(Daniel I. Khomskii) Basic Aspects of The Quantum

This page intentionally left blank
BASIC ASPECTS OF THE QUANTUM THEORY OF SOLIDS

Order and Elementary Excitations
Aimed at graduate students and researchers, this book covers the key aspects of
the modern quantum theory of solids, including up-to-date ideas such as quantum
fluctuations and strong electron correlations. It presents the main concepts of the
modern quantum theory of solids, as well as a general description of the essential
theoretical methods required when working with these systems.
Diverse topics such as the general theory of phase transitions, harmonic and
anharmonic lattices, Bose condensation and superfluidity, modern aspects of magnetism including resonating valence bonds, electrons in metals, and strong electron
correlations, are treated using the unifying concepts of order and elementary excitations. The main theoretical tools used to treat these problems are introduced and
explained in a simple way, and their applications are demonstrated through concrete
examples.
daniel i. khomskii is a Professor in the II Physikalisches Institut, Cologne
University. His main research interests are the theory of systems with strongly
correlated electrons, metalinsulator transitions, magnetism, orbital ordering and
superconductivity.
BASIC ASPECTS OF THE QUANTUM

THEORY OF SOLIDS
Order and Elementary Excitations
DANIEL I. KHOMSKII
CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore,

So Paulo, Delhi, Dubai, Tokyo
Cambridge University Press
The Edinburgh Building, Cambridge CB2 8RU, UK
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521835213
D. Khomskii 2010
This publication is in copyright. Subject to statutory exception and to the
provision of relevant collective licensing agreements, no reproduction of any part
may take place without the written permission of Cambridge University Press.
First published in print format 2010
ISBN-13
978-0-511-78832-1
eBook (EBL)
ISBN-13
978-0-521-83521-3
Hardback
ISBN-13
978-0-521-54292-0
Paperback
Cambridge University Press has no responsibility for the persistence or accuracy

of urls for external or third-party internet websites referred to in this publication,
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
Contents
Foreword and general introduction

page ix
1 Some basic notions of classical and quantum statistical physics
1
1.1 Gibbs distribution function and partition function
1
1.2 Thermodynamic functions
2
1.3 Systems with variable number of particles; grand partition
function
4
2 General theory of phase transitions
6
2.1 Second-order phase transitions (Landau theory)
6
2.2 (Weak) First-order phase transitions
11
2.2.1 Another possibility of getting a first-order phase
transition
13
2.3 Interaction with other degrees of freedom
14
2.4 Inhomogeneous situations (GinzburgLandau theory)
16
2.5 Fluctuations at the second-order phase transitions
19
2.5.1 Critical indices and scaling relations
21
2.6 Quantum phase transitions
23
2.7 General considerations
25
2.7.1 Different types of order parameters
25
2.7.2 General principle
25
2.7.3 Broken symmetry and driving force of phase transitions
26
2.7.4 The Goldstone theorem
27
2.7.5 Critical points
28
3 Bose and Fermi statistics
31
4 Phonons in crystals
34
4.1 Harmonic oscillator
34
4.2 Second quantization
35
4.3 Physical properties of crystals in the harmonic approximation
38
vi
Contents
4.4 Anharmonic effects

4.4.1 Thermal expansion
4.4.2 Melting
4.4.3 Another approach to melting. Quantum melting
4.4.4 Low-dimensional solids; why is our world
three-dimensional?
General Bose systems; Bose condensation
5.1 Bose condensation
5.2 Weakly interacting Bose gas
5.3 Bose condensation and superfluidity
5.3.1 Landau criterion of superfluidity
5.3.2 Vortices in a superfluid
Magnetism
6.1 Basic notions; different types of magnetic response
6.1.1 Susceptibility of noninteracting spins
6.2 Interacting localized moments; magnetic ordering
6.2.1 Mean field approximation
6.2.2 Landau theory for ferromagnets
6.2.3 Antiferromagnetic interactions
6.2.4 General case
6.3 Quantum effects: magnons, or spin waves
6.3.1 Magnons in ferromagnets
6.3.2 Antiferromagnetic magnons. Zero-point oscillations
and their role
6.4 Some magnetic models
6.4.1 One-dimensional models
6.4.2 Resonating valence bonds, spinons and holons
6.4.3 Two-dimensional models
6.5 Defects and localized states in magnetic and other systems
Electrons in metals
7.1 General properties of Fermi systems
7.1.1 Specific heat and susceptibility of free electrons
in metals
Interacting electrons. Green functions and Feynman diagrams
(methods of field theory in many-particle physics)
8.1 Introduction to field-theoretical methods in condensed
matter physics)
8.2 Representations in quantum mechanics
8.3 Green functions
8.4 Green functions of free (noninteracting) electrons
41
43
45
48
51
54
54
58
62
65
67
70
70
74
76
77
80
84
87
91
92
98
104
105
109
117
123
127
127
129
133
133
136
139
141
Contents
10
11
12
8.5 Spectral representation of Green functions

8.5.1 Physical meaning of the poles of G( p, )
8.5.2 Physical meaning of the spectral function A( p, )
8.6 Phonon Green functions
8.7 Diagram techniques
8.7.1 Dyson equations, self-energy and polarization
operators
8.7.2 Effective mass of the electron excitation
Electrons with Coulomb interaction
9.1 Dielectric function, screening: random phase approximation
9.2 Nesting and giant Kohn anomalies
9.3 Frequency-dependent dielectric function; dynamic effects
Fermi-liquid theory and its possible generalizations
10.1 The foundations of the Fermi-liquid theory
10.2 Non-Fermi-liquid states
10.2.1 Marginal Fermi liquid
10.2.2 Non-Fermi-liquid close to a quantum critical point
10.2.3 Microscopic mechanisms of non-Fermi-liquid
behaviour; Luttinger liquid
Instabilities and phase transitions in electronic systems
11.1 Peierls structural transition
11.1.1 Qualitative considerations
11.1.2 Peierls instability in the general case
11.1.3 Different theoretical ways to treat Peierls distortion
11.1.4 Peierls distortion and some of its physical
consequences in real systems
11.2 Spin-Peierls transition
11.3 Charge-density waves and structural transitions,
higher-dimensional systems
11.4 Excitonic insulators
11.5 Intermezzo: BCS theory of superconductivity
11.6 Spin-density waves
11.7 Different types of CDW and SDW
11.8 Weakly and strongly interacting fermions. Wigner
crystallization
Strongly correlated electrons
12.1 Hubbard model
12.2 Mott insulators
12.3 Magnetic ordering in Mott insulators
vii
143
144
146
147
149
153
156
159
159
166
169
175
175
183
183
184
186
188
188
188
190
192
198
202
206
207
212
216
220
222
229
230
230
234
viii
Contents
12.4 One-particle spectrum of strongly correlated systems

12.4.1 Aproximate treatment (Hubbard I decoupling)
12.4.2 Dealing with Hubbard bands. Spectral weight
transfer
12.4.3 Motion of electrons and holes in an
antiferromagnetic background
12.5 Ferromagnetism in the Hubbard model?
12.6 Phase diagram of the Hubbard model
12.7 Phase separation
12.8 tJ model
12.9 Orbital ordering in the degenerate Hubbard model
12.10 Charge-transfer insulators
12.11 Insulatormetal transition
13 Magnetic impurities in metals, Kondo effect, heavy fermions
and mixed valence
13.1 Localized magnetic moments in metals
13.2 Kondo effect
13.3 Heavy fermion and mixed-valence systems
13.4 Kondo insulators
13.5 Ferromagnetic Kondo lattice and double exchange
mechanism of ferromagnetism
Bibliography
Index
235
236
238
239
244
244
247
251
252
258
265
272
272
276
282
288
290
296
298
Foreword
There are many good books describing the foundations and basics of solid state
physics, such as Introduction to Solid State Physics by C. Kittel (2004) or, on a
somewhat higher level, Solid State Physics by N. W. Ashcroft and N. D. Mermin
(1976). However there is a definite lack of books of a more advanced level which
would describe the modern problems of solid state physics (including some theoretical methods) on a level accessible for an average graduate student or a young
research worker, including experimentalists.
Usually there exists a rather wide gap between such books written for theoreticians and those for a wider audience. As a result many notions which are widely
used nowadays and which determine the face of modern solid state physics remain
hidden and are not even mentioned in the available literature for non-specialists.
The aim of the present book is to try to fill this gap by describing the basic
notions of present-day condensed matter physics in a way understandable for an
average physicist who is going to specialize in both experimental and theoretical
solid state physics, and more generally for everyone who is going to be introduced
to the exciting world of modern condensed matter physics a subject very much
alive and constantly producing new surprises.
In writing this book I tried to follow a unifying concept throughout. This concept,
which is explained in more detail below, may be formulated as the connection
between an order in a system and elementary excitations in it. These are the
notions which play a crucial role in condensed matter physics in general and in
solid state physics in particular. I hope that this general line will help the reader
to see different parts of condensed matter physics as different sides of a unified
picture and not as a collection of separate unrelated topics.
The plan of the book is the following. After discussing the general theory of
phase transitions (Chapter 2) which forms the basis for describing order in solids,
ix
I go step by step through different specific situations: systems of bosons (phonons

in crystals Chapter 4, and general Bose systems, including Bose condensation
and superfluidity Chapter 5). Then follows the important chapter on magnetism,
Chapter 6 (strictly speaking dealing neither with bosons, nor with fermions), and
after that we switch to the discussions of fermions electrons in solids, Chapters 7
13. In each topic I have tried to follow the general line which I have already
described above: to discuss first the type of order we have in one or the other
situation, then introduce different types of elementary excitations in them, first
independent excitations, but then paying most attention to the interaction between
them and to their quantum nature. Thus altogether the material presented in the
book is supposed to cover the main situations met in solids.
The theoretical methods used to describe these phenomena are introduced not
so much separately, as such, but in the appropriate places where they are needed,
and in a way which immediately shows how they work in specific problems. Thus,
in studying Bose systems I introduce the widely used Bogolyubov canonical transformation, which later on is also used for treating magnons in antiferromagnets and
for certain problems for electrons. Discussing spin waves, I introduce the method
of equations of motion with corresponding decoupling, later on also used, e.g. for
studying correlated electrons (the Hubbard model). When going to electron systems, I describe the Green function method and the Feynman diagram technique
without complete and rigorous derivations, but with the aim of demonstrating how
these methods really work in different situations.
I hope the material covered in this book will give the reader a relatively complete
picture of the main phenomena in modern solid state physics and of the main
theoretical methods used to study them. But of course it is impossible to cover in
one book of modest size this whole field. The most important and evident omissions
are:
I do not practically touch on the broad and important field of transport phenomena (resistivity, thermal conductivity, thermopower, the Hall effect, etc.) This is a
very big topic in itself, but it lies somewhat outside the main scope of this book. I
also do not discuss specific features of such important, but well-known materials
as semiconductors, ferroelectrics, etc. Also the wide field of superconductivity is
touched upon only to the extent it is required to illustrate the general treatment.
Yet another relatively recent and very beautiful topic is missing the phenomenon of the quantum Hall effect. Hopefully I can repair this omission later.
On the theoretical side probably two important methods are not sufficiently
discussed in the book. One is the renormalization group method used to treat
complicated situations with strong interaction. I only briefly mention this method,
but do not describe it in detail. Interested readers may find its description, e.g. in
the books by Chaikin and Lubensky (2000) or Stanley (1987).
xi
Another theoretical technique widely used nowadays is the use of different types
of numerical calculations. This is a very broad and rapidly developing field which
proved its efficiency for studying real materials and for theoretical investigations
of many situations not accessible to analytical calculations. This is quite a special
field, and it requires special treatment although when appropriate I present some
of the results obtained in this way.
With all these omissions, I still hope that the material which is included will be
useful for a broad audience and will give the reader a relatively complete picture
of the main phenomena and main problems in modern solid state physics.
A few words about the style of the presentation. This book has grown out of
a lecture course which I gave for several years at Groningen University and at
Cologne University. Therefore it still has some features of the lecture notes. I
present in it all the main ideas, but often not the full derivations of corresponding
results. This is also caused by the fact that the material touched upon in this book
in fact covers a huge field, and it is impossible to present all the details in one
book. There are many monographs and textbooks discussing in detail the separate
subfields presented below. However I have tried to choose the topics and present
them in such a way that the general ideas underlying modern solid state physics
and the internal logic of this field become clear. For more detailed discussions of
particular problems and/or corresponding methods of their theoretical treatment
the readers should go to the specialized literature.
In accordance with this general concept of the book, I did not include in it a
special problems section. In some places, however, especially in the first part of
the book, I formulate parts of the material, as Problems. Those who want to get a
deeper understanding of the subject are recommended to stop reading the text at
these places and try to find the answers themselves; the material presented before
usually makes this task not too difficult. The answers, however, are usually given
right after the problems, so that readers can also go on along the text if they do
not have a desire, or time, to do these exercises themselves. Actually most of the
problems, with their answers, form an integral part of the text.
In several places in the text I have also put some more special parts of the text
in smaller type. These parts usually relate to more specialized (although useful)
material.
In addition to the main material I have also included three very short chapters
(Chapters 1, 3 and 7) with a short summary of some of the basic facts from statistical mechanics. I think it would be useful for readers to have this information at
hand.
Some important notions are mentioned several times in different parts of the text.
I did this intentionally, so that different chapters would become somewhat more
independent although of course there are a lot of cross-references in the text.
xii
I hope that the book gives a coherent presentation of the main ideas and methods
of the quantum theory of solids.
There are many good books which cover parts of the material contained in the
present book (and actually much more!). I can recommend the following, already
classical books:
1. J. M. Ziman (1979), Principles of the Theory of Solids. A very good and clear
book covering the main topics in solid state physics. Highly recommended.
However, it does not contain more modern methods.
2. N. W. Ashcroft and N. D. Mermin (1976), Solid State Physics. Also a very good
and widely used book, covering the topics in more detail, on a somewhat more
elementary level. Very transparent and useful.
3. M. P. Marder (2000), Condensed Matter Physics. A rather complete book
describing the main phenomena in solid state physics, but not going into much
theoretical detail.
4. C. Kittel (1987), Quantum Theory of Solids. Contains detailed discussion of
many problems in quantum theory, using more modern methods such as diagram
techniques. Somewhat more theoretical.
5. G. D. Mahan (2000), Many-Particle Physics. Gives a very complete treatment
of the topics discussed; it is a kind of encyclopedia. It uses the Green function
method all the way through. Very useful for theoreticians, and contains all the
necessary details and derivations, etc. However not all topics are discussed there.
6. L. D. Landau and I. M. Lifshits, Course of Theoretical Physics, especially
Statistical Physics (old one-volume edition 1969, or new edition v. I 1980), and
Quantum Mechanics (1977). These classical books contain virtually all the basic
material necessary, and many particular topics important for our problems. If
one can call the book by Mahan an encyclopedia, then the course of Landau and
Lifshits is a bible for all theoreticians, especially those working in condensed
matter physics. But these books are very useful not just for theoreticians, but
for everyone looking for clear and precise description of all the basic ideas of
theoretical physics.
7. J. R. Schrieffer (1999), Theory of Superconductivity. A very clear book; contains
in particular a very good and condensed treatment of the Green function method
and diagram technique, in a form used now by most theoreticians.
8. A. A. Abrikosov, L. P. Gorkov and E. Dzyaloshinsky (1975), Methods of the
Quantum Field Theory in Statistical Physics. One of the first (and still the best)
books on the modern methods applied to condensed matter physics. It gives
a very detailed treatment of the theoretical methods and a good discussion of
specific problems (Fermi and Bose liquids; plasma; electronphonon interaction
and the basics of the theoretical treatment of superconductivity).
xiii
9. P. M. Chaikin and T. C. Lubensky (2000), Principles of Condensed Matter

Physics. A very good book containing in particular detailed discussion of different questions connected with phase transitions. The accent is on general statistical mechanics; specifically it contains a lot of material on soft condensed matter
physics, but does not discuss such topics as electrons in metals, magnetism, etc.
Some other references will be given later, in the main body of the book. But,
keeping in mind the character of the book (which is practically expanded lecture
notes and which still retain that character), I deliberately refrained from including
too many references it would be simply impossible to cite all the relevant works.
Therefore I mostly refer not to original publications but rather to monographs and
review papers. Interested readers may find more detailed information on particular
subjects in these references.
General introduction
The unifying concept in this book is the concept of order and elementary excitations;
these are the key words.
One can argue as follows. In general in macroscopic systems with many degrees
of freedom the internal state, or internal motion on the microscopic scale, is random.
However as T 0 the entropy of the system should go to zero, S 0; this is the
well-known Nernst theorem, or the third law of thermodynamics. Accordingly, at
T = 0 there should exist perfect order of some kind.
Such ordering sets in at a certain characteristic temperature T , often with a
phase transition, but not necessarily.
At T T we can describe the state of the system as predominantly ordered,
or maybe in a weakly excited state. Such relatively weakly excited states will be
thermally excited, but can appear also due to small external perturbations. Usually
in such a state we can speak of a small number of elementary excitations, or
quasiparticles. Examples are, e.g. phonons in crystals, magnons or spin waves in
ferromagnets, excitons in semiconductors, etc.
Sometimes such elementary excitations are rather strange: instead of electrons
they may be excitations with spin, but no charge (spinons), or vice versa (holons).
There exist also topological excitations (solitons, vortices, etc.).
In a first approximation we can consider these excitations as noninteracting.
Such are, e.g. phonons in a harmonic crystal, etc. However in the next step we have
to include in general an interaction between quasiparticles.
There may exist interactions between the same quasiparticles. They lead, e.g. to
anharmonic effects in crystals (phononphonon interactions); they are included in
the Landau Fermi-liquid theory, and give rise to screening for electrons in metals;
the magnonmagnon interaction can lead, e.g. to the formation of bound states
xiv
of magnons, etc. There also exist interactions between different quasiparticles:

electronphonon interactions, pairing in conventional superconductors, and interactions between many other elementary excitations. Often due to these interactions
the properties of quasiparticles are strongly changed, or renormalized: an example is the formation of polarons (electron + strong lattice distortion). Also new
quasiparticles may be formed (plasmons due to the Coulomb interaction of electrons; excitons bound states of electrons and holes in semiconductors). Even the
ground state itself, the very type or ordering, may change because of such interactions. An example is the superconducting state instead of the normal state of a
metal.
It is important that these quasiparticles, or elementary excitations, are quantum
objects. Consequently, one should not visualize the order as completely classical:
there are quantum fluctuations (zero-point motion, or zero-point oscillations) even
at T = 0. Sometimes they lead only to minor numerical changes, but there are cases,
especially in low-dimensional or frustrated systems, when they can completely
modify the properties of a system, e.g. destroying the long-range order totally.
They can also modify the properties of the phase transitions themselves, e.g.
leading to quantum phase transitions. Thus the classical picture is always very
useful, but one should be cautious and aware of its possible failures in some cases
but very often these cases are the most interesting!
In treating these problems a lot of different approaches were used, and different
methods developed. These methods are often used not only in solid state physics or
condensed matter physics in general; many of them are also widely used (and often
have been developed!) in other parts of physics: in elementary particle physics, in
field theory, and in nuclear physics. Methods such as the Green function method
and Feynman diagrams were introduced in field theory, but are now widely used in
condensed matter physics. On the other hand, some methods and concepts which
first appeared in solid state physics (the mean field, or self-consistent field method,
the concept of a phase transition) are now used in nuclear physics, in elementary
particle physics (e.g. quarkgluon plasma), and even in cosmology.
In this book I will try to describe the main concepts and ideas used in modern
many-particle physics, and the methods used to study these problems, such as
the equation of motion method, canonical transformations, diagram techniques,
and the Green function method. Once again, the key words will be elementary
excitations, in connection with order. The illustrations will be predominantly given
for the examples of solid state systems, although, as I said before, many concepts,
notions and methods are also applicable in different fields, and even not only in
physics! Thus, some of the ideas of many-particle physics are now widely used for
treating problems in biology, and even economics and sociology. I hope that such
a general view will help readers to form a unified concept of the main phenomena
in condensed matter physics and related fields.
1
Some basic notions of classical and
quantum statistical physics
1.1 Gibbs distribution function and partition function

In this short chapter some of the basic notions from thermodynamics and statistical
physics are summarized.
The probability to observe a state |n with energy En is
wn = AeEn /T ;
(1.1)
this is called the Gibbs distribution. (Here and below we put the Boltzmann constant
kB = 1, i.e. the temperature is measured in units of energy, and vice versa.) The
normalization constant A is determined by the condition that the total sum of
probabilities of all states is 1:
wn = 1 ,
(1.2)
from which we find

1
=
eEn /T Z .
A
n
(1.3)
Here Z is the partition function

Z=

eEn /T = Tr eH/T ,
(1.4)
where H is the Hamiltonian of the system. Thus

wn =
eEn /T
.
Z
1
(1.5)
Basic notions
The entropy is defined as

S = ln wn =
ln wn eEn /T
Z
(1.6)
( . . . is the symbol for the average). When we put (1.5) into (1.6), we obtain
S = ln Z +
E
,
T
where E is the average energy of the system, E =

duce the quantity
(1.7)
1
Z

n
En eEn /T . We can intro-
F = E T S = T ln Z ,
(1.8)
which is called the (Helmholtz) free energy:

F = T ln Z = T ln
eEn /T .
(1.9)

The Helmholtz free energy, F , is a function of the temperature T and of the density
n = N/V , or of the volume: F = F (V , T ). One can also introduce other so-called
thermodynamic potentials, expressed as functions of different variables. These are:
At fixed pressure and temperature the Gibbs free energy
(P , T ) = E T S + P V = F + P V .
(1.10)
If instead of the temperature T we chose as free variable its conjugate, the entropy,
then we obtain the enthalpy
W (P , S) = E + P V .
(1.11)
Enthalpy is often used in discussions of chemical reactions, thermodynamics of

formation of different phases, etc.
The energy itself is also one of the thermodynamic potentials; it is a function of
volume and entropy, E(V , S).
Similar to mechanics, where the system at equilibrium tends to a state with
mimimum energy, many-particle systems at finite temperature tend to minimize
the free energy, i.e. the corresponding thermodynamic potential F or .
From these definitions it is clear that, e.g.
dF = S dT P dV ,
(1.12)
from which we obtain

F
S=
,
T V

F
P =
.
V T

(1.13)
(1.14)
Similarly
d = S dT + V dP ,
(1.15)

S=
,
T P

.
V =
P T

(1.16)
(1.17)
Other useful thermodynamic quantities are, e.g. the specific heat at constant volume,
cV , and at constant pressure, cP :

E
S
=T
,
(1.18)
cV =
T V
T V

cP =
W
T

=T
S
T

.
(1.19)
One can express cP , cV through F , , using (1.12), (1.15).

Using the expressions given above, one can obtain useful relations between
different thermodynamic quantities, e.g. between the specific heat, the thermal
expansion coefficient (the volume coefficient of the thermal expansion = 3,
where is the linear thermal expansion)
=+
1 V
,
V T
(1.20)
and the compressibility

1 V
.
(1.21)
V P
The resulting connection has the form (see, e.g. Landau and Lifshits 1980,
Section 16):
=
cP cV = T
(V /T )P2
2
= VT
.
(V /P )T
(1.22)
Similarly one can also find relations between other thermodynamic quantities;
some examples will be given below, especially in Chapter 2.
Basic notions
1.3 Systems with variable number of particles; grand partition function

One can also introduce thermodynamic quantities for systems with variable number
of particles N. The thermodynamic potentials introduced above depend on the
particle density N/V , i.e.

V
,T
F = N f1
N
= N f2 (P , T )

S V
.
E = N f3
,
N N
(1.23)
From these equations we get:

dF = S dT P dV + dN
d = S dT + V dP + dN
dE =
(1.24)
T dS P dV + dN .
Here we have introduced the chemical potential which is defined by

F

E
=
=
.
=
N S,V
N T ,N
N P ,T
(1.25)
From (1.25) and (1.23) we obtain

,
(1.26)
N
i.e. the chemical potential is the Gibbs free energy per particle.
One important remark is relevant here. If the number of (quasi)particles N is not
conserved, such as for example the number of phonons in a crystal, then the value
of N is determined by the condition of minimization of the free energy in N, e.g.
F /N = 0, etc. One sees then that in such cases the chemical potential is = 0.
This fact will be used in several places later on.
The chemical potential and the number of particles N are conjugate variables
(like T and S; P and V ). One can introduce a new thermodynamic potential with
as a variable; it is usually denoted (V , T , ). Using equations (1.3), (1.8) we
can write down the distribution function (1.1) as

eEn /T
F En
.
(1.27)
= exp
wn =
Z
T
=
For a variable particle number N, it takes the form

+ N EnN
wnN = exp
,
T
(1.28)
1.3 Systems with variable number of particles
where we have used this new thermodynamic potential , instead of the free energy:
(V , T , ) = F N .
(1.29)
Thus is a generalization of the free energy to the case of variable number of

particles. Similar to (1.24), we have:
d = S dT P dV N d ,
(1.30)
i.e. the total number of particles is connected to the chemical potential by the
relation

N =
.
(1.31)
T ,V
Problem: One can show that = P V ; try to prove this.
Solution: From (1.29) = F N. But, by (1.26), N = , and, by (1.10),
= F + P V . Thus = F N = F = P V .
Analogously to (1.5), (1.9), we can write down
= T ln ZGr ,
where ZGr is called grand partition function:

eEnN /T .
eN/T
ZGr =
N
(1.32)
(1.33)
2
General theory of phase transitions
The state of different condensed matter systems is characrerized by different quantities: density, symmetry of a crystal, magnetization, electric polarization, etc.
Many such states can have a certain ordering. Different types of ordering can be
characterized by order parameters.
Examples of order parameters are, for instance: for ferromagnets the magnetization M; for ferroelectrics the polarization P ; for structural phase transitions
the distortion u , etc. Typically the system is disordered at high temperatures, and
certain types of ordering may appear with decreasing temperature. This is clear
already from the general expressions for thermodynamic functions, see Chapter 1: at
finite temperatures the state of the system is chosen by the condition of the minimum
of the corresponding thermodynamic potential, the Helmholtz free energy (1.8) or
the Gibbs free energy (1.10), and from those expressions it is clear that with
increasing temperature it is favourable to have the highest entropy possible, i.e.
a disordered state. But some types of ordering are usually established at lower
temperatures, where the entropy does not play such an important role, and the
minimum of the energy is reached by establishing that ordering.
The general order parameter depends on temperature, and in principle also
on other external parameters pressure, magnetic field, etc. Typical cases of the
dependence of the order parameter on temperature are shown in Fig. 2.1. The
situation shown in Fig. 2.1(a), where the order parameter changes continuously,
is called a second-order phase transition, and that shown in Fig. 2.1(b), where
changes in a jump-like fashion, is a first-order phase transition. The temperature Tc
below which there exists order in a system (
= 0) is called the critical temperature
(sometimes the Curie temperature, the notion coming from the field of magnetism).
2.1 Second-order phase transitions (Landau theory)
For the second-order phase transitions close to Tc the order parameter is small,
and we can expand the (Gibbs) free energy (P , T , ) in a Taylor series. This
6
2.1 Second-order phase transitions
(a)
Tc
(b)
Tc
Fig. 2.1
approach was first developed by Landau, and in this section we largely follow the
classical presentation of Landau and Lifshits (1980).
The expansion of the free energy in small is, in general,
= 0 + + A2 + C3 + B4 + .
(2.1)
(It will be clear below why we have chosen such an unnatural notation with the
sequence of coefficients A, C, B.) As mentioned above, the state of the system,
in particular the value of the order parameter (magnetization, or spontaneous
polarization, or distortion, etc.) is determined by the condition that the free energy,
in this case , has a minimum. The coefficients , A, C, B are functions of P , T
such that the minimum of (P , T , ) as a function of should correspond to
= 0 above Tc (disordered state), and to
= 0 (and small) below Tc . From this
requirement it is clear that the coefficient in a system without external fields
should be = 0, otherwise
= 0 at all temperatures: in the presence of the linear
term in (2.1) the free energy would never have a minimum at = 0, which should
be the case in a disordered system at T > Tc .
The same requirement that = 0 above Tc , but
= 0 for T < Tc , leads to the
requirement that the first nonzero term A2 in the expansion (2.1) should obey the
condition
A(P , T ) > 0 for T > Tc
A(P , T ) < 0 for T < Tc .
(2.2)
As a result the dependence of () would have the form shown in Fig. 2.2.
Thus at the critical temperature Tc the coefficient A(P , T ) should pass through
zero and change sign. (We assume that it changes continuously with temperature.
We also assume that the other coefficients in equation (2.1) are such that C = 0,
which is often the case, see Section 2.2 below, and B > 0.) Again, making the
simplest assumption, we can write close to Tc :
A(P , T ) = a (T Tc ) ,
(2.3)
T > Tc
T < Tc
Fig. 2.2
~ Tc T
Tc
Fig. 2.3
with the coefficient a > 0. Then

= 0 + A2 + B4 = 0 + a (T Tc ) 2 + B4 .
(2.4)
The behaviour of (T ) can be easily found from (2.4) by minimizing the free
energy with respect to :

= 0 = 2A + 4B3 = 2a(T Tc ) + 4B3 = 0 ,
2 =
A
a
=
(Tc T ) .
2B
2B
(2.5)
(2.6)
This behaviour is shown in Fig. 2.3.

Here in principle all coefficients may be functions of pressure (or other external
variables), a = a(P ), B = B(P ), Tc = Tc (P ). But in practice the dependence of
Tc (P ) is the most important one; the coefficients a and B can usually be taken as
constants.
The equilibrium free energy itself at T < Tc is obtained by putting the equilibrium value of the order parameter (2.6) back into the free energy (2.4):
A2
a2
= 0
(Tc T )2
(2.7)
4B
4B
(and = 0 for T > Tc ). Thus (and other thermodynamic potentials e.g. the
Helmholtz free energy F if we work at fixed volume V and have a second-order
phase transition) are continuous, see Fig. 2.4(a). However the derivatives (/T ),
min = 0
2.1 Second-order phase transitions
cP
Tc
Tc
(a)
(b)
Fig. 2.4
etc. have kinks at Tc , and the second derivative would have jumps; this is typical
behaviour of thermodynamic functions at the second-order phase transitions in the
Landau theory.
Problem: Using the definition of specific heat cP (1.19), equations (1.16), (2.6),
(2.7), find the behaviour of specific heat at the second-order phase transition.
Solution: The entropy S, by (1.16) and (2.7), is
(T > Tc )
S
0
S=
=
S a 2 (T T ) (T < T )
T
0
c
2B c
and, by (1.19), the specific heat is

0
S
=
cP = T
T P
a 2 T /2B
(2.8)
(T > Tc )
(2.9)
(T < Tc ) .
Note that this expression is valid only close to Tc ; at lower temperatures the specific
heat may and will deviate from this simple linear behaviour.
At Tc the entropy has a kink, see (2.8), and there exists at Tc a jump in the
specific heat at the second-order phase transition:
cP =
a 2 Tc
.
2B
This behaviour is shown in Fig. 2.4(b).

The total entropy connected with the ordering is
Tc
1
Sord =
cP (T ) dT .
T
0
(2.10)
(2.11)
The experimental measurements of specific heat and of the total entropy of the
transition give very important information: the observation of the behaviour of
cP of the type shown in Fig. 2.4 proves that we are dealing with a second-order
10
phase transition (see however Section 2.5 later), and the measurement of the total
entropy of the transition (part of the total entropy, connected with the ordering)
tells us which degrees of freedom participate in ordering. Thus, e.g. if we have a
magnetic ordering of spins 12 , the total entropy of the transition in the ideal case
should be Stot = kB ln 2 (or kB ln(2S + 1) for spin S, where 2S + 1 is the number
of possible states of spin S in a fully disordered state, and this entropy has to be
removed in the ordered state at T = 0). If experimentally one finds Stot smaller
than this value, then this means that there is still a certain degree of ordering (or
short-range correlations) above Tc . If, however, one finds the value of Stot larger
than the expected one, one can conclude that some other degrees of freedom also
order at Tc , not only the ones initially assumed. This is an important test, often used
experimentally.
Problem: Find the connection between the specific heat jump
cP and other
properties of the solid (compressibility, thermal expansion).
Solution: By definition, second-order phase transitions are continuous, so that
along the transition line there is no jump in volume and in entropy,
V = 0,
S = 0
(these would be nonzero at the first-order phase transition). Let us differentiate these
relations along the curve Tc (P ): we thus obtain, e.g.

T
V
V
+
=0.
(2.12)
P T
P
T P
Tc
Remembering that the thermal expansion coefficient is = 3 =

, we can rewrite equation (2.12) as
compressibility = V1 V
P
= 3
dTc
dTc
.
dP
dP
1 dV
V dT
, and the
(2.13)
Similarly, from the condition

S = 0, we obtain:

T
S
S
+
=0.
P T
P
T P
(2.14)
Tc
As

S=
T

S
P

T
2
=
=
T P
T

,
(2.15)

P
V
=
T

,
P
(2.16)
this gives the relation
V
T

=
P
dTc
cP
dP T
11
(2.17)
(here we have used the expression (1.19) for the specific heat cP ). In effect we
obtain
3
=
=
dTc 1
cP .
dP V Tc
(2.18)
The relations (2.13), (2.18) are known as the Ehrenfest relations. They are the analogues for the second-order phase transition of the well-known ClausiusClapeyron
relations valid for first-order transitions, e.g. the relation between the jump in volume and the latent heat of transition
Q = T
S:
V =
Tc
1 dTc
S =
Q .
P
Tc dP
(2.19)
One can easily check that in the limit in which all the jumps at the first-order
transition go to zero, i.e. when the first-order phase transition goes over to the
second-order one, this expression gives equation (2.18): one can obtain this by
=
cp , and = 3 = V1 V
.
applying the operation V1 T to (2.19) and using (
Q)
T
T

Until now we have ignored the cubic term in the expansion (2.1). Very often it is
indeed zero, just by symmetry. Thus, in an isotropic ferromagnet the states with
positive and negative magnetization should be equivalent, which means that the
free energy may contain only terms even in M (which in this case is the order
parameter), i.e. the term C3 in this and similar cases should be absent. But there
may be other situations, in which such terms are allowed by symmetry and should
be present in the expansion (2.1).
Suppose now that the term C3 (cubic invariant) in the free energy (2.1) is
nonzero. One can easily analyse the resulting equation for , analogous to (2.5),
and find the properties of the solution. It is also very instructive just to look at the
dependence (, T ) in this case, which will immediately tell us what happens. The
corresponding set of curves () for different temperatures is shown in Fig. 2.5 for
the case C < 0 (the case C > 0 can be studied similarly). For high enough T , when
the coefficient A in (2.1) is large, we have only one minimum of (), at = 0
(curve 1 in Fig. 2.5.) With decreasing temperature and, consequently, decreasing
coefficient A, the set of curves () would look as shown in Fig. 2.5, curves 26.
At a certain temperature T there will appear a second minimum in (), curve 2.
12
~ T*
3
4
~ Tc(I)
~ T **
h
decreasing T
Fig. 2.5
T ** Tc(I)
T*
Fig. 2.6
Initially it is just a metastable state, the absolute minimum is still at = 0. But with
further decrease of temperature, at T < Tc(I) , this new state will have free energy
lower than the disordered state with = 0, curve 4. Note that this will happen
when the coefficient A is still positive. If we wait long enough, at this temperature
the system would jump from the state = 0 to a new minimum with nonzero , i.e.
we will have a first-order transition. The disordered state = 0 will still exist as a
local minimum, i.e. as a metastable (overcooled) state. At a still lower temperature
T the coefficient A in (2.1) will itself become negative, and the metastable state
= 0 will cease to exist. The temperatures T and T are the limits of overheating
and overcooling in such a first-order transition; they determine the maximum width
of hysteresis at such a transition. These points are called spinodal points, and if Tc(I)
changes, e.g. under pressure, they will form spinodal lines, or simply spinodals.
Thus the behaviour of (T ) has here the form shown in Fig. 2.6. We see that
the presence of cubic invariants in the expansion of the free energy (2.1) always
leads to first-order phase transitions. Note that here the real critical temperature, the
point Tc(I) in Fig. 2.6, is not a singular point of the free energy it is just the point
at which two minima of the free energy in Fig. 2.5 have equal depths. On the other
hand, the temperatures T and T (the limits of overheating and overcooling) are
13
Fig. 2.7
singular points, so that, e.g. the susceptibility would not diverge at Tc(I) , but it would
at T and T .
When do we meet such a situation? The general principle is the following: we
should write down in the free energy functional (2.1) all the terms permitted by
symmetry. Usually one has to check it using group theory. And if cubic invariants
are allowed by symmetry, then generally speaking they should be present in the
expansion (2.1), so that the corresponding transition should be first order.
There are many examples in which this is the case. One is the structural phase
transition or charge (or spin) density wave transition in a triangular or honeycomb lattice with the formation of three equivalent waves n = ei Q n r , with the
wavevectors Q 1 , Q 2 , Q 3 at 120 to one another, so that Q 1 + Q 2 + Q 3 = 0. As a
result there exists an invariant C1 2 3 = C3 ei( Q 1 + Q 2 + Q 3 )r = C3 , so that such a
phase transition should be first order. This is, for example, the situation in transition
metal dichalcogenides such as NbSe2 , TaS2 , etc.
One can also show in general that at crystallization (formation of a periodic
structure from a liquid) the situation is the same and consequently crystallization,
or melting, should be a first-order phase transition (Landau).
2.2.1 Another possibility of getting a first-order phase transition
Suppose the coefficient C in (2.1) is zero, and we have
(P , T ) = A(P , T )2 + B(P , T )4 .
(2.20)
We have assumed previously that the coefficients A(P , T ) = a(T Tc ) and

B(P , T ) > 0. But the coefficient B is itself a function of pressure, and in principle it may become negative. Then at T < Tc , instead of the behaviour shown in
Fig. 2.2, we would have the behaviour of the type shown in Fig. 2.7, i.e. the free
energy (2.20) would have no minima at finite . To stabilize the system we then
have to include higher order terms in the expansion for the free energy. Thus, e.g.
we should write
= A2 + B4 + D6
(2.21)
14
decreasing T
h
Fig. 2.8
T
T*
first order
second
order
P*
Fig. 2.9
with D > 0. Suppose that the coefficient B < 0, and that the temperature is slightly
above the original Tc (i.e. the coefficient A is positive but small). Then the behaviour
of at different temperatures would have the form shown in Fig. 2.8. When we
approach the old Tc from above, so that A = a(T Tc ) becomes small but is still
positive, the term B4 with B < 0 starts to be important at small , leading to the
situation with three minima of . That is, in this case the phase transition also
becomes first order. Thus at the point where B(P , T ) changes sign, a second-order
phase transition changes to a first-order one. (The point (T , P ) in the (T , P ) plane
is indeed a point: it is the simultaneous solution of two equations A(P , T ) = 0,
B(P , T ) = 0.) Such a point, at which a second-order phase transition changes to
a first-order one, is called a tricritical point. Close to it the phase diagram of the
system has the form shown in Fig. 2.9. Here we have marked the second-order
phase transition by a bold solid line and the first-order transition by a bold dashed
one; thin solid lines are spinodals the lines of overheating and overcooling.
The Landau method of treating phase transitions is very simple, but extremely
powerful. It permits us, e.g. to study the influence on the phase transition of the
15
interaction with other degrees of freedom. Thus, for instance, we can study phase
transitions (e.g. magnetic) in a compressible lattice. If u is the deformation, we can
write in general (this is a typical situation)
bu2
+ 2 u .
(2.22)
2
Here the third term is the elastic energy (b is the bulk modulus, i.e. the inverse
compressibility), and the last term is the coupling of our order parameter with
the deformation, for example the magnetoelastic coupling in the case of magnetic
phase transitions.
Minimizing (2.22) with respect to u, we obtain
= A2 + B4 +

= bu + 2 = 0 ,
u
u=
2
.
b
(2.23)
Now, we put it back into (2.22):

2 4 2 4
2
2
= A + B
4 .
= A + B +
2b
b
2b
2
(2.24)
Thus we see that if the coupling to the lattice is sufficiently strong, or if the lattice
compressibility is large (bulk modulus b small), the renormalized coefficient of the
4 term in (2.24) may become negative and this, according to Figs. 2.8 and 2.9,
makes the transition first order. This is a general rule: coupling to other degrees of
freedom gives a tendency to make a phase transition first order (although it may
remain second order if this coupling is not strong enough1 ).
Note that for the coupling included in equation (2.22) the resulting deforma
tion is u 2 , see (2.23). Thus, whereas below Tc we have Tc T (if the
transition remains second order), the corresponding distortion changes linearly,
u (Tc T ). In principle this effect can be measured directly. In particular, if
the corresponding distortion breaks inversion symmetry and leads to ferroelectricity (i.e. if the polarization P u), then the polarization will be proportional to
the square of the primary order parameter and close to Tc would also behave
as Tc T . Such systems are known as improper ferroelectrics, in contrast to the
ordinary (proper) ferroelectrics in which the polarization itself is the main, pri
mary order parameter, = P Tc T . A similar situation can exist also in
other systems with coupled order parameters. The resulting properties depend on
the specific type of coupling between such order parameters and : coupling of
the type of equation (2.22), 2 , or of the type 2 2 , etc. (The detailed type of
1
A more detailed treatment shows that the tendency to make a first-order transition due to coupling of the order
parameter to the lattice is actually much stronger than that obtained above, especially when we take into account
coupling to shear strain (Larkin and Pikin).
16
such coupling is determined by the symmetry properties of corresponding order

parameters.)
2.4 Inhomogeneous situations (GinzburgLandau theory)

Up to now we have considered only homogeneous solutions, i.e. situations in
which the order parameter e.g. the magnetization is taken as constant, the
same throughout the whole sample, independent of the position r. Often one has to
consider inhomogeneous situations, for example in the presence of external fields,
or close to the surface, etc. Then = (r), and the total free energy should be
written as an integral over the whole space, of the type

= d 3 r {free energy density (r)} .
(2.25)
Fig. 2.10
The free energy density (r) contains terms of the type (2.1), with the order
parameter (r) taken at a given point. But the variation of in space also costs
some energy. Thus, e.g. in a ferromagnet all spins prefer to be parallel, .
Variation in space means, e.g. the formation of structures of the type shown in
Fig. 2.10 spirals, or domain walls, etc. It is clear that the canting of neighbouring
spins costs a certain energy (the usual exchange interaction in a ferromagnet preferring to keep neighbouring spins parallel). For slow variation the cost in energy
should be proportional to the gradient of the order parameter, d/d r = . The
only invariant containing (for scalar ) is ()2 . Thus, in the same spirit as
before, we can write the simplest generalization of equation (2.1) in the form

2
= d 3 r A 2 (r) + B 4 (r) + G (r)
.
(2.26)
This is called a GinzburgLandau (GL) functional (sometimes also Ginzburg
LandauWilson (GLW) functional) the functional of the function (r). It is widely
used, e.g. in the theory of superconductivity (where it was actually introduced), in
discussion of domain walls in magnets, etc. Minimizing this functional in the order
parameter (r) (which in the theory of superconductivity is usually a complex scalar
denoted , and which can be viewed as a wavefunction of the superconducting
condensate) gives not the ordinary algebraic self-consistency equation (2.5), but,
due to the presence of the gradient term in (2.26), a differential equation the
famous GinzburgLandau equation. In the theory of superconductivity it is usually
2.4 Inhomogeneous situations
17
G>0
G<0
Fig. 2.11
written as

1
2e 2
i h
A + A + 2B||2 = 0 .
2m
c
(2.27)
(Here A is the vector potential, introduced to describe the behaviour of superconductors in a magnetic field.) We see that the GinzburgLandau equation (2.27) has
a form similar to the Schrodinger equation, but with a nonlinear term 3 . Close to
Tc , when the order parameter 0, this equation can be linearized, and then it is
indeed exactly equivalent to the corresponding Schrodinger equation. This analogy
is very useful for the treatment of many problems in superconductivity, such as the
upper critical magnetic field Hc2 , the formation and properties of vortices, etc.
Similar equations can also be written down for other physical systems. Thus, the
corresponding equation for ferromagnets (known as the LandauLifshits equation)
is widely used for treating the domain structure of ferromagnets, the dynamics of
domain walls, etc.
As we have discussed above, in the usual situation the homogeneous solutions
correspond to the minimum energy, which means that the coefficient G next to
the gradient term in equation (2.26) is positive. But what would happen if the
coefficient G(P , T ) becomes negative?
Let us suppose that G < 0. Then it is favourable to have
= 0, i.e. the
homogeneous solution becomes unstable, and there should occur a transition to an
inhomogeneous state. For instance, instead of a ferromagnet we may have
a transition to a spiral state. To find such a solution and to describe its properties we
have to minimize now with respect to . Again, if the term G()2 becomes
negative, we have to write down the next terms in the gradient (but still lowest order
in itself), e.g. E( 2 )2 with a positive coefficient E, to stabilize the system.
Thus we would have in the free energy the terms
G()2 + E( 2 )2 .
(2.28)
As a function of , has a form similar to the one shown in Fig. 2.2, see Fig. 2.11.
It is convenient to go to the momentum representation: ()2 q 2 2 . Then
18
T
disordered phase
homogeneous
phase
Lifshits point
*
inhomogeneous
phase
certain parameter
Fig. 2.12
would contain the terms

Gq 2 2 + Eq 4 2 .
(2.29)
Now we can find the value of q which minimizes :

= G2 + 2Eq 2 2 = 0 ,
q 2
(2.30)
G
2E
(2.31)
2
Q2 = qmin
=
(G < 0) .
That is, in this case the structure with the wavevector Q, or with the period
l = 2/Q, will be formed. In general the period of this superstructure (lattice,
or magnetic, etc.) is incommensurate with respect to the underlying lattice period.
The point where the coefficient G changes sign and at which there is a change
from homogeneous to inhomogeneous ordering is called the Lifshits point. The
typical phase diagram in this case looks like Fig. 2.12. Here the solid lines are
second-order phase transitions, and the dashed line is a first-order transition. Such
is, for example, the phase diagram of CuGeO3 (spin-Peierls system, see Section 11.2
below) in a magnetic field.
In general, for certain symmetry classes, there may exist in the free energy also
invariants linear in the derivatives (the so-called Lifshits invariants), for example
the term M curl M in magnetic systems. These terms will give helicoidal structures even for the case when the coefficient G > 0 (instead of (2.29) we will have
an equation for the wavevector of a new periodic structure of the type const q +
Gq 2 2 = 0). This is actually the situation in many rare earth metals having different types of such incommensurate magnetic structures. These terms also play
a crucial role in magnetically driven ferroelectricity in the so-called multiferroics.
The microscopic origin of such linear terms is a special problem which we do not
discuss here.
19

The theory of second-order phase transitions described above is essentially a mean
field theory. However, close to Tc , fluctuations become important. They modify the
behaviour at Tc of all thermodynamic functions, e.g. the specific heat c, the thermal
expansion , the compressibility , etc. For example, instead of a jump in c, there
may appear a real singularity, e.g. c(T ) ln(|T Tc |/Tc ) or |T Tc | . One
can estimate the region |T Tc |/Tc in which fluctuations become very strong
and the behaviour of many quantities, such as the specific heat, will deviate from
that described in Section 2.1. This region is determined by the condition that
the average fluctuations
order parameter become comparable to the order
of the

2
2
parameter itself, (
) . The corresponding criterion for the width of this
region, due to Ginzburg and Levanyuk, has the form
|T Tc |
B2
B 2 Tc

=
.
Tc
8 2 aG3
8 2 a 4 Tc2 06
Here
(2.32)

0 =
G
aTc
(2.33)
is the so-called correlation length at zero temperature; the parameters a, B, G

are the parameters of the free energy expansion (2.1), (2.26). The correlation
length (2.33) determines, e.g. the typical length-scale at which, in the ordered state
at T = 0, the order parameter recovers its equilibrium value when disturbed at
some point, or shows at which length-scale the order parameter changes close to
the surface of the sample.
Inside the temperature interval (2.32) fluctuations are important and modify
the behaviour of all physical characteristics of the system, but outside it (but still
close to Tc ) we can use the Landau theory. The parameter Gi = B 2 Tc /8 2 aG3 is
called the Ginzburg number; if Gi 1, one can use mean field results practically
everywhere without restrictions.
One sees from (2.32) that Gi 1 if the correlation length 0 is large enough.
This is, for example, the case in the usual superconductors, in which 0 (also called
much larger than the lattice parameter a.
there the coherence length) is 0 104 A
4
However, for example, in liquid He, 0 is of atomic order, and the fluctuation region
is very large, so that the specific heat at the transition to the superfluid phase has the
shape shown in Fig. 2.13, which is sometimes called the -anomaly. The qualitative
explanation of this behaviour is that due to fluctuations, e.g. above Tc , the ordering
does not disappear immediately, but there exists a certain short-range order, so
that the whole entropy is not released at Tc , but part of it remains also above Tc .
20

c
T
Fig. 2.13
This explains the tail of the specific heat at high temperatures. On the other hand,
very close to Tc the anomalies of thermodynamic functions may differ from those
assumed in the Landau expansion (2.1), see below.
One can show that the size of the region in which fluctuations are important is
determined microscopically by the spatial extension of the interaction leading to
the ordering, such as the exchange interaction in a ferromagnet: if the interaction
is long-range, the Ginzburg parameter is small, and consequently the fluctuation
region is small as well, and the mean field description, which the Landau theory in
fact is, is applicable. If, however, the corresponding interactions are short-range (as,
for example, the interactions between helium atoms in liquid He), then fluctuations
are important in a broad temperature interval; this is actually the microscopic reason
for the -anomaly in helium.
The quantity has a definite physical meaning, which gave rise to its name
correlation length. If one considers fluctuations of the order parameter at different
points, they are not completely independent but are correlated, very often of the
form

T
(0)
(r) er/ ,
r
(2.34)
where is the correlation length. In general it depends on temperature:

(T ) =
G
.
a |T Tc |
(2.35)
The quantity 0 in (2.33) is equal to (T = 0). Close to Tc , becomes very

large both when approaching Tc from below and from above. For T Tc + 0 the
average equilibrium value is = 0, but there exist fluctuation regions in which
order is already established; and (T ) is the typical size of such regions, which
tends to infinity for T Tc . Below Tc there exists long-range order,
= 0, but on
approaching Tc from below there appear disordered regions, whose size is again of
the order (T ).
21
It is often convenient to describe fluctuations not in real, but in momentum space.

Equation (2.34) then looks as follows:

q
q = |q |2 =
1
2
T
T
=
. (2.36)
2a|T Tc | (1 + q 2 2 )
2G (1 + q 2 2 )
This is the famous OrnsteinZernike theory of fluctuations. One can also obtain
the expression for the susceptibility close to the second-order phase transition. It is
given by similar expressions:

1
1

,
(T ) =
,
(2.37)
(T ) =

2a(T Tc )
4a(Tc T )
T >Tc
T <Tc
i.e. the susceptibility also diverges when T Tc (see also later, Chapter 6,
(6.40), (6.44)).
In principle the generalized susceptibility is a function of q and , (q, ) (e.g.
the well-known dielectric function (q, ), which is a response function to the
external electric field E(q, ), just as the usual susceptibility of magnetics is a
response function with respect to the magnetic field). There exists a very important
between (q, ) and the corresponding correlation functions

general connection
(r, 0) (r , t) or their Fourier transforms. For static susceptibility this relation
has the form

d 3 r iqr
1
(q) =
e
(0,
t)
(r,
t)
.
(2.38)
T
(2 )3
For the usual susceptibility, e.g. in magnetic systems for which the order parameter
is the magnetization, or the average spin S, it gives a convenient relation:
1
S0 Sn .
(2.39)
=
T n
In general close to Tc , e.g. for T > Tc , one obtains from (2.36) and (2.38), that

1
2 (T )
1
1

=

, (2.40)
=
(q)
T >Tc
2a(T Tc ) 1 + q 2 2 (T )
2G 1 + q 2 2 (T )
which for q = 0 gives (2.37).
2.5.1 Critical indices and scaling relations
All the considerations presented above are valid when we proceed from the
expansion (2.26) for the free energy and treat a second-order phase transition
in essentially a mean field way. However, as we have already seen above, close
to Tc the fluctuations are always strong (and the width of the region in which
22
they are important, given by the GinzburgLevanyuk criterion, may be sufficiently

broad). In this region the description of second-order phase transitions should be
modified, and all the anomalies of thermodynamic functions such as specific heat,
compressibility, thermal expansion, susceptibility, etc. are very different from those
predicted by the Landau theory. This is really the field of the theory of second-order
phase transitions, which was especially active in the 1960s and 1970s and which
resulted in a rather deep understanding of the phenomena close to Tc for such
transitions. This is a very big field in itself, which we cannot cover here; one can
find corresponding results and references in many books and review articles, e.g.
Chaikin and Lubensky (2000) and Stanley (1987). The basic conclusion of these
very elaborate studies is that close to Tc the properties of the system usually have
singularities of the type |T Tc |i , where the exponents i (different for different
quantities) are called critical indices. Thus, e.g. the specific heat behaves as ,
where = |T Tc |/Tc ; the order parameter itself changes close to Tc as , the
susceptibility as , the correlation length diverges as , etc. (This is the
standard notation for these critical exponents.)
The exact values of the critical indices are known only in very few cases;
a notable example is the exactly soluble two-dimensional Ising model, see
Section 6.4.3 below. However, despite the absence of exact solutions in most
cases, there exist very powerful general results in this field known as scaling relations (Kadanoff; Patashinskii and Pokrovskii). The underlying idea is that when the
system approaches Tc , the correlation length diverges and becomes infinitely large.
In this case all microscopic details, important at short distances, become irrelevant,
and the properties of the system become universal. Moreover, as T Tc all lengthscales become equivalent, and, simply from dimensional arguments, one can show
that for instance when one changes all distances by a certain factor, L kL, e.g.
doubling the size of the system, different quantities would change accordingly.
From these arguments one can find the relations between different quantities in the
vicinity of the critical temperature. Some examples of these relations are:
+ 2 + = 2 ,
(2.41)
d+ =2
(2.42)
(here d is the space dimensionality), etc. The critical indices may be different from
system to system, but relations of the type (2.41), (2.42) are universal. The values of
the indices depend only on the dimensionality of the system d and on the symmetry
of the order parameter (real or complex scalar; vector; isotropic or anisotropic
system, etc.), and for each particular case the critical indices should be the
same, independent of the specific physical situation. These are called universality
classes.
23
As mentioned above, the values of the critical indices for different universality
classes are known exactly only in very few cases. But numerical calculations using
powerful computer algorithms have given pretty accurate values of these indices
in many cases. Thus, for example, we have:
1. For the 2d Ising model: = 0 (logarithmic divergence, which is weaker than
any power law divergence); = 18 ; = 74 ; = 1.
2. For the 3d Ising model: = 0.10; = 0.33; = 1.24; = 0.63.
3. For the 3d Heisenberg model: = 0.12; = 0.36; = 1.39; = 0.71.
4. Compare this with the mean field behaviour following from the Landau theory:
= 0 (the specific heat has a jump, equation (2.9), but no divergence at Tc ;
= 12 , see (2.6); = 1 (the well-known CurieWeiss law, see equation (6.40)
below); = 12 , see (2.35).
We see that indeed the specific behaviour (divergence) of different quantities
as T Tc is different for different cases (different universality classes), but the
scaling relations (2.41), (2.42) and others are fulfilled for each of them.
Theoretical methods used to obtain these results are both analytical (those giving scaling relations) and numerical (giving specific values of critical indices for
different quantities in different situations). One of the very powerful methods used
to treat these problems, and also many others in which the interaction is strong
and we do not have any small parameter which would allow us to use perturbation
theory or expansion of the type of (2.1), is the renormalization group method. It
originates from field theory, but is now widely used in many difficult problems in
condensed matter physics, such as the theory of second-order phase transitions,
the quantum Hall effect, etc., see, e.g. Chaikin and Lubensky (2000) and Stanley
(1987).
Until now we have considered phase transitions occurring at finite temperatures.
The description we used was actually that of classical statistical mechanics. Indeed,
even if the physical nature of some phase transition is of essentially quantum origin,
such as the superconducting transition in metals or the superfluid transition in liquid 4 He, close to Tc the correlation length becomes very large, see equation (2.35),
i.e. it becomes much larger than the distance between particles or the radius of the
interaction. In this case the behaviour of the system is essentially classical.
However, there may exist situations in which the critical temperature can be
suppressed, e.g. by application of pressure, a magnetic field or some other control
parameter g, see Fig. 2.14 (the grey region is here the region of classical fluctuations). In this case, if Tc tends to zero, quantum effects start to play a more and
more important role. The state of the system for g > gc (e.g. pressure P > Pc )
24
T
Tc
ordered
phase
QCP
Fig. 2.14
may be disordered even at T = 0, not because of classical, but because of quantum

fluctuations we can speak of a quantum disordered phase.
There is here no contradiction with the Nernst theorem, which requires that the
entropy of the system should tend to zero as T 0: the state of such a system
at T = 0 and P > Pc is a disordered one from the point of view of standard
phase transitions (e.g. magnetic ordering may be suppressed for P > Pc ), but it
is a unique quantum state described by a, maybe very complicated, but unique
wavefunction. Consequently the entropy of such state is zero. A simple example
is given by an ordinary metal (Fermi liquid), as compared, for example, with the
ferromagnetic metallic state: although, in contrast to the latter, a normal metal is
paramagnetic (Pauli paramagnetism, see Chapter 6 below), it is a unique state the
filled Fermi surface, which has zero entropy. As discussed below, in Section 2.7.3,
this determines, for example, the slope of the high-temperature insulatormetal
transition in V2 O3 .
The behaviour of the system close to the point (T = 0, g = gc ) in Fig. 2.14 is
determined by quantum fluctuations, and such a point is called a quantum critical
point (QCP). One can show that quantum fluctuations dominate the behaviour of
the system above QCP in the region of the phase diagram marked by hatching in Fig. 2.14 (of course the intensity of these fluctuations and consequently
the amplitude of all the anomalies decreases with increasing temperature). The
proximity to a quantum critical point may strongly influence not only thermodynamic, but also transport properties of the corresponding systems, leading, e.g. to
a non-Fermi-liquid behaviour, see Chapter 10 below. There are even suggestions
that the proximity to QCP may be important for the appearance of unconventional
superconductivity, in particular for the high-Tc superconductivity in cuprates. More
detailed descriptions of these phenomena one can find, e.g. in Sachdev (1999) and
in von Lohneisen et al. (2007).
25

2.7.1 Different types of order parameters
There may exist different types of order parameters in different systems. For example, they may be scalars, e.g. periodic density in crystals as compared to a homogeneous liquid; vectors (or pseudovectors), e.g. magnetization in ferromagnets or
spontaneous polarization in ferroelectrics. Order parameters may also be complex. An example is a complex scalar the electron (condensate) wavefunction
(r) in superconductivity. There may also exist more complicated types of order
parameters, for example tensor order parameters (liquid crystals, anisotropic superconductivity and superfluidity, e.g. in 3 He and most probably in high-temperature
and in heavy-fermion superconductors).
2.7.2 General principle

The general principle is that the free energy should contain invariants built in from
the order parameter. Thus,
for the vector order parameter M in ordinary magnets,

may contain (M)2 or div M 2 = ( M)2 , but, in the absence of an external field,
no terms linear in M or terms M 3 because they break the inversion symmetry
(equivalence of the states M and M). 2 Similarly, for the complex scalar order
parameter in isotropic systems the free energy should contain ||2 , but not
2 , etc.
Symmetry considerations are very powerful, and they determine the form of
the Landau free energy expansion. In the general treatment one should always
keep all the terms allowed by symmetry, even if we do not know in detail the
physical mechanisms responsible for their appearance. This was, for example,
the reason why we have included the term 2 u in (2.22), although we did not
specify the type of the order parameter or its dependence on the lattice distortion. Of course, the coefficients of such terms do depend on the microscopic
nature of the ordering, and to calculate these coefficients is a separate, often
very difficult problem. However we can deduce many general conclusions even
without such a microscopic treatment, using only very general properties such as
symmetry.
For magnetic states there is another, even more powerful restriction. All magnetic states break time reversal
invariance; magnetic moments are odd with respect to time inversion. (This becomes clear when one remembers
that the standard way to create a magnetic moment is by a current running in a coil; and time inversion means
that the current would run in the opposite direction, changing accordingly M to M.) But the free energy
should of course remain the same under time inversion, from which it follows that magnetic vectors should
always enter in even combinations; e.g. M 2 , M curl M, or M L where L is the antiferromagnetic vector, see
Section 6.2.3.
26
2.7.3 Broken symmetry and driving force of phase transitions

At second-order phase transitions the symmetry changes (it decreases in the ordered
phase). Thus, e.g. in a paramagnetic phase there exists spherical symmetry (free
rotation of spins), whereas in a ferromagnet all spins point in the same direction,
the spin orientation is fixed, i.e. we have broken spin rotation invariance.
The situation is similar in liquids compared with crystals: in liquids there is
continuous translational symmetry (shift by an arbitrary vector), whereas in a
crystal there remains only a shift by a vector equal to the lattice period. This means
that we have broken the continuous translational symmetry. The low-temperature
phase usually has lower symmetry than the high-temperature one.
One should keep in mind an important distinction: there exist cases of a broken
continuous symmetry, or of a discrete one. The examples given above correspond
to broken continuous symmetry. But there are also cases of a broken discrete
symmetry. Such is, e.g. the case of strongly anisotropic magnets. Suppose that
spins can take not an arbitrary orientation, but only two: or . Again, in the hightemperature disordered phase there is equal probability of finding spins or at
a given site. The low-temperature ordered phase would correspond, e.g. to spins
being predominantly (at T = 0, only ). This is the so-called Ising ferromagnet.
Here a discrete symmetry (spin inversion ) is broken.
As we have already discussed, it is easy to understand why the system goes to
a more disordered phase with increasing temperature. If we start from the lowtemperature ordered phase, we see that the transition to a disordered phase is
driven by the entropy. According to (1.10) the free energy (for fixed pressure) is
= E T S. At low temperatures, e.g. T = 0, to minimize the free energy we
should make the (interaction) energy as low as possible, which is reached in the
ordered phase. However at high enough temperatures to decrease we should
make the entropy nonzero, and this drives the transition to a disordered phase.
We thus gain in entropy, losing in energy, which, according to (1.10), becomes
favourable with increasing temperature.
We have to include all contributions to the entropy. Sometimes it is not easy to
understand at first glance why the transition occurs in a certain way. A good
example is the insulatormetal transition (Mott transition), e.g. in V2 O3 , see
Chapter 12 below. Briefly, this is a transition between the state with electrons
localized each at their own site (MottHubbard insulator) and a metallic state with
itinerant electrons. It seems that in this case we should expect the insulating state
to be the stable state at low temperatures (electrons localized at their sites, an
electronic crystal), and the metallic phase to be the high-temperature phase (delocalized, moving electrons an electron liquid), similar to melting of an ordinary
crystal with increasing temperature. However the actual situation is not so simple.
27
paramagnetic
insulator
paramagnetic
metal
A
antiferromagnetic
insulator
P
Fig. 2.15
The experimental phase diagram of V2 O3 looks schematically as shown in

Fig. 2.15. We see that the transition between an antiferromagnetic insulator and
a metal does indeed occur with increasing temperature. However, the pure Mott
transition (line AB), not associated with the destruction of magnetic order, is such
that with increasing temperature we go from a metal to an insulator! Why? The
explanation is most probably that in the metallic phase we have a Fermi surface,
and in a sense it is an ordered state (a unique state, the entropy of which is small
and at T = 0 would be zero). But the insulating state with localized electrons
also has localized spins, which in the paramagnetic phase are disordered, i.e. the
paramagnetic insulator has entropy (per site) kB ln(2S + 1), higher than that of
the metallic Fermi sea. Therefore, according to the general rule, with increasing
temperature the system goes over from the state with lower entropy (metallic state)
to the one with higher entropy (paramagnetic insulator), although at first glance this
looks strange and counter-intuitive. This is not a unique situation, and in each such
case we have to think which entropy drives the observed transition with increasing
temperature.
2.7.4 The Goldstone theorem

There exists one general result which is known as the Goldstone theorem. According
to this, when there is a broken continuous symmetry at the phase transition, there
should exist in the ordered state of the system (without long-range interaction)
a collective mode, an excitation, with gapless energy spectrum (the energy starts
continuously from 0). There exist many examples of such excitations. Here I give
only a few: in an isotropic ferromagnet these are spin waves, with the spectrum
shown qualitatively in Fig. 2.16. In crystals these are the usual phonons, Fig. 2.17.
We will meet many other such examples later in this book.
28
q
Fig. 2.16
q
Fig. 2.17
Qualitative explanation
In the case of broken continuous symmetry in the ordered phase there exist infinitely
many degenerate states (e.g. in a ferromagnet all directions of spontaneous magnetization are possible). All these states have the same energy and are equivalent.
A collective mode with q = 0 describes a transition from one such state to another
(e.g. rotation of the total magnetization of the sample as a whole) and it should
cost us no energy, hence the spectrum of such excitations (q) should start from
= 0, i.e. there exists gapless excitation. This is the content of the Goldstone
theorem; the corresponding gapless modes are often called Goldstone modes.
2.7.5 Critical points

In principle there may exist phase transitions without a change of symmetry, but
only the first-order ones. Such is, for example, the liquidgas phase transition. The
symmetry of both these phases, gas and liquid, is the same, but they differ in density
(and of course in many other properties, not related to the phase transition itself).
Such a transition can end at a critical point (P , T ) (Fig. 2.18): here the thick dashed
line is a first-order transition, and the thin solid lines are the limits of hysteresis. As
in this case there is no change of symmetry across the first-order phase transition,
one can in principle go from one phase to another continuously, e.g. from point
A to point B in Fig. 2.18, moving around the critical point without crossing the
transition line. For the second-order phase transitions this is impossible, because
the symmetry of the two phases, the disordered one and the ordered one (or two
different ordered phases) is different.
The same situation in (P , V ) coordinates has a familiar form, Fig. 2.19, where
above the critical point (for T > T , P > P ) the state of the system is unique, and
below it the system with fixed total density, or fixed volume in the grey region, would
phase-separate (decompose) into different phases (here gas and liquid), the relative
volumes of which will be determined by the well-known Maxwell construction. The
point (P , T ) is the critical point. Above it the P V isotherms have only negative
29
T
~
T
B
P
~
P
Fig. 2.18
P
~
P
~
T>T
~
T=T
~
T<T
Fig. 2.19
slope, dP /dV < 0, the compressibility (1.21) is positive, and the transition (crossover), e.g. from a liquid to a gas state, is smooth. Inside the grey region in Fig. 2.19
there are parts of P V isotherms with dV /dP > 0, which would imply negative
compressibility, i.e. an absolute instability of the corresponding homogeneous state.
This is in fact the reason for the phase separation in this region.
Critical points can appear even in solids. This is the situation, e.g. for the
isomorphous phase transitions like the transition in Ce or in a high-temperature
metalinsulator transition in V2 O3 shown in Fig. 2.15, where the point B is such a
critical point.
Concluding this chapter, one general remark is in order. We have seen that the
Landau approach to phase transitions, although conceptually and technically rather
simple, is nevertheless very powerful in describing many quite different situations.
It uses the most general arguments such as those of symmetry, etc. and if there are
no special indications otherwise, makes the simplest assumptions possible, such
as the use of a Taylor expansion in the small parameter or , to give quite
general and very successful descriptions of very complicated phenomena. Such an
approach is often very fruitful also in many other fields of physics.
30
One good example is the treatment by Landau of the extremely complicated

phenomenon of turbulence (see Landau and Lifshits, Fluid Mechanics, 1987). In
this problem it is known that the flow of a liquid remains homogeneous, laminar,
if the so-called Reynolds number R is less than a certain critical value Rc , and
the flow develops instability (transition to turbulence) for R > Rc . Landau again
used here, in the absence of a complete theory (which still does not exist!) an
expansion in terms of the amplitude of a new mode of the motion, which
appears for R > Rc and breaks the laminar flow. The expansion again has the form
A2 + B4 , with A = (Rc R), so that the new mode is absent for R < Rc
and the laminar flow is stable, and

= 0 ( R Rc ) for R > Rc . The analogy
with the treatment of the second-order phase transition (cf. (2.1)(2.6)) is of course
apparent. Again, this approach is limited, and the real theory of turbulence should
be much more complicated, e.g. it may resemble the theory of second-order phase
transitions going beyond Landaus approach and briefly discussed in Section 2.5.1,
with critical exponents, etc. or be even more complicated. But the Landau theory
gives at least a first orientation in this nontrivial problem. Such an approach is
widely used also in many other fields of physics, including elementary particle
physics and cosmology.
3
Bose and Fermi statistics
This chapter is a short reminder and a collection of basic formulae on Fermi and
Bose statistics. For noninteracting particles with the spectrum k , the energy of the
quantum state, in which there are nk particles, is Ek,nk = k nk (we incorporated in
the index k also spin indices, and other indices if they exist). Then, according to
(1.32), (1.33), the thermodynamic potential for this quantum state is equal to

nk
k = T ln
e(k )/T
.
(3.1)
nk
For bosons (the occupation numbers nk can take any value, i.e. the summation
in equation (3.1) goes from nk = 0 to ), the expression (3.1) converges if
e(k )/T < 1, i.e. we necessarily have < k (or < 0 if k k 2 /2m), and after
summation we obtain

Bk = T ln 1 e(k )/T .
(3.2)
Then from (1.31) the average occupation of the state k is
n Bk =
Bk
1
.
= ( )/T
e k
1
(3.3)
For fermions, due to the Pauli principle, we can have no more than one particle in
a given quantum state, i.e. nk = 0 or 1. Then the thermodynamic potential (3.1) is

Fk = T ln 1 + e(k )/T ,
(3.4)
and
n Fk =
Fk
1
.
= ( )/T
e k
+1
31
(3.5)
32
Fig. 3.1
In both cases the total number of particles N =

the chemical potential (T ).

k
nk ; this condition determines
Problem: Discuss the cross-over to Boltzmann statistics. Find the asymptotic

behaviour of (T ) at high temperatures.
Solution: Boltzmann statistics (the classical case) corresponds to the situation in
which all nk 1 (no double occupancy, etc.), i.e.

e(k )/T 1
(k )/T
.
(3.6)

nk = e
(k )/T
or e
1
The chemical potential is obtained from the condition

N
d3k
n=
.
= nk
V
(2
h)3
(3.7)
Taking the energy spectrum k = k 2 /2m, and going over to the variable (and then
to the dimensionless variable z = /T ), we obtain:

3/2 3/2
m2/3
T
()/T
/T m
=
n=
e
d
e
z ez dz .
3
3
2
(/T
z)
2 h 0
2 2
h 0
(3.8)
3
The integral in (3.8) is the gamma function, 2 = 2 .
In effect n = e/T (T / )3/2 , where is a constant (a combination of m, ,
h, . . . ). Inverting this equality, we find

3/2
T
/T
=n
,
(3.9)
e

T
3
(3.10)
= T ln n T ln
2
which is valid at T . This behaviour is schematically shown in Fig. 3.1; we

will need these results later on, in Chapter 5.
33
Special consideration is required for (quasi)particles, whose number is not conserved. Such is, for instance, the situation for phonons in crystals or for photons
quanta of the electromagnetic field. As mentioned in Chapter 1, in this case the
number of particles N itself is determined by the condition of minimization of the
corresponding thermodynamic potential, e.g. the free energy, with respect to N.
From the definition of the chemical potential (1.25) we then see that in this case
the corresponding chemical potential has to be taken as zero for all temperatures.
4
Phonons in crystals
In this chapter we will discuss the first, and probably the best-known example of
bosonic systems phonons in crystals. According to our general scheme, after
briefly summarizing the basic facts about noninteracting phonons in a harmonic
lattice, we will pay most attention to the next two factors: quantum effects in the
lattice dynamics, and especially the interaction between phonons which leads, e.g.
to such phenomena as thermal expansion, explains the features of melting, etc. But
for completeness we give, at the beginning, a very short summary of the material
well known from standard courses of solid state physics.
4.1 Harmonic oscillator

The classical equation for a harmonic oscillator is
M x = Bx .
(4.1)
x = x0 eit ,
(4.2)
Its solution is:
where the frequency is

=

B/M .
The Hamiltonian of the harmonic oscillator has the form

h
p2
Bx 2
H=
+
p =
.
2M
2
i x
In quantum mechanics the energy levels are quantized:

n = n + 12
h .
34
(4.3)
(4.4)
(4.5)
35

The operators of the coordinate x and momentum p obey the commutation relation
p]
= x p p x = i
[x,
h.
(4.6)
It is convenient to introduce the annihilation and creation operators of phonons

(in the following we omit the sign for the operators):
(M x + ip)
b=
2 hM
(4.7)
b =
(M x ip).
2
hM
Problem: Check that b, b obey the commutation relation
[b, b ] = 1
(4.8)
(in future we also omit the minus sign on the commutator [ ] ).

