Ruben Serral Gracia - Spin Bosons and Spin Glasses

Spin Bosons and Spin Glasses
ACADEMISCH PROEFSCHRIFT
ter verkrijging van de graad van doctor
aan de Universiteit van Amsterdam
op gezag van de Rector Magnicus
prof. mr. P.F. van der Heijden
ten overstaan van een door het college voor promoties ingestelde
commissie in het openbaar te verdedigen in de Aula der Universiteit
op donderdag 25 november 2004, te 14:00 uur.
door
Rubèn Serral Gracià
geboren te Xerta, Spanje
Promotor: prof. dr. B. Nienhuis
Copromotor: dr. Th. M. Nieuwenhuizen
Promotiecommissie: prof. dr. F. A. Bais
prof. R. Balian
dr. J.-S. Caux
prof. dr. L. F. Cugliandolo
dr. L. Leuzzi
dr. L. G. Suttorp
prof. dr. C. J. M. Schoutens
Faculteit der Natuurwetenschappen, Wiskunde en Informatica
This thesis is part of the research programme of the Sticht-
ing voor Fundamenteel Onderzoek der Materie (FOM), which
is nancially supported by the Nederlandse Organisatie voor
Wetenschappelijk Onderzoek (NWO).
Paranimfen:
Jordi Busquets Blanco
Daniele Moroni
This thesis is based on the following papers:
R. Serral Gracià and Th. M. Nieuwenhuizen
Quantum spherical spin models
Phys. Rev. E 69 056119 (2004) [cond-mat/0304150]
R. Serral Gracià, Th. M. Nieuwenhuizen and I. V. Lerner
Concentration dependence of the transition temperature in metallic spin glasses
Eur. Phys. Lett. 66 (3) 419-422 (2004) [cond-mat/0311491]
A. E. Allahverdyan, R. Serral Gracià and Th. M. Nieuwenhuizen
Work extraction and cooling in the spin boson model
In preparation
A. E. Allahverdyan, R. Serral Gracià and Th. M. Nieuwenhuizen
Cooling spins via bosonic bath
In preparation
R. Serral Gracià, L. Leuzzi and F. Ritort
Numerical simulation on the hypercubic cell spin glass
In preparation
A la meua famlia,
Contents
Introduction 11
I Spherical Model 15
1 Phase transitions and critical phenomena 17
1.1 Path Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.1.1 Bosonic coherent state representation for a single oscillator . 22
2 Quantum spherical model 25
2.1 Classical spherical model . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.1.1 Vector spherical spins . . . . . . . . . . . . . . . . . . . . . . 27
2.1.2 Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.1.3 Spherical constraint on the length of the total spin . . . . . . 28
2.1.4 Spherical constraint on the number of spin quanta . . . . . . 28
2.1.5 Comparison of the two constraints . . . . . . . . . . . . . . . 29
2.2 Ferromagnetic Hamiltonians with creation and annihilation operators 30
2.2.1 Coherent state representation for spherical spins . . . . . . . 30
2.2.2 General solution . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.2.3 Ferromagnetic couplings with transversal eld in d dimensions 33
2.3 Hamiltonians involving spins but not their momenta . . . . . . . . . 40
2.3.1 Ferromagnetic couplings in the presence of a transversal eld 42
2.4 Generalization and mapping from Heisenberg spins . . . . . . . . . . 46
2.5 Summary and discussion . . . . . . . . . . . . . . . . . . . . . . . . . 49
II Spin-Boson Model 53
3 Introduction 55
3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.1.1 Work Extraction . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.1.2 Cooling of spins . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.2 Open quantum systems and spin boson model . . . . . . . . . . . . . 57
8 Contents
4 Solvable model 61
4.1 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.1.1 Heisenberg equations and their exact solution . . . . . . . . . 63
4.1.2 Factorized initial conditions . . . . . . . . . . . . . . . . . . . 64
4.1.3 Correlated initial conditions . . . . . . . . . . . . . . . . . . . 66
4.2 Pulsed Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2.1 Parametrization of pulses . . . . . . . . . . . . . . . . . . . . 70
4.3 Setup of pulsing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.4 Realizations of the model . . . . . . . . . . . . . . . . . . . . . . . . 71
4.4.1 Why do we insist on the exact solvability of the model . . . . 72
4.5 The spectral density . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.5.1 Ohmic spectrum of the bath . . . . . . . . . . . . . . . . . . 73
4.5.2 Other types of environment . . . . . . . . . . . . . . . . . . . 74
4.A Correlated versus factorized initial conditions . . . . . . . . . . . . . 76
4.B Quantum noise generated by ohmic bath . . . . . . . . . . . . . . . 78
5 Work extraction 81
5.1 Denition of work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.2 General restrictions on work extraction . . . . . . . . . . . . . . . . . 82
5.2.1 Eciency of work-extraction . . . . . . . . . . . . . . . . . . 84
5.3 Work-extraction via two pulses . . . . . . . . . . . . . . . . . . . . . 85
5.3.1 Formulas for work . . . . . . . . . . . . . . . . . . . . . . . . 85
5.3.2 Work extraction for T > T
S
. . . . . . . . . . . . . . . . . . . 87
5.3.3 Work extraction for T < T
S
. . . . . . . . . . . . . . . . . . . 89
5.3.4 Eciency of work extraction . . . . . . . . . . . . . . . . . . 90
5.4 Work-extraction via spin-echo pulses . . . . . . . . . . . . . . . . . . 91
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.5.1 Lasing without inversion . . . . . . . . . . . . . . . . . . . . . 95
5.5.2 Quantum heat engine . . . . . . . . . . . . . . . . . . . . . . 96
5.A Derivation of Eqs. (5.9, 5.10) . . . . . . . . . . . . . . . . . . . . . . 97
6 Cooling of spins 99
6.1 Cooling via pulses . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.2 Cooling using spin-echo pulses . . . . . . . . . . . . . . . . . . . . . . 101
6.3 1/f spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
III Spin Glasses 105
7 Introduction to spin glasses 107
7.1 Theory and mean eld models . . . . . . . . . . . . . . . . . . . . . . 109
8 Transition temperature in metallic spin glasses 113
8.1 Concentration dependence of the transition temperature . . . . . . . 113
Contents 9
9 The Ising Spin-Glass on a Hypercubic Cell 117
9.1 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
9.2 Simulation and Data Analysis . . . . . . . . . . . . . . . . . . . . . . 118
9.2.1 Exchange Monte Carlo . . . . . . . . . . . . . . . . . . . . . . 118
9.2.2 Multispin coding . . . . . . . . . . . . . . . . . . . . . . . . . 122
9.2.3 Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
9.3 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
9.3.1 Phase Transition . . . . . . . . . . . . . . . . . . . . . . . . . 129
9.3.2 Low temperature behaviour . . . . . . . . . . . . . . . . . . . 131
9.3.3 Evidence for Ultrametricity . . . . . . . . . . . . . . . . . . . 135
9.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Bibliography 137
Samenvatting 147
Acknowledgements 151
Introduction
The last centuries showed the birth of, among others, statistical mechanics and
quantum mechanics; theories which are the basic building bricks of todays physics.
By now we are familiar with concepts like phase transitions, critical phenomena, or
quantum coherence; and with quantities like the partition function or the density
matrix. It is during the last two centuries that physicists realized that a complete
description of reality is intractable from a mathematical point of view. Not only
this, they realized that it was unnecessary, as is the case of statistical mechanics, or
that it was impossible, as is the case for quantum mechanics. Indeed, the answer
to problems in physics changed from a given deterministic solution to a solution in
terms of averages, or a set of probabilities for each of the possible events. Commonly,
full solutions, even in terms of averages, were not possible since the mathematics
at disposal was not sucient to tackle the forthcoming problems. Approximations
were called for.
Presently theoretical physics is mainly based on approximations. We can only
nd approximate solutions, or at most, solutions valid in a restricted range of the
parameters involved in the problem. In spite of that, these solutions can give us
most of the needed information and understanding. Another successful approach is
to use simplied models which, however, keep the key features of more complicated
and more realistic ones. From these toy models, we can infer the behavior of
realistic models, we can test general ideas like, for example, critical phenomena and
universality. This approach is, in fact, followed in the rst part of this thesis.
This theoretical development was followed, or even sometimes pushed, by ex-
perimental development. For instance, the range of temperatures of contemporary
experiments is much larger than only a few decades ago. Quantum eects are at
the moment seen in many laboratories in the world. Very low temperatures, ap-
proaching T = 0, are now accessible and controllable. Understanding the physics
at such low temperatures is important and new problems and worries arose from
that. New phases of matter were discovered, such as Bose-Einstein condensates.
New phase transitions were found, which brought a better understanding of critical
phenomena. In addition, a new type of transitions was proposed, quantum phase
transitions. Because they occur strictly at T = 0, they are rarely observed di-
rectly, but their presence can be inferred indirectly since many eects in very low
temperature physics arise from lying in the vicinity of a quantum critical point.
One of the most interesting features discovered in the last century, was the exis-
12 Introduction
tence of spins. We describe spins in many dierent ways depending on the purpose
and the situation. For instance, spins are responsible for magnetic eects. Many ap-
proximations were needed to understand their behavior. Often, in a magnet, details
on the dynamics of each of the spins are not of leading importance. Furthermore,
the crystalline structure can break the symmetries at disposal, leading then, in the
case of spins 1/2, to Ising or XY spins. The interactions between each pair of spins
throughout the lattice can sometimes be treated as classical, though their origin
is purely quantum mechanical. In spite of all these approximations, the problems
typically remain unsolvable. To overcome that, several approximations and gener-
alizations in the algebra describing spins have been performed, during many years
of study. A very successful generalization is the so-called spherical spin approxima-
tion, to which the rst part of this thesis is devoted. A quantum spherical model
is presented and solved there in order to study quantum phase transitions. This is
one of the simplest models showing non-trivial critical phenomena. We will show
how spherical spins can arise as a limit of Heisenberg spins and that leads to an
analytically tractable problem. With this model we can analytically study critical
phenomena of a quantum phase transition, i.e., we can calculate exactly the in-
volved critical exponents, the critical amplitudes and the relation with the classical
counterpart.
As technology improved, new ways to study spins appeared. Nowadays, for
instance, it is common to deal with SQUIDs (superconducting quantum interference
device) which can be understood as a two level system, thus analogous to a spin
1/2. In this situation, with only one SQUID, no spin-spin interaction is present but
the dynamics play an important role. Furthermore, these types of systems cannot
be isolated from the external world around them, therefore the interaction with the
environment must be taken into account. Thus we jump from a static approach,
i.e. in equilibrium, like the one we often have in magnetism, towards a dynamical
one. Many other systems can be approached in this manner. In nuclear magnetic
resonance (NMR), one deals with an ensemble of spins that can, sometimes, be
considered as non-interacting. There, by using very strong magnetic pulses, it is
possible to tune the dynamics of the spin up to the point to be able to perform
controlled rotations on the Bloch sphere. Experimentally the dynamics is analyzed
in free induction decays (FID). The environment also plays an important role here.
The dynamics of the system is biased by the surroundings of the sample, bringing
decoherence and damping. This is the subject of part II (chapters 3 to 6) of this
thesis. In this part, an exactly solvable limit of the spin-boson model is presented
and solved. The model is a description of the dynamics of an ensemble of spins
on which NMR pulses are applied. As in the previous part, the approximations
are performed when posing the problem, and not half way as is most frequently
done. This reduces its applicability but in exchange, it leads to an analytically
solvable model. The rst objective of this part is to use this model to study new
mechanisms of work extraction, for instance lasing, in such microscopic systems.
These techniques are expected to have a number of advantages over the standard
ones. The second objective is to cool spins, i.e. to obtain purer spin states, since
the eciency of, e.g., NMR studies depends strongly on the initial polarization of
13
the spins. Enhancing that quantity enhances the output signal in the experiment.
The problem of cooling is well-known. However, some important objectives are not
yet met with the existing methods. This led us to propose a new method of cooling,
where the bath is involved in an essential manner.
Advances in the elds of chemistry and physics of materials allowed precision cre-
ating new materials. For instance, creating alloys of dierent metals at a given pre-
cise concentration of their constituents. Examples of that are for instance Au
1c
Fe
c
,
a transition metal embedded in a noble metal, which for certain concentrations c is
a spin glass. The discovery of spin glasses boosted an at that time young branch of
statistical mechanics, the study of disordered systems. Indeed, spin glasses can be
understood as disordered magnets which at low temperatures condense in a frozen
disordered conguration instead of a periodic one.
Part III (chapters 7 to 9) is focused on spin glasses. In this part our aim is more
modest than to start from a simplied Hamiltonian and evolve analytically till the
end. In chapter 8 departing from the fact that at nite temperature the range of
the RKKY interaction is cut-o, we give a description of the dependence of the
transition temperature on the concentration of magnetic impurities in metallic spin
glasses.
In chapter 9, nally, another way to tackle intractable problems is considered,
namely computer simulations. With the help of computers, simulations of unsolvable
models can be carried out and important results can be found. This brings a whole
new dimension to physics, since the problems are no longer the unsolvability but, for
instance, the too small size of the simulated systems that has to be chosen, because a
computer has nite memory and limited speed. In chapter 9 a numerical simulation
of a spin glass on a hypercubic cell is presented. With this simulation we study how
this nite size and nite connectivity model approaches the well known innite size,
innitely connected mean eld model of the spin glasses.
Part I
Spherical Model
1
Phase transitions and critical
phenomena
In this chapter we introduce fundamental concepts of phase transitions and criti-
cal phenomena. There are many good text books on critical phenomena where the
reader can nd an extensive treatment of the problem. In order to write this chap-
ter the author has mainly beneted from Refs. [Ma82, Bel91, Sac99, NO98, ID89,
BGZJ76] to which the reader is refered for further reading.
When describing a piece of material, we are mainly interested in quantities such
as, the energy, the mass, the total magnetic moment, or the corresponding intensive
counterparts, the energy density, mass density, magnetization. These are refered to
as mechanical variables [Ma82]. In most cases they are properly dened once we
characterize the environment of the material of interest, say, we know the external
elds applied to it, the temperature, magnetic eld, pressure, etc. These external
elds dene the parameter space, and the response or state of the system in each of
the points of this parameter space dene the phase diagram. At some points or lines
of the dierent external eld variables, the system undergoes a phase transition. At
this stage, one of the mechanical variables, or a combination of them, will dene an
order parameter. The order parameter is a variable which is zero in one phase and
non-zero in the other. At the transition it starts to grow. Most studies in critical
phenomena are devoted to nd the order parameter or to understand how it evolves.
Examples of order parameters are for a ferromagnet: the magnetization, for a liquid
gas transition: the relative density with respect to the one at the critical point, for
superconductors: the complex eld density amplitude of Cooper pair bosons.
Since in this thesis we study the critical phenomena of magnetic systems, we
proceed using them as the example for dening the critical exponents.
The dierent critical exponents are dened as:
: Specic heat divergence [BGZJ76]:
18 I.1. Phase transitions and critical phenomena
At zero eld h = 0, the specic heat diverges at T
c
as a power law
C =
_
A(T T
c
)
+B for T > T
c
,
A
t
(T
c
T)
+B
t
for T < T
c
.
(1.1)
: Order parameter m as a function of temperature T:
For zero external magnetic eld, h = 0, the magnetization is a decreasing
function of T and vanishes at T
c
m (T
c
T)
for T < T
c
. (1.2)
: Magnetic susceptibility divergence:
The magnetic eld susceptibility =
m
h
T
diverges at h = 0 at the transition
temperature
=
_
C (T T
c
)
for T > T
c
,
C
t
(T
c
T)
for T < T
c
.
(1.3)
: Order parameter m as a function of the external magnetic eld h:
At the transition temperature, T
c
, the magnetic eld follows a power law with
the external magnetic eld, vanishing for vanishing eld
m [h[
1/
for T = T
c
. (1.4)
: Spatial dependence of the correlation function G
ij
Near the transition temperature T
c
the correlation function
G
ij
= S
i
S
j
) S
i
)S
j
), (1.5)
can be expressed for long enough distances r
ij
= r
i
r
j
G(r) =
g(r/)
r
D+2
,

G(q) q
2+
. (1.6)
where

G is the correlation function in Fourier space, D corresponds to the
dimension of the system, g is a function that at r decays exponentially
g(r/) exp(r/), and is the correlation length.
: Divergence of the correlation length
At the transition temperature T = T
c
and for vanishing eld h = 0, the
correlation length diverges
=
_
D(T T
c
)
for T > T
c
,
D
t
(T
c
T)
for T < T
c
.
(1.7)
19
z: Divergence of the correlation time
At the transtition temperature T = T
c
and for vanishing eld h = 0, the
correlation time diverges

z
[T T
c
[
z
; [T T
c
[
z
, (1.8)
where corresponds to the characteristic energy uctuation.
Each system has its own critical exponents. However, although being very dif-
ferent, many systems share the same critical exponents. Such systems are said to
belong to the same universality class. At the critical point and in nearby surround-
ings, the scaling limit can be taken. The scaling limit of an observable is dened
as its value when all corrections involving the ratio of small to large lengths are
neglected. In the case of magnetic spins on a lattice, the small length would be the
lattice spacing a, while the large ones would be the correlation length and the
system size. The assertion of universality is that the results of the scaling limit are
not sensitive to the precise microscopic model used. This can be seen as a formal
consequence of the physically reasonable requirement that correlations at the scale
of large should not depend upon the details of the interactions on the scale of the
lattice spacing, a. In this situation, we can map our Hamiltonian onto a continuum
eld theory that represents the universality class. Typically, only essential quali-
tative features, such as the symmetry of the order parameter, the dimensionality
of the space, and constraints placed by conservation laws, survive the continuum
limit, and the structure of the quantum eld theory is severely constrained by these
restrictions.
The critical exponents depend strongly on the

0 1/2 1 3 0 1/2
Table 1.1: The mean eld
critical exponents
spatial dimension of the system. In next chapter
we will solve anallytically the critical phenomena of
the quantum spherical model for arbitrary spatial
dimension. We will see there that below a certain
dimension the phase transition does not occur any-
more, which is a general phenomenon. The dimension where this happens is called
lower critical dimension. Qualitatively this can be understood because cooperative
eects are not strong enough to overcome the disorder created by temperature. On
the contrary, above a certain dimension, the cooperative eects are so strong that
uctuations can be ignored and the mean eld approximation turns out to be valid.
This dimension is called upper crticial dimension. The system then has the mean
eld critical exponents, see table 1.1. Between the upper and the lower critical
dimension, there are the hyperscaling relations, i.e. the critical exponents are func-
tions of the space dimension and relations among them can be found in terms of this
dimension d. Some examples can be
=

2
, (1.9)
= 2 d, (1.10)
=
1
2
(d 2 +), (1.11)
=
d + 2
d 2 +
. (1.12)
So far in these denitions we were considering classical critical points. Classi-
cal critical points occur at nite temperature. In such situations, all uctuations
have a thermal origin. No matter how quantum the system behaves, for instance the
transition to a superconductor phase, as long as the transition occurs at a nite tem-
perature, thermal uctuations are responsible for the transition and the transition
is classical. On the contrary, quantum phase transitions occur at zero temperature,
T = 0, [Sac99, Voj00a, Voj00b]. Classical systems at zero temperature usually freeze
into a uctuationless ground state. In contrast, quantum systems have uctuations
driven by the Heisenberg uncertainty principle even in the ground state. These are
responsible for the quantum phase transitions since in such situations, temperature
and the corresponding thermal uctuations cannot play a role, and the only possible
source of uctuations is quantum mechanical. A simple model that exhibits a quan-
tum phase transition is the Ising model with transversal eld. This model reproduces
the main ideas of quantum phase transitions. We will reproduce it qualitatively here
not only for pedagogical reasons, but also because it was a source of inspiration for
the work described in the following chapter.
The Ising model is one of the simplest models to study magnetism. It consists of
pairwise interactions among the Ising spins. Ising spins can only take 2 values, +1
and 1. Physically, they describe strongly anisotropic magnetic moments on which
the typical 3 directions available to the microscopic magnetic moment, only one is
available leaving as the only degree of freedom the orientation, whether positive
or negative. A physical realization of such model are the low-lying excitaitons of
the insulator LiHoF
4
which consist of uctuations of the Ho ions between two spin
states that are aligned parallel or antiparallel to a particular crystaline axis [Sac99].
Dipolar interactions among these ions at low temperatures align all spins parallel
forming thus a ferromagnet. We will simplify these interactions to only nearest
neighbours. For this study of quantum phase transitions the relevant situation is
having a transversal eld, say in the x-direction. The Hamiltonian reads
H = J
ij)

z
i

z
j
g
i

x
i
, (1.13)
where J is the coupling constant, positive for ferromagnetic couplings, i.e. when the
spins tend to aling with respect to each other, g > 0 is a coupling proportional to the
external magnetic eld, and
a
i
are the Pauli matrices (a = x, y, z) that correspond
to the spin on the lattice site i with
z
=diag(1, 1). The transversal eld induces
I.1.1 Path Integrals 21
quantum tunneling between the two states of each Ho ion, and a suciently strong
tunneling rate can eventually destroy the long range order.
In the limit of zero temperature and zero external eld, T 0, g 0, the
system nds itself in its ground state: all spins point parallel or antiparallel to the
z direction. Conversely, at zero temperature but in the limit of very large magnetic
transversal eld, T 0, g the ground state consists of having all spins pointing
in the x direction. Then by tuning the transversal eld g, one is able to go from a
phase where all spins are ordered in the z direction towards one where the spins are
pointing mostly to the x direction. In between these two limiting situations there
must be a critical value of the external eld g = g
c
where the transition takes place
since the system cannot analytically go from one situation to the other.
In this quantum critical point, all
Paramagnet
T
Tc
g
g
gc
T
Figure 1.1: Phase diagram exhibiting a
quantum phase transition at T = 0, g = g
c
denitions of critical exponents made
above must be reformulated. In this sit-
uation, the temperature plays the role
the transversal magnetic eld was play-
ing in the classical critical point. Figure
1.1 a typical phase diagram is sketched.
One can see how temperature pulls the
system away from the critical point in
the same manner the transversal mag-
netic eld g does in the classical crit-
ical point. The role played before by
temperature is now played by the con-
trol parameter of quantum uctuations,
which drive the phase transition. In this case it is the transversal magnetic eld g.
Then, in all critical exponent denitions above, temperature has to be changed by
transversal eld, T g and T
c
g
c
. On the contrary, since the longitudinal
magnetic eld h always helps the transition in the same way, its role is unchanged.
Using renormalization group arguments, Hertz [Her76] proved that the critical
exponents in the quantum critical point are related with their counterparts in the
classical critical point. The relation is set by the dynamical critical exponent z. The
quantum critical point behaves as the classical one for dimensions D = d+z. So the
upper and lower critical dimensions are shifted down in the critical point by z and
the hyperscaling relations in between these two dimensions are found by changing d
by D.
1.1 Path Integrals
In this section we give a brief introduction on how to compute the partition function
of a quantum system using Feynman path integrals [FH65]. To nd the partition
function, using one technique or another, is usually the starting point of any cal-
culation in statistical mechanics. Here, we will only focus on nding the partition
function of a system of bosons in their coherent state representation. This tech-
nique will be exploited and extended in the following chapter. The reader is refered
to e.g. [NO98] or [Sto] for a review of path integrals and coherent states and to
e.g. [KS85] for a complete study of coherent states. These books served as source
and inspiartion for the elaboration of this section.
1.1.1 Bosonic coherent state representation for a single oscil-
lator
Fock space is the Hilbert space of states labeled by the number of oscillator quanta.
Coherent states are dened as the eigenstates of the anihilator operator a. Then it
can be proven that for a system with many particles
[ ) = e
P
[ 0 ) =
)
n
!
_
[ 0 ), (1.14)
is a coherent state, where [ 0 ) is the vacuum representation in Focks space, and
stands for an index of a mode of the system. Indeed, because of the identity
a
( a
)
n
= n
( a
)
n
1
+ ( a
)
n
, it holds that a
[ ) =
[ ). The scalar
product of two coherent states gives
[
t
) = e
P
. (1.15)
A crucial property of the coherent states is that they form an overcomplete set of
states. Any vector in Fock space can then be expanded in terms of coherent states.
This is expressed by the closure relation [NO98]
_

d(
)d'(
[ ) [ = 1, (1.16)
where the measure in the integral comes from gaussian integration with complex
variables ( stands for imaginary part and ' for real part and these parts are
integrated from to ) and the exponential term is due to the fact that coherent
states are not normalized. Let us check that eq. (1.16) is indeed a representation of
the identity of Fock space. We insert it in the left hand side of eq. (1.15) and we get
[ 1 [
t
) =
_

d(
)d'(
[ ) [
t
) =
_

d(
)d'(
e
P
)
= e
P
,
(1.17)
which indeed is the right hand side of eq. (1.15).
The partition function of any quantum system Z = tr
_
e
H( a
, a)
_
can be com-
puted by the Trotter approach. The exponential has the same form as a time evo-
lution operator in imaginary time; thus it is possible to create a path integral over
I.1.1 Path Integrals 23
closed paths. The procedure is to split the exponential in a product of M equal
terms. Between each pair of them a representation of the identity, eq. (1.16), is
inserted. The partition sum then has the following shape
Z = tr
_
_
e
H( a
, a)
_
M
_
= tr
_
e
H( a
, a)
1e
H( a
, a)
1 1e
H( a
, a)
_
, (1.18)
where = /M and each 1 is the above identity operator. Each of these identities is
given an index; they represent the steps the system passes through in a discretized
path. By using the identity dened in Eq. (1.16) only the following matrix element
is needed in the calculation:

j
[ e
H( a
, a)
[
j1
). (1.19)
Provided the Hamiltonian is normal ordered, the outcome is [NO98]

j
[ e
H( a
, a)
[
j1
)
j
[ 1 H( a
, a) [
j1
) =
e
j1
(1 H(
j
,
j1
)) = e
j1
H(
j
,
j1
)
+O(
2
).
(1.20)
Correction terms can be neglected in the limit M [NO98]. Each identity
brings an integral at each time step. These integrals cover any path between its
initial and its nal state. The trace will nally tie the ends giving a closed path.
The partition function nally reads
Z =
_
()=
(0)
D(
()
())
exp
_
=0
d
_
()
d()
d
+H(
(), ( d))
_
_
,
(1.21)
where the subindex of the integral reects the trace structure of the partition function
since it gives a closed path integral; stands for the imaginary time step, so () =
i
; d is the imaginary time dierence between steps, so ( d) =
i1
; and
d()
d
=
() ( d)
d
=

i
i1
/M
. (1.22)
Despite the fact that the nomenclature used in these formulas suggests a continu-
ous time, it should always be understood as being discrete. The limit M should
always be taken at the end of the calculations, otherwise some indeterminacies may
arise. Continuous notation is used nevertheless because it is more compact.
2
Quantum spherical model
In this chapter we are going to study the quantum spherical model. This model is a
perfect ground up to understand the concepts of the previous chapter and generally
the ideas of critical phenomena, since it is an exactly solvable model that reproduces
non-trivial critical phenomena. The fact of being quantum mechanical allows us to
study quantum phase transitions analytically. A denite expression is found for the
dierent critical exponents in terms of the dimensionality of the system.
Originally the spherical model, in its classical version, was introduced by Kac.
After being introduced in 1947 to Onsagers rather intricate solution of the 2d-Ising
model, he desired to formulate a simpler spin model. As a rst step he took the
spins to be continuous Gaussian variables, nowadays called the Gaussian model.
This had unphysical behavior at low temperatures which led Kac to consider the
spherical model. The spherical model has continuous spins that are restricted by the
spherical constraint

N
i=1
S
2
i
= N, which represents the hypersphere intersecting all
vertices of the hypercube sustained by the Ising spins, S
i
= 1. In the end, the
spherical model is formally the same as the Ising model, i.e. the Hamiltonian is the
same, with a global constraint instead of a local one: the sum of the square of the
length of the spins is constrained rather than each of them. At that time the saddle
point method, needed in the solution, was not widely known, and here Berlin came
in, leading the celebrated joint publication on the spherical model in 1952 [BK52].
Kacs personal reminiscence of this history is presented in Ref. [Kac64].
The spherical model for a ferromagnet has been considered in great detail in
literature. Actually, the paramagnetic to ferromagnetic transition is similar to an
ideal Bose-Einstein condensation. As already pointed out, the critical behavior can
be solved exactly. Critical exponents and scaling functions can be derived. In par-
ticular, the model with short range interactions exhibits d
lc
= 2 as the lower critical
dimension; for d 2 no stable ferromagnetic phase occurs. Likewise, d
uc
= 4 is the
upper critical dimension; for d > 4 critical exponents take their mean-eld values,
see previous chapter. These analytic results have been used to test approximations
and general ideas of phase transitions for a wide range of interactions, short and
26 I.2. Quantum spherical model
long range. For a review on the classical spherical model see Ref. [Joy72].
As said, the spherical model was introduced for its mathematical simplicity.
However, Stanley [Sta68] proved that the free energy of a model of arbitrary spin
dimension , incorporating thus the Ising model (for spin dimension = 1), the
x y model (spin dimension = 2) and Heisenberg model ( = 3), approaches that
of the spherical model in the limit of innite spin dimensionality . Hence,
it gives a geometrical interpretation to the spherical model. Since various critical
properties where proven to be monotonic functions of the spin dimensionality , the
critical properties of the Heisenberg model appeared to be bounded on one side by
those of the Ising model and on the other by those of the spherical model.
The spherical model for antiferromagnets was studied by Knops. The spherical
constraint imposes < S
2
i
>= 1 for ferromagnets. However, this does not work for
antiferromagnets because of the lack of translational invariance. To recover this,
Knops added a second constraint; more generally, one constraint has to be added
for each translationally invariant set, which in the case of antiferromagnets means
each of the two sublattices. He found that the two constraints reduce to a unique
one provided the staggered external eld is zero. The fact that the spherical spins
are scalars makes it impossible to dene an order parameter that can be identied
with the spontaneous staggered magnetization. To solve that and get the proper
order parameter Knops used a vector version of the spherical model [Kno73a]. He
also generalized Stanleys arguments to non-translational interactions [Kno73b].
The spherical model has also been applied to disordered systems. Though, in
view of Knops nding, perhaps an innite number of spherical constraints should
be used, typically no analog of the staggered external eld is applied, and one may
expect that all constraints collapse into a single one. Therefore spherical spin glass
models may still give insight in the physics of the problem which would be more
dicult to study e.g. with Ising spins. In the case of pair couplings the exact solution
exhibits no breaking of replica symmetry and the replica trick need not be used
[KTJ76]. The family of p-spin spin glasses (p-spin models) [CS92] has been shown
to exhibit one step replica symmetry breaking by studying the spherical version.
For spin glasses with random pair and quartet interactions (p = 2 + p = 4,
p = 2 +4), Nieuwenhuizen showed that an exact solution exists, exposing the full
replica symmetry breaking scenario. The simplicity of spherical models thus may
give insight in dicult problems for which otherwise no exact solution is available.
For an early review on the use of the spherical model in disordered systems, see
Ref. [KKPS92].
So far the discussion has been classical. The classicality can be understood in
particular because the entropy diverges at low temperature as ln T, just as for a
classical ideal gas. Dierent quantum versions of the spherical model have been pro-
posed. In this chapter we will discuss the two main approaches and the dierences
between them. The spins have to become quantum mechanical operators and their
adjoint operators also have to be included in the formalism. The spherical constraint
can then be kept the same though in operators language, as it was started by Ober-
mair [Obe72], or, conversely, one can constrain both the total spin length as before
plus its global kinetic energy, as developed by Nieuwenhuizen [Nie95a, Nie95b]. The
I.2.1 Classical spherical model 27
fact of having the dynamics inside the spherical constraint allows one to consider
Hamiltonians without an explicit a kinetic part. Both approaches, though in dier-
ent setups, exhibit a quantum phase transition [Voj96, SN04]. Obermairs model,
belongs to a universality class, i.e. O(n) non-linear sigma model for large n, while
Nieuwenhuizens can, depending how the dynamics is considered, belong to the same
universality class or to another, i.e. SU(n) Heisenberg ferromagnet in the limit for
large n.
2.1 Classical spherical model
The spherical constraint was conceived as a relaxation of the Ising constraint. In-
deed, Ising spins, S
i
s
i,z
=
1
2
, obviously satisfy it. Adjusting the coecients
from the original version it may be written as
1
2
N
i=1
S
2
i
= N, (2.1)
with =
2
/8 having dimension (Js)
2
. So the spins are treated as real variables
constraint by Eq. (2.1). The Berlin-Kac spherical model is dened by the partition
sum
Z =
_
DS e
H
(
1
2
N
i=1
S
2
i
N) =
_
DS
_
i
i
d
2i
e
H
1
2

P
N
i=1
S
2
i
+ N
(2.2)
where
DS =
i
_

dS
i
. (2.3)
2.1.1 Vector spherical spins
For vector spins the generalization of Eq. (2.1) in the case of m spin dimensions
reads
1
2
N
i=1
m
a=1
(S
a
i
)
2
= Nm. (2.4)
It is worth mentioning that the spin dimensionality in eq. (2.4) is not related to
the approach of Stanley, who started with vector spins and ended up with scalar
spherical spins. We only introduce vector spherical spins to avoid the restriction
scalar spins have. We benet from the fact that the vector character allows to study
the behavior in a transverse eld. A similar step allowed Knops to dene a proper
order parameter for the antiferromagnetic spherical model [Kno73a].
2.1.2 Quantization
It is natural to consider the S
i
analogous to position variables of harmonic oscillators.
In quantum mechanics they become hermitian operators

