Spin glasses:

an experimental
introduction
Schrtig in dem eisernen Gitter
der drehende Spin
am Drehsinn
errtitst du die Seele

Frei nach Paul Celans ‘Sprachgitter

Spin glasses:
an experimental
introduction

J. A. Mydosh

Taylor & Francis

London Washington, DC
l

1993
UK Taylor & Francis Ltd, 4 John St., London WClN 2ET
USA Taylor & Francis Inc., 1900 Frost Road, Suite 101, Bristol PA
19007

0 J. A. Mydosh 1993
All rights reserved. No part of this publication may be reproduced,
stored in a retrieval system, or transmitted, in any form or by any
means, electronic, electrostatic, magnetic tape, mechanical,
photocopying, recording or otherwise, without the prior permission
of the copyright owner and publishers.

British Library Cataloguing in Publication Data

are available from the British Library
ISBN o-7484-0038-9

Library of Congress Cataloging-in-Publication Data

are available

Cover designed by Amanda Barragry

Printed in Great Britain by Burgess Science Press, Basingstoke on

paper which has a specified pH value on final paper manufacture
of not less than 7.5 and is therefore ‘acid free’.
Contents

. ..
PREFACE Vlll

1 INTRODUCTION TO RANDOMNESS AND FREEZING 1

1 .l What is a spin glass? 1

1.2 How to create randomness? 4
1.3 Magnetic interactions (are necessary) 6
1.3.1 Direct exchange 6
1.3.2 RISKY 7
1.3.3 Superexchange 9
1.3.4 Dipolar 10
1.35 Mixed superexchange (random bonds) 10
1.4 Magnetic anisotropy (plays a role) 11
1.5 Frustration (a far-reaching concept) 13
1.6 The freezing process (a first look) 15
1.7 Some historical background 16
Further reading 18

2 DESCRIPTION OF RELATED CONCEPTS 20

2.1 Virtual-bound-state model 20

2.2 Kondo effect and weak moments 24
2.3 Giant moments 28
2.4 Spin glasses (a second visit) 30
2.5 Mictomagnetism (cluster glass) 32
2.6 Percolation and long-range magnetic order 34
2.7 Concentration regimes 35
2.8 Metallic glasses: amorphous magnets 37
2.9 Re-entrant spin glasses 39
2.10 Superparamagnetism 40
2.11 Magnetic insulators and semiconductors 42
Further reading and references 44
vi Contents

3 BASIC SPIN-GLASS PHENOMENON 45

3.1 High-temperature (T % Tr) experiments 45

3.1.1 Susceptibility 46
3.1.2 Specific heat 50
3.1.3 Resistivity 53
3.1.4 Neutron scattering 57
3.2 Experiments spanning Tf 63
3.2.1 AC-susceptibility 64
3.2.2 Specific heat 76
3.2.3 Resistivity 78
3.2.4 Neutron scattering 80
3.2.5 $SR (muon spin relaxation) 83
3.2.6 Mossbauer effect 86
3.3 Low-temperature (T 4 Tf) experiments 88
3.3.1 Magnetization 89
3.3.2 Specific heat 98
3.3.3 Resistivity 100
3.3.4 Neutron scattering 101
3.3.5 Torque and ESR 105
3.4 Spin glasses in a field 111
3.5 Experimentalist’s (intuitive) picture 113
References 116

4 SYSTEMS OF SPIN GLASSES 118

4.1 Transition-metal solutes 119

4.1 .l Noble metals with transition-metal impurities 119
4.1.2 Transition-metal/transition-metal combinations 120
4.1.3 Collection of characteristic temperatures for AuFe
and CuMn 122
4.2 Some rare-earth combinations 125
4.3 Amorphous (metallic) spin glasses 127
4.4 Semiconducting spin glasses 129
4.5 Insulating spin glasses 131
4.6 What is a good spin glass? 133
References 134

5 MODELS AND THEORIES 135

5.1 Some historical perspectives 136

5.2 The Edwards-Anderson model 139
5.3 Sherrington-Kirkpatrick model 144
5.4 Instability of the SK solution 147
Contents Vii

5.5 The TAP approach 148

5.6 How to break the replica symmetry 150
5.7 Physical meaning of RSB 153
5.8 Dynamics of the mean-field model 161
5.9 Droplet model 164
5.10 Fractal-cluster model 167
5.11 Non-Ising spin glasses 171
5.12 Computer simulations 174
Further reading and references 179

6 IDEAL SPIN GLASSES: COMPARISONS WITH THEORY 181

6.1 What is an ‘ideal’ spin glass? 182

6.2 RbZCu&oxF~, an ideal 2D spin glass 184
6.3 Ideal 3D spin glasses 194
6.4 Future possibilities 200
References 203

7 SPIN-GLASS ANALOGUES 204

7.1 Electric dipolar and quadrupolar glasses 205

7.2 Superconducting-glass phases 212
7.3 Random-field Ising model 216
7.4 Other ‘spin-offs’ 222
Further reading and references 231

8 THE END (FOR NOW) 233

8.1 Symmetry breaking and phase space 233

8.2 Mesoscopic spin glasses 237
8.3 Recent theoretical progress 241
8.4 A final word 246
References 248

INDEX 251
Preface

Phenomena that exist in nature should not only be observed and classified
but describable by models or theories which then lead to a better
understanding and predictability of the phenomenon. Often, as is the case
with the spin glasses, experiment has serendipitously uncovered some
unusual or unexpected effect which at first glance baffles the anticipations
of our physical intuition. As a result of the novel experimental properties,
the phenomenon requires further study and a preliminary picture or concept
to represent it. Here a name should be given. With the accumulation of
data mathematical notions are needed to quantify the results and an incipient
theory is born. The phenomenon, its various experimental manifestations,
the many (hopefully) specific examples which can be found, and the
embracing theory, all developing over the course of many years, create a
problem area, in our case the ‘spin-glass problem’. Fresh input comes from
the drawing of various analogies with other areas, new anomalies are
discovered by more sophisticated experiments, and theoretical progress
leads to a deeper understanding and predictive power. Hence the problem
is nicely on its way to being solved.
Frequently there is a long pause after an important discovery and the
recognition of its significance. Time and ideas are necessary to set the stage
for further advancement and the delay can be great. With some luck the
given phenomenon will become a major effect, i.e., an important topic of
contemporary research. Hectic efforts, measured by the large number of
workers presenting their results at countless conferences and workshops,
are demanded. And all this could continue for many years before a (near)
solution is found or the ideas and curiosity run out. Thus, for a particular
period the area becomes highly fashionable or ‘in-physics’, and then after
its fame and topicality diminish, the problem, if still unsolved, may remain
as an interesting, yet pmsC research topic. Nevertheless during these intense
years, a great deal was learned and this knowledge will remain as an
integral part of physics and even appear in tomorrow’s text books.
So it was and is with the spin glasses. Around the late 1960s there were
some unusual effects observed on a series of magnetic alloys (rather simple
and traditional ones such as iron in gold and manganese in copper).
X Preface

However, then the time was not right but at least a name was coined. Only
in the mid 1970s did theory using some radically new ideas come to grips
with the experiments and the possibility of a sharp phase transition. Around
1975 ‘all hell broke loose’ and this began the fame period of the spin glasses
with great activity and much progress. Here analogies and spin-offs increased
the pace. After 1985 the tempo slowed down and a third phase of tranquility
appeared with more sophisticated theories, subtle new experiments and
‘ideal’ spin-glass materials. The phenomenon had run its course to
establishment so that now is perhaps a good time to try and tell its story,
naturally in a phenomenological way.
The purpose of this book is indicated in the title. It is meant to introduce
the phenomenon and physics of spin glass (Chapters 1 and 2) via an
experimental approach (Chapter 3). The method is to use simple physical
pictures and concepts based on the empirical observations, rather than
employing theoretical models with the associated mathematics. Hence,
figures and tables will be more common than equations and derivations.
Materials, i.e., the real spin glasses, will also play a major role (Chapter
4). In spite of this particular emphasis the theory of spin glasses is an
essential part of the story and its ideas, techniques and advances will be
portrayed in an elementary fashion (Chapter 5). How close does experiment
come to theory? Chapter 6 considers the recent advances in finding and
measuring spin glasses which closely mimic the theoretical models.
Why write such a book on spin glasses? At least four good reasons can
be given.

1 The spin glass has become a fundamental and general form of magnetism;
its general occurrence as ‘magnetic ordering’ occupies the third place
after ferromagnetism and antiferromagnetism.
2 Randomness, frustration, glassiness and amorphousness represent very
important phenomena in contemporary physics.
3 Radically new concepts and ideas are needed to describe the basic
experimental behaviours.
4 There is a richness of analogies with many other areas extending from
astrophysics via molecular evolution to zoology. A few of these analogies
are examined in Chapter 7.

The spin glasses were surely one of the first precursors to the physics of
complexity - a new area which has recently been established.
For the above reasons a simple, less theoretical treatment of the problem
should be made available for the non-theorist or even non-physicists, and
also the younger (graduate) students in order to serve as an introduction
to the subject and its basic physics. Thus, the level of the book is low; only
a general background in solid-state physics and elementary magnetism is
required beyond the undergraduate courses, for the remaining necessary
concepts are explained, many in Chapter 2 and others as needed. Emphasis
Preface Xl

is placed on experimental facts and physical notions with always a real spin-
glass system in mind. The powerful developments in the theory - the new
statistical mechanics - will be sketched and compared to experiment but
not discussed in mathematic details or via long derivations. Concerning the
theory of spin glasses there are a number of texts and review articles already
available here and it would be unwise and unnecessary to compete with
these excellent theoretical exposes. What is missing is the elementary level,
rudimentary explanation of the spin-glass problem for just about everyone.
More so the novice or newcomer to the field and the expert from another
area are the welcome readers. Perhaps the real spin-glass connoisseurs
would value the book as a reference to the salient experiment features and
an indication of the theoretical lines and directions of progress.
The author has been too long occupied with the phenomenon of spin
glasses. First there were the quiet years of experimentation, then the hectic
decade when it seemed like everyone was working on the problem, and
finally, now that the dust has settled a bit, a chance to look back at the
past 25 or so years and see what knowledge has been gained and what we
really understand. Although my research emphasis has changed to other
topics and other problems, nevertheless, my ‘first love’ remains the spin
glasses. Before forgetting everything it’s time to pause for a moment and
in retrospection write this book as quickly as possible. The season is
certainly ripe, the need is great and a new generation of students is moving
into this research area. But such an undertaking is not meant as the final
word; the field is not finished. The spin-glass problem has not been fully
solved, there is still movement, albeit slow, towards a more complete
theoretical description with yet newer ideas. Also experimentation is
continuing with novel forms of spin glasses and even more subtle effects.
Hopefully then a more finalized version of this spin-glass text will appear
in the next ten years. Thus my apologies now for presenting the yet
incomplete and modifiable ‘truth’.
Although the main ideas and threads of thought for this book were
conceived at the Kamerlingh Onnes Laboratory in Leiden, most of writing
took place at the University of California in Santa Barbara during a
sabbatical visit of six months. I must thank V. Jaccarino for his kind
hospitality at UCSB. Also I would like to acknowledge the support and
encouragement of M. H. Vente during the rather long period of gestation,
writing and correcting. M. Nieuwenhuijzen-Verschoor did a superb job in
typing the manuscript. There were simply too many discussions and
interactions some of them long-forgotten during the past 25 years which
moulded my views on the spin glasses. I should mention the experimental
years at Fordham University in New York and Institut fur Festkorperfor-
schung in Jtilich, and the changing times at Leiden. All these places and
their staffs, my colleagues, contributed to a rather particular and sometimes
personal interpretation of the spin glasses which is related in this book. I
hope it reads a bit like an adventure novel as so many parts have seemed
to me, none the least of which was the actual writing. Read well.
Introduction to randomness
and freezing

All beginnings are difficult, but we must somehow start, so let us do it as

simply as possible. Pictures are worth many words, thus, some oversimplified
sketches are used to illustrate what a spin glass is and how we create the
all-important randomness. In addition to the disorder we must have magnetic
interactions, and these are also treated in a pictorial way. The other
important ingredients of a spin glass, anisotropy and frustration, are
introduced via easy examples. After formulating a working definition of a
spin glass we take a first look at the freezing process. The spin glass does
indeed constitute a new state of magnetism, distinctly different from the
long-range ordered ferro- and antiferromagnetic phases. Yet similar to these
magnets, the spin glass also has a co-operative or collective nature in the
frozen state. And this is what we must try to understand and put into a
clear physical presentation with the aid of ‘solvable’ theoretical models. In
order to satisfy the biographic aspects we give a very brief account of the
historical development of the spin-glass problem and how the name was
coined in 1970. With such an outline of the goals of Chapter 1, let us
plunge into the spin glasses.

1.1 WHAT IS A SPIN GLASS?

We shall begin by giving a simple physical picture of a frozen spin glass.

This can be accomplished by playing the ‘coupling game’ between randomly
distributed magnetic moments, i.e., spins on a simple, two-dimensional
(2D), square lattice. Once we know what the frozen configuration (say at
temperature T = 0 K) looks like, we can attempt to track down the basic
ingredients which have created this state. Afterwards we can increase the
temperature to study the formation and precursors of our frozen spin glass.
At least then, we shall know the type of physical phenomenon we are
dealing with.
Let us imagine the square lattice whose sites are represented by dots as
2 Introduction to randomness and freezing

depicted in Fig. 1.1. A certain fraction x of the sites are occupied by a

magnetic moment or spin. These are randomly distributed throughout the
lattice and are designated by arrows. We keep x small (4) to avoid too
many nearest-neighbour occupancies. Now let us ‘turn on’ a magnetic
interaction among the moments. The coupling algorithm is ferromagnetic
(parallel) between nearest neighbours and antiferromagnetic (antiparallel)
between next-nearest neighbours. Such a situation is illustrated in Fig. 1.1
where we have removed any thermal disorder by reducing the temperature
to a negligible value. To begin our coupling game we arbitrarily choose the
alignment of the top right moment to point upwards. As the reader can
verify by proceeding downwards, the given configuration then results. Note
there are some spins or arrows which are completely decoupled from the
larger clusters (3 fully alone and 2 pairs). Most spins enjoy a happily
connected environment that has propagated from top to bottom. Further a
nice balance (zero net moment) exists between the connected up and down
spins (9 up versus 9 down). However, a small number of spins are unhappy,
for they cannot satisfy their second-neighbour bonds. In particular, for the
marked spin (large dot) at the lower right of Fig. 1.1 there is one first-
neighbour coupling with strength Jr > 0 (ferromagnetic) and two second-
neighbour couplings with strength J2 < 0 (antiferromagnetic). If IJr( = 21J,(,
this spin will not know which way to point. Thus, it is ‘frustrated’ and will
communicate its discontent to the two spins below itself.
A striking example of a completely frustrated spin-lattice is a triangular

r( 0 0 0 0 + Origin

Fig. 1.1 Frozen spin glasswithJ, > 0 (ferromagnetic) andJz < 0 (antiferromagnetic).
The zig-zag links indicate a broken bond; the spin in the heavy circle is frustrated.
1.1 What is a spin glass? 3

one with all antiferromagnetic interactions as shown in Fig. 1.2. Here there
is total confusion among the surrounding spins once the direction of the
centre spin is chosen. In terms of energy consideration, it will not make
any difference whether the frustrated spin points up or down. You can
imagine the complicated ‘frozen-frustrated’ situation in a real 3D lattice.
With such simple pictures we can visualize the ground (T = 0 K) or
frozen state of a spin glass. Its basic ingredients are randomness (here in
site occupancy), mixed interactions (we shall return to this point later), and
frustration: many possible ground state configurations. In order to gain a
deeper insight into this new state of magnetism, we should consider how it
is formed out of the randomly churning spins at high temperature. This
‘freezing process’ and its associated dynamics will be introduced in a later
section of this chapter.
Let us now give a working definition of a spin glass based upon the
above schematic picture. We may define a spin glass as a random, mixed-
interacting, magnetic system characterized by a random, yet co-operative,
freezing of spins at a well-defined temperature Tf below which a highly
irreversible, metastable frozen state occurs without the usual long-range
spatial magnetic order. A key word in our definition is ‘co-operative’ which
usually implies ‘ordering’ or even a phase transition. The question that the
spin glasses pose is a paradoxical one: order out of randomness? How? As
we shall see later there are a number of caveats to our definition, but in
any event we now have a foundation. From now on, the term spin glum is
used to generically represent the class of materials exhibiting the frozen-
state transition. We shall become more specific as we progress.
The two most important prerequisites of our spin glass are firstly the
randomness in either position of the spins, or the signs of the neighbouring
couplings: ferro ( t t ) or antiferromagnetic ( t J ). There must be disorder,
site or bond in order to create a spin glass, otherwise the magnetic transition
will be of the standard ferromagnetic or antiferromagnetic type of long-

e l 0 l
Fig. 1.2 Completely frustrated triangular (hexagonal) lattice with all antiferromag-
netic interactions (Jr < 0).
4 Introduction to randomness and freezing

range order. The combination of the randomness with the competing or

mixed interactions causes frustration, as illustrated in Fig. 1.1. This second
prerequisite of frustration is a unique attribute of the spin-glass ground
state, which is disordered, but still exhibits many interesting properties
related to its co-operative nature. For example, a macroscopic anisotropy
is created and peculiar metastabilities (time and aging dependences) occur
especially in reaction to an applied magnetic field. These matters will be
treated further by examining experiments in Chapter 3.

1.2 HOW TO CREATE RANDOMNESS?

As so often happens and long before the name spin glass was coined,
nature already provided us with the magnetic alloys. They are composed
of magnetic impurities bearing a moment or localized spin and occupying
random sites in a non-magnetic host metal. The spin glasses were accidentally
discovered with such binary alloys. Here we want to control the concentration
x of these impurities so that they can interact with each other. Now if we
take Fig. 1.1, make it a 3D metal, with a real lattice structure, e.g. face-
centred cubic, and then randomly distribute the little arrows on a proper
fraction of sites, we will most probably find a spin glass at low temperatures.
The host can be just about any non-magnetic metal that dissolves the ‘good
moment’ elements such as Mn, Fe, Gd, Eu, etc. The archetypal specimens
of the metallic (site-random) spin glasses are Cur-,Mn, and Aur-,Fe,.
These noble-metal alloys are also called canonical spin glasses. Site
substitution of magnetic solute for non-magnetic solvent should occur
completely randomly, i.e., without even short-range order of a chemical or
atomic nature. For then we can treat the system as statistical and use
gaussian probabilities. With a proper choice of elements this can be
accomplished to a reasonably fair approximation.
Chemical compounds both insulating and conducting can also be made
random. This is attained by diluting into one of the sub-lattices a magnetic
species in place of a non-magnetic one. Some prime examples here are
Eu,SrI-,S (a semiconductor) and LaI-,Gd,A12 (a metal). Many, many
hundreds of additional combinations can be formed in the laboratory. And
this plethora of systems has made the spin glasses a general and common
type of magnetic ordering.
Moreover, the above random-site occupancy of the alloys, mixed insulating
compounds and ‘pseudo-binary’ intermetallic compounds can be mimicked
in a more modern and violent way. One takes a standard intermetallic
compound, e.g. GdA12 and YFe2, and destroys the crystalline lattice by
making it amorphous. Techniques such as splat-cooling, quench-condensation
and sputtering are employed to de- or un-crystallize the compound. Figure
1.3 illustrates the loss of a periodic lattice with an integer ratio (or fraction)
magnetic to non-magnetic sites. A l-to-l ratio was chosen for simplicity in
1.2 How to create randomness? 5

Fig. 1.3 Amorphous spin glass for which the disordered lattice sites are 50%
occupied with magnetic moments.

Fig. 1.3. Really what matters now is the random apportionment of distances
between the magnetic impurities. The same gaussian distribution function
P(r) occurs here as in our original random-site alloys and the mixed
compounds. The variance or distribution width may be narrower for the
amorphous materials sketched in Fig. 1.3, nevertheless, it is the positional
randomness that creates an impurity-spin distance distribution which is the
first and basic ingredient of our would-be spin glass.
So much said for site disorder. Let us now consider another possibility
to produce a spin glass. Suppose we view the perfect lattice of Fig. 1.4,
every site dot is occupied with a spin. Assume that the magnetic interaction
is only between nearest neighbours (short-range) and imagine that this
single coupling J alternates randomly in sign with +J values as we move
through the lattice. Thus, there are only parallel or antiparallel bonds as
illustrated by the ‘dash’ or ‘zig-zag’ in Fig. 1.4. This is the random-bond
type of system and it was only recently that their existence in real materials,
e.g. the compounds Rb,Cu,-,Co,F, and FeI-,MnXTi03, were unearthed
to give reasonable approximations of 9 couplings. But more about these
random-bond spin glasses when we arrive at the theoretical models for a
spin glass. Suffice it to say, there must be disorder in the constitution of a
spin glass: either site randomness with a distribution of distances between
the magnetic spins, or bond randomness where the nearest-neighbour
interaction varies between parallel coupling +.I and antiparallel coupling
-J. If the system is truly random, the average value of all the exchange
bonds is zero. So we now know how to create the all-important randomness.
 Introduction to randomness and freezing

Fig. 1.4 Random-bond spin glass where -- * ferromagnetic coupling and+ +

antiferromagnetic coupling. The number of ferromagnetic bonds is equal to the
number of antiferromagnetic ones.

And once established, the randomness remains forever fixed or constant,

consequently we give it the name quenched-disorder.

1.3 MAGNETIC INTERACTIONS (ARE NECESSARY)

1.3.1 Direct exchange

The spin-glass problem is one of interactions between the spins and it is

through these exchanges and the above randomness that a unique type of
ground state is formed. Magnetic-exchange interactions go back to the early
days of quantum mechanics and are usually written in the form of a spin
hamiltonian as Xij = -JijSi - Si parameter which couples the spins at lattice
sites i and j. Direct exchange involves an overlap of electronic wave
functions from the two sites and the Coulomb electrostatic repulsion. The
Pauli exclusion principle keeps the electrons with parallel spin away from
each other, thereby reducing the Coulomb repulsion. The energy difference
between parallel and antiparallel configurations is the exchange energy.
Usually an antiparallel configuration is favoured as in the simplest case of
the hydrogen molecule and as often found in practice. However, since the
wave functions of the magnetic d or f electrons decrease exponentially with
distance from the nucleus, the Jij obtained from the overlap integral is far
too small to provide the necessary coupling. There was an unsuccessful
1.3 Magnetic interactions (are necessary) 7

attempt to create a direct exchange spin glass by doping phosphorus into

silicon and trying to observe the freezing of the P-moments at very low
(mK) temperatures. So another mechanism must be involved for our spin
glasses.

1.3.2 RKKY

For the magnetic alloys the conduction electrons lead to stronger and
longer-range indirect-exchange interaction. This is the now-famous Ruder-
man, Kittel, Kasuya, Yosida (RKKY) interaction whose hamiltonian is
X = J(r)Si * Sj. Embedding a magnetic impurity, i.e., a local moment (spin
Si), in a sea of conduction electrons with itinerant spin s(r) causes a damped
oscillation in the susceptibility of the conduction electrons, and thereby a
coupling between spins (Si and Sj) according to

J(r) = 673J2N&)
cos (2k,r)
(w9’ 1
where Z is the number of conduction electrons per atom, J the s-d exchange
(1.1)

constant, N(E,) the density of states at the Fermi level, kF the Fermi
momentum and r the distance between two impurities. This reduces to
J(r) = Jo COS(%r+d’)
C&r)”
(1.2)
at large distance. A phase factor 4, has been included to account for the
charge difference between impurity and host and the former’s angular
momentum. Such oscillatory behaviour of J(r) or really the Pauli suscepti-
bility, which in the free-electron model has spherical symmetry, is illustrated
by the two coupling schemes in Fig. 1.5. Notice that the (l/r)” fall-off is
sufficiently long-ranged so that it can effectively reach a number of nearest-
neighbour sites. Now suppose we put a second magnetic impurity with spin
Sj at one of the neighbouring sites. This spin will produce its own RKKY
polarization and the two conduction-electron-mediated polarizations will
overlap in such a way as to establish a parallel or an antiparallel alignment
of the two spins. These situations are sketched in Fig. 1.5 by plotting the
Pauli susceptibility of the conduction electrons. Note that the sign (+ = t t
and - = t J ) of the impurity coupling varies with distance. If we combine
this property with the site disorder (various separations between the spins),
discussed in the previous section, we have generated a random distribution
of coupling strengths and directions. A computer simulation for some half-
a-million RKKY bonds, Jij, coupling a random-site magnetic alloy generates
the probability function P(Jij) shown in Fig. 1.6. Important is the near
symmetry in the number of + and - bonds. Here the required factor of
‘competition’ among ferro and antiferromagnetic exchanges is obtained in
8 Introduction to randomness and freezing

Fig. 1.5 Ruderman-Kittel-Kasuya-Yosida (RKKY) interaction between two

impurities in terms of the Pauli susceptibility xp.

2000-

1000-

500.

xx)-
-FT
.z 100 -
3
=.
5 50-
&

2
7 1
a mI

10 1
I

5 t

2.10-4 -1.10‘* G 1.W -2.10-"

J/Jo

Fig. 1.6 Probability distribution

of coupling strengths for about 5 x lo5 bonds
resulting from a RKKY interaction in a dilute magnetic alloy. From Binder and
Schrijder (1976).
1.3 Magnetic interactions (are necessary) 9

a natural and common way, namely via the oscillating RKKY. This is why
the whole spin-glass problem started with the magnetic alloys. We must
once again emphasize the combination of site disorder with the + and -
RKKY interaction causing a mixture of competitive bonds that will eventually
lead to frustration in some of these bonds.

1.3.3 Superexchange

How about insulating or semiconducting materials? Since there are no

conduction electrons, in what way can we couple the spins if the direct
exchange is so weak? The prime mechanism in insulators is that of
superexchange. In this case an intervening ligand or anion transfers an
electron (usually in a p state) to the neighbouring magnetic atom. A sort
of covalent mixing of the p and d (or f) wave functions occurs with spins
pointing in the same direction. Because the two anion p-spins must be
opposite in direction (Pauli exclusion principle), they will cause antiparallel
pairing with the d-electrons on the magnetic atoms to the left and to the
right. This situation is shown in Fig. 1.7 and leads to an antiferromagnetic
coupling via the ligand situated between the two magnetic atoms. The net
result is a stronger and more long-ranged [usually at some large power IZ
of (l/r)“] interaction than the exponential fall-off of the direct coupling.
Superexchange can also create a ferromagnetic coupling if we take into
account the exchange polarization of the anion orbitals. In the magnetic

d-element ligand d-element

Fig. 1.7 Superexchange between two magnetic d-ions and a p-state Iigand (U-
transfer). The shaded regions represent the covalent mixing of the different wave
functions.
10 Introduction to randomness and freezing

semiconductors the proximity of the valence band can lead to non-localized

states which can propagate the exchange interaction even further. So, once
again we can have a parallel or antiparallel orientation between the magnetic
spins which, if randomly spaced, results in the competing interactions
necessary to form the spin-glass state.

1.3.4 Dipolar

Another mechanism that causes spins to interact is the dipolar energy.

Although weak it is always present and takes the form:

(1.3)

Note that in addition to the (llrij)3 distance dependence there is also a

built-in anisotropy to the dipolar interaction which can orient the spins
ferro- or antiferromagnetic. For example, if the spins are oriented along
rij, they will couple parallel; however, if the spins are oriented perpendicular
to Fri.3then they will couple antiparallel. Try it by seeing which configuration
gives the lowest (negative) energy in the above equation. The anisotropy
of the interaction plays an important role in the resulting ordering. Up
until now our RKKY and superexchange couplings were considered as
isotropic interactions, i.e., only a function of distance r with no angular
dependences. The dipolar interaction introduces an angular dependence
into the coupling scheme and we will delve more into magnetic anisotropy
in the next section.

1.3.5 Mixed superexchange (random bonds)

Thus far we have listed the mixed or competing interactions which are to
be combined with the random site occupancy to form a spin glass. But
what about the -+.I random-bond mechanism mentioned at the end of
section 1.2 and illustrated in Fig. 1.4? Now we must rely on using two
different magnetic species, in our previous examples Cu (which in an
insulator, but never in a metal, can be magnetic S = i) and Co, and Fe
and Mn. Let us examine the insulating, superexchange Rb&u,-,Co,F,
system. Here the tetragonal crystal electric field experienced by the Co
(S = s) splits the various spin states such that only the lowest S = $ doublet
is populated at low temperatures. Hence both the Cu and Co ions have
effectively S = i. The crystal-field splitting also causes a single-ion anisotropy
(see next section) which forces the spins to point only up or down along
the z (c-axis) direction. This simple, up or down, collinear orientation of
the moments we call an Ising character. At this point we bring the nearest
1.4 Magnetic anisotropy (plays a role) 11

neighbour superexchange into action. Because of the different ground state

orbitals of the Cu, their wave-functions alternate lobes along the x and y
axes. This generates a ferromagnetic coupling. By substituting the Co on
the Cu sites, two possibilities exist for the Co-Cu superexchange. Firstly a
Cu-lobe pointing along the distance to the Co ion results in an anti-
ferromagnetic coupling; and secondly a Cu-lobe perpendicular to the Co
site forms a ferromagnetic pair. Finally, if two Co atoms happen to be
nearest neighbours the usual antiferromagnetic superexchange is the outcome.
Therefore, we have four distinct interactions among the mixed Cu-Co
system, two are parallel and two antiparallel. Figure 1.8 shows the bond-
probability function P(J) for a particular choice of x = 0.29. While not
exactly the +J distribution used in the creation of Fig. 1.4, we do have a
favourable approximation to a random-bond spin glass with a realizable
material. More will be said about this system when we deal with the ‘ideal’
spin glasses in Chapter 6.

1.4 MAGNETIC ANISOTROPY (PLAYS A ROLE)

We have already been forced into using the notion of magnetic anisotropy
which offers a preferred direction for aligning the spins. The simplest form
of this anisotropy previously used in Fig. 1.2 is the Ising model in which
all the spins are constrained to point either up or down. This is caused by
a very strong uniaxial anisotropy in one direction, usually. taken as the z-
axis. Alternatively, the spins can also be forced to lie in a 2D plane, then
we have x-y anisotropy. Remember that the Heisenberg model is, of course,
fully isotropic, and the spins can align in any direction, even non-collinear,
of the 3D space. These various types of magnetic anisotropies will become

0.6 r-

Fig. 1.8 Bond probability function P(J) for RbzCuI...,CoxF4 with x = 0.29. From
Dekker (1988).
12 Introduction to randomness and freezing

very important when we treat the possibilities of a phase transition using

the predictions of the theoretical calculations. In addition the anisotropy
plays a major role in governing the frozen-state properties of a spin glass.
Let us briefly review the various mechanisms for generating the magnetic
anisotropy. As previously stated the dipolar interaction gives an intrinsic
preferred coupling as a function of angle, i.e., orientation of the moments
at sites i and i relative to rii. This will certainly have a bearing on the
particular type of magnetic ordering at low temperatures. Since the dipolar
energy is usually so weak compared to the various exchange energies, it is
a secondary effect and must be used in conjunction with an (isotropic)
exchange interaction to invoke its influence, for instance, on the particular
direction of the moments as the spin glass freezes.
Single-ion anisotropy, produced from the local crystalline electric fields,
is another important source of anisotropy. We already used this mechanism
in explaining the random-bond Rb2CuCoF4 system. On a microscopic basis,
it is directly related to the crystal symmetry of the given material and the
spin and orbit angular moment of the magnetic species. Certain preferred
directions are imposed on the (non-S state) magnetic ions by the crystalline
electric field, thus creating the magnetocrystalline or single ion anisotropy.
Its energy is usually written in spin-only hamiltonian form as

xg = -D~((s:)’ (1.4)

for an uniaxial crystal,

F&I = -Dc[(s:)4 + (sg4 + (Sf)“] (1.5)

i

for a cubic crystal; and

SITS,
= -CDi(Sf)’ (l-6)
i

for an amorphous material. The prefactor D is the local anisotropy constant,

except for the amorphous material, where Di has a distribution of magnitudes
and directions (axes) which are different at every site. Consequently, in a
non-crystalline solid there can exist a random anisotropy defined through a
locally varying ‘easy axis’.
Yet another form of anisotropy arises in the metallic spin glasses and is
due to the Dzyaloshinskii-Moriya (DM) interaction. Its energy (hamiltonian)
is given by XijDM = -D, - (Si X Sj) and describes the following process: a
conduction electron is initially scattered by spin Si, then it interacts with a
non-magnetic scatterer having a large spin-orbit coupling, and finally, the
conduction electron scatters off another spin Sj. It is this intermediate non-
magnetic, spin-orbit scattering that is the essence of the DM anisotropy.
Here there appears in D, a cross product (Ri X Rj) of the positions of the
two spins with respect to the non-magnetic scatterer placed at the origin.
1.5 Frustration (a far-reaching concept) 13

The net result of this anisotropy is to establish a unidirectional character

which aligns the spin in a single direction, quite different from the
uniaxial (parallel/antiparallel) forms previously discussed. If we perform the
summations for these two distinct types (DM plus uniaxial) of anisotropy,
we arrive at a macroscopic anisotropy energy
E AN = -K,cosO - ;K,cos20 (1.7)
where 0 is the angle which rotates the spins from their frozen direction.
K1 is the unidirectional anisotropy constant, which changes sign with a
rotation of 7~, and K2 is the uniaxial anisotropy constant, which remains the
same under a rotation of T. For the CuMn spin glass K1 can be made
predominant and a unidirectional character appears.
To summarize this section we have tried to collect and outline the relevant
forms of anisotropy or freezing-direction preference. Unique to the spin
glasses are the possibilities of random anisotropy and the DM interaction
creating a unidirectional anisotropy. Also noteworthy is that they can
manifest themselves as a macroscopic energy in certain experiments (Chapter
3). We shall use the notion of anisotropy frequently as we dig further into
the mysteries of the spin glasses.

1.5 FRUSTRATION (A FAR-REACHING CONCEPT)

The word frustration and its illustration have already appeared at the very
beginning of this chapter. Figure 1.2 sketched the basic triangle and
antiferromagnetically bonded Ising spins. The result was six degenerate
ground states (all with the same energy) for which there is always one
unsatisfied bond. Try this exercise yourself. The concept of frustration,
implicit in many physical and biological phenomena, was first clearly used
and quantified in the description of the spin glasses. It is the essential
ingredient to establish the special ground state. So let us look a bit closer
at frustration.
Suppose we have a square or plaquette of four Ising spins with four &J
bonds as shown in Fig. 1.9. For the left-hand configuration (Fig. 1.9(a)) of

+ +

++ +

-I u
(a) lb)
Fig. 1.9 Square lattice with mixed interaction +:F, -:AF. (a) unfrustrated plaquette
@ = 1 and (b) frustrated plaquette Q, = -1.
14 Introduction to randomness and freezing

two + bonds and two - bonds, all the bond energies are satisfied and
there will only be a two-fold-degenerate ordered state. This arbitrariness is
caused by the initial choice of the first spin which could be set either up
or down. Here we say the plaquette is unfrustrated. Moving over to the
right-hand configuration in Fig. 1.9(b) (three + and one - bonds), all the
bond energies cannot simultaneously be satisifed. One spin remains frustrated
or one bond is broken no matter what we do. Notice how the degeneracy
increases to &fold (count them for yourself). For this trivial example eight
energetically equivalent ground states are possible. In a mathematical way
we may write the frustration function as @ = II sign (Jii) = ? 1, where the
minus sign denotes a frustrated system. For the case shown in Fig. 1.9,
Q = J1J2J3J4 = +l for the unfrustrated left-hand side and -1 for the
frustrated right-hand side.
If there is a great deal of frustration in our system, many of the plaquettes
will be frustrated. Connecting such plaquettes with a set of minimal distance
lines gives a measure of the excess energy caused by the frustration. The
task, and it is not an easy one, is to find the minimum length between the
various frustrated plaquettes and thereby determine the lowest ground-state
energy. There is great similarity between this ‘optimization’ problem and
others such as the intriguing ‘travelling salesman’ problem (see Chapter 7).
Note that the frustration creates a multidegenerate, metastable, frozen
ground state for the spin glass. And as we shall soon see such a ground
state leads to many interesting and unique experimental properties. But
now we should ask the question - ‘is frustration alone enough to generate
a spin glass according to our working definition?’ The answer is ‘No, it
must be coupled with disorder or randomness, site or bond, our first basic
ingredient’. An antiferromagnetically coupled, regular triangular lattice
(which is full frustrated) will not exhibit the co-operative freezing transition
of a spin glass. More likely, a slow ‘blocking’ will occur to a ground state
which is dominated by large magnetic fluctuations. Furthermore, there
seems also to be the need for mixed (+ and -) interactions supplementary
to the randomness in order to attain our spin glass. Thus, and this aspect
is still under contention, a system with only antiferromagnetic interactions
plus disorder will not yet be a good, co-operative spin glass. The ‘extra
necessary’ ingredient would be to include some ferromagnetic bonds.
Support for our standpoint comes from very recent experiments on KagomC
(triangular-like) lattices with only antiferromagnetically coupled ions which
do not show generic spin-glass behaviour.
Let us recapitulate the essence of the previous sections to secure our
point of view. Randomness or disorder is indispensable and can be created
by varying the distances between the magnetic species or by randomizing
the bonds among a periodic set of magnetic ions. The mixed
(ferro-antiferromagnetic) interactions are essential to install the competition
and ensure co-operativeness of the freezing process. Anisotropy seems to
be required to establish preferred directions along which the local spins can
1.6 The freezing process (a first look) 15

freeze. And frustration is a direct consequence of the disorder and mixed

interaction, while a necessary condition for a spin glass, it is not a sufficient
one.

1.6 THE FREEZING PROCESS (A FIRST LOOK)

At this point it would be instructive to try to present our first naive picture
of the evolution of a spin glass as we reduce the temperature from far
above the freezing temperature Tf to far below. Our description is based
on experimental observation which will be explicated and discussed in
Chapter 3. We now shall hope to gain some rudimentary insights into the
underlying physics of the freezing process.
Of course, as T + CQ,there will simply be a collection of paramagnetic
spins, i.e., independent and rapidly rotating arrows in the ‘chaos’ caused
by the high temperature. When the temperature is lower from T % Tf many
of these randomly positioned and freely rotating spins build themselves into
locally correlated units or clusters, even domains, which can then rotate as
a whole. These ‘clusters’ may simply be ferromagnetically coupled resulting
in ‘giant’ or superparamagnetic moments. Or they may form from a strongly
localized overlapping of RKKY interactions, in which case, because of the
(+) parallel, (-) antiparallel oscillations, the net moment will be proportional
to the square root of the number of spins taking part. The remaining spins,
those not belonging to any cluster, are independent of each other, but help
to transmit interactions between the clusters allowing for changes in the
cluster sizes and their relaxation or response times. This formation of
clusters can be tracked with various experimental techniques and is a direct
consequence of the randomness and mixed interactions.
Now as T + Tf the various spin components begin to interact with each
other over a longer range, because the temperature disorder is being
removed. The system seeks its ground-state (T = 0) configuration for the
particular distribution of spins and exchange interactions. This means a
favourable set of random alignment axes generated by the local anisotropy
into which the spins or clusters can lock. However, as we learned in the
previous section frustration plays its role and a multi-degenerate array of
ground states presents itself for the system to choose from. Since there is
a spectrum of energy differences between the frozen states, the system may
become trapped in a metastable configuration of higher energy.
What type of transition or transformation occurs at the freezing of a spin
glass? Is it a phase transition with Tf a critical temperature (T,), treatable
within the theories of critical phenomena ? Or is there a gradual blocking
of the cluster rotations so that the system slowly settles into the frozen
state? A third possibility exists somewhere in between these two extremes,
where a new type of ‘random’ transition occurs requiring a very different
theoretical description. How are the spin glasses and their freezing related
16 Introduction to randomness and freezing

to the real ‘window’ or ‘wine’ glasses and their transition? Is it possible to

combine the glassy analogies into one unified treatment? Yet, at the
moment, the real glass transition does not have a generally acceptable
microscopic model and new efforts (e.g., mode coupling theories) are
constantly going into creating one. Consequently, we cannot answer these
questions completely, because the exact nature of the freezing process both
in real glasses and spin glasses is presently unresolved. There are, however,
certain directions, which became clear during the intervening twenty years
of research, leading to the conclusion that the spin glasses are unique and
necessitate an unconventional description both phenomenologically and
mathematically. And the whole purpose of this monograph is to introduce
you to these strange and intriguing properties and their underlying physics.
Before concluding this chapter with some historical background, we
should mention the low-temperature (T << Tr) frozen state. Here unusual
magnetic behaviour appears related to the ‘glassy’ condition. Most of these
effects are caused by the application and cycling of an external field which
also probes the anisotropy. Additional metastabilities and relaxations have
been attributed to the rate dT/dt at which the sample is cooled through T,,
and the ‘waiting time’ tw below T,, with T constant, before the field is
changed.
Another salient feature of the frozen state is the appearance of a
unidirectional (see section 1.4) anisotropy which has a microscopic basis
and creates randomly distributed focal orientations throughout the sample.
With various field-cooling or cycling procedures this quantity can be made
to manifest itself as a macroscopic variable and is of particular importance
in governing the low-temperature collective properties, such as magnetization,
torque, ESR, etc. The excitation spectra of the frozen state can also be
probed via various scattering techniques and more localized than propagating
modes are found.
Finally it should be emphasized again that no usual ferro- or antiferromag-
netic state with long-range magnetic order is formed. Hence, one does not
observe magnetic Bragg peaks with scattering experiments due to a lack of
translational invariance. The spin-glass condition for the randomly frozen
state is C& = 0 where the sum is over all spin sites. Another convenient
way of describing a frozen spin glass is by its time correlation function
(Sj(t)Si(0)) which remains finite for T < Tf as t + ~0.

1.7 SOME HISTORICAL BACKGROUND

Magnetic alloys have a long history dating back a hundred or so years.

Especially interesting are the noble-metal/transition-metal combinations in
which the first spin-glass effects were discovered. Already in the 1930s
experimental investigations on AuFe and CuMn indicated some strange
I. 7 Some historical background 17

magnetic properties. The first had to do with the magnetism or lack of it

caused by the introduction of a magnetic impurity into a non-magnetic
matrix. Would the moment remain, or would it lose its magnetism when
dissolved in the sea of conduction electrons? Moment formation was the
issue and the virtual-bound-state model (c. 1960) resolved the matter. On
the other hand, the transport properties exhibited anomalous effects;
resistivity minimum and giant thermoelectric power were discovered and
given the name Kondo problem. The main research emphasis during the
1960s was on the moment formation and the Kondo effect. There were
indeed studies of more concentrated alloys, but the inherent interactions
were often incorrectly neglected to get at the single-impurity Kondo
behaviour.
A number of investigations in the 1950s did consider the interaction of
well-formed moments; some experimental behaviour was unusual but not
striking. For example, supplementary linear term in the specific heat at low
temperatures was observed in various magnetic alloys. In addition, the
susceptibility measured in rather large fields showed a broad maximum at
a temperature where roughly the magnetic remanence disappeared. The
latter and its hysteresis were different from the conventional ferromagnetism;
more like a collection of mutually interacting ferromagnetic and antiferromag-
netic domains. Mossbauer experiments with its advent during the same
period indicated there was some type of random magnetic order developing
at low temperature. These were the empirical fruits of the 1960s. But then
the main concern especially theoretically was with the single (Kondo)
impurities.
The situation changed around 1970, at that time the sharp cusps in the
low-field susceptibility of spin glasses were discovered and correlated with
a magnetic-ordering temperature as determined from other measurements.
Moreover, the RKKY interaction was brought to bear, and from its l/r3
decrease, the invariance of the product xr3 (= const.), where x is
the magnetic-impurity concentration, was discerned and enabled the
magnetization MIX and the specific heat C,,lx to be represented as universal
functions of the reduced variables H/x and T/x. A ‘scaling’ (or reduction
of much data to a single curve) of many experiments was thereby possible.
These advances marked the beginning of the spin-glass problem as we know
it today.
But how did the name ‘spin glass’ arise and adhere? The term was first
suggested by B. R. Coles in 1970 to be applied to the strange magnetic
behaviour of the weak moment Ado system. Simultaneously, this expression
appeared again at Coles’ instigation in a 1970 paper by P. W. Anderson
linking localization in disordered systems with the magnetic-alloy problem.
How it took hold is more difficult to answer. Other terms were competing,
mictomagnetism was suggested around 1971 by P. A. Beck. Later the
designation ‘cluster glass’ was used. But anyway what’s in a name? It’s
18 Introduction to randomness and freezing

more important to learn what physical phenomenon the nomenclature

represents and how we can understand its observable facts. Then we can
give a ‘catch-all’ appellation to this body of knowledge.
Returning to our history and the 1970s finally the theorist became
interested around 1975 and ‘all hell broke loose’ with a number of exciting
new models and their attempted solutions, viz., the Edwards-Anderson
model and the Sherrington-Kirkpatrick solution. In Chapter 5 we shall
delve into these theories. The five year interval between 1975 and 1980 was
a time of great activity - theory led the way with experiment offering a
series of new surprises. Naturally many papers (thousands) were published.
Around the end of this period, the order parameter describing the mean-
field phase transition of spin glasses was shown to be inadequate. The
frozen state according to theory was unstable. Henceforth, there began a
renewed quest for the proper ground state or replica-symmetry-breaking
scheme - a term with which we shall later become acquainted. A suitable
scheme was discovered a few years later, but its physical meaning was
initially obscure. Nevertheless, a new form of statistical mechanism
with ‘ultrametricity’ was created requiring a novel, unconventional order
parameter. And with great popularity the spin-glass models and mathematics
were applied to other diverse areas. One especially interesting new field is
that of neural networks. Dynamics (time and frequency dependences) was
also treated within the mean-field theory and recognized as a key element.
But as more and more materials were called spin glasses, the contact
between theory and experiment diminished. It was difficult to compare the
empirical behaviour of the many real spin glasses to the new ideahzed
theories. Circa 1985 the post-replica age began, for, another theory based
upon a ‘droplet’ (Zstate) model was proposed for short-range spin glasses.
Here the insulating materials could be treated and compared with the
model’s predictions, if ideal glasses could be fabricated in the laboratory.
After 1985, the pace of spin-glass research slowed down and while the
basic phenomena have been established and catalogued, a thorough
understanding or full description via the existing theories is still lacking.
Into the 1990s there remain many unresolved or grey areas of the spin-
glass problem. As we conclude our treatise in the final chapter we could
perhaps look into some of these. If history is our guide, usually in times
of peace and quiet, complete solutions are obtained for difficult questions.
So let us wait and see what happens during the writing/reading of this book.

FURTHER READING AND REFERENCES

General solid state physics

Kittel, C. (1986) Introduction to solid state physics, (6th edn), Wiley.
Ashcroft, N. W. and Mermin, N. D. (1976) Solid state physics, Holt, Rinehart and
Winston.
tittel, C. (1964) Quantum theory of solids, Wiley.
Further reading and references 19

General magnetism

Morrish, A. H. (1965) Physical principles of magnetism, Wiley.

Crangle, J. (1977) The magnetic properties of soliah, Arnold.
White, R. M. (1983) Quantum theory of magnetism (2nd edn), Springer.
Mattis, D. C. (1988) The theory of magnetism Vol. I and II, Springer.

Introductory articles to spin glasses

Stein, D. L. (1989) Spin glasses, Scientific America, July, 36.

Fisher, D. S., Grinstein, G. M. and Khurana, A. (1988) Physics Today, Dec.
Ford, P. J. (1982) Contemp. Phys., 23, 141.
Hurd, C. M. (1982) Contemp. Phys., 23, 469.
Mydosh, J. A. (1978) .I. Magn. Magn. Mater., 7, 237.

Historical background

Anderson, P. W. Physics Today. Series of six articles in Reference Frame, Jan.

1988, March 1988, June 1988, Sept. 1988, July 1989, and Sept. 1989.

References

Binder, K. and SchrBder, K. (1976) Phys. Rev., B14, 2142.

Dekker, C. (1988) PhD Thesis, University of Utrecht.
2
Description of related
concepts

In order to better present the spin glasses and establish their place in
magnetic phenomenon, we should first consider how they evolved and
where they belong with respect to other allied magnetic effects. Some
perspective at this point is important and a number of cognate terms are
in need of introduction. This chapter will briefly survey the related or
surrounding phenomena from which the spin glasses developed. The main
portion of Chapter 2 is devoted to the dilute magnetic alloys. Moment
formation via the virtual-bound-state model and the Kondo effect along
with the giant moments will be introduced. Afterwards we will discuss the
various types of behaviour which appear in the different concentration
regimes extending from very dilute to fairly concentrated. This is to show
exactly where the spin glasses lie with respect to their disorder. Here a
valuable new concept for us is percolation which distinguishes the spin
glasses from the more-concentrated, long-range-ordered, random systems.
Then we shall view the magnetic glasses, since amorphous magnetism is
certainly a part of the spin-glass problem and many of these materials
exhibit the novel ‘re-entry’ phenomenon. Finally, the notion of superparamag-
netism is explored and its applicability to the spin glasses perused, and to
conclude, the insulating and semiconducting systems will be mentioned and
identified.

2.1 VIRTUAL-BOUND-STATE MODEL

The simplest case of a magnetic ion in a solid is that of a rare-earth

magnetic insulator. Most rare-earth ions have similar chemical properties
with valency +3 and the 4f (magnetic) electrons are situated well to the
interior of the atom and surrounded by 5, .5p and 5d electrons. Therefore,
the ions in many insulating compounds are non-interacting and simply
paramagnetic. This means the rare earths nicely obey the Curie law
2.1 Virtual-bound-state model 21

ivJ(J + l)g$$
X= 3kB T (2.1)

p,(THEO) = g[J(J + l)]* and p,(EXP)= (3k;:)* . (2.2)

J-B
Here iV is the number of paramagnetic ions, g is the Land6 g-factor, ~~
the Bohr magneton and kB the Boltzmann constant. For rare-earth ions J
is a good (total) angular-momentum quantum number and can be constructed
out of the atomic ground-state L and S quantum numbers invoking Hund’s
rules. Once we know J it is an easy task to define a theoretical effective
Bohr magneton number P,(THEO) and compare it with experiment via
the measured p,(EXP). The agreement is usually good and the few
exceptions Pm+3, Sm+3 and Eu+~ can be corrected for by using the concept
of induced-moment (Van Vleck) magnetism. Crystalline electric fields will
influence the ionic behaviour at low temperatures. However, such effects
can be taken into account knowing the crystal symmetry and the energy
splitting.
The situation is somewhat different regarding the other source of magnetic
ions - the 3d transition metals. In this case there is a much larger radius
for the 3d-orbital wave functions and the 4s electrons are usually lost to
the ionic bonding of the compounds. Thus, no outer electrons exist to
shield the 3d electrons and the crystalline electric fields cause significant
energy splittings, much larger than that induced by the spin-orbit coupling.
So J is not a good quantum number and the ionic model no longer works.
There is, however, one simplifying hope, namely that the lowest orbital
level, when split by a cubic crystal field, is a singlet L = 0. In many cases
this is the only level populated and there can be no further splitting due to
the spin-orbit coupling. Accordingly, we say the orbital moment has been
quenched and we have ‘spin-only’ magnetism. In these instances we can
compare theory to experiment merely using
P,(THEO) = g[S(S + l)]” (2.3)
and observe sound agreement. Nevertheless, there are many exceptions to
the above simplified singlet case. Sometimes doublet or triplet orbital levels
are lowest, and these are further split by the spin-orbit coupling. Also
incomplete quenching can arise and the spin-orbit interaction could be
large. All these cases must be handled on an individual basis compound by
compound, and while complicated combinations of crystalline electric field
and spin-orbit coupling arises, notwithstanding, they are all treatable within
an individual atomic model which gives the local atomic moment.
The metals are completely different, for now we have an interaction
between the magnetic ions and the sea of conduction electrons. Whether a
magnetic moment survives is a fundamental question. In order to treat this
problem we must use the virtual-bound state (vbs) or Anderson model.
22 Description of related concepts

Let us consider a single 3d magnetic ion immersed in the sea of conduction

electrons produced by a host metal. To keep things simple we regard the
host as a free-electron-like metal. There are four terms which arise in the
hamiltonian describing such a system:
X = KE(s-elec) + KE(d-elec) + V&mixing) + U&Coulomb). (2.4)
The first two terms represent the individual kinetic energies of s and d
electrons, respectively. The third term is a mixing or hybridization interaction
between the s and d electrons, and the final term gives the intra-atomic
Coulomb repulsion of the lowest d orbitals of opposite spin. Finding the
solution of this vbs-hamiltonian is a long and arduous task. We shall not
go into the mathematical details, but simply view some sketches of the
physical situations.
Figure 2.1 (left-hand side) shows a two-band (spin up/spin down) density
of states N(E) plot for the host’s s-band - filled up to Fermi level EF and
a localized (atomic-like) d-impurity level with energy less than EF. Now we
turn on the mixing or hybridization and the intra-atomic Coulomb repulsion.
The former causes the ‘localized’ level to be firstly shifted downwards in
energy with respect to EF:Ed = 2N(E,))V,$/D (D is the s-bandwidth), and
then to be broadened with width 2r = 27rh7(EF)IVJ2. The latter (Coulomb
repulsion) causes an energy shift U between the spin-up level and the spin-
down level. The right-hand side of Fig. 2.1 delineates the case where
jVsdl 4 1Uj. Note how the two virtual-bound states (vbs) are formed on
each of the sub-bands, one filled, one empty. Here the spin-down vbs-level
is simply pushed above the Fermi energy and is therefore unoccupied.
Thus, we have created a strongly magnetic case via the filling of the spin-
up vbs, i.e., a localized magnetic moment is formed.
-But what happens if we make IV,,1 comparable to U? The spin-up and
spin-down vbs levels become broader and could partially overlap with the

Fig. 2.1 Schematic diagram (left-hand side) of a 3d impurity immersed in a sea of

conduction electrons separated into spin-up and spin-down electrons. Right-hand
side illustrates the formation of the vbs. An occupied spin-up vbs at Ed with width
2r is attached to the conduction electron density of states. A similar, but unoccupied,
vbs split by energy U above EF occurs with down spin.
2.1 Virtual-bound-state model 23

Fermi energy. This means they will be occupied with different amounts of
spin-up d-electrons and spin-down d-electrons. A reduced Coulomb repulsion
results in an upwards shift of the up-vbs relative to the down-vbs. This
further increases the overlap and is shown in the left-hand side of Fig. 2.2.
The net vbs occupancy of up-spin minus down-spin will determine the ‘local’
d-moment which could easily be non-half integral. Thus, any value of spin
or moment could appear in a dilute magnetic alloy. Or, if we even push
our parameters further IV,,] + UN,, = 0, we arrive at the symmetric case
of up-spin occupancy equal to down-spin occupancy, and no net spin, i.e.,
the impurity loses the magnetism; it is effectively dissolved in the sea of
conduction electrons. Figure 2.2 (right-hand side) depicts this non-magnetic
possibility.
The main consequence of the virtual-bound-state model is that an atomic
magnetic impurity may not even form a moment when introduced into a
metal. Everything depends on the given set of parameters. If it does, then
any (not just half integer) value of spin or moment may be generated. The
model had its heyday in the early 1960s when the problem of moments
formation was a paramount issue. What the model does not answer is how
the moment forms as a function of temperature and what effect this
temperature dependence will have on the experimental properties.
We shall look further into these questions in the next section. But before
concluding this brief sketch of the vbs model, let us mention the
Schrieffer-Wolff (S-W) canonical transformation which allows us to combine
the various model parameters into an effective exchange interaction J. In
1966 S-W showed that in the strongly magnetic limit [the case displayed in
Fig. 2.1 (right-hand side)]

(2.5)

Fig. 2.2 Schematic diagram (left-hand side) of partially magnetic vbs. Note how
the magnetic moment can form with any spin value. Right-hand side represents the
symmetric case of equal spin-up and spin-down occupancies (U = 0) and thus no
net magnetic moment.
24 Description of related concepts

The underlying physics of this antiferromagnetic coupling between local-

moment and conduction electrons may be gained from the intuitive picture
given in Fig. 2.3. However, to make the draught realistic we should place
the vbs levels close to EF so we can invoke the process of ‘covalent mixing’.
Now what happens when, due to the covalent mixing, there is an electron
transfer from the occupied vbs to the up-conduction electron band and
from the down-conduction electron band to the unoccupied vbs? The answer
is that a difference in the Fermi levels or fillings will appear between up
and down spin bands. In order to correct for this non-equilibrium situation,
up-spin conduction electrons flow into down-spin conduction-electron states
(see Fig. 2.3). The net result of this mixing and siphoning process of up-
spin electrons is to produce slightly more down-spin conduction electrons
than up-spin conduction electrons at the expense of the occupied local
moment. If we realize that the covalent mixing between unlike spins is
‘attractive’, then a small conduction-electron polarization of opposite
spin will be created surrounding the local moment (up-spin vbs). The
corresponding admixture of like spins is forbidden by the Pauli exclusion
principle and a repulsion or delocalization results. Although the processes
described in Fig. 2.3 are rather schematic, they do lead to an antiferromag-
netic coupling between the local moment and the conduction electrons.
And this indeed is the origin of the Kondo effect.

2.2 KONDO EFFECT AND WEAK MOMENTS

The subject of dilute magnetic alloys may be divided into two classes of
interest. Firstly, non-interacting or very-dilute magnetic impurities dissolved
into a non-magnetic host can be classified under the general heading ‘Kondo
effect’ - a localized antiferromagnetic interaction of an isolated or single-

Fig. 2.3 The effect of covalent mixing on the vbs which siphons off the moment
and causes antiferromagnetic coupling (J < 0) between the local moment and
conduction electrons.
2.2 Kondo effect and weak moments 25

impurity spin with the surrounding conduction electrons. Secondly and of

prime importance in this book are the impurity spin-spin interactions which
lead to spin-glass freezing and long-range magnetic order.
Returning to the first (Kondo) class, we must treat the magnetic spins as
isolated from each other. In practice, such a condition is rather difficult to
realize, since the previously mentioned RKKY interaction is present and
falls off rather slowly as (ll~)~. Hence, there is always a weak coupling
between even far-distant spins. A reasonable empirical approximation is to
work at impurity concentrations below 100 ppm and not too low a
temperature. Assume now that the RKKY is negligibly small. The only
remaining interaction (since the dipolar one can be completely disregarded)
is between the local-moment spin S and the conduction-electron spins s,
and as just described it may be treated by the so-called s-d or s-f exchange
hamiltonian %e= -JS - s. Here J(< 0) is the antiferromagnetic exchange
coupling introduced in the previous section. At high temperatures the
impurities behave like free (paramagnetic) moments. However, below a
characteristic temperature, specific for each alloy system (see Table 2.1)
the isolated impurity becomes non-magnetic due to its interaction with the
conduction electrons. This temperature is known as the Kondo temperature
TK and it theoretically signifies the breakdown of higher-order perturbation
theory which is used to calculate the physical properties from the s-d or
s-f hamiltonian. In the calculations

TK = (D&J exp [ - lIZV(E,)IJJ]

where D is the bandwidth (usually taken to be of order of the Fermi energy

EF or Fermi temperature TF) and N(E,) is the conduction-electron density

Table 2.1 Isolated impurity Kondo (spin fluctuation) temperatures for important
dilute alloy systems

Noble metal-transition metal

Host Impurity: V Cr Mn Fe co Ni

cu 1000 K 2K 10 mK 30 K 500 K >lOOO K

Ag
Au 300
- K 10
1 mK <10e6
<10d6 K 0.25 K 500
- K >lOOO
- K

Transition metal-transition metal

MO
Rh 50 K 50 K 1000 K simple
- 10 K 1K 25 K I
Pd 100 K 10mK 20 mK 0.1 K I exchange
Pt 200 K 0.1 K 0.3 K 1K enhanced
26 Description of related concepts

of states at the Fermi surface. For T < TK there is a gradual loss of local
moment as the conduction electrons begin to form a surrounding cloud of
oppositely (antiferromagnetically) polarized spin. This situation, i.e., that
of a quasi-bound state, is sketched in Fig. 2.4. Its formation is not a phase
transition, but a slow transformation on a logarithmic temperature scale.
In the calculations, this broad-temperature transition to the quasi-bound
state is heralded by the appearance of troublesome logarithmic divergences
which reduce the response of the moment to external fields and enhance
the electron scattering cross-section. Note that the entity comprising the
quasi-bound state does not remain fixed in its orientation. Instead, it churns
about as a whole, a rotating, yet compensated, moment.
Experimental manifestations of the Kondo effect T < TK are:

(i) a loss of magnetic moment, the magnetization falls below its free-
moment value and the susceptibility is less than its Curie value
NP2 x (2.7)
XK = 3kB(T + TK) < xc = 3k,T
where p = gp,n[S(S + l)];.
(ii) The enhanced scattering rate creates a logarithmic upturn in the
resistivity at low temperature.

Accordingly, the total resistivity becomes for T > TK

pTOT(T) = p. + AT5 + xpl + xp,[l + 4J N(&)ln (kBTID)] (2.8)
when the phonon T5 dependence is included (pO is the residual resistivity
of the host metal and pi the residual resistivity of the impurity). For J < 0
a minimum appears in pTOT(T) since with T 6 D the last term in the above
equation diverges logarithmically as the temperature is lowered. This now-
famous Kondo minimum is illustrated in Fig. 2.5 and signifies the competition
between decreasing phonon and increasing Kondo contributions for T + 0.
Notice that TMIN a xf where x is the magnetic-impurity concentration.