Solution:

1
2
[x,
x]
+
iM[p,
x]
iM[x,
p]
+
[p,
p]
(M)
2
hM
(2i)i
h
1
=1.
(4.9)
= 2i[x, p] =
2h
2h
[b, b ] =
We can express x, p through b, b , using (4.7):

h
hM
i 2
x=
p=
(4.10)
(b + b) ,
(b b) .
2M
2
From (4.4), (4.10) and using (4.8), we obtain

H = 12
h(bb + b b) =
hb b + 12
h =
h n + 12 ,
n = b b .
(4.11)
For the eigenstates (the states with certain particular value n) equation (4.11)
gives the energy levels (4.5); this is actually the simplest way to obtain the energy
spectrum of the harmonic oscillator. Note the presence of 12 in (4.11); this term
corresponds to the so-called zero-point oscillations and it describes a real physical
effect, which, as we will see later, often has very important physical implications.
The states with n excitations |n obey the relations
b |n = n |n 1
(4.12)
b |n = n + 1 |n + 1 ,
that is, b is indeed an annihilation operator and b is a creation operator; they
respectively decrease and increase the number of phonons by 1. For the number
36
Phonons in crystals
M1
M2
M1
M2
M1
Fig. 4.1
operator n we have, as we should,
n|n
= b b |n = b n |n 1 = n |n .
(4.13)
For the ground state |0 (the state with zero phonons)
b |0 = 0 .
(4.14)
1
|n = (b )n |0 .
n!
(4.15)
From (4.12) we then obtain
Let us now consider not an isolated oscillator, but a linear chain, consisting of
atoms with the harmonic interaction: its Hamiltonian is

M u 2
B
n
2
+ (un un+a ) .
(4.16)
H=
2
2
n
(Here we have introduced the deviation of the n-th atom from its equilibrium
position un = xn xn0 .) The equations of motion have the form
M u n = B(2un un+1 un1 ) .
(4.17)
We seek the solution in the form un = uq eiqn . Then from (4.17) we get
M u q = n(2 eiqa eiqa ) = 2B(1 cos qa) uq .
This is also the equation for the harmonic oscillator, with frequency

B
B
qa
q = 2
sin
qa
M
2 (q0)
M
(4.18)
(4.19)
(compare with (4.1), (4.3)).

Problem: Consider a linear chain with the basis (two atoms per unit cell e.g.
with different masses M1 and M2 , see Fig. 4.1). Find the phonon spectrum.
Solution: For convenience we introduce two variables, un and vn , for the atoms of
two kinds, see Fig. 4.2. The equations of motion for un and vn are:
M1 u n = B(vn + vn1 2un )
M2 vn = B(un+1 + un 2vn ) .
(4.20)
n+1
u n+1
}
}
}
un
37
cell:
n+1
n+2
Fig. 4.2
w
w+
w
p / 2a
Fig. 4.3
Let us seek the solution in the form un = ueiqnait , vn = veiqnait . The solution
of equation (4.20) is now reduced to the diagonalization of the matrix

B(1 + eiqa )
2B M1 2
=0.

(4.21)

2B M2 2
B(1 + eiqa )
The eigenenergies are

2
(q)
=B
1
1
+
M1 M2

B
1
1
+
M1 M2
2
4 sin2 qa
.
M1 M2
(4.22)
Thus there exist in this case two branches of phonons, see Fig. 4.3: optical phonons
with the spectrum + , which at q = 0 have finite frequency,

1
1
2
,
(4.23)
+
+ (q = 0) = 2B
M1 M2
and ordinary acoustic phonons, whose spectrum is obtained by taking the minus
sign in equation (4.22),

(q = 0) = 2B/(M1 + M2 ) sin qa ,
(4.24)
which for M1 = M2 coincides with (4.19).
Problem: The the same for equal masses, but alternating spring constants, Fig. 4.4.
38
Phonons in crystals
B1
B2
B1
M
B2
M
Fig. 4.4
Solution:
2
(q) =
B1 + B2
1
M
M

B12 + B22 + 2B1 B2 cos qa .
(4.25)
In the general case the Hamiltonian of the harmonic lattice takes the form
H=

hq bq
bq + 12 + const.
(4.26)
q,
where q is the (quasi)momentum (in the one-dimensional case a < q < a ; in

general q lies in the first Brillouin zone), and is the mode index (denoting
acoustic or optical modes and the corresponding polarization one longitudinal,
two transverse modes).
4.3 Physical properties of crystals in the harmonic approximation

Phonons are bosons. The number of phonons is not fixed, which means that the
phonon chemical potential should be taken as zero, see Chapter 3:
ph = 0 .
(4.27)
Consequently, the phonon occupation number is (cf. (3.3))

n q =
1
e
hq /T
(4.28)
The energy of the mode q is Eq = (nq + 12 )

hq . Thus the total energy (including
1
the zero-point energy the term with 2 in (4.26)) is
E=
hq

hq
q
e hq /T
1
+
1 2

=

hq e hq /T + 1
2 e hq /T 1
q
coth
hq
.
2T
(4.29)
4.3 Physical properties of crystals
39
General rule
The transformation of the sum in q to an integral is done as follows:

in 3d-case d = 3;

dd q
similarly in one-dimensional and in . (4.30)
=
(2 )d
q
two-dimensional cases (d = 1, 2)
By (1.18) the specific heat (per unit volume) is

hq )2 e hq /T
d 3 q (
E
1
cV =
= 2
.
T V
T
(2 )3 (e hq /T 1)2
(4.31)
h = 1 (as well as the Boltzmann constant kB = 1).

Later, in most cases we put
One can introduce the phonon density of states D(): d 3 q/(2 )3 = D() d,
where D() d is the number of phonon
states at energy in an interval d.
#
Normalization should be such that d D() over the Brillouin zone is equal to
the total number of phonon modes per unit cell, i.e. equal to the total number of
degrees of freedom. If there are m atoms per unit cell and Nc unit cells, there are
3mNc = 3N modes. Then

(/T )2 e/T
cV = 3N
D() d .
(4.32)
(e/T 1)2
Different models for the phonon spectra (q) give different forms of the phonon
density of states D() and consequently in general different phonon specific heat:
(1) Einstein model, (q) = 0 = const. (this is not such a bad approximation for
optical phonons). The phonon density of states is then a delta function, and the
specific heat is
cVE = 3N
(0 /T )2 e0 /T
(e0 /T 1)2
(4.33)
or, with
h0 E , where E is the corresponding temperature (we then put
E in (4.33) instead of 0 , or simply take 0 in kelvin), it is

2

E
E
E
(4.34)
T 0 :
cV (T E ) 3N
exp
T
T

E
T E :
.
(4.35)
cVE (T E ) 3N + O
T
The expression (4.35) is the DulongPetit law, well known from classical
physics (we recall that above we put kB = 1; the standard form of the Dulong
Petit law is c = 3kB N).
(2) Debye model: we approximate the total phonon spectrum by the spectrum
of acoustic phonons (q) = sq, with the appropriate
upper cut-off, see the
#
schematic picture in Fig. 4.5. The integration in d 3 q is carried out not over
40
Phonons in crystals
D()
D()
~ 2
(a)
(b)
Typical real density of

states for phonons
Debye model (with the

approximation = sq)
Fig. 4.5
the real Brillouin zone, but over a sphere, with the volume equal to the volume
of the Brillouin zone.
The number of q-points in the Brillouin zone is N ; their density is V /(2 )3 ;
thus
V 4 3
(4.36)
N=
q
(2 )3 3 0
where q0 the Debye wavevector is the maximum wavevector of the equivalent sphere in q-space.
The maximum frequency D = sq0 is called the Debye frequency (s is the
sound velocity); D =
hD is the Debye temperature.
In effect

D /T
x 4 ex
T 3
3
dx .
(4.37)
cV = 3N
D
(ex 1)2
0
(Here we have used that D() d in (4.29) is
D() d =
4 q 2 dq
32
=
d ,
4
D2
qD2
3
hD .)
and substituted kB D for
The limiting behaviour of the specific heat in the Debye model is

T 3
12 4
N
cV
,
T 0 (T D ) :
5
D
T (T D ) :
cV 3N .
(4.38)
(4.39)
(4.40)
At high temperatures we have the DulongPetit law again. (Actually the crossover from the low-temperature behaviour cV T 3 to the high-temperature limit
occurs not at T D , but at approximately T 14 D .) One has to remember
that at low temperatures the phonon specific heat is T 3 ; this relation is very
important for many experiments.
41
Problem: Find the free energy of the harmonic crystal.

Solution: The energy of the harmonic crystal is

hq nq + 12 ,
nq = 1, 2, 3, . . . .
E=
(4.41)
The partition function is

Z=
$
q
nq
%
& $ hq /2T

hq
e
1
nq + 2 =
exp
.
T
1 e hq /T
q
(4.42)
The free energy (1.9) is then given by the expression

hq
hq /T
1
.
hq + T
ln(1 e
)=T
ln 2 sinh
F = T ln Z = 2
2T
q
q
q
(4.43)
Here we have used a simple transformation

h/2T

e h/2T
h/T
h/2T e
ln(1 e
) = ln 2e
2

h
h
.
(4.44)
=
+ ln 2 sinh
2T
2T

In general the interatomic interaction, i.e. the potential energy of two atoms with
coordinates x1 , x2 is v (x1 x2 ). Previously, in the harmonic approximation, we
used an expansion in small deviations from the static positions
v (x1 x2 ) = v 0 + 12 B (u1 u2 )2
x1 = x10 + u1 ,
(4.45)
x2 = x20 + u2 ,
see, e.g. (4.16).

In general there exist anharmonic terms, proportional respectively to (u1 u2 )3 ,
(u1 u2 )4 , etc.:
v = v0 +

(u1 u2 )2 + (u1 u2 )3 + (u1 u2 )4 + .

2
3!
4!
2v
,
B=
x 2
3v
=
,
x 3
4v
=
,
x 4
(4.46)

...
(4.47)
The term with the coefficient gives rise to cubic anharmonism, the one with to
quartic, etc.
42
Phonons in crystals
harmonic
harmonic
anharmonic
anharmonic
Fig. 4.6
q3
p
z
q4
(a)
q1
n
q2
(b)
Fig. 4.7
Usually the coefficient is negative, < 0, and the potential looks like the one
shown in Fig. 4.6 by the solid line, i.e. it is steeper for negative relative distortion
(when two atoms approach one another) and less steep for positive u1 u2 , when
the two atoms move further apart. (This is quite natural: the interaction between
atoms becomes stronger when they approach the overlap of atomic or ionic cores
leads to a very strong, almost hard core repulsion; and this interaction is weaker
when the distance between atoms becomes large.)
One can also write down the corresponding anharmonic terms in second
quantization form, through the operators b, b , and add them to the phonon
Hamiltonian (4.26). Cubic anharmonism will give rise to terms with the struc
ture p,q (bpq bq bp + h.c.), and quartic anharmonism, to terms of the type
(
'

q1 q2 q3 q4 (bq1 bq2 bq3 bq4 + h.c.) + k1 k2 k3 k4 (bk1 bk2 bk3 bk4 + h.c.) , where the momenta involved obey conservation laws q1 + q2 = q3 + q4 , k1 = k2 + k3 + k4 (the
total momentum of created phonons is equal to the total momentum of annihilated
ones).
Anharmonic interactions have several consequences. One is that they lead to
phononphonon interactions, schematically illustrated in Fig. 4.7. These processes
are important, e.g. for thermal conductivity. These graphs, Fig. 4.7(a), 4.7(b), at
this stage can be treated simply as pictorial representations of certain processes.
Thus, if we depict the phonon by a wavy line, Fig. 4.7(a) corresponds to a process in
which in the initial state there was one phonon with momentum p, which as a result
of anharmonic interaction is transformed into two other phonons, with momenta q
and p q; such processes are allowed by cubic anharmonism. Correspondingly,
43
Fig. 4.7(b) describes the process of scattering of two phonons with momenta q3
and q4 into two others, with momenta q1 , q2 ; such processes are contained in the
quartic term of the Hamiltonian. But one can give these pictures much more
meaning: they are actually Feynman diagrams, which allow one really to calculate
the probabilities of corresponding processes. This method will be discussed in
more detail in Chapter 8, and in Chapters 911 we will show how it works in many
specific problems.
Another important consequence of anharmonic interactions is their role in thermal expansion and in the melting of crystals. This will be discussed in the following
sections.
4.4.1 Thermal expansion
By (4.43) the free energy of a crystal in the harmonic approximation is (putting
h = 1)
)
q *
F =T
ln 2 sinh
.
(4.48)
2T
q
With the inclusion of anharmonicity, i.e. of the phononphonon interaction, this is

no longer true. But one can still use this expression approximately in the so-called
quasiharmonic approximation, accounting for anharmonicity in the following way.
The phonon frequencies q in the anharmonic crystal in general depend on the
specific volume, q (V ) (see Fig. 4.6: the curvature of the potential v (x), d 2 v /dx 2 ,
which according to (4.3), (4.4) determines phonon frequencies, in the anharmonic
case depends on x). Usually this dependence is described by the phenomenological
relation
d ln
V d
=
= ,
(4.49)
dV
d ln V
which is called the Gruneisen approximation; is the Gruneisen constant (usually,
in ordinary crystals, 12).
The total free energy as a function of volume can then be written as

q (V )
1 V 2
.
(4.50)
F (V ) =
+T
ln 2 sinh
2 V
2T
q
The first term in (4.50) is the elastic energy after deformation V , and is the
lattice compressibility (inverse bulk modulus). In equation (4.50) we considered
the situation when we (artificially) fix the volume of the system V , which may
differ from the equilibrium volume without phonons by the distortion V . Then
we indeed should include in the total energy the first term of equation (4.50)
describing such deformation. The second term gives the phonon contribution to the
44
Phonons in crystals
free energy, which is calculated by treating the lattice as harmonic, but at a given
volume V (the expression (4.43) for the phonon contribution, strictly speaking, is
valid only for the harmonic crystal). However, whereas in the purely harmonic case
the phonon frequencies are constant, independent of the distance between atoms
or of the volume, here we assume that they do depend on V , by the relation (4.49).
Thus we effectively take into account anharmonic effects, albeit approximately,
via the phenomenological Gruneisen relation. Such a scheme is often used in
treating anharmonic lattices, and it is called the quasiharmonic approximation. The
equilibrium volume V in this scheme should be determined by minimizing the free
energy (4.50); due to the presence of the second term the resulting volume would
depend on temperature, and this dependence gives thermal expansion.
Differentiating (4.50) in V and using (4.49), we obtain
q
q
1 V
),
=
coth
= E(T
(4.51)
V
2
2T
q
) is the average energy of the lattice (cf. (4.29)).
where E(T
Thus the volume thermal expansion is:
=
)
E(T
cV
1 (V )
=
= cV =
,
V T
T
B
(4.52)
where B = 1/ is the bulk modulus.

The linear thermal expansion coefficient is

l
1
1 V
1 (l 3 )
3 l
=
= as =
= 3
=
= 3 .
T
3
V T
l T
l T
Thus
=
cV
.
3B
(4.53)
This is called the Gruneisen equation. We see that we get nonzero thermal expansion only when we include anharmonic effects (the dependence of the phonon
frequencies on interatomic distance or on volume, which is phenomenologically
described by the relation (4.49)). For constant phonon frequencies (Gruneisen
constant = 0) thermal expansion would be absent.
The Gruneisen equation (4.53) is very useful, as it establishes the relation
between different measurable quantities and allows us, e.g. to calculate the thermal expansion if the specific heat and bulk modulus are known. It is often used in
practice, to check the consistency of different thermophysical data of real materials.
Why is anharmonism necessary for thermal expansion? The physics of the
expansion is illustrated in Figs. 4.8 and 4.9. At low temperatures the system is
45
3
2
Fig. 4.8
Fig. 4.9
at the lowest quantum level of the oscillator (state 1). At higher temperatures the
average distance between atoms is given by averaging over all occupied higherlying quantum levels. In the harmonic case we have the situation shown in Fig. 4.8,
and the average displacement at each level, i.e. the average lattice parameter, is the
same at each quantum level. Thus it does not change with temperature, and in the
purely harmonic crystal we would not have any thermal expansion.
On the other hand, in an anharmonic crystal the interaction potential qualitatively looks as shown in Fig. 4.9 (cf. Fig. 4.6). Then with increasing temperature,
when the higher levels become occupied, the average distance between atoms
(lattice parameter) increases with temperature and this is the conventional thermal expansion.
4.4.2 Melting
From (4.49) we see that the dependence of the phonon frequencies on the specific
volume can be rewritten as

V
.
(4.54)
= 0
V0
As the change of volume (due to thermal expansion) is usually small, we can
approximately write instead of (4.54) (and using the definition of thermal expansion
)
= 3 = V1 V
T
V V0 (1 + 3T ) .
(4.55)
(T ) 0 (1 3 T ) .
(4.56)
This means that
Actually, as follows from a more complete treatment, see Section 8.6, the description of phonons always contains not the phonon frequency , but rather 2 . Consequently, the equation describing the change of phonon frequencies with temperature
46
Phonons in crystals
w2
T~*
Fig. 4.10
has not the form (4.56), but is rather

2 (T ) = 02 (1 6 T ) .
(4.57)
The more general equation for (T ) is

2 (T ) = 02 1 cn(T ) ,
(4.58)
where n(T ) = 1/(e/T 1) is the average number of excited phonons (this will be
shown later); here c is some constant ( 2 /B 3 ). Then for 0 < T

cT
2
2
,
(4.59)
= 0 1
0
i.e. for high temperatures the expression (4.58) gives the same temperature dependence as (4.56), (4.57).
These expressions give the dependence of on T shown in Fig. 4.10. At a
certain temperature T the phonon frequency 2 = 0, and for T > T it becomes
negative, which means an instability of our system and the transition to a new state,
e.g. melting of the crystal.
I remind readers that in quantum mechanics the time dependence of the
wavefunction is given by (t) = (0) exp it. When 2 becomes negative, i.e.
t = i||t, this would give an exponential growth of the corresponding quantities. In our case, when the phonon frequencies cross zero, this would mean an
exponential growth of the number of corresponding phonons, or of the respective distortion, which means absolute instability of the initial state, in this case
a crystal. Whether such instability would indeed correspond to melting or to a
structural transition to a different crystal structure, depends on which particular
phonon mode becomes unstable. If this is the phonon with momentum Q, then this
implies the change of the crystal structure with the formation of a superstructure
with this wavevector, u(r) exp(i Q r). Melting corresponds in this language to
the softening of transverse phonons at q = 0 (in other words, the shear modulus
becomes negative).
47
w
w0
T*
T~*
Fig. 4.11
In general, the description of structural phase transitions as an instability of the

original crystal structure due to softening of particular phonon modes is known as
the soft mode concept (Ginzburg, Cochran, Anderson). It is widely used in solid
state physics and leads to definite predictions which can be checked experimentally.
Returning to melting, we see that in this approximation the transition would be
continuous, i.e. second order (the phonon frequency goes continuously to zero). But
on general grounds melting should be a first-order phase transition (one can show
that in this case there exist cubic invariants in the Landau free energy expansion,
see Section 2.2 above). How can we correct this drawback? It is more correct to
take in (4.58) not n(T ) = 1/(e0 /T 1) (which leads to (4.59)), but the number
of phonons with an already renormalized, new frequency , 1/(e/T 1). Then
instead of (4.59) we would obtain

cT
(4.60)
= 0 1 cn() = 0 1
(here we have simplified the mathematics by writing the equation not for 2 , but
for ; this is sufficient for our qualitative treatment).
The equation (4.60) is a self-consistent equation for . Its solution is

0
=
02
cT 0 ,
4
(4.61)
see Fig. 4.11. We see that in contrast to the previous treatment, the self-consistent
solution with real exists only up to a temperature T = 0 /4c (smaller than T ),
after which it becomes complex, i.e. the time dependence of the vibration amplitude (4.2) e+it is again diverging, which indeed signals an instability of the
lattice, in our case melting. And in this theory melting would be a first-order phase
transition, with a jump in , as it should be. The results are qualitatively the same
if we had proceeded from equation (4.58) for 2 .
48
Phonons in crystals
Fig. 4.12
The softening of phonons means at the same time a decrease of the bulk modulus
of the lattice, B: as 2 = B/M, we get from (4.57)
B(T ) B0 (1 6 T ) .
(4.62)
But, from the Gruneisen relation (4.53), 1/B, which again gives a selfconsistent equation for B (or for 2 ), equivalent to (4.58):

2 2 cV T
,
(4.63)
B(T ) = B0 1
B(T )
i.e. we again obtain a quadratic equation for B(T ) similar to (4.60), which will
again give first-order melting.
Once again, strictly speaking an instability of a phonon does not mean necessary
melting; it can signal, e.g. a transformation into another crystal structure. To check
that we will indeed have melting, we have to show that the shear modulus (shear
modes for q 0) becomes unstable. But qualitatively the picture described is
correct, and it is also consistent with other approaches to melting, described below.
Why does melting occur at all? How can one explain it qualitatively? Let us
extend Fig. 4.6 a bit. From (4.46), v = v 0 + 12 Bu2 3!1 | |u3 , i.e. this potential
actually looks as in Fig. 4.12. We see that there exists an infinite motion at high
enough temperature! The lattice is no longer stable, which means melting. Thus
we again see that the melting is intrinsically connected with the anharmonicity of
the lattice.
4.4.3 Another approach to melting. Quantum melting

From equation (4.10) we see that the vibration amplitude can be expressed through
the phonon operators as

h
(4.64)
(b + b)
u=
2M
49
(for one phonon mode). The average shift from the equilibrium position is of course
zero, u = 0, but the square of the average amplitude of vibrations is nonzero:

h
h
(b + b)2 =
b b + bb + b b + bb
2M
2M

h
h
2 b b + 1 =
n + 12
=
2M
M
u2 =
(4.65)
(we use the commutation relation bb b b = 1).

Thus
u2 =
2

h
h
n + 12 =
n + 12
M
MD
(4.66)
(there may enter some numerical factors such as 3 because of the presence of several
phonon modes). The limiting values of this mean square vibration amplitude are:
T 0 : u2
T D :
h
2MD
hT
u2 M
2
2
(important: note that u2 T =0

= 0;
(4.67)
these are the famous zero-point oscillations)

2
hT
9
(4.68)
actually it is equal to
.
M2D
When the vibration amplitude becomes comparable to the lattice spacing itself,
u2 a 2 (actually when u2 /a 2 0.2), melting occurs; this is the Lindemann
criterion of melting.
The factor u2 also enters into the intensity of X-ray scattering in crystals, in the theory
of the Mossbauer effect, etc. It enters through the factor e2W , where
W u ,
2
2 2
3hK
8 M
D
W =
2 2
3h K T
2 M2D
(T 0)
(4.69)
(T D ) .
h/a; the quantity W is called the DebyeWaller factor.

K is the Umklapp wavevector
Thus, in melting an important parameter is

=
u2
,
a2
(4.70)
50
Phonons in crystals
the so-called de Boer parameter (usually this term is used at T = 0, for zero-point
vibrations; this is then the quantum de Boer parameter). At T = 0

h
B
2
with =
.
(4.71)
u
M
M
The stiffness of the lattice B (the bulk modulus) can be estimated as follows: when
an atom is shifted from its equilibrium position by a distance a (a is the lattice
parameter), the change in potential energy Ba 2 is of the order of the typical
interaction between atoms v (a), i.e.
B
v
a2
(4.72)
As a result the quantum de Boer parameter (4.70) becomes

h 1
.
a Mv
(4.73)
If >
0.20.3, the crystal is unstable with respect to zero-point motion even at
T = 0, which means quantum melting. This will also be important for electrons
(see the discussion of Wigner crystals and cold melting below, and in Section 11.8).
When will a substance melt by quantum fluctuations and remain liquid down
to T = 0? According to (4.73), better chances for this exist if:
the mass of the atom M is small;
the interaction v is weak.
The best candidate for this is helium: it is light, and He atoms have filled 1s
shells (inert atoms), so that the HeHe interaction is weak. That is why helium
remains liquid down to T = 0 (at normal pressure).
Another good candidate could have been hydrogen. It is even lighter than
helium. But the effective interaction between hydrogen atoms and even between H2
molecules is too strong. Nevertheless there have been active experimental attempts
to stabilize hydrogen in a liquid phase down to the lowest temperatures, i.e. to
prevent its crystallization, with the idea that it would then experience Bose condensation and possibly would become superfluid, as 4 He. These attempts have not yet
succeeded; instead Bose condensation was reached in completely different systems,
in optically trapped and supercooled alkali atoms (Rb, Cs, etc.).
As we have said, the condition for quantum melting is determined by the value of
the quantum de Boer parameter . Typical values of are: 3 He = 0.5, 4 He 0.4,
H2 0.3, Ne 0.1. (In the book by Ashcroft and Mermin (1976) other values
for are given, which is due to the different normalization used there, but the ratio
for different elements is the same.)
51
T
solid
electronic
liquid
(metal)
liquid
superfluid
He II
electronic solid
(Wigner crystal)
gas
(a)
(b)
Fig. 4.13
Under pressure the average distance a between atoms decreases, the atom
atom interaction v (a) increases, and may decrease or increase depending on
the behaviour of v a 2 (see (4.73)). For neutral particles (helium) the interaction
v increases faster, decreases, and He becomes solid under pressure (the phase
diagram of 4 He is shown schematically in Fig. 4.13(a)). For electrons the opposite
is true: thecharacteristic Coulomb interaction between electrons is v = e2 /a, and
h/a me2 /a 1/ a), where a is the average distance between electrons.

=
Under pressure this distance decreases, the de Boer parameter increases, and there
occurs cold melting of the electronic (Wigner) crystal (Fig. 4.13(b)).
When helium crystallizes under pressure, it still has a large value of , i.e. large
quantum fluctuations. This means that it is a quantum crystal, for which quantum
effects are important. Thus, for instance, the vacancies in solid He are very mobile,
they behave like quasiparticles. There is even discussion of a possible superfluidity
of vacancies in solid He (superfluidity in solid supersolid). In 2004 there
appeared the first experimental indications that it could indeed be true, although
this question is still controversial.
Yet another manifestation of quantum effects in solid He is their importance
for the exchange interaction, especially in solid 3 He. The exchange interaction in
solid 3 He (3 He is a fermion!) is usually due to ring exchange in the presence of
vacancies, see Fig. 4.14. As a result of this process, after three steps the atoms 1 and
2 interchange their positions. This can be shown to lead to a partial ferromagnetism
of solid 3 He.
4.4.4 Low-dimensional solids; why is our world three-dimensional?
Problem: Using the approach described above (see, e.g. equation (4.65)), discuss
what would be the situation with crystals in the one-dimensional (1d) and twodimensional (2d) cases.
52
Phonons in crystals
vacancy
Fig. 4.14
Solution: Similar to the single oscillator case, see (4.65), in a crystal the average
vibration amplitude u2 is

h
h
1
1
2
=
bk bk +
nk +
u =
Mk
2
Mk
2
k
k

dd k
h
1
1
.
(4.74)
=
+
(2 )d Mk ek /T 1 2
1. First consider the case of zero temperature. At T = 0, nk = 0, and

dd k
h
1
2
u =
.
d
(2 ) Mk 2
(4.75)
1(a) At k 0, k = sk (acoustic phonons, s is the sound velocity). Thus for

d = 1 (one-dimensional system)

dk
2
,
(4.76)
u
sk
i.e. u2 is logarithmically divergent! This means that in a 1d system even
at T = 0 u2 , and as a result there is no long-range crystalline
order in this case! Zero-point vibrations are so strong that they destroy the
ordered state even at T = 0! Mathematically this divergence is due to the
behaviour of the integral (4.75) at the lower limit of integration k 0,
or 0; this is what is called an infrared divergence. The upper limit
of integration is determined by the upper edge of the spectrum, which in
solids is finite, so that usually there are no divergences there. Therefore
here and in the future we do not specify this upper limit of integration.
1(b) T = 0, 2d system. Here d 2 k k dk, and everything is OK,

k dk
2
u
,
(4.77)
sk
the integral is convergent, the mean square vibration amplitude u2 is
finite, and in general there may exist long-range crystalline order at T = 0
in 2d case.
1(c) And of course everything is fine in 3d systems, where d 3 k k 2 dk.
53
2. T
= 0. Again the behaviour as k 0 is critical; thus we consider the region
k = sk < T . This part of the spectrum exists if the spectrum is gapless. In this
region
nk =
1
ek /T
and from (4.74) we get

u
2
1
T
=
,
k /T
sk

dd k T
1
.
+
k k
2
(4.78)
In this expression again the most dangerous part is that close to k = 0 or = 0,

and we can ignore the term with 12 in the integrand.
2(a) 1d case: u2 was divergent already at T = 0. At T
= 0 it is even more
2
divergent
# dk the corresponding expression for u would be proportional
to T s 2 k2 and would diverge not logarithmically, but linearly.
2(b) 2d case: The most dangerous term has the form

k dk T
k dk
u2
T
,
(4.79)
2
s 2k2
k
i.e. in the 2d case, u2 is logarithmically divergent at any finite T . Thus
at finite temperature there is no long-range order either in 1d or in 2d
systems.
2(c) 3d case. Here all is quiet even at nonzero temperatures:
2
k dk
,
(4.80)
u2 T
s 2k2
which is convergent, so that the fluctuations are finite. Luckily for us!
Otherwise everything surrounding us, and maybe we ourselves, would not
be stable. Our bones, and all other tissues would melt. (If you like, this
may be the physical explanation of why we exist in a three-dimensional
world.)
There exists a general theorem the MerminWagner theorem which states
that whenever an ordering corresponds to a breaking of continuous symmetry,
there is no long-range order in one-dimensional and two-dimensional cases at
any nonzero temperature. Actually this theorem is intrinsically connected with the
Goldstone theorem about the presence of gapless Goldstone excitations for a broken
continuous symmetry, mentioned above in Section 2.7.4: we saw above that for the
divergence of the mean square vibration amplitude u2 and consequently for the
instability of the crystal, it is crucial that the energy spectrum k which stands in
the denominator in equations (4.74)(4.80) should be gapless, k 0 for k 0.
All these features are especially important in different magnetic systems, see below,
Chapter 6, but also in low-dimensional superconductors, etc.
5
General Bose systems; Bose condensation

There exist in nature different kinds of bosons. These may be phonons or photons.
Their number is not conserved, and consequently their chemical potential is = 0,
see the previous chapter.
We meet a different situation in the case of systems of bosons with conserved
particle number. These are, for example, atoms or molecules with an even spin,
such as 4 He. For these cases in general the chemical potential is
= 0; it is
determined by the condition (1.25), that is by the requirement that the total number
of particles, or particle density, is fixed. In these cases we meet the phenomenon of
Bose condensation.
Consider ideal noninteracting bosons with the spectrum p = p2 /2m. From
equation (3.3) we obtain the number of bosons in a unit of the phase space
d 3 p/(2
h)3 :
dn p =
1
d3 p
,
3 ( p )/T
(2 h) e
1
(5.1)
or
m3/2
n=
3
2 2
h
d
.
()/T
e
1
(5.2)
To show that the phenomenon of Bose condensation considered below is intrinsih in an apparent
cally a quantum phenomenon, we keep here the Plank constant
way, and do not put it equal to 1, as elsewhere in this book. For a given density
(given number of particles) equation (5.2) is an equation for the chemical potential (T ). At high temperatures, (T ) < 0 (cf. (3.10) and Fig. 3.1). As discussed
in the problem in Chapter 3, the chemical potential increases with decreasing
temperature, and it tends to zero at a certain temperature T0 , given by the condition
54
55
T0
Fig. 5.1
(we use below the dimensionless variable z = /T )

m3/2 T 3/2 z dz
,
n=
3
2 2
h 0 ez 1
(5.3)
see Fig. 5.1. The integral in (5.3) is finite; it can be expressed through the Riemann
function,
x1
z
dz
= (x) (x) ,
(5.4)
z
e 1
0
where (x) is the gamma function.
Similarly, more general integrals of this type are
x1
z
dz
= (1 21x ) (x) (x) .
z
e +1
0
(5.5)
These formulae will also be useful for fermions later on.
In our case (x = 32 )
3
2
,
2
3
2
2.61. This finally gives
h 2/3
3.31
(5.6)
n .
m
As follows from the general principles of statistical mechanics, the chemical potential (T ) for Bose particles should always be negative (or, in general, should lie
below the bottom of the corresponding band); otherwise the sum over N (from 0
to ) in (1.33), (1.32) would not converge, see Chapter 3 (for free bosons we have
to put in (1.33) EnN = p N, where p = p2 /2m).
For T < T0 there is no solution of equation (5.2) with < 0, but that is a
necessary requirement for a Bose system! Indeed the left-hand side of equation (5.3)
is constant, whereas the right-hand side goes to zero as T 0.
The solution of this apparent paradox is the following: there is a macroscopic
occupation of the state with p = 0 or with =# 0. The transition from summation

over discrete values of p, p , to the integral d 3 p/(2 )3 , which we usually do
2
T0 =
56
General Bose systems
when going from a finite system to a system with infinite volume and which we
automatically did in writing equation (5.2), is not valid in this case. Due to the
macroscopic occupation of one particular state, here the state with p = 0 (which
is allowed for bosons!), the number of particles in this state is infinite, and in the
corresponding summation over p the first term ( p = 0, = 0) tends to infinity. The

remaining sum p
=0 can then be transformed into an integral in (5.3), which then
can be finite. There is no contradiction any more: the number, or density of particles
for all > 0, given by the expression (5.3), can indeed go to zero as T 0, but
the total number of particles can still be conserved: the missing particles are now
in the condensate, in the state with p = 0 and = 0.
Thus at T < T0 , where the chemical potential is identically zero, we have for
the states with > 0,

d
V m3/2

,
(5.7)
dN
=
3
/T
>0
1
2 2
h e
and the total number of particles with > 0 is
3/2

V (mT )3/2 z dz
T
N>0 =
dN =
.
=N
z
3
2
e 1
T0
2 h
0
The remaining
+
N=0 = N N>0 = N 1
T
T0
(5.8)
3/2 ,
(5.9)
particles are in the state p = 0 in a condensate. This is the Bose condensation

(or BoseEinstein condensation).
Problem: Check what would be the situation with Bose condensation of an ideal
Bose gas in one-dimensional and two-dimensional systems.
Solution: Why does Bose condensation occur in 3d systems? In equations (5.2),
(5.3) (with the minimal possible value of the chemical potential = 0) the integral
converges. Then as T 0 we cannot fulfil the condition (5.3), and we have to put
an infinite number of particles in one particular state in a condensate.
It turns out that in 1d and 2d cases the corresponding integrals diverge, and there
exists a solution of similar equations with
= 0 for all T > 0. As a result there
occurs no Bose condensation at any nonzero temperature in 1d and 2d cases!
Indeed, the equation similar to (5.2) has, in general, schematically the form

d () d
(5.10)
n = const.
()/T
e
1
0
57
where at small the density of states d () for the 3d case (d = 3), 2 ()
2 = const., and 1 () 1/ .1
Let us make a change of variables, /T = z, /T = .
We then obtain

dz
n = Tc
(2d case) (5.11)
ez 1
and

dz

n = c T 1/2 z
(1d case) (5.12)
z e
1
(c is a certain constant). The integral in (5.3) is finite even for = 0, and when
T 0 we have a contradiction: the left-hand side of equation (5.3) is finite, and
the right-hand side goes to zero. The resolution of this paradox leads to Bose
condensation. However, for the 2d case the integral in (5.11) is logarithmically
divergent for = 0. Thus when T 0 we can compensate the small factor T
in (5.11) by the corresponding increase of the integral,
choosing the appropriate
#
dependence (T
) (
= 0) so that the product T c dz/(ez 1) remains finite
(equal to n).
The same is true also for the 1d case: the integral in (5.12) diverges even more
strongly than in the 2d case, which means that there should be no Bose condensation
in one-dimensional systems either.
In an ideal Bose gas one can calculate all thermodynamic functions at the Bose
condensation transition. It turns out that thermodynamic functions E, F , , S, cV
are continuous, i.e. this transition is not even a second-order phase transition (at
the second-order phase transition there is a jump in c). Here, not cV but dcV /dT
has a jump, i.e. it is a third-order phase transition.
But:
(1) The situation would be different if we were to work not at fixed volume (or
fixed density of particles), as we have until now, but at fixed pressure. In this
case Bose condensation becomes a real second-order phase transition even for
an ideal Bose gas.
(2) The interaction between bosons is especially important it will also change
the order of the transition. This we will consider in the next section.
1
Why do we have this form of the density of states d () at the edge of the spectrum 0? Say, for the 3d case
we had initially d 3 p, with d 3 p p 2 dp which we transformed to ()d. In the 3d case, with the spectrum
(p) = p 2 /2m, this gives

p2
p2

()
p .
3d
d/dp
p
Similar considerations show that:
In the 2d case: d 2 p p dp = () d, () p/(d/dp) const.
In the 1d case: d 1 p dp = () d, () 1/(d/dp) 1/p 1/ .
We will often use these asymptotics of the density of states later on.
58

Let us include an interaction between bosons
v (r r ) n(r) n(r ) = v (r r ) (r) (r) (r ) (r ) ,
(5.13)
which we will treat as weak.

In the second quantization form it is convenient to work in the momentum
representation. The Hamiltonian in this representation is
H = H0 + Hint =
(or

p1 + p2 = p3 + p4
p2

v (q) a p+q a p q a p a p
a p a p +
2m
p
p p q
(5.14)
v a p a p a p a p ). Here, in the beginning, we keep the notation for

1
operators.
For simplicity we take the interaction as constant, v = U/2V (the volume V in
the denominator is needed for normalization). Physically this corresponds to the
assumption that the interaction is point-like, and that it does not depend on the
angle, so-called s-wave scattering.
In the Bose condensed state there is a macroscopic occupation of the state p = 0
by N0 particles. This means that for the momentum p = 0,
a 0 a 0 = N0 N ,
i.e. a0
(5.15)
N 1. In this case the commutator
[a 0 , a 0 ] = a 0 a 0 a 0 a 0 = 1
(5.16)
is small compared with a0 itself, and therefore commutation relations for the zero
momentum operators a0 are not important. This means that for this particular state
we may treat a0 , a0 as ordinary c-numbers and not as operators.
Note: a0 , a0 are complex numbers, that is
N0 ei ,
a0 = N0 ei .
a0 =
(5.17)
The phase is in general very important, as we will discuss later. In our present
discussion, however, we always have bilinear combinations of the type a a, and at
this stage we do not have to worry about
the phase.
Thus we can treat a0 , a0 as (large, N) c-numbers, and a p , a p as (small) operators. Let us make this substitution and keep the leading terms in the Hamiltonian,
of order N 2 and N:
Hint =
59

U )
(2a p a0 a p a0 + 2a p a0 a p a0
a 0 a0 a0 a0 +
2V
p
=0
*

+ a p a p a0 a0 + a0 a0 a p a p )
*

U ) 4
(4a p a p + a p a p + a p a p ) .
a0 + a02
2V
p
=0
(5.18)
For noninteracting particles at T = 0 all particles are in the condensate, N0 = N.

For weakly interacting bosons almost all particles will be there, N0 N, N
N0 (N, N0 ). In the second term in (5.18) we can put a02 = N (the second term
is already of first order in the small parameter, due to the presence of ap ap ). The
term a04 should be treated more accurately, keeping all terms of the same order. As

a02 +
(5.19)
a p a p = N ,
p
=0
we should write
2

a0 = N
a p a p = N 2 2N
a p a p +
a p a p a p a p . (5.20)
p
=0
p,p
=0
The last term here is only of order 1, thus it can be omitted. In effect equation
(5.18) becomes
*

U ) 2
Hint =
(2a p a p + a p a p + a p a p ) ,
(5.21)
N +N
2V
p
=0
and the full Hamiltonian (5.14) takes the form

UN
U N 2 1 p2
UN
+
+
(a p a p + a p a p ) +
(a p a p + a p a p ) .
H=
2V
2 p
=0 2m
V
V
(5.22)
The expression (5.22) is nondiagonal in the operators a p , a p . But it is a quadratic

form and can be easily diagonalized using the so-called Bogolyubov canonical
transformation, or uv transformation:
a p = up bp + vp bp
a p = up bp + vp bp
(here we assume the coefficients up , vp to be real).
Problem: Write down b , b in terms of a , a.
(5.23)
60
Solution: bp = up ap vp ap (up , vp are even functions of p), and the corre

sponding equation for bp is obtained from this one by Hermitian conjugation.
We require that (we now omit the notation for operators) b , b are bosons:
bp bp bp bp = pp ,
b p b p bp b p = 0 .
(5.24)
In order for this to hold, the coefficients up , vp should satisfy the condition
u2p vp2 = 1 .
(5.25)
Thus we can write

ap =
ap
=
1
1
A2p
(bp + Ap bp )
(5.26)
(bp
+ Ap bp ) ,
1 A2p

where up = 1/ 1 A2p , vp = Ap / 1 A2p (or, equivalently, we can write down
up = cosh p , vp = sinh p , and express everything through p ). We can determine the coefficients up , vp , or Ap from the condition that the nondiagonal terms
in (5.22), after making the canonical transformation, drop out. This gives the
equation
2

p
+ nU 2Ap + nU (1 + A2p ) = 0 ,
(5.27)
2m
or

2
2
2
p
1 p
Ap =
(5.28)
nU +
+ nU (nU )2 ,
nU
2m
2m
where we have introduced the density n = N/V . The Hamiltonian then takes the
form

2

2
p
U N 2 1 p2
H=
+ nU
+ nU (nU )2
2V
2 p
=0
2m
2m

2
p2
+ nU (nU )2 bp bp .
+
2m
p
=0
(5.29)
The first two terms give the energy of the ground state at T = 0, and the third term
describes elementary excitations (cf. phonons, (4.26)),

p bp bp ,
(5.30)
p
=0
61
ep
p
Fig. 5.2
with the spectrum

p =

nU
2
p
p2
m
+ nU (nU )2 =
2m
p2
2m
(p 0)
(5.31)
(large p).
Thus the elementary excitations in a weakly interacting Bose condensed system

have the character of sound at small momenta p (Bogolyubov sound), and they
continuously go over to free particles with the spectrum p2 /2m at large p, see
Fig. 5.2.
We should make here several remarks:
The sound velocity (5.31) coincides with the standard sound velocity of a gas
with density n = N/V and interaction U .
One can find the coefficients of the canonical transformation (5.23) or (5.26) not
from the condition that nondiagonal terms in the Hamiltonian cancel, but from
the condition of the minimum of the ground state energy.
Problem: Check this: put (5.26) in (5.22), collect the terms without operators,
and minimize the resulting expression in Ap .
One can calculate the total energy of this weakly interacting Bose gas. Usually
one expresses all the quantities through the scattering length a given by
U=
4
a
m
(5.32)
(a is the scattering amplitude for energy tending to zero). The resulting energy
is given by the expression

E
2 a 2
128 3/2 1/2
=
n 1+ a n
(5.33)
V
m
15
(T. D. Lee, C. N. Yang, 1957).
62
ep
p0
Fig. 5.3
p0
Fig. 5.4
From (5.33) one can find the sound velocity in the usual way, as we are doing
in gases:

V 2 2E
4 an
,
(5.34)
=
s=
mN V 2
m
which coincides with (5.31) (taking into account (5.32)).
The main system for which, for a long time, one applied the concept of Bose
condensation (before the recently observed Bose condensation of optically trapped
supercooled atoms) is superfluid 4 He. 4 He atoms are bosons, and helium remains
liquid down to T = 0 (see Section 4.4.3). It goes over to a superfluid state at
Tc = 2.4 K (P. L. Kapitza). Superfluidity in 4 He is attributed to a Bose condensation
of He atoms into the state with p = 0.
There exists in real helium one important difference relative to the previous
treatment: He atoms interact strongly, as a result of which there is a number of
modifications and because of which there exists actually no microscopic theory of
superfluid 4 He (there is no small parameter weak interaction which we have
used in the theoretical treatment above). Nevertheless, the basic concepts described
above apply, with some modifications, also to this case. The main modifications are:
1. The actual number of atoms in the condensate N0 is finite, but even at T = 0 it
is rather small, N0 /N 810% at most.
2. The spectrum of elementary excitations (originally postulated in a slightly
different form by Landau) has the form shown in Fig. 5.3. As p 0 the
excitations remain phonons (spectrum linear in p). At larger p there exists
an extra minimum the so-called rotons. Physically the roton minimum is a
consequence of strong interaction in the liquid; it reflects the tendency to crystallization (under pressure the roton minimum becomes deeper, and when p0
approaches zero, Fig. 5.4, the corresponding mode becomes unstable, and there
appears a positive increment, i.e. a large standing wave with wavevector p0 , or
h/p0 , will develop; this signals the formation of a crystal). Cf.
with period a =
the discussion of the reverse process of melting in Section 4.4.3, Fig. 4.13(a).
63
p /a
p /a
Fig. 5.5
period of the crystal!
2 p/a
p/a
Fig. 5.6
Problem: Compare this situation with ordinary phonons in crystals.

Solution: For an ordinary crystal with period a the phonon spectrum in the
first Brillouin zone has the form schematically shown in Fig. 5.5. In the
extended zone scheme it may be represented as shown in Fig. 5.6, with the plot
extended to larger p. If we do not have real long-range crystalline order, but
only short-range order typical for liquids, the spectrum would not go to zero
at p = 2/a, but would retain a minimum at this wavevector (and of course
would acquire an imaginary part, i.e. would be strongly damped). Thus there is
a close analogy between this situation and that of rotons in liquid helium shown
in Figs. 5.3 and 5.4; that is, the spectrum of Fig. 5.4 with the roton minimum
is not something specific to liquid 4 He, but in principle such a minimum could
exist in any liquid close to the melting point. An important difference is that in
normal liquids these excitations with short wavelengths are usually extremely
strongly damped, so that there is not much sense in speaking about them there,
whereas they are well-defined excitations in the superfluid phase of 4 He.
These arguments are of course only qualitative. But one can make them
more accurate. R. Feynman has shown (1954) that the spectrum of elementary
excitations in 4 He can be written as
p =
p2
,
2mS(p)
(5.35)
64
where S(p) is the static structure factor describing spatial correlation of atoms
in a liquid. At small p S(p) = p/2ms, where s is the velocity of Bogolyubov
sound, s = nU/m, see (5.31), so that the spectrum (5.35) goes over to (5.31).
h/d0 where d0 is
For large p, however, S(p) has a maximum at the value p0
the typical distance between He atoms (of the order of the lattice parameter of
solid 4 He at high pressures); this tells us that the probability of finding an atom
at distance d from a given atom is a maximum at d d0 . Correspondingly, the
spectrum (5.35) would develop a minimum at p p0 which is nothing else but
the roton minimum of Fig. 5.3.
3. An important point: in Bose condensation, and in superfluidity, the order
parameter (cf. (2.1)) is a 0 (it is zero above Tc , and nonzero in the Bose
condensed phase). This is a complex scalar (see the remark (5.17) in Section 5.2):

a 0 = a0 = N0 ei .
(5.36)
At the phase transition, at T < Tc , its phase becomes fixed, i.e. the superfluid
state is a coherent state (phase coherence). But the number of particles in the condensate N0 fluctuates. The operators N and are conjugate variables in quantum
they obey the uncertainty relation (Heisenberg relation)
mechanics, like x and p;
N0

h.
(5.37)
Bose condensation is thus a phase transition with breaking of a continuous

symmetry gauge symmetry (fixing of the phase which is a continuous
variable). Correspondingly, Bogolyubov sound is the Goldstone mode for this
broken symmetry. This also gives another explanation of the absence of Bose
condensation at any T
= 0 in the 1d and 2d systems discussed in the Problem
in Section 5.1: phase fluctuations caused by the excitation of this gapless mode
are so strong that in the 1d and 2d cases they destroy the long-range order at
any finite temperature. This is yet another application of the MerminWagner
theorem mentioned at the end of Chapter 4.
4. The fact that the order parameter in Bose condensed and superfluid systems
is a complex scalar (5.36) permits one also to establish the correspondence
with certain other systems with the same symmetry, notably the anisotropic
spin system the so-called xy model, in which spins are confined to the
xy-plane. In this case the order parameter takes the form s = |s|ei , where
the angle determines the orientation of spin in the xy-plane (see below,
Section 6.4.3). One can use this analogy by borrowing the concepts and results
from one field and applying them to the other. Thus one often speaks now about
Bose condensation of magnons, although one must be careful in applying this
65
concept. Also, the concept of vortices, first introduced in superfluid 4 He, is now
translated to magnetic vortices, see Section 6.4.3(b), p. 118.
5. Generalizing the description of a Bose condensed state to the case with
spatial inhomogeneities (cf. Section 2.4 above), we should treat the order
parameter (5.36) as a function of position. The most important is the change in
space of the phase (r). It turns out that the gradient of the phase determines
the local superfluid velocity in the system:
h
.
(5.38)
m
Actually the collective mode we have described above Bogolyubov sound is
predominantly the oscillations of the phase and velocity v s (or local currents)
in the liquid.
vs =
5.3.1 Landau criterion of superfluidity

Why is 4 He superfluid? The answer to this question is given by the Landau criterion
of superfluidity: it turns out that for the excitation spectrum of Fig. 5.3 there is no
dissipation at velocities not exceeding a certain critical value. The corresponding
arguments are straightforward, but require a bit of concentration.
Suppose that a liquid flows with velocity v through a capillary (thin tube).
The elementary processes leading to friction are the creation of excitations in the
liquid, one after another. These excitations should reduce the total momentum of
the flowing liquid. If such processes are allowed, friction would appear, and there
would be no superfluidity.
Let us first consider this situation in the system of coordinates where the liquid
is at rest but the walls of the tube move with velocity v. Suppose that one (the
first) such an excitation appears, with momentum p and energy ( p). Now let us go
back to the laboratory coordinate frame, in which the liquid moves with velocity v.
In this coordinate system the total energy of the liquid with one excitation is2
E=
Mv 2
Mv 2
+ ( p) + p v =
+ E ,
2
2
E = ( p) + p v .
(5.39)
2
Here Mv /2 is the initial kinetic energy of the moving liquid, and E is the change
in the total energy due to the creation of the excitation. For such a process to
occur we need E < 0, otherwise the excitation would cost us energy, and such
2
where
This follows from the well-known formulae of mechanics: energy and momentum are transformed from one
reference frame to another, moving with velocity v, as E = E0 + p0 v + Mv 2 /2, p = p0 + Mv, where E0
and p0 are the energy and momentum in the frame with the liquid at rest. In our case E0 = ( p), p0 = p, which
gives (5.39).
66
ep
ep
crit. =
Fig. 5.7
Fig. 5.8
ep
crit. roton
p
Fig. 5.9
an excitation would not be created spontaneously. But for that p should, first of
all, be antiparallel to v, and, most important, the change of the energy E (5.39)
should be negative, i.e. pv < 0. Thus for such a process (friction) to begin, the
velocity of the liquid should satisfy the condition
v>
( p)
p
(5.40)
(at least for some p). That is, the excitations can be created only if the velocity of
the flow exceeds the critical velocity,

v vcrit =
,
(5.41)
p min
starting from which the flow can slow down, and dissipation, or friction, appears.
If we have a usual liquid or gas, consisting of noninteracting particles, the
excitation spectrum is p = p 2 /2m, Fig. 5.7, and there is no superfluidity at any
velocity (the critical velocity vcrit = 0). But if the spectrum has the form shown
in Fig. 5.8, see equation (5.31), then there exists a finite critical velocity vcrit = s,
which in this case is equal to the sound velocity; for smaller v such excitations
cannot be formed, and the motion is dissipationless, i.e. superfluid.
For real 4 He the sound velocity s = 2.4 104 cm/sec, and this condition would
give the critical velocity which is too large. Experimentally dissipation starts much
earlier. One of the possible explanations is that the real spectrum looks as shown
in Fig. 5.9 and the critical velocity could be determined by rotons. Actually even
67
this velocity is too large, and the excitations determining the critical velocity in the
bulk helium in most of the actual experiments are special topological excitations
vortices (similar to smoke rings from a pipe) see below, Section 5.3.2. Only in
cases when the formation of vortices is suppressed due to a restricted geometry, for
example in thin films of 4 He or in very thin capillaries, can we reach the critical
velocity determined by rotons, which is indeed much higher than in bulk 4 He.
We have seen that if the velocity is below a certain critical value, new excitations
cannot be spontaneously created in the moving liquid. This conclusion is valid not
only at T = 0 and not only for the ground state. However, at finite temperature
there are always present thermally excited elementary excitations in the liquid. And
when the liquid flows through a capillary, these excitations can collide with the
walls and can change their momentum. Therefore these excitations, which initially
flow with the liquid, will gradually slow down, exactly like ordinary gas flowing
through a tube. In effect it looks as though a part of the liquid experiences friction,
whereas the remaining part moves without any resistance. In other words, it seems
as though there exist two components in the liquid; a normal component and a
superfluid one, with the total density = norm + s . Such a two-fluid picture
(L. Tisza) gives a very useful phenomenological description of many properties
of superfluid helium. It is also widely used for the description of many properties
of superconductors. But one has to realize that it is only a way to interpret the
properties of these systems; in no way should we take this picture too literally and
think that indeed some of the atoms are moving without dissipation, whereas the
others experience friction. In fact, it is the same atoms which display both types of
behaviour, and the real meaning of these two fluids is the one explained above: the
normal fluid consists of elementary excitations collective modes of the liquid
as a whole.
5.3.2 Vortices in a superfluid
As mentioned in the previous section, there is yet another very important type of
excitation in a superfluid liquid: topological excitations, or vortices. If one starts to
rotate a vessel containing a superfluid, initially the liquid remains at rest. However,
starting from a certain critical angular velocity, a vortex will be formed in the liquid:
a circular motion of the superfluid around a certain line which is called the vortex
core. As a result some circulation is transferred to the liquid, i.e. the liquid starts
to participate in the rotation. In a cylindrical vessel vortices start at the bottom
and go all the way up to the upper surface of the liquid, Fig. 5.10; they cannot be
interrupted and cannot simply end inside the liquid.
Using equation (5.38), we can show that the circulation of velocity around a
vortex should be quantized. Let us integrate (5.38) over a contour c surrounding
68
Fig. 5.10
Fig. 5.11
the vortex, Fig. 5.11. The circulation is

1
h
v s dl =
,
m
(5.42)
where
is the total change of phase along the contour c. However the total
wavefunction of the superfluid has to be a single-valued function, which means
that
= 2 n, where n = 0, 1, 2, . . . . Thus the circulation is
1
h
(5.43)
v s dl = 2 n ,
m
i.e. it is indeed quantized. Such quantization is a manifestation of the quantum
nature of superfluidity and confirms its interpretation as a macroscopic quantum
phenomenon
(the whole system is described by one quantum wavefunction (r) =
N0 ei(r) ). The above-mentioned fact that the vortex cannot end inside the liquid
is actually connected with this property: if it were to, we could continuously deform
the contour c in equation (5.43) in such a way that it would be above the end of the
vortex, after which we could contract it to zero, with zero circulation in contrast
to the finite (quantized) value (5.43) we started with. This is why the vortices are
actually topological excitations.
The situation may be much more intricate in superfluid systems with a more
complicated order parameter, e.g. such as that of liquid 3 He. In these cases the
space of the order parameter is different from the case of 4 He (which lives on
the circle = ||ei with 0 2 ). Consequently the types and properties of
69
Fig. 5.12
Fig. 5.13
topological excitations in such systems, including vortices, may be very different

from those of 4 He.
As mentioned above, in most of the real experiments the critical velocity in
liquid helium is determined not by the sound velocity and not even by rotons, but
by excitation of vortices in moving liquid He. Vortices are created before the critical
velocity for excitations of sound quanta or rotons is reached, and their creation and
motion leads to dissipation. In a rotating cylinder vortices are parallel to the rotation
axis, Fig. 5.12. In a liquid flowing through a capillary the vortices form vortex rings,
Fig. 5.13, very much like smoke rings from a pipe (as we mentioned before, vortices
cannot be simply interrupted inside a liquid: they either end at the surface of the
liquid, or form closed loops, as in these rings).
Many features and phenomena described above are also met (and are even much
better known) in superconductors. Thus, the general description of superconductivity as a macroscopic quantum phenomenon is rather similar to that given above; the
Landau criterion of the existence of supercurrents without dissipation works there
as well, and the notion of quantized vortices leading to dissipation plays a very
important role in the physics and application of superconductivity (in the so-called
type-II superconductors, to which all practically important superconducting materials, including high-temperature superconductors, belong). Theoretical work by
Abrikosov and Ginzburg laying the foundations of the description of these systems
earned them a Nobel Prize in 2003.
6
Magnetism
6.1 Basic notions; different types of magnetic response

I will again start by briefly recalling the basic notions from general physics and
quantum mechanics about different sources and types of magnetic response, which
are actually covered in most of the corresponding textbooks. For simplicity I will
mostly ignore in this chapter the details of atomic structure leading to different
contributions to atomic magnetic moments (e.g. orbital contributions, with the
often important role of spinorbit coupling), see e.g. Goodenough (1963) and
Kugel and Khomskii (1982), and will consider only the spin contribution, mostly
illustrating the results on the example of spin 12 . More detailed treatments can be
found, e.g. in the books by White (2006) and Yosida (1996), and in many others.
The Hamiltonian of electrons in a magnetic field has the form
e 2
1
p A
S H .
(6.1)
H=
|S|
2m
c
The first term in this Hamiltonian describes the response due to the orbital motion
of electrons, and the second one is due to the spin of the electron. Here is the
magnetic moment of spin S. [Often the last term in the Hamiltonian (6.1) is written
h/2mc is the Bohr magneton, and g is the so-called gas gB S H, where B = e
factor, which for free electron spins is gspin = 2 (and for orbital moments gorb = 1).]
In general the vector potential A is a function of the coordinate, A = A( x ), and it
does not commute with the momentum p , p A A p = i
h div A. However,
we can make a gauge transformation of A, and if we choose a gauge such that
div A = 0, the commutator of p and A will be zero. For example, we can take
[H r] .
(6.2)
e2
1 2
e
2
p
p A +
A
S H .
|S|
2m
mc
2mc2
(6.3)
A=
1
2
Then
H=
70
6.1 Basic notions
71
As to the magnetic response of various systems, there exist different situations:

(1) Closed shells with orbital moment L = 0 and spin S = 0. In the ground state
0| p A|0 = 0, and here first of all the third term in (6.3) works (see also (2)
below); the energy is
E = H =

(2
e2
e 2 '
e2
2
2

A
=
H
r
=
H
ra2
a
2mc2
8mc2 a
12mc2
a
(6.4)
after averaging over angles, i.e. over all directions of r.

The energy of a system in a magnetic field is
E
.
(6.5)
H
As the magnetic moment is M = H, the susceptibility for N such atoms with
charge Z is, using (6.4),
Ze2 N r 2
e2 2
r
=
.
(6.6)
=
6mc2 a a
6mc2
E = M H,
i.e.
M=
This is the standard diamagnetism classical diamagnetic screening of the

external field. Note again that it is due to the term A2 in the Hamiltonian (6.3);
this is also the case in other situations, e.g. in the famous Meissner effect, the
ideal diamagnetism of superconductors.
(2) The term in (6.3) linear in A can admix the excited states with L
= 0 to
the ground state: in second order in perturbation theory (using the second term
in (6.3) as a perturbation) we then have the change of the energy in the magnetic
field
2

0|H
M
|n

L
E =
(6.7)
n 0
n
e
L, L = r p, and we have used the gauge
(here we have used that M = 2mc
(6.2)).
This would give positive susceptibility, which in a first approximation does
not depend on temperature. This is the temperature-independent Van Vleck
paramagnetism.1
The two terms described above, the usual diamagnetism and Van Vleck paramagnetism, are always present in all materials, even in those with much stronger
magnetic response due to localized spins; these contributions are responsible
1
Note that in certain cases, where there exist low-lying magnetic excited states close to the nonmagnetic ground
state, the Van Vleck paramagnetism may become temperature dependent. This, for example, is the case for
many compounds containing Eu3+ .
72
Magnetism
Fig. 6.1
for the temperature-independent background in magnetic susceptibility. In

metals there are two extra contributions to the magnetic response.
(3) Free electrons in metals. There exists a diamagnetic contribution to the susceptibility due to the orbital motion of electrons, which is called Landau
diamagnetism:
Landau =
(N/V )e2
e2 pF
=
.
12 2 mc2
4mc2 pF2
(6.8)
The calculations giving this expression are not simple; the physics is connected
with the boundary effects. In classical physics one can show that there is no
magnetism in thermodynamic equilibrium (Bohrvan Leeuwen theorem): if
we apply an external field H to a metal with classical electrons, the diamagnetic currents created inside the sample will be compensated by the surface
current along the boundary flowing in the opposite direction, see Fig. 6.1. In
quantum mechanics there is no such compensation, which results in Landau
diamagnetism.
The physical reasons underlying the Bohrvan Leeuwen theorem may be
understood if we recall that classically the force acting on an electron in a
magnetic field is the Lorentz force F H v, i.e. it is perpendicular to the
velocity v and consequently it does not change the energy of the electron. In
effect the magnetic field does not enter the thermodynamic potentials and does
not induce a magnetic response in thermodynamic equilibrium. In quantum
mechanics this is no longer true.
(4) Spins of electrons in metals give rise to Pauli paramagnetism. Its origin is the
splitting and shift of the spin-up and spin-down subbands in a magnetic field,
(H ) = 0 H . This leads to the redistribution of electrons between these
subbands, see Fig. 6.2, because the chemical potential of both these components
should be the same. As a result there appears a net polarization of conduction
electrons, proportional to the magnetic field, M = Pauli H. Straightforward
calculations give for Pauli the expression
Pauli = 2B (F ) =
e2 pF
.
4 2 mc2
(6.9)
6.1 Basic notions
73
Fig. 6.2
It turns out that the Landau diamagnetism discussed above is determined by

the same parameters as in equation (6.9), i.e. by the density of states at the
Fermi level (F ), and is equal to
Landau = 13 Pauli .
(6.10)
(5) Localized electrons (localized spins, localized magnetic moments). Why localized electrons exist in certain systems will be discussed later, especially in
Chapters 12 and 13; briefly speaking, this is connected with the existence
of partially filled inner shells and with strong electronelectron interactions.
Typical systems of this type are those containing the following:
Transition metals with partially filled d shells. In the 3d series (Mn, Fe,
Co, Ni, . . . ) the d electrons are relatively strongly localized, especially in
compounds such as oxides, e.g. NiO, MnO, Fe2 O3 , etc. In the 4d and 5d
elements (Ru, Ir, Pd, Pt, . . . ) the d electrons are usually less localized than
those of the 3d series.
Rare earth elements, containing 4f electrons (Gd, Eu, Dy, . . . ). The 4f
electrons are almost always very strongly localized (see, however, Section
13.3).
Actinides with 5f electrons (U, Np, Am, . . . ). From the point of view of
electron localization these are rather analogous to the 3d transition metals.
However, there is one important difference here: due to the much larger
atomic mass the relativistic effects in actinides, notably the spinorbit interaction, are much stronger than in transition metals (this also applies to rare
earths).
The main interaction of localized electrons with the magnetic field is the
Zeeman term,
M H = gB S H .
(6.11)
This interaction gives rise to paramagnetism of localized spins; see the next
section.
74
Magnetism
6.1.1 Susceptibility of noninteracting spins

Here we consider the susceptibility of localized magnetic moments without an
exchange interaction between them. First we present the classical treatment,
using Boltzmann statistics. Suppose we have a collection of atoms with magnetic moment M. In the magnetic field H the probability of finding the moment
M is

MH
(kB = 1)
(6.12)
n(M) exp
T
(see (1.1), with the energy (6.5)). From this we find the average moment
#
Me MH/T d
M = # MH/T
,
d
e
where is the solid angle. From (6.13), the magnetic susceptibility is
3
2
N
M
1 Ng 2 2B S(S + 1)
= M 2 =
,
0 = N
H
T
3
T
(6.13)
(6.14)
which is the well-known Curie law; the index 0 denotes the fact that we are dealing
with noninteracting spins. Here we have already used the connection M = gB S,
see equation (6.11), and also the fact that in quantum mechanics S2 = S(S + 1).
The factor 13 comes from averaging over all directions, see e.g. Kittel (1987).
The susceptibility per unit volume is thus
0 =
C
N g 2 2B S(S + 1)
,
V
3T
T
(6.15)
where C is called the Curie constant.

Consider now the quantum case. Instead of the integral over all orientations of
M in (6.11) we have to take a sum over possible quantum states S z (M z = gB S z ,
S S z S).
Thus for S = 12 we would have (for the field in the z-direction and for the
moment parallel to the field; we omit the index z below):

1
1
gB H
1 egB 2 H /T egB 2 H /T
1
tanh
.
=
2 egB 12 H /T + egB 21 H /T
2
2T
(6.16)
Taking into account that for electrons g = 2, we can rewrite equation (6.16) as

B H
.
(6.17)
M = B tanh
T
M = gB S ,
S =
6.1 Basic notions
75
.....
S
S
S1
S z = +S
Fig. 6.3
M
Fig. 6.4
For small H (gB H T ) the moment M is linear in the field,

M=
N g 2 2B H
,
V 4T
and
0 =
M
N g 2 2B
=
,
H
V 4T
(6.18)
which coincides with (6.15) for S = 12 , S(S + 1) = 34 .