S
i
with the dimension of
, Js. The conjugate momentum operator

a
i
is dimensionless and postulated to
satisfy the commutation relation
[
S
a
i
,

b
j
] = i
i,j
a,b
. (2.5)
As for harmonic oscillators, this allows to dene creation and annihilation operators
a
i
=
1
S
a
i

i
a
i
,

a
i
=
1
S
a
i
+
i
a
i
(2.6)
satisfying the commutation relation
[
a
i
,

b
j
] =
i,j
a,b
. (2.7)
2.1.3 Spherical constraint on the length of the total spin
There is some freedom to choose the spherical constraint, which amounts to describ-
ing dierent physical situations. The standard quantum constraint considered in
literature is just the quantized version of the mean of eq. (2.4),
Constraint 1 :
1
2
i,a
< (
S
a
i
)
2
>= Nm, (2.8)
where < ... > denotes the quantum expectation value. Obermair took as the quan-
tum Hamiltonian the classical H(S) with spins replaced by operators, and added
the kinetic term that one expects for physical rotors,
H(
S,

) =
1
2
g
2
i
+H(
S), (2.9)
where g
1
is the rotors moment of inertia. An eective Hamiltonian which includes
the constraint can be derived with a Lagrange multiplier. One ends up with
H
tot
=
1
2
g
2
i
+H(
S) +(t)[
1
2
i,a
(
S
a
i
)
2
Nm], (2.10)
where is the Lagrange multiplier that enforces the constraint. In equilibrium its
value is given by the equation of the spherical constraint
F
= 0.
2.1.4 Spherical constraint on the number of spin quanta
The constraint can restrict additionally to spin length also the global dynamics of
the system. This is possible by using
Constraint 2 :
i,a
<

a
i
a
i
>=
i,a
< n
a
i
>= Nm

2
. (2.11)
I.2.1 Classical spherical model 29
A constraint analogous to this one was introduced by Nieuwenhuizen [Nie95a,
Nie95b]. The kinetic energy is also constrained here since this constraint includes
the momenta as well. This can be seen by writing it in the form
Constraint 2 :
1
2
i,a
(
S
a 2
i
) +
2
a 2
i
)) = Nm( +

2
2
). (2.12)
For a Hamiltonian

H(
S,

) that may, but need not, depend explicitly on the mo-
menta, the eective spherical Hamiltonian is
H
tot
=

H(
S,

) +
1
2
i,a
_
(
S
a
i
)
2
+
2
(
a
i
)
2
_
Nm( +

2
2
). (2.13)
Now, the situation where the Hamiltonian does not depend explicitly on the
momenta (no kinetic term),

H(
S,

)

H(
S), still leads to sensible dynamics,

since the constraint already depends on the momenta. Dierent constraints describe
dierent physics. However, at high temperatures one expects the dierences to
become small.
In the remaining of this paper we will simplify the notation by taking units in
which = 1.
2.1.5 Comparison of the two constraints
The main dierence between the two constraints is obviously the presence or ab-
sence of momenta. In the second case eq. (2.11), the spherical constraint can carry
all the dynamics of the model. On the contrary, using the rst constraint eq. (2.8),
a kinetic term, with an external parameter g, has to be added to the Hamiltonian
[Obe72]. This parameter determines the strength of quantum uctuations; the clas-
sical model can be recovered for g = 0. This fact makes models with the rst
constraint describe quantum rotors, as was pointed out in Ref. [Voj96]. The rst
constraint, eq. (2.8), brings actions which are invariant under orthogonal transforma-
tions. Conversely, using the second constraint, eq. (2.11), the choice of Hamiltonian
can bring symmetry under unitary transformations or orthogonal ones depending
on the question whether the Hamiltonian contains momenta or not. Hamiltonians
with unitary transformation symmetry yield free energies analogous to the large ^
limit of the generalization of SU(2) Heisenberg spins to SU(^). Hamiltonians with
orthogonal transformation symmetry share the critical phenomena with the large
^ limit of O(^) non-linear sigma model and describe therefore quantum rotors as
occurs by using the rst constraint, eq. (2.8).
Each of the symmetries belongs to dierent universality classes in the quantum
regime, yet classical critical phenomena are always the same as in the classical model,
consistent with the expectation that quantum eects do not lead to qualitative
changes at nite temperatures. We will see that the dynamical critical exponent z
is dierent in both symmetries, causing the dierence in critical exponents at the
quantum critical point as was pointed out in Ref. [Her76].
2.2 Ferromagnetic Hamiltonians with creation and
annihilation operators
We want to study the Hamiltonian
H(
,

) =
i,=j
J
ij
i
(
i
+

i
)
2
=
1
2
i,=j
J
ij
(
S
i
S
j
+

j
)
S
i
,
(2.14)
where in the second equality we inserted Eq. (2.6). The i
S
i
j
cancelled since we
assumed symmetric couplings, J
ij
= J
ji
. Obviously, the momentum operators do
occur in this expression. The couplings J
ij
can in principle express any kind of inter-
action, ferromagnetic, antiferromagnetic, spin glass... The
i
represent an external
eld, that can be constant, variable, random... Later on, we will focus on ferromag-
netic couplings in the presence of constant magnetic eld. This Hamiltonian without
the external magnetic eld is symmetric under unitary transformations, a fact that
will determine the critical behavior.
The rst step to get the partition function is to diagonalize the couplings,
i
() =
()e
i
,
() =
i
()e
i
,
(2.15)
where e
i
is the normalized eigenvector of the coupling matrix J
ij
.
2.2.1 Coherent state representation for spherical spins
In this section we explain, following Ref. [NR98], how to add the spherical constraint
to a quantum Hamiltonian using the path integral formalism for models with the
second constraint Eq. (2.11). In second quantization the spins are given a bosonic
algebra. Then we can use the formalism introduced in section 1.1. We can deal with
spherical spins using almost the same approach described there. The operator a
i
is identied with

a
i
, where the index a denotes the spin vector direction, and the
corresponding elds
i
are denoted as
a
i
.
In order to impose this constraint, Eq. (2.11), in the path integral formalism,
the identity denition Eq. (1.16) is modied to adopt to the spherical case, in a way
inspired by Ref. [Nie95a, Nie95b, NR98]: one restricts the path integral to states
which exactly satisfy the constraint by employing the truncated identity
1 1
spherical
C
_

ia
d(
a
i
)d'(
a
i
)
[ ) [ ( n Nm), (2.16)
I.2.2 Ferromagnetic Hamiltonians with creation and annihilation operators 31
where the number operator n,
n =
i,a
a
i
a
i
, (2.17)
counts the total number of spin quanta. We insert
( n Nm) =
_

d
2
e
i ( nNm)
=
_
i
i
d
2i
e
( nNm)
(2.18)
where = i is imaginary. (Strictly speaking, we should insert a Kronecker- func-
tion, rather than the Dirac-, but for large N this amounts to the same.) Repeating
the same procedure with this new identity we get
Z =
_
()=(0)
DD
Dexp(A), (2.19)
with the action
A =
=0
d
_
()
d()
d
+()(
() ( d) Nm) +H(
(), ( d))
_
(2.20)
and integration measures dened as
_
D
D =
ia
_

d(
a
i
())d'(
a
i
())
, (2.21a)
_
D = C
_
i
i
d()
2i
, (2.21b)
so () is the Lagrange multiplier introduced to impose the spherical constraint and
the prefactor C
(M)
is added to ensure, if needed, a proper normalization. Details on
this factor were given in Ref. [NR98].
It should be noted that the particle number operator in the denition of the
identity eq. (2.16) will be surrounded, as is the case for the Hamiltonian, by spin
operators on dierent timesteps; therefore its creation and annihilation operators
will also be projected on dierent timesteps,

a
i
a
i

a
i
()
a
i
( d). In
Refs. [Nie95a, Nie95b, NR98] the spherical constraint was slightly dierent from
the one presented here but equivalent up to an additive constant.
It is worth remarking that, up to now, we imposed the spherical constraint
strictly, no thermal average has been performed. In the following section will be
integrated over by the method of steepest descent. This approximation allows the
particle number to uctuate and therefore the satisability of the constraint in the
end remains only in average.
2.2.2 General solution
We may write the partition function sum as a continuum expression,
Z =
_
DDD
exp
_
_
d
a,
_
()
d
a
()
d
+() (
a
()
a
( d) Nm)
J
()
a
( d)
1
(
a
() +
a
( d))
__
.
(2.22)
In discrete notation, the action of Eq. (2.19) reads
A =
_
1
a,
_
a
,j
a
,j

a
,j
a
,j1
_
+(j)
_
a,
a
,j
a
,j1
Nm
_
a,
J
a
,j
a
,j1
a,
(
a
,j
+
a
,j1
)
2
_
,
(2.23)
where = d is the imaginary time step, j the time index and
=

i

i
e
i
is the
eld in the basis of eigenvectors of J
ij
. Collecting all terms we have
Z =
_
D
,a
_
_

j
_
d
a
,j
d
a
,j
2i
_
exp
_
_
ij
a
,i
B
ij
a
,j
+
(
a
,j
+
a
,j1
)
2
_
_
_
e
P
j
Nm(j)
,
(2.24)
where B
ij
=
ij
(1 + J
(j))
t
i,j+1
; here the prime stands for the fact that
t
1,M+1
1 due to the trace structure of the partition function. We can now integrate
over the spins
Z =
_
Dexp
_
_
,a
_
_
_
mln det B
ij
+

2
ij
B
ij
1
+m
j
(j)
_
_
_
_
_
. (2.25)
As usual, in thermodynamics, the saddle point value of one-time quantities like
() can be taken independent of . We will employ this simplication throughout
the rest of this chapter. The determinant and the matrix inversion can then be
performed [NO98]. Integrating over by the saddle point method we obtain
F = m +
1
N
,a
_
ln(1 a
)
M
2
2(1 a
)
_
, (2.26)
where a
= 1 ( J
). Sending M we nally get

F = m +
m
N
_
ln(1 e
(J
)
)

2
2( J
)
_
=
m( +
1
2
) +m
_
dJ
(J
)
_
ln
_
2 sinh
_
2
( J
)
__
2( J
)
_
,
(2.27)
where in the last equality we have assumed that the couplings satisfy
1
N
= 0.
Otherwise, we would get an additional term 1/2mJ). The saddle point equation
reads
+ 1 =
1
N
_
1
1 e
(J
)
+

2
2( J
)
2
_
=
_
dJ
(J
)
_
1
1 e
(J
)
+

2
2( J
)
2
_
.
(2.28)
The sums over the dierent eigenvalues of the coupling matrix have been changed
into integrals. Each J
has a weight in this integral given by (J
). The actual form

for this weight function will depend on the type of couplings. A set of weight
functions for ferromagnets in dierent cubic lattices can be found in Ref. [Joy72],
and for spin glasses with long range interactions in Refs. [NR98, KTJ76].
At large temperatures these equations reduce to
F = m +m
_
dJ
(J
)
_
ln ( J
)

2
2( J
)
_
, (2.29)
+ 1 =
_
dJ
(J
)
_
T
J
+

2
2( J
)
2
_
. (2.30)
Apart from a factor two, these are exactly the equations of the classical spherical
model, see e.g. [Joy72]. This factor two arises because the momenta double the
degrees of freedom, see e.g. [Nie95a, Nie95b]. Near the phase transition they are
already approximate, but the transition stays within the classical universality class.
2.2.3 Ferromagnetic couplings with transversal eld in d di-
mensions
In this section we will use the results given in the previous one for the concrete case
of ferromagnetic couplings with uniform transversal eld. The Hamiltonian in this
case diers from the one before Eq. (2.14) in the fact that the couplings only act
in the z-direction while the external eld only acts in the x-direction (we restrict
ourselves therefore to m = 2). The free energy reads
F =(2 + 1) +
_
d
d
k
(2)
d
ln
_
2 sinh
_
2
( J(k))
__
+ ln
_
2 sinh
_
2
__

2
2
(2.31)
and the saddle point equation
2( + 1) =
_
d
d
k
(2)
d
1
1 e
(J(k))
+
1
1 e
+

2
2
2
, (2.32)
where we have applied the changes J
J(k) and
_
dJ
(J
) =
_

d
d
k
(2)
d
. (2.33)
We choose J(k) J
0
J
t
[k[
x
for [k[ 0. In the case of short range couplings,
for instance, one has x = 2 since J(k) =

J cos k
i
J(0)
1
2
J[k[
2
. A long range
coupling that decays as J(r) 1/r
at large r gives x = d.
As in the theory of Bose-Einstein condensation, the saddle point equation xes
the dependence of on temperature. There should be a solution at any T. In
order to have a real free energy, cannot be smaller than the maximum value for
J(k). Therefore, we should investigate the convergence of the integral in the limit
J
0
. If the integral diverges, must go to innity before reaches J
0
in order
to satisfy the saddle point equation, so there exists a for all temperatures and no
phase transition occurs. If the integral converges, however, there will be a range
of temperatures in which the saddle point as it stands cannot hold. This indicates
that we have overlooked a macroscopic occupation of the ground state, as occurs in
Bose-Einstein condensation. The relevant integral behaves as
_

d
d
k
(2)
d
1
1 e
(J
0
J(k))

d
(2)
d
_
0
dkk
d1
1
1 e
J
k
x

_
0
dkk
d1x
(2.34)
where
d
is the hypersurface of a sphere in d dimensions. At k = 0, this integral
converges for d > x, hence there will be a phase transition for dimensions larger
than x.
At low temperatures, may get stuck at J
0
and the saddle point equation as it
is in Eq. (2.32) is no longer valid. This is because, as in Bose-Einstein condensation
calculations, the ground state is not properly included in the integral. It should
be taken out of the sum before this one is converted to an integral. This causes
a change in the free energy by a factor ( J
0
)q, where q =
1
N
z
k=0
z
k=0
) is the
ground state occupation, and the saddle point equation becomes
2( + 1) =
_
d
d
k
(2)
d
1
1 e
(J(k))
+
1
1 e
+

2
2
2
+q, (2.35)
q can be evaluated from the saddle point equation ( J
0
)
q = 0. Thus when
= J
0
the occupation of the ground state can take non-zero values that can be
determined using Eq. (2.35). Hence the ground state occupation is macroscopic in
the ordered phase.
A transversal eld will lower the transition temperature. Above a certain value
c
, the transition does not exist anymore, thus T = 0, =
c
is a quantum critical
point, see chapter 1 or for a complete study over quantum phase transitions, see e.g.
[Sac99]. We will now rst study the classical critical point, at = 0.
Finite temperature phase transition
For the dimensions where the phase transition exists, the critical temperature is
found by solving the equation
2( + 1) =
_
d
d
k
(2)
d
1
1 e
c
(J
0
J(k))
+
1
1 e
c
J
0
. (2.36)
The dependence of the chemical potential (the Lagrange multiplier introduced
in Eq. (2.18) is equivalent to the chemical potential and both terminologies will be
used) on temperature near the transition is the rst thing needed. To get it, we
expand the saddle point equation around the critical point T = T
c
+, = J
0
+.
The integral gives, up to rst order in and
_

d
d
k
(2)
d
1
1 e
(J(k))

_

d
d
k
(2)
d
_
1
1 e
c
(J
0
J(k))
+
J
0
J(k)
4T
2
c
sinh
2
_
J
0
J(k)
2T
c
_
1
4T
c
sinh
2
_
J
0
J(k)
2T
c
_
_
.
(2.37)
The coecient of is an integral that diverges for d 2x. This means that
for these dimensions the leading term in the expansion of Eq. (2.37) has a power
smaller than one. For dimensions d > 2x we will have which will lead to
the mean eld exponents, thus d
uc
= 2x is therefore the upper critical dimension.
To study the system near the critical point we subtract Eq. (2.36) from the saddle
point equation, a procedure that will cancel the zeroth order term in the expansion
in and , giving nally
a
d<2x
dx
x
, a
d<2x
=
4
d
T
3
c

(2)
d
J
t
d
x
xsin
_
(dx)
x
_ for x < d < 2x
a
d=2x
ln , a
d=2x
=
4
d
T
3
c
(2)
d
J
t
2
x
for d = 2x (2.38)
a
d>2x
, for d > 2x.
a
d>2x
= T
c
_
_
_
d
d
k
(2)
d
1
sinh
2
_
J
0
J(k)
2T
c
_ +
1
sinh
2
_
J
0
2T
c
_
_
_
where
=
_
_
_

d
d
k
(2)
d
J
0
J(k)
sinh
2
_
J
0
J(k)
2T
c
_ +
J
0
sinh
2
_
J
0
2T
c
_
_
_
1
(2.39)
is a nite, positive number.
The internal energy of the system reads
U =(2 + 1) +
_

d
d
k
(2)
d
J(k)
2
coth
_
( J(k))
2
_
+

2
coth
_
2
_

2
2
.
(2.40)
The specic heat close to the transition from the paramagnetic side can be written
as
C
_
C
0
+
x
a
d<2x
(dx)
C
1
2xd
dx
for x < d < 2x,
C
0
+
1
a
d>2x
C
1
for d > 2x,
(2.41)
where
C
0
=
1
4T
2
_

d
d
k
(2)
d
[ J(k)]
2
sinh
2
_
J(k)
2T
_ +

2
4T
2
sinh
2
_

2T
_, (2.42)
C
1
=2 1 +
_

d
d
k
(2)
d
_
_
_
1
2
coth
_
J(k)
2T
_
J(k)
4T sinh
2
_
J(k)
2T
_
_
_
_
+
1
2
coth
_

2T
_

4T sinh
2
_

2T
_ +

2
2
2
,
(2.43)
where a
d
are the prefactors in Eq. (2.38) for the corresponding dimension. In the
ordered phase is stuck in its minimum value ( = J
0
) for any temperature. Hence,
C = C
0
( = J
0
) in the ordered phase. The critical exponent , see section 1 for
dention of the critical exponents, is the expected one: =
d2x
dx
for x < d < 2x,
and the mean eld value = 0 holds for d > 2x, which describes a jump in the
specic heat.
Adding a small longitudinal eld h, the free energy reads
F =(2 + 1) +
_
d
d
k
(2)
d
ln
_
2 sinh
_
2
( J(k))
__
+ ln
_
2 sinh
_
2
__

2
2

h
2
2( J
0
)
(2.44)
and the saddle point equation becomes
2( + 1) =
_
d
d
k
(2)
d
1
1 e
(J(k))
+
1
1 e
+

2
2
2
+
h
2
2( J
0
)
2
(2.45)
By dierentiating the free energy with respect to h can be seen that the magne-
tization is
M
z
=
h
J
0
. (2.46)
In the limit h 0, it is proportional to the square root of the occupation of
the ground state, since by comparing Eq. (2.45) with Eq. (2.35) one nds q =
1
N
z
k=0
z
k=0
) =
M
2
z
2
. The factor
1
2
appears because it is actually the real part of
the spin eld that is macroscopically occupied. This factor comes in from the trans-
formation Eq. (2.6). From Eq. (2.45), we can approach the transition by sending the
longitudinal eld to zero at the critical temperature. The saddle point equation now
accounts for the dependence of the chemical potential on the eld. The calculation
is similar, yielding nally
h
_
_
2
d
T
c
(2)
d
J
t
d/x
xsin
_
(dx)
x
_
_
_
1
2
d+x
2x
for x < d < 2x,
h
_
2
d
T
c
(2)
d
J
t
2
x
ln
_1
2
3
2
for d = 2x, (2.47)
h
_
_
d
d
k
(2)
d
1
2T
c
sinh
2
_
J
0
J(k)
2T
c
_
+
1
2T
c
sinh
2
_
J
0
2T
c
_
_1
2
3/2
for d > 2x.
Inserting these expressions in Eq. (2.46), the critical exponent is found to be
=
d+x
dx
for dimensions x < d < 2x and the mean eld value = 3 is recovered for
d > 2x. From the magnetization, the susceptibility follows as
1
. Therefore
we nd =
x
dx
for x < d < 2x and = 1 for d > 2x. In the ordered phase, the
expansion of the saddle point equation, Eq. (2.45), for T near the transition yields
M
2
z

1
2T
2
c
_
_
_

d
d
k
(2)
d
J
0
J(k)
sinh
2
_
J
0
J(k)
2T
c
_ +
J
0
sinh
2
_
J
0
2T
c
_
_
_
. (2.48)
Since M
2
z
for all dimensions where the phase transition exists, one has =
1
2
.
For other critical exponents the correlation function is needed. It can be com-
puted adding the right source term to the Hamiltonian,

(g
(
q
)
+ g
(
r
)
)
and dierentiating
G(,
q
[
t
,
r
) =T [
(
q
)
(
r
)])
=
,
2
g
(
q
)g
(
r
)
Z(g
, g)
Z
0
=g=0
(2.49)
where T stands for the time ordered product. Z(g
, g) is the partition function of

the Hamiltonian including the source terms and Z
0
is the partition function without
them. This procedure is carefully explained in Ref. [NO98] giving the result
G(k, ) = G(k, [k, 0) = e
(J(k))
( )(1 +n
k
) +( +)n
k
(2.50)
where is a Heaviside step function and is a positive innitesimal that indicates
that the second term is the relevant one at = 0. Furthermore,
n
k
=
1
e
1
(2.51)
is the boson occupation probability, with = (J(k)). Fourier transforming this
last result to Matsubara frequencies we get
G(k,
n
) =
1
J(k) i
n
k0
1
J
t
[k[
x
+ +i
n
. (2.52)
So when we approach the critical point, we can see from this equation that
x
, then using Eq. (2.38) we nd that =
1
dx
for dimensions x < d < 2x
and =
1
x
for d > 2x. = 2 x due to the fact that the couplings depend on k
x
,
and z = x because in the denominator
n
appears as a linear term. Both and z
are valid for any dimension.
This nally gives all the critical exponents of this nite temperature phase tran-
sition, which are exactly the same as in the classical model. This is expected from
renormalization group arguments [Her76]. The critical behavior is controlled by a
classical xed point, therefore quantum dynamics does not play a qualitatively new
role. Hence, the results are the same as in the classical spherical model [Joy72]
or other models with dierent quantum dynamics considered at nite temperatures
[Voj96].
T=0 quantum phase transition
In this section we analyze the behavior of the system at T = 0. As it can be seen from
Eq. (2.32), when the transversal eld increases, the temperature of the transition
decreases till it reaches zero. This denes a quantum critical point T
c
= 0 at =
c
.
In order to study it, an analogous procedure as before should be followed. At T = 0
everything happens to be rather simple. The free energy reduces to
F = 2

2
2
. (2.53)
The saddle point equation turns out to be
2 =

2
2
2
in the paramagnetic phase,
2 =

2
2J
2
0
+q in the ferromagnetic phase. (2.54)
where q =
1
2
M
2
z
is the occupation of the ground state for small transversal elds.
Since the temperature vanishes, quantum uctuations, controlled by , give rise to
the phase transition. Therefore, the parameter that should be used to control the
transition is the transversal eld and not the temperature. Then the proper analog
of the specic heat will be proportional to the second derivative of the free energy
with respect to the source of uctuations, the transversal eld.
C

2
F
2
=
1
. (2.55)
As before, we must know the dependence of ( = J
0
+) on the distance to
the critical point () in the paramagnetic phase. The upper critical dimension will
be d
uc
= x. Between those two dimensions, 0 < d < x, the analysis of the saddle
point equation, Eq. (2.32), yield that the product goes to a nite, strictly
positive value for T 0. This leads to a scaling form
x/d
. On the T = 0
line, the analog of the specic heat goes continuously from the paramagnetic value
C
1/J
0
+
c
/J
2
0
(xd)/d
to the simple ferromagnetic value C
= 1/J
0
. This
implies that = (d x)/d. Adding a longitudinal eld we nd the dependence
h
2x/(d+2x)
bringing = (d + 2x)/d and = x/d. Subtracting the saddle
point equation near the transition in the ferromagnetic phase from the one at the
transition, we get q = (
2
2
c
)/2J
2
0
, which in the lowest order gives M
2
z
and
therefore, as always, =
1
2
. Equation (2.52) can be used here once it is transformed
to real frequencies, i
n
= +i. Then we nd = 1/d, = 2 x and z = x.
For dimensions d > x, from eqs. (2.55,2.54), it can be seen that the analog of
the specic heat has a jump discontinuity, implying = 0.
C
= 0 in the paramagnetic phase,

C
=
1
J
0
in the ferromagnetic phase. (2.56)
where the minus sign comes from the fact that the T = 0 free energy, Eq. (2.53), is
negative. Adding a small longitudinal eld as before, we nd the critical exponent
= 3, since M
z

h
and h
2
3
at
c
. For the susceptibility, we nd = 1
since
1

1
. As before we nd =
1
2
and since we nd =
1
x
,
= 2x and z = x. Hence we nd always the mean eld values. This occurs because
the quantum critical point of a d-dimensional model shares the critical exponents of
a classical critical point of a (d+z)-dimensional model, as it was shown by general
renormalization group arguments [Her76].
Summary of the critical exponents
Finally all the critical exponents together read
Exponent Classical Quantum Classical (d > 2x)
(x = d
lc
< d < d
uc
= 2x) (d < d
uc
= x) Quantum (d > x)
d2x
dx
dx
d
0
1
2
1
2
1
2
x
dx
x
d
1
d+x
dx
d+2x
d
3
1
dx
1
d
1
x
2 x 2 x 2 x
z x x x
Table 2.1: Critical exponents for the Hamiltonian with momenta
We can see how the third column corresponds to the table 1.1 for the normal
nearest neighbors interaction where x = 2.
2.3 Hamiltonians involving spins but not their mo-
menta
In this section we extend the analysis of section 2.2 to a Hamiltonian which only
depends on the spin operators

S and not on the momenta

. When going from the
classical to the quantum model, we have to keep in mind that the Hamiltonian must
be Hermitian. To be precise, the Hamiltonian we will deal with is
H =
1
2
ij
J
ij
S
i
S
j

S
i
(2.57)
with real valued J
ij
and
i
and where

S = (
)/
2 is the real part of the former

spin eld. Hence, the Hamiltonian does not involve momenta, but the spherical
constraint does, see Eq. (2.11). This changes the symmetry of the problem from
invariance under unitary transformations to orthogonal ones. In terms of boson
creation and annihilation operators the coupling term for symmetric interactions,
J
ij
= J
ji
, is proportional to J
ij
(2
j
+

j
+

j
), where we can notice the
symmetry of the problem. We will see that this model reproduces the O(^) quantum
rotor model.
I.2.3 Hamiltonians involving spins but not their momenta 41
We can get the partition function in many ways. A similar procedure using
discrete imaginary time path integrals can be done as before. This gives us many
problems due to the fact that creation and annihilation operators are projected on
dierent time steps which is a lengthy and tedious procedure. However, the form
of the Hamiltonian makes it suitable to apply a Bogoliubov transformation (for
details see e.g. [Aue94]). Due to that, we get the same Hamiltonian as before but
with dierent coecients. In order to do so, we must add the spherical constraint
directly to the Hamiltonian. The procedure is as follows: rst the couplings matrix
is diagonalized as done before in Eq. (2.15) and then the

Ss are shifted to absorb
the eld term, i.e.

S
. This nally gives

H =
_
(
J
2
)
4
(
)

2
2( J
)
_
Nm. (2.58)
Performing the Bogoliubov transformation it turns into
H =
_
_
( J
2( J
)
_
Nm, (2.59)
where the canonical transformation consists of an imaginary rotation
= cosh
+ sinh
= cosh
+ sinh
(2.60)
where
are real and even in . The condition to get 2.59 from 2.58 is that
tanh 2
=
J
2
_

J
2
_. (2.61)
Finally Eq. (2.59) is a Hamiltonian analogous to Eq. (2.14). So it can be diago-
nalized as explained, giving nally
F =m( +
1
2
)
+m
_
dJ
(J
)
_
ln
_
2 sinh
_
2
_
( J
)
__
2( J
)
_
(2.62)
where we have put back the factor m standing for the number of components of the
vector spin. The saddle point equation is obtained as
+
1
2
=
_
dJ
(J
)
_
2 J
4
_
( J
)
coth

2
_
( J
) +

2
2( J
)
2
_
. (2.63)
At large temperatures these equations reduce to
F = m( +
1
2
) +m
_
dJ
(J
)
_
ln
_
( J
)

2
2( J
)
_
, (2.64)
+
1
2
=
_
dJ
(J
)
_
T
2
+
T
2( J
)
+

2
2( J
)
2
_
. (2.65)
These equations are very similar to the standard ones of the classical spherical
model (up to a factor 2), see Eq. (2.29), but they are only identical where they
should be, namely at large T, where also T is very large, see also Ref. [Joy72].
2.3.1 Ferromagnetic couplings in the presence of a transversal
eld
Analyzing the phase transition of the Hamiltonian that does not contain momenta is
analogous to the previous case. We begin again by choosing the coupling term in the
z direction and the external eld in the x and we assume ferromagnetic couplings.
The saddle point equation gives a phase transition via a macroscopic occupation of
the ground state, which in the present case is a bit more complicated. The critical
exponents are dierent, due to the fact that the symmetries of the system have
changed. The free energy reads
F =(2 + 1) +
_
d
d
k
(2)
d
ln
_
2 sinh
_
2
_
( J(k))
__
+ ln
_
2 sinh
_
2
__

2
2
,
(2.66)
and the saddle point equation
4 + 2 =
_
d
d
k
(2)
d
2 J(k)
2
_
( J(k))
coth
_
2
_
( J(k))
_
+ coth
_
2
_
+

2
2
.
(2.67)
We now analyze this model in detail.
Finite temperature phase transition
Following the same procedure as before we can nd that the transition exists for
d > x and that the upper critical dimension is d = 2x. The critical temperature is
the solution of
4 + 2 =
_
d
d
k
(2)
d
2J
0
J(k)
2
_
J
0
(J
0
J(k))
coth
_
2
_
J
0
(J
0
J(k))
_
+ coth
_
J
0
2
_
.
(2.68)
The dependence of the chemical potential in the temperature near the classical
critical point reads
a
d<2x
dx
x
, a
d<2x
=
2
d
T
3
c

(2)
d
J
t
d
x
xsin
_
(dx)
x
_ for x < d < 2x
a
d=2x
ln , a
d=2x
=
2
d
T
3
c
(2)
d
J
t
2
x
for d = 2x (2.69)
a
d>2x
, for d > 2x
a
d>2x
=
_
_
d
d
k
(2)
d
2 J(k)
2
_
( J(k))
coth
_
2
_
( J(k))
_
+ coth
_
2
_
__
=J
0
where
=
_
_
_

d
d
k
(2)
d
2J
0
J(k)
2 sinh
2
_
J
0
(J
0
J(k))
2T
c
_ +
J
0
sinh
2
_
J
0
2T
c
_
_
_
1
. (2.70)
The internal energy of the system reads
U =(2 + 1) +
_

d
d
k
(2)
d
_
( J(k))
2
coth
_
2
_
( J(k))
_
+

2
coth
_
2
_

2
2
.
(2.71)
The specic heat has the same expression as in Eq. (2.41) where a
d
now corre-
spond to the prefactors of Eq. (2.69) and with coecients
C
0
=
1
4T
2
_

d
d
k
(2)
d
( J(k))
sinh
2
_
(J(k))
2T
_ +

2
4T
2
sinh
2
_

2T
_, (2.72)
C
1
=(2 + 1) +
_

d
d
k
(2)
d
_
2 J(k)
4
_
( J(k))
coth
_
2
_
( J(k))
_
2 J(k)
8T sinh
2
_
(J(k))
2T
_
_
+
1
2
coth
_

2T
_

4T sinh
2
_

2T
_.
(2.73)
This is analogous to the previous model and gives the same exponent, =
d2x
dx
for x < d < 2x and = 0 for d > 2x. Adding a small magnetic eld longitudinal to
the couplings, the free energy becomes
F =(2 + 1) +
_
d
d
k
(2)
d
ln
_
2 sinh
_
2
_
( J(k))
__
+ ln
_
2 sinh
_
2
__

2
2

h
2
2( J
0
)
,
(2.74)
therefore the magnetization is M
z
=
h
J
0
, which is as before the square root of the
occupation of the ground state q =
1
N
<

S
z
([k[ = 0)
2
>= M
2
z
. The saddle point
equation is now
2(2 + 1) =
_

d
d
k
(2)
d
2 J(k)
2
_
( J(k))
coth
_
2
_
( J(k))
_
+ coth
_
2
_
+

2
2
+
h
2
( J
0
)
2
,
(2.75)
With all these and following the algebra of the previous section one nds the
same critical exponents for the magnetization for the same dimensions since we are
in the classical critical point.
The time ordered correlation function T
S
z
k
()
S
z
k
(0)) diers from the previous
one, Eq. (2.50) since in this case the

S are not the variables that diagonalize the
Hamiltonian in Eq. (2.59). We must write

