second and fourth terms are odd
in Φ> and hence vanish when averaging. The last term provides a constant
shift of the free energy which we ignore. Remains the terms Φ< 2 Φ> 2 which
has the form
Z Λ/b
>
hVi |Φ< 2 Φ> 2 = u (dd k1 )(dd k2 ) Φ~< Φ~< (16.40)
k1 k2
Z Λ
× (dd k3 )(dd k4 ) (2π)d δ(Σ4i ~ki )hΦ~> Φ~> i> .
k3 k4
Λ/b

The correlator hΦ~k3 Φ~k4 i> is known from (12.26),

hΦ~k3 Φ~k4 i = (2π)d δ(~k3 + ~k4 ) [(r + ck 2 )]−1 , (16.41)

and hence
Z Λ/b
>
hVi |Φ< 2 Φ> 2 = uI1 (dd k1 )(dd k2 )(2π)d δ(Σ2i ~ki ) Φ~< Φ~< (16.42)
k1 k2

with
Z Λ
I1 = (dd k) [r + ck 2 ]−1 . (16.43)
Λ/b

Finally, we carry out the integral over ~k2 and rescale to arrive at
Z Λ
hVi> |Φ< 2 Φ> 2 = uI1 b2 (dd k)|Φ~0k |2 (16.44)

This term has the identical shape as the usual mass term in the Landau-
Ginzburg (or Gauss) theory. There are 6 ways to choose the two wavevectors
of the fast modes out of the four wavevectors ~ki . To lowest order in V, we
find that these terms add to the renormalization of r (note that the mass
term appears with a factor 1/2),

r0 = b2 r + 12 b2 uI1 . (16.45)

Proceeding with the second-order contribution [hV 2 i> − hVi> 2 ], we need

to be more systematic, as we get too many terms—this is the entry point to
Feynman diagrams for a classical field theory.
268 CHAPTER 16. K-SPACE RG

(a) (b) (c) (d)

φk1 φk 3 φk>1
>
φk 3 G> G> G>
u u
φk 2 φk φk<2 φk<
4 4
φ< φ<
k 1 + k 2 + k 3+ k 4 = 0 tree level one−loop cluster

Fig. 16.1: (a) four-leg vertex representing the term uφ4 in ~k-space. The
contact interaction translates into the condition of conserved momenta for
the fields on the legs. (b) Same with distinction between fast (dashed legs)
and slow (solid legs) modes. (c) one-loop diagram renormalizing r; the
paired up fields generate a Green’s function G> (~k ) and the momentum ~k
is integrated over. (d) Cluster diagram shifting the energy, of relevance in
the discussion of the free energy density, but irrelevant in the context of
renormalized couplings.

The Feynman diagrams are a mnemonic for properly organizing the per-
turbation theory. A perturbation typically involves a coupling constant g
with p fields, e.g., for our contact interaction we have g = u and p = 4. This
defines the vertex shown in Fig. 16.1(a), a point with four legs. The contact
nature in real space transforms into the conservation of momenta in ~k-space,
hence, the momenta ~ki attributed to the four legs add up to zero. When
separating modes into slow and fast ones, we draw solid and dashed lines for
these fields in order to distinct them, see Fig. 16.1(b). Taking averages over
the fast modes means joining up two dashed lines into a loop that gener-
ates correlators or propagaters, i.e., Green’s functions. Wick’s theorem tells
us, how to bind up lines/fields into Green’s functions: Taking the average
h. . . iG of a 2n-correlator/propagater over a Gaussian weight (or quadratic
Hamiltonian), we have to sum over all possible pairings,
X
hΦ1 Φ2 . . . Φk . . . Φ2n−1 Φ2n iG = Πn Gij , (16.46)
pairs ij

with Gij = hΦi Φj iG , what reduces the task of calculating a horrible 2n-
correlator to finding all allowed pairings and properly counting them (the
clue is that this works only for Gaussian/quadratic weights, this is why we
need perturbation theory at all). Correlators with odd number of fields
vanish.
In the calculation of the average hVi> , we have three different types
of diagrams: The one with four solid lines is the tree-level diagram and
generates the trivial scaling term b u in Eq. (16.39). The six diagrams with
two solid and two dashed lines generate the loop diagram in Fig. 16.1(c)
16.2. Φ4 THEORY 269

(actually, 1-loop) that give rise to the term 12b2 uI1 in Eq. (16.45)—we have
seen how the calculation of this type of diagram is done. The six versions of
this diagram arise from 4 possibilities to place the first dashed line on the
four momenta ~ki and the 3 possibilities to select the second moment, gives
12 possibilities, divided by two for dividing out trivial interchanges in what
is first and second, gives 6 (note the factor 1/2 in front of r, which takes the
factor 6 to 12/2). The last diagram is a cluster diagram, see Fig. 16.1(d).
It involves the integration over two Green’s functions, actually the square
of an integral over a Green’s function
Z Λ 2
>
hVi |Φ> 4 = u d ~
(d k) G(k ) (16.47)
Λ/b

and is a mere number producing a shift in the free energy (note that there
are 3 versions of this diagram).

(a) (b) (d)

[72]

[96]
[144]

[24]

[72]
[144]

(c)
[16]
[96]

Fig. 16.2: (a) Fully contracted (cluster) diagrams shifting the free energy.
(b) Two-loop diagrams with 2 external legs provide corrections to r in order
u2 (or 2 ) and are beyond the requested precision. (c) The six slow-fields
diagram generate a Φ6 term which is irrelevant at the Gaussian fixed point.
(d) Of the two diagrams with four slow fields the first is zero and the second
generates the renormalization of u to order . The numbers in square brack-
ets denote the symmetry factors (multiplicity of appearance of the term in
the perturbation series); note that a factor 1/2 has to be included in the
cumulant expansion of Eq. (16.36).

In going to the second order in V, we start with two 4-leg vertices; the
corrections will be of order u2 ∼ 2 and contribute only to the RG equation
270 CHAPTER 16. K-SPACE RG

for u. We divide the legs into two classes describing slow (solid lines) and
fast (dashed lines); in performing the average hV 2 i> we can pair dashed
lines (fast modes) on individual vertices or between vertices. The former
ones appear in the product hVi> 2 as well and exactly mutually cancel one
another. What remains are the connected diagrams and we write
2 2
hV 2 i> − hVi> = hV 2 i> >
c = hhVii , (16.48)

where the last bracket symbol hh. . . ii (double brackets) usually denotes a
cumulant, hence, the index c can be read as ‘connected’ or ‘cumulant’. So
the great news is that only diagrams with dashed lines contracted between
vertices matter. However, there are still many of them. A number of them is
not relevant in our discussion: the fully contracted (cluster) diagrams in Fig.
16.2(a) are mere numbers shifting the energy. The two-loop diagrams with
2 external legs, see Fig. 16.2(b), provide corrections to r in order u2 and are
beyond the requested precision. The six slow-fields diagram in Fig. 16.2(c)
generates a Φ6 term which is irrelevant at the Gaussian fixed point (see our
power counting argument above). Remain the four slow-fields diagram in
Fig. 16.2(d), of which the first is zero and the second (which is of a bow-tie
form), relevant one, is (cf. (16.38))

hV 2 i> 2  ~ ~
c |bt = u b I2 (k3 + k4 ) (16.49)
Z Λ
× Π4i (dd ki )(2π)d δ(Σ4i ~ki ) Φ~k1 Φ~k2 Φ~k3 Φ~k4

with (accounting for all conserved momenta)

