T H E O RY O F S I M P L E G L A S S E S

This pedagogical and self-contained text describes the modern mean field theory
of simple structural glasses. The book begins with a thorough explanation of
infinite-dimensional models in statistical physics, before reviewing the key
elements of the thermodynamic theory of liquids and the dynamical properties
of liquids and glasses. The central feature of the mean field theory of disordered
systems, the existence of a large multiplicity of metastable states, is then
introduced. The replica method is then covered, before the final chapters describe
important, advanced topics such as Gardner transitions, complexity, packing
spheres in large dimensions, the jamming transition and the rheology of glasses.
Presenting the theory in a clear and pedagogical style, this is an excellent resource
for researchers and graduate students working in condensed matter physics and
statistical mechanics.

g i o r g i o pa r i s i is a professor of physics at Sapienza University of Rome. His

research is broadly focused on theoretical physics – from particle physics to glassy
systems. He has been the recipient of numerous awards, including the Boltzmann
Medal, the Enrico Fermi Prize, the Max Planck Medal, the Lars Onsager Prize
and an ERC advanced grant. He is president of the Accademia dei Lincei and a
member of the collaboration Cracking the Glass Problem, funded by the Simons
Foundation.
p i e r f r a n c e s c o u r b a n i is a CNRS researcher. His research activity focuses
on statistical physics of disordered and glassy systems. After a joint PhD between
Sapienza University of Rome and University of Paris-Sud, he joined the Institut
de Physique Théorique of CEA, first as a post-doctoral researcher and then as a
permanent researcher.
f r a n c e s c o z a m p o n i is a CNRS research director and an associate professor
at Ecole Normale Supérieure. His research is broadly focused on complex
systems, ranging from glasses to agent-based models for macroeconomy. He
has been awarded an ERC consolidator research grant, and he is a member of the
collaboration Cracking the Glass Problem, funded by the Simons Foundation.
T H E O RY O F S I M P L E G L A S S E S
Exact Solutions in Infinite Dimensions

G I O R G I O PA R I S I
Sapienza University of Rome

P I E R F R A N C E S C O U R BA N I
Institut de physique théorique, Université Paris Saclay, CNRS, CEA

FRANCESCO ZAMPONI
Ecole Normale Supérieure
 University Printing House, Cambridge CB2 8BS, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre, New Delhi – 110025, India
79 Anson Road, #06–04/06, Singapore 079906

Cambridge University Press is part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit of
education, learning, and research at the highest international levels of excellence.

www.cambridge.org
Information on this title: www.cambridge.org/9781107191075
DOI: 10.1017/9781108120494
© Giorgio Parisi, Pierfrancesco Urbani and Francesco Zamponi 2020
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2020
Printed in the United Kingdom by TJ International Ltd, Padstow Cornwall
A catalogue record for this publication is available from the British Library.
ISBN 978-1-107-19107-5 Hardback
Cambridge University Press has no responsibility for the persistence or accuracy
of URLs for external or third-party internet websites referred to in this publication
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
 Contents

Preface page ix
1 Infinite-Dimensional Models in Statistical Physics 1
1.1 The Ising Model 1
1.2 Large Dimension Expansion for the Ising Model 6
1.3 Second-Order Phase Transition of the Ising Ferromagnet 10
1.4 Low-Temperature Ferromagnetic Phase 18
1.5 Metastable States 26
1.6 Wrap-Up 34
2 Atomic Liquids in Infinite Dimensions: Thermodynamics 37
2.1 Thermodynamics of Atomic Systems 37
2.2 The Virial Expansion 46
2.3 Liquids in Large Dimensions 50
2.4 Wrap-Up 62
2.5 Appendix: Rotationally Invariant Integrals 64
3 Atomic Liquids in Infinite Dimensions: Equilibrium Dynamics 67
3.1 Properties of Equilibrium Dynamics 68
3.2 Langevin Dynamics of Liquids in Infinite Dimensions 76
3.3 Dynamical Glass Transition 82
3.4 Critical Properties of the Dynamical Glass Transition 89
3.5 Wrap-Up 93
3.6 Appendix: Reversibility for Langevin Dynamics 97
4 Thermodynamics of Glass States 99
4.1 Arrested Dynamics and Restricted Thermodynamics 99
4.2 Restricted Thermodynamics in Infinite Dimensions 109
4.3 Replicated Free Energy and Replica Symmetry 120

v
vi Contents

4.4 Replica Symmetric Phase Diagram of Simple Glasses 128

4.5 Wrap-Up 132
4.6 Appendix 135
5 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes 140
5.1 An Introduction to Replica Symmetry Breaking 140
5.2 The Algebra of Hierarchical Matrices 146
5.3 Probability Distribution of the Mean Square Displacement 154
5.4 Replicated Free Energies and Hierarchical Matrices 164
5.5 De Almeida–Thouless Transition and Marginal Stability 172
5.6 Wrap-Up 176
6 The Gardner Transition 180
6.1 State Following in the Replica Symmetry Broken Phase 180
6.2 Gardner Transition and Replica Symmetry Breaking 185
6.3 Gardner Transition of Simple Glasses 187
6.4 Critical Properties of the Gardner Transition 193
6.5 Wrap-Up 195
6.6 Appendix: Numerical Resolution of the RSB Equations 197
7 Counting Glass States: The Complexity 199
7.1 Equilibrium Complexity and the Kauzmann Temperature 199
7.2 Out-of-Equilibrium Complexity: The Monasson Method 204
7.3 The Monasson Construction in Infinite Dimensions 210
7.4 The Phase Diagram of Hard Spheres 214
7.5 Replica Symmetry Breaking Instability in the Monasson
Construction 222
7.6 Wrap-Up 227
8 Packing Spheres in Large Dimensions 231
8.1 Statement of the Problem 231
8.2 Review of Rigorous Results 235
8.3 Review of Non-rigorous Results 240
8.4 Liquid, Glass and Packings in Infinite Dimensions 243
8.5 Wrap-Up 248
9 The Jamming Transition 251
9.1 The Jamming Transition as a Satisfiability Threshold 251
9.2 Criticality of Jamming 260
9.3 The Unjammed Phase: Hard Spheres 273
9.4 The Overjammed Phase: Soft Harmonic Spheres 280
9.5 Wrap-Up 287
 Contents vii

10 Rheology of the Glass 290

10.1 Perturbing the Glass by a Shear Strain 290
10.2 Linear Response 295
10.3 Stress–Strain Curves 299
10.4 Wrap-Up 302
References 305
Index 322
 Preface

State of the Art: In Search of a Fully Microscopic Theory of Glasses

Most classical solid state textbooks are almost entirely devoted to crystals, see,
e.g., [21]. The main reason is that, while the theory of crystalline solids is fully
developed, that of amorphous solids is still very incomplete. As a first approxima-
tion, crystals can be understood as perfectly symmetric periodic lattices, around
which particles undergo small vibrations. A low-temperature harmonic expansion
can then be constructed to obtain the thermodynamic properties in terms of har-
monic excitations – i.e., phonons. Defects (mostly dislocations) can then be added
to the theory to describe crystal flow (or plasticity) and melting [180, 212]. Crystals
are well understood mainly because they can be thought of as small perturbations of
a perfectly symmetric lattice, the small parameters being the amplitude of thermal
vibrations and the density of defects.
Yet most of the solid matter in nature is not crystalline but amorphous: glasses,
foams, pastes, granulars and plastics are but a few examples. These materials are not
only ubiquitous but also extremely important for practical, everyday applications.
For simplicity, in the rest of this book, we call these materials ‘glasses’. Glasses
display all kind of anomalies with respect to crystals: in particular, their vibrations
cannot simply be understood in terms of plane waves, their flow is not mediated
by well defined defects, and their dynamics is extremely complex. Unlike crystals,
glasses offer no guiding symmetry principle to construct a microscopic theory, and
no natural ‘small’ parameter can be used to organise a perturbative expansion.
Constructing a complete first-principle theory of glasses has then turned out to be
an extremely difficult task. Yet a lot of progress has been made and, recently, several
books on glasses written (or edited) by theoretical physicists [37, 53, 168, 357]
have appeared. These books are largely devoted to the phenomenology of real (or
realistic models of) materials, as known from experiments and numerical simula-
tions, and the theoretical approaches they discuss mostly make use of approximate

ix
x Preface

methods. This is, of course, an excellent idea given the complexity of the problem,
and it is typical of theoretical physics.
The aim and style of this book is, however, quite different and, in our opinion,
complementary to previous efforts. We discuss here the exact solution of a micro-
scopically well-defined model which, we believe, can be taken as the simplest real-
istic model of a glass. By ‘exact solution’, we mean that one is able to compute in
a mathematically exact way all the relevant observables of the model. Although the
solution is not mathematically rigorous, we argue that it is exact from a theoretical
physicist’s perspective. We believe, as we discuss in the rest of this preface, that the
material presented in this book constitutes a useful first step towards reaching the
goal of constructing a complete and fully microscopic theory of glasses.