S in terms of and then compute the
correlations. This brings
G(k, [k, 0) =
J(k)
4
_
( J(k))
_
n
k
cosh
_
_
( J(k))
_
+e
]]
(J(k))
_
,
(2.76)
where
n
k
=
1
e
1
(2.77)
is just like before in Eq. 2.51 but with =
_
( J(k)). The correlation function
in frequency space reads
G(k, i
n
) =
J(k)
2[
2
n
( J(k))]
. (2.78)
When approaching the critical point we nd that
x
as before and we
get the same critical exponent =
1
dx
for dimensions x < d < 2x and =
1
x
for d > 2x. Since couplings appear in the same way as before we also get the
same critical exponent, = 2 x for all dimensions.. The dierence appears in the
dynamical critical exponent. Here
n
appears squared, therefore z = x/2. Here we
see how the model reproduces the critical exponents of the rotor model as in Ref.
[Voj96] bringing thus a dierent behavior at the T = 0 quantum critical point from
the model of section 2.2.3.
T=0 quantum phase transition
In this case the T = 0 phase transition the dynamical critical exponent z =
x
2
is
smaller than z = x of the previous section. The free energy reads
F = (2 + 1) +
_
d
d
k
(2)
d
_
( J(k))
2
+

2

2
2
(2.79)
and the saddle point is set by
4 + 1 =
_
d
d
k
(2)
d
2 J(k)
2
_
( J(k))
+

2
2
. (2.80)
We nd that the transition exists for dimensions larger than d >
x
2
and d =
3x
2
is the upper critical dimension. The chemical potential depends on the source of
uctuations =
c
as
a
d<3x/2
2dx
2x
, a
d<3x/2
=

d
J
5
2
0
2(2)
d
J
t
d
x
_
3
2

d
x
_
_
d
x
_
(2d x)x
for
x
2
< d <
3x
2
a
d=3x/2
ln , a
d=3x/2
=

d
(2)
d
J
5
2
0
8xJ
t
3
2
for d =
3x
2
(2.81)
a
d>3x/2
, for d >
3x
2
a
d=3x/2
=
J
2
0
2
_
2
2
c
J
3
0
_
d
d
k
(2)
d
_
1
_
J
0
(J
0
J(k))

2J
0
J(k)
4 (J
0
(J
0
J(k)))
3
2
__
where the s on the right hand side of the rst equality are Eulers Gamma functions.
The specic heat (see eq. (2.55) ) coming from the disordered region will behave as
C
_
1
J
0

4
c
a
d<
3x
2
J
2
0
(2dx)
2d+3x
2dx
for
x
2
< d <
3x
2
,
1
J
0
+
2
c
J
4
0
_
2
2
c
J
3
0
a
d>
3x
2
_
1
for d >
3x
2
,
(2.82)
where a
d
is the prefactor in Eq. (2.81) for the proper dimension. Coming from the
ordered region, conversely, C

1
J
0
. Therefore =
2d3x
2dx
for
x
2
< d <
3x
2
and
= 0 for d >
3x
2
.
The dependence of a small longitudinal eld on the chemical potential, in case
the transversal eld is at its critical value, reads
h a
1
2
d<
3x
2
2d+3x
4x
for
x
2
< d <
3x
2
,
h
_
a
d=
3x
2
ln
_1
2
3
2
for d =
3x
2
, (2.83)
h
_
2
2
c
J
3
0
a
d>
3x
2
_
1
2
3
2
for d >
3x
2
.
From these equations and the ones for the magnetization and the susceptibility
we can nd that =
2d+3x
2dx
and =
2x
2dx
for
x
2
< d <
3x
2
, while = 3 and = 1
for d >
3x
2
. As before =
1
2
for every dimension. For the correlation function the
calculation is the same as in the nite temperature case, projected into real time,
the exponents are =
2
2dx
for dimensions
x
2
< d <
3x
2
and =
1
x
above the critical
dimension, = 2 x and z =
x
2
for all dimensions.
Summary of the critical exponents
Finally the critical exponents read
Exponent Classical Quantum Classical (d > 2x)
(x = d
lc
< d < d
uc
= 2x) (
x
2
< d <
3x
2
) Quantum (d >
3x
2
)
d2x
dx
2d3x
2dx
0
1
2
1
2
1
2
x
dx
2x
2dx
1
d+x
dx
2d+3x
2dx
3
1
dx
2
2dx
1
x
2 x 2 x 2 x
z
x
2
x
2
x
2
Table 2.2: Critical exponents for the Hamiltonian without momenta
2.4 Generalization and mapping from Heisenberg
spins
In this section we generalize the two preceding Hamiltonians and we map the Heisen-
berg model onto the spherical model. In a more compact way, we can write the
former Hamiltonians in absence of external eld as
H =
ij
_
A
ij
j
+
B
ij
2
_
j
+

j
_
_
. (2.84)
I.2.4 Generalization and mapping from Heisenberg spins 47
If the matrices A
ij
and B
ij
can be diagonalized simultaneously, the techniques
from previous sections can be used. The free energy reads
F = m( +
1
2
) +
m
N
_
A
2
+ ln
_
2 sinh
_
2
_
( A
)
2
B
2
___
(2.85)
where satises the saddle point equation
+
1
2
=
1
N
2
_
( A
)
2
B
2
coth
_
2
_
( A
)
2
B
2
_
. (2.86)
The coecient B
ij
in Eq. (2.84) is the responsible for a change in the symmetries
of the problem. If B
ij
is zero, the action is symmetric under unitary transformations
while if it is non-zero the symmetry is reduced to orthogonal.
The mapping from Heisenberg spins comes as follows. The Hamiltonian can be
written in terms of Schwinger bosons [Aue94]. The Schwinger boson transformation
for SU(2) spins reads
S
+
= a
1
a
2
, S
= a
1
a
2
, S
z
=
1
2
(a
1
a
1
a
2
a
2
). (2.87)
This can be generalized to SU(^) spins and expand around the large-^ limit
[AA88]. In a path integral formalism for ferromagnetic interactions
H =
1
2
ij
J
ij
S
i
S
j

1
2N
ij,mn
J
ij
a
jm
a
im
a
in
a
jn
, (2.88)
where i, j represent lattice sites and m, n represent the boson avor. The Hilbert
space spanned by Schwinger bosons is much larger than the one given by Heisenberg
spins. The constraint needed to restrict it to the physical Hilbert space is that
the number of Schwinger bosons at each site has to be kept xed

A
m
n
m
= ^S.
This is inserted into the formalism in the same way as we have done it with the
spherical constraint, a Lagrange multiplier
i
() appears. The biquadratic terms
can be decoupled by a Hubbard-Stratonovich transformation (see e.g. [NO98]). In
the case of a ferromagnet, the transformation at each time step and for each avor
in the path integral reads
exp
_
_
_

2N
ij
J
ij
a
j
a
i
a
i
a
j
_
_
_
_

i,j
dQ
ij
exp
_
_
_
^
2
i,j
Q
ij
J
ij
Q
ji
ij
Q
ij
J
ij
a
j
a
i
_
_
_
(2.89)
where a eld Q
ij
() has been generated. In the mean eld approximation, one puts
Q
ij
() = Q and
i
() = . Hence, one gets up to a non interesting constant
H
FMB
MF
(^) =
i,m
a
im
a
im
Q
ij,m
J
ij
a
jm
a
im
+
NQ
2
2
ij
J
ij
^NS (2.90)
where we have already added the Schwinger boson constraint

i,m
a
im
a
im
= ^NS.
The free energy per particle reads
F =
^
N
k
ln(1 e
(QJ(k)
) +
^Q
2
2
J(k = 0) S^ (2.91)
and the saddle point equations are
1
N
k
n
k
= S, (2.92)
1
N
k
J(k)n
k
= QJ(k = 0), (2.93)
where n
k
is the boson occupation number Eq. (2.51) with = QJ(k). Sub-
tracting the two saddle point equations we can see that for large S and small T
we can approximate Q S recovering then the spherical model Eq. (2.27,2.28) for
zero external eld or Eq. (2.85,2.86) for B
ij
= 0. From this approach we thus see
that the free energy of a SU(^) Heisenberg ferromagnet for large ^ is formally the
same as the quantum spherical model proposed in Eq. (2.14) in the thermodynamic
limit, so when the radius of the hypersphere that denes the model (N in eq. (2.11)
) is also very large. Thus the large ^ limit is analogous to Stanleys large spin
dimensionality limit.
In the case of an SU(^) antiferromagnet the procedure is more or less the same
but the symmetries are dierent. The lattice is divided in two sublattices A, B.
In one of the sublattices a spin rotation is performed that allows us to write the
Hamiltonian in the form [AA88]
H =
1
2
ij
J
ij
S
i
S
j

1
2N
ij,mn
J
ij
a
im
a
im
a
jn
a
jn
. (2.94)
Performing a Hubbard-Stratonovich transformation as before, the Hamiltonian
with the Schwinger boson constraint in the mean eld approximation nally reads
H
AFMB
MF
(^) =
i,m
a
im
a
im
Q
2
ij,m
J
ij
(a
im
a
jm
+a
im
a
jm
)
+
NQ
2
2
ij
J
ij
^NS.
(2.95)
I.2.5 Summary and discussion 49
It is important to stress that here the SU(^) symmetry has been reduced to a
residual O(^). The free energy per particle reads
F =
^
N
k
ln
_
2 sinh
_
2
_
2
Q
2
J
2
(k)
__
^
_
S +
1
2
_
+
^Q
2
2
J(k = 0)
(2.96)
and the saddle point equations read
1
N
2
Q
2
J
2
(k)
_
n
k
+
1
2
_
= S +
1
2
, (2.97)
1
N
k
J
2
(k)Q
_
2
Q
2
J
2
(k)
_
n
k
+
1
2
_
= QJ(k = 0), (2.98)
where n
k
is Eq. (2.51) for =
_
2
Q
2
J
2
(k). Subtracting the rst equation times
from the second times Q we get
1
N
k
_
2
Q
2
J
2
(k)
_
n
k
1
2
_
=
_
S +
1
2
_
Q
2
J(k = 0). (2.99)
The rst term is proportional to T, so for very small temperatures and very large
S, near the transition where QJ(k = 0), we can approximate Q S +
1
2
. Then
eqs. (2.96,2.97) are analogous to eqs. (2.85,2.86) for A
ij
= 0. This will have the
same critical behavior as the model in section 2.3 due to the fact that it comes from
the term coth[
_
J(k)]/
_
J(k) which also appears here due to the equality
2n
k
+ 1 = coth[
_
( +QJ(k))( QJ(k))].
2.5 Summary and discussion
In this chapter we have discussed a way of working with quantum spherical spin
models using path integrals and coherent states. Some examples of the use of this
formalism are given, eqs. (2.14,2.57), and their critical phenomena are studied.
We propose a comparison with SU(^) Heisenberg models that gives a geometrical
interpretation to the quantum spherical spins. The spherical constraint we use, xes
the number of spin quanta

, Eq. (2.11); in other words, it xes both the average
length square of the spin operator,

S
2
, and the one of its conjugate momentum,

2
.
The usual version of the quantum spherical model, on the contrary, involves only the
spin part

S. The presence of momenta in the spherical constraint allows to consider
Hamiltonians that have no kinetic term, since it can be induced by the constraint,
a fact that can change the symmetries of the problem, and due to that, the critical
behavior.
The Hamiltonian in Eq. (2.14) yields an action invariant under unitary transfor-
mations. It brings formally the same free energy as a SU(^) Heisenberg ferromagnet
in the limit of large ^. The other Hamiltonian studied, Eq. (2.57), brings an action
invariant under orthogonal transformations; it gives the same critical behavior as
an SU(^) Heisenberg antiferromagnet in the limit of large ^, which is, in its turn,
analogous to an O(^) nonlinear -model or quantum rotor model [Voj96, Sac99].
The main dierence between these models lies in the dynamical critical exponent
z which brings a dierent behavior at the quantum critical point. Classical critical
phenomena are, as expected, the same in both models and equal to those of the
classical spherical model.
In the formulation of the model, the strict spherical constraint has been used
where uctuations on the particle number are not allowed. The constraint is added to
the action via a Lagrange multiplier. The strict approach has to be abandoned when
we integrate this Lagrange multiplier since we cannot perform the integration exactly
and we use the saddle point approximation. In this step, we automatically allow
uctuations on the particle number and therefore the constraint ends being satised
only in average. These eects are immaterial in the considered thermodynamic limit,
but do enter nite size corrections.
When performing the analogy between the spherical model and the Heisenberg
one, we did not include the external eld. Mapping such term was not possible since
in the spherical model the external eld comes linearly, as a source term. In spite
of that, the phase diagram follows the expected behavior for a spin model with an
external transversal eld. The critical exponents for the classical and the quantum
model are the ones expected by renormalization group arguments [Her76]. The
quantum critical point behaves as the classical one for dimensions D
quant
= d
class
+z
where z is the dynamical critical exponent.
Many works in literature are closely related to the one described here. For in-
stance, starting from the SU(^) Heisenberg model and to do the already stated
large ^ limit to get to a solvable model. In order to have a transversal eld that
competes with the ordering of the interacting spins one could introduce anisotropy
in the model. A study of this type has been done for 2 dimensions by Timm et al.
[TJ00] in terms of Schwinger bosons and in terms of Holstein-Primako bosons by
Kaganov et al. [KC87] for any dimension. The anisotropy term brings a residual
spin symmetry describing Ising or XY spins. The phase transition depends on the
type of this residual symmetry; an additional transversal eld decreases the transi-
tion temperature towards zero giving a quantum critical point, result qualitatively
reproduced by our model.
Additionally, Sachdev and Bhatt [SB90] represented pairs of spins in a square
lattice with a bond representation; they form either a singlet or a triplet. These
elements can be written down in terms of the canonical Schwinger boson repre-
sentation of the generators of SU(2)SU(2) = SO(4). Since a couple of spins either
form a singlet or a triplet, a constraint must be added s
s +
= 1, where s
represents the singlet annihilation operator, and t
represents a triplet annihilation

operator in the direction. Sachdev and Bhatt study using this formalism systems
with interactions up to third nearest neighbors. They make the further assumption
I.2.5 Summary and discussion 51
that the singlet part condenses and replace the s operator by its mean eld value
s) = s, and solve the rest for the triplets. The nal Hamiltonian is very close to
our Eq. (2.57), or, better, the generalization of our model Eq. (2.84) with the proper
couplings. A minor dierence is the role played by the non-constant mean value of
the singlet part.
Part II
Spin-Boson Model
3
Introduction
3.1 Motivation
A known feature of technological progress is the increase of human ability to control
and design the microscopic world. Recent eorts in manipulating simple quantum
systems, e.g. in the context of quantum computing or quantum chemistry, is one
aspect of this general trend. Another aspect is the eld of quantum thermodynamics
whose main objective is in designing and studying thermodynamical processes in the
domain where quantum features of small systems are relevant. In particular, this
activity aims to improve our understanding of phenomenological thermodynamics by
addressing its concepts from the rst principles of quantum mechanics. The current
activity in quantum thermodynamics includes quantum engines, general aspects of
work extraction from quantum systems, thermodynamical aspects of quantum infor-
mation theory and limits of thermodynamical concepts such as the second law and
temperature. There were also much earlier applications concerning, in particular,
thermodynamic aspects of lasers and masers.
Our present purpose is to study two complementary processes: work extraction
from a two-temperature system and cooling at the expense of work, on the basis
of the spin-boson model [BP02, L
+
87, Luc90, VL98]. The spin-boson model is one
of the simplest models in the study of quantum open systems. By open quantum
systems, we mean systems which are in contact with an environment, which for us
will be a thermal bath. Closed systems, on the contrary, are systems which can
be described on themselves, without giving any consideration to their surroundings,
though for instance, external elds can act on them. Furthermore, one could speak
over isolated systems where all dynamics is determined by the internal degrees of
freedom without any external inuence.
The study of an open quantum system consists in describing the dynamics of the
system together with the dynamics of the environment. The latter can produce qual-
itative changes on the dynamics of the system. In our situation, the necessity of the
bath has to be stressed since in the usual practice of quantum system manipulation,
56 II.3. Introduction
the bath is a serious hindrance one cannot get rid of. In contrast, both processes
of work-extraction and cooling, when starting from an equilibrium situation, cannot
be achieved without external thermal baths.
The second law in Thomsons formulation
1
which can be derived as a theo-
rem in quantum mechanics [Thi83, BK77, BK79, Bas78, PW78, Len78, AN02]
forbids work-extraction from an equilibrium system via cyclic processes gener-
ated by external elds. The easiest way to employ such an equilibrium system
in work-extraction is to attach it to a thermal bath having a dierent temper-
ature, which is the standard setting of heat engines. The entire system is then
out of equilibrium and work-extraction is not forbidden, at least in principle,
at the most general level.
The no-cooling principle, which is related to the second law, states that an
equilibrium system cannot be cooled that is for a two level system that the
occupation of its ground state cannot be increased by means of cyclic exter-
nal elds [KP92]. It is, however, possible to cool in the presence of the thermal
bath, even when having the same initial temperature. Then according to the
above statement of the second law such a cooling process will be accompanied
by a loss of work, which is the energy cost needed for cooling, that is, the
principle of a refrigerator.
3.1.1 Work Extraction
The general restrictions just mentioned determined the way how standard quan-
tum work extraction (also known as amplication or lasing/masing) processes are
designed [Sie71]. The most traditional lasers and masers operate by extracting
work from an ensemble of two-level systems having a negative temperature, in other
words, population inversion, which is a strongly non-equilibrium state. More recent
schemes of lasing without inversion employ non-equilibrium states of three (four,
multi-) level systems without population inversion of energy levels, but, in contrast,
they have initially sizable non-diagonal terms of the corresponding density matrix in
the energy representation, usually called coherences [SZ97, Koc92]. These schemes
attracted attention due to both their conceptual novelty and the fact that non-
zero non-diagonal elements represent a weaker form of non-equilibrium compared to
population inversion, and thus their preparation can be, in principle, an easier task.
3.1.2 Cooling of spins
Cooling, i.e. obtaining relatively pure states from mixed ones, is of central impor-
tance in elds dealing with quantum features of matter. Laser cooling of motional
states of atoms is nowadays a known achievement [EMSKB03]. The related problem
1
Thomsons formulation of the second law: Departing from equilibrium, so (0) = e

H
/Z, and
having a cyclic process

H(t), so

H(0) =

H() =

H, then W 0
II.3.2 Open quantum systems and spin boson model 57
of cooling spins is equally known: it originated as an attempt to improve the sensi-
tivity of NMR/ESR spectroscopy [AG82, Sr89, Sli90, Lam68, C
+
90, H
+
97, A
+
03,
B
+
02, I
+
03], since in experiments the signal strength is proportional to polariza-
tion. Recently it got renewed attention due to realizations of setups for quantum
computers in NMR physics [GC97, War97]. The very problem arises since the most
direct methods of cooling spins, such as lowering the temperature of the whole sam-
ple or applying strong dc elds, are not feasible or not desirable, e.g. in biological
applications of NMR, where strong static elds or low temperatures may destroy the
very studied material. Without any external inuence the polarization is really low,
for example, at temperature T = 1K and magnetic eld B = 1T the equilibrium
polarization of a proton is tanh
B
2k
B
T
= 10
3
where is the ratio
frequency
eld
and it is
equal to 42 MHz/T for a proton, 10
3
larger for electrons and 10 times smaller for
15
N.
Over the years, several methods were proposed to attack the problem of small po-
larizations. The polarization is generally increased via a dynamical process and it is
used before relaxing back to equilibrium [AG82, Sr89, Lam68, C
+
90, H
+
97, A
+
03,
Sli90, I
+
03, B
+
02]. Especially known are methods where a relatively high polariza-
tion is transferred from one place to another, e.g. from electronic to nuclear spins
[AG82, Sr89, Lam68, C
+
90, H
+
97, A
+
03, Sli90, I
+
03]. In this respect electronic
spins play the same role as the zero-temperature bath of vacuum modes employed
for laser cooling of atoms [EMSKB03] (this latter bath is typically useless for cooling
spins). Polarization transfer was studied in various settings both theoretically and
experimentally [AG82, Sr89, Sli90, Lam68, C
+
90, H
+
97, A
+
03, I
+
03]. However,
this scheme is limited besides requiring already existing high polarization by
the availability and eciency of the transfer interaction.
A related method, polarization compression, consists in manipulating a set of n
spins, each having a small polarization, in such a way that the polarization of one
spin is increased at the expense of decreasing the polarization of the remaining n1.
These spoiled spins can be recycled and used again [B
+
02]. This method cools spins
one by one, thereby taking a long time for cooling a large ensemble of spins, and
requires carefully designed inter-spin interactions.
3.2 Open quantum systems and spin boson model
The following chapters are devoted to quantum open systems, in particular to the
spin-boson model, more precisely, an exactly solvable limit of it. In this introduction
we give the basic ingredients of the spin-boson model. Only a general background
is given. In chapter 4, a concrete situation will be given, motivated and solved.
As already stated, in quantum open systems we describe the subsystem plus its
environment. We can write then the full Hamiltonian as
H =

H
S
+

H
B
+

H
I
(3.1)
where

H
S
corresponds to the Hamiltonian of the subsystem S,

H
B
corresponds to
the Hamiltonian of the environment or reservoir (which is typically a heat bath) and
H
I
correspond to the interaction between them. Then the Hilbert space where

H
works is composed in the tensorial product of the subsystems Hilbert space with
the baths Hilbert space: SB.
The state of the subsystem S will change as a consequence of its internal dynamics
and of the interaction with the surroundings. The interaction with the environment
can be such that the resulting state changes in the subsystem S alone can no longer
be represented by unitary, Hamiltonian dynamics. To get the subsystems dynamics,
one traces out the environment degrees of freedom (tr
B
), e.g. one sums over all
possibilities for the environment. By doing this we get an eective description of the
dynamics of the subsytem, the so-called reduced dynamics (S is called the reduced
system).
In the spin-boson model, the sub-
V
0
+
Figure 3.1: Double well potential
system is a two-level system, thus a spin.
By two-level system we can mean an
intrinsic two-level system, i.e. a 1/2-
spin, the polarization of a photon; but
we can also consider a system having
many levels where the two lowest ones
are the only accessible ones or a system
having a continuous degree of freedom
subject to a potential energy function
V (q) with two separated minima, see
g.(3.1). If the height of the barrier
V
0
is large enough compared with
+
,
the two minima can create an eec-

tive two-level system. The energy lev-
els though, do not correspond to each
of the minima but to a state which is
a superposition of both. Examples of such a situation could be for instance some
types of chemical reaction or the motion of defects in some crystaline solids. Eec-
tive two-level systems are also believed to be responsible for the linear specic heat
of amorphous solids at low temperatures 1K [Phi81, Esq98].
Therefore the motivation to use a two-level system is overwhelming, it is almost
everywhere and it is the minimal model having non-trivial quantum features. In the
spin boson model [L
+
87], the Hamiltonian for the subsystem reads
H
S
=
1
2

x
+
1
2

z
(3.2)
where
x
,
y
and
z
are Paulis matrices. In the picture of the two-wells potential,
is related with the transition probability between wells and with their energy
dierence, see g.(3.1). In the spin language, however, and correspond to
magnetic elds acting on the x or the z direction respectively.
Furthermore, in the spin-boson model the environment is modeled by a set of
harmonic oscillators, thus bosons. In some cases this may be taken in the literal
sense, when harmonic oscillators represent phonons or photons. However, it is also
II.3.2 Open quantum systems and spin boson model 59
known that rather general classes of thermal baths can be eectively represented
via harmonic oscillators. The motivation for that is the subject of sections and
appendices of many papers, appendix C in Ref. [CL83] or section 2 in Ref. [MS80]
to mention some.
The basic idea is that the bath is a macroscopic entity in a stable equilibrium
state which is only weakly perturbed by the interaction with the system. Therefore,
due to the weak perturbation noticed by the bath, the system sees only excitations
which can be considered as the ones of harmonic oscillators.
The Hamiltonian of the bath is thus taken as
H
B
=
k
a
k
a
k
, [ a
l
, a
k
] =
kl
, (3.3)
where a
k
and a
k
are boson creation and annihilation operators of the bath oscillator
with the index k with frequency
k
.
The last part of the Hamiltonian is the interaction between the systems degrees
of freedom and the environment surrounding it. According to the above assumptions
of harmonicity of the bath, the interaction, up to this approximation, can be taken
as linear in the variables of the bath. Due to the assumption that the bath is
weakly perturbed, higher order terms, i.e. anharmonicities, can be neglected. A
linear interaction suces to bring the system in equilibrium to the temperature of
the bath. For stronger arguments in this line, we refer the reader to [L
+
87, CL83].
Then in general
H
I
=

2
a=x,y,z

a

X
a
,

X
a
k
g
a
k
( a
k
+ a
k
), (3.4)
where g
a
k
is the strength of the coupling of the a component of the system to the k
oscillator. The coupling only depends on the position and not the momenta of the
oscillators. It can be proven that any linear coupling can, via appropiate canonical
transformations, be expressed only in terms of position operators [Leg85]. Further-
more, this interaction couples the bath to all spin components. This is usually
relaxed to couple the bath only to
z
. The reason for that depends on the system
studied. In the following chapter we will do so and motivate this approximation.
If the modes of the bath are dense enough, we can describe the eect of the
interaction of the bath to the system via a single spectral function J(). This is
the case, for instance, when the thermodynamic limit is taken for the bath. Then
the coupling to each of the harmonic oscillator g
k
is not important and a global
description for the interaction suces to obtain all interesting physics. The spectral
density function reads:
J() =
k
g
2
k
(
k
). (3.5)
All quantities involving the interaction with the bath will be composed by integrals
of this function.
The problem is completely dened by the parameters, , and the function
J(). However, this is untractable analytically. A lot of eord has been spent in the
last two decades to nd appropiate approximations that lead to reasonable physics,
though maybe, in a restricted area of application, for instance, high temperatures or
weak coupling strength to the heat bath. In the following chapter we will propose a
simplication of this general model that ends up in an analytically tractable model
for all temperatures and coupling strength to the heat bath. Further we will study
how, via external perturbations, one can achieve cooling and work extraction in this
setup.
4
Solvable model
The most general version of the spin boson model is intractable analytically. There
are many dierent approximations to solve the dynamics dened in Eqs. (3.1-3.4),
see e.g. [L
+
87, HG98]. Our approach consists in making all approximations in the
denition of the model. Then we end up having a model with restricted validity,
which, however, has an exact solution under the constraints imposed. This way we
have a better control on the approximations we perform. That is very important for
our purposes as it will be described later 4.4.1.
4.1 The model
In this section we will describe the precise version we take from the generic spin-
boson model and its motivation. In the following sections, we will show how to get
through the analytic details to nd the dynamics. As before, and in general for open
systems we start from Eq. (3.1). The two level system (spin
1
2
) is described by the
Hamiltonian:
H
S
=

2

z
, , (4.1)
so a spin system with energy levels
2
. This basically represents a spin under the
action of a constant magnetic eld where we take the direction of the magnetic eld
as the z direction.
The spin interacts with a thermal bath which is a set of harmonic oscillators as
before in Eq. (3.3). Later on, the thermodynamic limit for the bath will be taken.
The next important point is to specify the interaction between the spin and the
bath. In general, we should use Eq. (3.4) since the bath perturbs/damps the spin in
all directions. In this respect, for two-level systems, one distinguishes three types of
relaxation processes and the corresponding time scales:
i) T
2
-time scale, related to the relaxation of the average transversal components

x
) and
y
) of the spin (decoherence). Note that the very notion of the
62 II.4. Solvable model
transversal components is dened by the form (4.1) of the spin Hamiltonian.
ii) T
1
-time scale, related to the relaxation of
z
).
It is customary to have situations, where
T
2
T
1
, (4.2)
the main physical reason being that the transversal components are not directly
related to the energy of the spin and their decay is not directly connected with
energy exchange between the spin and the bath.
Our basic assumption on the relaxation times is (4.2). Moreover, to facilitate
the solution of the model we will disregard T
1
time as being very large, thereby
restricting the times of our interest to those much shorter than T
1
. The interaction
Hamiltonian is thus chosen such that it induces only transversal relaxation:
H
I
=

2

X
z
,

X
k
g
k
( a
k
+ a
k
), (4.3)
where g
k
are the coupling constants to be specied later, and where

X is the collective
coordinate operator of the bath.
iii) A third relaxation time T
2
has a completely dierent origin. It only appears
when dealing with an ensemble of non-interacting spins each having Hamilto-
nian (4.1) with a randomly distributed energy (dephasing). This randomness
can be caused by inhomogeneous elds contributing into energy , or by ac-
tion of environment, e.g., chemical shifts for nuclear spins or eective g-factors
for electronic spins in a quantum dot. Due to this dispersion in , dierent
spins will precess at dierent speed causing a relaxation in the transversal
components of the magnetization, i.e. the average of ). The relaxation T
2
is directly related to the distribution of if this is suciently large.
As a starting point we will consider situations where T
2
is innity, i.e. we
consider only one spin or identical copies of one spin. In sections 5.4 and 6.2 we
will consider work extraction and spin cooling respectively in the case of arbitrary
short T
2
. In order to cancel the eect of this relaxation we will use the spin-echo
technique [Hah50], bringing qualitatively the same results as with only one spin.
The eect, however, is reduced quantitatively.
The model up to now described has been considered in literature [Luc90, Unr95,
PSE96, VL98, SL04] to study decoherence. This model describes pure decoherence,
no damping is considered, since
) decay but
z
) does not. In order to study
work extraction and spin cooling we need to use external elds to perturb the system.
These external elds will be in the form of pulses in order to keep the model ana-
lytically solvable. The approximations needed and the formalism used is described
in section 4.2.
II.4.1 The model 63
4.1.1 Heisenberg equations and their exact solution
The Heisenberg equations for the operators
z
(t) and a
k
(t) read from Eqs. (3.1, 3.3,
4.1, 4.3):

z
= 0,
z
(t) =
z
(0), (4.4)
a
k
=
i
[H, a
k
] = i
k
a
k
i
2
g
k

z
. (4.5)
Eqs. (4.4, 4.5) are solved as
a
k
(t) = e
i
k
t
a
k
(0) +
g
k

z
2
k
_
e
i
k
t
1
_
, (4.6)
and then
X(t) = (t)
z
G(t), (4.7)
where
G(t)
k
g
2
k
k
(1 cos
k
t), (4.8)
quanties the reaction of the spin on the collective operator of the bath, and where
we denoted
(t) =
k
g
k
[ a
k
(0)e
i
k
t
+ a
k
(0)e
i
k
t
], (4.9)
for the quantum noise operator
1
. Recalling the standard relations

=
x
i
y
, [
z
,
] = 2
,
z

, (4.10)
and using (4.7) and [

X(t),
(t)] = 0 since they belong to dierent Hilbert

spaces, one derives
=
i
[H,
] = i
_
+

X
_

+
= i ( (t) G(t) )
. (4.11)
These equations are solved as:

(t) = exp [it iG

1
(t)]

(0, t)
(0), (4.12)
(t
0
, t
1
) T exp
_
i
_
t
1
t
0
ds (s)
_
, (4.13)
F(t)
_
t
0
ds G(s) =
k
g
2
k
k
_
t
sin
k
t
k
_
, (4.14)
1
Note that the commutator of the quantum noise is a c-number: [ (t), (s)] =
2i
P
k
g
2
k
sin
k
(t s) = 2i

G(t s).
where T stands for the time-ordering operator.
It is also useful for later to consider the Heisenberg time evolution of the

operator:
e
i1t/
(t
1
, t
2
)e
i1t/
=

(t
1
+t, t
2
+t) exp [i
z
(t
1
, t
2
, t)] (4.15)
where
(t
1
, t
2
, t) F(t
2
) F(t
1
) +F(t
1
+t) F(t
2
+t), (4.16)
with F(t) dened in Eq. (4.14).
4.1.2 Factorized initial conditions
We assume that initially, at the moment t = 0, the bath and the spin were in the
following factorized state:
(0) =
S
(0)
B
(0) =
S
(0)
e

H
B
tr e

H
B
(4.17)
where
S
(0) is the initial density matrix of the spin, and where the bath is initially
at equilibrium with temperature T = 1/.
Factorized initial conditions are adequate when the spin was prepared indepen-
dently from the equilibrium bath and then is brought in contact to the bath at the
initial time
2
. For example, injection of an electronic spin into a quantum dot, or
creation of an exciton by external radiation. Yet another situation where factorized
initial conditions are adequate is selective measurement of
z
by an external appa-
ratus. In this case
S
(0) is an eigenstate of
z
upon which the selection was done.
Non-factorized initial states are commented upon below, in section 4.1.3.
The equilibrium relation
a
k
(0) a
k
(0) + a
k
(0) a
k
(0)) = coth
_
k
2
_
, (4.18)
implies that the quantum noise is stationary Gaussian operator with
(t)) = 0, (4.19)
and with the following time-ordered correlation function:
K
J
(t t
t
) = T [ (t) (t
t
)])

=
k
g
2
k
_
coth
_
k
2
_
cos
k
(t t
t
) i sgn (t t
t
) sin
k
(t t
t
)
_
(4.20)
2
It is useful to note that this process of bringing spin in contact to the bath need by itself not
be connected with any fundamental energy cost. Imagine, fo example, a sudden switching of the
interaction Hamiltonian