Fig. 2.4 Static picture of the Kondo bound state with local moment surrounded
by oppositely polarized conduction electrons.
2.2 Kondo effect and weak moments 27

‘\--/ T min
T

Fig. 2.5 The resistivity minimum in the temperature dependence of a dilute alloy.
Notice that the Kondo contribution becomes saturated (‘unitary scattering limit’) as
T+= 0.

An alternative approach to the interaction between a local moment and

the conduction electrons is the localized spin-fluctuation model. Here the
local spin fluctuates in amplitude (recall that spin also changes its orientation
due to temperature disorder - Curie’s law) at a rate (~,~)-i where Tsf is the
spin-fluctuation lifetime. When this rate is greater than the orientational
changes produced by thermal fluctuations, the impurity spin appears non-
magnetic and there is additional conduction-electron scattering. As before,
a broad temperature interval separates the magnetic from the non-magnetic
regime, and an analogous temperature may be defined as Tsf = hl(kB7,f).
Table 2.1 collects both Kondo and spin-fluctuation temperatures for some
dilute alloy systems. Note how these temperatures span many orders of
magnitude.
These two models, quasi-bound state and localized spin fluctuation, give
quite similar experimental predictions, and for the present treatise we can
assume the two characteristic temperatures to be equal. Thus, for T %- TK
good magnetic moments exist, i.e., p = gp.,J[S(S + l)]* S(S + l), while for
T G TK, the moments become ‘weak’ and lose their magnetic character. It
is not just temperature which can remove or install the magnetic moment,
an external magnetic field H has a similar effect. Applying such large fields
to a Kondo system below TK will gradually restore the magnetization to its
high temperature value. The physics is as if the field breaks up and aligns
the antiparallel quasi-bound state or slows down the spins fluctuations
making T,, += 0. Only quite recently (early 198Os), after most interest in
the Kondo effect had subsided, an exact solution was found for the Kondo
problem. Perturbation theory cannot fully treat a single magnetic impurity
weakly coupled to a Fermi sea of electrons at low temperature. A rather
complicated and mathematical Bethe ansatz treatment was employed to
diagonalize exactly the Kondo hamiltonian and derive its thermodynamic
quantities. The contemporary physical picture is that of a many-body singlet
ground state and corresponding resonance in the density of states situated
28 Description of related concepts

at k,T, above the Fermi energy with a height a l/T,. It is this impurity
(Kondo) resonance which creates the many interesting and anomalous
properties.
What we neglected to mention before is that the specific heat of a Kondo
system is large and has a broad maximum at approximately TK. This along
with the magnetization and susceptibility has now been calculated precisely
from the above exact theory. The trouble was then to find a suitable, real,
ideal material whose thermodynamic properties could be directly compared
to theory. A similar difficulty exists, as we shall later see, with the spin
glasses. Finally, (La,-,Ce,)A12, where TK = 0.2 K and S = $ for the crystal-
field split ground state, was deemed appropriate. And a very nice,
quantitative fit between experiment and theory resulted, at least for the
measurable thermodynamic quantities, magnetization, susceptibility and
specific heat. Yet what remains to be calculated are the transport properties:
the resistivity increase and the giant thermoelectric power associated with
the Kondo effect. Up to now there have been no exact entire-temperature-
range solutions to compare with the large amounts of experimental data
available.
The occurrence of the Kondo effect is a hindrance to spin-spin interactions
and magnetic ordering. In Table 2.1 a collection of Kondo temperatures is
given for some transition-metal alloy systems which do indeed exhibit
interactions and ordering effects at sufficiently high concentrations. At very
low concentrations and below TK the moments are ‘weak’ and cannot simply
interact with each other. However, the trick here is to bring the impurities
sufficiently close to each other and turn on the RKKY interaction. For, we
can greatly eliminate the role of the Kondo effect and weak moments by
increasing the concentration and creating good moments via a strongly
magnetic local environment, since when interactions are taken into account
T,(single impurity) % T,(pair of impurities) % T,(triple), etc.

2.3 GIANT MOMENTS

The term giant moment describes the unique behaviour of dissolving a

single localized magnetic moment in a non-magnetic host and thereby
producing a net moment much greater than that due to the bare magnetic
impurity alone. This overall moment is the sum of the localized moment
plus an induced moment in the surrounding host metal. Notice now the
two important differences with our preceding Kondo effect. First of all the
interaction with the conduction electrons is dominantly ferromagnetic, i.e.,
a parallel coupling, and secondly, the host must be able to support this
positive coupling over quite a long range (not the +/- oscillations of the
RKKY) .
For the description of a large number of experimental properties, these
entities (impurity moment plus polarization cloud) can be considered as
2.3 Giant moments 29

one magnetic moment which is then called ‘giant’. A typical example of

such a system is PdFe in which the Fe moment is approximately 4kB, yet
experiments on PdFe reveal moments = 12~~ per Fe atom. Thus, via a
strong, parallel, itinerant-electron polarization extending for about 10 A,
the Fe moment induces a small average (- O.O5t+J moment on the 200 or
so surrounding Pd atoms. Figure 2.6 sketches a static giant moment. Since
this interaction is ferromagnetic, the Kondo effect is unimportant and long-
range ferromagnetism occurs at very low (- 0.1 at. % Fe) concentrations
due to the extended, indirect, ferromagnetic coupling. This behaviour
represents the simplest type of ferromagnetic ordering to appear in a dilute,
random alloy.
Sometimes in the literature the term giant moment is also used for the
large magnetic entities which are caused by chemical clustering of magnetic
atoms or groups or nearest neighbours needed to produce a moment via
the local environment. Here we shall reserve the term giant moment only
for those moments associated with isolated ‘good moment’ impurities
supported by large host polarization.
A very interesting feature of the giant moments is that at very low
concentrations (= few ppm Fe) and large separations (20-30 A) between
the Fe impurities, the ubiquitous RKKY oscillations of almost insignificant
amplitude will be felt, if the temperature is reduced sufficiently. Hence, a
spin glass is possible with the presence of mixed interactions. Tour de force

3pB(bare moment)

O.l- ’

F PdZnn

E _ Pd 3nn
7J
z PdSnn
2
.c
o-
Fe I I I 4 I I I I I I

0 L
2 !I, for f.cc. Pol 6 8 A

Fig. 2.6 Schematic picture of the giant-moment polarization cloud for an Fe

impurity in Pd.
30 Description of related concepts

experimentation has verified this with a freezing temperature in the micro-

Kelvin range.

2.4 SPIN GLASSES (A SECOND VISIT)

Since we had our first glimpse at what the spin glasses were in Chapter 1,
let us now consider what they are not. The moments must be good, not
weak - no significant Kondo effect. The concentration must be large enough
so that there is enough competition from the mixed interactions - not the
ferromagnetic predominance of the giant moments and not long-range,
spatial ordering of the usual ferro- or antiferromagnetic transitions.
This last condition is quite interesting. Mathematically, it may be expressed
as ZiSi = 0 (no ferromagnetism), but (S,), f 0 where ( ), represents a time
average over a long observational time t, and in addition

&~(SJ,eXp(ik’Ri)=O asiV+m
I
This equation states that there is 120 long-range antiferromagnetic ordering,
i.e., the directions of the spins are not related to a particular wave vector
k, as is true for even the most complicated of antiferromagnets with lattice
vectors Ri.
Another way of viewing the spin ordering is via their Ising +l spin-spin
correlation function

(SiSj): z exp (- I:si &I)

(SJj), = 0 but

where cSG is the spin-glass correlation length. The crucial questions here
are: does the spatial correlation length & + ~0 as T + T,, and does the
decay of the ‘order parameter’ governed by a characteristic time become
infinitely long signifying a phase transition? The first would imply the
second as in a conventional second-order phase transition. However, the
spin glasses are different and we do not really know the final answers to
the above questions. Long is not infinite. So let us now hypothesize a phase
transition and study its consequences. It follows that we can replace the
time averages by the ensemble averages of statistical mechanics and define
an order parameter for the frozen spin-glass phase:
4 = ((SJ$)c # 0 for T I Tf (2.11)
where the inner brackets represent ensemble (or thermal) averages while
the outer brackets are for a configurational average over the set of random
interactions. 4 is called the Edwards-Anderson spin-glass order parameter
and will appear often as we tread further into the models and mysteries of
spin-glass theory.
2.4 Spin glasses (a second visit) 31

Since the magnetic impurities have random positions in the lattice, the
effective molecular field, H, may be calculated as a superposition of
contributions from all the impurities in the alloy, due to the infinite range
of the RKKY interaction. Furthermore, H has a distribution of magnitudes,
P(H), and a random distribution of directions because of the oscillatory
nature of the RKKY interaction. A computer simulation has been made
for P(H) using an fee lattice with about 100000 sites, 324 being occupied
with spins (x = 0.3%), and an RKKY interaction performs the coupling
among these vector spins Si. The results for P(H) are shown in Fig. 2.7.
The reader should compare this figure to Fig. 1.6 which gives the probability
of RKKY interaction strengths. Notice that P(H) + 0 as H + 0 meaning
that at each lattice site occupied by a spin there will be a finite internal
field having both a magnitude and direction. Although P(H) generates the
same function for a large enough spin system, it is not required that all the
spins point in exactly the same direction for each simulation or ‘freezing’.
The freezing temperature Tf of a spin glass may be viewed from a very
simple model. By considering only the envelope of the RKKY oscillations,
J’(r) = J,lr3, we have the strength of the interaction as a function of
distance. For a low-concentration alloy the average distance of impurity
separation is r E (l/x)“. Now we use the thermal disorder energy, k,T, to
destroy the correlations between the spins, on average, P apart. In other
words, the exchange energy caused by the RKKY interaction from a spin
Sr, acting at a distance P on another spin S2 must be greater than thermal
disordering for freezing to occur. At the freezing temperature,

J’(r) = k,T, or Tf=x$ (2.12)

B

This means that no matter what the concentration is, there will always exist
a temperature for which the RKKY interaction will prevail and produce
this random (due to the oscillations) freezing. Here only good moments are
assumed, i.e., no Kondo effect or TK = 0.

0 Oil2 0.04 M6 WI3

h

Fig. 2.7 Probability distribution of ground-state, internal-fields magnitudes for an

RKKY interacting spin glass. From Walker and Walstedt (1977).
32 Description of related concepts

A further remark about these dilute alloys and the RKKY interaction
has to do with the relation (xp3) = constant. If a measure of length is used
which is related to the system of impurities, then we can ‘scale’ the
properties for one concentration x1 to those of another x2. Our length scale
is now the average distance between impurities J, so the volume f3 contains
a constant number of magnetic impurities equal to xf3. This has particular
significance for the thermodynamic properties. Magnetization, susceptibility
and specific heat can be expressed in the form
ii4
-= HT c
and p- - H T withT,ax.
x f(- x ‘x-) x fi- x ‘x-)
(2.13)
Experiments in the dilute limit [just above the concentration, x,, where the
Kondo effect plays a major role, viz. TK = T,(x,)] verify these ‘scaling
laws’ and thereby prove the efficiency of the RKKY interaction in creating
the spin-glass state. We shall return to the roots of this scaling at the
beginning of Chapter 5.

2.5 MICTOMAGNETISM (CLUSTER GLASS)

As the concentration of magnetic impurities is increased, there is a greater

statistical chance of the impurity being first or second nearest neighbour to
another impurity. Since the wave functions for the 3d electrons (transition
metal) have a finite extent, they can also carry a type of RKKY-polarization.
A similar situation occurs for the more localized 4f electrons (rare earths),
but now through an intermediate polarization of 5d electrons. Consequently,
there is a short-range RKKY-like interaction which can strongly couple
neighbouring impurities ferro- or antiferromagnetically depending upon
the particular magnetic element involved and the neighbour position.
Accordingly, magnetic clusters can form as a result of concentration
fluctuations in a random alloy. In addition, at these larger concentrations
short-range chemical or atomic order, and even chemical clustering of
longer-range (due to solubility problems of the given alloy) may occur.
Such deviations from randomness can greatly influence the local magnetism.
It has been experimentally found for a number of typical spin-glass alloys
(e.g. CuMn and AuFe) that both of the above mechanisms (statistical
fluctuations and chemical ordering) produce a predominance of ferromagnetic
couplings and very large effective moments of order 20-20000 l.+$. When
the magnetic behaviour is dominated by the presence of such large magnetic
clusters, the terms ‘mictomagnet’ or ‘cluster glass’ have been used in the
literature. Figure 2.8 shows some of these clusters within the dotted lines
embedded in the spin-glass matrix. At sufficiently low temperature, these
clusters will freeze with random orientation in a manner analogous to the
2.5 Mictomagnetism (cluster glass) 33

................................... ....
..\......~~..~.......r ...\
.. ..w.ir(........~~) ...... ..-CO..
b....-w.+.......~/#f.......~~~ ..
.. ..wb~......I...i#/p..f....~\.~* .
... ..h.~../....irf*b....~~.\.~ ..
..~..i..i...................~.\ ...
......... ................... ..*hv
../...~.~.q....f...........-..b ....
....... ..*~.j........~ ..........
... ..Y...../*: ..................
........ ..~.~.~~.............\ ...
.. ../.....C1’b......j ...........
\.......b.......f..b..~ ........
.... ../ ........................
..b.*
........ ................
..................... ..~~.....#..~
..~...............~~-~.....-......
. . . . . . . ..m.r....i~~.........

.......
.........
. ..#.b........b.-.

Fig. 2.8 Spin glass with mictomagnetic cluster for a 2D square lattice.

spin-glass freezing. The presence of such large ferromagnetic entities

simplifies detection of the freezing process, for, the magnetization and
susceptibility are very large. Also enhanced are the remanences and
hysteretic behaviour of the frozen state. However, the mictomagnetic phase
is especially sensitive to its metallurgical state. Heat treatments, with various
times, temperatures and cooling rates, along with plastic deformation can
greatly affect the cluster formation and magnetic behaviour. Of special
interest here are the hard magnetic properties such as the peculiar displaced
hysteresis loops and the remanences and irreversibilities in the magnetization.
The origin of mictomagnetism goes back to a series of early investigations
in the 1960s where an ‘ensemble of domains’ model was used. In the 1970s
the name mictomagnetism was coined for a continuation of such experiments
where careful consideration was given to the metallurgical state and the
effects of random freezing below a certain temperature. Other workers, a
few years later, used the name cluster glass to describe similar very large
moment effects in more concentrated alloys. Since mictomagnetic and
cluster glass refer to alloys of higher concentration where long-range
magnetic order is incipient, they represent a complicated melange which is
difficult to study or to model. Therefore, this topic will not receive great
attention in the following sections.
34 Description of related concepts

2.6 PERCOLATION AND LONG-RANGE MAGNETIC ORDER

For the sake of completeness, let us proceed and further increase the
number of magnetic atoms. At some given concentration for the particular
crystal structure, each magnetic site will have at least one magnetic nearest
neighbour. When such a macroscopic connection or uninterrupted chain
extends from one end of the crystal to the other, the percolation limit (that
concentration for which there is a non-zero probability that a given occupied
site belongs to an unbounded cluster) has been reached. The schematic 2D
lattice shown in Fig. 2.8 is far below the site percolation concentration
which, for a square lattice with only nearest neighbour coupling, requires
58% of the total sites to be occupied. Similar critical concentrations for
percolation may be obtained for other 2D and 3D crystal structures with
various near-neighbour interactions.
Another type of percolation is bond percolation where an interaction or
bond which couples two particular sites is present or absent. Then the
closed bonds must continue in a macroscopic path through the crystal. A
simple physical picture for bond percolation is conduction through a network
of on/off switches. How many switches between the sites of our square
lattices (Fig. 2.8) need be closed before a current flowing in one end of the
network will flow out the other? 50% is the answer.
Percolation applied to magnetic systems is more complicated. Let us
begin with a 3D fee lattice and simple nearest-neighbour interactions, and
randomly fill the lattice sites with magnetic atoms; if they are nearest
neighbours, they will align parallel; if not, they are uncoupled. When 17%
of the sites are occupied an infinite ferromagnetic cluster will form. This
represents long-range ferromagnetic ordering and a proper phase transition
at a Curie temperature TC. Certainly there will be modifications in the
critical phenomenon due to the percolation behaviour and its fractal nature,
and also because of the many ‘free’ spins which are not (or only weakly)
coupled to the infinite cluster. With nearest-neighbour antiferromagnetic
bonding in the same 3D fee crystal, a much larger concentration is needed.
Here the element of frustration or unsatisfied orientation enters, because
the fee triangular symmetry is unfavourable for long-range antiferromagnetic
order and many more spins are needed to build this coherent two sub-
lattice structure in fee space. Now the percolation limit is = 45% site
occupancy above which we have a long-range antiferromagnet with a distinct
T N*
We can take the following step in our percolation approach and allow
mixed interactions, remembering this is the essential ingredient for our spin
glasses. For example, return to Fig. 1.1 where the first neighbour was
ferromagnetic and the second, antiferromagnetic. There does exist an
‘infinite’ coupled cluster, but now without long-range ‘periodic’ order. The
spins are frozen randomly depending on the initial condition. Here we
might say that the spin glass correlation length &o + 03, but recall the
2.7 Concentration regimes 35

example of Fig. 1.1 is at T = 0 K as usual for percolation problems.

Nevertheless, we can still attempt to use the concept of percolation, even
cutting off the envelope of the long-range RKKY interaction with the
thermal disorder, as was done in section 2.3. Now we do not care how the
spins are coupled, only if they are. The main difficulty with the percolation
approach is that it explicitly neglects frustration - an essential element of
the frozen state and something that is built into the spin-glass transition.
So we conclude that percolation theory can only give an oversimplified and
incomplete picture of a spin glass. However, it becomes quantitative and
more useful in describing the various types of long-range magnetic order
appearing at larger concentrations above the spin-glass phase.

2.7 CONCENTRATION REGIMES

After all this discussion we should try to summarize the various concentration
regimes of different behaviour for a metallic magnetic alloy. It must now
be clear that the concentration is most important in determining the
magnetic state of the alloy. Figure 2.9 offers a schematic representation for
such a concentration regime division. At the very dilute magnetic concen-
tration (ppm) there are the isolated impurity-conduction electron couplings
which result in the Kondo effect. This localized interaction (if J < 0) causes
a weakening or fluctuation of the magnetic moment, and below the Kondo
temperature, the moment disappears and the impurity appears non-magnetic.
Thus, the Kondo effect prevents strong impurity-impurity interactions which
are a basic necessity of the spin glasses. Nevertheless, the local-environment
model tells us that there will be pairs (or triplets, etc.) which possess a
much lower TK than the singlets. These ‘good’ moment pairs may then
interact with other pairs and give rise to spin-glass behaviour. Therefore,
in principle, there is no lower concentration limit to spin-glass consequences,
until we run out of measurement response or temperature. Usually, however,
we take TK = T&J as the concentration, x0, below which the Kondo
effect plays a large role in modifying the spin-glass behaviour.
Up to a concentration of few thousand ppm (- 0.5 at. %), the scaling
approach (see section 2.4), based on the (1/r)3 decrease of the RKKY
exchange is particularly appropriate for describing the problem of interacting
impurities. This means that the measurable properties (magnetization,
susceptibility, specific heat, and remanences) are universal functions of the
concentration scaled parameters T/x and H,,,lx, and in addition Tf c( x.
While an effective way to scale the various data, such a treatment tells us
nothing about the freezing process and the possibility of a phase transition.
At low temperature T -@ Tf and with the incorporation of anisotropic dipolar
interactions, also m(llr)3 and thus scalable with x, the magnetic behaviour
was initially explained within the existing model of NCel superparamagnetism.
Superparamagnetism (see below section 2.10) is the temperature and field
 cluster formation

Exp. properties-

..,..,
::;,:.:: variables 8
= 50 ppm = i at. % = 10 at. % Magnetic
percolation ‘p9,
T, = To(c) limit %
3
RKKY RKKY + short range correlation direct

Dipolar
interactions

Fig. 2.9 Various concentration regimes for a canonical spin glass illustrating the different types of magnetic behaviour which occur.
2.8 Metallic glasses: amorphous magnets 37

activation of non-interacting ‘fine’ magnetic particles or large clusters with

respect to their anisotropy energy barriers. This served as a first attempt
to treat the frozen state and it will be considered in more detail at a later
stage.
Although the scaling regime and its companion model represent a
reasonable first-order approximation at low x, the random freezing persists
over an extended concentration range which exhibits the more dramatic
experimental effects. After the scaling laws break down around i at. %, a
non-scaling spin-glass regime occurs where Tf 0: x$. Here clusters (pairs,
triplets, etc.) begin to form and exert their influence. When the magnetic
behaviour is dominated by such clusters, at around lo-15 at. %, the
nomenclature mictomagnetism or cluster glass is used to emphasize the
metallurgy and anomalies generated by these very large magnetic clusters.
Finally, the percolation limit is reached for long-range, but inhomogeneous,
ferro- or antiferromagnetic order with a well-defined Tc or TN. The
particular characteristics of the infinite-cluster, long-range order, while of
interest in itself, will not concern us any further. In the subsequent chapters
we employ the generic term spin glass from the dilute limit almost up to
the percolation limit for long-range order. Hereby we can avoid, for the
sake of simplicity, the mostly unnecessary distinction of having three
or more spin-glass regimes. Once again Fig. 2.9 illustrates the various
concentration regions. Note that the different regions are not sharply
separated; there is a smooth and gradual transformation from one to the
other (except perhaps for a well-behaved alloy at its percolation threshold).
A temperature - second dimension - ordinate should be added to Fig. 2.9
in order to give it the character of a T-x phase diagram. In Fig. 2.10 such
a general scheme is shown. TK represents the ‘average’ Kondo temperature
which decreases as a function of the concentration. For TK > Tf we have
the weak moment concentration regime. Then as the local environment
builds good moments (X > x0), the spin-glass regime appears with first a
linear, followed by a less than linear Tf dependence on X. Finally, when
the percolation limit is surpassed x > .rr there is a rapid rise of Curie or NCel
temperatures as the long-ranged ordered state is formed and strengthened. So
by properly adjusting the values of X, and X, and scaling the three
temperatures of interest TK, Tf and Tc (or TN), a number of real alloy
systems may be described with the generic phase diagram of Fig. 2.10.

2.8 METALLIC GLASSES: AMORPHOUS MAGNETS

This name has been given to a new class of non-crystalline solids which are
conducting and contain magnetic elements. Most of the metallic glasses are
naturally based on rather large concentrations of transition elements (TM),
e.g. FesoP2, and COAST, so long-range ferromagnetism is a common
occurrence. It is only when the TM-concentration is reduced by non-
38 Description of related concepts

Fig. 2.10 General (schematic) T-x phase diagram for a dilute magnetic alloy.

magnetic substitutions, Pd or Ni, which now have lost their magnetic

tendencies due to their hybridization with the metalloid (P or B) atoms,
that a spin-glass phase appears. A similar situation may be created using
rare-earth atoms as the magnetic impurity of the amorphous (a) compound.
Some of the many examples here are a-Laso-XGdXAu20 and a-Dy,-,Cu,.
The spin-glass ingredients in these amorphous magnetic systems are
essentially the same as for our dilute magnetic alloys of the previous
sections. Since they are also metals, the RKKY interaction is operative in
its usual competing-exchange form. Very similar behaviour and regimes
exist as a function of x - the magnetic concentration. The ‘double disorder’
of random site occupancy within the non-crystalline lattice certainly increases
the overall randomness and eliminates the problem of chemical clustering.
Look-alike T-x phase diagrams can be mapped out for the numerous
amorphous magnetic alloys which further have the great advantage of being
able to be prepared over a wide concentration range, in many different
combinations, without change in (a) structure. We shall not discuss the
various elaborate techniques employed to ‘quench a liquid’ and thereby
form an ‘amorphous’ spin-glass.
Two novel features manifest themselves in the amorphous magnets.
Firstly, due to the crystallographic disorder, the electron mean free path I
is significantly reduced. This introduced a dampening factor, approximately
e --rl’, into the RKKY-interaction which effectively cuts off the (llr)3
polarization at a smaller length scale. Secondly, again due to the random
environment of the non-magnetic species, the various rare-earth impurities
(except Gd) will experience the effects of random anisotropy. Namely, the
2.9 Re-entrant spin glasses 39

direction of the uniaxial anisotropy becomes a random variable. This means

we must write the anisotropy energy in our hamiltonian (see Section 1.4)
as

“deRA = -CD,(n, . SJ2 (2.14)

where ni represents different, randomly oriented, easy directions at the

various sites i. The net result here is to enhance the spin-glass behaviour
to the detriment of the ferromagnetic ordering. Thus, the amorphous
magnets offer many ‘modern combinations’ (see Chapter 4) of spin glasses
without really adding basically new effects or fundamentally differing
insights.

2.9 RE-ENTRANT SPIN GLASSES

At the percolation boundary of phase diagram, recalling Fig. 2.10 and its
dashed lines, a strange behaviour takes place, if the long-range order is
ferromagnetic. The spin-glass phase does not simply stop but persists into
the ‘ferromagnetic’ phase. This effect is shown for a real material Au,-,Fe,
in Fig. 2.11. The strangeness is related to the observation that the more
disordered state appears from the more ordered one as the temperature is
reduced. According to experiment it would seem that the T = 0 phase is a
cluster glass created out of the ‘percolated’ infinite ferromagnetic cluster
which has broken up into large but randomly frozen clusters. Hence, the
apparent long-range order is destroyed although a gamut of large ferromag-
netic clusters is still present - conforming to our previous definition of a
cluster glass or mictomagnet. The word re-entrant, something of a misnomer,
means here that the cluster-glass phase develops from a ferromagnetic state,
and thus it re-enters the frozen (disordered) phase out of another, not
paramagnetic, state.
There are many different systems especially among the amorphous metallic
magnets which show this re-entrant behaviour. So the phenomenon is
general, yet a good theoretical model is hard to find. Various suggestions
have been proposed, one of which is the appearance of a temperature-
dependent random anisotropy (dipolar?). At low temperatures this anisotropy
grows sufficiently large that it severs many of the weak links holding
together the rather tenuous x z xP infinite cluster. Then the anisotropy
points these divided units along the randomly preferred directions. Another
model involves a freezing of transverse spin components. At the ferro-
magnetic transitions T, the longitudinal components order, but because of
the local mean field being inhomogeneous, a random freezing perpendicular
to the applied field occurs at a lower temperature Tf.
Recent experimental work indicates that the ‘infinite-cluster’ ferromagnetic
state is not truly long-ranged. Instead, it consists of many extremely large
40 Description of related concepts

l Mossbauer
200
l Susceptibility
0 Neutron scattering
0 Specific heat
f
A. Magnetic resonance

160

/
para
I
120 ferro
0 I
e-
l- I

4
I
80 -
b

: / 1

O 4 8 12 16 20 24
x(at%FeI

Fig. 2.11 Magnetic phase diagram of AuFe constructed from the anomalies observed
in various experiments according to symbol; from Coles et al. (1978).

clusters which are hard to tell from infinite. These large, but disordered
from each other, clusters then dissociate into the collection of smaller
clusters at the re-entry transition. Such processes occur without the usual
type of percolative phase transition.

2.10 SUPERPARAMAGNETISM

Suppose we ponder a family of magnetic clusters, also called domains or

fine particles. They may be composed of any type of internal magnetic
order, namely, ferromagnetism, antiferromagnetism, random, etc., and their
2.10 Superparamagnetism 41

ordering already has occurred at a temperature far above our considerations.

The clusters, in contradistinction to the previous cluster glass, do not
interact with each other. So, we ask, what happens to this collection of
independent particles, constituting magnetic moments proportional to either
n or An or dn (n is the number of local moments within a cluster) as a
function of applied field and temperature.
First of all at reasonably high temperature, the moments of the various
particles fluctuate due to the thermal disorder in a paramagnetic way. One
may, therefore, use the Brillouin function to describe the magnetization

M~+]=M,(n)[(S+t)coth((S+Q~)-Icoth(!!)]
(2.15)
Since each particle contains many spins, we require the limit of the Brillouin
function as S + 03. Can you remember how this function depends upon S?
For S becoming large the initial slope decreases and there is substantial
rounding as the argument increases making saturation more difficult.
Confirmingly, experimental data nicely fit the Brillouin function with S = 03,
at least if T is not reduced too far. Since the measured magnetization for
our assemblage of particles is zero when no external field is applied, and
follows the Brillouin function for infinite spin when a field is swept, the
name superparamagnetism aptly fits.
Now we lower the temperature to look for additional effects. Here we
must include the consequences of anisotropy in an energy-barrier model as
depicted in Fig. 2.12. There are many possible sources of the anisotropy:
magnetocrystalline, shape, dipolar, surface, etc. The anisotropy constant
times the particle volume forms the energy barrier height, E, = KV which
can be overcome with either field M . H or temperature kBT energy. Hence,
two minima appear with oppositely directed spin, i.e., two easy orientations
of magnetization. The relaxation time of the magnetization between
these two states is given by thermal activation (Arrehenius law)

0 IT
Fig. 2.12 Anisotropy barrier for an up/down superparamagnet. Note the thermal
activation energy (kBT) is too small to cause rapid barrier hopping.
42 Description of related concepts

T = 7, exp (E$knT), where E, is the energy barrier separating the states

and T, is a microscopic limiting relaxation time usually = 10e9s. Below a
blocking temperature TB of order E$25kB, the fluctuations between the
two states are becoming long enough to be observable on a laboratory time
scale. For a yet lower temperature the system will appear ‘blocked’ in one
of the states. With the energy of this state lowered by the external field a
stable remanence would emerge when averaged over the many particles.
Commensurate with the blocking process is a drop of the magnetization;
this manifests itself as a broad temperature-dependent maximum in M(T),
since the various particles are blocked out at different temperatures.
For spin glasses there are strong interactions between the moments. So
the concept of independent superparamagnetic particles is incorrect. Certainly
the freezing process of a spin glass is co-operated and phase-transition-like,
in stark contrast to the gradual blocking of the superparamagnetic particles.
Nevertheless, as a naive and crude approximation we can still use some of
the above ideas for the frozen state. Historically this was the first comparison.
There are always some correlated spin clusters which still have a random
distribution of barrier heights and anisotropy directions. The frozen spin
glass locks into this array and its ground state exhibits the magnetic
remanence while relaxation is associated with the thermal activation. At
low temperatures T < Tf the relaxation is slow, proportional to logarithm
of the time, and a metastable random state arises. As we shall soon see it
sometimes becomes arduous to distinguish experimentally between an
insulating spin glass and a superparamagnet.

2.11 MAGNETIC INSULATORS AND SEMICONDUCTORS

While we can discriminate reasonably well among the metallic materials as

to what is a spin glass and what is not, there are more difficulties with
doing the same for non-conducting materials. The main problems with the
latter systems are the short range and weakness of the interactions which
are present, and the predominance of only antiferromagnetic exchange
coupling in many of the disordered insulators. Of course, for the very
few materials, which have disorder, mixed (ferro and antiferromagnetic)
interactions and frustration, we have no qualms about grouping them
alongside the canonical RKKY spin glasses (see listing of spin glasses in
Chapter 4 and the ideal spin glasses in Chapter 6).
In a manner similar to the dilute magnetic alloys or the amorphous
magnets we can create site disorder or a distribution of impurity distances
in the form of insulating glasses. So there is no problem with randomness.
Yet for short-range exchange J(r) is proportional to either exp (-r) or (r)-p
where the power p is 8 to 10, thus a strong cut-off in the range of coupling
occurs after the first or second nearest neighbours. Here above the percolation
Limit xi, for the given crystal structure, the long-range antiferromagnetism
2.11 Magnetic insulators and semiconductors 43

manifests itself, albeit with reduced TN and modified critical behaviour

compared to the pure (nondiluted) system. Many of these cases are without
any exchange-frustration effects. The study of such alterations in the phase-
transition properties is called the random-exchange problem. One particularly
suitable case for experiment and comparison with theoretical predictions is
when the system has a simple Ising critical phenomenon; accordingly the
name random-exchange Ising model is used. Or, if a similar system is
placed in a large magnetic field, it becomes an experimental demonstration
of the so-called random-field Ising model (see Chapter 7).
What takes place below xP (the magnetically diluted limit)? Well,
most measurements indicate that the system behaves analogous to the
superparamagnets. The infinite antiferromagnetic cluster is broken up in
many various size domains of antiferromagnetic order which are independent,
i.e., do not interact with each other, and are therefore not correlated.
Consequently, even though the many clusters are composed of antiparallel
spins, there are still small moments to be associated with their irregularities
and surfaces, and gradually these are blocked out according to the model
introduced in the previous section. Note, as already mentioned, there are
a number of similarities with the frozen state of a spin glass and the blocked
superparamagnet. However, some major dissimilarities are that the time
scales become more smeared out and the blocking temperature TB cannot
be compared to the sharp and well-defined spin-glass freezing temperature
Tf. To see these differences clearly, requires a lot of fine experimentation
- it’s a tough job here to sort out the true spin glasses.
What happens if we do have a lattice favouring frustration and sufficient
antiferromagnetic exchange to form an infinite cluster? Will then the system
be a spin glass? These questions are genuine, for, real systems do exist
especially with the semimagnetic semiconductors such as Hg,-,Mn,Te and
Cd,-,Mn,Te. Even when doped their conductivity is not large so the RKKY
mechanism is inoperative and the antiferromagnetic-only exchange is
mediated by the valence electrons - a type of superexchange (see section
1.3.3). Once again these systems do have many spin glass properties
particularly in the frozen state. Yet they do not possess the sharp freezing
behaviour of the canonical spin glasses. It would seem that they are spin-
glass-like at low temperatures, but somehow enter this state via another
and more gradual type of freezing process. To know what we are talking
about we must distinguish between superparamagnets, canonical (RKKY)
or good spin glasses, spin glass-like and completely frustrated non-random
materials. These distinctions will only become clear when we consider in
detail the experimental properties of the mixed interacting spin glasses.
And such is our task for the next chapter.
44 Description of related concepts

FURTHER READING AND REFERENCES

Related concepts

Kondo, J. (1969) in Solid state physics, Vol. 23 (eds F. Seitz and D. Turnbull)
Academic Press, New York, 183.
Heeger, A. J. (1969) Ibid, 293.
Rado, G. T. and Suhl, H. (eds) (1973) Magnetism V, Academic Press, New York.
Nieuwenhuys, G. J. (1975) Adv. Phys., 24, 515.
Mydosh, J. A. and Nieuwenhuys, G. J. (1980) in Ferromagnetic materials, Vol. 1
(ed. E. P. Wohlfarth) North Holland, Amsterdam, 71.
Moorjani, K. and Coey, J. M. D. (1984) Magnetic glasses, Elsevier, Amsterdam.
Stauffer, D. and Aharony, A. (1992) Introduction to percolation theory, Taylor and
Francis, London.

Spin glasses

Mezard, M., Parisi, G. and Virasoro, M. A. (1987) Spin glass theory and beyond,
World Scientific, Singapore.
Fischer, K. H. and Hertz, J. A. (1991) Spin glasses, Cambridge.
Chowdhury, D. (1986) Spin glasses and other frustrated systems, World Scientific,
Singapore.
Binder, K. and Young, A. P. (1986) Rev. Mod. Phys., 58, 801.
Maletta, H. and Zinn, W. (1986) in Handbook on the physics and chemistry of rare
earths (eds K. A. Gscheidner, Jr and L. Eyring) North Holland, Amsterdam.
Rammal, R. and Souletie, J. (1982) in Magnetism of metals and alloys (ed. M.
Cyrot) North Holland, Amsterdam.
Mydosh, J. A. (1981) in Magnetism in solids: some current topics (eds A. P.
Cracknell and R. A. Vaughan) SUSSP, Edinburgh.
van Hemmen, J. L. and Morgenstern, I. (eds) (1983) Heidelberg colloquium on
spin glasses, Vol. 192 Lecture Notes in Physics, Springer, Berlin.
van Hemmen, J.L. and Morgenstern, I. (eds) (1987) Glassy Dynamics Vol. 275,
Lecture Notes in Physics, Springer, Berlin.

References

Coles, B. R., Sarkissian, B. V. and Taylor, R. H. (1978) Phil. Mag. B, 37, 489.
Walker, L. R. and Walstedt, R. E. (1977) Phys. Rev. Lett., 38, 514.
Walker, L. R. and Walstedt, R. E. (1980) Phys. Rev. B., 22, 3816.
3
Basic spin-glass
phenomenon

We have been considering for the past two chapters the spin glasses and
their entourage of cognate phenomena. After having spent so much time
and energy introducing the encompassing physics and basic principles of
the spin glasses in a rudimentary and qualitative way, we now should turn
to experiment in order to find out exactly how it supports and confirms
these notions, i.e., our physical picture, of a spin glass. Accordingly,
measurements on real systems and their proper interpretation are the
windows through which we gain our knowledge. Let experiment be our
guide - after all it did come first.
We shall attempt in this chapter to survey salient experimental features
of a spin glass beginning with the high-temperature measurements and then
reducing the temperature to T = Tf. Here we examine the novel critical
behaviour and the various manifestations of a phase transition. Finally, the
important experiments characterizing the low-temperature frozen state
(T < Tf) are viewed with respect to its unique properties. Our three
temperature regime approach is arbitrary, yet effective, since three very
different types of behaviour are encountered. Our choice of a few from the
many thousands of experiments published on spin glasses is geared to bring
our physical picture into focus and keep it generic - not a collection of
exceptions. Thus, we want to embed our notion of a spin glass firmly in
reality before proceeding with models and theories which, of course, should
give the ultimate explanation of the phenomenon and lead to a final
understanding.

3.1 HIGH-TEMPERATURE (Ta 7J EXPERIMENTS

Perhaps the dullest part of the spin glasses is their behaviour far above the
freezing temperature. For, we would expect simple paramagnetic effects in
this limit, or a collection of freely rotating, independent magnetic moments.
Using a paramagnetic model we can compare this basis to the generic spin-
46 Basic spin-glass phenomenon

glass measurements of susceptibility, specific heat, resistivity and neutron

scattering. Systematic data have been tediously accumulated over the past
years for these experimental quantities. We know what to anticipate so let’s
see what will be found as Tf is slowly approached from far above.

3.1.1 Susceptibility

For a paramagnetic, the susceptibility x is just x = M/H since we usually

measure it in the low-field limit of the magnetization’s Brillouin function.
There exist many techniques with sufficient sensitivity to detect either x
directly or M in a given small H. Fortunately, for the majority of the
canonical spin glasses, the host susceptibility is small, e.g. the noble metals
have a tiny temperature-independent diamagnetic contribution which is
known and can be subtracted.
Now we carry out a systematic study on many different concentrations
of a spin glass alloy, say CuMn. Figure 3.1 shows the results plotted as

temperature (K)

Fig. 3.1 The reciprocal susceptibility as a function of temperature: the dashed lines
are a linear extrapolation to determine 8 and the solid line represents the fit to
the model calculations; from Morgownik and Mydosh (1981).
3.1 High-temperature (T % TJ experiments 47

inverse susceptibility versus temperature, recall equation (2.1). Such a plot

permits a Curie-Weiss type of analysis
A$& g2S(S + 1)
x= 3k,(T - 0) (3.1)

Notice that the temperature dependence is asymptotically good as T

increases. There is the expected straight line behaviour whose extrapolation
gives the Curie paramagnetic temperature 0. The freezing temperature is
not seen clearly in these measurements (it is smeared out), since they were
performed in a rather large field (0.6 T) in order to increase the sensitivity
at high temperatures.
Further analysis is done by calculating the effective moment; recall
equation (2.2)

(3.2)

In the case of a Curie-Weiss behaviour p(T) is simply the usual effective

Bohr magneton number as given in equation (2.3)

p(T-+ w) =pc, = g[S(S + l)]* (3.3)

Figure 3.2 presents the results of this analysis for the same CuMn
concentrations. Note how all samples have very similar p,, values = 5 in
units of pu; see Fig. 3.3. Also, returning to Fig. 3.2, pay attention to the
upturns in p beginning at a ‘deviation temperature’ Td. At T < Td(* T,)
the effective moment begins to increase. The meaning is clear; ferromagnetic
clusters are beginning to develop out of the paramagnetic background. The
greater the concentration of Mn, the higher T,, and the larger p( T) becomes
at low temperatures. This growth of local ferromagnetism is further reflected
in O(x) and its rapid increase with concentration. In standard theory the
Curie paramagnetic 0 represents a sum of all the exchange interactions
which are present

O(x) = x

for a completely random alloy. Thus, the ferromagnetic exchanges are

becoming more probable as x increases. And Fig. 3.3 shows this behaviour
with its roughly linear x dependence of 0 (except at low concentrations).
One can repeat the same series of measurements on many different
systems with various concentrations. Interestingly, most of the canonical
(RKKY) spin glasses show a lot of local ferromagnetism developing far
above Tf. With so much data available one could even attempt a simple
ad-hoc cluster model and calculate with the aid of a computer the effects
of first and second nearest-neighbour pairs and triplets on the susceptibility
x(T, x). Some of these clusters are illustrated in the right-hand panel of
48 Basic spin-glass phenomenon

table of
flrst 11
:Onf IQUP
QPO”P5
OP D t.c c
lattlco

.
M
.-*
v

v
r
-
Yi

” ‘*

.-?4

100 150 200 ..*a

tempecoture (K)

Fig. 3.2 P as defined in the text versus temperature. The deviation from C-W
behaviours at Td is illustrated in the inset; also shown are the first 11 configuration
groups for an fee lattice. From Morgownik and Mydosh (1981).

Fig. 3.2. The hope is to obtain a better than Curie-Weiss fit, at least a bit
further down in temperature. But such a model cannot be accurate as Tf is
approached due to the co-operative interaction effects - pairs and triplets
will no longer suffice. A few years ago such a model was turned loose on
a variety of dilute alloys (i < x < 5 at. %) and even included the deviations
from randomness due to chemical clustering.
The model was a computer calculation of all possible energy levels for
the various cluster configurations according to the hamiltonian

(3.5)
i<j i

where Si = S, S - 1, . . . . -S. Since the experimentally determined S-value

(pO = 5.01~~) for CuMn is very close to 2, we set S = 2 with g = 2. The
applied magnetic field H is equal to 0.6 T as in the experiment. The
canonical ensemble is then calculated taking into account all states:
3.1 High-temperature (T + TA experiments 49

r I I I I I I

d , I I I
0 JO
0 2 4 6
concentration (at % Mn)

Fig. 3.3 Effective Bohr-magneton number p. and the paramagnetic Curie tempera-
ture 0 as a function of concentration in CuMn: p. and 0 are determined from
least-squares fitting of the susceptibility for temperatures above Td; from Morgownik
and Mydosh (1981).

5(=2S + 1) for the single impurity, 5 x 5 for each doublet group, 5 X 5 X 5

for the various triplet groups, etc. The total ensemble 2 is obtained as a
product of the group ensembles weighted with occupational probabilities
for the various cluster configurations at the given concentration and degree
of short-range chemical ordering. Only for a limited x-range (- -‘,to 5 at. %)
does the model work. Once the partition function 2 is known, it is a
straightforward task to calculate the thermodynamic properties from the
usual equations.

(3.6)

(3.7)

The fitting parameters for this configurational cluster model were the
various exchange interactions as a function of neighbour shell (or distance).
The solid lines in Figs. 3.1 and 3.2 are the model calculations. They do
indeed fit better than the Curie-Weiss law. However, they cannot be used
too close to Tf. The net outcome was an estimate of the exchange parameters
J,,, extended, with certain modifications, to fifth nearest neighbour. Such is
illustrated in Fig. 3.4 along with the expected RKKY conduction-electron
polarization for a Mn impurity in Cu. There is more ferromagnetism present
in the model calculations than in the RKKY interaction. It would seem
that for the close by impurities (first or second neighbours), d-like conduction
electrons are mediating the indirect exchange interaction, thus causing the
50 Basic spin-glass phenomenon

o Au
- Fe
V Cu
- Mn
1
l -Au Mn
o -Pt Mn

L s?L---\-.3 45671

-1 “”
0 0.5 1.0 1.5 2.0
interatomic distance (units of lattice constant)
Fig. 3.4 Exchange parameters J,, as function of neighbour distance: the dashed
line represents the RKKY-conduction electron polarization; from Morgownik and
Mydosh (1983).

extra ferromagnetic couplings and discrepancy with standard free-electron

(s-like only) RKKY theory, equation (11). Nevertheless, the oscillating
ferroiantiferromagnetic interactions are clearly and happily at hand. So
while experiment nicely verifies this competition of mixed couplings (except,
recall section 2.1 .l, for the insulating Mn and Co aluminosilicate glasses
where Jr and Jz are both antiferromagnetic), it warns about the presence
of ferromagnetic clusters - the building blocks out of which the spin-glass
state is established. A similar conclusion was reached for the more
concentrated CuMn alloys on the basis of a Brillouin function analysis. The
small number of large ferromagnetic clusters determined from the fitting
gave birth to the term mictomagnetism.

3.1.2 Specific heat

Another most-important thermodynamic property of the spin glasses is their

specific heat, also at high temperatures. Unfortunately, the main problem
with the specific heat is to separate out non-magnetic contributions, e.g.
phonons and electrons (in a metal). Above a few kelvin these effects
overwhelm the magnetic (Kondo) or spin-glass terms. Since the various
contributions all superimpose, elaborate procedures have been developed
3.1 High-temperature (T S Td experiments 51

to subtract the unwanted components and obtain the magnetic specific heat.
Both the measurement technique itself and the various subtraction procedures
are laborious and imprecise, thereby making the high-temperature spin-
glass specific heat a tough quantity to determine accurately. So we must
work with low T,s or dilute systems, and see how far we can increase the
temperature before the error bars become too large.
Very careful measurements and extraction of the magnetic component
do exist for a number of canonical spin glasses. Figure 3.5 shows a now
classic set of data for CuMn 0.3 at. % where Tf = 3.0 K. We shall return
to this figure a number of times in the coming sections. Suffice it for now
to note the broad maximum in C,,, above Tf (T,,, = 1.4 Tf) and the gradual
fall off at increasing temperatures T + 4 or 4.5 Tf. Too bad we cannot go
yet higher in temperature. The entire decrease with increasing T becomes
less pronounced, quite smeared, and the maximum disappears as an external
magnetic field H is applied. Qualitatively the data exhibited in Fig. 3.5 are
generic - always a broad maximum above Tf with a slow decrease at the
highest temperatures and field smearing.
In order to compare these long-range RKKY spin-glass results with those
of a short-range insulating spin glass, we display the measurements on

20.0 I I I
2790 ppm CuMn 1

0 0

0 75
I I I
0 5.0 10.0 15.0 20.0
T (K)
Fig. 3.5 Magnetic contribution of the specific heat of CuMn spin glasses with 0.3
at. % Mn plotted versus temperature in various magnetic fields (Tf = 3.0 K); from
Brodale et al. (1983).
52 Basic spin-glass phenomenon

EuO.$$,$ (Tr = 1.7 K) in Fig. 3.6. Be careful now for the data are plotted
&IT versus T in the main part and log C, versus log T in the inset. Once
again the same general type of behaviour is found (see inset): a rounded
peak at = 1.8 Tf followed by a drop at higher T. Now this decrease seems
stronger than the RKKY spin glass. The main portion of Fig. 3.6 exhibits
the field smearing as before and also indicates the magnetic entropy S,
from the areas under the various field-dependent curves. Remember that

TC
s, = “dT= cRln(2S + 1) (3.8)
I0 T
Even a rough glance at the C/T versus T plots would tell that a great deal
of the magnetic entropy is lost or frozen-out far above Tf. And the applied-
field shifts even more entropy from the low-temperature portion to the high
temperatures. We recall that there is a constant amount of entropy available
in the sample via equation (3.8), due to the fixed number of magnetic
degrees of freedom. The larger the field the more degrees of freedom lost
at a given tied T(> Tf). Crude estimates suggest that even in zero field
7O-gO% of the degrees of freedom are already forfeited above Tf. This
doesn’t leave too much entropy, only 20-30%, for a big effect at Tf.

-10
I;
t ke
05 t \
1' I '!.

L B(T)= ---- ----‘*---..:~-~>,,

I *r-Inn
-.-.. -. “‘.Z.
*..
I 0 0.05 . : +J

L T(K) 4
Fig. 3.6 Specific heat per Eu atom divided by k,T versus temperature Tin various
applied fields (units in tesla) for Eu,Sri-,S with x = 0.40. Inset shows the zero-
field behaviour in a log-log plot. The freezing temperature Tf is indicated by an
arrow. The upturn of the curves for C/T, seen at very low temperatures, is
interpreted as being caused by a Schottky term due to magnetically decoupled small
clusters of Eu spins. From Meschede et al. (1980).
3.1 High-temperature (T * Td experiments 53

But returning to our higher-temperature specific heat, what can we say

or how can we analyse its decrease? Perhaps the best, yet rough, fit to the
CuMn data is a l/T dependence. This was suggested, early on, by the
scaling analysis and by a virial expansion of thermodynamic quantities for
interacting dilute spins. However, the (EuSr)S spin glass seems to possess
a more rapid high-T fall off - the Cm(T) data unhappily do not lend
themselves to a more quantitative resolution. Nevertheless, such a result is
very reasonable. Since (EuSr)S has only two nearest-neighbour interactions
and the second is already quite weak compared to the first, temperature
should easily overcome these couplings. In other words, thermal disorder
will quickly wipe out the short-range magnetic clustering. Thus, the full
number of degrees of freedom are quickly reached and the magnetic specific
heat returns to zero. On the other hand, the canonical CuMn spin glass
has the stronger and longer-range RKKY interaction so one would expect
the short-range order or clustering to persist to much higher temperatures,
hence, the long tail.
In keeping with the above thinking we could at least try the configurational
cluster model employed to treat the high-T susceptibility of the previous
section. It should be easy, for, the specific heat is a thermodynamic quantity
derivable from the same free energy which is always calculated from the
partition function Z of the various configurations (see section 3.1.1). We
simply take temperature derivatives instead of field derivatives of the free
energy and determine C, via equation (3.7). By using the same values of
Ji, . . ., J5 found from the susceptibility we can fully calculate Cm(T) for
H = 0. The results are shown in Fig. 3.7 for CuMn (2.4 at. %). Unfortu-
nately, here Tf is too large and the experimental data are too limited and
scattered in the high-T-regime of interest. The 0.3 at. % CuMn of Fig. 3.5
is simply not concentrated enough for our model to work. At best, we can
say the trend looks right, however, nothing quantitative can be gained.
Despite all these experimental (not model) difficulties with the high-
temperature specific heat, we are drawn to the same conclusion as from
the susceptibility. Namely, magnetic-cluster formation or short-range order
is appearing far above Tf. And it is out of these clusters that the spin glass
state is formed, i.e., our ‘building blocks’. There is indeed a quantitative
difference in high-T behaviour between the long-range and short-range
interacting spin glasses. This can be understood by relating the interaction
energy as a function of distance J(r) to the thermal disorder energy kBT
and noting that the RKKY materials require a much higher temperature to
overcome their interaction energies.

3.1.3 Resistivity

We now move onto a transport property, viz., the electrical resistivity,

which is easy to measure yet hard to extract the exact spin-glass contribution
54 Basic spin-glass phenomenon

0.161 I I I I I I

T [Kl
Fig. 3.7 Open circles: measured magnetic specific heat for CuMn (2.4 at. %).
Experimental data from Wenger and Keesom (1976). Solid line: fit to the
configuration cluster model with the J,s determined from the susceptibility fit: (a)
low temperature regime; and (b) model prediction up to room temperature.

and even more arduous to calculate at high temperatures. Of course, at

very high temperature we have the constant spin-disorder scattering of the
fully paramagnetic state. Accordingly, we choose for examination a highly
conducting canonical spin glass.
Figure 3.8 shows the ‘magnetic part’, Ap, versus the temperature for a
series of CuMn alloys. This so-called magnetic part is determined from a
point-by-point subtraction of the pure-Cu resistivity. There are certain, but
hopefully small, discrepancies with this simple superposition and subtraction
of various contributions. For example, mixing or cross terms could occur
and modify the temperature dependences especially when these are rather
weak. However, note the large changes in Ap over the entire temperature
3.1 High-temperature (T >> TA experiments

100 200 30 0
T(K)

Fig. 3.8 Temperature dependence of the magnetic resistivity for spin glass CuMn:
the arrows represent Tf determined from the low-field a.c. susceptibility; from Ford
and Mydosh (1976).

range. There is always a temperature dependence, caused by the local-

spin/conduction-electron scattering.
What we seem to be detecting, at very high temperatures, is the long tail
(Fig. 3.8) of the Kondo effect which has a negative temperature coefficient
(dp/dT < 0) and Ap m ln(k,T/D), equation (2.8), extending over many
decades in temperature. This means that there are still isolated magnetic
impurities present and these can form the quasi-bound state. Yet we draw
attention to the maximum which develops and the rapid drop as the
temperature is lower towards Tf. Here the change in sign of the slope (-
to +) represents a competition between the single-impurity Kondo effect
which causes a resonant scattering and the RKKY interaction which wants
56 Basic spin-glass phenomenon

to couple the spins, and thereby reduce the spin-disorder scattering. All
this occurs far above Tf and as a precursor to it.
In order to gain some insight into the physics of the resistivity maximum
we must develop a theory of this competition, i.e., the Kondo effect in an
interacting spin system. Recall our formula (2.6), for the single impurity
Kondo temperature or energy

(3.9)

and let us introduce a mean RKKY-interaction energy

xPS(S + 1)
A RKKY = (3.10)
EF

with the same exchange J parameter. A lengthy calculation was performed

to relate T, (the temperature of the maximum in Ap) to TK and AnKkY.
The result for AnKky % TK is

T, = $ARKKy In

Thus, T, is roughly proportional to the concentration x as seen in

experiment. Note that the above theory is more sophisticated than our
intuitive idea (section 2.7) that increasing the concentration will simply
reduce TK. However, to obtain a quantitative comparison with theory, a
concentration variable, because of the above, makes everything too
complicated. A better experimental variable is pressure which in a known
manner can change J, then J(P) can be inserted into the previous equation
and T,,, is directly calculable and measurable at various pressures.
What have we learned from the many resistivity studies of canonical spin
glasses? First of all, there are the effects of magnetic correlation and
clustering far above Tf which reduce the spin-disorder scattering. But we
already knew this, for, the conclusion is perfectly consistent with the
susceptibility and specific-heat experiments. However, to describe such
cluster growth and short-range ordering is too complex for scattering theory.
Transport properties except in well-defined limits or simplified models are
usually very hard to calculate. Secondly, we have the Kondo effect present.
This is still offering a theoretical challenge, and while the complete
temperature dependence of Ap could not be obtained, at least the
temperature of the maximum in Ap and its pressure dependence was
extractable and understandable in terms of this competition Kondo versus
RKKY. Again, with the resistivity at high T, we must rely on qualitative
pictures rather than even simple theories.
3.1 High-temperature (T * TA experiments 57

3.1.4 Neutron scattering

We proceed now to a more dynamic and microscopic technique which

should open for our viewing the fine-scale (- A) spatial and short-time
(- lo-l2 s) temporal correlations in a spin glass. Neutron scattering is
completely different from the previous measurements of thermodynamic or
transport properties because it can directly probe these spin-spin correlations
on an atomic basis, and thus it should be easier to construct and verify
interpreting models.
A bit of theoretical background is needed before we discuss the spin-
glass experiments. The cross-section for neutron scattering due to a
momentum transfer between incoming and outgoing neutron q = k’ - k,
and an energy transfer ho = (h2/2mN) (k’* - kz) is

(3.12)

where N is the number of spins causing the scattering, (ye2/mNc2) is the

neutron coupling constant, and S(q, o ) is the neutron-scattering function,
which we can write, for an isotropic spin system, as

S(q, co) = ‘z/Im[ C(Si(O)Sj(t)) exp (iq * rij)] exp [- iot] dt

cc i, j
(3.13)
The pair correlation function (Si(0)Sj(t)) is for the local-moment spins
separated by a distance rij, and F(q) is the atomic-spin form factor which
is usually known. It is helpful to split up this function by adding and
subtracting (Si)(Sj). A ccordingly, we can distinguish a dynamic and a static
component.
(3.14)
where

S,(q) =
exp [iq .
Tz(Si)(Sj) rij]

131
This static part gives rise to an elastic scattering (w = 0) which may be
further divided into coherent (Bragg) at particular q-values or diffuse
(smeared over q-space) scattering contributions. If there is long-range
magnetic order, we find sharp Bragg peaks at the wave vector for the
ordering q. For a ferromagnetic q = 0 so we simply obtain Bragg peaks
commensurate with the given crystal structure. This is how the percolation
concentration for AuFe was established by detecting the appearance of
these Bragg peaks. Note they are never seen in a spin glass. If there is
short-range magnetic or chemical ordering, diffuse scattering will emerge
58 Basic spin-glass phenomenon

at different q-value or regions in the Brillouin zone which represents q-

space for the given solid. (For CuMn various scattering contours related to
both chemical clusters and magnetic short-range order were observed
throughout the Brillouin zone even at high temperatures. We discuss this
situation further below.) If there is a randomly frozen spin glass, the
Edwards-Anderson order parameter should be reflected in S,(q) = 2F2(q)
(SJ2. Unfortunately, a direct determination of the spin-glass order parameter
will be masked by several other q dependences especially those related to
the short-range order.
We now return to the dynamic or inelastic part which is

S,(q, w) = Fz irnm z [(Si(0)sj(r)> - (si>(sj>l

I ,
exp [i(q - rij-wt)] dt (3.16)
The correlation function within the square brackets can be written in terms
of a Bose-Einstein thermal factor and a generalized susceptibility x(q, w)
1
S,(q, w) = 2F20 x(9,4 (3.17)
ng2& [ 1 - exp (-ho/k, 2-jI Im
This in turn may be simplified for an isolated magnetic atom in a sea of
conduction electrons becoming

(3.18)

where x(q) is the static q-dependent susceptibility andf(w) is a q-independent

lorentzian,

f(w) = 5 (A) with r = $ (3.19)

and r is the Korringa relaxation time.