A Digression on the Structure of Scientific Theories

Let us make a philosophical digression to discuss what we can reasonably expect
from a theory of glasses. A convenient definition of a ‘scientific theory’, given1
by Lucio Russo in [312, section 1.3], is obtained by requiring the following three
properties:
1. Its statements do not concern concrete objects pertaining to the real world but
specific abstract mathematical objects.
2. It has a deductive structure: it is made by a few postulates concerning its objects
and by a method to derive from them a potentially infinite number of conse-
quences.
3. Its application to the real world is based on a series of ‘correspondence rules’
between the abstract objects of the theory and those of the real world.
According to Russo, a useful criterion for determining whether a theory has these
properties is to check if one can compile a collection of exercises that can be solved
within the theory. Solving a problem in the context of the theory is then nothing but
an (arbitrarily difficult) ‘exercise’.
As an example of this structure we can take Newtonian mechanics, where (1)
the abstract objects are point particles interacting via forces, (2) the postulates
are minimal but the theory is extremely powerful because from these postulates
one can deduce an enormous variety of results, and (3) the point particles of the
theory can be put in correspondence with many real-world objects, ranging from
atoms to planets, depending on the context in which one wishes to use the theory.2

1 Together with a nice discussion of its limits, that is not reproduced here.
2 Note that we are nowadays used to this kind of logical structure, which is, however, the result of an extremely
long historical process. Even the mathematical definition of ‘point’ has long been debated [312].
 Preface xi

This example highlights that the choice of correspondence rules is extremely deli-
cate. We know very well that atoms and planets are not point particles. They have a
complex internal structure, which limits the applicability of the theory, giving rise
to important physical phenomena.
‘Scientific theories’, as defined earlier, are powerful for two main reasons:
(i) Working on two parallel but distinct levels (the mathematical model and the real
world) allows for a very flexible reasoning. In particular, one can guarantee the
‘truth’ of scientific statements by limiting them to the domain of the model.
(ii) The theory can be extended, by using the deductive method and introducing
new correspondence rules, to treat situations that were not a priori included in
the initial objectives for which the theory was developed.
At the same time, it is important to keep in mind that any ‘scientific theory’ has a
limited utility. In general, it can only be used to model phenomena that are not too
‘far’ from those that motivated its elaboration. Theories that become inadequate to
describe a new phenomenology must, for this reason, be substituted. They remain,
however, according to our definition, ‘scientific theories’, and one can continue
to use them in their domain of validity [312].3 This last statement is particularly
important because it reminds us that ‘the’ theory of a given class of phenomena –
e.g., the glass transition – will never exist. Scientific theories are never unique or
everlasting. They are models of reality, and there is no problem in using different
models of the same phenomenon and in replacing current models with more pow-
erful ones when they are found.

Towards a Scientific Theory of Glasses

The aim of this book is to make some steps in the application of the programme
mentioned earlier to the problem of glasses. Let us state from the very beginning
that we will not be able here to complete this programme. This book is mainly
concerned with steps 1 and 2 – i.e., constructing an abstract mathematical theory
that has a deductive structure and describes at least the basic phenomenology of
glass formation and of the amorphous solid phase. The difficult problem of estab-
lishing a correspondence with the real world is left aside here. We give only hints
and references to the literature so that the reader can form their own opinion on the
proper correspondence rules and judge the quality of the theory according to their
own taste. Another book will have to be written on the subject in the future.

3 For instance, Newtonian mechanics did not become useless once quantum mechanics was developed, and the
fact that it gives incorrect predictions (e.g., the instability of the hydrogen atom) does not mean that it is plain
wrong.
xii Preface

We believe that the approach described above has important advantages, mostly
based on points (i) and (ii). Let us give two examples. Concerning point (i), the
problem of glasses is extremely complex and many different theories have been
proposed. In the attempt to describe real-world materials, most of these theories
make heavy use of approximations, to the point that it often becomes quite difficult
to establish whether the statements made by the theory are true even within the
logical structure of the theory itself. We instead introduce a simple and solvable
mathematical model of glass: a system of Newtonian point particles in the limit of
infinite spatial dimensions. Our aim is to discuss the mathematical solution of this
model, which is already extremely rich and complex. But, although we believe the
solution to be exact, a mathematically rigorous proof of its exactness is still lacking,
and we hope that presenting the non-rigorous solution in a clear and aspirationally
pedagogical way will help progress towards a rigorous proof. Our statements are
limited to this mathematical model, and one is then able to decide whether they
are true or false in a well-defined mathematical sense. Concerning point (ii), we
will see a spectacular example of its power in Chapter 9. The model, originally
designed to describe the liquid and glass phases of atomic materials, also displays
a phase transition that can be put in correspondence with the jamming transition of
granular materials. Thus, the model shows a potential to unify phenomena (glass
and jamming transitions) and materials (atomic glasses, colloidal glasses, granular
glasses) that were thought to be somehow distinct.
The main, and very important, drawback of this approach is that the infinite-
dimensional limit is quite abstract. Hence, the final step of establishing correspon-
dence rules between the different phases and observables of the abstract model and
their real-world counterparts (i.e., real liquids, real glasses, real granular materials)
is non-trivial and remains largely open to debate. In granular materials, for example,
the role of friction remains to be clarified. If successfully performed, this step would
ultimately correspond to constructing a scientific theory of real-world glasses (i.e.,
to implement step 3), but it requires a lot of additional discussion which goes
much beyond the scope of this book. Here, when discussing each specific aspect
of the mathematical solution of the model, we limit ourselves to a few hints and
references to direct the reader towards real-world phenomena that could potentially
be described by this solution. This is done at the end of each chapter, in the Further
Reading sections. We do not, however, specify completely the list of phenomena
that could be accounted for by such a theory, nor do we try to establish precise
correspondence rules between the mathematical model discussed here and real-
world objects. Discussing this issue with all the needed details will be part of the
follow-up publications to the present book.
 Preface xiii

Historical Note
The idea of using infinite-dimensional solvable models has been extremely suc-
cessful in condensed matter, appearing in the context of atomic physics [337],
liquids [154, 206], ferromagnetic systems [163] and strongly correlated electrons,
where it has led to the celebrated dynamical mean field theory [164]. See also [356]
for related ideas in high-energy physics.
In the context of the glass transition, the idea of solving the problem in the
infinite dimensional limit was first proposed in [206]. A complete and exact solu-
tion of the infinite-dimensional problem was then obtained more recently, in a
long series of research articles to which we contributed together with many other
colleagues [3, 88, 89, 220, 221, 239, 291, 292, 299]. The methods used in this
solution are deeply rooted in the theory of spin glasses [79, 254], using dynamical
methods [109, 111] and replica methods [144, 260] specifically developed for the
glass problem. These methods were first applied to approximate the behaviour
of glasses in finite dimensions [76, 252, 253, 256, 292, 366, 368] before it was
realised that they become exact in the infinite-dimensional limit [220, 291]. The
phase diagram turns out to be similar to that of a class of spin glass models (Ising
p-spin and Potts glasses [79, 160, 170, 171, 203, 204, 207]), which confirms the
main assumption behind the random first-order transition (RFOT) theory of the
glass transition [205, 208, 357]. The solution also reproduces the essential features
of the mode-coupling theory of the glass transition [168], as discussed in [239].
The aim of this book is to collect these results and organise them in a pedagog-
ically coherent way. We did not include here any new material (except for some
polishing of the original work), and we do not wish to take any additional credit for
the results. The original papers are thus carefully referenced along the book.

Target Audience
This book is written having in mind two distinct types of readers. The first are
young students (at the level of the final year of undergraduate studies or at the
beginning of their graduate studies). We expect these readers to have no or very little
background knowledge of the physics of glasses. Reading this book only requires
a basic background in statistical mechanics: the mean field (Landau) approach
to phase transitions in magnetic materials and some basics in liquid state theory
(as covered, e.g., in [175]). We ask these readers to believe our conjecture that a
system of infinite-dimensional atoms can be a good mathematical model of glass
and to follow us in the mathematical study of the model. Along the way, they will
learn advanced statistical mechanics techniques (e.g., the replica method) as well
xiv Preface

as many deep concepts that pervade the physics of disordered systems (e.g., long-
lived metastable states, complex free energy landscapes, ergodicity breaking). By
working out all the calculations in this example, they will learn methods that can
be used in many different contexts ranging from spin glasses [254] to optimisation
problems [251] and neural networks [10]. However, they will not acquire suffi-
cient background about the phenomenology of glasses and on the many different
approaches that are used in their theoretical description. For this purpose, we refer
to other existing excellent books and reviews [37, 40, 53, 80, 168, 357].
The second group of readers are experienced researchers working on the physics
of glasses. We expect them to be already acquainted with the main physical
concepts discussed in the book and to be familiar with the material contained
in [37, 40, 53, 80, 168, 357]. Yet we hope that these readers will find here a way
to put many different pieces of knowledge into a common perspective. Some of
these more experienced readers might also be interested in learning the details of
the methods mentioned earlier.

Structure of the Book

The book presents the main logical steps of the derivation of the exact solution
of amorphous infinite-dimensional particle systems. All the non-trivial steps are
presented, but leaving to the reader some trivial intermediate steps (for which we
provide references to the original work). At the end of each chapter, a Summary
section is provided to recapitulate the main points discussed in the chapter. The
correspondence with real-world objects, as well as the comparison with approxi-
mate theories of glasses, is logically separated from the main stream of the book.4
Some elements are presented in short Further Reading sections at the end of the
relevant chapters.
The structure of the book chapters is the following.

• Chapter 1 reviews classical results in the statistical mechanics of infinite-

dimensional systems, using the Ising model as a paradigmatic example. The
notion of metastability is introduced and discussed.
• Chapter 2 provides a short review of classical results in liquid state theory [175]
and then introduces the strategy to solve the thermodynamics of atomic liquids in
infinite dimensions.
• Chapter 3 presents the solution of the equilibrium liquid dynamics in infi-
nite dimensions by a series of simple arguments, following Szamel [339].

4 Except in Chapter 9, where the problem is easier and some discussion of this correspondence is provided in
the main text.
 Preface xv

The existence of a sharp dynamical glass transition, at which the equilibrium

diffusion constant vanishes, is discussed.
• Chapter 4 discusses the central feature of the mean field theory of disordered
systems, namely the existence of a large multiplicity of metastable states. The
replica method is introduced in this context, following Franz and Parisi [144].
The appearance of metastable states is directly connected to the dynamical arrest
of the liquid. The exact expression of the replicated free energy of an atomic
system is derived, following the ideas of [220, 299] and the detailed derivation
of [59]. From this basic object, the glass phase diagram of two model systems
is derived. This chapter also contains the core of the mathematical solution of
the model.
• Chapter 5 provides a short compendium of basic notions of replica symmetry
breaking (RSB) and the associated ultrametric distribution of states in phase
space. This chapter is a review of classical results in spin glass theory [254].
• Chapter 6 discusses RSB effects in the Franz-Parisi construction introduced in
Chapter 4. The phase diagram of two simple glass models is derived. The notion
of a Gardner transition is introduced, and its physical meaning is discussed using
the results of Chapter 5. This chapter reviews results originally presented in
[88, 299, 315].
• Chapter 7 introduces another replica scheme due to Monasson [260], which
allows one to study a different class of metastable states. The phase diagram of
hard spheres is derived and compared with that of Chapter 6. The connection
with the Edwards approach to the study of jammed packings is discussed. This
chapter reviews results originally presented in [88, 93, 291, 292].
• Chapter 8 introduces the sphere-packing problem. A brief review of mathemat-
ical and physical results is provided. It is shown how the results obtained in
Chapters 4, 6 and 7 provide insight on this problem. This chapter reviews results
originally presented in [240, 291, 292, 325].
• Chapter 9 focuses on the jamming transition. It is explained why jamming is a
critical point, and the associated critical exponents are described. The non-trivial
scaling form of the RSB equations in the vicinity of jamming is discussed in
detail. This chapter presents results originally derived in [88, 89, 91].
• Chapter 10 discusses the response of the glass to an applied shear strain. The
elasticity and yielding of the glass are discussed. This chapter presents results
originally derived in [58, 298, 299, 346].