H
I
=
1
2
X
z
. Since in the equilibrium state of the bath

X = 0, the work
done for the realization of this switching is zero.
II.4.1 The model 65
where the average ...) is taken over the initial state (4.17). It can be written as
K
J
(t) = K(t) i

G(t), (4.21)
where
K(t t
t
) = 'K
J
(t t
t
) =
1
2
(t) (t
t
) + (t
t
) (t))
=
k
g
2
k
coth
_
k
2
_
cos
k
(t t
t
),
(4.22)
is the symmetrized correlation function.
Since (t) is a gaussian random operator, one can use Wicks theorem for de-
composing higher-order products: Any correlation of an odd number of s vanishes.
A correlation of an even number of s is equal to the sum of products of pair
correlations, the sum being taken over all pairings. For example:
T (t
1
) (t
2
) (t
3
) (t
4
)) = T (t
1
) (t
2
))T (t
3
) (t
4
))
+T (t
1
) (t
3
))T (t
2
) (t
4
)) +T (t
1
) (t
4
))T (t
2
) (t
3
)).
Note that the similar Wick-decomposition of T (t
1
)... (t
2k
)) will be a sum of (2k
1)!! = (2k 1)(2k 3)...3 terms.
Due to the factorized structure (4.17) of the initial state, the common averages
of and various spin operators can be taken independently. For example, averaging
Eq. (4.12) and using Wicks theorem together with k! 2
k
(2k 1)!! = (2k)! one gets:

(t)) = e
itiG
1
(t)
_

(0, t)
_

(0)) = e
it(t)

(0)), (4.23)
where for t
2
t
1
:
_

(t
1
, t
2
)
_
=
k=0
(1)
k
(2k)!
_
t
2
t
1
...
_
t
2
t
1
ds
1
...ds
2k
T [ (s
1
)... (s
2k
) ])
= exp
_
1
2
_
t
2
t
1
_
t
2
t
1
ds
1
ds
2
K
J
(s
1
s
2
)
_
= exp[(t
2
t
1
) +i G
1
(t
2
t
1
)],
and where
(t) =
1
2
_
t
0
_
t
0
ds
1
ds
2
K(s
1
s
2
) =
_
t
0
ds
1
_
s
1
0
ds
2
K(s
2
). (4.24)
As seen from (4.23), (t) characterizes the decay of
due to the interaction

with the bath.
4.1.3 Correlated initial conditions
For some applications it is more sensible to use the correlated initial conditions for
the spin and the bath:
(0) =
1
Z
exp
_
S

H
S
(

H
I
+

H
B
)
_
, Z = tr e
S

H
S
(

H
I
+

H
B
)
, (4.25)
where
S
is the inverse temperature of the spin and is that of the bath. The
temperature of the spin is dened by Eq. (4.27) below or analogously by

z
) = tanh
_
2
_
. (4.26)
This dierence between the temperature of the bath and that of the spin will be used
as an initial out of equilibrium situation for the work extraction in chapter 5. For
cooling purposes, chapter 6, this is not needed and both temperatures can initially
be the same.
This initial condition with
S
,= can be generated from the overall equilibrium
equal-temperature state of the spin and the bath via cooling or heating the bath by
means of some superbath. During this process
z
is conserved, and the bath relaxes
to its new temperature under an external eld
1
2
X generated by the interaction

Hamiltonian

H
I
with
z
= 1. More details of this procedure are given in Appendix
4.A.
In the thermodynamical limit for the bath the correlated initial condition (4.25)
is equivalent to the factorized condition (4.17) with
S
(0) =
1
tr e
S

H
S
e
S

H
S
, (4.27)
that is, when starting from the factorized initial condition (4.17, 4.27), the dynamics
of the system builds up a correlated state which at times t much longer than the
response time of the bath, t 1/, is equivalent to (4.25). By saying equivalent we
mean that the initial conditions (4.17, 4.27) and (4.25) produce the same values for
spin observables and for collective observables of the bath (i.e., the ones involving
summation over all bath oscillators).
A formal reason for this can be seen as follows. One can write the full Hamiltonian
H dened in (3.1) as
H =
k
_
a
k
+
g
k

z
2
k
__
a
k
+
g
k

z
2
k
_
+

2

z
k
g
2
k
4
k
, (4.28)
and diagonalize it via a unitary operator:
U = exp
_
k
g
k

z
2
k
( a
k
a
k
)
_
,

U a
k

U
= a
k
g
k

z
2
k
,

U
z

U
=
z
. (4.29)
II.4.2 Pulsed Dynamics 67
Thus the operators
b
k
= a
k
+
g
k

z
2
k
, [
b
k
,
l
] =
kl
(4.30)
are distributed over the initial state (4.25) independently of the operators of
spin. Moreover, as follows from (4.28, 4.25), the operators

b
k
have on the state
(4.25) exactly the same statistics (i.e., the same correlators) as the corresponding
operators a
k
on the factorized state (4.17).
Now note that for the initial condition (4.25),
z
(0) and the quantum noise
operator (t) are in general not independent variables, in contrast to the case of
the factorized initial condition (4.17, 4.27). However, for t 1/ they do become
independent:
(t) =
b
(t) +
z
(G(t) G),
b
(t)
k
g
k
[
k
(0)e
i
k
t
+
b
k
(0)e
i
k
t
], (4.31)
where
G
k
g
2
k
k
(4.32)
is the limit of G(t) for t 1/. Taking the latter limit in (4.31), one sees that
(t) becomes equal to
b
(t) and thus independent of
z
. Recalling that
b
(t) has in
the state (4.25) the same statistics as (t) in the factorized state (4.17), nishes the
argument.
Note that the thermodynamical limit for the bath is essential for this conclusion.
Otherwise, G(t) will be a nite sum of cosines, and will not converge to G.
In the following we will use the factorized initial condition Eq. (4.17) since it is
technically simpler. The time limit t will be taken before any perturbation acts
on the system to ensure the equivalence with correlated initial conditions Eq. (4.25).
This is the so-called ergodic limit. In a more general model, the one considering T
1
,
this limit is in fact t 1/ and t T
1
4.2 Pulsed Dynamics
The external elds acting on the spin are described by a time-dependent Hamiltonian
H
F
(t) =
1
2
k=x,y,z
h
k
(t)
k
, (4.33)
with magnitudes h
k
(t), which is to be added to

H dened in (3.1) such that the
overall Hamiltonian is time-dependent:
H(t) =

H +

H
F
(t). (4.34)
Eq. (4.33) represents the most general external eld acting on the spin. We shall
concentrate on pulsed regime of external elds which is well known in NMR and ESR
physics [Sie71, Sli90, EBW87, AG82, Abr61, SS66, Hah50, Wau92]. For example, it
was used to describe spin-echo phenomena [Hah50, Wau92] or processes of switching
o undesired interactions, such as those causing decoherence [Sli90, EBW87, VL98,
SL04].
A pulse of duration is dened by sudden switching on the external elds at some
time t > 0, and then suddenly switching them o at time t + . It is well-known
that during a sudden switching the density matrix does not change [LL87], while the
Hamiltonian gets a nite change. Let us for the moment keep arbitrary the concrete
form of external elds in the interval (t, t +). The Schrodinger evolution operator
of the spin+bath from time zero till some time t +, > , reads:
T exp
_
_
t+
0
ds

H(s)
_
= e
i(t+t)

H/
T exp
_
_
t+
t
ds

H(s)
_
e
it

H/
(4.35)
= e
i

H/

U
P
(t) e
it

H/
. (4.36)
The LHS of Eq. (4.35) contains the full time-dependent Schrodinger-representation
Hamiltonian

H(s), while in the RHS of this equation we took into account that the
actual time-dependence is present only between t and t +. The terms e
it

H/
and
e
i(t+t)

H/
stand for the free (unpulsed) evolution in time-intervals (0, t) and
(t +, t +), respectively. In Eq. (4.36) we denoted
U
P
(t) e
i

H/
T exp
_
_
t+
t
ds

H(s)
_
(4.37)
= T exp
_
_
t+
t
ds e
i(st)

H/

H
F
(s) e
i(ts)

H/
_
(4.38)
= T exp
_
_

0
ds e
is

H/

H
F
(s +t) e
is

H/
_
, (4.39)
for the pulse evolution operator. The transition from (4.37) to (4.38) can be made by
recalling

H(t) =

H +

H
F
(t) and then by noting that the expressions in these equa-
tions satisfy the same rst-order dierential equation in with the same boundary
II.4.2 Pulsed Dynamics 69
condition at = 0, implying that they coincide:
e
i

H/
T exp
_
_
t+
t
ds

H(s)
_
=
i
H e
i

H/

H(t +) e
i

H/
_
e
i

H/
T exp
_
_
t+
t
ds

H(s)
_
=
i
e
i

H/

H
F
(t +) e
i

H/
e
i

H/
T exp
_
_
t+
t
ds

H(s)
_
,
(4.40)
T exp
_
_

0
ds e
is

H/

H
F
(s +t) e
is

H/
_
=
e
i

H/

H
F
(t +) e
i

H/
T exp
_
_

0
ds e
is

H/

H
F
(s +t) e
is

H/
_
.
(4.41)
We focus on pulses so short that the inuence of the spin Hamiltonian
z
/2
and the interaction Hamiltonian

H
I
can be neglected during the interval , that is,
in a Taylor-expansion we may keep only the leading term,
e
is

H/

H
F
(s +t) e
is

H/
=

H
F
(s +t) +
is
[

H,

H
F
(s +t)] +...
=

H
F
(s +t) +
is
2

z
+

H
I
,

H
F
(s +t)
_
+...

H
F
(s +t), (4.42)
for 0 < s < . Thus, for the pulse evolution operator one gets
U
P
(t) = T exp
_
_

0
ds

H
F
(s +t)
_
. (4.43)
The generalization of the evolution operator (4.36) to an arbitrary number of short
pulses is straightforward.
Note that in obtaining (4.43) we do not require that the bath Hamiltonian

H
B
during the pulse is neglected. Since the external elds are acting on the spin only,
the inuence of the bath Hamiltonian disappears automatically from e
is

H/

H
F
(s +
t) e
is

H/
once the interaction Hamiltonian

H
I
has been neglected.
Recalling the orders of magnitude and of the spin energy and the inter-
action energy, respectively, in particular, recall Eq. (4.7), G =
_
0
d J()/
and Eq. (4.55), one gets the following qualitative criteria for the validity of the
short pulsing regime
min
_
1
, []
1
_
. (4.44)
As should be, for very small and a xed , the second restriction on is weaker
than the rst one. More quantitative conditions for the validity of the pulsed regime
were studied recently in the context of decoherence suppression by external pulses
[SL04].
To deal with the pulsed dynamics in the Heisenberg representation, one intro-
duces the following superoperators:
c
t

A e
i

Ht/

Ae
i

Ht/
, (4.45)
T
t

A

U
P
(t)

A

U
P
(t) (4.46)
Then the Heisenberg evolution of an operator

A corresponding to Eqs. (4.36, 4.43)
reads
A(t +) = c
t
T
t
c

A. (4.47)
4.2.1 Parametrization of pulses
As seen from (4.33, 4.46), any pulse corresponds to the most general unitary oper-
ation in the Hilbert space of the spin, which would correspond to a rotation in the
classical language since SU(2)=O(3). It turns out to be convenient to parametrize
pulses as:
T
a
e
i

H
F
/

a
e
i

H
F
/
=
b=, z
c
a,b

b
, a = , z. (4.48)
For more detailed applications we will need the explicit for of e
i

H
F
/
as a 2
2 unitary matrix whose determinant can be taken equal to unity without loss of
generality:
e
i

H
F
/
=
_
e
i
cos e
i
sin
e
i
sin e
i
cos
_
, (4.49)
where
0 , 2, 0

2
. (4.50)
4.3 Setup of pulsing
The spin and the bath are prepared in the state (4.17), with the initial state
of the spin having the denite temperature T
S
as given by (4.27). Thus, for
the initial average population dierence one has:

z
) = tanh
_
2
_
< 0. (4.51)
II.4.4 Realizations of the model 71
One waits for a time t which for the present model with the Hamiltonian
(3.1) can be arbitrarily large, but in reality should be much smaller than the
T
1
time-scale. This waiting is done for ensuring the robustness of the setup.
For our purposes it will suce to take t 1/, this imposes T
1
1/, a
restriction much weaker than the ones already discussed to the applicability
of the model and then the setup does not depend on details of the initial
preparation, since the initial conditions (4.17) and (4.25) are equivalent, see
section 4.1.3.
We shall restrict ourselves with two-pulse or three-pulse setup. In the rst
case the pulses
T
t
= T
1
, T
t+
= T
2
, (4.52)
are applied at times t and t +, respectively.
In the second case the pulses
T
t
= T
1
, T
t+
1
= T
2
, T
t+
1
+
2
= T
3
, (4.53)
are applied at times t, t +
1
and t +
1
+
2
, respectively.
4.4 Realizations of the model
Once the model with all its ingredients has been dened, we discuss some of its
realizations and provide some numbers. A two-level system coupled to a thermal
bath is a standard model for practically all elds where quantum systems are studied:
NMR, ESR, quantum optics, spintronics, Josephson junctions, etc. Two particular
conditions are however necessary to apply this model: the condition T
1
T
2
on
the characteristic relaxation times and availability of suciently strong pulses. On
the other hand, we can admit rather short times T
2
, since this timescale can be
overcome with the spin-echo technique.
Here are experimentally realized examples of two-level systems, which have suf-
ciently long T
2
times, satisfy to T
1
T
2
, e.g. T
1
exceeds T
2
by several orders of
magnitude, and admit strong pulses of external elds. For atoms in optical traps,
where T
2
1s, 1/ 10
8
s, there are ecient methods for creating non-equilibrium
initial states and for manipulating atoms by external laser pulses [CZ95]. For an
electronic spin injected or optically excited in a semiconductor, T
2
1 s [KA00],
and for an exciton created in a quantum dot T
2
10
9
s [B
+
98]; in both situations
1/ 10
9
10
13
s, and femtosecond (10
15
s) laser pulses are available. In the
case of NMR physics T
2
10
6
10
3
s, 1/ 1 s, and the duration of pulses can
vary between 1ps and 1 s [Sli90, EBW87, A
+
03].
In all above examples the response time 1/ of the bath is much shorter than
the internal time 1/ of the spin. Sometimes it is argued that such a separation
is related to the large size of the bath and is something generic by itself. This is
clearly incorrect, since as seen from the derivation in section 4.1, the dimensionless
parameter / has to do with the bath-spin interaction, rather than with the size
of the bath. Moreover, several examples of the interaction of the spin with the bath
are known and were analyzed both experimentally and theoretically, where / 1.
For example, Ref. [DSDA77] focusses on relaxation of nuclear spins with hyperne
frequencies 700 MHz, T
2
< 90 MHz, and the ratio / may vary between 10
and 0.1.
Another important parameter that characterizes our setup is the initial polar-
ization [
z
)[ of the spin. It is known in NMR and ESR physics that the response
of magnetic atoms (nucleus) to external dc magnetic eld is best characterized by
the
frequency
eld
ratio, which for electronic magnetic moments can reach few MHZ per
gauss of applied magnetic eld, while for nuclear magnetic moments it is 2 10
3
times smaller due to the dierence between atomic and nuclear Bohr magnetons.
For example, one has
frequency
eld
= 4.2 kHz/gauss for a single proton.
4.4.1 Why do we insist on the exact solvability of the model
The model as stated above that is, with the Hamiltonian presented in chapter
3, Eq.(3.1) together with Eq.(4.34) is exactly solvable for all temperatures and
all bath-spin coupling constants. It is useful at this point to recall for the reader
what are the specic reasons to insist on this feature. The model with interaction
Hamiltonian Eq.(3.4) is a particular case of a more general spin-boson model, where
the inuence of the T
1
-time is retained either via an additional term
x
in the
Hamiltonian of the spin, or via an additional coupling in the interaction Hamil-
tonian. This model is in general not solvable, and, what is worse, there are no
realiable approximate methods which apply for a xed (may be weak) coupling to
the bath and all temperatures including very low ones. The standard weak cou-
pling theories both markovian leading to well-known Bloch equations, and non-
markovian ones are satisfactory only for suciently high temperatures, while at
low-temperatures weak-coupling series are singular, and dierent methods of their
resummation produce dierent results. This situation becomes even more problem-
atic under driving by external elds. The objects studied by us such as work, the
energy of the spin can be rather fragile to various not very well-controlled approx-
imations, since there is a general limitation governing their behavior: Thomsons
formulation of the second law. It is derived from the rst principles of quantum me-
chanics [Thi83, BK77, BK79, Bas78, PW78, Len78, AN02] and has to be respected
in any particular model.
4.5 The spectral density
The coupling with the bath can be parametrized via the spectral density J():
J() =
k
g
2
k
(
k
). (4.54)
In the thermodynamical limit the number of bath oscillators goes to innity,
II.4.5 The spectral density 73
and J() becomes a smooth function, whose form is determined by the underlying
physics of the system-bath interaction.
4.5.1 Ohmic spectrum of the bath
We shall be mainly working with the ohmic spectrum:
J() = e
/
. (4.55)
where is a dimensionless coupling constant, and where is the maximal character-
istic frequency of the baths response. This spectrum and its relevance for describing
open quantum systems was extensively discussed in literature; see, e.g. [L
+
87].
Quantum noise correlation function and decay times
The correlation function of the quantum noise in the ohmic case, using Eqs.(4.22,
4.54, 4.55), is given by :
K(t) =
_

0
d J() coth
_
2T
_
cos t (4.56)
=
_

0
d coth
_
2T
_
e
/
cos t. (4.57)
Recall that the decay factor (t) is related to K(t) via Eq. (4.24):

(t) = K(t).
Properties of these functions are worked out in Appendix 4.B. In particular, for (t)
one gets from Eqs. (4.B.13, 4.B.14) the following explicit expression:
(t) = ln
_

2
_
1 +
T
_
1 +
2
t
2
_
1 +
T
i
Tt
_
1 +
T
+i
Tt
_
_
, (4.58)
where is Eulers Gamma function. It is seen that the temperature enters via the
dimensionless parameter T/().
Let us now determine the behavior of this quantity for low and large tempera-
tures. Using Eq. (4.B.7) one obtains for /T 1 (low temperatures):
(t) = ln
_
t
sinh
_
t
__
+

2
ln
_
1 +
2
t
2
. (4.59)
This implies two regimes of decay: power-law and exponential,
t : e
(t)
= (1 +
2
t
2
)
/2
, (4.60)
t : e
(t)
= e
t/J
2
, T
2
=

T
. (4.61)
For /T 1 (high temperatures) one uses Eq.(4.B.12) to get
(t) =
2T
_
t arctan(t)
1
2
ln(1 +
2
t
2
)
_
. (4.62)
This time the possible regimes of decay can be approximated as gaussian and expo-
nential.
t 1/ : e
(t)
e
t
2
/J
2
2
, T
2
=

2
T
, (4.63)
t 1/ : e
(t)
= e
t/J
2
, T
2
=

2T
(4.64)
In this latter case, as seen from Eq. (4.B.12), K(t) behaves approximately as a delta-
function: K(t)
2T
(1+t
2
2
)
with strength 2T/ determined by the parameters
and T. Note that in all the above cases the characteristic decay times become shorter
upon increasing the temperature T or the coupling constant , as expected.
The G-factor
Finally we will indicate the form of the function G(t) in the ohmic case (see Eq.(4.8))
which characterizes the reaction of the spin to the bath. As seen below, this function
and its primitive are rather important for our purposes. Using
_

0
d e
/
cos(t) =
1
1 +
2
t
2
, (4.65)
one gets for the ohmic case:
G(t) =
_
1
1
1 +
2
t
2
_
. (4.66)
It is seen that G(t) becomes equal to a constant on the characteristic time 1/.
It is therefore justied to call the latter the response time of the bath. For future
references note as well that the function F dened in Eq.(4.14) takes the form
F(t) = [t arctan(t)] . (4.67)
4.5.2 Other types of environment
The main characteristic of the spectral density function is its behaviour at low
frequencies. That is characterized by the power dependence on . Generalizing
Eq. (4.55)
J() =
e
/
. (4.68)
This equation brings 3 qualitatively dierent regimes:
> 1 Superohmic situation: Underdamped case
= 1 Ohmic situation: Critical case
II.4.5 The spectral density 75
< 1 Subohmic situation: Overdamped case
where the overdamped, critical, underdamped language comes from the damped
harmonic oscillator jargon. Throughout this thesis we will only consider the ohmic
case since it is the most relevant situation and brings the most interesting physics.
When dealing with the cooling experiment an example of subohmic spectral density
function, the so called 1/f noise, will be briey presented and studied. In 1/f
noise one has = 1/2. This interaction has been found lately to be the relevant
interaction, for instance, to the background charges in a Cooper pair box [N
+
02].
Though the superohmic situation is also relevant, for instance = 3 or 5 for
a defect in a three dimensional solid interacting with acoustic phonons, it will not
be considered here since it does not bring any new qualitative phenomena for the
purposes studied here.
Appendix 4.A
Correlated versus factorized initial conditions
Here we shortly outline how the two-temperature state (4.25) can be prepared
starting from the overall equlibrium state
(0) =
1
Z
exp
_
S

H
S
S
(

H
I
+

H
B
)
_
, Z = tr e
S

H
S
S
(

H
I
+

H
B
)
, (4.A.1)
which has equal temperatures of the spin and the bath.
Assume that the bath was subjected to another much larger thermal bath (su-
perbath) at temperature T dierent from T
S
, so that the total Hamiltonian of the
spin, bath and superbath reads:
H
total
=

H +

H
sup
, (4.A.2)
where the latter operator

H
sup
characterizes the weak interaction of the bath with
the superbath and contains also self-Hamiltonian of the superbath. Thus
[

H
S
,

H
sup
] = 0. (4.A.3)
Now the statement of this appendix is that under the action of the superbath at
temperature T, the common state of the spin and the bath will relax to the state Eq.
(4.A.11 or 4.25) with dierent temperatures for the spin and the bath. The reason
is that due to Eq. (4.A.3),
z
is conserved during the whole evolution generated by
the superbath, so that
z
does not relax and keeps its value given by Eq. (4.A.1). In
contrast, the variables of the bath including

X do not have such a protection and
they relax under inuence of the superbath. Let us now substantiate this statement.
Since [

H
S
,

H] = 0, the initial equilibrium state (0) of the spin and the bath can
be represented as
(0) =
j=1
p
jj

jj
(0) [j)j[, (4.A.4)
where
p
jj
=
e
j
S
/2
2 cosh(
S
/2)
, j = 1, (4.A.5)
are probabilities for the spin to be up (j = +1) or down (j = 1) respectively, [j)
is the eigenstate of

H
S
=

2

z
with eigenvalue j = 1, and where
jj
(0) =
1
Z
j
exp
_
S
_
j
2

X +

H
B
__
, Z
j
= tr
B
e
S
(
j
2

X+

H
B
)
, j = 1, (4.A.6)
are conditional states of the bath.
The total initial state of the spin, bath and superbath thus reads:
total
(0) =
j=1
p
jj

sup
(0)
jj
(0) [j)j[, (4.A.7)
II.4.A Correlated versus factorized initial conditions 77
where
sup
(0) is the initial equilibrium state of the superbath. Note that due to
weak coupling between the bath and superbath, their initial states can be assumed
to be factorized.
As follows from (4.A.3, 4.A.4), the time-dependent state of the total system
consisting of spin, bath and superbath can be presented as
total
(t) =
j=1
p
jj

jj
(t) [j)j[, (4.A.8)
where
jj
(t) the conditional state of the bath and superbath satises the von
Neumann equation
i

jj
= [
j
2

X +

H
B
+

H
sup
,
jj
], (4.A.9)
with the initial condition:
jj
(0) =
sup
(0)
jj
(0). (4.A.10)
Thus,
jj
moves according to the Hamiltonian
j
2
X +

H
B
+

H
sup
. It is now
clear that in the weak coupling limit of the bath-superbath interaction the marginal
conditional state tr
sup
jj
(t) will for suciently long times t relax to Gibbs dis-
tribution at temperature T (equal to the one of the superbath) and with Hamiltonian
j
2
X +

H
B
. Thus the (unconditional) marginal state of the spin and the bath will
indeed relax to
exp
_
S

H
S
(

H
I
+

H
B
)
_
, (4.A.11)
as was claimed before.
Appendix 4.B
Quantum noise generated by ohmic bath
Here we discuss properties of the function:
K(t) =
_

0
d coth(/2) e
/
cos t. (4.B.1)
In the given integration domain one can use
coth(/2) = 1 + 2
1
e
1
= 1 + 2
n=1
e
n
(4.B.2)
and can get from (4.B.1):
K(t) =
2
1
2
t
2
(1 +
2
t
2
)
2
+ 2
n=1
(
1
+ n)
2
t
2
((
1
+ n)
2
+t
2
)
2
(4.B.3)
With help of a standard relation:
n=1
1
t
2
+y
2
(n +)
2
=
i
2ty
_
_
1 + i
t
y
_
_
1 + +i
t
y
__
, (4.B.4)
where (z) =
t
(z)/(z), one obtains
n=1
( +n)
2
t
2
(( +n)
2
+t
2
)
2
=
1
2
[
t
(1 + it) +
t
(1 + +it)]. (4.B.5)
Combining (4.B.5) with (4.B.3) and = 1/() one ends up with the following
formula
K(t) =
2
1
2
t
2
(1 +
2
t
2
)
2
+
T
2
2
_
t
_
1 +
1
i
t
_
+
t
_
1 +
1
+i
t
__ (4.B.6)
Let us now consider separately the cases of low and high temperatures. For 1
one uses the known relation
_
1 i
t
_
1 +i
t
_
=
t
1
sinh[t/()]
(4.B.7)
and obtains from (4.B.6):
K(t) =
2
1
2
t
2
(1 +
2
t
2
)
2
+

t
2

T
2
2
1
sinh
2
[t/()]
. (4.B.8)
II.4.B Quantum noise generated by ohmic bath 79
For small t (t 1/) K(t) is positive as it should be:
K(t) =
2
+
T
2
2
3
2
. (4.B.9)
In contrast, for t 1/ it becomes negative, namely the noise is anticorrelated:
K(t) = 3
1
2
t
4

T
2
2
1
sinh
2
[t/()]
. (4.B.10)
At the end it is again correlated in the limit of very large times t where the
rst term in the r.h.s. of Eq. (4.B.10) does dominate (this domain is small for low
temperatures).
In the high-temperature limit 1 one can use in Eq. (4.B.6) the Stirling
formula:
t
(z) =
1
z
+
1
2z
2
+..., z 1 (4.B.11)
and then the quasiclassical limit for the quantum noise reads (after some more
simplications):
K(t) =
2
1
2
t
2
(1 +
2
t
2
)
2
+
2T
1
1 +t
2
2
. (4.B.12)
In the purely classical limit the rst term in the r.h.s. can be neglected and we
return (for t 1) to the classical white noise with the strength 2T (recall that
what enters in the expressions is K(t)).
Finally in the context of Eq. (4.B.6) we notice the following useful relations:
(t) =
_
t
0
dt
t
K(t
t
) =
t
2
1 +
2
t
2
+
iT
_
1 +
1
i
t
_
1 +
1
+i
t
__
,
(4.B.13)
(t) =
_
t
0
dt
t
_
t
0
dt
tt
K(t
tt
)
= ln
_
_
_
1 +
1
i
t
_
1 +
1
+i
t
2
_
1 +
1
1 +t
2
2
_
_
,
(4.B.14)
which are used in the main text.
5
Work extraction
In this chapter we show how one can use the setup of the previous chapter to extract
work from the out of equilibrium situation dened in Eq.(4.25).
5.1 Denition of work
The action of external elds on the system is connected with ow of work. The work
done in the time-interval (0, t) is standardly dened as the increase of the average
overall energy of the spin and bath dened by the explicitly, i.e. in the Schrodinger
picture, time-dependent Hamiltonian

H(t) [LL87, Kei87, Bal92]:
W(0, t) = tr[ (t)

H(t) ] tr[ (0)

H(0) ]. (5.1)
Due to the conservation of energy of the entire system (spin+bath), work is
equal to the energy given by the corresponding work-source. Since external elds
are acting only on the spin, there is a dierential formula for the work which uses
only quantities referring to the local state of the spin and which thus illustrates that
the work-sources exchange energy only through the spin:
dW
dt
= tr
_
S
(t)

H
F
(t)
t
_
, (5.2)
where

H
F
(t) as dened by (4.33) is the contribution of the external elds into the
spin Hamiltonian, and where
S
(t) is the density matrix of the spin. Eqs. (5.1, 5.2)
are related with each other by the von Neumann equation of motion
=
i
[

H(t), (t)] =
i
[

H +

H
F
(t), (t)] (5.3)
for the common density matrix (t) of the spin and the bath.
In order to get (5.1) from (5.2), note that the external elds are acting only on
the spin and that
t

H
F
(t) =
t

H(t). Then in expression Eq. (5.2), we can change
82 II.5. Work extraction
the reduced density matrix
S
for the full density matrix since the only time
dependence of the Hamiltonian lives in the Hilbert space of the spin. Then Eq. (5.2)
can be written as
dW
dt
= tr
_
(t)

H
F
(t)
t
_
= tr
_
(t)

H(t)
t
_
. (5.4)
Now integrate this expression from 0 to :
_

0
dt
dW
dt
=W(0, ) =
_

0
dt tr
_
(t)

H(t)
t
_
= tr
_
()

H()
_
tr
_
(0)

H(0)
_
_

0
dt tr
_
(t)

H(t)
_
.
Note that the last integrand is equal to zero due to the equation of motion (5.3).
More specically, we are interested in the work due to a pulse. For the example
of a single pulse at time t this quantity reads from (4.36, 4.47, 5.1):
W(0, t +) = W(t, t +) = tr[ ((t +) (t))

H] = tr((0) [T
t

H

H] ). (5.5)
This expression is directly generalized to several successive pulses: assume that
the pulse T
t
at time t was followed by another pulse T
t+
at time t + with > 0.
The work done during the rst pulse is given by (5.5), while the work done during
the second pulse reads:
W(t+ +, t + + 2) = tr[ ((t + + 2) (t + +))

H]
= tr((t + +)[T
t+

H

H]) = tr((0) c
t
T
t
c
[T
t+

H

H]).
(5.6)
Summing this up with W(0, t +) one gets for the complete work for the 2 pulse
situation:
W(0, t + + 2) = tr((0) [c
t
T
t
c
T
t+

H

H]), (5.7)
just equal to the dierence between the nal and the initial energy, as it should be.
5.2 General restrictions on work extraction
At this point it is useful to recall that the setup of two systems having initially
dierent temperatures and interacting with a source of work allows to draw a num-
ber of general relations on work-extraction. Departing from the following general
assumptions:
1. Out of equilibrium initial conditions. The initial conditions at the moment
t = 0 are given by Eq. (4.25), where the bath and the spin have initially
dierent temperatures T and T
S
, respectively. Recall from discussion in section
4.1.3 that this initial condition is equivalent to the factorized one (4.17, 4.27).
We use the former one since it is more convenient when dealing with the general
restrictions on the work-extraction.
II.5.2 General restrictions on work extraction 83
2. Cyclic external elds. For the following derivation the Hamiltonian

H
F
(t) of
external elds acting on the spin is arbitrary. In particular, it need not be
composed by pulses. The only general assumption made on

H
F
(t) is that its
action is cyclic at some nal time t
f
:
H
F
(0) =

H
F
(t
f
) = 0. (5.8)
We can nd the following two relations (derived explicitly in Appendix 5.A):
W
_
1
T
T
S
_
H
S
, (5.9)
W
_
1
T
S
T
_
(H
I
+ H
B
), (5.10)
where

H
k
= tr
_
H
k
[ (t
f
) (0) ]
_
, k = S, I, B, (5.11)
are the changes of the corresponding average energies of the spin, bath and interac-
tion, with (t
f
) being the complete density matrix of the spin and bath at time ,
and where the total work reads:
W = H
S
+ H
I
+ H
B
. (5.12)
Then from Eqs.(5.9,5.10) we can draw the following conclusions:
If T
S
> T, then in order to extract work, W < 0, Eq.(5.9) implies that
H
S
< 0, H
I
+ H
B
> 0 : (5.13)
so the system looses energy, while the bath gains it and the amount of the
extracted work [W[ is then bounded from above by [H
S
[.
If T = T
S
both Eqs.(5.9, 5.10) produce: W 0, which is, in fact, the statement
of the second law in Thomsons formulation: no work can be extracted from
an equilibrium system by means of cyclic perturbations.
If T
S
< T inequalities in Eq.(5.13) are reversed: now work-extraction implies
that
H
S
> 0, H
I
+ H
B
< 0, (5.14)
and [W[ is then bounded from above by [H
I
+ H
B
[.
These conclusions are close to what one could have expected from the standard
(phenomenological) thermodynamical reasoning. However, it should be emphasized
that Eqs. (5.9, 5.10) were derived starting from rst principles (see Appendix 5.A),
and, moreover, their derivation is by no means restricted to a weak bath-spin cou-
pling, a restricted situation which need not be satised in practice.
5.2.1 Eciency of work-extraction
Another useful notion is the eciency of the work-extraction, which shows how eco-
nomically one uses for work-extraction the given non-equilibrium, two-temperature
resource [LL87, Kei87, Bal92]. The special importance of eciency comes from the
fact that in the standard thermodynamics it is bounded from above by Carnots
value, which is a system-independent, thus universal, quantity.
Though our system starts out of equilibrium due to dierent initial temperatures
of the spin and the bath, the notion of eciency should be studied for it anew, since
it does not automatically fall into the class of heat-engine models, as studied in
textbooks of thermodynamics and statistical physics [LL87, Kei87, Bal92]:
There is no working body which operates cyclically between two thermal baths.
With us cyclic processes are dened with respect to the work source.
The interaction between the systems having dierent temperatures in the
case discussed, here the spin and the bath is not weak.
We do not insist that our systems always stay very close to equilibrium. In
contrast, both during and immediately after the work-extraction process, the
spin is in a non-equilibrium state, which in general cannot be described in
terms of a time-dependent temperature.
However, in spite of all these dierences we can dene the notion of eciency
and this will be an equally useful characterization of the work-extraction process
[LL87, Kei87, Bal92].
Recall that external elds are acting exclusively on the spin variables and not on
those of the bath. This implies that when during work-extraction the source of work
receives energy [W[, this energy consists of a contribution coming directly from the
spin and of a part which comes to the work-source from the bath but through the
spin. In this context one can write the change of energy of the spin as
d
dt
tr
_
S
(t)