For the general case (arbitrary S, or total angular momentum J ):
M(H ) = gB J BJ (gB J H /T ) ,
where
1
BJ (x) = 1 +
2J

coth

1
1
x
1+
x
coth
2J
2J
2J
(6.19)
(6.20)
is the so-called Brillouin function (used here instead of tanh in equation (6.16),
which was valid for J = S = 12 ).
The physics leading to (6.14)(6.20) is illustrated in Fig. 6.3. In a magnetic field
there occurs the Zeeman splitting of the levels, and with decreasing temperature
more and more spins in our system will accumulate at the lowest level, with the spin
parallel to the field, which leads to the increase of total magnetization (6.13), (6.16)
and to the susceptibility (6.14).
Both equation (6.16) and equations (6.19)(6.20) describe the saturation of
magnetization in strong fields gB J H T , see Fig. 6.4. For very large spins the
76
Magnetism
quantum expressions go over to the classical ones, with the substitution S(S + 1)
S 2 . This is a general property of any spin system: it becomes classical for S ,
because for S the noncommutativity of spin operators is irrelevant, and
quantum effects disappear (cf. the treatment of Bose condensation in Chapter 5).
6.2 Interacting localized moments; magnetic ordering

The main interaction between localized spins is the exchange interaction:

H=
Jij Si Sj gB H
Si ,
ij
(6.21)
where we have introduced also the interaction with the external field H. This is
the Heisenberg exchange interaction, the simplest form of spinspin interaction. In
general exchange may be anisotropic, e.g. J S z S z + J (S x S x + S y S y ); short-range
(Jij = J j,i1 ) or long-range, etc. For higher spins S > 12 the exchange interaction
may also contain higher-order terms, for example biquadratic exchange Si2 Sj2 ; there
may also exist other, more special terms. We will treat below predominantly the
simplest interaction (6.21) or its anisotropic generalizations; the general case is
considered in specialized monographs or reviews on magnetism, e.g. in White
(2006) and Yosida (1996).
Note: often the exchange interaction
differently: some (our equation (6.21)) is defined
times with the opposite sign, Jij Si Sj , and sometimes as 2 Jij Si Sj ; this
corresponds to different definitions of the exchange integral Jij . For example, the definitions of exchange integrals in two of the most popular textbooks, those by Kittel (1987)
and by Ashcroft and Mermin (1976), differ by a factor of 2. It is also important to know
whether in the summation in the Hamiltonian (6.21) each pair ij is counted only once or
the summation is carried out for all i and j independently, i.e. each pair enters twice (we
use the latter convention below); this is in fact the reason for the difference in the factor
of 2 mentioned above. Thus you have to be careful when someone cites the value of the
exchange constant for a particular system; you should always check which definition of the
exchange integral is being used.
Depending on the sign and the detailed distance dependence of Jij = J (Ri
Rj ) the exchange interaction can give different types of magnetic ordering:
ferromagnetic: all spins parallel, ;
antiferromagnetic: in the simplest case, alternating spins, (so-called
Neel ordering);
spiral, e.g. helicoidal or cycloidal state:
;
etc.
6.2 Interacting localized moments
77
6.2.1 Mean field approximation

The simplest treatment of the influence of the exchange interaction (6.21) on
magnetic properties of the system, in particular on magnetic ordering, is the mean
field, or self-consistent field approximation. In this method we consider each spin
as being in an average field created by other spins of the system, which has to be
determined self-consistently. This is equivalent to the decoupling

(6.22)
Jij Si Sj = Jij Si Sj + Si Sj Si Sj ,
where
Si = S
(6.23)
is the average spin, in the ferromagnetic case independent of the site. The last
term in equation (6.22) is necessary to avoid double counting; this is important
in the calculation of the total energy. If we take z as the quantization axis, only
z-components of the spin remain. Consequently, later on we will omit vector
notation and denote the average spin simply as S. This mean field approximation
(the decoupling (6.22)) means that we consider each spin, say spin Si , as being in
a molecular field (also called the internal, or effective field) Hintern created by its
neighbours and given by the expression

gB Hintern = 2
Jij S .
(6.24)
j
(We have used here the standard form of the coupling (6.11); the factor of 2
comes from the fact that, according to our definition of the exchange Hamiltonian
(6.21), we sum over all indices ij independently, i.e. each pair of spins is counted

twice.) For the nearest-neighbour coupling j Jij = J z, where z is the number of
nearest neighbours. Having in mind ferromagnetic interactions and introducing for
convenience the notation J = J (so that J > 0), we get for the molecular field
the expression
gB Hintern = 2Jz S .
(6.25)
Putting this expression into equation (6.16), we now obtain the self-consistency
equation (the mean field equation) for the average magnetization M = gB S,
which for the case of S = 12 (g = 2) with ferromagnetic interaction takes the form

J z S
1
(6.26)
S = tanh
2
T
or
Jz(M/B )
M/B = tanh
.
(6.27)
2T
In the presence of an external field H that field would enter in the numerator in
tanh in (6.27) together with the molecular field (6.24).
78
Magnetism
S
S
1
c
in
sing
rea
~ Tc T
(a)
(b)
Tc
Fig. 6.5
The easiest way to analyse the mean field equations (6.26), (6.27) is the graphical
solution, see Fig. 6.5(a): we plot the curves for the left- and right-hand sides of these
equations and look at their crossings. As we see, for zero external field H = 0 there
is always a zero (nonmagnetic) solution M = 0, or S = 0. But for low enough
temperatures there also exists a nonzero solution with finite value M (and also M).
One can show that these nontrivial solutions, when they exist, correspond to the
minimum of the free energy, whereas the zero solution becomes the maximum.
The temperature dependence of the spontaneous magnetization M, or the average
spin S, given by equation (6.27), is shown in Fig. 6.5(b). The value of the critical
temperature in this case (S = 12 ) is
(6.28)
Tc = 1 Jz .
2
For an arbitrary spin S with nearest-neighbour interaction the corresponding

expression is
(6.29)
Tc = 23 JzS(S + 1) ,
which for S = 12 reduces to (6.28).
Using equations (6.24)(6.27) we can also easily obtain an expression for the
magnetic susceptibility of interacting electrons. We can write the magnetic moment
as
M = 0 (H + Hintern ) ,
(6.30)
where 0 is given by the expression (6.15). Using the expression (6.24) for Hintern ,
we finally obtain for a short-range interaction
CH
M=
,
(6.31)
T Tc
where we took into account the expressions (6.15) for C and (6.29) for Tc .
The full susceptibility , defined by the relation M = H , is given in this case
by the expression
C
.
(6.32)
=
T Tc
79
This is the famous CurieWeiss law, which is usually written as

=
C
.
T
(6.33)
In the mean field approximation in the case of only a nearest-neighbour interaction

J , the Weiss temperature is
= Tc = 23 S(S + 1)Jz = 23 S(S + 1)J z ,
(6.34)
where J = J is the exchange integral as introduced in equation (6.21) for the

nearest-neighbour interaction. The sign of tells us whether the corresponding
exchange interaction is ferromagnetic ( > 0) or antiferromagnetic ( < 0).
In the general case the Weiss temperature is given by the expression

Jij ,
(6.35)
= 23 S(S + 1)
j
where the summation goes over all neighbours with which a given spin interacts.
This general expression (6.35) has broader applicability than the nearest-neighbour
version (6.34). It can happen, for example, that for anisotropic and especially for
frustrated magnets, or for systems with long-range interactions, the value of is
very different from Tc or even has the opposite sign. It is important to realize that the
sign of is determined by the sum of all exchange interactions, with all neighbours
of a given site. If these interactions are predominantly antiferromagnetic, the sign of
would be negative, and the magnetic ordering would be antiferromagnetic. There
may be situations, however, when the strongest interactions are ferromagnetic,
e.g. a strong ferromagnetic interaction in a magnetic layer, but if the interlayer
coupling is antiferromagnetic (and, e.g. weaker), the resulting state would be also
antiferromagnetic (ferromagnetic planes stacked antiferromagnetically). In this
case at high temperatures we would have a ferromagnetic susceptibility ( > 0),
although the ground state is actually antiferromagnetic. On the other hand, for
frustrated systems with antiferromagnetic interactions the value of the critical
temperature (the Neel temperature TN , see below) may be much smaller than the
typical absolute values of the exchange interaction. In this case || may be much
larger than TN ; the small value of TN /|| (or large ||/TN ) is now often taken as
an empirical criterion, a fingerprint of strong frustrations.
A word of caution again: for a different definition of the exchange constant,
when, in contrast to our convention, each pair ij in the summation in the exchange
Hamiltonian (6.21) is counted only once, one would use the exchange integral J
which is twice our J , J = 2J ; this is, e.g. the definition used by Ashcroft and
Mermin (1976) and Ziman (1979). Then the results (6.29), (6.34) for the critical
temperature Tc or for the Weiss temperature , written through this J , would
contain, instead of the factor 23 , the factor 13 , and the value of the exchange constant
80
Magnetism
H =/ 0
H=0
T
Fig. 6.6
determined, e.g. from the high-temperature susceptibility, would differ from the
more conventionally determined one by a factor of 2.
6.2.2 Landau theory for ferromagnets
The mean field approximation gives results equivalent to the general Landau theory
of second-order phase transitions. The Landau expansion of the free energy has
here the form
= AM 2 + B M 4 H M .
(6.36)
We have introduced in (6.36) the interaction with the external magnetic field H to
point out two consequences:
1. We can obtain from this approach the behaviour of the magnetic susceptibility
at the phase transition, which turns out to be the same as that obtained above in
the mean field approximation. The equation for M is

= 2AM + 4B M(M)2 H = 0 .
M
As M H, we may omit vector notation here:
2AM + 4BM 3 H = 0 .
(6.37)
(6.38)
It is clear that, for finite H , M is always nonzero, even above Tc , Fig. 6.6. Close to
Tc (T > Tc ), where M 1, we have (with A = a(T Tc ), see (2.3))
M=
H
H
=
,
2A
2a(T Tc )
(6.39)
1
,
2a(T Tc )
(6.40)
i.e. with M = H , we get

=
the standard CurieWeiss law, cf. (6.32).
81
The expression (6.39) is valid for T > Tc . Below Tc there exists spontaneous
magnetization in the absence of the magnetic field H , so that one cannot define
the susceptibility as = limH 0 M/H , as is usually done. However one can
always define the differential susceptibility dM/dH . From Fig. 6.6, where the
temperature dependence of the magnetization of a ferromagnet is shown in the
presence and in the absence of an external magnetic field, it is clear that in
the finite external field the magnetization increases both above and below Tc .
From equation (6.38) we find for T < Tc (where MH =0
= 0):

dM
d
=
+
=0,
(6.41)
dH M
H M
M M dH
i.e.
dM
=0,
dH
(6.42)
1
dM
=
.
dH
2A + 12BM 2
(6.43)
1 + (2A + 12BM 2 )
or
diff =
For T < Tc the coefficient A = a(T Tc ) < 0, M 2 = A/2B = a(Tc

T )/2B, and
=
1
1
1
=
=+
.
2A 6A
4A
4a(Tc T )
(6.44)
Thus the differential susceptibility also diverges as 1/(Tc T ) from below, for
T Tc 0, but with the coefficient two times smaller than above Tc , cf. (6.44)
and (6.40) see also (2.37); as mentioned there, this is a real effect, measurable
experimentally.
In obtaining equations (6.39)(6.44) we have taken into account only terms
linear in M and H . However, the full equation (6.38) contains also nonlinear
terms. They become important for higher fields. One can use equation (6.38) as
a very convenient way to determine experimentally the critical temperature of a
ferromagnetic phase transition.
The point is that by just measuring the magnetization in weak fields it is very
difficult to get an accurate value of Tc . As is well known, there always appear
in such systems ferromagnetic domains with different spin orientations, which
change with field and temperature and which dominate the magnetic response at
small fields. One can get rid of this problem if one works at high enough fields,
sufficient to orient all spins in one direction. But in this case one has to take into
account nonlinear effects in the M(H ) dependence, described by the cubic term
82
Magnetism
M2
Tc
dec
r
eas
ing
H/M
Fig. 6.7
H<0
H=0
1
2
3 H>0
Fig. 6.8
in equation (6.38). Keeping all terms there, one can rewrite equation (6.38) in
the form
2BM 2 = H /2M A .
(6.45)
Thus if we plot M 2 vs. H /M at different temperatures, we should have a series

of straight lines, see Fig. 6.7. The slope of these lines would give us the value of
the coefficient B in the Landau expansion of the free energy, and the offset, the
value at H = 0, would give the coefficient A. Remembering that in the Landau
theory A = a(T Tc ), we see that the value of the temperature at which such a
straight line passes through zero would give us the value of Tc .
In real situations, due to domain effects, at small fields this dependence could
strongly deviate from that of equation (6.45), but it is usually satisfied at large
enough fields. Thus to determine Tc one has to use the extrapolation of the
curves of Fig. 6.7 measured at high fields to H = 0. This method is widely used
in practice. The plot (6.45), Fig. 6.7, is called the Arrott plot (or sometimes the
BelovArrott plot).
2. Let us now consider the free energy (6.36) as a function of M for different
temperatures and fields. For T > Tc it has the form shown in Fig. 6.8, i.e. as a
83
Fig. 6.9
H<0
1
H=0
2
3 H>0
Fig. 6.10
Fig. 6.11
function of H the magnetization changes continuously, Fig. 6.9. Now, consider

the situation for T < Tc . The set of curves for (M) is now given in Fig. 6.10.
We see that as the field changes sign, the minima with positive and negative
M change absolute values; for H > 0 the minimum with M > 0 lies deeper,
and for H < 0 that with M < 0. Thus there should be a jump of M, when H
passes zero, very much like in the first-order transitions discussed in Chapter
2! (cf. Figs. 2.5, 2.8). This is nothing else but the well-known hysteresis of
ferromagnets, Fig. 6.11. And it really is a first-order transition. However, not as
a function of temperature, but as a function of field. Thus, the phase diagram
of a ferromagnet in the (T , H ) plane looks as shown in Fig. 6.12. There is
indeed a first-order transition line, and the end point of this line (critical point,
84
Magnetism
H
line of
first-order phase
transition
overheating and overcooling

lines (limits of hysteresis)
Tc
Fig. 6.12
in the terminology of first-order phase transitions) is our old familiar Tc the

critical point of the second-order phase transition. All the anomalies of different
quantities at the second-order phase transition point are actually the same as in
the critical point of first-order phase transitions, cf. e.g. Fig. 2.9.
6.2.3 Antiferromagnetic interactions

The mean field description of antiferromagnetic ordering coincides with that of
ferromagnetic ordering, with the only difference that we change the spin direction in one of the sublattices, together with the opposite sign of the exchange
interaction. The corresponding self-consistency equations, analogous to equations
(6.26), (6.27), will now be written for a sublattice magnetization. Below we discuss
the response of an antiferromagnet to an external field, which is different from that
of a ferromagnet.
Let us write down the Landau free energy for an antiferromagnet. In an antiferromagnet in the absence of external field there are two sublattices with antiparallel
spins, M 1 = M 2 . The order parameter here is
L = M1 M2 ;
(6.46)
it is zero above the critical temperature (usually denoted as TN , the Neel temperature), and nonzero for T < TN . The free energy of an antiferromagnet in the
presence of an external field has the form
= 0 + AL 2 + B L 4 + D(H L)2 12 p H 2 .
(6.47)
It is important that in contrast to ferromagnets the external field H does not couple
to the order parameter L linearly (linear couplings with different sublattices cancel,
because M 1 = M 2 ), so that the lowest nonzero coupling allowed by symmetry is
quadratic. At the same time there would definitely appear a certain magnetization in
85
c
c
c
TN
Fig. 6.13
the external field, even in the absence of antiferromagnetic ordering, e.g. above TN ;
this effect is described by the last term in equation (6.47). (We recall that the
moment M = /H , i.e. here it is equal to p H , as it should be.) Here p is
the magnetic susceptibility of the sample in the paramagnetic phase, which is given
by the CurieWeiss law (6.33) with = 23 S(S + 1)J z (6.34).
On the other hand, the presence of antiferromagnetic order below TN should
modify the susceptibility. Let us consider the susceptibility below the Neel temperature. We have to discriminate two situations: the field parallel to L (or to the
sublattice magnetization) or perpendicular to it. In general the moment is
M=

= p H 2D L(L H) .
H
(6.48)
(a) For the perpendicular field, H L, we obtain from (6.48) that

= p ,
(6.49)
i.e. below TN the perpendicular susceptibility is constant, independent of

temperature.
(b) For the field parallel to the sublattice magnetization, H L, the second term in
equation (6.48) also contributes. We then obtain
Da
(6.50)
(TN T )
B
(by (6.49) p = ; we also took the standard expression for the order parameter
a
of an antiferromagnet L2 = 2B
(TN T ), see equation (2.6)).
= p 2DL2 =
Thus we see that below TN , and are different, see Fig. 6.13. The moment
in the ordered phase is smaller for the field parallel to the sublattice magnetization
than for the perpendicular field. In the parallel field we have to invert some spins,
Fig. 6.14(a), and this costs a large amount of energy, and the resulting susceptibility is small, whereas for the field perpendicular to the sublattice magnetization
moments the sublattices can cant, Fig. 6.14(b), which is much easier. And we
86
Magnetism
H
H
(a)
(b)
Fig. 6.14
Hspin-flop Hspin-flip
(a)
H
(b)
Fig. 6.15
gain more energy during this process. Therefore in an isotropic antiferromagnet

the sublattice magnetization always orients perpendicular to the external field: in
this case the direction of the sublattice magnetization does not matter and the antiferromagnetic exchange energy is the same, whereas we gain more energy due to
interaction with the external field for the case of sublattices oriented perpendicular
to the field ( > , see (6.50)).
What if there is a certain anisotropy in a material? For instance there can exist
uniaxial anisotropy which favours all spins being oriented, say, along the z-axis.
(This could be caused for example by single-site anisotropy, described by terms
in the Hamiltonian of the type K(S z )2 with K > 0, or by exchange anisotropy

y y
J ij Siz Sjz + J ij (Six Sjx + Si Sj ), with J > J .) Then at zero field and at
small field along the z-axis, when anisotropy dominates, spins will still be parallel
to z, H . However when the external field H z increases, at a certain
moment, when it exceeds the anisotropy field, it will become favourable to turn all
spins into the xy-plane H (and cant them in the direction of the field). We gain in
this process more energy on the interaction with the field than we lose on anisotropy.
Such a transition is called a spin-flop transition: instead of having the structure of
the type H , the structure now looks as H , i.e. it is
still antiferromagnetic, but with the sublattices perpendicular to the field and canted
in the direction of the field. At still larger fields the canting angle increases till all
the spins become parallel to the field. This transition to a forced ferromagnetic
arrangement is called a spin-flip transition. Thus altogether the behaviour of the total
magnetization in this case has the form shown in Fig. 6.15(a) (at low temperatures
87
T
paramagnetic state
easy-axis
antiferro. state,
spins || easy
axis z
forced
ferromagnetic
state
spin-flop
phase
(spins H
and canted)
Hz
Dashed line: spin-flop transition
Shaded region: gradual cross-over to a paramagnetic state
Fig. 6.16
T TN ). At higher temperatures this behaviour can be smeared out, Fig. 6.15(b).

The phenomenon illustrated here (a sharp superlinear increase of magnetization at
certain fields) is called metamagnetism; the situation described above gives one of
the possible mechanisms of metamagnetic behaviour. A typical phase diagram of
an easy axis antiferromagnet is shown in Fig. 6.16.
6.2.4 General case

Let us consider now the general case, with the exchange interaction Jij extending
over larger distances and possibly of different sign. To study magnetic ordering
in this case it is convenient to write down the Hamiltonian in the momentum
representation,

H=
Jij Si Sj =
J (q) S(q) S(q) ,
(6.51)
q
ij
where
S(q) =
eiqRn Sn
(6.52)
is the Fourier transform of Sn , and similarly for J (q).

The mean field approximation can also be done in the momentum representa
tion: S(q)S(q) = S(q) S(q). The energy E = q J (q)|S(q)|2 reaches the
minimum at a certain value of q = Q, the value for which J (q) is minimal. Then
we can leave only these values of q = Q in S(q):
S(q) = S (q Q) ,
(6.53)
88
Magnetism
(a)
(b)
Fig. 6.17

Sn
d 3 q eiqRn S(q) = Sei QRn .
(6.54)
Thus we see that for a general case the type of the resulting magnetic ordering (the
one giving maximum energy gain) will be the modulated spin structure (e.g. spiral)
with momentum Q corresponding to the minimum of J (q).
Let us consider several examples, taking the simplest one-dimensional case. For
the nearest-neighbour interaction
Jij = J j,i1 = J (q) = 2J cos qa .
(6.55)
(1) Let us first take J negative, J < 0. J (q) has a minimum at q = Q = 0.

By (6.54)
Sn = S ,
(6.56)
i.e. we automatically obtain a ferromagnetic solution, .

(2) For J > 0 the minimum of J (q) is at q = Q = /a. This gives
Sn = SeiQRn = Sein = S(1)n ,
(6.57)
which corresponds to antiferromagnetic ordering, , as it should.

Thus in these two cases this method reproduces the known results.
(3) In general (for longer-range exchange Jij ) J (q) can have a minimum at a certain
arbitrary Q, and S(q) will be given by (6.53). In the coordinate representation
this corresponds to the magnetization changing as
Sn = SeiQRn ,
(6.58)
so that it describes a periodic (e.g. spiral) spin structure with period d = 2/Q.
One should be more careful here in treating different projections of the spin; there
may exist different types of structures with period d. Thus, this may be the so-called
sinusoidal structure, with the magnetization at each site parallel say to the z-axis, with
Snz = S cos Q Rn , see Fig. 6.17(a). Here the absolute value of the average magnetization changes in space. This will be, e.g. the situation for strong uniaxial (easy-axis)
anisotropy. (We stress that we are speaking here about the thermodynamic average of the
spin at site n, thus the fact that Snz is less than the nominal spin S, and can even be zero,
89
does not imply any contradiction: one can visualize such a state as that with a finite average
z-projection of S, S x and S y components remaining fluctuating.)
y
Another possibility is the helicoidal structure Snx = S cos Q Rn , Sn = S sin
Q Rn . In this structure the value of the average momentum at each site remains constant,
but the spin rotates for example in the xy (or in another) plane, see Fig. 6.17(b). If spins
are rotating in the plane perpendicular to the wavevector of the spiral Q, e.g. the spins
are rotating in the xy-plane and Q = Qz , one can speak about a proper screw structure.
y
If, however, the spins rotate in a plane containing Q, e.g. with the same Snx and Sn as
written above but with the vector Q, e.g. in the x-direction, we will have a cycloidal spin
structure. (This one, in particular, can give rise to ferroelectricity, which is very important
in multiferroics materials combining magnetic and ferroelectric properties.)
All such structures are often called spin-density waves (SDW), see also
Section 11.6 below.
Note that in general the period of corresponding structures may have nothing to
do with the periodicity of the underlying crystal lattice. Thus, for example, it can
be incommensurate with respect to the lattice period.
Typical cases in which such magnetic structures can occur are those with competing (long-range) interactions. They are commonly met in metals with special
shapes of the Fermi surface, e.g. in many rare earth metals and compounds, or in
the metal Cr. Another typical example in which such spiral magnetic structures can
appear are the frustrated magnets, with geometrically frustrated lattices (triangular,
kagome, etc.) or with competing interactions.
Problem: Find the most favourable type of magnetic order in a one-dimensional
system with competing nearest-neighbour and next-nearest-neighbour antiferromagnetic interactions

H=J
Si Si+1 + J
Si Si+2 .
(6.59)
i
Crude solution: Let us find the value of q which minimizes the total J(q):
J(q) = 2J cos q + 2J cos 2q
d J
= 2J sin q 4J sin 2q = 2J sin q 8J sin q cos q
dq

J
= 2J sin q 1 + 4 cos q = 0 .
J
The solutions are:
(1) sin q = 0, i.e. q = 0 or q = .
(6.60)
(6.61)
90
Magnetism
~
J
J'/J < 14
q
J'/J > 14
Fig. 6.18
From (6.60) we see that q = 0 corresponds to the maximum of J(q), and q =

may correspond to a minimum (if = J /J < 14 ). The solution with q =
gives an antiferromagnetic state (the doubled period l = 2
= 2, ).
(2) For = J /J > 14 there exists yet another solution, which gives the absolute
minimum of J(q) in this case: this is given by the solution of the equation
1+4
J
cos q = 0 ,
J
cos q0 =
1
.
4J /J
(6.62)
The dependence J(q) has the form shown in Fig. 6.18. Indeed, as |cos q| < 1,
this solution exists only when J /J > 14 .
Thus one may expect that the simple two-sublattice antiferromagnetic structure (the Neel configuration) will become unstable at J /J 14 and could
be transformed into a spiral with wavevector q0 . (Note that a similar spiral
solution exists also for a ferromagnetic nearest-neighbour interaction J < 0 if
J /|J | > 14 .)
Real solution: The actual situation for S = 12 and for J , J > 0 is more complicated
and much more interesting than in our mean field treatment. Indeed the twosublattice Neel-like configuration becomes unstable for (J /J )crit 14 (numerical
calculations give the accurate critical value (J /J )crit = 0.241, not so far from our
mean field result 14 !). But due to the quantum nature of spins the actual state for
J /J > (J /J )crit does not have real long-range order, it consists predominantly
of singlets and is of spin-liquid type with a gap in the singlettriplet excitations
(this state is sometimes called a MajumdarGhosh state). But our simple treatment
at least correctly predicts an instability of the simple Neel-like state and gives
the value of the critical coupling close to the actual one. And this, in general,
incommensurate, spiral state can indeed be realized in real quasi-1d materials
when one takes into account the weaker interchain interaction. A similar treatment
may be done for other types of lattices with competing interactions, not necessarily
one-dimensional.
6.3 Quantum effects
91
6.3 Quantum effects: magnons, or spin waves

Again first I give a short reminder:
The commutation relations for spin operators are
y
[Six , Sj ] = iSiz ij ,
[Si , Sjz ] = iSix ij ,
[Siz , Sjx ] = iSi ij .
(6.63)
It is convenient to introduce the spin raising and spin lowering operators

S + = S x + iS y ,
so that
S = S iS ,
x
S+ + S
,
2
S+ S
.
Sy =
2i
Sx =
(6.64)
The commutation relations for these operators are

[Si+ , Sj ] = 2Siz ij ,
[Siz , Sj+ ] = Si+ ij ,
[Siz , Sj ] = Si ij .
(6.65)
Thus at different sites the operators S + , S commute, i.e. they behave as bosons.
But we cannot put two spin-excitations on one site for S = 12 !
S + | 12 = |+ 12 ;
(S + )2 | 12 = S + (S + | 12 ) = S + | 12 = 0 , (6.66)
i.e. (S + )2 = 0, as for Fermi operators!

Problem: Check that at one site for S = 12 the operators S + , S indeed obey
anticommutation relation for Fermi operators.
Solution:

{S + , S }+ = S + S + S S + = 2 (S x )2 + (S y )2 = 2 S2 (S z )2
+

2 ,
1 1
3 1
1
=1,
(6.67)
=2
=2
+1
2 2
2
4 4
as it should be for Fermi operators. (Note that this is not the case for S > 12 .)
Thus the spin operators, strictly speaking, are neither Bose nor Fermi operators.
For spin 12 they behave as Bose operators at different sites and as Fermi operators
at the same site. Sometimes they are called paulions (Pauli statistics, vs. Bose or
Fermi statistics).
But it is often sufficient to treat spin excitations as approximately bosons, when
their density is small and, on average, they are far from each other.
92
Magnetism
6.3.1 Magnons in ferromagnets

Let us now consider elementary excitations in a ferromagnet spin waves, or
magnons. It is convenient to rewrite the Heisenberg exchange Hamiltonian (6.21),
using (6.64), as
*
)
H=
Jij Siz Sjz + 12 (Si+ Sj + Si Sj+ ) .
(6.68)

= |0, and act on this state by the operWe start with the
state

ator Si : Si |0 = , i.e. in this state one spin is reversed.
i
The term Siz Sjz in (6.68) leaves the reversed spin at the same place, and only gives
an energy cost to reverse it. The terms Si+ Sj + Si Sj+ , on the other hand, move the
reversed spin to another site, i.e. they create a spin wave with a certain dispersion.
The traditional way to treat this problem is the HolsteinPrimakoff transformation. We introduce new operators a , a instead of S + , S:
Si+ = 2S (1 ai ai /2S) 2 ai
(6.69)
Si = 2S ai (1 ai ai /2S) 2
so that [ai , aj ] = ij , the standard commutation relation for bosons.

One can then show that
Siz = S ai ai .
(6.70)
Problem: Check this, starting from the expression (S z )2 = (S)2 (S x )2 (S y )2 ,

using equation (6.69), the commutation relations for a, a and expanding
a] =
(1 a a/2S)1/2 ; use the fact that n = a a commute, [a a, a a] = 0, and [n,
[a a, a] = a. See Kittel (1987).

Usually it is sufficient to keep only the lowest terms in the expansion of (6.69),
i.e.
Si = 2S ai ;
(6.71)
Si+ = 2S ai ,
other terms are small at least for large S. For S = 12 we simply replace Si+ and Si
by ai and ai . For convenience changing the sign of the exchange constant as in

Section 6.2, Jij = Jij , we obtain
*
)
H=
Jij (S ai ai )(S aj aj ) + 12 2S(ai aj + ai aj )
*
)
=
Jij S 2 + S(ai aj + ai aj ai ai aj aj ) + O(a 4 ) . (6.72)
Here we have omitted quartic terms in the Hamiltonian, originating from the term
Siz Sjz and describing the interaction between magnons. The resulting expression
6.3 Quantum effects
93
q
Fig. 6.19
is already a quadratic form, similar to the phonon case, cf. Chapter 4. We can
diagonalize it by the Fourier transform:

Jij 1 eiq(Ri Rj ) aq aq
H = const. +
(6.73)
2S
q
ij
or
H = const. +
q aq aq ,
(6.74)
where the spin-wave spectrum is

'
(
q = +2S
J(Ri Rj ) 1 eiq(Ri Rj ) = 2S J(q = 0) J(q) .
(6.75)
As q 0 the spectrum is quadratic in q; for the nearest-neighbour interaction
Jij = Ji,j 1 it is
q 2S Jq 2 a 2 ,
(6.76)
see Fig. 6.19. Here S is a nominal spin (e.g. for spin 12 , 2S = 1). (By the way,
from (6.75) we see again that the ferromagnetic ordering becomes unstable if J (q)
has a minimum not at q = 0 but at a certain nonzero q 0 : then the spin-wave energy
q becomes negative for certain values of q close to q 0 , signalling an instability of
the initial ferromagnetic state and a transition to another type of magnetic ordering,
e.g. the helicoidal one with momentum q 0 .)
Spin waves (magnons) aq , aq are approximately bosons (at low temperatures,

when there are only a few of them and they are, on average, far from each other).
Then their occupation number is
nq =
1
eq /T
(6.77)
(Their number is not conserved, thus, according to the treatment of Chapter 3, we

take their chemical potential in (3.3) as zero. But the external magnetic field in
(6.21) is linearly coupled to the magnetization, i.e. to the number of magnons, and
it can play the role of their chemical potential.)
94
Magnetism
The total number of magnons at temperature T is (in the 3d case)

3

T 2 x dx
N
1
3
.
n=
d q nq =
(2 )3
4 2 2S J
ex 1
0
(6.78)
Each magnon (spin reversal) reduces the magnetization, i.e. the magnetization is
4
5

T 3/2
.
(6.79)
M(T ) = M(0) 1 const.
2S J
This dependence is indeed observed experimentally in isotropic ferromagnets. Note
that the behaviour of the magnetization close to T = 0, given by the spin-wave
theory, is quite different from that of the mean field treatment of Section 6.2:
according to the latter, the magnetization would approach the saturation value as
T 0 exponentially.
How can we visualize a spin wave? Suppose we have created a magnon with
|0. We have to use a (quasi)classical description if

wavevector Q, aQ |0 SQ
we want to visualize what is going on and want to be able to draw pictures, etc.
Thus we assume that there is a macroscopic occupation of the spin wave with the
wavevector Q. Then the average is
Q|Sq |Q = a(q Q) ,
(6.80)
where a is some (in general complex, a = |a| ei ) constant, describing the amplitude and phase of the spin wave. (One can make this treatment rigorous by introducing the so-called coherent states, equivalent to the classical description of a
coherent electromagnetic wave instead of the description in terms of photons; actually this state |Q with the macroscopic occupation is such a coherent state.) Then
the average projection of the spin on the x-axis is
%
&
'
(
S + + Sn
|Q = Q| 12
dq eiqn Sq+ + eiqn Sq |Q
Snx = Q| Snx |Q = Q| n
2
&
% iQn+i
+ eiQni
e
= |a| cos(Qn + ) .
(6.81)
= |a|
2
Similarly,
Sn+ Sn
|Q = |a| sin(Qn + ) .
(6.82)
2i
And the z-projection of the spins is the same for each site and just decreases slightly,
Sny = Q| Sny |Q = Q|
S z =
1
2
an an =
1
2
|a|2 ,
(6.83)
cf. (6.70); i.e. the spins are actually only slightly tilted from the original direction.
Actually the phase is = t, and at each site one has a picture of precession
6.3 Quantum effects
z
y
z
y
95
y
x
Fig. 6.20
(a)
(b)
Fig. 6.21
around the average (molecular) field, Fig. 6.20: the spin at each site rotates on a
cone. But the instantaneous picture (snapshot) corresponds to a wave cos Qn
in S x and sin Qn in S y .
The terms omitted in (6.72) include, e.g. terms of the type a a aa, i.e. they
describe the interaction between magnons. Because actually magnons are not
exactly bosons, when two of them come together, there is a Pauli-like on-site
repulsion. But there also exists a certain attraction between spin deviations, illustrated in Fig. 6.21.
In Fig. 6.21(a) we show two independent magnons (reversed spins) far from each
other. Each of them spoils z bonds (here, in Fig. 6.21(a), four bonds, denoted
by the wavy lines). When two reversed spins come close together and become
nearest neighbours (Fig. 6.21(b)), they cost less energy, as they destroy fewer bonds
(here in total six wrong bonds instead of eight in the case of Fig. 6.21(a)). This
decrease of energy means an attraction between spin deviations; it can in principle
lead to the formation of bound states of two magnons. Such bound states always
exist for the Heisenberg interaction in the 1d and 2d cases, and under certain
conditions in three-dimensional systems too. If, instead of a reversed spin, we
simply remove an electron with a spin, i.e. make a hole-doping, similar arguments
tell us that these two holes may prefer to stay together forming a bound state. This
is, in a nutshell, one of the mechanisms of the formation of Cooper pairs suggested,
e.g. for the high-temperature superconductivity cuprates.
Another way to treat spin waves
There exists another, very powerful and convenient method to study excitation
spectra, which is applicable to many situations the method of equations of motion.
96
Magnetism
The general scheme of this method is the following. The Schrodinger equation for
a wavefunction | is
H | = E| .
(6.84)
Let | = A |0, where |0 is the ground state (vacuum), and the (yet unknown)
operator A creates a given state | out of |0. Equation (6.84) can then be rewritten
as H A |0 = E A |0, or, adding and subtracting A H , as (H A A H + A H ) |0 =
E A |0. In other words, the commutator satisfies the equation
|0 = E A |0 A H |0 .
[H , A]
(6.85)
We assume that the ground state |0 is an eigenstate of the Hamiltonian, H |0 =
E0 |0; choosing the beginning of the energy, we can put E0 = 0, which we will do
below. Then (6.85) takes the form
|0 = E A |0 .
[H , A]
(6.86)
We can now treat equation (6.86) not as a Schrodinger equation for the wavefunction, but can say that the operator A obeys the equation
.
E A = [H , A]
(6.87)
If this equation is satisfied, then also the original Schrodinger equation (6.84)
will be satisfied. (Until now we have not made any approximations and used only
identity transformations in going from (6.84) to (6.87).) The approach in which,
instead of the Schrodinger equation for the wavefunction (6.84), we consider the
equivalent equation (6.87) for operators, is called the Heisenberg representation (in
this method time- or frequency-dependence is transferred from the wavefunctions
to operators; see the more detailed discussion in Chapter 8).
Actually these two approaches go back to the early days of the formulation of
quantum mechanics. In the 1920s two approaches had been suggested. One was
the method of Schrodinger, which dealt with the wavefunctions and with equations
for them. But slightly earlier Heisenberg suggested the matrix formulation of quantum mechanics, which dealt with matrices, which are actually operators (or, vice
versa, one can represent operators as matrices in a given basis). The terminology
Schrodinger representation and Heisenberg representation reflects the origin of
these two methods.
Consider now a ferromagnet with spin S = 12 with the ground state |0 =
| . Let us write down the equation of motion (6.87) for the operator Sl , which, as we have seen above, creates spin deviations and consequently
spin waves. With the exchange Hamiltonian H (6.68) the equation (6.87) becomes,
6.3 Quantum effects
using commutation relations (6.65),

ESl = [H, Sl ] = 2
97

Jlj Sl Sjz + Slz Sj .
(6.88)
(The factor of 2 in (6.88) comes from the terms with l = i and l = j in (6.68);
interchanging the summation indices i j in the second term we obtain (6.88).)
Now, we have started from the operator Si , and obtained in the right-hand side
more complicated operators, the products Si Sjz , Siz Sj . Generally speaking, one
now has to write down equations similar to (6.87), (6.88) for these new operators.
One gets, in this way, a set of equations for more and more complicated operators
which one can solve by interrupting, or truncating, these equations at a certain
stage, making certain assumptions about the nature of the ground state. Here we
do it by making a decoupling similar to a mean field: we replace Siz by the average
magnetization Siz = Sjz = S. Then we obtain

ESj = 2 SSj
Jij + 2 S
Jij Sj .
(6.89)
i
In the ferromagnetic case Jij < 0; denote again Jij = Jij . After Fourier transforming we get
6
7
ES (q) = 2 S J(0) J(q) S (q) ,
(6.90)
which gives the excitation spectrum
6
7
E(q) = 2 S J(0) J(q) .
(6.91)
The spectrum (6.91) coincides with the magnon spectrum q , (6.75), with one
important modification:
S
q ,
(6.92)
S
i.e. it is self-consistent, and in this approximation the energy of the spin wave
changes with the average magnetization S and with the temperature: the spectrum
becomes softer (the energy of the excitation decreases) when T Tc . This is
indeed very reasonable: as each spin is kept in its direction by the combined action
(molecular field) of its neighbours, the spin stiffness (the cost of reversing a
given spin) should depend on the average spin of the surrounding ions and should
decrease with decreasing S. This is what equation (6.92) corresponds to.
If one now calculates the magnetization M(T ) using this spectrum, one obtains
the standard spin-wave law (6.79) as T 0, and one reproduces the mean field
results M Tc T close to Tc , i.e. it is a very good interpolation from T = 0

to Tc .
E(q) =
98
Magnetism
6.3.2 Antiferromagnetic magnons. Zero-point oscillations and their role

Let us consider now the antiferromagnetic case, J > 0. As the ground state we
take here the state with two sublattices | |0 (the Neel state). The
spin operators act on different sublattices as follows:
Sublattice j :
Sjz |0 = +S |0 ,
Sj+ |0 = 0 ,
Sublattice l:
Slz |0 = S |0 ,
Sl |0 = 0 .
(6.93)
Again we introduce spin deviation operators, as in (6.69), but in this case for each
sublattice separately:
Sjz = S aj aj ,
Sj+ 2S aj ,
Sj 2S aj ,
(6.94)
+
z
Sl 2S bl ,
Sl 2S bl .
Sl = S + bl bl ,
Because of the antiferromagnetic (Neel) structure of the ground state, in this case
the creation operators for spin waves are S for one sublattice and S + for the other.
Substituting (6.94) into the Hamiltonian (6.68) and making a Fourier transform,
we get (again omitting higher order terms):

H = E0 + 2S
J (0) aq aq + bq bq + J (q) aq bq + bq aq . (6.95)
q
In contrast to the ferromagnetic case, the expression (6.95), albeit a quadratic form,
is still nondiagonal (it contains terms a b , ab). But we know what to do with
such terms: we can diagonalize this expression using the Bogolyubov canonical
transformation (5.23), as we have done for the Bose gas:
aq = uq q + vq q ,
bq = uq q + vq q .
(6.96)
As u2q vq2 = 1, we can represent uq and vq , for example, as

uq = cosh q ,
vq = sinh q .
(6.97)
As in Chapter 5, the requirement of cancellation of dangerous terms , ,

, , or equivalently the minimization of the energy E = H, gives
tanh 2q = J (q)/J (0) ,
and in effect we get
H = E0 +
(6.98)

q (q q + q q )
q 2SJ (0) +
q
= const. +

q
)
q
q q +
1
2

*
+ q q + 12
(6.99)
6.3 Quantum effects
99
with the spin-wave spectrum

q = 2S J 2 (0) J 2 (q) .
(6.100)
For q 0 q |q|, i.e. we obtain here a linear spectrum (vs. quadratic in the
ferromagnetic case). There are two degenerate modes, and . The last two

terms in (6.99) combine and give the term 2 q q (nq + 12 ), cf. the case of
phonons, Chapter 4.
The presence of 12 in (nq + 12 ) in (6.99) tells us that, in analogy with phonons in
crystals, Chapter 4, there exist zero-point fluctuations in quantum antiferromagnets,
which reduce sublattice magnetization even at T = 0, and which can even suppress
long-range order completely. This is connected with the presence of nondiagonal
terms in (6.95): the initial Neel ground state |0 is not an eigenstate of H, and pairs
of spin excitations at neighbouring sites (in different sublattices) can be created.
(One can easily see that the fully ordered ferromagnetic state | is
an eigenstate of the exchange Hamiltonian.) In an antiferromagnet the term Sj Sl+
in (6.68) gives Sj Sl+ | = | , i.e. it reverses two spins and
j l
j l
creates two excitations, leading to quantum fluctuations. This means that the actual
ground state of the antiferromagnetic Heisenberg model (6.21), (6.68) is not just the
state | |0 (the Neel state), but contains an admixture of states
with different number of spin flips. In effect zero-point fluctuations are present even
at T = 0, and the sublattice magnetization S is less than S. Indeed, the deviation
of the average spin, e.g. in the b-sublattice from the nominal value S, is
Slz = S Slz = bl bl =
2
b b .
N q q q
(6.101)
Using (6.96)(6.98) we can express it in the following way (see the details in
Ziman (1979) and Kittel (1987)):
S z = const. + 2
J (0)

nq + 12 ,
q
q
(6.102)
where q is the spectrum (6.100).

Check this, using the relations (6.96), expressing the coefficients uq , vq through
tanh 2q (6.98), using the bosonic commutation relations for the operators q , q ,
and taking into account that the nondiagonal terms of the type , should
drop out.
The expression for spin deviations (6.102) is exactly analogous to the expression for the vibration amplitude of the lattice (4.74). Therefore we should expect
100
Magnetism
that all the consequences thereof, discussed in Section 4.4.4, would also apply to
antiferromagnets. Thus the deviations from the
sublattice magnetiza# maximum
d
tion at T = 0 in d-dimensional space are d q/q, i.e. they are finite for 3d
(d 3 q q 2 dq) or 2d (d 2 q# q dq) cases, but for the 1d case the integral diverges
logarithmically, bl bl dq/q, which means that there is no long-range order

in a 1d antiferromagnet even at T = 0 ! (cf. the case of crystals, (4.74)(4.77)).
This is due to quantum fluctuations, driven by the terms S + S in the Hamiltonian.

A model without these terms, H = Jij Siz Sjz , is different in this respect. This
Ising model is in a sense classical (vs. the Heisenberg model which is intrinsically
quantum).
It is instructive at this point to elaborate somewhat on the parallel with ordinary
lattice phonons. We have seen that many features of antiferromagnets remind
us of a harmonic lattice: the spectrum of collective excitations (antiferromagnetic
magnons vs. acoustic phonons) is linear as q 0; there exist zero-point vibrations,
cf. (4.65), (4.74) and (6.102); because of that, one-dimensional order becomes
unstable even at T = 0, etc. Physically this is connected with the fact that, in
contrast to a ferromagnet, for which a perfectly ordered state is an exact eigenstate
of the Hamiltonian, both in an antiferromagnet and in a lattice the ordered state we
usually start with is not an eigenstate; there are terms in the Hamiltonian which
lead to the creation of pairs of excitations.
Mathematically, as we have seen above, both the linear spectrum of antiferromagnetic magnons and the presence of zero-point deviations (6.102) follow

from the Bogolyubov canonical transformation (6.96) (giving the term (nq + 12 )
in (6.102)). We have seen that the same term 12 in (n + 12 ) which is responsible
for zero-point oscillations at T = 0 (where nq = 0) is also present in the case
of phonons. However we did not do any Bogolyubov transformation in Chapter 4!
What is really the connection here, then?
The answer is the following. When considering lattice vibrations, one usually proceeds classically, treating classical equations of motion for atomic coordinates xi , or for small deviations ui , see Chapter 4. Only at a later stage did we go to
the second quantization formalism and introduced phonon operators b , b. The cor
responding connection has the form (see (4.10)) ui (bi + bi ). If we had used this
representation from the very beginning, and written the initial Hamiltonian through
the phonon operators, then, e.g. the term 12 B(ui uj )2 in the lattice Hamiltonian

(4.16) (elastic energy) would contain both the terms bi bj and bi bj , bi bj , exactly
as in the case of weakly interacting bosons (5.21) or antiferromagnets (6.95). And
then we should have used again the same Bogolyubov canonical transformation to
get rid of these terms, even in ordinary crystals! Thus it is indeed not accidental that
the behaviour of antiferromagnets (antiferromagnetic magnons) resembles that of
lattice (phonons).
6.3 Quantum effects
101
On the other hand, one can proceed in the opposite way and try to obtain the
magnon spectrum classically, going to the second quantization description at a
later stage (as we have done with the lattice dynamics). This is indeed possible;
the corresponding classical description can be formulated and is actually widely
used in various problems of magnetism. The corresponding equations are the
LandauLifshits equations mentioned in Section 2.4. They are especially useful
for discussions of problems such as the dynamics of domain walls, etc.
Whereas at T = 0 the long-range antiferromagnetic order is unstable in onedimensional systems, at finite temperatures these deviations may diverge also in
two-dimensional antiferromagnets or ferromagnets.
Problem: For an isotropic ferromagnet show that the magnetization is zero at finite
temperatures in the 1d and 2d cases.
Solution: From equation (6.78), with nq = 1/(eq /T 1) (6.77), and with q given
by (6.75), (6.76), we obtain that the deviations from S at T
= 0 are
d1

dd q
q dq
S n
.
(6.103)
2 /T
cq
e
1 (q0) 0 q 2 /T
#
Thus for the 3d case S T q 2 dq/q 2 which is OK; spin deviations are finite
(and given by (6.79)).
#
But already for the 2d case S q dq/q 2 is logarithmically divergent! And
in the 1d case this divergence is even stronger.
Problem: Do the same for an isotropic antiferromagnet at T
= 0. Remember
that the antiferromagnetic spectrum is q q instead of q 2 for a ferromagnet.
But, on the other hand, the expression for spin deviations in this case is given by
equation (6.102), i.e. it contains an extra q in the denominator (and also the term
1
in the integrand, describing zero-point fluctuations).
2
Solution: All the calculations here are exactly the same as in the case of phonons,
Section 4.4.4. Thus the conclusions are the same: there is no long-range order in
1d either at T = 0 or T
= 0, and in 2d at T
= 0.
Thus for the isotropic (Heisenberg) interaction there is no long-range order at
T
= 0 in 1d and 2d systems. For 1d antiferromagnets there is no long-range order
even at T = 0. This is again connected with the fact that elementary excitations
(magnons or spin waves) are gapless, their spectrum obeys q 0 as q 0 (and
sufficiently fast).
We want to repeat here this important point, already discussed in Sections 2.6 and
4.4. The general statement (the Goldstone theorem) is: if there exists a continuous
102
Magnetism
broken symmetry, there should exist gapless excitations (if the interactions are not
long-range ones).
Once again, here are examples of this general statement in specific cases:
Magnets
The continuous symmetry that is broken in the low-temperature ordered state is the
continuous spin rotation: arbitrary directions, , or , or , etc. are allowed and
are equivalent for the Heisenberg interaction (6.21). As a result there exist gapless
spin waves. Spin waves with q 0 correspond to a rotation of magnetization
(or sublattice magnetization) of the whole sample. As all spin orientations are
equivalent, all such states are degenerate, it costs no energy to go from one to
another, and consequently (q = 0) = 0. But if the interaction is anisotropic (e.g.
J Siz Sjz this is called the Ising interaction), then the spin excitations have a gap,
and the conclusions would be different.
Crystallization
In liquids there exists a continuous symmetry: continuous translations r r + r
(and continuous rotations). In crystals, however, we have a periodic structure, and
the allowed translations are discrete, r r + an. Broken continuous symmetry
again leads to a gapless collective mode acoustic phonons, with the spectrum q =
sq. The mode with q 0 corresponds to a shift of the crystal as a whole, which
explains why these phonons are gapless.
Bose condensation
The order parameter here is a0 , or a0 . What is the broken symmetry in this

case? As explained in Chapter 5, there exists a gauge transformation of operators,
an arbitrary change of phase: aq ei aq , aq e+i aq . Usually all physical

observables contain all operators always in pairs, a a, a a aa, etc. and nothing depends on . But in the Bose-condensed state the average, not of a pair,
but of only one operator a0 , is nonzero and plays the role of an order parameter, which means a broken gauge symmetry (again it is a continuous symmetry
which is broken: all values of the phase 0 < 2 were possible). An ordered
phase is characterized by fixed (exactly as in ferromagnets, where below Tc we
chose some fixed orientation of M). Consequently in the Bose condensed state
there exists a mode q without a gap the Bogolyubov sound. One may thus
connect the Bogolyubov sound with the oscillations of the phase of the order
parameter.
Thus we again repeat: the conclusion is that, because of the presence of such
gapless excitations, there is no long-range order in models or systems with broken
6.3 Quantum effects
103
Table 6.1
Ferromagnetic
Antiferromagnetic
Ground state: |0 = |
Crudely: the long-range ordered state is

the Neel state |0 = |
It is the exact ground state of the

Heisenberg
exchange Hamiltonian
H = J i,j Si Sj
It is not an exact eigenstate of H

(mathematically, one has to make
a Bogolyubov transformation to
diagonalize the Hamiltonian and to
obtain the spectrum of elementary
excitations)
Spin waves: k = ck 2
Spin waves: k = ck
Spin deviations:
d
d

d k
d k
T
d d k nk T
k
k2
Spin deviations:
d

d k
nk + 12
k
1d divergent, no LRO
even at T = 0 !
T =0:
2d OK
3d OK

T
= 0 : 2d divergent, no LRO
3d OK
T = 0 : LRO, no spin deviations in 1d,

2d and 3d systems

T
= 0 : 2d divergent, no LRO
3d OK
continuous symmetry in 1d and 2d cases at T

= 0; this is the MerminWagner
theorem discussed in Chapter 4.
We summarize the behaviour of ferro- and antiferromagnetic systems (from the
point of view of the role of fluctuations) in Table 6.1 (LRO stands for long-range
order).
The behaviour at T
= 0 is a consequence of the MerminWagner theorem.
Problem (a real problem).
The excitation spectrum (gapless) may be, e.g. k k or k k 2 . In crystals
(sound waves) or in antiferromagnets, for which the simple ordered state is not
the exact ground state# and there exist zero-point oscillations, the spectrum is
linear, = sk, so that d d k/k is divergent for d = 1 (1d systems) but not in 2d
(at T = 0). (At T
= 0 the linear spectrum is sufficient to destroy ordering also in
the 2d case.)
For a ferromagnet the ground state of the type is an exact eigenstate of
the Hamiltonian, and there are no zero-point fluctuations; as a result there exists
long-range order at T = 0, S = S (e.g. ST =0 = + 12 ). At the same time the
spin-wave spectrum is quadratic, k = ck 2 .
104
Magnetism
Is it accidental that the situation with zero-point motion goes together with a
linear spectrum, and an exact state without zero-point oscillations gives quadratic
dispersion?
For instance if there would have been zero-point oscillations, but with excitations with the spectrum k k 2 , then the fluctuations (e.g. spin deviations in
magnets), given in this case
of the type of (4.74) or (6.102), at
#
# by an expression
T = 0 would behave as d d k/k = d d k/k 2 , i.e. they would diverge also at
d = 2 (2d systems) and there would be no ordering at T = 0 in both 1d and
2d cases.
But worse than that: then at T
= 0 we would have had, for example, for a crystal
(cf. (4.78))

u
2
dd k
=T
k Tk
dd k
T
2k
dd k
,
c2 k 4
(6.104)
and u2 would diverge, i.e. the crystal with such a spectrum would be unstable at
any temperature even in our real 3d world! (And even in the 4d case.)
Maybe just because of that, this is not the case? In other words, is there a theorem
that states:
In the presence of quantum fluctuations at T = 0 (or of the zero-point motion)
the excitation spectrum (spectrum of the Goldstone mode) is linear (or at least
k = ck with < 32 , so as to make a 3d system stable at T
= 0); and when
we have the exact ground state (an ordered state which is an eigenstate of the
Hamiltonian) and there are no quantum fluctuations at T = 0, then the spectrum
of the Goldstone mode may be (should be?) ck 2 ?
Note also: the spectrum of surface modes (vibrations of the surface of a liquid, for example) is k k 3/2 , just enough to#make the surface
at T = 0
# stable
(the amplitude of zero-point oscillations is d 2 k/k 3/2 dk/ k which is
convergent).

The study of different versions of magnetic models is now a rather well-developed
field. We often have a good feeling of what to expect; many mathematical methods
can be used for these systems and were actually developed in this context. It is
also very useful and instructive because it is often possible to map other physical
problems onto magnetic ones and then use the magnetic know-how to solve or at
least to better understand these problems.
105
6.4.1 One-dimensional models

These models are often exactly soluble (not all of them of course). The results of
the exact solutions are often different from the simple mean field results.
(1a) 1d Ising model with nearest-neighbour interaction
The Hamiltonian of the Ising model in one dimension can be written as

z
H=J
Siz Sjz = 2J
Siz Si+1
ij
(6.105)
( ij means summation over nearest neighbours nn). It is easy to solve this

model exactly to find the partition function Z(T , H ) at arbitrary temperature and
magnetic field H z (described by an extra term in the Hamiltonian (6.105) of the

type H i Siz ), see, e.g. Ziman (1979). Note that for the Ising model the situation
for ferromagnetic and antiferromagnetic couplings is practically the same: in both
cases there exists long-range ordering at T = 0, ferromagnetic in one case and a
simple two-sublattice Neel state in the other. One can formally go from one case
to the other by changing the spin quantization axes for every second site to the
opposite.
There exists an elegant method for solving this problem the so-called transfer
matrix method; it can also be used for treating some other 1d models. One can
write down the partition function as ( = 1/T )
$
$
Z = Tr e H = Tr
exp(2J Si Si+1 ) =
Tr exp(2J Si Si+1 )
i

S1
S2
N

exp 2J
Si Si+1 .
SN
(6.106)
i=1
Taking into account the fact that every Ising spin Si can take only two values, 12 ,
we can write down the partition function (6.106) as the trace of a product of matrices
TSS such that S|T|S = exp(2J SS ) ( +1|T| + 1 = 1|T| 1 = eJ /2 ,
1|T| + 1 = +1|T| 1 = e+J /2 ):

Z=
S1 |T|S2 S2 |T|S3 SN |T|S1

S1
S2
SN

N
S1 |TN |S1 = Tr(TN ) = N
=
+ + ,
(6.107)
S1
where + and are two eigenvalues of the matrix T, and we have taken the
periodic boundary condition SN+1 = S1 . One can easily generalize this treatment

to the Ising model in a parallel magnetic field (with the extra term Hext i Si
added to (6.105)). Thus the solution of the problem is reduced to finding the
106
Magnetism
T/J
~1
Fig. 6.22
(maximum) eigenvalue of the transfer matrix T, which in this case amounts to

solving a quadratic equation.
Using this method we can easily obtain all the thermodynamic properties of the
1d Ising model. Thus, for example, the specific heat is (exact result)

2
J
2 J
,
sech
cV =
T
T
(6.108)
see Fig 6.22.

The maximum of cV (T ) can be easily understood. The energy spectrum of the
Ising model consists of discrete levels. The energy of a given bond is 12 |J | for the
ground state (parallel spins for the ferromagnetic case) and + 12 |J | for the wrong
bond (antiparallel spins). Thus the creation of such a wrong bond costs us |J |; it
is the cheapest elementary excitation here. (One can easily see that in fact this
means the creation of a domain wall the spins are, say, to the left of
this defect, but to the right of it; this character of elementary excitation is
very common in 1d systems, as we will see below. Of course, in a general excited
state we can create many such wrong bonds.) The specific heat (6.108) in this
essentially two-level-like situation has a broad maximum at the temperature of the
order of level splitting in this case at T J ; this is called the Schottky anomaly.
Note once again that for the Ising model the ferromagnetic (J < 0) and antiferromagnetic (J > 0) cases behave in the same way (e.g. cV is an even function
of J ). The ground states of the kind for J < 0 (ferromagnetic case) and
for J > 0 (antiferromagnetic case) are both exact ground states of the Ising
Hamiltonian (6.105), in contrast to the antiferromagnetic Heisenberg model (6.21).
In this sense the Ising model is classical, it does not have significant quantum
effects. The broken symmetry here is discrete, it is the reversal of spin ;
the spectrum of elementary excitations contains a gap equal to |J | for both cases,
i.e. there is no Goldstone mode, and correspondingly the MerminWagner theorem
does not apply. However this system is ordered only at T = 0, and there is no
long-range order at any finite T > 0.
107
e
(a)
(b)
Fig. 6.23
(1b) 1d xy model, spins 1/2

In this model the spins are allowed to point in any direction in the xy-plane. The
Hamiltonian of the xy model is

y y
H=J
(Six Sjx + Si Sj ) = 12 J
(Si+ Sj + Si Sj+ ) .
(6.109)
ij
ij
Remembering (6.69), (6.71), one may hope to map this model onto a model like

J ij ai aj , with bosonic operators a, a . But in the standard HolsteinPrimakoff

representation (6.69) the operators a, a are only approximate bosons. In the 1d
case, and for S = 12 , quantum effects are especially important, in particular the
on-site anticommutativity of Si+ , Si , see (6.67). Thus one should expect that the
conventional HolsteinPrimakoff approximation would be a rather poor approximation in this case.
However, there exists another possibility: it is possible to map the model (6.109)
exactly onto a model of noninteracting fermions by the so-called JordanWigner
transformation. In effect Si+ is transformed into ci , and Si goes over to ci (with
some complicated phase factors), where ci , ci are spinless fermions; the Hamiltonian (6.109) is transformed into

H=J
ci cj ,
(6.110)
ij
which can be easily diagonalized by the Fourier transform (this is the tight-binding
model for spinless fermions):

H=
k ck ck ,
k = 2J cos k .
(6.111)
The ground state (the state with the minimum energy) is reached when all fermion
states with negative energy are filled. The wavefunction of the ground state is thus
the Fermi sphere of these spinless fermions ck , with kF = , see Fig. 6.23(a) for
the case J > 0, Fig. 6.23(b) for J < 0.
108
Magnetism
One can show that the projection of the total spin is

2

3

z
1
1
S = 2 N N =
c k ck 2 .
(6.112)
k=
Thus, e.g. if all k-states were filled, ck ck = nk = 1 for all k, then from
equation (6.112) S z = N 12 N = 12 N, i.e. this would correspond to all spins
being . If all k-states are empty, then S z = 12 N, which corresponds to all spins
being . The ground state (half of the states are filled) corresponds to the total
spin S z = 0. But it is not an ordered (LRO) antiferromagnet, although there are
antiferromagnetic short-range correlations, or short-range order (SRO).
Note also that, similar to the Ising model, for the one-dimensional xy model
the sign of the exchange interaction J in (6.109) does not matter: it is clear from
Fig. 6.23 that the change of sign of J corresponds simply to the transformation
k k + , and the properties of the solution remain the same.
This mapping of the one-dimensional xy model into spinless fermions is used
nowadays to discuss many problems, e.g. the physics of elementary excitations in
polymers, etc.
(1c) 1d Heisenberg model for S = 1/2
The Hamiltonian of the 1d Heisenberg model with nearest-neighbour interaction is
*

)
+
z
H = J
Si Si+1 = J
+ 12 (Si+ Si+1
+ Si Si+1
) . (6.113)
Siz Si+1
i
We use here the standard way to write down the Hamiltonian of the 1d Heisenberg
model, with the summation over one site index i (cf. (6.21)). Note that the exchange
integral J introduced thus differs by a factor of 2 from our previous definition, J =
2J . This definition here is convenient, because then the energies of the singlet and
triplet states of the pair of spins (S1 , S2 ) would be 34 J for a singlet and + 14 J for
a triplet, see also below, so that their difference is J .
The ferromagnetic case is simple: the ordered state with all spins, e.g. up is the
exact ground state at T = 0. The antiferromagnetic model is much more interesting
because of the presence of spin deviations in the ground state, as discussed above.
An exact solution of this model was obtained in the 1930s by H. Bethe, who
found an expression for the wavefunction of the ground state and calculated certain
properties. Thus the exact value of the ground state energy is E/N = J (ln 2 14 ),
i.e. the energy per bond is between the value 34 J which one would get for an
isolated singlet dimer, and the mean field value 14 J . The method proposed
by Bethe (which is called the Bethe Ansatz) is rather complicated; still not all
properties are calculated, and many results were obtained only relatively recently.
109
(a)
(b)
(c)
p
(d)
x
Fig. 6.24
It has a number of interesting features, some of which will be discussed in the next
section.
6.4.2 Resonating valence bonds, spinons and holons

The usual semiclassical picture of an antiferromagnetic state of a 1d chain would
be the following:
The ground state would be of the Neel type: .
The simplest excitation would correspond to one reversed spin:
.
In our previous treatment this approach gave spin waves, or magnons. Such
an excitation changes the total spin by 1, i.e. the quantum number of a magnon
is S = 1. This is actually why magnons could be treated as approximate bosons:
the general rule is that integer spins correspond to bosons, and half-integer spins
to fermions (although just in the 1d case this rule, strictly speaking, is no longer
true, and the situation may be more complicated).
However, as already discussed above, in the Heisenberg antiferromagnet due to
quantum fluctuations (caused by terms Si+ Sj + Si Sj+ in (6.113)) the following
process is possible. We start from the state with one reversed spin, Fig. 6.24(a),
and interchange, using Si+ Sj , two spins in the vicinity of the reversed one, as
shown by the dashed arrow in Fig. 6.24(a). As a result we would have the situation
110
Magnetism
shown in Fig. 6.24(b). We can continue this process further, and go to the state
of Fig. 6.24(c), etc. In effect one spin excitation (a reversed spin) is split into two
excitations, each of which is actually a domain wall. Between them we again have
good antiferromagnetic ordering, but with even and odd sublattices interchanged.
In other words the sign of the order parameter changes to the opposite, or its phase
changes by , see Fig. 6.24(d).
Such domain walls, or solitons, are the actual elementary excitations in the
1d Heisenberg antiferromagnet (although the picture presented above is strictly
speaking not valid, because there is no long-range Neel order in the ground state to
start with). As the initial excitation (reversed spin) had S = 1 and it is now split into
two equivalent defects, this means that each of these new excitations carries spin
1
, i.e. they are more like fermions. In the previous section we saw that the xy model
2
can be transformed into free fermions. These are related features, but the isotropic
Heisenberg case is much more complicated it rather gives interacting fermions.
(As mentioned above, strictly speaking, in 1d systems the standard connection
between spin and statistics is not valid, i.e. the fact that these excitations carry
spin 12 does not yet automatically imply that they are fermions.) But in any case
these solitons are really the elementary excitations in one-dimensional Heisenberg
antiferromagnets, and the treatment of the properties of such systems in terms of
these excitations is definitely much better than the one using ordinary magnons.
The fact that such objects are the actual elementary excitations does not only
follow from our approximate (and in fact not very realistic, see below) treatment,
but has been rigorously proven using the Bethe Ansatz method (Faddeev and
Tachtajan).
Now, the same quantum effects (S + S terms) which split the spin-wave excitation into two, lead to the fact that the Neel state is not really the ground
state: there appears an admixture of configurations with reversed pairs of spins. We
know from quantum mechanics that the ground state of an isolated pair of spins
with an antiferromagnetic interaction is not simply , but rather an antisymmetric
combination a real singlet. Therefore it may be worthwhile to study an alternative
trial state, making real singlets in the ground state.
The Neel state is . Its energy is (using the Hamiltonian (6.113))
1
ENeel = NJ .
4
(6.114)
Let us now form real singlets. The singlet state for a pair of spins 1 and 2 is
1
|Singl. = |1 2 1 2 .
2
(6.115)
111
Fig. 6.25
Fig. 6.26
The energy of a singlet is

Singl.| J S1 S2 |Singl. = J
3
.
4
(6.116)
Check: this directly and by using the expression for the total spin (S1 + S2 )2 and
the fact that S2 = S(S + 1).
Solution: S2tot = (S1 + S2 )2 = S21 + S22 + 2S1 S2 , from which S1 S2 =
1
(S2tot S21 S22 ). For a singlet state, Stot = 0 and for S1 = S2 = 12 , with S2 =
2
S(S + 1), this gives S1 S2 = 34 , from which we obtain the result (6.116).
Why does the factor 34 appear in (6.116)? The term S1z S2z gives 14 . In the Neel
state only this term contributes. But in a real singlet state (6.115) not only S1z S2z
y y
contributes, but also S1x S2x and S1 S2 in S1 S2 (the singlet state is spherically
symmetric). Each of them gives the same contribution 14 , that is why the total
energy is 34 J .
Now, let us take instead of the Neel state with the energy (6.114) another trial
state, the one made of singlets (actually we form valence bonds, as in the H2
molecule, see Fig. 6.25):
|VB = |Singl.12 |Singl.34 |Singl.56 . . . .
(6.117)
The energy of this state is

3
3
N
= J N .
EVB = J
4
2
8
(6.118)
In comparison with the Neel state (6.114) we have lost half of the bonds, but gained
a factor of 3 at each of the remaining singlet bonds. As a result the energy of this
valence bond state, or valence bond crystal, EVB is less than ENeel . Actually we
can decrease the energy still further by using an exchange between singlets (the socalled resonance, see Fig. 6.26), i.e. interchanging singlet and empty bonds. It
112
Magnetism
Fig. 6.27
Fig. 6.28
can be shown that this mixing indeed decreases the energy still further. This mechanism of extra stabilization of such a state due to resonance between
different config

urations is known to work, e.g. in the benzene molecule,
The very concept of resonance appeared in this context in chemistry and is largely
due to L. Pauling. In our problem this is what is now called the resonating valence
bond (RVB) state (the concept introduced to solid state physics by P. W. Anderson
in 1973); it is at present a very popular notion, in particular in application to high-Tc
superconductors, but not only there.
Let us now consider elementary excitations in the (R)VB state. Similar to the
Neel configuration, we can first create a spin 1 excitation by breaking the pair
exciting one pair from the ground state singlet into a triplet state, Fig. 6.27. Again,
as in Fig. 6.24, we can recommute the singlet bond several times, and in effect
we will get two separated spin 12 objects, between which there will still be a
paired state, but with the interchanged sublattices of singlet and empty bonds,
Fig. 6.28. We thus again have created two excitations with spin 12 solitons, or kinks
(domain walls between two different domains). From Fig. 6.28 it is clear that these
excitations are indeed S = 12 objects. These excitations are now called spinons.
One can show that excitations of this kind spinons are indeed the actual
elementary excitations in the exact treatment of 1d Heisenberg antiferromagnets.
They have a gapless spectrum. Note, however, that this is not a consequence of the
Goldstone theorem: although we have here a continuous symmetry, this symmetry
113
Fig. 6.29
is not broken in the ground state; there is no long-range order in the ground state in
this case. The singlet pairs in the RVB state are spherically symmetric objects, there
is no preferred orientation in the spin direction, and the continuous spin-rotation
symmetry is not broken.
This remark is important when we compare the properties of antiferromagnetic
spin chains with spin 12 and those with other spins, e.g. spin 1. For a long time it was
believed that all the main features in these cases are the same. However, relatively
recently D. Haldane has shown that this is not the case, and that there exists a gap in
the spectrum of elementary excitations in chains with integer spins S = 1, 2, . . . ,
whereas no gap exists for half-integer spins 12 , 32 , . . . . This Haldane gap is now
indeed observed in several quasi-one-dimensional antiferromagnets with spin 1.
A simple qualitative picture which can illustrate the origin of this difference is
the following. One can treat spin 1 at a site as composed of two parallel spins 12 .
Then one can visualize the following situation: one can form singlet pairs of these
S = 12 subspins with neighbours both to the left and to the right, see Fig. 6.29. In
contrast to the case of spin 12 considered earlier, here all bonds are equivalent, and
such a state is nondegenerate, whereas it was doubly degenerate for spin 12 (either
(12)(34)(56). . . , or (01)(23)(45)(67). . . singlets). Consequently, we cannot create
here the usual domain walls, as there are no different domains. This simplified
picture explains the appearance of the Haldane gap, although the real mathematical
proof of that is quite elaborate.
Let us now return to spin 12 antiferromagnets in higher dimensions and let us see
whether the concept of the RVB state can be used in these cases as well.
Problem: Compare, similarly to the 1d-case, the Neel and VB (or RVB) states in
(a) a 2d square lattice, (b) a 2d triangular lattice.
Solution: The case of a square lattice is simple: the energy of the Neel state shown
in Fig. 6.30 is
EN = J
1
4
2N = 12 J N ,
(6.119)
where N is the number of sites, each has four nearest neighbours, or four antiferromagnetic bonds per site, but each such bond belongs to two sites.
114
Magnetism
Fig. 6.30
Fig. 6.31
Fig. 6.32
A typical valence bond state is shown in Fig. 6.31. There exist actually many
possibilities; we can connect lattice sites by bonds in many different ways which
will give rise to RVB. The energy of each of these states is
3
N
3
EVB = J
= J N .
4
2
8
(6.120)
(the energy is 34 J per singlet bond, but there is one such bond per two sites).
Thus the simple valence bond state (without resonance switching of bonds) here
is worse than the Neel configuration. What will be the situation with resonance?
Numerical calculations show that the Neel state of Fig. 6.30 (with LRO, but with
a lot of fluctuations, of course) is still the ground state of the 2d Heisenberg model
on a square lattice.
The triangular lattice is a bit more tricky. Let us try to make the Neel state; we
immediately see that we are in trouble already at the first step, see Fig. 6.32. When
we consider the basic block of such a lattice a triangle with antiferromagnetic
interactions the situation is not trivial: if we only consider ordering of spins
115
S1
S3
S2
Fig. 6.33
Fig. 6.34
up and down, then whatever we do here, one bond is always wrong; this is called
frustration. One cannot subdivide a triangular lattice into two sublattices with spins
up and down so that every site of one is surrounded by the sites of the other (i.e.
when all nearest neighbours of one sublattice belong to the other). Thus the simple
Neel solution with collinear spins is not good, and we have to think of something
else. The best classical state is the one shown in Figs. 6.33 and 6.34. In this
state the spins on the triangle are directed at the angle 2/3 to each other. For such
a state the whole lattice is subdivided into three sublattices, and at each triangle
we have the same situation, all the angles between neighbouring spins are the
same, 2/3. The energy of such state is
2 1 Nz
3

= J N ,
E2/3 = J cos
3 4; 2
8
8 9:
(Si Sj )
(6.121)
(z=6)
where z is the number of nearest neighbours.