The reader has certainly noticed that the title of this book makes explicit refer-
ence to the classic Theory of Simple Liquids by Hansen and McDonald [175]. Our
intention is indeed to apply to glasses the same program that was applied to liquids
in the 1960s. The word ‘simple’ refers, here and in [175], to the fact that we only
xvi Preface

consider the simplest model of liquids and glasses: a collection of classical point
particles, modelling atoms. There are, of course, much more complex glass-forming
systems (e.g., polymer glasses and network glasses, or anisotropic granular systems
like hard ellipsoids), but their description falls beyond the scope of this book.

Acknowledgements
We would like to especially thank Patrick Charbonneau, who – besides participating
actively to the research effort that has led to this book – has carefully reviewed the
original manuscript and provided very useful comments.
The main results presented in this book are the outcome of a long and much
broader research effort that involved many important ideas and tools in this
and related fields. This has been possible thanks to the efforts of many people,
colleagues and students. It is a pleasure to warmly thank all of them. We would
like, in particular, to thank our most recent collaborators that have been directly
involved in the results presented in this book: Elisabeth Agoritsas, Ada Altieri,
Ludovic Berthier, Giulio Biroli, Jean-Philippe Bouchaud, Carolina Brito, Eric
Corwin, Silvio Franz, Sungmin Hwang, Atsushi Ikeda, Harukuni Ikeda, Hugo
Jacquin, Jorge Kurchan, Yuliang Jin, Thibaud Maimbourg, Marc Mézard, Corrado
Rainone, Federico Ricci-Tersenghi, Camille Scalliet, Antonio Sclocchi, Mauro
Sellitto, Guilhem Semerjian, Beatriz Seoane, Gilles Tarjus, Marco Tarzia, Matthieu
Wyart and Hajime Yoshino. We would also like to thank all the members of the
Simons Collaboration on Cracking the Glass Problem for many useful discussions
and for providing an ideal environment for developing this project.
Part of the research described in this book has received funding from the Euro-
pean Research Council (ERC) under the European Union’s Horizon 2020 research
and innovation programme (grant agreement n. 694925 – Low Temperature Glassy
System, Giorgio Parisi and n. 723955 – GlassUniversality, Francesco Zamponi),
by a grant from the Simons Foundation (#454949, Giorgio Parisi and #454955,
Francesco Zamponi) and by the Investissements d’Avenir program of the French
government via the LabEx PALM (ANR-10-LABX-0039-PALM).
 1
Infinite-Dimensional Models in Statistical Physics

Infinite-dimensional models are core to statistical physics. They can be used to

understand liquids and glasses, as they are in this book [153, 206, 292, 362], but
also strongly coupled electrons [164], atomic physics [337] and gauge field theory
[132], to name a few. The reason is that infinite-dimensional models are exactly
solvable using mean field methods. The aim of this chapter is to give an example of
this construction in the context of the Ising model of magnetism.
It will be assumed that the reader is already familiar with the basic properties
of the Ising model as presented, for example, in the first chapters of [69]: its main
observables (magnetisation, magnetic susceptibility), its phase diagram and phase
transitions, and its dynamics. The aim of this chapter is mostly to present these
properties in the context of a large dimensional expansion and to introduce the
concept of a thermodynamic (stable or metastable) state, identified with a local
minimum of a suitable free energy function.

1.1 The Ising Model

1.1.1 Definitions
Although some of the concepts presented in this chapter are fairly general, it is
instructive to focus on a specific setting. We thus consider a model of N Ising spins
{Si }i=1,...,N , with Si = ±1. The energy of a spin configuration S = {Si } is given by
the Hamiltonian function
1 
H [S] = − Jij Si Sj − Bi Si , (1.1)
2 ij i

where Jij = Jj i ∈ R denotes the (symmetric) exchange coupling between spins i

and j (with Jii = 0,∀i), and Bi ∈ R denotes an external magnetic field acting on
spin i. The factor 1/2 in front of the exchange energy ensures that each pair ij is
counted only once. The Hamiltonian in Eq. (1.1) summarises many cases of interest,

1
2 Infinite-Dimensional Models in Statistical Physics

such as each spin interacting with all others (the ‘fully connected’ model) [69], or
with a random subset of them (the ‘Bethe lattice’ or ‘random graph’ model) [251].
In general, the set of non-zero couplings defines the ‘interaction graph’; its nodes
are the spins and its edges are the pairs ij , such that Jij = 0. In the following, the
neighbourhood of spin i, denoted by

∂i = {j : Jij = 0}, (1.2)

is the set of all spins j that interact with spin i.

The ferromagnetic Ising model in d dimensions corresponds to a particular
choice of interaction graph: a d-dimensional cubic lattice of unit spacing and linear
size L, containing N = Ld lattice points. Each spin i sits at a point x(i) of the
lattice and interacts only with its nearest neighbours, located at unit distance away
in each principal direction. Because there is a bijective correspondence between
labels i = 1, . . . ,N and lattice points x, both labels can be used equivalently. In the
following, spins will thus be denoted either Si or Sx depending on which labelling
is more convenient. Each pair of nearest neighbour spins interact via a coupling
Jij = J = 2d1 , and Jij = 0 otherwise. In lattice notation, this corresponds to
Jxy = δ|x−y|,1 /(2d), where |x − y| is the Euclidean distance between points x,y and
δab is the Kronecker delta.
Because a cubic lattice has a boundary, one needs to specify boundary condi-
tions. Three choices are commonly used [158].

• Periodic boundary conditions – Each face of the lattice is identified with its
opposite face. In this case, the lattice is translationally invariant in all directions,
and each spin has 2d nearest neighbours, corresponding to displacements in all
possible directions on the lattice; hence, the size of ∂i is 2d. In d = 2, the
topology is that of a torus, generalised to a hypertorus in d ≥ 3.
• Open boundary conditions – The lattice is considered isolated. In this case, the
system is not translationally invariant. In particular, the spins on the boundary
have fewer interactions than those in the bulk.
• Frozen boundary conditions – A layer of external spins is added to each face of
the lattice. The external spins are frozen in prescribed positions: for example, they
are all fixed to +1 or to −1. In this case, each spin has 2d nearest neighbours, but
spins on the boundary interact with one frozen spin (or more, for those on cube
edges), acting as an effective external magnetic field. Also in this case, the system
is not translationally invariant.

Note that the overall magnitude of the couplings J only sets the energy and temper-
ature scales and is therefore irrelevant for the properties of the model. The choice
J = 2d1 guarantees that the exchange energy remains finite when d → ∞, for any
 1.1 The Ising Model 3

value of N . For instance, with periodic boundary conditions, the fully magnetised
spin configuration 1 = {Si = 1, ∀i} has
   N 
H [1] = − Jij − Bi = −NJ d − Bi = − − Bi , (1.3)
ij  i i
2 i

which explicitly shows that the exchange energy is finite for any N . Because the
exchange energy remains of the same order as the entropy and the magnetic field
energy, the model behaviour is interesting at finite temperature and magnetic field.1

1.1.2 Thermodynamic Free Energy and Observables

In equilibrium statistical mechanics, within the canonical ensemble, the probability
of observing a spin configuration S is given by the Gibbs–Boltzmann distribution
e−βH [S]
PGB [S] = , (1.4)
Z
 −βH [S]
where Z = Se is the partition function and β = 1/T is the inverse
2
temperature. Solving the model amounts to computing the free energy of the
system at temperature T :
 β 
F = −T log Z = −T log e 2 ij Jij Si Sj +β i Bi Si . (1.5)
S

Note that F is extensive – i.e., proportional to N . In the following, capital letters

are used for extensive thermodynamic quantities, while lowercase letters are used
for intensive quantities that remain finite in the thermodynamic limit, N → ∞ –
e.g., f = F /N. From the free energy, one can derive the statistical average of any
observable, which we denote by brackets •. For example, the average energy is
   e−βH [S] ∂(βF )
U = H [S] = H [S] = , (1.6)
S
Z ∂β

the entropy is

S=− PGB [S] log PGB [S] = β(U − F ), (1.7)
S

1 All the thermodynamic quantities are functions of T /J and B/J for dimensional reasons. Hence, different
choices of J require rescaling both T and B with d to avoid a trivial behaviour of the model dominated by
entropy or energy. Once T and B are properly rescaled, the result is invariant. For instance, the choice J = 1
leads to an exchange energy of order d and thus requires T ,B ∝ d.
2 The Boltzmann constant k is fixed to unity throughout this book. In other words, T stands for k T in such a
B B
way that temperatures are measured in the same units as energies. Similarly, entropy S stands for S/kB and is
thus adimensional.
4 Infinite-Dimensional Models in Statistical Physics

the local magnetisation is

 e−βH [S] ∂F
mi = Si  = Si =− , (1.8)
S
Z ∂Bi

and the local magnetic susceptibility is

∂mi ∂ 2F
χij = =− = β(Si Sj  − Si Sj ). (1.9)
∂Bj ∂Bi ∂Bj
By construction, χij is a positive matrix; i.e., all its eigenvalues are greater than or
equal to zero.3 The global magnetic susceptibility is the variation of the global mag-

netisation m = N −1 i mi with respect to a uniform magnetic field Bi = B,∀i.
Under a variation of the global field, dB, all local fields change by dBi = dB, hence
dBi /dB = 1,∀i and
dm 1  dmi 1  ∂mi dBj 1 
χ= = = = χij . (1.10)
dB N i dB N ij ∂Bj dB N ij

In general, any physical observable can be written as a linear combination of

products of spin variables,4

N 
O[S] = Oi1 ···in Si1 , . . . ,Sin . (1.11)
n=0 i1 ···in

Therefore, in order to characterise all thermodynamic averages, it suffices to

consider the average of spin products Si1 , . . . ,Sin , which can be computed as
multiple derivatives of the free energy with respect to the relevant magnetic fields.