We can test these formulas through a particularly simple experiment. For
a series of CuMn (1 to 7 at. %) the total scattering cross-section has been
measured for a constant q-value at 300 K. The data are exhibited in Fig. 3.9
where the central elastic peak related to S,(q)G(w) has been omitted. At
low concentrations a lorentzian fit gives good agreement using the single
relaxation time, derived from I? (the half-width parameter), of order
T = lo-l3 s, a reasonable Korringa value. However, as the concentration
of Mn is increased there are noticeable deviations from the single-T
lorentzian fit even at this high temperature. A distribution of relaxation
times IV(T) is needed (see below) to recover the accord. So now we have
necessitated an important new parameter, the spin relaxation time T and
its distribution function N(T), which is directly obtainable from neutron-
scattering measurements. Such are critical for the characterization of our
3.1 High-temperature (T S- TJ experiments 59

Cu-7.3 al.% Mn
1

Cu-52 a1.U Mn
r=lbrOZmeV

-4
@-3.Oat.XMn
r=l.9to.2 mcV
-2

r = 1.910.2 mev

Fig. 3.9 S(q, o) versus w for q = 0.08 b;-l at 300 K for several CuMn alloys. The
central region of the elastic peak is omitted in the diagram. The continuous curves
represent best fits to a lorentzian spectral function in the energy range -2.5
meV < o < 2.5 meV. From Murani (1978a).

spin-glass freezing. Let us incorporate into our dynamical structure factor

this distribution of relaxation times g(q, I’). The S,(q, w) formula (3.18),
therefore, becomes
ho
1 - exp(-ho/knT)
(3.20)
There is an upper theoretical limit (time of passage of a neutron through
the sample) to the longest relaxation time which can be measured with
neutrons. Without energy analysis and for a typical sample size, a relaxation
60 Basic spin-glass phenomenon

process longer than -lop6 s will be seen as a static event. The situation is
even worse with energy analysis, for, now the specific instrumental energy-
resolution, AE, determines the static onset. Using the uncertainty relation,
any relaxation process with characteristic time T 2 hlAE will be indistinguish-
able from a completely static or elastic one. For today’s highest resolution
spectrometers (smallest AE), T of 10-7-10-8 s is achievable. Consequently,
the dynamic or inelastic neutron-scattering technique is a ‘fast observer’ of
the spin glasses. In order to take into account the above distinction, we
separate our dynamic structure factor into a truly dynamic part and an
‘instrument-caused’ static or elastic part. Accordingly,

Sb (q, w) = 2F20 ho
g2p$ 1 - exp (-holksT)
(3.21)
and

(3.22)
Consider first of all the static structure factor S: (q, w) of equation (3.22).
For the I (= h/T) values which most contribute to the integral, ho 4 k,T.
Then

(3.23)

(3.24)

This part of the total structure factor we call the ‘elastic’ scattering, since
due to the instrumental resolution AE, it appears as a static o = 0 process.
Note that all relaxations, which are slower or longer than T = h/AE = lop8 s,
are lumped into this term. If there are no slow relaxations, then 9s (q) + 0
except if we are near a Bragg peak or a q-region of diffuse (short-range)
scattering.
Now consider the dynamic structure factor & (q, o) in equation (3.21).
Here we use the quasi-static approximation such that ho G kBT and
integrate over all o. Thus

s7, (9) = S7, (q, W) do = ‘3 k+

B

(3.25)
3.1 High-temperature (T * TJ experiments 61

(3.26)

where we have defined a wave-vector, q, and relaxation-time, 7, dependent

susceptibility as

x(4,7) = (3.27)

The $6 (q) we obtain after integrating over all energies is called quasi-
elastic scattering. It represents all the fast fluctuations and rapid relaxation
processes in the paramagnetic state. If this state is destroyed by spin-spin
interactions, then the quasi-elastic contribution will decrease.
The key concept compelled by experiment in all this formalism is the
distribution of relaxation times N(T) being very fast and b-function like at
high temperature, and then evolving as T + Tf to longer time scales via a
broad distribution. Well, what can we measure to get a better handle on
N(T)? One usually picks a series of momentum transfers or q-values at
certain arbitrary (see below) positions in the Brillouin zone. Very low q-
values mean small-angle scattering and here one could also obtain information
on the local spatial correlations and their distributions or ranges. With q
constant we then determine the scattering cross-section by sweeping the
energy transfer w. Now the big problem is separating the elastic (o = 0 _’ AE)
central peak scattering from the surrounding quasi-elastic cross-section. By
fitting the shoulders (outside the elastic peak) to a lorentzian function, we
arrive at a distinction: quasi-elastic cross-section, namely, what is contained
in the lorentzian; and elastic cross-section, namely, the remaining sharp
central peak.
Figure 3.10 shows such a separation for the scattering cross-sections of a
AuFe (10 at. %) sample. At high temperature (150 K), the elastic part is
practically zero, yet there is a lot of quasi-elastic (or fast) scattering present.
This increases with decreasing q as greater regions of real space, and
thereby more spins become involved. Then as the temperature is reduced
the elastic scattering grows at the expense of the quasi-elastic scattering
because more slow relaxation times are fed into the time window of the
instrumental resolution. The smooth elastic increase continues down across
T,, while the fast quasi-elastic scattering slowly decreases as the number of
fast relaxation processes is eliminated.
We have been emphasizing in the present discussion spin relaxation times
and their distribution which shifts to longer times and broadens. This is a
natural consequence of our neutron scattering analysis in energy (0) or
time (t) space. However, we can also relate the physical process to cluster
formation or the growth of correlated volumes of spins. The bigger the
cluster size, the slower its relaxation time. Hence, as the fast single spins
become bound into clusters, they lose their ‘fastness’. In any case, scattering
intensity in a spin glass is always transferred from quasi-elastic to elastic as
62 Basic spin-glass phenomenon

l
ELASTIC SCATTERING Se(4

QUASI-ELASTIC SCATTERING .&(q>

#
50 100 150 20
T (K)
Fig. 3.10 The elastic scattering and the integrated quasi-elastic scattering S,(q) as
a function of temperature for three q-values for the AuFe (10 at. %) alloy. The
vertical solid line marks the temperature of the maximum in ac-susceptibility x(O)
and the dashed line the temperature of the maximum in x(q) from the neutron
results. From Murani (19786).

these clusters are formed and slowed down. The neutron scattering sees
them directly. And as we traverse Tf in the next section we can even arrive
at some more quantitative results,
Similar neutron measurements have been performed for CuMn
(5-25 at. %), but now the total (elastic plus quasi-elastic) scattering cross-
section was determined at different temperatures as a function of momentum
transfer q. Here the whole of q-space, i.e., the Brillouin zone, was scanned
for both magnetic and nuclear (non-magnetic) scattering. This refinement
was made possible by neutron-polarization analysis for which the magnetic
scattering occurred with a spin-flip of the incoming polarized neutron and
vice versa for the non-magnetic scattering. By examining the scattering
contours one can distinguish regions of chemical (non-magnetic) clustering
3.2 Experiments spanning Tf 63

from regions of magnetic clustering. Figure 3.11 shows a 1D q-scan along

the (lo{)-direction for the magnetic cross-section of a rather concentrated
(25 at. %) CuMn alloy. There is a lot of magnetic scattering smeared out
over the Brillouin zone even at room temperature. A model calculation
confirms this to be ferromagnetic short-range order which continues to exist
at temperatures above 400 K. As the temperature is reduced two peaks at
two distinct q-values emerge from the diffuse-scattering plateau. Something
is certainly happening as Tf is crossed, but we must save this for a future
section. At this point in our journey through the high-temperature regime
of the spin glasses, we should content ourselves with the formation of these
mainly ferromagnetic clusters. Their existence in either time or space has
been established consistently from all four experiments considered in this
section. Next we must study the behaviour of spin glasses and these clusters
at the freezing temperature.

3.2 EXPERIMENTS SPANNING Tf

Here is where the action is! The possibility of a sharp freezing transition
to a unique phase must be investigated through fine experimentation on
the canonical spin glasses. While many different techniques have been

1.0 I I 1 I ’ I
II) (b) r-4
0.0
0.0
9 0.6
$ 0.6
:- 0.4
$j 0.4

0.2

0- -
0 0.2 0.4 0.6 OB

0.6
4
s
F 0.4
‘2

Fig. 3.11 Magnetic cross-section for CuMn (25 at. %) along (105) at elevated
temperatures; from Cable et al. (1984).
64 Basic spin-glass phenomenon

employed to study the ‘phase transition’, only a few come to grips with the
dynamics (wide range of measurement times) and zero (or small) external-
field requirement. For, as we shall soon see, dynamics plays a most essential
role and a field of only a few hundred gauss wipe out the sharpness of the
transition. We made a choice of four experimental methods in the last
section, so let us now carry on with these. Our dc-susceptibility becomes
mainly ac-susceptibility and its frequency dependence is the key variable.
Specific heat and resistivity are extended to span Tf and certain types of
critical analysis are performed. We view different relaxation-time extractions
and interpretations of the neutron-scattering technique including the very
novel spin-echo method. And to gain some experimental variation on a
fine scale near T,, we introduce, at the risk of increased complexity, two
‘local-probes’ experiments: muon spin relaxation (l&R) which tests the
inter-atomic molecular fields both spatially and temporally, and the
Miissbauer effect which is a local nuclear technique for measuring the
hyperfine-field reflecting the magnetism of the electronic system.

3.2.1 Ac-susceptibility

There are many methods to determine the magnetic susceptibility x. Usually

a static magnetization M is measured and divided by the applied field H.
However, when M is no longer directly proportional to x, then this method
fails. A different technique is to apply a small ac-field, called the driving
field h, and measure x by taking the derivative dM/dh at some frequency
w and even with a dc-H applied. This ac susceptibility is especially important
for spin glasses, since h can be made very small = 0.1 gauss and w can be
varied over a rather large frequency range, thereby permitting a full
determination of the real x’(w) and imaginary x”(w) parts.
It was in the early 1970s that sharp cusps were discovered in the ac-
susceptibility of AuFe and CuMn. Both of these canonical spin glasses were
studied for many years previous but only in large magnetic field. Since a
few hundred gauss of applied field smears out the cusp to a broad maximum
(see below), this is why the spin glasses, as an excitingly new and possibly
unique phase transition, were not identified a decade or two sooner.
Figure 3.12 reproduces the first measurements on AuFe which combined
the term spin-glass with the sharp cusps. The frequency was the low audio
range (50-155 Hz) and the driving field 5 gauss. Notice how the peak
increases in magnitude and shifts to higher temperatures as the concentration
of Fe is increased. Such distinct temperature effects were totally unexpected
from early speculation based upon a random molecular-field model and the
previously measured high-field susceptibilities. The height of the peak is
rather small, orders of magnitude below the typical ‘step height’ or
demagnetization cut-off of a ferromagnetic transition. It is more or less
comparable to that found in many long-range antiferromagnetic material,
3.2 Experiments spanning Tf 65

T PK)

Fig. 3.12 Low-field susceptibility x(T) of AuFe for 1 I x 5 8 at. %. The data
were taken every t K in the region of the peak, and every 4 or 1 K elsewhere. The
scatter is of the order of the thickness of the lines. The open circles indicate isolated
points taken at higher temperatures. From Cannella and Mydosh (1972).

even through the concentration of magnetic moments for a spin glass is

much less. The tail in Fig. 3.12, persisting to high temperatures, overlaps
with the dc-susceptibility in larger fields which was discussed in the previous
section. It may be fit with the same modified Curie-Weiss analysis as
before. However, the ac-susceptibility rapidly loses its sensitivity at high
temperature due to experimental reasons, and so, no longer has the desired
accuracy. On the low temperature side
x(T) = x(O) + bT” with m = 2 (3.28)
Thus x extrapolates to a finite value for T + 0 and a ratio
x(0)/x( Tf) = 0.5-0.6 is roughly obtained for magnetic concentrations between
0.1 and a few atomic percent. These measurements have been extended to
a great variety of spin glasses which show the same general characteristics.
Figure 3.13 shows how an applied dc-field rounds off the peak. Already
in 200-300 gauss the peak is smeared away and only a broad maximum
remains. The reader can imagine what will happen in the few thousand
66 Basic spin-glass phenomenon

I 0 0 0 100 G I\

.61 ’ I I . .,. I
6 16 24 32 40
T PK)

Fig. 3.13 Susceptibility data for AuFe with x = 5 and 8 at. %, showing the curves
for zero field, and for various applied fields; from Cannella and Mydosh (1972).

gauss fields typically employed in static magnetization measurements. The

field effect is surprising, for, we have a Tf of = 25 K being strongly affected
by a field of = 1000 gauss, although k,T, >> Feff H. (Assuming ~~~ = lot+,
then 0.1 T corresponds to a temperature of = 0.67 K.) This means the
spin-glass state has a peculiar sensitivity even to a small external field.
Certainly from looking at the sharpness of the ac-susceptibility cusp we
could exclaim ‘phase transition’ and immediately delve into the statistical
mechanics of critical phenomenon. But, be careful, and let us blow up the
X-T region around the ac-cusp and vary the frequency w - something which
took about ten years to accomplish with sufficient accuracy. Figure 3.14
presents one such set of data on CuMn. Over the wide-temperature range
the cusp is clearly seen, however, on the fine scale of the inset there is a
slight rounding of the cusp into a peak. Yet more important the peak is
shifted downwards in temperature with decreasing frequency of measurement.
For the frequency variation of 2; decades in Fig. 3.14, Tf is reduced by
about 1%. Other similar metallic spin-glasses have roughly the same or a
slightly larger frequency shift, a few per cent. For the insulating spin glasses
the frequency dependence is even larger (see below). A quantitative measure
of the frequency shift is obtained from (ATfIT,) per decade W. Table 3.1
shows this quantity for a number of spin glasses and also for a known
superparamagnet a-(Ho,O,)(B,O,). Notice the difference in magnitude and
how really little the canonical spin glasses are affected by frequency. This
frequency shift in Tf offers a good criterion for distinguishing a canonical
spin glass from a spin-glass-like material from a superparamagnet. A similar
frequency dependence of Tc (Curie temperature) for a ferromagnet or TN
3.2 Experiments spanning Tf 67

Fig. 3.14 Zero-field susceptibility x’ as a function of temperature for sample IIc

(CuMn (0.94 at. %) powder), measuring frequencies: Cl, 1.33 kHz; 0, 234 Hz; X,
10.4 Hz; and A, 2.6 Hz. From Mulder et al. (1981).

Table 3.1 Frequency shifts of Tf (defined as the

maximum in the ac-susceptibility, x’) for various
spin glasses

System ATd[T&log m)l

CuMn 0.005
AuMn 0.0045
AgMn 0.006
P&n 0.013
NiMn 0.018
AuFe 0.010
(LaGd)Al, 0.06
(EuSr)S 0.06
W4W2 0.08
a-Co0.A120$i02 0.06
a-WOK%) (B203) 0.28

(NCel temperature) for an antiferromagnet has not been observed at such

low w-values. Here one needs very high frequencies (mega to giga Hz) to
find T,(o) or TN(o) dependencies in these long-range ordered materials.
We shall return below to the significance and interpretation of this frequency
dependence.
Since we are using an ac technique, the susceptibility, as mentioned
above, will have two components a real part x’(o), the dispersion, and an
68 Basic spin-glass phenomenon

imaginary part x”(o), the absorption. They are usually related through a
relaxation time by a set of equations called the Casimir-du PrC equations
which we shall discuss later. For a spin glass there is a sudden onset of the
imaginary component near T,. We depict this behaviour in Fig. 3.15 for
(EuSr)S, an insulating spin glass. Since x” is so small it is difficult to
measure accurately and in a metal it will be distorted because of eddy
currents induced by the alternating field h. This limits x”-measurements to
about a thousand hertz even on powdered conducting samples. As with x’
the frequency shift of x” is downwards in temperature with decreasing
frequency. Now, however, the peak in x’(T) corresponds to the maximum
slope in x”(T), (dx”/dT),,,. The appearance of an imaginary component
means relaxation processes are affecting the measurement and by decoupling
the spins from the lattice they cause the absorption. Such effects are usually
not found at magnetic transitions except when hysteresis arises as in a
ferromagnet. So with an analysis of the complex ac susceptibility we have
a handle on the low-frequency or intermediate (1O-1-1O-5 s) relaxation

Temperature (K)

Fig. 3.15 Temperature dependence of the dispersion x’ (solid symbols) and

absorption x” (open symbols) for Eu,,.zSrO.sS:00, 10.9 Hz; Kl, 261 Hz; AA,
1969 Hz (applied ac field h = 0.1 Oe); from Hiiser et al. (1983).
3.2 Experiments spanning Tf 69

times. Remember that the neutron scattering tells us about the fast relaxation
processes (10e8 s and shorter).
Finally, with the newly developed SQUID techniques a highly sensitive
method exists for determining the magnetization in a very small applied
field. As H + 0 there must be a proportionality Xdc = M/H, and a quasi-
static (governed by the time of the measurement at a given T, e.g. = 100 s)
delineation of the susceptibility is gained. It is remarkable how many
decades of time or frequency may be probed using the various experimental
techniques, for, with the SQUID we have times available which can
approach lo5 s. There are two distinct ways to measure the susceptibility
with a small dc field. The first is to apply the field far above Tf and cool
the sample in a field to T 6 Tf all the while recording the magnetization.
This method we call field cooling (FC). Secondly, we can cool the sample
in zero field to T 4 Tf and at this low temperature apply the field. Then
we heat the sample while measuring M to T P Tf with the field constant.
Here we use the term zero-field cooling (ZFC). Figure 3.16 illustrates the
different temperature dependences (FC versus ZFC) for CuMn (1 and
2 at. %) with a field of 6 gauss. Notice what happens at Tf. The FC-x
becomes constant in value and to a great extent independent of time: if we
stop and wait at a given T(< T,), x remains unchanged. (Only at extreme
sensitivities and very long waiting times is there a drift in x.) Furthermore,
the process FC followed by field warming is reversible. If we cycle the
temperature back and forth (with H = const.) the FC-susceptibility traces

I ’ I I I I I

I-

I I 1 I I
O-3 IO 15 20 25
T (K)

Fig. 3.16 Field cooled [(a), (c)] and zero-field cooled [(b), (d)] magnetizations (x
= Ml6 gauss) for CuMn (1 and 2 at. %) as a function of temperature; from Nagata
et al. (1979).
70 Basic spin-glass phenomenon

the same path (see double-tipped arrows in Fig. 3.16). On the other hand,
the ZFC susceptibility is zero until the field is applied, in Fig. 3.16 at
= 5 K. Then with H on and T constant, the ZFC-x jumps to a value
comparable with that found from xac. However, now there is a slow, clear
drift upwards which continues over many decades in times. If we wait long
enough (and no one has yet achieved this - the average lifetime of a
graduate student is only about lo8 s), xzFc (t -+ m) = XFC at the given
T < Tf. So, when we proceed to increase the temperature at some reasonable
rate dT/dt, we arrive at the single-arrow curves in Fig. 3.16. These traces
are irreversible, XzFc is always drifting upwards. Therefore, the actual curve
or T-dependence is a function of dT/dt. Such does not occur with xac - it
is fully reversible. If we continue to increase the temperature above Tf and
then cool the sample backdown, there is first the reversible ‘paramagnetic’
regime above Tf and then the susceptibility will become the reversible Xrc
(double-tipped arrows) with subsequent cooling. In short, the dc field (albeit
small), if applied below Tf creates a metastable, irreversible state, and Tf
is nicely defined by the onset of these irreversibilities, i.e., difference
between FC and ZFC curves. More about the low-temperature effects in
the next section.
How can we better analyse and interpret the data near T,? Let us begin
with the frequency shift of Tf. First of all we could try the Arrhenius law
for thermal activation already used for a superparamagnet
T = T, exp [E&u T] (3.29)
which we can rewrite as
w = w, exp [ - E,/k, Tf] (3.30)
Here w is the driving frequency of our xa,-measurement and Tf its peak.
By plotting In (w/w,) versus l/Tf, we can from the slope and value of the
logarithm determine E, and w,. For the canonical spin glasses using the
data in Fig. 3.14, best-fitting to the Arrhenius law gives completely unphysical
values for w, and E,, e.g. o, = 1020° Hz and E, = 4400 K. This nonsense
is clearly due to the very small changes in Tf (maximum in xac), and it
distinguishes a spin glass from a superparamagnet where the Arrhenius law
does indeed hold. There is much more than simple energy-barrier blocking
and thermal activation in the transition of a spin glass.
A second method of analysis is ferreted out from the literature on real
glasses. It is the empirical law which describes the viscosity of supercooled
liquids, namely, the Vogel-Fulcher law. Written for use in describing T,
shifts this becomes
w = w, exp [ - E,lk,( Tf - TO)] (3.31)
where T,, is a new parameter (for real glasses it is called the ‘ideal glass’
temperature). With three fitting parameters (wO, E, and TO) the agreement
is naturally much better, except perhaps at low frequencies, using a
3.2 Experiments spanning Tf 71

more physical set of parameter values. Typically for CuMn (4.6 at. %)
WI = 1.6 x lo8 Hz, E, = 11.8 K and rO(< T, = 27.5 K) = 26.9 K. The
problem here is the precise physical meaning of the large T,,. Some attempts
have been made to relate it to the interaction strengths between the clusters
in a spin glass. Since a spin glass is not a non-interacting collection of
clusters something must be added to take into account the inter-cluster
couplings. To might also be related to the true critical temperature of an
underlying phase transition for which Tf is only a dynamic manifestation.
Thus far, a deeper understanding of To and its connection to the longer-
range interactions is lacking.
Finally, a third approach draws on the standard theory for dynamical
scaling near a phase transition at T,. The conventional result of dynamical
scaling relates the critical relaxation time r to the correlation length 5 as
r - 5’. Since 5 diverges with temperature as 5 - [T/(T - T=)]“, we have
the power-law divergence

(3.32)

where the product exponent zv is called the dynamical exponent. In the

language of spin glasses we can write

7AV=To (3.33)

where Tf is the frequency-dependent freezing temperature determined by

the maximum in xac. Once again with three parameters (r,, T, and zv) an
experimental extraction is attempted and now reasonable, and perhaps not
unphysical, values are obtained. As a typical example CuMn (4.6 at. %)
has T, = 10-l* s, zv = 5.5 and T, = 27.5 K which should correspond to
the dc (equilibrium) value of Tf (o + 0). The trouble here is that zv varies
between about 4 and 12 for the different spin glasses. For conventional
phase transitions, zv usually is around 2. This latter (power-law) approach
seems to be the acceptable one with which to compare to Monte Carlo
simulations of spin glasses - more of this when we get to the theory.
For the sake of completeness we include another relation for r based on
the dynamical-scaling exponent v. This expression was proposed for the
limit of a zero temperature [T, = 0 or T,(o + 0) = 0] phase transition

T = T* exp [E~lkBT)*“] (3.34)

Assuming the critical temperature is zero the Tf (maximum in xac) represents
a relaxational effect not related to a phase transition at finite temperature.
An attempt was made by fitting some real spin glasses to obtain the three
parameters (r*, E, and zv), but the results were totally meaningless and a
T, = 0 phase transition for a 3D spin glass can be eliminated.
We now consider the real and imaginary susceptibilities and their frequency
72 Basic spin-glass phenomenon

and temperature dependences. Let us start with the Casimirdu PrC

equations
XT - &
x’ = xs + 1 + &2 (3.35)

and

x” = WT (3.36)

where -r is a single relaxation time, XT is the isothermal susceptibility in the

limit o + 0, and xs is the adiabatic susceptibility in the limit o -+ m. If
x” = 0, then there are two possibilities: either wr + 0 and we measure an
equilibrium isothermal XT, or o-r + m and we measure the, totally decoupled
from the lattice, nonequilibrium adiabatic xs. At w = l/r the dispersion
x’(o) will have an inflection point, whereas the absorption X”(O) will show
a maximum. Thus, this maximum provides a method for determining an
average relaxation time r,, for each temperature. Also according to equation
(3.36), the absorption x” should follow a sech (ln 07) functional dependence
for single T and be considerably broadened if a distribution of relaxation
times g(~) is present. Figure 3.17 exemplifies this behaviour for the
(Eu,,,Sr,.,)S sample shown previously in Fig. 3.15. No clear maximum in
x” is observable over the frequency range investigated. Nevertheless, the
absorption at the lowest frequencies dramatically increases in the temperature
interval from 700 to 600 mK which spans 7’,(10 Hz) = 640 mK. For
T < 600 mK, the absorption is essentially flat with all isotherms remaining

10 lo2 10’
Frequency (Hz)

Fig. 3.17 Absorption x” as a function of frequency for different temperatures in

(EuO.* SR&S. The solid lines are visual guides; from Hfiser et al. (1983).
3.2 Experiments spanning Tf 73

parallel to one another. In order to illuminate further this relaxation-time

distribution, the susceptibility data can be plotted in the complex plane as
x’ versus x”, where each frequency represents a point. These so-called
Argzind diagrams are illustrated in Fig. 3.18 for the same (EuSr)S sample.
Note now that the expected half-circle for a single 7, which is given by the
maximum where OT = 1, is squashed into an arc length whose flatness
signifies a large distribution of relaxation times. Hence, the deviations from
a full half-circle are an estimate for the width of the distribution function.
A mathematical method exists to treat these deviations, if the frequency
range of measurements can be made sufficiently large. Such analysis we
shall reserve until our chapter on ideal spin glasses.
As for the present situation we must modify the Casimir-du PrC equations
to account for this broad distribution of relaxation times existing between
T,in and T,,. Consequently, equations (3.35) and (3.36) become

QiTk(T) W 7) (3.37)

and

(3.38)

0.2

X’ (emu/mot Eu)

Fig. 3.18 Argand diagrams for three different temperatures. The frequency of the
ac driving field h increases from right (10.9 Hz) to left (1969 Hz). The lines are
computer fits to the data points assuming a symmetric diagram (see text). From
Hiiser et al. (1983).
 74 Basic spin-glass phenomenon

where m,(7) is the time variation of cluster moment m, which responds to a

driving field h and which becomes blocked at its individual blocking temperature
TB(r). We remark in passing that there is a simple relationship between x’
and x” for frequencies o in the middle of a flat distribution, namely
T -z--
axf = m,(T,)
Xuz-- 2alno 2 ~ h d4
where T, = l/w is this middle relaxation time.
While the above x’ and X” equations work reasonably well in extracting
g(7) for superparamagnets and even for some insulating spin-glass-like
materials (e.g. cobalt aluminosilicate glass) and a temperature evolution of
g(~, T) may be calculated, they do not do as well for the canonical (RKKY)
spin glasses. These latter systems seem to possess such a wide range of
relaxation times lo-l4 s to CQ,so that experiments cannot deliver the many
decades of frequency for x’ and x” which are necessary to determine g(T,
T). A combination of various techniques is needed. Additionally, there
are indications here that g(r, T) exhibits a sudden shift to longer times as
T is reduced to and crosses Tf. Does this mean a phase transition? In any
case, for a metallic spin glass we can schematically sketch the distribution
function and its evolution with temperature in Fig. 3.19. The true paramag-
netic (noninteracting isolated spins) regime (T 2 10 Tf) is distinguished by
a nearly single Korringa relaxation time at =10-12 s. As T is lowered, the
spatial correlations begin to form and this perturbs the distribution function.

[SJ / /,/ / / / / /‘XT; .

: ‘., :, ,: ‘, - ; ( ( .; .:, 1. ..,..: : : .. ., ., .‘.,’ : : : _.,. ,. ,. ,. ,.
j$j / .:: ‘.. ..,. ..‘....‘.... . . . . . . . . . .._
I I I / I / I

T(K)
J lo-'2 10-'010-8 10-6 10-4 1o-2 loo lo2 log r (9

Fig. 3.19 Schematic representation of the probability distribution for spin relaxation
times with its evolution as a function of temperature.
3.2 Experiments spanning Tf 75

Then, as T + Tf there is a rapid shift of g(T) to slower time scales due to

the long-range competing interactions. Finally, at Tf a very long-time tail
suddenly appears and for all practical purposes the system is frozen or
‘transformed’ on a laboratory time scale. The static times continue to grow
below T,, but what do they mean if we cannot measure them? Remember
the lifetime of a graduate student.
Before completing this subsection on susceptibility a final word must be
said about the non-linear susceptibility xnl. This is an important new variable
which according to theory exhibits the critical susceptibility divergence and
exponent of a spin glass. Such a quantity may be defined in two different
ways:

(9 Xnl = 1 - JC (3.40)
x3
where x0 is the linear susceptibility M/H in the limit of H + 0. Since
we can expand M in odd powers of the field
M = xoH - u&,H)~ + a&, H)5 - . . . (3.41)
where the as are a function of temperature. Then
xnl = +a3(xoW2 - dxoH>” + * * * (3.42)
Hopefully, our power series rapidly converges so we are left with
only one or two leading terms.
(ii) Alternatively, by applying an ac driving field, h, at frequency w, we
perform a similar expansion, but now as a function of odd frequency
harmonics 30, 5w, . . . .

M(w) = c [Oh cos kwt + 0; sin kwt] (3.43)

k=odd

where
@;=x;h+%X;h3+iX;h5+. . .

0;=;X;h3+;x;hS+.. .
1 X;h5 + . . .
0; = 16 (3.46)

and similarly for the imaginary part 0”. If we assume that the driving
field is very small (h G 1 in our expansion), then
M(w) = xih cos WC+ x’;h sin wt +

+ t X$h3 cos 3wt + i xL;h3sin 3wt (3.47)

+ & X&h5cos 5wt + & $h5 sin 5wt

76 Basic spin-glass phenomenon

Thus, we can measure the various coefficients of the different harmonics.

The higher order or non-linear ones (3w, 50, etc.) through scaling theory
should reflect the critical divergences characterized by a set of critical
exponents (y and p). Neglecting the imaginary part, the non-linear
susceptibility becomes

(3.48)

where the Xs are the amplitudes of the divergences.

Highly sophisticated experimental techniques, e.g. fast Fourier transforms,
have been brought to bear on the determination of x. However, difficulties
exist with each definition (i) and (ii) of xnl. For the field expansion (i), the
rather large applied fields will definitely affect the spin-glass transition
smearing out the critical phenomenon. For the harmonic expansion (ii), the
relaxation times will soon become larger than w-l and thereby influence or
broaden the critical behaviour. For this reason only in the limit w + 0
should the static transition be accessible. Nevertheless, xnl, and nof xac, is
crucial to establish the properties of a phase transition in spin glasses. By
way of illustrating the typical behaviour of x,,i, we reproduce the data at
w = 10e2 Hz for an AgMn (0.5 at. %) sample in Fig. 3.20. The plot of xi
versus reduced temperature, (T - 7’&/Tg is made on a log-log scale in
order for the slope to give the (negative) critical exponent. Here Tg (=
2.95 K) is the critical temperature fixed as the peak in xi with w = 0.001 Hz.
The results do show what looks like a power law behaviour with y = 2.1
for at least a decade of reduced temperature. But take note of the rounding
beginning below 2 x lo-* in reduced temperature. This means we cannot
approach too close to Tg before the transition becomes smeared out, most
probably due to these relaxation time effect, even a 100 s measurement
time is not long enough at 2 x lo-* Tg. It would seem that the spin glass
begins by following a critical divergence. However, the distribution of
relaxation times becomes too great and the system drifts out of equilibrium.
For, in usual phase transitions, these divergences can be tracked much
closer to T,, say 10e3 or even lo- ‘. Well then, the spin glasses certainly
do not display an ordinary phase transition.

3.2.2 Specific heat

After all this action around Tf for the susceptibility, the specific heat is a
big disappointment. Already from Fig. 3.5 there is little to be seen at Tf.
As a matter of fact without further analysis Tf does not even seem to exist
in a C, versus T plot.
What can we do further to reanalyse or blow-up the specific-heat
behaviour? One suggestion for extracting Tf was to make plots for the most
3.2 Experiments spanning Tf 77

1 , I I I
10-2 lo-'
(T-Tg)/Tg

Fig. 3.20 Temperature dependence of --x3 above Tg (= T,) measured at lo-* Hz

in static fields of 0 (open circles) and 90 G (solid circles) as a function of reduced
temperature. The same slope (-y) indicates that the divergence of x3 in dc fields
less than 100 G remains well described by the same exponent (y = 2.3) as in zero
field. From Levy (1988).

accurate specific-heat data of C,/T versus T. The hope was to establish a

break or a rapid fall in the rate of change of entropy dS,/dT = C,,,/T at
Tf. By examining the meticulous results shown in Fig. 3.21 for three samples
of CuMn (0.08, 0.43 and 0.88 at. %) we conclude the following: while there
is some indication for a drop, it is certainly not sharp enough to relate to
a distinct and well-defined Tf. Consequently, on the basis of the best-
available data, no peak or singularity can be found in C,, in contrast to
the cusps and incipient divergences in xac and xn,, respectively.
We still need a reason for the absence of a clear effect in C, at Tf.
Perhaps the critical phenomenon is very subtle. A large negative critical
exponent, 01, for the specific heat would produce a broad maximum centred
at T,. And, if this is superimposed on a flat background of a non-critical
specific-heat contribution,it could be quite difficult to discern the critical
part. Perhaps the criticalregion for C, is so narrow that the finite
temperature steps AT, which are necessary for the specific-heat measurement
(remember experimentally C = AQ/AT), are enough to wipe out the critical
fine-peak structure. What we learned from our high-T specific-heat analysis
78 Basic spin-glass phenomenon

a
.Q*
l ..*
.
.- l ..v.,, l

4 *
c
‘;
0.0 ‘.‘.I. “.I. ‘.“..“I. .“““’
0 IO 20 30

TEYF’ERATlRE
(K1
Fig. 3.21 Spin-glass specific heat divided by temperature plotted against tempera-
ture. The vertical bars intersecting the plots mark the estimate of Tf for each
composition. From Martin (1980).

(section 3.1.2) was that most of the entropy (- 80%) is already frozen out
in short-range order above Tf. So conceivably there is simply not enough
entropy remaining to produce a significant peak at Tf. Recall how our
cluster configuration model nicely spanned the C, data through Tf (Fig. 3.7)
whereas it could not deal with the susceptibility peak.
Such conjectures are trying to preserve a phase-transition explanation for
C,. However, we must live with the empirical fact that little happens in
the specific heat. Either the phase transition is unconventional or maybe
we should mimic the real glasses and their freezing transitions (fall out of
equilibrium) as a viable model for the spin glasses. Before we choose let’s
look at some other experiments.

3.2.3 Resistivity

For usual phase transitions the critical behaviour of the electrical resistivity
is reflected in its temperature coefficient or derivative d(Ap)/dT. Theory
has even related this derivative to the specific heat and its critical divergence
for certain metallic cases. So it is not surprising that little can be found in
d(Ap)/dT at Tf. Figure 3.22 illustrates the point by plotting the T-coefficient
d(Ap)ldT for some CuMn alloys as a function of T. The arrows represent
3.2 Experiments spanning Tf 79

1 I I I I I I I I I
I n -.---.- Cu+S.'lot%Mn :
-----Cu+2.7ot%Mn ------ Cu+63ot%Mn '
-----Cu+l.Got%Mn - Cu+CSat%Mn 1
- Cu+O.7ot%Mn I

I \ I ---__ I I I II I

lOi w 40 50 10 20 30 40
! T(K) T(K?"
Fig. 3.22 Temperature dependence of the temperature derivative of the impurity
resistivity d(Ap)dT (rL? cm/K) for the CuMn alloys. The arrows represent Tf as
determined from the susceptibility measurements. From Ford and Mydosh (1976).

temperatures of the sharp cusps observed in xac. There is no correlation

with Tf only a broad maximum at a temperature significantly below Tf.
Once again we have the same smeared behaviour as C,. The transport
properties, really a scattering of conduction electrons from magnetic
impurities, exhibit no transition, only a gradual transformation. For,
certainly, the spin-disorder scattering is steadily reduced upon lowering the
temperature through Tf. A comparable situation occurs for the thermoelectric
power which is more complicated due to a variety of contributions - nothing
striking happens at Tf. Why is there no effect? Well, one suggestion is that
it is to do with the lack of a periodically ordered state, and thus, no long-
range spatial correlation function. The frozen state is random and frozen
disorder will still make a significant contribution to the scattering. Unques-
tionably the time correlations are progressively becoming slower, but even
a rapid change to longer time scales near Tf is not affecting the much
shorter local-moment/conduction-electron scattering times which determine
the transport properties. There are always some fast relaxation times
available in our g(r) distribution which will continue to scatter the conduction
electrons. It seems exclusively when there is the formation of a periodic
spin lattice that the transport properties will reflect the critical behaviour.
This periodicity may even create distortions of the Fermi surface, which,
of course, will be seen in dApldT and the thermopower. So our job of
80 Basic spin-glass phenomenon

experimentally characterizing the freezing process Tr is a most difficult task

without a periodic spin order. The intrinsic randomness of the freezing
masks many of the key experimental effects which might be related to
critical phenomenon, and we certainly have an unusual type of phase
transition.

32.4 Neutron scattering

Let us return to our structure-factor separation into static and dynamic

parts (section 3.1.4). We recall that the total scattering cross-section is
proportional to the sum of these two parts which, within our relaxation
time approximation, we can write as equations (3.24) and (3.26). The limits
of the integrals depend upon the instrumental resolution of our neutron
spectrometer. So we just repeat the measurements of the scattering cross-
section at a fixed value of momentum transfer 4 on as many different
spectrometers as are available. (Fig. 3.10 was limited to only one
spectrometer.) Easier said than done! Another hindrance is the exact
separation of elastic from quasi-elastic. But this is a fitting ‘detail’
disentangled by the experts.
For an arbitrary value of 4 = 0.13 A-l and a CuMn (3 at. %) sample
we show the results in Fig. 3.23 for two elastic scattering cross-sections
measured with two different spectrometer resolutions, 1.5 PeV (- 10e9 s)
and 230 PeV (- lo-l1 s). This means that only relaxation times longer than
lop9 s (or lo-l1 s) will be seen in the elastic window. Notice how the total
cross-section given by the squares in Fig. 3.23 remains relatively constant
over the entire temperature range. Below 75 K the fast scattering rate is
being shifted to slow scattering and this appears as ‘elastic’ intensity - it
contributes to a continuously increasing central peak. Tsg is the freezing
temperature (TJ determined via xac. Here the ‘a’ length or amount of
scattering cross-section represents the very fast quasi-elastic range
(< lo-l1 s), the ‘b’ amount is the scattering with relaxation times in the
interval between the two spectrometers (10-11-10-9 s), and ‘c’ is the slow
scattering (> low9 s) all lumped together. The next step is to construct a
schematic form for the distribution of relaxation times N(r). Figure 3.24
presents a sketch of N(log T) versus log T as expected from the data in
Fig. 3.23. Note the sudden and asymmetric shift of the distribution function
to much longer times at Tf. Surprising even for me is the remarkable
similarity between this distribution function and that independently obtained
from the x’(o) and x”(o) results (Fig. 3.19). While both are qualitative,
merely schematic, the same basic idea is at hand, namely the sudden shift
in the distribution of relaxation times to much longer values at Tf.
We could continue with our neutron scattering analysis and examine the
q-dependence of the scattering as a function of T. A bit of this was already
introduced in Fig. 3.11. But let us postpone this discussion until the low
3.2 Experiments spanning Tf 81

Cu-3at % Mn
i=
i= q = 0.13 2

I II ,, II 1
0 2s 50 75 100
T(K)
Fig. 3.23 The ‘elastic’ cross-sectionsmeasured with two different energy resolutions
along with the quasistatic part of the total cross-section (as defined in the text) for
the CuMn (3 at. %) alloy at q = 0.13 A-l. The dots, 0, represent the ‘elastic’
cross-section measured with energy resolution of 1.5 PeV. The triangles, A,
represent the elastic cross-section measured with energy resolution of 230 FeV. The
quasistatic part of the total cross-section is represented by the squares, W. From
Murani (1981).

Fig. 3.24 The possible forms of the density of relaxation times N(T) per unit
logarithmic interval of T, for the CuMn (3 at. %) alloy at (i) T * Tsg(=T,) and
(ii) T = Tsg. The regions marked a, b and c correspond to the lengths a, b and c
in Fig. 3.23 representing the spectral weights in the various time zones. From
Murani (1981).
82 Basic spin-glass phenomenon

temperature (T + Tf) section (section 3.3.4), and instead, briefly consider

a novel neutron scattering technique which directly measures the time-
correlation function at some average small q-value for longer time scales.
The method is called neutron spin-echo and essentially it enables us to
detect small inelastic scattering components by comparing the Larmor-
precession phase of a spin-polarized neutron beam before and after scattering
with a sample. Since this is essentially a resonance process, for which slight
deviations in the polarization phase are measurable, it can extend the time
detection range down to lop8 s. The formula for the average polarization
along a given axis is

(3.49)

where 0 is the phase angle. According to our original definition of

S,(q, 0) in equation (3.13), (P) is related to the time correlation function
which for a spin glass is (P) = c(q, t) a (&(O)&(t)). In the limit of q + 0
and F2(q 3 0) = 1, and substituting for S,(q, 0) via equation (3.25), we
have

~0) = (3.50)

exp (- II) dr

At t = 0 the last expression reduces to

5(O)
=‘s Ij(r) dr (3.52)

which is nothing but our old friend ST or equation (3.26) in the limits
q -+ 0 and T + QJ.If g(I) is a slowly varying function, then we can define
a correlation function at relaxation time 7 as

5~~) = ‘g :hg(r) dr = s, (4 = 0)
I
See equation (3.24). Therefore,
I;(O) - e(T) = S&q = 0) - SZ (4 = 0) = % (4 = 0) (3.54)
where

s, (q = 0) = $$f x(q = 097) a x,,(w)

So we can collect these results and plot the measured I& against log-
3.2 Experiments spanning Tf 83

time. Figure 3.25 shows how the neutron-spin-echo determined correlation

function (at (4) = 0.09 A-l) decreases at various temperatures for times
= 1012-10-8 s. Again the alloy is CuMn (5 at. %) with T, = 27.4 K. Clearly
one can discern the temperature evolution from exponential decrease
exp(-Tt) (I? = 0.5 meV) via power-law fall off to approximately log(t).
However, the neutron technique is not suited for really long time scales
> 1O-8 s. So what we have also included in Fig. 3.25 are the t(t) data
(i) for pSR (muon spin relaxation), discussed in the next section (3.2.5),
which tracks the intermediate times = lo-’ s, and (ii) from the x==(o)
according to equation (3.55). Using equation (3.54) we can write

1 - t(7) = s, (4 = 0) = 2$$ x(7) (3.56)

where we have assumed a phase transition and t(O) = 1. Note how the
correlation function jumps up at Tf and develops its long-time tail (right-
hand side of Fig. 3.25). Once more there is the same type of experimental
behaviour and phenomenological model. At least we have an empirical
picture of the freezing process that is unlike that of the standard phase
transitions which exhibit a rapid critical slowing down and disappearance
of the fluctuations, a small critical regime where all this happens, and a
becoming infinite correlation length (long-range order). Using these various
techniques we have traversed almost ten decades of time - something
necessary for the spin glasses. Yet needed are probes of the really long
times which characterize the frozen state (T + Tf). Section 3.3 will handle
this situation, as for now we persevere with two further experiments which
are useful for ‘around T,’ behaviour.

3.2.5 pSR (muon spin relaxation)

Here a spin-polarized beam of positive muons k+ (m = 250m,, Q = +e

and S = ?J, produced by an accelerator, is implanted in our spin glass. The

QMn (5at.%,Tg=27.4K)

i log,() (t) (sec)

Fig. 3.25 Comparison of the time correlation C;(t)of the Mn moment in CuMn (5
at. %) measured by neutron-spin echo (NSE), zero-field pSR (ZF @GR), and ac
susceptibility (xac); from Uemura et al. (1985).
84 Basic spin-glass phenomenon

u.+ will process about any local magnetic fields (internal or external) felt
at its rest site. After 2.2 us the muon decays by emitting a positron along
the instantaneous direction of its spin S. One measures the asymmetry in
the direction of the emitted positrons. Although various configurations of
fields and detectors are possible, we, for simplicity, treat one important
case for the spin glasses, namely, that of a weak (LF+R), or even no
(ZF-l&R), longitudinal field, H L, applied parallel to the incoming beam
and thus HL((pL. The difference between forward emitted positrons and
backward emission is related to a spin-relaxation function G,(t).
We assume for a spin glass that there exists a lorentzian distribution of
random internal fields which fluctuate as a markovian process ee”’ (v is the
rate of fluctuation). This model introduces two parameters: ‘a’ the half-
width at half-maximum of the lorentzian field distribution, and v which is
taken as the inverse average spin-correlation time reflecting the fluctuations
of the local field, v = l/7,. For v/a % 1 (slow modulations or quasi-static
local fields)

G,(t) = i exp (-i vt) + $ (1 - at) exp (-at) (3.57)

In the opposite limit v/a % 1 (fast modulation or rapidly fluctuating magnetic

moments)
G,(t) = exp (- &%) (3.58)

Hence, the LF-l&R experiments can determine the two important para-
meters, a as a spatial distribution of local fields, and 7, as a sort of spin-
relaxation time. An advantage of this technique is that it can be performed
in zero external field. We remark that the time window available in uSR
to study relaxation processes is = 10-10-10-6 s.
With the above oversimplified parametrization we obtain results for G,(t)
in the different limits. Then T= can be extracted at each temperature and
plotted versus l- TJT, as is shown in Fig. 3.26 for various alloy systems
and concentrations. This figure illustrates the rapid upturn in 7, as Tf is
approached from above followed by a saturation very close to Tf. The
effective correlation time appears to move through the l&R-time window
with a power law dependence. Best fitting of the data shows that

(3.59)

where 7, = lo-l2 s and II = 2.6-2.9. Note that rounding does seem to

occur for reduced temperatures less than 5 X 10e2. Analogously, we have
a consistent picture with our average or effective relaxation time slowing
down, but not diverging, as Tf is spanned.
Regarding the spatial distribution of local fields (parametrized in ‘a’)
which are sampled by the muons, two extreme cases have been considered.
3.2 Experiments spanning Tf

&Fe

Fig. 3.26 Correlation time T, of Mn (or Fe) moments determined by zero-field

@R from the dynamic depolarization rate observed above Tf. The straight line
corresponds to the power law of critical slowing-down. From Uemura et al. (1985).

(i) Spatially inhomogeneous dynamics, where various sized clusters

independently fluctuate with the particular correlation time of the
given volume. In other words, each (rigid) cluster has its own
chracteristic T, and there are many different cluster sizes and thus 7,s.
(ii) Spatially homogeneous dynamics, the local field at a given site is a
superposition of various contributions from spins belonging to different
clusters with different 7,s. Said otherwise, the separate clusters are
really coupled via the long-range interactions and each site feels a
superposition of these many fields which in turn lead to a similar
fluctuation spectrum of the local field at different sites. Hence there
is no spatial inhomogeneity.

Since we are not dealing with a superparamagnet we would expect the

latter (ii) model. After a lengthy and systematic analysis of the muon-spin-
relaxation function, (ii) was confirmed around and slightly above Tf.
Therefore, we have gained new and corroborative information from the
local probe of $3R about our experimental model of interacting cluster out
of which the spin-glass state is formed.
86 Basic spin-glass phenomenon

Another achievement of the pSR technique is to relate its measured

relaxation time to the spin correlation function determined from neutron
spin echo and xac. If we set c(t = 7,) = l/e (e is the base of the natural
logarithms), then we obtain the triangular points in Fig. 3.25. On the other
hand, if we bring an order parameter Q, as some sort of static component
(with linewidth a,), into the (ISR random-field distribution via Q = (u&z)*,
then c(t = T,) = (1 - Q)/e + Q. And these are represented by the two
solid dots in Fig. 3.25. Hence, we have a few agreeing points in the
intermediate time range spanned by the muons.
It is the combination of spatial (a local probe) and temporal ( 10-10-10-6 s)
resolution that has made this very sophisticated method so powerful for the
spin glasses. l.rSR seems to have been the contact point with a fractal cluster
model (see section 5.10) which will give us a little more quantitative insight
into our phenomenological picture. What does t&R say about the question
of a phase transition. 3 First of all, the static linewidth a, remains finite at
T,, i.e., all the spins are not statically frozen. Secondly, the T, m (T-Z’,)/T,
does not really diverge in the allowed time window, and even when one
can extract or estimate a critical exponent, it is much different from those
values obtained using other techniques. At the risk of repeating oneself, if
there is a phase transition, it is certainly unconventional: the usual theories
and intuitions are not instructive.

3.2.6 Miissbauer effect

Another microscopic or local-probe technique is the Mossbauer effect which

involves the resonance absorption of a ‘recoil-less’ emitted gamma rate
from a radioactive nucleus. 57Fe turns out to be one such Mossbauer
nucleus. If a local magnetic field is present, then there will be an induced
hyperfine field which splits the 57Fe nuclear spin, I, into a doublet ground
state (I = $J and a quartet (I = t) excited state. It follows that the emission
of a y-ray from this latter state with the selection rule Am, = +-1 or 0 has
six paths. Accordingly, the single-line emission in zero field becomes six
lines in a field. Both the amount of hyperfine splitting and the intensity
ratios give valuable information about the magnetic state of the sample in
which the 57Fe is embedded.
This hypertine field can be related to the local spontaneous magnetization.
For, say, ferromagnetic (pure) iron below its T, = 1040 K, the full
spontaneous magnetization and thus the Mossbauer splitting obeys the
standard (order-parameter) temperature dependence. What about a spin
glass? First we need to know and include the time scale of the Mossbauer
effect. The decay processes have an intrinsic lifetime of about lo-’ s. So if
a spin in a AuFe alloy is frozen on a time scale longer than = lo-’ s, the
hyperfine splitting will appear, regardless of the orientation of the frozen
spin. The technique is a zero-field one and measures the local freezing
3.2 Experiments spanning Tf 87

without determining the various alignments. In Fig. 3.27 we show a very

early measurement (Violet and Borg 1966) of the hyperfine field versus the
reduced temperature for a series of AuFe spin glasses. While there is a
lack of data near To, now our T,, the solid line through the data has a
form which does not differ much from that of pure Fe. This was one of
the first clues that perhaps a phase transition was occurring.
By adding an external field to the Miissbauer effect, the changes in the
intensity ratios can be used to determine the spin alignment. Then it was
called a peculiar type of antiferromagnetic (weak, canted) ordering. Further
analysis pointed out the existence of a random distribution of magnetic
moments and various probability distributions for the internal field could
be established. The randomness of the local-spin arrangements was greatly
affected by the external field - a ferromagnetic component could be induced.
We shall have much more to say about these low-temperature properties
in the next section.
Modern Miissbauer experiments have focused on the smearing of the
transition and relaxation effects near Tf. There is indeed a finite temperature
interval of some degrees which signifies the Miissbauer transition. The best
estimate of a Miissbauer ‘transition temperature’ is systematically higher
(- 20%) than Tf from the susceptibility cusps. Local-moment relaxation
times may be deduced from the paramagnetic Mdssbauer (lorentzian)
linewidths. For a typical AuFe (3 at. %) spin glass TM, the average relaxation
time, begins to deviate from its paramagnetic value (lo-l2 s) far above Tf.

0 0.2 0.4 0.6 0.6 I .o

T/TO

Fig. 3.27 Normalized Miissbauer-effect splitting versus T/T, for a series of AuFe
alloys. From Violet and Borg (1966).
88 Basic spin-glass phenomenon

As T + Tf, TM rapidly increases to = 5 x 10e9 s and then the single-line

spectrum evolves into the six-finger one of the ‘static’ field state, i.e.,
slower than low7 s. Our analysis and interpretation follow the previous lines
of thought, namely, a distribution of relaxation times which progressively
broadens and shifts via an ‘average’ time to slower and slower values. In
addition, a careful examination of the six-finger pattern and its evolution
suggests an inhomogeneous ‘static’ broadening related to non-zero average
spins (3) with a distribution of magnitudes \(&I). This conclusion which is
a direct result of local-probe or microscopic measurement technique, is
certainly connected to our collection ferromagnetic clusters of different size
proposed previously. It is interesting to note that the inhomogeneous static
broadening from the Miissbauer effect (not related to the spatially
inhomogeneous dynamics proposed and rejected for the uSR) goes through
a maximum very close to Tf from the susceptibility cusp. Such behaviour is
what we would expect at a percolation transition so once again we have
another, now Mdssbauer effect, confirmation of our spatial cluster and
lengthening relaxation-time model.

3.3 LOW-TEMPERATURE (T 4 &I EXPERIMENTS

Now we must confront the experimental consequences of our frozen state.

Certainly there is a new and unique magnetic phase, but how can we
succinctly characterize its basic properties? In this section we select a few
simple, yet representative, experiments that can maintain the thread of
continuity from the previous temperature regimes, and use them to delineate
the ground state of a spin glass. We fully realize that there is no long-
range periodic order, so conventional ordering behaviour and elementary
excitations will be transmogrified or non-existent. Dynamical effects persisting
down into the frozen state govern many properties, especially how they are
modified with an external field - the key variable of the new phase. Our
choice of measurement techniques will be to persevere with our four basic
experiments: magnetization will take the place of susceptibility, since the
proportionality M m x no longer exists, specific heat and resistivity our two
standard thermodynamic/transport properties, and neutron scattering, where
we consider some recent results and their iconoclastic interpretation. Finally,
two new techniques will be introduced, namely, torque and electron
paramagnetic resonance (EPR) which have clarified the particular importance
of anisotropy but requires rather large field sweeps. As always we try to
treat only generic behaviour, i.e., we carefully select our examples to
correspond to the most basic and general characteristics of the canonica1
spin glasses.
3.3 Low-temperature (T Q Ta experiments 89

3.3.1 Magnetization

Let us use one of the many magnetization techniques and zero-field cool
(ZFC) the spin-glass sample to T < Tf. We begin with sweeping the applied
field slow enough that the measurement is effectively dc, however, the
sweep rate is sufficiently rapid to avoid large relaxation or creep effects
mentioned previously. A generic sketch is given in Fig. 3.28. The small
initial slope near the origin is comparable with our ac-susceptibility, limit
(H + 0) M(ZFC)/H = xac, and before discernible time dependences occur
a small finite H is required. These usually appear around the start of the
S-shaped form in the virgin M versus H curve. Thus, already in small field,
the magnetization slowly drifts upwards roughly following a log-time scale.
The inset of Fig. 3.28 illustrates the upward creep for increasing field and
a downward one for decreasing field by a thick line.
As we reach very large fields (= 40 T), now shown in Fig. 3.29, there is
a tendency towards saturation, but full saturation Ms,, = g&S (- 4 FB for
CuMn) is never attained even at the lowest temperatures and concentrations.
Notice also how the magnetization at constant field decreases with increasing
concentrations (Fig. 3.29). What does this all mean? Well we certainly do
not have a superparamagnet which is easy to polarize! Instead, not only

ZFC
T<<Tf

‘H

Fig. 3.28 Schematic of zero-field cooled hysteresis loop for T 6 Tf illustrating the
S-shape, the lack of saturation, and the isothermal remanent magnetization (IRM).
The time dependences are depicted in the inset.
90 Basic spin-glass phenomenon

,gIY 14 at %

0 10 20 30 40
B (V
Fig. 3.29 Magnetization per Mn-atom for the various concentrations of CuMn
studied at 4.2 K; from Smith et al. (1979).

the long-range RKKY interactions, but also the peculiar anisotropy associated
with CuMn play a major role and prevent the external field from fully
rotating the frozen moments into its direction. One could try to calculate
the various exchange interactions based upon a superposition of three
Brillouin functions fitted to a universal (concentration scaled) magnetization
curve. However, for CuMn too much antiferromagnetic coupling is required
for a reasonable fit. The results totally disagree with our cluster configuration
model’s estimates of the .Jis from section 3.1.1 and Fig. 3.4. Since the
present magnetization versus field behaviour is a low-temperature T G Tf
measurement, there must be a lzew aspect to the frozen state. And this is
the random anisotropy which is formed below Tf and is particularly strong
in CuMn. The external field has to overcome an array of local anisotropy
axes, randomly oriented, before the various clusters can point along the
field direction and fully align, i.e., saturate, the spins. It takes field energy
to rotate the cluster moment away from its anisotropy-pinned ‘frozen’
orientation. Such an energy barrier for rotation was already introduced for
a superparamagnet, but now for a frozen spin glass the barrier orientations
are randomly distributed and the barrier heights, also a distribution, are
temperature dependent disappearing above Tf. We shall delve more closely
into this anisotropy when we consider the torque and ESR experiments
(section 3.3.5).
The AuFe system is somewhat different. The magnetization at constant
field increases with increasing concentration and saturation, while not
completely reached, appears a bit closer, here Ms, = &J+$ = 3 PB, while
3.3 Low-temperature (T Q Ta experiments 91

M(40 T) = 2-2.5 l.~n. In addition, the Brillouin-function fitting gives a more

reasonable estimate for the exchange parameters Jr and J2 which compare
favourably with that found from the cluster configuration model. The
random anisotropy also is much weaker and field cooling (FC) does not
induce a unidirectional component (see below).
If we FC CuMn to T G T,, a most unusual ‘displaced’ hysteresis loop
occurs upon reducing the field through zero to negative values and back,
thereby tracing out the loop. Figure 3.30 presents this behaviour for CuMn
(0.5 at. %) after a few kgauss FC to 1.35 K + Tf = 6.5 K. Metastable with
respect to time, point A represents the thermal-remanent magnetization
(TRM) - a field-cooling effect where, after the temperature change, the
FC field has been reduced to zero. Note what happens when the applied
field is increased in the negative direction. The magnetization remains
positive. At point B (or B’ dependent on the field sweep in different runs)
there is a sharp switching to negative magnetization - a jump in M to point
C (C’). Upon proceeding to point D and then returning the field back
through zero, the reversal -M to +M of remanent magnetization, E to F,
(at H 5 0) is approximately equal to the initial value of the remanence
(neglecting the small time dependence). Therefore, we have a displaced
hysteresis loop or memory effect whose width is determined by the weak

cu Mn
cu Mn qsqs D/.
D/.
T = 1,35OK
+loo-
+loo

Fig. 3.30 Hysteresis for field-cooled CuMn with saturated remanence magnetization.
Different symbols are used for different runs. From Monod et al. (1979).
92 Basic spin-glass phenomenon

uniaxial anisotropy and whose displacement (position of the loop centre

with respect to the origin) by the unidirectional anisotropy.
A simple physical picture here is that the FC induces a preferred direction
along which a small net ferromagnetic moment is frozen. This causes the
TRM. However, a small negative perturbing field can flip the preferred
direction and thus the TRM from + to -. We visualize an external field-
dependent energy barrier which has its lowest lying minimum along the
field direction (Fig. 3.31 a). When the field is reversed (H < 0), so is this
the minimum and now a thermal fluctuation or ‘noise’ will push the TRM
over the small energy hump into this opposite-spin minimum: -M
(Fig. 3.31 b). Finally, when the field is reduced towards zero, the return
process takes place (Fig. 3.31 c) and we have generated a hysteresis loop.
The value of the return field in Fig. 3.30 is slightly positive and depends,
in general, on the value of the uniaxial anisotropy, which determines the
width, and the unidirection anisotropy which gives the displacement. The
asymmetry in our naive potential-barrier sketch is created by the unidirec-
tional component. The entity which spin flips between + M and -M is our
small TRM. Since these switches of magnetization are so sharp, there must
be a co-operative nature to the frozen spin-glass state, i.e., a long correlation
length between the frozen spins such that they can collectively turn the
TRM back and forth on a macroscopic scale. For AuFe there is simply not
enough unidirectional anisotropic which can be induced by field cooling.
By including a third element, e.g. Pt, to increase the spin-orbit coupling of
the alloy system, a displaced hysteresis loop can be established.
Before we proceed to the time dependence of the various magnetizations,
let us classify the two types of remanent magnetizations which occur in a
spin glass. As already mentioned, the thermoremanent magnetization (TRM)
is a field-cooling effect for which the sample temperature is reduced from
T > Tf to T < Tf in an FC field H. At this lower temperature, H is set

t H,,H=O H<O
l TRM l TRM l =TRM
(4 (b) (4
Fig. 3.31 Memory effect as represented by energy-barrier deformations for the
field-cooled magnetization of CuMn.
3.3 Low-temperature (T e Td experiments 93

equal to zero and the magnetization measured. Here saturation of the TRM
is quickly reached for low values of H. Figure 3.32 illustrates this behaviour
for a canonical spin glass AuFe (0.5 at. %). For the TRM an irreversible
susceptibility may be defined as Xir = TRM(H)IH. By adding xir to x =
u?$(M/H), called the reversible susceptibility, xFc may be indirectly obtained.
From this procedure XFC is found to be constant for T 5 Tf - the result
agrees nicely with the direct method of simply FC a spin glass in field H.
The second type of remanence is called isothermal remanent magnetization
(IRM) and is obtained by cooling the spin glass in zero field (ZFC),
followed by cycling an external field 0 + H + 0, and then measuring the
magnetization. A more difficult saturation of the IRM is observed for
sufficiently large H values. Figure 3.32 also shows the IRM(H) with its
slower approach to reaching the already saturated TRM.
We have frequently mentioned relaxation and time dependences of the
magnetization within the frozen state. Let us now examine these in some
detail. With the presently available, highly sensitive SQUID magnetometer,
it is quite easy to detect very small changes (10-100 ppm) in magnetization.
What is more difficult is to prepare the sample magnetically for a meaningful
and reproducible experiment. We treat here a few simple well-defined
examples, which have created a great deal of interest and introduced a
totally new concept, viz., the waiting or ageing time, t,.,.
Magnetization of a spin glass will react in time to any change of external
field while in the frozen state. There are various routes to prepare a spin
glass for systematic measurements of these long-time relaxational effects.
Since the behaviour is irreversible and complicated by ageing processes (see

Au- Fe 0.5 %

Fig. 3.32 Field dependences of the thermoremanent magnetization (TRM) obtained

after cooling from Ti (+ Tf = 5.4 K) to T = 1.2 K in a field H, and of the isothermal
remanent magnetization (IRM) obtained when a field H applied at 1.2 K is
suppressed; from Tholence and Toumier (1974).
94 Basic spin-glass phenomenon

below), it is imperative to employ a well-defined H-T cycling procedure.