The valence bond state shown in Fig. 6.35 has the energy
3
N
3
EVB = J
(6.122)
= J N .
4
2
8
Thus in a 2d triangular lattice the energy of the valence bond state EVB is equal to the
energy of the best classical long-range ordered state E2/3 , i.e. the valence bond
state (random covering of the lattice by singlets), even without their resonance
116
Magnetism
Fig. 6.35
(which would decrease the VB energy still further) is degenerate with the best
classical, Neel-like state, and the resonance could have made the RVB state the
ground state.
These considerations are of course only suggestive; they do not yet prove that
the RVB state is indeed the ground state of the Heisenberg model on a 2d triangular
lattice. The point is that both the RVB state and the state with some long-range order
would be modified (and their energies would be reduced) by quantum fluctuations.
Thus a priori we cannot say which state would have the lowest energy. The most
detailed numerical calculations carried out until now seem to show that there still
exists in this case a long-range order of the 120 degrees type. However, for
example, already a small doping can make corresponding systems RVB-like.
The RVB state is a spin liquid (it is made of singlets, but without any long-range
order). The notion of RVB is at present an important concept in different fields,
e.g. in the theory of high-Tc superconductivity and in the physics of frustrated
magnets such as, e.g. the kagome lattice, a 2d lattice of corner-sharing triangles, or
the pyrochlore lattice, a 3d lattice of corner-sharing tetrahedrons. What is important is the fact that in such spin-liquid states with short-range RVB (short-range
antiferromagnetic correlations) the excitation spectrum is gapful; this is the typical
situation for such quantum liquid states (although there may exist also spin liquids
with a gapless spectrum). This is also important for many situations with quantum
critical points, see Section 2.6 and Section 10.2 below.
One extra interesting and important feature of the RVB state is the unusual
character of elementary excitations in it which we already encountered, when
considering the 1d case. The creation of such excitations is illustrated in Fig. 6.36:
starting from a particular VB state, Fig. 6.36(a), we first break one singlet bond,
making it a triplet, Fig. 6.36(b), and then recommute singlets, moving spins
apart, Fig. 6.36(c). Thus, similar to the 1d Heisenberg model we get two independent
excitations, each with spin 12 , which are in fact very similar to the spinons introduced
above for the 1d case. Note that they have no charge (at each site we have a positive
nucleus with charge +e, and a localized electron with charge e). Let us now
remove an electron, i.e. create a hole (e.g. by doping). It is much easier to remove
(a)
(b)
117
(c)
Fig. 6.36
Fig. 6.37
an electron from a broken bond (e.g. from the site in Fig. 6.36(c)) we do not
have to spend energy to break a singlet. Then we will have the situation shown in
Fig. 6.37, that is we will create a site (a hole) + with charge +e and without
spin a charged but spinless excitation, which is now called a holon.
This kind of decoupling of spin and charge degrees of freedom, or spincharge
separation, is definitely valid in 1d models. As to the 2d-case, this is still questionable and not proven. But this is a very appealing picture; it is widely used in
describing systems with strong electron correlations, such as cuprate superconductors (the main proponent of this idea is P. W. Anderson).
6.4.3 Two-dimensional models

Now, we have already gradually switched from 1d to 2d models. What about them?
(3a) 2d Ising model
The two-dimensional Ising model is exactly soluble on a square lattice and in some
other cases; it was the first example of an exact solution of a nontrivial 2d problem
(the famous solution of Onsager). For the Ising model on a square lattice there is a
real phase transition, and the long-range order appears at
Tc =
|J |
.
arc tanh( 2 1)
(6.123)
118
Magnetism
(a)
(b)
Fig. 6.38
?
?
?
?
Fig. 6.39
It is a second-order phase transition, but not mean-field-like; thus, e.g. the specific
heat has not a jump, Fig. 6.38(a), but a logarithmic divergence, Fig. 6.38(b), i.e.
close to Tc we have c(T ) ln |T Tc |.
One can also calculate exactly some other properties for this case. For example
the order parameter (magnetization or sublattice magnetization) behaves close to
Tc as M (Tc T )1/8 . Note that these exact results for the specific heat and for
the order parameter satisfy the scaling relations of Section 2.5, as they should (the
logarithmic dependence of the specific heat in scaling relations means that the
corresponding exponent should be taken as zero).
An interesting situation exists in the two-dimensional Ising model on a triangular
lattice. As discussed above, see, e.g. Fig. 6.32, in such a case we have frustration,
and many states have the same energy (the three-sublattice solution with 120 spins,
discussed in the previous section, cannot be formed in the Ising model, where all
spins must be either up or down). And indeed the exact solution of this model
(G. Wannier) shows that there exists a macroscopic number of degenerate ground
states, so that this system has finite entropy at T = 0.
The simplest way to understand this is to form first a two-sublattice antiferromagnetic ordering on a honeycomb sublattice of a triangular lattice, Fig. 6.39. Then
the spins at the centres of each hexagon are in zero molecular field, and we can put
their spins at random. All such states will have the same energy. The number of
119
M
1
1
3
13
Fig. 6.40
Fig. 6.41
such states is 2N/3 , i.e. the ground state entropy can be estimated as S 13 N ln 2.
The exact value is not far from this simple estimate.
Yet another interesting consequence of this simple picture is that, as one can
easily see, we expect in this case that in an arbitrary small external field, say up,
all these free spins immediately orient along the field, so that there appears a
net magnetization of the sample equal to 13 B per site. Similarly, at any small
H < 0 the magnetization will be negative, 13 B . Thus there will be a jump of
magnetization at H = 0, and the second jump to full magnetization 1B will occur
at still higher critical field (at T = 0). As a result the magnetization at T = 0 will
change with (parallel) field as shown in Fig. 6.40. Such magnetization plateaux are
very typical for frustrated magnetic systems.
(3b) 2d xy model. Topological excitations (vortices)
Let us now consider the two-dimensional xy model with the Hamiltonian of the
type (6.109) on a square lattice. In contrast to the 1d case, here one does not yet
know the exact solution. An interesting feature of this model is that it possesses
interesting types of excitations topological excitations. Thus let us take the simple
ferromagnetic state shown in Fig. 6.41. One possible type of elementary excitation
is ordinary spin waves. But we can also create configurations, e.g. like those shown
in Figs. 6.42 and 6.43.
120
Magnetism
(a)
(b)
Fig. 6.42
2
1
Fig. 6.43
Far from the centres of such excitations the spins are locally parallel, as they
should be in a ferromagnet, but when one approaches the centres this rule is
of course violated. We cannot destroy such excitations (defects) by continuously
rotating the spins they are topological defects. They are very similar to vortices in
liquid He. (Actually, starting from the ground state of Fig. 6.41, we cannot create
one such excitation, because it has a different topological quantum number than
the ground state and will necessarily lead to a distortion of the spin structure at
infinity. But we can create a pair of such excitations, similar to a vortexantivortex
pair, so that the corresponding field (distribution of arrows) will look like the field
of a dipole, and the corresponding distortion will decrease with distance rapidly
enough.)
In a Bose gas the broken symmetry is the gauge symmetry, the phase of the order
parameter, a aei , 0 < < 2. The order parameter there is a0 = |a0 |ei . In
the xy model there exists the same symmetry: the order parameter is here the
two-dimensional vector S = |S|ei , with the phase showing its orientation in the
xy-plane being 0 < 2 . Thus there exists an exact mapping of the Bose gas
onto the xy model, and not only in the 2d case! The connection between these two
models is the following:
Bose condensation ordering in the xy model;
Bogolyubov sound spin waves (magnons);
Vortices topological excitations in the xy model.
121
Fig. 6.44
Fig. 6.45
Fig. 6.46
In a superfluid the current (or velocity of superfluid motion) is v sf , and

in a vortex the current circulates around the core, see Fig. 6.44. In the spin model
this would correspond to a spin configuration shown, e.g. in Figs. 6.42 and 6.43,
i.e. this spin configuration is characterized by nonzero gradient of such that the
integral along a contour surrounding the centre is
1
dl = 2 .
(6.124)
A defect or excitation of the opposite sign, Fig. 6.45, would correspond to an
antivortex (opposite circulation equal to 2 ). The circulation around the vortex
is, in general, 2 n, where n is called the winding number. In previous figures
we had n = 1. One can draw spin configurations for example corresponding to
n = 2, Fig. 6.46, etc.
There is no usual phase transition in the 2d xy model at nonzero temperatures,
at which real long-range order would appear (this also follows from the Mermin
Wagner theorem, see Sections 4.4.4 and 6.3.2), but there is a point T at which
the correlation between spins S0 Sr , which above T behaves as r (T ) with
= 2(T /T ), below T , still decaying with distance, becomes so strong that
122
Magnetism
the susceptibility r S0 Sr diverges. This is also a special kind of phase

transition, which is called a BeresinskiiKosterlitzThouless (BKT, or often simply
KT) transition. Although strictly speaking there is no long-range order at 0 < T <
T (the average S = 0, in accordance with the MerminWagner theorem), the
fact that the susceptibility diverges means that the magnetic response of such a
system is in many respects similar to that of an ordered state. For example, as
we mentioned already, the Bose condensation phenomenon may be mapped onto
the xy model. It turns out that below the BKT transition the two-dimensional
Bose system (e.g. thin films of liquid 4 He) would be superfluid, despite the formal
absence of long-range order.
Actually the KosterlitzThouless transition is connected with the topological
excitations vortices which we have just discussed. One can show that the
energy of an isolated vortex with one flux (winding number 1) is equal to

R
,
(6.125)
Ev = J ln
where is the coherence length (crudely speaking it is the radius of the inner part,
the core of the vortex, where the modulus of the order parameter goes to zero the
so-called normal core in superfluids or superconductors), and R is the dimension
of the sample. At the same time each core can be at any point of the sample with
the area R 2 , i.e. the entropy of such a system is
S = ln
R
R2
= 2 ln .
2
(6.126)
The total free energy of the system is then

F = E T S = ( J 2T ) ln
R
.
(6.127)
When T < T = J /2, the free energy is minimized when there are no vortices
in the system; such a state would have magnetic stiffness, or, in a 2d Bose system,
superfluid properties. For T > T it is favourable to create vortices, and these
properties would disappear.
One can also give a somewhat different interpretation of this phenomenon. One
can show that the interaction between vortices separated by distance r is also
logarithmic, ln(r/ ). This interaction is repulsive for the vortices with the same
circulation and attractive for vortex and antivortex. If there exist in a system vortices
and antivortices (e.g. thermally excited), at T > T they will be unbound, but for
T < T they will be bound in pairs (the existence of free vortices is unfavourable at
T < T by the same arguments as follow from (6.127)). Unbound vortices destroy
superfluidity, whereas the bound ones do not.
One more remark is relevant here. The topological character of vortices in
the 2d xy model means that we cannot destroy such excitations by a continuous
6.5 Defects and localized states
123
deformation of the spin structure. It turns out that despite the apparent similarity, in
this respect the Heisenberg model is principally different from the xy model. For
the Heisenberg model we can also draw spin configurations similar to those shown
in Fig. 6.42. However in this case we can annihilate this defect by continuously
deforming the spin structure, moving spins out of the plane, so that in the end,
for example, all spins would point up, restoring perfect ferromagnetic ordering.
This remark actually has a rather deep meaning: what is important here is the
space of the order parameter. Whereas in the xy model (or, to this end, for the
Bose condensation) the order parameter is a complex scalar, and its possible values
live on a circle (are characterized by an arbitrary phase 0 2 ), the order
parameter S of the Heisenberg model is a vector characterized by its modulus
and two Euler angles, i.e. it can take any value on a sphere with the radius |S|.
The vortex of Fig. 6.42 is a state represented by a circle on this sphere, e.g.
its equator. But we can continuously deform this circle, moving it e.g. to the
North Pole (all spins pointing up), where it would reduce to a point and finally
disappear. It is clear that this process is forbidden if the space of the order parameter
is not a sphere, but a 2d circle, as for the xy model. Different spaces of possible order
parameters, which can exist in different ordered states in condensed matter systems,
determine possible types of excitations in them, and finally their thermodynamic
and transport properties.
6.5 Defects and localized states in magnetic and other systems

An interesting problem arises when there exist external defects in an otherwise
regular material (e.g. impurities with different spin S in a system with spin S,
or a site with different exchange interaction J
= J ). A similar (mathematically)
problem arises in lattice theory if there exists an impurity atom with mass M
different from the mass M of the regular lattice, or in a system of tight-binding
electrons with an impurity.
Often there appear in this case localized states, or impurity states: localized
phonons, or localized magnons, or localized (impurity) electron states.
Mathematically the problem is the following: using the Hamiltonian of the
type of (4.26), or (6.68), or the standard tight-binding description of electronic
systems, see, e.g. the term H in (12.4), and adding an impurity at the origin of
the coordinates, we can write down the equations for the corresponding operators
(phonon operators, or the spin ones, or the electron creation and annihilation
operators) in the form
ci = 0 ci +

j
tij cj + i,0 c0 ,
(6.128)
124
Magnetism
1 l
( > 0)
l
e*
emax e**
1
l
( l < 0)
Fig. 6.47
cf. (6.89), see also the standard description of the tight-bound model for electrons.
Here the last term describes a perturbation at the site i = 0 (e.g. different spin,
or different mass, or different potential for electrons).
The standard solution of this problem is the following. Let us make a Fourier

transform ci = q eiqRi cq . Then equation (6.128) takes the form

ck .
(6.129)
[ 0 t(q)] cq = i,0 c0 =
k
From equation (6.129) we find

k ck
k ck
,
cq =
0 t(q)
q
(6.130)
where q = 0 + t(q) is the spectrum of the system without impurities. Now, take

q of both the left- and right-hand sides of (6.130):

q
or, denoting A =

q
cq =

k
ck

q
cq , we see that A = A
1
,
q
1
q q ,
1 1
=
.
q
q
(6.131)
or
(6.132)
Let us consider a typical situation in which the bare spectrum is confined, 0 <
q < max . If we initially consider the system in a box so that the values of q are
discrete (and q too), the graphic solution of (6.132) looks as shown in Fig. 6.47.
The solutions are given by the crossing of the straight line 1/ with the curves
representing the right-hand side of equation (6.132). Thus we see that there exist
solutions inside the former band of excitations (each energy is slightly shifted, but
for the volume V these solutions fill the entire initial band 0 < < max ,
6.5 Defects and localized states
125
1
w emax
0
emax
1
w
Fig. 6.48
the wavy line in Fig. 6.47). But besides that there may appear solutions outside the
band the localized modes.
For the situation described by equation (6.129) such states, if they do exist, lie
below the band (at < 0) for < 0 (attraction), or at > max , i.e. above the
continuum of the band states for > 0 (repulsion).
Now, in which cases do such split-off localized states exist, and how does their
energy depend on the strength of the perturbation ? The answer is different for
1d, 2d and 3d systems.
#

Let us take a continuous limit of (6.132) (i.e. take dq instead of q ),
1
=
1
dq
=
q
max
() d
.
(6.133)
The result depends on the behaviour of the density of states () at the edge of the
band ( 0, max ).
(1) One-dimensional systems. Here () 1/ as 0 (if the spectrum is

quadratic, q q 2 , as for electrons or for ferromagnons). The integral in equation (6.133) is divergent, and we would have the situation shown in Fig. 6.48.
Thus there is a bound (or antibound) state for any , however small. For
= || 0 the energy of the localized state behaves as
2 .
(6.134)
The result depends also on the initial spectrum (q), of course.

(2) Two-dimensional systems. In the 2d case the typical density of states at the edge
of the band is ( 0) = const. (for electrons). Then the integral in (6.133)
is logarithmically divergent. The bound state then exists also at any , and for
small its energy is
e1/ .
(6.135)
126
Magnetism
Fig. 6.49
(3) Three-dimensional case. Here, for electrons, () and the integral is

finite. This means that Fig. 6.47 actually looks like that shown in Fig. 6.49, and
the bound state exists only for exceeding a certain critical value, || > c .
For different types of spectrum (e.g. if q q, and not q 2 ) the criteria for the
existence, and the energies of the bound states may be different (and the form of the
perturbation may also be different from that introduced in (6.128)), but the general
scheme is always like the one described above (it is often called the SlaterKoster
method). One can also use the same technique for considering possible bound state
formation due to interaction between excitations (e.g. the formation of the bound
state of magnons, cf. the qualitative discussion in Section 6.3.1, Fig. 6.21, or the
famous solution of the Cooper pair problem which led finally to the Bardeen
CooperSchrieffer theory of superconductivity).
7
Electrons in metals

This very short chapter serves as a reminder of the main properties of Fermi systems
and a useful collection of respective formulae. We consider here noninteracting
electrons with the spectrum ( p) = p2 /2m. In the degenerate Fermi gas the number
of states in the interval (p, p + dp) is (per spin)
d3 p
4p 2 dp
p 2 dp
=
V
=
V
.
(7.1)
3
(2
h)3
(2
h)3
2 2
h
Fermions obey the Pauli principle, and at T = 0 the electrons in metals occupy all
states up to a certain momentum the Fermi momentum pF , so that the electron
density (for both spins) is
pF
1
pF3
N
2
=
p
dp
=
,
(7.2)
n=
3
3
V
2
h 0
3 2
h
or the Fermi momentum
V
hn1/3
pF = (3 2 )1/3
(7.3)
and the Fermi energy
h 2/3
pF2
= (3 2 )2/3
n .
(7.4)
2m
2m
The distribution function of electrons n(p) has the well-known form shown in
Fig. 7.1. At T = 0, n(p) = 1 for |p| < pF and n(p) = 0 outside this region; |p| =
pF determines the Fermi surface. At finite temperatures some of the electrons
are excited from the states below the Fermi surface to empty states above it,
i.e. there appear electronhole pairs. Consequently the distribution function n(p)
is broadened, see Fig. 7.1. The width of the region around F in which n(p)
significantly changes is T .
2
F =
127
128
Electrons in metals
n(p)
T=0
T=0
p
Fig. 7.1
Often one introduces the quantity rs (the average distance between electrons) by
the relation
1
4 rs3
V
= =
.
N
n
3
Then one can express different quantities through rs :
1/3
1
9
1.92
3.63 1
A ,
= =
pF =
4
h rs
h rs
rs /a0
(7.5)
(7.6)
where
2
h
(7.7)
me2
is the Bohr radius. Usually one uses the dimensionless parameter rs = rs /a0 , measuring rs in units of the Bohr radius a0 , and writes all formulae in terms of rs ,
omitting a0 . Then the situation with rs 1 corresponds to a high-density electron
gas, and that with rs 1 to the low-density limit. Typically in metals rs 23
(see Table 2.1 in Ashcroft and Mermin (1976)). From (7.6) the Fermi velocity is
a0 =
vF =
cm
pF
4.2
108
=
,
m
rs
sec
(7.8)
50.1 eV
,
rs2
(7.9)
F
58.2
= 2 104 K .
kB
rs
(7.10)
and the Fermi energy is

F =
or the degeneracy temperature TF is
TF =
Note that the parameter rs is actually a measure of the relative importance of the
Coulomb interaction in a metal. Omitting numerical factors, we see that the ratio of
the typical value of the Coulomb interaction VC = e2 /rs (the Coulomb interaction
129
at the average distance between electrons) to the characteristic value of the kinetic
energy, which is of the order of F , is
< 2
h
e2
me2
rs
=
= rs .
(7.11)
rs =
VC /F =
2
2
a0
rs mrs
h
(We have used here equations (7.4), (7.5) and (7.7).) Thus the small values of rs
(dense electron systems) correspond to weak interactions, VC /F 1, whereas
large values of rs (low-density electron gas) correspond to systems with strong
interactions. These considerations will be very important for us later on, in
Chapters 1012. Note also that for the typical values of rs for ordinary metals,
rs 23, the electronelectron interaction is of the same order as the kinetic
energy or the bandwidth, which is actually the most difficult case for theoretical
treatment. Luckily, in most normal metallic systems one can still use the description very much resembling that of a Fermi gas; the justification of this possibility
is given by the Landau Fermi liquid theory, see Chapter 10.
The density
of electrons which, in particular, is used in the transforma# of states
#
tion from dp to () d, for free electrons in the three-dimensional case has
the following form:

2m
3n
m
=
.
(7.12)
() =
2
2
2 F F
2
h
h
At the Fermi level
F = (F ) =
mpF
23
3n
.
2 F
(7.13)
7.1.1 Specific heat and susceptibility of free electrons in metals

The total energy of the electron system at temperature T and the density of electrons
are given by the following general expressions:

E = d () f ()
(7.14)

n = d () f () .
(7.15)
Here () is given by (7.12), and f () is the Fermi distribution function (3.5),
f () =
1
e()/T
+1
(7.16)
The chemical potential (T ) should be determined from (7.15) and put into (7.14),
and then the specific heat, entropy and other characteristics can be calculated.
130
Electrons in metals
The integrals in (7.14), (7.15) the so-called Sommerfeld integrals are

calculated as follows (cf. Landau and Lifshits (1980)). The general form is

() d
I=
,
(7.17)
()/T + 1
e
0
where is a function such that the integral is convergent. Let us make a change of
variables: = T z,
/T

( + T z)
( T z)
( + T z)
T
dz
=
T
+
T
.
dz
dz
I=
z+1
z + 1
e
e
ez + 1
/T
0
0
(7.18)
1
1
In the first integral we write ez +1 = 1 ez +1 ; then

( + T z) ( T z)
I=
() d + T
dz .
(7.19)
ez + 1
0
0
For low temperatures T F we can make an expansion in T z in the second
integral and integrate by parts:

3

z dz
z dz
2
1 4
+ 2 T ()
+
() d + 2T ()
I=
z
e +1
ez + 1
0
0

2 2
7 4 4
=
T () +
T () .
() d +
(7.20)
6
360
0

1
(x) =
is the Riemann function
nx
0
n=1
( 32 ) = 2.61, ( 52 ) = 1.34, . . . ; ( 32 ) = 12 , ( 52 ) = 34 , . . .
(7.21)
zx1 dz
= (1 21x ) (x) (x) ,
ez + 1
As a result, from (7.14)(7.20) we obtain:

%
&
F
2 2
T (F )
n=
() d + ( F ) (F ) +
6
0
(with F = (T = 0)). From the condition that n(T ) = n(0) we find
4

5
1 T 2
2 2 (F )
T
= F 1
.
= F
6
(F )
3 2F
(7.22)
(7.23)
For the energy, finally we get

E = E0 +
2 2
T (F )
6
(7.24)
131
and the specific heat of free electrons is

2/3 mT 1/3
E
2
2
T
=
cV =
=
n T .
T (F ) =
n
2
T V
3
2
F
3
h
(7.25)
Here we have used (7.13), (7.3), and introduced the standard notation = c(T )/T .
Problem: Find the entropy of a Fermi gas at low temperatures.
Solution:
dS
c=T
dT

= S =
0
1
c(T ) dT =
T
1
T dT = T ,
T
(7.26)
i.e. the entropy behaves exactly as the specific heat itself.

Problem: Estimate cP cV for T 0.
Solution:
S V
cP cV = +T
P T T
V
P T

,
and
V
T

P
S
=
P

.
(7.27)
If as T 0, S T n (the entropy should go to zero as T 0, according to the

Nernst theorem, or the third law of thermodynamics), then
S 2
P T
T 2n+1 .
cP cV T V
(7.28)
P T
Thus for electrons, for which c = T and S = T , i.e. the coefficient n in equation
(7.28) is equal to 1, cP cV T 3 , i.e. it is much smaller than the specific heat c
itself. Therefore we can ignore the difference between cP and cV for electrons at
low temperatures.
Yet another conclusion can be obtained from the treatment presented above.
According to general thermodynamic relations, see e.g. (7.27), the thermal expansion satisfies

1 dV
1 S
=
,
V dT
V P
thus the electronic contribution to the thermal expansion is T , i.e. in metals the
thermal expansion (T ) at low temperatures is linear in temperature,
(T ) const T .
(7.29)
This result is very similar to the Gruneisen equation for phonons (4.53): here also
the thermal expansion is proportional to the specific heat (and both are T ). (Do not
132
Electrons in metals
confuse the Gruneisen constant in (4.53) with the coefficient in the electronic
specific heat (7.25)!) Sometimes one speaks about the Gruneisen constant for
electrons e , determined by the same relation as (4.53), with being the electronic
contribution to the thermal expansion, and cV being the electronic specific heat.
Note also that as the phonon specific heat at low temperatures is T 3 , see equation
(4.39), the phonon contribution to the thermal expansion, according to (4.53), is
also T 3 , thus indeed at the lowest temperatures both the specific heat and thermal
expansion in metals are determined by the electron contribution.
Similarly (even simpler) one can calculate the Pauli (spin) susceptibility of free
electrons:
p = 2B (F ) .
(7.30)
Thus both the specific heat c and magnetic susceptibility are proportional to
(F ), and we have the ratio (the so-called Wilson ratio)
RW =
2
=1,
32B
(7.31)
i.e. the Wilson ratio for free electrons is 1.

When we include an interaction between electrons, both cV and are modified,
in principle differently; and the Wilson ratio tells us a lot about what is going
on. For example, for metals close to ferromagnetism (e.g. Pd, Pt) is enhanced
(exchange enhancement). Thus the Wilson ratio is 12 for normal metals (Fermi
liquids), and may be 1 in nearly ferromagnetic metals (see Chapter 10). The
notion of the Wilson ratio is widely used, in particular, in treating heavy-fermion
systems, see Chapter 13.
8
Interacting electrons. Green functions and
Feynman diagrams (methods of field theory in
many-particle physics)
In this chapter I will give a short summary of the Green function method and
Feynman diagram techniques in application to condensed matter physics. This in
itself is quite a big field, and the full description of it, with all derivations and
all details, requires a quite lengthy discussion. I will only present the main ideas
of this method and give recipes which can be used in practical calculations. The
detailed discussion of these points, with all the derivations and proofs, can be found
in many books specially devoted to these problems, such as the books by Mattuck
(1992) and Fetter and Walecka (2003). This method is also discussed in detail in
the books by Abrikosov, Gorkov and Dzyaloshinskii (1975) and by Mahan (2000),
mentioned in the Introduction, where also many applications can be found. A very
good summary of this method is also contained in the book by Schrieffer (1999).
8.1 Introduction to field-theoretical methods in condensed matter physics
In general, electrons can interact with external forces or potentials, e.g. impurities,
with phonons, with magnons, and also interact between themselves. Postponing the
discussion of interaction with other excitations till later and concentrating on the
electronelectron interaction, we can write the electron Hamiltonian in ordinary
quantum mechanics as

H=
Hi + 12
V (r i r j )
(8.1)
ij
2
h 2
Hi =
+ v (r) .
2m i
(8.2)
Here v (r i ) is the external potential and V (r r ) is the electronelectron interaction. Thus, e.g. for electrons this is the Coulomb interaction,
V (r r ) =
133
e2
.
|r r |
(8.3)
134
Interacting electrons
In the second quantization the interaction with the external potential has the form
v (r) (r) = v (r)
(r)
(r) ,

(8.4)
(r) are creation and annihilation operators for electron with spin
(r),
where
at position r. The interaction between electrons is
V (r r ) (r) (r ) =
(r)
(r)
(r )
(r ) . (8.5)
V (r r )
Here , = 12 are spin indices.

It is usually convenient to work in the momentum representation. Making the

Fourier transform (r)

= 1 p ei pr c p , we get finally for the total Hamiltonian
H=
p cp, c p, +
p,
1
2
v (q) cp+q, c p,
pq,

p p q,
V (q) c p+q, c p q, c p , c p, .
(8.6)
Here is the total volume (needed for normalization in the Fourier transform; we
use here the notation to distinguish it from the interaction V ).
The electron density operator (r) = (r) (r) is transformed into
(q) =
cp, c p+q, .
(8.7)
p,
One can conveniently describe the interaction terms in (8.6) using pictures
diagrams. Thus, for example, the interaction of an electron with an external (impurity) potential v will look like
p,
Particle in the state ( p, ) is

annihilated (operator c p, in
the second term in (8.6))
p+q,
In the vertex (cross)

stands v (q)
Particle in the state ( p + q, )

is created (operator cp+q, in
the second term in (8.6))
(8.8)
In this process the momentum is not conserved, because here we describe scattering
by an external potential, e.g. by an impurity which we assume to be infinitely
p,
thus describes the electron with
heavy and which has no recoil. The line
momentum p and spin , and the cross corresponds to the interaction with the
external potential.
8.1 Introduction to field-theoretical methods
135
The last term in (8.6) (the interaction between electrons) may be depicted as
p
p+q
(8.9)
q
p
p q
Here the solid line

again describes the electron, and the dashed line
q
represents the interaction V (q). Thus in fact here we describe this
process (Coulomb interaction) as an exchange of a virtual photon: one electron
q
, and another absorbs it. There exists another, alternative
emits a photon
form to depict the electronelectron interaction, which we will also use sometimes:
(8.10)
In this form the interaction corresponds to the vertex with two electron lines going
in and two out.
In the process (8.9) the total momentum is conserved (this is why in (8.6) we
have these particular momenta, p, p , p + q, p q: the total sum of momenta of
incoming and outgoing particles should be the same). Sometimes one writes the
interaction term as

1
V ( p1 p4 ) cp1 , c p2 , c p3 , c p4 , ,
(8.11)
2 p1 , p2 , p3 , p4
( p1 + p2 = p3 + p4 ),
p4
p1
p1 p4
p3
(8.12)
p2
(electrons with momenta p3 , p4 which were present in the initial state are annihilated, and two other electrons, with momenta p1 and p2 , are created instead).
Similar pictures can also be drawn for other types of interaction (e.g. for the
electronphonon interaction, see below).
The most important fact is that such diagrams, pictures, are not just a way
to depict different terms of the Hamiltonian and different processes; they may be
really used to calculate many properties of the material. This is one of the virtues of
the Green function method; the corresponding technique is the Feynman diagram
technique.
136

The following section is of a more technical character and can be omitted at
first reading, although it is necessary if one wants to understand better how the
techniques used below can be really derived.
To explain the method of Green functions and Feynman diagrams we have to
go back for a while and discuss different formulations of quantum mechanics
so-called different representations.
1. Standard quantum mechanics is usually described in the so-called Schrodinger
representation in which the wavefunction (r, t) is time dependent, and operators are constant in time:
i
S (t) = HS (t)

t
(8.13)
(in the future we often put

h = 1).
The formal operator solution of (8.13) is
S (t) = ei Ht S (0) .
(8.14)
2. An alternative is the Heisenberg representation, in which we take wavefunctions

as time independent and ascribe all time dependence to operators, with the
dependence
ei Ht ,
A H (t) = ei Ht A(0)
(8.15)
H = S (0) = ei Ht S (t) .
(8.16)
and the wavefunction
Then the physically measurable quantities (averages) do not change:

S (t) A S (t) dr
A = S (t)|A|S (t)
from (8.14)
S (0)|ei Ht A ei Ht |S (0) = H |A H (t)|H .
(8.17)
Now the function is taken as constant, but the operator A H (t) obeys the
equation (cf. equation (6.87))
i
'
(

AH (t) = A H (t), H .
t
(8.18)
3. Yet another convenient form is the mixed representation the so-called interaction representation, which is actually the basis of the Feynman diagram
technique.
137
If the Hamiltonian consists of two parts,

H = H0 + H ,
(8.19)
where H0 is the bare Hamiltonian of noninteracting particles (or anything

we can solve simply), and H is an interaction, then we can introduce the
representation where operators depend on time as in the case of free particles,
i.e. as in (8.15), but with H = H0 ,
ei H0 t ,
A I (t) = ei H0 t A(0)
(8.20)
and the wavefunctions also are time dependent, but only because of the presence
of the interaction H :
I (t) = ei H0 t S (t) = ei H0 t ei Ht S (0) .
(8.21)
If H = 0, then indeed in (8.21) we have ei H0 t ei H0 t = 1, and I (t) (8.21)

becomes identical with H (8.16), i.e. time independent. In general, the operators

H0 and H in (8.21) do not commute, and ei H0 t ei(H0 +H )t
= ei H t !

Problem: Try to find the expression for eA eB when the operators A and B do not
commute.
Solution:
eA eB = eA+B e 2 [A,B] ,
(8.22)
(
B]
is nonzero; if the next commutators, e.g. A,
[A,
B]
,
if only one commutator [A,
are nonzero, then the corresponding formula looks more complicated.
'
From the previous definitions one can show that
I (t) = HI (t) I (t) ,

t
(8.23)
HI (t) = ei H0 t H S ei H0 t
(8.24)
i
where
according to the general rule (8.20).

The interaction representation is convenient because we can usually choose H0
such that the time dependence, or -dependence in a Fourier transform, is simple,
for example starting from the noninteracting electrons (the ideal Fermi gas, see
Chapter 7). Then we can solve equation (8.23) or an equivalent one by perturbation
theory, which in turn can be conveniently represented by Feynman diagrams. This
is actually the main reason for introducing the interaction representation.
138
Let us try formally to solve equation (8.23) with HI given by (8.24). We cannot
write the solution as
I (t) = const ei
#t
HI (t ) dt
()
because now the operators HI (t) do not commute with themselves at different
moments (t1 , t2 ). But we can try to solve it by perturbation theory. For that we write
down an integral equation for I (t), equivalent to (8.23):
t
I (t) = I (t0 ) i
HI (t ) I (t ) dt
(8.25)
t0
and seek the solution as I (t) = I(0) (t) + I(1) (t) + . In lowest order I(0) =
I (t0 ) (as though H in (8.25) were absent). Then
t
(1)
I (t) = i
HI (t1 ) dt1 I (t0 ) ,
t0

I(2) (t)
=
t0
HI (t1 ) dt1
(8.26)
t1
HI (t2 ) dt2
t0
I (t0 ) , etc.
One can write the general solution in terms of the so-called S-matrix:
I (t) = S(t, t0 ) I (t0 ) ,
(8.27)
and for the S-matrix we have the series

t1
HI (t1 ) dt +
S(t, t0 ) = 1 i
t0
+ (i)
n
t0
HI (t1 ) dt1

t0
t1
HI (t2 ) dt2
+ .
t0
tn1
HI (tn ) dtn
(8.28)
It is important that here we have the# sequence

such that t > t1 > t2 . . . tn >
# # of times

HI (tm1 ) HI (tm2 ) . . . HI (tmn ) all operators H
t0 , or, in a general term of the type
should be time-ordered (time increases from right to left, tm1 > tm2 > . . . > tmn ).
Thus the formal solution for the S-matrix should be written as
&
% t

(8.29)
HI (t ) dt ,
S(t, t0 ) = T exp i
t0
where by T we mean time-ordering under the integral. Thus in general

1 )A(t
2 )A(t
3 )] is A(t
1 )A(t
2 )A(t
3 ) if t1 > t2 > t3 , or A(t
2 )A(t
1 )A(t
3 ) for
T[A(t
t2 > t1 > t3 , etc. (Do not confuse the symbol T here with the temperature!) Note
8.3 Green functions
139
also the signs: for bosons we always have the plus sign, but for fermions one interchange of fermion operators gives a minus, two interchanges give plus, etc., which
is the usual rule for fermions.
The connection between operators in the Heisenberg and in the interaction
representation can now be written as
A H (t) = S 1 (t) A I (t) S(t) .
(8.30)
8.3 Green functions

In this section we give a brief introduction to the method of Green functions. For a
more complete treatment, see e.g. the books by Abrikosov et al. (1975), Doniach
and Sondheimer (1974), Fetter and Walecka (2003), Mahan (2000), Mattuck (1992)
and Schrieffer (1999).
Instead of working with the full many-body wavefunctions, let us introduce
certain combinations thereof Green functions, which of course contain not the
full information, but nevertheless describe a lot, and through which one can directly
express many different measurable quantities.
The definition of the most common type of Green functions (they are sometimes
called causal Green functions, to discriminate them from the retarded and advanced
Green functions) is:
6
7
(r 1 , t1 )
(r 2 , t2 ) |0 .
G (r 1 , t1 ; r 2 , t2 ) = i 0|T
(8.31)
(r, t) is the electron operator in second quantization form in the Heisenberg
Here
representation (do not confuse it with the wavefunction in (8.13)(8.16)), 0| . . . |0
is the average over the ground state (in principle unknown!), and T stands for the
so-called T-product the time-ordered sequence of operators standing after this
symbol (cf. (8.29)):

(r 2 , t2 )|0 for t1 > t2
(r 1 , t1 )
i 0|
G (r 1 , t1 ; r 2 , t2 ) =
(8.32)
(r 2 , t2 )
(r 1 , t1 )|0 for t1 < t2 .
i 0|
For nonmagnetic systems Green functions are diagonal in the spin indices, G =
G ; in future we will mostly omit the indices , . Different signs for t1 > t2 and
t1 < t2 in (8.32) are needed because of Fermi statistics; for bosons (e.g. phonons)
the sign will be the same.
How can we understand qualitatively why we have introduced such a strange
object, with the T-product? What is the physical meaning of these Green functions?
The proof of the pudding is in the eating, according to a British proverb. We will
see soon that Green functions are indeed very convenient objects. One can express
a lot of physical quantities through them. Green functions also give information
140
about the spectrum of elementary excitations in the system. But here we will first
present some qualitative arguments:
Consider the initial system in its ground state |0, and let us create an extra elec (r 2 , t2 )|0. This electron will propagate
tron at the moment t2 at point r 2 : |2 =
in time, and we want to know the fate of this electron, i.e. to look at the probability
that at time t1 it will be at point r 1 (note that in the process it will interact with
other electrons, the system may be excited, etc.). The corresponding wavefunc (r 1 , t1 )|0. The amplitude of the probability we are interested in is
tion is |1 =
(r 2 , t2 )|0.
1 , t1 )
1|2 = 0|(r
Similarly, we can consider the process in which we initially create a hole, e.g.
= (r
2 , t2 )|0, and consider its propagation to the state |1
1 , t1 )|0.
|2 = (r
2
= 0|
2 , t2 )|0.
(r 1 , t1 )(r
Then we want to know 1|
Note that it is reasonable to consider such processes if t1 > t2 (first we create a
particle, and then it propagates to a new state).
2
It turns out that it is convenient to combine these two objects, 1|2 and 1|
(which are both functions of (r 1 , t1 , r 2 , t2 )), into one, ordering the times as defined
in (8.31), (8.32). The Green function thus defined describes the motion of an
added electron from (r 2 , t2 ) to (r 1 , t1 ), with t1 > t2 , or the motion of an added
hole from (r 1 , t1 ) to (r 2 , t2 ) (and here t2 > t1 ; to combine the description of both
these processes in one function we interchanged (r 1 , t1 ) (r 2 , t2 ) for the hole).
Actually this definition goes back to Feynman who described positrons (in our case
holes) as electrons moving backwards in time.
From the definition (8.31) the particle density n(r) (also denoted sometimes
(r)) is
n(r) = 2i lim G(r, t; r , t )
r=r
t t+0
(8.33)
(the factor of 2 comes from summation over spins). For spatially homogeneous systems G(r, t; r , t ) = G(r r , t t ). Its Fourier transform (the Green function
in the momentum representation) G( p, ) is given by
3
d p d

G(r r , t t ) =
G( p, ) ei[ p(rr )(tt )] .
(8.34)
4
(2 )
One can show that the momentum distribution function for electrons n( p) is
expressed through the Green function as

d
,
(8.35)
n( p) = i lim
G( p, ) eit
t0
2
cf. (8.33).
One can also express through the electron Green function G( p, ) all thermodynamic properties of the system. There are several ways to do that. One of them
8.4 Green functions of free electrons
141
? (we denote it
is to use the general expression for the thermodynamic potential
?
here by so as not to confuse it with the volume ) (1.29), (1.30). At T = 0 the
entropy is S = 0, and
? = dN .
d
(8.36)
Using the relation

n = 2i lim
t0
G( p, ) eit
d d 3 p
,
(2 )4
(8.37)
following from (8.35) (the factor of 2 again comes from summation over spins)
and integrating (8.36) over , we obtain

d d 3 p

?
d lim
G( p, ) eit
.
(8.38)
() = 2i
t
(2 4 )
0
Thus we see that many important properties of the interacting system may be
expressed through the Green function.
By analogy with the one-particle Green function G(x; x ) or G( p, ), we can
introduce also two-particle and higher-order Green functions. In particular, the
two-particle Green function is important for treating the response of the system
to an external perturbation such as an electromagnetic field, for the discussion
of eventual instabilities of the system and for treating transport properties of the
system, such as resistivity or thermal conductivity.
8.4 Green functions of free (noninteracting) electrons

For free electrons the one-electron operators (r, t), (r, t) in (8.31), (8.32) are
connected with the creation and annihilation operators c p , c p by the usual relation
1
c p ei[ pr0 ( p)t] .
(r, t) =
p
(8.39)
Let us put this in (8.32): for Green functions of free fermions G0 (r, t) we then get
%
i i[ pr p t]
1 f p (t > 0)
G0 (r, t) =
e
(8.40)
f p
(t < 0) ,

p
where by p we denote the bare spectrum of noninteracting electrons 0 ( p), and
f p = n p = c p c p is the Fermi distribution function at T = 0,

%
1 for | p| < pF
fp =
0 for | p| > pF .
142
Using this, one can show that

G0 ( p, ) =
1
p + i sign(| p| pF )
(8.41)
where is an infinitesimally small positive number. (Alternatively we can write in

the denominator i sign( p ).)
Problem: Check (8.41) using (8.34)(8.40).
Hint: Work back from (8.34) using (8.41).
The bare Green function G0 (8.41) obeys the equation

i p G0 ( p, t) = (t) ,
t
(8.42)
i.e. it really is a Green function as introduced in mathematics (the solution of the

corresponding differential equation with a point source on the right-hand side). But
it will no longer be so simple for interacting systems.
Often one counts the energy from the chemical potential :
p = p .
(8.43)
Then
G0 ( p, ) =
1
p + i p
p = sign p .
(8.44)
From equations (8.41) and (8.44) we immediately see that the poles of the Green
function correspond to energies of (quasi)particles. This is true not only for the
Green functions of noninteracting electrons, but also in general; this is one of the
reasons why the notion of Green functions is actually very useful. One can show
(see below) that the real part of the poles gives the energy, and the imaginary part
describes the damping (finite lifetime) of quasiparticles. For free electrons this is
evident: Im G0 is proportional to 0, and the lifetime is infinite, as it should be
for a noninteracting system. For interacting systems the excitations in general have
finite lifetimes. However one can still speak about quasiparticles as well-defined
objects if their damping is not too strong, i.e. if the real part of the pole p is
Re p > Im p . And in this case the poles of G( p, ) describe such excitations.
143

One can show that the Green function G( p, ) is an analytical function in the
complex -plane. In this case one can represent it as

A( p, ) d
G( p, ) =
(8.45)

+ i
( = +0). This is called the spectral representation; the function A( p, ) is the
spectral function. For an ideal Fermi gas (with the Green function G0 ( p, ) (8.44))
the function A( p, ) is

(8.46)
A( p, ) = [ p ] = ( p ) .
Thus for free noninteracting electrons the spectral function is the delta function
at the position of the pole at the energy of the elementary excitation (here at
p = p ). In general A( p, ) may strongly differ from this simple form, but
if it still contains a relatively narrow peak, it can be interpreted as a quasiparticle
(with a finite lifetime determined by the width of the peak). But in general the
spectral function can also contain broad features which describe the incoherent
part of the spectrum, see Section 8.5.2 below.
One can obtain an important expression for the Green function G and for its
spectral function A. From the definition of the Green function one can show that
G( p, t) = i 0|T{c p (t) cp (0)}|0

N
i 0|c p (0)ei Ht c p (0)|0 eiE0 t

=
N
i 0|c p (0)ei Ht c p (0)|0 eiE0 t
for t > 0
(8.47)
for t < 0 ,
where we have used the definition of the Green function (8.31), (8.32), with the
Fourier transforms of operators c p , c p , and we took into account (8.17); E0N is the
energy of the ground state of the system with N electrons. Putting between the c and

c operators in (8.47) the complete system of functions m |mN1 mN1 | = 1
(here m is the index of the quantum state, N 1 denotes that we are dealing with
a system with N 1 particles, i.e. one electron added to or removed from our
system), we get

N +1
i m |(c p )m,0 |2 ei(m +)t for t > 0

G( p, t) =
(8.48)

N 1
+i m |(c p )m,0 |2 ei(m +)t for t < 0 .
For the spectral function A( p, ) we then have:

N+1
N1
|(cp )m,0 |2 ( m
)+
|(c p )m,0 |2 ( m
) , (8.49)
A( p, ) =
m
144
i.e. the positive-frequency part of G and A ( > 0) describes the creation (addition) of one electron to our system, whereas the negative-frequency part ( < 0)
describes annihilation (removal) of an electron; for more details see Section 8.5.2.
In general one can express A( p, ) through the imaginary part of G( p, ) itself:
%
A( p, ) for > 0
Im G( p, + ) =
(8.50)
A( p, ) for < 0 .
Here we have used the identity

P
1
= i (x) ,
x i 0
x
(8.51)

where P is the symbol for the principal part of an integral, P

g(x) g(x).
With equation (8.50), the spectral representation (8.45) gives the Green function
through the integral of Im G it is the so-called dispersion relation (similar to the
KramersKronig relation in optics).
The general properties of the spectral function A( p, ) are the following: it is
real, positive
A( p, ) = A ( p, ) > 0 ,
and obeys the sum rule
A( p, ) d = 1 .
(8.52)
(8.53)
8.5.1 Physical meaning of the poles of G( p, )

We have already mentioned that the poles of the Green function describe excitations
(quasiparticles). In general, close to a pole (with some finite (small) imaginary part)
one can write
Zp
.
(8.54)
G( p, )
p + i
(More accurately one has to use here the so-called retarded and/or advanced Green
functions.) Here p is the energy of the quasiparticle, is its damping, and the
residue Z p 1 gives the weight of the real electron in the quasiparticle (in an
interacting system the excitations quasiparticles are not bare electrons, but are
renormalized, dressed by the cloud of other electronic excitations). The factor Z p
is called wavefunction renormalization, or simply the Z-factor.
One can show that there exists an important relation: there exists in general a
jump in the distribution function of electrons n( p) at pF (the same Fermi momentum
145
n(p)
free electrons
Z
interacting electrons (T = 0)
pF
Fig. 8.1
A
Zp d ( w ep)
incoherent part
ep
w
Fig. 8.2
pF (7.3) as in noninteracting systems), and this jump is given by Zp :

n (pF 0) n (pF + 0) = Z
(8.55)
(for electrons with isotropic spectrum ZpF = Z is a constant at the Fermi surface;
in general, for a complicated Fermi surface, it may depend on the direction of p).
Thus the distribution function of electrons n( p) looks as shown in Fig. 8.1. In
systems in which Z
= 0 the nature of single-particle excitations is similar to those
of real electrons or holes. However, there may be situations in which Z = 0 these
are the non-Fermi liquid systems (see next chapter).
Problem: Using the properties of the spectral function A( p, ), show that 0
Z 1.
Solution: The pole contribution (8.54) corresponds to the spectral function
A( p, ) Z p ( p ). But according to the general properties of the spectral
function A( p, ) it is positive, and the sum rule (8.53) gives that Z 1 (Z = 1
for an ideal Fermi gas, and it is less than 1 if there exist, besides the pole, other
(positive) incoherent contributions to A( p, )).
Actually what happens is that due to interactions a part of the spectral weight
A( p, ), which for the noninteracting system was all contained in the -function
( p ), is now transferred into the incoherent part, Fig. 8.2, and the weight or
intensity of the remaining pole at p is decreased. As mentioned above, ZpF also
146
gives the jump of the electron distribution function n( p) at the Fermi surface; it is
also reduced by interaction.
There is one more very important general result in this field: one can prove that
when one starts from the noninteracting case and then adds interactions, the volume
of the Fermi surface, or in other words the value of the Fermi momentum pF , does
not change. This is known as the Luttinger theorem (see Chapter 10 for more
details).
8.5.2 Physical meaning of the spectral function A( p, )
As discussed above, from the definition of the Green function (8.31), (8.32) and
from equations (8.45), (8.49), one can show (see, e.g. Schrieffer (1999)) that the
positive-frequency part of A( p, ) (for > 0) describes the process when we add
one electron to our system. This is the process studied for example by inverse
photoemission (IPES or BIS). If we started from the ground state of an N-particle
system |0N and created an electron with momentum p, c p |0N = | p , then

the state | p is not an eigenstate of the system of N + 1 electrons. One can
decompose | p into eigenstates |mN+1 . Then the probability of finding the
system (after one electron was added) in the state with energy is
+d
P ( p) d =
A( p, ) d .
(8.56)
Thus for > 0 the spectral function A( p, ) gives directly the intensity of the spectrum of angle-resolved inverse photoemission (IPES) the probability of finding
the system in the state with energy and momentum p after adding one electron.
Similarly, for < 0, A( p, ) describes the probability of extracting, removing
an electron, leaving the system in the state with energy , which is the process
of (angle-resolved) photoemission (ARPES). For the ideal gas there is a one-toone correspondence of and p, i.e. given p we know the energy = p ; this is
described by (8.46). In this case the state c p |0N is an eigenstate with energy p (or
p + ), and corresponding PES or IPES spectra would formally contain -function
peaks. But in a system with an interaction this is no longer the case, and there is in
general an incoherent part in the spectral function A( p, ) besides the -functions
(or Lorentzians with small width) describing quasiparticles. Thus the schematic
form of spectral function shown in Fig. 8.2 (with corresponding broadening) is
actually a typical form of photoemission (or inverse photoemission) spectra.
Thus, an important short summary:
The one-electron Green function describes quasiparticles (if they exist!) these
are the poles of G( p, ). The real part of the pole gives the energy of the
147
quasiparticle, and the imaginary part describes its decay (it is the inverse lifetime of the excitation). The residue of the pole (or its intensity) Z p shows what is
the relative weight of a single electron in the total wavefunction of the excitation.
For Fermi liquids there is a jump of the distribution function of (real) electrons
at the Fermi momentum | p| = pF , coinciding with pF of noninteracting electrons;
ZpF Z is the magnitude of this jump. The spectral function A( p, ) describes
the probability of observing the system in a state with energy after an electron
with momentum p was added ( > 0) to or removed ( < 0) from the system;
A( p, ) is directly related to the spectrum measured by photoemission and inverse
photoemission.

The Green function method and diagram techniques can also be formulated for
a system of interacting electrons and phonons. The electronphonon interaction
originates from the change of ionic charge density when ions shift from their
equilibrium positions. Generally speaking, the interaction is proportional to

N

eel (r)V (r, r ) div Ze u(r ) ,
(8.57)
V
where el is the electron density at point r, N/V is the density of ions, Ze is their
charge and u(r) is the shift of ions from their equilibrium position. The kernel
V (r, r ) describes the Coulomb interaction between an electron charge at point r
and an extra ionic charge at point r . In metals, due to strong Debye screening,
this interaction is very short-range, and we can replace this interaction by one at the
same point, r = r . By using second quantization and going into the momentum
representation, one can finally reduce this interaction to the standard form (often
called the Frohlich interaction):
Heph =
gq c p+q, c p, (bq + bq ) ,
(8.58)
p,q,
so that the total electronphonon Hamiltonian (Frohlich Hamiltonian) has the form
H=

p,
( p) cp, c p, +

q
0 (q) bq bq +
gq c p+q, c p, (bq + bq ) .
p,q,
(8.59)
Here, for simplicity, we have left only the interaction of electrons with one phonon
mode, e.g. with longitudinal phonons. The interaction with other phonon modes
transverse acoustical phonons, optical phonons can be also written in the same
way as (8.58); all the specifics will be contained in the particular form of the
148
coupling constant, or the corresponding electronphonon matrix element g (which

can in principle depend on both momenta p and q).
For the interaction with longitudinal acoustical phonons one can estimate the
coupling constant g as
g
ea 2 Ze N
V
,
(8.60)
where a is the lattice constant, s is the sound velocity and is the density of the
metal. One often introduces the dimensionless electronphonon coupling constant
mpF g 2
,
(8.61)
2 2
where pF is the Fermi momentum, and, according to equation (7.13), the quantity mpF / 2 = (F ) is the electron density of states at the Fermi level (do not
confuse this with the density of the metal in (8.60)). The dimensionless electron
phonon coupling constant thus defined is, in typical cases, 1. It enters into
many important expressions describing different properties of the metal, such as
resistivity due to electronphonon scattering, etc. Probably the most famous is the
expression for the critical temperature of conventional superconductors in which
superconductivity is due to electronphonon interactions:
=
Tc = 1.14 D e1/ .
(8.62)
Analogously to the electron Green function one can introduce the phonon Green
function
D(r 1 , t1 ; r 2 , t2 ) = i 0|T{(r 1 , t1 ) (r 2 , t2 )}|0 ,
(8.63)
where the T-product has the same meaning as before, and (r, t) are the phonon
operators in coordinate space.
In the harmonic approximation, taking for free operators and making a Fourier
transform, we can obtain the Green function of free phonons D0 (q, ). Depending
on the normalization of phonon operators, there are two different forms of D0 used
in the literature:
With the normalization
q = bq + bq
(8.64)
D0 takes the form (see e.g. Schrieffer (1999))

20 (q)
1
1
= 2
,
D0 (q, ) =
0 (q) + i + 0 (q) i
02 (q) + i
(8.65)
where 0 (q) is the bare phonon spectrum.
Another often used normalization is

0 (q)
(bq + bq ) ;
(q)
=
2
149
(8.66)
then using the definition (8.63) but with ,

we obtain the phonon Green
function in the form
D 0 (q, ) =
02 (q)
;
2 02 (q) + i
(8.67)
see, e.g. Abrikosov et al. (1975).

Accordingly, there will be certain differences in the matrix element of the
electronphonon interaction in the diagram techniques which compensate for this
difference and make the physical results independent of the scheme used, as it
should be; but one should be aware of these different conventions.
The fact that it is 2 which enters into the expression for the phonon Green
function, in contrast to in the electron Green function (8.41), (8.54), is actually
connected with the fact that phonons are bosons, and in particular that the number
of phonons is not conserved, so that phonons can be created and annihilated
independently. Consequently, for example, for the electronphonon coupling the
results would be the same if, say, a phonon with momentum q is emitted or if a
phonon with momentum q is absorbed; these processes always enter together,
see equation (8.59). Therefore it is always the combination bq + bq that enters

the expressions (8.64), (8.66), (8.59), and in the Green function (8.65), (8.67) both
terms, with +0 and with 0 in the denominator ((q) = (q)!), enter on an
equal footing.
The Green function for free electrons is known, see (8.41). For interacting systems
one has to calculate it, usually in a certain approximation.
There exist several methods to do this. One is the method of the equations of
motion, similar to the one used in Chapter 6, equations (6.87)(6.92): one writes
p, )
down the equations of motion for the Heisenberg operators (r,
t) or (
entering the definition (8.31), (8.32) using the rule (8.18). Usually one obtains,
after commutation with the Hamiltonian, the Green functions of higher order (twoparticle Green functions, etc.). To solve this, in principle infinite, set of equations,
one has to truncate these equations, making certain decouplings dictated by some
physical arguments (such as the mean field decoupling used in going from (6.88)
to (6.89)). Usually this is an uncontrolled approximation (there is no small parameter here), although it may be physically quite sound. This method is described e.g.
150
by Zubarev (1960). (Yet another example of such an approach will be given later
in the treatment of the Hubbard model, see Section 12.4.)
Another, very widely used method is perturbation theory, treating the interaction
as weak. Here we can use the interaction representation, see (8.20), (8.21), and use
diagram techniques which permit us to represent separate terms of the perturbation
expansion by Feynman diagrams and which greatly simplifies the calculations.
The idea of the method is the following. The Green functions (8.31), (8.32) are

in the Heisenberg representation, i.e. according
defined using the operators ,
i Ht
H (t) = ei Ht (0)e
. For free, noninteracting electrons H = H0 , and
to (8.15),
H (t) and the corresponding Green function G0 . Now, suppose we
we know both
have an interacting system, H = H0 + H , and we want to treat the interaction H
H and G, and make an
as a perturbation. Thus we can keep H0 in the exponent in

expansion in H , e.g. writing in the lowest order
i H t i H0 t
H (t) = ei H0 t ei H t (0)e

e
'
(
i H0 t

i H0 t i H0 t
= e
(1 + i H t) ei H0 t ei H0 t (0)[e
e
](1 i H t)ei H0 t
I (t)(1 i HI t) ,
= (1 + i HI t)
(8.68)
etc. where we have inserted the unity operator [ei H0 t ei H0 t ] and used (8.20). (Note
that we have to take care of the order of operators H0 , H , because in general
they do not commute.) Keeping H0 in the exponent means that we can deal with
i H0 t
the operators ei H0 t (0)e

and with the corresponding known noninteracting
Green functions (G0 ) and can build our perturbation theory using these as the basis.
Different terms of the perturbation theory expansion can then be conveniently
depicted as different diagrams, n-th order terms containing n interaction lines, or
n interaction vertices.
i Ht
H = ei Ht e
Note also that what we did here, e.g. the transition (8.68) from
i H0 t i H0 t
e
to e
is nothing else but the transition to the interaction representation
described above. Indeed, by putting products ei H0 t ei H0 t = 1 in (8.68) we see
that in each term of the perturbation theory we have both -operators

and the

i H0 t
i H0 t
i H0 t i H0 t
(0)e
perturbation H in the form e
and e
He
, but these are
I (t), HI (t), cf. (8.20), (8.24).
exactly the operators in the interaction representation
The possibility to formulate conveniently perturbation theory in the interaction
representation was actually the main reason for introducing this representation.
The rigorous derivation of the corresponding rules is described in many books,
e.g. Abrikosov et al. (1975), Fetter and Walecka (2003), Mahan (2000), Mattuck
(1992) and Schrieffer (1999); see also (8.23)(8.30). For us here it is sufficient just
to formulate the corresponding rules, so as to be able to use them afterwards. These
rules are the following:

p ,
p,
p,
151
p,
p ,
p,
p ,
p p
k =k+ p p ,1
k,1
(a)
(b)
(c)
q= p p , =
q,
p ,
p,
(d)
(e)
Fig. 8.3
Each term of the perturbation theory expansion for a Green function is represented by a certain diagram. The contribution of each diagram, for an interaction
between electrons or an interaction with an external field, is calculated in the
following way:
1. Write down all topologically inequivalent diagrams with one external electron
line (one going in and one going out). Only connected diagrams are to be
included. Electrons are denoted by solid lines (Fig. 8.3(a)).
2. Associate with each electron line with momentum p and energy the bare
Green function iG0 ( p, ) (8.41) or (8.44).
3. The Coulomb interaction is denoted by a dashed line; it connects two electron
lines, Fig. 8.3(b). In each vertex the momentum and energy are conserved.
Associate with each interaction line the factor V ( p p ); for the Coulomb
interaction (8.3) the Fourier transform is
V ( p p ) =
4 e2
.
( p p )2
(8.69)
4. The interaction with the external potential v (8.2), e.g. that of an impurity, is
denoted by a cross , or a dotted line with a cross (Fig. 8.3(c)). Associate with
it the factor v ( p p ) (the Fourier transform of the potential v (x)). Note that
at this vertex the electron momentum p is not conserved (an impurity can take
extra momentum); for elastic scattering the energy is of course conserved.
5. Similarly, one can also include phonons (denoted by wavy lines, Fig. 8.3(d)):
one associates with each phonon line with (q, ) the phonon Green function
D0 (q, ) (8.65) or (8.67). In the electronphonon vertex, Fig. 8.3(e), one puts
the corresponding matrix element gq .
152
6. Multiply by the extra factor (i)n , where n is the number of internal interaction
lines, and by (1)l , where l is the number of closed electron loops in the diagram.
7. Multiply all these factors together and integrate over all intermediate momenta
and energies,
3
d p1 d1 d 3 p2 d2
,
(2 )4
(2 )4
and sum over spin directions at the internal lines.
These rules and examples of the simplest diagrams are presented in many
books (e.g. Abrikosov (1975), Mahan (2000), Schrieffer (1999), and in many
others). Thus, e.g. the lowest-order diagrams for the electron Green function for
the electronelectron interaction have the form
q
q
pp =0
p
p =p
pq
(8.70)
Note that this is actually a formula, an expression for the Green function
(the double line, or thick solid line denotes the full Green function, in contrast
to the thin solid line which represents the bare Green function G0 ). Thus
equation (8.70) may be rewritten as

d 4q
G(p) = G0 (p) + (2i)G0 (p) V (0) G0 (q)
G0 (p)
(2 )4

d 4q
G0 (p) .
(8.71)
+ iG0 (p) G0 (p q) V (q)
(2 )4
In the second term one minus sign comes from the existence of one loop; the
factor (i) comes from one interaction line, and the factor of 2 from summation
over spins. We have introduced here the shorthand notation p = ( p, ), using the
four-vector p.
The second term (diagram
) contains formally V (0) which for the
Coulomb interaction (8.69) is infinite. Actually this term describes the Coulomb
interaction of a given electron with the average density of all other electrons
this is the Hartree term of the HartreeFock approximation. In real systems the
condition of electroneutrality should be obeyed, i.e. there should exist an equal
positive charge density (ions, or structureless positive background jellium). It
can be incorporated in the external potential v in (8.1), and it will compensate the
contribution of this diagram in a homogeneous system. Therefore in the following
we will always omit this and similar diagrams. One should be aware, however, that
153
they may be important in inhomogeneous systems (e.g. in an atom, or at a surface,

or close to a contact of different materials).
As mentioned above, the second term in (8.70) is the Hartree, or densitydensity
, represents an
term of the mean field approximation. The last term,
exchange, or Fock contribution in the HartreeFock approximation.
8.7.1 Dyson equations, self-energy and polarization operators

The diagrams presented in equation (8.70) are only the lowest-order contributions
to the full Green function. One should in principle consider also higher order
contributions, the choice of which is dictated by the concrete situation. Often it is
necessary to perform even the summation of an infinite number of diagrams of a
certain class.
There exists a convenient way to carry out such a programme. It is called
the method of Dyson equations; it uses objects known as the self-energy and
polarization operators. Suppose we start from the lowest-order diagrams
=
(as explained above we can omit the Hartree diagram). At the next step we can
repeat this contribution yet another time, and so on, and we will get a series
=
(8.72)
or
G = G0 + G0
0 G0
+ G0
where we have denoted the contribution

0 ( p, )
p,
0 G0
0 G0
p,
+ ,
by
V (q, ) G0 ( p q, )
0 ( p, ),
(8.73)
so that
d 3 q d
.
(2 )4
(8.74)
One immediately sees that the sum (8.72) is a geometric series which can be easily
summed to give
G( p, ) =
G0
1
= 1
1 0 G0
G0
=
0
1
.
p 0 ( p, ) + i p
(8.75)
154
Another way to perform this summation is to notice that equation (8.72) can be
rewritten as
=
+
.
(8.76)
G = G0 + G0 0 G ;
The solution of this equation again gives the result (8.75). Equation (8.76) already
has the form of a Dyson equation. However, diagrams of the type (8.72) do not
exhaust all the possibilities. We can have, e.g. terms of the type
+
which can also be repeated,

or
which
would
give
the
,
+ ;
sum
+ , or
+
+
, etc. Each of these series can be summed up, giving a result similar to (8.75),
but with different s.
One can formally incorporate all such terms by introducing the object ( p, )
which is called the self-energy, or mass operator, and which is the sum of all
diagrams of the type
+
The term 0 (8.74) is the lowest-order contribution to the total self-energy.