1.1.3 Free Energy as a Function of the Local Magnetisation

In order to investigate the high dimensional limit, it is convenient to introduce the
free energy function F [m] which gives, roughly speaking, the free energy of a sys-
tem constrained to have a prescribed set of local magnetisations m = {mi }i=1,...,N .

 
3 Writing equivalently Eq. (1.9) as χ = β (S − m )(S − m ) , it follows, for any vector v, that
ij i i j j
     2
vT χ v = vi χij vj = β vi (Si − mi )(Sj − mj )vj = β vi (Si − mi ) ≥ 0.
ij ij i

This result holds in particular for the normalised eigenvectors of χ , for which χ v = λv, with λ the
corresponding eigenvalue; hence, v T χ v = λv T v = λ ≥ 0. This relationship proves that the eigenvalues are all
positive.
4 Equation (1.11) is easily justified by noting that O[S] can take at most 2N values, one for each spin

configuration. The expression on the right-hand side of Eq. (1.11) contains N n coefficients for each n, which
 N
gives N N
n=0 n = 2 coefficients in total.
 1.1 The Ising Model 5

The local magnetisations are fixed by introducing a set of auxiliary local magnetic
fields b = {bi }i=1,...,N , and defining the corresponding auxiliary free energy:
 
[b] = −T log e−βH [S]+β i bi Si . (1.12)
S

Note that the ‘physical’ magnetic fields Bi are included in H [S], while the auxiliary
fields are denoted explicitly as we will use them to constrain the local magneti-
sations. Obviously, Eqs. (1.8) and (1.9) hold equivalently if one takes derivatives
of [b] with respect to the auxiliary fields bi . Because the matrix of its second
derivatives is −χij , which is negative, [b] is necessarily a concave function. The
Legendre transform of [b] defines F [m] as

−βF [m] = −β max [b] + bi mi
b
i
⎡ ⎤ (1.13)
 
= min ⎣log e−βH [S]+β i bi (Si −mi ) ⎦,
b
S

which can be justified as follows. For any finite N , [b] is everywhere differen-
tiable, and its concavity ensures that the maximum over b exists and is unique.
In this case, the value b = b[m] that corresponds to the maximum is the unique
solution of ∂[b]/∂bi = −mi ; i.e., it is the set of local fields bi [m] needed to
enforce the magnetisations mi . Once these values are computed, the value of
[b[m]] gives the free energy of the system with field b[m]. By subtracting the

additional magnetic energy due to the external field, − i bi mi in Eq. (1.13), one
obtains the free energy F [m] of the system constrained to have local magnetisa-
tion m. However, in the thermodynamic limit, [b] can develop singularities (in the
vicinity of a phase transition) and become non-differentiable. This complicates the
discussion of the Legendre transform, as will be detailed in Section 1.4.
Note that the derivative of F [m] (when it exists) is the auxiliary field b[m] that
corresponds to the maximum in Eq. (1.13):
∂ ∂bi ∂ 2 F [m]
bi = F [m], = = (χ −1 )ij . (1.14)
∂mi ∂mj ∂mi ∂mj
The matrix χ −1 is positive, because χ is positive. F [m] is thus a convex function,5
and the free energy [b] can be recovered as its inverse Legendre transform:

−β[b] = −β min F [m] − bi mi . (1.15)
m
i

5 Note that the convexity of F [m] also directly follows from its definition, Eq. (1.13), because the maximum
(over b) of linear functions (of m) is a convex function.
6 Infinite-Dimensional Models in Statistical Physics

The stationarity condition implies that m[b] is a solution of Eq. (1.14), and
it must be a minimum because the second derivative of F [m] is positive. This
result leads to an important observation: if there are no auxiliary fields, bi = 0,
then [b = 0] = F is equal to the thermodynamic free energy, as defined in
Eq. (1.5). We obtain from Eq. (1.15) that

F = min F [m]. (1.16)

m

The thermodynamic free energy thus corresponds to the minimum of F [m] over
all possible sets of local magnetisations m, and the set of local magnetisations that
achieves the minimum in Eq. (1.16) corresponds to the equilibrium magnetisations
in absence of any auxiliary field.

1.2 Large Dimension Expansion for the Ising Model

The function F [m] contains a lot of information but is unfortunately impossible
to compute explicitly for the Ising model when d is finite. One can nonetheless
try to obtain information by constructing a perturbative expansion. The simplest
perturbative expansion of F [m] is the high-temperature expansion [296], which
here also coincides with the large d expansion [163], as shown in this section.

1.2.1 Infinite Temperature

At infinite temperature (β = 0), the Gibbs–Boltzmann probability distribution
e−βH [S] /Z is uniform over all spin configurations. In order to fix the magnetisations,
we need strong magnetic fields; we thus rescale the fields bi introducing λi = βbi .
Because the entropy S at infinite temperature is finite, while the energy U vanishes,
the free energy F = U − T S diverges proportionally to T = 1/β when T → ∞. It
is thus better to consider A = −βF = S − βU which remains finite even at β = 0.
With these rescalings, Eq. (1.12) becomes
  
A0 [λ] = −β0 [b] = log e i λi Si
= log[2 cosh(λi )], (1.17)
S i

where the suffix 0 highlights that this is the zeroth order (β = 0) expression.
The free energy is an analytic function of the auxiliary fields, as obviously there
is no phase transition at infinite temperature. The Legendre transform can thus be
computed by differentiation. The condition that determines λi is

∂0 [b] ∂A0 [λ]

mi = − = = tanh(λi ), (1.18)
∂bi ∂λi
 1.2 Large Dimension Expansion for the Ising Model 7

and, according to Eq. (1.13),


−βF0 [m] = A0 [λ[m]] − λi [m]mi
i
 
= {log[2 cosh(λi [m])] − λi [m]mi } = s0 (mi ), (1.19)
i i
 
1 + mi 1 + mi 1 − mi 1 − mi
s0 (mi ) = − log + log .
2 2 2 2
The function s0 (m) is the entropy of a single spin constrained to have magnetisa-
tion6 m; at infinite temperature, spins are independent, and the total entropy – which
coincides with −βF0 [m] – is the sum of the single-spin entropies.
The first small β correction can be computed easily. We only sketch here the
derivation. First, one expands A[λ] in powers of β as follows:
   
A[λ] = log e i λi Si e−βH [S] = log e i λi Si (1 − βH [S]) + O(β 2 )
S S
 
λi Si
S e i H [S]
= A0 [λ] − β   + O(β 2 ). (1.20)
i λi Si
Se

The first correction is then given by the average of H [S] over independent spins
subjected to magnetic fields λi . The average of Si is tanh(λi ), and, therefore,7 the
average of H [S] is H [tanh(λ)], where tanh(λ) = {tanh(λi )}i=1,...,N . We thus have
A[λ] = A0 [λ] + βA1 [λ], with A1 [λ] = −H [tanh(λ)]. The equation for λ then
becomes mi = tanh(λi ) + β∂A1 [λ]/∂λi , and λi = λ0i + βλ1i with λ0i = atanh(mi ).
The Legendre transform is obtained by writing

−βF [m] = A0 [λ0 + βλ1 ] + βA1 [λ0 + βλ1 ] − (λ0i + βλ1i )mi
i
  (1.21)
= A0 [λ ] + βA1 [λ ] −
0 0
λ0i mi = s0 (mi ) − βH [m].
i i

6 Because an Ising spin has only two states, the probability distribution p(S) of a single spin is expressed by two
real numbers p(1) and p(−1) satisfying p(1) + p(−1) = 1. It is thus specified
 by one real number, which can
be conveniently chosen to be the average magnetisation of the spin, m = S Sp(S); hence,
p(S) = (1 + mS)/2. With this choice, the single spin entropy

s0 (m) = −p(1) log p(1) − p(−1) log p(−1)

coincides with Eq. (1.19).

7 Note that this is true only if H [S] is explicitly written as a linear combination of products of spin variables,
which is true for our reference expression Eq. (1.1). Any general spin Hamiltonian can also be written in this
form, using Eq. (1.11).
8 Infinite-Dimensional Models in Statistical Physics

Note that the contributions of λ1 have disappeared from Eq. (1.21) for the
following reasons. In A1 , λ1 can be eliminated because it gives contributions

of order β 2 . In the remaining terms, A0 [λ] − i λi mi , the derivative with respect
to λ vanishes identically due to the Legendre transform condition, and, therefore,

at first order, the correction to λ disappears. Finally, the terms A0 [λ0 ] − i λ0i mi

give the infinite-temperature result, i s0 (mi ), while the correction is βA1 [λ0 ]
= −βH [tanh(λ0 )] = −βH [m].