H
S
(t)
_
= tr
_ _
d
dt
S
(t)
_

H
S
(t)
_
+ tr
_
S
(t)
_

t
H
S
(t)
__
=
d
dt
Q+
d
dt
W,
(5.15)
where in our case the Hamiltonian of the spin reads from Eqs. (4.1, 4.33):
H
S
(t) =

2

z
+
1
2
k=x,y,z
h
k
(t)
k
, (5.16)
and the partial derivative in the middle expression is to stress that we employ the
Schrodinger representation and the time dependence comes only from external elds,
in particular
t

H
S
(t) =
t

H
F
(t). In the last equality of Eq. (5.15) we have used
Eq. (5.2) and then we could use this expression as a denition of heat (dQ). This
II.5.3 Work-extraction via two pulses 85
is indeed the case, one can see that Eq. (5.12) is just the statement of the rst law
of thermodynamics with
Q = (H
I
+ H
B
) , (5.17)
as found by integrating Eq. (5.15) from 0 to .
Note that in the above denition of heat, the average interaction energy is at-
tributed to the heat received from the bath although it by itself depends on the
variables of the spin; see Eq. (4.3). The reason for this asymmetry is clearly con-
tained in the very initial statement of the problem, where we quite in accordance
with the usual practice of statistical physics restricted the work source to act only
on the spin.
All this being said, one can now proceed for W < 0 (work-extraction) with the
usual denition of eciency as the ratio of the useful energy [W[ to the maximal
energy involved in the work-extraction:

[W[
max ( [H
S
[, [H
I
+ H
B
[ )
. (5.18)
For T
S
> T Eqs. (5.12, 5.13) and W < 0 imply [W[ = [H
S
[ [H
I
+ H
B
[,
and then (5.18) results in
=
[W[
[H
S
[
. (5.19)
Analogously, for T
S
< T we have
=
[W[
[H
I
+ H
B
[
=
[W[
[W[ +[H
S
[
, (5.20)
from [W[ = [H
I
+ H
B
[ [H
S
[.
It is now seen from Eqs. (5.9,5.10, 5.13, 5.14) that the eciency is always bounded
by the Carnot value:
1
min ( T, T
S
)
max ( T, T
S
)
. (5.21)
5.3 Work-extraction via two pulses
5.3.1 Formulas for work
The work done for the rst pulse reads (as dened in Eqs.(5.5)):
W
1
=

2
T
1

z

z
)
t
+

2
_
(T
1

z

z
)

X
_
t
, (5.22)
where
_
A
_
t
= tr[

A(t) ], (5.23)
for any operator

A. The value of W
1
is read o from Eqs.(4.6, 4.7, 4.17):
W
1
= ( 1 c
(1)
z,z
)
_
2
G

2

z
)
_
, (5.24)
with G as dened in Eq. (4.32), and where c
(1)
z,z
is the corresponding parametrization
coecient of the rst pulse as dened by (4.48).
It is seen from Eq. (4.51) that the work W
1
is always positive. This is in
agreement with the thermodynamical intuition: the second term in the RHS of Eq.
(5.24) is the contribution from the spin energy and it is positive, since the spin was
in equilibrium before the application of the rst pulse (remember that T
S
> 0).
Another positive term
1
2
( 1 c
(1)
z,z
) G in the RHS of Eq. (5.24) comes from the
interaction Hamiltonian, as the bath operators, and thus the bath Hamiltonian, are
not inuenced by this rst pulse. Again, it is intuitively expected that the interaction
Hamiltonian should make the average energy costs higher.
The work done for the second pulse reads analogously to Eq. (5.22):
W
2
=

2
T
2

z

z
)
t++
+

2
_
(T
2

z

z
)

X
_
t++
. (5.25)
For a detailed derivation of this expression see [ASNb]. The resulting total work
W = W
1
+W
2
is:
W =

2
_
1 c
(2)
z,z
c
(1)
z,z
_

z
) + e
()
'
_
c
(1)
+,z
c
(2)
z,+
e
i

e
i
2

z

z
_
_
(5.26a)
+
G
2
(1 c
(1)
z,z
) +

2
(1 c
(2)
z,z
)
_
G() +g
2
() c
(1)
z,z
_
(5.26b)
+e
()
'
_
c
(1)
+,z
c
(2)
z,+
e
i
_
i

()

e
i
2

z

z
_
+ g
2
()

e
i
2

z
_
__
. (5.26c)
The detailed explanation of various terms in this expression and of their physical
meaning is as follows.
The rst term in the RHS of Eq. (5.26a) is the contribution from the initial spin
energy. The second term comes from the transversal degrees of freedom excited by
the rst pulse. The factor e
()
accounts for the reduction of this terms in time
interval between pulses . Recall that the parametrization coecients c
(1,2)
+,z
and
c
(1,2)
z,z
for the rst and the second pulse are dened in Eq. (4.48):
The terms in Eqs. (5.26b, 5.26c) are the joint contribution from the bath Hamil-
tonian (3.3) and from the interaction Hamiltonian (4.3). The last of them couples
to the transversal degrees of the spin, as reected by the presence of e
()
. Recall
that the averages ...) in Eqs. (5.26a, 5.26c) refer to the initial state Eqs. (4.17,
4.27).
Finally, the factors
g
2
() GG(), (5.27)
2
(, t) =
_

0
ds [ G(s) G(t +s) ] = F(t) +F() F(t +). (5.28)
(the lower index 2 refers to the two-pulse situation) come from the backreaction of
the spin to the bath. For the ohmic case and in the limit t 1/, g
2
() and
2
()
read from Eqs. (4.66, 4.67)
g
2
() =

1 +
2
2
, (5.29)
2
() = arctan(). (5.30)
Next we note that the behavior of W = W
1
+ W
2
is controlled by a few dimen-
sionless parameters, see [ASNb] for a derivation. For the ohmic case it reads
W = W
1
+W
2
=

2
w
_
T
, ,

,
z
),
_
. (5.31)
Let us state them explicitly: i) The ratio
T
of the bath temperature to its

response energy; this can be viewed as the dimensionless temperature of the bath.
ii) The dimensionless coupling constant . iii) The ratio of the spin frequency
= / to the response frequency of the bath. iv) The initial average population
dierence
z
) dened in Eq.(4.51); the spin temperature T
S
enters only through
this parameter. v) The dimensionless time of the free evolution between the two
pulses.
There are two situations within the present setup, where work-extraction is not
prohibited: T > T
S
and T < T
S
. We deal with them in separate sections, since for
these cases the work-extraction eect exists in dierent ranges of the parameters.
5.3.2 Work extraction for T > T
S
It was seen above that the rst pulse always costs work, since it is applied on the
spin whose state is (initially) in local equilibrium at temperature T
S
. However, the
rst pulse can do more than simply wasting work. For example,

2
in y-direction:
T
1
= T
_
2
; y
_
, (5.32)
where
T (; y)
z
e
i
y
/2

z
e
i
y
/2
=
z
cos
x
sin , (5.33)
T (; y)
x
e
i
y
/2

x
e
i
y
/2
=
z
sin +
x
cos . (5.34)
excites the transversal component
x
. This degree of freedom starts to decay under
the action of the bath, and thus correlations between the spin and the bath are
established. If now a proper second pulse is applied at a time , for instance a
rotation

2
in the x-direction:
T
2
= T
_
2
; x
_
, (5.35)
0 0.2 0.4 0.6 0.8 1
-1
-0.5
0
0.5
1
1.5
w
Figure 5.1: Dimensionless total work w (see Eq. (5.31) in the text) versus dimen-
sionless time in between two pulses in the regime T > T
S
. The parameters are:
T
= 10, = 1,

= 0.01,
z
) = 0.8, 0.5, 0.4, 0.3 (from bottom to top).
The two pulses are given by Eqs. (5.32, 5.35) with: c
(1)
+,z
= 1, c
(2)
z,+
=
1
2i
, c
(1)
z,z
= 0
and c
(2)
z,z
= 0. Work-extraction (W < 0) disappears for larger
z
), that is, for closer
(initial) temperatures of the spin and the bath.
where
T (; x)
z
e
i
x
/2

z
e
i
x
/2
=
z
cos +
y
sin , (5.36)
T (; x)
y
e
i
x
/2

y
e
i
x
/2
=
z
sin +
y
cos , (5.37)
then not only some work can be extracted by the second pulse, but the overall work
by the two pulses can be negative:
W = W
1
+W
2
< 0, (5.38)
as seen in Figs. 5.1, 5.2.
Note that the time needed for work-extraction should be neither too short
otherwise the two pulses will eectively sum into one, and we know that no work-
extraction is achieved by a single pulse, nor too long, otherwise the transversal
degree of freedom excited by the rst pulse will decay, and we will have two isolated
single pulses. This is seen in Figs. 5.1, 5.2. Note that the choice of pulses is obviously
important for having work-extraction. Eqs. (5.32, 5.35) represent only one particular
example leading to work-extraction in the regime T > T
S
.
As for the magnitude of the extracted work, one notes from Eq. (5.31) and Fig. 5.1
that it is of order of /2, which is basically the response energy of the bath. This
is not occasional, since as seen from (5.14) the work in this regime T > T
S
is coming
from the bath.
Noting the ratio /() = 0.01 in Fig. 5.1 this and even smaller ratios are usual
for the realizations of the model as we discussed in section 4.4 we conclude that
the extracted work can be of several orders of magnitude larger than the energy
(or free energy) of the spin. On the other hand, the extracted work is limited
0 0.2 0.4 0.6 0.8 1
-0.5
0
1
2
3
w
2
Figure 5.2: The ratio
W
=
w
2
(see Eq. (5.31) in the text) versus dimensionless time
for two pulses in the regime T > T
S
T
= 10,

= 0.01,

z
) = 0.8 and = 4 (upper solid curve), = 2 (lower solid curve), = 0.5 (thick
curve), = 0.1 (dotted curve), The two pulses are given by Eqs. (5.32, 5.35) with:
c
(2)
z,+
=
1
2i
, c
(1)
+,z
= 1, c
(1)
z,z
= 0 and c
(2)
z,z
= 0. It is seen that the maximal extracted
work is a non-monotonous function of the dimensionless coupling constant .
by T which is the characteristic thermal energy available in the bath. In Fig. 5.1,
T/() = 10. Not unexpectedly, the possibility for work-extraction disappears when
the temperatures T and T
S
are close to each other; see Fig. 5.1.
5.3.3 Work extraction for T < T
S
0 1 2 3 4 5 6
-1
0
1
3
5
w
Figure 5.3: Dimensionless work w (see Eq. (5.31) in the text) versus dimensionless
time for two pulses in the regime T < T
S
in the case of = 0.1,
z
) = 0.01,
c
(1)
+,z
= i, c
(2)
z,+
=
1
2
, c
(1)
z,z
= 0 and c
(2)
z,z
= 0. Normal curve:
T
= 0.1,

= 3, Dashed
curve:
T
= 0.1,

= 2. Thick curve:
T
= 1,

= 3.
Let us now turn to scenarios of work-extraction in the regime T
S
> T. As
seen from (5.13), the specicity of this situation is that if there is work-extraction
at all, the work should come from the average energy dierence of the spin, while
the quantity H
I
+ H
B
is then necessarily positive. Since the latter quantity is
of order of (response energy of the bath), and the spins energy dierence is
obviously of order , there are two ways to try to achieve work-extraction, that is,
to get W = [H
I
+ H
B
[ [H
S
[ < 0: One should either take /() 1 or
take the dimensionless coupling constant very small. The second way did not lead
to work-extraction, since the required coupling constants are so small that the spin
eectively decouples from the bath. In contrast, the rst case with /() 1 led
to a sizable work-extraction, as see in Fig. 5.3. Recall in this context that systems
with /() = / 1 are well-known; see section 4.4 for details.
5.3.4 Eciency of work extraction
0 0.1 0.2 0.3 0.4
0.98
0.99
1
1.01
1.02
Figure 5.4: Eciency versus dimensionless time for two pulses in the regime
T > T
S
T
= 10, = 1,

= 0.01,
z
) = 0.8, c
(1)
+,z
= 1,
c
(2)
z,+
=
1
2i
, c
(1)
z,z
= 0 and c
(2)
z,z
= 0. The eciency is slightly below the corresponding
Carnot value and is maximizal at a value for coinciding basically with the maximal
dimensionless work w; see Fig. 5.1.
Let us now turn to the eciency of work-extraction as dened by Eqs. (5.18,
5.19, 5.20). To calculate it one needs to know the total work given by Eqs. (5.24,
5.26a, 5.26c), and the contribution H
S
to the work W coming from the average
energy of the spin, which is read from the RHS of Eq. (5.26a).
The eciency as a function of is presented by Figs. 5.4, 5.5 for T > T
S
and
T < T
S
, respectively. There are two important things to note. First, for T > T
S
the eciency can be very close to one, if the temperatures T and T
S
are suciently
separated from each other, which is the case in Fig. 5.4. It is, however, always limited
by Carnots value, as given by Eq. (5.21). For T < T
S
the eciency is sizable, but
is rather below the corresponding Carnot value. Second, the work and eciency are
maximal in almost simultaneously.
It should be recalled in this context that in the standard thermodynamics ef-
ciencies close to the optimal value are connected to very small work per unit of
II.5.4 Work-extraction via spin-echo pulses 91
2.9 3 3.1 3.2 3.3 3.4
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
Figure 5.5: Eciency (upper curve) and dimensionless work w (lower curve) versus
dimensionless time for two pulses in the regime T < T
S
T
= 0.1, = 0.1,

= 3,
z
) = 0.01, c
(1)
+,z
= i, c
(2)
z,+
=
1
2
, c
(1)
z,z
= 0 and c
(2)
z,z
= 0.
The eciency is quite below the corresponding Carnot value 0.99 and it is maximal
in almost simultaneously with the dimensionless work w.
time (zero power of work), since they are achievable for very slow processes. This is
not the case with the presented setup. As seen from Figs. 5.1, 5.2, 5.3, the work is
extracted on times which are of order of 1/ (response time of the bath), which is
typically much smaller than the internal characteristic time 1/ of the spin. Thus,
in Fig. 5.4 we have nearly optimal eciencies together with the maximal work and
a nite power of work.
5.4 Work-extraction via spin-echo pulses
So far we assumed that we deal either with a single spin coupled to the bath, or,
equivalently, with an ensemble of identical non-interacting spins each coupled with
its own bath. However, many experiments especially in NMR physics are done
on ensembles of non-interacting spins which are not identical. Their dierence lies in
the dierent energies . Moreover, these energies are randomly distributed, so that
the collective outcomes from such ensembles are obtained by averaging over =
the corresponding expressions for a single spin. This induces a T
2
relaxation as we
already discussed in section 4.1.
We shall assume that the distribution of is Gaussian with average
0
and
dispersion d:
P() =
1
2d
e
(
0
)
2
/(2d)
. (5.39)
It is now clear that the averaging over P() the oscillating terms e
i
will produce
a factor e
d
2
/2
that is, a strong decay on characteristic times
T
2
1/
d. (5.40)
For /T
2
1 all the terms containing e
i
will be zero after averaging, and
the corresponding averaged work for two pulses will always be positive as seen from
Eqs. (5.24, 5.26a, 5.26b, 5.26c). Indeed, all possible negative values of the full work
W were related to transversal degrees of freedom excited by the rst pulse. These
terms come with the factor e
i
which is connected to the free evolution in between
the two pulses. The decay of these terms in between pulses after /T
2
1 prevents
work extraction.
Let us now notice that quantum coherence is not lost on the timescale T
2
. To
extract work in the strongly-disordered situation where T
2
is short, we combine
our work-extraction setup with the spin-echo phenomenon [Hah50, Wau92]. For our
present purposes this amounts to applying a -pulse, for instance in x-direction:
T

z
=
z
, T

y
=
y
, T

x
=
x
, (5.41)
at the moment right in the middle of two pulses. Thus we have the setup described
by (4.53), the rst and the third pulses being at the moment arbitrary, while T
is
given by Eq. (5.41). The work done by the rst pulse reads from Eq. (5.24) after
averaging over P() given by Eq. (5.39):
W
1
= ( 1 c
(1)
z,z
)
_
2
GE
_
, (5.42)
where
E =
2
_
dP()tanh

S
2
< 0, (5.43)
is the average energy of the ensemble of spins. The work done by the second -pulse
is found by using Eqs. (5.25, 5.41), for details see [ASNb], with the result being
W
2
= G() + g
2
() c
(1)
z,z
2E c
(1)
z,z
, (5.44)
where g
2
() is dened in Eqs. (5.27, 5.29). It is seen that W
2
> 0 which is connected
to the fact that the -pulse does not couple properly with the transversal degrees of
freedom excited by the rst pulse.
Ultimately, the total work W = W
1
+W
2
+W
3
done by the three pulses together
is derived in Ref. [ASNb]
W =
G
2
_
1 +c
(2)
z,z
c
(1)
z,z
_
+ G() (2 c
(2)
z,z
c
(1)
z,z
)
G(2)
2
_
1 +c
(2)
z,z
c
(1)
z,z
c
(1)
z,z
c
(2)
z,z
_
(5.45a)
+e
4()+(2)
'
_
c
(1)
,z
c
(2)
z,+
_
[ 2

()

(2) ] [ sin
3
+i mcos
3
]
g
3
[cos
3
i msin
3
]
__ (5.45b)
E
_
1 +c
(2)
z,z
c
(1)
z,z
_
+
e
4()+(2)
'
_
c
(1)
,z
c
(2)
z,+
( 2 E cos
3
i
0
sin
3
)
_
,
(5.45c)
II.5.4 Work-extraction via spin-echo pulses 93
where
g
3
() = GG(2), (5.46)
3
() = 2F() F(2), (5.47)
are the backreaction factors for the considered setup of pulses, and where
m =
_
dP() tanh

S
2
< 0 (5.48)
is the average magnetization of the ensemble.
For the Ohmic case g
3
and
3
reduce to
g
3
() =

1 + 4
2
2
, (5.49)
3
() = [ arctan(2) 2 arctan() ] . (5.50)
As compared to Eqs. (5.26a, 5.26b, 5.26c), which present the work for two pulses,
Eqs. (5.45a, 5.45b, 5.45c) are dierent in several aspects:
1. There are no oscillating factors e
i
which after averaging over the distribution
P() would produce damping on times T
2
. This is due to the -pulse (5.41)
right in the middle of the two pulses. A simple explanation why the terms
e
i
are absent is as follows: Assume that the interaction with the bath is
absent and the spin moves under dynamics generated by the free Hamiltonian
H
S
=

2

z
. Denote by c
(0)
t
the corresponding Heisenberg evolution operator:
c
(0)
t

A = exp
_
it

H
S
_

A exp
_
it

H
S
_
. It is now seen with help of (5.41) that
the factor e
i
drops out:
c
(0)
T
2
c
(0)

+
= e
i
c
(0)
T
2

+
= e
i
c
(0)
= e
i
e
i

. (5.51)
2. The decay factor e
4()+(2)
in Eqs. (5.45b, 5.45c) is dierent from e
(2)
.
The last decay factor being the one generated by the free (unpulsed) evolution
during the time 2. Only in the exponential regime (t) t/T
2
the two
situations are equivalent: e
4()+(2)
e
(2)
. Recall that the exponential
regime is present for the ohmic spectrum at long times, see section 4.5.1.
For Gaussian decay however, (t) t
2
/T
2
2
, e
(2)
predicts sizable decay in
contrast to e
4()+(2)
1.
3. Now there are two independent parameters which characterize the initial state
of the ensemble of spins: E and m. The work W in Eqs. (5.45a, 5.45b, 5.45c)
can be expressed in the dimensionless form similar to Eq. (5.31):
W =

2
w
_
T
, ,

0
,
T
S
,
d
2
,
_
. (5.52)
It is now more convenient to account for the temperature of the spin via
T
S
and not via the average occupation number

z
) as before in Eq. (5.31), since
0 2 4 6 8 10
-2
-1
0
1
2
3
4
w
Figure 5.6: Dimensionless work for three spin-echo pulses with parameters:
T
=
10, 5, 1, 0.5 (from top to bottom),
T
S
= 10
3
,
d
2
= 10
2
, = 0.1,

0
= 8, c
(1)
,z
= i,
c
(2)
z,+
=
1
2
. Work-extraction is poor or disappears for smaller
d
2
or

0
, because there
appears to be to much random thermal energy in the ensemble.
there is a new dimensionless parameter
d
2
, which quanties the ratio of the
response time 1/ to T
2
= 1/
d.
Fig. 5.6 presents a scenario for work-extraction in the regime T
S
> T. It
is seen that the initial high-temperature ensemble of spins is strongly disordered:
d
2
= 10
2
1. Still this disorder cannot be much larger, since there will be too
much random energy in the ensemble, that is, the positive term E
_
1 +c
(2)
z,z
c
(1)
z,z
_
in Eq. (5.45c) will be too large and cannot be compensated by potential negative
terms. For analogous reasons there are no interesting scenarios of work-extraction for
strongly disordered ensemble in the regime T
S
< T: the average magnetization [m[
appears to be too low. Recall that in the regime T
S
> T there is a positive contribu-
tion to the total work coming from the bath, and sizable average frequencies

0
5
are needed to overcome this contribution, as seen from Fig. 5.6. This restriction on
the (average) frequency is similar to the one present in the two-pulse work-extraction
scenario for the non-disordered ensemble of spins in the regime T
S
> T.
5.5 Conclusion
This chapter describes several related scenarios of work-extraction based on the
spin-boson model dened in chapter 4: spin-
1
2
interacting with external sources of
work and coupled to a thermal bath of bosons. The work-sources act only on the
spin, since the bath is viewed as something out of any direct access. The work
is extracted from an initial local-equilibrium state of the spin at temperature T
S
,
not equal to the equilibrium state of the bath at temperature T. As we recalled
several times, the Thomson formulation of the second law prohibits work-extraction
via cyclic processes, thus includes any sequence of pulses, from an equilibrium state
of the entire system: T = T
S
[Thi83, BK77, BK79, Bas78, PW78, Len78, AN02].
II.5.5 Conclusion 95
In this spirit one would expect that work-extraction is also absent when external
elds are acting only on the spin in a local equilibrium state [Sli90, AG82]. We have
shown however, that this is not the case. It is possible to extract work in this latter
setup due to the common action of the following factors: i) backreaction of the spin
to the thermal bath; and ii) generation of coherences (i.e., transversal components
of the spin) during the work extraction process. It is also possible to extract work
from a disordered ensemble of spins having random frequencies. This ensemble can
even be strongly disordered in the sense that the relaxation time T
2
induced by the
disorder is the smallest characteristic time in the problem.
Two conditions are necessary for the existence of the work-extraction eect.
First, the transversal relaxation time T
2
has to be shorter than the longitudinal
relaxation time T
1
. This condition allows the notion of local equilibrium, because
once the transversal components decay after time T
2
, the spin can be described via
a temperature dierent from the one bath; after time T
1
the spin relaxes to the
temperature of the bath. This setup would not work for T
2
T
1
even though when
both are large because, then the coupling to the bath would be too weak and the spin
would not feel much of its interaction. Therefore it would be close to equilibrium,
which means after all, no work extraction. Second, the action of external elds
has to be suciently strong. We demonstrated the work-extraction eects for a
version of the spin-boson model which is exactly solvable thanks to pushing the above
two physical conditions to their extremes: the T
1
-relaxation time is taken innitely
large and the duration of pulses of external elds is taken innitely small. Both
these idealizations are well-known in NMR/ESR physics and related elds, and were
numerously applied and discussed in literature [Sli90, EBW87, AG82, VL98, SL04].
It may be of interest to see how precisely nite T
1
-times and nite pulsing-times
inuence the work-extraction eect; to rst order a perturbation expansion will
suce.
As to provide further perspectives on the obtained results, let us discuss them in
two related contexts, those of lasing without inversion and quantum heat engines.
5.5.1 Lasing without inversion
As we discussed in the introduction, besides the standard lasing eect, where work
is extracted from a spin having population inversion (i.e. having a negative tem-
perature), there are schemes of lasing which operate with a weaker form of non-
equilibrium since they employ three or higher-level atoms which are initially in a
state with non-zero coherence (i.e. non-zero o-diagonal elements of the density
matrix in the energy representation). There are numerous works both theoretical
and experimental, partially reviewed in [SZ97, Koc92], showing that in such systems
one can have various scenarios of lasing without inversion in the population of en-
ergy levels. In quantum optics lasers without inversion are expected to have several
advantages over the ones with inversion.
The eects described by us also qualify as lasing without inversion or more
precisely gain or work-extraction. There are, however, several important dierences
compared to the known mechanisms. First, we do not require coherence to be present
in the initial state. Our mechanism operates starting from initial local equilibrium
state of the spin, which by itself is stable with respect to decoherence (i.e., T
2
time-scale). It does employ coherence however, but it is generated in the course of
the work-extraction process, which, in particular, means that the energy costs of
its creation is included in the extracted work. Second, we do not need three-level
systems: the eect is seen already for two-level systems. Third, in our scheme the
work comes from the bath if its temperature is higher than the initial temperature
of the spin. Due to this, the extracted work can be much larger than the energy
change of the spin.
5.5.2 Quantum heat engine
The standard thermodynamic model of a heat engine is a system (working body)
operating cyclically between two thermal baths at dierent temperature and deliv-
ering work to an external source [LL87, Bal92]. The work produced during a cycle,
as well as the eciency of the production, depends on the details of the operation.
The upper bound on the eciency is given by Carnot expression, which is system-
independent (universal). This eciency is reached for the Carnot cycle during very
slow (slower than all the characteristic relaxation times) and therefore reversible
mode of operation [LL87, Bal92]. Though the Carnot cycle illustrates the best ef-
ciency attainable, it is rather poor as a model for a real engine, in particular due
to very long duration of its cycle: the work produced in a unit of time is very small
(zero power). This problem initiated the eld of nite-time thermodynamics which
studies how precisely the eciency is to be sacriced in order to reach a nite power.
In a similar spirit a number of researchers transferred these ideas into quantum
domain designing models for engines where the basic setup of the classic heat engine
is retained, while the working body operating between the baths is quantum [Ali79,
GK94, FK00, Scu01, HCH02, H
+
02, FK03, S
+
03].
Our setup for work-extraction can also be viewed as model for a quantum engine.
It is, however, of a nonstandard type since there is no working body operating
between two dierent-temperature systems (in our case these are the bosonic thermal
bath and the ensemble of spins). The two systems couple directly and the work-
source is acting on one of them. From the economical perspective, the absence of
a working body is an advantage. In spite of this dierence, the notion of eciency
can be dened along the standard lines, and it is equally useful as the standard
one; in particular, it is always bound from above by the Carnot value. We have
shown that the eciency can approach this value basically at the same time when
the extracted work approaches its maximum. Moreover, the whole process of work-
extraction takes a nite time of order of the response time of the bosonic bath, which
is actually much smaller than relaxation times of the spin. Thus, the three desired
objectives can be achieved simultaneously: maximal work, maximal eciency and a
large power of work.
II.5.A Derivation of Eqs. (5.9, 5.10) 97
Appendix 5.A
Derivation of Eqs. (5.9, 5.10)
Assume that the initial state of the spin and bath is:
(0) =
1
Z
exp
_
S

H
S
(

H
I
+

H
B
)
_
, Z = tr e
S

H
S
(

H
I
+

H
B
)
, (5.A.1)
with dierent temperatures for the spin and the bath.
An external eld

V (t) is acting on the system,
H(t) =

H +

V (t) (5.A.2)
such that it is zero both initially and at the moment t = :
V () =

V (0) = 0. (5.A.3)
This condition denes cyclic process. The total work which was done on this system
reads:
W = H
S
+ H
I
+ H
B
, (5.A.4)
where
H
k
= tr
_
H
k
[ () (0) ]
_
, k = S, I, B, (5.A.5)
are the changes of the corresponding energies, and where () is the density matrix
of the entire system (spin+bath) at time .
Recall that the relative entropy is dened as (see, e.g. [BP02]):
S[[[] tr( ln ln ) 0, (5.A.6)
is non-negative for all density matrices and . One now uses:
S[()[[(0)] = tr( () ln () () ln (0) ) = tr( (0) ln (0) () ln (0) )
=
S
H
S
+(H
I
+ H
B
) 0, (5.A.7)
where we used (5.A.1) and tr() ln () = tr(0) ln (0) is due to the unitarity of
the overall dynamics generated by the time-dependent Hamiltonian

H(t).
Combining (5.A.7) with (5.A.4) one gets Eqs.(5.9, 5.10)
W
_
1
T
T
S
_
H
S
, W
_
1
T
S
T
_
(H
I
+ H
B
). (5.A.8)
Finally note that would we use the separated initial condition
(0) =
S
(0)
B
(0) =
1
tr e
S

H
S
e
S

H
S
1
tr e

H
B
e

H
B
(5.A.9)
we would not be able to conclude from the above derivation that eciency is limited
by the Carnot value. Indeed, instead of Eqs. (5.A.8) one has, respectively:
S
H
S
+H
B
0, W H
I
+
_
1
T
T
S
_
H
S
. (5.A.10)
The latter inequality is not informative with respect to Carnots bound, since it
cannot and should not in general be excluded that H
I
is not small.
However, for the model studied in the present paper, the equivalence of the initial
conditions (5.A.1) and (5.A.9) is known from other places, see section 4.1.3.
Let us emphasize the main points by which the present derivation diers from
the standard textbook one:
No postulates were used: the whole derivation is based on the quantum-
mechanical equations of motion and certain assumptions on the initial con-
ditions.
It was not assumed that the interaction between the system and the bath is
small, which indeed need not be satised in reality.
The choice of using the initial condition in Eq. (5.A.1) is important in the
present derivation, though presumably Carnots bound is valid in certain more
general cases, such as, in our case, the factorized initial condition from Eq.
(5.A.9).
6
Cooling of spins
In this chapter we will use the formalism described in chapter 4 in order to achieve
cooling. Cooling amounts to decreasing the nal polarization
z
)
f
, i.e. make it
more negative, as compared to its initial value. In this case, as compared to the
work extraction situation, the initial state of the whole system is chosen to be in
equilibrium. Therefore, the initial state is described by Eqs. (4.17, 4.27) using the
same temperature for the spin and the bath
S
= . Then Eq. (4.51) becomes

z
)
i
= tanh
1
2
(6.1)
and we recall that
x
)
i
=
y
)
i
= 0.
6.1 Cooling via pulses
A single pulse cannot achieve cooling since it sees the initial local equilibrium state of
the spin, and then according to the no-cooling principle [KP92], statement related to
the Thomson formulation to the second law, it can only heat the spins state up: for
an arbitrary pulse T
1
applied at time t, one can write the ground state occupation
probability a time after the pulse, using Eq. (4.47) and Eq. (4.49), with some
algebra [ASNb, ASNa]

z
)
f
c
t
T
1
c

z
) = c
t
T
1

z
) =
z
)
i
cos 2
1

z
)
i
. (6.2)
Recall that
z
is a constant of motion when no pulse is acting on the spin and that

z
)
i
0 since we consider positive temperatures
S
, see Eq. (6.1). Thus we have
to employ at least two pulses to achieve any cooling.
The nal polarization after one pulse at t and one at t +, given by
P = [
z
)
f
[ = [c
t
T
1
c
T
2

z
)[, (6.3)
100 II.6. Cooling of spins
reads from Eqs. (4.12-4.14) and Eq. (4.49):

z
)
f
=
z
)
i
cos 2
1
cos 2
2
+s
2
sin 2
1
sin 2
2
, (6.4)
s
2
= e
()
'
_
e
i+i
2
(
z
)
i
cos +i sin )
_
, (6.5)
where = (0, t, ) in Eq. (4.16), and
2
=
1