One can express G through this new function as
=
G( p, ) = G0 ( p, ) + G0 ( p, ) ( p, ) G( p, ) ,
(8.77)
which can be formally solved, giving

G1 = G1
0
(8.78)
or
G( p, ) =
Here the self-energy
1
p ( p, ) + i p
(8.79)
( p, ) contains all the diagrams which cannot be cut
across one electron line. Thus, e.g. the diagrams

should not be included in
or
they are already taken
155
into account in the summation leading to (8.75) or to its generalization (8.79). By

comparing (8.75), (8.79) with (8.41), (8.44) one can immediately understand the
origin of the terms self-energy and mass operator: the contribution is added
to the bare energy of the electron, and, if the pole structure of the Green function (8.79) is preserved, it will modify the effective mass of the quasiparticle (see
below, (8.85)(8.91)). The equation (8.77) or (8.78), (8.79) is called the Dyson
equation. The Dyson equation (8.77) is not yet a closed equation for the Green
function: the self-energy should be calculated separately. Despite this, the use
of the Dyson equation proves to be very useful: the virtue of this approach is that it
permits one actually to sum up an infinite number of terms in perturbation theory
even by including only the lowest-order terms in the self-energy .
This treatment can also be generalized for the case of electronphonon interactions: one should only substitute everywhere instead of the electron interaction
with which we associate the phonon Green funcline the phonon line
tion (8.65) or (8.67), and at each electronphonon vertex
we put the
electronphonon coupling constant gq .
Analogously to the electron Green function and self-energy, one can introduce
similar objects also for the interaction lines and for phonons
=
(8.80)
or
=
(8.81)
Here
is the screened Coulomb interaction (which now becomes also
is the full (dressed) phonon
frequency dependent, or retarded), and
(q, ) is the so-called polarization
Green function. The bubble
operator. The simplest diagrams for it are
=
(8.82)
etc. (and the same with phonons). Again, it is an irreducible operator (in the
which can be
sense that it should not contain terms like
cut through one Coulomb (or phonon) line, because these diagrams are already
or
in
accounted for when we put the full Green functions
the right-hand side of (8.80), (8.81)).
The diagrams shown in (8.80), (8.81) are again equations which can be formally
solved similarly to equation (8.79). We do not present the corresponding results
here, but they will be discussed in detail in Chapter 9.
156
8.7.2 Effective mass of the electron excitation

From the general treatment given above we can obtain some important relations even without detailed calculation of particular diagrams. Let us start from
equation (8.79) expressing the one-electron Green function through the selfenergy . The bare Green function G0 ( p, ) (8.41) has a pole at the bare spectrum
p = p2 /2m . Suppose that there is a pole (renormalized electron spectrum)
p also in the interacting system, so that
G( p, )
1
.
p
(8.83)
From (8.79) the spectrum p should be a solution of the equation

p Re ( p, ) = 0 ,
or p p Re ( p, p ) = 0 .
(8.84)
We ignore for a while the imaginary part of , which in principle will determine
the finite lifetime of the excitation (assume that close to the pole Im < Re , or
that the lifetime is long enough).
Let us expand the Green function close to the pole p :
G( p, ) =
1
1
)
=
p ( p, )
p
( p, p ) +
p
1
( p, p )
As, according to (8.84), p +
p )
*
( p )
p
(8.85)
( p, p ) = p , we have (cf. (8.54))

1
G( p, ) =
1
p
p )
= p
Zp
.
p
(8.86)
Thus we see that because of the interaction the spectrum is renormalized, p

p , and there appears a factor Z p : the strength of the pole, the residue, is no longer
1 but is
Zp =
1
1
( p,)
=
(8.87)
We have thus managed to express the Z-factor, introduced phenomenologically

in (8.54), through the self-energy ( p, ), which gives in principle a way to
calculate it.
157
One can also obtain a useful expression for the change of the effective mass of
the electron due to interaction. From (8.84), assuming that
p2
,
2m
p =
(8.88)
with the bare spectrum being p = p2 /2m (actually this is the definition of the
effective mass m ), we obtain:

p
p
p
1
=
+
+
=
2m
( p2 )
( p2 )
( p2 ) p ( p2 )

p

1
2 +
+
(8.89)
=

2m 2m p

( p2 )
= p
2m
or, collecting different terms, we obtain

1

1
1+
1
=
,

m

m
p
(8.90)
m
=
m
1+
1
1
( p,)
Z p 1 +
p
(8.91)
This is a very important formula which connects the effective mass renormalization
m /m with the pole strength Z p . One can show that if the interaction leading to
mass renormalization is retarded (for instance, is carried out by low-energy excitations, e.g. is an electronphonon interaction with ph F ), then one can neglect
the momentum dependence of ( p, ),1 and we have m /m = 1/Z p , which is
necessarily 1 (as Z p 1). In particular, for the electronphonon interaction we
get (see below)
m = m(1 + ) ,
(8.92)
where is the dimensionless electronphonon coupling constant (8.61), entering

also, e.g. into the expression (8.62) for the superconducting Tc in ordinary superconductors, in which superconductivity is due to electronphonon interactions.
1
This is connected with the Migdal theorem which states that in typical metals, due to the existence of a small
parameter D /F 1, one can keep in the electronphonon self-energy only the simplest diagrams of the type
and ignore the so-called vertex corrections, e.g.
. Similarly, we can keep only the
simplest bubble diagrams (but in general with the full electron Green functions) in the polarization operator, i.e.
we can ignore the diagrams of the type
which are small if D /F 1.
158
Even more interesting is the situation in the so-called heavy fermion systems, see
below, Chapter 13. In this case experimentally m /m (and the coefficient in the
linear specific heat c = T which is proportional to m , cf. (7.25)), is extremely
large, 102 103 . One can also describe the renormalization of the effective mass
here as occurring due to interaction with very soft spin fluctuations, and the value
m /m is also directly related to the pole strength Z 1.
If, on the other hand, both and the p-dependence of ( p, ) are important
(e.g. for the Hubbard on-site Coulomb interaction), one cannot draw any general
conclusion about the mass renormalization m /m.
9
Using the techniques described in the previous chapter, we can in a unified way
discuss properties of the electron gas with Coulomb interaction and consider such
effects as optical response, screening, plasmons, etc. In many textbooks these
properties are obtained using a variety of methods. The virtue of the Green function
method is its universality and, I would say, not much simpler, but standardized form.
This method permits one to obtain all the properties mentioned above in the form
of one general expression, and it also gives the possibility to generalize the results
quite easily to the cases of low-dimensional (1d, 2d) systems, or to take into account
the details of the band structure of the material, etc. But more important is the fact
that it leads naturally to a number of special interesting consequences which would
be rather difficult to obtain with the usual classical methods. In this and in the
next two chapters I will demonstrate how to reproduce, using this method, the
familiar results such as Debye screening or the plasmon energy, but I will mostly
concentrate on less frequently discussed effects which are quite naturally obtained
using this technique.
9.1 Dielectric function, screening: random phase approximation
We start by studying the form of the effective electronelectron interaction in
metals. The ordinary Coulomb interaction V (q) = 4 e2 /q 2 is modified by the
reaction of the electronic system. The first, well-known effect is just the screening
of the electric charge. But there are other interesting effects as well.
One can describe the modification of the Coulomb interaction using the corresponding Dyson equation (8.80)
=
(9.1)
or
v (q, ) = V (q) + V (q) (q, ) v (q, ) .

159
(9.2)
160
Here (q, ) is the irreducible polarization operator (8.82), and v (q, ) is the
renormalized interaction which now becomes frequency dependent (this corresponds to retardation effects: the reaction of electrons has a certain characteristic
time-scale, which leads to the retardation of the effective interaction, in contrast
to the initial Coulomb interaction V (q) which is, in nonrelativistic theory, instantaneous or frequency independent). Note that sometimes the polarization operator
is defined with the opposite sign, e.g. our definition of (q, ) differs by a sign
from that in the book by Schrieffer (1999): (q, ) = PSchrieffer (q, ). Solving
equation (9.1) formally we obtain:
v (q, ) =
V (q)
V (q)
.
1 V (q) (q, )
(q, )
(9.3)
Here we have introduced the dielectric function

(q, ) = 1 V (q) (q, ) .
(9.4)
Taking the lowest approximation for the polarization operator

p+q
0 (q, ) =
= 2i
G0 (p + q) G0 (p)
d 4p
,
(2 )4
(9.5)
we obtain what is known as the random phase approximation (RPA) for the dielectric function and for the effective interaction. The Dyson equation (9.1) in this
approximation corresponds to the summation of an infinite set of diagrams of the
type
=
+
(9.6)
i.e. we dress the Coulomb line
by electronhole bubbles
One can show that this is a good approximation for high-density electron systems,
rs 1 (where rs is the dimensionless parameter characterizing the electron density,
see Chapter 7). The justification of this approximation is connected with the fact
that the Coulomb interaction V (q) is large for q 0, so that in the perturbation
theory (9.6) higher-order terms contain extra factors e2 but also (1/q 2 )2 , and we
have to sum all terms of this kind which gives (9.3) with = 0 given by (9.5),
i.e. the RPA result.
The RPA form for the dielectric response function, which corresponds to keeping
only the sum of bubble diagrams, may be presented in two equivalent forms:
9.1 Dielectric function and screening
161
One can say that the total interaction is

=
+
,
i.e.
v=V
(9.7)
+ V 0 v
(9.8)
from which we get the result (9.3) with 0 . Or we can write this also as
=
(9.9)
i.e.
v=V
,
+ V V
(9.10)
where
=

(9.11)
= 0 + 0 V
.

(9.12)
or
=
(Do not confuse
(9.11) with the irreducible polarization operator
we
=
defined by equation (8.82)! As is seen from (9.11), in
include diagrams which can be cut across one electron line, which are excluded
we
in the irreducible polarization operator (8.82). On the other hand, in
included only simple electronhole bubbles, ignoring vertex corrections included
in (8.82).)
From (9.11) we find
=

0
,
1 0 V
and then from (9.9) we obtain the same expression (9.3) for v :

V
V

V
0
)=V 1+
=
v = V (1 + V
=
.
1 0 V
1 0 V

(9.13)
(9.14)
The expression (9.13) describes the polarizability of the system (still in the RPA).
One can obtain a similar expression also for the magnetic susceptibility. We have
to consider the response of our system to a magnetic field, causing, e.g. a reversal
162
of spin directions. Then the response function is given by electronhole diagrams

with opposite spins
+ ,
(9.15)
denotes the effective electronelectron interaction (not necessarwhere

ily long-range Coulomb interaction). Thus, for a local interaction U ni ni (the
Hubbard interaction) one can easily sum these diagrams and obtain
(q, ) =
0 (q, )
,
1 + U 0 (q, )
(9.16)
where 0 (q, ) is proportional to the same expression (9.5) (the block repeated in
(9.15) is in fact the same product of electron and hole Green functions as in (9.5)),
0 (q, ) = 12 g 2 2B 0 (q, ) .
(9.17)
Note the change of sign in the denominator of (9.16) as compared to (9.13): it

comes from the fact that for densitydensity correlations described by 0 or the
closed loops in the diagrams give an extra factor 1, according to the general rules
of Chapter 8. Here, for the magnetic response, we have rather the so-called ladder
diagrams (9.15). Closed loops here are forbidden because of spin conservation:
electron and hole lines in (9.15) correspond to opposite spins, and these cannot
simply recombine. Nevertheless for a local (or q-independent) interaction the basic
expressions are similar. The expressions (9.16), (9.17) will be very useful later on
in the discussion of magnetic instabilities of electron systems.
The calculation of the polarization operator 0 (9.5) is a straightforward but
tedious task. The first step is relatively easy. We write 0 (9.5) as
0 (q, )
2i
=
(2 )4
d 3 p d
)

*)

*
+ ( p+q)+i sign ( p+q) ( p)+i sign ( p)
(9.18)
and integrate over using contour integration, closing the contour in the upper
half-plane. The result has the form

2
n( p) n( p + q)
0 (q, ) =
d3 p
,
(9.19)
3
(2 )
+ ( p + q) ( p) + i
where n( p) and n( p + q) are the usual Fermi functions, n( p) = 1 for | p| < pF
and = 0 for | p| > pF . Thus the integration region is the hatched area in Fig. 9.1,
163
Fig. 9.1
i.e. it lies inside the sphere | p| = pF , but outside | p + q| = pF , or vice versa.

The origin of the factor n( p) n( p + q) in the numerator of equation (9.19)
mathematically follows from the position of the poles in the Green functions
entering (9.18) (here the imaginary parts in the denominators of (9.18) play a
role!). Physically this is easy to understand: the polarization operator (9.5) is given
by the electronhole bubble, and if one particle, e.g. with momentum p, is a hole,
i.e. | p| < pF , then the second one, with momentum p + q, should be an electron,
i.e. | p + q| > pF . The factor n( p) n( p + q) reflects just this fact.
Further integration is straightforward but rather elaborate. The result is known
as the Lindhardt function. I will present here only the important limiting cases; for
the full expression see, e.g. Schrieffer (1999).
The static dielectric function (q, 0) is given by the expression
+
,1/3

rs
4
4me2 pF
q
=1+
u
u(x)
(q, 0) = 1 0 V = 1 +
2
4
q
2pF
9
x2
2

pF
q
,
(9.20)
= 1 + 0.66 rs
u
q
2pF
where x = q/2pF and

(1 x 2 ) 1 + x
1
1+
ln
.
u(x) =
2
2x
1 x
(9.21)
From this expression we can easily obtain the Debye screening length rD = D1 :

(q, 0)
q0
=1+
D2
,
q2
D2 = 4 e2 (F ) =
6 ne2
,
F
(9.22)
(see (7.13)), so that the screened Coulomb interaction (9.3), (9.14) would have the
usual form
v (r) =
e2 D r
.
e
r
(9.23)
164

1(q,0)
2pF
Fig. 9.2
The screening length D1 coincides with that obtained in the usual semiclassical
approximation (ThomasFermi screening). (Note also the convenient result:

0 (q, 0)
= (F ) ,
(9.24)
q0
where (F ) is the density of states at the Fermi level (7.13).)

However, this is not really the full story. If we look at the detailed behaviour of
(q, 0) for arbitrary q, we see from equations (9.20), (9.21) that there is a singularity
in (q, 0) at q 2pF : 1 /q at this point, see Fig. 9.2. Mathematically
this singularity is connected with the fact that starting from q = 2pF the region of
integration in (9.19) no longer changes, see Fig. 9.1. Correspondingly, when we
study how the screened potential v (q, 0) (9.3) behaves in real space, v (r r ),
we have to make a Fourier transform of v (q) = 4 e2 /q 2 (q, 0) with (q, 0) given
by (9.20), (9.21), and the presence of the singularity of (q, 0) at q = 2pF modifies
the behaviour of the screened Coulomb potential.
Usually the asymptotic
of Fourier transforms (the behaviour of inte# bbehaviour
iqx
grals of the type f (x) = a e f (q) dq at large x) is determined by the behaviour

of f (q) at the limits of integration (a, b) (here at q 0), but also by the special
points (singularities in f (q) and its derivatives) inside the interval (a, b), if they
exist. Thus, e.g. if f (q) itself at q = q0 , for example, if f (q) contains a
-function, c(q q0 ), then f(x) would evidently have a contribution ceiq0 x , or it
will be an oscillating function not decaying at large x at all. A similar situation
exists in our case: because of the singularity of 1/q at q = 2pF the screened
Coulomb interaction v (r) is not simply
e2 D r
e
(9.25)
r
as one would expect classically and as is usually taken, but there exists also a
long-range oscillating term

cos(2pF r + )
v (r)
,
(9.26)
r
r3
where is a certain phase.
v (r) =
165
Correspondingly, the screening charge around a charged impurity in a metal

would also follow the law (9.26) at large distances. This is known as Friedel
oscillations, and they occur due to the sharp cut-off of the Fermi distribution
function n( p) at p = pF . These oscillations are a real effect; they were observed
experimentally.
As at finite temperatures the sharp Fermi surface is somewhat washed out, these
oscillations will start to decay with increasing temperature. Impurities in the metal
also lead to a similar effect: in the presence of impurities the Friedel oscillations
start to decay as er/ l , where l is the mean free path.
This phenomenon is even more important not for charge but for magnetic interaction in metals. It turns out that in the first approximation (RPA) practically the
same formulae describe the screening of a localized magnetic moment inserted
into the metal (i.e. the magnetic polarization of conduction electrons by a magnetic
impurity). The magnetic susceptibility (q, ) is also given in this approximation
by the summation of the same bubble-like (in fact ladder) series (9.15) (see,
however, the discussion of the Kondo effect below, Chapter 13). The result has the
form (in the static case = 0):

q
3g 2 2B n
(9.27)
u
(q, = 0) =
8F
2pF
where g is the g-factor of the electron (usually g = 2), B = e
h/2mc is the Bohr
magneton, and u(x) is the same function (9.21) as in the expression for the dielectric
function. Correspondingly, the spin density s(r) at distance r from the impurity
spin Sa ,

J
(q)eiqr Sa ,
(9.28)
s(r) = 2 2
g B V q
will behave as (9.26) (here J is the contact exchange interaction between an
impurity and a conduction electron, J Sa s). If we now put another spin Sb
(another magnetic impurity) at some point, it will have a similar local exchange
interaction with the conduction electrons polarized by the first impurity Sa . As a
result there will be an effective exchange interaction between Sa and Sb mediated
by the (oscillating) spin polarization of conduction electrons. This interaction is
called the RudermanKittelKasuyaYosida or simply RKKY interaction, and it
has the form
J2
J 2 cos(2pF r + )
HRKKY 2 2
(q) eiqr Sa Sb
.
(9.29)
F
r3
g B V q
This oscillating long-range interaction has a lot of consequences. It gives, e.g. a
broadening of NMR lines. A special form of interaction (9.29) with alternating
166
0(q, 0)
0(q,0)
2 pF
(a)
0(q, 0)
2pF
2pF
(b)
(c)
Fig. 9.3
positive (antiferromagnetic) and negative (ferromagnetic) parts gives rise to often

complicated magnetic structures in many metallic magnets, in particular to spiral
or helicoidal structures in rare earth metals. It is also in most cases responsible for
the formation of the so-called spin-glass phase: when we put random magnetic
impurities in a metal, due to the oscillating form of the RKKY interaction (9.29)
the effective exchange interaction, or the effective (exchange) field produced at the
position of a given impurity by all the others, has random orientation. This leads to
a specific state: at low temperatures all spins are frozen, but in random orientations,
so that there is neither total magnetization (no ferromagnetic moment), nor any
other spatial order. However, the susceptibility in this case usually has an anomaly
at the spin freezing temperature, which is different for zero-field and field-cooled
measurements. This anomaly also depends on the measuring frequency. These are
the fingerprints which allow one to identify the spin glass state.

The same feature a singularity in the response functions becomes even more
pronounced in low-dimensional systems or for special shapes of Fermi surfaces.
Both (q, ) and (q, ) are called response functions, because they describe the
response of our system to external perturbations external electric or magnetic
fields, charged or magnetic impurities, etc. (By the way, one can easily understand
why we have summation over q in the RKKY interaction: the response to an external
field, e.g. to the electric field E q in optics, is described by the corresponding
Fourier component (q, 0); however when we are dealing with a perturbation
localized in space, as in the case of impurities, we have to integrate over all
harmonics.)
Whereas in the usual 3d systems the polarization operator 0 (q, 0) has a logarithmic singularity in the derivative 0 /q|q=2pF , see equations (9.20), (9.21)
and Fig. 9.3(a), the singularities in 2d and in 1d cases have the form shown in
Figs. 9.3(b), (c) (note that with our convention (9.3), 0 (q, 0) is negative). Thus,
167
for example, in the 1d case the polarization operator itself has a logarithmic
divergence,
0 (q, 0) ln|q 2pF | .
(9.30)
This divergence has important consequences, in particular it leads to instabilities

of one-dimensional metals with respect to the formation of superstructures with
wavevector Q = 2pF or period l = 2 a/Q. There will be a special chapter devoted
to the detailed discussion of this and other instabilities (Chapter 11); here, anticipating this material, we will only point out that, e.g. the renormalization of the
phonon spectrum is determined in the first approximation by the same diagrams
entering the expression for (q, ),
=
(9.31)
(cf. (9.7)), i.e.

D(q, ) = D0 + D0 g 2 0 D ,
where g is the electronphonon coupling constant (vertex
D(q, ) =
1
D01 (q, ) g 2 0 (q, )
(9.32)
). This gives
02 (q)
2 02 (q) g 2 02 (q) 0 (q, )
(9.33)
(with the normalization (8.67)). Thus, the new renormalized phonon spectrum (pole
of the Green function (9.33)) is given by the equation
)
*
2 (q) = 02 (q) 1 + g 2 0 (q, ) ,
(9.34)
so that if 0 (q, ) (as it does at q = 2pF in the 1d case), the phonon
frequency becomes imaginary (2 < 0) which means absolute instability of the
system (the phonon mode with q = 2pF would accelerate, leading to macroscopic occupation of the corresponding mode, i.e. to a real deformation).1 Such
instability in one-dimensional systems is called Peierls instability (see also
Section 11.1 below). The corresponding Peierls distortion is often observed in
real one-dimensional materials, e.g. in some long organic molecules or polymers,
1
By the way, we see here why in discussing possible instabilities of an anharmonic lattice and melting in
Section 4.4.2, I said without proof that the actual equation for phonon softening contains not the phonon
frequency , but 2 : the phonon Green function D(q, ) and the renormalized spectrum always contain 2 .
Equation (9.34) is written for the phonon renormalization due to electronphonon coupling, but a similar
expression can also be obtained for the case of the anharmonic phononphonon interaction of Chapter 4.
168

4 5
,p
(p
2
1
Fig. 9.4
the best-known example being polyacetylene Cn Hn . Note also that the deformation
with the wavevector Q = 2pF gives new periodicity with a new Brillouin zone
and, according to the usual picture of nearly free electrons in a periodic potential,
it produces a gap exactly at the position of our initial Fermi surface (pF , +pF ),
so that the quasi-1d material after such transition acquires a gap and becomes an
insulator.
The anomaly of the phonon spectrum (9.34) in the 1d case is really extreme;
but even if the anomaly is not so strong (in 2d or 3d cases, for instance), still the
singularity of 0 at q = 2pF is reflected in the phonon spectrum (q) at the same
wavevector. This anomaly in the phonon spectrum is called a Kohn anomaly. If
such an anomaly is strongly enhanced, as happens in low-dimensional materials,
one speaks of a giant Kohn anomaly. As is clear from the preceding discussion,
this Kohn anomaly is closely related to Friedel oscillations and to the oscillating
long-range character of forces in metals, see (9.26), (9.29).
Up to now when considering 1d, 2d or 3d systems we always had in mind the
situation with isotropic electron spectrum, e.g. ( p) = p 2 /2m. However, in real
crystalline materials the spectrum may be anisotropic, and the corresponding Fermi
surface need not be a sphere (or a circle in the 2d case); it can and often does have
a very complicated shape. In particular there may be cases when the whole or
parts of the Fermi surface are flat. Such is, for example, the Fermi surface of the
two-dimensional square lattice in the tight-binding approximation for a half-filled
band (one electron per site). In this case the spectrum is (we put here the lattice
constant a = 1)
( p) = 2t(cos px + cos py ) ,
(9.35)
and the Fermi surfaces for different band fillings have the form shown in Fig. 9.4;
the numbers 1, . . . , 5 correspond to increasing electron concentration, such that
the case of the half-filled band (one electron per site, n = 1) is represented by the
square Fermi surface 3. One can check that in this case the polarization operator
0 (q, 0) will behave for q (, ) exactly as in the one-dimensional case:
0 (q, 0) ln |q Q| ,
Q = (, ) .
(9.36)
9.3 Frequency-dependent dielectric function
169
Fig. 9.5
(This is connected with the fact that for this Fermi surface the region of integration in (9.19) (cf. Fig. 9.1) will change very rapidly as q (, ), as in the
1d case.) Consequently, many properties of such systems will resemble those of
one-dimensional systems, in particular they will be very susceptible to different
instabilities, there may exist giant Kohn anomalies in them, etc.
More generally one needs for this not necessarily flat Fermi surfaces (or flat
parts of Fermi surfaces) but what is known as nesting. Nested Fermi surfaces are
such that different parts of them will coincide when shifted by a certain wavevector.
For example, the Fermi surface for the spectrum (9.35) for n = 1 is a nested Fermi
surface, with the nesting vector Q = (, ). But in general we can have nesting
even for curved Fermi surfaces, for example, such as the one shown in Fig. 9.5.
Mathematically what we need is that the energy spectrum satisfies the condition
( p + Q) = ( p)
(9.37)
where the energies are counted from the chemical potential. Thus the spectrum (9.35) definitely obeys this condition at half-filling ( = 0), with Q = (, ).
Similarly, the spectrum of a three-dimensional cubic lattice in the tight-binding
approximation,
( p) = 2t(cos px + cos py + cos pz ) ,
(9.38)
has the same property, with Q = (, , ), and its Fermi surface for n = 1 is
also nested, although it does not actually contain flat parts. In this sense a onedimensional metal always has a nested Fermi surface: the Fermi surface consists
simply of two Fermi points which of course will coincide when shifted by the
nesting wavevector 2pF . The property of nesting and respective divergence of the
polarization operator will play a crucial role in discussions of different instabilities
in metals, see Chapter 11.
9.3 Frequency-dependent dielectric function; dynamic effects

Let us now turn to the opposite limit: consider the polarization operator and dielectric function at small q as functions of . This limit is especially important in
170
optics, as the wavevector of light (photons) is small (it can usually be taken as
zero), and we are mostly interested in the optical spectrum, i.e. in the frequency
dependence.
For large (|| (q 2 + 2qpF )/2m) one can get from the general expression
for , obtained from (9.19), that
Re (q, ) = 1
2
pl
(9.39)
where pl is the usual plasma frequency

2
pl
=
4 ne2
.
m
(9.40)
(In the case of finite we have to be careful: the polarization operator 0 (q, )
and (q, ) have imaginary parts.)
Actually the plasma frequency is an eigenfrequency of the density oscillations
of electrons on a positive background. This is how it is usually obtained in most
textbooks; it is essentially a classical notion. However, one can also give it another
interpretation. We see from (9.39) that the plasma frequency is a pole of the
corresponding response function or of the effective interaction v (q, ) (9.3).
Let us consider the process of scattering of an electron and a hole. During this
process they can recombine, emitting a photon, which in turn can be absorbed,
creating again an electron and a hole. One can represent this process by the
diagram
Seemingly another but actually equivalent process is the usual scattering, when an
electron emits a photon, and a hole absorbs it:
Topologically these diagrams are equivalent. They represent the first (in the interaction) term in the perturbation expansion of the so-called two-particle Green
171
function
The next most important terms are our old bubbles:

=
+ .
(9.41)
Or, using the summation (9.6), we can write this contribution (RPA contribution)
as
p1
p3
q= p1 p2 = p3 p4
p2
p4
i.e. we put inside the full renormalized interaction v (q, ) = v ( p1 p2 , 1 2 ).

Consequently the poles of v (q, ) are simultaneously poles of the two-particle
Green function in the corresponding total momentum of the incoming electron
( p1 ) and hole ( p2 ). They describe actually something like bound (or antibound)
states of an electron and a hole.
The diagrams (9.41) are of course not all possible contributions to the twoparticle Green functions; there may be, for example, diagrams of the type
or
etc. Thus the full solution of the two-particle problem in the presence of an interaction is a very complicated problem which can be solved only in a few simple cases.
However, already the simple RPA approximation gives a lot of information and
provides if not the full description, then at least a very useful starting point (and
often it is really sufficient for many purposes).
The plasmon pole does not exhaust all eigenstates of a two-particle (two excitation) system, or all zeros of (q, ). Actually from the expression (9.19) we see that
for a fixed q there are also eigenstates (poles, or zeros of the denominator, which
172
0(q,0)
1
V(q)
w pl
Fig. 9.6
in the limit of an infinite system come close together and form mathematically
a cut in the complex -plane) at all ( p) ( p + q) for different p. In terms of
the two-particle system these states correspond to scattering states of an electron
and a hole, having a continuous spectrum, whereas a plasmon pole is a bound (or
rather antibound) state. The situation here and the mathematical treatment is rather
similar to that of impurity states in Section 6.5. One can write the expression (9.19)
for 0 (q, ) in the discrete case as

1
1
,
0 (q, ) =
( p + q) + ( p) + i + ( p + q) ( p) + i
| p|<p
F
| p+q|>pF
(9.42)
and then the poles of (q, ) (9.4) are the solutions of the equation
1 = V (q) 0 (q, )
(9.43)
which can be analysed graphically, Fig. 9.6, as we did for equation (6.132), cf.
Fig. 6.47: all states with 0 < < max = pF q/m + q 2 /2m describe scattering
states and form for an infinite system a continuum spectrum (wavy line), and the
special solution lying above the continuum is the plasmon. One can show that
the plasmon pole exists (i.e. does not decay) only for a certain limited range of q.
The total spectrum has the form shown in Fig. 9.7: plasmons exist as real excitations
only for q < q , and for q > q plasmons merge with the continuum (the hatched
area in Fig. 9.7) and decay into independent particlehole excitations.
It is also interesting and instructive (and important practically) to consider what
would become of plasmons in low-dimensional (e.g. quasi-two-dimensional) materials, such as layered systems or thin films. It turns out that the general treatment
remains qualitatively the same, but the plasmon spectrum starts from zero, Fig. 9.8;
in a pure 2d system it is pl (q) q, and in quasi-2d systems, depending on
173
2qp F + q2
2m
w pl
q*
Fig. 9.7
w
w pl
(q)
q
Fig. 9.8
q
Fig. 9.9
the perpendicular component of the total wavevector q = (q , q ), it occupies the

whole region, Fig. 9.9 (the curves above the hatched electronhole continuum represent plasmon dispersion for different values of q ). And the most straightforward
and standard way to study this problem is to use the Green functions and diagram
technique described above.
One can extend this approach and use the results obtained in the RPA to discuss
the modifications of the electron spectrum brought about by the interaction. This
spectrum is given by the poles of the one-electron Green function (8.79) and can
be expressed through the self-energy or mass operator , see (8.85)(8.91). We
174
can calculate
in the RPA using the diagrams

q,
+
or
( p, ) = i
+ =
v (q, ) G0 ( p q, )
pq,
d 3 q d
.
(2 )4
(9.44)
(9.45)
The calculation finally gives the value of the effective mass

(
1 '
1
1 0.083 rs (ln rs + 0.203)
=
(9.46)
m
m
which shows that the Coulomb interaction leads to a certain increase of the effective
mass.
One can also calculate in this limit the total energy of the electron system
(M. Gell-Mann and K. Brueckner); the result is
%
&
2.21 0.916
E0
=
+ 0.062 ln(rs ) 0.096 Ry

(9.47)
N
rs2
rs
where 1 Ry = 13.6 eV. The first term in equation (9.47) represents the kinetic
energy of electrons (it is equal to 35 F , with F given by (7.9)), and the second is
the exchange part of the average Coulomb energy (the direct part of the electron
electron interaction is cancelled by interaction with the positive background). The
first two terms together represent the HartreeFock energy of the electron gas,
and the remaining terms give the so-called correlation energy (by definition the
correlation energy is everything beyond HartreeFock).
These results, and the RPA in general, are valid for dense systems, rs 1. In the
opposite limit of low-density electron systems the situation is changed drastically,
and perturbation theory in the interaction is no longer valid. In this case we enter
the field of strong electron correlations, which will be extensively discussed in the
last section of Chapter 11 and in Chapter 12.
10
Fermi-liquid theory and its possible generalizations
As we have seen in the previous chapters, interactions, if they are not too strong,
preserve many features of the electronic system which are present in the noninteracting case (Fermi gas, Chapter 7). In general, however, the interactions are
not at all small: for instance in typical metals, rs 23, and not rs 1 as was
implicitly assumed in Chapter 9 and which was actually the condition for the
applicability of perturbation theory used there. Nevertheless we know that the
description of normal metals using the concepts developed for the free Fermi gas
or Fermi systems with weak interactions (such as Drude theory, for example) is very
successful.
An explanation of the success of the conventional theory of metals, and the
generalization of the corresponding description to a more general situation, was
given by Landau in his theory of Fermi liquids. This theory is very important
conceptually, although in the usual metals there are only few special effects which
indeed require this treatment. However, there exist also systems (3 He, or rare earth
systems with mixed valence and heavy fermions) for which this approach is really
vital. Also the emerging new field of non-Fermi-liquid metallic systems requires
first an understanding of what is the normal Fermi liquid.

The main assumptions of the Landau Fermi liquid theory are completely in line with
our general approach. Landau postulated that, given the ground state of a metallic
system, weakly excited states can be described in terms of elementary excitations,
or quasiparticles, which are very similar to the usual electrons and holes in a Fermi
gas, despite strong interactions. In particular, these quasiparticles are fermions
with spin 12 and with charge e, which can be characterized by momentum p and
energy ( p). A crucial assumption is that there exists a well-defined Fermi surface,
175
176
Fermi-liquid theory and its generalizations
with the radius pF satisfying the same relation

n=
N
pF3
=
3
V
3 2
h
(10.1)
as for the free Fermi gas, cf. (7.2). Thus, the value of the Fermi momentum pF
and the volume of the Fermi surface are determined by the total particle density
N/V and do not depend on the interaction. This statement can actually be proven
in every order of perturbation theory, using the Green function technique, and is
known as the Luttinger theorem. Despite its apparent simplicity, this is a very strong
statement, which leads to a number of experimental consequences and which can be
really checked experimentally, e.g. by angular-resolved photoemission. One should
say right away that it is not the only possible state of an electronic system. Thus, we
know for sure that there exists an alternative state the superconducting state in
which the Luttinger theorem formally does not hold and there is no Fermi surface
left (there is an energy gap at the position of the former Fermi surface). There are
also nowadays a lot of discussions about possible breakdown of the Fermi liquid
description and of non-Fermi-liquid states in cases of strong electron correlations
which may (but also may not) be accompanied by a breakdown of the Luttinger
theorem. And finally the whole big field of strongly correlated electrons, discussed
in Chapter 12 below, is a bona fide example of the inapplicability of Fermi liquid
theory. These questions will be discussed later on; now however, having made
these remarks, we continue to discuss positive aspects and the formulation of the
Landau theory of Fermi liquids, in cases when this description is indeed valid.
What are the physical reasons which permit one in principle to retain many
essential features of free fermions in a strongly interacting system? The main factor
is the very fact that electrons are fermions, i.e. the Pauli principle. One can show that
because of that the Fermi surface remains sharp, and quasiparticles are well defined
close to the Fermi surface. If we consider one electron with momentum p1 > pF
excited above the Fermi surface, one can show that despite interactions with other
electrons its lifetime becomes infinite at the very Fermi surface, when p1 pF .
Indeed the simplest process which can lead to a finite lifetime is the excitation of
another electron from below F to above it due to the electronelectron interaction,
i.e. the creation of an electronhole pair, Fig. 10.1. (We use here a simplified form
of diagram, contracting interaction lines to a point; this is rather convenient in
the case of a four-fermion interaction.) In this process we have as usual p1 + p2 =
p3 + p4 , and 1 + 2 = 3 + 4 (here | p1 |, | p3 |, | p4 | pF , | p2 | pF ; we also
took 1 , 3 , 4 F , 2 F ).
From these conservation laws it is clear that when | p1 | pF , all other particles
should also have | p2 |, | p3 |, | p4 | pF (this is especially easy to see if we count all
177
p2
p1
p3
p4
Fig. 10.1
energies from F : then 1 , 3 , 4 > 0, 2 < 0, and the condition 1 + 2 = 3 + 4

implies that when 1 +0, all other should 0).
When p1 is slightly larger than pF , all other momenta | p | pF should be
of the same order as | p1 | pF . Then by the golden rule the probability of the
process shown in Fig. 10.1, i.e. the probability of the decay of an electron (which
determines its lifetime ), is proportional to

1
(10.2)
W = (1 + 2 3 4 ) d 3 p2 d 3 p3
( p1 is given, and p4 = p1 + p2 p3 , thus there remain two independent integrations). As both p2 and p3 lie close to pF , (| p | pF ) (p1 pF ), this integral is
(| p1 | pF )2 , i.e. the inverse lifetime of the electron with momentum p goes to
zero as p pF quadratically in (p pF ):
p1 (p pF )2 .
(10.3)
(Here and below p = | p|.) This is essentially the phase space argument: because
of the Pauli principle the scattering of weakly excited electrons is confined to
the vicinity of the Fermi surface, which gives (10.3). In the language of Green
functions the condition (10.3) means that the pole = (p) of the one-electron
Green function is well defined when p pF : its real part Re (p) vF (p pF )
is much larger than its imaginary part Im (p) 1 (p pF )2 . In other words
this means that the quasiparticles (quasi-electrons and quasi-holes) are well defined
at the Fermi surface and in its vicinity. This does not imply, however, that they
are well-defined objects far away from the Fermi surface: in general they are
not, and the structure of the spectrum far from the Fermi surface often does not
have the character of quasiparticles. However, at low energies and at low enough
temperatures it is sufficient to consider only excitations close to F , and for these
purposes the Fermi-liquid description usually works fine.
As to the finite temperature, we have seen in Chapter 7 that the Fermi-distribution
function n(p) is spread over the region T around F . This gives that, besides
the factor (p pF )2 in 1 (10.3), there may also be contributions of thermally
excited quasiparticles. As we have seen in (7.23), (7.24), actually the corrections to
178
the energy and to other properties due to such thermally excited quasiparticles are
T 2 (F ) T 2 /F . Consequently, combining this with (10.3) (with the natural
scale F in this term too) we get the estimate

( F )2 T 2
1
=a
,
(10.4)
+
F
F
where the constant a is of order 1.
Already from this simple treatment, which actually used only very general
arguments such as the Pauli principle and energy and momentum conservation
laws, we can make several important conclusions. First, as the low-temperature
excited states have nearly the same nature as ordinary electrons, such quantities as
specific heat and entropy will be given by the usual relations (7.25), (7.26), with
the same pF , with the only difference being the substitution of the effective mass
m instead of the bare mass m:
c = T ,
(10.5)
(cf. e.g. (9.46)). In general this mass renormalization can be substantial, and sometimes even extremely large. Thus in heavy-fermion compounds such as CeCu6 ,
CeAl3 , UBe13 , m 103 m, i.e. the effective Fermi energy F = pF2 /2m is 103
times smaller than the typical values for conventional metals like Al; whereas in
these simple cases F 110 eV, i.e. 104 105 K, in heavy-fermion compounds
the effective Fermi energy is F 103 eV 10 K.
From the expression for the electron lifetime (10.4) one can also make interesting
conclusions about some transport properties of the system, in particular electrical
resistivity. From the usual Drude expression for the conductivity
=
ne2
m
(10.6)
one gets, using 1 T 2 /F and also using (10.1), that the resistivity behaves as
2
T
T 2m
1
1
R=
2
.
(10.7)
3 2
e pF F
F pF e
One can further transform this expression using the fact that pF 1/rs (see (7.6)),
2
h /me2 . Then
so that typically pF 1/a0 , where a0 is the Bohr radius, a0 =
in (10.7) e2 pF e2 /a0 ( Rydberg). In general the functional dependence of the
resistivity due to electronelectron scattering is
R = AT 2 ,
A 1/F2 .
(10.8)
This is an important result; sometimes it is called the Baber law. In ordinary

metals F is very large, and the contribution of electronelectron scattering is
179
masked by other, much stronger contributions, such as electronphonon scattering,

which typically behaves at low temperatures as eph T 5 , or impurity scattering, giving residual resistivity. The T 2 -dependence of resistivity becomes noticeable in transition metals where the 3d band is narrower, and it is very clearly
seen in heavy-fermion compounds. The scaling between the coefficient A in the
resistivity (10.8) with the effective mass or effective Fermi temperature is also
confirmed there experimentally: as F = pF2 /2m , A is proportional to m2 , or
to 2 (10.5).
Up to now we have presented some arguments which helped to justify the
applicability for interacting electrons of ordinary relations valid for free fermions.
However, one can substantially deepen the description of the system, making the
next step and taking into account the fact that the Landau quasiparticles are really
not free fermions (with a different mass), but are still strongly interacting objects.
It is this next step which permits one to get new results and improve the description
considerably. We will describe it only schematically, and will qualitatively discuss
some of the consequences; one can find more detailed discussions, e.g. in Abrikosov
et al. (1975) and Baym and Pethick (1991). Suppose we somehow changed the
distribution of quasiparticles, n( p) = n( p) + n( p). As quasiparticles interact with
each other, the energy of each of them depends on the distribution function of the
others. Taking into account only the first, linear term of this dependence, we can
write:

f ( p, p ) n( p ) .
(10.9)
( p) = 0 ( p) +
p
This expression, which looks extremely (and deceptively) simple, actually permits one to treat a lot of effects using only the most general properties, such as
Galilean invariance, symmetry, conditions of stability of the system, etc. without
specifying the detailed nature of the interaction. The function f ( p, p ) introduced
phenomenologically in (10.9) is called the Landau interaction function; one can
show that it is connected with the scattering amplitude of electrons.
In general the quasiparticle energy, the electron distribution function and the
interaction function f may depend also on spin indices. For a nonmagnetic system
one can decompose f ( p, ; p , ) into two terms:
f ( p, ; p , ) = ( p, p )1 + ( p, p )( )
(10.10)
where 1 is a unit matrix in spin space and , are the Pauli matrices. As the
whole theory is valid close to the Fermi surface, one can also put | p| = | p | pF
and keep only the dependence on the angle between p and p . Then one can
use a standard expansion in terms of spherical harmonics (Legendre polynomials
180
Pn (cos )), writing

3
( p, p ) =
h
h
2
2
A()
=
[A0 + A1 P1 (cos ) + ]
m pF
m pF
( p, p ) =
h
h
2
2
B()
=
[B0 + B1 P1 (cos ) + ] .
m pF
m pF
(10.11)
(10.12)
(Here we have introduced the dimensionless interactions An , Bn , normalizing the

3
h (7.13).)
interaction with the use of the density of states (F ) = m pF / 2
The coefficients An , Bn are in general unknown; they are treated as phenomenological parameters, which are usually determined from experiment. The virtue of
the approach described above is that just a few parameters are usually needed to
explain a lot of experimental data. For electrons in metals, when we more or less
know and can describe important interactions, this approach is, maybe, less significant, but it is really necessary and extremely useful for strongly interacting Fermi
systems such as, e.g. 3 He and nuclear matter.
There are several slightly different ways to introduce the Landau parameters.
Often one uses a different notation instead of (10.10)(10.12):
f ( p, ; p , ) = f s + ( )f a ,
f s,a () =
2 3
h
m pF
Fls,a Pl (cos ) ,
(10.13)
(10.14)
l=0
where s, a stands for symmetric and antisymmetric. Sometimes the expansion

is also written with different normalization (note the factors (2l + 1) in the sum)
f
s,a
3
h
2
() =
(2l + 1)Fls,a Pl (cos ) .
m pF l=0
(10.15)
In a one-component Galilean-invariant system such as 3 He, the effective mass is

expressed as
m
Fs
A1
=1+ 1 =1+
= 1 + F1s .
m
3
3
(10.16)
(Note that in crystals, if the interaction is predominantly local but retarded, such as
the interaction with phonons, the expression for the effective mass may be different
and may contain not F1s but F0s , cf. (8.91), (8.92). In terms of Fermi-liquid theory
/ gives the terms F0s , and /k contains F1s .)
One can also get expressions for some other properties of Fermi liquids. They all
look similar to the corresponding formulae for an ordinary Fermi gas, but contain
the effective mass m (10.16) instead of m, and are renormalized by the respective
181
Landau parameters. Thus, the specific heat is

m
c0 ,
m
(10.17)
m /m
0 ,
1 + F0a
(10.18)
m /m
0 ,
1 + F0s
(10.19)
c=
magnetic susceptibility
compressibility
etc.
Already from these expressions one can draw certain general conclusions. Thus,
from (10.18), (10.19) one gets the criteria of stability of a homogeneous Fermi
liquid without any extra ordering (the so-called Pomeranchuk criteria1 ):
1 + F0a > 0 ,
1 + F0s > 0 .
(10.20)
Indeed, if, e.g. 1 + F0s < 0, this would give negative compressibility, which means
an instability of the system and will lead to a certain phase transition (e.g. structural
phase transition in metals). One can show that this is actually a general requirement:
the system remains stable if
1+
1
Fls,a = 1 + Fls,a > 0
2l + 1
(10.21)
for all l. For instance, for l = 1 the condition (10.21) guarantees that the effective
mass (10.16) is positive. Systems close to magnetic instability have 1 + F0a 1.
One can experimentally measure this parameter by comparing susceptibility and
specific heat: the Wilson ratio (7.31) becomes in this case
RW =
1
2 T
=
1.
2
1 + F0a
3B c
(10.22)
More interesting are the consequences of quasiparticle interactions, described by

Landau parameters, for collective modes and for transport properties. Thus, e.g.
by considering the nonequilibrium states one can show that there can exist in a
Fermi liquid (with short-range interaction) a collective mode which is called zero
sound. Usually there exists in liquids a sound mode in which the equilibrium is
restored due to collisions this is the ordinary hydrodynamic sound which exists for
1
Very often one speaks of Pomeranchuk instability in a narrow sense, as in the magnetic instability of the normal
Fermi liquid, when F0a < 1.
182
1, where is the sound frequency, and is the collision time. With increasing frequency, when 1, this mode becomes strongly dissipative, and sound
ceases to exist. In an interacting system, such as a Fermi liquid, however, another
mechanism of equilibration exists: collective interactions. As a result the sound
mode exists also in the opposite limit 1, and its velocity u0 or dimensionless
velocity s = u0 /vF (where vF is the Fermi velocity) is given by the expression
1
s s+1
ln
=1+ s .
2 s1
F0
(10.23)
For F0s 0, s 1, i.e. the velocity of zero sound approaches the Fermi velocity,
u0 vF . This value differs from the velocity of classical hydrodynamic sound
modes, which is v = 13 vF for a classical electron liquid, where the equilibrium is
established by electronelectron collisions and where every elementary volume is
in thermal equilibrium. Physically zero sound may be viewed as oscillations of the
shape of the Fermi surface, and not just a density wave.
There may also exist in a Fermi liquid analogous oscillations connected with
fermion spins. They may be called spin collective modes, although formally there
is no magnetic ordering in the usual sense. The velocities of such spin modes
are in general different from the sound velocity (10.23) because they contain
other Landau interaction parameters. Such spin waves were indeed observed
experimentally in some simple nonmagnetic metals.
One may ask the following question: we have learned before and stressed several
times that the presence of gapless collective modes, in particular sound modes (and
maybe of the electron and hole quasiparticles themselves?) is usually connected
with some ordering of the system and with broken continuous symmetry. And we
see that we have such gapless modes in the normal Landau Fermi liquid. Can we,
then, view the normal Fermi liquid as such an ordered system too, and if so, what
kind of order does exist in it?
This question is now under hot discussion, and it is not finally solved, or at least
the solution is not universally accepted, but one can indeed give some arguments
that this is really the case. The very ground state of the Fermi liquid the filled
Fermi sea is a unique well-defined state with zero entropy, and that is what we
usually have in an ordered state! What is then the order parameter, and which
symmetry is broken here? The possible answer is that the order parameter is
the Fermi surface itself. It is a well-defined object, a sharp surface in momentum
space, and the continuous symmetry corresponding to the possibility of arbitrary
shift of all momenta k is thus broken by the presence of this Fermi surface. As
the mathematical quantity characterizing this ordered state we can for example
take the Z-factor (8.54), (8.55), which gives the jump in the electron distribution
function n(p) at the Fermi surface, see (8.55); it can play a role similar to that of
183
the order parameter in the Landau theory of phase transitions. At high temperatures
the Fermi distribution is broadened, and a sharp Fermi surface, formally speaking,
disappears, but at T = 0 we may speak of an ordered state. In this picture the
zero sound discussed above the oscillations of the shape of the Fermi surface
may be naturally treated as the corresponding collective excitation (the Goldstone
mode).
This general point of view, the notion of Fermi liquid as a special type of
ordered state, does not give us directly any new results, but it is rather important
conceptually. It puts the normal Fermi liquid state in the same row, same category
as other types of ordered states that usually exist in nature when the temperature
goes to zero.

There exist, besides normal Fermi liquids, several other possible ground states of
interacting electron systems. The system may become superconducting; different
types of magnetic ordering are possible; there may occur states with inhomogeneous
charge distributions. Some of these states may become insulating. We will discuss
several such possibilities later, but now we ask the following question: is the
conventional Landau Fermi liquid the only possible state of a normal metal without
long-range order of some kind?
This problem has been actively discussed in recent years, in the beginning mostly
in connection with some unusual normal state properties of high-temperature superconductors, but now in a much broader context. There are several indications in
the experiment that these and possibly some other materials do not obey the usual
rules and behave in a way that is much different from the normal Fermi liquid.
10.2.1 Marginal Fermi liquid

There exist at present several suggestions or scenarios which can in principle give
a non-Fermi-liquid ground state. One of the first was the concept of the so-called
marginal Fermi liquid, formulated phenomenologically mainly by C. M. Varma. In
this theory it is assumed that the polarization operator (q, ), see Sections 8.7.1
and 9.1, does not strongly depend on q, and its -dependence is such that
%
(F ) /T for T
(10.24)
Im (q, ) =
for T c .
(F )
Here c is a certain cut-off frequency, which is supposed to be c F . From the
dispersion relation for (similar to the KramersKronig relations well known in
184
optics) one then gets

Re (q, ) (F ) ln
.
T
Putting these expressions into the self-energy of electrons, we get
=

x
+ i x sign()
() = ln
c
2
(10.25)
(10.26)
(10.27)
where x = max(, T ). This gives, using equation (8.87), the quasiparticle weight Z
(residue of the pole of the Green function)
1
1

,
(10.28)
1 Re /
1 ln y/c
(
'
where y = max (p ), T . Thus we see that the assumption (10.24) alone tells
us that the quasiparticle weight goes to zero at the Fermi surface, which means
that the jump in the electron distribution function n(p) at the Fermi surface, proportional to Z, disappears, and that the single-particle density of states tends to
zero for F . Simultaneously the quasiparticle lifetime, given by the imaginary
part of (10.27), is linear in energy, 1 (p ). As a result the quasiparticles are only marginally defined at the Fermi surface: in ordinary Fermi liquids
it is required (and assumed) that the lifetime is long enough, 1 (F )2 ,
see (10.3). If 1 were to remain constant at F , the Fermi surface would be completely smeared out; here it is still defined, but the properties of quasiparticles are
much different, and the quasiparticles themselves are not so well-defined as in an
ordinary Fermi liquid.
Z=
10.2.2 Non-Fermi-liquid close to a quantum critical point

The marginal Fermi liquid discussed above is only one of several possible types of
non-Fermi liquids. In general there can exist many other situations in which a metal
behaves as a non-Fermi liquid. The most widely discussed situation is that with
quantum critical points (QCP) in metallic systems, cf. Section 2.6. As explained in
that section, this is the situation in which the critical temperature of some ordering
is suppressed (goes to zero) by some external parameter (pressure, magnetic field,
etc.). When Tc 0, besides thermal fluctuations always present in the vicinity
of Tc , quantum fluctuations start to play an important role. They can significantly
modify the properties of the system in the vicinity of such QCP. In particular,
if we are dealing with a metallic system, which is, say, a normal Fermi liquid
185
non-Fermiliquid
region
Fermi-liquid
phase
ordered phase
Pressure
(or other control
parameter)
QCP
Fig. 10.2
at high pressures, see Fig. 10.2, often the conventional Fermi-liquid behaviour is
violated in the vicinity of QCP, see e.g. von Lohneisen et al. (2007). Notably,
different thermodynamic and transport properties behave in an abnormal way: e.g.
the specific heat, instead of the standard linear temperature dependence c = T ,
behaves as c T ; the resistivity, instead of the usual Fermi liquid asymptotic
behaviour AT 2 , becomes linear in temperature, T , or T 4/3 , etc.
Qualitatively the general explanation of these deviations from the normal Fermiliquid behaviour is that close to QCP the electrons strongly interact with collective
modes specific for a particular type of ordering which is especially strong when
Tc 0 and when these modes have strong quantum character. Specific features of
corresponding couplings in general depend on the particular ordering in question,
so that here, similarly to second-order phase transitions, cf. Section 2.5, we may
have in different situations different laws, with different exponents. A microscopic
theory of most of these phenomena is still absent.
Yet another very surprising phenomenon was discovered during the study of
metallic systems close to a quantum critical point. It was found that in many
such cases not only do we have non-Fermi-liquid behaviour, but there may appear
in the vicinity of QCP a novel phase a superconducting phase, see Fig. 10.3.
Such is, e.g. the situation in some itinerant magnets, e.g. UGe2 , URhGe, where
the magnetic ordering can be suppressed by pressure. A similar phenomenon is
observed in several heavy-fermion compounds (see Chapter 13), e.g. in CePd2 Si2 ,
and in some organic compounds, for example in (TMTTF)2 PF6 . There are also
ideas that this is what can happen in high-Tc cuprates, although this question is
still very controversial. In any case, there are strong reasons to believe that the
superconductivity which appears in this situation may be not of the conventional
s-wave BCS (BardeenCooperSchrieffer) type, but is unconventional, e.g. with
the singlet d-wave or triplet p-wave pairing.
186
ordered phase
normal
Fermi liquid
SC
QCP
SC superconducting phase
Fig. 10.3
10.2.3 Microscopic mechanisms of non-Fermi-liquid

behaviour; Luttinger liquid
It is not really clear what are the different possible microscopic mechanisms of
non-Fermi-liquid behaviour. Several microscopic models have been suggested:
singular electronelectron interaction originating from scattering on very soft excitons; an interaction with quadrupolar Kondo-centres; etc. None of these mechanisms has been really shown to result in the desired behaviour and to work, e.g.
in high-Tc cuprates. Probably the most reliable case is the one-dimensional electron system, where one can indeed show that there exists a state resembling a
Fermi liquid, but different from it. One can show that in 1d systems of interacting electrons formally there exists a Fermi surface, but it is not a Fermi surface
of ordinary electrons, but rather of spinons, cf. Section 6.4.2. The real electron
distribution function does not have any jump at F (recall that the existence of
such a jump is a prerequisite for ordinary Fermi liquids), but behaves instead as
n(p) n(pF ) |p pF | sign(p pF ), i.e. the momentum distribution and the
single-particle density of states have a power-law singularity at F . The exponent
depends on the electronelectron interaction; in typical cases 18 . Physically
this behaviour is connected with the spincharge separation: there exist in this
case two types of elementary excitations, spinons carrying spin but no charge, and
holons which have charge but no spin (the situation here is similar to the RVB
picture discussed in Chapter 6). Such a state is often called a Luttinger liquid.
Simple arguments which show that there should be spincharge separation in the
one-dimensional case may be given using the example of one-dimensional Hubbard
model with strong interactions (more detailed discussion of the Hubbard model in
general will be given later, in Chapter 12). For one electron per site the ground state
with localized electrons is almost antiferromagnetic, . If we now
187
Fig. 10.4
create one hole, then due to the exchange process the hole could move, say, to the
left, see Fig. 10.4, so that as a result the hole, which initially was surrounded by two
spins , i.e. , finally is surrounded by the correct antiferromagnetic spins,
, so that if we contract the system by removing the hole, the remaining
spin structure close to the location of the hole would be undisturbed. However,
at the same time in another part of the chain a pair of wrong spins, , would
remain. Thus the hole would live without disturbing the spin order of the remaining
electrons in its vicinity, but another purely spin excitation a spinon would be
created; this is what is meant by spincharge separation. This is the typical situation
in Luttinger liquids in general. This is the best-proven case of non-Fermi-liquid
behaviour unfortunately only in the 1d case!
Whether the behaviour of 2d or 3d systems may be in any respect similar to the
one-dimensional case, is not clear at present, although one may think that the RVB
state described above, Section 6.4.2, would lead to a similar picture. However, one
can easily see that in the 2d case the same process as illustrated in Fig. 10.4 would
not lead to simple spincharge separation; there will be a trace of wrong spins
connecting the hole and the remaining spin defect (the original location of the
hole). This will be discussed in more detail in Section 12.4. Thus, unfortunately,
even for the 2d case, not to speak of real 3d systems, there is no rigorous proof that
the state with spincharge separation and with the properties similar to a Luttinger
liquid can be realized in practice.
11
Instabilities and phase transitions in
electronic systems
We have already mentioned several times before (e.g. while discussing the giant
Kohn anomalies in Section 9.2 or the Pomeranchuk criteria of stability of the Fermi
liquid (10.20) in Chapter 10) that there may exist situations when the usual Fermiliquid state of electrons in metals becomes unstable, even for weak interactions. In
this case a transition to some new state, often an insulating state, can take place.
In this chapter we will discuss several such cases, using different approaches to
illustrate how the general theoretical methods described above really work. We
start with the Peierls transition.

One of the first historically, and conceptually simplest and best-known examples
of possible instabilities of the usual metallic state even for weak interactions is
the Peierls instability. We first give a simple qualitative picture, and then present
several equivalent ways to treat it, thus illustrating some of the approaches described
above.
11.1.1 Qualitative considerations

We start with the simplest case of a one-dimensional chain with one electron per
site, treating electrons in the tight-binding approximation. The energy spectrum in
this case is
(k) = 2t cos ka ,
(11.1)
where t is the electronelectron hopping between neighbouring sites, and a is

the lattice period (see Chapter 12 for more details). This spectrum is shown in
Fig. 11.1(a), and for one electron per site the corresponding band will be exactly
188
p
a
p
2a
189
2a
p p
2a
a
(a)
2a
p
a
(b)
Fig. 11.1
a
2a
Fig. 11.2
half-filled (Fermi momentum kF = /2a).1 If we now double the unit cell, shifting every second atom and forming dimers, Fig. 11.2, the new period will be 2a,
and the new Brillouin zone boundaries, instead of /a, will be /2a, i.e. they
will coincide with the original Fermi wavevectors. As always, there appear energy
gaps at the Brillouin zone boundaries, thus the spectrum would change from that
of Fig. 11.1(a) to that of Fig. 11.1(b). Thus we see that in effect the energies of
all occupied electronic states decrease, i.e. we gain electron energy in this process.
Of course we have to deform the lattice, which costs us elastic energy u2 , where
u is the lattice distortion in going from the regular to the dimerized arrangement
of atoms or ions. But if we gain more than we lose, this process will be energetically favourable, and the homogeneous chain will be unstable with respect to
dimerization.
These were the arguments first given by R. Peierls (2001) in a short section of
his book (first published in 1955). He had indeed found that the electron energy
gain always exceeds the elastic energy loss (the electronic energy goes as u2 ln u
whereas the elastic energy is quadratic, u2 ). However, in that book Peierls had
not presented a mathematical proof, thinking that it is either rather self-evident
or too simple (the mathematics is presented in a very nice short book by Peierls
(1991), devoted to some problems he had encountered in the course of his long
life in physics). We will give a mathematical treatment of this situation below,
using several different approaches. This problem, conceptually rather clear, can
1
In the physics of one-dimensional systems it is more common to denote momentum by k, and to speak about
2kF - or 4kF -instabilities, etc.; we follow this convention in this chapter.
190
Instabilities and phase transitions

e
/2
/2
Fig. 11.3
give us a good opportunity to illustrate different theoretical approaches which we

have discussed in general above and which are often used to treat this and similar
problems. This is why some of the conclusions of this section are repeated several
times below just to see how one can obtain them using different methods.
First, however, we note that the conclusion about the Peierls instability of onedimensional metallic systems does not apply only to the case of one electron per
site and to tight binding (where it gives Peierls dimerization): one would get the
same instability also for free electrons in a weak periodic potential, and for arbitrary
band fillings.
11.1.2 Peierls instability in the general case
Let us consider in more detail a one-dimensional metal in a nearly free-electron
approximation. As is well known, in the presence of a weak periodic lattice potential
with Fourier component UK the electron wavefunction (x) obeys the Schrodinger
equation

2
i Kx

k
i Kx
+ UK e
+e
(x) = (x) .
(11.2)
2m
Here K are the Umklapp wavevectors; we have also used the fact that due to
symmetry UK = UK . Equation (11.2) has solutions in the form of energy bands,
with energy gaps at k = Kn/2, Fig. 11.3, which appear due to mixing of states
with k and k nK (later on we put the lattice constant a = 1).
For weak interactions, when UK is treated as a perturbation, the energy spectrum
E(k) close to the points K (Brillouin zone boundaries) has the form

k kK 2
k + kK
+
2 ,
(11.3)
E(k) =
2
2
where k = k 2 /2m and the energy gap is
= UK .
191
eF
kF
kF
Fig. 11.4
kF
kF
Fig. 11.5
Let us start now with free electrons with the energy band filled up to the Fermi
momentum kF , Fig. 11.4. If we now create a periodic potential with period a =
2 a/2kF , i.e. with Fourier harmonic K = 2kF , e.g. if we shift all the atoms of our
one-dimensional chain according to the rule
ux = Rx Rx0 = u cos 2kF x ,
(11.4)
then we create energy gaps at kF , i.e. exactly at the Fermi surface, see Fig. 11.5.
After such a distortion our system will become an insulator with the gap
2
= 2U2kF . The strength of the extra periodic potential U2kF thus created will
be proportional to the amplitude of the distortion u.

We see that, similar to the case discussed in Section 11.1.1, after distortion the
energies of all occupied states decrease. Thus we gain some energy in this process.
But by making the lattice deformation (11.4) we also lose elastic energy 12 B u 2 ,
where B is the bulk modulus. Therefore the energy balance between the electronic
energy gain and the elastic energy loss determines whether such distortion will be
energetically favourable, and if it is, what will be the amplitude of corresponding
distortion.
We can easily calculate the change in the electronic energy using equation (11.3):
kF
'
(
Eel =
E(k) (k) dk .
(11.5)
kF
192
This straightforward calculation shows that the dominant term in Eel for weak
distortion u or small energy gap
= U2kF u has the form
Eel =
2
(F ) F
ln
.
2
(11.6)
Problem: Check this, using the spectrum (11.2).

Solution: The easiest way to get this result is to linearize the spectrum (k) in
the vicinity of the Fermi points kF , (k) = v#F (|k| k#F ), or to transform the
integral (11.5) to an integral in by the usual rules dk () d; for simplicity
we consider here the case corresponding to the doubling of the period.
11.1.3 Different theoretical ways to treat Peierls distortion
We see from (11.6) that it is always favourable in 1d metals to make a distortion
with the wavevector q = 2kF : the electronic energy gain
2 ln(F /
) u 2 ln u is
always bigger than the elastic energy loss 12 B u 2 . To find the value of this distortion,
we have to minimize the total energy
2 (F ) F B u 2
ln +
,
(11.7)
2
2
but for that we first have to find the connection between u and
. We consider
below the simplest case of one electron per site, i.e. half-filled tight-binding bands,
in which case 2kF = and the Peierls transition corresponds to dimerization
often one uses the term Peierls transition just for this case. (We put here the
original lattice constant a = 1.)
The exact relation between U2kF =
and the distortion u can be found from the
following considerations. Usually one writes the electronlattice interaction in the
form
Heph =
g(q) ckq, ck, (bq + bq )
(11.8)
E (
) =
k,q,
(Frohlich Hamiltonian, see Section 8.6). The classical treatment corresponds

to treating lattice displacements with period q = 2kF as static, i.e. substituting
bq + bq = 2b (q 2kF ). Separating terms with this q and denoting g(2kF ) = g,

(2kF ) = ,
we see that the Fourier component of the extra potential U2kF , equal
to the energy gap
, is
.
U2kF =
= 2gb
(11.9)
Remembering now that the distortion u is connected to the phonon operators by

1
(b + b) (see (4.10)), we have the connection between the energy gap
u = 2M
193
short, strong bonds, higher electron density
or
weak bonds
Fig. 11.6
and the distortion u(2kF ) = u in the form
= g 2M u .
(11.10)
Putting (11.10) into (11.7) and minimizing E (

) with respect to
, we obtain
finally in the weak coupling case (
F , g 0)
= F e1/2 ,
(11.11)
where the dimensionless electronphonon coupling constant is

=
g 2 (F )
(11.12)
(here we have taken into account that 2 = B/M).