1.2.2 High-Temperature Expansion


At first order in β, the free energy function F [m] = H [m] − T i s0 (mi ) is thus
given by the free energy of independent spins with magnetisations m, as shown by
Eq. (1.21). The computation of higher-order corrections in β goes along the same
lines as the first order correction, as described in [163]. At third order, for instance,
the result is
  1 + mi 1 + mi 1 − mi 1 − mi

−βF [m] = − log + log
i
2 2 2 2
1 
+β Jij mi mj + β Bi mi
2 ij i

β2  2
+ J (1 − m2i )(1 − m2j )
4 ij ij (1.22)
⎡
β3 ⎣  3
+ 2 Jij mi (1 − m2i )mj (1 − m2j )
6 ij
⎤

+ Jij Jik Jj k (1 − m2i )(1 − m2j )(1 − m2k )⎦ + O(β 4 ),
ij k

and
⎡ ⎤
∂βF [m] 
βbi [m] = = atanh(mi ) − β ⎣ Jij mj + Bi ⎦
∂mi j
(1.23)

+ β 2 mi Jij2 (1 − m2j ) + O(β 3 ).
j

Imposing the absence of auxiliary fields, bi = 0, is equivalent to minimising

F [m]. In this case, one obtains the so-called Thouless–Anderson–Palmer (TAP)
equations [342],
 1.2 Large Dimension Expansion for the Ising Model 9

mi = tanh βhi ,
  (1.24)
hi = Bi + Jij mj − βmi Jij2 (1 − m2j ) + O(β 2 ),
j j

where hi is the effective magnetic field provided by the spins neighbouring site i.
Note that the TAP Eq. (1.24) are examples of a general class of mean field equations
for disordered systems. These equations can be derived through high-temperature
expansions, as discussed here, or alternatively via probabilistic methods called the
‘cavity method’ or ‘belief propagation’ [251, 254, 372].

1.2.3 Large Dimension Expansion

The high-temperature expansion in Eq. (1.22) can be specialised to the ferromag-
netic d-dimensional Ising model, with periodic boundary conditions and with a
uniform external magnetic field Bi = B. In this case, the system is translationally
invariant. It is, therefore, reasonable to expect that the free energy minimum is
realised by a uniform magnetisation mi = m. In the following, the free energy per
spin corresponding to a uniform magnetisation, also called ‘potential’, is denoted8
by v(m). Equation (1.22) gives
F [{mi = m, ∀i}] 1
v(m) = = −T s0 (m) − m2 − Bm
N 2
(1.25)
β β2 2
− (1 − m ) −
2 2
m (1 − m ) + O(β ).
2 2 3
8d 12d 2
With the choice of coupling scale J = 1/(2d), the first three terms, which represent
entropy, exchange energy and magnetic field energy, respectively, remain finite for
d → ∞, as discussed in Section 1.1.1. The correction terms form a series in powers
of 1/d: the high-temperature (small β) expansion can be used to construct the

8 To clarify the rationale behind this notation, it is convenient to anticipate here briefly a few notions that will be
clarified later in the chapter. In a coarse grained representation the spatial dependent magnetisation profile is
m(x). The thermodynamic free energy F (m) corresponding to a global magnetisation m is the minimum of
 
F [m] ∼ F [m(x)] over all configurations m(x) such that m = N −1 N i=1 mi ∼ V
−1 dxm(x). Although the
minimum is often realised by uniform configurations, this is not always the case. Examples are
antiferromagnets or systems in the phase coexistence region. In a coarse grained representation, corresponding
to Landau theory, one can approximate

F [m(x)] ∼ dx{c[∇m(x)]2 + v[m(x)]},

where v(m) is the free energy of a uniform configuration, which then provides a ‘potential’ term in the total
free energy. If v(m) has a unique minimum, then the profile is uniform; the gradient term disappears and
F (m) = N v(m). Otherwise, phase coexistence can lead to F (m) < Nv(m) for some values of m. See
Section 1.4 for a more detailed discussion.
10 Infinite-Dimensional Models in Statistical Physics

large d (small 1/d) expansion.9 The same result holds also in presence of disorder,
although the proper scaling of the couplings might then differ.10
It is important to stress that while the true function F [m] is guaranteed to be con-
vex, any truncation at a finite order of its high-temperature expansion in Eq. (1.22)
does not necessarily share this property. For example, it is very easy to check
that Eq. (1.25), truncated at any order in β, is not a convex function of m if the
temperature is low enough. This point will be further discussed in Section 1.4.

1.3 Second-Order Phase Transition of the Ising Ferromagnet

The large d expansion in Eq. (1.25) can be used to investigate the ferromagnetic
phase transition, a second-order phase transition that is observed in many magnetic
materials and is well described by the Ising model [11, 69, 282]. In this section, we
discuss the nature of this transition in the limit d → ∞, the corrections in 1/d and
the nature of spin–spin correlations around the transition.

1.3.1 Mean Field: The Curie–Weiss Model

For a uniform magnetisation profile mi = m, keeping only the terms that remain
finite for d → ∞, the potential is
1
v(m) = −T s0 (m) − m2 − Bm, (1.26)
2
and the TAP equations that determine m simplify to reproduce the mean field result,
m = tanh[β(B + m)]. (1.27)
The global magnetic susceptibility χ = dm/dB is then obtained by differentiating
Eq. (1.27) with respect to B:
 
dm dm β(1 + m2 )
= {1 + tanh[β(B + m)] }β 1 +
2
⇒ χ= . (1.28)
dB dB 1 − β(1 + m2 )
Similar expressions have been derived by Curie and Weiss under a mean field
approximation, and for this reason, the infinite-dimensional Ising model is also
known as the Curie–Weiss model [69]. In fact, Eq. (1.27) can be equivalently
obtained by assuming that the neighbours of a given spin i (whose set we denote ∂i,
see Eq. (1.2)) are uncorrelated and replacing the spins by their average. The local

field, hi = Bi + 2d1 j ∈∂i Sj , is then on average equal to hi = B + m. The mag-
netisation of an isolated spin in a field hi is m = tanh(βhi ), from which one

9 See [163, figure 3] for additional correction terms.

√
10 For instance, if J are independent Gaussian variables with zero mean, the natural scaling is J ∝ 1/ d.
ij ij
 1.3 Second-Order Phase Transition of the Ising Ferromagnet 11

obtains Eq. (1.27). This construction is the origin of the name ‘mean field’. The
high-temperature expansion shows that this approximation actually becomes exact
when d → ∞. The local field hi then becomes the sum of a large number of terms,
and by the central limit theorem, it can be replaced by its mean. Neighbouring spins
then also become independent, as will be shown in Section 1.3.3.
In absence of magnetic field, B = 0, the Curie–Weiss model can undergo
a phase transition. According to Eq. (1.16), the absolute minimum of F [m]
corresponds to the thermodynamic equilibrium state of the system. Because of
the translational invariance of the ferromagnetic Ising model, it is reasonable to
expect that in equilibrium all spins have the same magnetisation and mi = m.
Under this assumption, and using Eq. (1.25), the thermodynamic equilibrium is
given by the absolute minimum of v(m). At high temperature, the potential has a
unique minimum in m = 0, the ‘paramagnetic’ state. This state has zero energy and
entropy density s = S/N = log 2; hence, v(m = 0) = −T log 2. One can study
the stability of this state by expanding the free energy around m = 0. Expanding
s0 (m) = log 2 − m2 /2 − m4 /12 + O(m6 ) gives
a 0 2 b0 4
δv(m) ≡ v(m) − v(m = 0) = −Bm + m + m + O(m6 ), (1.29)
2 4
with a0 = T − 1 and b0 = T /3. From this expression one deduces that for B = 0, the
minimum in m = 0 becomes unstable when a0 < 0 – i.e., T < 1. At low tempera-
tures T < 1, the function v(m) has two degenerate minima at m = ±meq (T ,B = 0),
which correspond to ‘ferromagnetic’ states. Close to the critical temperature
Tc = 1, the magnetisation behaves as

meq (T → Tc−,B = 0) ∼ −a0 /b0 ∼ (Tc − T )1/2 . (1.30)
Moreover, from Eq. (1.28), the magnetic susceptibility in the paramagnetic phase
is given by
1
χ (T ,B = 0) = ⇒ χ (T → Tc+,B = 0) ∼ (T − Tc )−1, (1.31)
T −1
and diverges upon approaching the critical temperature. Finally, if one keeps the
temperature fixed at the critical value T = 1, and adds a small magnetic field B,
the magnetisation is then
meq (T = Tc,B → 0) ∼ (B/b0 )1/3 ∼ B 1/3 . (1.32)
The phase transition is said to be of second order, because the free energy and its
first derivatives with respect to T and B are continuous at Tc ; the first singular term
in the free energy appears in its second derivatives.
Many physical observables scale as power laws in the vicinity of the critical
temperature Tc = 1. We have seen that meq ∝ |T −Tc |β with β = 1/2, the magnetic
12 Infinite-Dimensional Models in Statistical Physics

susceptibility χ ∝ |T − Tc |−γ with γ = 1, and meq ∝ B 1/δ with δ = 3. The power-

law scaling of physical observables close to Tc is dubbed the critical behaviour, and
β, γ , δ are the corresponding critical exponents. It is easy to see that adding other
terms in the expansion of s0 (m) in the vicinity of m = 0 (the next term is propor-
tional to m6 ) does not change the critical exponents, which are entirely determined
by the first terms in the expansion that have been kept in Eq. (1.29), as recognised by
Landau. For this reason, Eq. (1.29) is known as ‘Landau free energy’ and the result-
ing theory of the phase transition in the Ising model is the ‘Landau theory’ [69].