Let us consider two such preparation recipes as sketched in Fig. 3.33. The
first utilizes field cooling through Tf in a weak field, and after the FC, one
waits a time t,,. before reducing the field to zero and measuring M(t).
Alternatively, we can ZFC (or T-quench) the spin glass and wait t, before
applying a small field, see Fig. 3.33 b. Afterwards M(t) is tracked over
many decades of time for different t,.,. Note in both cases that the waiting
time begins running after the sample is cooled to its final temperature
T < Tf and stops at the change of field when the measurement time t starts.
There is also a field-quench route whereby a very strong field is applied in
the frozen state, then switched off; and after t,, a weak field is restored
and M(t) measured. Here the behaviour is similar to the ZFC or T-quench
method.
Figure 3.34 shows some typical data for CuMn (10 at. %) after the ZFC
procedure of Fig. 3.33 b (H is usually limited to fields less than 20 gauss).
The base line of the figure is arbitrary and the relative scale is given in
units of the FC-magnetization divided by H. It is interesting how M(t)

TA 4
t

3-

I
I
1 t
*
TIME

I I
I I
IL----____-----___---__ I _t
‘TIME

Fig. 3.33 (a) FC-TRM MR(t) measurement: solid line T-profile, dashed line ff-
profile; (b) ZFC or T-quench M(t) measurement: solid line T-profile; dashed line
H-profile.
3.3 Low-temperature (T 6 TJ experiments 95

(a)

loo 10’ lo2 lo3 lo4 t(sec)

(b)

0
loo 10’ lo2 lo3 lo4 t(sec)

Fig. 3.34 ZFC susceptibility (lIZ-Z)M(t) and the corresponding relaxation rate S(t)
at different wait times, tw = 102, 3 x 102, 103, 3 X 103, lo4 and 3 X lo4 set plotted
versus log t: (a) (l/H)M(t), 5% of (1IH)M Fc is indicated; and (b) S(t), 1% of
(l/H)k& is indicated. CuMn (10 at. %), T/T, = 0.91, Tf = 45.3 K. From Sandlund
et al. (1988).
96 Basic spin-glass phenomenon

always increases, has an inflection point, and is a strong function of t,. The
latter behaviour is due to the ageing processes of the spin glass below Tf
which are continuing even though H = 0. These are then reflected in the
different time dependences of M(t) after each f,. The lower plot gives the
relaxation rate S(t) = (l/H)aM/a In t. Here the inflection point of M(f)
corresponds to a maximum in S(f) which shifts to longer observation times
(t) with increasing t,. The peak in S(t) occurs at a time f which is
approximately equal to f,. For the small fields used in the experiments the
ageing process proceeds unaffected by the field. Also for small H (< 20 gauss)
there is a linear-response regime where the data scales as M(t)IH and the
principle of superposition applies for different field-cycling procedures. For
example, the isothermal remanent magnetization (IRM) may be measured
by ZFC, applying a field after a waiting time t,, returning the field to zero
at t; and then tracking IRM(f). The same quantity can be calculated by
applying a +H at t, and superimposing a -H at t&, and then subtracting
the two measured M(t)s. Needless to say the agreement for the decay of
IRM(t) from direct measurement and this subtraction is excellent.
Various functional forms have been proposed to describe the magnetization
as a function of observation time and waiting time. One of the most popular
relations is the stretched exponential
M(t) = MO exp [-(t/tr)l-n] (3.60)
where il4, and tP (the time constant) depend upon T and t,, while n is only
a function of T. If n = 0, we have the Debye, single time-constant,
exponential relaxation. On the other hand, for n = 1, M(t) is constant. So
the value of n will critically govern the exact relaxation rate from very
strong to none at all. In order to verify such dependences sensitive
experiments must be carried out over many decades of observation time
and waiting time for a well-defined H-T procedure. Some illustrative
measurements, which have attempted the above, are displayed in Fig. 3.35.
Once again we have CuMn (5 at. %) and the FC-cycle has been used to
create a TRM. In fact, M,(t) is tracked for different waiting times and
temperatures over four decades of observation time (l-lo4 s). This is
indicated by the bold solid line in the figure. Next an attempt is made to
accommodate the data with a stretched exponential using a two parameter
(tP and n) fit of the above equation. The results of the best-fit are denoted
by the fine solid lines. While the agreement is good at short times, clear
deviations (see Fig. 3.35) commence as the time interval is extended. In all
four cases the stretched exponential function is insufficient to match the
data over a broad range. Something more is needed and this has been
provided by the Uppsala group, Nordblad et al. (1986). Superimposed upon
the stretched-exponential contribution is a purely logarithmic-decay term

Mk=SHlnt (3.61)
where S is the relaxation rate in ‘dynamical equilibrium’ which only weakly
3.3 Low-temperature (T Q Ta experiments 97

t w =lOOO set

(_i1o-4 1o-2 loo lo* lo4 10 ’ t(sec)

t w =lOO set

lo-’ lo+ loo lo2 lo4 10 ‘t(sec)

Fig. 3.35 Relaxation of the remanent magnetization (MR) in the time interval
10m4< t < 108. Thick full lines are experimental data. Dashed lines are the
relaxation according to power-law plus stretched-exponential dependences. Thin full
lines represent the stretched exponent alone: (a) t, = 103 sec., temperatures 21 K
(top) and 25 K (bottom; and (6) t, = 10r sec., temperatures 21 K (top) and 25 K
(bottom); from Nordblad et al. (1986).

depends upon the time and waiting time. With the addition of equation
(3.61) to equation (3.60) the fitting is greatly improved as illustrated by the
dashed line in Fig. 3.35. Therefore, the time decay of MR(t) is logarithmic
for t Q t, and t + t,,,, and the ageing process superimposes a stretched-
exponential term on the relaxation around t = t,,,. At this point a bit of
theoretical interpretation is required to distinguish the quasi-equilibrium
dynamics represented by the log(t) decay at very long or very short time
scales from the non-equilibrium dynamics causing an ageing-time dependent,
stretched-exponential behaviour spanning t = f,. We shall defer such
explanations until the theoretical chapter (section 5.9) when we will introduce
a domain-growth model.
98 Basic spin-glass phenomenon

3.3.2 Specific heat

Here we can return to the low-temperature portion of our previously

discussed measurements on the magnetic specific heat, e.g., Figs. 3.5, 3.6
or 3.21. Below Tf to a first approximation the temperature dependence of
C, in these figures seems linear. This means that C,,,IT versus T should
stay constant in a more severe test of the experimental data. For CuMn
and (EuSr)S there is a substantial low-temperature regime where &IT is
indeed constant. Yet at the lowest temperatures definite deviations occur
particularly in CuMn. A better example of the linear dependence is AuFe
(Fig. 3.36) which shows a 1 at. % sample with only a very slight deflection
around 0.4 K from the line extending down from Tf = 8.5. A positive
curvature must take place to force C, to zero as T + 0 (3rd law of
thermodynamics). In this particular case less than half the entropy
S, = CR In (2s + l), where S = $ for Fe, appears between T = 0 and Tf.
Such is consistent with our high-temperature C,,, conclusion that most of
the available entropy is already lost above Tf.
This strong (recall there will always be a weak yT-term for a normal
metal) linear T-dependence of the magnetic specific heat at low temperatures
(known already since the early 1960s) marked the starting point of the spin
glasses in the early 1970s. For, then it was discovered that real glasses or
amorphous solids also had an excess specific heat with respect to their

Au - I- at.% Fe

TEMPERATURE (K)

Fig. 3.36 Plot of the spin-glass specific-heat data fro AuFe (1 at. %). Note that
the linear part extrapolates to a positive temperature intercept. From Martin (1980).
3.3 Low-temperature (T 4 Td experiments 99

crystalline counterparts that was directly proportional to the temperature.

And this led to the now-famous two-level tunnelling model. At low
temperatures, where thermal activation is unimportant, various positional
or spin configurations in an amorphous material or spin glass are separated
in phase space by energy barriers with similar energy minima. Figure 3.37
gives a sketch of a two-level energy diagram with a tunnelling barrier
separating the two nearly equal energy minima. Also shown is the constant
density of states for the two-level excitations. These configurations exist in
a spin glass because of the frustrated, degenerate nature of the ground
state. Hence, quantum-mechanical tunnelling through the barrier may occur
between the nearly similar spin clusters - a small rearrangement or partial
reorientation of some spins modifies the configuration by a tunnelling
process which is temperature independent. The two slightly different
configurations represent the two levels, and assuming a constant density of
level states (see Fig. 3.37), a linear proportionality at low T naturally arises

I
O.lev AE

E(X)’

Fii. 3.37 Two-level energy-barrier construction and density of states at very low
energies AE = El - E2.
100 B&c spin-glass phenomenon

for the excess ‘amorphous’ specific heat or for the randomly frozen spin-
glass C,,,. It was this connection with the real glasses via the similar specific-
heat behaviours that caused the word ‘glass’ to be attached to the random
magnetic systems whether metallic or insulating. So an approximately linear
specific heat is the low temperature signature of a spin glass.
Since the specific heat is in principle a rather difficult measurement, it
has not been fully used to explore the dynamic of spin glasses by introducing
a frequency or time of measurement into the technique. A related experiment
can be performed to determine the energy flux (or heat flow) dQldt for
T 4 Tf as the external field is changed. Here the spin-glass sample is weakly
connected to a reservoir by a constant heat leak. For both FC and then
setting H = 0 or ZFC and then turning on a field H, heat flows out of the
sample with a long-time relaxation (similar to the magnetization). At very
low T and for a given AH, the energy flux dQ/dt 0: l/t and the energy
Q = (const) In t. One can picture this behaviour as a lowering of the spin-
glass energy by a change of field, either removal after FC or application
after ZFC. Such, alterations drive the system slowly towards a new
equilibrium state. A more complete interpretation may be gained from the
two-level tunnelling system with a distribution of both energy-barrier heights
P(V) as in our NCel superparamagnetic model and splittings of the double-
well (nonequal depths) P(AE) caused by the field change AH (see Fig. 3.37).
Note that the tunnelling processes are essential at these low temperatures
(T < T,), since thermal activation (or surmounting the barrier) is mainly
ineffective. Therefore, we maintain our nexus among relaxation, energy
barriers and tunnelling, all participating in the quest for equilibrium.

3.3.3 Resistivity

At low T the magnetic resistivity of a metallic spin glass, Ap, has a large
residual (T = 0) component - a combination of the disorder of the alloy
due to random A and B site occupancies, and the spin disorder of the
randomly frozen state. We call this contribution xAp, since these disorders
are roughly proportional to X, the concentration. As the temperature is
increased
Ap(T, x) = xAp, + A(x)Ts (3.62)
where A(x) decreases very slowly with increasing concentration. We illustrate
this behaviour, which seems rather general, in Fig. 3.38 for a series of
CuMn alloys.
Some evidence has been offered for the appearance of a limiting T2
dependence at the very lowest temperatures (T < 0.3 K). However, this
must quickly crossover to the Ti behaviour for a larger T-range approximating
Tf. The interpretation of these A(x) and T%dependences has been discussed
in terms of long-wavelength elementary excitations which are diffusive in
3.3 Low-temperature (T -+ TJ experiments 101

- T(K)
1.5 3.0 LO 5.0 6.0 7.0 6.0 9.0
0.3

0.0
0 5 IO 15 20 25
- 1% fK-%

Fig. 3.38 The tfmperature variation of the resistivity for 4.4, 6.3 and 9.5 at. % Mn
in Cu versus Ti. The lines through the data represent a fit to this temperature
dependence. From Mydosh and Ford (1974).

character. The excitations are thought to be highly damped, non-coherent

(independent), localized spin fluctuations which scatter the conduction
electrons. Such a model is against propagating excitations, e.g. spin waves
which occur in long-range order magnets. Once again as in the two-level
system, we resort to a localized description for our frozen state. The state
may be a co-operative one with correlation length, &, large, but because
of the disorder the excitations cannot propagate.

3.3.4 Neutron scattering

What does inelastic neutron scatter say about the above conjecture? Usually
this technique provides a powerful probe for the study of collective spin
excitations. If we recall our dynamic structure factor S&q, w) given by
equation (3.17), then we could look for propagating spin waves, namely,
two inelastic lines at o = +o,. Hence,
or
Im x(9, 0) = (0 - w,)’ + I-2 + (w + $2 + r*
where r is the inverse lifetime of the spin wave which has a typical
dispersion relation wq = Dq* (with D the spin-wave stiffness constant).
Much experimental effort has gone into searching for these excitations in
the metallic spin glasses - all in vain. There are no sharp inelastic lines to
102 Basic spin-glass phenomenon

be found, meaning the modes are overdamped, thus localized. A similar

conclusion was reached from our resistivity measurements. For the insulating
spin glasses, e.g. (EuSr)S, some indication was found of broad quasi-elastic
to inelastic scattering intensity as weak shoulders to the elastic central peak.
The broadness implies I is very large, distinctive of short lifetime or highly-
damped excitations. Here the experimental and analytic difficulties become
so great that we can only conclude that propagating spin waves do not, in
general, exist in the canonical spin glasses.
More recently a quite radical interpretation has been proposed for the
various neutron scattering results in the noble-metal-based spin glasses (e.g.
CuMn, AuFe, etc.). We alluded to this explanation in section 3.1.4 and
Fig. 3.11, where single-crystal samples must be used to determine the exact
place in reciprocal lattice (q) space from which the scattering occurs. So let
us consider in more detail the two relatively sharp peaks in Fig. 3.11 which
emerge at low T out of the broad in q-space magnetic scattering. These
peaks, called satellites, appear at q-values of (1, 0, $ 2 6) and other
symmetry-related positions. In Fig. 3.11 8 takes the value of 0.19. The
position indicates that there is a modulation, A, of spin correlations which
has a period longer than the unit-cell spacing (a, = 3.6 A for Cu), namely
A = a,/($ - 6) = 11.6 A. The halfwidth I of the satellite peaks is not
resolution limited - they are not true Bragg peaks. This means the
modulation of spin correlations is not a long-range propagating sinusoidal
wave with period A. In other words, its correlation length 5 is finite and
may be estimated from I. For the case of q-dependent peaks a gaussian fit
is usually made to determine the halfwidth in units of 2&,. Hence
5 = 2Vln 2/I = 6u, or 22 A dependent on the concentration x of Mn or
Fe. This corresponds to a domain of roughly 2000 atoms of which the
fraction x are magnetic impurities.
We call this incommensurate (having a different periodicity A than that
of the lattice which is an integer times a,) modulation of spin correlations
a spin density wave (SDW). A standard example of an SDW material is
the antiferromagnetism of Cr, others are the helical ordering of certain
rare-earth elements or alloys. SDW represents a collective, excited state of
the conduction-electron gas with an energy greater than the ground-state
energy. This higher energy can preclude its formation in a pure non-
magnetic metal, e.g. Cu or Au. However, in a dilute magnetic alloy at low
T, the orientation of the solute spins in the effective field of the SDW
provides an interaction energy which may more than compensate the higher
energy of the SDW. Here the spin polarization of the electron gas s varies
with position R as
s=Nbccos(q - R) (3.64)
where N is the number of atoms per cm 3, b is the order-parameter amplitude
of the SDW, E its polarization and q its wave vector. Paramagnetic solute
spins, e.g. Mn or Fe, will interact with the SDW via s-d (or s-f for the rare
3.3 Low-temperature (T 4 Td experiments 103

earths) exchange interactions. Hence, a solute spin Sj at Rj will experience

an effective field
Hj = Gk COS (Q * Rj) (3.65)
where G is related to the exchange interaction between conduction electrons
and solute spins. The interaction energy is -Hi - Sj. It tries to orient the
solute spin along the effective field, and thereby compensate for the positive
energy necessary to form the SDW. Upon removing the thermal disorder
by going down to low T, b, the SDW order parameter, will become non-
zero at an ‘antiferromagnetic’ critical temperature TN.
This phase transition had already been derived by Overhauser in the late
1950s and within the mean-field approximation TN and the various
thermodynamic quantities may be calculated. The Overhauser theory leads
to long-range antiferromagnetic order via a second order phase transition
of the randomly distributed solute spins according to the SDW polarization
s(R). The ordering wave vector q is usually incommensurate with the lattice,
i.e., unequal to the reciprocal lattice vectors G, since a SDW is a Fermi-
surface effect, most favourable when 9 = 2k, (Fermi momentum) unrelated
to the lattice.
Now, what are the neutron-scattering results telling us about the possible
SDW in CuMn et al.? First of all, the ordering q-wave vectors have been
determined by the satellite peaks which clearly emerge below Tf. However,
the linewidths I of these peaks remain too broad for a long-range
propagation of the SDW. In CuMn, our example in Fig. 3.11, the sinusoidal
propagation extends only about 20 A (its correlation range) then breaks off
and with different amplitude and direction a new SDW begins and traverses
= 20 8, before the process is repeated. We sketch this short-range (SR)
SDW in Fig. 3.39 (lower part) where there is an additional complication
due to the formation of very short-range ferromagnetic clusters embedded
within the SR-SDW. That the chemical or atomic short-range order leads
to these ferromagnetic clusters has also been established from the polarized
neutron-scattering experiments and its separation into nuclear and magnetic
scattering, so, the ‘chunks’ of ferromagnetic spins simply follow the SDWs
sinusoidal polarization of the electron gas.
As a basis for comparison we have shown the results for a typical long-
range (LR) SDW (linearly polarized) in Fig. 3.39 (top part). The alloy YGd
is similar regarding concentration, but it has a preferred SDW propagation
direction along the c-axis of its hexagonal crystal structure. Here the
correlation range 5 + 03 and A = 20 A with an incommensurate periodicity.
The Gd 7 pn moments align themselves in the u-b plane according to the
modulation of the SDW.
This picture of a SR-SDW is entirely consistent with the arbitrary q-value
scattering on polycrystalline samples discussed in section 3.2.4. Then the
model was a distribution of relaxation times that comes from the fluctuations
of the ferromagnetic clusters which are a direct result of the chemical SR
104 Basic spin-glass phenomenon

0-v plane v Gd (hcp)

LR- SDW (incomm.)

Cu Mn (fee)
I -

hzz 1oA
I - t- -II-+
+-2oii ‘-I gz208 etc. _....
SR-SDW (incomm.)

Atomic-SRO

-‘j
-,) Ferromagnetic
clusters embedded
with in SR-SDW

Fig. 3.39 Schematic of linear SDW propagation, upper part: long-range for YGd
and lower part: short-range for CuMn. From Mydosh (1988).

order. The satellite magnetic peaks can only be observed on single-crystal

samples with a separation of nuclear and magnetic scattering. And indeed
our clusters do slow down and become static as the SDW forms below Tr.
We conclude by mentioning that the thermodynamic and transport properties
of our LR-SDW (YGd) have all the ‘sharp’ behaviours of a good second-
order, long-range phase transition. There are nolte of the time dependences,
irreversibilities, remanences, etc. characteristic to the spin-glass state. Why
a LR-SDW does not form in CuMn, AuFe etc. has to do with a number
of specific reasons, such as the weak polarizability of the noble-metal
matrix, the more itinerant 3D moments, the strengths of the RKKY
interaction between the moments and the cubic crystal structure which by
symmetry allows twelve different propagation directions each associated
with a specific q-vector. For these reasons the spin-glasses become unique
because they are disordered at long range and this makes them so very
difficult to describe, even if we realize an incipient SDW is starting to form.
3.3 Low-temperahue (T 4 Td experiments 105

3.3.5 Torque and ESR

According to our experimental model, a random anisotropy should appear

in the frozen spin-glass state. For metallic systems the origin is thought to
be the Dzyaloshinsky-Moriya interaction, while for the insulators the dipole
interaction seems to provide the random anisotropy (see section 1.4). In
any case, the problem with a random (both in magnitude and direction)
and local anisotropy is to measure it, since on a macroscopic scale it would
average out to zero. As was seen from the magnetization studies, we can
via FC or ZFC, then a field, induce a remanent magnetization M, along
the field direction. The field, in whatever direction it is applied, will probe
the local random anisotropy for each frozen spin by trying to rotate it. If
the spin-glass state is a cooperative, collective one, i.e., all spins locked
into a ‘rigid body’, then the rotation of this stiff spin system will be governed
by a single macroscopic anisotropy parameter K. But how can we probe
the ‘stiffness to rotation’ of our rigid collection of randomly frozen spins?
The trick is to use the induced remanence M, as a handle fixed in our spin
system and then to turn Mr with the external field. Yet Mr and K are
independent quantities. Usually a combination of unidirectional (K,) and
uniaxial (&) anisotropies are required to describe the magnetization results
for a displaced, broad hysteresis loop. Thus, in general, the total anisotropy
energy may be written for rotation 0 as

E*(O) = Ki(1 - cos 0) + $K2 sin20

A very effective way of investigating the anisotropy is through torque
measurements. The torque I is defined as the vector product of magnetization
times external field, or for field-prepared spin glasses
I = MJZsin (0, - 0) (3.67)
where OH is the rotation of H, and 0 is the rotation of Mr from their
original (FC) direction. With respect to the energy

The measurement of rotation or torque may be carried out very accurately

by determining the corresponding displacements of two metal plates using
a capacitance bridge. For simplicity, we choose a sample (large Mn
concentration in Cu) and a field-cooling procedure such that K2, the uniaxial
component, is negligible. With our applied field along the z-axis, the FC
direction, we have induced a value of Mr, which at the very low temperatures
of the measurement remains roughly time-independent, and created a
macroscopic value for K1 by our rigidly frozen spin system. Next we rotate
the field H about the x-axis, this in turn drags along the Mr at some delay
angle and we can measure the torque. This process is sketched in Fig. 3.40
106 Basic spin-glars phenomenon

Fig. 3.40 Rotation of the anisotropy triad during a torque experiment. A rigid
rotation of the spin system is induced from 0 = 0 to 0 = 7~when H is rotated in
the y-z plane from OH = 0 to Or., = 7~. When a subsequent rotation cpis induced
in the x-y plane, the total rotation angle of the triad is still 7~with respect to the
0 = 0 starting position. However, a non-controlled irreversibility P rotation of the
anisotropy triad around a brings back the rotation angle to n - cp. From Alloul
(1983).

where u is used to represent M,, and 0 is the angle of rotation of h4, with
respect to the z-axis. The field angle OH is somewhat larger due to the
‘pinning’ of it4, by the unidirection anisotropy Kt.
Figure 3.41 shows the measured torque r, versus the field-rotation OH
in curve (a), while the calculated torque derived from equations (3.66) to
(3.68) is also plotted as a dashed line. Notice how the two curves nicely
agree until about OH = 50” (or 0 = 40”). The measured torque is here
reversible and independent of time. This means a rigid rotation of the
entire spin system is occurring with M,. However, at larger angles small
rearrangements are taking place with the corresponding irreversibilities and
time dependences. Nevertheless, the deviations of the experiment from the
calculated curves are not too bad (see Fig. 3.41) for the particular system,
and the remanent magnetization turns over for a field rotation of 180” and
comes back into phase with H.
Now what happens, if we rotate the applied field back to its original +z-
axis, not via a return rotation about the x-axis, but instead via a rotation
+ about the y-axis? This results in a r,, which is measurable and shown in
Fig. 3.41 as curve d. Here there is a large difference in the r,, values
compared to r, at least for the first 20” of rotation. Therefore, a return
rotation about the y-axis after an x-axis rotation is non-identical to a direct
3.3 Low-temperature (T -6 TJ experiments 107

. . . . . . .

*e---w_
. l.
,’ \
I’ \
,’ ‘\
I \
#’ \
#’ \
\
#’
8’
,’ \ ‘\
a’P “,

%(deg)
L

-180 -160 -120 -80 -40 0 'f,,(degI

Fig. 3.41 Torque measurements in a CuMn 20% sample at 1.5 K. The torque r,
and r, are measured after a 7~rotation of u in the y-z plane. These two responses
are quite different and point out the triad character of the anisotropy. r,, is found
to be very weak for small cp as the total anisotropy energy is then cpindependent.
From Fert and Hippert (1982).

rotation from the FC z-axis about the y-axis and back. For, we must attach
a threefold orthonormal coordinate system or triud to our iI4, (or a) as is
done in Fig. 3.40. Note the dissimilarity between an x-axis followed by a
y-axis rotation and a simple y-axis, back and forth, rotation. Although o
always points in the same direction after these rotations the two non-u triad
axes are oriented in opposite directions along the direct and return (after
X) y-axis rotations. This coordinate-system difference or triad property is
manifested in the two unlike values of rY measured around 180”.
Our external field controls the rotation of the remanent magnetization
M,, but we cannot control the rotations of the collective rigid spin system
around M, - slippage will occur at large angles and the spin system will
revert back to its direct-rotation and lower-energy form and r,, = r,. Since
the random freezing is isotropic, our spin system comprises all orientations
in the 3D space, i.e., is non-collinear or Heisenberg-like. This means the
random local anisotropy is also distributed over all directions as an ‘isotropic
anisotropy’. When we probe the macroscopic anisotropy via a FC-field, we
establish the ‘memory’ direction ti of our fixed triad independent of the
108 Basic spin-glass phenomenon

crystal axes. Moreover, we have also induced M, and its triad vector fi
which rotates with M, as we turn i&ii away from the FC orientation by
applying and rotating the external field I-I. With the proper choice of
system, temperature and fields, we can only hope that the randomly frozen
spin system remains rigid and moves bodily and in phase with i&f, which is
being controlled by H. This seems to be roughly true for the x-axis rotation
in Fig. 3.41, but, if followed by a y-axis rotation, the spin system readjusts
or inverts two of its triad axis and returns to the torque and lower energy
of the original x-axis rotation. All these contrasting processes are illustrated
in Fig. 3.40 using the triad coordinate that turns with M, (or a). Note the
striking difference with a collinear (Ising) spin system where all of the
randomly frozen spins point either ‘up’ or ‘down’ along a given unit-vector
axis. For this ‘vector model’, unidirectional anisotropy only, there is no
difference in the spin system after the various axes rotations. It remains
completely the same no matter which path is followed. Thus, we would
expect to see a dramatic difference in the rotational properties between a
Heisenberg and an Ising (collinear) spin glass.
Another powerful experimental technique which can be used to study the
consequences of the spin-glass anisotropy is electron spin resonance (ESR).
Here one detects the spin resonance frequency of the Mn atoms. It has
long been known (late 1950s) that Mn impurities in a noble metal satisfy
the condition for ESR and many interesting physical properties can be
derived from such experiments in the dilute limit. But now we wish to
investigate the co-operative behaviour of the frozen Mn spin system for
T < Tf- For this purpose we can apply a hydrodynamic approach and
construct an appropriate free energy. Hydrodynamic theory represents a
collective approximation at low frequency and long wave-length using
macroscopic variables, and as a result it neglects the microscopic details of
the systems under consideration. For a canonical spin glass we have a non-
collinear, isotropic, frozen, spin structure with its magnetization, M(= xH),
a possible remanent magnetization M, and the unidirectional (K,) and
uniaxial (K,) anisotropies. Accordingly, the free energy (FE) becomes

FE=~(M-AYrfi)z-M~H-KlcosO- ~K,cos20 (3.69)

where the notation is similar to before. The triad coordinates must be used
with N the FC direction and il the rotated ikf, direction. One can check
this FE by minimizing it with respect to 0 and obtain the equilibrium
rotation angle 0, after field cooling I& and rotating the external field H.
Substituting M, = XH + M,ii, as shown in Fig. 3.42, we arrive at the
equilibrium condition with KZ = 0:

sin 0, = MH
__K sin (0, - 0,) (3.70)
1

which is identical to that used in the torque experiments. Figure 3.42

3.3 Low-temperature (T Q Ta experiments 109

Fig. 3.42 Equilibrium rotation angle 0, for a field-cooled spin glass followed by a
rotation of the field H to an angle OH.

sketches the various vectors and their equilibrium angles. These conditions
have been verified by several magnetization and torque measurements as
long as the frozen spin system remains rigid; however, often such is not
the case.
For ESR experiments more is needed. Equations of motion must be
derived for the dynamical variables M and ii. This is usually carried out by
forming Poisson brackets of these variables with the free energy. Let us
consider a special case, namely, M,ii is parallel to H which is parallel to
H,. Now three resonance frequencies occur because of the three orthonormal
triad vectors. For a collinear (Ising) spin system there would be only two
resonances. The solutions of the eigenvalue equations for the (3) frequencies
are

and %= +(H- F) + ;[(H+ Jy + 4(:)]’ (3.71)

Y
where y is the gyromagnetic ratio. Taking the simplest case of h4, = 0, i.e.
ZFC, the two 0% equations reduce to

(3.72)

if H 4 KJx. The first ESR experiments, shown for this special case in
Fig. 3.43, gives good agreement for the o+ mode (linear slope = $ and
expected intercept) even with resonance fields as low as 300-400 gauss. The
predicted situation for the three modes develops according to Fig. 3.44,
where the longitudinal mode oL remains at a tied frequency (no dispersion),
thereby making it difficult to find. This mode was finally detected indirectly
110 Basic spin-glass phenomenon

3300 - .i
-
t . .
U-J
: 3200 - .* /
L9 .
h
/
‘47 - aH, + Hi
2
a = 0.51 f 0.02
Hi = 3023 i IO
3000 -

2900 I I I I 1
0 200 400 600 800 1000
Field for resonance, H, (gauss)-,

Fig. 3.43 ESR (spectrometer frequency versus field for resonance) for a ZFC spin
glass, the resonance corresponds to the w + mode according to the equation given
in the text. From Schultz et al. (1980).

Fig. 3.44 A plot of the three expected resonance modes as a function of H for
zero remanence.

by using the mode-crossing repulsion phenomenon which divaricates two

distinct frequencies for a given field-for-resonance at a particular rotation
angle of M,. There were several experimental attempts to verify the various
angular dependences of the different ESR models. Reasonable results in
3.4 Spin glasses in a field 111

accord with hydrodynamic theory were mostly gathered, provided the spin
system rotated bodily, that is, stayed rigid. At large angles without a proper
preparation the CuMn spin system usually broke apart and along with it
the hydrodynamic description. The crowning achievement of all this low-
temperature effort was to demonstrate the co-operative nature of the spin-
glass phase. Clearly a new and unique state exists in the frozen spin glasses,
so we can conclude that there must be a type of phase transition which
transforms the system from the paramagnetic state at high temperatures to
the collective entity at T 4 Tf. And this state cannot be a mere assortment
of superparamagnetic particles or domains.

3.4 SPIN GLASSES IN A FIELD

Now that we have surveyed many of the important experimental results in

the three temperature regimes of a spin glass, we should briefly explore
what happens to a spin glass in a large applied field. The only regime of
real interest is that spanning T,, since not much will be different at high
temperatures - only more deviations from paramagnetism and we have
already considered in the previous section the effects of a large field on the
frozen state. While there are a certain group of theoretical predictions
which give an H-T, phase diagram, let us first view the problem
experimentally.
We know already from the ac-susceptibility that the cusp is smeared out
even in a small field. The broad maximum of the specific heat above Tf is
made even less visible by the field-generated spreading. The various
relaxation times just below Tf seem to increase and become faster as a
larger field is applied and the system seeks its new equilibrium state. One
conclusion of these experiments would be that the field removes the
criticality of the phase transition, yet it does not fully prevent the formation
of the frozen state, although with a useful handle - the remanent
magnetization. This means there are strong driving forces, e.g. competing
exchange and random anisotropy, whose average energies are greater than
that created by the field. Nevertheless, when the phase transition is trying
to take place at Tf the field can have its potent effect.
In order to avoid all the complications with relaxation times and
irreversibilities, a good experiment to perform is that of FC. We saw in
section 3.2.1 that the FC magnetization displayed a clear kink in the plateau
at Tf and time dependences were minimal. For certain of the canonical spin
glass even a small peak develops with FC at Tf. In Fig. 3.45 a plot is given
of the inverse magnetization divided by FC field (HIM) versus the
temperature. At small fields < 100 gauss the inverted cusp denotes Tf
(similar value to that obtained by xac). However, note how the reverse
peak rapidly disappears with increasing field. Now we must employ another
criterion to determine T,, for example, the onset of the plateau, as
112 Basic spin-glass phenomenon

I5 - kilogauss

1 _

0.5- \ Tg = 37.4 K
-L T
o- 0 .l,. . x
10 20 30 40 50 60
8 1 I . , I
0 10 2o 3o 4o 50
T(K)
Fig. 3.45 Inverse of the FC-susceptibility (HIM) for AgMn (10.6 at. %) as a
function of temperature for various magnetic fields (indicated on each curve in
gauss). Data were obtained by slow cooling in constant applied field. The onset of
the ‘plateau’ (marked by arrows) is taken arbitrarily as the point of the M(T) curve
departing by 3% from its low-temperature value thereby defining Tf. The resulting
boundary of the spin-glass phase H,(T) is shown as an inset. From Monod and
Bouchiat (1982).

represented by the arrows in Fig. 3.45. These not only shift downward in
T with increasing H, but become greatly smeared, an estimate of which is
given in the figure’s inset where an H-T phase diagram is attempted, albeit
with large error bars. At larger fields Tf is simply not well defined.
If we use other measurement techniques, first of all there are significant
differences in establishing an H(T,) line between the methods and analytic
procedures. And secondly, the time scale of the measurement enters and
gives different forms for H( Tf) depending on the experimental time window.
Hence, the dynamics of the transition are playing the more important role
3.5 Experimentalids (intuitive) picture 113

and a static H( Tf) phase diagram seems impossible to generate based upon
general experimental criterion. When the strong influence of dynamics is
coupled with the large smearing effects on a so-called critical temperature,
the whole meaning of a line of phase transitions created by a field becomes
moot.

3.5 EXPERIMENTALIST’S (INTUITIVE) PICTURE

We summarize this lengthy chapter with an experimentalist’s model. Or, in

other words, how a spin-glass researcher, who has never had any contact
with the theory, would interpret his measurements. Two important concepts
have emerged and been repeatedly employed in our discussions of the spin-
glass experiments. The first is the local correlations or ‘clusters’ which are
detected at high temperatures (T > Tf). These represent the ‘building
blocks’ of spin for the spin-glass state.. The second is our need for not just
a single relaxation time, but a broad distribution of such times to describe
the measurements. Both of these concepts go hand-in-hand, for, big clusters
have long relaxation times.
Our simple intuitive picture of a spin glass is as follows. There exists
compelling experimental evidence that for T > Tf short-range magnetic
interactions are present. Thus we divide the system up into dynamical,
evolving magnetic clusters (the building blocks) which develop out of the
high-temperature paramagnetic collection of spins. A competition occurs
between the short-range exchanges J(r& (where rii is the separation between
two spins) and the disordering effect of temperature kBT. When J(r& > kBT
for a given group of spins a cluster is formed. The clusters need not be
fully ferromagnetic entities, they can be mostly random with a dn small
net moment. Yet, they possess a correlation length (so which rigidly couples
together two or more spins at ‘ij of arbitrary orientation. As T is decreased
the clusters will grow in size (tso increases) and take on diverse shapes
based upon the distribution of competitive interactions among the various
spins in its immediate neighbourhood. As an illustration of this cluster
behaviour, the results of a computer simulation using the J(rij) versus kBT
model are shown in Fig. 3.46. Note the complex structures which are built
via the various range contours. There is even an isolated spin that does not
belong to any cluster in the lower right of Fig. 3.46.
Concomitant to this spatial magnetic clustering will be the temporal
relaxation rate UT. We can easily describe a collection of isolated non-
interacting spins in terms of a single Debye lifetime. But our spin-glass
situation is more complex, the various sized clusters all have different
relaxation times. And clearly the temperature evolution will be towards
longer and longer time scales as T is reduced to Tf. Nevertheless, our
isolated spin in Fig. 3.46 will still keep its orientation fluctuating at a rapid
rate, only when this spin finally joins a cluster does its turning rate slow
114 Basic spin-glass phenomenon

Fig. 3.46 Computer simulation for a random lattice of points (spins) with the
contours and their overlap representing the different ranges of spin-spin interactions.
From Verbeek et al. (1980).

down, Hence, there is a natural emphasis on the dynamics especially near

Tf which results in a wide distribution of relaxation times shifting with
temperature.
Since these time dependences are so intimately connected to the
spatially rigidity, the experimental model requires both and the particular
measurement techniques will highlight one or the other of the correlations.
As T + T,, the random anisotropy seems to take hold and preferred
orientations are established throughout the crystal. This anisotropy creates
the random-freezing directions. At Tf a sort of percolation ensues which
generates an infinite cluster of rigidly frozen spins. The infinite cluster
comprises many, many much smaller clusters that are randomly frozen in
orientations. Spins inside the infinite cluster are correlated, but they have
their local direction governed by the random shape or exchange anisotropy
of the smaller clusters. Therefore, the small clusters retain their original
identity, however, they are no longer able to react to an external field,
since they are firmly embedded or fied within the infinite cluster. In
general, for a percolation transition the order parameter is represented by
the fraction of spins belonging to the infinite cluster. Consequently, at
T < Tf there are still many free spins or small clusters which behave
superparamagnetically. These latter entities contribute a fast-relaxation
component and cause the wide distributions of relaxation times to be also
3.5 Experimentalids (intuitive) picture 115

observed below Tf. So the freezing process in our depiction is a percolation-

like one, but the infinite cluster is composed of many randomly frozen
smaller clusters. Quite an unusual ‘mosaic-fractal’ picture is required for
the spin glasses in order to elucidate the preceding experimental results.
Such a phenomenological description may be carried down to the lowest
temperature and used to explain the collective and co-operative properties
where the anisotropy is all important. In addition, the various irreversibilities
and metastabilities can occur in an inherent way within the inhomogeneously
percolated state. Here thermal activation is sufficient to flip some special
clusters or possibly tunnelling processes reorient other cluster configurations.
Energy barriers are a natural consequence of our model and this concept
nicely describes the remanence, hysteresis and long-time relaxations.
Spatial inhomogeneities and a vast spectrum of relaxation times both
over a wide temperature range are the experimental conclusions. Our
special percolation model seems to accommodate these deductions in a
reasonable qualitative description. In such a model the freezing temperature
is a function of the time or frequency of measurement. Since the infinite
cluster is a dynamical entity consisting of weak and strong links with many
fluctuations and excitations, it will appear percolated according to the
particular time window in which it is viewed. Thus, with a very rapid
measurement time the slowly occurring breaks in the m-cluster will not be
observed and Tf will be discerned at a higher temperature than by slower
measurements. We can represent this schematically according to Fig. 3.47.
In the limit of very long times of observation, a single Tfo or critical
temperature T, should be measured without any time dependences.

Tfo- -_________ ---------- -----

/s
I I I I I I I I
10-g 1o-6 1o-3 loo 1o+3
Fig. 3.47 Schematic representation of the freezing temperature as determined by
different measurement techniques having different ‘time constants’ of measurement.
116 Basic spin-glass phenomenon

This then would indicate that an equilibrium state has been reached.
Experimentally, it is very difficult to know if Tfo has been attained, one
can always try for another decade of time, and see if there is a small
change in a measurable quantity. Most likely, there is and this then
makes the question of an equilibrium phase transition unresolvable. How
can we rid ourselves of these dynamical processes to determine the basic
properties of the underlying phase transition? At present an answer has
not yet appeared.

Alloul, H. (1983) Heidelberg colloquium on spin glasses, Vol. 192 Lecture Notes in
Physics (eds J. L. van Hemmen and I. Morgenstern), Springer, Berlin, 18.
Brodale, G. E., Fisher, R. A., Fogle, W. E., Philips, N. E. and van Curen, J.
(1983) J. Magn. Mugn. Muter., 31-34, 1331.
Cable, J. W., Werner, S. A., Felcher, G. P. and Wakabayshi, N. (1984) Phys.
Rev. B, 29, 1268.
Cannella, V. and Mydosh, 3. A. (1972) Phys. Rev. B. 6, 4220.
Fert, A. and Hippert, F. (1982) Phys. Rev. Lett., 49, 1508.
Ford, P. 3. and Mydosh, J. A. (1976) Phys. Rev. B, 14, 2057.
Htiser, D., Wenger, L. E., van Duyneveldt, A. J. and Mydosh, J. A. (1983) Phys.
Rev. B, 27, 3100.
Levy, L. P. (1988) Phys. Rev B, 38, 4983.
Martin, D. L. (1980) Phys. Rev. B, 21, 1902.
Meschede, 0.) Steglich, F., Felsch, W., Maletta, H. and Zinn, W. (1980) Phys.
Rev. Lett., 44, 102.
Monod, P., PrCjean, J. J. and Tissier, B. (1979) J. Applied Phys., 50, 7324.
Monod, P. and Bouchiat, H. (1982) J. Phys (Paris) Len., 43, 145.
Morgownik, A. F. J. and Mydosh, J. A. (1982) Physica, 107B, 305.
Morgownik, A. F. J. and Mydosh, J. A. (1981) Phys. Rev. B, 24, 5277.
Morgownik, A. F. 3. and Mydosh, J. A. (1983) Solid State Commun., 47, 321.
Mulder, C. A. M., van Duyneveldt, A. J. and Mydosh, J. A. (1981) Phys. Rev.
B, 23, 1384.
Murani, A. P. (1978 a) Phys. Rev. Lett., 41, 1406.
Murani, A. P. (1978 b) .I. Appl. Phys., 49, 1607.
Murani, A. P. (1981) J. Mugn. Magn. Mater., 22, 271.
Mydosh, J. A. and Ford, P. J. (1974) Phys. Lett., 49A, 189.
Mydosh, J. A. (1988) J. Magn. Magn. Mater., 73, 247.
Nagata, S., Keesom, P. H. and Harrison, H. R. (1979) Phys. Rev. B, 19, 1633.
Nordblad, P., Svedlindh, P., Lundgren, L. and Sandlund, L. (1986) Phys. Rev. B,
33, 645.
Sandlund, L., Svedlindh, P., Granberg, P., Nordblad, P. and Lundgren, L. (1988)
J. Appl. Phys., 64, 5616.
Schultz, S., Gullikson, E. M., Fredkin, D. R. and Tovar, M. (1980) Phys. Rev.
Lett., 45, 1508.
Smit, J. J., Nieuwenhuys, G. J. and de Jongh, L. J. (1979) Solid State Commun.,
31, 265.
Tholence, J. L. and Toumier, R. (1974) J. Phys. (Paris), 35, C4-229.
Uemura, Y. U., Yamazaki, T., Harshman, D. R., Senba, M. and Ansaldo, E. J.
(1985) Phys. Rev. B, 31, 546.
References 117

Verbeek, B. H., Nieuwenhuys, G. J., Mydosh, J. A., van Dijk, C. and Rainford,
R. D. (1980) Phys. Rev. B, 22, 5426.
Violet, C. E. and Borg, R. J. (1966) Phys. Rev., 149, 545.
Wenger, L. E. and Keesom, P. H. (1976) Phys. Rev. B, 13, 4053.
Systems of spin glasses

For the sake of generality and completeness we should try at this stage to
collect our various systems of spin glasses. The previous chapters have, in
their generic approach, dwelt on a few very specific materials, e.g. CuMn,
AuFe and to a lesser extent (EuSr)S. Yet the spin-glass phase is, as stated
previously, a very general phenomenon. Over 500 different systems have
been claimed to be spin glasses, i.e., exhibit at least some (even one) of
the experimental characteristics mentioned before for the canonical (RKKY)
spin glasses. After ferro- and antiferromagnetism, spin-glass freezing is the
third type of magnetic ‘order’ or, better put for the latter case, co-operative
magnetic state.
While it is not the purpose of this chapter to consider all 500-plus
materials, nevertheless, we wish to list the most important spin-glass
categories and to distinguish their class according to the strength of the
magnetic interactions which are present. Remember the conduction-electron-
mediated RKKY interaction of the transition-metal impurities have the
strongest coupling. Oppositely, superparamagnets (rock magnets or Co0
particles) have little or no coupling and are, therefore, not spin glasses.
We begin with the transition-metal-solute alloys, work our way through
the rare-earth binary or pseudo-binary alloys and add some of the vast
number of multi-component, amorphous magnetic alloys. Then we move
onto the semiconducting materials for which there is an enhanced superex-
change. Such a coupling may be sturdy enough to create a good spin glass
at a reasonable temperature. And finally, we reach the magnetic insulators
where the superexchange is weak and, thus, the concentration must be
increased to promote a possible spin-glass state. Or, if the coupling is
insufficient, we return to our (super) paramagnetic, at least, down to the
lowest available temperature of measurement (see section 2.10). Here the
blocking of individual spins or clusters at T, would take precedent over a
cooperative and collectively frozen ground state which might begin to form
at Tf < TB.
4.1 Transition-metal solutes 119

4.1 TRANSITION-METAL SOLUTES

4.1.1 Noble metals with transition-metal impurities

We start with the well-studied noble-metal/3d transition metal alloys, our

canonical spin glasses. Since 4d or 5d transition metals are non-magnetic,
i.e., do not form local moments, they cannot be used as impurities in a
noble metal to form a spin glass. Two criteria can be invoked to find the
simplest spin-glass behaviour:

(a) ‘good’ moment systems, meaning that the Kondo temperature should
be less than = 1 K, so that no complications are encountered with
weakening of the local moments at low temperatures; and
(b) a favourable solubility such that at least 10 at. % of the 3d metal may
be dissolved in the noble-metal host.

This latter criterion, in conjunction with a proper homogenization process,

provides for a random distribution of impurities and eliminates difficulties
with chemical clustering. In Table 4.1 the various combinations of noble-
metal solvent 3d solute are given. Besides the archetypal examples of CuMn
and AuFe, there are only three other uncomplicated spin-glass systems,
denoted by a ‘good’ in Table 4.1: A&r, AuMn and AgMn. By referring
to the table, problems are encountered for many of the other combinations,
either with the Kondo temperature or especially with the solubility limit,
represented by XT and XS, respectively.
However, a number of special cases exist as, for example, with CuFe
which possesses a rather poor solubility and a TK = 30 K but, nonetheless,
shows the magnetization characteristics of a mictomagnetic. The high TK
means that isolated or single Fe atoms cannot participate in the magnetic
interactions, yet Fe pairs or triplets may. The deviations from randomness
(precipitation of Fe for x 2 1 at. %) are reflected in the low-field suscepti-
bility by the expected sharp peaks at Tf changing into broad maxima. Also

Table 4.1 Spin glass combinations

Noble metal-transition metal

Host Impurity: V Cr Mn Fe co Ni

cu xs xs GOOD XS+T XS XT
& xs xs GOOD XS xs xs
Au XT GOOD GOOD GOOD XS+T XT

‘GOOD’ represents the most favourable combinations, XS or XT means that the

spin glassbehaviour is limited by lack of solubility or too high a Kondo (fluctuation)
temperature, respectively.
120 Systems of spin glasses

in this special subset we should include AuCo for which the Co solubility
is poor, but not as bad as CuFe. The difficulty here is with the very large,
few hundred kelvin, Kondo (or spin-fluctuation) temperature. As before
TK is strongly dependent upon the local environment, i.e., the number of
Co nearest neighbours (nn). In order to reduce TK to less than 1 K, a Co
triplet is required and then modified spin-glass behaviour is observed. For
this system an effective magnetic concentration of Co-nn triplets should be
used instead of the actual atomic percent.
A few non-noble-metal host systems may be grouped with the above
collection. Examples of which include ZnMn and C&In, whose crystal
structure, being hexagonal, adds a crystalline single-site anisotropy to restrict
the orientations of the freezing spins. This breaks the isotropic distribution
of frozen-spin directions discussed in section 3.3.5 and creates Ising
(up/down) or x-y (spins confined to lie in a plane) types of spin glasses.
Nevertheless, for these systems the solubility of Mn in Zn or Cd is rather
limited, in the thousand ppm range. A final extreme example of a local-
environment spin glass is Feo.3Al,,7 where the very large Fe concentration
is necessary to make the Fe magnetic, since Fe in Al has a very high spin-
fluctuation temperature of some thousands of degrees. Note that in all of
the above systems the competing ferro-/antiferromagnetic RKKY interaction
is active and results in a strong and oscillating exchange coupling.

4.12 Transition-metal/transition-metal combinations

By using the previous criteria of good moments and high solubility, binary
combinations of two transition metals may be used to create a spin glass.
That is, if a giant-moment system does not sweep away the random freezing
and leave a dilute, but long-range ordered, ferromagnet. Table 4.2 offers
a collection of the various possibilities. We begin with the giant-moment
alloys whose formation requires an exchange-enhanced host (see section
2.3). This means that, firstly, there is an abundant density of itinerant (d-

Table 4.2 Transition metal-transition metal spin glass/giant moment combinations

Host Impurity: Cr Mn Fe co

MO XT SG SG XS+T simple
Rh XT SG XT XT
Pd XT GM+SG GM GM exchange
Pt XT SG GM SG enhanced

SG and GM represent favourable combinations for spin glass or giant moment

behaviour. Note the strong, mixed behaviour of PdMn, XS or XT means solubility
or too high a Kondo (fluctuation) temperature limits the appearance of both the
spin glass or giant moment states.
4.1 Transition-metal solutes 121

states near the Fermi energy N(Er) such that the Pauli susceptibility is
large: x,, = 2&N(&). And secondly, an intra-atomic exchange, 1, should
be present on the host sites. This results in an exchanged-enhanced
susceptibility x = x&l - IN(E,)]-1 where the factor [l - IIV(E,)-1 = 0 is
known as the Stoner enhancement factor. For Pd, 0 = 10, while for Pt,
0 = 3. Consequently, Pd or Pt hosts alone will produce the giant-moment
polarization. As mentioned previously (section 2.3) only in the very dilute
limit of Fe impurities (ppm) in Pd will a spin-glass state form at very low
temperatures (< 1 mK).
An especially interesting situation is with PdMn where for x < 3 at. %
Mn, the giant-moment ferromagnetism prevails. However, upon further
increasing the Mn concentration (X > 4 at. %), the probability of having
two Mn atoms as first or second nearest neighbours increases. This then
supplies the essential element of competing exchange for the appearance of
the spin-glass phase. Since Mn-nn couple antiparallel, they thereby
produce the antiferromagnetic coupling which mixes into the longer-range
ferromagnetic, giant-moment polarization. By making a ternary alloy
PdMnFe and utilizing different concentrations of Mn and Fe we can
independently vary the ratio of ferromagnetic to antiferromagnetic exchange.
This procedure has become very important in comparing with theory and
studying the re-entry spin glasses of section 2.9.
Cr impurities, when introduced into the Pd or Pt host, locally ‘blot-out’
the uniform exchange enhancement, but in this process the Cr loses its
magnetic moment and becomes a weak moment with a high TK. Only when
the Cr concentration is large enough to provide a suitable Cr local
environment does a stable (or good) Cr moment appear, now however, in
a non-enhanced Pd or Pt matrix. Accordingly, the RKKY-oscillating
interaction can occur leading to a spin-glass state at low temperature. Here
there is no giant-moment phase, if anything, the term ‘dwarf’ or ‘destroyed’
moment is more applicable. CrFe, a combination of two magnetic elements,
represents another exceptional case. Fe when diluted into the Cr-bee host
first eradicates the antiferromagnetic (itinerant) spin-density wave. Yet, the
Fe moments remain intact, and again, due to RKKY interaction a spin-
glass phase appears around 15 at. % Fe. At slightly lower concentrations,
there is a re-entry transition back to the spin-density-wave antiferromagne-
tism. At slightly higher Fe concentrations, a series of transitions occur:
paramagnetic * ferromagnetic + (re-entrant) spin glass.
At this point in our survey of binary, crystalline alloys we should also
mention the various Ni-solute systems. Either in noble-metal or transition-
metal matrix, Ni moments require a local environment of other Ni nearest
neighbours. For PdNi three or more nearest neighbours are needed to
produce a giant-moment ferromagnet. For CuNi at least seven or eight
nearest neighbours give a narrow spin-glass regime before ferromagnetism
takes over. The all-important concept, used time and time again for moment
formation, is the local-environment model that a particular number of weak-
122 Systems of spin glasses

moment atoms are present as nn to produce a sufficiently strong magnetic

surrounding and thereby generate the local magnetic moment which is per
se a magnetic cluster.

4.1.3 Collection of characteristic temperatures for AuFe and CuMn

We now return to our two most-popular canonical spin glasses, AuFe and
CuMn. There has been an enormous amount of effort put into studying
these materials and it seems appropriate to gather the various experimental
results in a temperature-concentration phase diagram. According to Larsen
(1978), different characteristic temperatures are defined in the following
way:

(a) the freezing or ordering temperature usually defined by a susceptibility

cusp or for a ferromagnet the ‘demagnetizing-limit’ plateau;
(b) the temperature T, of the maximum in the magnetic resistivity Ap,
see equation (3.11);
(c) freezing or ordering temperatures derived from other experimental
techniques with different time windows such as Mossbauer spectroscopy
and neutron scattering; and
(d) a ‘noise’ temperature A, calculated from A, = T,Q[ln (T,/TK)] where
TK is the Kondo temperature and Q[.] is the noise function derived
by theory.

Figures 4.1 and 4.2 display the many data points collected over the years
for AuFe and CuMn, respectively. The labelling corresponds to the above
definitions while the solid lines represent Larsen’s theoretical calculation of
the freezing T&x). Here the model relates Tf to the root-mean-square
RKKY interaction energy. Exponential damping of the RKKY interaction
is included to take into account a finite mean-free path via a parameter
r(x). Some salient features of these data, which extend over five decades
in concentration, are listed below.

The magnetic-resistivity maximum always occurs at a significantly larger

temperature than the freezing temperature.
Within the limits of a log-log plot there is a reasonable agreement
between the various experimental determinations of critical (freezing or
ordering) temperatures.
The noise model can nicely relate the temperature of the resistivity
maximum to the freezing temperature, except in the l-10 at. %
concentration regime of AuFe.
While CuMn continues to follow a smooth curve over its entire
concentration range, AuFe has the above-mentioned anomaly, followed
at larger x by an upward jump in the ordering temperature beginning
4.1 Transition-metal solutes 123

Au c

.- Irn

bO
0
01
0

0
/

Fig. 4.1 Collection of characteristic temperatures for AuFe alloys over the complete
concentration range; from Larsen (1978).

at about 15 at. %. This deviation corresponds to the onset of percolation

ferromagnetism. The precursor effect is probably due to the Apmax being
shifted to much larger temperatures by the formation of big ferromagnetic
clusters.
5 The theoretical calculation of Tf from the RKKY energy seems to work
well with the proper choice r(x), the damping parameter. It looks like
this parameter must be made a function of the concentration above
about 10 at. % for both systems in order to maintain the fit.

The important conclusion of all this data plotting and theoretical analysis
is that the RKKY-interaction is the driving force behind the spin-glass state
of these many transition-metal alloys. The coupling strength is large, its
spatial oscillations create the mixed couplings of the randomly distributed
124 Systems of spin glasses

""1
IO4 10-5 10.' 10-a 10‘) 10-l 1
c

Fig. 4.2 Collection of characteristic temperatures for CuMn alloys over the complete
concentration range; from Larsen (1978).

impurities, and frustration is a natural consequence. So we have a simple

recipe for creating a new spin glass: take a transition metal, alloy it into a
non-magnetic metal, and, with the caveat of insufficient solubility or too
high a Kondo temperature, we will discover another spin-glass material.
But what have we really learned by adding another spin glass to our long
list? Answer: not much about the essence of the phase transition or
transformation, how and out of what the frozen state is formed, and what
governs its metastable properties. Such crucial points have not been touched.
For we must work (measure everything) on a few ideal systems which
closely correspond to the theoretical models. Only then can we use the
microscopic theory to explain and delve deeper into the nature of the spin
glasses. And experiments on these sparsely chosen materials will point the
4.2 Some rare-earth combinations 125

way as to which model or solution is the proper and correct one. However,
in this chapter we push forward with our spin-glass menagerie.

4.2 SOME BARE-EARTH COMBINATIONS

The same game can be played with magnetic rare-earth elements: dilute
them into a non-magnetic host metal and let the RKKY interaction
perform its coupling. Nevertheless, there are some advantages and certain
disadvantages of the rare earths when compared to the transition-metal
solutes. First the good news, for most rare earths, with the notable exception
of Ce, there is no Kondo effect, because the coupling between local moment
and conduction electrons is a positive (or ferromagnetic) one. The rare-
earth elements, which do exhibit a Kondo effect (antiparallel coupling),
e.g. Ce, usually have a very low TK.
And now the bad news. In many cases limited solubility is a problem
one cannot simply alloy over a large concentration range, e.g. LaGd exists
as a solid solution for only a few percent. A way of overcoming this
difficulty is to use a pseudo-binary intermetallic compound, e.g. La,-,Gd,Al,.
Here there are a vast number of such combinations, a few of which are
listed in Table 4.3. Further bad news is that the strength of the RKKY,
i.e., its coefficient J, in equation (1.2), becomes much weaker for rare-
earth impurities. This smaller polarization of the conduction electrons results
in much lower T,s even at increased concentration of impurities. A final
complication with the rare earths is the crystalline-electric-field splittings
which are rather low in energy = 100 K. This means that the effective
moment will change as a function of temperature and Schottky anomalies
will influence the various experimental behaviours. On the positive side,
the single-ion crystal field may be used to generate a crystalline anisotropy.
Such, in turn, can create Ising (up-down) and x-y (spins confined to a
plane) types of spin glasses which should have different critical phenomenon

Table 4.3 Some good rare-earth spin-glass combinations

LaGd, LaEu
Binary
ScGd, ScTb, YGd, YTb and PrTb

Lal-,Gd,Alz, Lal-,CeA&
La,-,Gd,I,, La3-,Ce,I,
Lal-,GdxBb, Lal-,Ce,B,
Pseudo-binary
Lal-,Ce,Ruz, Th,-,Gd,Ru,
Cel-,GdxRu2, Ce,-,Tb,Ru*
Thl-,NdxRhz
126 Systems of spin ghses

associated with their freezing behaviour than our isotropic (Heisenberg),

transition-metal spin glasses.
After all these pluses and minuses of the rare earths are taken into
account, we have essentially similar ingredients and behaviours as with the
transition-metal systems. The common property here is the conducting host
and the RKKY interaction connecting the site-random impurities. We
might have to measure at lower temperatures or higher concentrations
(renormalized T/,X), but the basic effects will be the same. And we can
include via the above recipe many new systems to our ever expanding
collection of spin glasses.
A very interesting question arises concerning the mutual coexistence of
magnetism and superconductivity. This query has been the subject of much
effort for the past 30 years. The conventional wisdom is such that
superconductivity can coexist with all the various types of antiferromagnetic
long-range order (even spiral sorts with rather long periods). But ferromag-
netic order destroy the superconductivity. So what about a spin glass? After
a long search, finally a suitable rare-earth system, (Thr-,Nd,)Ru,, was
fabricated which was a good (T,S = 4 K) superconductor at x = 0 and a
ferromagnet (T,F = 22 K) at x = 1. For intermediate concentrations x = 0.35
a superconducting/spin glass developed. Studies of the critical fields for
superconductivity, and the freezing behaviour and remanence of the frozen
state proved that there was indeed a coexistence of these two states.
However, the spin-glass transition did weaken the superconductivity by
depressing T,S and reducing the critical fields. If Tf 5 Tz, a remarkable
quenching of the superconductivity occurs at Tf followed by a recovery or
‘re-entry superconductivity’ at very low temperatures. This unique behaviour
results in one spin-glass (or more accurately a cluster-glass) and three
superconducting-normal transitions at a single concentration (X = 0.35)
value. Figure 4.3 illustrates the exceptional phase diagram of Thl-,Nd,Ruz
where the coexistence regime is shaded. Thus, we conclude that a weakened
superconducting state can coexist within a frozen spin-glass state.

4.3 AMORPHOUS (METALLIC) SPIN GLASSES

As already mentioned in section 1.2 an amorphous solid, viz., an alloy or

an intermetallic compound without crystallographic order, can also be a
spin glass. In the cases considered below all the samples are metallic with
a reduced conductivity due to the disordered matrix. We represent such
materials with an ‘a’ before their elemental symbols. There are many
combinations of these systems starting with a single magnetic species of a
transition metal, e.g. a-Fe,Pda+,Pzo or a-Fe,Snr-,. Table 4.4 collects a
few additional examples. Notice how complex the element manipulation
and compositions can become in order to ‘fine tune’ the material. Then we
can proceed to more complicated double transition-metal glasses such as
4.3 Amorphous (metallic) spin glasses 127

0.1 0.2 0.3 0.4 0.5

X
Fig. 4.3 Superconducting (TJ and magnetic (T,,, = Tf) phase diagram for
ThI-,Nd,Ru*. The lines are a visual guide; from Hiiser et al. (1983).

Table 4.4 Some good amorphous metallic spin glasses

a-Fe,B,-,, a-Fe&,-,; a-FexZrI-,

a-Mn,Si,-,, a-Mn,GeI-,
Binary
a-Fe,Y1-,, a-Ni,Y,-,
a-Gdo.37A10.63

Pseudo-binary

a-(FexNil--x)79P13B8 or even a-(Fe,Ni1--x)75P16B6A13. These multi-element

compounds (see Table 4.4) formed by melt-quenching can be tailor-made
to give the desired amounts of ferro- and antiferromagnetic exchange, in both
ratio and absolute magnitudes. Hence, the spin-glass freezing temperature can
be adjusted to a convenient range and re-entry (ferromagnetism + spin
glass) behaviour can be extended over a wide composition region.
In addition, amorphous materials can be fashioned out of the rare-earth
elements: a-Gd3,Alh3 and a-Las,-,GdXAuz, are good examples. Particularly
with the latter system, a wide x-range of composition (a few percent up to
50 at. %) is available to trace the spin-glass phase with T,s spanning
= Z-50 K. We could easily employ melt-quenching or splat-cooling and
sputtering techniques to fabricate, at will, other rare-earth spin glasses. The
same caveats as for the crystalline alloys apply for the amorphous ones.
A special property of the amorphous spin glasses is their higher resistivity
128 Systems of spin glasses

which will dampen the range of the RKKY-interaction. Therefore, large

concentrations of the magnetic elements are required to form a spin glass.
This reduced interaction range places such system intermediate between the
short-range insulators and the long-range undamped RKKY alloys. From
the sharp, well-defined freezing temperatures of these amorphous com-
pounds, the range of the competing interactions does not seem to affect
the freezing characteristics.
In general, the spin-glass aspects of these amorphous magnetic materials
are mainly similar to those of their crystalline counterparts. One advantage
of using them is the vast region of magnetic concentration available in an
amorphous structure. This permits an examination of cluster glass and re-
entry (ferromagnet + spin glass) behaviours. Another benefit comes from
the damped RKKY interaction which allows the coupling strength and
range to be varied. A third is that any crystalline or single-ion anisotropy
will be averaged to zero, due to the disordered lattice, leaving a fully
isotropic (Heisenberg) spin glass. In summary, we would gather these
amorphous alloys into our universal class of metallic, competing exchange,
random-site spin glasses.