Z Λ
I2 (~k3 + ~k4 ) = (dd k5 ) G(k~5 ) G(~k3 + ~k4 − ~k5 ), (16.50)
Λ/b

hence, the correction is dispersive (i.e., depends on momentum). To the

precision/order we are treating the problem here, we can ignore this com-
plication and substitute ~k3 + ~k4 = 0, hence,
Z Λ
I2 = (dd k)[r + ck 2 ]−2 . (16.51)
Λ/b

This diagram appears in 6 × 6 × 2 = 72 versions (6 for the choice of fast

modes in each vertex and 2 ways to pair); including the factor 1/2 and the
minus sign of the cumulant expansion, this provides a factor −36 and hence
the RG equation for u takes the form (we copy the equation for r as well)

r0 = b2 r + 12 b2 uI1 , (16.52)
0   2
u = b u − 36 b u I2 .
16.2. Φ4 THEORY 271

Finally, we have to evaluate the integrals I1 and I2 . These integrals

depend on the non-integer parameter d, what gives rise to additional de-
pendencies on . Fortunately, in the present case, we are not interested in
such further corrections (of higher order in ) and we can evaluate the inte-
grals in dimension d = 4. The first integral provides 2 terms (we use that
1 − b−2 ≈ 2` = 2 ln b with b = 1 + ` and ` small; furthermore, we set c = 1),
Z Λ Z Λ
4 2 −1 S4
dkk 1 − r/k 2
 
I1 = (d k)[r + k ] ≈ (16.53)
Λ/b (2π)4 Λ/b
1
≈ [Λ2 ln b − r ln b]
8π 2

with Sd = 2π d /Γ(d/2) the surface of a sphere in d dimensions, S4 = 2π 2 .

The second integral reads
Z Λ Z Λ
2 −2 S4 1
I2 = 4
(d k)[r + k ] ≈ dkk −1 ≈ ln b. (16.54)
Λ/b (2π)4 Λ/b 8π 2

The differential (with ` = ln b  1) RG equations take the form

3
r` − r = 2r + 2 u(Λ2 − r) `,
 
(16.55)
2π
h 9 i
u` − u = u − 2 u2 `.
2π

Rescaling r → r0 = r/Λ2 (cf. (16.27)) and dropping the prime, we arrive at

the final form of the differential RG equations,

3  
∂` r` = 2r` − 2 u` r` − 1 , (16.56)
2π
h 9 i
∂` u` = u`  − 2 u` .
2π
The first terms on the right hand side of these equations, 2r` and u` , derive
from field rescaling, while the second terms account for the integration over
short-wave degrees of freedom. Furthermore, the term ∝ u` derives from
large k ∼ Λ; it is responsible for the shift in Tc , see also the discussion in
Sec. 12.3, and is the relevant term in the determination of the non-trivial
fixed point. The term ∝ u` r` acts only as a correction to this shift, but is
the relevant one in the determination of the critical exponent λτ , see below.
Besides the Gaussian fixed point at rG∗ = 0, u∗G = 0, these equations have
a second non-trivial (Wilson-Fisher) fixed point of order  at
∗
rWF = −/6, (16.57)
∗ 2
uWF = (2π /9) .
272 CHAPTER 16. K-SPACE RG

The exponents associated with these fixed points follow from the linearized
equations

2 − (3/2π 2 )u∗ (3/2π 2 )(1 − r∗ )

 
B(r, u) = . (16.58)
0  − (9/π 2 )u∗

At the Gaussian fixed point we have

2 3/2π 2
 
B(r, u) = (16.59)
0 

with eigenvalues and eigenvectors

λτ = 2 and λu =  (16.60)
τ u 2
~v = (1, 0) and ~v = (−3/4π , 1).

Similarly, the Wilson-Fisher fixed point is characterized by

2 − /3 (3/2π 2 )(1 + /6)

 
B(r, u) = (16.61)
0 −

with eigenvalues and eigenvectors

λτ = 2 − /3 and λu = − (16.62)

~v τ = (1, 0) and ~v u = (−3/4π 2 , 1).

Hence, the flow directions at the two fixed points align and the Wilson-
Fisher fixed point lies on the ~v u direction of the Gaussian fixed point, see
Fig. 16.3.
Let us then have a look at the flow of the Landau-Ginzburg theory. This
depends crucially on the dimension. When d > 4 the Wilson-Fisher fixed
point resides in the unphysical region u < 0 and the only fixed point re-
maining is the Gaussian one. So the system’s behavior follows that dictated
by the Gaussian fixed point, where our results tell that u is an irrelevant
variable since λu =  < 0 (although dangerously irrelevant, see below), while
r or τ is relevant.
When the dimension drops below 4, d < 4,  = 4 − d > 0 and things
change quite appreciably. Now the Gaussian fixed point has become fully
repulsive and the critical behavior is governed by the Wilson-Fisher fixed
point which now resides in the physical region with u > 0. The ~v u di-
rection now is attractive with respect to the WF fixed-point and defines a
critical manifold—all Hamiltonians on this line flow to the WF fixed-point.
Deviating from this line, the parameters first flow towards the WF fixed-
point and then turn away to follow the (shifted) r or τ axis to the high- or
low-temperature phase. E.g., viewing the LG Hamiltonian as a continuum
16.2. Φ4 THEORY 273

cri
tica
u
d< 4

lm
flow

ani
under RG

fol
d
WF fixed point

broken symmetry disordered phase

ordered phase T

8
T 0

G fixed point
r

unphysical region

u
d >4
G fixed point

unphysical r
region WF fixed point

Fig. 16.3: Sketch of the flow of the φ4 - or Landau-Ginzburg theory. For

d < 4 (main figure) the Gaussian fixed-point (G) is unstable and a new
Wilson-Fisher (WF or Heisenberg for the On -model) fixed-point appears rep-
resenting the non mean-field like phase transition. All Hamiltonians starting
on the critical manifold flow into the WF fixed-point. Parameters away from
the critical manifold flow towards the ordered-phase at low temperatures or
towards the disordered phase at high temperatures. For d > 4 the WF fixed-
point disappears in the unphysical region (with u < 0) of the diagram and
the (stable) Gaussian fixed-point describes the corrections to the mean-field
behavior.

description of the Ising model, the flow is towards the disordered paramag-
netic or ordered ferromagnetic phase. The transition then is governed by the
exponents ν = 1/λτ ≈ 1/(2 − /3) ≈ 1/2 + /12 > 1/2 and has an (anoma-
lous) correction θ ≈ −/6, see Eq. (13.20). Since elasticity is not normalized
at this order, the anomalous exponent η ≈ 0 vanishes—it will become finite
when going to order 2 , see below. Knowing the two exponents ν and η,
we can determine the remaining ones using the scaling laws: The exponent
γ = ν(2−η) ≈ 1+/6 follows from Fisher’s relation and α = 2−dν ≈ /6 fol-
lows from Josephson’s relation. The exponent β = (2 − γ − α)/2 ≈ 1/2 − /6
is derived with the help of Rushbrooke’s law and δ = 1 + γ/β ≈ 3 +  is the
result of Widom’s law. In d = 3,  = 1, we get the exponents α ≈ 0.166,
274 CHAPTER 16. K-SPACE RG

β ≈ 0.333, γ ≈ 1.166, δ ≈ 4, ν ≈ 0.5833, η ≈ 0.