1.3.2 Corrections to Mean Field and Landau Theory

The next task is to investigate the corrections to Eq. (1.29) that come from higher-
order terms in the 1/d expansion. Expanding the term of order 1/d for small m in
Eq. (1.25), we obtain
a 1 2 b1 4
δv(m) = m + m − Bm,
2 4 (1.33)
β T β
a1 = T − 1 + , b1 = − .
2d 3 2d
Because all orders of the 1/d expansion are analytic functions of m and preserve
the inversion symmetry of the original model at B = 0, the form of the free energy
remains a polynomial in m2 . The coefficients are modified by 1/d corrections, so
their numerical values at a given d are modified, but they remain analytic functions
of T at all orders. Hence, at any order in 1/d, the free energy has the form of the
Landau theory, and all the consequences of the Landau free energy remain valid.
In particular, a second-order phase transition takes place with the same critical
exponents β, γ , δ. The numerical values of the observables are, however, changed.
For example, at first order in 1/d the critical temperature Tc , defined by a1 = 0, is
Tc = 1− 2d1 . One concludes that while the critical behaviour of the large d expansion
is independent of the order in d as a consequence of analyticity and symmetry, as it
was understood by Landau, its quantitative results strongly depend on the order at
which the expansion is truncated.
How do these results compare with the true behaviour of the Ising model in finite
d? In d = 1, there is no phase transition [69, 223], so the large d approximation
fails badly, and this choice of d will not be further considered. In d = 2, the
ferromagnetic Ising model has been solved exactly by Onsager for B = 0 [69].
In d > 2, while there is no exact solution, many properties of the model are
known very accurately by a combination of Monte Carlo numerical simulation,
renormalisation group and conformal bootstrap methods. In short:

• The critical properties of the model are correctly described by the 1/d expansion
for all dimensions d ≥ du = 4, where du is the upper critical dimension of the
 1.3 Second-Order Phase Transition of the Ising Ferromagnet 13

model. For d = 3 and d = 2, the phase transition remains of second order, but
the critical exponents are different. Around the critical point, the potential has the
‘scaling’ form11 [69, chapter 7]
δv(m) = −Bm + m1+δ F[(T − Tc )m−1/β ], m → 0, T → Tc, (1.34)
where F(x) is a scaling function that is analytic in x. At the critical point T = Tc ,
in order to recover that mδeq ∝ B for small B, we must have F(0) > 0. For
m → 0 at T = Tc – i.e. slightly away from the critical point – the susceptibility
χ ∝ 1/v (m = 0) is finite. All higher-order derivatives of δv(m), being related
to finite higher-order susceptibilities, remain finite too. Hence, δv(m) must be an
analytic function of m2 . This condition is satisfied if the function F(x) has the
form, for x → ±∞,
∞

F(x) = cn± |x|γ −2(n−1)β = c1± |x|γ + c2± |x|γ −2β + · · · , (1.35)
n=1

with γ /β = δ − 1. In fact, plugging this result in Eq. (1.34), one obtains for
12

small m and T = Tc :
δv(m) = −Bm + c1± |T − Tc |γ m2 + c2± |T − Tc |γ −2β m4 + · · · , (1.36)
where the coefficients cn+ correspond to T > Tc and cn− to T < Tc . One concludes
that δv(m) ∝ m2 with a finite coefficient a ∝ 1/χ ∝ |T − Tc |γ , and the
susceptibility13 is χ ∼ |T − Tc |−γ . For T < Tc , assuming c1− < 0, this form
implies that meq ∼ |T − Tc |β . Equations (1.34) and (1.35) thus encode correctly
the critical behaviour around Tc . They also imply that for d < 4, in which
case some critical exponents are not integers (Table 1.1), the free energy is not
an analytic function of T and m around the critical point. Such a non-analytic
behaviour cannot be derived from the 1/d expansion at any finite order. One must
therefore use other techniques, such as the renormalisation group, to understand
it in detail14 [11, 69, 282].
• The quantitative properties of the model – that is, the numerical values of many of
its observables of interest – can be obtained by truncating the 1/d expansion at a
suitable order. As an illustration, we consider the value of the critical temperature

11 For the free energy, Eq. (1.34) holds for T ∼ T and m ∼ 0, but with the additional constraint that |m| > m
c eq
for T < Tc . When T < Tc and |m| < meq , the free energy is given by the Maxwell construction
(Section 1.4). This constraint, however, does not affect the derivation of the scaling relations from Eq. (1.34).
Note that for d → ∞ one has F (x) = x/2 + T /12, which satisfies all the requirements discussed below and
reproduces Eq. (1.29).
12 This result is an example of a scaling relation that is obeyed by the critical exponents in all dimensions.
13 From Eq. (1.36) one can show that the nonlinear susceptibility χ , defined by m = χ B + χ B 3 + · · · ,
3 eq 3
diverges as χ3 ∼ |T − Tc |−γ3 with γ3 = 3γ + 2β, which is another scaling relation.
14 For the ferromagnetic Ising model, appropriate resummations of the 1/d expansion can nonetheless be used to
obtain reasonable estimates of the critical exponents. See [163] for a more complete discussion.
14 Infinite-Dimensional Models in Statistical Physics
Table 1.1 Critical properties of the Ising model as a function of spatial dimen-
sion d. The critical exponents are independent of d for all d ≥ 4, and for this
reason, they are not reported explicitly for d > 4. The values in d = 2 are
from Onsager’s exact solution. The critical exponents in d = 3 are obtained from
conformal bootstrap [137].

d β γ δ Tc (exact/MC)
2 1/8 7/4 15 0.567296. . .
3 0.326419(3) 1.237075(10) 4.78984(1) 0.75192(2) [54, eq. (3.1)]
4 1/2 1 3 0.835033(3) [237]
5 0.877844(2) [54, figure 5]
6 0.90290(5) [165]
7 0.91921(5) [165]
8 0.9311(6) [4]

[142, 163]. Fisher and Gaunt [142] have obtained the following result at the fifth
order in 1/d, and additional terms up to order twelve are given in [71]:
1 4 13 979 2009 176749
Tc = 1 − − 2− 3− − −
q 3q 3q 45q 4 15q 5 189q 6
6648736 765907148 5446232381 829271458256
− − − − (1.37)
945q 7 14175q 8 14175q 9 467775q 10
164976684314 6495334834824112
+ + ..., q = 2d.
22275q 11 638512875q 12
This result is compared with the true value of Tc (obtained from the exact Onsager
solution in d = 2 and from Monte Carlo simulations for d > 2) in Figure 1.1. The
fourth-order result gives a remarkably good approximation to the true result in all
d ≥ 2, but adding more terms to the expansion actually worsens the agreement
at the lowest dimensions. The convergence properties of the series are still poorly
understood. See [71, 142, 163] for a more detailed discussion.15

15 There are essentially two possible scenarios. One possibility is that the series converges for d ≥ 4, the upper
critical dimension. The other is that the series is asymptotic, as in the spherical model or in the O(n) model in
the infinite n limit [71, 142]. An asymptotic series in 1/d is a formal series expression of the form

Tc (d) = ∞ ak d −k such that, truncating the series to a given order K, one has
k=0
|Tc (d) − k=0 ak d −k | ≤ CK d −(K+1) for some K-dependent constant CK . In other words, by truncating
K
the series to a given fixed order, the error decreases upon increasing d. If one instead fixes d and increases K,
the error first decreases but then diverges for large K because the series is divergent for all finite d.
 1.3 Second-Order Phase Transition of the Ising Ferromagnet 15

Figure 1.1 The critical temperature of the ferromagnetic Ising model as a function
of d. Monte Carlo results are compared with different orders in the 1/d expansion,
from 1/d to 1/d 12 [71, 142, 163]. The vertical dotted line indicates the upper
critical dimension du = 4.

1.3.3 Correlations in Large Dimensions

The spin–spin correlation function is a very important quantity for analysing the
phase transition of the Ising model. According to Eq. (1.9), it is proportional to χij ,
and its explicit expression in the large d limit can be obtained as follows. As shown
in Section 1.3.1, only the first two lines of Eq. (1.22) remain finite for d → ∞ in
the ferromagnetic case. Then, from Eq. (1.14), we have

∂ 2 F [m] T
(χ −1 )ij = = δij − Jij . (1.38)
∂mi ∂mj 1 − m2i

Specialising to the paramagnetic phase, where mi = 0, one has (χ −1 )ij = T δij

− Jij , which must be inverted to obtain the correlation matrix.
To perform the inversion, one can use the periodicity of the lattice, working
in Fourier space. This standard procedure is here only sketched; see [282] for
details. Following the discussion of Section 1.1.1, we consider for simplicity a
d-dimensional cubic lattice of side L with periodic boundary conditions in all direc-
tions. This choice conveniently guarantees that translational invariance is preserved,
and other choices of boundary conditions would, in any case, lead to the same
result in the L → ∞ limit. We label lattice points by d-dimensional vectors –
e.g., x,y. The normalised eigenvectors of (χ −1 )xy are plane waves of the form
vx = L−d/2 eiqx . The periodic boundary conditions impose that the wavevectors
(q)
16 Infinite-Dimensional Models in Statistical Physics

q have the form q = 2π L

(n1, . . . ,nd ) where the nμ are integers that belong to the
interval (−L/2,L/2). The eigenvector equation is then
 
(χ −1 )xy vy(q) = (T δx,y − Jx,y )L−d/2 eiqy
y y
⎡ ⎤
d (1.39)
1
= L−d/2 eiqx ⎣T − cos(qμ )⎦ = λq vx(q),
d μ=1
d
with associated eigenvalues λq = T − 1
d μ=1 cos(qμ ). Denoting complex conju-
gation by an overline, one then obtains
 1 (q) 1  eiq(x−y)
χxy = vx(q) vy = d d . (1.40)
q
λq L q T − 1
μ=1 cos(qμ )
d

In the thermodynamic limit L → ∞, the sum over q becomes an integral, and one
obtains

dq eiq(x−y)
χxy =  d
= βSx Sy . (1.41)
[−π,π]d (2π) T − d
d 1
μ=1 cos(qμ )

The spin–spin correlation function obviously respects the translational symmetry


of the lattice and is therefore a function of x − y. Note that x eiqx = (2π)d δ(q),
 
and thus, χ = L−d xy χxy = x χx,0 = 1/(T − 1). As expected, we recover
the Curie–Weiss expression for the susceptibility given in Eq. (1.28). Note that the
critical temperature is here Tc = 1 because only the leading terms in d → ∞ have
been retained. The expression of χxy is thus only correct for T > Tc = 1.
To evaluate χxy , we choose y = 0 for convenience. To facilitate the integra-
tion over q, we also choose to evaluate the correlation along one of the principal
directions of the lattice, say 1, without loss of generality. Calling x = x1 , while
xμ = 0,∀μ = 2, . . . ,d, then

dq eiq1 x
Sx S0  =  d
[−π,π]d (2π) 1 − d
d β
μ=1 cos(qμ )
  ∞ 
dq iq1 x −λ[1− βd dμ=1 cos(qμ )]
= d
e dλ e (1.42)
[−π,π]d (2π) 0
 ∞
= dλ e−λ Ix (βλ/d)I0 (βλ/d)d−1,
0

where In (z) are the modified Bessel functions of the first kind. This last expres-
sion can be easily evaluated to visualise the shape of the correlation function; see
Figure 1.2. A detailed analysis of its asymptotic properties can be found in [282].
At a given T > 1, there are two distinct regimes of correlation decay.
 1.3 Second-Order Phase Transition of the Ising Ferromagnet 17

logSx S0 

Figure 1.2 Spin–spin correlation function of the ferromagnetic Ising model,

evaluated using Eq. (1.42), in dimension d = 25 at temperature β = 0.96.
The short-distance exponential decay corresponding to Eq. (1.43) is well visible,
while at large distances, the correlation slowly converges to the critical exponential
decay given by Eq. (1.44).