4.4 SEMICONDUCTING SPIN GLASSES

We commence this section with a most important family of magnetic

semiconductors, namely the rare-earth monochalcogenides. As discussed
earlier, Eu,Sr,-,S is a prime example of a short-range (only first and second
nearest-neighbour exchanges are important) Heisenberg spin glass. Let us
briefly consider the phase-diagram of this system which is reproduced in
Fig. 4.4. The J1 and J2 exchange couplings are competing, ferro- and
antiferromagnetic, respectively, with ratio J2/J1 = -0.5. Since the Eu sites
occupy an fee lattice, there will be a two neighbour percolation threshold
at x = 0.13. According to the phase diagram in Fig. 4.4, this is exactly
where the spin-glass phase ends. For x < 0.13 a sort of superparamagnetic
cluster blocking begins without any of the collectivity or co-operativeness
of inter-cluster coupling - a prerequisite for the spin-glass transition. This
behaviour nicely illustrates the need for mixed interactions and the
disadvantage of a short-range random site system which cuts off the spin
glass at the percolation limit.
Other members of this family include Err.&-,Te and Et.@-,Se which,
also due to the competing +Ji exchanges, show spin-glass-like characteristics.
An exception to this trend is Eu,Sri-,O where both J1 and J2 are
ferromagnetic. Hence no spin-glass phase occurs, only a ferromagnetic one
down to the percolation threshold. We should at this point mention the
superexchange mechanism in these chalcogenides. Theoretical work has
proposed a novel f-d overlap of wave functions between two Eu-ions, which
are either first or second nearest neighbour. For the closest Eu-Eu nearest
4.4 Semiconducting spin glasses 129

r; lo-

t"
F

5-

0 A
0 0.5 1.0
x

Fig. 4.4 Magnetic phase diagram of EuXSrl-,; from Maletta (1982).

neighbours a virtual transfer of a 4f electron to the 5d-t,, excited state of

its neighbour results in a strong ferromagnetic coupling. In contrast, an
indirect (or more typical) superexchange via the ligand (see section 1.3.3)
creates the next-nearest neighbour J2 which involves the antibonding 5d-e,
orbitals of the Eu and the p-orbitals of the ligand. Here the coupling is
either ferro- or antiferromagnetic depending on the distance.
Another interesting, but more troublesome, class of ‘dilute’ magnetic
semiconductors (DMS) are the II-VI group compounds (the II and VI refer
to columns of the periodic table). These include the sulfides, tellurides,
sellurides and arsenides of Zn, Cd, Hg, Ge, Pb, and Sn, the latter sites
being doped with Mn (or Fe) as the magnetic impurities. A prime example
is Cd,-,Mn,Te, others are listed in Table 4.5. So we have a random-site
problem and must consider the exchange coupling between the Mn2+ ions.
Again theoretical work has established that superexchange is the main
source of the magnetic coupling. Accordingly, the filled valence band of
the semiconductor exchanges electrons with the half-filled 3d band of the
Mn. With such 2-electron (or hole) processes the interaction between the
Mn will always lead to an antiparallel (AF) orientation and this seems to
be true for both first and second nearest neighbours. Although, the second
neighbour J2 is usually much weaker in magnitude, since J(r) falls off either
as exp(-r2) or llr7. Experiment seems to support these conclusions
concerning the exchange coupling. Therefore, we must pose the question
130 Systemsof spin glasses

Table 4.5 Some dilute magnetic semi-conductors (DMS) as

possible spin glass combinations

Zn,-,Mn,S, Znl-,Mn,Se; Zn,-,Mn,Te

Cdl-,Mn,S, Cdl-,Mn,Se; Cdl-,Mn,Te
Hg,-,Mn,Se; Hg,-,Mn,Te
Pbl-,Mn,S, Pb,-,Mn,Se; Pbl-,Mn,Te
(Znl-M&Ass (C4-,MnMs~
Zn,-,Fe,Se, Cdl-,Fe,Se; Hgl-,Fe,Se
Sn,-,Mn,Te; Gel-,Mn,Te

with these II-VI DMSs: can an all antiferromagnetic, random-site systems

produce a good spin glass? We let experiment be our guide.
Many measurements have been performed on the DMS systems listed in
Table 4.5. The following characteristics are generally found. A Curie-Weiss
(CW) law is obeyed at high temperatures with a paramagnetic temperature
indicative of the predominantly antiferromagnetic coupling. Deviations from
CW occur at low temperatures depending on the Mn concentration usually
resulting in smooth kink or rounded plateau. But there is no cusp or sharp
maximum except at large Mn-x values (x 2 0.4). A difference in FC versus
ZFC magnetization is also observed, however, this is very gradual without
the sharp features of the canonical (RKKY) spin glasses. The magnetic
specific heat exhibits a broad maximum shifting to higher temperatures with
X. Attempts at scaling the static and dynamical behaviour of such DMSs
via x’(o) and x”( o ) measurements give roughly similar results and exponents
with respect to the canonical or even ideal (Chapter 6) spin glasses. In spite
of this there remain the troublesome features referred to earlier, viz., the
lack of sharpness in the ac-susceptibility and magnetization. It would seem
tha the freezing behaviour is non-co-operative at low x-values. A significant
po Ii ion of the spins remains independent and only joins the randomly
frozen, infinite cluster at a much lower temperature than the ‘kink
temperature’. This absence of finely-defined freezing temperature is probably
caused by the short-range nature of the exchange interaction J(r) and the
enormous frustration present. Too much frustration appears not to be good
for a co-operative spin-glass freezing. With most two or three nearest
neighbours participating in the coupling scheme, the percolation cut-off
should be around 10%. For x < 0.1 only superparamagnetic clusters would
become blocked as in the case (Eu,Sr,-JS. However, for the DMS this
seems not to occur around 10% but around 40%. Is the frustration taking
its toll? Furthermore, the exclusively antiferromagnetic alignment makes
the susceptibility and magnetization very small. With such a weak magnetic
response, experiment will have difficulty tracking the subtleties of freezing
or blocking. Thus, a major problem may be unresolvable, i.e., how
to distinguish an all-antiferromagnetic spin glass from its two possible
counterparts: a ? J spin glass or a non-interacting ‘super’-paramagnet. More
4.5 Insulating spin glasses 131

work is needed to sort out these nuances and to arrive at the fine (or
significant?) distinctions between a competing or mixed-interaction spin
glass and an only antiferromagnetic one. As we shall soon see theory only
treats the mixed-interacting spin glasses.

4.5 INSULATING SPIN GLASSES

Chemical compounds with one magnetic element have long been a testing
ground for magnetic critical phenomenon. The chemists can almost at will
produce a material which has the desired spatial (l-, 2- and 3D) and
order parameter (Ising, x-y and Heisenberg) dimensionalities. Exchange
interactions can be varied (ferro, mixed and antiferromagnetic) with
superexchange and sometimes dipolar mechanisms. In the 1960s and 1970s
such pure compounds dominated the subject of magnetic phase transitions.
For the chemist there is no great difficulty in randomly replacing a
magnetic element by a non-magnetic one. Thus, in principle, we can create
multitude spin glasses. However, what is to be gained by discovering yet
another spin glass? There must be something particularly interesting or
‘ideal’ about the given material. As a case in point let us consider the
mixed compound Fe,Mg,-,Clz. Pure FeCl, is a hexagonal-lattice layered
compound which has ferromagnetic a-b planes (with a triangular lattice)
that stack antiferromagnetically along the c-axis; for this reason we have a
long-range ordered antiferromagnet. A strong uniaxial anisotropy aligns the
moments along the c-axis so the material is Ising-like. The in-plane
interactions are ferromagnetic for nearest neighbours, but antiferromagnetic
for next-nearest neighbours (nnn) with J2/J1 = -0.13. In diluted Fe,Mg,_,Cl,
one expects for x I 0.4 a random site, 3D, Ising spin glass which should
evolve out of an ordered antiferromagnetic structure. Very interesting here
is the coexistence (at intermediate x = 0.45), proven by neutron-scattering
experiments, of an infinite antiferromagnetic network with a collection of
randomly frozen spins in the various a-b planes. At lower X, until two-
neighbour percolation cuts off the spin-glass phase, the usual competing-
interaction spin-glass behaviour is observed, but now in an Ising system,
i.e., predominantly along the c-axis are there magnetic responses (freezing).
Hence, this material is a good candidate for studying the dynamics of an
Ising random-site spin-glass.
Another example is the Fe,Zn,-,F,, also an antiferromagnetic Ising
system, which for x 2 0.5 represents a prototypical 3D random-exchange
Ising model (REIM). In the presence of a field applied along the Ising axis
it exhibits crossover phenomenon to random-field Ising-model (RFIM)
behaviour. The above two models (REIM and RFIM) represent unusual
critical phenomenon and crossovers of different types of phase transitions,
albeit both to a long-range-ordered state which will be discussed in Chapter
7. Now what happens when the concentration is reduced? Before attempting
132 Systems of spin glasses

to answer this question we must first consider the superexchange present

in the body-centre-tetragonal (rutile) FeF2 structure. Best determinations for
these parameters give zlJl = +0.14 K, .z2J2 = -42 K, and z3J3 = -1.15 K
where Zi is the coordination number: for the ith nn, z1 = 2, z2 = 8 and
z3 = 4. If we neglect the tiny ferromagnetic J1, then the ratio of z2J2/z3J3
is about 40 and this small J3 exchange is the only source of frustration. The
site-percolation threshold for the Fe,Zni-,Fz system with nnn interactions
is xp = 0.24. So below this concentration we would expect our old nemesis,
namely, independent antiferromagnetic domains gradually becoming blocked
at low temperatures in analogue to super-paramagnetism. But what about
concentrations slightly above xp, e.g. = 0.3? At present there is much
interest in this query. Can a spin glass be formed without frustration? Some
experimental indications exist for spin-glass behaviour, but others seem to
suggest a non-co-operative freezing more like the above blocking of discrete
domains. Physically, if no other exchange were present except J2, then we
would anticipate a long-range, percolative antiferromagnetic transition at
some low temperature. And neutron scattering with the observable Bragg
peaks has shown this long-range order at x = 0.31. Notice that this situation
is somewhat similar to that discussed previously (section 4.4) with our all-
antiferromagnetic dilute magnetic semiconductors. However, for Fe,Zni-,Fz
there is only ooze predominant superexchange coupling, not two. Again,
with better samples, time and experiment will tell us the correct answer as
to the low-temperature phase.
We could continue further with the many additional combinations of
mixed or random compounds which are insulating and have well-defined
superexchange. Such will certainly increase the generality of our spin-glass
phenomenon, but with our chemists able to supply ‘custom-made’ compounds,
we should restrict ourselves to truly ideal or model spin glasses. This topic
will be dealt with in Chapter 6 once we have explained the various
theoretical models for spin glasses. Only then can we judge what real
materials come closest to the theories. And this will afford us the luxury
of making a direct and meaningful experimental-theoretical comparison
without the numerous complications and approximation of the many, many
non-ideal spin glasses.
In concluding this section we must mention the sundry magnetic insulators
which exhibit little or no exchange or dipolar coupling. Most of these
superparamagnets are distinguished by their peculiar chemical structure.
For instance, rock magnets are tiny magnetic particles diffusely dispersed
within a non-magnetic ‘rock. An arrangement of Co0 particles sufficiently
spaced is another example; ferrofluids, aerosol particles, nanocrystals, etc.
are more possibilities. In what follows we touch upon two specific
exemplifications.
An insulating helium-borate glass has been intensively studied via ac-
susceptibility and its frequency and temperature dependences (see Table
3.1). The resulting behaviour was fully describable within the NCel
4.6 What is a good spin glass? 133

superparamagnetic model using a gaussian distribution of relaxation times.

On the other hand, when similar measurements were performed on cobalt-
aluminosilicate glasses definite indications were found for interaction effects.
For example, as the temperature is lowered there was a more rapid shift
of the (average) relaxation times than simple Arrhenius activation. Yet,
when these relaxation-time shifts were compared to CuMn et al. as in Table
3.1, they were not nearly as swift. So we conclude that the helium-borate
glass is essentially non-interacting, while the cobalt-aluminosilicate glass is
a weakly coupled system. We remove these two glasses and similar, non or
flimsily interacting, spin systems from further consideration in this treatise.

4.6 WHAT IS A GOOD SPIN GLASS?

We revisit this question and try to pose an answer with respect to the
different amounts of magnetic coupling between the randomly distributed
magnetic entities. Table 4.6 collects, in summary version, the results of the
Table 4.6 Examples of differently coupled random magnetic systems

Strong Medium Weak None

TM I
II ,6
.u_
= z
2 RE I z
{
2
<
Amor. I ; I

2 Semi. cn
0
t;
2 Insul.
5
s Part.
z
0
Exchange or dipolar coupling

Exchange or dipolar coupling

TM = CuMn, AuFe, PtMn, etc.
RE = LaGd, (YGd)Al*, (CeGd)Ru*, etc.
Amor. = a-(FePd)80 PzO,a-FeSn etc.
Semi. = (EuSr)S, CdMnTe etc.
Insul. = (MgFe)Cl*, ZnFeF2, a-CoO.Al203Si02 etc.
Part. = COO, a-(Ho20,)(B,0,), etc.
134 Systemsof spin glasses

previous five subsections. A rough estimation is given for the range of the
exchange coupling in the different classes of materials which we have
encountered. The interactions are basically RKKY versus superexchange.
We list the material entries according to the two types of magnetic elements,
3d and 4f, that go into the RKKY random alloys. Here the 3d-impurities
are noticeably stronger than the 4f ones. The amorphous magnetic metals
follow with their similar expanse of interaction strengths. Then come the
not-so-dilute magnetic semiconductors with a more narrow range comparable
to the rare earths. The insulators appear next, occasionally the superexchange
can be rather large, but it is too often only antiferromagnetic. Finally, we
have the particles or glasses, which possess little or no coupling energy and
thus can be ignored with respect to their spin-glass properties.
So once again a spin glass is a random, mixed-interacting, with sufficient
strength, magnetic system characterized by a random, yet co-operative,
freezing at a well-defined Tf below which a highly irreversible, metastable
frozen state occurs without the usual long-range magnetic order. As this
chapter has demonstrated, there are myriad systems which satisfy this
definition. And we have been calling such systems the good spin glasses
where the special case of canonical denotes the infinite-range RKKY noble-
metal alloys upon which so much experimentation has been performed. The
word ‘ideal’ we reserve for those materials which closely conform to a
specific theoretical model as we shall enumerate in Chapters 5 and 6.
Nevertheless, there are other materials, discussed previously, which lie in
the grey area, and when further progress is gained through more experimen-
tation they may possibly be shifted into the good spin-glass category.

REFERENCES

Hiiser, D., Rewiersma, M. J. F. M., Mydosh, J. A. and Nieuwenhuys, G. J. (1983)

Phys. Rev. Lett., 51, 1290.
Larsen, U. (1978) Phys. Rev., Bl8, 5014; and unpublished.
Maletta, H. (1982) J. Appl. Phys., 53, 2185.
5
Models and theories

In this chapter we wish to introduce the llavour of spin-glass theory and

the associated models. It is not our purpose to review in detail the entire
development of the various theories but to offer the salient features. There
are simply too many contributions mainly spanning the decade 1975-1985
to present a complete account in one chapter. Besides several monographs
and review articles (see Further Reading and References at the end of this
chapter) have focused upon an in-depth description of the theory including
the necessary mathematical manipulations. Instead, we shall survey the
important theoretical concepts without mathematical rigour, yet with
particular emphasis on the physical content. This means the mainstream
models and concepts will be explained simply by words and pictures
employing a limited number of equations. Derivations will not be carried
out, but hopefully the basic physics of the spin glasses and its unique nature
can be gleaned from our treatment.
As we shall see present-day theory is still developing and has a long
journey to make before the ‘solution’ of the spin-glass problem is reached.
Also many of the new ideas and calculations are as yet preliminary and
certainly there has been no complete contact with experiment to establish
the final validity of a particular model. The spin-glass problem is a tough
one and requires some radical new concepts and sophisticated insights.
Essentially, a new form of statistical mechanics needs to be constructed for
even the first-order (mean-field) approximation of the freezing transition.
Well with all this fanfare as a preamble, let’s sketch how we shall proceed
with the various sections of this chapter.
We begin by listing four historical developments which preceded the
name ‘spin glass’ or the possibility of a phase transition. These then set the
stage for the Edwards-Anderson (EA) model and order parameter, to be
immediately followed by the Sherrington-Kirkpatrick (SK) mean-field model
and solution. These theories directly confront the freezing process as a
phase transition with a special order parameter. Proceeding chronologically
we discuss the instability of the SK solution and the de Almeida-Thouless
line. A different approach is the Thouless-Anderson-Palmer (TAP) solvable
model of a spin glass. Here the solution gives a profusion of equivalent
136 Models and theories

ground states. The cause of the instability of the SK model lies in the
replica-symmetric order parameter and a replica-symmetry-breaking (RSB)
scheme must be incorporated into the theory. At this point, an unconven-
tional, yet mean field, order parameter arises with fundamentally new and
subtle physics. We try to dwell on the pictorial representation of RSB and
its physical meaning. Afterwards, we move onto the dynamics of the mean-
field model and the time-dependent interpretation of the order parameter.
We hope these seven sections will present an introductory overview of the
mean-field theory (only a first approximation) for the spin-glass transition.
Remember the theory took more than five years to evolve and at least a
thousand publications to elucidate.
The remaining sections of the chapter summarize two more recent
treatments of spin glasses, namely the droplet model which is diametrically
opposed to the RSB standpoint, and the fractal-cluster model which closely
resembles our experimentalist’s interpretation, yet can quantitatively derive
the results. Then, since all of the above models and calculations have
usually relied on the Ising simplicity, we briefly consider non-Ising spin
glasses and what kinds of new effects can be expected. Finally, the all-
important computer simulations of spin glasses are described and the various
results collected for comparison (in Chapter 6) with experiment.

5.1 SOME HISTORICAL PERSPECTIVES

In the old days (1960s) before the name ‘spin glass’ was coined, a number
of theoretical ideas were afloat helping to explain random magnetic systems.
Four of these early approaches are now recapitulated, particularly because
we have already utilized them in the previous chapters. So we begin this
theory chapter with a bit of history by collecting the four notions and
briefly outlining their physical content.
The concept of concentration and range scaling for the RKKY interaction
was exploited by Blandin (1961) to derive some universal properties of the
noble-metal alloys. All thermodynamic quantities follow from the partition
function
Z= Trexp(-X/&T) (5.1)
where for an RKKY system, recall (1.2), the Hamiltonian can be
approximated with leading term and external field H
Jo COS (2kFrij)
(2kFrij)3 si ’ % - @BHxsi (5.2)

If we simultaneously multiply H and T by the same coefficient (l/x), the

inverse concentration, the partition function remains unchanged. This follows
from the introduction of a length scale R,, which is that of the impurities
5.1 Some historical perspectives 137

- not the lattice,so defined that its volume always contains a constant
number of impurity spins, xR%. If we divide this invarient by the constant
volume V, determined by the number of conduction electrons N, we can
write
xR%x(k&)”= const or (kFRc)3 m i
vo- 37r2N
using

(5.4)

Therefore, the first (RKKY) term in our Hamiltonian (5.2) also scales with
the inverse concentration, and after dividing by x it becomes an invarient
with respect to (rq/R=). Since all terms scale with x-l, the partition function
may be expressed in scaling form as

H )I
zzzT””
X

Hence the free energy obeys the scaling relation

‘X

P’=-kn~lnZ=-k~,ln[,(~,~)]=-knx2~(~,~) (5.6)

By taking the appropriate deviatives with respect to H or T, the

thermodynamic properties for all concentrations can be calculated, if they
are known for a single x. And experimentally we used these scaling
equations already in sections 2.4 and 2.7.
The second of our historical ideas is related to the first and was proposed
by Larkin and Khmel’nitskii (1970). It utilized a virial expansion of the
same Hamiltonian in powers of x/T or H/T. The first term (linear in x)
gives the usual paramagnetic contribution to the free energy F(l), and
higher-order terms P m, reflect the effects of impurity-impurity interactions.
Summing over all the terms establishes the total free energy

where Q is the scaling function of the virial expansion. Fc2) and Fc3) may
be calculated in certain limits and, accordingly, various prognostications are
acquired for the specific heat, magnetization and susceptibility. These are
only valid at very low concentrations and rather high temperatures for
diverse values of the applied field. By studying sufficiently dilute magnetic
alloys, a number of these predictions have been verified. For example, the
magnetization should approach saturation in a large field as

M=M,,,(l-y) (5.8)
138 Models and theories

and the specific heat at high temperature is C, CCl/T. Nevertheless, it

becomes very difficult to calculate the F(“) virial coefficients beyond the
third (m = 3) term and this limits the X, T and H ranges of application.
Furthermore, there is no occurrence in this method of the phase-transition-
like freezing phenomenon of our canonical spin glasses.
Proceeding to the third approach, we return to the early 1960s when
Marshall (1960), and Klein and Brout (1963) attempted to calculate the
free energy of what is now known as a spin glass by using a distribution
P(H) of local, random magnetic fields. The non-trivial problem is how to
determine the P(H) prescribed by the random impurities and RKKY
interactions. At T = 0 a static molecular field at the spin-site i can be
written as
Jo cos (2kFrij) s.
Hi = 7 (2kFrij)3 ’ (5.9)

For these competing interactions in an Ising (up/down) model, the modulus

of the field and its energy may be related to the random-walk problem. So
it is not surprising that a gaussian P(H) results. However, we should go a
step further and include the effects of correlations, i.e., the influence of
the original orientation of Si on its neighbouring spins, usually called the
Onsager reaction field. Thus, we must split up the RKKY into near (reaction
field) and far (cavity field) parts:

ii- (COS (2k,rij)( + 2 COS (2kFrij)

JcSi - ri,zRc (2kFrij)3 ,.ij,R, (2kFrU)3

where R, is used to limit the range of correlated spins, and, due to the
previous scaling arguments, varies like (x))*; therefore, contains a constant
number of impurities. The final distribution P(H) is a convolution of
reaction-field distribution with the cavity-field distribution and results in a
broadened, more lorentzian-like shape with a significant value at H = 0.
For the Heisenberg case all 3D orientations are possible and we must use
a 3D random walk. But now the probability density P(H) (note the vector
H), which is constant near H = 0, should be integrated over the appropriate
volume element 47rH2 dH to obtain the scalar molecular-field distribution
P(H). This gives P(H) + 0, as H + 0, and the full distribution is shown
in Fig. 5.1 a. Here P(H) begins as H2 and reaches symmetric maxima at
+H,. Thus, most of the moments are aligned in finite field (H,I. Nevertheless,
for the Heisenberg case there will be a transverse response to a field H
and this creates elementary excitations (localized magnons) of the random
system. Employing numerical simulations of randomly distributed spins
coupled via the RKKY interaction, the density of these excitations at
various energies may be estimated. This N(o) is shown in Fig. 5.1 b from
the computer results. Note that now a finite density of excitations is found
as w + 0. Such an N(O) # 0 requirement is essential for the agreement of
5.2 The Edwards-Anderson model 139

H
Fig. 5.1 (a) Expected shape for the internal (molecular) field distribution P(H)
for a Heisenberg model in d = 3; note the initial Hz dependence. (b) Density of
elementary excitations from the computer simulation for an RKKY interacting
x = 0.3 at. % fee lattice. From Walker and Walstedt (1977).

experimental data, e.g. susceptibility and specific heat. And it further

justifies the ad hoc postulate of a finite (constant) density of excitations as
w + 0 used in the two-level model. See below and section 3.3.2 with
Fig. 3.37.
The fourth and final ideal has already been considered. It is the above
two-level system of Anderson, Halperin and Varma (1972), and Phillips
(1972) which sought to explain the linear dependence of the specific heat
at low temperatures for all disordered materials, glasses as well as spin
glasses. The two asymmetric wells model of Fig. 3.37 with a constant density
of low lying excitations and quantum mechanical tunnelling provides the
necessary ingredients to generate this special linear term. Excess contributions
to the specific heat (and other experimental quantities) have been ascertained
for a variety of real glasses and spin glasses and they served as a bridge to
connect spatial (lattice) disorder with spin disorder.
The above-discussed, four theoretical approaches (scaling, virial expansion,
random molecular field and two-level systems) all came into being before
the acceptance by the theorists of a possible phase transition demonstrated
by the ac-susceptibility cusp. Up until the early 1970s there were no
indications of sharp or dramatic behaviour in any of the spin-glass properties.
Most measurements of the susceptibility were smeared out with the use of
too large an external field and the specific heat does not exhibit any acute
effects. Consequently, it was only in the mid-1970s that a flurry of theoretical
activity began to treat critical phenomenon in a disordered spin system.
But we have said all this before so let’s move on to the theories which
confront this possibility.

5.2 THE EDWARD%ANDERSON MODEL

How can we describe the sudden random freezing of a spin glass? In 1975
Edwards and Anderson (EA) proposed the following picture. Each spin Si
140 Models and theories

becomes locked into a preferred direction whose orientation is random over

the distribution of sites i. This transition from a paramagnet to the frozen
ground state requires an order parameter. Since there is no long-range
order, conventional order parameters reflecting spatial correlations are
unusable. So EA focused on the time order. If a spin at site i is frozen
into its ground state, its orientation will remain the same tomorrow as it
was today. Thus, we can try a time autocorrelation function:
q = lim (CW> - si(f) MC (5.11)
t4m
where the inner angular brackets represent a thermal averaging (T) and the
outer a configurational average (C) over all spins. Naturally, q = 1 at T = 0
and q + 0 as T + Tf. A way of visualizing this state is to compare snapshots
taken at different time intervals of our spin glass. If the various Si point in
the same direction in the different photos, then the spin glass is frozen and
q # 0. If there are changes of orientation between the photos, then q = 0
and we have a paramagnet. For ergodic systems the local time correlation
is identical to
4 = (@iK)C (5.12)
where we have squared the thermal (ensemble) average before taking the
configurational one. Recall section 2.4 for our spin glass: ((SJT)C = 0, so
our system is neither ferro- nor antiferromagnetic on any scale. q, then,
represents the Edwards-Anderson order parameter which will characterize
our spin-glass phase. Now we need a model which will allow the free energy
to be calculated in terms of q and EA have also provided us with such a
model and its molecular-field approximation.
The EA model innocuously starts by writing the standard hamiltonian

X = - ZJijSi . Sj - XHi * Si (5.13)

ij I
for a random-bond, 3D, square lattice. The classical spins on site i and j
interact via the exchange coupling Jij.These are randomly chosen according
to a gaussian distribution
1 JS
p(JiJ = (2TA2)1 exP i 2~2 (5.14)
1
where A is the variance. We now need to determine the free energy which
is given in terms of the partition function 2

F= -k,TlnZ= -kBTTr (5.15)

The model system has quenched disorder, i.e., the impurity degrees of
freedom are rigidly frozen, meaning there is no change in the randomness
‘.2 The Edwards-Anderson model 141

)f the spin sites (sample structural disorder is frozen in), only the spin
orientations can vary. For such a system one must average In 2 over the
listribution P(J& which is a difficult task. However, it is relatively easier
o average the partition function raised to some power Z” where IZ is an
nteger. Here the so-called replica trick can be employed

In Z = lim +Zn - 1) (5.16)

?I+0 [ n I
;or positive it, we can express Z”{X} ({x} represents the set of bonds
lescribing the disorder) in terms of n identical replicas of the system

Z”{x} = fi Z,(x) = fi exp [- X(x, Sy}lk,T]

or=1 CC=1

= exp - i %{x, S~}lk,T] (5.17)

[ a=1
vhere Z, is the partition function of the ath replica.
The configurational average of Z”{x} = (Z”), over the disorder is
:omputed by

(Z”), = Tr exp
CS3 [
- i X{x, Sy}IRBT
a=1 1 (5.18)

iubstitution of our starting hamiltonian (with Hi = 0 for simplicity) leads

0

(Zn)c = Tr n dJ,P(J,) exp ST i SrST] (5.19)

{SF} (i, i) B or=1

W)c = c

{SPJ
I_‘:[(rl)
I,
dJij P(Jij) exp SPSF
II (5.20)
The integral over Jij can be performed for the gaussian P(J,) by completing
he square

rhis four-spin product in the exponential arises from the addition and
ubtraction of a term

(5.22)

o complete the square. Now the free energy may be obtained in the mean-
142 Models and theories

field approximation, assuming (($S$),)c = 0, and setting q = ((SJ&. An

expression F(q) results from which q(r) can be determined via the condition
dF/dq = 0. These original EA equations are not simple and are only soluble
in the limits T + 0 and T + T,. The results are

(5.23)

and

q(T+ Tf) = - ;

We can write the susceptibility

[ 01
1- $ ’

via the fluctuation-dissipation

(5.24)

theorem as

(5.25)

where all i # j terms have to be set equal to zero (no long- or short-range
correlations). Since ((S&)o = 1 and ((SJc)o = q we obtain

Substituting the limiting results for q(T) and since H = 0, we write for the
ac-susceptibility

xa,(Ts
(&wBY
TJ = 3k,Tf o(Tf - T)2

and

(5.28)

According to these equations an asymmetric cusp is formed in the

susceptibility at T,, while x should approach a constant value at very low
temperatures.
Once we have the free energy F(q), the internal energy U can easily be
calculated. Since the specific heat C, is just aulaT, C,(T) follows from
differentiating and knowing q(T). In the original EA estimates a cusp was
implied at Tf for the specific heat. The calculations of the EA model were
extended by Fischer (1975) who employed a different technique within the
mean-field approximation and used quantum spins (S = 5) instead of classical
ones (S = 03). In Fig. 5.2 we show the results of the susceptibility and
specific heat for these two spin values. Note the quantum theory gives
sharp cusps in both these quantities at Tf. The x(T) for S = $ nicely
resembles the generic measured behaviour of xac even down to the low-
temperature constant. In contradistinction, the theoretical specific heat for
5.2 The Edwards-Anderson model 143

-Y
0’

0.5
Fig. 5.2 (a) susceptibility and (b) specific heat versus reduced temperature calculated
from the EA model for two spin values; from Fischer (1975).

S = $ completely disagrees with experiment except for the low-temperature

linear dependence. So we clearly have a problem with the EA model and
its mean-field approximation. Nevertheless, this was the first provocative
attempt at coming to grips with the ‘cusp’ and a possible phase transition.
The model is really very simple, yet elegant: replace the site disorder and
RKKY interaction by a random set of bonds which satisfy a gaussian
distribution. The clever definition of an order parameter related to a time-
correlation function makes the freezing transition tractable with a statistical
mechanics treatment. What still remains is to establish the true mean-field
theory for this model.
144 Models and theories

5.3 SHERRINGTON-KIRKPATRICK MODEL

A mean-field theory (MFT) is usually a first-order approximation for

describing a second-order phase transition. For ferromagnetism the MFT
becomes exact in the limit of infinite-range interactions. Another applicable
example is superconductivity where the interaction between the electrons
forming the Cooper pair is long range. Once the MFT has been established,
the necessary additions or corrections, e.g. fluctuations, short-range interac-
tions, etc., can be included to treat real systems. Sherrington and Kirkpatrick
(SK) in 1975 proposed that the proper MFT of spin glasses should be the
exact solution of an infinite-range EA model where every spin (N is their
number) couples equally with every other spin. This means the probability
distribution P(J,) is assumed (unphysically) to be the same for all i-j pairs
of spins independent of how far they are apart.
SK considered an Ising spin glass with a gaussian distribution

P(Jij) = 1 exp [- (Jij - JA)2/2A’2]

J5z

where a mean, J& has been included for the possibility of ferromagnetism
in the gaussian function. Infinite-range interactions require the scaling of
the variance and the mean according to A’ = A/N+ and JA = JOIN, so that
both our new A and JO are intensive quantities. Thus

Repeating the ‘replica trick’, i.e., calculate (Z”), instead of (ln Z),, as with
the EA calculations and after a lot of mathematics, which nobody likes to
show, SK arrived at a rather complicated expression for (Z”),, the
configurational-averaged, replica partition function. Substituting this
expression into the replica equation, we obtain

F = -knT(ln Z), = - kBT liliO i ((Z”), - 1) (5.31)

frdy(““‘, (%Trdx’..’
B
5.3 Sherrington-Kirkpatrick model 145

+$Tc x(a)p) - 1 (5.32)

l3 (a) II I
Here y(“P) and x@) are the dummy variables of integration where (c+) label
distinct pairs of replicas cx and l3 which take on values from 1 to IZ. The
trace is over 2” values of S” = ? 1 at a single site. SK first took the
thermodynamic limit N + ~0, then the replica limit 12+ 0 in order to
perform more easily the integration (method of steepest descent). They
considered the various replicas to be indistinguishable. This is called the
replica-symmetric solution with
4 = qap = ((~amT)c (5.33)
for the spin-glass order parameter and
m = m, = (W& (5.34)
for the ferromagnetic one.
The problem is now to calculate F(q, m) in the limit n + 0 and by
differentiating with respect to q and m determine the self-consistent
simultaneous equations for q and m. The SK results are

q=&/exp($]tanh’[;z+$]dz (5.35)

m=&/exp($]tanh[sz+g]dz (5.36)

Hence for given ratios of JO/A, q(T) and m(T) may be calculated and a
magnetic phase diagram is thereby established. Figure 5.3 shows this T

1.25 - Para

Ferro -

0.50 - Spin glass

0.25 -

0.00 0.25 0.50 0.75 1.00 1.25 1.50

Jo/A

Fig. 5.3 Magnetic phase diagram for Ising spins interacting via an infinite-ranged
gaussian distribution of exchange forces with variance A and mean J,; from
Sherrington and Kirkpatrick (1975).
146 Models and theories

versus J,/A plot for Ising spins interacting via an infinite-ranged gaussian
distribution of exchange forces centred at Jo with width A. Note the three
possibilities for phase transitions which are predicted from the SK model:
(i) paramagnetic * spin glass; (ii) paramagnetic + ferromagnetic; and
(iii) double (or re-entry) transitions paramagnetic + ferromagnetic + spin
glass.
The differential (or ac-) susceptibility may be computed from the q(T)
function using
[l - q(T)1 (5.37)
X(T) = kBT - JO[l - q(T)]
By including the applied field H from our original hamiltonian in all the
SK calculations (something we omitted for simplicity), we can also obtain
the field dependence of x(T). Figure 5.4 exhibits the susceptibility behaviour
for J,lA = 0 and 0.5 with and without a field H = 0.1 A. Once again,
conforming to experiment, we have a cusp in the ac-x which becomes
rounded and shifted downward in a dc field. But, when the specific heat is
calculated according to the same energy-derivative procedure used in the
EA model, there is also a cusp in the predicted C,(T) at Tf. For
T > T,, Cm = NkBA2/2(kBT)Z, hence a tail in C, 0: l/T2 persists to higher
temperatures. (Remember from the virial expansion treatment of the specific
heat at high temperatures (section 5.1) C, CCl/T.) The SK result is in stark
contrast to the usual mean-field-theory conclusion for a pure system where
C, = 0 for T > Tf. The leading term for the spin-glass specific heat at low
temperatures is proportional to T. However, the entropy S, when determined

O.OO
I
0.00 1.00 1.50 2.00
kT

Fig. 5.4 Differential susceptibility as a function of temperature calculated from the

SK model: solid curves H = 0; dashed curves H = 0.1 J,; lower curves J,/A = 0;
upper curves J,lA = 0.5. From Shetington and Kirkpatrick (1975).
i.4 Instability of the SK solution 147

‘ram the SK model, goes to a negative limit: -Nkg/2T at T = 0. This is a

nost unphysical and disturbing feature of the model.
One of the impressive experimental realizations of the SK model is
he ternary Pdl-,-,Fe,,Mn, system. Here Fe,, controls the amount of
trromagnetism present, viz. the J,, and essentially Mn, creates a mixed
nteraction which governs A. With a suitable choice of y and X, a phase
liagram closely resembling (mirror image) that of Fig. 5.3 is generated and
:xperiment clearly shows the three distinct types of ‘phase-transitions’ listed
above. Furthermore, there is also a sound qualitative agreement between
he SK-model results and the measured susceptibility and its field dependences
‘or the various concentration regions.

i.4 INSTABILITY OF THE SK SOLUTION

rheoretically the first warning that something was wrong with the SK
solution came from the negative, low-temperature entropy. Other difficulties
)ecame apparent with the free energy which turns out to be a maximum
vith respect to q for the solutions q = 0, T > Tf and q # 0, T < Tf.
tioreover, the q = 0 solution, if analytically continued below Tf has the
ower free energy than the spin-glass state (q # 0). These results are
jpposite to what is expected for a conventional second-order phase transition.
Finally, after a few years, de Almeida and Thouless (1978) performed a
letailed analysis of the SK solution and showed it to be unstable at low
emperatures both in the spin-glass and ferromagnetic phases. According to
he SK solution the spin-glass susceptibility is really negative. This is not
mly in contradiction with experiment, but also with the correlation-function
#usceptibility which is positive definite. In the presence of an applied field
H # 0) the instability line of the SK-solution for the spin-glass phase
:xtends all the way from Tf down to 0. Such behaviour is shown in Fig. 5.5
vhere we plot the H-T line (also called the AT line) that gives the stability
imits of the SK solution. The functional form is

Tf - T,,(H) = 3 f H :
(5.38)
A 4
O( a 1
‘or Tf = TAT(H) and an Ising spin glass. This becomes exponential in -H2
IS T + 0. An experimental consequence is that below the stability line
rreversibilities in the magnetic properties should appear. As we know from
3hapter 3 these irreversibilities are indeed a decisive characteristic of the
rozen spin-glass state.
The reason for the instability has to do with treating all the replicas as
ndistinguishable. Remember SK set q = qup in (5.33) - a replica symmetric
,olution, and such an assumption leads to an invalid solution of the mean-
ield EA model. Despite these theoretical difficulties, the SK model offers
148 Models and theories

I I

1.00

0.00
0.00 0.20 0.60 1 .oo
kT/A
Fig. 5.5 H-T phase diagram (or AT line) illustrating the stability limits of the SK
solution for the case of J, = 0; from de Almeida and Thouless (1978).

a reasonable first basis for comparison with experiment. As we mentioned

above (section 5.3), the predicted phase diagram can be nicely mimicked
by real spin-glass materials, and the calculated susceptibility is in qualitative
agreement with measurement. This accord seems to be due to the SK-
solution being correct above TAT(H). And for certain experimental quantities
(e.g. the ac-susceptibility), which are not strongly dependent on the
irreversibilities or energy of the frozen state, the theory is valid. Yet for
the majority of other measured parameters (specific heat, magnetization,
etc.) the above instability causes the SK-model to break down and more is
needed to describe the subtleties of the frozen spin glass.
Theoretically what do we do now? After using so much mathematics to
derive a simple mean-field theory that looked attractive from an experimental
point of view, we find the solution to be unstable. One answer is to find a
scheme and solution which breaks the replica symmetry. Nonetheless before
proceeding with devising an RSB scheme, we consider a different approach
that does not rely on the replica trick.

5.5 THE TAP APPROACH

Let us recall what the usual mean-field equations for a magnet (ferro or
antiferro) are
5.5 The TAP approach 149

where the thermal averaging of the hyperbolic tangent has been replaced
by the thermal average of Sj, namely, (f(Sj))T = f((Sj)T). And such is the
standard mean-field approximation. However, for a spin glass we must
subtract from the ‘effective field’ given above a reaction term. This reaction
field represents the influence of the ith spin on the polarization of the
neighbouring spins at site j. A spin (SJ, produces a ‘field’ J,(SJT at site j
and induces a moment Jij(S&Xjj at that site (where xii is the local
susceptibility). The induced moment in turn produces a reaction field back
at site i:

Jij * J&Si)TXjj = J$(S& $T [I - (Sj>%] (5.40)

using the linear response form for the susceptibility

X jj = &iG [l - <Sj%] (5.41)

By including this correction term in our original mean-field equation, we

now have

CQ= tad &T C Jij<Sj>,

- ( bT[zJ$[ 1 - (s%]@A]
B I I
(5.42)
where i = 1, 2, 3, . . . , N and we have neglected the local applied field Hi.
This set of N coupled equations has been proposed and studied by Thouless,
Anderson and Palmer (1977) as a way of avoiding the replica method. The
resulting free energy is

FTAP= - CJij<S&(Sj>, - i($T)zJ$ [l - (SJQI[l

- (Si>CI
ij ‘I

(1 + C9-r) In* (1 + @A)

+ (1 - (SJ,) In* (1 - WT)

which may be verified by taking aF,,pIa(Si)T =

1 (5.43)

0 and arriving back at the

(SJT equation. Also the TAP approach is only valid near Tf around the
region 1 - CJI T/T, and in the low-temperature limit (T + 0). Here 4 is
the EA order parameter defined now [compare with (5.12)] as

The physical meaning of F -rAp in (5.43) is the following. The first term
represents the energy of the frozen spin glass, the second is the correlation
150 Models and theories

energy of the fluctuations which are reduced from the paramagnetic result
by 11 - (S&l f or each spin; and the last term gives the entropy of N Ising
spins possessing the mean values (S&e
At T 4 T,, TAP find that the order parameter behaves like

9(T) = 1 - or ; * (5.45)
0
which is different from the linear SK dependence 4 = 1 - a,T/A derived
from (5.35). Near Tf there is a saddle-point behaviour defining the accessible
region q0 1 q = 1 - (T/T,), below which divergences occur in the FTAP.
Thus, a linear temperature dependence of q occurs just below Tp According
to these q(T) variations we can calculate the low-T susceptibility as
(Xi)T = aiT/A which goes to zero as T + 0 (recalling Xsk remains finite as
T + 0 in agreement with experiment). The entropy S may also be
determined: S,,, = -aFTAPIaT, and in the low-T limit

fL,(T+ 0) = (5.46)

Substituting the low-T quadratic form for q(T) in (5.45), S,,, + 0 as

T + 0, in stark contrast to the negative constant entropy at T = 0 for the
SK result. Therefore, the TAP model has a physically meaningful entropy
and its configurational-averaged free energy also tends to zero as T + 0,
i.e., the TAP ground-state energy vanishes at T = 0. These distinctions
between the SK and TAP models lie in the two dissimilar temperature
dependences of q(T) and in the breakdown of linear-response theory for
spin glasses. Even so both models incorrectly predict a cusp in the specific
heat at Tf.
While the TAP approach does not have the fundamental theoretical
drawbacks of the SK solution, it offers many physical and stable solutions
below the AT-line with different free energies. We can interpret each of
them as a locally stable thermodynamic state whose free energy is not
necessarily a global minimum. But what are the exact meaning and relevance
to experiment of all these different thermodynamic states for real spin
glasses? Our attempt at an answer will emerge below. Nevertheless, above
the irreversibility or AT line (shown in Fig. 5.5), similar results for the
‘paramagnetic’ phase are obtained from both SK and TAP models. The
hint here is clear, the frozen spin-glass state is a unique phase that requires
new physics, e.g. an unconventional order parameter and statistical treatment.
Consequently, a new approach to the theory and interpretation is needed.

5.5 HOW TO BREAK THE REPLICA SYMMETRY

Since we know where the trouble lies let’s try to overcome the replica
symmetry by introducing a scheme that breaks it. After a number of
i.6 How to break the replica symmetry 151

Ittempts, the correct one was finally proposed by Parisi in 1979. His
lrocedure for breaking the symmetry of the order-parameter matrix was as
~0110~s. Consider an n x 12 replica symmetric matrix q0 (for illustrative
Imposes we take n = 8). We then divide the (n x n) matrix of elements
I,, into constant blocks [(n/ml) x (n/ml)] of size ml x ml. Along the
diagonal we introduce sub-matrix ql. The off-diagonal blocks are left
unchanged with elements qO. This process of dividing 8 x 8 matrices into
4 blocks of 4 X 4 matrices is portrayed in Fig. 5.6 and is only the first of
nany such steps. The division is again repeated along the diagonal:
n1/rn2 X ml/m2 (or 4) new blocks of m2 X m2 (2 X 2) sub-matrices are
:reated with order parameter q2 for the new diagonal ones. And so on (see
Fig. 5.6) the procedure is reiterated with smaller and diagonal blocks each
with its own order parameter qk. (In our 8 X 8 matrix shown in Fig. 5.6,
:he next iteration results in all zeros along the diagonals.) Finally, for a
nuch larger matrix the process is terminated after R such iterations with
:he smallest diagonal block of size mR X m& Throughout this construction,
:he successive sizes of the blocks are
n~mI~m2~~~~rnR~l (5.47)
which in the limit as n + 0 becomes ‘turned around’
OSmlIm2-~ -ImRI1 (5.48)
3y taking R, the number of iterations, to be very large, we obtain a
:ontinuous variation and so

$$+, BLOCKS

n x n OF $lZE m, x ml

8x8 4 BLOCKS
OF SIZE 4 X 4

Fig. 5.6 Replica symmetry breaking (RSB) scheme for qae with two levels of
breaking (n = 8, ml = 4; m2 = 2).
152 Models and theories

mklmk+1 3 1 - dxlx and qk + q(x) (5.49)

where x is defined over the unit interval 0 I x ‘: 1. In other words, we
now have an infinite number of order parameters. Note that the SK solution
corresponds to the k = 0 step or original q. matrix which means that q(x)
is independent of X.
There is an alternative way of visualizing the Parisi replica-symmetry-
breaking scheme. The order-parameter matrix qap can be represented by a
tree with emanating branches. Again we utilize the (8 x 8) or 8 replica
matrix of Fig. 5.6 and sketch our tree construction in Fig. 5.7. At lowest
order q. all the elements are equal and contained in one big branch. The
sub-division process breaks up the branch into two smaller ones which in
turn may be further broken up into two yet smaller branches, etc. The
integers represent the it = 8 replicas and as they are divided q increases
towards 1. In order to find the value of qap for any two of our replicas (in
Fig. 5.7, (Y is replica 1 and l3 is 4), we trace these back until they join at
ql. The overlaps of the different replicas are all equal if the point of
encounter lies along the same horizontal line. For example in Fig. 5.7, q12,
q34, q56, q7s all have equal overlap with value q2, however, qig, i.e., replicas
1 and 8, has a different overlap since it finally joins branches at qo.
The concept of replica overlap has already been treated in pure
mathematics and is known as an ultrametric space. It is defined by the so-
called ultrametric inequality
44 C) 5 Max{d(A, B), W, c)> (5.50)
where the ds represent the distances AC, AB and BC. This is a much
stronger inequality than the usual triangular one with the same three
distances
d(A, C) 5 d(A, B) + d(B, C). (5.51)
For a spin glass, which was shown to satisfy the ultrametric inequality, we
mean the ‘overlap’ (equivalent to the above distance) between two spin
configurations or replicas

(5.52)

(Ising spins). The ultrametric inequality now becomes

q”p 2 Min (4-P , qp’} . (5.53)
This inequality implies that the space of states can be divided into clusters
of states, each cluster of state may be subdivided into subclusters, and so
on. Figure 5.8 depicts such a space of clusters with separation into smaller
and smaller clusters. Note there are no overlappings among the different
clusters of similar order, i.e., each point lies in just one cluster. The sketch
in Fig. 5.8 is nothing but another way of looking at our tree diagram of
5.7 Physical meaning of RSB 153

I 12345678 I

Fig. 5.7 Tree representation of Parisi’s RSB scheme. To find qms trace back along
the branches of the tree from cxand from p until they join: qae = q1 is the value
of q at this point.

Fig. 5.7, both of which are a direct result of ultrametricity. And this new
‘space’, called ultrametric, is a natural consequence of the replica-symmetry
breaking required to remove the defects of the SK solution

5.7 PHYSICAL MEANING OF RSB

Returning briefly to experiment, let us try and see how this highly
sophisticated mathematical theory makes contact with what we have
154 Models and theories

Fig. 5.8 Sketch of the cluster structure of ultrametric space. n = 8 {1,2,3,4,5,6,7,

S}; m, = 4 {1,2,3,4} and {5,6,7$X}; m2 = 2 {1,2,}, {3,4}, {5,6} and {7,8}; m3 = 1
{l}, {2}, {3}, {4}, {5}, {6}, (7) and (8). Note how closely this subdivision follows
the tree representation and the replica symmetry breaking scheme of our 8 x 8
matrix.

measured on real spin glasses. The key information after the RSB has been
accomplished is contained in our now-continuous Parisi order parameter
q(x). This is obtained in the interval 0 to 1 via the matrix construction and
limit taking procedures

(5.54)

A number of calculations to determine q(x) have been performed at

different temperatures and fields. In Fig. 5.9 we illustrate the various
behaviours of q(x) and Z'(q) = dxldq (see below) for several interesting
cases spanning a range of T and H values. A careful perusal of this figure
is necessary to understand its significance.
According to the RSB model, the susceptibility and the internal energy
are, respectively, given by

x = &TI’ I1 - +)I ‘b (5.55)

B o
and

u = - kTI’ [l - qZ(x)] d.x (5.56)

0

We call attention to the breakdown of linear-response theory, and its

corresponding susceptibility x = (C/T) (1 - q), in the RSB scheme. By
substituting the results for q(x) at several different temperatures, we can
calculate x(T) by evaluating the integral. The Parisi and SK results are
compared in Fig. 5.10 for the low-field susceptibility. Notice the dissimilarity
with respect to the SK form, for, x(RSB) is constant for all T < T,, and
this T-independent susceptibility certainly resembles our field-cooled one,
 5.7 Physical meaning of RSB
155

q-

q P
t

(L(.
% %I q

Fig. 5.9
156 Models and theories

1 x 9
”

%A
II) (I)

Fig. 5.9 Continued Behaviour of the order-parameter function q(x) and the
probability distribution of overlaps P(q) for various cases. (a) and (b) just below
T, in zero field: qM a T, - T and the initial slope of q(x) is roughly independent
of temperature. (c) and (d) at low-temperatyre T Q T, and in zero field: qM = 1.
(e) and (f) for small magnetic fields: q,,, 0: Hs. (g) and (h) at larger magnetic fields
near the AT line: (qw - qm) is proportional to the distance from this line. And (i)
and (i) in the replica-symmetry region T > Tf at non-zero magnetic field. Note the
value of qEA vanishes when H + 0. From Mezard et al. (1987).

xFc, not the ZFC or ac-x. Further (computer) computations show that
U(T = 0) reaches a negative value while S(T = 0) = 0, so we have removed
via RSB the unphysical negative entropy. Finally, the stability of the Parisi
solution has been carefully analysed and has been found to be at least
marginally stable.
In order to avoid the subtleties of the RSB approach we digress for a
paragraph and use two results of RSB in combination with the stable SK
solution (above the instability line) which we project into the spin-glass
phase. This is called the ‘projection hypothesis’ which begins with valid SK
mean-field theory. The two features of RSB are that x(T) = constant in
the spin-glass phase and that linear response is destroyed at low T.
Accordingly, we can assume for T < T,:
W, H> = W)
MT, H) = M(H)
0, W = dT) (5.57)
5.7 Physical meaning of RSB 157

I
RSB

F
0 1
T/T f
Fig. 5.10 Difference in the temperature dependence of the susceptibility as
calculated from the RSB-solution of the SK model and the SK-solution without
RSB.

where the latter represents the field independence of the EA order

parameter. A Maxwell relation unites the first two assumptions via
(W3H)r = (dM/aT)n. Starting from the instability line, where the replica
symmetric (SK) solution is valid, and using the above projections, expansions
can be developed for the magnetization M(T, H) around Tf.

3T2
HZ T2+2T:+
T2-T; ‘-*
1
forT’Tf (5.58)

M=; IHI...I
f[l--a+ for T = Tf (5.59)

&[l-($@+. . .] for T< Tf (5W

This last equation, describing a non-analytic regime and being independent

of T, suggests that a line of singularities extends over the entire interval
from Tf down to 0. Equation (5.58) for T > Tf shows that, while there is
a cusp (or really a plateau) in x = M/H; the non-linear susceptibility
a3M 2 T2 + 2Tf
a (T-T&l (5.61)
Xnl = aH3 = T3 (T + Tf) (T- Tf)
diverges according to the mean-field critical exponent y = 1. In addition
the above equations exhibit the field rounding of x (and also for C,)
around Tf. So here we have established a few preliminary points of contact
with experiment, some in agreement as with the FC susceptibility and field
158 Models and theories

smearing, and others clashing as with the specific heat. But much more is
needed in the interpretation of RSB.
Is there a simple physical picture for the replica symmetry breaking
(RSB) and the resulting ultrametricity? Well let’s attempt to sketch one.
During the freezing of a spin glass, the various spins can align their
directions in many different ways such that all these arrangements have
similar values of the free energy. Hence, a frozen spin glass possesses a
multi-degenerate ground state. The configurations which have lower free
energies correspond to pure equilibrium states. Other configurations with
higher free energies correspond to the metastable states. How do we then
characterize the ensemble of equilibrium state? Answer: use a weighting
probability w, for the state (Y to appear in the ensemble

w-p LF
i kBT p)
where F, is the free energy of equilibrium state 0~. Another question is
how does one equilibrium state (Y differ from a second one B. Answer:
define a ‘distance’ d& between these states as the difference in their average
spin values square

(5.63)

In a similar manner we can construct an ‘overlap’ between two states OL

and B

Here the EA order parameter q uA represents states which have the same
overlap with themselves, i.e., qaa = qpp = qnA. Thus
k$3 = %,A - qa& (5.65)
So we have arrived at a simple relation between distance and overlap via
(5.65).
Now we must consider the probability distribution P(q) of the various
overlaps among the pure equilibrium states

P(q) = P.&l) = c%wB WL, - 4) (5.66)

4
where the bar averaging is over different values of the couplings J. In other
words, Z’(q) is the probability of finding two states with overlap q after
weighting each state with their ensemble probabilities w,, a* The physical
interpretation of RSB is grounded in the non-trivial result that P(q), defined
above, and P(q) = dxldq (introduced at the beginning of this section and
shown in Fig. 5.9) are identical.
5.7 Physical meaning of RSB 159

Let us examine a most simple example via this formalism, namely, the
ferromagnetic Ising model. In zero magnetic field at T < T,, P(q) contains
two delta functions at q = + m2 where m is the spontaneous magnetization.
If an up (or +) external field is applied, then only one delta function at
q = +m2 remains. For the former case there are two equilibrium states;
for the latter only one. Now we return to the spin glasses. At T > Tf one
delta function at q = 0 exists (the paramagnetic state). However, as T is
lowered below Tf in zero field, a delta function survives but along with it
is a smooth region of probabilities in between 0 and qw - see Fig. 5.9.
Here, when P(q) is not an array of discrete delta functions, we say that
the replica symmetry is broken. This usually occurs in a very smooth way
as the temperature is reduced through Tf. Smeared out transitions?
Consequently, there are a significant number (not just one or two) of pure
equilibrium states which compose the spin-glass ground state. And the
weight and overlap of these states generate the continuous function P(q).
Figure 5.11 gives a schematic, 1D picture of the free-energy landscape
that exists in multi-dimensional configurational space. The lowest-lying
minima represent the pure equilibrium states. At higher energies are the
many metastable states. Upon cooling, a spin glass may become stuck in
one of these states. If the system cannot explore all of phase space, then
we call it non-ergodic. RSB leads naturally to ultrametricity which in turn
is pictured by the multi-valley free-energy landscape. And when the barriers
between the valleys become infinite, ergodicity is lost. Note that the time
it takes to go from one valley to another is the exponential of the height
of the lowest saddle point between these two valleys. So it is not surprising

CONFIGURATION SPACE
Fig. 5.11 Multi-valley ‘landscape’ for the free energy according to the RSB scheme.
160 Models and theories

that there are many different relaxation times present, and we should now
move on to examine the dynamics of spin glasses within the SK and RSB
models. However, before we do, we must mention a final consequence of
the RSB. This is the lack of self-averaging.
Usually in a random system we can average over the extensive quantities
by dividing it into a large number of macroscopic subsystems, each containing
a different set of random variables {x}. If we then average over the
subsystems, we expect to obtain a result very similar to performing a
complete configurational average over all choices of {x}. Taking the
magnetization per spin as our extensive quantity, self-averaging may be
written as
M(x) - (M)c 3 0 for [N-spins] + uJ (5.67)
where the chosen set {x} occurs with reasonable probability. In other
words, one large random system gives the same result for the magnetization
as a configurational average over all choices of {x}. Clearly, when we
measure the magnetization on various samples of CuMn having the same
concentration, there are no differences in the experimental values, even
though each sample has a different set {x}. The configurational-averaged
magnetization is expressed as

But after performing the configurational average on the right-hand side,

translational invariance is restored to the spin system. So each term in the
summation is equal. Thus (M)c = ((SJ&. And using our definition in
(5.67), of self-averaging M(x) = (MC) for N+ 03, we have

where the summation runs over all the spin sites for a single large sample.
Although self-averaging is intuitive and seems almost obvious, it does
not work when a phase transition occurs which makes the system depend
crucially on the boundary conditions. For our spin glass the RSB solution
creates the existence of many degenerate ground states and this causes the
lack of self-averaging. Such quantities as (r& - q2 (overlaps) and
(PJ(ql)PJ(q& - P(ql)P(q2) (weights) do not equal zero as RSB calculations
have shown. Here the weights of the thermodynamic states are the important
factor. Extensive quantities depending on these weights will not be self-
averaging. This means different bond configurations will give different
results even for N + ~4. It turns out in RSB theory that the magnetization
is self-averaging, but the susceptibility is not:
5.8 Dynamics of the mean-field model 161

And this is because the thermodynamics weights change rapidly as the

held is varied. It is too bad experiment cannot distinguish these two
susceptibilities.

5.8 DYNAMICS OF THE MEAN-FIELD MODEL

Since the SK-model considers Ising spins that have no natural intrinsic
dynamics, we must introduce time dependences artificially via Glauber
dynamics or the soft-spin gaussian-noise term. Near a second-order phase
transition time-dependent correlations take the form predicted by dynamical-
scaling theory. The autocorrelation function is defined by
(5.71)
where the thermal average ( )* can be replaced by an average over
sufficiently long observation times. Dynamical scaling predicts that the
decay of q(f) is governed by a characteristic time T which (power-law)
diverges at Tf. Accordingly, we can write

q(t) 0: t-“Q*(k) (5.72)

where Q=(x) are universal scaling functions and A is a critical exponent.

We now relate the characteristic time to a characteristic length &o, the
correlation length for our spin glass,

Here v is the correlation-length critical exponent and z is the dynamical

one. In mean-field (SK) theory v = $. Taking the limit of t + ~0 for the
autocorrelation function q(t) we have

lim q(t) = qnA a

*-am
where p is the order-parameter exponent. In order for the scaling function
Q(X) to cancel the tP dependence in the limit t j ~4, we obtain a relation
between the exponents A = pl(zv).
Above the AT-line where SK is valid, various calculations have all shown
for H = 0

with the proper scaling function Q+(x) a exp (-x). Thus, for T > Tf and
ast+w
162 Models and theories

4(t)Oc
(T+t $exp[-t(T)‘] (5.76)

Tf
The scaling function reduces to q(t) a t-i at Tf. For H # 0 one can begin
with the dynamical-scaling form and define a critical temperature according
to [T - TAT(WYTATW). I-I owever, both the scaling functions and the
exponents are non-universal and vary along the AT-line. This suggests a
marginality of the transition in a field. By examining the dynamical
susceptibility near the AT line, theory predicts the frequency dependence
of x”(o) a o for T %=Tf crossing over to x”(o) a & at Tf.
Below the AT-line, instabilities also occur in the dynamics of the SK
model, so a new and different approach is required here. And Sompolinsky
(1981) has provided the method with a stable solution entirely within a
dynamical framework. We now write our continuous order parameter as
q(x) = ((Si(O)Si(Tx)),>, 0s X5 1 (5.77)
where (.)= and (.)o represent thermal and configurational averages,
respectively, and T~ represent a time in the broad continuous spectrum of
relaxation times. Note that T,.. % T,~, if x2 > x1. So as x + 0 we obtain the
very long relaxation times necessary to equilibrate the system. This gives
the statistical-mechanics order parameter
q = q(0) = lim lim q(t) (5.78)
N-SW f-vu
which Sompolinsky assumed goes to zero in zero field, i.e., all correlations
have decayed away. On the other hand, for x + 1 we have the short times
scales characteristic of the EA (or confined to one valley of Fig. 5.11) order
parameter
(5.79)

It is which limit comes first that distinguishes these order parameters.