For arbitrary band filling (2kF
= ) the period of the superstructure in general
will be incommensurate with the underlying lattice period; in this case there will be
not 1/2, but 1/ in the exponent in (11.11).2 The instability of a one-dimensional
metal towards lattice distortion with wavefactor 2kF the Peierls instability
simultaneously means the creation of charge-density wave (the short notation,
which is now widely used, is CDW). For instance when the period is doubled,
the electron density alternates at consecutive bonds, see Fig. 11.6; this will be
mathematically shown below.
The fact that our one-dimensional metallic system is unstable with respect to the
formation of CDW or to Peierls distortion is connected with the logarithmic divergence of the response function or the polarization operator (giant Kohn anomaly),
cf. (9.20), (9.21), (9.30):

1 2kF 1 + q/2kF
ln
.
(11.13)
(q, = 0)
vF q
1 q/2kF
One can see this instability if one considers phonon renormalization: as discussed
in Chapter 9 (see (9.34)) the phonon frequency (q) becomes imaginary at q = 2kF ,
which tells us that if we start from a homogeneous undistorted metallic chain, the
2
The factor of 2 in equation (11.11) is characteristic of a half-filled band (Peierls dimerization) and is due to
Umklapp processes (in this case the processes with momentum transfer 2kF = are equivalent to the Umklapp
process with K 2kF = 2 = , which doubles the coupling constant in equation (11.11)).
194
increment of the vibrations with q = 2kF will be positive, and a macroscopic

standing wave with this period will be created, i.e. Peierls distortion will occur.
There are different theoretical methods to consider this situation. One of them
is the classical treatment described above. We can also proceed directly from the
electronphonon Hamiltonian

H=
(k) ck, ck, +
g(q) ckq, ck, (bq + bq ) +
(q)bq bq .
k,
k,q,
(11.14)
We expect that the phonon mode with q = 2kF will be macroscopically occupied,
i.e. it will behave classically. Accordingly we replace operators bq=2kF , b2kF by

c-numbers, b = b2kF . Keeping only these main terms, we have

2
H=
(k) ck, ck, +
(11.15)
g ck2kF , ck, 2b + b .
k,
k,
Here g = g(2kF ), = (2kF ); for simplicity we take b to be real (it is quite

We also omit for simplicity signs at 2kF ,
easy to include also the phase of b).
remembering only that we should take symmetrized sums when necessary, and
take care of momentum conservation.
We can proceed in several formally different but equivalent ways. One is to
use the Bogolyubov canonical transformation: the Hamiltonian (11.15) is already
quadratic, but not yet diagonal; it contains terms ck2kF ck . This is done below,
see equations (11.24)(11.25). Another possible way is to write down the average
energy E = H which will have terms of the type
3
2
2
E = g
ck2kF , ck, 2b + b + E0
(11.16)
k,
we obtain
(and similar terms with k + 2kF ). Minimizing E with respect to b,
2
3
.
(11.17)
c
c
b =
k, k2kF , k,
We see from (11.17) that if, as we have assumed, b = bq=2kF
= 0, then the average

k, ck2kF, ck, is also nonzero. One can show that this implies modulation of
the electron density with the same wavevector 2kF , i.e. the formation of a chargedensity wave mentioned above. Indeed the electron density is

(x)
=
(x) (x) =
q eiqx .
(11.18)
q
195
Its Fourier transform is

=
ckq, ck, .
(11.19)
k,
(The product in coordinate space, here (x) (x), goes over

to a convolution in

momentum space.) Thus the average density q is equal to

k, ckq, ck, .
If we were to have, as in the normal case, that only ck ck

= 0, i.e. ckq ck =

0 (q), then, from (11.18), (x) = 0 = const. In our case, however, we have
bq=2kF
= 0, and, from (11.17), also 2kF
= 0 (11.19). In other words
q = 0 (q) + 1 (q 2kF ) ,
1 =
2
ck2kF , ck, . (11.20)
k,
From (11.17) we see then that there appears modulation of the electron density
with period (2kF )1 :

(x) = 0 + 1 ei2kF x = 0 + 1 cos(2kF x) .
(11.21)
(

'
(In a more accurate treatment of course (x) = Re 0 + 1 ei2kF x , or, in other
words, if the Fourier harmonic with q = 2kF enters in (11.20), so also does the
harmonic with q = 2kF , which gives (x) cos(2kF x) as in (11.21).)
From (11.20), (11.21) and (11.17) we also see that the electron density wave,
or CDW, is proportional to the lattice distortion. Thus indeed the Peierls transition
goes hand in hand with charge-density wave formation; these are different sides
of the same phenomenon. Usually people speak about a Peierls transition when
dealing with one-dimensional systems, and use the terminology CDW for similar
phenomena in 2d or 3d materials (such transitions may occur in these cases when
special conditions nesting of the parts of the Fermi surface are fulfilled, see
equation (9.37) and discussions thereof).
There is yet another approach to our problem. We can exclude phonon variables
and go over to an effective electronelectron interaction. The correct way to do this
would be, e.g. by writing down the electronelectron vertex due to the exchange
of a phonon,
196
One can also proceed in a less rigorous way, starting from equation (11.15).
Putting the expression for b (11.17) back into (11.15) we would get
H=
q2
ck2kF , ck, ck +2kF , ck ,
k,
kk ,
*
)
+ (2kF ) (+2kF ) .
(k) ck, ck,
(11.22)
One may treat this as a mean field decoupling of the four-fermion interaction
Hint =
g 2
ckq, ck, ck +q, ck , ,

kk ,
(11.23)
with q = 2kF ; this is exactly the interaction we would obtain by excluding

phonons as mentioned above.
We see that the exchange of phonons leads to an effective electronelectron
attraction. (Actually such interactions will be retarded, or frequency-dependent;
our treatment is carried out in the static limit, 0. This attraction is in fact the
mechanism of Cooper pair formation and of superconductivity in ordinary superconductors.) We thus see that we can proceed in two ways; either treat electron
phonon interactions in an apparent way, or first exclude phonons and then consider
electrons with an effective attractive interaction. We would get essentially the same
instability, which in the first method would be ascribed to lattice distortions here
we use the term Peierls transition; in the second approach we would speak of a
CDW transition. If there exists an electronelectron attraction of any other origin
besides the electronphonon interaction, it will also lead in the 1d case to CDW
instability; the lattice will of course then follow. Note right away that if there
exists not an attraction but an effective repulsion between electrons, 1d systems
would still be unstable, however not with respect to charge-density wave, but to
spin-density wave (SDW) formation see below, Section 11.6.
The treatment given above is not yet the complete solution. We have to
make only one more step. Actually we have already obtained the Hamiltonian
(equation (11.15) or (11.22)) which is quadratic in the electron operators. It can be
written as
*
)
H=
(11.24)
(k) ck, ck, +
ck2kF , ck, + h.c.
k,

2
with (as yet undetermined)

= g b = g
k, ck2kF , ck, , which should be
found later self-consistently.
We see that the situation now resembles the case of Bose condensation, Chapter 5,
or the treatment of antiferromagnetic magnons, Chapter 7: the Hamiltonian (11.24)
197
is quadratic, but still nondiagonal; it contains terms ck2kF , ck, . Thus we again
have to use the Bogolyubov canonical transformation to diagonalize it:
k, = uk ck, + vk ck2kF ,
k, = vk ck, uk ck2kF , .
(11.25)
With a proper choice of coefficients uk , vk the Hamiltonian (11.24) now takes the
form
H=
E(k)(k, k, + k, k, ) + const. ,
(11.26)
k,
with the spectrum (11.3) (with K = 2kF ), as of course it should. This transformation is also very similar to the well-known transformation in the theory of
superconductivity, see below, Section 11.5, equation (11.50). The difference is that
whereas in the latter case we mix electrons and holes, uk ck, + vk ck, , here
we mix two electrons with different wavevectors. Canonical transformations in
the theory of superconductivity imply nonzero averages of the type ck, ck, ,
the famous Cooper pairs. Analogously, the transformation (11.25) corresponds
to nonzero averages ck2kF , ck, , see (11.17). We may call them electronhole
pairs, or excitons. Thus we can speak here about an electronhole, or exciton
condensate.
Note also that this procedure (mixing electron states with momenta k and
k 2kF ) is essentially the same procedure as is usually done in treatments of
the energy spectrum and band formation in a periodic potential, e.g. in the weak
coupling approximation, see (11.2), (11.3) and any textbook on solid state physics.
Similar to this case, we have obtained here that after this transformation the spectrum acquires a gap at k = 2kF , i.e. exactly at the position of the initial Fermi
surface, in agreement with equation (11.3) and Fig. 11.5.
Thus we see that there are several, technically somewhat different, but conceptually very close ways to treat Peierls distortion: a purely classical treatment
of equations (11.7)(11.12); a treatment using the electronphonon (Frohlich)
Hamiltonian with mean field decoupling; the transition to the effective electron
electron interaction by excluding phonons, with similar mean field decoupling at
this level. At this (mean field) level all these methods are equivalent and give the
same results. However, if one would want to go beyond a self-consistent treatment
and consider fluctuation effects, both classical and quantum, the different starting points (electronphonon model, or the effective electronelectron one) would
become more, or less appropriate, depending on the effects studied.
198

H
Fig. 11.7
H
Fig. 11.8
11.1.4 Peierls distortion and some of its physical consequences in

real systems
Peierls distortion is observed experimentally in a number of quasi-one-dimensional compounds. Although formally in a pure 1d system there should be no phase
transition, in systems consisting of many chains, with weak interchain interaction,
the Peierls transition actually occurs as a real phase transition.
Probably the best-known and the clearest example of a system whose properties
are described using this picture is met in organic materials with conjugate bonds.
Carbon is usually four-valent, but often its s- and p-electrons, due to sp2 hybridization, form three strong chemical bonds (called -bonds) lying in one plane at 120
to one another, and the remaining electron occupies the perpendicular pz -orbital.
This extra electron (one per carbon ion) is called a -electron. These extra electrons can either localize at certain bonds, giving what is called in chemistry double
bonds, or -bonds; or these electrons may be delocalized, forming broad bands.
They are in fact responsible for electric conduction in graphite, or in (very popular
nowadays) graphene, single-layer sheets of graphite. In small organic molecules
these double bonds can give rise to resonating structures, e.g. in the benzene
molecule C6 H6 , see Fig. 11.7 (this actually gave rise to the notion of resonating
valence bonds, discussed above, in Chapter 6). And in some cases these half-filled
-bands (one -electron per site!) may show the phenomenon of Peierls distortion. A good example is polyacetylene Cn Hn , which is usually depicted as shown
in Fig. 11.8. In chemists notation double bonds correspond to shorter distances
between carbon atoms. This structure can be visualized as originating from the
homogeneous one
as a result of Peierls distortion with the
199
doubling of the period (which actually would have occurred at temperatures much
above the temperature of decomposition of this substance).
There is one very interesting aspect of the physics of polyacetylene and other
similar systems. In this particular case (doubled period) there exist two equivalent configurations:
and
can form a defect a domain wall
. One
. Similarly to one-
dimensional magnets, this defect may propagate along the chain, i.e. it forms an
elementary excitation, a soliton. One can see that there are different possible states
of such an object. It may be neutral: each carbon usually has four electrons (four
bonds in Fig. 11.8). When we create a defect of the type
an unpaired electron may remain at the appropriate carbon atom (actually it will
be delocalized over a certain distance ). In this case the object, the soliton, will be
neutral: there will again be four electrons at this site. However, in contrast to the
initial case, when all electrons participate in valence bonds and form singlets, such
a neutral soliton will have an unpaired spin 12 . Thus it is an elementary excitation
which carries spin but no charge.
But we can also create charged excitations. Let us start from the neutral soliton
described above, and let us remove an electron (take it to infinity). The remaining
object will have uncompensated positive charge +1, but no spin. Similarly, we can
put onto a neutral soliton an extra electron which will form a singlet pair with the
existing unpaired one. The resulting object will have charge 1 and again no spin.
Thus we see that, in contrast to the usual electrons in metals or semiconductors,
which carry both charge and spin, there is here a spincharge separation: neutral
solitons carry spin 12 , and charged solitons carry charge 1 but no spin. Such
solitons were indeed observed in polyacetylene. (This picture was developed by
Su, Schrieffer and Heeger, and these solitons are sometimes called SSH solitons.)
As we see, there exists a close analogy between the properties of one-dimensional
systems with Peierls distortion (neutral solitons carrying spin 12 , charged soliton
with spin 0) and the properties of elementary excitations (spinons and holons) in the
RVB state discussed in Chapter 6. This is not accidental: actually the dimerization
in one-dimensional materials leads in the strong coupling case to the formation of
the usual valence bonds (singlet states of two electrons 12 (1 2 1 2)) well
known in chemistry; we have actually already used this when we depicted possible
states in polyacetylene. Thus one can say that the Peierls dimerization of a halffilled one-dimensional metal is a first step towards the formation of a molecular
crystal, like molecular hydrogen, consisting of H2 molecules. Actually, this property
of spincharge separation can be really proven only in 1d systems, whereas its
existence in higher dimensions (e.g. in the 2d case discussed in Chapter 6) is still
200
hypothetical. And, as mentioned in Chapter 10, such spincharge separation is

responsible for the formation of the Luttinger liquid state, one of the very few cases
in which the existence of a non-Fermi-liquid state is really established theoretically.
(Note that the treatment presented above, which gave an insulating state due to a
Peierls or CDW transition, was actually a mean field treatment implicitly relying on
the interchain coupling in a quasi-one-dimensional system. Formally in the purely
1d case quantum fluctuations would be so strong that they would destroy this mean
field long-range order, and consequently there will appear no gap in the energy
spectrum. But the spectrum itself would be very different from the usual electron
spectrum; there will be spinon and holon excitations. This resulting state is just the
Luttinger liquid state mentioned above.)
One more interesting consequence of the picture described above becomes clear
if we consider Peierls transitions for a different band filling, e.g. corresponding not
to dimerization, but to tripling of the period. It is clear that there will now be three
possible degenerate ground states, e.g.
= =
= =
= =
(1)
(2)
(11.27)
(3)
and correspondingly more (essentially three independent) types of domain walls,

or solitons. Arguments similar to those given above lead to a striking conclusion:
the charge of these solitons will not be e, but will be fractional ( 23 e), like quarks
in elementary particle theory!3
Returning to the case of double periodicity one can show that energetically the
solitons are indeed the most favourable excitations. The spectrum of one-particle
excitations (electrons and holes) is given by equation (11.3) which in the case of a
half-filled tight-binding band looks like the one shown in Fig. 11.9, i.e. it consists
of two subbands, Fig. 11.10. Consequently, the lowest electronhole excitations
(for a fixed lattice) have minimum energy 2
. It turns out that the soliton level
lies exactly in the middle of the energy gap, Fig. 11.11, thus it costs less energy
to create such an excitation. In the symmetric case the chemical potential also
lies in the middle of the gap, so it seems that formally the creation of a soliton
would cost no energy. However, to make the soliton we also have to change the
lattice configuration, e.g. create two neighbouring long bonds; this also costs some
3
Indeed the simplest excitation which keeps the state the same at the right and left ends of the molecule would
be the one in which we remove one double bond from one of the states, e.g. from the state (1) in (11.27). But
then we can move the double bonds so as to create a domain wall of the type = = = (three
single bonds in a row). We go back to the original domain (1) after three such steps, such domain walls, i.e. the
missing charge 2e (two removed electrons) will now be equally split into three solitons, or domain walls, and
thus the charge of each of them will be 23 e.
p /2
201
p /2
or
p /2
p /2
Fig. 11.9
Fig. 11.10
soliton
Fig. 11.11
energy, so that the resulting energy of the soliton is not

, but 2
. Nevertheless it
is still more favourable to create two charged solitons (or soliton and antisoliton)
instead of an independent electronhole pair; it costs 4
instead of 2
. Similarly
charge carriers introduced by doping also create solitons and are bound to them.
An interesting situation may occur if such a CDW superstructure is incommensurate with the underlying lattice period. In this case the density wave which has
the form (x) = 0 cos(2kF x + ) may have an arbitrary phase, or may be located
at an arbitrary position relative to the lattice. There exists a collective mode
oscillations of the phase , the so-called phasons, with the gapless spectrum
(q) = a vF q ,
a 1 = 1 +
4
2
.
02
(11.28)
Physically such a mode (it is also called a Frohlich collective mode) for q 0
corresponds to a CDW which slides along the chain, Fig. 11.12. Such a sliding mode
contributes to the conductivity of the system, because it carries with itself an extra
charge; at finite frequency it is optically active. This mechanism of conductivity,
202
Fig. 11.12
due to sliding CDWs, is called Frohlich conductivity; in fact Frohlich suggested this
picture in the 1950s as a possible explanation of superconductivity. It turned out later
that the real superconductivity is explained differently, but Frohlich conductivity
was observed recently in quasi-one-dimensional system NbSe3 and in several other
materials. (Usually a CDW is pinned to the lattice both by commensurability effects
and, e.g. by impurities, defects, etc., thus one needs a certain critical electric field to
set it in motion. Thus experimentally the fingerprint of conductivity due to sliding
CDW is the existence of such a threshold, or nonlinear I V characteristics. Another
experimental check is the presence of a specific noise spectrum in conductivity.)

There exists another class of systems for which Peierls instability may be important:
these are (quasi-)one-dimensional antiferromagnets. We have already discussed in
Chapter 6 the properties of one-dimensional antiferromagnets and have shown that
the valence bond wavefunction, the structure of the type
, gives a
lower energy than the usual two-sublattice Neel structure even without resonance,
see (6.114)(6.118). In chemical language this corresponds to the formation of
valence bonds, 12 (12 12), as in molecular hydrogen H2 . Actually in the
spin chain with a fixed lattice and equal distances between spins there is an even
better state in which there exists resonance between these bonds. However, if we
release the lattice and permit it to adjust to the spin structure, it may turn out
that the state of the type drawn above may become preferential: there may occur a
dimerization of the lattice, leading to an alternation of the exchange constants,
Jn,n+1 = J0 + J (1)n = J0 J .
(11.29)
One can easily see that this is indeed the case in the one-dimensional xy
model (6.109). We have seen in Chapter 6 that this model can be mapped by a
JordanWigner transformation onto the model of noninteracting spinless fermions
with the spectrum k = 2J cos k, see (6.111). As discussed there, the ground state
in this case corresponds to a half-filled band, Fig. 11.13. It is evident that when we
now take into account interaction with the lattice and allow the lattice to distort,
there will occur the same Peierls transition, which in this case is called a spin-Peierls
transition. (Of course, here again we have to include coupling to neighbouring spin
203
p /2
p /2
Fig. 11.13
chains, otherwise quantum fluctuations would suppress such long-range order. Usually one invokes here an elastic interaction between distortions on different chains;
lattice distortion is here treated classically.)
An effective spinlattice interaction (or fermionlattice interaction) originates
from the distance dependence of the exchange interaction

J
u Si Sj
J (R)Si Si+1 = J (R0 + u)Si Si+1 = J0 +
R
= J0 Si Si+1 + g Si Si+1 u
(11.30)
i.e. the spinlattice coupling constant is g J /R. Similarly to (11.11), there

will be a distortion and a gap in the spectrum of spin excitations
J e1/ ,
g2
.
0 J
(11.31)
Consequently, e.g. the magnetic susceptibility of such a system will behave as

e
/T at temperatures below the spin-Peierls transition temperature. The
typical behaviour of susceptibility in this case is shown schematically in Fig. 11.14.
The susceptibility follows the CurieWeiss law at high temperatures, then passes
through the maximum at about T 0.7J (called the BonnerFisher maximum),
and at lower temperature Tc
(11.31) the system has a spin-Peierls phase
transition, below which the susceptibility decreases exponentially. One can show
(this was done by Cross and Fisher) that the Heisenberg model is even more
unstable with respect to the spin-Peierls transition than the xy model, so that the
energy gap is not proportional to the lattice displacement u as in equation (11.10),
but to
u4/3 .
An interesting situation occurs when we put a spin-Peierls system in an external
magnetic field H . As discussed in Chapter 6, such a field plays the role of the
chemical potential of JordanWigner fermions,

H=
2J cos k ck ck H
ck ck .
(11.32)
n
204
~ 1/T
~ e /T
Tsp
T
Fig. 11.14
k F
kF
Fig. 11.15
Consequently, for H
= 0 the situation would correspond to a fermion band filled
by more (or less) than one fermion per site, Fig. 11.15. This means that the
wavevector of the instability Q = 2kF will be different from , i.e. there will
occur not a simple dimerization, but there will appear instead a superstructure
with Q = 2kF , in general incommensurate with the initial crystal lattice, and the
period will depend on the magnetic field. Actually due to the fact that for the
dimerization the effective coupling constant is twice as large as that for other
periods (see (11.11) and discussions there), the energy gain for the dimerization
is larger than for other periods. Therefore, for small enough fields (band close to
being half-filled) the dimerized state will still be the most favourable one (pinning
of the distortion in the commensurate structure), but away from this region of
attraction to dimerization the period will indeed change and become in general
incommensurate. This problem may be solved at least close to Tc , and the resulting
phase diagram has the form shown in Fig. 11.16, where denotes the point of
cross-over to the incommensurate structure, which is an example of the Lifshits
point discussed in Section 2.4.
The phenomenon of the spin-Peierls transition was experimentally observed in
several quasi-one-dimensional organic compounds and recently in the inorganic
compound CuGeO3 which contains chains of ions Cu2+ with spin 12 . The phase
205
disordered phase
commensurate
structure
(dimerized)
Lifshits point
*
incommensurate
structure
first-order phase transition

Fig. 11.16
dimerized phase
3p
2p
Qx
(a)
domain wall
(soliton)
(b)
Fig. 11.17
diagram of the type shown in Fig. 11.16 (which was already presented in the general
discussion in Section 2.4) has indeed been observed experimentally.
There is yet another very interesting aspect of this story. In the first approximation
the distortion in the incommensurate phase is indeed simply u(x) = u eiQx , with the
period l = 2/Q in general incommensurate with the underlying lattice period.
We can also write this down as u(x) = u ei(x) , with the phase (x) = Qx, see
Fig. 11.17(a). However, in a more detailed treatment the picture is different. On
average the distortion is similar to the previous case, but actually there appears
inhomogeneous distortion, with the domains of the dimerized phase divided by
domain walls. This is illustrated schematically in Fig. 11.17(b).4 These domain
walls are actually the same solitons discussed above, with the difference that here
they are not excitations, but exist in the ground state and form a soliton lattice. Inside
the commensurate domains, spins are paired into singlets, but at each domain wall
there is an unpaired spin. All these spins are parallel to the external magnetic
4
Interestingly, such a curve appeared in 1904 on the first pages of the science fiction novel The Food of the Gods
by H. G. Wells, and with a rather similar scientific content!
206
ky
kx
Fig. 11.18
field, and it is this energy gain which in fact leads to the formation of such a state.
Interestingly enough, there may exist also new excitations oscillations of this new
superlattice, similar to ordinary phonons in crystals, etc. Such a soliton lattice
seems to be observed experimentally in CuGeO3 .
11.3 Charge-density waves and structural transitions,

higher-dimensional systems
The idea of the Peierls transition was first put forth for one-dimensional systems.
It was realized later that in principle one may have similar situations also in 2d
and 3d systems with special features of the electron energy spectrum. As already
mentioned in Chapter 9, there may exist situations with nesting of the whole or
parts of the Fermi surface, see equation (9.37). If the energy spectrum satisfies the
condition (k + Q) = (k) (with energies counted from the Fermi energy), the
response functions and polarization operator will behave essentially similarly to
the one-dimensional case, see (11.13). Consequently, such systems will be unstable
with respect to CDW formation and lattice distortion with the wavevector Q (if the
dominant interaction is the electronphonon interaction; however if the dominant
electronelectron interaction is repulsive, we will have similar instability in the
spin channel, and a spin-density wave SDW will be formed instead, see Section
11.6 below).
If the condition (k + Q) = ( Q) is not fulfilled over the whole Fermi surface,
but only parts of it are nested, such transitions may still occur; this will lead to
the creation of an energy gap at appropriate parts of the Fermi surface, leaving the
remaining parts intact; see the schematic picture in Fig. 11.18. This phenomenon is
observed in some layered transition metal dichalcogenides (NbSe2 ; TaS2 ). In some
of them (e.g. 1T TaS2 ) when there is strong nesting and the gap covers nearly the
whole Fermi surface, the material becomes practically insulating. In others, e.g. in
2H NbSe2 , there occurs partial gapping of the Fermi surface, but the remaining
parts of the Fermi surface make these materials still metallic below TCDW (NbSe2
207
Tc
TCDW
Fig. 11.19
e1(k)
e2(k)
(a)
Q
(b)
(c)
Fig. 11.20
even becomes superconducting at still lower temperatures). One may expect that
the typical behaviour of resistivity in this case may look as shown in Fig. 11.19:
the removal of a part of the Fermi surface (reduction of the number of carriers)
can increase resistivity at TCDW (and the system may still become superconducting
at Tc < TCDW ). However, the results may be even opposite and the resistivity
may decrease below CDW formation, if in this process the scattering of electrons
decreases; this seems to be the case in NbSe2 .

(The following section has a somewhat more technical character. But the methods
of the theoretical treatment of the situation discussed below are widely used now
in different situations, and I think this is a good opportunity to learn how these
methods work.)
Phenomena very similar to those considered in the previous sections may occur
in another situation in the so-called excitonic insulators. These are systems in
which the initial band structure consists of two overlapping bands, one of which
is filled by electrons and the other by holes, see Fig. 11.20. (If the band overlap is
208
small, such materials are called semimetals.) There exist in this case electron and
hole pockets of the Fermi surface. If the electron and hole Fermi surfaces coincide,
as in Fig. 11.20(a), or will coincide after a shift by a certain wavevector Q as in
Fig. 11.20(b), then this will be exactly equivalent to the nesting situation (9.37), and
there will exist an instability resembling the Peierls instability in one-dimensional
systems: the attraction between electrons and holes will lead to the formation of
excitons or rather of an excitonic condensate. As a result the material will become
insulating, with the new band structure shown in Fig. 11.20(c) by thick lines. (I put
excitons in quotation marks because it turns out that for weak coupling the radius
of such excitons is much larger than the average distance between them, so that
such an excitonic insulator should be visualized as a collective state with excitonic
correlations, but not actually consisting of real excitons. The situation here very
much resembles that in weak coupling BCS superconductors, where Cooper pairs
strongly overlap and the ground state is a collective state with appropriate pairing
correlations.)
The mathematical treatment of this problem is also very similar to that of the
Peierls transition and of the BCS theory of superconductivity. First of all one can
check that there indeed exists an instability of the normal state. One can see this
by looking at the two-particle (electronhole) Green function
1
where
denotes electrons in the conduction band, and
denotes
holes in the valence band (in the following discussion I mostly follow the work by
Keldysh and Kopaev who were among the first to study this problem).
(An option: from here on you may first go to Intermezzo, Section 11.5, and then
come back to this page.)
Consider the situation shown in Fig. 11.20(a) and take the spectra

k2
k2
1,2 (k) =
+ F
2m1,2 2m1,2

(11.33)
(below, for simplicity, we will consider the case of equal masses, m1 = m2 ). The
calculation of the electronhole two-particle Green function, or, equivalently, the
scattering amplitude for electronhole scattering in the usual RPA approximation
(we use here the alternative form of Feynman diagrams for the electronelectron
interaction, rather than the one used in Chapter 9, because it is more convenient,
209
but of course equivalent, for four-fermion interactions (11.37))

1
=
2
(11.34)
gives the following result:

(k = 0, ) =

.
1 + ln 2 0 i 2
(11.35)
Here 0 is a cut-off of the order of F , k and are the total momentum and
energy, and is the electronhole coupling constant which depends on the detailed
interaction involved. The origin of the logarithm in the denominator of (11.35) is
actually the same as in the case of the one-dimensional Peierls instability or its
three-dimensional analogue, the giant Kohn anomaly in the case of nesting, cf.
Section 9.2. Here the corresponding expression is written accurately, and one
sees that the electronhole Green function, or the scattering amplitude, has an
imaginary pole at = i, if < 0 (which is the case here, as we have electron
hole attraction). The position of this pole is given by the expression
= 20 e1/|| .
(11.36)
The existence of this pole shows that our system is unstable towards the formation
of electronhole pairs (excitons) from different bands close to the Fermi surface.
The position of the pole = 20 e1/|| gives the binding energy of the pair, i.e.
it determines the magnitude of the energy gap in the spectrum, see Fig. 11.20(c).
To find this spectrum and to discuss the properties of the resulting state one can
proceed in different ways. One of the methods is similar to the one which we
have used, e.g. in the treatment of the Peierls transition: one writes the interband
interaction as
Hint =
V (k, k , q) c1 (k + q) c2 (k q) c2 (k ) c1 (k) (11.37)
and keeps the most divergent terms which correspond to the scattering of electrons
from band 1 and holes from band 2 with the same wavevector, i.e. the terms with
k + q = k . Then we make a decoupling

Hint c1 (k ) c2 (k) c2 (k ) c1 (k) = H int c1 (k ) c2 (k ) c2 (k) c1 (k)
(11.38)
210
and diagonalize the resulting quadratic Hamiltonian H = H0 + H int by the

Bogolyubov canonical transformation. The spin structure of the average
c2 (k) c1 (k)
(k) ,
(11.39)
which plays the role of the order parameter, may be in principle different, i.e. exciton
correlations may occur either in a singlet or in a triplet channel. We consider below
the singlet case = and omit the spin indices.5
The simplified interaction term H int (11.38) is still nondiagonal in the band
indices 1, 2, and we have to use the transformation practically identical to the one
used for treating Bose condensation in Section 5.2 or for studying antiferromagnetic
spin waves in Section 6.3.2, with the only difference that because of the Fermi
statistics here the signs in corresponding linear transformations should be different.
Thus we go from the operators c1 , c2 to new operators , , defined as
c1 (k) = uk (k) + vk (k) ,
(11.40)
etc. and require that in terms of these new operators the Hamiltonian should be
diagonal. We leave the details to the reader (the treatment coincides with that of
Sections 5.2 and 6.3.2; see also the next section). Finally we get the following
spectrum of one-particle excitations:

'
(2
1 (k) 2 (k)
1 (k) + 2 (k)
+
2 ,
(11.41)
1,2 (k) =
2
4
where
is equal to (11.36). This is the anticipated result shown in Fig. 11.20(c).
Note that the ground state thus obtained is an eigenstate of the new operators
, , which are linear combinations of the original electrons c1 , c2 . As a result in
the ground state we have a nonzero anomalous average c1 (k)c2 (k)

= 0, which
describes the excitonic correlations (electronhole pairing) discussed above. This
anomalous average serves as the order parameter of the excitonic insulator.
One can use this model to illustrate yet another method widely used for treating
similar problems the method of anomalous Green functions. It was introduced by
Gorkov for treating superconductivity, and the corresponding equations are called
Gorkov equations. In addition to the usual Green functions G1 (k, ), G2 (k, )
for the two bands, one introduces new mixed Green functions F , F : in the
coordinate representation they are defined as (cf. (8.31))
6
7
F (r 1 , t1 ; r 2 , t2 ) = i T 2 (r 2 , t2 ) 1 (r 1 , t1 )
(11.42)
6
7
F (r 1 , t1 ; r 2 , t2 ) = i T 1 (r 1 , t1 ) 2 (r 2 , t2 ) .
5
This order parameter can also be imaginary, in which case the resulting state would correspond to that with
orbital currents (B. Halperin and T. M. Rice), see Section 11.7 below.
211
These functions are necessary because, according to our expectations and to the
physical picture, there will occur some mixing between the two bands 1 and 2
(proportional to the excitonic average (11.39)). The new Green functions F , F
introduced above describe just this effect. Writing down the equations of motion
for the ordinary Green function, e.g. G1 , we obtain on the right-hand side terms of
higher order which should be decoupled in some way. Similar to (11.38) we keep
the interband terms, which will give the anomalous Green functions F (11.42). The
resulting equations have the form

1 (k) G1 (k, ) +
F (k, ) = 1
(11.43)

2 (k) F (k, )
G1 (k, ) = 0 ,
where the quantity (c-number)
is

d 3 p d
.
(k) = i F ( p, ) V ( p k)
(2 )4
(11.44)
The solution of equations (11.43) has the form

G1 (k, ) =
F (k, ) =
2 (k)

2

2 1 (k) + 2 (k) + 1 (k) 2 (k)
(k)

(k)
(11.45)

2 ,

2 1 (k) + 2 (k) + 1 (k) 2 (k)
(k)
from which we again find the spectrum (11.41), and the solution of the selfconsistent equation (11.45) gives the same energy gap
= (11.36).
We have used this model to illustrate how the different methods discussed above
work in a particular case. By studying the scattering amplitude in a normal state
we have detected an incipient instability of this system. Then, guided by the kind
of instability, we devised the method(s) to treat the new state which will appear.
There are different ways to describe this new state (the Bogolyubov canonical
transformation, or the introduction of anomalous Green functions and solutions of
the Gorkov equations) which give coinciding results and are actually rather close
physically, although technically different.
Returning to the excitonic insulators themselves, one should say that we have
obtained an insulating state which at first glance is rather different from conventional insulators: we treated it in a rather sophisticated way, and used such terms
as electronhole bound states, exciton condensate, etc. The question arises, are
the physical properties of this state very special?
The model we have used is theoretically indeed very nice and rich. One can
use here all the machinery of modern theoretical physics; there are many different
212
possible types of orderings accompanying this transition. We did not specify them
above, but in general one can get here structural phase transitions, or spin-density
waves see the next section or even orbital antiferromagnetism. However, it turns
out that in most respects such a state is rather similar to the usual insulator. Thus,
there are no supercurrents connected with an excitonic condensate, although
there were attempts to obtain some superproperties. Probably the most interesting
feature of this state is its close proximity to a prototype metallic phase, so that
there may occur here an insulatormetal transition at not too high temperature.
The influence of impurities on such states may also have some specific features.
Nevertheless, most of the properties of excitonic insulators are (unfortunately)
rather similar to ordinary insulators, which nevertheless does not make this model
less attractive and beautiful.

It is instructive at this point to compare the treatment presented above with the corresponding treatment of superconductivity (BardeenCooperSchrieffer, or BCS
theory). Superconductivity is a big field in itself, which is well covered in many
textbooks and monographs see e.g. Schrieffer (1999); we partially follow in
this section the presentation by Abrikosov et al. (1975). Without going into much
detail, we present here only some of the results to illustrate how the general methods described above work in this case. (Actually some of these methods, like the
Gorkov equations, were introduced first in the theory of superconductivity, and
later on borrowed by other fields.)
We start with the electronelectron interaction and first consider the lowest-order
terms in the electronelectron scattering (as in Section 11.4, we use here the form
of Feynman diagrams convenient for the interaction (11.49)):
+
The second diagram can be shown to give the term

mkF 20
(11.46)
ln
2 2
where is the coupling constant (the vertex in these diagrams, see (11.49)), and 0
is a cut-off frequency (since in the BCS theory, devised for conventional superconductors, the actual electronelectron attraction is due to the exchange of phonons,
this cut-off is usually taken equal to the average phonon energy, the Debye frequency D , although it may be different for other situations, with other microscopic
mechanisms of pairing). As this lowest-order loop gives a logarithmically large
2
213
contribution, we have to sum the whole series:

+
which will give us, analogously to (11.35), the effective scattering amplitude, or
electronelectron (two-particle) Green function
(q = 0, ) =
)
*.
F
20 + i
1 + mk
ln
2
2
2
(11.47)
Thus for an attraction between electrons (coupling constant < 0) () has a pole
at = i,
= 20 e2
/||mkF
(11.48)
(cf. (11.35), (11.36)), where is the cut-off frequency. These results are quite
similar to those presented above for the excitonic insulators, the difference being
that in the case of excitonic insulators this pole corresponded to an electronhole
bound state (exciton) whereas here it is a Cooper pair (electronelectron pair)
which exists when the electronelectron interaction is attractive ( < 0).
Further treatment of these two problems is also similar. One of the methods
often used is the Bogolyubov canonical transformation. To take into account the
electronelectron correlations (the formation of Cooper pairs pairs of electrons
with opposite momenta k, k, and opposite spins,6 so that the total momentum q
in (11.47) is zero) one has to keep the averages ck, ck, . Making a corresponding decoupling in the electronelectron interaction term

ck, ck, ck , ck , + h.c.

Hint ck1 , ck2 , ck3 , ck4 ,
=

(k1 +k2 =k3 +k4 )
k1 =k2 =k
k3 =k4 =k
=
(11.49)
we diagonalize the resulting Hamiltonian by the transformation (for simplicity we
drop spin indices):
bk = uk ck vk ck ,
bk = uk ck + vk ck .
bk = uk ck vk ck ,
bk = uk ck + vk ck ,
6
(11.50)
We consider here the simplest case of singlet s-wave Cooper pairs. This is the situation in most conventional
superconductors. In principle more complicated types of pairing may exist, e.g. singlet pairings in different
orbital states (thus, such d-wave pairing most probably exists in high-Tc superconductors), or spin-triplet pairing
which is realized in 3 He and probably in superconducting Sr2 RuO4 . According to the general rules of quantum
mechanics singlet pairs may have orbital momenta l = 0 (s-wave pairing), l = 2 (d-wave pairing), l = 4, etc.
whereas triplet pairing with Stot = 1 may exist with odd values of l (l = 1, the p-wave pairing, l = 3 etc.).
214
Note that the transformation (11.50) corresponds to the appearance of anoma
lous averages ck ck
= 0, similar to the excitonic averages c1k c2k in excitonic
insulators, cf. (11.39).
The resulting spectrum is the well-known spectrum of superconductors:

2
2
(11.51)
E(k) = (k ) +
k2 +
2 ,
where k = k and
is the energy gap equal to (11.48). The coefficients of
the Bogolyubov transformation (11.50) have the form

k
1
1
k
2
1+
=
1+
uk =
2
2
Ek
2 +
2
k

1
k
k
1
2
=
1
1
;
vk =
2
2
Ek
k2 +
2
(11.52)
these expressions are very important and often used, e.g. in the treatment of coherence effects in superconductors.
Now we illustrate on the same example the use of anomalous Green functions

and the Gorkov equations. As the anomalous averages ck ck are nonzero in this
case, we introduce the corresponding anomalous Green functions
@ 6
7A
(11.53)
F (r 1 , t1 ; r 2 , t2 ) = T (r 1 , t1 ) (r 2 , t2 ) ,
where , are spin indices. The coupled equations of motion for the normal
Green function G(k, ) and the anomalous Green function F (k, ) have the form
(cf. (11.43)):
( k ) G(k, ) i
F (k, ) = 1
( + k ) F (k, ) + i
G(k, ) = 0 .
(11.54)
Just these equations were originally known as the Gorkov equations (although
now this term is used in a broader sense). Their solution has the form
G(k, ) =
+ k
,
k2
2
where, similar to (11.44),
F (k, ) = i
F (k, )
d 3 k d
(2 )4
, (11.55)
k2
2
(11.56)
(note that, as compared to (11.42)(11.44), here the definition of the functions

F, F , and
differs by a factor of i). From (11.55) we see that we obtain the
215
same spectrum (11.51) (poles of the Green functions (11.55)). The Green functions
themselves can be written in the form
vk2
u2k
+
E(k) + i + E(k) i

F (k, ) = i
E(k) + i + E(k) i
G(k, ) =
(11.57)
where the functions u2k and vk2 are given by the expressions (11.52), i.e. they
coincide with the coefficients of the Bogolyubov canonical transformation.
By putting the expression (11.57) into (11.55), and performing integration over
with the help of residues, we obtain the self-consistency equation for the energy
gap
:
1=
2(2 )3
d3k

,
2
2
k +
(11.58)
from which we finally obtain the energy gap
= 20 e1/ ,
mkF
=
.
2 2
(11.59)
This expression coincides with (11.48), as it should. Here we have again introduced a cut-off 0 in the integral (11.58), which depends on the physical situation
considered. Thus, for the superconductivity induced by the electronphonon interaction (as in most of the conventional superconductors) 0 is the typical phonon
frequency (Debye frequency D ). If pairing is due to some other mechanism of
attraction, 0 may be different, coinciding usually with the characteristic energy of
the intermediate quanta carrying the interaction (e.g. instead of phonons we can
consider electronic excitations excitons, or magnons, etc.).
One must say that the BCS theory of superconductivity is one of the most
successful theories in the modern quantum theory of solids. Many concepts and
methods were actually first introduced here and later on transferred to other fields.
This is true in particular for the use of Green functions for treating different nontrivial situations, where, depending on the specific problem, one has to introduce
different anomalous Green functions and solve the corresponding coupled equations. The treatment of excitonic insulators presented above gives a good illustration
of that. But we can also use this method, e.g. for considering Peierls transitions or
the formation of CDW and SDW states.
216

We have seen above that in certain cases (specifically when the Fermi surface has
nested parts) there may occur a structural phase transition or formation of CDWs
which creates energy gaps at parts or at the whole of the Fermi surface and which can
sometimes lead to metalnonmetal transitions. Similar phenomena occur in twoband systems with coinciding electron and hole Fermi surfaces leading to excitonic
insulators. In these cases we had an effective attraction between relevant particles:
the effective attraction via phonons in one-band models, cf. (11.23), or the usual
(Coulomb) attraction between electrons and holes in the two-band case. However,
even if the resulting interaction is repulsive, the system satisfying the nesting
condition (9.37) will still be unstable, however, not in the charge but in the spin
channel, i.e. not CDW, but a spin superstructure with the corresponding wavevector
will be created, known as a spin-density wave (SDW). The corresponding instability
was first discussed in the one-dimensional case by Overhauser and is sometimes
called by his name.
One can see that the same nesting condition leads to an instability in the spin
channel if one looks, e.g. at the expression for magnetic susceptibility for local
repulsion (9.16):
(q, ) =
0 (q, )
0 (q, )
=
.
1 + U 0 (q, )
1 U2 0 (q, )
(11.60)
(Remember that with our definitions 0 < 0, and 0 = 2B 0 .) One sees that in
one-dimensional systems, as well as in 2d and 3d systems with nesting, we have
the same divergence of magnetic susceptibility in the case of repulsion U > 0 as
we had in the charge response function (dielectric function) for an attraction; this
again signals an instability, here towards the formation of SDW. It will occur in
this case at arbitrarily small repulsion U . In principle the corresponding instability
may exist even in the case of incomplete nesting or even without it; in this case,
however, we will need the interaction strength to exceed a certain critical value, so
that the condition

(11.61)
U 0 (q, 0) > 1
is satisfied. This condition is well known for ferromagnetism (q = 0); as
(q 0, 0) = (F ), see (9.24), it takes the form
U (F ) > 1 ,
(11.62)
which is known as the Stoner criterion of ferromagnetism. In general (q, 0) may

have a maximum at different values of q = Q, and the corresponding instability
would first appear at the corresponding wavevector Q. In the case of nesting, where
217
there is a real divergence of 0 (q, 0), the Stoner-like criterion (11.61) will be
definitely fulfilled at the wavevector q equal to the nesting wavevector, even for an
arbitrary weak interaction.
One can consider the resulting problem similarly to the treatment of the Peierls
transition or CDW (as an example we look below at the 1d case). We start from the
Hamiltonian of the so-called Hubbard model (see Chapter 12) in which one takes
only on-site electronelectron interactions U ni ni . In the momentum representation the model has the form

U
H=
k ck, ck, +
ak ak+q ak ak q .
(11.63)
V

k,
kk q
The spin density at point x is given by the general expression
*
1)
S z (x) = c (x) c (x) c (x) c (x)

2
*
1 )
=
ck ck ck ck ei(kk )x .
2V kk
(11.64)
As the susceptibility (11.60) diverges at q = 2kF , we keep in (11.64) only terms

with k k = 2kF . Then we have
z
*
1 )
S (x) =
(11.65)
ck ck+2kF ck ck+2kF ei2kF x + c.c.
2V k
(The term with k = k 2kF becomes the complex conjugate of the first term
in (11.65) after a change of the summation index k k + 2kF .) Denoting

1
ck ck+2kF ck ck+2kF = s = |s|ei

V k
(11.66)

S z (x) = 12 Re sei2kF x = |s| cos(2kF x + ) .
(11.67)
we obtain

Making the corresponding mean field decoupling in the Hamiltonian (11.63) (keeping again terms with q = 2kF ) we obtain

H=
(11.68)
k, ck, ck, +
ck+2kF , ck, + h.c. ,
k,
U
where
= N
S z .
Diagonalizing (11.68), we obtain the new energy spectrum

Ek = (k )2 + |
|2 .
(11.69)
218
kF
kF
Fig. 11.21
Density of
electrons
Density of
electrons
Total spin density
Fig. 11.22
Fig. 11.23
All these results are quite similar to those obtained earlier for the Peierls transition
or excitonic insulators; we obtain the spectrum with the gap at kF , Fig. 11.21 (in
principle there will also be smaller gaps at 2kF , 4kF , etc.). One can visualize the
resulting structure as two density waves, for electrons with spin and spin , but
being in antiphase, so that the total charge density remains uniform, and the spin
density oscillates with the period l = 2/2kF , see (11.67) and Fig. 11.22. Such
spin-density waves are called sinusoidal SDW.
There exists yet another solution, the so-called helicoidal SDW, in which the
magnitude of the spin remains constant, but its direction rotates in the xy-plane,
S x = |S| cos(2kF x + ), S y = |S| sin(2kF x + ), see Fig. 11.23. Usually this
structure is energetically more favourable, because the length of spins at each
site here remains the same, which, e.g. optimizes the exchange interaction (6.21).
Sinusoidal SDW may be stabilized if there is strong uniaxial anisotropy in the
system. Note also that the plane of spin rotation may lie in different directions in
the crystal and should not be perpendicular to the spiral axis: in the absence of
spinorbit interaction the axes in spin space have nothing to do with the real crystal
219
electron pocket around -points of the

Brillouin zone
H
hole pocket around H-point

Fig. 11.24
2k
(p
,p
ky
kx
Fig. 11.25
axes! (This is true for all magnetic systems, including those with localized spins,
discussed in Chapter 6.)7
One can also study thermodynamic properties of this model. It turns out that in
the mean field approximation there will be a phase transition in which SDW and the
energy gap in the electron spectrum simultaneously disappear. The corresponding
formulae are very similar to those in the BCS theory of superconductivity.
A similar treatment can be carried out also for higher dimensions with nesting.
Experimentally SDW of this kind were observed in a number of quasi-one-dimensional organic compounds. But probably the most important and the best-studied
example is the magnetic structure of the metal Cr. The explanation of the magnetic structure of Cr was provided by this model (or rather by its two-band analogue, similar to the model of excitonic insulators), with the Fermi surface having
nearly nested electron and hole pockets, schematically shown in Fig. 11.24. Similar
physics seems to work also in many recently discovered FeAs superconductors.
It is also instructive to consider a prototype higher-dimensional system with
nesting the two-dimensional square lattice in the tight-binding approximation.
As shown in Fig. 9.4, for the half-filled band (n = 1) the Fermi surface is a
square, i.e. it is perfectly nested with the wavevector Q = (, ), see Fig. 11.25.
Consequently in the case of repulsive interactions SDW will appear here with the
wavevector (, ) which will open a gap in the whole Fermi surface, i.e. the system
7
Thus the spins can rotate in the plane containing the wavevector of the spiral; the resulting structure may be
called a cycloidal spiral. Interestingly enough, cycloidal spirals give rise to ferroelectricity (Katsura, Nagaosa
and Balatzky; Mostovoy), i.e. the corresponding magnets will be multiferroic.
220
will become insulating. But an SDW with this wavevector is nothing else but the
usual two-sublattice antiferromagnetism, . Thus one can obtain this state
both in the model of localized electrons (localized spins) with antiferromagnetic

exchange, or proceeding from the band picture with appropriate conditions (one
electron per site, full nesting).
There exist many systems with this type of magnetic ordering. Probably the bestknown nowadays is the insulating La2 CuO4 , the parent compound for high-Tc superconductors. Usually the insulating and antiferromagnetic structure of La2 CuO4 is
explained in the model of localized electrons, starting, e.g. from the Hubbard model
with strong coupling, see below, Chapter 12. However, if we did not know that the
electrons in La2 CuO4 were localized, and tried to describe the electronic structure
of this material in the usual band theory, using the tight-binding approximation, we
would obtain exactly the situation described above a metal with perfect nesting
and with the Fermi surface shown in Fig. 11.25. But then this situation would
be unstable, SDW would appear, and we would end up in an antiferromagnetic
insulator of exactly the same type as in the usual description! Thus these two
approaches have much in common, and many features are similar, although there
are also important differences, such as the value of sublattice magnetization, which
in the weak-coupling approach used here may be (much) less than the nominal
spin 12 , and also in the behaviour at finite temperatures: the SDW theory predicts
an insulatormetal transition when the magnetic ordering disappears, whereas in
the localized picture the system remains insulating (Mott insulator) even in the
paramagnetic phase, see Chapter 12.

In the previous sections we have discussed several cases of instabilities of normal
metals, many of which we could describe as the formation of charge or spin
superstructures (Peierls distortion, CDW, SDW, excitonic insulator formation). We
discussed these cases separately, but actually in all these cases we are dealing with
a very similar situation: due to specific features of the energy spectrum, typically
with nesting, the system is unstable towards the formation of a novel state with
electronhole
described by the appearance of anomalous averages

correlations,
of the type ck ck+ Q , either describing the modulation of charge density, or spin
density; see equations (11.16), (11.39), (11.65). What then is the relation between
these seemingly different phenomena?
The situation with CDW seems to be the simplest. Let us consider the situations
with density modulation (CDW, Peierls distortion); spin superstructures can be
(a)
221
(b)
Fig. 11.26
discussed similarly. Here we describe a new ordering by the anomalous average

= ck ck+ Q
= 0 ,
(11.70)
where we either assume summation over spins, or, for simplicity, consider spinless
fermions. The vector Q is determined by the nesting condition; let us consider,
as an example, the half-filled one-dimensional band treated in Section 11.1, for
which Q = . As discussed in that section, the appearance of the anomalous
average (11.70) leads to dimerization of the system and to the opening of the gap
at the Fermi surface.
Generally speaking, the average (11.70), for a fixed value of Q (e.g. equal
to ) can still depend on the momentum k. It can also have an arbitrary phase, in
particular it can be real or imaginary. It turns out that all these possibilities in fact
describe different physical properties of the resulting state.
Let us first take
constant and real,
(k) =
0 . We can calculate (the modulation of) the electron density at site n in the state with this order parameter:

1 ikn ipn
e
e
ck cp .
n = cn cn =
N k,p
(11.71)

With the only nonzero average ck ck+Q =
0 , we immediately get that the electron
density will oscillate as
n =
0 ein =
0 (1)n .
(11.72)
Thus this solution describes a site-centred CDW the electron density alternates
from site to site, see Fig. 11.26(a). As the order parameter is here a constant, it
may be called an s-wave CDW (analogous solutions in the 2d and 3d cases would
be spherically symmetric).
How then can one describe Peierls distortion and the modulation of the electron
density connected with it? It is clear that in this case all ions remain identical
and have similar charge, but the bonds are no longer identical: some of them
become shorter, and some longer. Consequently we expect that the electron density
will increase at the short bonds and decrease at the long ones, see Fig. 11.26(b).
(Actually we may think of the process of Peierls distortion as a first step in forming
molecules, like forming H2 molecules out of a row of hydrogen atoms.)
222
Fig. 11.27
One can show that we can describe this process if we take the anomalous
average (11.70) in the form
k = i
0 sin k .
(11.73)
One can easily check that in this case the electron

density at each site remains the
same, n = 0. But the bond average n,n+1 = cn cn+1 + cn+1 cn is now nonzero.
Check this making
transform of this bond density and taking the
a Fourier

anomalous average ck cp = i
0 (p k ).
As a result we obtain exactly what we wanted: with the order parameter (11.73)
we see that the bond density alternates from bond to bond, n,n+1 =
0 (1)n ,
Fig. 11.26(b). Thus such a bond-centred CDW, associated with a Peierls transition,
is described by the imaginary k-dependent order parameter (11.73), which because
of its sine-like k-dependence may be called a p-wave CDW.
Similarly, one can also consider other possible types of k-dependence of the
excitonic-like anomalous averages (11.70). Thus, e.g. in the 2d case with a square
lattice, the order parameter
k
0 (cos kx cos ky ), which may be called a
d-wave CDW, in fact describes orbital currents running around plaquettes, see
Fig. 11.27. This state is better known as a flux phase there will be magnetic fluxes
alternatingly piercing these plaquettes. Similar states were discussed in connection
with high-temperature superconductivity, with currents running on CuOO
triangles (C. M. Varma). Analogously, if we consider other spin structures of the
averages (11.39), (11.70), we would obtain either states with different spin structures (site-centred CDW of different kinds, see above), or states with spin currents.
11.8 Weakly and strongly interacting fermions. Wigner crystallization

Yet another situation in which the usual description of electrons in metals (ordinary
band theory, the Fermi surface, etc.) breaks down is the case of strong electron
electron interactions. We will mostly discuss this topic in the next chapter, but here
11.8 Weakly and strongly interacting fermions
223
we want to give a general consideration of when we should consider an electron

system as weakly or strongly interacting and which qualitative consequences we
can expect.
It turns out that the answer strongly depends on the character of the interaction
we are dealing with. Consider first the case of short-range interactions; this may
not be very realistic for electrons which always experience long-range Coulomb
interactions, but we can meet this situation in other Fermi systems such as 3 He or
nuclear matter.
In all these cases we cannot simply take the interaction as weak. Thus, e.g.
both in 3 He and for neutrons in nuclear matter or in neutron stars, where particles
interact weakly at large enough distances, the interaction definitely becomes very
strong when they come close together: there is always what we may call a hard
core repulsion. But despite this in certain situations we can consider such a system
as effectively weakly interacting, like the so-called nearly ideal Fermi gas. The
condition for this is that the radius of the interaction (e.g. the radius of hard cores)
r0 is small compared to the average distance between fermions rs (V /N)1/3 =
n1/3 , see (7.5). This will definitely be the case for low-density fermion systems
with short-range interactions.
When the condition

r0 rs
V
N
1/3
(11.74)
is satisfied, we have simultaneously the inequality

kF r0 /
h1
(11.75)
1/3
)
h, see (7.3).
for characteristic values of the momentum k kF , because kF ( N
V
This means that we are dealing only with slow fermions. It is known in quantum
mechanics that for the interaction (collision) of slow particles we can replace the
interaction matrix element by the scattering amplitude, which for small momenta
describes only s-wave scattering and which tends to a constant (a), called the
scattering length:
a =
M
4
h
U0 ,
U0 =
U (r) d 3 r .
(11.76)
The actual small parameter which can be used in perturbation theory is

kF a
1.
h
(11.77)
224
In terms of diagrams this corresponds to the summation of the so-called ladder

diagrams. We want to describe the scattering of two particles,
.
The lowest (Born) approximation corresponds to the diagram
where the dashed line describes the interaction U . When we have a low-density
system, the main effect is the multiple (repeated) scattering of the same pair of
particles, which is described by the ladder diagrams
+
+ .
All other processes will be connected with the creation of electronhole pairs, e.g.
,
or
which is equivalent to
,
and for low density they can be neglected. Such a ladder summation leads to
effective replacement of the interaction U by the scattering amplitude which for
a contact interaction is
Ueff =
U
.
1 + U(F )
(11.78)
Thus even for very strong (hard core) local repulsion the effective vertex
remains finite, equal to a constant a (the scattering length),
and for low density one can carry out all these calculations. It turns out that in this
limit the Fermi liquid (or, rather, weakly interacting Fermi gas) picture is valid,
225
even far away from the Fermi surface. I will only give here a couple of results
which are obtained in this limit:
The ground state energy of the system is (K. Huang, C. N. Yang)
4
5

10 kF a 4(11 2 ln 2) kF a 2
3kF2
1+
+
.
E0 = N
10m
9
h
21 2
h
The effective mass is (A. A. Abrikosov, I. M. Khalatnikov)

m
kF a 2
8
(7 ln 2 1)
=1+
m
15 2
h
(11.79)
(11.80)
(cf. the case of Coulomb interaction (9.46)).

We see that these expressions are the first terms of a series in the small parameter
h) (11.77). From the treatment given above it is clear, in particular, that
(kF a/
one should not expect any instabilities for a low-density system with repulsive
interactions. In particular one should not expect magnetic instabilities which could
be present in systems with repulsion: e.g. the
(11.62) will not be
(
B' Stoner criterion
satisfied, as we should replace there U by U 1 + U(F ) . Note that the situation
with an attraction is drastically different: in this case, as we know, there may
exist two-particle bound states, and as a result the system becomes unstable for
arbitrarily weak attraction; this is the superconducting instability.
Let us now turn to the case of long-range Coulomb interactions. It turns out
that the situation here is just the opposite. Whereas for short-range interactions
the low-density limit corresponds to a weakly interacting regime, for the Coulomb
interaction, in contrast to that, the low-density situation corresponds to the strongcoupling case, and the high-density system behaves as weakly interacting. This can
be understood as follows (see also Chapter 3):
We should compare the average potential energy
Epot =
e2
rs
(11.81)
with the average kinetic energy of the electrons

Ekin =
h
,
mrs2
(11.82)
where rs is the average distance between electrons. The system can be treated as
weakly interacting if Epot < Ekin , or
2
rs <
h
= a0 ,
me2
i.e.
rs =
rs
<1,
a0
(11.83)
226
where a0 is the Bohr radius. In the opposite limit, rs > 1, the interaction becomes
comparable to or bigger than the kinetic energy, and one can expect that the standard Fermi liquid description breaks down. This indeed happens: it was shown
by Wigner that the better state, the one with lower energy, is reached if electrons, instead of being in a liquid state filling the Fermi surface, were localized
in space, forming the so-called Wigner crystal. By doing this we increase the
average kinetic energy of the electrons (due to the uncertainty relation, when
the position of an electron is confined to a certain volume rs3 , its momentum
and corresponding kinetic energy will increase). However, by keeping electrons
apart, as far away from each other as possible, we gain more in potential energy,
if rs > 1.
What is the critical value of rs , beyond which a Wigner crystal becomes stable,
is still a matter of debate. Different calculations give for 3d systems different
values, from rs 10 to 170; the most probable value seems to be rs 80.
There were many attempts to observe Wigner crystallization experimentally. Up
till now it was observed for ions at the surface of liquid He, and there were reports
about observation of Wigner crystallization of electrons in some semiconducting
structures with low electron concentration (the conditions for Wigner crystallization
are less stringent in the two-dimensional case, to which both these examples actually
belong). A phenomenon very similar to Wigner crystallization was also observed in
several transition metal and rare earth compounds with mixed valence: magnetite
Fe3 O4 , Eu3 S4 , La0.5 Ca0.5 MnO3 . The best-known example is magnetite, in which
the ions Fe2+ and Fe3+ order below 119 K (this transition is known as the Verwey
transition). This ordering can also be visualized as ordering of the extra electrons
on Fe2+ as compared to Fe3+ . However, it is not really clear whether the main
driving force for this ordering is in this case the Coulomb interaction or, e.g. the
electronlattice interaction.
One can easily calculate the energy of a Wigner crystal. Crudely it is given
by an expression similar to equation (11.81) with the coefficient depending on the
detailed type of the lattice formed. The calculations give for the best of such lattices
the value (in Rydberg)
EW =
1.8
Ry .
rs
(11.84)
By using two limits, that of a high density (9.47), and the low-density limit giving
Wigner crystals, one can attempt to make an interpolation for the total energy or
for the correlation energy Ec defined (see the definition of Ec in Chapter 9, after
equation (9.47)) as the difference between the exact and the HartreeFock values.
There are several interpolation formulae of this type suggested, e.g. the formula of
227
E
dense electron system
(Fermi liquid)
1
rs
Wigner crystal
Fig. 11.28
T
electron liquid
(metal)
Wigner
crystal
electron liquid
cold
melting
Wigner
crystal
n ~ rs3, or P
rs
Fig. 11.29
Wigner himself:
0.88
Ry.
(11.85)
rs + 7.8
The total energy as a function of rs thus looks as shown in Fig. 11.28 (see (9.47) and
(11.84)) where the dashed line represents the interpolation line (11.85). There is a
certain danger in this interpolation, stressed by P. W. Anderson: actually high- and
low-density curves cross, which in the general thermodynamic treatment implies
a first-order phase transition. This is consistent with the general statement made
in Chapter 2, that the transition between crystal and liquid should always be first
order. Melting of the electronic Wigner crystal is no exception.
The melting of the Wigner crystal presents a rather interesting situation. As
always we can induce melting by increasing the temperature. But there exists yet
another possibility: the Wigner crystal becomes unstable and melts even at T = 0
with increasing density of electrons, for example under pressure (so-called cold
melting). The resulting phase diagram looks as shown in Fig. 11.29. (This cold
melting of the Wigner crystal presents yet another example of a quantum critical
point, cf. Sections 2.6 and 10.3.)
One can understand the general shape of the phase diagram from the following
simple arguments: at very small density (large rs ) the Wigner crystal is stable at
Ec =
228
T = 0, but its energy is 1/rs , see (11.84). Consequently the temperature needed
to melt it will scale as Tm 1/rs . However, for high density (small rs ) the Wigner
crystal will become unstable even at T = 0, so that Tm should go down. One can
see a certain analogy with the phase diagram of the insulatormetal transition, see
the case of V2 O3 , Fig. 2.15. One can also note the similarity of this phenomenon
with the question of the existence of liquid due to quantum effects as compared to
a crystal state, cf. the discussion in Section 4.4.3 (the discussion of the quantum
de Boer parameter, (4.73), and nearby). In this sense we see that the low-density
Coulomb system behaves more or less as a classical one, whereas the high-density
limit (Fermi liquid) is indeed a quantum liquid.
Three extra remarks are in order here. First, when considering the criteria (rs crit )
for Wigner crystallization, one should be aware of other possible instabilities as
well. Thus, a simple treatment shows that a 3d electron gas may develop magnetic
instabilities at the values of rs smaller than those necessary for Wigner crystallization; one has thus to consider these different possibilities simultaneously.
The second problem concerns possible magnetic states of the Wigner crystal.
When we form a crystal made of electrons, there will be localized spins 12 at each
lattice site which should order in some fashion. Most probably this ordering will
be antiferromagnetic (cf. the next chapter). However, this question has not really
been investigated well enough.
And the third point concerns a possible regular description of Wigner crystallization, or at least the instability of the normal Fermi liquid towards it. When
describing different possible instabilities in the first parts of this chapter, in most
cases we could use (and did use) the standard Feynman diagram technique, looking
for some signatures of instabilities, e.g. in the appearance of imaginary poles in
certain Green functions or response functions (imaginary frequencies of certain
collective excitations). As far as I know there is no similar treatment for the Wigner
crystallization, and it is not clear which response function and which subset of
Feynman diagrams could show the instability towards it. It may well be that this
instability is beyond the standard treatment of most of this chapter: the instabilities
discussed above existed already at weak coupling and were related to the special
features of the energy spectrum such as nesting, and with the corresponding divergences of certain response functions, or susceptibilities. At the same time Wigner
crystallization appears only when the effective coupling reaches a certain critical
strength, i.e. it is actually a signature of strong electron correlations, the main topic
of the next chapter.
12
In the previous section we have already seen that in the case of strong electron
electron interactions, when the average interaction energy becomes larger than
the corresponding kinetic energy, one can expect drastic changes of the properties
of the system. Notably, the electrons will have a tendency to localize, so as to
minimize their repulsion at the expense of a certain increase in kinetic energy.
Materials and phenomena for which this factor plays an important role are now at
the centre of activity of both experimentalists and theoreticians; this interest was
especially stimulated by the discovery of high-Tc superconductivity in which electron correlations play a very important role. But even irrespective of the high-Tc
problem, there are a lot of other interesting phenomena which are connected with
strong electronelectron interactions. These phenomena include electron localization, orbital ordering and certain structural phase transitions, insulatormetal
transitions, mixed valence and heavy fermion behaviour. The very existence of
localized magnetic moments in solids, both in insulators and in metals, is actually determined by these correlations. That is why this is one of the most actively
studied classes of phenomena at present.
Real materials to which one applies the models and the treatment presented in
this chapter are mostly transition metal and rare earth compounds, although general
ideas developed in this context are now applied to many other systems, including
organic materials, nanoparticles or supercooled atoms. The typical situation in transition metal compounds is the one with partially filled d-shells. The corresponding
wavefunctions are rather localized; their spatial extension may be smaller than the
distance between these atoms or ions, especially if we are dealing with transition
metal compounds such as oxides, in which typically there are oxygen ions between
transition metal ions. As the effective hopping of electrons between sites determines the electron bandwidth and their kinetic energy, in these cases we can have
a situation discussed in Section 11.8, in which the kinetic energy is smaller than
the electron interaction. This situation is thus equivalent to a low-density electron
229
230
system, and we can expect that the standard description of electrons as a weakly
interacting Fermi liquid breaks down. What will be the outcome in this case and
what are the properties of corresponding systems constitutes the field of strong
electron correlations.
12.1 Hubbard model

The very idea that in the case of a strong electronelectron interaction one can
expect drastic changes in the properties of the system has already been explained in
the previous section. After the first treatment by Wigner (1937), similar approaches
were applied to this situation by Landau and Zeldovich (1943) and by Mott (1948),
see below. In applications to most of the relevant cases one now usually uses a
somewhat different approach and starts not from free electrons with long-range
Coulomb interactions, but rather from the tight-binding model with only on-site
interactions, the so-called Hubbard model. We consider the system with a fixed
lattice and nondegenerate band. In coordinate space the model has the form

H = t
ci cj + U
ni ni ,
(12.1)
ij ,
where ni = ci ci and the summation in the first term goes over nearest neighbours ij .1 The negative sign in equation (12.1) is here chosen simply for convenience, so that the bottom of the corresponding tight-binding band (12.2) would be
at k = 0 (although in more complicated cases the signs of different hopping matrix
elements sometimes have to be fixed and can modify the results). The limitations of
this model (omission of longer-range interaction, etc.) are of course rather severe;
however it turns out that even this seemingly so simple model describes very rich
physics and is at present far from being completely understood.

There are essentially two parameters in the Hubbard model (12.1). These are the
interaction strength (dimensionless parameter U/t) and the electron concentration,
or band filling n = Nel /Nsite . Thus, e.g. the often studied case of one electron per
site n = 1 corresponds to a half-filled band.
The main virtue of the Hubbard model is that it permits us quite naturally to
describe two opposite limits: that of weakly interacting electrons U t, and the
case of strongly interacting, or strongly correlated electrons, U t.
1
More generally we can write down the first term in equation (12.1) as ij, tij ci cj , where tij describes
hopping between sites i and j which are not necessarily nearest neighbours. For nearest-neighbour hopping we
usually take tij = t, as in equation (12.1).
231
Fig. 12.1
Fig. 12.2
In the first case one expects that the standard Fermi-liquid picture will be valid,
at least in the three-dimensional case. In this situation we can treat the interaction
term in (12.1) as a small perturbation and use all the techniques developed in
Chapter 9.
Consider now the opposite case of strong interactions, U t, for a half-filled
band, n = 1. Formally the standard band theory would give a metal even in this
case: the first term in (12.1) (kinetic energy) will in the momentum representation
have the form
Hkin =
k ck, ck, ,
k = t(k) = 2t(cos kx + cos ky + cos kz ) ,
k,
(12.2)
where t(k) is the Fourier transform of the hopping matrix elements tij , which for
nearest-neighbour hopping and for a simple cubic lattice gives the spectrum (12.2).
Thus we have here a simple tight-binding band, and for n = 1 this band will be halffilled, i.e. we would have a metal, irrespective of the distance between the ions and of
the value of the hopping matrix element t and of the corresponding bandwidths W =
2zt (z is the number of nearest neighbours), which may be very small! However,
this is a rather unphysical conclusion: one can argue that for a sufficiently narrow
band t U (which is definitely the case if the lattice parameter, i.e. the distance
between sites, is much larger than the radius of corresponding orbitals) the electrons
will be localized at each site, and there will be no metallic conductivity. Indeed,
such a localization of electrons helps to minimize the Coulomb repulsion U at the
expense of the kinetic energy t, and for n = 1 and U t such a state is clearly
preferable. (The situation here is very similar to the case of the Wigner crystal,
cf. Section 11.4.)
One can depict such a state as in Fig. 12.1, where we have chosen certain but
arbitrary directions of the spins of the electrons ( or ). We see that if we start
from this state, then to create charge-carrying excitations we have to transfer one
electron from its site to another, Fig. 12.2. After such an excitation is made, the
electron (doubly occupied site) and the hole (empty site) will already propagate
232
freely (if we ignore initially the background magnetic structure). This will give an
energy gain t (both electron and hole will now be at the bottom of respective
bands (12.2)). However to create such a pair one has to spend the energy U , the
repulsion of two electrons at the doubly occupied site. And if U t, this process
costs energy U t (or rather U W = 2zt, where z is the number of nearest
neighbours), i.e. such a system will behave as an insulator with the energy gap
Eg U 2zt. Thus the resulting state should be an insulator this is the famous
Mott insulator.
Actually in his original publication in 1948 Mott used different arguments
and proceeded from the picture of electrons with long-range Coulomb interactions. He argued that if we start from an insulator, and create electron and
hole excitations, they would be bound to excitons, so that the system would
remain insulating. And only if the concentration of these electrons and holes
exceeds a certain critical value, then the screening of the Coulomb attraction
would be sufficiently strong so that the excitonic bound state would disappear.
(Recall that in quantum mechanics one needs a certain minimum strength of shortrange attraction for the bound state to appear in the three-dimensional case.) The
condition of that is qualitatively that the Debye screening length rD , given by
2
h becomes smaller than the
equation (9.22), rD2 = D2 = 6 ne2 /F 4 e2 n1/3 /
2
Bohr radius a0 =
h /me2 . From this condition Mott obtained his famous criterion
for the transition from an insulator to a metal:
n1/3 a0 0.25 .
(12.3)
(Note here a certain relation of this picture to that of the excitonic insulators,
Section 11.4, in which, however, due to the assumed nesting, an excitonic instability and the transition to an insulating state with a gap appeared at arbitrary
small band overlap and respective electron and hole concentration.) As we see
from this estimate, very narrow bands (large values of the effective mass m and
small Bohr radius a0 ) always favour the formation of such bound states and prevent the transition to a metallic state, so that such narrow-band materials would
remain insulating.2 And only somewhat later, when dealing with transition metal
2
In fact already five years earlier, in a paper published in 1943 (and which largely remained unknown) practically
the same arguments were presented by Landau and Zeldovich [ZhETF 32, 1944 (1943); Acta Phys.-chem.
URSS 18, 194 (1943)] who, in turn, referred to an earlier remark by Peierls. They wrote: A dielectric differs
from a metal by the presence of an energy gap in the electronic spectrum. Can, however, this gap tend to zero
when the transition point into a metal is approached (on the side of the dielectric)? In this case we should have
to do with a transition without latent heat, without change of volume and of other properties. Peierls has pointed
out that a continuous transition in this sense is impossible. Let us consider the excited state of the dielectric
in which it is capable of conducting an electric current: an electron has left its place, leaving a positive charge
in a certain place of the lattice and is moving throughout the latter. At large distances from the positive charge,
the electron must certainly suffer a Coulomb attraction tending to bring it back. In a Coulomb attraction field
there always exist discrete levels of negative energy, corresponding to a binding of the electron; the excited
conducting state of the dielectric must therefore always be separated from the fundamental one, in which the
electron is bound, by a gap of a finite width. This is the same picture as the one first put forth by Mott in 1948,
only the estimate (12.3) had not been made by Landau and Zeldovich.
233
compounds, Mott himself used arguments of the type which we started with, which
became standard after the very important contribution of Hubbard in 1964, who
formulated in a clear way and solved in certain approximations (see Section 12.4
below) the model (12.1) which now carries his name.
Actually the situation described by the Hubbard model with t U is met in
many transition metal and rare earth compounds, typical examples being oxides
such as NiO, CoO. They have partially filled 3d levels, and according to the band
theory they should be metals. Experimentally, however, they are good insulators.
This is explained by the mechanism described above: the radius of 3d orbitals
is much smaller than the metalmetal (e.g. NiNi) distance (3 A),
so
( 0.6 A)
that the effective dd hopping t (which actually goes via an intermediate oxygen)
is much smaller than the on-site Coulomb repulsion U .3
Thus the ground state of the Hubbard model for U t and n = 1 is an insulating
state with electrons localized one at each site. Note that this state is an insulator
of a completely different type from those described by the standard band theory:
in the usual cases the conventional band scheme is applicable, and the material
is insulating, or semiconducting, if the valence band(s) is completely full, and
conduction band(s) empty. The energy gap in these cases is determined by the
interaction of electrons with the periodic lattice potential, and interactions between
electrons do not play a crucial role there, i.e. this conventional gap is obtained
already in the one-electron picture. Here, however, the very insulating character
is determined by the interaction between electrons. This is completely different
physics, and consequently many properties of such states are quite different from
those of ordinary band insulators.
In the situation considered (U t) the theoretical description should be reversed
as compared to the standard one: whereas usually one takes the first term in
the Hamiltonian (12.1) as the zero-order Hamiltonian and treats the interaction
as a perturbation, here we should invert the description and treat the second
term as the zero-order Hamiltonian, the kinetic energy playing the role of the
perturbation:

H0 = U
ni ni ,
i
H = t
(12.4)
ci cj .
ij ,
It turns out that in the oxides of heavier transition metals such as NiO and CoO another process is important:
charge transfer between metal and oxygen, so that just these oxides which are usually cited as typical examples
of Mott, or MottHubbard insulators, are actually insulators of a somewhat different type they are called
charge-transfer insulators (Zaanen, Sawatzky and Allen), see Section 12.10 below. However, many of their
properties are similar to those of Mott insulators, and very often one also uses for their description the Hubbard
model (12.1) (although there are situations for which this reduction is not valid).
234
E = 0
E =
(a)
2t 2
U
(b)
Fig. 12.3
12.3 Magnetic ordering in Mott insulators

The state with localized electrons (Mott, or MottHubbard insulator) is a ground
state of our zero Hamiltonian H0 . However, it is not yet a unique state. There is
still spin degeneracy left: each localized electron (localized spin) can have two
orientations, or , thus the degeneracy is 2N where N is the number of sites.
This degeneracy is lifted by electron hopping (the term H in (12.4)). This term,
treated as a perturbation, in the first approximation creates a nearby electronhole
pair (or a doublonhole pair), with energy U , i.e. it creates a polar state which
lies outside the degenerate ground-state manifold. However, a second application
of H can return us to the nonpolar state, so that the second-order terms in the
perturbation expansion in t/U give nonzero average in the subspace of localized
states. It is these terms which lift the spin degeneracy and give magnetic (here
antiferromagnetic) order.
One can explain the tendency to antiferromagnetism very simply. Consider
two neighbouring sites with one electron at each. There are two possibilities:
their spins may be parallel or antiparallel. The processes of virtual hopping of
electrons (the second-order contributions in H ) giving the energy change E may
be illustrated as shown in Fig. 12.3. There we show the second-order contribution
to the ground state energy for each case. Here t 2 comes because of applying the
hopping term H twice, and U in the denominator is the energy of the intermediate
state, as always in perturbation theory. In the case of parallel spins such hopping is
forbidden by the Pauli principle, thus the corresponding contribution is zero. But
for antiparallel spins it is allowed, and this extra delocalization of each electron
(virtual excursions to neighbouring sites) leads to a decrease of kinetic energy,
according to the Heisenberg uncertainty relation. (Recall that in quantum mechanics
the second-order contribution to the ground state energy is always negative.)
Mathematically one has to calculate the second-order term in perturbation theory
|H E0 1H0 H |, where |, | are states of the 2N -degenerate ground state manifold. In this subspace we obtain the secular equation with the effective Hamiltonian
Heff = H
1
t2
H =
c c c c .
E0 H0
U ij ,, i j j j
(12.5)
235
In our subspace with one electron per site (ni + ni = 1) the electron operators
can be expressed via the spin operators:
ci ci = ni =
1
2
+ Siz ,
ci ci = ni =
1
2
Siz ,
ci ci = Si+ ,
ci ci = Si .
(12.6)
Putting these into (12.5) we obtain that the effective Hamiltonian is4
Heff = const. +
2t 2
Si Sj .
U ij
(12.7)
Check this using commutation relations and the expressions (12.6). Obtain the
value of the constant in (12.7). Express the resulting exchange Hamiltonian through
the projection operators onto the singlet state of the pair of spins i, j , PStot =0 =
1
Si Sj , and onto the triplet state PStot =1 = 34 + Si Sj , where Si Sj is
4
the spin correlation function, Si Sj = 34 for the singlet state (total spin of the
pair Stot = 0), and Si Sj = 14 for the triplet state (Stot = 1).
The result (12.7) indeed confirms our qualitative considerations and shows that
the ground state of our system is antiferromagnetic. Such an exchange mechanism
is called superexchange (sometimes one also uses the term kinetic exchange).
It is the main mechanism of antiferromagnetism in insulators. One can calculate
from (12.7) all standard properties such as spin-wave spectrum, etc. as in Chapter 6.
The Hamiltonian (12.7) also describes thermodynamic properties of our system
when T U . At T = TN t 2 /U the antiferromagnetic order disappears, and
the material becomes paramagnetic. Nevertheless the material remains insulating,
because strong electron correlations are still present and still prevent the formation
of charge carriers. This is an essential difference from the case of spin-density
waves (SDW) which exist in the case of weak interaction due to nesting, see
Chapter 11; in this latter case the material can also be insulating in the SDW state,
but above TSDW , when the SDW disappears, it would become a metal. Here, due to
strong correlations, the situation is qualitatively different (although the description
of the ground state itself may be rather similar in both cases).