• If the distance x is kept fixed while d → ∞, one can use the small-argument
expansion of the Bessel functions to show that16
 x
β 1
Sx S0  ∼ = e−x/ξ1 , ξ1 = − . (1.43)
2d log(β/(2d))
In this regime, the correlation decays exponentially on a length scale ξ1 . When
d → ∞, ξ1 → 0, and hence, for any x = 0, the correlation vanishes. One
concludes that in the d → ∞ limit, each spin is uncorrelated from all the other
spins, consistently with the mean field construction. Remarkably, ξ1 does not
diverge at the critical point β = 1. In other words, correlations decay quickly
even at the critical point.
• If instead d is kept fixed, while x → ∞, one can use the large-n asymptotics of
In (z) to show that
1
Sx S0  ∼ e−x/ξ2 , ξ2 = ∝ (1 − β)−1/2 . (1.44)
arccosh(d/β − d + 1)
In this large-distance regime, the correlation decays exponentially with a length
scale ξ2 that diverges at the critical point, with critical exponent 1/2.

16 This result can also be derived either at the first order in the high-temperature expansion or by considering a
model defined on a Bethe lattice in the large connectivity limit [251].
18 Infinite-Dimensional Models in Statistical Physics

Therefore, the limits d → ∞ and x → ∞ do not commute. For strictly infinite d,

correlations vanish even for nearest-neighbour spins. For very large but finite d,
there are two regimes: at small distances, the decay is exponential with a length ξ1 ,
which remains small even at the critical point; for much larger distances (that
diverge upon increasing d), the decay is characterised by a second length ξ2 , which
diverges at the critical point. At a second-order phase transition, even in the mean
field approximation, which keeps only the leading terms in d → ∞, a correlation
length diverges in any finite d. Including additional terms in the 1/d expansion does
not qualitatively change this picture.
Note that when d is large, the total magnetic susceptibility is dominated by the
short range decay of the spin–spin correlation. The reason is the following. The
correlation between a spin and its nearest neighbours (x = 1) is β/(2d) and,
thus, vanishes quickly. However, there are 2d nearest neighbours, so their total
contribution to the susceptibility is β/(2d) × 2d = β, which remains finite for
d → ∞. The Curie–Weiss expression of the total susceptibility, χ = 1/(T − 1),
which is exact for d → ∞, can be recovered from the short-distance regime in
Eq. (1.43) by a careful computation, in which the correlations are also evaluated
off the principal axes. The large-distance regime gives negligible contributions in
large d. Hence, while correlations between individual spins vanish, the number
of neighbouring spins diverges. The two effects compensate one another, giving
rise to an overall finite susceptibility. In mean field, one does not need a diverging
length scale to obtain a diverging χ , because χ diverges while ξ1 remains finite.
Remember, however, that ξ2 diverges even if its associated correlations do not
contribute to χ .

1.4 Low-Temperature Ferromagnetic Phase

In Section 1.3, we discussed the simplest possible situation, that in which v(m) has
a single minimum, corresponding to the paramagnetic state at m = 0. We studied
how this minimum becomes unstable and bifurcates into two minima, thus giving
rise to a second-order phase transition at Tc . In this section, we study in more detail
what happens in the ferromagnetic phase, for T < Tc , in which v(m) has more than
one minimum.
Recall that in Section 1.1, it was shown that the thermodynamic free energy is
the absolute minimum of F [m], as expressed by Eq. (1.16). In Section 1.2.3, the
special case of uniform magnetisation was considered, defining the potential
1
v(m) = F [{mi = m, ∀i}], (1.45)
N
which is the free energy of a system constrained to have uniform magnetisation m
and is given by Eq. (1.25). It is also useful to introduce
 1.4 Low-Temperature Ferromagnetic Phase 19
1
f(m) = min F [m], (1.46)
N m:m=N −1 i mi

which is the true thermodynamic free energy of a system constrained to have a given

global magnetisation m = N −1 i mi and otherwise free to choose the set of local
magnetisations that minimises F [m]. From these definitions, it follows that f(m)
≤ v(m) and that the thermodynamic free energy per spin is f = minm f(m). Note
that there is no guarantee that the absolute minimum of v(m) gives the thermody-
namic free energy because the absolute minimum of F [m] could be realised by a
non-uniform configuration, as it happens in antiferromagnets.

1.4.1 Multiple Equilibrium States and Phase Coexistence

The structure of the equilibrium states of the ferromagnetic d-dimensional Ising
model is very well understood and is rigorously proven in most cases [158, 297,
310, 328]. We thus know for certain that the following properties hold.

• An ‘equilibrium state’ of an infinite system is defined as follows. To construct the

infinite system size limit, one considers a sequence of systems of N spins with
some prescribed boundary conditions BN for each N (examples have been given
in Section 1.1.1). If the limit
   
Si1 · · · Sin = lim Si1 · · · Sin N, BN (1.47)
N→∞

exists for all possible sets of n spins,17 the resulting set of correlation functions
defines an equilibrium state of the infinite system.
• A ‘translationally invariant equilibrium state’ is an equilibrium state such
that
 the expectation
  value of  any observable is translationally invariant – i.e.,
Sx1 · · · Sxn = Sx1 +y · · · Sxn +y ,∀y.
• B = 0, or if B = 0 and T > Tc , there is only one equilibrium state, and it is
If
translationally invariant. In other words, in the thermodynamic limit, all possible
sequences of boundary conditions give the same correlation functions and, in
particular, the same uniform magnetisation m = Si .   
• For B = 0 and T < Tc , let us denote by Si1 · · · Sin + and Si1 · · · Sin − the equi-
librium states obtained by fixing frozen boundary conditions (Section 1.1.1), with
external spins all fixed to S = 1 and S = −1, respectively. The thermodynamic
limit with these boundary conditions exists and defines equilibrium states that
are
 translationally
 invariant
 and are obviously related by inversion symmetry:
Si1 · · · Sin + = (−1) Si1 · · · Sin − . In particular, meq = Si + = − Si − . These
n

17 According to the discussion of Section 1.1.2, this implies the existence of the thermodynamic limit for all
observables.
20 Infinite-Dimensional Models in Statistical Physics

two states thus correspond to uniform states with positive and negative magnetisa-
tion ±meq . By symmetry, they have the same free energy v(meq ) = f(meq ), which
is the lowest possible free energy. It therefore corresponds to the thermodynamic
free energy – i.e., F = Nf(meq ) = N f(−meq ).
• Both the unique equilibrium state for B = 0 or B = 0, T > Tc , and the two states
+ and − for B = 0 and T < Tc , enjoy the so-called ‘clustering property’. That is,
the spin–spin correlation function (Si − m)(Sj − m) = Si Sj  − m2 that appears
in Eq. (1.9) vanishes when the distance between spins diverges.18 Well-separated
spins are thus uncorrelated – i.e., Si Sj  ∼ Si Sj  = m2 . This property is

necessary, for example, to ensure that the susceptibility χ = N −1 ij χij is finite.
Moreover, recalling that χij = ∂mi /∂Bj , the clustering property ensures that the
response of spin i to a variation of the local field Bj acting on a faraway spin
vanishes. This condition is a minimal stability requirement for a physical system.
Equilibrium states that satisfy the clustering property are said to be ‘pure states’.
• For all T < Tc , the two pure states + and − can also be constructed by taking
the limit of zero magnetic field of the unique equilibrium state that exist for
B = 0 – i.e.,
•± = lim •B . (1.48)
±B→0

This relation is very important. It implies that pure states can be constructed
equivalently by a sequence of appropriate boundary conditions or by adding a
weak external magnetic field that is sent to zero after taking the thermodynamic
limit.
• For B = 0 and T < Tc , any other translationally invariant equilibrium state can
be written as a convex combination of the two pure states:
     
Si1 · · · Sin α = α Si1 · · · Sin + + (1 − α) Si1 · · · Sin − , (1.49)
with α ∈ [0,1]. In particular, the magnetisation mα = αmeq − (1 − α)meq =
(1 − 2α)meq can take any possible value in [−meq,meq ]. Note that according
to Eq. (1.49), the system can be found with probability α in state + and with
probability 1 − α in state −, uniformly in space. This behaviour is unphysical;
a classical spin system cannot be in two states simultaneously. Another, perhaps
more transparent, reason why these states
 are unphysical
 is that theydo not enjoy

the clustering property. In fact, Si Sj α = α Si Sj + + (1 − α) Si Sj − = Si Sj +
   
by symmetry. Then, at large distance, Si Sj α ∼ Si Sj + ∼ (meq )2 = (mα )2
unless α = 0 or α = 1. This result implies that the response of spin i to a
weak field Bj remains large even when i and j are very distant. Equivalently, a

18 A similar property is true for all n-spin ‘connected’ correlation functions. For a detailed discussion, see, e.g.,
[158, 282].
 1.4 Low-Temperature Ferromagnetic Phase 21

+ −

+ −

L +meq −meq

+ −

ρL (1 − ρ)L
+ −

Figure 1.3 Schematic illustration of phase coexistence in the ferromagnetic Ising

model in d = 2. Here x ∈ [0,L] is the coordinate of the horizontal direction.
Frozen boundary conditions are used, with S = 1 (S = −1) on the left (right)
vertical faces, while on the horizontal faces S = 1 for x < ρL and S = −1 for x >
ρL. A fluctuating interface separates the two coexisting states with equilibrium
magnetisation ±meq . Here it is assumed that L is large enough that the interface
looks smooth at this scale. Note that the interface is not sharp for d = 2 because
its fluctuations diverge for L → ∞, contrary to what happens for d ≥ 3 and low
enough temperature.