16.3 Dangerously irrelevant variables

An irrelevant variable (or operator) is called dangerous when its vanishing

causes problems with the order parameter expectation value below Tc . Such
is the case for the uΦ4 term in Landau theory, which is indeed irrelevant in
d > 4 and one might thus have the idea to get the exponents from Gauss
theory plus scaling. However, this idea is doomed, as, e.g., Josepson’s law
is not satisfied. The point is, that u is a dangerously irrelevant parameter
in the sense that Φ ∼ (|r|/4u)1/2 → ∞ and Φ ∼ (h/4u)1/3 → ∞ as u → 0,
thus potentially causing problems in the calculation of β and δ. As a way
out, consider the scaling of the free energy density under the inclusion of
the variable u, a generalization of (13.42)
f (τ, h) = b−d f (bλτ τ, bλh h, bλu u), (16.63)
with λu = . The magnetization then scales as
m(τ, h, u) = −∂h f = b−d+λh m(bλτ τ, bλh h, bλu u). (16.64)
Setting h = 0 and b = ξ = τ −1/λτ , we obtain
m(τ, 0, u) = τ (d−λh )/λτ m(1, 0, τ −λu /λτ u). (16.65)
While previously we have assumed m(1, 0, bλu u) to be regular (i.e., u is
irrelevant variable), in the context of the Landau Ginzburg or Φ4 theory we
should account for the singular behavior of m(1, 0, u) ∝ u−1/2 when u goes
to zero, see Eq. (11.8). Hence we obtain the corrected scaling
m(τ, 0, u) ∼ τ (d−λh )/λτ τ λu /2λτ . (16.66)
Similarly, setting τ = 0 and choosing b = h−1/λh , we obtain
m(τ, 0, u) = h(d−λh )/λh m(0, 1, h−λu /λh u) (16.67)
and using the singular behavior m(0, 1, u) ∝ u−1/3 of Eq. (11.9), we obtain
m(0, h, u) ∼ h(d−λh )/λh hλu /3λh . (16.68)
These corrected scaling laws provide us with the correct exponents for β and
δ,
d − λh + λu /2 1
β = = , (16.69)
λτ 2
1 d − λh + λu /3 1
= = ,
δ λh 3
where we have used the Gaussian exponents λτ = 2, λh = 1 + d/2, and the
exponent λu = 4 − d deriving from power counting.
16.4. β-FUNCTION AND ‘CHARGE’ 275

16.4 β-function and ‘charge’

Above, we have carried out the renormalization on the level of the Hamil-
tonian—this is one of the great achievements of Ken Wilson. An important
element in this machinery, besides integration, is the process of rescaling.
But even before Wilson’s RG, renormalization was used in an alternative
manner without rescaling, the idea being, that a physical quantity is renor-
malized by the elimination of degrees of freedom through integration; the
elimination of these (short-scale) degrees of freedom then is compensated by
a change in the coupling or in the quantity under consideration. A famous
example of this type of ‘old-style’ RG is the renormalization of the charge
in QED, corresponding to the renormalization of u in the present context of
the φ4 -theory. In this RG, the integration is not over a small k-intervall, but
is usually carried out over all k-values from Λ until a large distance cutoff
R ≤ ξ(τ ) is reached; the flow of the quantity is then studied as a function
of the scale R, see also the discussion at the end of Sec. 16.1 above.
The RG equation (16.56) for the coupling u as a function of scale R
involves only u itself and we write it in the form

∂ln R u = β(u). (16.70)

This is the Gell-Mann Low equation for the ‘charge’ u and β(u) is the Beta-
function providing the flow of u. For d = 4, we have β(u) = −(9/2π 2 )u2
(see Eq. (16.56) with  = 0)

9 2
∂ln R u = − u , (16.71)
2π 2
and a simple integration provides us with the solution
u0
u(R) = . (16.72)
1 + (9u0 /2π 2 ) ln(R/a)

Hence, when integrating out degrees of freedom, the initial charge u0 de-
creases and renormalizes to zero as R/a → ∞—as a result, the theory
becomes ‘free at large distances’.
In a similar spirit, we can integrate out the degrees of freedom at scales
below R and obtain a renormalized mass r or a renormalized temperature
τ : in the absence of rescaling, there is no trivial term 2r in (16.56) and we
drop the u term in (16.56) by absorbing it in a shifted Tc , i.e., we replace
Σ1 → Σ1 (r) − Σ1 (r = 0) in the derivation of the RG equation. These
manipulations leave us with the equation

3
∂ln R r = − ur (16.73)
2π 2
276 CHAPTER 16. K-SPACE RG

that is solved by
r0
r(R) = . (16.74)
[1 + (9u0 /2π 2 ) ln(R/a)]1/3

As R → ∞, we should stop the flow at R/a ∼ ξ/a; alternatively, note that

r = c/ξ 2 ∼ τ = 1 − T /Tc cuts off the k → 0 divergence in G(k) ∝ 1/(r + k 2 ).
We then can rewrite this solution as a renormalized temperature τren
τ
|τren | = . (16.75)
[1 + (9u0 /4π 2 ) ln(1/a2 τ )]1/3

With the susceptibility χ ∼ G(k = 0) ∼ 1/r ∼ 1/τ , we find that renor-

malization produces logarithmic corrections in d = 4,

1 [− ln |τ |]1/3
χ∼ ∼ , (16.76)
τren |τ |

in agreement with (12.63). For the coherence length, we should compare

massive (rΦ2 ) and elastic (c∇Φ)2 ∼ c(Φ/ξ)2 ) terms in the Hamiltonian to
find that |τren | ∼ 1/ξ 2 , from which we obtain

1 [− ln |τ |]1/6
ξ∼ 1/2
∼ p . (16.77)
τren |τ |

We can use the above ‘old-style’ analysis also away from d = 4. We start
from Eq. (16.52) without rescaling cwcouplings and fields,

r0 = r + 12 uI1 , (16.78)
0 2
u = u − 36 u I2 ,

i.e., we only account for the integration over short-wave degrees of freedom.
For the integral I2 , we have (in d = 4)
Z Λ
S4 1
I2 ≈ dkk −1 ≈ ln(RΛ). (16.79)
(2π)4 1/R 8π 2

cf. Eq. (16.54). Going to d = 4 − , we obtain the new result

Z Λ
Sd 1 1 Λ
I2 ≈ d
dkk −1− ≈ − 2  (16.80)
(2π) 1/R 8π k 1/R
1 1 Λ 1 R 
= − 2  ≈ 2 .
8π k 1/R 8π 

The RG equation for u, rather then ∂ln R u = −(9/2π 2 ) u2 , then changes to

∂R u = −(9/2π 2 ) u2 . (16.81)