The dynamical susceptibility (measured at frequency o, = l/7,) may be
calculated with the help of another function A(X)

kBT I 7XXii(t)dt = 1 - qnA + A(X)

Jo
We assume that linear-response theory (or equivalently, the fluctuation-
dissipation theorem) is valid at finite (short) x + 1 times, but not at the
very long times x + 0 needed for equilibrium. Hence, at x = 1, A(1) = 0
and (5.80) reduces to its EA-form. By establishing self-consistent equations
for q(x) and A(x), Sompolinsky has determined the following stability
condition for his solution, using M = tanh (tilk,T) where A is a local field
made up of components which are frozen on a time scale 7,:
i.8 Dynamics of the mean-field model 163

& 11- &l(l) + (M”>]= 1 (5.81)

The limit is A(x) + 0 (or x + 1) and J is now the width or variance of

)ur gaussian-distribution of exchange interactions. The above equation
:orresponds to the TAP stability condition which is marginally stable. For
l(x) = 0 the SK solution is revived since q(1) = qEA.
As the external field H goes to zero, the only dynamically stable solution
s given by the equilibrium (or long-time) susceptibility x = xii
k,Tx = l-q(l) + A(0) (5.82)
vhere in the limit of x + 0, A(x) = q(l)-l+(k,T)/J. Hence, x = l/J (a
:onstant below Tf) and we have recovered a dynamical form of Parisi’s
susceptibility. The interpretation here is that in the short-time limit (e.g.
tc-susceptibility) we do obtain a quasi-equilibrium (linear-response) x with
ts cusp

x(quasi-equil) = kT (1 -qEA) (5.83)

lowever, for the true equilibrium susceptibility

x(eqW = &- P-qm,+Wl (5.84)

The time scales necessary for (5.84) to be valid may be far beyond our
txperimental reach of maximum lo6 s. Nevertheless, we can measure a
ield-cooled susceptibility which although not the true equilibrium one, does
ndeed mimic it in a field. The FC-field seems to drive the spin glass
owards its long-time equilibrium, and thus, XrC f constant at T 5 Tf. For
he intermediate cases we must return to the local dynamical susceptibility
neasured at frequency w, = l/t, according to (5.80) and explicitly calculate
w.
For our practical purposes the Parisi and Sompolinsky results may be
llaced within the model of RSB and benefit from its consequences.
rherefore, we can return to our rugged landscape of Fig. 5.11 and use it
o describe the spin-glass dynamics. Two main dynamical processes can be
Iistinguished in the frozen state. Firstly, confined to a single deep valley
tre the non-exponential-relaxation processes within one phase, i.e., the
ransversing of small bumps composing the floor of the ‘equilibrium’ valley.
n contradistinction, there are the hopping processes between the equilibrium
chases or from one deep valley to another. These appear on much longer
ime scales and they are estimated to diverge with the size of the system
IS log 7 - IV:. Now according to RSB, hierarchical distributions of time
scalesmanifest themselves and the hopping has taken place in an ultrametric
ipace.
164 Models and theories

5.9 DROPLET MODEL

During the mid-1980s a series of numerical studies of domain walls and

their scaling properties by McMillan (1980, 1984), and Bray and Moore
(1984) stimulated a completely different approach to the theory of spin
glasses. Fisher and Huse (1986) advanced these methods and presented the
dropret model, as a phenomenological scaling theory of droplet excitations
in short-range, Ising spin glasses. Central to this theory is the belief that
an understanding of the spin-glass phase (which should exist at T = 0) can
be obtained from its ground-state properties. The basic idea is to define a
‘droplet’ in the ground state as the lowest-energy excitation of length scale
L around a particular point Xi. Figure 5.12 illustiates the ‘droplet’ which
consists of having all its spins placed in the opposite direction with respect
to those in the ground state l?. So r is a global reversal of spins (or spin
flips) inside the length scale L. Note that the model assumes a phase
transition below Tf where the global spin-reversal symmetry is broken and
there are only two pure-equilibrium states related by this global symmetry.
The above droplets are the dominant low-lying excitations in each state.
Now we study these low-lying droplet excitations with this scale L. They
have a broad distribution of free energies FL m Y(T) Le where Y is the
stiffness constant and 8 a new exponent. If 0 < 0, as is expected for d = 2,
then F cc L-be1 and the low-lying excitations can be created on longer and
longer length scales, thereby destroying the frozen spin-glass phase.
Computer simulations indicate that 8 becomes positive just below d = 3,
so a marginal phase transition is anticipated for a 3D spin glass. A
distribution of droplet free energies p(F,) can be determined from the
above scaling ansatz, and then correlations between the droplet free energies

Fig. 5.12 Schematic picture of the droplet model. A droplet of length scale L
(containing site j) has all its spins reversed (global spin flip) creating ground state
r. Outside the droplet the spins are aligned as in ground state r. The surface of
the droplet is fractal. From Fisher and Huse (1988).
5.9 Droplet model 165

are calculated. This leads to a non-linear susceptibility x,,i which is infinite

when d > (1 + $)0 (4, is an exponent governing the dependence of p on
FL). Hence, we would expect a true phase transition if this condition is
satisfied.
The effect of a uniform magnetic field on the Ising spin-glass state is to
destroy it by breaking the global spin-flip symmetry. So argue Fisher and
Huse that the two pure-equilibrium state becomes unstable, xnl is finite,
and the system falls out of equilibrium even in the smallest fields.
Consequently, in their interpretations there is no true AT line, only a
dynamical line which may be accessible to experiment depending on the
available measurement time where non-equilibrium behaviour sets in.
Although Tf scales with H in the same way as the AT line, the position of
this freezing line is weakly dependent on the time allowed for equilibration.
Annihilation and creation of droplet excitations determine the equilibrium
low-frequency dynamics of the ordered phase. Here thermal activation over
energy barriers is assumed for those droplets which are thermally active,
i.e., their free energy is anomalously low: FL = kBT G YLe. The typical
free-energy barriers according to the length-scaling ansatz are B - L+ where
JI is a new independent exponent 8 5 + 5 d- 1. The characteristic time for
the droplet to form or grow to scale L is that necessary to surmount this
energy barrier. Thus, a droplet of scale L will last for

T=r,exp(&) or In(~)-& (5.85)

When the autocorrelation function C(t) is derived, the above thinking leads
to an extremely slow logarithmic decay of temporal correlations
C(t) - (ln t)+‘* (5.86)
Further calculations within the droplet model give a ‘l/f-noise’ spectrum
proportional to l/olln oJ1+@‘JI) and real and imaginary parts of the
susceptibility

x’(o) - (In o[-~‘+ and x”(o) - Iln w(-~+(~‘+) (5.87)