The structure and properties of one-particle excitations in the strongly interacting
Hubbard model is highly nontrivial. This is connected with the electronelectron
4
Here, according to our convention of Chapter 6, the summation goes over all site indices i, j independently, i.e.
every bond is counted twice. If one should count every bond only once, the exchange constant in (12.7) would
be 4t 2 /U .
236
interaction, which manifests itself in the strong interaction of electron and hole
excitations with the underlying magnetic structure, or, in other words, interaction
between charge and spin degrees of freedom.
12.4.1 Aproximate treatment (Hubbard I decoupling)
Before discussing these particular questions, we present here the general treatment
of the electron spectrum, using the method of decoupling of the equations of motion
(cf. Section 6.3.1), first used for this problem by Hubbard (it is often called the
Hubbard I decoupling scheme). It will also give us a good opportunity to illustrate
how the method of equations of motion works in general.
Hubbard himself used the so-called double-time Green functions which are
slightly different from the usual Green functions introduced in Chapter 8. For our
purposes it is sufficient to write down the equations of motion for operators.
Let us write down the standard equation of motion for the operators in the Heisenberg representation. For the Hubbard model (12.1) (with arbitrary hopping tij ) we
have

tij cj, + U ni, ci, .
(12.8)
ci, = [ci, , H] =
j
Check this using the standard anticommutation relation for fermions ci, cj, +
cj, ci, = ij .
On the right-hand side of this equation we have a new operator ni, ci, . If we
make a decoupling at this stage, ni, ci, ni, ci, , we would have a closed
equation for ci, which can be solved by Fourier transform. This would correspond
to the HartreeFock (mean field) approximation with the bare electron energies
k, k, + U n = k + 12 U n (for the nonmagnetic case, ni, = 12 n).
However this approximation does not account for the correlation effects; the system
would remain, in this approximation, a metal.
To correct for this deficiency, we have to keep at least the terms describing
on-site correlations, i.e. we keep the term ni, ci, and write down the equation
of motion for this term; we will make a decoupling in similar terms containing
operators at different sites. Thus we write
ni, ci, = [ni, ci, , H]

tij ni, cj, + U ni, ci, +

tij ci, cj, ci, cj, ci, ci, .
=
j
(12.9)
237
w
N states
U
N states
Fig. 12.4
The second term in (12.9) takes this form when we take into account that n2i, =
ni, (ni, is equal to 1 or 0). Now, we make a decoupling in the first term,
ni, cj, = ni, cj, (i
= j !). The last term drops out after a similar decoupling,
and instead of (12.9) we have the equation

tij ni, cj, + U ni, ci, .
(12.10)
ni, ci, =
j
Equations (12.8) and (12.10) form a closed set of two equations for ci, and
ni, ci, . Solving them (e.g. finding ni, ci, from (12.10) and putting it into (12.8),
then performing a Fourier transform) we find the energy spectrum consisting of
two branches (we consider here the paramagnetic state with ni, = ni, independent of the site index i):

2
U + t(k) 1
=
U t(k) + 4U t(k) n
(12.11)
2
2
which for the half-filled band n = 12 n =
1
2
takes the simple form
U + t(k) 1 2
U + t 2 (k) .
(12.12)
2
2
In contrast to the initial spectrum with the single half-filled band, the spectrum (12.12) describes two bands, one centred around = 0 and another at = U ,
see Fig. 12.4. One can show that whereas the original band contained 2N places
(two per site, with spins and ), these new so-called upper and lower Hubbard
(sub)bands each contain (for the half-filled case n = 1) only N states (N places),
so that the lower band will be completely filled, and the upper one empty, see
Fig. 12.4. Thus this solution gives an insulating state, which is the MottHubbard
insulator. One can check that the solution obtained is exact for the case of an isolated atom: indeed it gives two energy levels, = 0 (one electron in the atom) and
= U (two electrons).
=
238
es
Fig. 12.5
One can also show that in general the one-electron Green function for electrons
with spin in this approximation has the form
G (k, ) =
n
1 n
+
.
(k) + (k)
(12.13)
For n = 1 in the paramagnetic phase n = n = 12 n = 12 , and we indeed

obtain that there are two poles in the Green function, the strength of each being 12 ,
which corresponds to the number of possible electron states in each subband
being N instead of 2N. Physically the form of the Green function (12.13) is quite
transparent: in accordance with the general considerations of Chapter 8, it describes
the spectrum of strongly correlated systems, and the probability to add an electron
with spin at site i with energy 0 is equal to the probability that there is
no electron with spin at this site (the term with 1 n in (12.13)). If, on
the other hand, there is an electron with spin at site i, ni, = 1, then when
we add an electron with at this site, it will have energy + U (the term with
n in (12.13)), see Fig. 12.5.
There are several appealing features in the Hubbard I solution; it indeed describes
the insulating state, which appears due to strong electron correlations, and qualitatively its properties are reasonable. However, there are also several drawbacks: the
metallic state for n
= 1 does not obey the Luttinger theorem, so it does not describe
an ordinary Fermi liquid; for n = 1 there is a gap in the spectrum (12.12) for any U ,
including U t, whereas we expect a transition to a metallic state in this case.
Thus, this approximation does not describe the insulatormetal transition (Mott
transition). It is also noteworthy that this is not a variational state: the energy of the
ground state in this approximation is lower than the exact value.
12.4.2 Dealing with Hubbard bands. Spectral weight transfer

One has to be very careful in using the picture of upper and lower Hubbard
bands and the semiconducting-type analogy of Fig. 12.4: these bands should not
be treated as ordinary bands. Thus, as we see from the expression (12.13), the
residues of corresponding parts of the Green function, which describe the capacity
of each subband, i.e. the number of states in them, depend on the occupation of
corresponding states, in contrast to the usual bands, each of which has 2N places.
239
2(N Ne)
Ne
Ne
m
(a)
(b)
Fig. 12.6
Here, when we add electrons to our system, the number of states in each subband
changes, or there occurs a transfer of spectral weight between subbands. This is
illustrated in Fig. 12.6. Suppose that there are Ne electrons in a lattice with N sites.
In this case there will be Ne occupied sites with energy 0 (below the chemical
potential), i.e. there will be Ne ways to extract an electron from the system, e.g.
by photoemission. On the other hand, there will be (N Ne ) empty sites. At each
empty site we can add two electrons with energy 0, with spins and , i.e.
there will be 2(N Ne ) empty states in the lower Hubbard band, with energies
above the chemical potential. Thus altogether in the lower Hubbard band there will
be Ne + 2(N Ne ) = 2N Ne states. And if we add an electron to an already
occupied site, we can do it on Ne sites, each time adding an electron with the
opposite sign. Thus there will be Ne empty places in the upper Hubbard band with
the energy U . In effect, the total structure of the density of states would look
schematically as shown in Fig. 12.6(b), where we mark occupied states below the
chemical potential by the grey region, and unoccupied states are empty. In total
both these bands contain 2N places, as they should. But we see that, as we change
the number of electrons Ne , the capacity of each Hubbard band, i.e. its spectral
weight, changes; only at Ne = N, i.e. n = Ne /N = 1, would the total weights of
the lower and upper Hubbard bands be equal. This spectral weight redistribution is
a very characteristic feature of systems with strongly correlated electrons. It shows,
in particular, that it may be very difficult in this case to isolate only low-energy
degrees of freedom, as one does, e.g. in the Fermi-liquid theory; there may be a
significant admixture to them of the higher-energy states with energies U . Such
effects are indeed seen in many systems with strongly correlated electrons, for
example in their optical properties.
12.4.3 Motion of electrons and holes in an antiferromagnetic background

One of the most important drawbacks in the original treatment of Hubbard presented
in Section 12.4.1 is the assumption of a nonmagnetic ground state, whereas we have
seen that actually for n = 1 and U t the ground state is antiferromagnetically
240
Fig. 12.7
ordered. This factor can significantly influence the motion of electrons and holes
and it modifies the one-particle excitation spectrum.
Qualitatively one can understand this from Fig. 12.7: if we start from the simple
Neel configuration and add one electron to it, e.g. with spin , then this extra electron cannot hop onto nearest neighbouring sites because there are already electrons
with the same spin on these sites, and the Pauli principle forbids this. This hopping
would be allowed if there are spin deviations at neighbouring sites of a thermal
or quantum nature. Thus we can expect that the presence of an antiferromagnetic
background strongly hinders the motion of current-carrying excitations electrons
or holes.
The simplest way to treat this problem is the following (L. N. Bulaevskii and
D. Khomskii, 1967). Let us start from the antiferromagnetic ground state |0
and consider one extra electron (or hole) at site i, |i = ci, |0 . Such states are
degenerate, and we seek a solution in the form

ai ci, |0 .
(12.14)
| =
i
From the usual Schrodinger equation

H| = E|
(12.15)
we find, with (12.4),

ai j |H |i = (E E0 ) j |j aj
(12.16)
(the wavefunction |i is an eigenfunction of H0 with the energy E0 = U ). The

matrix elements of H entering (12.16) may be expressed through spin operators
via (12.6), and we get

cl, cm, ci, |0 = t 14 + 0|Si Sj |0 . (12.17)

j |H |i = t 0| cj,
lm,
Taking into account also the normalization of the wavefunctions |i, given by
i|i = 0|ci, ci, |0 =
1
2
0|Siz |0
(12.18)
241
Fig. 12.8
and putting (12.17), (12.18) into equation (12.16), we obtain finally the spectrum
of the extra electron added into the antiferromagnet with n = 1:
E(k) = U + t(k)
1
4
+ S0 S1

.
1
2
S
4
(12.19)
Here we have denoted the nearest-neighbour correlation

function 0|Si Sj |0
z

as S0 S1 , and the sublattice magnetization 0|Si |0 = S. In the mean field
approximation S0 S1 = S 2 , and the spectrum of electrons (12.19) takes the
form

(12.20)
E(k) = U + t(k) 14 S 2 .
We can obtain a similar equation also for a hole in the otherwise half-filled lower
Hubbard band, which will have the same form as (12.20) without the constant
term U .
We see that as a result of antiferromagnetic ordering the motion of charge carriers
is indeed severely
hindered: the electron and hole bands significantly narrow,
1
k = t(k) t(k) 14 S 2 where S
< 2 (only slightly less in the 3d case due to
quantum fluctuations, cf. Chapter 6). Thus we see that there is very strong interplay
between electron or hole motion and the underlying magnetic structure, so that
charge transport may be strongly reduced or completely suppressed. One sees also
that above the Neel temperature, where the sublattice magnetization S in (12.20)
disappears or the spin correlator S0 S1 in (12.19) is strongly reduced, electron
and hole bandwidths may strongly increase, and conductivity may be significantly
enhanced. This factor may play a certain role in the insulatormetal transition
from an insulating antiferromagnet to a paramagnetic metal (L. N. Bulaevskii and
D. Khomskii, 1970); such transitions occur in many transition metal compounds,
e.g. V2 O3 .
There is one drawback in the treatment given above. In the picture described
above, see Fig. 12.7, we have considered the motion of an extra electron added
to the system. However, there is essentially no difference between extra and
own electrons. Thus, instead of hopping of the extra electron which is forbidden
by the Pauli principle, an own electron may hop instead at each new step, see
Fig. 12.8. It may seem that this process would lead to a free charge transfer on the
242
etc.
Fig. 12.9
~J
R
R
Fig. 12.10
antiferromagnetic background. However, the actual situation is more complicated

and much more interesting.
In the one-dimensional case this is what does indeed happen, and we have a
separate spinon (a pair of reversed spins at the point where we started this process)
and a holon (or, in this case, a doublon an extra electron, forming one doubly
occupied spin singlet site), which could propagate along the chain. But if we were
to carry out this process in two- or three-dimensional cases, we would see that the
charge carriers could indeed move by this process, but they would leave a tail
along their trajectory, a trace of wrong spins: each of the sites through which the
hole or electron has passed will have its spin reversed; each such reversal would cost
exchange energy J = 2t 2 /U , see Fig. 12.9 where the wrong bonds are marked
by thick wavy lines. Thus the farther away from the initial position the electron
moves (nowadays, in connection with high-Tc superconductivity, one speaks more
often of holes), the bigger is the energy loss: it will be proportional to J l, where
l is the length of the trajectory (L. N. Bulaevskii, E. Nagaev and D. Khomskii,
1970). As a result it looks as though the hole moves in a potential which increases
at least linearly with the distance R between the sites, see Fig. 12.10 (if we chose
the shortest straight trajectory leading from point 0 to R). Thus there will be
a constant force pulling the hole back to the origin, which leads to a confinement
of the hole (the hole is localized in the vicinity of its initial position). The actual
motion of the hole may be visualized as a wandering with predominant return
to the origin,
, and at each step the spin direction is reversed, so that as a
result in this local region the antiferromagnetic order will be strongly disordered or
243
Fig. 12.11
modified. The situation somewhat resembles a polaron state, and it may be called
a magnetic polaron: the electron or hole distorts the antiferromagnetic background
and is itself localized in this distorted region. The detailed form of the magnetic
state inside this region is rather complicated and is actually not well known.
In fact the extra energy of the electron which moved distance R from the origin
is not simply a function of R, as with an ordinary potential: it depends on the whole
trajectory along which the electron has travelled, and it is proportional to J l, where
l is the length of the trajectory, see Fig. 12.11. However, if we take into account
only the least costly, the shortest, i.e. straight trajectory (dashed line in Fig. 12.11),
then we indeed reduce our problem to the motion of a particle in an ordinary
potential, linearly increasing with distance, V (R) J R, as in Fig. 12.10. One can
solve the resulting Schrodinger equation, which for the s-wave is reduced to the
Airy equation. The energy of the ground state is (in a 3d cubic lattice)
E0 = 6t + 9.32 t 1/3 J 2/3 .
(12.21)
Thus only for J = 2t 2 /U 0 does the extra electron or hole reach the bottom
of the free-electron conduction band zt (which it would do in the ferromagnetic
case); for a finite antiferromagnetic exchange there is a certain loss of kinetic
energy5 t 1/3 J 2/3 t 5/3 /U 2/3 .
This notion of hole confinement is very important in the modern theory of
strongly correlated electron systems. The term confinement here is not accidental:
one can indeed establish a close mathematical correspondence with quark confinement in quantum chromodynamics (in particular there the energy also depends on
the trajectory the famous Wilson loops). There are several important implications
of this picture. One of the most spectacular is probably that it gives a mechanism of
electron pairing which may be relevant for high-Tc superconductors. The picture is
qualitatively the following: one electron moving through an antiferromagnet leaves
a trace of wrong spins and becomes localized, which decreases its kinetic energy.
However, if we have two electrons or holes, the second one can move after the first
5
The description presented above is somewhat simplified. This would be more or less true for the case of Ising
interactions between localized spins. In reality, however, in the case of the Heisenberg interaction the damage,
the trace of wrong spins left along the trajectory of the electron or hole, can be repaired due to terms of the type
S + S in the Heisenberg Hamiltonian. Also one can shift a hole along the diagonal of a plaquette by moving
it around this plaquette 1.5 times (Trugmans trajectories). Both these effects, however, are weak and do not
change the main qualitative conclusions.
244
one, repairing the damage done, so that such a motion of the pair would not cost
this exchange energy and the pair would not be confined. As a result the pair would
move freely, and the corresponding gain in kinetic energy may well be the physical
source of pairing in high-Tc superconductors.
12.5 Ferromagnetism in the Hubbard model?
The same physical factors described above were used as an argument that for large
enough U/t one can get ferromagnetic ordering in the Hubbard model. Indeed, the
motion of electrons or holes is completely unhindered if the background magnetic
ordering were ferromagnetic instead of antiferromagnetic. In this case the energy
spectrum is equal to the bare spectrum k = t(k), and the electron can gain
energy W , where W is the bandwidth, W = 2zt (the electron occupies the state
at the bottom of the band, which for a ferromagnet has the full width 2zt, without
narrowing due to spin correlations). Nagaoka has proven that this is indeed the
case if we consider the system with U = and with one extra electron or hole
(Ne = N 1) in a bipartite lattice (which can be subdivided into two sublattices).
This rigorous proof is, however, valid only in this somewhat artificial case; it is not
known whether such a state would exist in the thermodynamic limit, with a small
but finite concentration of holes when the total volume of the system or the total
number of sites goes to infinity.
After the work of Nagaoka the question of the possible existence of ferromagnetism in the simple nondegenerate Hubbard model (12.1) attracted considerable
attention. Very detailed numerical calculations have shown that there may indeed be
a region in the (n, U/t) phase diagram in which the ferromagnetic state is a ground
state. However, the exact boundaries of this region are not really known. They can
also depend on the type of underlying lattice, so that for some lattices, such as
simple cubic or bcc, ferromagnetism may even be absent altogether, whereas for
other types of lattices, e.g. for the fcc lattice, it can exist for one type of doping,
e.g. for electrons, but be absent for holes, or vice versa.
There exist special cases (e.g. with flat bands) for which one can rigorously
prove that the ferromagnetic state is the ground state. But these cases are somewhat
artificial. Thus the question of the existence of ferromagnetism in the nondegenerate
Hubbard model in different situations is still not completely solved (see also the
next section).
Let us discuss the possible states of electronic systems described by the Hubbard
model (12.1) in the whole (n, U/t)-plane. We have seen that for n = 1 the system
245
is insulating and antiferromagnetic for U/t 1. However this state can extend
down to small U for bipartite lattices and for nearest-neighbour hopping: in this
case we have the band structure with the nested Fermi surface (condition (9.37)
is satisfied), and according to the treatment in the previous chapter, an SDW state
will exist even for weak interactions. (In this case the deficiency of the Hubbard I
solution mentioned above the fact that it gives an insulator even for U t is
not really a drawback.) However, if we break the nesting, including, e.g. further
neighbour hopping, the insulating antiferromagnetic state will exist only above a
certain critical value of U/t. An open question here is whether the magnetic order
and the energy gap would disappear simultaneously; there are some arguments that
this need not be the case, and an intermediate metallic antiferromagnetic phase
may exist.
When we go away from n = 1, there will be competition between the kinetic
energy of extra electrons and holes, which does not like antiferromagnetism and
tends to destroy it, and the exchange interaction of remaining localized spins (this
was partially discussed in previous sections). In this situation several possibilities exist: antiferromagnetism may exist up to a certain doping = |1 n|, after
which kinetic energy starts to dominate. Antiferromagnetic order is stabilized by
the exchange interaction J = 2t 2 /U . However, as we have seen, in the antiferromagnetic state we lose the kinetic energy of extra electrons. If we change the order,
e.g. to ferromagnetic order, we gain kinetic energy t|1 n| = t. Thus one can
expect a cross-over to ferromagnetic ordering at
= |1 n|
t
.
U
(12.22)
For U this critical value of doping tends to zero, 0, which is consistent

with the Nagaoka theorem. Nagaoka himself obtained the corresponding criterion
in the form
t
.
(12.23)
0.246 >
U
However, there exist other possibilities as well. One can show that if one considers
only homogeneous solutions, this cross-over from antiferromagnetic to ferromagnetic behaviour occurs not abruptly, but proceeds via intermediate phases. One
such phase may be canted antiferromagnetism, Fig. 12.12 (D. Khomskii, 1970). In
this state two antiferromagnetic sublattices are preserved, but their moments are
canted so that a ferromagnetic component appears. The canting angle increases
with doping , and at a certain critical (close to the Nagaoka estimate (12.23))
/2, i.e. the ordering becomes ferromagnetic. The resulting phase diagram
has the form shown in Fig. 12.13. Yet another possibility is that the intermediate phase, instead of a canted two-sublattice structure, would correspond to spiral
246

a
Fig. 12.12
U/t
ferromagnetic ordering
canted antiferromagnetism
antiferromagnetic state
Fermi liquid?
Fig. 12.13
ordering with the wavevector Q decreasing with increasing ; at = c , Q 0,

i.e. the spiral state becomes ferromagnetic. This possibility may be understood
using the arguments given in the previous chapter in treating charge and spin density waves: for
= 0 the Fermi momentum of noninteracting electrons deviates
from the value kF0 = 2 , 2 , . . . which it had for n = 1 (or = 0), and the SDW
with the period determined by 2kF () may be the most stable solution.
A lot of effort was devoted to the study of the opposite case the limits of
stability of ferromagnetism when the doping becomes large, or the concentration of electrons n becomes small. From the arguments presented in Section 11.7,
cf. equation (11.78), one can conclude that at least in the low-density limit n 0
our system should be in the nonmagnetic Fermi liquid state even for very strong
interactions. If this is true there should be a certain critical concentration nc depending on U/t, below which ferromagnetism should disappear. Different approximate
schemes give for nc (U/t ) the values 0.20.4. However, this question is still
open: as already discussed in the previous section, first of all it is not yet rigorously
proven that ferromagnetism ever exists in the thermodynamic limit (the Nagaoka
state may be just a singular point). There are some arguments that ferromagnetism
may be stabilized only in the presence of orbital degeneracy or of the other bands.
Second, the applicability of the low-density approximation (ladder summation) of
Chapter 11 in the case of very strong interaction can also be questioned: it may
turn out that the results actually depend on whether we first take the limit n 0
and then U , or whether we take these limits in the opposite order. Thus the
247
real form of the phase diagram of the Hubbard model is actually still unknown.
The situation is even more complicated because of the possibility of the existence
of spatially inhomogeneous solutions; see the next section.
In the previous treatment we have considered only spatially homogeneous solutions. However, this is not the only possibility. We have already seen that due
to competition between kinetic energy and antiferromagnetic exchange an electron may become localized in a magnetically distorted region. For example, this
microregion may have ferromagnetic ordering. Such a ferromagnetic microregion
is called a ferromagnetic polaron, or ferron (E. L. Nagaev). Recently the same
object, introduced in the context of high-temperature superconductivity, was called
a spin-bag (J. R. Schrieffer).
One can easily estimate the size of such a ferromagnetic polaron. If we create
a ferromagnetic region of radius R, the loss of exchange energy is 4
R 3 J . The
3
lowest electron energy level in such a potential well is approximately tz + t/R 2 ,
so that the total ferron energy is
E(R) =
4 3
t
R J tz + 2 .
3
R
(12.24)
Minimizing (12.24) in R, we find

1/5
U
,
E0 = E(R0 ) tz + t 7/5 U 2/5 = tz + J 3/5 t 2/5 .
R0
t
(12.25)
Thus the size of such a ferromagnetic region becomes infinite when U , so
that it will occupy the whole sample, in agreement with the Nagaoka theorem.
The general situation can then be visualized as an antiferromagnetic background,
with doped electrons or holes localized in ferromagnetic bubbles created by
themselves.6
Now we can take one step further. If one electron is localized in a ferromagnetic
region, why not two, or three, etc.? In other words, would it not be more favourable
to create one big ferromagnetic region, containing all the doped electrons, leaving
the remaining part of the crystal doping-free and antiferromagnetic? This is the
6
Comparing (12.25) with (12.21) we see that for J t (or t U ) the energy of the ferromagnetic microregion
is slightly higher than that of the string solution (12.21), which makes the ferron state less favourable,
see Fig. 12.14. Thus one should conclude that most probably the distorted spin microregion would not be
simply ferromagnetic; only for J 0 or U may they become degenerate. Nevertheless the picture of
ferromagnetic droplets accounts for at least a part of the physics involved, and it can be used for qualitative
arguments and for crude numerical estimates (mathematically it is much simpler than the picture of confinement
of Section 12.4).
248
E0ferron tz + t 3/5 J 2/5 , see (12.25)

string
E0
tz
tz + t 1/3 J 2/3 , see (12.21)
Fig. 12.14
idea of phase separation in the Hubbard model, first put forth in 1974 by Visscher
and rediscovered later by many researchers, notably by Emery and Kivelson.
One can indeed give simple arguments that this can be the case in the partially
filled Hubbard model. Suppose that phase separation takes place, and let us calculate
the energy of such a state. If we have the average doping concentration = Nh /V =
(N Nel )/V (Nh is the number of holes), and out of the total volume of the
system V the part Vf is ferromagnetic, with all Nh holes (or extra electrons) inside
it, their actual concentration in this region will be f = Nh /Vf = (V /Vf ). The
remaining part of the sample Va = V Vf is antiferromagnetic. The energy of this
antiferromagnetic region is

Vf
.
(12.26)
Eaf J Va = J V 1
V
The energy of the ferromagnetic region consists of two parts: the magnetic energy
+J Vf and the energy of holes or extra electrons in this region, which is tzf Vf +
5/3
tVf f : holes move freely on the ferromagnetic background and occupy the states
from the bottom of the hole band, tz, up to the Fermi energy, tz + tpF2 , with pF3
#p 2
f . Then the energy of electrons occupying the states up to pF is 0 F pm d 3 p
1/3
pF5 /m, which, with m t 1 and pF f , gives the energy presented above. Thus
the total energy of the ferromagnetic region is
5/3
Ef = +J Vf tz Vf f + tVf f
(12.27)
Denoting Vf /V by y and minimizing the total energy Eaf + Ef in y, we obtain

finally (we recall that f = (V /Vf )):
3/5
Vf
U
y0 =
=
,
(12.28)
V 0
t
and the total energy of this phase-separated state is
2/5
t2
t
E0 /V = tz + t
U
U
(12.29)
249
(cf. equation (12.25)). Thus we see from (12.28) that formally the doped system
described by the Hubbard model (with only short-range interactions!) would phase
3/5
separate for all doping = |1 n| <
(t/U ) , until the ferromagnetic fraction
y = Vf /V occupies the whole sample.
One should compare this state with other possible states. The possible homogeneous states are, e.g. the antiferromagnetic one, the ferromagnetic one and the state
with canted sublattices. The antiferromagnetic state is evidently not good: in the first
approximation electrons in it are immobile, see, e.g. (12.19), (12.20). As we have
seen above, the ferromagnetic state may become favourable for the concentration
of extra charge carriers exceeding a certain critical value, which for the homogeneous case is given by (12.23). The energy of the homogeneous canted phase may
be easily calculated using the scheme similar to the one used above, see (12.27),
0 S21 , see (12.19).
but with the effective hopping matrix element t teff t 1/4+ S
1/4S
For the canting angle we find S0 S1 =

2 sin ), S = S z =
and teff = t sin . Putting these expressions into the total energy
14 (1
E canted /V = J S0 S1 teff z + teff 5/3
1
2
cos ,
(12.30)
(cf. (12.27)), and minimizing in , we find for small

sin
U
,
4 t
(12.31)
which gives the critical concentration c for the transition from the canted to the
ferromagnetic state ( /2), a value consistent with (12.23). For < c the
energy of the canted state is
E0canted
t2
U z2 2
=
.
V
U
8
(12.32)
If one compares (12.32) with (12.29), one sees that the phase-separated state
definitely has lower energy than the best homogeneous state the canted one
for small .
From (12.28) we find that with increasing the volume of the ferromagnetic
state increases, and this phase will occupy the whole volume, y = Vf /V 1,
when the doping concentration is
3/5
t
ph. sep.
> c
.
(12.33)
U
Comparing this expression with (12.23), we see that the phase-separated state
survives up to higher values of than the homogeneous antiferromagnetic state.
The energies of different states discussed have the form shown schematically in
Fig. 12.15. Thus we can conclude from these arguments that the phase-separated
250
d cant
~ t/u
c
d ph.sep.
~ (t/u)3/5
c
canted
state
d
ferro.
state
phase-separated
state
Fig. 12.15
state is indeed quite feasible for the partially filled Hubbard model at least in the
strong interaction case U t. Thus instead of the phase diagram of Fig. 12.13
we may have not the canted intermediate phase but phase separation in this (and a
somewhat wider) region.
The treatment given above is of course not rigorous. We have not taken into
account possible local spin distortions close to electrons or holes, and we have not
considered quantum fluctuations. We have also compared the energies of only a few
possible states. Nevertheless the conclusion that there may occur phase separation
in the Hubbard model is very plausible. This is also confirmed by recent numerical
calculations (which, however, are mostly carried out in the so-called tJ model,
which is often used as a substitute for the strongly interacting Hubbard model
see the next section).
In real materials there exists a factor which is missing in the Hubbard model
the long-range Coulomb repulsion between electrons. It is clear that it will strongly
counteract the tendency to phase separation discussed above: the requirement of
electroneutrality is usually very strong and it would prevent large-scale phase
separation. However, such phase separation can still be present if there exist, e.g.
mobile ions in the material which can guarantee electroneutrality; this seems to
be the case in oxygen-rich high-Tc superconductors, e.g. in La2 CuO4+y , where
phase separation is observed experimentally. Although it is not yet completely
clear whether the mechanism of phase separation in this case is the one described
above, this is most probably the case.
Another option could be that the intrinsic tendency to phase separation is still
there, but long-range Coulomb forces will prevent the formation of large charged
regions, and phase separation will be stopped at a certain length-scale charged
regions forming, e.g. small droplets each containing a finite number (30, or 130,
or some other number) of doped electrons and holes (frustrated phase separation
12.8 tJ model
251
of Emery and Kivelson). These phase-separated regions can also have different
shapes (e.g. linear stripes), they can have their own dynamics, etc.
Note also that the homogeneous canted state described above is actually absolutely unstable. Indeed, according to equation (12.32) its energy is a concave
function of the electron density = N/V or of the volume V . Correspondingly,
this state has negative compressibility, 1 = V ( 2 E/V 2 ) < 0, and formally (in
the absence of long-range Coulomb forces) it would be absolutely unstable with
respect to phase separation.
12.8 tJ model
When considering strongly interacting systems, one often uses instead of the Hubbard model the so-called tJ model. In the Hubbard model each site can have
four possible configurations, |0, |, | and |. For strong interactions, U t
and n 1, the last state the doublet has much higher energy than the other
three ( U instead of t). Thus one can project it out and consider only the
states with occupation one (|, |) and zero (|0).7 Of course, as we have seen
above, these doubly occupied states are actually very important: they determine
the antiferromagnetic exchange (12.7), and they participate in the spectral weight
transfer at doping, Section 12.4.2. However, in the superexchange process they
are only virtually occupied; the Hamiltonian (12.7) acts on the subspace of singly
occupied states. Thus we can get rid of these states, taking them into account
only implicitly, in the exchange interaction (12.7). As a result we have to keep
in our model terms of two types: those describing magnetic exchange between
neighbouring sites each having one electron, and those describing the motion of
the holes, if present. The resulting Hamiltonian is written as

H=
tij ci cj + J
Si Sj .
(12.34)
ij
ij
Here we can treat the parameters t and J as independent parameters, instead of t and
U in the Hubbard model. We can even formally consider the situation with J >
t
(although strictly speaking one can obtain the exchange interaction the second
term in (12.34) from the initial Hubbard model only in perturbation theory in
t/U , i.e. in reality J t 2 /U should be definitely less than t).
One should only be careful in dealing with the model (12.34) in that the operators
c , c are strictly speaking not the usual Fermi operators (although they are often
treated as such): one should remember that they describe the motion of electrons,
or rather holes, on the background of other electrons. Nevertheless with careful
enough treatment the tJ model (12.34) may be used to calculate certain properties
7
Technically one can do this, e.g. by using the so-called Gutzwiller projection, see, e.g. Fazekas (1999).
252
(a)
(b)
Fig. 12.16
of strongly interacting systems. This is especially useful in numerical calculations,

because the reduction of the total number of states in this model as compared to
the Hubbard model is very desirable for such calculations.
The possibility of phase separation exists also in the tJ model, and it is even
much easier to see it here, especially for the somewhat unphysical case of large J /t.
Indeed, when we have, e.g. two holes far from each other, we lose z antiferromagnetic bonds around each of them, see Fig. 12.16(a) (in this case four bonds
per hole). However, if we put these holes close together, we lose fewer bonds
(here only six bonds instead of eight), Fig. 12.16(b). Thus if J /t is large, we
gain energy by putting holes together, i.e. we have a strong tendency to phase
separation.

The main physical objects for the application of the Hubbard model are transition
metal compounds. If one wants to make their description more realistic, one has
to take into account the fact that, besides the spins, there exist also orbital degrees
of freedom for d electrons. In atoms the d levels (l = 2) are five-fold degenerate
(l z = 2, 1, 0). This degeneracy is partially lifted in a crystal: e.g. in a cubic
crystal field the five-fold degenerate levels are split into a triplet so-called t2g levels, and a doublet eg . Thus, in an octahedral coordination this splitting is as shown
in Fig. 12.17(a). When one fills these levels by electrons, for instance according
to Hunds rule (i.e. having maximal total spin possible, or, in other words, putting
one electron after the other with parallel spins until they fill all five levels, and
only then starting to fill these levels with electrons with opposite spin), one can
still have situations with orbital degeneracy. This is, for example, the case for Cu2+
(configuration d9 , one hole in a doubly degenerate eg level) or for Mn3+ (d4 , three
electrons with parallel spins in t2g levels and one electron with the same spin in an
eg level), see Fig. 12.17(b).
There exists the well-known JahnTeller theorem which states that the situation
with such degeneracy is unstable (it does not correspond to the energy minimum).
253
eg
t2g
(a)
(b)
Fig. 12.17
u0
(a)
u0
(b)
Fig. 12.18
The only degeneracy permitted in the ground state is the Kramers degeneracy, i.e.
spin-degeneracy, spins and .8 This may be quite simply explained as follows:
suppose we make a certain distortion of the atoms surrounding our transition metal
ion, for example we make an elongation of the oxygen octahedron surrounding
the transition metal ion in an oxide (this is the typical situation in systems such as
LaMnO3 etc.), see Fig. 12.18(a). Due to this distortion there will appear a certain
noncubic potential V which will be a perturbation in our problem. It will be
proportional to the distortion u, V u.
8
Interestingly enough, as Teller himself wrote in the Historical Note in the preface to the book of R. Englman
The JahnTeller Effect in Molecules and Crystals, John Wiley, London, 1972, the idea of the JahnTeller effect
could be attributed to Landau. Teller wrote:
In the year 1934 both Landau and I were in the Institute of Niels Bohr at Copenhagen. I had many discussions.
I told Landau of the work of one of my students, R. Renner, on degenerate electronic states in the linear CO2
molecule. . . . He said that I have to be very careful. In a degenerate electronic state the symmetry on which this
degeneracy is based . . . will in general be destroyed. . . .
I proceeded to discuss the problem with H. A. Jahn who, as I, was a refugee from the German university.
We went through all possible symmetries and found that the linear molecules constitute the only exception. In
all other cases Landaus suspicion was verified. . . .
This is the reason why the effect should carry the name of Landau. He suspected the effect, and no one has
given a proof that mathematicians would enjoy. Jahn and I merely did a bit of a spade work.
Of course Teller underestimated the importance of his contribution: this spade work was extremely useful
and played a crucial role in starting quite a big field. But it is also striking how often the name of Landau appears
in this book: Landau theory of phase transitions; Landau criterion of superfluidity; GinzburgLandau equations
for superconductors; Landau Fermi-liquid theory. And as we see now, his name is also intrinsically connected
with the JahnTeller effect, and also with the insulatormetal (Mott) transitions, see Section 12.2.
254
|x 2 y 2 orbital
|3z2 r 2 orbital
(a)
(b)
Fig. 12.19
It is known that in the degenerate case the splitting of energy levels is linear in
the perturbation, Eel = E0 u. At the same time the change of the lattice energy
at the deformation is quadratic, Elattice = 12 Bu2 (dashed line in Fig. 12.18(b)). The
resulting energy levels as a function of distortion look as shown in Fig. 12.18(b).
As a result there will appear a nonzero distortion at any coupling strength . The
instability towards this distortion is called the JahnTeller effect. Due to coupling
between different sites in the concentrated system we have a cooperative Jahn
Teller effect, which can lead to a structural phase transition which decreases the
symmetry of the crystal and lifts the orbital degeneracy.9 (Note the similarity of
this effect and the Peierls instability discussed in Chapter 11, see (11.6), (11.7).)
After this structural transition the degenerate orbital levels at each site are split in
a certain way, and electrons occupy the lowest orbitals. Consequently we can call
this process orbital ordering.
Different orbitals have different wavefunctions and different distributions
of electron density. Thus the two degenerate eg orbitals have wavefunctions
1 (3z2 r 2 ) = 1 (2z2 x 2 y 2 ) and 1 (x 2 y 2 ) (or their linear combinations).
6
6
2
Correspondingly, in the first one the electron density is predominantly oriented
along the z-axis, whereas in the second one it has the form of a flat cross in
the basal plane, see Fig. 12.19. It is clear that occupation of the orbital |3z2 r 2 ,
Fig. 12.19(a), would cause a tetragonal elongation of the O6 octahedra around such
an ion, like that shown in Fig. 12.18(a): the negative electron density extended along
the z-direction would push away negatively charged apical oxygens O2 (and to
conserve the volume the in-plane oxygens would move in). Vice versa, after such
a distortion the Coulomb energy of this orbital would be smaller than that of the
|x 2 y 2 orbital of Fig. 12.19(b), i.e. the doubly degenerate eg levels of Fig. 12.17
would be split by such tetragonal elongation so that the |3z2 r 2 level goes down,
9
For an isolated centre of this type, e.g. for a JahnTeller impurity, the situation may be more complicated:
because of the degeneracy of several minima, e.g. two minima at u0 in Fig. 12.18(b), there may occur
quantum tunnelling between them. The resulting situation is known as the dynamic JahnTeller effect.
255
and |x 2 y 2 level goes up. For the orbital |x 2 y 2 of Fig. 12.19(b) the situation
would be the opposite: it will be stabilized by the compression of the O6 octahedra along the z-axis.10 (For one hole in eg orbitals, as, e.g. in Cu2+ with nine
d electrons, the sign of the charge on the hole orbital and corresponding distortion
would be reversed, so that, e.g. the |x 2 y 2 hole orbital on Cu2+ would coexist
with elongated octahedra.)
When an electron occupies one of these orbitals, the ion acquires a quadrupole
moment; thus we can also say that such an orbital ordering is simultaneously a
quadrupolar ordering (this terminology is often used for similar phenomena in
rare earth systems). Strictly speaking the order parameter for such a transition, in
the sense of Landau expansion, is indeed a second rank tensor the quadrupolar
moment, although in specific cases one often uses different, technically simpler,
descriptions, see below.
Let us consider, for example, orbital ordering for the case of double degeneracy
for U t and with one electron per site, n = 1. In this case each localized electron
will be characterized not only by its spin ( z = 12 ), but also by the index of
the orbital occupied, = 1 or 2. One can map this extra double degeneracy into
an effective pseudospin = 12 , so that, e.g. orbital 1 corresponds to z = + 12 , and
orbital 2 to z = 12 . Orbital ordering in this language corresponds to ordering of
these pseudospins . In different situations it may be, for example, a ferro. orbital
ordering (the same orbital is occupied at each site), i.e. local distortions around
each transition metal ion are the same we have then the total ferro. distortion of
the whole crystal. Or it may be that different orbitals are occupied at neighbouring
sites, and we can then speak about antiferro. orbital ordering.
The mechanisms leading to such orbital ordering may be different. It may be the
electronlattice interaction which was invoked when we discussed the local Jahn
Teller effect, Fig. 12.18: the coupling between local distortions on different sites
will finally lead to a cooperative structural transition and to corresponding orbital
ordering. Or it may be an exchange interaction, similar to the one discussed for
the nondegenerate case in Section 12.3, which would couple orbital and magnetic
orderings. The model describing all such situations is the degenerate Hubbard
model

H=
tij ci cj + U
ni ni JH
( 12 + 2Si1 Si2 ) .
ij ,
i,
(12.35)
10
Besides the Coulomb, or point-charge contribution, there exists another factor, adding to the coupling between
local distortion and splitting of degenerate orbitals with corresponding orbital occupation: the covalency
between d orbitals of transition metals and p orbitals of ligands (e.g. oxygens). Usually these factors, Coulomb
repulsion and covalency, work together and add up, stabilizing the same orbitals.
256
Same orbital
same spin
E = 0
(a)
Same orbital
opposite spins
E =
2t 2
U
(b)
Different orbitals
same spin
E =
2t 2
U JH
(c)
Different orbitals
different spins
E =
2t 2
U
(d)
Fig. 12.20
Here , = 1, 2 are the orbital indices. We also included the last term which
describes the intra-atomic exchange responsible for Hunds rule (it is absent in the
nondegenerate case, in which two electrons at the same site may only have opposite
spins).11
Suppose that there exists only hopping between the same orbitals, t 11 = t 22 = t,
12
t = 0. Then we have the following four situations for the pair of doubly degenerate
sites, shown in Fig. 12.20. (These four possibilities replace the two configurations
which existed in the nondegenerate case, cf. Fig. 12.3.) As in the nondegenerate
case, we have here also an extra energy decrease due to virtual hopping of electrons
into neighbouring sites.
Whereas in the nondegenerate case these virtual hoppings provided the mechanism for antiferromagnetic ordering, we see that here the same superexchange will
lead simultaneously to both spin and orbital ordering: in our simple model the third
configuration (c) (same spin/different orbitals) has the lowest energy, because in
this case in the intermediate state two electrons at the same site have parallel spins,
which decreases the energy of this state (the denominator in Fig. 12.20(c)) due
to Hunds rule interaction. Thus, one may expect that in a concentrated system in
such a situation there will be ordering ferromagnetic in spin and antiferro. orbital
ordering. This mechanism of orbital ordering (Kugel and Khomskii, 1972, 1982)
acts simultaneously with that due to lattice distortion (the conventional JahnTeller
mechanism), and there are some indications that in some cases it can even be
the dominant mechanism. (Of course, if there occurs orbital ordering due to the
exchange interaction, the lattice would follow, so that such ordering will always
be accompanied by the corresponding structural change.)
Mathematically one can describe this situation writing down the effective exchange Hamiltonian analogous to (12.7). Here, however, each site is
11
Actually the Hunds rule coupling is not really an exchange interaction: it is rather the difference between
direct densitydensity Coulomb repulsion of electrons with parallel and antiparallel spins. Indeed, due to the
Pauli principle the electrons with parallel spins avoid each other (they have an antisymmetric coordinate
wavefunction), so that on average they are further away from each other and consequently experience weaker
Coulomb repulsion, which decreases the energy of such a state. But phenomenologically one can describe this
effect by the exchange interaction (the last term in equation (12.35)).
257
characterized not only by the usual spin, but also by the orbital index which can be
mapped onto the pseudospin = 12 , as explained above. The resulting Hamiltonian
has schematically the form

H=
JS Si Sj + J i j + JS (Si Sj )(i j ) ,
(12.36)
i.e. it describes both spin and orbital degrees of freedom which are coupled by the
last term in (12.36). (The actual form of this Hamiltonian in real systems may be
more complicated, e.g. anisotropic in operators.)
One of the interesting consequences of this treatment is the possibility of having,
besides the ordinary spin waves, also a new type of elementary excitations: orbital
waves orbitons (they will, however, be strongly mixed with phonons), and
possibly also coupled spinorbital excitations.
Orbital ordering has important manifestations in many properties of corresponding systems, but especially in their magnetic behaviour. Orbital occupation to a large
extent determines the magnitude and even the sign of the exchange interaction
and, consequently, the type of magnetic ordering. This constitutes the essence of
the so-called GoodenoughKanamoriAnderson (GKA) rules see, for example,
Khomskii (2001).
The importance of orbitals for magnetic exchange is already clear from the
simplified Hamiltonian (12.36). Indeed, if for example JS > JS , then, depending
on the type of orbital ordering, characterized by the correlation function i j ,
the sign of the total magnetic exchange JS + JS i j may change.
In general, the GKA rules determine the type of exchange interaction for different
local coordinations and for different orbital occupation. Without going into details,
one can formulate simplified GKA rules in the following way:
(1) If on neighbouring magnetic sites the half-filled orbitals (having one electron)
are directed towards one another, we would have a rather strong antiferromagnetic coupling. (In fact, this case is equivalent to the case of the simple
nondegenerate Hubbard model of Section 12.3.)
(2) If, however, on one site we have a half-filled orbital, but on the neighbouring site the corresponding orbital directed towards the first site is empty or
full (has two electrons), then the exchange interaction between these two
ions will be ferromagnetic, but weaker. This is actually the case schematically shown in Fig. 12.20(c), (d): the ferromagnetic ordering, case (c), is
more favourable, but the energy difference between this and the antiparallel spin configuration, case (d), is small: instead of the conventional antiferromagnetic exchange J t 2 /U , see equation (12.7), here the exchange
(proportional to the energy difference between parallel and antiparallel spin orientations) is J (t 2 /U )(JH /U ) where we have used the expansion in JH /U in
258
denominators of the energies in Fig. 12.20(c), (d) (typical values for transition
metal ions are U 45 eV and JH 0.8 eV).
Interestingly enough, the second GKA rule gives the possibility to get ferromagnetic ordering in magnetic insulators. This is not a common phenomenon:
most often insulators with strongly correlated electrons are antiferromagnetic, and
we usually get ferromagnetism in the metallic state. This is also clear from the
discussion of Sections 12.5 and 12.6: we saw there that ferromagnetic ordering
appears usually when we have partial occupation of corresponding states, n
= 1,
which are metallic (see also below, Section 13.5). And in fact the main mechanism
of making ferromagnetic insulators is that with the appropriate orbital ordering,
described above.
In dealing with real magnetic materials one has to take into account many specific
details, such as the type of crystal lattice, the detailed character of the active
orbitals, etc. Accordingly, the actual GKA rules are much more detailed than the
simplest cases described above, see, e.g. Goodenough (1963) and Khomskii (2001).
But the general conclusion remains the same; specific types of orbital occupation
largely determine the magnetic properties of corresponding systems.

Until now, even when discussing systems with strongly correlated electrons, we
have only considered the correlated electrons themselves for example, electrons of
partially filled d shells in transition metal compounds. When we want to describe
real materials, e.g. transition metal oxides like NiO or La2 CuO4 the parent
compound of high-Tc superconductors we may need to take into account the other
constituent atoms in these compounds, notably oxygen ions. We already partially
touched upon this problem in the previous section when we spoke about crystal
field splitting of d levels due to interaction of d electrons with the surrounding ions.
Since the main objects to which this treatment is usually applied are transition
metal oxides, one often speaks about oxygen ions surrounding a transition metal
ion, see Fig. 12.18(a), although in other cases these may be halogen ions such as
F , Cl , . . . , or S2 , Se2 , etc. The general term used for such anions surrounding
a given metal ion is ligands, and the splitting of d levels by the crystal field due
to interaction with these anions is also called ligand field splitting. However, the
role of these ligands does not reduce only to determining the detailed character of
d-levels; it may be much more significant.
Consider, as an example, transition metal compounds with the general formula
ABO3 , of the type of LaMnO3 or GdFeO3 (they are called perovskites). These systems basically consist of a simple cubic lattice of transition metal ions (Mn, Fe, . . . )
259
Fig. 12.21
Fig. 12.22
with oxygens located in the middle of the edges, see Fig. 12.21. Here crosses are
transition metal ions, and circles are oxygens. A-ions (La, Gd) are located at the
centre of the cube shown in Fig. 12.21. (In real systems there often appear certain
weak distortions of this structure, which, however, are not important for us at the
moment.)
We see that here transition metal ions are separated by oxygen ions, which,
generally speaking, reduce the dd-hopping t in the Hubbard model, and can make
the corresponding system a Mott insulator, with all the consequences thereof. But
this also leads to another consequence: in fact in this case the distance between
d ions is so large that one can practically neglect direct overlap of the d functions
themselves. However, instead we have a relatively large overlap of d orbitals of
transition metals with 2p orbitals, the valence orbitals of oxygen. The hopping of
d electrons from one site to the other actually occurs via intermediate oxygens, see
Fig. 12.22.
In many cases one can exclude these oxygen p orbitals and reduce the problem
to that of d electrons themselves, described by the Hubbard model (12.1) or its generalization (12.35). However, this is not always possible: there may exist situations
in which one has to take into account oxygen p states in an apparent way.
The general model would then be the one which includes both d electrons of
the transition metal ions and p electrons of oxygens, with the hybridization (or
260
hopping) between them. According to our general arguments, for d electrons we

definitely have to take into account the electronelectron (Hubbard) repulsion. As
the p electrons of oxygen have larger radius, one can in the first approximation
neglect the corresponding Coulomb (or Hubbard) repulsion of electrons on oxygens
(but generally speaking one has to include it as well).
Thus the general Hamiltonian describing this situation can be written as

H = d
di di + p
pj pj +
(t(pd)ij di pj + h.c.)
i,
+ Udd

i
j,
ndi ndi + Upp
ij,
npj npj + Upd
ndi npj .
(12.37)
ij,
Here we have used the self-evident notation d , d for d electrons and p , p for
oxygen p electrons. We also consider the simplest case, ignoring orbital degeneracy
of d electrons, and taking into account only one p orbital that directed towards
neighbouring transition metal ions in Fig. 12.22. In real systems one has to use
different p orbitals for different oxygens surrounding a given d ion. We have also
included not only the dd interaction, but also pp and pd repulsion, although in most
of this section we will only take into account Udd , denoting it simply U .
The elementary hopping process described by the Hamiltonian (12.37) is the
transfer of an electron from the filled 2p shell of O2 to the d ion. This costs the
energy
= d p + Udd .
(12.38)
This is called the charge-transfer energy. It can also be defined as the energy
required to go from the initial configuration d n p 6 to the excited configuration
d n+1 p 5 .
Alternatively one could formulate the model in terms of d and p holes. Then
the hopping process would be the hopping of a d hole from a transition metal to
an oxygen. This is sometimes more convenient, and the corresponding description
is usually used, for example, to describe high-Tc cuprates. The corresponding
Hamiltonian would look the same as equation (12.37), but one has to interpret d ,
d, p , p as the creation and annihilation operators not for d and p electrons, but for
d and p holes (and one has to redefine the corresponding energies, see Fig. 12.23).
In the electronic representation the typical energy scheme for the simplest case
with one d electron looks as shown in Fig. 12.23(a), with the charge-transfer
energy marked. One has to remember, however, that this is not a one-electron
energy diagram, and the level d + U is the energy of a state with two electrons on
the same d ion (thus, in the nondegenerate case, necessarily with opposite spins).
An alternative energy level scheme in the hole representation would look simpler,
Fig. 12.23(b). Here the arrow denotes the hole, i.e. one missing electron on the d
ed + U
ed
261
~
ephole
~
e dhole
ep
(a)
(b)
Fig. 12.23
level, and there are initially no holes on the oxygen (O2 with filled 2p6 shell). The
excitation energy will then be simply the energy difference between oxygen and
transition metal hole levels d and p , and
= p d ,
(12.39)
which makes it much simpler to draw and to deal with. Thus we will mostly use
this scheme from now on, using the Hamiltonian (12.37), but having in mind that
it describes holes with corresponding hopping and interactions, i.e. Udd , Upp , etc.
would denote the holehole interaction. (One has to be aware that the definitions of
the hole energies d and p in fact should contain the information on the interactions
included in the Hamiltonian (12.37), which become especially important when we
are dealing with the real situation with possibly many d electrons or d holes at a
site.)
Let us consider again the standard situation with one d electron (or, which is
the same for a nondegenerate d level, with one d hole) per site. Let us first treat
the case when both U ,
1 (recall that here U is Udd for d holes). In this
case the situation would strongly resemble that of the simple Hubbard model of
Sections 12.112.3 in the case of a strong interaction: in the ground state the d
electrons (or holes) would be localized, one at each site, and we would have a Mott
insulator. We will again have localized spins, and the magnetic interaction will
appear when we take into account virtual hopping of d holes, which we can again
treat in perturbation theory.
In contrast to the usual Hubbard model (12.1), however, here virtual hoppings
occur via an intermediate oxygen which is taken into account in an apparent way.
Thus we have to use the hopping term tpd p d in the Hamiltonian (12.37) not twice,
but four times, see Fig. 12.24(a). In this figure we show (in the hole representation)
two neighbouring d sites with an oxygen ion in between, and by wavy lines we show
virtual hoppings of a d hole from one site to the other and back via intermediate
p levels (the numbers near the wavy lines show one particular sequence of consecutive hoppings). Once again, this process, if allowed, decreases the total energy,
262

e~p
~
ep
2
1
4
~
ed
ed
2
4
~e
d
j
(a)
~
ed
j
(b)
Fig. 12.24
but it is allowed only if the spins of neighbouring d ions are antiparallel. Thus in
effect we again obtain the effective antiferromagnetic Heisenberg exchange (12.7),
but the exchange integral will be given by the expression
J =
4
2tpd
2 Udd
(12.40)
Indeed, each of four hoppings gives a factor tpd . After the first hopping the energy
of the intermediate state with one hole on oxygen is d p =
. After the second
hopping this hole moves to the neighbouring d ion, and the energy of this intermediate state is Udd . When the hole moves back, we again have the oxygen hole with
the energy
. These three values of intermediate energies stand in the denominator
of equation (12.40).
Note that we can also rewrite the expression (12.40) in exactly the same form
as the exchange integral (12.7) in the simple Hubbard model, if we introduce the
effective dd-hopping
t = tdd =
2
tpd
.
(12.41)
This is how, it seems, we can reduce the pd model (12.37) to the conventional
Hubbard model (12.1).
However, besides the process described by Fig. 12.24(a), in this case there is
yet another process which can also contribute to the antiferromagnetic exchange.
This process is illustrated in Fig. 12.24(b). At first glance it looks identical to that
of Fig. 12.24(a), but in fact it is different: here after first transferring a d hole to
the oxygen from, say, the left d ion, in the next step we transfer to the same oxygen
another d hole, from the right. In effect the second intermediate state does not have
the energy Udd , but instead we have here two holes at the same oxygen. The energy
of this state is 2
, or, if we take into account the repulsion of two holes on the
same oxygen, 2
+ Upp . Again, as we can put on the same oxygen p orbital only
two holes with opposite spins, this process will also lead to an antiferromagnetic
263
interaction, with the exchange constant

J =
4
4tpd
2 (2
+ Upp )
(12.42)
(the factor of 4 instead of 2 in (12.42) comes from the fact that there are twice as
many different routes in Fig. 12.24(b) than in Fig. 12.24(a): one can interchange
the sequence of electron hops from sites i and j to the oxygens, and of hops back).
The total antiferromagnetic interaction is the sum of both these contributions,
Jtot = J + J , i.e. introducing the effective dd-hopping (12.41), we can write it as

4
2tpd
1
1
2
1
2
= tdd
. (12.43)
Jtot = 2
+
+
Udd
2
+ Upp
Udd
+ Upp/2
Depending on the ratio of the effective charge-transfer energy
and dd Hubbard
repulsion Udd , either one or the other term in equation (12.43) will dominate. If
(or
+ Upp/2 ) is much bigger than Udd , we can keep in (12.43) only the first term,
and then our problem will indeed completely reduce to the ordinary Hubbard model.
In the opposite limit we should rather keep the processes of the second type, with
two oxygen holes, and this regime is indeed somewhat different from the first one.
The difference is not so much in the properties of the ground state: in both these
cases, if only tpd (Udd ,
), the ground state is a Mott insulator, with localized
spins and with antiferromagnetic exchange between them. The real difference is in
fact in the nature of the lowest charge-carrying excitations (here again we follow
our general line: after discussing the type of the ground state we go over to the
excitations). In the case of the Hubbard model such excitations correspond to the
transfer of an electron from one d site to the other, dn (p6 )dn dn+1 (p6 )dn1 , i.e.
to the formation of a d hole and a doubly occupied d state. This excitation costs
energy Udd . The state of the O2 sitting in between remains here the same, (p6 ).
The extra d electron (the state dn+1 ) and the d hole (dn1 ) thus created can now
move through the crystal and carry current.
However, there exists another possibility: if
< Udd , then the lowest chargecarrying excitations would be those with charge transfer between a d ion and an
oxygen: dn p6 dn+1 p5 , which costs energy
. Thus this is in a sense a different
type of insulator: the ground state is in principle the same as in the Mott, or
MottHubbard case, but the lowest charge excitations are quite different; they
correspond to the creation of a doubly occupied d state and an oxygen hole. Such
systems are called charge-transfer insulators. This classification was first proposed
by J. Zaanen, G. Sawatzky and J. Allen, and it is often called the ZSA scheme. One
can illustrate this on the phase diagram of Fig. 12.25 (the ZSA phase diagram). Here
we show different possible states of the general dp model (12.37) as a function of
and U , or of corresponding dimensionless quantities

/tpd , U/tpd (U = Udd ).
264
Upd /tpd
charge-transfer
insulator
MottHubbard
insulator
metal
1
/tpd
Fig. 12.25
If both U,
tpd , we have an insulating state with localized electrons, but now
this phase is separated into two regions: for
U we can exclude virtual oxygen
states and go over to the Hubbard model (12.1), dealing only with d electrons, and
this region is a usual MottHubbard insulator. In the opposite limit U
tpd
we still have an insulator with localized spins and with antiferromagnetic interaction
between them, but the lowest charge excitations are those with the formation of
oxygen holes, and the exchange integral will be given by the expression (12.42).
This is the region of charge-transfer insulators.
In reality, e.g. among transition metal oxides, the tendency is such that the oxides
at the beginning of the 3d series (Ti, V), if insulating, belong to the MottHubbard
class, whereas the oxides of heavier transition metals (Co, Ni, Cu) typically are
charge-transfer insulators.
For large values of
, when we decrease the Hubbard interaction U , we expect
a Mott transition to a metallic state. The situation in the left-hand part of the ZSA
phase diagram of Fig. 12.24 (large U , small
) is much less clear. Most probably
when we reduce
, going to the limit of small and possibly negative charge-transfer
gap, we would also have here a metallic state, but it would be an oxygen metal,
with mobile oxygen holes as charge carriers, but with still strongly correlated
d electrons, hybridized with oxygen states. This situation resembles that of mixed
valence or heavy fermion systems, see the next chapter, and the properties in this
regime may be quite nontrivial. There exist also other options in this regime (e.g.
different types of insulating states); all these questions remain largely open.
The difference between MottHubbard and charge-transfer insulators becomes
especially important if we dope the system by holes. In the MottHubbard regime
these holes would go to d sites, and we would have a situation of partially filled
Hubbard bands, discussed in Sections 12.412.8. In the charge-transfer case,
however, it is more favourable to put an extra hole to an oxygen instead of a
265
d level. Of course there is always some mixing, hybridization between d and p

states, caused by the third term in the Hamiltonian (12.37), so that the total wavefunction of the hole would always be a superposition | = |d + |p, but if in
the MottHubbard regime the main weight in this wavefunction is on the d site,
|| ||, in the charge-transfer case the situation would be the opposite, and in
the first approximation the hole would be predominantly located on the oxygen.
And although even in this case sometimes we can reduce the description to an
effective one-band model, treating these hybridized states as new basis functions
(this is the picture of ZhangRice singlets widely used for high-Tc cuprates), one
has to be aware that the real nature of corresponding states, and consequently some
of their properties, e.g. such as the distribution of electron and spin density (the
magnetic form-factor), details of magnetic interactions, or transport phenomena,
may be rather different in doped charge-transfer insulators as compared to doped
MottHubbard systems. This one has to keep in mind, especially because many
important materials, such as cuprates with high-temperature superconductivity or
colossal magnetoresistance manganites, belong to this category.
Some of the most interesting phenomena occurring in solids are the insulator
metal transitions which exist in certain materials. These transitions can occur due
to doping, as discussed above. But they may occur even in stoichiometric systems
with a change of temperature, pressure, magnetic field, etc.
There are several possible types of insulatormetal transitions.
1. There may occur insulatormetal transitions which can be explained by the
standard band theory, without invoking any special electronelectron interaction.
Thus the system having an even number of electrons per unit cell may be an
insulator or semiconductor simply because the filled valence band is separated
from the empty conduction band by an energy gap, see Fig. 12.26. Such is, e.g.
the situation in typical semiconductors like Ge and Si. The structure of the energy
spectrum in this case is determined by the type of crystal lattice. If there occurs
a structural change, new bands may in principle overlap, so that the energy gap
would disappear, and the material would undergo an insulatormetal transition.
This indeed happens when Ge and Si melt: the short-range order in liquid Ge and
Si is quite different from the tetrahedral coordination in solids (it is actually similar
to the structure of liquid Pb or Sn), and as a result liquid Ge and Si are good metals.
Another well-known example is the case of tin (Sn). It exists in two forms:
ordinary metallic tin (white tin), and the so-called grey tin which is actually the
stable form of Sn at low temperatures (below 13 degrees Celsius) and which is
an insulator (strictly speaking it is a zero-gap semiconductor, i.e. it lies on the
266
Fig. 12.26
borderline between metals and insulators; but in any case it is not a good metal like
white tin). Grey tin has a diamond structure, like that of Ge or Si. The transition of
white to grey tin is accompanied by a large change in the lattice and is known as
tin plague: in the old days sometimes soldiers tin spoons or tin buttons of their
uniforms suddenly decayed, becoming grey powder especially in winter, after
being stored for some time in cold weather. In the Middle Ages tin organ pipes
also sometimes disintegrated in winter, which was thought to be the work of the
Devil. Thus the whitegrey tin transition is one of the earliest known examples of
metalnonmetal transitions.12
2. In a certain sense the Peierls transition discussed in Chapter 11 is similar to
the band structure transition, although we described it completely differently, using
all the powerful methods of modern many-body theory. In essence what happens
there is a change of the crystal lattice such that the new energy bands in the new
crystal potential acquire a gap, this gap appearing at the position of the former
Fermi surface. In this case the metalnonmetal transition is accompanied by the
appearance of long-range order new lattice periodicity.
3. A similar treatment may be given also to insulatormetal transitions accompanying, e.g. spin-density wave formation. An SDW, like the charge-density wave
(CDW), does not necessarily lead to an insulating state, e.g. antiferromagnetic Cr,
in which SDW exists, remains a metal. However, there are also cases in which
the SDW state is really insulating, the SDW energy gap covering the whole
12
Tin plague could lead not only to unpleasant, but just curious events such as transformation of a box of
tin spoons into a box of grey powder. Sometimes it can even have tragic consequences. This is at least one
of the explanations of the tragic fate of the famous polar explorer Captain Robert Falcon Scott and his three
companions. They perished in the Antarctic in 1912 on the way back after reaching the South Pole, and one
of the reasons could have been that in such a cold climate the tin-soldered cans in which they stored their fuel
(kerosene) started to leak because of the whitegrey tin transition. As a result they lost part of their fuel, which
was one of the reasons for the disaster. I am not sure if this is a true explanation, but in any case in his notes,
found after his death, Scott himself wrote that loss of fuel played a crucial role in their fate.
267
Fermi surface. Such is, e.g. the situation in certain quasi-one-dimensional organic
compounds.
An important difference here is that in contrast to previous cases the extra
potential creating the energy gap is not the external potential due to the lattice, but
a self-consistent potential of the electrons themselves. This self-consistent potential
is spin-dependent, and it leads to the splitting of the electron subbands. Thus the
difference is that here the electronelectron interaction is already essential. The
similarity with the first two situations is, however, that the energy gap exists only
to the extent to which there exists new long-range order in the system; above the
Neel temperature, when magnetic ordering disappears, this energy gap would close
and the material would become metallic.
4. And finally there are the situations in which the insulating character of the system is completely determined by the strong electronelectron interactions. This is
the case of MottHubbard insulators.13 In this case the materials are insulating even
without any extra order in the system. Thus, e.g. one can in principle describe the
insulating nature of the ground state of materials such as NiO and CoO as connected
with the antiferromagnetic order which can split the original partially filled (i.e.
metallic) bands and create an energy gap. This treatment would then be similar to
the treatment of SDW systems. It is indeed possible in some cases to carry out such
a programme. Thus the standard band structure calculations, using spin-polarized
bands (e.g. LSDA local spin density approximation) can sometimes give an
13
There exists yet another, fifth case of insulatormetal transitions transition to an insulating behaviour in
strongly disordered systems. Electron localization in this case (called Anderson localization) occurs in systems
of electrons interacting with a random potential (e.g. random impurities); electronelectron interactions do
not play a crucial role here. The physics of disordered systems is a big special field which I will not discuss
here. Suffice it to say that if the disorder is strong enough (for example, if average fluctuations of the random
potential 2 1/2 exceed the bandwidth), electrons become localized and do not go to infinity as t ,
which means the absence of metallic conductivity. The electron density of states in this case remains finite at
the Fermi level, i.e. the energy spectrum does not have an energy gap, but the electron mobility is zero there
exists a so-called mobility gap.
The proof of the Anderson result (Anderson localization) is rather involved, but one can qualitatively explain
what is going on using the following arguments. When disorder in a metal increases, the electron mean free
path l decreases. However, it is hardly possible to imagine the situation when, e.g. l becomes less than the
interatomic distance a; it is clear that the standard description of electron transport (Boltzmann equation, etc.)
would fail in this case. One can show that the conductivity in this case (l a) will be
= e2 /3ha .
(12.44)
The limit of the standard description of conductivity related to the condition l a is known as the IoffeRegel
limit. If disorder is so strong that formally we would have l < a, the conventional description breaks down,
and the electron state would become localized there will occur Anderson localization and the material will
become insulating. The borderline between these two regimes (still a metallic one, in which (T ) remains
finite as T 0, albeit with short mean free path and small conductivity, and an insulating one, with (T ) 0
or R(T ) for T 0) corresponds to a value of similar to the one given by equation (12.44), with a
somewhat smaller prefactor,
min 0.03 e2 /ha
(12.45)
corresponds to 300 1 cm1 . This value is known as the Mott minimum metallic
which for a = 3 A
conductivity.
268
insulating state (although with a quite wrong value of the energy gap). However, this
approach predicts that in the disordered state, at T > TN , the gap would close and
the system would be metallic. Experimentally this is definitely not the case in these
materials, and it is just this fact which actually led to the development of the whole
big field of the physics of Mott insulators, strongly correlated electrons, the Hubbard
model, etc.
There are several systems among transition metal compounds in which insulator
metal transitions are experimentally observed. These are, for example, many oxides
of V and Ti: VO2 ; V2 O3 ; many of the so-called Magneli phases Vn O2n1 ; Ti2 O3 ;
and several systems Tin O2n1 (e.g. Ti4 O7 ). In these materials metalnonmetal
transitions occur as a function of temperature, and can also be induced by pressure,
doping, etc. Some of these transitions are accompanied by magnetic ordering, but
often such ordering occurs at a temperature not coinciding with Tinsulatormetal . There
is usually a structural change at such transitions, but it is often not clear whether
this structural change is the cause and the main driving force of such transition, or
it is only a consequence of it: in any case the lattice structure is sensitive to the
state of the electronic subsystem, and any change in the latter, especially such a
strong change as a metalinsulator transition (e.g. localization of electrons) should
be reflected in the structure.
There were a lot of theoretical attempts to describe insulatormetal transitions
starting from the Hubbard model. Most of them use certain poorly controlled
approximations, and it is difficult to judge how reliable they are. I will only present
here some of the results of probably the most successful approach developed by
Brinkman and Rice. They used a certain variational scheme (called the Gutzwiller
method) and studied, at T = 0, the behaviour of the metallic state when the Hubbard
interaction U increases. They have shown that in this method the system becomes
insulating (electrons become localized) when U Uc = 8| 0 |, where 0 is the
average kinetic energy of electrons (of the order of the hopping matrix element t).
When U Uc , the electron effective mass diverges as
m
1
,
=
m
1 (U/Uc )2
(12.46)
i.e. the electrons (or rather quasiparticles) become infinitely heavy, and for U > Uc
they are localized at particular sites. Using the relations m /m = 1/Z where Z
is the renormalization factor in the one-electron Green function (8.54), (8.91),
one sees that (12.46) corresponds to Z 0, i.e. the strength of the quasiparticle
pole goes to zero at this point. Simultaneously, according to (8.55), the jump
of the electron distribution function n(pF 0) n(pF + 0) goes to zero, i.e. the
very Fermi surface disappears, and with it the Fermi liquid. In the insulating
state at U > Uc the electron distribution function is smooth, Fig. 12.27(a), and it
tends to a constant value, Fig. 12.27(b), for U (the state localized in space,
269
1
2
(a)
(b)
Fig. 12.27
(x) (x x0 ), gives a constant in momentum space after a Fourier transform,

i.e. all momenta are equally represented in the localized state).
One can also show that at least in this approximation the magnetic susceptibility
also diverges at U Uc :
=
2B (F )
m
,
1 (U/Uc )2
m
(12.47)
i.e. it is enhanced mainly because of the increase of the effective mass. In terms
of the Fermi-liquid theory this means that the Landau parameter F1s diverges, but
(1 + F0a )1 remains finite as U Uc , in contrast to the usual magnetic transitions,
cf. (10.16), (10.18). This is consistent with our general idea that a Mott transition is
a transition to localized electrons (and consequently localized magnetic moments)
but not necessarily a transition to a magnetically ordered state.
Actually of course the localized magnetic moments which appear in a Mott
insulator should somehow order at T = 0. This effect is missing in the original
treatment of Brinkman and Rice. The hope is, however, that this ordering, albeit
important, does not play a crucial role in the very phenomenon of electron localization and the corresponding Mott transition, so that the basic physics is accounted
for well enough.
Other methods of treating Mott transitions have also been developed recently.
One of the most successful of these is the so-called dynamical mean field theory (DMFT), see, e.g. Georges et al. (1996). In this method, which is somewhat
similar to the conventional mean field theory, one reduces the description of the
concentrated system to that of one site interacting with the mean field of the surrounding. But, in contrast to the usual treatment, one takes into account dynamic
effects both in the singled-out impurity and in the bath. One of the outcomes
of this treatment is a rather appealing picture of what happens when a metallic
system approaches a Mott transition, e.g. with increasing U/t in the Hubbard
model with n = 1. This method gives the following sequence of events, see
Fig. 12.28: far from the Mott transition we have an ordinary metal, with the density
of states shown in Fig. 12.28 and with the usual Fermi surface. With increasing
electron correlations (increasing U/t) there appear wings in () both below
270

r
U/ 2
+U/2
U/2
+U/2
U/2
+U/ 2
U/t
Fig. 12.28
and above the original metallic band (the position of the Fermi energy here is
taken as zero energy). These broad peaks in () appear at U/2 below F and
at U/2 above F . When U/t becomes still larger, the main weight is gradually transferred to these side peaks, which actually represent the lower and upper
Hubbard bands with the gap U between them. Still, on the metallic side of
the transition there remains a quasiparticle peak at the original Fermi level, with
the constant density of states at F , but with the width (and correspondingly the
total weight Z) going to zero as we approach the Mott transition. And at a certain critical value of U/t this quasiparticle peak disappears, and we have only the
lower and upper Hubbard bands with the gap between them. This state is a Mott
insulator.
Actually a lot of questions in this field still remain open. Can a Mott transition
occur by itself, without being accompanied by a real ordering of some type (magnetic ordering, structural distortion)? If so, will this transition be first or second
order? What is the symmetry (if any!) discriminating between insulating and metallic phases (in the sense of the Landau theory of second-order phase transitions, see
Chapter 2)? What would be the corresponding order parameter? It is not even clear
which phase, insulating or metallic, should be considered as the ordered and which
as the disordered phase. The first impulse is usually to treat the insulating state as
the ordered one. However this may not be the case. We have already given earlier
the arguments (see the discussion at the end of Section 10.1) that the very existence
of the Fermi surface in metals may be a kind of ordering. From this point of view
the metal should be treated as a unique ordered phase with zero entropy, and the
Mott insulator without magnetic ordering as a state with higher symmetry, or a disordered phase (at least it has spin disorder cf. the discussion in Section 2.7.3 of the
high-temperature metalinsulator phase transition in V2 O3 as driven by magnetic
entropy).
All these questions are actually rather deep. It is at present not at all clear
whether one can indeed treat Mott transitions on the same footing as the usual
second-order phase transitions. Thus, e.g. in the conventional theory we define the
order parameter as the average of a certain operator over the ground state (e.g.
271
the magnetization M = 0|M|0, etc.). It is not at all clear whether there exists
such a notion for the insulatormetal transitions. Actually the physical definition
of the difference between an insulator and a metal is connected with the static
conductivity ( 0) being zero or finite. But this is not a characteristic of the
ground state, but rather of the lowest excited states: one may define insulators as
systems in which the first excited states are separated from the ground state by
a finite gap, whereas there is no such gap in a metal.14 In this sense there may
be formally no order parameter in the ordinary sense which would discriminate
Mott insulators from metals. If so, the pure Mott transition can only be a first-order
transition, if nothing more intricate takes place here.
Concluding this chapter one should say that the whole field of correlated electrons in general and insulatormetal (MottHubbard) transitions in particular is
still an active field of research, and a lot remains to be understood.
14
But be careful with superconductors! In this sense they are more similar to insulators than to metals. Also in
strongly disordered systems the electronic states can be localized and consequently the conductivity will be
zero, but the energy spectrum may be continuous and have no gap see the discussion of Anderson localization
above. Thus even this definition applies, strictly speaking, only to nonsuperconducting systems without strong
disorder.
13
Magnetic impurities in metals, Kondo effect, heavy
fermions and mixed valence
In the previous chapter we have considered the properties of strongly correlated

electrons. The systems we mostly had in mind were the compounds of transition
metal and maybe rare earth elements with partially filled inner d or f shells. We
have discussed only the correlated electrons themselves, the prototype model being
the Hubbard model (12.1).
When turning to real materials, several extra factors missing in the model (12.1)
are important. One of them is the possible influence of orbital degrees of freedom,
especially in cases with orbital degeneracy, treated in Section 12.9.
In many situations there is yet another very important factor. There may exist in
a system, besides correlated electrons, also electrons of other bands, e.g. electrons
in wide conduction bands, responsible for ordinary metallic conductivity. Such is
for instance the situation for magnetic impurities in metals, or in the concentrated
systems like rare earth metals and compounds in which localized f electrons coexist
with the metallic electrons in broad spd bands. The interplay between localized, or,
better, strongly correlated electrons and itinerant electrons of the wide bands can
lead to a number of very interesting consequences; these will be discussed in this
chapter.
When we put transition metal impurities in ordinary metals (e.g. Mn or Fe in
Cu, Au), the result may be two-fold. In certain cases the impurities retain their
magnetic moment, but in others they lose it. Qualitatively this second possibility
is connected with the following process: when the localized electron level, e.g. the
d level of the impurity d , overlaps with the continuous spectrum, electrons on
this level may have a finite lifetime, i.e. the d electron with spin can escape into
the conduction band, and in its place another electron from the conduction band
may be transferred, possibly with the opposite spin . As a consequence of this
272
273
e
eF
ed
k
Fig. 13.1
process the localized level acquires a finite width, Fig. 13.1, and there appears a
tendency to reduce or completely quench the moment.
The situation, however, is not so simple. When the d level lies deep below the
Fermi level, all states in the conduction band with energies close to d are occupied,
and this process is suppressed. The large Coulomb interaction of localized electrons
Udd also acts to preserve the localized magnetic moment.
The problem of the appearance of localized magnetic moments at impurities in
metals was treated in detail by P. W. Anderson. He used the model qualitatively
discussed above, with the Hamiltonian

H=
k ck, ck, + d
d d + U nd nd +
Vk ck, d + h.c. .
k,
k,
(13.1)
Here d , d are creation and annihilation operators for electrons on the d level
of the impurity, ck, , ck, are those of conduction electrons, and Vk is the matrix
element of dc hybridization which describes the process of mixing of d and c
electrons. (Note that here and below V is the dc hybridization, not the volume!)
This model is known as the Anderson model of magnetic impurities.
The approximation used in the first paper by Anderson was the unrestricted (in
other words, spin-dependent) HartreeFock approximation, in which the following
decoupling was made:
U nd nd U nd dd + U nd nd U nd dd .
(13.2)
With this decoupling one can easily find the energy spectrum by diagonalizing the
resulting quadratic Hamiltonian. A convenient way to do this, used by Anderson,
is to write down the equations of motion for the operators ck, , d , or for the
corresponding Green functions, which have the form

d = d d + U nd, d +
Vk ck, ,
(13.3)
k
ck, = k ck, + Vk d .
274
Magnetic impurities in metals
From these equations, or from the equivalent equations for the corresponding Green
functions, we find for the Green function of the impurity electrons the expression
Gdd =
1
d U nd,
or, in other words, (Gdd )1 = d
d ()
d (),
= U nd, +
|Vk |2
k k
(13.4)
where the self-energy is
|Vk |2
.
k
(13.5)
When the renormalized d level d = d + U nd, lies within the energy band
k , d has both real and imaginary parts, which determine the width of the d level.
Thus one can write down the density of states of d electrons as

1

,
d U nd, 2 + 2
d () =
where the width is given by

|Vk |2 ( k ) = |Vk |2 (F ) .
=
(13.6)
(13.7)
The average occupation of the d level with spin projection is then given by

F
d + U nd, F
1
nd, =
. (13.8)
d () d = arctan
One can also obtain this result directly using the formula (8.37) expressing the
electron density through the Green function (however, one should not integrate
in (8.37) over d 3 k/(2 )3 because we are now studying an isolated impurity).
The equation (13.8) and a similar equation for nd, constitute two selfconsistent equations which should be solved together. There exist in general two
types of solutions depending on the values of the parameters F d , U , . There
is always a nonmagnetic solution nd = nd = 12 nd . However for large U
and small this solution corresponds not to a minimum, but to a maximum of
the energy, and there exists another, magnetic solution with nd
= nd . This
solution is doubly degenerate ( nd > nd or vice versa). For the symmetric
case F d = U/2 (i.e. when the levels d and d + U are situated symmetrically
below and above the Fermi level) the condition for the appearance of localized
moments is
Ud (F ) > 1 ,
(13.9)
where d is given by (13.6), i.e. it coincides with the Stoner criterion (11.62) for
ferromagnetic instability of itinerant electrons. The total density of states then has
275
e
ed + U nd
eF
ed + U nd
k
Fig. 13.2
eF = ed +
U
2
k
Fig. 13.3
the form shown in Fig. 13.2, i.e. we have two virtual levels with finite width (13.7)
and with the Lorentzian density of states (13.6), nd being given by the (grey)
area of the peak on the right-hand side of the figure, with energies up to F , and
nd being given by the corresponding grey area ( < F ) of the left peak. The
nonmagnetic solution would correspond here to the situation shown in Fig. 13.3, i.e.
to equal occupations of the d levels with up and down spins. (We stress once again
that this situation is realized for the symmetric Anderson model F = d + U/2;
the conditions for the existence of localized magnetic moments for asymmetric
situations are more stringent.)
For the symmetric case the criterion (13.9) may also be rewritten in the form
U
>1.