16.5. TWO-LOOP EXPANSION FOR G AND FINITE η 277

Redefining
 a 
u = ũ , (16.82)
R
this equation assumes the form of (16.56),
h 9 i
∂R ũ = ũ  − 2 ũ (16.83)
2π
with a fixed point ũ∗ ∝ . We find several new results: First, the original
u ∝ /R flows to zero algebraically in R rather than logarithmically u ∝
1/ ln R, see Eq. (16.72). Second, we note that the u in (16.56) has been field-
rescaled with a factor b , what corresponds exactly to the step in (16.82),
hence our ũ above corresponds to the u in (16.56). So while it is the fixed
point plus exponent defined nearby that is of central interest in the Wilson
RG, it is the flow of the coupling that is in the focus of the present approach.
Repeating the above analysis for the coupling r, we note that the reduced
integral I1 (r) − I1 (0) that is relevant after absorption of the Tc shift has a
similar form as I2 ,
Z Λ
1 ru R
I1 (r) − I1 (0) = −ru (d4− k) 2 ≈ − , (16.84)
1/R rk + k 4 8π 2 
that leads to the equation
3
∂R r = − rũ. (16.85)
2π 2
Its solution near the fixed point ũ∗ assumes the form
 a /3
r(R) = r0 . (16.86)
R
Comparing massive and elastic terms in the Hamiltonian, we find that
r(ξ)/c ∼ 1/ξ 2 , from which we obtain the scaling for ξ in the form,
1/(2−/3)
ξ ∼ 1/τ0 , (16.87)
implying a new exponent ν = 1/(2 − /3) ≈ 1/2 + /12, in agreement with
the result of the Wilson approach. Similarly, we obtain for the susceptibility
χ(R) ∼ 1/τ (R) ∼ R/3 /τ0 at the scale R and stopping the flow at the scale
1+ν/3
ξ, we find that χ ∼ 1/τ0 , resulting in an exponent γ = 1 + /6.

16.5 Two-loop expansion for G and finite η

As already explained in Sec. 12.3, we can find a perturbative series for the
correlator G(~k ). Expressing the Green’s function with the help of the self-
energy Σ,
G−1 (~k ) = G−1 ~ ~
0 (k ) − Σ(k ), (16.88)
278 CHAPTER 16. K-SPACE RG

we find in lowest order the non-dispersive result

dd q
Z
Σ1 = −12u G0 (~q ), (16.89)
(2π)d
while the next order (most conveniently written in real space) contributes
the non-local term (see Fig. 12.4 bottom for the relevant diagrams)
Z
~ 2 ~
Σ2 (k ) = −96u dd r G30 (~r ) eik·~r . (16.90)

The symmetry factor 96 is obtained by choosing the first, second, and third
leg in each vertex for connection with the second vertex, and dividing by 6
to avoid double-counting (we can order the sequence of 3 lines in 6 ways),
resulting in (4 · 3 · 2)2 /6 = 96. The first term Σ1 contributes a constant
and shifts Tc downwards due to large-k fluctuations, see Sec. 12.3. Here,
we are interested in the renormalization of the elastic energy k 2 and hence
determine the dispersive (i.e., ~k-dependent shift δΣ(k) = Σ2 (k) − Σ2 (0).
Since this term is already ∝ u2 ∼ 2 , we can work in d = 4. At Tc , we
have G0 (q) = 1/q 2 and transforming to real space, we obtain (using spatial
isotropy and ~r = (r, 0, 0))
d4 q ei~q·~r dp d3 k eipr
Z Z
G(r) = = (16.91)
(2π)4 q 2 (2π)4 p2 + k 2
π
Z −|k|r Z
3 e 1 1
= d k = dk k e−kr = 2 2 .
(2π)4 |k| 4π 2 4π r
For the dispersive shift, we obtain
96u2
Z Z
~
δΣ(k) = −96u2 d4 r G3 (r) [eik·~r − 1] ≈ d4 r (~k · ~r )2 G3 (r),
2
96u2
Z Z
3 3 dΩ ~
= dr r G (r) (k · ~r )2 (16.92)
2 2π 2
96u2 k 2
Z
= dr r5 G3 (r), (16.93)
8
where we have expanded exp(i~k · ~r ) − 1 ≈ i~k · ~r + (~k · ~r )2 /2 and performed
an angular average (~k · ~r )2 = k 2 r2 /4. The remaining integral is logarithmic;
the divergence at small x is cut on 1/Λ ∼ a, the one at large x is cut on 1/k
~
(due to the factor [eik·~r − 1]). We then find the result (with S 4 = 2π 2 the
surface area of a 4D sphere)
2π 2 Λ
δΣ(k) = 12u2 k 2 6
ln . (16.94)
64π k
The correlator thus picks up a logarithmic correction that we can transform
to a small power k η (since 1 − η ln k ≈ exp(−η ln k) = k −η ),
G−1 (k) = k 2 [1 − (3u2 /8π 4 ) ln(k/Λ)] = k 2−η Λη (16.95)
16.5. TWO-LOOP EXPANSION FOR G AND FINITE η 279

with
3u2 2
η= = , (16.96)
8π 4 54
where we have used the fixed point u∗WF = (2π 2 /9) in the last equation, see
Eq. (16.57). Note the numerical smallness of η. The anomalous scaling of
G(k) ∼ aη /k 2−η is in agreement with the discussion in Chap. 13, see Eq.
(13.17); as expected, the microscopic scale a enters the result through cut-
ting off an integral in the perturbation series. Translating the k-dependence
to a real-space dependence, we find that G(r) ∼ 1/rd−2+η , what produces
the field scaling zr = b(d−2+η)/2 , see Eq. (14.7).
Philosophically, the above renormalization of G(k) should be viewed as
a renormalization of the elasticity. Let us assume that τ is small but finite,
i.e., we are slightly away from the critical point. Then, the renormalization
produces a dispersive elasticity c(k) ∝ k −η on scales k > 1/ξ (‘short’ dis-
tances), what leads to a non-zero anomalous exponent η. On the other hand,
on scales k < 1/ξ (‘large’ distances), we should stop the flow at k ∼ 1/ξ and
the elasticity is renormalized to a new constant c → c[1 + (2 /54) ln(Λξ)].
Finally, the anomalous exponent η is defined at τ = 0 or ξ = ∞ and c is
dispersive on all scales.
The systematic calculation of the RG equations for the couplings r and
u (at ~k = 0) to second order (2-loop RG) is pretty involved. In particular,
care has to be taken not to double-count logarithmic terms that already
have been accounted for in the one-loop calculation.
280 CHAPTER 16. K-SPACE RG
Chapter 17

Berezinskii-Kosterlitz-
Thouless Transition

In the last chapter, we study an example of a continuous real-space RG,

the Berezinskii-Kosterlitz-Thouless transition that we have already encoun-
tered in section 8.7.2. The low-energy Hamiltonian of this system involves
spin-waves and topological excitations or vortices, where the latter are a
consequence of the compactness of the angle-variable ϑ. The Hamiltonian
of this system can be written in the form (see Eq. (8.87))

d2 R ~
Z
K
KXY ≈ [∇ϑsw ]2 (17.1)
2 a2
Z Z
− πK d2 R d2 R0 nv (R ~ 0 ) ln[|R
~ )nv (R ~ 0 |/a] + Ec
~ −R ~ ),
d2 R nv (R

with K = J/kB T and the vortex density nv (R ~ ) describing the singular

~
component of ϑ(R ). The parameter Ec = ec /kB T describes the core energy1
ec of a vortex and usually appears within the present context through the
(inverse) fugacity

z = exp(−Ec ). (17.2)

We assume ec to be large, implying a small density of vortex pairs at low

temperatures (otherwise, the transition turns first-order). The parameter
K describes the spin stiffness (in the context of superfluidity this is the
superfluid density, in the context of melting this is an elastic constant).
1
This energy arises from the immediate vicinity of the vortex center where the usual
continuum description breaks down. It is a central quantity determining the density of
vortex pairs at low temperatures. At small ec , the BKT transition turns first-order as
vortex-pairs become too dense, see A. Jonsson, P. Minnhagen, and M. Nylen, Phys. Rev.
Lett. 70, 1327 (1993).