Finally, the ac non-linear susceptibility scales as
~~~(30; 0) - Iln w(ld-(l++)el’JI (5.88)
and this diverges for o + 0 if d > (1 + +)f3. For the 3D spin glasses the
droplet model has assumed that conventional dynamical scaling is operative.
As we shall see in the next chapter there exists a non-conventional dynamical
scaling for 2D spin glasses.
Since the theory is so recent there have been very few meaningful
experimental tests of the above predictions. One problem is the lack of an
‘ideal’ 3D short-range Ising spin glass (see Chapter 6). Another is the
popularity of the Parisi RSB solution to the mean-field-SK model on which
166 Models and theories

most measurements have concentrated. So at this stage it is impossible to

say which of the two gives a better description of the spin-glass phase,
although the ‘noise’ people are strictly for RSB.
We proceed now to consider the non-equilibrium behaviour and its
approach to equilibrium of the ordered spins within the droplet model. If
we quench a spin glass in zero magnetic field from T = m to T a T,, a
mixed of I’ and r (reversal) domains are formed. After the quench, the
system will try to lower its free energy by decreasing the amount of interface
between I and r, thereby growing larger and larger domains of I+ and p.
For a spin glass this is a tediously slow procedure because of the randomness-
induced free-energy barriers. Such barriers must be surmounted before
sections of the wall between I and r can be moved. Since B = A(T)L’J’, in
a time t after the quench, the characteristic length scale of the domains
Rta grows as

(5.89)

where t, is the total age of the system after the quench and f, is a
microscopic unit of time (to = 1 is the usual basis). This growth of domains
is the fundamental process of the long-time, non-equilibrium dynamics of
the spin glasses below Tf. A corresponding ‘quench’ can be performed by
turning off (H = 0) an infinite magnetic field at T G Tf. The magnetization
is expected to decay with the formation of domains according to

(5.90)

where A is a new dynamic exponent unrelated to the equilibrium ones.

Now that we have established the length scale that determines the
approach to equilibrium, we can consider waiting-time or ageing effects.
These are the experiments described in section 3.3.1, where, for example,
a sample can be cooled in zero field, and after waiting a time t, a weak
field is turned on and the magnetization ‘measured as a function of time,
m(t). Or, the sample can be field cooled, and after waiting t,,, the weak
field is switched off and m(t) tracked in time. In these experiments,
relaxation processes are probed on a length scale L which begins at time t
when the field is changed. Here
$
(5.91)

Let us treat the processes which occur in the two limiting cases: In t 4 In t,
and In t % In t,. For the first case we have

(5.92)
5.10 Fractal cluster model 167

where now ta = t,,, + t. For all practical purposes Rta will be constant in
this ‘early epochs’ regime, and since L,, or experiment, probes length scales
which are much smaller, we use the term quasi-equilibrium dynamics. In
other words, our experimental scale is so small that it does not see the
domain walls, only the pure states I and r. So to a first approximation the
relaxation processes are characteristic of equilibrium. And the magnetization
should decay as

-- H
m(t) (In t)W
with 8 and JI the equilibrium exponents. Figure 5.13 sketches the various
functions R’J’, L+ and the expected derivative of m(t) with respect to In t.
For the second case: In t % In t, (‘late epochs’), and therefore L, = R,
since t = ta = t + t,. Now the experiments are clearly seeing the domain
walls and thus the non-equilibrium dynamics are detected with

40

where naturally the non-equilibrium

- H x
[ 1
In t, 3

exponent A enters the magnetization

(5.94)

decay.
There are two crossover regimes of interest. The first is at In t = In t,
or where L, + R,, i.e., a transformation from quasi-equilibrium to non-
equilibrium dynamics. Here the measured magnetization decay shows an
‘S-shape which, when plotted as dm(t)ld(ln t), reflects a peak at t = t,; see
Fig. 5.13. The second crossover occurs in the limit of infinite time which
of course cannot be reached by experiment. At this point L and R have
reached macroscopic length scales so the number of domains and their walls
become insignificant. True equilibrium has finally been attained and the
magnetization and its slope have decayed to zero. In general, based on this
picture of the domains and their growth, we have a simple and qualitative
model for interpreting the long-time and ageing behaviours of the spin
glasses presented in section 3.3.1.

5.10 FRACTAL CLUSTER MODEL

A scaling theory, which makes particularly close contact with our experimen-
talist’s interpretation (Chapter 3), is the fractal-cluster model of Malozemoff
and Barbara (1985). They proposed a scaling theory of the spin glasses by
considering clusters of correlation length 5 which diverge as [(T- Tf)/7’r]-“.
These coherent regions have a cluster size sE on which all relevant physical
quantities depend and whose volume is SD, where D is the fructuZ dimension
because the clusters are expected to be highly irregular and branched. The
clusters have a magnetization mg - si or s, if the given cluster consists of
168 Models and theories

w
InO, 1
MACROSCOPIC
/ ------- LENGTHS
/’
04

0 I 0 EOUILIBRIUM
to tw In(t)
Fig. 5.13 Growth of domain size R and experimental-probing-length scale L versus
logarithm of time (+ is the barrier exponent); (a) t, = t+t,,, the total age of the
system; (b) t is the time of measurement, tw, the waiting time, and to, a microscopic
time; and (c) relaxation rate aM/a In t versus In t. From Lundgren (1988).

a random distribution of spins or a ferromagnetic one. Above Tf the regions

can be viewed as superparamagnetic clusters which rotate in response to
the applied field H giving a net magnetization

(5.95)

where n, is the number of clusters with spin s and 3 is the Brillouin

MYDOSH: spin glasses - an experimental mtroduction mydosh$p05 08-03-93 12:50:54

5.10 Fractal cluster model 169

function. The criterion m,H = kBT corresponds to a crossover permitting

the clusters to align along the field. Using

we can define the crossover exponent + as C$ = uD, and since T is not

critical, the crossover H-T line follows immediately

T-T, -$H = const. or

T- T,
T 0~Hz (5.97)
t-1Tf f

Now if + = 3, as is suggested by experiment, we obtain a new, scaling

derivation of the AT line without recourse to a mean-field instability. At
Tf an infinite percolating cluster forms which we call the frozen matrix.
Embedded in the matrix or exterior to it are the remaining correlated
regions or clusters which govern the low-temperature behaviour.
Below Tf, these correlated regions can be visualized as rotating in a
frozen matrix. Here the matrix imposes internal energy barriers which
hinder the rotations, but which may be overcome via thermal activation.
For further decrease of temperature the barrier heights stay fixed while the
size of the remaining clusters grows or becomes part of the matrix. Using
an Ising fractal-cluster model the net magnetization at T and H is an
average over all cluster sizes:

(5.98)

where we have applied a weak field so as to use the small-argument limit

of the Brillouin function. According to percolation-scaling theory, the
cluster-size distribution n, is given by

where se is the characteristic cluster size, r is another critical exponent and

fis the scaling function, i.e., f(x) = const. as x + 0 and f(x) + 0 for x > 1.
We now calculate the time decay of the magnetization for T < Tf when
the weak field H is switched off at t = 0.

M(t) [+f(:) (ep+)~ (5 loo)

-=
M(O) [s*-Tf(;) (!&) * *

This equation is derived assuming that the characteristic relaxation time t

of a given cluster is exponential and that tanh x s X. It is further
170 Models and theories

assumed that t’ depends on the size of the cluster as t’ = t, sX where

t, = t,, exp (EB/kB7) and x is a new dynamical critical exponent. Here the
effects of energy barrier EB activation have been placed in the &-constant.
Rewriting (5.100), we obtain

9-7 exp [ - (S/Q)“] exp [ - (t/fJ~-~] &

__ = (5.101)
cc
M(O)
sl-7 exp [ - (s/sJ”] d.s
I0
The percolation-theory result has been substituted for

S
= exp [ - (S/S&~] (5.102)
f(k)St
with T 5 Tf and sufficiently large clusters. In the long-time limit and
ignoring algebraic prefactors, the integration of the numerator gives

M(t) - exp [- (t/tl)‘-“1 (5.103)

where n = x/(x + u) and t1 - se. This is our old friend, the stretched
exponential, which has been used to describe the slow relaxation of real
glasses and also certain regimes of spin-glass relaxation. If we repeat the
above calculations for M(t), but at T = T,, we can use the percolation-
theory result right at the percolation threshold: n,(T,) = s-‘l. In this case
the integral for M(t) becomes
D
M(t) a exp [- (t/t&-“] ds a : 6r (5.104)
0

i.e., a power-law behaviour is obtained with exponents D/z = l/x and

l/6 = T - 2. Comparatively, there is a similar expression found in the
dynamic of the mean-field theory and some experimental indications for
such intermediate behaviour have been reported.
We could proceed further with this model and determine the dynamical-
scaling behaviour of the H-T phase boundary, the so-called critical line.
As the flavour is clear we stop. The utility of this fractal-cluster approach
is that it begins very practically, employing the percolation picture demanded
by experiments. Then using various scaling functions it attempts to calculate
a number of physical quantities which seem to agree with measurement.
With greater efforts the model could probably be extended to addition
parameters both at and below Tf. Since the clusters and percolation are so
helpful in visualizing the physical phenomenon in a spin glass, any theory
done with them provides a viable alternative to the rather abstract mean-
field approach.
5.11 Non-lsing spin glasses 171

5.11 NON-ISING SPIN GLASSES

Up until now we have been mainly considering Ising spin glasses. We will
now relax this restriction and briefly deal with m-component vector spins
Si, such that Si - Si = tn. The spin-glass order parameter is now a tensor in
spin space which for H = J, = 0 reduces to a simple scalar, i.e., the system
will be isotropic in spin space. And here one can repeat the standard
procedures of RSB and obtained modified solutions and properties for the
static and dynamical behaviours.
A more interesting case is the isotropic spin glass in non-zero field with
J, = 0. Now M = ((S$)c # 0 and qll = ((S~l’)T(S~l)),), # 0 for the
parallel (to the field) magnetization and spin-glass order parameters,
respectively. In addition, there are various transverse order parameters
which allow the possibility for transverse freezing

bc/&jl = (($%($%>c (5.105)

with k, k’ = 2, 3, . . . . m. The line in the H-T plane where q1 first becomes
non-zero is called the Gabay-Toulouse (GT) line (1981) and it varies for
small fields as
k,K-k(H)I = in + 4 (5.106)
A
recalling that A is once again the variance of our gaussian P(J,) distribution.
A sketch of the GT-line is shown in Fig. 5.14 along with the AT line.
Notice the different H(T) dependences between the two lines representing

Fig. 5.14 Sketch of field-temperature phase diagram for the mean-field model with
vector spins: (a) freezing of transverse spin components with weak irreversibilities
(GT line); and (b) onset of strong irreversibilities with RSB (AT line).
172 Models and theories

the (GT) transverse freezing ifiT, 0: HZ, followed by the (AT) longitudinal
freezing 6Tf 0: H% RSB occurs at both lines and the transverse spin-glass
susceptibility is expected to diverge as [T- T&H)]-l for T + T&(H). If
we additionally let J, (the mean of the gaussian P(J,) distribution) be non-
zero, Gabay and Toulouse predict the more complicated ‘re-entry’ T-J,
phase diagram shown in Fig. 5.15. Here there is a set of transitions:
paramagnetic to ferromagnetic to the first mixed phase Mr where ferromag-
netic and spin-glass (frozen transverse components) phases coexist. Finally
Mi + M2 signifies another transition where a crossover occurs to a new
mixed state with strong irreversibilities as opposed to the weak irreversibilities
in Mi.
Continuing with the vector spin glasses we next discuss systems with a
uniaxial anisotropy. Here the hamiltonian becomes from equation (1.4)

X = - {&J& . Si - Do’ (5.107)

I, I
where the Sl” spin component is taken along D, the magnitude of the
uniaxial anisotropy which has a lixed direction. If D > 0, ordering is
preferred along this direction (longitudinal ordering), whereas transverse
ordering is favoured for D < 0. The phase diagram can be calculated within

Para

Ferro
/

Spin glass

0 1

Jo/A
Fig. 5.15 Phase diagram for vector spins interacting with infinite-ranged gaussian
distribution (JO is the mean and A the width or variance). M1 and M2 represent
mixed magnetic phases of the re-entry transition. From Gabay and Toulouse (1981).
.ll Non-king spin glasses 173

<SB theory setting J, = 0. The results are shown in Fig. 5.16. ‘P’ represents
he paramagnetic phase at high temperature, ‘T’ the transverse phase with
fll = 0 and ~7~ # 0, ‘L’ the longitudinal phase with qll # 0 and q1 = 0,
.nd a mixed phase, ‘LT’ is also present, where both order parameters are
lot equal to zero. A variety of different phase transitions is predicted with
wo possible types of entry into the mixed LT phase. Experiments, mainly
sing crystalline anisotropy as the source of D, seem to exhibit the L or T
ransitions, however, values of D are not accessible in real materials to test
or the LT phase.
We conclude this section by mentioning the theoretical conjecture that
he Dzyaloshinskii-Moriya (DM) interaction (see section 1.4) and its
orresponding random anisotropy in an isotropic (or Heisenberg) spin glass
an be converted by an applied field into an Ising spin glass. The small
leld reorganizes the random DM directions to make a component of them
loint along the field. This then establishes an Ising-like (longitudinal) frozen
lignment and changes the universality class from Heisenberg to Ising. With
urther increasing field the behaviour transforms to a new regime which
xhibits both a transverse freezing with RSB at higher temperatures followed
bya crossover to longitudinal freezing (strong irreversibility) as T is lowered.
he two H-T lines are similar to the GT and AT lines - see Fig. 5.14 -
nd note the different H(T) dependences. Experimental searches for these
arious lines and crossovers have had to rely on a precise determination of
he freezing temperature over a wide range of magnetic fields. As we have

‘lg. 5.16 Phase diagram of an infinite-range vector spin glass with uniaxial
nisotropy D; from Roberts and Bray (1982) and Cragg and Sherrington (1982).
174 Models and theories

previously discussed (section 3.4) this is impossible. Furthermore, dynamical

effects are playing an important role near T,, hence the resolution of a so-
called critical temperature is strongly contingent on the type and time of
measurements. And the controversy remains that the different lines (AT
or GT) are mostly dynamical in nature and not the result of a mean-field
phase transition as predicted by RSB and the SK model.

5.12 COMPUTER SIMULATIONS

Spin glasses are a natural choice for numerical simulations and the Monte
Carlo method lends itself readily for computational studies. These computer
experiments serve to span the gap between the measurements done on non-
ideal materials and the analytic calculations performed via approximate and
oversimplified theoretical models. But first what is the Monte Carlo (MC)
method as applied to a spin glass?
We pick a lattice in arbitrary dimension usually 2 or 3, but depending
on computer power we can go even higher, e.g. d = 5. Sizes of the lattice
or number of spins can reach 64d using fast special purpose machines
especially built for simulations of random spin systems. One then selects
the number of spin components: m = 1 Ising, m = 2 x-y, and m = 3
Heisenberg. Finally, we need a probable distribution P(J,) for the random-
bond configurations. This is customarily taken as the EA or SK gaussians
or simply +J. Although a more realistic model, the canonical RKKY-
random site ‘CuMn’ situation, has also been MC simulated (see Chapter
8). Once we have decided upon the desired spin-glass ‘system’ via the above
choices and with a given initial configuration of energy U,,, the game starts
by flipping one spin. If the energy change U1 - U, < 0, the spin flip is
accepted and we repeat the process for a second spin giving a U,. However,
if the energy change U1 - U, > 0, then we calculate the exponential
B = exp [-(UI- U,)Ik,T] and compare the result with a random number
R between 0 and 1. If B > R, then the spin flip is accepted (Metropolis
acceptance criterion) and we go on to a U,. If B < R, the spin flip is
rejected and we return to our original initial condition U,, and start all over
again by flipping another spin. Each time we attempt a spin flip, we set
the time scale with MC step/spin (MCS). This for a good equilibrium
ensemble average can reach lo7 or 108 MCS. For better averaging we
should repeat the calculations with 30 to 100 realizations {Jij} of the random-
bond configuration. In addition, the initial condition should be checked for
not influencing the results.
Such recursive procedures give a nearly exact calculation of the partition
function Z{Jij} averaged over the various sets of {Jij}. MC simulations have
been employed from almost the beginning of the spin-glass problem. The
early computations (late 1970s) focused upon 2D lattices with Ising spins
(Glauber dynamics) and a gaussian distribution of near-neighbour bonds.
5.12 Computer simulations 175

The temperature was normalized to the distribution width. Interestingly,

these pioneering efforts found susceptibility and specific-heat behaviours
mimicking those of experiments on the canonical spin glasses. Furthermore,
the dependences at T < Tf for the remanent magnetizations (IRM and
TRM) on field closely correspond to those measured on AuFe - see
Fig. 3.32. These initial results were hampered by insufficient lattice size
(finite-size effects) and not enough MCS to obtain true equilibrium
(dynamical effects or relaxation times are longer than MC ‘observational
times’). Nevertheless, in recent years such difficulties have been surmounted
and meaningful conclusions have been drawn and reasonable predictions
can be made.
The time-dependent EA order parameter q(t) may be directly obtained
from the standard MC time averaging (t is measured in MCS)

(5.108)

Accordingly, the time-dependent (normalized) susceptibility becomes

x(t)= bT 11- 4(01 (5.109)

and can be interpreted in terms of the Sompolinsky/Parisi (section 5.8)

dynamical susceptibility. Furthermore, we can proceed with the two-spin
correlation function and define the time-dependent, non-linear susceptibility

)(“I = C((SJj)$), = (5.110)

i
where

((SiSj)+), - y -’ - exp (-Rgl&) (5.111)

i )
and esG is the spin-glass correlation length

(y and u are the usual critical exponents associated with a possible spin-
glass transition). Again using MC simulations these quantities may be
numerically computed. And in a similar manner just about any quantity of
physical or theoretical interest could be calculated for the various models
of spin glasses.
Rather than delve into the many such MC experiments we limit ourselves
to discussing the main and accepted (by the MC experts) results. Remember
these numerical simulations are a middle route between theory and
experiment, and they are most useful as a test or proving grounds for one
176 Models and theories

or the other. Hopefully, as is the case in simpler problems, all three should
agree.
In huge simulations on a specially built processor Ogielski and Morgenstern
(1985) considered the Ising, +J, 3D spin glass. The static properties showed
a sping-glass ordering transition at Tf = 1.175J with critical exponents for
the correlation length u = 1.3 and for the non-linear susceptibility y = 2.9.
Working within the mean-field RSB approach, Bhatt and Young (1985)
compared order-parameter distributions P(q) computed in different ways.
A finite-size scaling function was derived from the moments of P(q) which
should be independent of size at Tf and then splay out again at lower
temperatures. The various size curves do indeed cross at T = 1.755 for
d = 4. In the case of d = 3 they come together at T z 1.21, but stay
merged as T is reduced below Tf. This indicates that d = 3 is very close to
the lower critical dimension dl below which T, = 0. Additional scaling fits
yield y = 3.0.
In another series of computer simulations called domain-wall renormaliza-
tion-group techniques, Bray and Moore (1984) and McMillan (1984) further
studied the dimensionality of the spin-glass transition. In this method one
computes the variation of the ground-state energy 6E, by changing the
boundary conditions from periodic to antiperiodic in one direction. For a
spin glass (6E.J = 0, but (6E2,); and (ISE,I) are not zero. So if the energy
change decreases with system size, then there is no order because at some
length scale SE, = kBT, and this will disorder the system. The results
indicate order for d = 3, but not for d = 2.
Naturally, these techniques can be repeated for isotropic (Heisenberg)
vector spin glasses with short-range *J or gaussian bond distributions in 2,
3 and 4 dimensions. Even the long-range RKKY interaction may be included
in the simulations. The results suggest that Tf = 0 for d = 2 and 3 isotropic
spin glasses and that dl = 4 (above which Tf should be finite). Also the
long-range RKKY interacting spin glasses are at, rather than below, their
lower critical dimension.
MC simulations (Ogielski, 1985) have also been used to determine the
spin-glass dynamics via the behaviour of q(t), the dynamic correlation
function defined above. Figure 5.17 exhibits how q(t) decays in time (MCS)
at various temperatures spanning Tf. Note the system size is 323 = 32768
spins and about 5 x lo7 MCS have been performed. Once one knows the
q(t) dependence various analytic fits can be made. For example, above Tf
the functional form is
q(r) = C,t-” exp [ - (wt)a] (5.113)
where both exponents x and B depend on the temperature. Above the
Curie temperature of the non-random 3D Ising model (T, = 4.55) x = $
and B = 1. While from T, down to Tf (the Griffiths phase) j3 is between
0.4 and 1 indicating that the time decay is a stretched exponential or
Kohlrausch function. Below Tf it changes to power law
5.12 Computer simulations 177

2.

loo IO’ 102 I03 I04 105 I06 IO’ 106

nhtE.wcs)
Fig. 5.17 Dynamic correlation function q(t) calculated by Monte-Carlo simulation
(lattice size 323) for temperatures around and below Tf = 1. From bottom to top
T = 1.30, 1.25, 1.20, 1.10, 1.00, 0.90 and 0.70. From Ogielski (1985).

q(t) = Cot-” T 5 Tr (5.114)

with x very small (- 0.1 at Tf decreasing to = 0.01 at $T,, the lowest
temperature of the simulation). This is such a tiny exponent, i.e., the time
dependence is exceedingly slow, that a In t or even a ln(ln t) behaviour
may also be a suitable fit to the computer plots in Fig. 5.17.
If we extract from q(t) the relaxation time T used in the dynamical scaling
of equation (5.73), then we can estimate the dynamical exponent z. This is
accomplished by the proper integration of q(t), namely
m
t q(t) dt
Pi.= I 0 (5.115)
co
4(t) dt
I0
By performing the above integrals at various temperatures approaching T,,
we can plot log 7 versus log (T - TJT,. Figure 5.18 shows one such graph
for a 643 system where the slope gives zu = 7.9 with the ‘proper’ choice of
Tf (= 1.175J). The determination of Tf (here it is a critical temperature) is
very difficult also in the computer experiments as it is in a real experiment.
Note that T starts to exceed the maximum MCS (here = 5 x 1Os) at about
T = 1.305. This means that a reduced temperature of only
(T - T&T, = 10-l can be reached. So the fit in Fig. 5.18 is restricted to
less than a decade. Not very good for the extraction of meaningful
178 Models and theories

IO'
IO0 J-L&l_ I1 III
10-I IO0
(T/T,,-I)

Fig. 5.18 Power-law fits for the relaxation time T versus (T- 7’,)-‘” from the
Monte-Carlo simulations for lattice size 643 and T 2 1.30; from Ogielski (1985).

exponents, which also depend on the exact ascertainment of Tf. Nevertheless,

the above simulations are the biggest and longest of the 1980s. And they
show a rapid increase described by an unusually large exponent (zu) power
law as T+ T;.
Various computer studies all agree that Tf = 0 for the 2D Zsing spin
glass. The dl lies between 2 and 3, and the correlation length SSGand non-
linear susceptibility xn, should both diverge at T = 0 K. The critical
exponents v and y are different depending upon which distribution is used.
For the k.Z distribution v is about 2.6 and y between 4 and 5.5. From the
scaling relation y = (2 - q)v we can estimate q the length exponent of the
correlation function. For the gaussian distribution u is around 3.5 to 4.4
while y can be above 7 even reaching 9. Substituting these ‘value spreads’
into the above scaling-equation results in an estimate for q = 0. Presumably,
the large ground-state degeneracy of the +.Z discrete model places it in a
different universality class from models with a continuous Jij distribution,
if Tf = 0.
Now we proceed down into the frozen spin-glass state. Here the question
is how to probe the equilibrium state by computer experiments which are
always out of equilibrium. We rely on the two main theoretical models
which have been previously proposed. The first is the mean-field SK model
with Parisi RSB solution. This model leads to many thermodynamic states
unrelated by symmetry and differing little in their ground-state free energy.
Further reading and references 179

The distribution of order parameters P(q) is an important property of the

solution. Below T, it should exhibit a peak at finite q and a tail with finite
weight at q = 0. Using computer simulations of the SK model such a form
for P(q) has indeed been found. The alternative model is the droplet
approach of Fisher and Huse where there is only a single thermodynamic
state I whose excitations are droplets of globally reversed spin directions
r. For a finite system these states can produce a finite P(q) down to q = 0.
However, P(0) should vanish as Lme as the linear size of the lattice L goes
to infinite. Recent MC calculations for a d = 4 Ising spin glass displayed
very little reduction in the P(0) tail as L was increased. Therefore, at the
present state of numerical simulation, the results seem to be favouring the
SK-Pa&i theory over the droplet model. The above examples have nicely
illustrated the power and effectiveness of special-purpose computers in the
study of spin glasses.
With this cursory review of the spin-glass models and theory we conclude
the chapter. The reader is referred to the listings in Further Reading and
References for more details and full derivations of the theoretical results.
As mentioned earlier the spin-glass problem has not yet been solved and
new developments should be expected, perhaps requiring a future addendum
of the most recent progress. We now proceed to the ideal spin glasses and
their comparison with theory.

FURTHER READING AND REFERENCES

Statisical mechanics

Tolman, R. C. (1962) Principles of Statistical Mechanics Oxford.

Huang, K. (1963) Statistical Mechanics Wiley, New York.
Landau, L. D. and Lifshitz, E. M. (1989) Statistical Physics Pergamon, London.
Ma, S.-K. (1985) Statistical Mechanics World Scientific, Singapore.
Ziman, J. M. (1979) Models of Disorder Cambridge.
Binder, K. (editor) (1986) Monte Carlo Methods in Statistical Physics Springer,
Heidelberg; (1987) Applications of the Monte Carlo Method in Statistical Physics
Springer, Heidelberg.

Phase transitions

Brout, R. H. (1965) Phase Transitions Benjamin, New York.

Stanley, H. E. (1971) Introduction to Phase Transitions and Phenomena Oxford.
Domb, C. and Green, M. S. (eds) (1972) Phase Transitions and Critical Phenomena
Academic, New York.
Hohenberg, P. C. and Halperin, B. I. (1977) Rev. Mod. Phys., 49, 435.
Ma, S.-K. (1976) Modern Theory of Critical Phenomena Benjamin, New York.
180 Models and theories

Amit, D. J. (1984) Field Theory, the Renormalization Group and Critical Phenomena
McGraw-Hill, New York.

Theoretical spin-glass reviews

Chowdhury, D. (1986) Spin Glasses and Other Frustrated Systems Princeton.

Binder, K. and Young, A. P. (1986) Rev. Mod. Phys., 58, 801.
Fisher, K. H. and Hertz, J. A. (1991) Spin Glasses Cambridge.
Mezard, M., Parisi, G. and Viraso, M. A. (1987) Spin Glasses Theory and Beyond
World Scientific, Singapore.
Moorjani, K. and Coey, J. M. D. (1984) Magnetic Glasses Elsevier, Amsterdam.

References

Anderson, P. W., Halperin, B. I. and Varma, C. M. (1972) Philos. Mug., 25, 1.

Bhatt, R. N. and Young, A. P. (1985) Phys. Rev. Lett., 54, 924.
Blandin, A. (1961) Thesis, University of Paris (l%l), unpublished; (1978) J.
Physique, 39 C6, 1499.
Bray, A. J. and Moore, M. A. (1984) J. Phys., C17, L463 and L613.
Cragg, D. M. and Sherrington, D. (1982) Phys. Rev. Lett., 49, 1190.
de Almeida, J. R. L. and Thouless, D. J. (1978) J. Phys., All, 983.
Edwards, S. F. and Anderson, P. W. (1975) J. Phys., F5, 965; (1976) ibid 6, 1927.
Fischer, K. H. (1975) Phys. Rev. Lett., 34, 1438.
Fisher, D. S. and Huse, D. A. (1986) Phys. Rev. Lett., 56, 1601; (1988) Phys.
Rev., B38, 386.
Gabay, M. and Toulouse, G. (1981) Phys. Rev. Lett., 47, 201.
Klein, M. W. and Brout, R. (1963) Phys. Rev., 132, 2412.
Larkin, A. I. and Khmel’nitskii, D. E. (1970) Sov. Phys. - JETP, 31, 958; (1971)
ibid, 33, 458.
Lundgren, L. L. (1988) J. Physique, 49 - C8, 1001.
Malozemoff, A. P. and Barbara, B. (1985) J. Appl. Phys., 57, 3410.
Marshall, W. (1960) Phys. Rev., 118, 1519.
McMillian, W. L. (1984) J. Phys., C17, 3179; (1980) Phys. Rev., B30, 476.
Mezard, M., Parisi, G., and Viraso, M. A. (1987) Spin Glasses Theory and Beyond
World Scientific, Singapore.
Ogielski, A. T. (1985) Phys. Rev., B32, 7384.
Ogielski, A. T. and Morgenstern, I. (1985) Phys. Rev. Lett., 54, 928.
Parisi, G. (1979) Phys. Rev. Lett., 43, 1754; (1980) J. Phys., A13, 1101; (1980) ibid
13, 1887; (1980) 13, L115; (1980) Phys. Rep., 67, 97.
Parisi, G. and Toulouse, G. (1980) J. Physique Lett., 41, L-361.
Phillips, W. A. (1972) J. Low Temp. Phys., 7, 351.
Roberts, S. A. and Bray, A. J. (1982) J. Phys., C15, L527.
Sherrington, D. and Kirkpatrick, S. (1975) Phys. Rev. Lett., 35, 1792; (1978)
Kirkpatrick, S. and Sherrington, D. (1978) Phys. Rev., B17, 4384.
Sompolinsky, H. (1981) Phys. Rev. Lett., 47, 935; (1981) Phys. Rev., B23, 1371.
Thouless, D. J., Anderson, P. W. and Palmer, R. G. (1977) Philos. Mug., 35, 593.
Walker, L. R. and Walstedt, R. E. (1977) Phys. Rev. Lett., 38, 514.
6
Ideal spin glasses:
comparisons with theory

Now that we have toured the main spin-glass experiments, listed the various
materials and scanned the highlights of the theory, how close can we relate
theory to experiment on the most ideal systems? We do not wish to make
a detailed comparison of every aspect, but chiefly to show which models
can be mimicked by experiment and whether their predictions are valid
especially regarding the freezing transition. Remember we have our
experimentalist’s rendition of the basic phenomenon: growing and freezing
clusters with a broad and shifting distribution of relaxation times. Theory,
if it is correct, can take us much further with a quantitative description of
experiment and the predictions of new effects. Once we gain a sufficient
rapport between the two, experimentation and theory, the problem is solved
and the physics is said to be understood. However, for the spin glasses the
full solution has not yet been attained. There is no complete quantitative
comparison or detailed set of predictions directly relating to experiment.
Great progress has indeed been made and is continuing with the various
models and calculations, particularly the computer simulations. On the
other hand, the majority of the measurements have been performed on
systems which are too complicated and greatly differ from the simplified
theoretical models. So how can you expect to compare experiments on a
complex material with an idealized theory?
In this chapter we adopt the point of view that real spin-glass systems
must be made or chosen to resemble intimately the theoretical model. And
these special materials we call the ideal spin glass. We distinguish an ideal
spin glass from a canonical (RKKY) or good spin glass - these generic
terms were frequently used in the previous chapters - as follows. An ideal
spin glass completely conforms to the theoretical model under consideration.
In contrast, a canonical or good spin glass exhibits the main experimental
features of a spin glass exposed in Chapter 3. There is no question about
the latter being a spin glass with a sharp freezing transition. However,
complications are present due to its universality class, deviations from
randomness, departures from pure RKKY behaviour, higher spin values,
orbital moments, undefined anisotropy, etc. Such aberrations add to the
182 Ideal spin glasses: comparisons with theory

complexity and prevent a meaningful comparison with the simple theoretical

models.
First, we explain what is an ideal spin glass in the abstract. Then, we try
to come as close as possible with some recently discovered mixed compounds.
Here space dimensionality will play an important role as we are only
interested in Ising spin glasses. For the 2D and 3D cases we examine the
experiments and draw a comparison with the existing theoretical results
which are mainly related to the static and dynamic scaling behaviours of
the freezing transition. As we shall see there are weaknesses from both
sides, but such is the present state of the science of spin glasses. Finally,
some future possibilities are offered for the fabrication of new ideal spin
glasses using the many sophisticated techniques of modern materials research.

6.1 WHAT IS AN ‘IDEAL’ SPIN GLASS?

The many canonical or good spin glasses, which exist in nature and possess
the prerequisite experimental criteria, are not at all ‘ideal’, i.e., closely
conform to the basic theoretical model (see below). For example, most of
the natural spin glasses are Heisenberg systems with site randomness. Very
many, especially the metallic alloys, are severely complicated with local
chemical clustering that leads to atomic, and in turn, to magnetic short-
range order. While the RKKY interaction is long range, it is greatly
modified by the magnetic clusters and may even taken on different distance
dependences for real metals. Another possibility is that a truncated spin
density wave will form. Molecular-field theory is inappropriate to treat
these effects.
Other spin glasses do not have fully localized moments and are intricated
with spin fluctuations or the Kondo effect. Still more materials are unknown
regarding their exchange interactions and, in the case of metals, are
arbitrarily defined as proper RKKY systems. Nevertheless, despite these
difficulties, such systems as (canonical) CuMn and AuFe or (good) (EuSn)S
and a-GdAl, are often assumed to be ideal and compared in great
detail to Ising-model, random bond, mean-field theory. Insulating mixed
compounds also have their difficulties with only antiferromagnetic exchanges
or complicated crystal structures and Ising systems are usually hard to find.
We now propose to create an artificially structured spin glass which will
fully satisfy the theoretical models and their computer simulations. Figure
6.1 shows a 2D projection of our ‘ideal’ spin glass. The ‘material’, which
also could be 3D, is an insulating Ising system with good local moments
(S = !J occupying regular lattice sites of a simple crystal structure. In Fig.
6.1 the magnetic sub-lattice is square and the spins can point only up or
down (Ising). An interpenetrating sub-lattice is available for non-magnetic
elements. Here we randomly distribute two types of ligands which control
the superexchange coupling. The exact ligand configuration between two of
6.1 What is an ‘ideal’ spin glass? 183

Fig. 6.1 Schematic in 2D of a 3D king ideal spin glass formed by ?J distribution

of bonds. Note the randomly distributed ligands (small open and filled circles).

the moments will govern their alignment: parallel or antiparallel. This

results in a short-range, well defined +J exchange interaction which can
be adjusted by varying the relative concentrations of the two different
ligands represented by the small open and filled circles in Fig. 6.1. With
this abstract creation we have constructed an Ising, spin 1, random-bond
(Edwards-Anderson), +J spin glass. Such we dub ideal, as it closely mimics
the simplest of theoretical models and therefore merits experimental study.
It is all well and good to fantasize with sketches, but how close can we
come in practice? Presently, there are no artificially made spin glasses that
satisfy all of the above model criteria. The full power of modern material
science, e.g. sputtering deposition, molecular-beam epitaxy, etc., has not
yet been brought to bear on fabricating an ideal spin glass. We should add
that some of the above techniques have been used to examine the effects
of a reduced dimension, 3D + 2D, using thin films of mainly CuMn and
AuFe. However, nature and our chemist friends have provided us with two
insulating compounds that are reasonable approximations of the ideal spin
glass defined above. In the subsequent two sections we shall examine in
some detail these materials, 2D-Rb$u,-,Co,F, and 3D-Fel-,Mn,TiO,,
and see what comparisons and conclusions may be drawn from the theoretical
predictions. At the present moment this is the best we can do, but hopefully
in the future better ‘real-artificial’ materials will be concocted that do an
even better job of being ideal. For, only then can the question of a true
phase transition be fully resolved.
184 Ideal spin glasses: comparisons with theory

6.2 Rb,Cu,-$o,,, AN IDEAL 2D SPIN GLASS

The crystal structures of both RbzCuFd and Rb2CoF, are the K,NiF,-type,
meaning a body-centred tetragonal sub-lattice of magnetic Cu or Co atoms.
Since the spins are separated by such a large distance along the long or c-
axis, the compound is magnetically quasi-2D. In the decoupled (u-b) planes
a simple square lattice of strongly interacting moments is formed. Rb,CoF,
is an archetypal Ising (up/down spins along the c-axis) antiferromagnet with
T, = 103 K. Rb,CuF, possesses a small x-y anisotropy which causes the
spin to lie ferromagnetically aligned in the u-b plane, Tc = 6 K. The
magnetic interactions are superexchanged via the F-ligands and fall off very
rapidly such that the next nearest-neighbour coupling is only 1% of the
nearest neighbour one. Thus a nearest-neighbour exchange is fully sufficient
to describe the magnetic behaviour of this system.
Now instead of mixing ligands to create random bonds as in Fig. 6.1, we
substitute Cu for Co which effectively mimics bond randomness. The reasons
for this are as follows - recall section 1.35. The Cu2+ ion has two types
of ground state orbitals whose lobes alternate along the x and y axes. As
nearest neighbours two Cu ions are ferromagnetically coupled and their
lobes are always orthogonal. But a Cu lobe pointing towards a no-lobe Co
neighbour (remember the square-lattice symmetry) generates an antiparallel
alignment. For a Cu lobe perpendicular to its Co neighbour, a parallel spin
pair results. Finally, Co-Co nearest neighbours are always ferromagnetic.
A sketch of the lattice with its ‘random’ and competing bonds is given in
Fig. 6.2. Notice we have two ferro ( t t ) and two antiferromagnetic ( t 4 )
interactions. These can be varied by changing the concentration x, and for
a particular x = 0.29 the resulting P(J) was already shown as part of the
introduction to spin glasses in Fig. 1.8. In essence we are now repeating
what has been previously stated in section 1.3.5 for a random-bond spin
glass.

Fig. 6.2 Representative distribution of ferromagnetic (F): Cu-Cu and Cu-Co and
antiferromagnetic (AF): Co-Co and Cu-Co exchange interactions in RbZCul-$ZoxFd.
Also indicated are the in-layer d-orbital lobes of Cuz+. From Dekker (1988).
6.2 Rb2Cu1&oxFq, an ideal 2d spin glass 185

The Co, although in the minority, plays a very interesting role. Because
it experiences a tetragonal crystal field, Co*+-ions have only their lowest
Kramers doublet occupied. Thus its effective spin is i, exactly the same as
the Cu. Furthermore, the single-ion anisotropy of the Co*+ is such as to
create a good Ising, uniaxial-oriented system. Last but not least, when
single crystals are grown, and this procedure is not without difficulty, they
form without short-range chemical order. So there is a true distribution of
random bonds unhampered by any atomic clustering. Consequently, the
Rb2Cu1-$oxF4 compounds represent an Ising, spin i, random bond, short-
range (-+J), square-lattice spin glass. This set of adjectives closely conforms
to our ideal sketch in Fig. 6.1 with the exception of the symmetric +J
distribution. For the real case there are 2+Js and 2-Js, and, while the
relative probabilities Z’(J) (see Fig. 1.8) may be adjusted by varying x, the
asymmetric magnitude of the four Js is an aberration from our ideahzation.
The next step is to optimize the x-value and to demonstrate that the
system is not long-range ordered but indeed a random freezing occurs at
some temperature. In the present 2D case we expect Tf = 0. Experiment
establishes the spin-glass regime to exist for 0.18 < x < 0.40, with a good
working concentration x = 0.22. The usual measurements of ac- and dc-
susceptibility, magnetization and specific heat are carried out to prove the
absence of long-range order, i.e., a featureless specific heat, and the
beginning of a random freezing via the time (frequency) and field (FC
versus ZFC) dependences of the other quantities.
Once the spin-glass properties have been firmly established and a T-x
phase diagram mapped out, we assume a phase transition transpires at
some T,. Since we are being guided by theory which would predict a critical
temperature T,, we have changed our notation from Tf to T,. The static
critical behaviour can be studied through the non-linear susceptibility xnl.
Here a dc method in different applied fields (vibrating sample magnetometer)
was used to determine x,,i taking advantage of the magnetization expansion
indicated in equation (3.41).

(6.1)

The definition of the non-linear susceptibility was already given in (3.49).

Remember theory predicts the linear susceptibility x0 = q/T to remain
non-singular at T,, below which a cusp appears for fast times and a plateau
for very long times (equilibrium). The other a,-coefficients should diverge
and depend on the critical exponents according to theory as
a3 m (T-T,)-r (6.2)
u5 oc (T-T,)-(*y+P) (6.3)
etc.
Assuming that x0 for Rb2Cu0,78C00.22F4 obeys a simple Curie law, i.e.,
186 Ideal spin glasses: comparisons with theory

possesses a symmetry distribution of exchange couplings, we can rewrite

(3.42) as
Xnl = a3(xcJW2 - %(XoW4 + *** (6.4)
with the coefficients reflecting the critical divergence. A typical plot for the
experimental data is given in Fig. 6.3, where the solid lines are a fit to the
above equation. The net result of this analysis is a resolution of the
temperature dependences of a3 and u5, and thereby the y and p exponents,
once T, is established.
However, a severe problem remains with the exact determination of T,,
the critical temperature. From the ac-susceptibility (see below) a cusp is
visible at about 3 K and is strongly frequency dependent, shifting downwards
in T with decreasing o. What then is the true T, in the limit o + O?
Theory has already prognosticated the answer, namely, T,(w=O) = 0. But
we want to test this conclusion via experiment. So, since we cannot let

0.7 I I I I I I Ii I I I I I I I I I I I
~1.67 K
v
~2.30 K
h -
0.6 -*2.75 K 0
03.05 K A
l 3.35 K .
-03.65 K v
2 0.5 A .
l 3.95 K 0
x 04.25 K p
.- VA .
s 0.4 0
v d
z
PA . .
e 0
:: 0.3
t VA
v b .
. 0
. 1
4
I A l

. I

b 0.2
.-iz
f
2 0.1

0.0

III III I I I I II I I I I 11 I
-0.1
0 50 100 150 200
Magnetic Field (G)
Fig. 6.3 Non-linear susceptibility xnl = l-(M/x,H) versus the internal field for a
selection of temperatures. Solid lines are fits to determine the a-coefficients. From
Dekker et al. (1988).
6.2 RbZCul--xCo,Fq, an ideal 2d spin glass 187

o = 0 which may be less than lo- lo Hz or some geological time scale, we

analyse the xni data with T, as a free parameter. Figure 6.4 shows the
results. Especially important is the statistical analysis performed on the data
using the x*-test. Note how x2 decreases and converges nicely to 1 as T,
+ 0. We judge our best fit to be with T, = 0, y = 4.5 and B = 0.0 (see
Fig. 6.4).
In order to check further this delicate point we can perform a scaling
analysis of the x,i-data. The predicted scaling relation for T, = 0 is

where the scaling function G(x) goes to c,x2 in the limit x + 0 and to 1
- c,x-“~ in the limit x + 03, with A = 1 + ‘2 (-y+B). By plotting x,T-P
versus H/TA at a variety of temperatures, excellent overlap is obtained -

10

Y- 6
A-
5
C.
“x 4

0 1 2 3

T, 09
Fig. 6.4 Results for x2 test (best-fit for T,), y, and 2y + l3 versus T,, obtained
from fitting the non-linear susceptibility with various values of ‘T,‘. Solid lines are
guides to the eye. From Dekker et al. (1988).
188 Ideal spin glasses: comparisons with theory

see Fig. 6.5. Superimposing data at different values of an experimental

parameter (T, H, etc.) is called a scaling plot and the best overlap is used
to determine the values of T, and the critical exponents. For Fig. 6.5 a
log-log plot is utilized and the superposition looks good except at the lower
values of the abscissa and ordinate. Accordingly, the best overlap gives
T, = 0 K, A = 3.2 and p = 0.0. Furthermore, if we used the scaling relation
appropriate to a 3D, T, # 0 system, fits or data overlap can be obtained,
but at the expense of large and unphysical values of the critical exponents.
Here there appears a definite trend, the lower T,, the better and more
meaningful the fits. This once again confirms our conclusion that T, = 0 in
accord with the theoretical prediction. At this point a very disturbing
question arises: how do we treat the cusp in the ac-susceptibility which is
usually taken not only as a measure of Tf but also of T, - the critical
temperature of the supposed phase transition? This should serve as a
warning to experimentalists - a cusp in xac does not automatically mean a
phase transition, or in other words Tf # T,! One must really perform a

10”

T,=OK
7
A = 3.2 , /e?= 0.0
yx 10-l if
l- l
$ ‘4.10 K
,A 0 3.95 K
4 =3.80 K
a* ~3.65 K
- 4v

,s’ ~3.35 K
03.50
4
l 3.20 K
v I I 1111111 I 11111111 I I llll~
lo-'
10-l loo 10' lo*
xoH/TA-'
Fig. 6.5 T, = 0 scaling plot of the non-linear susceptibility data. External fields
range from 1.0 G to 12.5 kG. From Dekker et al. (1988).
6.2 RbzCul-$o,F4, an ideal 2d spin glass 189

detailed analysis of xnl on an ideal spin glass before jumping to conclusions

about phase transitions. To shed more light on this problem let us consider
the dynamical behaviour of Rb2Cu0.78C00.22F4.
Now we measure the ac-susceptibility (x’ and x”) over as wide frequency
and temperature ranges as possible. In practice five or six decades of o can
easily be generated and T must span the region of maximum response in
both x’ and Jo’. Figure 6.6 exhibits these two susceptibilities from 0.3 Hz
to 50 kHz. Note the ‘cusp’ in both x’ and $’ which becomes sharper and
shifts to lower temperature. There are three distinguishing features of this
cusp behaviour (see Fig. 6.6) compared to that usually found in the 3D
spin glasses. Firstly, the maximum value (height) of x’ is very strongly
dependent on the frequency; a factor of three change is noted in the above
frequency range. Secondly, the susceptibility deviates from isothermal
already at T,,, meaning a x” contribution appears at this ‘high’ temperature.
And thirdly, X” strongly decreases at the lower temperatures instead of
remaining constant. Well how can we interpret this different behaviour
within the contexts of a dynamical phase transition? Answer: analysis of
the susceptibility data is necessary to extract a relaxation time.
As revealed in section 3.2.1, an average relaxation time may be extracted
from a broad distribution g(r) by employing the Cole-Cole phenomenological
approach. Here the basic equation is

x(w) = x’(w) - ix”(o) = xs + x0- xs

1 + (iorc)l--U
where xs and x0 are the adiabatic and isothermal susceptibilities, respectively,

I 1 I / , , ,
(a)

0-
012345678 012345678
Temperature (K) Temperature (K)

Fig. 6.6 (a) In-phase linear susceptibility ~‘(0, T) for Rb&u0.,&00.218F4 versus
the temperature. Data were taken at frequencies 0/27~ ranging from 0.3 Hz to 50
kHz, and with the driving field along the c axis. Solid lines are guides to the eye.
(b) Same as (a), but out-of-phase linear susceptibility x”(o, T). From Dekker et al.
(1989).
190 Ideal spin glasses: comparisons with theory

and r, is an average relaxation time. The fitting parameter cr determines

the width of the distribution such that cx = 1 corresponds to an infinite-
width distribution while a = 0 reverts back to the Debye form of a single
relaxation time. The trick now is to plot x” (x’) for the various frequencies
at each temperature (Argand diagram - section 3.2.1). At the maximum of
x”, WT, = 1, thereby establishing the value of T, for the given T. A collection
of such 7,s is shown in Fig. 6.7 as a function of temperature. The special
feature of the Cole-Cole approach is that it allows average relaxation times
to be derived even though they lie far outside the measurement window.
The experiments spanned a few seconds to a few microseconds. Yet the
analysis determines 7,s from lo6 down to lo-” seconds. Once we have
T,(T) plot as in Fig. 6.7, theory can be invoked.
A prediction from the droplet model (section 5.9) is that for a 2D Ising
?J spin glass, activated dynamics governs the dynamical critical behaviour.
Activated dynamical scaling leads to the following equation, derived from
(5.85), for the characteristic relaxation time,

(6.7)

3 4 5 6 7
Temperature (K)
Fig. 6.7 Relaxation time T, versus the temperature. Solid line is a fit to the
activated-dynamics equation. From Dekker et al. (1989).
6.2 RbzCul-,CoxF4, an ideal 2d spin ghs 191

where B is the barrier height (IJJis its exponent) and 5 m t-” for a zero
temperature critical point. Thus, according to theory T, diverges much faster
with temperature than a power-law divergence. Returning to Fig. 6.7, we
have displayed the fit to the above equation by the solid lines with +I.J = 2.2,
70 = 2 x lo-l3 s and B = 11 K - a physical set of values. Other fitting
procedures were attempted, e.g. power laws with T, = 0 or T, as a
parameter. But in all cases poor x2-tests and unphysical values of the
parameters resulted. The best agreement with experiment is clearly via the
activated dynamics of the droplet model with critical exponent IJJU= 2.2.
One can also use the Cole-Cole analysis to generate the form of the
relaxation-time distribution g(~). Figure 6.8 exhibits the evolution of g(~)
as a function of temperature. Notice the many decades shift with T especially
in the long-time tails. These enormous displacements and broadening are
all occurring far above T, = 0, at a few kelvin. Once again the dynamics
is dominant and such huge changes within the measurement window between
5.8 K and 3.4 K could easily and incorrectly be taken for finite temperature
phase transition.
Alternative methods exist to derive T, from the spin-correlation function
and by use of the fluctuation-dissipation theorem:
do = &w&)> (6.8)
and

x(0) = -x0 om[exp (-iwt) d$$

i
Two choices exist for the form of q(t):

(i) an exponential-logarithmic decay

0.06

0.05

0.04

0.03
0.02

0.01

0.00

“Log(T) (-r in s)
Fig. 6.8 Distribution of relaxation times g(T) according to the Cole-Cole analysis
for a selection of temperatures. From Dekker et al. (1989).
192 Ideal spin glasses: comparisons with theory

(6.10)

and

(ii) a stretched exponential

s(t) = expHWpl (6.11)

When the numerical integrations are performed to determine optimum
values of T, (and also x0 and 6 or B keeping T, = lo-13s), the temperature
can be varied to generate a plot similar to Fig. 6.7. Best fitting leads to
slightly different values for the $v exponent. For (i) the result is @J = 1.6,
while for (ii) Jlv is 1.9, not too bad results considering the small experimental
window compared to the enormous range (15 to 20 decades) of T,( 7’).
In order to double check the above approaches a dynamical scaling
analysis was carried out using the scaling function suggested by the droplet
model
x”T-P = F[-ln(wr,)T+‘] (6.12)
where p = -l-y+@. Excellent scaling exists over the entire range, except
near 07, = 1, with p = -3, $v = 2.2, T, = lo-l3 and most important
T, = 0. Also the data overlaps remain valid even for temperatures where
the system is clearly out of equilibrium. So the Cole-Cole approach and
similar x’ and x” analyses give quantitative conclusions which are consistent
with the dynamical scaling method.
Let us summarize these results and their meaning for this ‘ideal’ 2D spin
glass. First of all, Rb,Cul-,Co,F, has been shown to possess practically all
the specific characteristics of an Edwards-Anderson model spin glass. Table
6.1 lists and compares these characteristics for the system and for the

Table 6.1 Comparison of Rb2Cu1-xCo,F4 and the Edwards-Anderson spin-glass

model for d = 2

Edwards-Anderson
SG model for d = 2

dimensionality d=2 d=2

lattice square square
spin value s+ s=;
type of randomness bond randomness bond randomness
type of interactions Ising Ising
range of interactions nearest neighbours nearest neighbours
distribution of discrete (four competing Gaussian (continuous) or
interactions interactions shown in *J (discrete)
Figure 1.8)
6.2 Rb2Cu1--xCo,Fq, an ideal 2d spin glass 193

model, only with the distribution of interactions is there a deviation. Then,

for a limit x-range this system was shown to obey all the experimental
criteria related to spin-glass freezing. With these properties firmly established
we can delve more systematically into the freezing phenomenon. Our
experiments are non-linear susceptibility and ac-susceptibility. The former
determines the static critical behaviour, the latter the dynamics. Measure-
ment, using different forms of analysis to confirm the results, demonstrates
that T, = 0; y = 4.5, l3 = 0.0 and A = l+i(-y+P) = 3.2 for the static critical
exponents. And for the dynamical behaviour a new type of activated dynamics
was established with exponents Jlv = 2.2 and p = -1--Y(z-q-+) = -3.0.
Although various scaling relations exist between the different exponents,
y = v(z-q), zp = v(d-z+q) and v = -l/8, we still have one unknown
too many. Here we revert to theory for the most innocuous choice, namely
the prediction of a very small positive q ranging from 0 to 0.28. So taking
q = 0.2, we then proceed to calculate the remaining critical exponents.
Table 6.2 collects the results for the six exponents, the first four are the
usual static ones, the last two are those based on the droplet model and its
activated dynamics. The theoretical predictions, which are also listed, have
been gained mainly through computer simulations (see section 5.12). A
study of Table 6.2 will reveal good agreement with the ?J model and a
self-consistent set of exponents, at least within the rather large error bars
of both the experimental and the simulation, results. Very important, T,
was found from various best fits to be zero, nicely confirming the theoretical
conclusion drawn from the 2D computer calculations. Furthermore, the
exponent 0 being negative corroborates that there is no finite-temperature
phase transition for a 2D Ising spin glass. In addition, the dynamics is
activated as forecast by the droplet model. All in all, we have a rather

Table 6.2 Comparison of critical exponents obtained for Rb2Ch.782Coa.218F4with

those of the %J and Gaussian Edwards-Anderson Ising SG models

Critical kJd=2 Gaussian d = 2

exponent Ising SG Ising SG

V 2.3 r 0.4 2.59 2 0.13 3.4 r 0.1

2.5 r 0.3 2.6 2 0.2 3.5 2 0.3
2.64 f 0.23 3.6 2 0.1
Y 4.2 2 0.6 4.5 k 0.5 7.0 * 0.2
4.5 5 0.2 5.3 * 0.3
P 0.0 * 0.1 0.2 2 0.1 0
rl 0.2 k 0.2 0.20 5 0.02 0
0.28 k 0.04
* 0.9 -+ 0.2 O~qJ~l O~qJ~l
e -0.42 k 0.06 -0.38 -c 0.03 -0.29 k 0.01
194 Ideal spin glasses: comparisons with theory

complete description for this ideal spin glass and a most favourable
experimental-theoretical agreement. Certainly more work is needed to
reduce the error bars to justify that the four real Js of Rb2Cuc7sCo,,22F4
are definitely in the same universality class as the simple ?J model. And
finally upon viewing Fig. 6.7 again, the warning is clear. Far above a
possible T, the relaxation shifts outside the realm of feasible experiment.
Here the spin glass is out of equilibrium and we must live with these
disturbing non-equilibrium effects.
With this in mind, let us move onto the available 3D materials where
the question of a finite T, is paramount and, if so does T, = T,? And what
about the behaviour below T,, once a phase transition has been established?
Keep in mind we are using T, to denote the critical temperature of a phase
transition as predicted by theory. In contrast we reserve Tf for the freezing
temperature as is clearly characterized by experiment. The latter may or
may not be related to the former.

6.3 IDEAL 3D SPIN GLASSES

Enormous effort has gone into attempts to compare experiments of xn,,

X”(O), etc. on various 3D spin-glass materials with theory using different
types of scaling functions and analyses which give the critical exponents.
The approach is similar to that followed in the previous section for the
ideal 2D spin glass. The hope is to gain some insights into the true nature
of a phase transition, if there is one. For example, CuMn, AgMn and AuFe
have all been leading candidates for such endeavours and once the scaling
analysis is completed, the values of the critical exponents are likened to
those obtained from Monte-Carlo simulations of specific model (mainly
Ising) spin glasses. Certain difficulties exist with this prescription. Firstly,
it is based on the assumption that a standard second-order phase transition
occurs with conventional static or dynamic critical behaviour. We have
already experienced in section 6.2 the unconventional activated dynamics
for 2D Rb2Cu1-JZo,F4. So would a 3D spin glass exhibit a completely
normal second-order phase transition? Probably not! The above assumption
assumes too much and begs the question of a phase transition. Secondly,
for the many different and definitely non-ideal or in some cases not even
good spin glasses, there is a vast range of critical exponents which have
been determined from the scaling analysis. These depend strongly on the
range and quality of the experimental data and the particular analytic
procedure. One jestingly speaks of the ‘right order of magnitude’ for a
critical exponent. Thirdly, the manifold values usually do not agree with
the Monte-Carlo results. And fourthly, more refined and recent scaling
analysis using linear plots instead of log-log ones has demonstrated the
ambiguity of even the finest of the older work. The latest scaling usually
6.3 Ideal 30 spin glasses 195

shifts the exponents upwards and T,, which is taken as a fitting parameter,
downwards. Furthermore, the ‘best’ exponent values obtained to date are
anomalous, i.e., much too large when compared to those observed in
normal phase transitions. These points when taken collectively, indicate
that a 3D spin glass is close to or just above its lower critical dimension.
In addition, we have all the complications of non-equilibrium setting in as
we approach T,, so that even if there is a phase transition, we could not
measure its equilibrium properties.
Rather than belabour the above obstacles (or shall we say challenges),
let us consider the present situation regarding a possible ideal, or model-
conforming (see Table 6.1 and dilate the dimensionality from 2 to 3), 3D
spin glass. To date our ‘best choice’ lies with the Fe,,,Mn,,STi03 mixed
compound, although strictly speaking it is not as ideal in 3D as the
Rb&uCoF, is in 2D. For the former compound we have a random
substitution of Fe and Mn spins situated in a hexagonal lattice; see Fig.
6.9. In the basal plane the Fe-Fe nearest neighbour superexchange is
ferromagnetic, while the Mn-Mn one is antiferromagnetic. However, it has
recently been verified that the Fe-Mn coupling is also antiparallel. Between
the planes, the interplanar coupling is always weakly antiferromagnetic as
shown in Fig. 6.9, and this gives the system its 3D character. So for a
50-50, Fe-Mn composition, there are three -Js with large weight and one
+J creating the short-range interactions. This represents a most asymmetric
distribution P( +J) of exchanges along with a large amount of frustration
due to the ‘missing-centre’ hexagonal (see Fig. 6.9) crystal structure.
Moreover, the Fe and Mn ions have different values of spin, both not 4,
which makes the system seem random-site-like due to the distinction
between Fe and Mn ions. In spite of these difficulties a strong unaxial
anisotropy keeps the spins aligned (up/down) along the c (or hexagonal)
axis, so Fe,,SMn,,,TiO, is indeed a 3D Ising short-range spin glass. And
this conclusion has naturally been substantiated by the usual measurements
of ac- and dc-susceptibility, FC versus ZFC magnetization, time dependences,
etc. These experiments have nicely mapped out the (T-x) phase diagram.
Equal amounts of Fe and Mn lies right in the middle of the spin-glass
regime.
Let us proceed with the scaling analysis of this quasi-ideal 3D spin glass.
By defining the non-linear susceptibility in a similar manner as in section
6.2, we expect the coefficient a3 to diverge as (T-T&Y. One then plots
xnl versus Hz (the small applied dc field) and in the limit of H,, + 0, i.e.,
the initial slope a3 is acquired at the various temperatures. Then a3 may
be plotted on a log-log scale against t = (T- T,)/T, for various choices of
T,. Figure 6.10 shows such dependences where xn, m u3, and Tg = T,. Note
there are two parameters here T, and y. A ‘best fit’ is curve B, T, = 20.70 K
and y = 4.0. The reduced temperature interval is very small because,
according to FC-magnetization, the system loses equilibrium below = 22.0
196 Ideal spin glasses: comparisons with theory

C-axis
0: F&&f+ FeTio3 MnTi03
t

(4 (b)

(cl
Fig. 6.9 (a) Crystalline structure of ilmenite-type compounds in the hexagonal
lattice. (b) Magnetic structure of FeTi03 and MnTi03. (c) Arrangement of magnetic
ions in the hexagonal c-layer. From Ito et al. (1988).

or log,,(T/T,-1) = -1.2, and at high T the non-linear response is very

weak and the critical region may be exceeded. Approximately a decade of
reduced temperature is present in Fig. 6.10, which is not a scaling analysis.
The usual spin-glass scaling equation for the non-linear susceptibility is

(6.13)

where F(x) a x for small x and F(x + m) a x1/8 with 6 another critical
exponent. This scaling function is displayed in Fig. 6.11 as a log-log plot.
The data ‘collapsing’ is obtained using the previously determined values for
T, and y. l.3 was the only remaining parameter and a value of 0.54 was
optimized from the scaling overlap. It should be noted, however, the
6.3 Ideal 30 spin glasses 197

B C
. .
.\ .

TBW)
A 21.00

B 20.70

C 20.50

-1.4 -1.2 -I -0.8 -0.6

l~g,~(T~~-l)

Fig. 6.10 Log&-x2) versus loglo for three different values of Tg = T,. The best
fit, indicated by a solid line, is obtained for Tg = 20.70 K, yielding y = 4.0. From
Gunnarsson et al. (1991).

-3

3 4 5 6 7 8 9
log IO[ H 0’ / ++-’ (G2)]

Fig. 6.11 LoglO(&lt~) versus logIO(H~ltP+~). The points represent the data
collapsing obtained using Tg = 20.70 K, y = 4.0, and p = 0.54. From Gunnarsson
et al. (1991).

superposition of data point is not very good for lower magnitudes of the
coordinates.
Via ac-susceptibility experiments the dynamical behaviour can be analysed
using the conventional divergence of T, = T,(T- T,)-‘“. Here for
Fe0.5Mn0.,Ti03 eight decades of frequency are available from x’(T) and
198 Ideal spin glasses: comparisons with theory

x”(T) measurements. Defining an 7, according to a ratio of x”(w)/x’(o),

we can make the standard plots of log 7 av versus log (T- T,) and try for a
unique determination of T,, 7, and zv. The results are T, = 20.95 K,
70 = lo-l3 s and zv = 9.3. However, once again this is not a scaling
analysis, but the conventional dynamical critical-phenomenon relation for a
second-order phase transition.
A new improved scaling function has recently been suggested which takes
the form

(6.14)

Since we expect llzv and plzv to be much less than one, the above equation
considerably weakens the frequency dependences. Now a linear scaling plot
of x”T/wP’~” versus (T- T,)/w~‘~” can be made and this permits a much
better judgement of the scaling quality. Such refined analyses result in
T, = 20.5 K, zv = 11.5 and l3 = 0.58. Significant here is that as always
when a more sophisticated scaling analysis is made T, is reduced and the
critical exponents are increased.
As a final comparison we use the static form of this improved scaling
analysis to treat the non-linear susceptibility for Fe0.5Mn,,STi03. With the
same rationale as above in mind, we write

(6.15)

where the scaling function d(x) + 1 for x - 0 and G;(x) = X-Y for small
Ho. As before scaling plots are constructed by plotting on linear scales x,,,/
H iP’r+P versus (T- T,) TJH :‘+a. Figure 6.12 displays the scaling analysis
for a particular choice of parameters. The inset is an enlargement of the
scaled data around the upturn. Unfortunately, it is impossible to find a
uniquely optimum set of parameters (here T,, y and p) that gives a perfect
collapse of the data. Various sets of the three parameters produce equally
good ‘collapsibility’. T, can vary between 20.5 and 21.0, y between 3.6 and
4.3 and p = 0.5 to 0.7.
Now that we have summarized the latest attempts at scaling the most
ideal 3D spin glass, we should collect the results and relate them to the
theoretical predictions from the Monte-Carlo simulations of a 3D Ising
short-range spin glass (see section 5.12). Tables 6.3 and 6.4 list the salient
features and results. Regarding the ‘idealization’, notice in Table 6.3 the
inconsistency with respect to the different spin values and the distribution
of interactions. The unlike crystal structures should not be of great
significance, except that the hexagonal lattice will increase the amount of
frustration and strongly reduce the magnitude of superexchange coupling
-J4 along the c-axis.
Table 6.4 requires a close examination. Here we make a direct comparison
6.3 Ideal 30 spin glasses 199

0.3
0.06
_ t Tg=20.70K

0.2
- '. ;=I=,",",
*
I
0.1 t
i

0 0.05 t, H,“(d) 0.1 0.15

Fig. 6.12 x~JH~@‘(~+@) versus t/Hz(Y+a). Th e points represent the data collapsing
obtained using Tg = 20.70 K, y = 4.0, and l3 = 0.54. The insert figure displays a
magnified part of the plot. From Gunnarsson et al. (1991).

Table 6.3 Comparison of Fer-,Mn,TiOs and the Edwards-Anderson spin-glass

model for d = 3

Fel-,Mn,TiOs EA Model

dimensionality d=3 d=3

lattice hexagonal cubic
spin value S(Fe) = 2, S(Mn) = i s=g
type of randomness bond randomness bond randomness
type of interaction Ising Ising
range of interaction nearest neighbours only in plane nearest neighbours
distribution of discrete and anisotropic (+J,, -.I*; Gaussian or +J
interactions -J3 in plane, plus weak -J4
interplane)

between the experiments on Feo.5Mn0.5Ti03 and the large-scale Monte-

Carlo (MC) simulations for a 3D (cubic), Ising, +J spin glass. Table 6.3
has already suggested that the similarity (real system with MC model) is
reasonable and besides it is the best we have available at present. Both
(real and MC) experiments indicate a finite T,, i.e., a phase transition is
occurring at a non-zero temperature. The MC computations have determined
the full range of exponents, except for (Ythe specific-heat ‘divergence’ which
must be calculated from a scaling relation and B, the order-parameter
exponent (see below). Since the simulations cannot reach below or even
too close to T, (loss of equilibrium), they give a ‘best-fit’ function for the
order parameter q(t) which contains a temperature-dependent exponent
200 Ideal spin glasses: comparisons with theory

Table 6.4 Comparison of critical exponents obtained for Fe0.5Mn0.5Ti03 with those
of the 3D king +J Monte-Carlo simulations

Critical exponent MC simulations

directly IY 4.0 f 0.4 2.9 -1-0.1

determined P 0.5 +- 0.2 = 0.5
from 10.5 * 1.0 7.9 2 1.0
experiment :(K) 20.7 k 0.3 finite
6 = r/p+1 = 9.0 6.8 k 1.2
a = 2-2p-y = -3.0 -1.9 2 0.3
v = (2-cl)/d = 1.7 1.3 k 0.1
rl = 2-ylv -0.35 -0.22 k 0.05
z = 10.5/v = 6.2 6.0 _’ 0.5

P(T). Around T,, where the simulations may not even be valid, l3 = 0.5.
So this is a deficiency of the MC technique, namely that l3 cannot be
accurately determined. Returning to the real experiments only 3 exponents
can be directly evaluated via the diverse scaling analyses of the measurements:
y, the susceptibility exponent, p, and ZY, a product of the dynamical
exponent with the correlation-length exponent. These are shown with the
appropriate error bars in Table 6.4 In order to continue the comparison,
different scaling relations (equations connecting the various critical
exponents) must be used to calculate 6 (the magnetic field exponent at T,),
cx, v, T (the correlation function exponent), and z. With this set of exponents
we should be able to completely characterize a second-order phase transition.
In general, the agreement is at best fair. Particularly disturbing are the too
large values of y and v for our real (ideal) 3D spin glass; also the inability
of fixing T, to within 1% weakens the analysis. While it seems that there
is a phase transition, it is not of the conventional type. Either the lower
critical dimension is close to 3, e.g. as some theories predict 2.7, so that a
marginal phase transition occurs, or the non-equilibrium behaviour which
is certainly present hinders a meaningful characterization of the critical
phenomenon. At this time we cannot proceed and resolve these difficulties.
One may study Table 6.4 and draw his or her own conclusions about the
nature of the spin-glass freezing. To be sure we must await future experiments
and theory we can offer a definite and physical settlement. Perhaps such
will be possible in the second edition of this book.

6.4 FUTURE POSSIBILITIES

Never before has so much effort been put into the study of a possible phase
transition as with the spin glasses. The previous sections have only scratched
the surface of huge amounts of work done on the problem. For we have
5.4 Future possibilities 201

Intentionally omitted the many experiments and comparisons with theory

:egarding the abundance of non-ideal (canonical and good) spin-glass
naterials. Tables exist for their critical exponents and temperature, but we
nave maintained the viewpoint: first to understand the most ideal and
simple systems before proceeding with imperfect and complex examples.
For the case of a 2D Ising short-range (*J) spin glass (section 6.2) we
nave a reasonable material which closely matches the constraints of the EA
model. A series of accurate and systematic experiments have been performed
md firmly compared to theory which predicts two dramatic features: T, = 0
and activated dynamics. Both of these highlights were borne out by the
measurements and a relatively consistent set of exponents was found to
agree (within the large error bars) with the Monte-Carlo simulations. Our
ife is made quite easy in all the fitting procedures by setting T, = 0. Hence,
we seem to now possess a basic understanding of the 2D situation.
Clearly, for the 3D Ising short-range spin glass (section 6.3), the present
&ate of the science is less satisfactory. Our best-choice system, while not
Terfectly ideal, does meet many of the constraints of the EA model. Here
:heory is more mundane: a finite T, and conventional dynamics are inferred.
However, these predictions cause a lot of trouble, T, cannot be determined
accurately from experiments and becomes a fitting parameter. And the
zonventional dynamics results in an exceptionally large value of ZY which
nakes one suspect that something is wrong with the fitting function itself.
ln addition, the static critical exponents do not agree with those found from
:he Monte-Carlo simulation. Since T, = 0 for d = 2, we would expect T,
+ 0 for d = 3 with both systems in the Ising universality class. While a
chase transition in 3D spin glass seems likely, it probably requires a more
iophisticated (beyond-the-conventional) type of critical-phenomenon theory
to describe its properties. Mean-field-type theories do not help us here
Jecause they always predict a phase transition with a standard set of
zxponents. Now, finally, we have arrived at the ‘cutting-edge’ of present-
day spin-glass research. What will the future bring? Let’s try to glance a
5it into the uncertain prospects for future progress.
Above all, an ideal 3D Ising ?J spin glass system must be discovered or
deliberately fabricated. The latter could be intentionally created using the
modern techniques of materials research. A material resembling that shown
m Fig. 6.1 and discussed in section 6.1 would then be measured completely:
specific heat, ac- and dc-susceptibility, magnetization, neutron scattering,
:tc. The accumulated data with temperature, field and concentration
dependences could be fully compared with existing theory and the results
3n Fe0.5Mn,,5Ti03. Primarily, we would more severely test the theory and
hopefully there would be a better agreement than now appears. The key
Iuestion here is how to obtain the 3D sample. Again it lies in the hands
af our materials or chemist friends.
A hint along these lines comes from the recent work to fabricate thin
films of metallic spin glass, e.g. CuMn, AgMn and AuFe. Various deposition
202 Ideal spin glasses: comparisons with theory

techniques have been employed to make samples of these alloys thinner

and thinner. To increase the sample mass usually a multilayer or superlattice
is constructed with non-magnetic interlayers (Cu or Ag) separating the thin
spin-glass films. Different experimental methods try to detect the freezing
temperature as a function of thickness, e.g. SQUID magnetometry or
anomalous Hall effect to determine the magnetization. The problem is to
get enough sample mass so that clear signs of the freezing, not simply
relaxational effects related to the loss of equilibrium, may be detected.
Remember it is rather difficult and subtle to distinguish the 3D spin-glass
cusps from the 2D spin-glass peaks caused by non-equilibrium effects. This
impediment becomes even more severe, if signal sensitivity is low due to
small sample size. The initial results of such experiments are ambiguous.
One group sees a strong reduction of Tf + 0 with d (the sample thickness)
+ 0. Another claims Tf decreases but remains non-zero down to thicknesses
of 1 and 2 monolayers. Also T,(d) depends strongly on what the interlayer
is in the superlattices. The experiments are hard to perform and the
interpretation of data is even more troublesome when it comes to establishing
Tf or T,. So far, accurate ac- or non-linear susceptibility has not yet been
obtained although attempts are presently being made. Thus it remains a
pending task to overcome these difficulties and settle the freezing behaviour
of truly 2D (Heisenberg) spin glass. How thin is 2D? What is the length
scale governing the crossover from 3D to 2D? At what thickness will we
observe a change in the dynamics? These and many other questions can be
posed concerning the artificial reduction of spin glass dimensionality. Future
research will certainly provide clear answers to these questions. And the
same sophisticated deposition techniques as used for the multilayers could
be upgraded to produce the artificially-structured ideal 3D spin glass shown
in Fig. 6.1.
The theory is far from complete and improvements are necessary.
Tremendous exertion has gone into the mean-field theory and obtaining its
solution. And we have tried to outline the physical interpretation of its
concepts and calculations (section 5.7) which are not obvious and took a
number of years to develop. Moreover, for this most-simple starting point
of theoretical approaches, enormous complexities lie in the continuous order
parameter and its probability distribution of overlaps. Even here experiment
is left behind. The first suggestions for mesoscopic systems are now beginning
to be made on how to measure the overlap distance or to estimate the
order-parameter probability distribution function based upon fluctuations
(noise), time dependences and ageing effects in the frozen spin-glass state.
But the basis is only mean-field theory - a first order approximation. What
happens when we advance to include fluctuations, or short-range interactions,
or remove the Ising up/down restriction? We call all this future exertion:
beyond the mean-field theory. Also the droplet and fractal cluster models
need to be pursued and extended with more and closer experimental contact
and comparisons. There is still a vast amount to be accomplished with the
Further reading and references 203

theory of spin glasses and despite the past successes new approaches, not
so heavily based on second-order phase transitions, and models from other
areas should be tried ensuring future progress not just in spin glass, but
more extensively with random and glassy systems in general.

FURTHER READING AND REFERENCES

Bass, J. and Cowen, J. A. (1992) Recent Progress in Random Magnets (Edited by

D. H. Ryan) World Scientific, Singapore.
Dekker, C. (1988) Ph.D. Thesis University of Utrecht.
Dekker, C., Arts, A. F. M. and de Wijn, H. W. (1988) Phys Rev., B38, 8985.
Dekker, C., Arts, A. F. M., de Wijn, H. W., van Duyneveldt, A. J. and Mydosh,
J. A. (1988) Phys. Rev. Lett., 61 1780; (1989) Phys. Rev., B40, 11243.
Gunnarsson, K., Svedlindh, P., Nordblad, P., Lundgren, L., Aruga, H. and Ito,
A. (1991) Phys. Rev., B43, 8199.
Ito, A., Aruga, H., Kikuchi, M., Syono, Y. and Takei, H. (1988) Solid State
Commun.. 66. 475.
7
Spin-glass analogues

The previous six chapters have dwelt upon the spin-glass phenomenon with
emphasis on the word ‘spin’. Except for a few cursory remarks pertaining
to real glasses, we have remained fully on the track of magnetic effects and
spin models. Naturally, magnetism does have its advantages since we can
easily couple into such systems with a magnetic field, the conjugate field
of the problem, and there exists a vast range of sensitive experimental
techniques to measure this coupling under various conditions. However,
now that we have outlined the basic spin-glass behaviour and theory, and
hopefully have gained an understanding of the rudimentary physics, the
obvious question arises: are there non-magnetic analogues? Answer: yes,
many! So let’s at this point take a look at some of these and venture
beyond the spin and even beyond the glass into life sciences.
We begin with the electrical analogues where electric-dipole moments
giving rise to a polarization P replace the spin moment causing the
magnetization M. Higher-order moments are also mentioned in the form
of electrical quadrupoles. By substitution of a moment bearing species for
a non-moment or a different-interacting moment, we introduce the elements
of randomness and competition. Thus, we can create electric dipolar and
quadrupolar glasses.
Our analogues are continued with various superconducting phases. First,
there are granular superconductors which behave like a network of randomly
coupled tunnel junctions, Then, the high-T, oxide superconductors have
been modelled on the spin glasses, initially because of their ceramic nature
and later due to the unusual field penetration (H,, < N < H,,) creating a
flux-line ‘lattice’ which becomes glassy.
In order to contrast the spin-glass phenomenon with a random magnetic
system which does exhibit a good second-order phase transition, we briefly
discuss the random-field Ising model. Here we focus upon the critical
behaviour and the H-T phase diagram whose critical line must be traversed
in a certain way in order to observe the phase transition.
Finally, we conclude the chapter with three non-physical analogues to
the spin glasses. These ‘spin-offs’ are combinatorial optimization, popularly
known as the travelling-salesman problem, neural networks or how our
7.1 Electric dipolar and quadrupolar glasses 205

Jrain can be modelled by a spin glass, and finally biological evolution. Is

.he origin of life a spin glass?

1.1 ELECTRIC DIPOLAR AND QUADRUPOLAR GLASSES

The electrical analogue of a magnetic spin is a small displacement d of two

mlike charges e creating a dipole moment p = ed. Such dipoles can be
brmed by the displacement of dopant atoms in certain compounds from
:he centrosymmetric host sites or by molecules with their bonds not having
1 spherical symmetry. Along these lines we can make direct contact with
)ur magnetic systems via the long-range order of the ferroelectrets (all the
:lectric-dipole moments are aligned parallel) or the antiferroelectrets
yantiparallel sublattices). In mixed (ferrojantiferro) or diluted (ferro) electric
naterials we have a close analogue of a spin glass which we call a dipolar
Ilass. Here the electric field serves as the conjugate field. Theoretically,
ve can use a ‘pseudospin’ formalism to represent the electric dipoles and
heir interactions, measure the dielectric polarization and susceptibility in
)lace of the magnetic ones, and apply an electric field for the FC or ZFC
lrocedures.
A random-site example of a dipolar glass is K,-,(Li,, Na, or Nb,) Tao,
vhere the displacement of the x-dopant (Li, Na or Nb) forms an electric-
lipole moment. For sufficiently large Li, Na or Nb concentrations, collective
:ffects occur via dipole-dipole or lattice-stress-field interactions, and through
:xperiment we can investigate the freezing behaviour. Figure 7.1 illustrates

FH

25 30 35 40 T (IO

rig. 7.1 Polarization versus temperature in a T-sweep experiment for K1-,Li,Ta03.

qote that the temperature where P becomes independent of time, aP/at = 0 (not
hown), coincides with the departure of the FC curve from the ZFC curve.
: = 0.016, dT/dt = 3 mWsec, and E = 30 kVlm. Inset: Field-temperature cycle.
Tram Hiichli et al. (1985).
206 Spin-glass analogues

the typical static, electric polarization P (analogue of M) versus temperature

curves for I(0.984Li0,016Ta03 according to the different (electric) E-field-
cooling procedures shown in the inset. Note the ZFC-field heating (FH)
cusp found at the freezing temperature Tf = 37 K. This nicely corresponds
to the ZFC magnetization for a spin glass. However, in the electric system
the FC plateau is missing and there is noticeable hysteresis below Tf between
FC and FC-FH. Another attempt at comparison is displayed in Fig. 7.2
where the non-linear part of the ‘static’ (1 kHz) dielectric susceptibility e,r
(obtained from the higher-order E-field dependence) is plotted against the
temperature for &,,Na0.zTa03. The raw data are shown as E,,, but the
1 kHz measurement frequency prohibits their use below = 20 K (where a
dispersive l “-component begins to appear). A special analysis tries to
extract the divergent coefficient of the non-linear susceptibility: a(T) cc
[(T-T&T&y. Th’IS is exhibited as the curve (a) in Fig. 7.2 and yields Tf

10 20 30 40
T (K)

Fig. 7.2 Non-linear dielectric susceptibility E,,(T) and critical coefficient a(T) in
K1-,Na,TaO,. Solid curves: best fits to power law. Dotted curve: extrapolation of
a(T) to show criticality. From Maglione et al. (1986).
7.1 Electric dipolar and quadrupolar glasses 207

= 14.5 K and y = 2.7. Unfortunately, better fitting closer to Tf could not

be accomplished due to the loss of equilibrium represented by the peak
and decrease in E,,.
Another dipolar glass is the mixed system Kr-, or RbI-,(NH4)XH2P0,.
These polar materials are characterized for sufficiently high temperature by
a disordered network of O-H--O bonds corresponding to randomly rotating
dipole moments. Here we have the paraelectretic phase. At low temperatures
the ordering of these H-bonds or dipoles leads to a ferroelectric polarization
in pure K or RbH2P04, but in pure NH4H2P04 an antiferroelectret develops.
These long-range transitions are suppressed by the random substitution of
K or Rb by (NH,),, i.e., competing or mixed interactions. A glass freezing
results below i= 30 K for x = 0.35. The quantity to measure once again is
the dielectric constant, real and imaginary (dielectric loss) parts, E’ and E”.
The data are shown in Fig. 7.3. While there is no cusp in the real part,
the dielectric loss exhibits a frequency-dependent peak which can be related

-wu 160
Fmq. Hz
0 106

. 3260 1 (a)
0 10640 fI
t l 33700

I I 1 I 1 I 1 I

0 10 20 30
TEMPERATURE (K)
Fig. 7.3 (a) The real part of the dielectric constant ELfor Rb,-,(NH&H,PO,. The
lines are guides to the eye. (b) The imaginary part of the dielectric constants E:
fitted via the lines to a broad distribution of relaxation times. From Courtens (1984).
208 Spin-glass analogues

to a broad distribution of relaxation times g(r, T). This function, similar

to that of a spin glass, extends to longer and longer times as the temperature
is reduced. So at a particular T, the characteristic or average relaxation
time becomes much greater than any time of measurement (loss of
equilibrium) and freezing results with the associated glassy dynamics.
Nevertheless, from the curves in Fig. 7.3, it does not seem that the freezing
transition displays sharp co-operative effects.
Quadrupolar glasses are formed when the elastic quadrupole-quadrupole
interaction is more dominant than the electric dipole-dipole interactions.
A good example is the mixed molecular crystal (KCN),(KBr),-,. In this
case the CN molecules possess not only a small dipole moment, but also a
quadrupolar one since they are rod- or cigar-shaped compared to the
spherical (meaning no moments of any order) Br-ions. Both the dipoles
and the quadrupoles couple individually via the quadrupolar stress field
generated by the lattice and its elastic constant. Pure KCN undergoes a
transition at 168 K from a completely disordered (paraelectret) cubic phase
to an orthorhombic phase where the axes of the CN molecules become
oriented along a given (b-axis) direction without any associated dipolar
ordering, i.e., the arrows or heads of the dipoles point randomly up or
down along the b-axis. At a lower temperature 83 K, there is an order-
disorder transition which creates an antiferroelectric structure of the dipole
heads along the b-axis. Now we mix (KBr)l-X into (KCN), and around
x = 0.5 for T 5 60 K, a random isotropic freezing of both the molecular
orientation (the axes) and the dipoles (the arrows) simultaneously occurs.
Figure 7.4 sketches this glassy state for (KCN),.,(KBr),.,. The experimental
proofs for such a phase come from measurements of specific heat (a linear
low-temperature dependence), neutron scattering (appearance of a central
peak) and dielectric susceptibility (low-frequency peak in E’ and l ” as a
function of T). Figure 7.5 shows the latter for various frequencies. At least
around the peak temperature the behaviour of the (dielectric) susceptibility
is standard spin-glass-like, except for broadness of the maximum in E’ .
Moreover, the shift of the peak temperature with frequency seems to follow
an Arrhenius law: w = o, exp(- EBIkBT), for the accumulated data. So
once again it would seem that with the mixed molecular crystals a reasonable
analogue reveals itself to a spin glass, albeit without the sharpness or
criticality of a good phase transition.
We could mention other examples of dipolar or quadrupolar glasses
[(W1-xArx, etc.] with the same general properties. But instead, let us
conclude this section by reviewing the interesting case of solid, molecular
(ortho/para) hydrogen, which is a quadrupolar system without a dipole
moment.
In molecular hydrogen because of the fermion statistics of the protons,
two forms are possible: (ortho o-Hz) and para (p-H,) hydrogen. o-H, has
a total nuclear spin of unity and odd angular-momentum values. Oppositely,
p-H, has zero total nuclear spin and even values of the angular momentum,
7.1 Electric dipolar and quadrupolar glasses 209

Glassy x=0.5 K <, 60K

Fig. 7.4 Sketch of the various molecules in a (1, 0, 0) plane for 50%
(KCN),(KBr),-,. The arrows on the cyanide molecules indicate the random freezing
of the dipoles. From Sethna and Chow (1985).

. :
.v .. :
1.2

2 0.9
g
d 0.6
2
w” 0.3

0.0
0 40 60 120
T 09
Fig. 7.5 ~1 (=e’) and Ed (=E”) versus temperature at several frequencies for
(KBr),.,(KCN),.,: o = 6~10~s~’ (A), 6~10~s~’ (m), 60-l (O), and 12s-’ (V).
From Birge et al. (1984).
210 Spin-glass analogues

which vanishes in the ground state of the solid phase. As a result of this
difference, o-HZ molecules are non-spherical, while p-H, molecules have
perfect spherical symmetry. Therefore, the asymmetry of the o-HZ offers
the possibility for various types of quadrupolar orientational order (dipole
moments are negligible) whereas the no-moment p-H2 can be used as an
inert host. Figure 7.6(a) illustrates the situation for the long-range-ordered
phase of pure o-Hz. A planar projection of the fee-like crystal structure is
shown with the novel type of periodic long-range orientational ordering at
low temperatures. More important for us is the mixed crystal where sufficient
amounts p-H2 have been substituted for o-HZ to form the mixed phase.
This is displayed in Fig. 7.6(b) where the p-Hz molecules are indicated by
the ‘dotted’ spheres and there are various shapes and orientations for the
o-HZ molecules. The latter is caused by a probability distribution for
different molecular shapes which occurs in the mixed phase. Here molecules
can be orientated preferentially along an axis (‘cigars’) or preferentially in
a plane (‘discuses’) with a continuum of intermediate shapes. At high
enough temperatures these dissimilar-shaped quadrupoles can freely rotate,

Fig. 7.6 Comparison of the orientational ordering for (a) a plane section of the
ordered structure and (b) the quadrupolar glass phase; from Suilivan (1983).
‘.l Electric dipolar and quadrupolar glasses 211

Jut as the temperature is lower they seem to gradually freeze out into a
,andom static alignment. Such glassy state is sketched in Fig. 7.6(b).
The main experimental technique used to measure these effects has been
VMR on the protons; resonance-line shapes and nuclear-relaxation times
ue determined as a function of temperature and concentration. Based upon
he results a phase diagram (T-x) is proposed in Fig. 7.7 for the mixed
lo-H,), (p-H&, system. Notice at the very low temperatures (below
1.3 K) a frozen (orientational glass) phase appears to form for a restricted
.ange of X. The coupling between the o-H, molecules is the weak electric
luadrupole-quadrupole interaction. Its conjugate field is an electric-field
Iradient which would be required for the measurement of the quadrupolar
susceptibility. However, it is quite difficult experimentally to produce a
sufficiently large electric-field gradient in order to perform accurate
susceptibility measurements. Consequently, most of the information per-
aining to o-HZ/p-H2 mixed crystals has been obtained from NMR which
should be complemented by other techniques. The generally agreed
nterpretation of the NMR results is a gradual freezing of quadrupole
noments with a distribution of relaxation times which seems to continuously
:volve as the temperature is lowered. The quadrupoles appear frozen when
he time of measurement is smaller than the average relaxation time,
vhereas no freezing is observed for experiments with longer observation
imes. As above we have a spin-glass analogue, but without any indications
)f a sharp and critical phase transition which seems to be generally missing
n these ‘electrical’ systems.

I I I I
- wqx (P-$1,-)(

I
0 0.2 0.4 0.6 0.8 1 .O
ortho-concentration
i‘ig. 7.7 Phase diagram for solid solutions of o-HZ and p-H*. For large concentrations
)f the orientationally anisotropic component (o-H,) the orientationally ordered
chasehas face-centred-cubic structure, the disordered one is hexagonal close packed.
Ihe orientational glass phase (denoted by SG) forms a narrow low-temperature
wedge. From Sullivan et al. (1978).
212 Spin-glass analogues

7.2 SUPERCONDUCTING GLASS PHASES

Consider a granular superconductor that consists of superconducting islands

randomly embedded in a non-superconducting host. The host can be an
insulator, semiconductor, metal or even another superconductor with a
lower T,. The various superconducting grains or islands interact via
Josephson tunnelling of Cooper pairs, i.e., a phase coherence of the electron
pair across a narrow insulator or the proximity effect, i.e., an induced
but weakened superconductivity in the host material. There are many
experimental realizations of such systems. Examples are In or Pb with
insulating Ge forming a random continuous medium, Pb (T, = 7.2 K)
spheres randomly dispersed in Zn (T, = 0.8 K) and Nb composites with
the normal metal Cu. In addition, a more modem technique is to artificially
fabricate a network of Josephson junctions out of Al/Al-oxide with different
size ‘weak links’ connected to each other. And indications exist that in a
magnetic field these disordered superconducting systems may freeze into a
state which exhibits spin-glass-like order: a random orientation of the
supercurrents in the various grains. Here we need to define our order
parameter which for a superconductor is the complex energy gap (wave-
function of the Cooper pairs) in the ith grain Jli,
Jli = MT) exp (i4J (7.1)
where hi(T) is the temperature-dependent real energy gap in the density
of states and & is the phase factor.
The hamiltonian describing the collection of grains and their interactions
can be written as

X= -CJijcOs($i-+j-Aii) (7.2)
{ii}
where

dl (7.3)

A is the vector potential and the integration is along a path connecting the
centres of the ith and jth grains. Jij varies between 0 and J without reversing
sign, but it falls off with distance between the grains as m l/r, for a
Josephson coupling or a exp(-rij) for the proximity effect. A random *
(or ferro/antiferro) interaction is caused by the variable A, as part of the
argument of the cosine in equation (7.2). Let us position the applied
magnetic field along the z-axis, B = BZ and use the gauge for the vector
potential, A = Bxf. Then

where x,Yi and XjYj are the centre coordinates of the ith and jth grains,
7.2 Superconducting glass phases 213

respectively. Disorder is introduced by an amorphous structure of the grains

such that the x and y become a set of random variables. This creates a
variation in A, which may be further tuned by increasing the field B.
Another way is to vary the Josephson-tunnelling coupling between the
grains such that Jij = J with occupation probability p, and Jij = 0 with
probability 1 -p. Because of these random, yet weak, interactions, frustration
results and causes random orientations of the supercurrents among the
collection of coupled grains.
The thermodynamic properties of this disordered systems are governed
by the free energy and associated partition function. In order to calculate