2

1

2
from Eq. (4.49).
There are now two factors that come from the bath: e
()
in s
2
accounts for the
decoherence in the time-interval (t, t +) of transversal terms generated by the rst
pulse, while is the backreaction factor from Eqs. (4.15, 4.16). We send t
(more precisely, 1/) as explained in section 4.3, i.e. we wait long enough
before applying the rst pulse, since then the factorized initial conditions Eq. (4.17)
are equivalent to the correlated ones Eq. (4.25).
Note that the sign of s
2
can always be made negative by tuning
2
. Then,
minimizing
z
)
f
over
1
,
2
produces min [
z
)
i
, s
2
]. If the initial polarization is
already high, [
z
)
i
[ > [s
2
[, it is better to apply no pulses at all. However, in the
relevant situation
z
)
i
0, the minimum
z
)
i
= s
2
is reached for
1
=
2
=

4
.
All together, using Eq. (4.16) and , yields

z
)
f
= s
2
= e
()
sin
2
sin [ arctan()] , (6.6)
where the backreaction remains contained inside the last sine-function. This is, nat-
urally, the same coecient we found in the work extraction, see Eq. (5.29). Eq. (6.6)
is minimal for
2
=

2
. The choice of optimal pulses can be, e.g., a

2
-pulse along
the x-axis followed by a
_
2
_
-pulse along the y-axis:
T
1

z,x
=
y,x
, T
1

y
=
z
; (6.7)
T
2

z,y
=
x,y
, T
2

x
=
z
. (6.8)
The results are plotted in Fig. 6.1. One can see that max
[
z
)
f
[ decreases with
(weaker backreaction) and with = 1/k
B
T (larger decoherence).
0.2 0.4 0.6
0
0.2
0.6
0.8
P
Figure 6.1: Final polarization P = [
z
)
f
[ given by Eqs. (6.4, 6.5) versus dimen-
sionless time for two optimal pulses for the ohmic interaction. Bold curve: = 5,
= 0.2. Normal curve: = 2, = 0.5. Dashed curve: = 1, = 1.
z
)
i
and
are taken to be 0.
II.6.2 Cooling using spin-echo pulses 101
An obvious way to improve these results is to apply more complex pulse se-
quences. First, we consider three successive pulses and optimize over their parame-
ters. The analytic expression for
z
)
f
is too complex to present here, but numerical
results obtained using Mathematica 5 are presented in Table I. A third pulse leads
to a further improvement, especially when the two-pulse cooling is not very ecient.
We could continue by adding more pulses, but the analysis becomes dicult. There
is still a particular, tractable situation: apply two pulses, wait for a time T
2
, so
that the transversal components decay,
) 0, apply another set of two pulses,

and so on. The calculation of the nal polarization for such an n2-pulse sequence
amounts to the recursive use of Eqs.(6.4, 6.5). The nal [
z
)[ is then maximized
over all free parameters of the involved pulses. We found that this maximization can
be carried out locally, i.e., by maximizing the output [
z
)[ after each pair of pulses.
For a suciently large n the recursion converges, and, as seen in Table I, the nal
results indeed improve.
Table 6.1: Maximum value of [
z
)[ for 2 pulses, 3 pulses and 50 2 pulses and
various values of and = k
B
T/() (for the ohmic case). The lowest two lines
represent the nal collective polarization [m
f
[ of a strongly disordered ensemble with
0
= 0 and arbitrary large disorder dispersion d under spin-echo setup of 2 pulses
plus a pulse and a sequence of 30 sets of such pulses.
0.1 0.1 1 1. 10. 10.
0.1 1. 0.1 1 0.1 1.
2 0.1126 0.0826 0.4930 0.3516 0.8910 0.7931
3 0.2036 0.1369 0.7250 0.4732 0.9550 0.9025
50 2 0.2779 0.1895 0.6948 0.5430 0.9509 0.8995
2 + 0.0834 0.0604 0.2498 0.2026 0.4634 0.4155
50 [2 +] 0.1396 0.1055 0.3119 0.2639 0.5214 0.4834
6.2 Cooling using spin-echo pulses
So far we assumed that we deal either with a single spin coupled to the bath, or,
equivalently, with an ensemble of identical non-interacting spins each coupled to
its own bath, which is a standard setup of NMR/ESR [AG82, Sli90]. However,
many experiments are done with ensembles of non-interacting spins having random
frequencies due to the action of their environment or due to an inhomogeneous
external eld, see sections 4.1 and 5.4. The collective variables are then obtained by
averaging over the corresponding expressions for a single spin:
m
f
=
_
dP()
z
)
f
. (6.9)
We assume for simplicity, as we did for the work extraction setup, that the
distribution of is Gaussian with average
0
and dispersion d, Eq. (5.39):
P() =
1
2d
e
(
0
)
2
/(2d)
. (6.10)
Averaging over P() the term e
i
in Eq. (6.5) produces factors e
d
2
/2
, that is,
a strong decay on characteristic times T
2
1/
d. After this decay, s

2
0 and any
set of two pulses will only heat the ensemble up as seen from Eq. (6.4).
It is however possible to employ the spin-echo phenomenon and cool, i.e., to
increase the collective nal polarization [m
f
[ as compared to the initial [m
i
[, even for
a completely disordered ensemble with T
2
being very short: Apply precisely in the
middle of two pulses an additional -pulse in x-direction:
T

z
=
z
, T

y
=
y
, T

x
=
x
, (6.11)
Now
z,f
= c
t
T
1
c
T
2

z
, and m
f
is worked out analogously to Eqs. (6.4, 6.5):
m
f
= m
i
cos 2
1
cos 2
2
+s
3
sin 2
1
sin 2
2
, (6.12)
s
3
= e
4()+(2)
'
_
e
i
3
(m
i
cos
3
i sin
3
)
_
, (6.13)
3
= (0, , t) (, 2, t), (6.14)
where
3
=
1

1

2

2
. As compared to (6.5), both the decoherence
e
4()+(2)
and the backreaction
3
term are dierent; this is analogous to what
happened in the case of work extraction, see page 93 for details. In particular, in
the Gaussian regime t
2
decoherence is absent, while the exponential regime
t is unchanged, e
4()+(2)
= e
(2)
. Due to the T
pulse, the T
2
-decay has
been eliminated, so no term like e
it
in Eq. (6.5) appears here. Now
0
and d
enter only via m
i
. The general structure of Eqs. (6.12-6.14) is close to the one of
Eqs. (6.4,6.5), in particular, the optimization over
1
,
2
goes in the same way. To
facilitate comparison, we take
0
= 0 thus m
i
= 0 and disorder strength d that
can be arbitrary large and recalling Eqs. (4.16, 4.55) and the ergodic condition
t leads to
[m
f
[ = e
4()+(2)
sin [2 arctan() arctan(2)] ,
where we already inserted the optimal values
3
=
2
,
1
=
2
=

4
: the choice of
optimal pulses can be the same as for the two-pulse scenario. As seen in Table I the
maximal [m
f
[ can exceed 0.4 for suciently strong coupling and/or low temperatures,
though it is smaller than the maximal polarization of the non-disordered system
under two pulses. The results are also seen to improve by applying a sequence of
three (spin-echo) pulses separated by a time much larger than T
2
, which amounts to
recursive use of Eqs. (6.12, 6.13).
6.3 1/f spectrum
Another relevant situation of the bath-spin interaction is the case of the 1/f spec-
trum. It has been recently observed in a two-level system (spin) of charge states
II.6.4 Conclusion 103
in Josephson-junction circuit (Cooper-pair box) [N
+
02]. The interaction of the spin
with the bath of background charges is modeled as
J
f
() =
b
f
e
/
( ), (6.15)
where b
f
is a coupling constant, the step function, and where and are, re-
spectively, the largest and smallest frequencies of bath response. In this spectrum
an infrared cuto is needed to avoid divergences. As compared to the Ohmic sit-
uation, the upper frequency is not relevant: the integrals in Eqs. (4.14, 4.24)
converge for . Within the realization of [N
+
02]: ( 10
11
Hz),
= 314Hz, = 34eV, while the relevant dimensionless coupling constant is
very large
f
b
f
/
2
= 2.3 10
11
. Thus the cooling eect will be especially visible
here. We worked out Eqs. (4.14, 4.24) using Eq. (6.15) and applied them to the two-
pulse cooling situation described by Eqs. (6.4, 6.5) under the following conditions: 1)
high temperatures
f
= k
B
T/() 1 (pessimistic case!); 2) 1 and
f
1
(experimentally relevant regimes [N
+
02]); 3)
z
)
i
0 due to large temperatures,
4)
f
due to large
f
. In analogy to Eq. (6.6), the result reads (
2
=

2
):

z
)
f
= s
2
= e
f
y
2
sin
_
y
y
2
4
f
_
, y
f
, (6.16)
where y is a dimensionless time, and where we omitted terms O[y
3
/
2
f
]. As seen
in Fig. 6.2, the polarization can increase from its original value zero to 0.997 for
f
/
f
= 0.01, which in numbers of Ref. [N
+
02] corresponds to the temperature
T = 15K. Lower temperatures only increase this value, while the eect is still visible
for a very high temperature T = 1500K, P
max
0.3.
1 2
0
0.2
P
0.8
1
Figure 6.2: Final polarization P = [
z
)
f
[ given by Eqs. (6.4, 6.5) versus di-
mensionless time for two optimal pulses for the 1/f interaction.
f
= 10
3
and
f
/
f
= 0.01, 0.1, 1 (bold, normal, dashed).
z
)
i
and are taken to be 0.
6.4 Conclusion
We described a method for cooling spins that uses the standard bath-spin interac-
tion. In contrast to existing methods [AG82, Sr89, Sli90], it assumes neither an
already existing high polarization, nor controlled spin-spin or bath-spin interactions.
Our basic assumptions are i) a decoherence time T
2
much smaller than the energy
relaxation time T
1
, and ii) the availability of sharp and strong pulses acting on the
spin. While a suciently long T
1
time is necessary for any dynamical polarization
method [AG82, Sr89], we have chosen strong pulses for a clean demonstration of
the cooling eect. It is likely that the eect exists for more general types of external
elds.
The origin of the present mechanism is in the backreaction of the spin to the
bath, which via coupling to transversal components of the spin can cool its state
provided the proper sequence of pulses is chosen. The bath is needed, since external
elds alone cannot cool [KP92]. However, within our method the cooling takes a
nite time and once it is over, the spin can be decoupled from the bath, e.g., via
applying a sequence of -pulses [Sli90, VL98]. The advantages of the present method
are that it works even for very weak dc elds and that it applies to an ensemble of
spins having completely random frequencies (inhomogeneous broadening). We are
not aware of other methods which may achieve such a goal. The cooling is ecient
already for small to moderate couplings between the spin and the bath, and is
especially visible for situations where a strong bath-spin coupling is inherent (1/f-
noise). This latter situation is experimentally realized [N
+
02], and we expect the
present cooling mechanism to apply there.
Part III
Spin Glasses
7
Introduction to spin glasses
During the last decades a lot of eort has been devoted to understand the physics
of spin glasses. This development was not only interesting to better understand this
type of materials, but it also served to bring new insights in the eld of complex sys-
tems as a whole, including applications to amorphous systems in condensed matter
physics, computer science, biology and economy [MPV87].
Spin glasses are materials whose ground state at low temperature +
+
+
Figure 7.1:
Example of
frustration
is a frozen disordered state instead of an ordered one, as it is the
case for ferromagnets, or a periodic one, as it is the case for antifer-
romagnets. This is a consequence of a feature in this systems called
frustration. Frustration means that the dierent interactions are in
conict, so not all of them can be mutually satised. For instance,
in gure 7.1 a sketch of a square lattice in 2D is given, with spins 1
on the edges. 3 of the couplings among the 4 spins are ferromagnetic
and the fourth coupling is antiferromagnetic. In this situation, there
is always one coupling which is not satised.
In spin glasses, frustration is usually a consequence of the randomness in the
interactions, i.e. quenched disorder. In spin glass literature, quenched disorder means
that the interactions between spins and/or their locations have constrained disorder.
The spins are free to move, however, due to this disorder in the interactions, the
spin glass phase is an example of spontaneous cooperative freezing of the (random)
spin orientations. This freezing in random orientations can be seen with NMR or
neutron scattering measurements.
In chapter 8, the transition temperature of metallic spin glasses is studied. Metal-
lic spin glasses are composed of noble metals (Cu, Ag, Au) having a small concen-
tration c of transition metals (Mn, Fe). The magnetic transition metal atoms are
randomly distributed throughout the non-magnetic host, the noble metal. These
magnetic impurities interact via the RKKY interaction. The spin glass behavior
appears due to the fact that this interaction has an oscillatory behavior, for some
spins it is positive while for others is negative. In this chapter we study how the
108 III.7. Introduction to spin glasses
Figure 7.2: Field cooled [(a), (c)] and zero eld cooled [(b), (d)] magnetizations
( M/6 gauss for CuMn (1 and 2 at. %) as a function of temperature; from
[NKH79]
transition temperature to the spin glass phase depends on the concentration c of
magnetic impurities. The study is performed by means of a cut-o in the range of
the RKKY interaction due to thermal uctuations.
Other types of materials also display spin glass behavior. Examples of that
are compounds with magnetically ordered host metals (PtMn), compounds with
rare earths (YDy, ScTb) and ternary systems (La
1x
Gd
x
Al
2
). All of those can, in
principle, be described with RKKY interactions. There are other materials, called
substitutional spin glasses, that show spin glass behavior and are not describable
via RKKY interactions. Examples are magnetic insulators, such as Eu
x
Sr
1x
S, and
amorphous alloys, such as Al
1x
Gd
x
and Fe
x
Ni
1x
.
Experimentally there are few properties which are common to all spin glasses,
no matter what is the precise physics underlying the disordered interaction among
spins. We have to experimentally see that all the magnetic moments freeze in a
disordered conguration in the spin glass phase. To this end, one can study the
dynamic magnetic susceptibility (), i.e. the the response of the system to an
external AC eld perturbation of frequency . In presence of weak external elds,
the magnetic susceptibility displays a peak at the critical temperature due to the
fact that the system freezes in a given conguration in the spin glass phase.
The magnetic susceptibility depends on whether the system has been cooled down
in presence of a magnetic eld or not. The zero eld cooled (ZFC) susceptibility is
measured by cooling down the sample in zero eld and heating it up in a small eld,
while the eld cooled (FC) susceptibility, the measurement is performed by cooling
down in a eld and repeating this for a slightly dierent eld, these results can be
seen in Fig. 7.2.
Many other experimental facts are exhibited by spin glasses, for a complete
description the reader is referred to Ref. [Myd93]. One last remark about all ob-
servables in spin glasses is that they are self averaging. In spite of the fact the
III.7.1 Theory and mean eld models 109
spin glass phase is disordered, the observables given by dierent samples are the
same. Therefore, any theory aiming to describe spin glasses has to give self aver-
aging observables. Self averaging means that sample to sample uctuations vanish
in the thermodynamical limit, then any intensive quantity does not depend on the
realization of the quenched disorder.
7.1 Theory and mean eld models
An Ising ferromagnet, below the Curie temperature, can take two opposite magne-
tizations, separated by an overall spin ip. These two minima of the free energy are
separated by an energy barrier that is innitely large for an innitely large system,
thus the ergodicity has been broken. In spin glasses, on the contrary, there are not
only two of these minima, there are many. Furthermore, inside these minima, or val-
leys, there can be many metastable states or sub-valleys. Each of these (sub-)valleys
brings a dierent time scale.
The properties given by each of these valleys will, in principle be dierent, and
dierent from the equilibrium ones. The equilibrium properties will involve, in the
end, an average over all possible valleys with the appropriate thermal weights. In
order to know the properties of each valley one should be able to apply innitesimal
magnetic elds h
i
proportional to the m
i
of that single valley. Since, of course, we
do not know a priori these m
i
the task becomes extremely dicult. Due to that,
the study of spin glasses is very dicult even in the mean eld approximation.
The rst mean eld model of spin glasses appeared in 1975 by Sherrington and
Kirpatrick [SK75], the SK model. The SK model is a fully connected Ising spin
system with quenched random interactions. In spite of being mean eld, it gives a
good insight in the understanding of the spin glass phase. To solve it, the use of
new theories was needed, the most known is the Replica Theory. A few years later,
Parisi proposed [Par79] a marginally stable solution for the SK model which is still
nowadays accepted by means of the Replica Symmetry Breaking [MPV87].
In ferromagnets the order parameter is the magnetization m
i
= S
i
). In spin
glasses however the magnetization stays at zero value in the low temperature phase
for symmetric bond distributions, i.e. random couplings symmetrically distributed
around zero. Higher order moments are needed, Edwards and Anderson proposed
[EA75]:
q
EA
= lim
t
lim
N
[S
i
(t
0
)S
i
(t
0
+t))]
av
(7.1)
where the average is over a long set of reference times t
0
[FH91]. This quantity
is zero if the system is ergodic and non-zero if the system is trapped in a single
phase. Since the N limit is taken rst, q
EA
measures the mean-squared single-
valley local spontaneous magnetization, averaged over all possible valleys. Then as
the equilibrium or statistical mechanics order parameter one takes the mean square
local equilibrium magnetization:
q =
_
S
i
)
2
av
=
_
m
2
i
av
. (7.2)
110 III.7. Introduction to spin glasses
P
q
P
q
P
q
m
2 2
m
P
q
q
EA
q
EA
Spin Glass Ferromagnet
Low T
High T
Figure 7.3: Comparison of the overlap distribution P(q) in ferromagnets and spin
glasses
Since there are many phases, it can be interesting to know, apart of this mean
value, the correlation between states, the overlap, between two states a and b,
q
ab
=
1
N
i
m
a
i
m
b
i
(7.3)
and consider also its distribution
P
J
(q) (q q
ab
))
ab
P
a
P
b
(q q
ab
) (7.4)
where P
a
, P
b
are the probabilities of the states a and b respectively.
This distribution in a ferromagnet will be composed by two deltas at the cor-
responding two possible values of the magnetization (M) because there are two
phases. In spin glasses however it might have a continuous part showing the large va-
riety of phases involved in the spin glass phase. A sketch of these phases is depicted
in gure 7.3
In the SK model, the order parameter grows continuously from zero as the tran-
sition is crossed. It is described by a Full Replica Symmetry Breaking (FRSB)
solution. In other models, like the p-spin model [Gar85, CS92], the order parameter
is a step function taking two values q
0
and q
1
> q
0
, with q
1
q
0
jumping discon-
tinuously from zero to a nite value as the transition is crossed. No discontinuity
appears in the thermodynamic functions however. These models are described by
one step of Replica Symmetry Breaking (1RSB). This last scenario has been used in
the last years for the study of structural glass transition observed in fragile glasses.
III.7.1 Theory and mean eld models 111
In chapter 9 a numerical simulation of an Ising spin glass on a hypercubic cell
is presented. This model has nite connectivity (not all spins are neighbors) but it
is expected to converge to the SK model in the thermodynamic limit. By studying
it, we can get information on the nite size eects in spin glasses by analyzing how
the simulation tends to the known mean eld limit. Moreover it provides a tool to
analyze whether the mean eld picture given by Full Replica Symmetry Breaking
theory holds for nite systems, a strongly debated question in literature.
8
Transition temperature in
metallic spin glasses
In this chapter, the dependence of the transition temperature T
g
in terms of the
concentration of magnetic impurities c in spin glasses is explained on the basis of
a screened RKKY interaction. The two observed power laws, T
g
c at low c and
T
g
c
2/3
for intermediate c, are described in a unied approach, [SNL04].
8.1 Concentration dependence of the transition tem-
perature
Metallic spin glasses such as Cu
1c
Mn
c
, Ag
1c
Mn
c
, Au
1c
Fe
c
are alloys formed by
magnetic impurities embedded in a noble metal. The transition temperature of such
materials depends on the magnetic impurity concentration c. Dierent phenomena
dominate for dierent concentrations. The mutual interaction between magnetic
impurities is mediated by electrons, the RKKY interaction. It can be understood
as follows: the sea of electrons interact with an impurity and the scattered wave
interferes with the incoming one. This creates a pattern of spin polarizations that
brings an oscillatory behaviour and a 1/r
3
fall o of the form J(r) = Acos(2k
F
r)/r
3
at T = 0. For very low concentrations, less than 50 ppm, the interaction can be
neglected and the magnetic impurities act independently bringing the Kondo eect.
For larger concentrations though less than 10 at. % the RKKY interaction is the
dominant interaction and the spin glass phase appears. The oscillatory nature of the
interaction and the position randomness of the impurities form a disordered magnet.
Above this concentration, the chance of having a signicant amount of impurities
as rst or second neighbours is high, and consequently clusters are formed. For
even larger concentrations the percolation limit is reached and ferromagnetism or
antiferromagnetism, depending on the type of magnetic impurities, appears. For a
review on these dierent regimes see Ref. [Myd93].
114 III.8. Transition temperature in metallic spin glasses
The spin glass region of concentrations (excluding the cluster region) exhibits
two dierent behaviours. On one hand, for concentrations lower than 1/2 at. %,
the data points approach a linear curve, T
g
c, while on the other hand, for higher
concentrations, up to 10 15 at. % the t turns to T
g
c
2/3
. In this chapter we
explain these scaling laws in a unied treatment.
The key point lies in the fact that the RKKY interaction is not innite ranged.
At nite temperature, phonons interact with the electron sea smearing out the spin
polarization pattern at large distances, consequently the interaction is cut o at a
length
T
. It has the form
J(r) = Ae
r/
T
cos(2k
F
r)/r
3
(8.1)
For pure metals this cut o is the thermal coherence length
T
= v
F
/ 1/T.
For disordered metals where non-magnetic impurities exist alongside with magnetic
ones, the former start to play a role [ZS86, BP86, Ber87]: the electron wave that
scatters from the magnetic impurity diuses around the non-magnetic ones before
it reaches another magnetic impurity. The contributions of all such diusive paths
add coherently as long as the distance between the magnetic impurities is smaller
than the thermal coherence length
T
. This leads to cutting o the typical value of
the interaction [ZS86, BP86, Ber87] as in Eq. (8.1), but with
T
= (D/)
1/2
1/T
1/2
, where D = v
2
F
/3 is the diusion constant of an electron in a disordered
metal and is the mean free time for elastic scattering. Thus for T (assuming
also that the elastic scattering dominates over the inelastic,
inel
) the eective
range of the interaction for a disordered metal is shorter than for the pure case.
This denes a limit between the two situations that can bring dierences in the
large concentration regime for certain materials since the above inequalities may
well be reached in some cases.
Shegelski and Geldart [SG92b, SG92a] have derived the range of an indirect-
exchange interaction in disordered metals which takes into account the RKKY in-
teraction and sd scattering. They could well describe a wide range of experiments
by tting new length scales that appear in the problem. We shall not aim at tting
the data but at giving the basic mechanism. We focus on the case where the con-
centration of magnetic impurities is changed with no other added impurities. Our
approach does not have adjustable parameters.
We use a Hamiltonian that takes into account the fact that the magnetic impu-
rities, i.e. spins, are present in some sites of the lattice and not in all of them
[Nie93, NvD99]
H(s) =
1
2
r,r
J(r r
t
)s
r
s
r
c
r
c
r
H
r
s
r
c
r
(8.2)
where s
r
represents the spin on site r and c
r
= 1, 0 whether a spin is present on site
r or not. J(r r
t
) is the RKKY interaction between sites r and r
t
. This model is
used since it contains from the beginning a dependence on the concentration via an
average of the c
r
. It is a random site problem, since the randomness comes from the
distribution of spins in a lattice, and not a random bond problem where each bond
III.8.1 Concentration dependence of the transition temperature 115
has a random strength, as is for example in the case of the SK model. This model
can be solved in the low concentration limit via a mean eld approximation in the
replica scheme [Nie93, NvD99]. The transition temperature was found to satisfy the
following condition
c
r
tanh
2
[J(r)] = 1 (8.3)
where r represents each of the sites of the lattice since the factors c
r
have been taken
in average. This relation was postulated long before by comparison with the random
bond problem [SS75]. Combining the RKKY interaction in Eq. (8.1) with Eq. (8.3)
we basically have an equation of the form
4c
_

T
0
dr r
2
tanh
2
_
A
r
3
_
= 1 (8.4)
where the upper limit of the integral accounts for the interaction cut o, and, as
a rst approximation, we can consider that the oscillations of the cosine are not of
qualitative importance in the range where the hyperbolic tangent squared has signif-
icant values. Eq. (8.4) gives the transition temperature T
g
for a given concentration
c. For small concentrations, in order to satisfy the equation, the transition temper-
ature has to be small, which allows us to extend the upper limit of the integral to
innity. It yields
4c
g
_

0
dx x
2
tanh
2
_
A
x
3
_
= 1 (8.5)
which leads to the expected result [SS75, Nie93, NvD99]
T
g
c (8.6)
For larger concentrations, the transition temperature has to decrease. Then the
upper limit remains nite and the tanh
2
in Eq. (8.4) can be approximated by 1. We
get
4c
_