281
282 CHAPTER 17. BKT TRANSITION

Fig. 17.1: The interaction between two charges (constituting a large dipole)
is screened by smaller dipoles. The latter are polarized in the field of the
charges of the large dipole with angles θ preferentially in the forward direc-
tion; temperature fluctuations reduce this polarization.

The basic idea of the RG is that the coupling K becomes renormalized

due to the presence of vortex pairs. This can be easily understood from
the Hamiltonian (17.1) which describes a CoulombPgas, i.e., logarithmically
interacting charges in two dimensions with energy i6=j qi qj ln(Rij /a). Con-
sider two vortices/charges with opposite polarity at a distance d, then vortex
pairs of size R < d will align such as to screen the interaction between the
two charges, see Fig. 17.1. Such a screening is conveniently described by
a dispersive dielectric constant ε(R) which reduces the interaction between
charges, πK ln(R/a) → (πK/ε) ln(R/a), motivating our replacement of the
coupling K by a new screened coupling K(R) = K/ε(R).
Our task then is to find an expression for the dielectric constant ε(R).
The dielectric constant ε relates to the polarizability of the medium via the
usual relation ε = 1 + 4πχ with χ = nα, n the density of dipoles and α the
polarizability of these dipoles. Here, dipoles are given by vortex–anti-vortex
pairs and their polarizability is determined by their ability to align in an
electric field.
Given the ‘charges’ q = ±1 of the vortices, a pair of extension R consti-
tutes a dipole of size qR. Given an electric field E, the polarizability in the
field is given by

α(R) = ∂E hqR cos θi|E=0 (17.3)

with θ the angle between the dipole- and the electric-field vector and the
average has to be taken with the Boltzmann weight exp(−E qR cos θ/kB T ),

q2 2 q 2 R2 π
α(R) = R hcos2 θi = = KR2 , (17.4)
kB T 2kB T 2

where we have used that πK = q 2 /kB T (as follows from comparing the
prefactors of the log-interaction). Such pairs of extension R appear with a
density

~ d2 R = z 2 −2πK ln(R/a) d2 R
n(R) e . (17.5)
a2 a2
 283

Collecting all elements, we find that the dispersive dielectric constant is

given by
Z R
ε(R) = 1 + 4π d2 R0 n(R0 )α(R0 ) (17.6)
a
3−2πK
R
dR0 R0
Z 
3 2
= 1 + 4π Kz .
a a a
The renormalized coupling K(R) = K/ε(R) then can be related to the
original coupling K via
dR0 R0 3−2πK
Z R  
1 1 3 2
= + 4π z . (17.7)
K(R) K a a a
Equation (17.7) is the starting point of our RG analysis, as it relates the
coupling K(R) at scale R to the coupling K at scale a. Equation (17.7) can
be obtained in other physical ways, e.g., by considering the renormalization
of the superfluid density ρs ↔ J due to the presence of vortices, however, this
path involves a subtle analysis of longitudinal versus transverse response,
see Nelson’s book (Defects and Geometry in Condensed Matter Physics).
Furthermore, Eq. (17.7) can be lifted to a self-consistency equation for the
function K(R), see below.
In the end, we want to integrate out all vortex–anti-vortex pairs and
find the renormalized coupling K∞ , hence we set R → ∞ in (17.7). The
result (17.7) with R → ∞ is well-behaved for large K > 2/π, i.e., at low
temperatures, as the integral remains finite in the limit R → ∞. Indeed,
at low temperatures, vortex–anti-vortex pairs are bound and our small z
expansion should be fine. On the other hand, at small values of K < 2/π,
i.e., high temperatures, the integral diverges at large distances R0 → ∞ and
the perturbation theory breaks down. This coincides with the unbinding
of vortex–anti-vortex pairs, a large increase in the dielectric response and a
collapse of the renormalized K.
The idea then is to renormalize K such as to deal with this divergence
at high temperatures. In order to do so, we break up the integral into two
parts,
Z ∞ Z ab Z ∞
= + , (17.8)
a a ab
with b > 1 and rewrite (17.7) as
∞ 3−2πK
dR0 R0
Z 
1 1
= + 4π 3 z 2 (17.9)
K∞ Kb ab a a
with
3−2πK
ab
dR0 R0
Z 
1 1
= + 4π 3 z 2 . (17.10)
Kb K a a a
284 CHAPTER 17. BKT TRANSITION

In (17.10), we integrate out pairs only in a thin shell a < R0 < ab, a typical
decimation step where we integrate over ‘fast’ degrees of freedom, here, small
vortex–anti-vortex pairs. Next, we perform the usual rescaling step in the
cutoff ab → a in order to recover the original relation (17.7)
∞ 3−2πKb
dR0 R0
Z 
1 1
= + 4π 3 zb2 (17.11)
K∞ Kb a a a
but with renormalized couplings Kb and zb ,

zb = b2−πK z. (17.12)

Note that we have replaced K in the exponent of (17.10) by Kb in (17.11),

a result that can be derived by pushing the perturbative analysis to higher
order in the fugacity z.
The results (17.10) and (17.12) are our RG equations. We bring them
into differential form by choosing b = e` with a small correction ` (or ` = ln b
with b > 1 close to unity) and arrive at

∂` K`−1 = 4π 3 z`2 , (17.13)

∂` z` = (2 − πK` )z` .

Before analyzing (17.13), we discuss an alternative derivation of these

equations that goes beyond perturbation theory. In order to do so, we go
back to equation (17.5) and note, that the exponent weighting pairs of size
R should actually contain the screened potential rather than the unscreened
one. This screened potential can be found by integrating the screened force
2πK/Rε(R),
Z R Z ln R
0 1
Usc (R) = 2πK dR 0 = 2π d ln R0 K(R0 ), (17.14)
a R ε(R0 ) ln a

and the density n(R, θ) is modified to

~ d2 R = z 2 −2π R ln R d ln R0 K(R0 ) d2 R
n(R) e ln a . (17.15)
a2 a2
The perturbative RG equation is transformed into a self-consistent equation
for KL (we rewrite R → L = ln(R/a)),
Z L  Z L0 
1 1 3 2 0 0 00
= + 4π z dL exp 4L − 2π dL KL00 . (17.16)
KL K 0 0

Defining the ‘screened’ fugacity on scale L,

 Z L 
0
zL = z exp 2L − π dL KL0 , (17.17)
0
 285

and choosing a small value L → `, the equation (17.16) can be rewritten

into the differential form (17.13). This derivation is obviously more direct,
however, it does not make explicit the usual RG steps of decimation and
rescaling and lacks the insight how the renormalization of coupling constants
deals with the divergence appearing in the perturbative scheme.2

z separatrix

critical
line of fixed point
initial points

separatrix

0
0 π/2 −1
K
T 0 T
8

Fig. 17.2: Sketch of the flow pattern of the Berezinskii-Kosterlitz-Thouless

transition. The flow in the K −1 -z (temperature T versus vortex density)
is characterized by a critical fixed point at K −1 = π/2 and two separa-
trixes (dash-dotted). At small temperatures below the separatrix the flow
is towards z = 0, i.e., no vortices/charges. The resulting phase exhibits
stiffness and quasi–long-range order with ξ = ∞. At high temperatures
and above the separatrix the flow goes towards the disordered phase with
free vortices/charges—vortex–anti-vortex pairs unbind at kB TBKT = πK/2
or K −1 = π/2.√ As T → TBKT from above the coherence length diverges as
ξ ∼ a exp(π/b T − TBKT ). The dashed line is the location of physical initial
conditions.