these quantities from the above hamiltonian both thermal and configurational
averaging is required, necessitating the replica trick. Thus, the order
parameter for the set of superconducting grains becomes analogous to the
Edward-Anderson one
q = ((Jl&>c = A’(T) ((exp i+&)c (7.5)
And the state it represents is a randomly orientated distribution of frozen-
in supercurrents instead of the EA spins.
The mean-field phase diagram based upon these model calculations is
shown in Fig. 7.8. Note the figure has three coordinates (T, p, H)
with applied magnetic field (B = H) as the third axis. For a granular
superconductor just above its percolation threshold @ L pC), there are the
collective Meissner phase - a complete expulsion of the flux up to Z&i, and
the Abrikosov one - above H,, a penetration of the field in a regular
triangular lattice which has a uniform orientation from grain to grain. Really

(H>O)

It_
N
(P’PJ
SC
is.

Fig. 7.8 Left-hand side: mean-field phase diagram as a function of temperature T,

applied magnetic field H, and Josephson bond-occupation probability p near
percolation threshold pC. Four distinct phases are exhibited: normal, Meissner, glass,
and Abrikosov (A) phases. Right-hand side: phase diagrams for fixed applied
magnetic field H, showing spin-glass (S.G.), superconducting (S.C.), and normal
(N) phases for (a) H > 0 and (b) p > pC. From John and Lubensky (1985).
214 Spin-glass analogues

new here is the prediction of a glassy phase just beyond pc for sufficiently
large fields. Glassy means a penetration of the magnetic field into the
superconductor via a completely random alignment of frozen-in supercurrents
among the various grains. This leads to long-range fluctuations in the
magnetic-field profile which decay as a power law with distance. So we
have a highly inhomogeneous field distribution within the superconductor
which yet remains to be detected. Unfortunately, the superconducting order
parameter is experimentally inaccessible. However, the differences between
FC and ZFC of the magnetization and the time dependences (dynamics) of
the glassy state should be measurable characteristics.
Recently there has been renewed interest in the superconducting/glass
concept due to the high-Tc oxide superconductors and their inability to
carry large currents in a field. Firstly, the ceramic form of these materials
is a natural granular superconductor. Thus, the model of the preceding
paragraphs may be directly applied. And, indeed, the experiments just
mentioned have borne out the spin-glass-like behaviour. Secondly, for the
oxide superconductors there are an enormous number of microscopic
defects, which, when combined with the unusually small superconducting
coherence length and large anisotropy, make even bulk singZe crystals a
candidate for glassy behaviour. The name ‘vortex glass’ has been given to
describe the random penetration of magnetic flux in such single-crystal
materials.
It is instructive to reconsider our superconducting order parameter in
order to elaborate our magnetic analogue. Recall that JI is a complex
quantity with two components: amplitude and phase. This corresponds to
the order parameter of a two-components (x-y) magnetic system. Returning
to Fig. 7.8 in the Meissner phase (H < Z&r), + is uniform as in a ferromagnetic
(uniform magnetization). In the Abrikosov phase (H,, < H < Z-Z,,) a
periodic flux-line lattice is formed without any disorder in which +(r) = $(r
+ R) where IR( = a, =(h/2eH). However, if some quenched disorder exists,
as it does in the high-T, materials through the assortment of defects, the
flux-line lattice will be destroyed on long-length scales. Figure 7.9 illustrates
the difference: long-range ordered Abrikosov phase versus short-range
ordered domains of flux lines and finally complete disorder (glass). It is
expected that the latter case is more vulnerable to displacements from the
Lorentz force, I x II, if the superconductor carries a current 1. When such
motions of the previously-rigid flux-line lattice occur, a resistance appears
and the superconductivity will effectively disappear.
As explained before it is the magnetic field via its random vector potential
A + A,, which arises in our hamiltonian that creates the ‘competing
interactions’ and thereby frustrations. Such are the basic ingredients for
forming an analogue spin glass. In a type-II superconductor a vortex glass
(see Fig. 7.9) may, therefore, exist and, since the flux-line lattice is rigidly
frozen, have no resistance at least for small currents. Although our initial
discussions have dwelt upon granular superconductors, now with the advent
7.2 Superconducting glass phases 215

l l l l l l l l l

l l 0 l l l l l l

l l l l l l l l l

l l l l l l l l l

a l eeeeee l

l l l l l l l / aOe

l l l l l l l l l

l l l l l l l l l

l l l l l l l l l

l l l l l l l l l

l l l l ;e l l l l

l l l l /e l l l l
I
l l l l ie l l l l

l l l l /e l l l l
- - ----------I
l l l l
/e l l l l

b l l l ‘/e l l l l
l l l l ’ /e l l l l

~.l l
-.--------4 l l i l l l l l
l l l r--e---c--3-------
e
l l l l
je l l l l
l l l l I
l l l l l

l e l l l e l l
l
l l l
l l l
l e l
l l l l eee l l

l *e l e l l l l

l l l e l
l
l e l

l l l
’ l l l l
l l
l l l l l
l e l
l l l
l
l l l l l l l
l
l e
l
l l
l l e
l eeeee l l m

Fig. 7.9 Superconducting flux-line lattice for H,, < H < &: (u) long-range
ordered, Abrikosov triangular lattice; (b) short-range ordered lattice with domains;
and (c) vortex glass. For the magnetic field into the page the supercurrents circulate
clockwise around each dot.
216 Spin-glass analogues

of the high-T, oxide materials, we can apply the model to a distinct phase
of bulk, single-crystal superconductors. We continue by drawing a direct
analogue with our spin glasses. The hamiltonian, if the A, values are
constrained to the range between 0 and 7~, is similar to the EA x-y +-J
spin glass. Based upon our previous review of the theory, especially the
computer simulations, this universality class is not expected to have a phase
transition in d = 3. For a superconductor a more realistic model would be
to let the A, vary from 0 to 27~. Such a model is called a ‘gauge glass’.
This means that the frustration is a serious disorder which is invariant under
the gauge transformations Si + -Si and J, -+ -Jij. The disorder cannot be
‘gauged away’ using the word gauge in analogy with the invariant
transformations of electro-dynamics. A gauge glass does not have reflection
& + -& as one of its symmetries because the magnetic field breaks the
time reversal invariance. And this lack of symmetry distinguishes it from
the x-y (0 to n) model. So present theoretical thinking is that the gauge
glass belongs to a different universality class and may very well have a
phase transition in 3D. Extensive Monte-Carlo computations indicate that
the gauge-glass model in 3D has behaviour comparable to an Ising spin
glass (d = 3) which should have a finite T, rather than to an x-y spin glass
which should not.
Experiment has tried to determine the various properties of the flux-line
lattice, especially to establish the existence of an intermediate phase with
zero resistance corresponding to the vortex glass (see Fig. 7.9). Initial
indications seem to be favourable for its formation accompanied by a sort
of phase transition and there is experimental agreement with other effects
predicted by the model. Since a flux-line lattice is a mesoscopic phenomenon
with larger than atomic dimensions (typical lattice constants are of order
microns), it can be directly observed via scanning tunnelling microscopy or
decoration techniques. To find such a glassy structure as sketched in Fig.
7.9 would be a challenging experiment and complete confirmation of the
model, as would the observation of the melting behaviour at higher fields
and temperature.

7.3 RANDOM-FIELD ISING MODEL

Before continuing with our spin-glass analogues and our transversal into
areas far different from our spin systems, let us pause to examine another
type of random magnet which is not a spin glass. The model, first proposed
by theorists (Imry and Ma, 1975), is called the random-field Ising model
(RFIM). Here an Ising ferromagnet of various dimensions is placed in a
site-random magnetic field. The model hamiltonian becomes

X = - C Jij SiSj - z HiSi (7.6)

ij i
7.3 Randomlfield Ising model 217

where the spins are all z-components, Jij is a constant (J), representing
positive ferromagnetic interactions between nearest-neighbour spins, and
the Hi are independent, local, random variables with
[HilAV = 0 (mean) [HflAv = Hg (variance) (7.7)
Using the following simple argument Imry and Ma concluded that such
RFIM systems would exhibit a phase transition for dimension d > 2 (d, = 2,
the lower critical dimension). The net-energy cost for the formation of an
oppositely-oriented domain of volume Rd out of the ferromagnetic state is
E(R) = JRd-’ - HRRdf2 (7.8)
The first term represents the domain-wall energy where the spin bonds are
broken along a surface R d - l . The ferromagnet in the presence of random
fields has a mean-square energy of HRR 2 d. For a particular choice of an
oppositely-oriented domain, this energy (after taking the square root) can
lower the total energy of the domain formation by a more favourable
alignment with the random fields. And thus we arrive at the second term
which is of opposite sign to the first. For HR * J the total energy is positive
for d > 2 (domains cannot form), but for d < 2 it becomes negative for
sufficiently large R. This means that domains are energetically favourable
and will be created thereby destroying the long-range order phase and its
phase transition. The d = 2 case is unresolved, hence giving rise to the
lower critical dimension.
Therefore, in 3D a special phase transition and critical phenomenon are
expected for the RFIM but not in 2D. And the H-T mean-field phase
diagram for this transition should be as sketched in Fig. 7.10. However, in
order to verify such behaviour we need a real system to measure. How do

ZFC
--*_
* +..,

.
T
TN TN
Fig. 7.10 Zero-field-cooling (ZFC) and field-cooling (FC) procedures for the RFIM.
The dotted line indicate the AF phase boundary in the H versus T phase diagram.
The arrows show the direction of the field-temperature cycling in each procedure.
P represents paramagnet and AF the RFIM-state which develops at sufficient field.
TN is the REIM Neel temperature (H = 0).
218 Spin-glass analogues

we in the laboratory turn on a site-random field in a ferromagnet? Answer:

we do not! Instead, we use another theoretical result that maps the above
RFIM onto an Ising antiferromagnet with random exchange (site dilution
of the spins staying more concentrated than the percolation threshold) in a
uniform magnetic field H. These systems exist, for example, in 3D:
FexZnl-xF2, Mn,Zn,-,F,, and Fe,Mg,-,Cl,, and in 2D: Rb&o,Mg,-,F,
are well-studied representations. The theory has demonstrated that such
dilute antiferromagnets in a uniform field lie in the same universality class
as RFIM, i.e., they both exhibit identical critical phenomena. Careful
experimentation has established that the above materials are, in single-
crystal form, ideal RFIM systems (closely corresponding to the theoretical-
model requirements).
At constant field there are two ways to cross the phase-transition line as
indicated in Fig. 7.10. One path is a direct field cooling (FC); the second
is ZFC and then applying the field and heating. These paths are illustrated
by the bold lines with arrows. Dramatic differences occur in the static
experimental properties between the two crossings. For ZFC all measure-
ments (specific heat, Mossbauer effect, neutron scattering, etc.) evidence a
sharp phase transition to a long-range ordered antiferromagnetic state with
unique critical exponents which are distinct (crossover effects have occurred)
from those of the random-exchange Ising model (REIM) or the zero-field
behaviour. Of particular importance are the resolution-limited Bragg-peaks
of the neutron scattering which track the long-range order right up to
T,(H). In stark contrast, FC experiments exhibit broadened transitions
characteristic of domains (short-range order) being frozen in below T,(H).
Here the FC procedure causes the system to lose equilibrium and enter a
metastable, frozen, domain state without long-range order. Above an
‘equilibrium’ boundary T&H) > T,(H), the two procedures (ZFC and FC)
yield the same experimental properties. So the domain formation and loss
of equilibrium with its peculiar dynamics begin already at T,,(H) for FC
and this masks the determination of T,(H). Figure 7.11 collects the various
results into a schematic H-T phase diagram for d = 3. Now H2’+ is plotted
as ordinate since this gives a straight line to the scaling behaviour of T,
with Hz’+, where I$ (= 1.42) is the crossover exponent. At sufficiently large
H for ZFC there is a crossover from REIM critical phenomenon to RFIM
critical phenomenon. The RFIM region of H-T space is delineated by the
two dashed lines labelled T; and T,+,. Within this region critical behaviour
and exponents corresponding to the RFIM will be observed. As we approach
the equilibrium phase-transition line [T,(H)] the critical slowing down
associated with the RFIM transition is encountered. The T,(H) line (see
Fig. 7.11) is weakly time or frequency dependent as indicated by the shaded
region surrounding it (TZ to Tz) with the various measurement times given.
Outside these narrow triangles of the experimental time window the system
is always in equilibrium on all accessible time scales. If a FC procedure is
used to approach the T,(H) line, the system willfirst confront an equilibrium
7.3 Random-field Ising model 219

Paramagnetic

i 1 FC ! / RFIM ,

Random exchange
‘,\ /’ REIM

TN
Temperature

Fig. 7.11 Schematic H-T phase diagram for the diluted AF, d = 3, RFIM system
at a fixed concentration, following a ZFC procedure. New RFIM critical behaviour
is observed for T,; < T < TA. All boundaries, both static and dynamic, obey REIM
to RFIM crossover scaling HU+, with Q, = 1.42. T*(H) indicates the onset of the
dynamic rounding of the critical behaviour associated with the extreme slowing
down as T + T,(H). It is only weakly time (frequency) dependent, as indicated by
the characteristic times associated with particular measurements. Outside of the
shaded region, the system is in equilibrium on all accessible time scales. In an FC
route (inset), the system falls off equilibrium at T,,(H), preventing the correlation
length 5 from diverging further for T < T,,(H) and ultimately leading to a frozen-
in (non-equilibrium) domain state below T,(H). From Jaccarino and King (1990).

line labelled T,,(H) in the ellipsoidal inset of Fig. 7.11. Now the system
falls out of equilibrium for T < T,,(H), thereby preventing the correlation
length 5 from further diverging and ultimately leading to a non-equilibrium
frozen-in domain state which forms below T,(H).
In 2D RFIM specific-heat and neutron-scattering measurements of
RbzCoXMg,-xF4 display a systematic rounding of the critical behaviour,
independent of the cooling procedure. There is no hysteresis in the vicinity
of the ‘destroyed’ phase transition at ‘T,(H)‘. From these data it was
concluded that d, 2 2. By utilizing the field scaling of the non-equilibrium
properties well below ‘T,(H)‘, the crossover exponent + is determined to
be 1.74, indistinguishable from the pure Ising value of 1.75. Figure 7.12
gives the schematic H-T phase diagram for the diluted antiferromagnetic
d = 2 case. Since hysteresis only occurs below ‘T,(H)‘, the FC and ZFC
procedures yield identical results in the surrounding ‘T,(H)’ region which
represents an equilibrium domain state, i.e., no long-range order or 5
remains finite. When hysteresis finally begins (see ellipsoidal inset of Fig.
7.12), it is to a non-equilibrium, metastable, frozen state that depends on
220 Spin-glass analogues

d=2
Domain state Paramagnetic

‘N
TEMPERATURE
Fig. 7.12 Schematic H-T phase diagram for the diluted AF, d = 2 = dl RFIM
system at a fixed concentration. The destroyed transition ‘Z’,(H)’ broadens and
shifts with HZ+ crossover scaling; it separates a domain state below ‘T,(H)’ from a
paramagnetic one above. Here + = 1.75. A freezing (metastability) boundary T,(H)
exists well below ‘T,(H)’ and also scales as Hz’+; its position depends on the
measurement time scale 7. Notice that destruction of the phase transition is an
equilibrium event. Unlike the d = 3 case, non-equilibrium behaviour at d = 2 has
nothing to do with extreme critical slowing down. From Jaccarino and King (1990).

the time scale of the measurement as indicated in the figure. Here 7’r(~) is
the time-dependent freezing temperature which tracks this loss of equilibrium.
Note the distinction between this freezing (non-equilibrium) behaviour and
the extreme critical slowing down in equilibrium of the rounded or destroyed
phase transition at ‘T,(H)‘.
Returning to the 3D RFIM systems, we briefly examine the dynamics of
the phase transition. Theory (Villian, 1985; Fisher, 1986) predicts an
activated-dynamics model where now the relaxation time becomes

CO
1
7
- = exp (ke) = exp 1 (7.9)
70 L[ T- T,(H)]“e
here we have used the standard result for the RFIM correlation length

5= 50 (7.10)

and 8 is a new ‘violation of hyperscaling’ critical exponent which controls

the anomalous growth of the free energy in a correlation volume via the
scaling relation 2-o = (d-8)~. The above application of activated dynamics
7.3 Random-field Ising model 221

should be compared to our previous usage in the 2D, Ising spin glass
(Tc = 0) of section 6.2. According to the RFIM theory, the free energy
and the associated thermodynamic quantities vary logarithmically with time
or frequency rather than linearly. For example, the frequency-dependent
Free energy scales as

(7.11)

and the ac-susceptibility as

x(w, T) = cy-” x
ln(o/o,)
____
Se [ 1
Hence, there exists a dynamical rounding temperature which limits the
(7.12)

divergences of the experimental quantities to a region of reduced temperature

-1
T- Tc(H

TN
)

I I-
> t*(w) cf ln- w ev
WO
since dynamical effects (rounding) should start appearing for ln(w/w,)/[e = 1.
(7.13)

In addition, the peak height of the ac-susceptibility under activated scaling

becomes a function of frequency

[X’WI, - /In(fj 1: (7.14)

which in the limit cx (the specific-heat exponent) + 0,

[X’WI, = In [ln (w/w,)] (7.15)

Measurements on (FeZn)F, have been performed employing ac-suscepti-
bility for the higher frequencies (kHz) and Faraday rotation for the lower
(mHz). Some typical low-frequency results are displayed in Fig. 7.13 where
the critical part of x’(w) is extracted as shown in the inset. Here it is
important to note, in contradiction to the spin glasses, the peak position
does not shift in temperature as a function of frequency. This critical x:(w)
begins to diverge at the lower reduced temperatures (solid straight line)
but breaks away or rounds at t*(w) given by the arrows. Also the peak
height [x’(w)], may be determined from the figure in arbitrary units. For
the lowest reduced temperatures (shaded area) the concentration gradient
af even the best single-crystal sample of Fe0.46Zn,,54F2 prevents any
meaningful interpretation. When the above data are combined with the
Faraday-rotation susceptibility, six decades of frequency can be spanned.
Comparison may then be made between the predictions of conventional
dynamic scaling: t*(w) = wl’*” and [x’(o)]r = wmulr”, and the above logarith-
mic behaviours of the activated dynamic scaling. The latest experimental
results suggest the activated-dynamics model yields a better fit to the above
222 Spin-glass analogues

OCICOI o.001 0.01

Fig. 7.13 x:(w) versus (t(, the reduced temperature, at three frequencies after
background subtraction, see inset. The open and closed symbols refer to T < T,(H)
and T < T,(H), respectively. Rounding of the transition due to the concentration
radient occurs only within the shaded region, i.e. ItI < 2 x 10d4. The expected In
f t( critical behaviour of X;(W) is used to determine a dynamic rounding temperature
t*(w) as indicated. From King et al. (1986).

two parameters with 8 = 1.0 and o, = lo7 rad/s. Since Y has also been
found from neutron scattering to be equal to one, the product 8v = 1.0
and hence the relaxation form is a simple Vogel-Fulcher type:

7 = r,exp (7.16)

Further effects of the ‘extreme’ critical slowing down can be observed in

the neutron-scattering experiments. For example, there is a rounding of
scattering intensity at the antiferromagnetic Bragg peak. The width of the
peak scales with reduced temperature as ZP’+ as expected from RFIM
dynamical crossover scaling. And a time dependence in the height of the
peak appears near T,(H) in ZFC experiments. Here a logarithmic increase
in peak-height occurs for times 10 to lo4 s upon warming through T,(H).
The good agreement between the various experiments and the theoretical
models has led to a basic understanding of both the 2D and 3D RFIM and
their ‘destroyed’ and real phase transitions.

7.4 OTHER ‘SPIN-OFFS

In this section we cull three non-physics analogues of the spin-glass problem.

Our choices for discussion while arbitrary are governed by our search for
7.4 Other ‘spin-off? 223

interesting and practical examples of how the spin-glass concepts and

mathematical methods can be applied to totally different types of complex
systems. Here the complexity of the problem plays a major role and has
even forged open a new area of physics called ‘complexity’. We consider
the three spin-offs: combinatorial optimization, neural networks and
biological evolution. While our treatment will be cursory we, nevertheless,
hope to illustrate the usefulness and direct contact of the spin glasses. We
begin with the combinatorial-optimization topic, or for us the ‘travelling
salesman’ problem: What is the shortest distance a salesman must travel to
visit a given number of cities and return to the origin city?
Given a set of N points (cities) and the ‘distance’ between them -
iN(N-1) numbers - find the shortest path which traverses all of the points
and returns to its starting point (origin). The correspondence with a
travelling salesman is clear when we define the cost function as the total
length of the ‘tour’. The cheapest tour (optimization) has the shortest
length, and the basic problem is to construct an algorithm which is able to
find the configuration of lowest cost for any instance (the N cities and their
various distances) with as little computing effort as possible. A useful
classification of the various optimization problems can be made according
to the time it takes to solve them on the computer. The ‘easy’ problems
are those that can be solved by a polynomial algorithm. This P-class employs
an algorithm which finds an optimal configuration in a computing time
growing like a power 12 of size N. Algorithms, where time complexity
functions are not bounded by such polynomials, generally require exponential
time functions. The problems in this exponential category (no algorithm is
known that provides an exact solution using a computation time that is
polynomial in the size of the input) are usually called NP-complete problems
where the NP refers to nondeterministic polynomial. If the running time
necessary to construct an optimal solution increases faster than any power
n of the problem size N, i.e., the computing time increases exponentially
with size, the problem is intractable. The travelling salesman is one such
problem since it becomes completely intractable (computing time + 00) for
N of the order a few hundred. Other non-physics cases of NP-complete
optimization problems are the placement and wiring of computer chips,
image processing, code design, digital-signal processing, seismic-data analysis,
neural networks, and many more.
In any event, connection of optimization and statistical (spin-glasses)
physics is obvious. The cost function corresponds to the energy of a
configuration. The exercise is to find the lowest cost or ground-state energy.
This is the optimum configuration or the lowest valley in the multi-
valley energy-configurational-space landscape of Fig. 5.11. In NP-complete
problems we deal with disorder and frustration, so that it is not possible to
satisfy simultaneously all the constraints which would locally minimize the
energy. To see the analogue more clearly let us consider the following
simple computer-design example. Suppose there are N circuits to be divided
224 Spin-glass analogues

between two chips placed one above the other. The number of signals that
pass between the ith and jth circuits is denoted by Kiie Next we attach a
two-value variable Si to each circuit i, such that Si = +l for the upper chip
and -1 for the lower one. The number of signals crossing from one chip
to the other is

C Kij a (Si-sj)* (7.17)

(ii)
The difference between the number of circuits on the two chips is Zsi. In
order to optimize the ‘eventual performance’ of the computer, e.g. minimize
the wiring length, reduce the noise and pick up, etc., the quantity

C (A-Kq/2)s,sj (7.18)
(ii)
should be minimized. Here A (a constant) takes into account the imbalance
in the number of circuits on the two chips and in the number of boundary
crossings. The above quantity to be minimized resembles an Ising, spin-
glass hamiltonian where the ‘interactions’ are the short-range ferromagnetic-
like Kij and the long-range antiferromagnetic-like A. Due to the competing
interactions and the intrinsic randomness, frustration results, and thus, we
have all the basic ingredients of a spin glass.
In order to secure the analogy of combinatorial optimization with the
spin glasses let us re-examine the SK-model in the former language. An
instance of this problem is a sample consisting of a given number of spins
N and a set of couplings Jij where 1 < i < j < N. The domain of possible
instances is defined by a probability distribution which reproduces the
choice of couplings. In the SK-model each Jij is taken as an independent
random variable with a gaussian distribution: exp [-(N/2).3. A configuration
is characterized by the values of the N Ising spins, Si = +l. And finally
the cost function is the energy given by the hamiltonian
%e= - ~ JijSiSj (7.19)
Iti<j<N

Barahona (1982) has demonstrated that finding the ground-state energy

(lowest or optimal cost function) is an NP-complete problem. In addition,
Parisi using the RSB solution has determined what the ground-state energy
should be in the limit of large N. Thus, the SK model has become one of
the most studied of NP-complete problems and the knowledge gained here
is very useful in treating other situations in combinatorial optimization.
For such NP-complete problems there are two main ways of attempting
a near-optimum solution beyond that of simply trying clever algorithms. In
the first, the so-called ‘iterative improvement technique’, a plausible
configuration of spin orientations, the circuit locations for the above ‘chip’
example or the cities route for the travelling salesman, is chosen and various
rearrangements of the bonds are tried. Only those rearrangements which
7.4 Other ‘spin-offs’ 225

lower the energy are accepted. This procedure is equivalent to a rapid

quench of the system down to T = 0. Accordingly, it can often become
stuck in one of the local minima and the spin, bond, et al. rearrangements
are not sufficient to eliminate this trap in a metastable configuration. Hence,
we are ‘back to employing special algorithms and tricks to free the system
and continue the search for the true ground state.
A second and more appealing general method is that of stimulated
annealing (Kirkpatrick et al., 1983). Here one introduces a fictitious
temperature as a parameter into classical optimization problems. This
artificial parameter permits a slow cooling Monte-Carlo simulation to be
carried out. Remember from section 5.12 such MC-simulations use a
Metropolis acceptance criterion exp [-AE/kBT] which has the temperature
as a natural variable. The stimulated-annealing algorithm begins by creating
a high-temperature state where the system can probe every possible region
of configurational space. Then, we sZow cool the system so that it settles
into a state of lower energy. Visualize the rough multi-valley landscape of
Fig. 5.11 and slowly move downwards a horizontal line or bar corresponding
to a reduction in E = kBT. If, at an early stage, the system becomes stuck
in a high-lying valley, then by slightly heating and slowly retooling it has a
good chance of escaping over the nearest ‘mountain pass’ to seek a deeper
valley or lower energy state. After many cycles of reheating and slow
cooling, the simulated annealing algorithm has a high probability to yield
a good solution, namely, a low-energy state which comes reasonably close
to the very lowest one. To find the latter is extremely unlikely because of
the huge amount of configuration space and the many nearly degenerate
ground states.
The annealing schedule is the most difficult part of the procedure. One
must choose a proper cooling rate r = AT/t for decreasing the T (t is the
number of Monte-Carlo steps, MCS). If the cooling rate is too rapid, the
final state will give too high an energy. Oppositely, too slow a rate is
wasteful of computer time. Grest et al. (1987) have performed a detailed
computer analysis on the ground-state properties for six model spin-glass
systems and the travelling salesman problem using simulated annealing. The
results showed the ground-state energy or cost function to be sensitive
function of the cooling rate r. For 1D nearest-neighbour gaussian and 2D
nearest-neighbour +-J and gaussian models
E(r) = E, + C,P (7.20)
In contrast for the 3D ? J, a two-layer + J and the infinite-range SK models
along with the N-city travelling salesman problem, the dependence on r is
much slower

E(r) = E, - C, sr (7.21)
( 1
The conclusion is that for the latter four models the slow logarithmic
226 Spin-glass analogues

dependence of finding the ground-state energy E,, makes the problem

intractable, i.e., an exponentially large amount of computer time is needed
to offset the very weak (ln r)-l dependence of E(r), thus it is NP-complete.
Conversely, the polynomial r-dependence for the former three models would
place them in P-class, even though they are not expected to exhibit a finite-
temperature phase transition.
Let us examine the simulated-annealing results for the travelling-salesman
problem given in Fig. 7.14. The plot shows the various distances travelled,
1, of a U&city and a 400-city tour as a function of l/hi r (r = AT/r). 1 is
defined as the total distance traversed between the N-cities randomly
distributed on a unit square divided by the square root of N:

+pk (7.22)
(ii) J-N
This is the quantity to be optimized by finding the shortest length lmin. Als
are equivalent to AEs and the ‘temperature’ equivalent generates the
acceptance criterion for a given configuration. Initially, at high temperature,
the algorithm accepts deteriorations in the cost function with a high
probability. However, in the course of the algorithm’s running time the
probability is gradually decreased to zero. This is accomplished by a control
parameter or temperature c which is a function of t, the number of MCS,
whereby c + 0 at the end of the computer run. Notice that we are still far
away from lmin with the finite rs used in the algorithm. So we must
extrapolate to r (or l/in r) + 0 as done in Fig. 7.14. The extrapolated
results of fmin agree nicely with the ‘exact’ value for N = 100 and the
‘expected’ value for N = 400.

1.0

A 400 City Tour

0 100 City Tour

0.61 ’ 0.2
0 0.1 ’ 0.3
’ 0.4
’ 0.5
’ 0.6
’ 0.7
’ 0.6
’ 0.9’ ’
1.0
’
1.1 . 2
-(l/lnr) l 10

Fig. 7.14 Optimal tour length 1versus - (ln r)-l for the N-city ‘travelling salesman’
problem for N = 100 and 400. From Grest et al. (1987).
7.4 Other ‘spin-offs’ 227

In 1982 Hopfield proposed a neural-network model which relied heavily

on spin-glass mathematics. He realized that his system of interacting neurons
was similar to a spin-glass freezing and could perform computations and
store information provided it had the appropriate dynamic rules. This model
is especially interesting since it simulates the operation of the brain more
accurately than a standard digital computer could. Let us first consider a
little of the background to neurons and their functions, and then move on
to draw the analogue with such networks and spin glasses. Finally, we shall
illustrate one of the unique properties of these neural networks related to
associative memory. In recent years the problem of neural-network models
has evolved into a rapidly developing field of its own.
The neurons are nerve cells which for simplicity have two discrete states,
namely active (fires or passes on an electrical impulse to another neuron)
and inactive or dormant. The process of firing is called synapses and is
accomplished via dendrites which reach out from the neuron or cell body
to thin axons which connect the various neurons. Whether a given neuron
remains in its present state or switches to the opposite one depends on the
states of all the other neurons connected to it. If the sum total of the
external stimuli exceeds a certain threshold, the neuron will change its
state. The nature of the network’s function, e.g. its computational task, is
determined by the pattern of neural connections or interactions. We put
this firing process of a neuron a bit more mathematically as follows. Denote
the state of the ith neuron at time t by variable vi which takes two values:
1 if the neuron is active and 0 if it is inactive. We let the strength of the
coupling or impulse connection (synaptic efficacies) between the ith and jth
neurons be Cij. Then, the sum total of the stimuli or impulses at the ith
neuron from all the other neurons is ZC, vi. Ti represents the threshold
for the firing of neuron i. Accordingly,
Vi= 1 ifZC,vj > Ti (7.23)
Vi=0 ifZC,Vj < Ti (7.i4)
To facilitate the analogy with the magnetic spin glasses we define

Si = 2vi- 1 Jij = $Cij and Hi = C$Cij- Ti (7.25)

i
This given the Ising-like pseudo-spins
Si=+l ifZJ,Sj-Hi>0 (7.26)
Si=-1 ifZJ,Sj-Hi<0 (7.27)
which can be expressed by the single formula
Si (Z Jij Sj - Hi) > 0 (7.28)
Therefore we can write a hamiltonian for the neural network as
228 Spin-glass analogues

Here the Jij need not be symmetric, yet both + and - (competing ferro-
and antiferromagnetic) pseudo-exchange interactions do exist since the
neuron synopses can be ‘excitatory’ or ‘inhibitory’. Also the various couplings
Jij may change with time, i.e., the system can learn in addition to store.
The set of all the spin values represents the collective state of the system
and such a situation closely resembles our SK, Ising hamiltonian in a
random external field. The solution of this hamiltonian, based upon our
previous spin-glass experience is expected to be represented by the multi-
valley, phase-space landscape of the many nearly-degenerate ground states.
In the present case these local energy minima of the collective system
correspond to patterns to be recognized or memories to be recalled or
other types of mental behaviour. It is the set of neural connections that
fixes the number location and meaning of the valleys. A given external
stimulus (input) determines the initial state of which neurons are firing and
which are not. Afterwards the system via the particular set of couplings
evolves to a special low valley (memorized pattern) that represents the class
to which the input stimulus belongs. Thus, the memorized patterns are
retrievable by their association with a given input pattern or addressable
by the content of the input pattern. A ‘basin of attraction’ surrounds the
lowest point in a valley such that if the system finds itself anywhere near
the valley it will be attracted to the lowest point - a non-ergodic process.
Again stimuli from the external world generate the selection of a special
solution or ground state. Note that with N-neurons there are 2N different
patterns for the network.
The choice of the Jij usually takes the form

where the 67 are quenched random variables assuming the values +l and
-1 with prescribed probabilities. p corresponds to the p-learned patterns
of N-bit words which are fixed by the p sets of { @} . From a straightforward
stability analysis, the stored patterns would only be stable at T = 0 K for
negligibly small values of p/N. However, we must recompute the stability
criterion using an upper limit of retrieval error of learned patterns. This
greatly increases the maximum number of patterns able to be stored in an
N-neuron network. As a first step towards the calculation of retrieval error,
we define a so-called Hamming distance d between a pure learned pattern
Sk and the pattern S’ as

d(S’, Sk) = $ [1-q(S’, Sk)] (7.31)

The Hamming distance represents the overlap between the two patterns
7.4 Other ‘spin-offs’ 229

and is the number of common bits. Compare it to the overlap distance of

pure states in a spin glass defined in section 5.7 where

q(S”, 9) = is- * SD (7.32)

It is exactly the same concept resulting from RSB and ultrametricity. The
average number of errors is given by the number of spins which do not
align with the learned or embedded pattern, and mathematically this
becomes

N, = :(1-q) (7.33)

Recent calculations have shown that for NJN = 1%) p/N = 0.14 and for
smaller error limits, the ratio p/N goes rapidly to zero. In fact, if reasonable
retrieval errors are permitted, the maximum storage capacity of learned
patterns may be increased to pmax = O.l4N, a rather high percent.
Relaxation of the neutron or spin system from the initial stimulus or state
to the final steady-state pattern or valley state is governed by the dynamics.
Any model of associated memory requires a well-defined dynamical process.
The Hopfield model assumes a single spin-flip Glauber dynamics which was
discussed in section 5.12 on Monte-Carlo spin-glass simulations. Here, after
every updating of a spin, it is the new configuration which is used to update
the next one. This is nothing but the heat-bath version of the Monte-Carlo
process required for the Ising (up/down) spin glasses. The other model,
called Little dynamics, updates all N-spins independently of each other.
Now at each time step all the spins simultaneously check their states against
the corresponding local fields. Hence, such an evolution of parallel
relaxations is called synchronous whereas the Hopfield ones are asynchronous
or series relaxations. Different forms are derived for the firing probability
of the neuron (or spin) at site i and t+Aht given the configuration {Si} at
time t. But it must be remembered that the dynamics are part of the model
and not a fundamental stochastic process. Various neuron-like electronic
circuits with the above properties have been built to study collective and
complicated computations and this is just the beginning of an exciting new
field.
We conclude our brief treatment of neural networks with a simple
illustration of associative memory whereby pieces of the stored information
serve to retrieve the entire memory. Consider the six-neuron network of
Fig. 7.15 and let the neurons have the + 1 or - 1 features corresponding to
the description of a person according to name, height, age, weight, hair
and eyes. See Table (a) as inset to Fig. 7.15. We wish to store three sets
of characteristics for three different people as memories A, B and C. See
Table (b) as inset to Fig. 7.15. Table (6) determines the wiring diagram of
the 6-neuron network. For example, the link between neurons (1) and (2)
must be for A + (+l)(+l), for B 3 (+1)(-l), and for C + (+l) (+l).
230 Spin-glass analogues

1 2 3 4 5 6

name height a&F weight hair eyes

-1 Smith ta11 old thin brown blue

+1 Jones short Young fat blond brown

T&b (b1 Node&

1 2 3 4 5 6

A +1 +1 +1 -1 -1 -1
I
B +1 -1 +1 +1 -1 +1

C +1 +1 -1 +1 -1 -1

Fig. 7.15 Associative memory with six nodes, or ‘neurons’, is linked by excitatorv
(solid line) and inhibitory (broken line) connections. The number of lines in eat
link indicates the strength of the connection; each solid line represents a connectio
strength of +l and each broken line represents a strength of -1. Each node migl
denote a characteristic of a person, as is shown in the Table (a). Suppose one wan
to store three memories, or sets of characteristics as given in Table (b). The node
that are supposed to be in the +l state are given an excitatory link to the othc
+1 nodes and an inhibitory link to the -1 nodes. To store information about a
three memories one simply adds up the connections. From Tank and Hopfiel
(1987).
7.4 Other ‘spin-offs’ 231

Since the sum of these products is +l we place one excitatory wire on a

solid line to connect (1) and (2). Repeating this procedure for neurons (2)
and (4), we obtain the sum equal to -1 which means one inhibitory wire
or a dashed line should connect (2) and (4). Note for connections (1) and
(5) and (2) and (6), the sum is equal to -3. Thus, 3 inhibitory wires must
be used for these connections and by following the same routine we can
establish the full wiring diagram for the six neurons of Fig. 7.15 according
to the characteristics of Table (b). Now when the circuit is turned on for
person or memory A, the network produces the correct pattern of current
flows to ensure that neurons (l), (2) and (3) have their +l states (solid
circles) while (4), (5) and (6) are in their -1 states (open circles). The
pattern is self-consistent, for, at each neuron the positive (solid arrow)
currents and negative (dashed arrow) currents always add up to have the
same sign as the neuron itself. Check this in Fig. 7.15! Note the associative-
memory character. If the network is given only partial data, e.g. a thin,
short Jones, it will immediately fall into a stable state from which one can
retrieve the entire memory, namely A of a thin, short, young Jones with
brown hair and blue eyes.
We conclude this section by mentioning another link between the spin
glasses and biology. Naturally, we could use the ultrametric tree of RSB (see
Fig. 5.7) as a classification scheme for the evolution of a species. However,
we wish to go further and consider the central question of biological evolution:
how does a ‘soup’ of small molecules evolve into highly-organized, information-
carrying macro-molecules? Here attempts have been made to construct
mathematical models of the first critical stages of molecular evolution where
an information transition occurs from little to much information. Two basic
elements, stability and diversity of the species, are required for their evolution.
Without stability one species would quickly mutate into another. If there is
one only stable state, thus no diversity, no possibility would exist to create
more complex structures with greater amounts of information. A survival
function selects the molecule’s chance of formation and is equivalent to a
random interaction between the components. Once again we have the basic
elements of a spin glass: competing interactions, frustration, complexity, etc.
Simple mathematical models are made to account for these factors statistically.
While the words are different the calculations and pictures of these models
closely resemble those of the spin-glass problem.
One important common denominator constantly appears in all these spin-
offs. This is the rugged mountainous landscape in phase space. For the spin
glasses the vertical axis is the energy states of the peaks and valleys, for
the combinational optimization it is the cost function, for the neural network
it is the memories to be recalled or patterns to be recognized. And finally
for the biological evolution we have put into the model the survival-
probability function with its many peaks and valleys in the ‘state space’ of
the molecules. So perhaps we should all return to Fig. 5.11 for another
glance because a great deal of our modern science of complexity depends
232 Spin-glass analogues

upon it. And as mentioned at the beginning of this chapter we have

progressed far beyond the spin and the glass and entered into totally foreign
areas of life and evolution. Hence we stop.

FURTHER READING AND REFERENCES

Spin glass analogues

van Hemmen, J. L. and Morgenstem, I. (eds) (1983) Heidelberg Colloquium on

Spin Glasses, Lecture Notes in Physics 192, Springer, Heidelberg; (1987)
Heidelberg Colloquium on Glassy Dynamics, Lectures Notes in Physics 275,
Springer, Heidelberg.
Hochli, U. T., Knorr, K. and Loidl, A. (1990) Orientational Glasses, Advances in
Physics, 39.
Stein, D. L. (ed.) (1989) Complex Systems, Addison-Wesley, New York.
Kauffmann, S. A. (1989) Origins of Order: Self Organization and Selection in
Evolution.
Kandel, E.R. and Schwartz, J. H. (1985) Principles of Neural Science, Elsevier,
Amsterdam.
Stein, D. L. (ed.) (1992) Spin Glasses and Biology, Directions in Condensed Matter
Physics, Vol. 6, World Scientific, Singapore.
Zallen, R. (1983) The Physics of Amorphous Solids, Wiley, New York.

References

Barahona, F. (1982) J. Phys., A15, 3241.

Birge, N. O., Jeong, Y. H., Nagel, S. R., Bhattacharya, S. and Susman, S. (1984)
Phys. Rev. B30, 2306.
Courtens, E. (1984) Phys. Rev. Lett., 52, 69.
Fisher, D. (1986) Phys. Rev. Len., 56, 416.
Grest, G. S., Soukoulis, C. M., Levin, K. and Randelman, R. E. (1987) Heidelberg
Colloquium on Glassy Dynamics: Lecture Notes in Physics 275 (edited by J.
L. van Hemmen and I. Morgenstern) Springer, Heidelberg.
Hochli, U. T., Kofel, P. and Maglione, M. (1985) Phys. Rev., B32, 4546.
Hopfield, J. J. (1982) Proc. Nat. Acad. Sci. (USA), 79, 2554.
Imry, Y. and Ma, S. K. (1975) Phys. Rev. Leff., 35, 1399.
Jaccarino, V. and King, A. R. (1990) Physica, A163, 291.
John, S. and Lubensky, T. C. (1985) Phys. Rev. Len., 55, 1014.
King, A. R., Mydosh, J. A. and Jaccarino, V. (1986) Phys. Rev. Left., 56, 2525.
Kirkpatrick, S., Gelatt, Jr., C. D. and Vecchi, M. P. (1983) Science, 220, 671.
Maglione, M., H(ichli, U. T. and Joffrin, J. (1986) Phys. Rev. Lett., 57, 436.
Sethna, J. P. and Chow, K. S. (1985) Phase Transitions, 5, 315.
Sullivan, N. S., Devoret, M., Cowan, B. P. and Urbina, C. (1978) Phys. Rev.,
B17, 5016.
Sullivan, N. S. (1983) Proceedings of the Symposium on Quantum Fluidr and Solids
(edited by E. Adams and G. G. Ihas) American Institute of Physics,New York.
Tank, D. W. and Hopfield,, J. J. (1987) Scientific American, 257, 62.
Villain, J. (1985) J. Phys (Paris), 46, 1843.
8
The end (for now)

If you have come this far you must either be an expert on the spin glasses
or slightly disappointed and still confused. In order to improve and solidify
the first and to alleviate the second possibilities we offer this concluding
chapter. It will not be a succinct recapitulation of the preceding seven
chapters. Hopefully after a rereading they will become sufficiently clear
and to the point. The phenomenon and physics of ‘spin glass’ have been
explained via an experimental approach - the purpose of this monograph.
In this final chapter we wish to reiterate briefly, using other words, the
basic and key concept of the spin glasses, namely, symmetry breaking
resulting in the rough multi-valley landscape of the system’s energy in phase
space. Then, we proceed to one of the latest (1992) developments in the
experimental area that of mesoscopic or nanostructured spin glasses and
their special fluctuations. And afterwards we take a last quick look at the
theory: what does it say about dimensional crossover (3D + 2D) and how
does the freezing temperature depend on the length scale? This is followed
by some of the most recent computer simulations. Here there are appearing
some novel and altered interpretations. These final considerations will
clearly illustrate to us the ongoing stream of progress in a relatively dormant
field of physics. At last we end the book with a few compositional thoughts
about our journey through the spin glasses and where to find what.

8.1 SYMMETRY BREAKING AND PHASE SPACE

All second-order phase transitions, especially ferro- and antiferromagnetic

ones, break the symmetry of their respective hamiltonians, and thereby,
are characterized by an order parameter. For the above two possibilities
the spontaneous magnetization and the staggered magnetization, M, represent
the order parameters. This means an enormous reduction in the system’s
possible spin configurations to simply one (with the help of an infinitesimal
symmetry-breaking field in an Ising ferromagnet, see below). Such a
situation is illustrated in Fig. 8.1 where we plot the free energy versus
magnetization as a function of temperature. The figure symbolizes a Landau-
234 The end (for now)

UJ
T<Tc

1
-tvl- +M
Fig. 8.1 Free energy as a function of the magnetization (order parameter) for a
ferromagnetic above and below its Curie temperature Tc.

theory approach to a second-order phase transition. Above T, the free

energy has its minimum at M = 0 and all possible microscopic spin
configurations are available which satisfy this ‘null’ macroscopic condition.
Here the symmetry of the magnetic (Ising) hamiltonian

me= -JCSiSj-H~Si (8.1)

ii I
is preserved along with the ergodic hypothesis. However, below T, there
is the ferromagnetic transition and the spins must either all point up or all
down and thus finite values of + M appear. With our tiny symmetry-
breaking field one of these two possibilities is chosen. Hence, only a single
configuration of spins is allowed as lowest energy or ground state. Such a
drastic constraint breaks the symmetry of the hamiltonian (8.1) and further
violates the ergodicity by limiting the phase space to either the + M or the
-M valley. Remember restriction of phase space is called broken ergodicity,
in this case it is ‘trivial’. The above description has now become standard
in the theory of second-order phase-transition. Here the order is long-range
and periodic. And we can calculate the ground-state energy, for there is
only one, and further how the order grows with temperature M(T). The
critical-phenomenon problem has also been solved and the various critical
exponents are known for the different universality classes - end of story.
Using this powerful theoretical framework let us consider the spin glasses.
Attempts at devising a single order parameter, e.g. the Edwards-Anderson
order parameter, have failed. In addition, to calculate the true ground-state
energy is impossible or better said intractable. And all of these difficulties
occur already within the mean-field or simplest theoretical approximation.
The net result after many years of work was a continuous order
parameter q(x) and the corresponding multi-degenerate ground-states. So
the transitional behaviour is very much different in a spin glass than in a
ferromagnet.
We again look at our multi-valley picture which is now repeated in Fig.
8.2 as a function of temperature. First there is the case of T > Tf where
the paramagnetic state is modified by the various clusters that introduce
small, local minima into the otherwise symmetric behaviour situated about
8.1 Symmetry breaking and phase space 235

(a>

(b)

-Mj/M

Fig. 8.2 Sketch of the multi-valley landscape as a function of temperature for a

spin glass.
236 The end (for now)

M = 0. The local minima represent the finite values of the cluster

magnetization, however, because of the high temperature they are metastable
and give only a fluctuation contribution to the free energy. It is the large
temperature which causes the various energy barriers to be surmounted,
and thus, the equilibrium thermodynamic state will be an average over a
vast region of phase space with zero net magnetization. Equilibrium is now
attained because the cluster barriers are much too small compared to the
temperature.
But what happens as we lower the temperature to just under T,? Here
the variations in the free-energy have increased, i.e., the magnitudes of the
energy barriers have become larger but remain finite. This is a direct
consequence of our interacting system of spins and their co-operative
freezing. Note now the appearance of many low-lying valleys possessing
different internal structures of sub-valleys with various magnetizations.
Remember the figure is only a schematic ‘bare-bones’ illustration, the real
CuMn alloy is much more complex. Yet for this case the temperature is
sufficient and the barrier heights not too intimidating so that the local
metastabilities are quickly removed and the macroscopic free energy can
span a reasonable region phase space within the available time. We would
say because of the relatively rapid relaxation at Tf a quasi-equilibrium is
maintained with, however, an increased chance for the system to become
stuck in one of the low-lying valleys. The problem here is we do not know
which one or even if it possesses the lowest possible energy. Furthermore,
an external field will greatly influence the barrier-height distribution causing
certain magnetization values to the preferred. All these processes are
difficult to calculate since we do not know the free-energy landscape, or to
which state the system will migrate, or how it depends on field.
If the number of these local magnetizations is large (multidegenerate) as
it will be at low temperatures (T -+ T,), then some of the energy barriers
separating the different local mp will tend towards infinite height - see Fig.
8.2(c). Accordingly, we can partition the landscape into mutually inaccessible
valleys where each of the valleys corresponds to a thermodynamic phase
similar to the two ferromagnetic minima in Fig. 8.1. Within each of the
inaccessible valleys there will be many sub-valleys or smaller structures
separated by finite (small) barriers. Since these sub-valleys have various
non-equal energy minima, they represent the metastable states of the frozen
spin-glass and are responsible for the unique experimental effects at T Q
Tf. When a system becomes trapped in one such valley, it will exhibit
properties specific to that particular valley, i.e., non-trivial broken ergodicity.
Such properties differ from those in true equilibrium, the latter involve
averages scanning all valleys with the appropriate thermal weights. It is the
EA order parameter qnA which measures the mean-square single-valley
local spontaneous magnetization averaged over all possible single-valleys.
Note how @?A will be non-zero for the distribution of inaccessible valleys
given in Fig. 8.2(c) since the + m,s are all squared eliminating the possibility
8.2 Mesoscopic spin glasses 23’7

of + cancellations. qEA is not, as already noted in Chapter 5, related to

the equilibrium magnetization. Such requires a new order parameter q
which takes into account the intervalley contributions. This necessitates the
accumulation of a large region of phase space with the incorporation of
many low-lying ‘almost inaccessible’ valleys. A dynamic picture is useful
here. We let the barrier heights be large but not infinite so that low-
probability transitions over the barriers are allowed as illustrated in Fig.
8.2(c). This will simply take a very long time according to T = r,, exp
(+ EB lkB T). If we use a time-scale long enough for the system to surmount
the barriers and to pass statistically many times through all the valleys with
a significant thermal weight, then true equilibrium is attained by including
these intervalley contributions. Granted it would take a much longer time
than is accessible to experiment, but that’s the price we pay for our complex
multi-valley landscape. These concepts were alluded to in Chapter 5 and
can be phrased in mathematical terms. For example, a third order parameter
A can be introduced which measures the degree of broken ergodicity:

A = qEA - 4 (8.2)
Notice that A + 0 when we return to our single equilibrium phase of Fig.
8.1 (T < T, with a small +H), even though one half of the phase space
has been removed.
As stated previously the key to the spin glass and its many analogues is
this peculiar and undeterminable multi-layer landscape - a unique but
necessary conception. This is an important bit of new physics with significant
consequences that can be tracked experimentally. For example, if we
measure the ac-susceptibility we are effectively probing (IuA and its short-
time quasi-equilibrium state. However, if we try to ‘mimic’ the very long-
time behaviour by field cooling, we obtain a measure of q which gives the
average equilibrium susceptibility. Can you remember this discussion in
section 5.8? More important is not to forget the picture and underlying
concepts of the many valleys, so you should look once again at Figs. 8.1
and 8.2 for here is the heart of the spin-glass problem.

8.2 MESOSCOPIC SPIN GLASSES

Very recently (1990s) the study of spin glasses has turned to samples with
reduced size. This means one or more dimension has been constrained to
less than c. 1000 8, lengths. The structure may be a very thin film (2D), a
narrow ‘wire’ or strip (1D) with both reduced thickness and width, or a
‘clump’ of the material usually called a dot (OD). (The latter has not yet
been tried for a spin glass.) Such samples are fabricated using the modern
sub-micron and nanostructure techniques currently available. The word
applied to describe these samples and their associated physics is mesoscopy.
238 The end (for now)

At the present time it is a very popular and active area especially with
semiconductors where its consequences and benefits are obvious.
The unique effects of mesoscopic physics (which lies between the very
small atomic scale, = 1 A, microscopy and the very large bulk behaviours,
2 few l.r,rn, macroscopy) are caused by the smallest sample length L being
comparable to one or more of critical length scales of the given problem.
Thus, for a mesoscopic conductor, L could be the same order of magnitude
sas

(i) the elastic mean free path e = vg, where vF is the Fermi velocity
and T is an elastic scattering time;
(ii) the inelastic mean free path [in which depends on the inverse
temperature to some power and includes now magnetic inelastic
scattering processes, e.g. spin flip;
(iii) the phase-coherence length Q, which permits ‘time-reversal’ interference
between the electron states.

For the latter we can write in the diffusive-scattering regime &, = (Dr+);
where D is the diffusion constant and T+ is related to the inelastic scattering
processes. A typical metal has 4 = 50 nm and 4, = 1000 nm at 1 K, hence,
L could easily be made much less than e, by sub-micron fabrication and
going to yet lower temperatures. Such a condition (L < Q,) eliminates a
complete cancellation of the different quantum interference terms (remember
the electron is a wave) between the various diffusive (4 4 L) paths. This
in turn produces finite fluctuations in the conductivity 6G, which depend
sensitively (in both magnitude and sign) on the specific placement of the
scattering centres or ‘impurities’. These direct interferences between multiply-
scattered electron waves give the so-called universal conductance fluctuations
(UCF) of amplitude

G UCF = (3 (@Q2) - ($ (8.3)

where the ( ) brackets represent ensemble averaging over different samples

with different impurity configurations. If a magnetic field is applied to such
a mesoscopic specimen, then the phase of the electron waves is changed
by 2n each time a flux quantum cPO= h/e passes through the phase-
coherent area. This gives rise to (Aharonov-Bohm-like) conductance
oscillations with field sweep of wavelength h/e. Such behaviours have
been investigated already for many years in small non-magnetic metals
mostly using magneto-resistance and noise-spectrum (l/f) techniques. Some
pertinent and introductory references are listed at the end of this chapter,
if one wishes to delve further into this area. Let us now consider what
happens if we replace the above ‘impurities’ by magnetic ones, e.g. CuMn
and AuFe, and examine the possibility of spin scattering both above and
below Tf.
8.2 Mesoscopic spin glasses 239

First of all we must convert the UCF equation to include spin-dependent

scattering. Theory (Feng et al., 1987) has dealt with this problem, and in
particular, how a small temperature change 6T (thermal cycling) will affect
the conductance within the frozen state of a spin glass, where most of the
inelastic spin-flip scattering has ceased.

(8.4)

Here the 4, e, and L have been previously defined, and kF and TF are the
Fermi momentum and temperature. Note that the freezing temperature Tf
enters this equation. If k&T,IT, 2 P/e& then we have the saturation
regime

which should be measurable as l/f-noise in the conductance. The 6T

necessary to induce such conductance changes is related via the droplet
model to a length scale that causes significant spin reorganizations. Or one
could simply cool the spin glass through Tf and tract how the noise spectral
density changes with temperature. Another method would be to detect the
UCF via field sweeps in magneto-conductance experiments and compare
the fluctuation spectrum above and below Tf. Here in a mesoscopic sample
one would attempt to find spin-glass fingerprints of the particular frozen
configuration.
The noise-spectrum measurements do indeed show a steep increase in
their spectral density below Tf for various concentrations. These experiments
were carried out on thin films of CuMn (5-10 at. %) Xl-70 nm thick and
AuFe (0.1-l at. %) 12-25 nm thick. A typical plot of the ‘noise-parameter’
4f, 0 K f%(f), w h ere S,(f) is the spectral density of fluctuations at
frequency f in the resistance R, is given in Fig. 8.3 for CuMn. Note how
the freezing temperatures determined from max (dddT) are very close to
the susceptibility-cusp temperature. Similar results are obtained for both
CuMn and AuFe in zero external field. A disturbing disagreement occurs
between the two systems when the external field is applied. For CuMn
little happens to the temperature dependence of SR (1 Hz) spanning Tf in
2 T field, however, for AuFe the increase of the fluctuation spectrum is
washed out already in 1 T. One possible reason for this disparity is the
factor of 10 larger concentration used in the CuMn film; another is the
single low (1 Hz) frequency of the noise parameter. Yet it does seem
strange that the relatively large fields will have no effect on the spin
fluctuations of a spin glass.
The magneto-conductance experiments were performed on strip-like
nanostructures of CuMn (0.1 at. %) with thickness 40 nm, 90 nm width
(w) and c. 2 km length with intervening potential-contact strips. At this
240 The end (for now)

1o-6
0 20 40 60 80
T(K)
Fig. 8.3 Electrical resistance ‘noise-parameter’, a(f, T), for various concentrations
of CuMn; from Israeloff et al. (1989).

concentration the average distance between the Mn atoms is 2.5 nm and

an electron mean free path of 30 nm was estimated. So the net structure is
approaching a 1D wire whose resistance can be measured at = 0.5 pm
intervals. Some salient features of these pioneering studies are displayed in
Figs. 8.4 and 8.5. The first shows the magneto-resistance for the wire with
the above dimensions compared to a codeposited CuMn film of width 2 pm
and much larger separations of potential contacts. Note the clear quantum
fluctuations in the wire at low temperature (23 mK). Such fluctuations are
called a ‘magneto-fingerprint’ and are attributed to the magnetic field
inducing a phase shift in the electron wave function. Recall the phase
change is related to the flux quantum a0 = h/e; thus fluctuations on a field
scale of h, = (h/e)l(w&,), where w is the width of the wire and Q, is the
phase coherence length, can occur. A second and larger field scale is created
by the distortions or rotations of the frozen spins away from the initial
(ZFC) configuration by the external field. At 25 mK the conductance
fluctuations have an average peak-to-peak amplitude of 6G = 0.2 e2/h.
In Fig. 8.5 the mesoscopic wire is fabricated into a square ring 0.4 p,rn
on a side. As the field is swept in different ranges Aharanov-Bohm
conductance oscillations are observed with period h/e for the AH = 0.25 T
in the intervals of 0.1-0.35 T and 5.0-5.25 T. Notice the sharpness of the
oscillation which gives h, = 260 gauss and &, = 0.5 brn. When the field is
repeatedly swept back and forth at T Q Tf the same fluctuation spectrum
is retraceable, i.e., there exists a large correlation coefficient C = 0.95.
8.3 Recent theoretical progress 241

-1

“0 -2
C
x
-3

5
a -4

-5 -2.5

-6 -3.0
0 15 45 60
;i” (kG)
Fig. 8.4 Upper three curves magneto-resistance as a function of field for a CuMn
(0.1 at. %) film (420 8, thick by 2 Frn wide). Lower two curves same as above but
for a CuMn (0.1 at. %) wire (420 A thick by 900 A wide by 2 km long). From de
Vegvar, Levy and Fulton (1991).

However, once the wire is heated above Tr and retooled a new fluctuation
spectrum will appear in the magneto-conductance related to the new
microscopic orientations of frozen spin. It is estimated that within the above
wire dimensions there are lo5 Mn spin and 25% of these must reorient to
produce of 6G = e2h. The question here is what does reorientation mean,
certainly not a spin flip, perhaps a small collective distortion would be
sufficient. But this is just the beginning of the new and exciting field of
mesoscopic spin glasses.

8.3 RECENT THEORETICAL PROGRESS

So what has been happening with the theory? In this subsection we mention
one ‘older’ (late 1980s) and one brand-new (1992) pieces of progress. The
former has to do with the droplet model and its approach towards answering
the question of dimensional crossover: 3D + 2D and this influence on T,.
Certainly it is related to the previous topic of mesoscopic or restricted-
dimension effects. The second concerns some very recent numerical
simulations carried out in improved ways or on more realistic models which
indicate that specific revisions are perhaps necessary in the long-accepted
theoretical conclusions.
242 The end (for now)

0.20

0.00
0.0 0.5 1.0 1.5 2.0 2.5
AH (kG)
Fig. 8.5 Aharonov-Bohm conductance oscillations (AG versus AH) for a CuMn
$.l;t. %) square ring 0.4 pm sides at 23 mK; from de Vegvar, Levy and Fulton

The first question is how does Tf change as we reduce one of the sample
dimensions beginning with a bulk 3D system. For example, we can decrease
the thickness of the sample until the 2D limit is reached, all the while
tracking Tf. The simplest theoretical approach is to use finite-size scaling
which introduces the new variable y = L/c(t). Here L is the sample
dimension that we are changing and t[t = (T- T,)/T,] is the correlation
length for an equilibrium phase transition. For the spin-glass case we have
a reduction in T,(L) at a given constant value of y. Thus, setting
T,(L = 03) = T, we obtain
T,(L) - T,(L=~) cc -$
T,(L=m) (8.6)

where v is the correlation-length exponent 5 0: t-” and L, is the length

scale at which T,(L,) = 0.
In 1987-88 Fisher and Huse extended these ideas via the droplet model
and included the effects of the long relaxation times in the spin glasses.
Their result was a measurement time factor T, multiplying the previous
L-l”‘. Accordingly
T,(L) - T,(L=m) a L-& In T,
T,(L=m) [ HIL=3
[(b3+v~@33)v31-’
(8.7)
8.3 Recent theoretical progress 243

Now the subscripts on the exponents denote their dimensionality 2D or

3D. Recall from section 5.9 that z, + and 8 represent the length-scale
critical exponents for dynamical scaling, activation barriers and free energy,
respectively. Notice six (v3, z3, G3, f13, & and YJ critical exponents
appear in the above equation. In the limits L + 0; 7, + 03, and hence,
T,(L)IT,(L=a) e 1

which for a constant long time of measurement simplifies to

T,(L)
T,(L=m) a L” (8.9)

with a being a combination of the other four (3r3, f13, I& and v2) critical
exponents.
The above equations suggest the ways in which the experimentalists
should plot their data. The different critical exponents may be estimated
from MC simulations or the measurements on ideal spin glasses - see
Chapter 6. And as mentioned in section 6.4 a number of groups are quite
active in preparing thinner and thinner films, usually of the canonical
(RKKY) spin glasses, in order to confront the theory. Figure 8.6 illustrates
one such ‘multilayer’ attempt. Here various Mn alloys of Cu or Ag, which
are kept separated from each other by a sufficiently thick interlayer of the
pure host metal (Cu or Ag), have their thicknesses reduced. T,(L) is taken

0.8- r’
.a
,600
0.67 12: ’ cuo.a~oo.,,/cu
’ cuo.oo~no.,,/Cu
0.4- ,9J A Cu0.0dno.dA8
0 Cu0,0dn0.0&u

* cuo.“~o..o,/cu

0 A80.01~n0.m./Cu
’ A8o.oo~oo.oo/cu

A A8o.dndA8

I .,.14

102 1

Fig. 8.6 Normalized freezing temperature plotted as a function of film thickness

for various CuMn and AgMn multilayers. Ws, (= L) represents the spin-glass layer
thickness on a logarithmic scale. The dashed curve is a fit to equation (8.6). From
Bass and Cowen (1992).
244 The end (for now)

as the peak in the ZFC magnetization or as the beginning of irreversibility

between ZFC and FC.
In Fig. 8.6 a reasonable fit to the collected data on the thicker films is
obtained using (8.6) with v = 1.6 and L, = 10 A. These latter values seem
physically meaningful - see Table 6.4. However, the very thin films (below
20 A) deviate from the universal curve with Tf remaining finite. Other
experiments on single-layer films of AuFe exhibit a different behaviour. If
(8.6) is used, v = 0.8 with L, = 70 A. Here also there are deviations for
the thinnest film thicknesses. Note the above spin glasses are not the ideal
Ising systems of the droplet model. Despite the severe experimental
difficulties inherent in producing the required high-quality and homogeneous
alloy layers, work is continuing and a more complete comparison with
theory should be shortly expected.
Before we leap into the latest of numerical simulations, let us succinctly
re-examine the computer folklore of the spin glasses and its conventional
wisdom or conclusions. Section 5.12 treated this issue in more detail. Up
until now, computer simulations and other numerical studies agree that a
phase transition occurs in Ising spin glasses of d = 3 for both short-range
(I’J) and the infinite-range (RKKY or mean-field) interactions. For these
cases Tf = J or A (the variance of the interaction distribution). In 2D the
above Ising model has Tf = 0 and 2 5 de (the lower critical dimension).
Remember the computer models are random bond ones where ferromagnetic
and anti-ferromagnetic couplings are randomly distributed on a regular
lattice of spins. For the Heisenberg case the simulations tell us that there
is no phase transition for d = 3 at finite temperature regardless of the type
of interaction, e.g. for RKKY, de = 3 and Tf = 0. Now, in order to induce
a phase transition, a bit of anisotropy is needed to break the Heisenberg
isotropy. Recall the use of the Dzyaloskinsky-Moriya interaction in section
3.3.5 for this purpose in the canonical (RKKY) CuMn and AuFe spin
glasses. The x-y universality class lies somewhere between the Ising and
Heisenberg models, with less numerical work having been performed and
no finite Tf phase transition anticipated in 2D and 3D. Nevertheless, after
all this use of computer time and energy, no one has yet resolved the
freezing temperature for an infinite-range (SK), ?J, Ising spin glass. It was
always expected to be simply Tr = J. Recently two different simulation
studies have considered this question, and surprisingly, one answer found
that Tf was significantly less than J. And, touching on another separate
issue, the Heisenberg models employed in the simulations are random
bonds, not the random sites of the canonical RKKY spin glasses. This was
one of the main reasons for experimentalists to create the ideal spin glasses
of Chapter 6. Now it is the (numerical) theorist’s turn to construct ‘more
realistic’ models, so how about random sites?
We briefly examine these two new types of numerical simulations. The
first (Campbell, 1992 and Bhatt and Young, 1992) investigated an infinite-
range Ising *J spin glass. Infinite range means that each of the N (up/down)
8.3 Recent theoretical progress 245

spins interacts with all the others (N-l) spins according to a constant
coupling strength J independent of the distance separating the spins. Only
the sign of J changes randomly. Here dimensionality does not play a role
and it is the size of the system given by the N spins that counts. Large-
scale (to N = 4096), long-time Monte-Carlo procedures were used and
averaged over various sets of bond configurations, i.e., sample averaging.
When the number of bond configurations averaged over was large enough,
the results confirmed Tf = J (Bhatt and Young) as predicted by the SK
continuous infinite-range model and reinforced by RSB theory. However,
for insufficient number of samples, the generated (non-linear) susceptibility
values were widely distributed and highly skewed. Thus the conclusion of
Campbell that Tf = 0.75 was biased because of the large fluctuations or
“rare samples” present in his simulations. Such effects can only be gotten
rid of by proper sample averaging in these finite size samples. So you see
there is still life, danger and contention in the spin glasses and certainly
other paragraphs of this book will require future revision as they continue
to evolve.
The ‘more-realistic’ model (Matsubara and Iguchi, 1992) is a random
alloy of RKKY-coupled, Heisenberg spins and the numerical simulation is
a hybrid MC spin-dynamic method. Here the model hamiltonian

X = -C J(r) SiSj
ij

is used with J(r) being given by equation (1.2). The classical Heisenberg
spins ISi\ = 1, are randomly distributed on the sites of an (L)3 fee lattice.
Parameters are chosen to represent CuMn with, for example, 5 at. % Mn.
The spin-glass susceptibility is calculated as

(8.10)

where ( 1 )= represents an average over K MCS per spin and ( . )o denotes

a sample average. This non-linear susceptibility is shown in Fig. 8.7 for T
in units of JolufjkB and with various L-size lattices. The error bars represent
an attempt to define the equilibrium values of Xso. Note that there is rro
anisotropy in the simulation, and yet Xso seems to diverge at low T for
large L (= 20) with K reaching 15 000 MCS. This indicates that a finite-
temperature phase transition is occurring. Such a conclusion may be put on
a more firm basis by employing finite-size scaling from which Tf = 0.07
and the v and q critical exponents may be derived (v = 0.74 and q = 0.69).
Previous studies of Heisenberg spin glasses, all resulting in no phase
transition, have limited themselves to bond models, i.e., a random
distribution of bonds in a periodic lattice of spins. The RKKY model is
necessarily a random site one with a J(r) interaction between the various
distance sites. So it would now seem that because of these recent MC
246 The end (‘for now)

+ L= 6
* L= fJ
...& L=1’)
80k o Lz1.2
ae o L=16
t. ;..:.\ : 0 L=20

E....
Il...g
-. +...+ :::*::: *::::,8:zax,...,..

o~.‘..“‘*‘.‘..*.“...l 0.1 0.2

0
T
Fig. 8.7 Temperature dependence of the spin-glass susceptibility for different size
(L) lattices of the RKKY coupled Heisenberg, random-site spin glass; from
Matsubara and Iguchi (1992).

simulations a substantial difference exists between a random-bond model

and a random-site one.
Another view on the Heisenberg spin glass is to consider chirul ordering.
This is the property of vector spins, which because of frustration can possess
a non-collinear or canted spin ordering. ‘Chirality’ designates the right- or
left-handedness of the spin structure and introduces a new universality class
with novel critical properties. Using numerical studies there are indications
that chirality (designated by +l for right-handed and -1 for left-handed)
often behaves like an Ising variable. Thus, the conjecture has been made
(Kawamura, 1992) for a chirality driven mechanism to create a finite-
temperature phase transition in the canonical (RKKY) spin glasses. As is
now crystal clear there is a lot more to come in the theory of spin glasses.

8.4 A FINAL WORD

Before we stop (for now) let us briefly look back at the composition of
this book. If you want to learn, put most simply, what a spin glass is and
8.4 A final word 247

how, in words and pictures, it freezes, you should return to Chapter 1 -

its beginning and its end. In between are the basic spin-glass ingredients:
randomness, magnetic interactions, anisotropy and frustration.
The spin glasses did not suddenly appear in the magnetic repertoire.
They are affiliated with various other (and sometimes older magnetic)
phenomena such as moment formation, the Kondo-effect, giant moments,
mictomagnetism or cluster glass, percolation behaviour and superparamagnet-
ism. Chapter 2 introduces these related concepts and offers a simple
description, or better said a physical picture, of them. The spin glasses are
then placed in their proper context according to their concentration
regime and interaction strength. Some different spin-glass species, namely,
amorphous, insulator and semiconductor are offered and compared to the
canonical (RKKY) ones. A really new type of behaviour is depicted under
the name re-entrant spin glass.
Chapter 3 moves on into the realm of experiment. The basic spin-glass
phenomenon is divided into three temperature regimes, i.e., above, at, and
below T,, and the salient and generic experimental features are illustrated.
Here we mainly use the existing measurements on the canonical spin glasses
CuMn and AuFe as specific examples, and data out of the original literature
are separately presented for the susceptibility, specific heat, resistivity,
neutron scattering @R, Mlissbauer effect, torque and ESR. With this vast
collection of behaviours, an experimentalist’s model or picture is constructed
to compact the results. Thus, we can talk nicely about the generic spin
glasses and interrelate the measurements within the framework of the cluster
model. It requires a rather lengthy chapter even to summarize the large
amount of experimental data which has been accumulated. But such is
necessary to acquaint the reader with the basics of the real, oft-studied
systems. And an abundance of plots and sketches are compiled and shown
for reference and posterity.
Before beginning with the weighty subject of theory, Chapter 4 serves as
a pause that allows us to list in table form, and rather completely, the
various combinations of spin glasses. Here we sail through the noble-metal
magnetic alloys, onto the rare-earth combinations and into the amorphous
metallic spin glasses. The going gets a bit rough when we discuss
semiconducting and insulating spin glasses. Nevertheless, we finally arrive
at our concept of what a good spin glass is.
Chapter 5 is perhaps the most difficult, yet the one with the most
overview references. So if you wish to really master spin-glass theory, skim
through Chapter 5 which will give you the flavour and historical thrust, and
then turn to one of the many theoretical reviews. Our survey has been to
look back at the precursors, i.e., the first attempts to confront this then
unnamed phenomenon of the spin glasses; then move into the E-A model,
follow with the S-K solution and subsequently its instability. Contact with
experiments is frequently made wherever possible. The TAP approach is
considered, and afterwards, a simplified description of RSB and its physical
248 The end (for now)

meaning are given via a few highly helpful sketches. Hopefully the above
acronyms are still familiar to the reader. The all-important dynamics enters
our discussion as we wind up with the mean-field model. It is then further
onwards into the more recent theoretical issues of the droplet and fractal-
cluster models. Finally, we switch from Ising (as with the preceding topics)
to non-Ising vector spin glasses and conclude Chapter 5 with the computer
simulations to which we returned briefly in Chapter 8.
In order to confront the experiment of Chapter 3 with the theory of
Chapter 5, we require ideal spin glasses. In Chapter 6 we first explain what
is an ideal spin glass, and then consider a 2D example and a 3D less ideal
one for detailed comparisons. A demanding (perhaps overly so) examination
is carried out to evaluate the correspondence of experimental results to
theoretical predictions. The emphasis here lies with the nature of the phase
transition and its critical properties. Chapter 6 closes with the conclusions
and some future possibilities.
At long last in Chapter 7 we turn our attention to non-spin analogues.
This area has become a vast out-growth of the original magnetic alloys.
Yet the ideas and notions are very similar only without the word ‘spin’. A
few examples are electric dipolar and quadrupolar glasses, and supercon-
ducting glass phases. For the sake of comparison, we then briefly revert
back to a spin system, the random-field Ising model, whose phase transition
is well understood. Finally and succinctly, we introduce three non-
physical analogues, namely, combinatorial optimization, neural networks
and biological evolution. Books can and are being written about these
topical subjects. However, we wish to show the reader that they owe a
great deal to the spin glasses for their recent development. For they all
stem from the basic spin-glass hamiltonian and possess the same jagged
phase space.
In this final chapter, which for obvious reasons is called ‘The end (for
now)‘, we have attempted to place the spin glasses within the notions of
symmetry breaking and phase space. The unique feature here is the multi-
valley landscape of the spin-glass phase space and it is the key to
understanding much of the unusual behaviour. In order to prove that there
is still life and new physics in the spin glasses we discuss the nascent subject
of mesoscopic spin glasses to which significant experimental interest is
turning. Also from the theory side we related how this reduced geometry
(or nanostructure) should influence the freezing temperature. Since the
theoretical efforts are continuing to generate progress, we have selected
some very recent and undigested Monte-Carlo numerical simulations and
tried to outline their provocative conclusions. And last but not least, here
we are attempting to summarize the book’s composition by giving the
reader some tips on where what is and how to find it in a ‘final word’.
REFERENCES 249

References

Bass, J. and Cowen, J. A. (1992) Recent Progress in Random Magnets (edited by

D. H. Ryan) World Scientific, Singapore. See also Hoines, L., Stubi, R.,
Loloee, R., Cowen, J. A. and Bass, J. (1991) Phys. Rev. Lett., 66, 1224.
Bhatt, R. N. and Young, A. P. (1992) Phys. Rev. Lett., 69, 3130.
Campbell, I. A. (1992) Phys. Rev. Lett., 68, 3351; 69, 3131.
Feng, S., Bray, A. J., Lee, P. A. and Moore, M. A. (1987) Phys. Rev., B36, 5624.
Fisher, D. S. and Huse, D. A. (1987) Phys. Rev., B36, 8937; (1988) ibid, 38, 373.
Israeloff, N. E., Weissman, M. B., Nieuwenhuys, G. J. and Kosiorowska, J. (1989)
Phys. Rev. Lett., 63, 794.
Kawamura, H. (1992) Phys. Rev. Lett., 68, 3785.
Matsubara, T., and Iguchi, M. (1992) Phys. Rev. Len., 68, 3781.
de Vegvar, P. G., Levy, L. P. and Fulton, T. A. (1991) Phys. Rev. Lett., 66, 2380.
Index

Abrikosov phase 213, 214-15 AuMn 119

absorption 68-9, 72
ac-susceptibility Bethe ansatz treatment 27
Edwards-Anderson model 142-3 biological evolution 230-l
freezing-temperature experiments blocking process 42, 43
64-76 bond percolation 34
ideal spin glasses 188-9, 193 bond randomness 5-6
random-field Ising model 221 Bose-Einstein thermal factor 58
Sherrington-Kirkpatrick model 146-8 Bragg peaks 16, 57, 60
ageing time 93-7, 167-8, 169 Brillouin function 41
AgMn 76, 119, 243 Brillouin zone 58, 61-3
amorphous magnets 38-9, 126-8, 134
Anderson model see virtual-bound-state canonical spin glassessee RKKY spin
model glasses
anisotropy, magnetic 11-13 Casimir-du PrC equations 68, 72, 73-4
experimentalist’s model 114-15 categories of spin glasses see systems of
low-temperature experiments 90-2, spin glasses
105-8 GfMn 120
metallic glasses38-9 Ce 125
re-entrant spin glasses39 chalcogenides 128-9
annealing, stimulated 225-6, 245 chemical clustering 62-3
chiral ordering 245-6
Argand diagrams 73
Arrhenius law cluster glass 17, 32-3, 37, 39
clusters
quadrupolar glasses 208 experimentalist’s model 113-15
superparamagnetism 41-2 fractal cluster model 167-70
susceptibility 70 magnetization 90-l
associative memory 229-30 Miissbauer effect 88
AuCo 120 neutron scattering 61-3, 103-4
AuCr 119 resistivity 56
AuFe specific heat 53
characteristic temperatures 122-5 susceptibility 47-50, 71
dimensional crossover 244 symmetry 234-6
magnetization 90-l Co 10-11, 184-5
mesoscopic spin glasses 239 cobalt-aluminosilicate glass 133
Miissbauer effect 86-7 coherent scattering 57
neutron scattering 61, 62, 102 Cole-Cole approach 189-91
re-entrant spin glasses 39, 40 collinear spin glass 108, 109
specific heat 98 combinatorial optimization 223-6
susceptibility 64 computer simulations 174-9, 244
252 Index

concentration 35-7, 56, 136-7 random-field Ising model 221-2

cooling see field cooling; zero-field susceptibility 71
cooling Dzyaloshinskii-Moriya interaction
Cooper pairs 212 12-13, 105, 173
Coulomb repulsion 22-3 DM interaction 12-13, 105, 173
covalent mixing 24
Cr 121 Edwards-Anderson model 139-43
CrFe 121 ideal spin glasses 192-3, 199, 201
critical line 170 Edwards-Anderson order parameter 30,
crystalline electric field 21 140, 236-7
Cu 10-11, 184-5 computer simulations 175
CuFe 119-20 neutron scattering 58
CuMn superconducting glass phases 213
characteristic temperatures 122-5 elastic scattering 57, 60-2, 80
dimensional crossover 243 electric dipolar glasses205-8
electron spin resonance 111 electric quadrupolar glasses208-11
magnetization 89-90, 91-2, 94-6 electron spin resonance (ESR) 108-11
mesoscopic spin glasses239-41, 242 energy
Monto Carlo simulation 174 flux 100
neutron scattering 58-9, 62-3, 80-1, see also free energy
83, 1024 entropy 52, 78, 98
resistivity 54-5, 78-9, 100-l EuSrO 128
specific heat 51, 53, 54, 77, 98 (EuSr)S 128-9
susceptibility 46-50, 64-7, 69, 71 neutron scattering 102
torque 107 specific heat 53, 98
CuNi 121 susceptibility 68, 72-3
Curie law 20-1, 185-6 evolution, biological 23&l
Curie-Weiss behaviour 47, 48, 65, 130 experimentalist’s model 113-16
exponential logarithmic decay 191-2
dc-susceptibility 46-50
deviation temperature 47, 48 fast scattering see quasi-elastic
differential susceptibility see ac- scattering
susceptibility Fe,Mg,-,Clz 131
diffuse scattering 57, 60 Fe,Mn,-,TiOs 195-200
dilute magnetic semiconductors 129-30 Fe,Zni-,F1 131-2
dimensional crossover 241-4 field cooling
dipolar glasses 205-8 magnetization 91-2, 94, 96
dipolar interaction 10, 12, 105 random-field Ising model 217-19
direct exchange 6-7 semiconducting spin glasses 130
dispersion 67 specific heat 100
diversity, natural 231 spin glassesin a field 111-12
domain-wall renormalization-group susceptibility 69-70
techniques 176 torque 105-9
driving fields 64 field, spin glasses in a 111-13
droplet model 164-7 fluctuation-dissipation theorem 191
computer simulations 179 fracal cluster model 86, 167-70
dimensional crossover 242-4 free energy
history 18 droplet model 164-6
ideal spin glasses 190-1, 193 Edwards-Anderson model 141-2
mesoscopic spin glasses239 electron spin resonance 108
dynamic correlation function 1767 historical developments 137-8
dynamical scaling replica-symmetry breaking scheme
ideal spin glasses 192 159
mean-field model 161-3 Sherrington-Kirkpatrick model 147
‘ndex 253

ieezing process 15-16, 31, 32-3 resistivity 55-6

reezing-temperature experiments 63-88 Kondo temperature 25-8, 119-20
requency dependence 66-7 Korringa value 58, 74
kozen matrix 169
iustration 2-3, 13-15 Larmor-precession phase 82
insulators 132 Little dynamics 229
percolation 35 localized spin-fluctuation 25, 27
semiconducting spin glasses 130 long-range magnetic order 34-5
long-range spin density wave 104
3abay-Toulouse (GT) line 171-2, 173 lorentzian distribution 58-9, 84
Iauge glass 216 low-temperature experiments 88111
Iiant moments 28-30, 120-l
Ilassy phase 213-14 magnetic anisotropy see anisotropy,
3lauber dynamics 174 magnetic
Iood spin glasses 119, 133-4, 181 magnetic clustering 63
magnetic insulators 42-3, 131-3, 134
3amming distance 228-9 magnetic interactions 6-11
Ieisenberg model 11, 138-9, 2446 magnetism and superconductivity 126,
lolnium-borate glass 132-3 127
righ-temperature experiments 45-63 magnetization
listory of spin glasses 16-18, 136-9 isothermal remanent 93, 96
Hopfield model 227-9 low-temperature experiments 89-97
lybridization 22 replica-symmetry breaking scheme
lydrodynamic theory 108, 111 157
rydrogen, molecular 208-l 1 semiconducting spin glasses 130
lyperfine field 86-7 thermal-remanent 91-3, 96
iysteresis 91-2 magneto-conductance experiments
239-41
deal glass temperature 70 mean-field model 161-3
deal spin glasses 181-203 mean-field theory 144, 146, 202
nsulators, magnetic 42-3, 131-3, 134 Meissner phase 213, 214
[sing spin glass 10 memory, associative 229-30
computer simulations 176, 178 mesoscopic spin glasses 237-41
droplet model 164-5 metallic glasses37-9, 126-8, 134
electron spin resonance 109 Metropolis acceptance criterion 174,
historical developments 138 225
insulating 131 mictomagnetism 17, 32-3, 37, 39, 50
Sherrington-Kirkpatrick model 144 mixed superexchange 10-l 1
torque 108 mixing 22
.sothermal remanent magnetization covalent 24
(IRM) 93, 96 molecular hydrogen 208-11
.sotropic anisotropy 107 moments
Isotropic spin glasses 171 giant 28-30, 120-l
Iterative improvement technique 224-5 weak 25-8
monochalcogenides, rare-earth 128
Iosephson tunnelling 212-13 Monte Carlo method 174-9, 198-200,
201
[KCN),(KBr),-, 208-9 Mossbauer effect 86-8
Ki -XLXTa03 206 muon spin relaxation @SR) 83-6
Ki -,Na,Ta03 206-7
Kondo effect 24-8 neural networks 227-30
concentration regimes 35 neutron scattering
history 17 freezing-temperature experiments
rare-earth spin glasses 125 8&3
 254 Index

high-temperature experiments 57-63 quenched-disorder randomness 6, 140-l

low-temperature experiments 101-4
random-field Ising model 222 random-exchange Ising model 43, 131,
neutron-spin-echo 82-3, 86 218-19
Ni 121 random-exchange problem 43
noble metals 119-20 random-field Ising model 43, 131,
noise-spectrum measurements 239-40 216-22
nomenclature 17 random magnetic fields 138-9
non-Isling spin glasses 1714 randomness 4-6
non-linear susceptibility 75-6 rare-earth monochalcogenides 128
ideal spin glasses 185-8, 193, 195-6, rare-earth spin glasses 125-6, 134
198 metallic glasses 38, 127
non-scaling 37 virtual-bound-state model 20-l
NP-complete problems 223, 224, 226 Rb,Co,Mgi mXF4219
numerical simulations 244-6 Rb2Cu1-$oXF4 184-94
Rb,-,(NH.,),H,PO, 207
Onsager reaction field 138 re-entrant spin glasses39-40
order parameter 213, 214 re-entry superconductivity 126
see also Edwards-Anderson order reaction term 149
parameter relaxtion times
ortho hydrogen 208-11 computer simulation 177
Overhauser theory 103 droplet model 166-7, 168
experimentalist’s model 113-15
para hydrogen 208-11 ideal spin glasses 189-91
paraelectretic phase 207 insulators 132-3
paramagnetic model 46-7, 74-5 magnetization 93, 95-7
Parisi replica-breaking symmetry Mossbauer effect 88
scheme 151-61, 178-9 muon spin relaxation 84, 86
Pd 121 neutron scattering 58-61, 80
PdFe 29 resistivity 79
Pdi-,-,Fe,Mn, 147 susceptibility 68-9, 72-4, 76
P&In 121 remanent magnetization
PdNi 121 computer simulations 175
percolation 34-5 isothermal 93, 96
experimentalist’s model 114-15 thermal 91-3, 96
fractal cluster model 169-70 replica-symmetric solution 145, 147
limit 34, 37 replica-symmetry breaking (RSB)
semiconducting spin glasses 128 scheme 150-3
phase space 233-7 computer simulations 178-9
phase transition 30-2 history 18
ideal spin glasses 188-9, 194, 200 mean-field model dynamics 163
Mossbauer effect 87 non-Ising spin glasses 171-3
muon spin relaxation 86 physical meaning of 153-61
Overhauser theory 103 replica trick
random-field Ising model 220 avoidance of 149
susceptibility 76 Edwards-Anderson model 141
power-law divergence 71, 76 Sherrington-Kirkpatrick model 144
pressure 56 superconducting glass phases 213
projection hypothesis 156-8 resistivity
Pt 121 amorphous spin glasses 127-8
freezing-temperature experiments
quadrupolar glasses208-11 78-80
quasi-bound state 26, 27 high-temperature experiments 53-6
quasi-elastic scattering 61-2, 80 low-temperature experiments 100-l
Index 255

RKKY spin glasses7-9, 119-20 semiconducting spin glasses 130

amorphous spin glasses 128 spin density wave 102-4
giant moments 29-30 spin-disorder scattering 79
good spin glasses 134 spin glasses
history 17, 136-9 characteristics 30-2
ideal spin glasses 181 definition l-4
Kondo effect 25, 28 spin-orbit coupling 21
metallic glasses 38 spin scattering 238-9
phase transition 31-2 SQUID 69, 93
rare-earth combinations 125 stability, natural 231
resistivity 55-6 stimulated annealing 225-6, 245
specific heat 51-2, 53 Stoner enhancement factor 121
susceptibility 47, 49-50, 66, 74 stretched exponential 96-7, 170, 192
transition-metal solutes 120, 121, superconducting glass phases 212-16
122-3 superconductivity and magnetism 126,
rock magnets 132 127
Ruderman, Kittel, Kasuya, Yosida see superexchange 9-10
RKKY spin glasses good spin glasses 134
insulators 132
satellites 102 magnetic insulators 43
saturation 8%90, 93 mixed 10-11
scaling 32 semiconducting spin glasses 128-9
computer simulations 176 superparamagnetism 35-7, 4&2, 66
concentration regimes 35-7 superparamagnets 132-3
fractal cluster model 167-9 susceptibility
ideal spin glasses 187-8, 194-8, 200 computer simulations 175
RKKY interaction 136-7 droplet model 165
see also dynamical scaling Edwards-Anderson model 142-3
Schrieffer-Wolff (S-W) canonical freezing-temperature experiments
transformation 23-4 64-76
self-averaging 1661 high-temperature experiments 46-50
semiconductors 42-3, 128-31, 134 ideal spin glasses 185-9, 193
Sherrington-Kirkpatrick (SK) model low-temperature experiments 88, 93
144-7 mean-field model 162-3
combinatorial optimization 224 numerical simulations 245-6
computer simulations 178-9 quadrupolar glasses 211
instability 147-8 random-field Ising model 221
mean-field model 161-3 replica-symmetry breaking scheme
replica-symmetry breaking 154-7 157, 160-l
TAP approach 150 semiconducting spin glasses 130
short-range spin density wave 103-4 Sherrington-Kirkpatrick model 146-8
single-ion anisotropy 12 symmetry breaking 233-7
site percolation 34 systems of spin glasses118, 133-t
site randomness 4-5, 244-5 amorphous (metallic) spin glasses
solubility limit 119-20, 125 126-8
Sompolinsky model 162-3 insulating spin glasses 131-3
spatial correlations 57, 79 rare-earth combinations 125-6
specific heat semiconducting spin glasses 128-31
Edwards-Anderson model 142-3 transition-metal solutes 119-25
freezing-temperature experiments
76-8
high-temperature experiments 5&3, T-quenching 94
54 TAP approach 148-50
low-temperature experiments 98-100 temporal correlations 57, 79, 82
256 Index

thermal-remanent magnetization (TM) virial expansion 137-8

91-3, 96 virtual-bound-state model 17, 20-4
Thi-,Nd,Rhz 126, 127 Vogel-Fulcher law 70-1, 222
Thouless, Anderson and Palmer (TAP) vortex glass 214, 215, 216
approach 148-50
torque 105-9 waiting time 93-7, 166-7, 168
transition metal solutes 21, 37, 119-25 weak moments 25-8
transition-metal/transition-metal spin
glasses 120-2 YGd 103-4
travelling salesman problem 223-6
triad, anisotropy 106, 107-8
two-level tunnelling model 99-100, 139 zero-field cooling
electron spin resonance 109-10
magnetization 89, 93, 94-6
ultrametricity 152-4, 158-9 random-field Ising model 217-19
uniaxial anisotropy 13, 39, 92, 172-3 semiconducting spin glasses 130
unidirectional anisotropy 13, 92 specific heat 100
universal conductance fluctuations susceptibility 69-70
238-9 ZnMn 120