T
0
dr r
2
= c
4
3

3
T
= 1 (8.7)
From the middle expression we see that the transition takes place when there
starts to be on the average more than one impurity in the range of attraction of
the RKKY interaction, as one might have expected. Since for a disordered metal
T
T
1/2
, the scaling law then reads
T
g
c
2/3
(8.8)
We want to stress that the 2/3 power law is intimately related to the fact that
the system is considered to be a disordered metal, having
T
T
1/2
. In this
sense, we consider metals with a small amount of non-magnetic impurities (for a
study on the eect of a variation on the concentration of non-magnetic impurities
116 III.8. Transition temperature in metallic spin glasses
see Refs. [SG92b, SG92a, VS85]). For a pure metal, i.e. for T
g
, where the
range of the RKKY interaction is proportional to the inverse of the temperature,
T
T
1
, the scaling law in Eq. (8.8) becomes T
g
c
1/3
. In the intermediate case,
i.e. for T
g
, to consider the system as being disordered is not anymore a good
approximation, a fact that may bring values of the exponent lower than the 2/3
predicted for the disordered case, i.e. for T
g
. An example of that is the case
for Au
1c
Fe
c
where the exponent 0.58 [CM72].
We have explained here the two pure scaling laws T
g
c and T
g
c
2/3
(or
T
g
c
1/3
for pure metals) corresponding to the canonical spin glass [Myd93]. These
are the limiting situations and eective exponents found in experiments may lie
between 1/3 and 1. We suspect that in the experiments performed till now, a non-
negligible amount of non-magnetic impurities were always present in the sample. We
therefore propose to perform new rened experiments in order to test the presence
of lower exponents .
9
The Ising Spin-Glass on a
Hypercubic Cell
In this chapter we describe a numerical simulation of the statics of an Ising spin-
glass on a hypercubic cell. This model was proposed a decade ago by Parisi, Ritort
and Rub [PRR91] in order to have a better understanding of the solution of the
Sherringtion-Kirkpatrick model proposed by Parisi [Par79]. Analytic progress on
the stability of the solution was intricate and numerical solutions were a hope to
have more information. The SK model however, is extremely CPU time consuming
since all spins are connected. Therefore some nite range models that converge in
the thermodynamic limit to the mean-eld SK were proposed. Hypercubic lattices
in nite dimensions are good candidates to this end.
In the case of ordered systems, the Ising model on a hypercubic lattice converges
to mean eld when the dimension goes to innity [Par88]. The same happens for
spin glasses [GMY90]. We deal with an hypercubic cell, i.e. an hypercubic lattice
with lattice size L = 2, with connectivity z = D and free boundary conditions. The
hypercubic lattice, on the contrary, has connectivity z = 2D. It can be proven that
the hypercubic cell in ordered systems converges towards mean eld. We expect the
same to happen in disordered systems. In the cell the number of spins is N = 2
D
a xed parameter for a given dimension, while in the hypercubic lattice N and
D are two independent parameters. In ordered systems a high coordination (high
dimension/high temperature) expansion in 1/z for the thermodynamic potential
shows that the hypercubic cell (in dimension D) and the hypercubic lattice (in
dimension 2D) are similar also at nite dimensions [Par88]. If that holds, as we
expect, for spin-glasses the study of the hypercubic lattice can give us information
over the spin glass in hypercubic lattices, thus the Edwards Anderson (EA) model.
We study the equilibrium properties as a function of the dimension (which in this
case is the same as the size of the system). We look at the behaviour as dependent
on both nite size and nite connectivity (dimensional) eects and try to distinguish
dierent regimes separated by a crossover and we carefully look, in the range of sizes
118 III.9. The Ising Spin-Glass on a Hypercubic Cell
we are capable to reach with our numerical simulations, the way the system tends
to its mean-eld limit: the Sherrington-Kirkpatrick Model.
9.1 The Model
The hypercubic cell is an hypercubic lattice of lattice size L = 2. Thus in D = 2 is
a square, D = 3 is a cube and so forth. It consists of 2
D
Ising spins s
i
, each with D
nearest neighbors (nn) whose interaction is described by the Hamiltonian
H[s] =
<ij>
J
ij
s
i
s
j
, (9.1)
where < ij > denotes nn couples i,j. Having D nn is due to the geometry of
the system. One can have the pictorial image in 3D, where the spins lie on the
vertices and the couplings lie on the edges of a cube. This view can be generalized
algebraically to higher dimensions. The J
ij
are binary quenched random couplings
with probability distribution
P(J
ij
) =
1
2
(J
ij
J) +
1
2
(J
ij
+J) (9.2)
of zero mean and variance J
2
ij
= J
2
= 1/D (this is chosen in order to properly
normalize extensive observables in the limit D ). The overline denotes the
average over the disorder, thus over the distribution Eq. (9.2).
9.2 Simulation and Data Analysis
9.2.1 Exchange Monte Carlo
Motivations
In a statistical mechanical system at equilibrium the probability that a certain con-
guration can take place at a given temperature is given by the Boltzmann-Gibbs
distribution. Any proper dynamics must lead to such a distribution in the long
time limit and must satisfy detailed balance at any step of relaxation. Based on
such general properties is the Metropolis algorithm, the kernel of the Monte Carlo
method widely used in numerical physics.
In a complicated model in statistical mechanics such an algorithm allows to
explore most of the phase space even though dierent states are often separated by
high energy barriers in the free energy landscape. Initially, the sequences of Monte
Carlo steps move the system around the region of the phase space next to the initial
conguration. If that region is separated by other regions of the phase space by high
barriers or narrow paths it can take a lot of time to cross to another region and
continue the probe of the systems states. To reach equilibrium can become a very
hard task. This is the case in disordered systems such as spin-glasses.
III.9.2 Simulation and Data Analysis 119
To be operative we can look at the computer time needed to achieve equilibrium,
that increases with the size of the simulated system exponentially. In the more precise
computer science language this heuristic observation translates into saying that our
thermalization problem is a Non-deterministic Polynomial Complete (NP-complete)
problem [GJ79]. The Exchange Monte Carlo method is an improved Monte Carlo
method which in many cases incredibly reduces the thermalization time and it is
extremely eective in spin-glasses.
The goal is to make the system explore the whole phase space accessible for a
given temperature. In the case of spin-glasses however, it is known that for low tem-
peratures the free energy displays a very rough landscape. This landscape consists
of a large number of valleys, i.e. minima, separated by large energy (or entropy)
barriers. Moreover, these energy barriers grow exponentially with the size of the
system. Then, for low enough temperatures and/or high enough sizes, the system
will most likely get stuck in one of these valleys not being able to overcome the
barriers. With diculties a barrier will be crossed and a new minima will be visited,
notwithstanding, the dynamics will be too slow and the number of Monte Carlo
steps needed to reach thermal equilibrium will be too large. The Exchange Monte
Carlo method circumvents this problem by using the fact that at high temperatures
the energy barriers decrease and ultimately disappear above the phase transition to
a paramagnet.
Let us consider a spin glass at low temperature. It is trapped in a minimum, not
visiting other minima within a reasonable number of standard Monte Carlo steps.
In order to make it move in the free energy landscape the system has to be heated
enough for these barriers to decrease and then cooled down again expecting that the
valley visited this time will be dierent from the previous one. By repeating these
heat kinks many times, the phase space accessible at a given temperature is much
easier explored, and thermalization is much less time consuming.
The above described procedure is called Simulated Tempering, where one system
moves back and forth around dierent temperatures. In order to do this guaranteeing
that the true Boltzmann-Gibbs measure is yielded for long times, one needs to set
certain parameters and this requires a certain number of tuning simulations before
starting the actual analysis. More ecient is the Exchange Monte Carlo method, also
called Parallel Tempering method, that uses the same idea of heating and cooling
but is applied to more (n) systems in parallel. At the beginning of the procedure
these n systems are distributed over n dierent temperatures, see g. 9.1. They are
heated and cooled by swapping them among adjacent thermal baths.
T
2
T
3
T
3
4
T
4
5
T
5
6
T
6
7
T
7
T
1 2
1
Figure 9.1: Schematic picture of the dierent systems in an Exchange Monte Carlo.
Here n = 7
The procedure is as follows:
i) Initial random systems are placed each one at a dierent temperature, see
g. 9.1.
ii) The dynamics of each system is simulated simultaneously and independently
as canonical ensemble for a few Monte Carlo steps using the standard Monte
Carlo method.
iii) A pair of systems at adjacent temperatures is interchanged according to some
rules that will be described below.
iv) Repeat steps ii and iii until equilibrium is reached in the thermal bath of lowest
temperature.
Finally, and if the procedure has been properly carried out, each of the systems
has visited all temperatures, which eectively means that it has been warmed up and
cooled down exploring then dierent regions of the phase space for each temperature.
If, on the contrary, we focus our attention to just one temperature, one system,
though a dierent one each time, it has visited most of the phase space and then
equilibrium has been reached for that temperature. So in the end we have systems
at equilibrium at all temperatures of the grid.
The exchange is performed two by two, so the conguration at
0
tries to be
swapped with the one at
1
. After that, the swap between the conguration at
1
and the one at
2
is proposed, and so forth. After a few iterations of this
procedure the starting systems are completely spread in the temperature space.
They walk around the dierent heat baths. An intermediate situation could be the
one described in gure 9.2
T
3
T
1
T
3
7
T
4
4
T
5
5
T
6
2
T
7
T
1 2
6
Figure 9.2: Example of the spreading of the systems among the dierent tempera-
tures
Exchange probability computation
In order to be physical, the exchange of systems at dierent temperatures has to
satisfy detailed balance. Otherwise, we would be inducing a biased evolution. Let
P(X,
m
) = Z
1
(
m
)e
m
1(X)
be the probability of having the conguration X
at temperature 1/
m
and let W(X,
m
[ Y,
n
) be the probability of exchanging the
conguration X at temperature 1/
m
with the conguration Y at temperature 1/
n
.
Detailed balance means that the transition in one direction, say from X,
m
Y,
n
,
has the same chance as in the other direction, Y,
n
X,
m
. Formally it reads:
P(X,
m
)P(Y,
n
)W(X,
m
[ Y,
n
) = P(X,
n
)P(Y,
m
)W(Y,
m
[ X,
n
), (9.3)
where we have used the fact that the probability of each conguration at each tem-
perature is independent of each other, i.e. P(X,
m
; Y,
n
) = P(X,
m
)P(Y,
n
).
We can write Eq. (9.3) as
W(X,
m
[ Y,
n
)
W(Y,
m
[ X,
n
)
= e
exp (
n
m
) [H(X) H(Y )] (9.4)
We then use the Metropolis algorithm to evaluate the exchange probability. This
amounts to say that
W(X,
m
[ Y,
n
) =
_
1 if < 0,
e
if > 0.
(9.5)
This basically states that both systems are always in equilibrium, in both involved
thermal baths, even after the exchange. The congurations X and Y are valid
congurations for both
n
and
m
otherwise no exchange is performed. The more the
two energy distributions overlap the higher is the exchange probability. Figure 9.3
reproduces a schematic representation of the region of energies where the exchange
successfully takes place.
T
1
T
2
T
3
E
P(E) P (E) P (E) P (E)
Figure 9.3: P(E) for dierent temperatures. The overlap has to be large enough to
have good exchange rates
For this method to be eective, we need to have not too low exchange rates
since otherwise it would be just standard Monte Carlo of many systems, bringing no
improvement. Also, we need to guarantee that each system visits all thermal baths
with the same frequency. We can tune the exchange rates by properly choosing
the temperature grid. If the distance between temperatures is too large, there will
almost be no overlap between adjacent P(E)s so there will almost be no exchange
of systems.
These exchange rates depend on the size of the system. The larger the system the
shorter the interval between temperatures. In practice, we check the exchange rates
for dierent temperature grids at the highest size to be simulated before starting the
real simulation. We choose a minimal exchange rate of 30% in order for the method
to be eective.
The Parallel Tempering algorithm has proved very eective in the statistical
mechanics of disordered systems. It oers the possibility to deal with low temper-
ature regimes whose equilibrium could not be reached for reasonable sizes with the
standard Monte Carlo. There are other methods that improve the Monte Carlo
technique, we already commented on simulated annealing for instance. Other re-
lated methods could be the scaling approaches based on Umbrella sampling, see for
instance [TV77], or the so-called Multicanonical Approaches [BN91, BN92]. Mul-
ticanonical Approaches are sometimes more ecient than Exchange Monte Carlo,
e.g. they work better for rst order phase transitions, but are much more dicult
to implement. In general, when the Exchange Monte Carlo method is applicable, it
is recommended since it is easy to implement and very powerful.
9.2.2 Multispin coding
In order to calculate the energy we use the multispin coding technique. This is
based on the fact that, since Ising spins (1) can take only two variables, they can
be represented by a 1-digit binary number. Then, when performing a simulation,
each spin in the system corresponds to a single bit in the computers memory. That
reduces enormously the memory consumed by the program and increases its speed
due to the fact that multiple operations can be done at the same time.
The way we distribute the spins in the computer memory is as follows. We
make use from the fact that the computer is prepared to work with words, i.e.
sequences of bits that create an integer, a oat... We use, long long integers, which
are sequences of 64 bits.
Position 0 1 2 3 4 5 6 7 8 9 10 11 ..... 62 63
Value 0 0 1 0 1 1 0 1 1 1 0 1 ..... 0 1
We write in each of these bits the value of one spin of a dierent system in the
parallel tempering technique. That limits the number of systems to the length of
the word (though, more words could be used), but it improves, as we will see, the
speed of the simulation. We will have then an array of words corresponding to each
of the spins of all systems. Therefore, the bit in position 0 of the 0th word of the
array will correspond to the rst spin of the rst system (so the rst temperature
in the parallel tempering technique) note that usually in computers language
enumerations start at 0 and not at 1, the bit in position 3 of the 4th word of the
array will correspond to the 4th spin of the 5th system, and so on. Couplings are
written in the same way, each bit corresponds to each of the dierent systems in the
parallel tempering (which is always the same value for all systems in one sample,
but using this we can compute various samples in parallel) and there is an array of
such words corresponding each of the dierent couplings inside the hypercube.
This is dierent from what it is commonly done, say when simulating the SK
model. There, usually each bit of one word corresponds to dierent spins of the same
system and not of dierent ones. The main dierence is that in our situation the
geometry of the system matters. One has to nd the correct neigbours for a given
spin and the couplings among them. By choosing this way of distributing the spins
in memory, one has to look only once for the neighbours for all samples involved
and this, to our understanding, reduces considerably the time. In the SK, however,
since there is no geometry, all spins are neighbours, so they are all coupled to each
other, and no intricate search for neighbours is needed. Having then all spin of one
system aligned in the computer memory can be useful.
To evaluate the energy of the system we have to use H
i,j
J
ij
S
i
S
j
. For
that we have to nd the neighbours of each spin and the corresponding coupling.
To nd the neigbours we have to keep in mind the geometry of the hypercube,
see gs. 9.4, 9.5
(0,0) (0,1)
(1,1) (1,0)
Figure 9.4: Hypercube D = 2
(0,0,0) (0,1,0)
(1,0,0) (1,1,0)
(0,0,1) (0,1,1)
(1,0,1) (1,1,1)
Figure 9.5: Hypercube D = 3
It is important to notice that an hypercube in D dimensions has 2
D
spins and
D2
D1
couplings. The spins lie in the vertices of the hypercube while the couplings
can be represented by the edges between vertices. Then, if we think in cartesian
coordinates with a vertex of the hypercube in the origin and the length of the edges
being 1, each of the vertices is dened by a set of D 0s or 1s. These denes an integer
written using binary numbers. Each of the D neighbours of the, say (0,0,0,0,...,0)
spin, can then be found by changing each time a 0 by a 1, see gs. 9.4, 9.5. To nd
the coupling linking the spin with one of its neighbours is a bit more complicated.
There are D2
D1
couplings, which means that there are 2
D1
couplings for each
of the D directions. Since we change one bit from the spin coordinates to nd one
neighbour, that denes which direction we have the coupling (the edge between these
two vertices), then the rest of the binary sequence of the spin coordinate (taking out
the one bit that changes) is a number ranging from 0 to 2
D1
1 so the residue of
the division of such number and 2
D1
yields the couplings that we have to use along
that direction. Then the rst 2
D1
couplings are in the 1st (x) direction the next
2
D1
in the 2nd (y) and so on till the direction number D.
Energy calculation
Now we have the couplings and each of the spins, then by just using two bitwise ex-
clusive-OR: , we compute the energy contribution of each couple of spins connected
by it.
E
ij
= J
ij
(S
i
S
j
) i = 1, . . . , N (9.6)
Since this operation is performed with each word, the energy attributed to a link
between two spins can be computed with a single operation for the whole set of
dierent systems in the exchange Monte Carlo method. Finally, the energy related
to one spin can be found by summing all the contributions from its neighbours. And
the total energy is found by summing this value for all the spins and dividing by
two, because of double counting.
In the Monte Carlo procedure each spin is asked to turn around. If that favours
the energy, e.g. the total energy decreases, the turn is accepted, otherwise it is
accepted only with some probability that depends on how large is the energy increase
proposed and on the temperature of the system. Such a probability only depends
on that spin and its couplings to its neighbours. Then to perform the switch of one
spin we only have to calculate the energy locally.
9.2.3 Data Analysis
Equilibration
One of the main issues when performing numerical simulations is to be sure that the
system has thermalized before starting making measurements. If equilibrium is not
reached, the observables will drift and yield systematic errors.
In the Monte Carlo evolution, one waits a number of timesteps letting the system
evolve towards equilibrium. Only afterwards the measurements are started. This
number of steps drastically depends on the size of the system and on the temper-
ature. The larger and the colder is the system, the longer we must let the system
evolve to reach thermalization. Many dierent methods have been created to ensure
thermalization, all of them having drawbacks, the worst being that most of the ef-
fective methods are a posteriori, so one has to run the simulation up to a point and
then observe whether the system has thermalized or not.
In this work we have used two thermalization checks:
i) Specic heat check. In this case one checks whether the specic heat computed
as
c =
de)
dT

e
n+1
) e
n
)
T
n+1
T
n
, (9.7)
where e
n
) is the time averaged energy in the thermal bath at temperature T
n
and the last expression gives the discretized derivative, is the same as the one
computed via the uctuation formula of the canonical ensemble
c = N
e
2
) e)
2
T
2
. (9.8)
that holds exclusively at equilibrium.
ii) Symmetry of P
J
(q) under q q for each sample. In equilibrium, the proba-
bility distribution of the overlaps has to be symmetric for our model because
of the Hamiltonian invariance under s
i
s
i
. This check is very strong in
standard Monte Carlo since it corresponds to a ip in the whole sample of
s
i
s
i
, which is the slowest mode of the dynamics. Even if in the Paral-
lel Tempering dynamics this mode needs not to be the slowest one, the check
turns out to be still quite reliable. This method is specially useful and powerful
since it works for one sample (one J
ij
realization) quantities and one does
not need very long runs to perform it.
With these two methods, we obtain a lower boundary for the number of time
steps needed to thermalize for each size to simulate. A number far above this one
is chosen to be sure, and it is used in the successive runs as the starting time of the
measurements.
Critical Point
In numerical simulations of systems of nite size, good tools to determine the critical
temperature are quantities that, in the critical regime where correlation length are
of the order of the system extension, do not depend on the size. Exactly at the
phase transition, therefore, these parameters are expected to take the same value,
irrespectively of the specic size simulated. Actually, in systems of small size, for
which the correlation length soon overcomes their characteristic length, such an
indication is not as precise as for large systems, and nite size corrections should
be taken into account. However, as the size increases, such eects are less and less
important and a proper t of the crossing points of these observables as functions
of temperature with increasing N should provide the true critical temperature.
Any dimensionless observable satises the nite size scaling (FFS) assumption:
O(N, T) = g
_
(T T
c
(N))N
1/
_
(9.9)
with
T
c
(N) = T
c
() +A N
1/
(9.10)
where N is the number of spins (N = L
D
) and is the critical exponent related to
the divergence of the correlation length [T
c
T[
.
The rst parameter we will consider is the Binder cumulant
B =
1
2
_
3
q
4
)
q
2
)
2
_
(9.11)
where (. . .) represents the average over quenched disorder and . . .) the thermal
average.
In systems such as the one we study here, whose Hamiltonian satises the time-
reversal symmetry (TRS) and displaying a Replica Symmetry Broken phase at low
temperature, the Binder cumulant is the proper quantity to identify the transition
point. It grows from zero, in the paramagnetic phase, to unity for T 0, displaying
a steep increase around T
c
. This steepness is enhanced as N .
If a magnetic eld h is added, however, TRS is broken and, due to the lack of
self-averaging in the P(q), the Binder cumulant behaves strangely and is no more a
suitable quantity to investigate the transition to the spin glass phase.
In order to bypass the problem of non-self-averaging two other parameters were
then introduced [M
+
98a, BBDM98, M
+
98b, PPR99]:
A =
q
2
)
2
q
2
)
2
q
2
)
2
, (9.12)
G =
q
2
)
2
q
2
)
2
q
4
) q
2
)
2
(9.13)
Their relationship with B is the following
B = 1
1
2
A
G
(9.14)
For RSB systems such as the one we consider, using the Guerra identity [Gue95]
q
2
)
2
=
1
3
q
4
) +
2
3
q
2
)
2
(9.15)
One can show that G is equal to 1/3 in the whole frozen phase (in the thermodynamic
limit). The parameter A is, instead, a measure of the uctuations of the spin glass
susceptibility
sg
= V q
2
). It vanishes at zero temperature, being a non trivial
increasing function of the temperature as it increases towards T
c
.
Even though for the system at h = 0 this does not provide new information on
the transition temperature, we have also analyzed the behavior in temperature and
size of A and G for the hypercubic spin-glass cell (see Figs. 9.8-9.9), to check and
improve the T
c
estimate.
Ultrametricity in the frozen phase
Studying the structure of the states in the RSB spin-glass phase by means of P(q) one
nds a very special hierarchical structure: the ultrametric structure. In the Parisi
Ansatz this becomes evident in the computation of the joint probability distribution
of the three overlaps among three real replicas, i.e. three dierent congurations of
spins exchanging among them the same interactions: P
J
(q
12
, q
23
, q
13
).
Averaging over quenched disorder we nd
P
J
(q
A
, q
B
, q
C
) = lim
n0
1
n(n 1)(n 2)
(abc)
(Q
ab
q
A
)(Q
bc
q
B
)(Q
ac
q
C
) (9.16)
where (abc) are triples with dierent values. From the analysis of the overlap matrix
Q
ab
in the Parisi Ansatz it comes out that this probability is zero unless, taken three
replicas, at least two overlaps are equal to each other and the third one is equal or
smaller.
If we dene the distance
d

1
N
i
(m
i
m
i
)
2
= 2(q
EA
q
) (9.17)
the ultrametric property is expressed as the inequality
d
max(d
, d
) (9.18)
This allows to set a metric in the states space such that there can be either isosceles
or equilateral triangles.
We can alternatively represent the ultrametric structure setting the correspon-
dence between a pure state and the leaves of a branch of a tree. The distance between
two states will then be proportional to the height at which the two branches, to which
the two states belong (see Fig. 9.6).
q
q
q
0
1
2
L=0
L=1
L=2
1 2 3 4 5 6 7 8
cluster
states
configurations
Figure 9.6: An example of partition of the ultrametric structure on three levels.
The nodes at level 1 correspond to three pure states. The leaves at the lower level
connected to each node are congurations belonging to the corresponding state. The
overlaps q
12
, q
34
, q
45
, q
35
, q
67
, q
78
, q
68
between congurations belonging to the same
pure state take the value q
1
= q
EA
, while overlaps of congurations belonging to
dierent states take the value q
0
< q
1
. q
2
= 1 is the value of the overlap of a
conguration with itself.
We can otherwise separate the space in groups formed by all the points occurring
inside at a certain distance the ones from the others. Each of this groups (clusters)
can still be separated in subgroups containing all the points occurring at a smaller
distance, and so on.
In the ultrametric space, such a partition does not lead to overlapping and each
subgroup exclusively belongs to one group hierarchically above it. The sequence of
decreasing distances taking place in this renement of the states space partition is
expressed, in terms of overlaps, as the increasing sequence
q
0
q
1
... q
A
l
1
q
A
l
= 1. (9.19)
where ^
l
is the number of levels. If we consider, for instance, three levels (see Fig.
9.6), in general q
2
will be equal to 1 (overlap of a conguration with itself) and q
1
will be equal to q
EA
(overlap of a pure state with itself). The second level is the
one of pure states: at each state belong the congurations represented by the leaves
branching from it.
The computation of Eq. (9.16) leads to
P
J
(q
A
, q
B
, q
C
) =
1
2
P(q
A
) x(q
A
)(q
A
q
B
)(q
A
q
C
) + (9.20)
+
1
2
P(q
A
)P(q
B
)(q
A
q
B
)(q
B
q
C
) +
+
1
2
P(q
B
)P(q
C
)(q
B
q
C
)(q
C
q
A
) +
+
1
2
P(q
C
)P(q
A
)(q
C
q
A
)(q
A
q
B
).
clearly showing the tree properties above described, with x(q
A
) =
_
q
A
0
dqP(q).
Ultrametricity is a very peculiar property of disordered systems and it is yet an
open and widely discussed problem to understand whether such highly non-trivial
structure exists also beyond mean-eld theory, in terms of which it was originally
derived. It is in any case based on a strong time-scale separation.
We now introduce useful tools to identify and analyze the possible ultrametric
properties of the spin-glass phase in numerical simulations of any kind of model
without assuming any mean-eld behaviour. To this aim, it is necessary to simulate
simultaneously three independent replicas of the system and to measure the three
overlaps one can build with them.
A rst method is to verify the Guerra relations [Gue95]:
q
2
)
2
=
2
3
q
2
)
2
+
1
3
q
4
)
q
2
q
t2
) =
1
2
q
2
)
2
+
1
2
q
4
) (9.21)
where q e q
t
are any two of the three possible overlaps.
An alternative study can be performed introducing other observables, func-
tions of the three overlaps. For instance those introduced as Binder cumulants
in Ref. [IPRL96]:
B
qqq

[q
12
q
13
q
23
[)
q
2
)
3/2
B
t
qqq

q
12
q
13
q
23
)
q
2
)
3/2
(9.22)
B
qq

_
([q[ [q
t
[)
2
_
q
2
M
)
B
t
qq

_
(q q
t
sign(q
M
))
2
_
q
2
M
)
(9.23)
where q
M
is dened as
q
M
maxq
12
, q
13
, q
23
(9.24)
and q and q
t
are the other two overlap values (q, q
t
q
M
).
III.9.3 Numerical Results 129
Since they are size independent observables, at criticality these can also be em-
ployed to give an estimate of the critical temperature and of the critical index.
Above all, however, their low T behavior can be compared with the theoretical
prevision for ultrametric systems.
As the original Binder cumulant Eq. (9.11) also these cumulants do not depend
on the index and the generic scaling behavior at nite size goes as in Eq. (9.9).
Let us consider the cumulant B
qqq
. Its behavior can also be computed analyt-
ically assuming an RSB phase, starting from P(q). Carrying out the average over
the joint probability distribution of three overlaps, given in Eq. (9.16) one obtains:
B
qqq
=
3
2
_
1
0
dxq
2
(x)
_
1
x
dyq(y) +
1
2
_
1
0
dx x q
3
(x)
_
_
1
0
dxq
2
(x)
_
3/2
. (9.25)
This is a function of T tending asymptotically to 1 as T 0. If the structure of the
space of the states is actually ultrametric the values of B
qqq
computed directly from
the data of numerical simulations would also display the same behavior in the deep
frozen phase.
9.3 Numerical Results
We studied systems with dimensions D = 6, 7, 8, 9, 10, 12 and 14 (number of spins
N = 2
D
). We were running three replicas in parallel since we wanted to study the
ultrametricity properties of the model. In order to have good acceptance rates in
the parallel tempering technique even for large sizes we have employed 40 dierent
temperatures in the range [0.49, 1.06]. The range is such that we have some systems
in the disordered phase and we can reach temperatures as low as 0.5. The number
of samples for the system of dimension 6, 7, 8, 9, 10 is a few hundreds (600 800),
for dimension 12 a hundred. The D=14 run is in the temperature range [0.69, 1.06]
and we simulated 50 samples.
9.3.1 Phase Transition
In order to study the phase transitions in the gures 9.7, 9.8, 9.9 we show the plots
for B, A and G for all sizes studied. From the crossing points of the cumulants at
dierent sizes we can infer the transition temperature as D . We show the
results for the Binder cumulant in table 9.1.
The estimates of the critical temperature and the index are: T
c
= 0.96(3)
and = 2.4(6). We perform a nite size scaling of the crossing points of the
Binder cumulant to get the phase transition temperature in the thermodynamic
limit, N . It scales as Eq. (9.10). From the analogue crossings of the parameters
G and A at dierent sizes we respectively get the estimates
T
c
= 0.98(1) = 3.2(5) (9.26)
T
c
= 0.96(5) = 3(1) (9.27)
Dimension T
c
(Nspins)
6-8 0.763 0.002
7-9 0.808 0.002
8-10 0.849 0.002
10-12 0.911 0.005
12-14 0.917 0.005
0.96 0.03
Table 9.1: Size dependent transition temperature.
The data are taken from the crossing point of the
Binder parameter, Eq. (9.11). The thermody-
namic limit value is yielded by tting the data with
the FSS Eq. (9.10).
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.4 0.5 0.6 0.7 0.8 0.9 1 1.1
B
T
D=6
D=7
D=8
D=9
D=10
D=12
D=14
Figure 9.7: The Binder cumulant at sizes
2
D
, D = 6, 7, 8, 9, 10, 12. As the size in-
creases the steepness of B in the frozen
phase increases.
0
0.05
0.1
0.15
0.2
0.25
0.3
0.4 0.5 0.6 0.7 0.8 0.9 1 1.1
A
T
D=6
D=7
D=8
D=9
D=10
D=12
D=14
Figure 9.8: The A Parameter for the
same cases as in g 9.7
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4 0.5 0.6 0.7 0.8 0.9 1 1.1
G
T
D=6
D=7
D=8
D=9
D=10
D=12
D=14
Figure 9.9: The G Parameter for the
same cases as in g 9.7
The G parameter becomes steeper and steeper as D increases (see Fig. 9.9) and
the curves behaviour is consistent with the limit of 1/3 for T < T
c
except for the
higher sizes, showing the lack of enough sampling [PRS02]. The SG susceptibility
uctuations A, already reach a maximum in the critical region here investigated (for
D > 8 in Fig 9.8) and start decreasing for decreasing T, as it is expected. The peak
shift towards the crossing region as D increases.
9.3.2 Low temperature behaviour
We expect the hypercubic spin-glass cell to approach the SK model for large dimen-
sions, i.e. to have the same limit of the hypercubic lattice (the Edwards-Anderson
model). We have computed the overlap distribution function
P
N
(q) = (q q
12
(t)) (9.28)
with
q
12
(t) =
1
N
N
i=1
1
i
2
i
(9.29)
where
(1,2)
i
the ith spin in the replica (1, 2). We can compute three overlaps at
each Monte Carlo step since we have three replicas. We use two of them (using all
three could induce correlations in the results) to increase statistics in all calculations
involving one overlap quantities. In gures 9.10-9.13 we plot P
N
(q) at T = 0.69 for
dierent dimensions ranging from 6 to 14 and at T = 0.649, 0.610, 0.488 for D ranging
from 6 to 12. At higher temperature, as soon as a peak is distinguishable, one notices
that it is at a value of q below the SK value of q
EA
at that temperature (also shown in
the gure 9.10) as opposed to what happens in the SK model as N increases [You83].
Looking at the P
N
(q) it is possible to see that it eventually tends, for large N, to the
analytical P(q) computed for the SK model in the thermodynamic limit. In Fig. 9.14
the evolution of the maximum of P
N
(q) (q
max
) is plotted for dierent temperatures.
One can see there clearly the trend towards q
EQ
. Conversely, for lower temperature,
e.g. T = 0.488 in Fig. 9.10, we can see that, at the simulated sizes, the peak of
P
N
(q) tends to its innite dimension limit from above (in magnitude), but becomes
lower than it around D = 8.
A crossover between the two dierent behaviors has to take place in between. To
understand better the situation we have to look at intermediate temperatures, trying
to see evidence of a turning point in the range of D values at our disposal. In Figs.
9.11-9.12 we show the probability distributions at T = 0.61 and T = 0.649 and in
Fig. 9.15 the behavior of the q
max
as a function of the size around these temperature
values. Even though the uncertainty on the peak is large, it is still possible to observe
that a crossover occurs, but data are not rened enough to determine the D
cross
(T)
at which it takes place.
What happens is that, at a given temperature, the system starts evolving towards
the mean-eld limit as in the fully connected model. In the hypercubic cell the
connectivity is z = D, whereas the number of particles is N = 2
D
. In the low
D regime, where the dilution is small (z/N not too small), nite size corrections
happen to be qualitatively similar to the nite size correction of the SK model.
What matters in this case is, thus, the number of particles. At some (temperature
dependent) value D, the peak of the P(q) overcomes the fully connected limit of the
hypercubic lattice. Increasing the dimension further, the system eventually stops its
monotonous tendency and turns back towards the SK value. In this second regime,
thus, the geometry (meaning here the way the spins are connected, or, equivalently,
the dimension of the cell) plays the most important role.
0
0.5
1
1.5
2
2.5
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
P
(
q
)
q
T = 0.488
q
EA
= 0.65824
x
c
= 0.48225
Figure 9.10: The overlap probability dis-
tribution P(q) at T = 0.488 from D = 6
(widest curve) to D = 12 (most peaked
curve). Also the analytical P(q) for the
SK model at this temperature is plotted.
The delta functions at q
EA
are repre-
sented by vertical lines.
tribution P(q) at T = 0.61 for D =
6, 7, 8, 9, 10, 12 (from widest to narrow-
est). Also the analytical P(q) for the SK
model at this temperature is plotted. The
delta functions at q
EA
are represented
by vertical lines.
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
P
(
q
)
q
T = 0.610
q
EA
= 0.49308
x
c
= 0.43313
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
P
(
q
)
q
T = 0.649
q
EA
= 0.43877
x
c
= 0.41027
6, 7, 8, 9, 10 and 12 (from widest to nar-
rowest). Also the analytical P(q) for the
SK model at this temperature is plotted.
The delta functions at q
EA
are repre-
sented by the vertical lines.
At T = 0.69 the crossover D is so small that cannot be even distinguished and the
system already for D = 6 is in the regime driven by nite dimensional corrections.
At T = 0.488 the temperature is so low that, for the sizes that we were able to
simulate, we always stay in the regime for which the corrections to the values of the
system mostly depend of its nite size. For the size range we have, the crossover is
visible only for T [0.61 : 0.662].
Already in Ref. [PRR91] such a behavior was conjectured, even though no precise
evidence could be found numerically. This very particular feature of the spin-glass
hypercubic cell, as opposed to the lattice usually studied, shows up in almost every
thermodynamic observable. Rather than in P(q), the just mentioned crossover is,
6, 7, 8, 9, 10, 12 and 14 (from widest to
narrowest). Also the analytical P(q) for
the SK model at this temperature is plot-
ted. The delta functions on q
EA
are rep-
resented by the vertical lines.
0
0.5
1
1.5
2
2.5
3
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
P
(
q
)
q
T = 0.69
q
EA
= 0.38338
x
c
= 0.38328
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
100 1000 10000
q
m
a
x
N
T = 0.488
T = 0.61
T = 0.69
Figure 9.14: The maximum value of
the P(q) versus N = 2
D
for D =
6, 7, 8, 9, 10, 12, 14 at T [0.488 : 0.690].
The value of q
EA
for the SK model is also
plotted as reference on the right side of
the gure. A crossover takes place be-
tween T = 0.61 and T = 0.69.
0.2
0.25
0.3
0.35
0.4
0.45
0.5
0.55
100 1000 10000
q
m
a
x
N
T = 0.61
T = 0.649
T = 0.676
Figure 9.15: Crossover region for the
maximum value of the P(q) versus N =
2
D
for D = 6, 7, 8, 9, 10, 12 at T [0.610 :
0.676]. The value of q
EA
for the SK model
is also plotted as reference on the right
side of the gure.
however, clear in self-averaging quantities, that we will now analyze.
To study the possible convergence to the SK model, as D , we compute the
local susceptibility
=
_
1
_
dq P(q) q
_
(9.30)
which in the SK model at vanishing external eld is always equal to unity for any
T T
c
, and the internal energy that, in the SK model, is known to be equal to
U
N
=
2
_
1
_
dq P(q) q
2
_
(9.31)
In Ref. [PRR91], the susceptibility versus D (at T = 0.5) was increasing monoton-
ically. Rening the simulation we have found that the increase is blocked at some
given size and a slight turning occurs also in this observable towards the SK value.
The phenomenon is clearer at T = 0.69 than at T = 0.488, since more statistics
and an extra size D = 14 is available. At T = 0.69 the turnover is clear around
D = 10, at lower temperature this turning point moves to higher dimensions and
we cannot see it anymore due to the lack of higher sizes for that temperatures. See
g. 9.16. To better show this eect, in g. 9.17 we plot the temperature where the
turnover occurs versus the dimension. As one notices, also for high temperature the
dimension at which the (D) turns downwards increases outside the range of the
simulated sizes.
0.94
0.96
0.98
1
1.02
1.04
1.06
1.08
1.1
6 7 8 9 10 11 12 13 14
S
u
s
c
e
p
t
i
b
i
l
i
t
y
D
T=0.5
T=0.6
T=0.69
Figure 9.16: Susceptibility for the dier-
ent sizes studied at T = 0.5, 0.6, 0.69
9
10
11
12
13
14
15
0.6 0.65 0.7 0.75 0.8 0.85
D
T
Max. Susceptibility
Figure 9.17: Sketch of the turnover of the
susceptibility. Both ends of the line ex-
ceed the limiting sizes of the simulation
we performed thus are only indicative
For the internal energy behaviour, Eq. (9.31) we show the behaviour of U +
(1 < q
2
>)/2 between T = 0.488 and T = 0.69. This value has to go to 0 for
innite dimensions if it has to approach the SK model. We show the plot in gure
9.18. At T = 0.488 the deviation from the SK value is small but yet not decreasing,
whereas at T = 0.69 one can clearly see that the turning point in the size dependence
has been overcome, and the deviation decreases to zero. In g. 9.19 we plot the
values of D (at each possible T) at which the deviation from the energy value of the
SK model starts to decrease to zero. Also in this case the maximum deviation point
in D shifts towards high D both for low and high T.
The energy dependence on the size of the system shows the expected behaviour.
The energy is self-averaging and so it does not suer from big uctuations. Already in
Ref. [PRR91] a very good agreement with SK was found. Its behaviour is U N
1/2
(Fig. 9.20). The t gives a value for the innite size system of U = 0.7048(3) at T =
0.5, 0.6683(6) at T = 0.61 and 0.6364(8) at T = 0.69. The theoretically predicted
values for the SK model are U = 0.71147, 0.67648 and 0.64558 respectively (see
Fig. 9.20).
In gure 9.21 we plot the size dependence of the spin-glass susceptibility per unit
volume, q
2
), again at T = 0.5, 0.61 and 0.69. In the SK model this value is equal
to q
2
) = 0.28853, 0.17469, 0.10901 respectively. The behaviour is the one expected
for the SK model, as it scales with N
1/2
. However, it does not extrapolate to the
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05
0.07 0.08 0.09 0.1 0.11 0.12 0.13 0.14 0.15 0.16 0.17
U