When integrating the differential equations (17.13), one arrives at the

flow pattern sketched in Fig. 17.2. The initial conditions for this flow are
determined by the line (dashed in Fig. 17.2)
−1
z0 = e−Ec = e−(ec /J)K = e−(ec /J)/K (17.18)
which has an essential singular form e−1/x . The fixed-points of the flow are
manifold: there is a critical fixed-point at
K ∗ = 2/π, z ∗ = 0, (17.19)
2
This divergence appears because the chosen perturbation theory with bound pairs is
actually a ‘bad’ starting point at high temperatures.
286 CHAPTER 17. BKT TRANSITION

determining the Berezinskii-Kosterlitz-Thouless critial temperature

πJ
kB TBKT = , (17.20)
2
in agreement with the result (8.90) derived from the simple free energy (or
energy–entropy) argument. Below [K ∗ ]−1 (at low temperatures T < TBKT )
extends a line of fixed-points, with every point characterized by an infinite
correlation length ξ(T < TBKT ) = ∞, hence, the system is critical or scale
invariant (or algebraic, see Eq. (8.76), note that the exponent in C(R) ∝
1/RkB T /2πJ goes from zero at T = 0 to 1/4, C(R) ∝ 1/R1/4 at TBKT , hence,
correlations weaker than 1/R1/4 cannot sustain a stiff phase any more) at all
temperatures below TBKT . Finally, there is a high-temperature fixed-point
at K −1 , z → ∞ describing the disordered or liquid or normal phase. The
low temperature phase T < TBKT exhibits stiffness and quasi long-range
order (hence, no finite order parameter); so there is spin-stiffness J in a spin
system, shear c66 in a 2D crystal, or superfluid density ρs in a superfluid
film, but no order, e.g., no δ-function Bragg peaks in a 2D crystal (where
the topological excitations are dislocations).
Let us finally discuss the RG equations and their solution in the vicinity
of the critical fixed-point. In order to do so, we define the new variables
x = x` = 2 − πK` and y = y` = 4πz` , then Eqs. (17.13) take the form

∂` x = y 2 , (17.21)
∂` y = xy.

Taking their ratio, we find that xdx = ydy or d(x2 − y 2 ) = 0, which defines
a hyperbola

y 2 = x2 + C (17.22)

with C an integration constant depending on the initial condition x0 , y0 .

The parameter ` then is a curve parameter taking us through the hyperbo-
las. The parameter C = 0 defines the two separatrices, the left separatrix
y = −x, x < 0 is the critical manifold (or line) with all points (initial
Hamiltonians) flowing into the critical fixed-point, hence C = 0 corresponds
to T = TBKT . Depending on the sign of C, the fate of the trajectories are
very different: For C < 0, the trajectory falls √ into the critical line below
the fixed-point and reaches y = 0 at x = − −C, implying that vortex-
anti-vortex pairs remain bound, hence, we are below TBKT . For C > 0,
the trajectory first approaches the critical fixed-point and then takes off to
infinity—vortex–anti-vortex pairs unbind and the system flows to the disor-
dered high-temperature phase, hence, we are above TBKT and we make the
ansatz (with b a numerical)

C = b2 (T − TBKT ). (17.23)
 287

Next, we solve the flow for both cases above and below TBKT and derive
some physical results. For T > TBKT , we deal with a positive parameter C
and the flow takes off to infinity, after approaching the critical point. We
solve the differential equationR for x = x` , ∂` x = y 2 = x2 + C, by simple
2
R
integration, dx/(x + C) = dl and obtain (with a trajectory starting at
x0 , ` = 0 and ending at x, `)
√ √ √
C` = arctan(x/ C) − arctan(x0 / C). (17.24)

For T < TBKT , √

the parameter C is negative and the flow approaches a finite
point x∞ = − −C,
√ y∞ = 0 as ` →√∞. The integralR is now a combination
of logs, dx [(x − −C)−1 − (x + −C)−1 ]/2 = dl, and we obtain
R

√ √
√ x − −C x0 + −C
2 −C` = ln √ √ . (17.25)
x + −C x0 − −C
The first solution above
√ TBKT provides us with an expression
√ for ξ(T ). On
approaching TBKT , C  |x0 | and we have arctan(x0 / C) ≈ −π/2 since
x0 < 0. As ` increases, the trajectory approaches the critical point but then
deviates away from it on a scale ξ. We √ choose a symmetric trajectory
√ with
xξ = x(`ξ ) = −x0 such that arctan(xξ / C) ≈ π/2 and `ξ = π/ C. Using
the relation between distances R and the curve parameter `, ` = ln(R/a),
we find that the coherence length
√
T −TBKT
ξBKT (T ) ∼ a eπ/b (17.26)

rapidly diverges on approaching TBKT from above. As T drops below TBKT ,

ξBKT remains infinite. Indeed, the fixed point at x∞ < 0, y∞ = 0 is reached
at ` → ∞ and hence ξBKT /a ∼ e` = ∞.
Furthermore, the solution (17.26) below TBKT provides√us with an ex-
pression for the renormalized stiffness: as ` → ∞, x∞ → − −C, and using
the definition of x, x = 2 − πK = 2 − π(J/kB T ), we find that vortices reduce
the stiffness on approaching TBKT as

J J(TBKT ) bp 2 bp
= + TBKT − T = + TBKT − T . (17.27)
kB T kB TBKT π π π
The second relation makes use of the universal jump in the stiffness at TBKT ,

J(TBKT ) 2 J∞
= ≡ , (17.28)
kB TBKT π kB TBKT
i.e., the stiffness collapses at TBKT from a finite value J∞ = (2/π)kB TBKT to
zero (Nelson-Kosterlitz universal jump, in the superfluid density).
A nice application is the establishment of true order and a finite Tc
in a layered system, e.g., a layered superconductor as is the case in the
288 CHAPTER 17. BKT TRANSITION

cuprate (or high-Tc ) materials. Assume a bulk 3D material that is made up

of 2D layers. Let us assume an anisotropic coupling, with J⊥ small in the
transverse direction, J⊥  J. If J⊥ were zero, there would not be a bulk
transition, rather, each layer would undergo a quasi-ordering transition at
TBKT = πJ∞ /2kB . On decreasing T towards TBKT , the coherence length ξBKT
diverges. Once the coupling energy (ξBKT /a)2 J⊥ reaches the temperature T ,
the layers lock up and order, hence the condition (ξBKT (Tc )/a)2 J⊥ ≈ Tc
defines the location of the 3D phase transition into the ordered (not quasi-
ordered) state at
 2
2π
Tc = TBKT + . (17.29)
b ln(TBKT /J⊥ )
Index