(13.10)
(Here we approach the instability from the nonmagnetic side, n = n = 12 ,

and we have used (13.6) and the condition F = d + U/2, which gives d (F ) =
1/ .) The region in the (d , U/ ) plane where there exists a magnetic solution is
shown schematically in Fig. 13.4. We see that the existence of localized magnetic
moments of the impurity is facilitated in cases of large Coulomb interaction, small
df mixing V , small density of states of the conduction band (F ) (as we see
from (13.7), = (F )V 2 ) and symmetric position of the localized level.
Thus the qualitative picture of the appearance (or of preservation) of the localized
magnetic moment in the Anderson model is the following: if the state at energy
magnetic
solution
ed eF + U
276
U
p
Fig. 13.4
d with spin is below the Fermi level and is occupied, then to put another
electron with spin at the same d level would cost energy U , i.e. one may say
that the second electron will occupy the state with the energy d + U , which for
large enough U will lie above the Fermi level and consequently should be empty.
In this case the impurity has localized moment. When, due to dc hybridization
the number of spin-up electrons decreases, so does the energy of the spin-down
state located at d + U nd . In effect these two levels, with spins and , would
move towards each other and approach F , and if this process is strong enough (if
the hybridization V is large and U small), one may end up in the nonmagnetic state
of Fig. 13.3. However, if the conditions (13.9), (13.10) are satisfied, the solution
with localized magnetic moments, illustrated in Fig. 13.2, would be stable.
13.2 Kondo effect

The treatment of the Anderson model given above is still a mean field one. When
one goes beyond the mean field approximation, new effects appear which make the
situation much more difficult but also much more interesting. Notably, it turns out
that the remaining interactions between the impurity electron (or spin) and conduction electrons, which were not taken into account in the HartreeFock treatment
of the previous section, lead to effective screening and eventual disappearance of
localized magnetic moments as T 0. Thus the very moment which we created
with such difficulties only two pages ago would disappear! This is the famous
Kondo effect.
Actually Kondo considered not the Anderson model (13.1), but the so-called sd
exchange model with localized spin S interacting with the conduction electrons via
an exchange interaction
Hsd =

k,k
Jkk ck, ck , S J s(0) S
(13.11)
13.2 Kondo effect
277
e
ed + U
1
eF
ed
k
Fig. 13.5
where are the -matrices and s(0) is the spin density of conduction electrons at
the position of the impurity. Actually this interaction can be obtained from the
Anderson model (13.1) in the limit in which there exist localized magnetic
moments, see the previous section. In this case, if the d level is deep enough and
is small, the total occupation of the d level is close to 1, e.g. nd 1, nd 0.
However, as we have already mentioned, there is another degenerate solution with
the same energy, corresponding to nd 1, nd 0. Thus, the average occupation of the localized levels in this regime is close to integral (in our case 1), but
there remains spin degeneracy, i.e. spin degrees of freedom. Consequently we may
project out charge degrees of freedom and keep only the spin ones, obtaining the
effective exchange interaction (13.11). This procedure is similar in spirit, and actually in mathematics, to the one which led us from the Hubbard model (12.1) to the
superexchange Hamiltonian (12.7); here it is known as the SchriefferWolf transformation. In the simplest form one can obtain the sd exchange interaction (13.11)
starting from (13.1) and treating the hybridization term in perturbation theory up to
second order in V . The processes which contribute to the exchange are virtual transitions of the d electron to the conduction band and back (process 1 in Fig. 13.5),
and the transition from the conduction band to the already occupied d state and back
(process 2 in Fig. 13.5); to put the second electron at the d level costs extra energy U .
As in our case n 1, n 0, we have d d , d d + U . Again, similar to
the Hubbard model, the resulting exchange interaction is antiferromagnetic, with
the exchange constant

U
1
1
2
.
= 2V 2
+
J = 2VkF
F d
d + U F
(F d )(d + U F )
(13.12)
Let us consider now the consequences of the exchange interaction (13.11). First
we do it more formally. One of the processes is the scattering of electrons on the
localized spin. There are two possible processes contributing to this scattering:
ordinary potential scattering and exchange, or spin-flip scattering, in which the
electron exchanges its spin with the impurity.
278

p1 1
p2 2
p2 2
p1 1
p
(a)
(b)
Fig. 13.6
In the lowest approximation (the first Born approximation) nothing special happens. The first-order contribution to the scattering amplitude A(1) will be proportional to J S, and after summation over all spin orientations we obtain for the
scattering rate W |A(1) |2 J 2 S(S + 1). However, in second order the spin-flip
scattering gives a term in the scattering amplitude which is logarithmically divergent. The corresponding processes may be represented by the two diagrams shown
in Fig. 13.6. The diagram 13.6(a) describes the process in which an incoming
electron with momentum p1 and spin 1 is scattered into an intermediate state p
and then goes to a state p2 2 . The summation over intermediate states should be
carried out over unoccupied states above the Fermi energy, i.e. it should contain a
factor (1 f ( p )) where f ( p) is the Fermi factor. This term thus has the form

d 3 p 1 f ( p ) 1 S 2 S
2
J
.
(13.13)
3
)
(2
)
(
p
)
(
p
1
The diagram 13.6(b) describes the process in which initially an electronhole pair
is created, the electron being in the state p2 2 and the hole in p (left vertex in
Fig. 13.6(b)), and then the initial electron is scattered from the state p1 1 into the
state p which is now empty. Correspondingly, now | p | < pF , which is taken
care of by the function f ( p ), which gives the term

d 3 p f ( p ) 1 S 2 S
2
J
.
(13.14)
3
) ( p )
(2
)
(
p
1
It is important that the second contribution has the opposite sign which is a consequence of the antisymmetric character of the electron wavefunction.
After summation over spin indices we obtain finally the contribution to the
scattering amplitude

d 3 p 2f ( p ) 1
(2)
2
S
.
(13.15)
A p1 p2 = J
2
1
(2 )3 ( p1 ) ( p )
The integral in (12.18) is logarithmically divergent for | p1 | pF or ( p1 ) F
(note that the scattering is elastic, i.e. | p2 | = | p1 | and ( p2 ) = ( p1 )). Thus the
13.2 Kondo effect
279
Rtotal
R0(T)
RK 1 + c ln TF
)
T
Fig. 13.7
second-order term in the scattering amplitude is proportional to ln max{|( p1F)F |,T } ,

and the total scattering amplitude up to second order in J is
4
5
F
6
7 . (13.16)
A p1 p2 = J ( S)2 1 1 + J (F ) ln
max |( p1 ) F |, T
This scattering amplitude determines the electron lifetime 1 |A|2 , which enters
the resistivity through the usual formula R 1 = = ne2 /m. As a result the
resistivity behaves as

F
,
(13.17)
R(T ) = R0 (T ) + RK 1 + 2J (F ) ln
T
i.e. it contains a term logarithmically increasing with decreasing temperature. This
term, together with the usual scattering R0 (T ), gives the famous minimum of the
resistivity, schematically shown in Fig. 13.7. When the second order in perturbation
theory diverges, as in our case, it means that we have to sum many terms of higher
order, which will diverge as (ln F /T )2 , (ln F /T )3 , etc. Summation of the simple
geometric series of this type gives an expression of the type
J
.
1 J ln(F /T )
(13.18)
This expression formally diverges for antiferromagnetic interactions J > 0 at a

certain critical, or rather characteristic temperature
TK = F e1/J(F ) ,
(13.19)
which is called the Kondo temperature cf. the case of the BCS theory of superconductivity, equations (11.47), (11.48). Of course, in contrast to superconductivity
in the case of a single impurity in a metal there can be no real phase transition.
280
Rather TK (13.19) gives the temperature scale at which there occurs a change of
the behaviour of our system: above TK the perturbation theory described above
is valid, and below TK it is no longer applicable. Thus, e.g. above TK there is a
logarithmic term in the resistivity; below TK the behaviour of R(T ) changes, and
actually resistivity saturates. Similarly, there is a cross-over in the behaviour of the
impurity susceptibility: it also increases with decreasing temperature at T > TK ,
but saturates at the value 1/TK at T < TK .
The divergent contribution we have obtained can be traced back to the fact
that the spin operators which enter into the spin-flip processes do not commute.
Thus the quantum nature of spins, together with the presence of the Fermi surface
which make scattering processes essentially two-dimensional, are responsible for
the Kondo effect (as T 0 the electrons scatter only at the Fermi surface, i.e.
their momenta are confined to a two-dimensional manifold, to the Fermi surface).
Actually it is this two-dimensionality which gives the logarithm in (13.16); in that
sense the situation is analogous to the formation of the well-known Cooper pairs
in superconductivity the bound state of two electrons which attract each other
but can scatter only above the Fermi energy.
This analogy can actually be made even closer. One can interpret the results
obtained above as a tendency to form a singlet bound state of the conduction
electrons with the localized spin. Indeed we can obtain such a bound state with
the binding energy
Eb F e1/J
(13.20)
if we consider the scattering of an electron outside the Fermi surface on a localized spin. Mathematically the corresponding treatment is analogous to the general
treatment of the formation of impurity states given in Section 6.5. The difference
is that here we have to sum in the equation of type (6.133) over unoccupied states
q > pF , which makes the problem essentially two-dimensional. Of course, this
bound state is not an actual bound state: the energy Eb (13.20) would correspond
to a decrease of energy relative to F , and for Eb F this level would still overlap
with the continuum and would have a certain width. It should rather be treated as
a resonance which for TK Eb F lies actually at the Fermi surface or very
close to it. This resonance is called Kondo resonance, or sometimes Abrikosov
Suhl resonance. Actually it describes the screening of localized spins: due to the
antiferromagnetic interaction (13.11) the electrons with spins opposite to S have a
tendency to come closer to the impurity, forming a screening cloud with the typical
hvF /TK . At high temperatures such screening is not efficient, and the
size =
system behaves as a real localized magnetic impurity, but with decreasing temperature the screening becomes more and more efficient until it looks as though
the localized magnetic moment completely disappears at T TK (when we look
13.2 Kondo effect
281
r ( e)
TK
eF
Fig. 13.8
from outside). Thus qualitatively we may say that the essence of the Kondo effect
is the screening of the localized magnetic moment of an impurity by conduction
electrons, occurring at low temperatures T <
TK .
What then is the state of the system as T 0? It is now established that in normal cases (e.g. for an impurity spin S = 12 interacting with a nondegenerate band)
the limiting behaviour is again that of the Fermi liquid, but with strongly renormalized parameters. The Kondo temperature and the corresponding energy scale TK
actually play the role of a new effective Fermi energy. As a result all the usual
expressions for thermodynamic properties presented at the beginning of Chapter 7
remain valid, but with F TK , (F ) 1/TK , etc. Thus, e.g. the specific heat due
to such Kondo impurities tends to c(T ) T /TK , and the susceptibility |T <TK
2B /TK , so that the Wilson ratio RW = 2 /(32B c/T ) remains 1. (Of course
all these contributions are proportional to the concentration of Kondo impurities.)
Qualitative interpretations of this behaviour may be given using the picture of
Kondo resonance: the low-temperature behaviour of this system is that of a Fermi
system with a narrow peak in the density of states (Kondo resonance) very close
to the Fermi energy, see Fig. 13.8 (of course the intensity, the area, or the number
of states in this peak is proportional to the concentration of such impurities).
The width of this peak is TK . At T TK we feel a large density of states,
which is reflected, e.g. in the large -value ( = c(T )/T 1/TK ). However, when
the temperature becomes larger than TK , the region T around F contributes to
the thermodynamic properties such as specific heat, etc., i.e. we have to average the
density of states shown in Fig. 13.8 over the interval T around F , and the average
density of states decreases with increasing temperature (actually the Kondo peak
itself starts to disappear).
The theoretical problem of how to describe the cross-over from the magnetic
behaviour at high temperatures to a nonmagnetic Fermi liquid as T 0 turned out
to be very difficult, and a lot of effort was required to reach a full understanding.
282
And even now the Kondo effect still brings about some surprises. In particular,
relatively recently it was realized that there exist certain special situations (e.g.
the so-called multichannel Kondo effect the situation when the total number of
degrees of freedom of conduction electrons is larger than the spin of the impurity) in
which the behaviour is very different from the conventional one. This multichannel
Kondo effect can lead, e.g. to non-Fermi-liquid behaviour as T 0, a problem
which now attracts considerable attention (see Section 10.2).

In the previous section we have considered the case of an isolated magnetic impurity. We have discussed predominantly the situation when the impurity level lies
deep below the Fermi energy, so that the total occupation of this level is close to
integer (e.g. 1), and only spin degrees of freedom remain. It is very interesting
to lift these restrictions and to study concentrated systems in which the d or f
levels lie relatively close to F or even cross it. It turns out that in this situation we
meet many very interesting phenomena, not all of which are now understood well
enough. These are the phenomena of mixed valence and heavy fermions. There
exist a lot of materials belonging to this class, especially rare earth (4f) and actinide
(5f) compounds. Typical examples of such systems are, e.g. CeCu6 , CeAl3 , UBe13 ,
U2 Zn17 , and many others. As the heavy fermion and mixed-valence states are predominantly found in rare earth (4f) and in actinide (5f) compounds, we will speak
below about f electrons, using also the corresponding notation.
The properties of these materials are very rich. There are among them normal
metals with very large effective mass, m 102 103 electron mass, and with huge
linear specific heat c = T , ( m ) 103 mJ/mole K2 (in ordinary metals
like Al or Cu 1 mJ/mole K2 ); examples are CeCu6 , CeAl3 . In other materials heavy fermion behaviour coexists with magnetic ordering, but often with
very small magnetic moment (CeAl2 , U2 Zn17 ). There also exist heavy fermion
superconductors (CeCu2 Si2 , UPt3 , UBe13 ) with rather unusual properties, most
probably having unconventional (d-wave or p-wave) pairing. (Actually the study
of superconductors with unconventional pairing, which is now so popular for
example in high-Tc superconductors, was started before the discovery of HTSC in
cuprates, in connection with superconductivity in heavy fermion systems.) There
exists among these systems also a relatively small number of insulating materials
with small energy gap the so-called Kondo insulators (SmB6 , YbB12 , CeNiSn).
Some of these substances display very interesting phase transitions as a function
of temperature, pressure, etc. at which the electronic structure and other properties
change drastically (the valence-change transitions), e.g. the transition in Ce,
283
blackgold transition in SmS, etc. And finally, many of these systems display
quantum critical points and non-Fermi-liquid behaviour.
All specific properties of these materials are connected with the interplay of
strongly correlated electrons and the electrons of wide conduction bands. Thus we
are dealing here with the generalization of the Anderson or Kondo models (13.1)
and (13.11) of magnetic impurities to the case of concentrated systems. We thus
obtain the so-called Anderson lattice

H=
k ck, ck, + f
fi fi + U
fi fi fi fi
i,
k,

+
Vik ck, fi + h.c. ,
(13.21)
i,k,
or Kondo lattice
H=
k ck, ck, +
Jikk ck, ck , Si ,
(13.22)
k,
where i is the lattice index, and where we have used for correlated electrons the
notation f , f to stress that we predominantly have in mind f electron systems.
We see that here, besides all the difficult problems we have met for the usual
Anderson or Kondo impurities, we also have to take into account the effects connected with the presence of many such f sites. As a result the properties of these
systems are much richer than in the impurity case.
We cannot cover here all aspects of the theory of these systems; they can be
found in special books and reviews, e.g. Hewson (1993). We will give here only
the basic scheme which permits one to systematize the behaviour of these systems
in different regimes.
The ff repulsion U is usually very large, and we can take it as the biggest
parameter in our model, as well as the bandwidth W (F ) of conduction electrons.
Consider first the case of a deep f level, f F . In this case the f level is occupied
by one electron, and we can go over from the Anderson lattice model (13.21) to
the Kondo lattice (13.22).
There are two physical effects (and, respectively, two possible regimes) in this
case. One of them is the RKKY exchange interaction between f electrons mediated by conduction electrons, see (9.29). This interaction would lead to a certain
magnetic ordering at the critical temperature
Tc JRKKY (F ) J 2 J 2 /F ,
(13.23)
where J is the sf exchange in (13.22). However, there also exists the opposing
tendency the tendency to screen out the magnetic moment and to create a nonmagnetic ground state due to the Kondo effect at each site. The scale of this effect
284
is given by the Kondo temperature

TK F e1/J(F ) ,
(13.24)
see (13.19).
When Tc > TK , the RKKY interaction wins, and the ground state will be
magnetically ordered. Such are most of the rare earth metals (Gd, Er, etc.) and their
compounds. However, if TK > Tc , then the magnetic moments disappear with
decreasing temperature at T TK before they have a chance to order. In this case
we would end up in a nonmagnetic state. In analogy with the impurity Kondo effect
the energy scale will be given here by TK , so that the thermodynamic properties
(c(T ), (T )) will be similar to those of the Kondo impurities, with the important
difference that we now have such impurities at each site, i.e. we have not 1, but
1022 of them. Consequently all anomalous properties will be strongly enhanced (in
dilute systems they are of course proportional to the impurity concentration).
Indeed, as we have discussed in the previous section, cf. Fig. 13.8, now these
systems behave at low temperatures as systems with extremely high density of
states at the Fermi energy. Consequently these systems behave at low temperatures
as Fermi liquids, but with a huge enhancement of the effective mass, reaching
103 m0 , and all this in systems with 1022 such states. This is the heavy fermion
state. All the anomalies in this regime are really giant, and, as one used to say, these
systems present a paradise for experimentalists but simultaneously a nightmare
for theoreticians.
An important difference with respect to the impurity case is seen in the behaviour
of the resistivity and in other transport properties. Whereas for the Kondo impurity
the resistivity behaves as shown in Fig. 13.7, i.e. with decreasing temperature it
increases and saturates at T TK , in regular systems such as the Kondo lattice
there should be no residual resistivity (if there are no extra impurities). Thus, R(T )
should decrease and formally should go to zero for T 0. Experimentally this is
indeed the case: the typical behaviour of the resistivity in heavy fermion systems
looks as shown in Fig. 13.9.
Indeed at high temperatures T TK each f site acts independently, and the
whole system may be visualized as a collection of independent impurities, with the
Kondo-type behaviour of resistivity of Fig. 13.7. However, at low enough temperatures a new coherent regime will be formed, and as T 0 the system behaves
as a Fermi liquid with very low degeneracy temperature T and high effective
mass m m. The resistivity at T < T is mainly determined by the electron
electron scattering, and according to the general treatment given in Chapter 10,
it behaves as R(T ) = (T /T )2 = AT 2 , see (10.7). The coefficient A (1/T )2
scales with the coefficient in the linear specific heat, 1/T . An important
problem, yet unsolved, is whether the coherence energy scale T coincides with
285
R
Kondo regime
coherent
regime
T*
TK
Fig. 13.9
the Kondo temperature for an isolated impurity TK , or whether they are different.
Experimentally, in most cases, the T 2 -law of resistivity is observed at temperatures
much lower than the single impurity TK , but whether it is because of the existence
of two different energy scales with T < TK , or is simply due to a large cross-over
region between the well-developed Fermi-liquid behaviour with R T 2 and the
Kondo behaviour at T > T , is not really clear.
We have stated above that the Kondo-like, or heavy fermion behaviour can be
observed if TK > Tc JRKKY . As JRKKY J 2 and TK exp(1/J ), this is possible for large enough exchange integral J ; at small J the magnetic regime always
wins. If we now remember that according to our previous treatment, see (13.12),
for large U (e.g. U ) we have
J =
2V 2
,
F f
(13.25)
we see that J increases when the f level approaches F . Of course, we cannot use
the expression (13.25) for F f 0; in this case the corresponding derivation
is not valid, but the qualitative tendency given by equation (13.25) is still correct.
Thus we can draw a phase diagram of our system in the plane (f , T ) (the Doniach
phase diagram), see Fig. 13.10. In our previous treatment we have considered the
situation with the (almost) integral occupation of f levels, and we have used the
concepts of localized moments, Kondo effect, etc. This picture is valid if the f
levels lie still relatively deep below F . What would happen if we increase f ? This
is illustrated in the Doniach phase diagram.
We start with a very deep f level, f F . As we have discussed above, in this
case J (13.25) is very small, and Tc JRKKY (F )J 2 J 2 /F (13.23) is much
larger than the Kondo temperature TK F e1/J(F ) (13.19). In this regime the
system is magnetically ordered at low temperatures. Such are the majority of rare
earth metals and compounds.
286

T
TKondo
TRKKY ~ J RKKY
normal metal
magneticaly ordered
state
heavy
fermion
phase
mixed valence
regime
ef
Fig. 13.10
When we move the f level closer to the Fermi level, the Kondo energy scale
starts to gain, and it may become comparable to or even exceed Tc . One can get this
regime for a still relatively deep f level, with occupation nf close to one. This is the
heavy fermion regime. The critical temperature of magnetic ordering by then starts
to decrease because the magnetic moments themselves are reduced due to Kondo
screening, and we see that we can naturally get here the situation with Tc 0,
which is the typical situation with quantum critical points. That is why many heavy
fermion systems show quantum critical behaviour.
When we move the f level still further up and reach the situation when the f level
f itself approaches the Fermi level and F f becomes comparable with the f
level width = V 2 (13.7), we cannot use the Kondo lattice model any longer,
and have to go back to the Anderson lattice Hamiltonian (13.21). The occupation
of the f level in this case will be less than 1 (or in general will be noninteger) we
enter the regime of mixed valence. (By mixed, or intermediate valence we mean
here just that noninteger occupation of the usually deep f level with a small radius
of f orbitals, which can normally be treated as belonging to the ionic core and which
do not participate in chemical bonding.) Here we have the situation intermediate
between two different occupations of the f shell. For example, in compounds like
CePd3 the average f shell occupation is neither 4f 1 , corresponding to valence Ce3+ ,
nor 4f 0 , which would correspond to Ce4+ , but has an intermediate value.
Intermediate valence means that there are quantum fluctuations between different
configurations, so that the ground state is a superposition of states with different
distributions of electrons between the f orbitals and the conduction band, e.g.
| = |f 1 cn + |f 0 cn+1 . The average occupation nf then is the weight ||2
with which this configuration is represented in the total wavefunction. The mixing
is due to the hybridization term V c f in (13.21). However, the difficult theoretical
problem is how to hybridize correlated bands. The f and c electrons have completely
different properties. The f electrons are strongly correlated: because of strong
IIa
IIb
III e F
287
IV
ef
Fig. 13.11
Hubbard ff repulsion we can put at each f level not more than one electron. At
the same time, conduction electrons are ordinary fermions, and each level may be
occupied by two electrons, with spins up and down. Thus if we have to hybridize two
such different bands, with different capacities and actually with different statistics
(ordinary Fermi statistics for c electrons and atomic statistics for f electrons),
then we cannot say in advance, for example, how many electrons we can put in the
hybridized band: is it one electron? or two? or 1.5? This is of course a very crude
way to formulate the emerging problems, but it essentially reflects the difficulties
encountered in the theory of mixed valence and heavy fermion compounds. This
field, as well as other topics in the physics of correlated electrons, is now at the
forefront of condensed matter theory.
The properties of the mixed-valence phase resemble those of the heavy fermion
(dense Kondo) phase. Notably, the magnetic susceptibility is Curie-like at T >
T (13.7), and saturates at the value 2B / for T 0. The difference is that
usually, according to equations (13.12), (13.19), the Kondo scale TK increases with
decreasing F f , but remains small. In the mixed-valence regime the appropriate
scale is larger: if for heavy fermions usually TK 110 K, for mixed-valence
compounds typically T 103 K.
And finally, if we continue to raise the f level, sooner or later it will be above
the Fermi energy F ; all f electrons would then spill out into the conduction band,
and we will end up with a normal nonmagnetic metal with empty f levels. Thus,
the phase diagram displaying the sequence of events as the f level moves upwards
may be finally drawn as in Fig. 13.10, and the evolution of the energy structure
with f would look as shown in Fig. 13.11:
Region I: Magnetic metal with localized magnetic moments. The f level lies
deep below the Fermi level. Usually this state is magnetically ordered at low
temperatures.
Region II: Dense Kondo system. The f level with width = V 2 is closer
to F . There exists a collective Kondo peak in the density of states close to F ,
with width TK F e1/J . This region is divided into two subregions, IIa
288
and IIb. In region IIa, Kondo-like behaviour may coexist with magnetic
ordering of the remaining (often strongly reduced) magnetic moments. In
region IIb, the nonmagnetic (heavy Fermi liquid) state becomes stabilized
down to T = 0. The f level f itself is still below the Fermi level, but there
appears Kondo resonance at F .
Region III: Mixed-valence region. The ground state is the nonmagnetic Fermi
liquid, with moderately enhanced susceptibility and specific heat. It may be
visualized as a fluctuating valence state (quantum fluctuations). This regime
is reached when the f level itself is close to the Fermi level.
Region IV: Normal nonmagnetic metal with an empty f shell (f level above the
Fermi level).
This is a qualitative scheme which describes the general behaviour of concentrated
systems with coexistent correlated and ordinary itinerant (metallic) electrons. We
cannot discuss here the many very interesting phenomena occurring in the heavy
fermion and mixed-valence compounds. Suffice it to say that due to the very large
effective mass and low characteristic temperature TK or T 1100 K (which
plays the role of the Fermi temperature TF F for normal electrons) we have here
a unique possibility to study experimentally both the regimes with T T (degenerate electrons) and T T (nondegenerate regime) in one system, which for other
systems would require extreme conditions (extremely high pressures and temperatures, magnetic fields, etc.). Thus, besides their own interest, the heavy electrons in
these systems give us a unique opportunity to model in experimentally accessible
conditions the behaviour of general condensed matter in extreme regimes.

Despite the large variety of properties, all different possible states of the heavy
fermion and mixed-valence systems described above, see, e.g. Fig. 13.11, were
metals maybe with very strongly renormalized parameters, very large effective
mass, with rather unusual properties, but still metallic. There exists, however,
a small group of these compounds which have very small gaps in their energy
spectrum, i.e. which are strictly speaking insulators. Several examples were already
mentioned above. They are usually called Kondo insulators, although some of
them, e.g. gold SmS, should be better called mixed-valence insulators. What is
so specific in these materials, which makes them fundamentally different from the
other mixed-valence and heavy fermion systems?
When analysing their electronic structure, we can notice one common feature:
these are the materials which, if the f level had integral occupation, would have
been insulators. Thus, in ordinary conditions SmS (black SmS), containing Sm2+
289
ef
ef
Fig. 13.12
ef
k
Fig. 13.13
with the configuration 4f 6 , is a semiconductor with the energy gap Eg 0.2

0.3 eV. Under pressure it undergoes a phase transition into a gold modification
which looks like, and for a long time was treated as, a metal; the corresponding
transition was treated as an insulatormetal transition. However, more detailed
studies have shown that the high-pressure gold phase is in fact a small-gap
semiconductor. The standard explanation of the transition of SmS is that under
pressure the f level crosses the bottom of the initially empty conduction band, see
Fig. 13.12, and there appear electrons in the conduction band. Note, however, that
the situation here is rather specific: all the electrons in the conduction band are
actually former f electrons, so that their number is always equal to the number of
empty places, holes in the f level, ncond. = nf holes . This is just the situation which
is most favourable for the creation of a gap in the spectrum.
There are several factors which can lead to the opening of the gap at the Fermi
surface in this situation. If we simply use the fc hybridization (the term V c f
in (13.21)) we may think that the bands would simply hybridize and repel each
other, as shown in Fig. 13.13. And in the case considered (the total number of electrons is integer) the lower hybridized band may be completely filled and separated
by a gap from the upper empty band. That is why the gap in these insulators is
sometimes called the hybridization gap.
However, this explanation has several drawbacks. First of all we cannot treat
different electrons on an equal footing and consider the usual band hybridization
as though f electrons are uncorrelated. As we have already mentioned above, the
290
problem of how to hybridize correlated electrons is one of the main problems

in the entire field of heavy fermion and mixed-valence materials. The second
difficulty is that even if we could obtain the hybridized spectrum of this kind,
usually due to different dispersion in different directions the actual gap would not
form.
Thus we have to look for an alternative explanation of the insulating nature of
these materials. The fact that the number of conduction electrons here is equal
to the number of the remaining f holes is still very important, but it can play
a different role. We notice that the situation here is rather similar to the one in
excitonic insulators, considered in Section 11.4. As in that case, we can expect
here the formation of excitons due to the attraction of s electrons to f holes. As
we have here an equal number of them, each of the conduction electrons can find a
mate, and the formation of such a state would produce an insulator. In this picture
Kondo insulators would be very similar to excitonic insulators, and the gap would
have a collective nature.
There is yet another picture which can in principle lead to an insulating ground
state. When we considered above the Kondo lattice, we did not specify what was
the relative concentration of the localized moments nspin and of the conduction
electrons nc ; or rather we always assumed that nc > nspin . In that case, even if
the Kondo effect dominates, we would get a (nonmagnetic) metallic ground state,
albeit with strongly renormalized parameters. However, if nc = nspin , again we
would have a similar situation: each conduction electron would bind to one spin,
and in effect we may get an insulator. (This is actually what gave the name Kondo
insulator to these systems.) This would definitely be the case if the fs exchange
interaction J and the binding energy of the Kondo singlet TK were very large,
larger than the bandwidth of conduction electrons. Usually, however, we are in the
opposite limit, and the result is not so evident. Theoretically this problem is still
far from being solved: however, qualitatively it seems that the condition discussed
above (equal number of conduction electrons and localized spins of f holes) is a
necessary (and sufficient?) condition for the formation of Kondo insulators, and that
the physical mechanisms discussed above are indeed responsible for the creation
of this state.
13.5 Ferromagnetic Kondo lattice and double exchange

mechanism of ferromagnetism
As we have seen in the previous sections, there exists in the Kondo lattice a competition between the nonmagnetic (e.g. heavy fermion) and magnetically ordered
ground state. Magnetic ordering in this case is realized due to the RKKY exchange
13.5 Ferromagnetic Kondo lattice
291
interaction (9.29) of localized spins via conduction electrons

H JRKKY S0 Sr ,
JRKKY (r)
J 2 cos(2kF r + )
.
F
r3
(13.26)
What if we have a system with a small number of conduction electrons, kF 0?

It is clear from (13.26) that the interaction between spins will be nonoscillating (or
rather oscillating with a very large wavelength), and at least for the ferromagnetic
sign of JRKKY we can expect the formation of net ferromagnetic ordering.
Actually in this case we cannot use the standard treatment leading to the RKKY
interaction: this treatment was based on perturbation theory in J /F , and for small
electron density (small F ) this approach breaks down. One can nevertheless treat
this case proceeding from the opposite end treating the on-site exchange between
conduction electrons and localized spins (often called sd-exchange) as strong and
taking into account the electron motion (the kinetic energy of localized electrons)
later. Typically one considers in this case not an antiferromagnetic, but a ferromagnetic on-site exchange
HH JH Si i ;
(13.27)
physically this corresponds to the Hunds rule exchange which tends to make the
spins at the same atoms parallel.1 In this case we have a system which can be
called a ferromagnetic Kondo lattice; it is described by the model (13.22), but
with the opposite (ferromagnetic) sign of the exchange coupling. In application to
ferromagnetic metals, this model is known as the double exchange model.
Suppose that in this case the interaction JH is larger than the conduction electron
bandwidth. It is convenient in this case to go back to the coordinate representation
and rewrite the Hamiltonian (13.22) (in the tight binding approximation) as

ci cj Si .
H = t
ci cj JH
(13.28)
One can also have an interaction between localized spins on different sites, due to
some other mechanism not connected with the conduction electrons; it often has
the antiferromagnetic sign. Then the total Hamiltonian will have the form

ci cj Si + J
H = t
ci cj JH
Si Sj .
(13.29)
If there are no conduction electrons, the material would be antiferromagnetic due
to the exchange interaction J > 0. However, one can easily see that the inclusion
1
This approach is definitely valid for systems in which both localized and itinerant electrons belong to the
same partially filled d shell (to different d subbands thereof). This is the situation we are dealing with when
considering, e.g. ferromagnetism in metals like Fe or Ni and in compounds of the type La1x Cax MnO3 . When,
however, localized and itinerant electrons belong to different ions or have a completely different nature, e.g.
f electrons of Ce and conduction electrons in materials like CeAl3 , the exchange is of SchriefferWolf
type (13.12) and is antiferromagnetic.
292
s
q
S
1
2
(a)
2
(b)
2
(c)
Fig. 13.14
of conduction electrons (e.g. due to doping) would lead to an opposite tendency,

trying to make the system ferromagnetic.
The easiest way to see this is to consider the situation with t JH , treating the
spins quasiclassically. Due to the strong Hunds rule exchange (13.27) the spins
of conduction electrons tend to be parallel to the localized spins. In this case we
can have two situations; either the neighbouring spin is parallel, or antiparallel
to the first one, see Fig. 13.14. In the first case (Fig. 13.14(a)) the conduction
electron can easily hop to the nearest neighbour (this hopping is shown by the
wavy line in Fig. 13.14(a)). Thus in this case the conduction electron can delocalize and decrease its kinetic energy. However in the second case this process
is forbidden: the electron at site 2 would have spin opposite to the local spin
there, which for JH t would prevent such hopping (the energy loss JH would
exceed the energy gain t). To decrease its kinetic energy, the conduction electron should make all localized spins parallel, i.e. it will induce ferromagnetic
ordering.2
We can also use slightly different language and explain the tendency to ferromagnetism due to electron motion using Fig. 13.14(c): the electron from site 1
hops to the neighbouring site 2 and there, due to strong Hunds rule interaction,
pulls up the spin S2 , orienting it parallel to itself, i.e. in the result again making
all its neighbours and finally the total sample ferromagnetic (of course we need a
finite concentration of conduction electrons to accomplish this). This mechanism
of ferromagnetism in metals is called double exchange, it was first suggested in the
2
Strictly speaking, these arguments are not completely correct: if one takes into account the quantum nature of
spins, one can see that the on-site exchange (13.27) does not really require that the spins of the conduction
and localized electrons are parallel, but it tells us that the lowest state should be the one with the maximum
total spin (e.g. for S = 12 , a triplet state with Stotal = 1). This state, however, besides the classical possibilities
z
= 0; the corresponding wavefunction is
S and S, can have the z-projection of the total spin Stotal
1
( S + S). We see that there is a part of this wavefunction which corresponds to the state ( S)
2
obtained in the situation of Fig. 13.14(b) after electron transfer.
Thus theelectron hopping will not be completely
forbidden in this case, but will be reduced by the factor 1/ 2 (or 1/ 2S + 1 in the general case) due to the
normalization factor above. This can lead to certain observable consequences, though the gross picture of
ferromagnetism caused by kinetic energy gain of conduction electrons remains the same.
293
q
Fig. 13.15
1940s by Zener, and then studied in detail by Anderson, de Gennes, Nagaev, and
others.
One can notice that the physics of ferromagnetism in this case is rather similar
to the one we saw in the single-band model (Hubbard model) in Chapter 12: in both
cases the motion of electrons is hindered by the antiferromagnetic ordering of the
background (due to the same electrons in Chapter 12, but due to other, localized
electrons here), and the tendency to decrease the electron kinetic energy leads to
the suppression of antiferromagnetic ordering and eventually to the establishment
of ferromagnetism (or possibly some more complicated magnetic structure in the
case of the Hubbard model, e.g. RVB-like states).
When there exists direct SS exchange (the last term in (13.29)), these two mechanisms of magnetic ordering compete with one another. As a result a compromise
may be reached, e.g. in the form of a canted state: a two-sublattice structure, but
with the spins of the sublattices not exactly antiparallel, but canted at a certain
angle
= , see Fig. 13.14(c) and Fig. 13.15. As we have seen, the electron hopping is strongly influenced by the background magnetic structure. Treating spins
classically, one can get the following expression for the effective hopping matrix
element:
teff = t cos
.
2
(13.30)
Thus for the ferromagnetic ordering ( = 0) we have the full hopping teff = t, and
correspondingly the full bandwidth of conduction electrons, and for the antiferromagnetic case ( = ) we have teff = 0 (this approximation is definitely valid
for large spin S, but, strictly speaking, should be modified for smaller spins due
to quantum effects see the footnote above; we shall nevertheless ignore these
corrections and use below the expression (13.30)).
One can easily write down the energy of the system with electron concentration
x ( 1) and with canted spin structure. From (13.29), (13.30) one gets (taking
JH t, J ) the energy per site
E
= J S 2 z cos zxt cos .

N
2
(13.31)
Here z is the number of nearest neighbours. The first term in (13.31) is the energy of
localized spins forming two sublattices at an angle , Fig. 13.15. The second term
294
describes the kinetic energy of doped electrons with concentration x, moving in a

tight binding band with the hopping matrix element t cos 2 (13.30) (for x 1 we
can put all electrons at the bottom of the band and ignore the filling of higher-lying
states in the band, giving the energy x 5/3 ).
Minimizing the expression (13.31) in , we obtain
tx
.
(13.32)
=
2
4J S 2
Thus in this approximation already at small doping x the antiferromagnetic sublattices start to cant (de Gennes). At x = xc = 4J S 2 /t, the value of cos 2 reaches 1,
i.e. the angle between the sublattices becomes zero, = 0, and the system goes
over to a ferromagnetic state. This mechanism of ferromagnetism is essentially
due to the same tendency which we already saw before in Sections 12.5 and 12.6:
the ferromagnetic state is here stabilized by the decrease of the kinetic energy of
electrons which move much more easily in the ferromagnetic background than in
the antiferromagnetic background.
The double exchange mechanism of ferromagnetism described above may well
be the main mechanism of ferromagnetism in metallic ferromagnets such as iron
or nickel. It is also widely used to describe the properties of metallic ferromagnetic oxides, e.g. systems with colossal magnetoresistance La1x Mx MnO3
(M = Ca, Sr). And although many details of the behaviour of such systems are still
not clear3 the tendency which we saw above that the ferromagnetic state usually
goes hand in hand with metallic conductivity, and the antiferromagnetic state is
more typical for insulators is definitely observed in most of the magnetic materials
(although of course there are also many exceptions to this general tendency).
cos
Concluding this book, I want to say once again that I have tried to give here at
least an impression of the problems physicists are now working on in condensed
matter physics, the language used and the methods employed. Of course I could not
3
Besides the possible importance of quantum effects mentioned above, there exists in the double exchange
model (13.29) the tendency to an instability of the homogeneous canted state towards phase separation into
metallic ferromagnetic regions containing all the doped electrons, and antiferromagnetic insulating regions. This
tendency is again similar to the one we saw in the single-band Hubbard model, Section 12.7: one easily sees
that the total energy of the homogeneous canted state which we would obtain by putting the value cos 2 (13.32)
into the expression (13.31), would be such that d 2 E/dx 2 < 0. But this is nothing else but the inverse compressibility of the system. A negative value of the compressibility is forbidden by the general rules of thermodynamics;
it means absolute instability of the corresponding homogeneous state. There are, however, other contributions
to the total energy of the system besides the one considered above, notably the long-range Coulomb forces
which favour electroneutrality and thus oppose the tendency to phase separation (which nevertheless seems to
persist, albeit in a modified form) see the discussion in Section 12.7.
295
cover all the interesting and exciting problems in this big field, as well as introduce
all the theoretical methods; for instance I have left out the big and very important
field of transport phenomena. I hope, however, that I have been able to present some
of the unifying concepts underlying modern condensed matter physics, especially
the concepts of order and elementary excitations. I have also tried, and I hope was
able to succeed to some extent, to show, that this relatively old and classical part
of physics is still full of surprises, alive and kicking.
Bibliography
A. A. Abrikosov, L. P. Gorkov and I. E. Dzyaloshinsky, Methods of Quantum Field

Theory in Statistical Physics, Dover, New York, 1975.
N. W. Ashcroft and N. D. Mermin, Solid State Physics, Harcourt College, New York,
1976.
G. Baym and Ch. Pethick, Landau Fermi-liquid Theory: Concepts and Applications,
Wiley-Interscience, New York, 1991.
P. M. Chaikin and T. C. Lubensky, Principles of Condensed Matter Physics, Cambridge
University Press, Cambridge, 2000.
S. Doniach and E. H. Sondheimer, Greens Functions for Solid State Physicists,
Benjamin/Cummings, New York, 1974.
P. Fazekas, Lecture Notes on Electron Correlations and Magnetism, World Scientific,
Singapore, 1999.
A. L. Fetter and J. D. Walecka, Quantum Theory of Many-Particle Systems, Dover, New
York, 2003.
A. Georges, G. Kotliar, W. Krauth and M. Rozenberg, Rev. Mod. Phys., 68, 13 (1996).
J. B. Goodenough, Magnetism and the Chemical Bond, Interscience, New York, 1963.
A. C. Hewson, The Kondo Problem for Heavy Fermions, Cambridge University Press,
Cambridge, 1993.
D. I. Khomskii, in Spin Electronics, ed. M. Ziese and M. J. Thornton, Springer, Berlin,
2001, p. 89.
C. Kittel, Quantum Theory of Solids, John Wiley, New York, 1987.
C. Kittel, Introduction to Solid State Physics, John Wiley, New York, 2004.
K. I. Kugel and D. I. Khomskii, Sov. Phys. Uspekhi, 25, 231 (1982).
L. D. Landau and I. M. Lifshits, Statistical Physics, Addison-Wesley, Reading, MA, 1969;
3rd edn Part 1, ButterworthHeinemann, London, 1980.
L. D. Landau and I. M. Lifshits, Quantum Mechanics: Non-Relativistic Theory, 3rd edn,
ButterworthHeinemann, London, 1977.
L. D. Landau and I. M. Lifshits, Fluid Mechanics, ButterworthHeinemann, 1987.
L. D. Landau and Ya. B. Zeldovich, ZhETF, 32, 1944 (1943a).
L. D. Landau and Ya. B. Zeldovich, Acta Phys. Chem. USSR, 18, 194 (1943b).
H. von Lohneisen, A. Rosch, M. Vojta and P. Wolfle, Rev. Mod. Phys., 79, 1015 (2007).
G. D. Mahan, Many-Particle Physics, Springer, Berlin, 2000.
M. P. Marder, Condensed Matter Physics, Wiley-Interscience, New York, 2000.
R. D. Mattuck, A Guide to Feynman Diagrams in the Many-Body Problem, Dover, New
York, 1992.
296
Bibliography
297
R. E. Peierls, Quantum Theory of Solids, Oxford University Press, Oxford, 2001.

R. E. Peierls, More Surprises in Theoretical Physics, Princeton University Press,
Princeton, NJ, 1991.
S. Sachdev, Quantum Phase Transitions, Cambridge University Press, Cambridge, 1999.
J. R. Schrieffer, Theory of Superconductivity, Perseus Books, New York, 1999.
J. S. Smart, Effective Field Theories of Magnetism, Saunders, New York, 1966.
H. E. Stanley, Introduction to Phase Transitions and Critical Phenomena, Oxford
University Press, New York, 1987.
R. M. White, Quantum Theory of Magnetism, Springer, New York, 2006.
K. Yosida, Theory of Magnetism, Springer, New York, 1996.
J. M. Ziman, Principles of the Theory of Solids, Cambridge University Press, Cambridge
1979.
D. N. Zubarev, Usp. Fiz. Nauk, 71, 71 (1960).
Index
AbrikosovSuhl resonance, 280

Actinide compounds, 73, 282
Anderson
lattice, 283, 286
localization, 267
model, 273277, 283
Angle-resolved inverse photoemission, 146
Angle-resolved photoemission (ARPES), 146
Anharmonic effects, 41
Anisotropy, 86
Anomalous average, 210, 214, 220, 222
Anomalous Green functions, 210, 214
Antiferromagnetic ordering, 88, 98, 101, 109, 186,
220, 239, 245, 257
in Mott insulators, 234
in the Heisenberg model, 108
in the Ising model, 106
Landau theory, 84
Arrott plot, 82
Baber law, 178
BardeenCooperSchrieffer theory, 126, 185, 208,
212215, 219, 279
Benzene molecule, 112, 198
Bethe Ansatz, 108, 110
Bogolyubov
canonical transformation, 59, 98, 100, 103, 194,
197, 210, 213215
sound, 61, 64, 65, 102, 120
Bohrvan Leeuwen theorem, 72
Boltzmann statistics, 32, 74
BonnerFisher maximum, 203
Born approximation, 224, 278
Bose
condensation, 50, 5458, 6265, 102, 120, 196, 210
statistics, 31
Bose gas, 5658, 6062, 120
Brillouin function, 75
Canted antiferromagnetism, 245, 249
Charge-density wave, 193, 195, 196, 201, 206,
215217, 220222, 266
Charge-transfer energy, 260

Charge-transfer insulators, 233, 263, 264
Chemical potential, 4, 32, 38, 54, 129
Coherence length, 19, 122
Cold melting, 227
Colossal magnetoresistance, 265, 294
Compressibility, 3, 10, 181, 251, 294
Confinement, 242, 243, 247
Cooper pairs, 95, 126, 196, 197, 208, 213, 280
Correlation energy, 174
Correlation length, 19
Coulomb interaction, 128, 133, 135, 147, 151, 152,
155, 158160, 163, 164, 174, 216, 223, 225,
230232, 250, 255, 256, 260, 273
Creation and annihilation operators, 35, 134
Critical exponents, 30
Critical indices, 22
Critical points, 28
Critical temperature, 6, 7, 84, 148, 184
Critical velocity (superfluidity), 66, 69
Curie
constant, 74
law, 74
temperature, 6
CurieWeiss law, 23, 79, 80, 85, 203
de Boer parameter, 50, 228
Debye
frequency, 40, 212, 215
model, 39
screening, 147, 163
temperature, 40
wavevector, 40
DebyeWaller factor, 49
Diamagnetism, 71
Landau, 72, 73
Dichalcogenides, 13, 206
Dielectric function, 160
Dimerization, 189, 190, 192, 193, 199, 202, 204, 221
Dispersion relation, 144, 183
Domain walls, 106, 110, 112, 113, 199, 200, 205
Doniach phase diagram, 285
298
Index
Double exchange model, 291294
Drude conductivity, 178
DulongPetit law, 39, 40
Dynamical mean field theory, 269
Dyson equations, 153155, 159, 160
Ehrenfest relations, 11
Einstein model, 39
Elementary excitations, ix, xiii
Enthalpy, 2
Entropy, 2, 6, 9, 19, 24, 27, 118, 122, 131, 178, 182,
270
of second-order phase transition, 9
Equations of motion method, 95, 149
Excitonic condensate, 197, 208, 212
Excitonic insulators, 207, 208, 210, 211, 213, 214,
216, 218, 220, 232, 290
Excitons, 197, 208, 209, 213, 215, 232, 290
Fermi gas, 127129, 131, 143, 145, 175, 223
Fermi liquid, 24, 147, 175186, 188, 226, 228, 230,
231, 238, 239, 246, 268, 269, 281, 284, 288
Fermi statistics, 31
Ferroelectricity, 15, 89, 219
Ferromagnetic insulators, 258
Ferromagnetic ordering, 88, 92, 96, 101, 245, 246,
257
in the Heisenberg model, 108
in the Hubbard model, 244
in the Ising model, 106
in the xy model, 119
Landau theory, 80
Ferromagnetic polarons, 247
Ferrons, 247
Feynman diagrams, 43, 134, 150157, 160, 162, 170,
171, 176, 208, 212, 224, 228
First-order phase transition, 6, 11
Flux phase, 222
Free energy
Gibbs, 2, 4, 6, 8, 11, 16, 19, 80, 84
Helmholtz, 2, 41, 43, 122
Friedel oscillations, 165, 168
Frohlich
collective mode, 201
conductivity, 202
Hamiltonian, 147, 192
interaction, 147
Frustrated phase separation, 250
Frustration, 89, 115, 118
Giant Kohn anomaly, 168, 169, 188, 193, 209
Gibbs
distribution function, 1
free energy, 2
Gibbs free energy, 4, 6, 8, 11, 16, 19, 80, 84
Ginzburg number, 19
GinzburgLandau
equation, 16
functional, 16
theory, 1618
GinzburgLevanyuk criterion, 22
299
Goldstone
modes, 28, 64, 106, 183
theorem, 27, 53, 101, 112
GoodenoughKanamoriAnderson (GKA) rules, 257
Gorkov equations, 210, 214
Grand partition function, 5
Graphene, 198
Green functions, 139144, 146153, 155, 156, 159,
162, 163, 167, 170, 171, 173, 176, 177, 184, 208,
213215, 228, 236, 238, 268, 273
anomalous, 210, 214
spectral representation, 143
Gruneisen
approximation, 43
constant, 43, 44, 132
equation, 44, 131
Gutzwiller method, 251, 268
Haldane gap, 113
Harmonic oscillator, 34
Hartree term, 152, 153
HartreeFock approximation, 152, 174, 226, 236, 273
Heavy fermions, 158, 178, 185, 264, 282, 284287
Heisenberg
model, 23, 76, 100, 106, 108, 123, 203
representation, 96, 136, 150
Helmholtz free energy, 2, 41, 43, 122
High-temperature superconductivity, 24, 95, 112, 116,
185, 186, 213, 220, 229, 242, 243, 247, 250, 258,
260, 282
Holons, 117, 186, 199, 242
HolsteinPrimakoff transformation, 92, 107
Hubbard
interaction, 158, 162
model, 186, 217, 220, 230, 231, 233252, 254263,
268, 277, 293, 294
subbands, 237, 238, 241, 264, 270
Hubbard I decoupling scheme, 236, 238, 245
Hunds rule, 252, 256, 291, 292
Hybridization gap, 289
Improper ferroelectrics, 15
Impurities, 123, 272276, 280284
Insulatormetal transitions, 265271
Interaction representation, 136, 150
Inverse photoemission (IPES), 146
IoffeRegel limit, 267
Ising
ferromagnets, 26
interaction, 102, 243
model, 23, 100, 105, 108, 117
Isomorphous phase transitions, 29
Itinerant magnets, 185
JahnTeller
effect, 254256
theorem, 252
JordanWigner transformation, 107, 202
Kagome lattice, 89, 116
Kohn anomaly, 168
300
Index
Kondo
effect, 165, 276278, 280, 281, 283285
insulators, 282, 288290
lattice, 283, 284, 286, 290, 291
resonance, 280, 281
temperature, 279, 281, 284, 285, 287
KosterlitzThouless transition, 122
Kramers degeneracy, 253
Landau
criterion of superfluidity, 65
diamagnetism, 72, 73
interaction function, 179
theory of phase transitions, 716, 19, 80,
270
LandauLifshits equation, 17, 101
Lifshits
invariant, 18
point, 18, 204
Ligand field splitting, 258
Ligands, 258
Lindemann criterion, 49
Lindhardt function, 163
Local spin density approximation, 267
Luttinger
liquid, 186, 200
theorem, 146, 176, 238
Magneli phases, 268
Magnetic susceptibility, 21, 71, 74, 78, 80, 122, 161,
203, 216, 269, 287
differential, 81
of Fermi liquid, 181
Pauli, 132
Magnetite, 226
Magnons, 64, 9297, 99104, 109, 120, 215
MajumdarGhosh state, 90
Marginal Fermi liquid, 183
Mass renormalization, 157, 158, 178
Mean field approximation, 77
Meissner effect, 71
Melting, 46, 47
of Wigner crystal, 227
quantum, 50
MerminWagner theorem, 53, 64, 103, 106, 121, 122
Metamagnetism, 87
Migdal theorem, 157
Mixed valence, 264, 282, 286288
Mobility gap, 267
Mossbauer effect, 49
Mott
insulator, 26, 220, 232235, 237, 259, 261, 263, 270
minimum metallic conductivity, 267
transition, 26, 238, 264, 269, 270
Multiferroics, 18, 89, 219
Nagaoka theorem, 244, 245, 247
Neel
state, 76, 90, 98, 99, 103, 105, 110, 111, 113, 202,
240
temperature, 79, 84, 241, 267
Nernst theorem, xiii, 24, 131

Nesting, 169, 206, 208, 216, 219, 220
Non-Fermi-liquid systems, 145, 175, 183, 184, 186,
187, 200, 282, 283
Onsager solution, 117
Orbital ordering, 254258
ferro-orbital and antiferro-orbital, 255
Orbitons, 257
Order, ix, xiii
Order parameter, 6, 25, 64, 68, 84, 102, 118, 123, 182,
210, 222, 255, 270
OrnsteinZernike theory, 21
Overhauser instability, 216
Paramagnetism
localized spins, 73
Pauli, 24, 72
Van Vleck, 71
Partition function, 1
Pauli
paramagnetism, 24, 72
principle, 31, 127, 176178, 234, 240, 241,
256
Paulions, 91
Peierls
distortion, 167, 194, 198, 199, 220, 221
instability, 167, 188, 190, 193, 202, 208, 209
transition, 188202, 206, 208, 215, 217, 218, 222,
266
Phase separation, 248250, 252
frustrated, 250
Phonons, 3749, 62, 63, 100, 131, 147, 151, 155,
167
Plasma frequency, 170
Plasmons, 171, 172
Polarization operator, 153, 155, 160162, 166, 169,
183, 193, 206
Polarons, 243, 247
Polyacetylene, 168, 198, 199
Pomeranchuk criteria, 181, 188
Pyrochlore lattice, 116
Quadrupolar ordering, 255
Quantum critical point, 24, 184, 185, 227, 283,
286
Quantum crystal, 51
Quantum melting, 50
Quantum phase transitions, 23
Quarks, 200, 243
Quasiharmonic approximation, 43
Raising and lowering operators, 91
Random phase approximation, 160, 165, 171, 173,
174, 208
Rare earth compounds, 73, 89, 175, 226, 229, 233,
255, 272, 282, 284, 285
Renormalization
mass, 157, 158, 178
wavefunction, 144
Renormalization group, 23
Index
301
Resonating valence bonds, 112116, 186, 187, 198,

199, 293
Reynolds number, 30
Rotons, 62, 66, 69
RudermanKittelKasuyaYosida (RKKY)
interaction, 165, 166, 283, 284, 290, 291
high-temperature (high-Tc ), 24, 95, 112, 116, 185,

186, 213, 220, 229, 242, 243, 247, 250, 258, 260,
282
type-II, 69
Superexchange, 235, 251, 277
Superfluidity, 51, 6269, 121
sd exchange model, 276

Scaling relations, 22, 118
Schottky anomaly, 106
SchriefferWolf transformation, 277, 291
Schrodinger representation, 96, 136
Second-order phase transition, 611, 19, 22, 47, 57,
80, 83, 118, 185, 271
Self-energy, 153156, 184
Semimetals, 208
SlaterKoster method, 126
Soft mode, 47
Soliton lattice, 205
Solitons, 110, 199, 200, 205
Sommerfeld integrals, 130
Specific heat, 3, 9, 39, 40, 106, 118, 131, 178, 181,
185, 284
Spectral function, 143, 145147
Spin-density wave, 89, 196, 206, 212, 215222, 218,
235, 245, 246, 266
Spin glass, 166
Spin liquid, 90, 116
Spincharge separation, 186, 187, 199, 200
Spin-bag, 247
Spin-flip transition, 86
Spin-flop transition, 86
Spin-Peierls systems, 18
Spin-Peierls transition, 202204
Spinodals, 12, 14
Spinons, 112, 186, 187, 199, 242
Statistics
Boltzmann, 32, 74
Bose, 31
Fermi, 31
Stoner criterion, 216, 225, 274
Stripes, 251
Strongly correlated electrons, 176, 230, 243,
258
SuSchriefferHeeger (SSH) solitons, 199
Superconductivity, 69, 95, 112, 116, 126, 148, 157,
176, 183, 185, 197, 202, 208, 212, 214, 219, 220,
229, 244, 247, 250, 258, 265, 280, 282
tJ model, 251, 252

Thermal expansion, 3, 10, 43, 131
ThomasFermi screening, 164
Tight binding approximation, 107, 168, 169, 188, 219,
220, 230, 291
Tin plague, 266
Topological excitations, 68, 120, 122
Transfer matrix, 105
Transition metal compounds, 13, 73, 206, 226, 229,
232, 241, 252, 258, 264, 268, 272
Tricritical point, 14
Trugmans trajectories, 243
Turbulence, 30
Umklapp wavevector, 49, 190
Universality classes, 22
Valence bonds, 111, 114, 202
Valence-change transitions, 282
Van Vleck paramagnetism, 71
Verwey transition, 226
Vortices, 6769, 120, 122
Wavefunction renormalization, 144
Weiss temperature, 79
White and grey tin, 265
Wigner crystal, 50, 51, 226228, 231
Wilson
loops, 243
ratio, 132, 181, 281
xy model, 64, 107, 119, 202, 203
ZaanenSawatzkyAllen (ZSA) scheme, 233, 263
Zeeman
splitting, 75
term, 73
Zero-point oscillations, 35, 49, 50, 52, 99101, 103,
104
Zero sound, 181
ZhangRice singlets, 265

(Daniel I. Khomskii) Basic Aspects of The Quantum

Uploaded by

Document Informationclick to expand document information[Daniel I. Khomskii] Basic Aspects of the Quantum

Document Informationclick to expand document information

Copyright:

Available Formats

(Daniel I. Khomskii) Basic Aspects of The Quantum

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Daniel I. Khomskii) Basic Aspects of The Quantum

Uploaded by

Copyright:

Available Formats

This page intentionally left blank

BASIC ASPECTS OF THE QUANTUM THEORY OF SOLIDS

BASIC ASPECTS OF THE QUANTUM

CAMBRIDGE UNIVERSITY PRESS

Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore,

Cambridge University Press has no responsibility for the persistence or accuracy

Foreword and general introduction

4.4 Anharmonic effects

8.5 Spectral representation of Green functions

12.4 One-particle spectrum of strongly correlated systems

Foreword and general introduction

Foreword and general introduction

I go step by step through different specific situations: systems of bosons (phonons

Foreword and general introduction

Foreword and general introduction

Foreword and general introduction

9. P. M. Chaikin and T. C. Lubensky (2000), Principles of Condensed Matter

Foreword and general introduction

of magnons, etc. There also exist interactions between different quasiparticles:

1.1 Gibbs distribution function and partition function

from which we find

Here Z is the partition function

where H is the Hamiltonian of the system. Thus

The entropy is defined as

where E is the average energy of the system, E =

En eEn /T . We can intro-

which is called the (Helmholtz) free energy:

1.2 Thermodynamic functions

Enthalpy is often used in discussions of chemical reactions, thermodynamics of

1.2 Thermodynamic functions

from which we obtain

One can express cP , cV through F , , using (1.12), (1.15).

and the compressibility

1.3 Systems with variable number of particles; grand partition function

From these equations we get:

Here we have introduced the chemical potential which is defined by

From (1.25) and (1.23) we obtain

For a variable particle number N, it takes the form

1.3 Systems with variable number of particles

Thus  is a generalization of the free energy to the case of variable number of

2.1 Second-order phase transitions

General theory of phase transitions

with the coefficient a > 0. Then

This behaviour is shown in Fig. 2.3.

2.1 Second-order phase transitions

This behaviour is shown in Fig. 2.4(b).

General theory of phase transitions

Remembering that the thermal expansion coefficient is = 3 =

Similarly, from the condition

2.2 (Weak) First-order phase transitions

this gives the relation

2.2 (Weak) First-order phase transitions

General theory of phase transitions

2.2 (Weak) First-order phase transitions

We have assumed previously that the coefficients A(P , T ) = a(T Tc ) and

General theory of phase transitions

2.3 Interaction with other degrees of freedom

Thus is a generalization of the free energy to the case of variable number of

and the laminar flow is stable, and

which is valid at T . This behaviour is schematically shown in Fig. 3.1; we