-

b
e
t
a
/
2

(
1
-
<
q
2
>
)
1/D
T=0.488
T=0.57
T=0.69
Figure 9.18: The parameter U +
/2(1 < q
2
>) at T = 0.488, 0.57 and
0.69. versus 1/D = ln 2/ ln N.
8
9
10
11
12
13
0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 1.05
D
T
Figure 9.19: Sketch of the turnover of the
parameter U +/2(1 < q
2
>)
-0.72
-0.7
-0.68
-0.66
-0.64
-0.62
-0.6
-0.58
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14
E
n
e
r
g
y
N
-1/2
T=0.5
T=0.61
T=0.69
Figure 9.20: Energy at T=0.5, 0.61, 0.69.
The lines are linear ts and the arrows
correspond to the SK value
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14
<
q
2
>
N
-1/2
T=0.5
T=0.61
T=0.69
Figure 9.21: The mean square overlap
q
2
) at T = 0.5, 0.61, 0.69 and the ar-
rows correspond to the SK value
correct SK value. At T = 0.69 since we have the point at D = 14 we can see this
turn towards the right value. This turn is seen clearly at D = 12 for T < 0.78. For
lower temperatures the turn is expected to appear at higher dimensions. No further
movement of the minimum can be seen with the present simulation. If we x T at
low D the scaling is clearly the one of SK while at higher D a crossover takes place.
9.3.3 Evidence for Ultrametricity
We have simulated the dynamics at equilibrium of three replicas at the same time
in order to study the possible ultrametric properties of the system. We have studied
both the cumulants given in Eqs. (9.22)-(9.23) and the Guerra relations, Eqs. (9.15,
9.21).
We show in gure 9.22 and 9.23 the plots for these cumulants for the dierent
system sizes. The curves approach unity as expected for low temperatures [IPRL96]
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
1.05
0.4 0.5 0.6 0.7 0.8 0.9 1 1.1
B
q
q
q
T
D=6
D=7
D=8
D=9
D=10
D=12
D=14
Figure 9.22: 3 replica cumulant B
qqq
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
0.4 0.5 0.6 0.7 0.8 0.9 1 1.1
B
q
q
q
T
D=6
D=7
D=8
D=9
D=10
D=12
D=14
Figure 9.23: 3 replica cumulant B
t
qqq
for all simulated sizes. The slope of the curve becomes steeper and steeper around
the transition region as the size increases, again agreeing with expectation.
9.4 Conclusion
For very high dimension, the spin glass cell model is expected to approach the mean-
eld SK model, corresponding to the innite dimensional hypercubic lattice spin-
glass. The mean-eld behaviour in the hypercubic lattice is known, more specically,
to hold for any D 8 [DKT91]. In the hypercubic cell the number of spins and
the dimension are not independent and in this respect it is dierent from a D-
dimensional Edwards-Anderson spin-glass. Moreover, no renement is possible at
small dimensions, since low D means few spins. This dierence should vanish as
both models tend to the thermodynamic limit in high dimension.
Analyzing the behavior of dierent observables as the size of the simulated sys-
tems increases we can identify two regimes. (i) For small sizes the system seems
to monotonously tend to the SK model limit, with deviations mainly due to the
nite size and that can be parameterized by means of nite size corrections. As D
continues to increase, however, the SK limit is overcome. (ii) At some temperature
dependent D the monotonous behaviour stops and the observables turn back, quite
likely to the corresponding value in the SK model.
We have been able to identify quite clearly this phenomenon for T down to 0.69
where we have data up to D = 14. As temperature is decreased the turning point
moves towards larger and larger systems. Thus, already at T 0.662 we cannot
determine any crossover dimension because it is at or beyond the highest simulated
dimension D = 12.
All by all, we observe a conrmation of the expected convergence of the Hyper-
cubic Cell model with 2
D
spins towards the SK model.
Bibliography
[A
+
03] J.H. Ardenkjet al. Increase in signal-to-noise ratio of > 10, 000 times
in liquid-state NMR. PNAS, 100:10158, 2003.
[AA88] D.P. Arovas and A. Auerbach. Functional integral theories of low-
dimensional quantum heisenberg models. Phys. Rev. B, 38(1):316,
1988.
[Abr61] A. Abragam. Principles of Nuclear Magnetism. Clarendon Press, Ox-
ford, 1961.
[AG82] A. Abragam and M. Goldman. Nuclear Magnetism. Clarendon Press,
Oxford, 1982.
[Ali79] R. Alicki. The quantum open system as a model of the heat engine. J.
Phys. A, 12:L103, 1979.
[AN02] A.E. Allahverdyan and Th.M. Nieuwenhuizen. A mathematical theo-
rem as the basis for the second law: Thomsons formulation applied to
equilibrium. Physica A, 305:542, 2002.
[ASNa] A.E. Allahverdyan, R. Serral Gracià, and Th.M. Nieuwenhuizen. Bath
assisted cooling of spins. Unpublished.
[ASNb] A.E. Allahverdyan, R. Serral Gracià, and Th.M. Nieuwenhuizen. Work
extraction in the spin-boson model. Unpublished.
[Aue94] A. Auerbach. Interacting electrons and quantum magnetism. Springer-
Verlag, 1994.
[B
+
98] N.H. Bonadeo et al. Coherent optical control of the quantum state of
a single quantum dot. Science, 282:1473, 1998.
[B
+
02] P.O. Boykin et al. Algorithmic cooling and scalable NMR quantum
computers. PNAS, 99:3388, 2002.
[Bal92] R. Balian. From Microphysics to Macrophysics. Springer-Verlag, 1992.
138 Bibliography
[Bas78] I.M. Bassett. Alternative derivation of the classical second law of ther-
modynamics. Phys. Rev. A, 18:2356, 1978.
[BBDM98] H. Bokil, A.J. Bray, B. Droseel, and M.A. Moore. Comment on: A
general method to determine replica symmetry breaking transitions.
Phys. Rev. Lett., 82:5174, 1998.
[Bel91] M. Le Bellac. Quantum and statistical eld theory. Oxford university
press, 1991.
[Ber87] G. Bergmann. Ruderman-Kittel-Kasuya-Yosida interaction in disor-
dered metals. Phys. Rev. B, 36:2469, 1987.
[BGZJ76] E. Brezin, J.C. Le Guillou, and J. Zinn-Justin. Field theoretical ap-
proach to critical phenomena. In C. Domb and M.S. Green, editors,
Phase transitions and critical phenomena, volume 6. Academic press,
1976.
[BK52] T.H. Berlin and M. Kac. The spherical model of a ferromagnet. The
Phys. Rev., 86(6):821, 1952.
[BK77] G.N. Bochkov and Y.E. Kuzovlev. On general theory of thermal uc-
tuations in non-linear systems. Sov. Phys. JETP, 45:125, 1977.
[BK79] G.N. Bochkov and Y.E. Kuzovlev. Fluctuation-dissipation relations for
nonequilibrium processes in open systems. Sov. Phys. JETP, 49:543,
1979.
[BN91] B.A. Berg and T. Neuhaus. Multicanonical algorithms for rst order
phase transitions. Phys. Lett. B, 267:249, 1991.
[BN92] B.A. Berg and T. Neuhaus. Multicanonical ensemble: a new approach
to simulate rst order phase transitions. Phys. Rev. Lett., 68:9, 1992.
[BP86] L.N. Bulaevskii and S. V. Panyukov. RKKY interaction in metals with
impurities. JETP Lett., 43:240, 1986.
[BP02] H.-P. Breuer and F. Petruccione. The theory of Open Quantum Sys-
tems. Oxford University Press, Oxford, 2002.
[C
+
90] G.D. Cates et al. Laser production of large nuclear-spin polarization
in frozen xenon. Phys. Rev. Lett., 65:2591, 1990.
[CL83] A.O. Caldeira and A.J. Leggett. Quantum tunnelling in a dissipative
system. Ann. Phys., 149:374, 1983.
[CM72] V. Cannella and J. A. Mydosh. Magnetic ordering in gold-iron alloys.
Phys. Rev. B, 6:4220, 1972.
[CS92] A. Crisanti and H.J. Sommers. The spherical p-spin interaction spin
glass model:the statics. Z.Phys. B, 87:331354, 1992.
Bibliography 139
[CZ95] J. I. Cirac and P. Zoller. Quantum computations with cold trapped
ions. Phys. Rev. Lett., 74:4091, 1995.
[DKT91] C. De Dominicis, I. Kondor, and T. Temesvari. Short-range corrections
to the order parameter of the ising spin glass above the upper critical
dimension. J. Phys. A: Math. Gen., 24:L301L308, 1991.
[DSDA77] S. Dattagupta, G. K. Shenoy, B. D. Dunlap, and L. Asch. Breakdown of
the white-noise approximation in the mossbauer relaxation spectra:
The case of cs2naybcl6. Phys. Rev. B, 16:3893, 1977.
[EA75] S.R. Edwards and P.W. Anderson. Theory of spin glasses. J. Phys.
(France), 5:965, 1975.
[EBW87] R.R. Ernst, G. Bodenhausen, and A. Wokaun. Principles of Nuclear
magnetic Resonance in One and Two Dimensions. Clarendon Press,
Oxford, 1987.
[EMSKB03] J. Eschner, G. Morigi, F. Schmidt-Kaler, and R. Blatt. Laser cooling
of trapped ions. J. Opt. Soc. Am. B, 20:1003, 2003.
[Esq98] P. Esquinazi, editor. Tunneling systems in amorphous and crystaline
materials. Springer-Verlag, 1998.
[FH65] R.P. Feynman and R. Hibbs. Quantum mechanics and path integrals.
McGraw-Hill, New York, 1965.
[FH91] K.H. Fischer and J.A. Hertz. Spin Glasses. Cambridge University
Press, Cambridge, UK, 1991.
[FK00] T. Feldmann and R. Koslo. Performance of discrete heat engines and
heat pumps in nite time. Phys. Rev. E, 61:4774, 2000.
[FK03] T. Feldmann and R. Koslo. Quantum four-stroke heat engine: Ther-
modynamic observables in a model with intrinsic friction. Phys. Rev.
E, 68:016101, 2003.
[Gar85] E. Gardner. Spin glasses with p-spin interactions. Nucl. Phys. B,
257:747, 1985.
[GC97] N.A. Gershenfeld and I. Chuang. Bulk spin-resonance quantum com-
putation. Science, 275:350, 1997.
[GJ79] M.R. Garey and D.S. Johnson. Computers and Intractability: a guide
to the theory of NP-completeness. Freeman, New York, 1979.
[GK94] E. Geva and R. Koslo. Three-level quantum amplier as a heat engine:
A study in nite-time thermodynamics. Phys. Rev. E, 49:3903, 1994.
140 Bibliography
[GMY90] A. Georges, M. Mezard, and J.S. Yedida. Low-temperature phase of
the Ising spin glass on a hypercubic lattice. Phys. Rev. Lett, 64:2937,
1990.
[Gue95] F. Guerra. About the overlap distribution in mean eld spin glass
models. J. Mod. Phys. B, 1995.
[H
+
97] D.A. Hall et al. Polarization-enhanced NMR spectroscopy of
biomolecules in frozen solution. Science, 276:930, 1997.
[H
+
02] T.E. Humphrey et al. Reversible quantum brownian heat engines for
electrons. Phys. Rev. Lett., 89:116801, 2002.
[Hah50] E.L. Hahn. Spin echoes. Phys. Rev., 80:580, 1950.
[HCH02] J. He, J. Chen, and B. Hua. Quantum refrigeration cycles using spin-
(1/2) systems as the working substance. Phys. Rev. E, 65:036145, 2002.
[Her76] J.A. Hertz. Quantum critical phenomena. Phys. Rev. B, 14(3):1165,
1976.
[HG98] P. Hanggi and M. Grifoni. Driven quantum tunneling. Phys. Rep.,
304:229, 1998.
[I
+
03] A. Imamoglu et al. Optical pumping of quantum-dot nuclear spins.
Phys. Rev. Lett., 91:017402, 2003.
[ID89] C. Itzykson and J.-M. Droue. Statistical eld theory. Vol 1. Cambridge
university press, 1989.
[IPRL96] D. I niguez, G. Parisi, and J.J. Ruz-Lorenzo. Simulation of three-
dimensional ising spin glass model using three replicas: study of binder
cumulants. J. Phys. A: Math. Gen., 29:4337, 1996.
[Joy72] G.S. Joyce. Critical properties of the spherical model. In C. Domb
and M.S. Green, editors, Phase transitions and critical phanomena,
volume 2, chapter 10. Academic Press, 1972.
[KA00] J.M. Kikkawa and D.D. Awschalom. All-optical magnetic resonance in
semiconductors. Science, 287:473, 2000.
[Kac64] M. Kac. The work of T.H. Berlin in statistical mechanics: A personal
reminiscence. Phys. Today, 17(10):4042, 1964.
[KC87] M.I. Kaganov and A.V. Chubukov. Interacting magnons. Usp. Fiz.
Nauk, 153:537, 1987.
[Kei87] J. Keizer. Statistical Thermodynamics of Nonequilibrium Processes.
Springer-Verlag, 1987.
Bibliography 141
[KKPS92] A.M. Khorunzhy, B.A. Khoruzhenko, L.A. Pastur, and M.V.
Shcherbina. The large-n limit in statistical physics and the theory of
disordered systems. In Phase transitions and critical phenomena, vol-
ume 15, chapter 2. Academic Press, 1992. C. Domb and J.L. Lebowitz.
[Kno73a] H.J.F. Knops. Antiferromagnetic spherical model and the spin dimen-
sionality limit. Phys. Rev. B, 8(9):42094218, 1973.
[Kno73b] H.J.F. Knops. Innite spin dimensionality limit for nontranslationally
invariant interactions. J. Math. Phys., 14(2):19181920, 1973.
[Koc92] O. Kocharovskaya. Amplication and lasing without inversion. Phys.
Rep., 219:175, 1992.
[KP92] W. Ketterle and D.E. Pritchard. Atom cooling by time-dependent po-
tentials. Phys. Rev. A, 46:4051, 1992.
[KS85] J. Klauder and B. Skagerstam. Coherent states: applications in physics
and mathematical physics. Singapore: World Scientic, 1985.
[KTJ76] J.M. Kosterlitz, D.J. Thouless, and R.C. Jones. Spherical model of a
spin-glass. Phys. Rev. Lett., 36:1217, 1976.
[L
+
87] A.J. Leggett et al. Dynamics of the dissipative two-state system. Rev.
Mod. Phys., 59:1, 1987.
[Lam68] G. Lampel. Nuclear dynamic polarization by optical electronic satura-
tion and optical pumping in semiconductors. Phys. Rev. Lett., 20:491,
1968.
[Leg85] A.J. Leggett. Quantum tunneling in the presence of an arbitrary linear
dissipation mechanism. Phys. Rev. B, 30:1208, 1985.
[Len78] A. Lenard. Thermodynamical proof of the Gibbs formula for elementary
quantum systems. J. Stat. Phys., 19:575, 1978.
[LL87] L.D. Landau and E.M. Lifshitz. Statistical Physics, I. Pergamon Press
Oxford, 1987.
[Luc90] J. Luczka. Spin in contact with thermostat: exact reduced dynamics.
Physica A, 167:919, 1990.
[M
+
98a] E. Marinari et al. A general method to determine replica symmetry
breaking transitions. Phys. Rev. Lett., 81:1698, 1998.
[M
+
98b] E. Marinari et al. Reply to comment on: A general method to determine
replica symmetry breaking transitions. Phys. Rev. Lett., 82, 1998.
[Ma82] S.K. Ma. Modern theory of critical phenomena. Benjamin/Cummings,
Reading, 1982.
142 Bibliography
[MPV87] M. Mezard, G. Parisi, and M.A. Virasoro. Spin glass theory and beyond.
World scientic, 1987. Book with introduction and reprints of the
original articles.
[MS80] K. Mohring and U. Smilansky. A semi-classical treatment of dissipative
process based on feynmans functional method. Nucl. Phys., A338:227,
1980.
[Myd93] J.A. Mydosh. Spin Glasses. An experimental introduction. Taylor &
Francis, 1993.
[N
+
02] Y. Nakamura et al. Charge echo in a cooper-pair box. Phys. Rev. Lett.,
88:047901, 2002.
[Nie93] Th.M. Nieuwenhuizen. Field theory for site-disordered spin glasses.
Europhys. Lett., 24:797, 1993.
[Nie95a] Th.M. Nieuwenhuizen. Exactly solvable models of a quantum spin glass.
Phys. Rev. Lett., 74:42894292, 1995.
[Nie95b] Th.M. Nieuwenhuizen. Quantum description of spherical spins. Phys.
Rev. Lett., 74:42934296, 1995.
[NKH79] S. Nagata, P.H. Keesom, and H.R. Harrison. Low-dcleld susceptibility
of CuMn spin glass. Phys. Rev. B, 19:1633, 1979.
[NO98] J.W. Negele and H. Orland. Quantum many-particle systems. Addison-
Wesley, 1998.
[NR98] Th.M. Nieuwenhuizen and F. Ritort. Quantum phase transition in spin
glasses with multi-spin interactions. Physica A, 250:845, 1998.
[NvD99] Th.M. Nieuwenhuizen and C.N.A. van Duin. Theory of site-disordered
magnets. Eur. Phys. B, 7:191, 1999.
[Obe72] G. Obermair. Dynamical aspects of critical phenomena. Gordon and
Breach, New York, 1972. J.I.Budnick and M.P. Kawatra (Eds.).
[Par79] G. Parisi. Innite number of order parameters for spin-glasses. Phys.
Rev. Lett, 43:1754, 1979.
[Par88] G. Parisi. Statistical eld theory. New York, Addison Wesley, 1988.
[Phi81] W.A. Phillips, editor. Amorphous solids. Low temperature properties.
Springer-Verlag, 1981.
[PPR99] G. Parisi, M. Picco, and F. Ritort. Continous phase transition in a
spin glass model without time-reversal symmetry. Phys. Rev. E, 60:58,
1999.
Bibliography 143
[PRR91] G. Parisi, F. Ritort, and J.M. Rub. Numerical results on a hypercubic
cell spin glass model. J. Phys. A Math. Gen., 24:5307, 1991.
[PRS02] M. Picco, F. Ritort, and M. Sales. Order-parameter uctuations in spin
glasses: Monte carlo simulations and exact results for small sizes. Eur.
Phys. J. B, 19:565, 2002.
[PSE96] G. M. Palma, K.-A. Suominen, and A. Ekert. Quantum computers and
dissipation. Proc. R. Soc. Lond. A, 452:567, 1996.
[PW78] W. Pusz and S.L. Woronowicz. Passive states and KMS states for
general quantum systems. Comm. Math. Phys., 58:273, 1978.
[S
+
03] M.O. Scully et al. Extracting work from a single heat bath via vanishing
quantum coherence. Science, 299:862, 2003.
[Sac99] S. Sachdev. Quantum phase transitions. Cambridge Univ. Press, 1999.
[SB90] S. Sachdev and R.N. Bhatt. Bond-operator representation of quantum
spins: Mean-eld theory of frustrated quantum heisenberg antiferro-
magnets. Phys. Rev. B, 41(13):9323, 1990.
[Scu01] M. Scully. Extracting work from a single thermal bath via quantum
negentropy. Phys. Rev. Lett., 87:220601, 2001.
[SG92a] M.R.A. Shegelski and D.J.W. Geldart. Indirect-exchange interactions
in disordered metals at nite temperature. Phys. Rev. B, 46:5318, 1992.
[SG92b] M.R.A. Shegelski and D.J.W. Geldart. Theory of impurity-
concentration dependence of freezing temperatures of metallic spin
glasses. Phys. Rev. B, 46:2853, 1992.
[Sie71] A.E. Siegman. Introduction to Lasers and Masers. McGraw Hill, 1971.
[SK75] D. Sherrington and S. Kirkpatrick. Solvable model of a spin-glass. Phys.
Rev. Lett., 35:1792, 1975.
[SL04] K. Shiokawa and D.A. Lidar. Dynamical decoupling using slow pulses:
Ecient suppression of 1/f noise. Phys. Rev. A, 69:030302, 2004.
[Sli90] C.P. Slichter. Principles of Magnetic Resonance. Springer, Berlin, 1990.
[SN04] R. Serral Gracià and Th.M. Nieuwenhuizen. Quantum spherical spin
models. Phys. Rev. E, 69:056119, 2004. [cond-mat/0304150].
[SNL04] R. Serral Gracià, Th.M. Nieuwenhuizen, and I.V. Lerner. Concentra-
tion dependence of the transition temperature in metallic spin glasses.
Eur. Phys. Lett., 66(3):419, 2004. [cond-mat/0311491].
[Sr89] O.W. Srensen. Polarization transfer experiments in high-resolution
nmr spectroscopy. Prog. NMR Spectr., 21:503, 1989.
144 Bibliography
[SS66] J. Schmidt and I. Solomon. High-sensitivity magnetic resonance by
bolometer detection. J. Appl. Phys., 37:3719, 1966.
[SS75] D. Serrington and B.W. Southern. Spin glass versus ferromagnet. J.
Phys. F, 5:L49, 1975.
[Sta68] H.E. Stanley. Spherical model as the limit of innite spin dimensional-
ity. Phys. Rev., 176(2):718, 1968.
[Sto] H.T.C. Stoof. Statistical eld theory. Downloadable from webpage:
http://www.phys.uu.nl/~stoof/SFT.pdf.
[SZ97] M. Scully and S. Zubairy. Quantum Optics. Cambridge University
Press, 1997.
[Thi83] W. Thirring. A Course in Mathematical Physics 4: Quantum mechan-
ics of large systems. Springer, Vienna, 1983.
[TJ00] C. Timm and P.J. Jensen. Schwinger boson theory of anisotropic fer-
romagnetic ultrathin lms. Phys. Rev. B, 62(9):5634, 2000.
[TV77] G.M. Torrie and J.P. Valleau. Nonphysical sampling distributions in
monte carlo free-energy estimation: umbrella sampling. J. Comp.
Phys., 23:187, 1977.
[Unr95] W. Unruh. Maintaining coherence in quantum computers. Phys. Rev.
A, 51:992, 1995.
[VL98] L. Viola and S. Lloyd. Dynamical supression of decoherence in two-
state quantum systems. Phys. Rev. A, 58:2733, 1998.
[Voj96] T. Vojta. Quantum version of a spherical model: Crossover from quan-
tum to classical critical behavior. Phys. Rev. B, 53(2):710, 1996.
[Voj00a] T. Vojta. Quantum phase transitions, September 2000. Contribution to
the lecture notes of the Heraeus summer school on Statistical Physics,
Chemnitz, cond-mat/0010285.
[Voj00b] T. Vojta. Quantum phase transitions in electronic systems. Ann. Phys.
Berlin, 9(6):403, 2000.
[VS85] D.C. Vier and S. Schultz. Evidence of multiple mechanisms contributing
to the transition temperature in metallic spin-glasses. Phys. Rev. Lett.,
54:150, 1985.
[War97] W.S. Warren. The usefulness of NMR quantum computing. Science,
277:1688, 1997.
[Wau92] J.S. Waugh. Spin echoes and thermodynamics. In G.G. Bagguley, edi-
tor, Pulsed Magnetic resonance: NMR, ESR and Optics (A recognition
of E.L. Hahn), page 174. Clarendon, Oxford, 1992.
Bibliography 145
[You83] A.P. Young. Direct determination of the probability distribution for
the spin-glass order parameter. Phys. Rev. Lett., 51:1206, 1983.
[ZS86] A.Y. Zyuzin and B.Z. Spivak. Friedel oscillations and Ruderman-Kittel
interaction in disordered conductors. JETP Lett., 43:234, 1986.
Samenvatting
Theoretisch fysici bestuderen met gebruik van mathematische modellen de eigen-
schappen van verschillende systemen in de natuur. In vorige eeuwen realiseerde men
zich dat we met de wiskunde die we hebben niet alles exact kunnen oplossen. Weten-
schappers beseften dat een complete beschrijving ook niet nodig is, zoals in de Sta-
tistische mechanica waar het niet nodig is om de beweging van elk van de atomen in
een gas te beschrijven om de interessante eigenschappen, zoals druk of temperatuur,
te bestuderen. Of het geval van Quantummechanica, waar een complete beschrijving
onmogelijk is. Dus de oplossing voor problemen in de Natuurkunde veranderde van
een deterministische oplossing naar een gegeven in termen van gemiddeldes of een
set van waarschijnlijkheden voor elk van de mogelijke gebeurtenissen. Verder was al
vroeg duidelijk dat een complete oplossing, zelfs in termen van gemiddeldes, met de
tot nu toe bekende wiskunde onmogelijk is. We moeten benaderingen gebruiken.
Tegenwoordig is theoretische natuurkunde gebaseerd op benaderingen. We kun-
nen alleen ruwe oplossingen vinden, of ten hoogste oplossingen die waar zijn in een
klein bereik van de betrokken parameters. Toch kunnen we met deze oplossingen
heel veel van de nodige informatie en het nodige begrip krijgen. Een andere manier
om het probleem te benaderen is het gebruik van vereenvoudigde modellen die de be-
langrijkste eigenschappen van moeilijker en realistischer modellen behouden. Vanuit
deze speelgoedmodellen, kunnen we iets zeggen over hoe realistischere modellen
zich gedragen en kunnen we ook algemene ideeen testen zoals kritieke verschijnselen
en universaliteit.
Deze theoretische ontwikkeling was gevolgd of soms gestimuleerd door experimen-
tele ontwikkelingen. We kunnen nu, bijvoorbeeld, makkelijk heel lage temperaturen
bereiken. Bij zulke lage temperaturen kunnen quantumeecten gezien worden. Dat
brengt nieuwe vragen en nieuwe problemen met zich mee. We moeten die nieuwe
eecten begrijpen. Nieuwe toestanden van materie werden gevonden, zoals de Bose-
Einsteincondensaten. Nieuwe fase-overgangen, zoals de quantum fase-overgangen
die bij temperatuur nul voorkomen, werden voorgesteld. We kunnen die niet direct
zien, maar veel eecten die we bij heel lage temperatuur waarnemen zijn het gevolg
van de nabijheid van een quantumkritiek punt.
Een van de grootste ontdekkingen was het bestaan van spin. We kunnen spins
beschrijven op veel verschillende manieren, afhankelijk van het doel en de situatie.
Spins zijn bijvoorbeeld verantwoordelijk voor magnetische eecten. Om hun gedrag
te begrijpen moeten heel veel benaderingen gebruikt worden. Vaak zijn in een mag-
148 Samenvatting
neet details van de dynamica van elke spin niet van het eerste belang. Verder kan
de kristalstructuur de symmetrieen breken zodat spins-1/2 als Ising- of XY-spinnen
beschreven kunnen worden. De wisselwerking tussen elk spinpaar in het rooster
kan klassiek beschreven worden, hoewel haar oorsprong quantummechanisch is. On-
danks deze benaderingen zijn dit soort problemen nog steeds niet oplosbaar. Om dat
te overkomen, zijn er veel benaderingen en generalisaties van de algebra van spins
uitgevoerd. Het eerste deel van dit proefschrift bestudeert een succesvolle generali-
satie van de spins, de genoemde sferische spins. We gebruiken een quantumversie
van dit model om quantum-fase-overgangen te bestuderen. We kunnen het helemaal
analytisch oplossen. Dit is een van de makkelijkste modellen die niet-triviale kri-
tieke verschijnselen vertoont. We laten zien dat de sferische spinnen uit een limiet
van Heisenbergspins kunnen komen. We berekenen alle kritische exponenten in het
klassieke en het quantum-kritische punt.
Terwijl we in een magneet, waar de spins aan een rooster vastzitten, vaak alleen
de evenwichtseigenschappen kunnen bestuderen, kunnen we in andere gevallen spins
(of systemen die als een spin beschreven kunnen worden, zoals een superconducting
quantum interference device, SQUID) los van de overige spins beschouwen. In kern-
spinresonantie (nuclear magnetic resonance, NMR), bijvoorbeeld, kan het gebeuren
dat spins niet in interactie zijn met andere spins. Met gebruik van lasers kunnen
we de richting van de spins precies draaien. We kunnen dan de dynamische eecten
bestuderen. Toch staan die spins niet los van de rest van de wereld en de dynamiek
is benvloed door de omgeving van de spin die demping en decoherentie veroorzaakt.
In het tweede deel van dit proefschrift stellen we een model voor om dat te beschrij-
ven. Ook hier gebruiken we een vereenvoudigde versie van het model die analytisch
opgelost kan worden. Nog een keer gebruiken we benaderingen in de gekozen vorm
van het model en niet halverwege, zoals meestal gebeurt om daarna alleen
maar analytische stappen te zetten. We kunnen dan in een gereduceerd bereik een
analytische oplossing vinden. We beschrijven een ensemble van spins onder het eect
van laserpulsen zoals in NMR. Met dit model bestuderen we nieuwe mechanismen
van arbeids-extractie, zoals lasing, en van afkoeling van spins. We stellen mechanis-
men voor die voordelen kunnen hebben boven de nu gebruikelijke. Het belang van
afkoeling van spins is dat resultaten in NMR-experimenten afhankelijk zijn aan de
aanvankelijke polarisatie van de spin. Hoe groter de polarisatie is des te sterker het
signaal.
Er is heel veel ontwikkeling op het gebied van scheikunde. Nu kunnen we legerin-
gen met precieze concentratie van elk element maken. Een voorbeeld is Au
1c
Fe
c
,
dus een transitiemetaal binnen een edelmetaal. Voor bepaalde concentraties c is
deze legering een spinglas. Spinglazen hebben de ontwikkeling van een jonge tak
van natuurkunde, de studie van ongeorderde systemen, erg bevorderd. Spinglazen
zijn ongeorderde magneten waarin de richting van de spins beneden de kritieke tem-
peratuur op een schijnbaar ongeorderde wijze invriest.
In de derde deel van dit proefschrift beschouwen we spinglazen. Om te beginnen
hebben we bestudeerd hoe in metalische spinglazen de fase-overgangstemperatuur
van de concentratie van magnetische onzuiverheden afhangt. Daarna hebben we een
andere benadering gemaakt om problemen op te lossen, namelijk computersimula-
149
ties. We hebben een spinglas model gesimuleerd in een computer. Het probleem hier
is dat we alleen kleine systemen kunnen simuleren doordat computers een eindig en
relatief klein geheugen hebben. Met behulp van deze simulatie kunnen we bestuderen
hoe een klein systeem een bekend oneindig gemiddelde-veldsysteem benadert.
Acknowledgements
Ten erste wil ik graag bedanken mijn begeleider Theo M. Nieuwenhuizen voor zijn
colaboratie, hulp en geduld die hij heeft gehad in de loop van de laatste vier jaar
en mijn promotor Bernard Nienhuis sinds hij de kans aan mij is gegeven om deze
proefschrijft te maken.
Voldria agrair tambe a Fèlix Ritort. Tu em vas portar aqui aquestic, sense tu
res daquest projecte no haguès pogut ni comencar.
Furthermore I am deeply indebted Armen E. Allahverdyan, Luca Leuzzi and
Stephane Peysson, without whom this thesis would not exist or at least would de-
nitely not be the same. Stephane, apart of being an ocemate, colleague and friend
you did a wonderful job in the rst two and a half years of my PhD. The many
hours and chalk you spent on me helped introducing me in the world of research
and to nd my way through it. Luca, I started sharing with you the supervisor and
we ended having a collaboration. I am extremely grateful with the job we did the
last month of writing this thesis and all the help throughout these years I got from
you. Armen, my last ocemate. I learned from you really a lot. You were always
patient and available, answering all my stupid questions. Further you kept me going
on the right track when I was starting to get lost which helped me a lot in nishing
on time.
I acknowledge too all the members of the institute of theoretical physics of the
University of Amsterdam for their support and for those coee breaks and chats
in the corridor. Specially I am very grateful to those who supported my programs
running on their computers for various months without a single complain. Estoy
especialmente agradecido con Alex, con quien uno poda desfogarse en su propio
idioma.
A qui primer haig dagrair, ara ja des dun punt de vista no professional es la
meva famlia. Especialment a la meva mare, per les innombrables hores telefòniques;
al meu pare, pel suport que em donava al tornar a casa; al meu germà, pel milio
de-mails enviats per resoldre tots els meus problemes tècnics; a la Caterin, pel munt
de xerrades via e-mail; als meus avis i padr, per la seva conanca, o fe, cega, ns i
tot quan jo amb mi mateix no en tenia.
Quiero agradecer a Laura Manzano Baena su apoyo, dedicacion y paciencia,
sobretodo en el primer a no de mi tesis. Por todo lo que contigo y de ti he aprendido,
Laura, por todos esos momentos, ya que en el fondo, esta tesis, lleva grabado tu
nombre.
152 Acknowledgements
Agradezco las mil horas de esta, charlas, cafes y demas con la Spanish crowd,
aunque muchos de ellos no sean espa noles: Annemarieke, Daniele, Diego, Jaime,
Jordi, Jorge, Juan, Nacho, Manu, Miriam, and others non-Spanish speakers like
Leonid and his amazing picutres, or the wonderful Fatima. With you guys I had the
non-physicist fun I needed to get along with all the work of this thesis.
Quiero agradecer a Silke las horas de discusiones sobre, ciencia, ya sea fsica o
sociologa, y la vida en su sentido mas general.
Entre los que solo han compartido un poco tiempo, solo un a no, pero que en este
han sido muy importantes aunque en a nos diferentes quiero agradecer a Montse y a
Rafa. Muy especialmente quiero tambien nombrar a Cati. Segueixo amb amics que
tot i no ser a Amsterdam, han estat sempre a prop meu ja sigui via e-mail o amb
visites. Especialment haig dagrair la companyia de Jordi Busquets, Ra ul Or us, i
Ester Sola. No puc deixar de nombrar a lAnna Batlle, Albert Royo, Llus Laguna,
Mireia Torralba...
Though short in time, I want to acknowledge the Kellerbar people, for the good
times and beers we had together.
I am specially grateful to my atmate Anders Tranberg. You resisted patiently
my hair, the outrageous armadas, and the presence of Laura for almost a year
without even complaining. Thanks and my best wishes for you in Brighton.
Ik wil graag ook even bedanken eerste Annemarieke Klein. Annemarieke jij had
de nodig geduld met mij en mijn Nederlands. Zonder jou zou ik nooit deze taal
kunnen spreken.
Ik heb heel veel plezier gehad in mijn Capoeira lessen in bijzonder met de mensen
van de Beneesta groep.

Ruben Serral Gracia - Spin Bosons and Spin Glasses

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ruben Serral Gracia - Spin Bosons and Spin Glasses

Uploaded by

Copyright:

Available Formats

Spin Bosons and Spin Glasses

S), still leads to sensible dynamics,

). Sending M we nally get

has a weight in this integral given by (J

). The actual form

, g) is the partition function of

= 0 in the paramagnetic phase,

2 is the real part of the former

. This nally gives

represents a triplet annihilation

the two minima can create an eec-

(t)] = 0 since they belong to dierent Hilbert

(t) = exp [it iG

due to the interaction

X generated by the interaction

of the bath temperature to its

and not via the average occupation number

) 0, apply another set of two pulses,

d. After this decay, s

You might also like