absolute zero, 19 Bose gas

action Stefan-Boltzmann law, 119
Euclidean, 90 broken symmetry, 139
adjabatic, 8
anomalous dimension, 189, 228 caloric equation, 4, 5
anomalous exponent, 228 caloric equation of state, 54
anti-ferromagnetic coupling, 135 Carnot
antiparticle, 65 efficiency, 15
arrow of time, 16 Carnot machine, 14
asymptotic series, 226 chain rule, 7
atoms chemical potential, 4, 20, 65
cold, 79 classical gas
average Boltzmann gas, 80
ensemble-, 46, 49 Clausius-Clapeyron, 27, 33
self-, 47 coarse-grained problem, 182
Avogadro’s number, 5 coherence length
in a superconductor, 149
Baker-Haussdorf, 75 cold atoms, 79
barion, 65 collision time, 42
basis, 67 compressibility
BCS transition, 187 adiabatic, 10
Berezinskii phase, 158, 189 negative, 37
Berezinskii-Kosterlitz-Thouless, 151, thermal, 10
162, 281 concave, 27
bicritical points, 200 configurational integral, 74, 85
BKT transition, 281 constraint, 38, 51
block spins, 253 convex, 25, 27
Kadanoff, 235 cooperative ring exchange, 94
Bogoliubov variational principle, 96 correlation length, 48
Boltzmann constant, 5 correlation time, 48
Borel, 226 correlator, 140
Bose Einstein condensation, 126 spin, magnetization, 140
critical density, 126 corresponding states, 166
critical temperature, 126 Coulomb gas, 282
fugacity, 127 coupling
transition line, 128 running, 263

289
290 INDEX

critical in phase space, 46

manifold, 242, 244 density matrix, 69
critical density density operator, 48
BEC, 126 diagrams, 35
critical dimension diamagnetismc, 107
upper, 221 dielectric constant, 282
critical endpoint, 32 differential, 5
critical exponent complete, 6
correlation length, 188 exact, 6
critical isotherm, 185 pfaffian, 7
order parameter, 144, 185 dimensional analysis, 225, 227
specific heat, 145, 186 dimensional reduction, 110
Gauss, 210 dimensionless variables, 226
susceptibility, 145, 186 dipoles, 282
critical exponents, 183 Dirac representation, 91
Gauss model, 209 dissipative systems, 48
in 3D, 219 domain wall, 149
critical isochore, 173 droplet, 40
critical isotherm, 144 Dulong-Petit, 123
critical nucleus, 41 dynamics
critical opalescense, 188 Langevin, 48
critical phase, 189
critical point, 34 effective field, 142
bicritical, 181 efficiencies
multicritical, 181 Carnot, 15
tetracritical, 181 elastic manifold, 78
tricritical, 181 endpoint
critical temperature critical, 32
BEC, 126 energy
magnetic systems, 148 free, 26, 58
crystallization, 194 internal, 11
cumulant, 97 ensemble
cumulant expansion, 254 average, 46
cumulants, 141, 267 canonical, 48, 56, 71
cyclotron frequency, 109 grand canonical, 48, 71
microcanonical, 47, 48, 50, 71
data collaps, 231 entropie
Debye temperature, 123 Sackur-Tetrode, 54
degree of freedom, 55 surface, 39
density entropy
canonical, 56 definition, 15
grand canonical, 59, 60 microcanonical, 50
microcanonical, 50 entropy jump, 192
density function equation of state, 4, 5
INDEX 291

caloric, 22, 54, 60 critical, 242, 250

thermal, 22, 55, 60 Gaussian, 271
van der Waals gas, 165 isolated, 250
equilibrium, 3 stable, 250
parameter, 31 trivial, 242
equilibrium parameter, 52 Wilson-Fisher, 271
equilibrium state fixed points, 241, 256
stability, 23 fixed-point
equipartition theorem, 55 critical, 285
ergodicity, 47, 49 line of, 286
Euclidean action, 203 flow
Euler’s theorem, 28 RG, 250, 256
exchange interaction, 135 fluctuation-dissipation theorem, 65,
expansion 157, 211
cumulant, 97 fluctuations, 133, 148
irreversible, 8 density, 64
reversible, 8 energy, 61
expansion coefficient, 10 particle number, 61, 64
exponent thermal, 42
gap, 239 flux quantum, 107
exponents Fourier transform, 207
critical, 183 free energy, 58
of vortices, 162
Fermi gas scaling, 239
degenerate, 101 frustration, 135
dilute, 101 fugacity, 59
equation of state, 101, 103 function
Fermi energy, 102 homogeneous, first order, 28
Fermi wave vector, 102 homogeneous, zero order, 29
in a magnetic field, 106 functional derivative, 141, 212
in a metal, 103
in a white dwarf, 105 gap exponents, 239
in neutron stars, 105 gas
internal energy, 103 classical, 80
Sommerfeld expansion, 100 classical limit, 83
specific heat, 103 density parameter, 100
Fermi Liquid Theory, 79 oversaturated, 168
ferromagnetic coupling, 134 quantum, 81
Feshbach resonances, 79 real, 165
Feynman diagrams, 267 undercooled, 37
field rescaling, 260 gas constant, 5
First Law, 11 gas–liquid transition, 186
fixed point Gase
attractive, 250 Boltzmannzählen, 80
292 INDEX

Bosegase configurational, 74
BE-Kondensation, 125 integrating factor, 6
gauge invariance, 106 interaction
Gauss integral, 206 statistical, 85
matrix, 206 inversion, 7
Gauss model, 183, 204 involutory, 25
Gaussian approximation, 212 irrelevant variable, 239
Gaussian fixed point, 271 dangerous, 272
generating function, 97 Ising model, 136
Gibbs surface, 32 RG, 249
Gibbs’ paradoxon, 50 isochore
Gibbs’ phase rule, 31 critical, 173
Gibbs-Duhem, 28 isotherm
Ginzburg criterion, 183, 221, 226 van der Waals, 166
Ginzburg temperature, 220
jumps, 34
glass, 47
Goldstone bosons, 157 Kadanoff
Goldstone modes, 133 block spins, 235
grand potential, 60 kink, 25
Grassmann variables, 94
group Lambda transition, 187
semi-, 240 Landau
Fermi Liquid Theory, 79
harmonic approximation, 96 Landau diamagnetism, 112
Hartree approximation, 218 Landau expansion, 181
heat, 11 Landau levels, 110
latent, 33 Landau-Ginzburg model, 263
specific, 18, 54 Langevin dynamics, 48
at constant p, 12 Larmor radius, 110
at constant V , 12 latent heat, 33
Heisenberg model, 134 lattice gas model, 136, 177
Helium layered system, 287
nucleation, 42 Lee-Yang
Helium mixture, 198 circle theorem, 179
Hilbert space, 67 Legendre transformation, 25
Hohenberg-Mermin-Wagner, 133, 151, Lifshitz transition, 201
191 linear response, 4, 48
homogeneous function, 233 linear response theory, 211
hysteresis, 193 Liouville theorem, 69
Liouville’s theorem, 46
imaginary time, 90 liquid, 30
information, 25 overheated, 37
integrability, 6, 21, 26, 28 superheated, 168
integral liquid–crystal transition, 194
INDEX 293

liquid–gas transition membrane

coexistence region, 173 elastic, 42
compressibility, 175 Mexican hat, 191
critical isotherm, 172 microscopic scale
Maxwell construction, 172 relevance, 228
specific heat, 175 microscopic theory, 3
transition line, 173 modes
volumes, 173 fast and slow, 261
liquid-crystal, 193 mol, 10
long-range order, 30 moment, 97
longitudinal suceptibility, 189 moulated phases, 201
loop diagrams, 216 MQP
low-energy physics, 183 macroscopic quantum phenomena,
lower critical dimension, 219 83
multicritical points, 200
magnetic field, 4
electrons, 106 Nelson-Kosterlitz jump, 287
magnetism nematic, 193
diamagnetic, 107 non-equilibrium physics, 48
paramagnetic, 107 non-linear sigma model, 163
magnetization, 4, 107 normalization
magnetization density, 136 grand canonical density, 59
majority rule, 253 nucleation, 40
manifold nukleus
critical, 242, 244, 286 critical, 41
elastic, 78
observable, 45, 67
marginal variable, 239
one-loop approximation, 216
matrix
Onsager, 151, 253
non-symmetric, 243, 245 operator
symmetric, 243 irrelevant, 244
Maxwell marginal, 244
construction, 167 relevant, 244
Maxwell Boltzmann distribution, 101 optical lattices, 79
Maxwell construction, 36 order
Maxwell relation, 26, 28 long-range, 30
mean field theory, 133 order parameter, 140
mean value order parameter correlator
over time, 45 scaling, 227
mean-field approximation, 142 order-parameter correlator, 211
mean-field equation, 143 Ornstein-Zernicke, 188
mean-field exponents overheated liquid, 193
table, 197
mean-field theory Pade, 226
real gas, 169 paradoxon
294 INDEX

Gibbs’, 50 pressure, 4
paramagnetismc, 107 primitve, 5
parameter probability
equilibrium, 52 a priori, 47, 69
particle number, 4 process
partition function adiabatic, 8
canonical, 57 adjabatic, 13
grand canonical, 60 cyclic, 11
path integral, 207 isobaric, 8
free particle, 90 isochoric, 8
particle in a potential, 92 isothermal, 8
Pauli reversible, 7
repulsion, 30 stochastic, 48
Pauli exclusion, 81 prozess
Pauli spin susceptibility, 115 irreversible, 8
Peierls, 152 pseudopotential, 86
percolation, 221
perturbation theory in u, 226 quantum gas
phase coexistence, 32 bosons, 81
phase diagram, 30, 32 fermions, 81
phase line, 35 partition sum, 81
phase rule quantum Hall effect, 110
Gibbs’, 31 quantum mechanics, 48
phase space, 45, 49 quantum simulation, 79
Γ-space, 49
µ-space, 74 random phase approximation, 218
average, 46 random phases, 68
probability, 46 Rayleigh-Jeans radiation law, 120
phase transition, 30 recursion
gas-liquid, 66 RG, 240
order, 179 relation
phenomenological theory, 3 Maxwell, 26, 28
photons, 117 thermal–caloric, 17
Planck relevant variable, 239
radiation law, 121 renormalization
Planck-Nernst, 17 in ~k-space, 260
polarizability, 282 renormalization group, 183
potential rescaling, 260
chemical, 20, 65 reservoir, 48, 67
grand, 60 response, 48
potentials linear, 4
convex, 168 resummation methods, 226
thermodynamic, 21 RG equations
power counting, 265, 266 differential, 245
INDEX 295

RG flow, 250, 256 state function, 5

RG recursion, 240 state space, 5, 50
rigitity, 30 state variable, 4
ruled surfaces, 39 state variables, 5
running coupling, 263 state vector, 67
states
Sackur-Tetrode formula, 54 corresponding, 166
saddle point, 212 Stefan-Boltzmann law, 119
scaling Stirling’s formula, 54
free energy, 239 suceptibility sum-rule
specific heat, 124 static, 211
tricritical, 197 superconducting transition, 187
scaling functions, 231 superconductor, 149
scaling laws, 183, 225, 229 superposition
scattering incoherent, 69
s-wave, 86 surface
screened interaction, 282 Gibbs, 32
Second Law surface tension, 41
Clausius, 14 susceptibility, 137, 140
Kelvin, 14 Curie, 115
self-averaging, 47 Landau, 112
self-consistent field approximation, 218 Pauli, 115
self-consistent ha, 96 symmetries
semi-group, 240 in Landau expansion, 181
separatrix, 286 Landau theory, 184
shear, 30 symmetry
Simplex, 39 particle–hole, 103
solid, 30 time reversal, 16
specific heat, 54
Debye, 124 Tc reduction, 218
delta-function, 193 temperature, 4
Dulong-Petit, 123 tension coefficient, 10
Gauss model, 209 tetracritical points, 200
sphere theorem
surface, 53 equipartition, 55
spin configurations, 159 Euler, 28
spin glass, 135 fluctuation-dissipation-, 65
spin waves, 281 Liouville, 46
spinodal, 37 multinomial, 80
spinodal temperatures, 191 virial, 56
spontaneous magnetization, 144 thermodynamic laws
spontaneously broken symmetry, 139 first, 11
stability second, 14, 16
of the equilibrium state, 23 third, 17, 72
296 INDEX

thermodynamic limit, 34, 178 pressure, 9

thermodynamic potentials, 4 substance, 10
enthalpy, 27 temperature, 9
entropy, 21, 53 universal jump, 287
free energy, 26 universality, 136, 239, 242, 245
Gibbs free energy, 27 upper critical dimension, 220, 221
grand potential, 28
internal energy, 22, 53 Van der Waals
Third Law, 17, 72 interaction, 30
time van der Waals equation
imaginary, 90 universal form, 166, 171
time average, 45 van der Waals gas, 136, 165
topological excitations, 133, 281 corresponding states, 166
trajectory entropy, 170
in phase space, 45 equation of state, 165
transfer matrix, 150, 248 free energy, 170
transformation Gibbs free energy, 171
adjabatic, 13 internal energy, 171
Legendre, 25 Maxwell construction, 167
transition mean-field theory, 169
continuous, 34 oversaturated, 168
first-order, 34 partition function, 170
infinite-order, 34 superheated, 168
second-order, 34 thermal equation, 170
sharp, 34 variable
smooth, 34 conjugate, 25
transition line extensive, 4, 21
BEC, 128 intensive, 4
transport, 48 irrelevant, 239, 244
transverse fluctuations, 191 dangerous, 272
transverse suceptibility, 189 marginal, 239, 244
triangular lattice, 252 natural, 26
tricritical point, 195 relevant, 239, 244
tricritical points, 200 state-, 5
tricritical scaling, 197 variational ansatz, 95
triple point, 32, 35 virial
Trotter theorem, 56
product formula, 91 volume, 4
vortex, 65, 159
ultra-violett cutoff, 208 vortex–anti-vortex pair, 160
undercooled gas, 193 vortices, 281
units unbinding, 286
heat, 10
mol, 10 Weiss field, 142
INDEX 297

Wenzel-Kramers-Brillouin, 76
Wick rotation, 90
Widom scaling, 231
Wien
displacement law, 120
Wilson-Fisher fixed point, 271
winding number, 159
WKB, 76
work, 11

XY model, 136