weak field applied on a single spin can perturb the entire system, which has a
catastrophically destabilising effect. Therefore, although translationally invariant
states defined in Eq. (1.49) formally exist, they are not physical. The only stable
and translationally invariant states are the two pure states19 + and −.
• For B = 0 and T < Tc , non-translationally invariant equilibrium states may exist.
This happens if one can separate the system in homogeneous regions separated by
sharp interfaces. Such sharp interfaces can only exist for d ≥ 3 and sufficiently
low temperatures, T < Tr < Tc [158], and the interfacial phase transition that
happens at Tr is called ‘roughening transition’. A particularly interesting example
of such states is obtained by choosing one particular direction, say x ∈ [0,L], and
setting frozen boundary conditions with S = 1 on one side, x < ρL, and S = −1
on the opposite side, x > ρL, with ρ ∈ [0,1], as illustrated in Figure 1.3. In

19 In general, by considering the class of translationally invariant states, one can prove that clustering states
cannot be written as a convex linear combination of other states, and that states which cannot be written as a
convex linear combination of other states are clustering [158].
22 Infinite-Dimensional Models in Statistical Physics

this case, for large L, the system has positive magnetisation meq for x  ρL
and negative magnetisation −meq for x  ρL. The two regions are separated
by an interface located around x ∼ ρL. In the thermodynamic limit, a fraction
ρ of the volume has positive magnetisation, and a fraction (1 − ρ) has negative
magnetisation. We denote mρ the local magnetisation of such a configuration, and
the corresponding total magnetisation is mρ = (1 − 2ρ)meq + O(1/L). Far from
the interface, the two regions of positive and negative magnetisation have the
same free energy per spin, v(meq ). Around the interface, the mismatch between
the two magnetisations introduces an additional free energy term. Because the
couplings are short range, this term only receives contributions from the region
of space around the interface and is thus proportional to the size Ld−1 of the
interface,20
F [mρ ] = Ld v(meq ) + Ld−1 σ (β) + o(Ld−1 ). (1.50)
The quantity σ (β) = limL→∞ {F [mρ ] − Ld v(meq )}/Ld−1 is called the ‘surface
tension’, and it is independent of ρ in the thermodynamic limit.21 From Eq. (1.50),
one deduces that
v(meq ) ≤ f(mρ ) ≤ v(meq ) + σ (β)/L. (1.51)
The right inequality follows from Eq. (1.46), which implies f(mρ ) ≤ F [mρ ]/Ld ,
with F [mρ ] given by Eq. (1.50). The left inequality follows from the fact that
v(meq ) = f(meq ) is the absolute minimum of F [m]/N, as stated above, hence
f(m) ≥ v(meq ), ∀m. One concludes that in the thermodynamic limit L → ∞,
f(mρ ) → v(meq ),∀ρ ∈ [0,1]. In other words, the function f(m) is constant and
equal to v(meq ) for m ∈ [−meq,meq ]. This is a particular case of the more general
‘Maxwell construction’ that will be discussed in Section 1.5. We have constructed
an example of non-translationally invariant states that have any possible magneti-
sation m ∈ [−meq,meq ] and have the same free energy as the pure states in the
thermodynamic limit. This mechanism, by which one can construct equilibrium
states with the two phases (+ and −) occupying different regions of the system,
is called ‘phase coexistence’.
In summary, for T < Tc and B = 0, the function f(m) has the form depicted
in Figure 1.4. The two homogeneous pure states minimise the free energy at
m = ±meq . In the interval m ∈ (−meq,meq ), phase coexistence ensures that f(m) is

20 For example, suppose that temperature is very low, such that m ∼ 1. In this limit, the interface is extremely
eq
sharp, and m(x) = sgn(ρL − x). The energy excess over the ground state energy, given by Eq. (1.3), due to
the interface is 2J Ld−1 . One deduces that Eq. (1.50) holds with σ (T → 0) = 2J = 1/d.
21 To be more precise, one should take into account in Eq. (1.50) the presence of the interface between the bulk
of the system and the boundary, which also gives a contribution of the order of the surface. The first term in
Eq. (1.50) should then be replaced by the free energy of a system with all + boundary condition [158].
 1.4 Low-Temperature Ferromagnetic Phase 23

v(m) v(m)

f(m) f(m)

m m

Figure 1.4 The functions v(m), Eq. (1.26), and f(m) for the Ising ferromagnet at
T = 0.8 < Tc , in the thermodynamic limit N → ∞ and for d → ∞. Left: B = 0.
Right: B = 0.02.

constant, f(m) = v(meq ). The main point of this section is that the presence of a flat
region in the free energy f(m) generally signals the presence of multiple equilibria
and phase coexistence. At the same time, the Legendre transform of f(m), which is
the free energy f(B) in presence of a uniform magnetic field B, has a discontinuous
derivative in B = 0 – i.e., m(B) jumps at B = 0.

1.4.2 Phase Coexistence in Large Dimension

We now examine the behaviour of the free energy in the phase coexistence region
in the large d limit. First of all, we note that in this limit, connected correlations of
distinct spins vanish in pure states. The discussion in Section 1.3.3 can indeed be
easily extended to the low-temperature phase and even to non-translationally invari-
ant pure states that appear in disordered systems. In other words, Si Sj  = Si 
Sj  = mi mj , ∀i = j . Therefore, in the infinite-dimensional limit, a pure state
is completely determined by the set of local magnetisations,22 mi , i = 1, . . . ,N,
which in the ferromagnetic Ising model are all equal by translational invariance.
For pure states in d → ∞, spins are thus independent, and one can write

N
1 + mi S
P (S1, . . . ,SN ) ∼ pi (Si ), pi (S) = , (1.52)
i=1
2

where pi (S) is a single-spin probability distribution6 with mi = Si .

We now specialise to the translationally invariant case with mi = m. The 1/d
expansion starts from a probability distribution of independent spins with fixed
magnetisation m for d → ∞ and perturbs systematically around it by introducing

22 Note that this is not true for non-pure states. For example, for the states • discussed in Section 1.4.1, one
α
has Si Sj  = (meq )2 = m2 .
24 Infinite-Dimensional Models in Statistical Physics

correlations in a controlled way. This approach makes sense for the translationally
invariant pure states at m = ±meq because the magnetisation then remains homo-
geneous at all orders in 1/d. The same reasoning holds for |m| > meq , when the
free energy v(m) = f(m) is also dominated by translationally invariant states.
But what happens when m ∈ (−meq,meq )? To illustrate the situation, one can
focus on m = 0. In this case, the starting point of the perturbation expansion in
1/d is the completely random distribution P (S1, . . . ,SN ) = 2−N , in which each
spin can equally likely be found in the state 1 or −1, with a local spin distribution
p(S) = 1/2. However, we know from the analysis of Section 1.4.1 that the free
energy f(m = 0) is dominated by completely different spin configurations, in which
half of the system has m = meq , while the other half has m = −meq , as illustrated in
Figure 1.3. This phase-separated configuration cannot be constructed perturbatively
from the homogeneous m = 0 state. The perturbative free energy v(m) given in
Eq. (1.25), truncated at any given order in 1/d, therefore represents the free energy
of the the homogeneous m = 0 state – i.e., the paramagnetic state. This state has a
higher free energy than phase-separated inhomogeneous states, and for this reason,
v(m = 0) > f(m = 0) in the phase coexistence region, as represented in Figure 1.4.
Moreover, v(m) is not convex in this region and thus cannot be the correct free
energy. The large-d expansion must therefore be handled with some care in the
phase coexistence region.
It is also possible to consider an inhomogeneous density profile as the starting
point of the 1/d expansion. For instance, we can consider Eq. (1.22) with ferro-
magnetic couplings Jij , divide the system in two by an interface and set mi = meq
for half of the system and mi = −meq for the other half. In that case, we correctly
get the thermodynamic free energy f = F [m]/N = v(meq ) in the thermodynamic
limit. We thus see that the convexity of f(m) holds (weakly) in the phase coexistence
region, as seen in Figure 1.4, but is lost when the perturbative expansion in 1/d
around the homogeneous state is truncated at any finite order.

1.4.3 Dynamics in the Phase Coexistence Region

To conclude the discussion of phase coexistence, and in preparation for the dis-
cussion of metastability in Section 1.5, it is important to mention the dynamical
properties of the Ising model in the phase coexistence region – i.e., at B = 0 and
T < Tc .
The dynamics in this region depends strongly on the boundary conditions. As an
illustration, consider a system of finite size L, with periodic boundary conditions,
undergoing a local dynamics23 that satisfies detailed balance (e.g., Metropolis or

23 A local dynamics is such that spins are updated according to a rule that involves a finite number of
neighbouring spins.
 1.4 Low-Temperature Ferromagnetic Phase 25

Glauber dynamics, which will be reviewed in more detail in Chapter 3), starting at
time t = 0 from the configuration with all spins Si (t = 0) = −1, hence with
magnetisation m(t = 0) = −1. The following statements can then be proven
rigorously, at least in d = 2 [246].

• In a time that remains finite for L → ∞, the magnetisation reaches a value

m(t) ∼ −meq , characteristic of the negatively magnetised pure state.
• At larger times, the magnetisation can diffuse randomly to values m > −meq .
In particular, the time needed to reach any positive value of the magnetisation,
m(t) > 0 (e.g., m(t) ∼ meq ), is
d−1
τ ∼ e2βσ (β)L , (1.53)
where σ (β) is the surface tension introduced in Section 1.4.1.
This result shows that the time needed to explore all the thermodynamically relevant
values of the magnetisation, m ∈ [−meq,meq ] (these are the values that minimise
the free energy), actually diverges exponentially in Ld−1 for L → ∞. In this limit,
the system is said to be non-ergodic, because it takes an infinite amount of time to
explore all the thermodynamically relevant part of the configuration space.
The scaling in Eq. (1.53) can be understood by the following simple argument. In
order to reach positive values of the magnetisation, at some point the system needs
to visit configurations with m = 0. We have already seen that the minimal possible
free energy at m = 0 is realised by phase coexistence. With periodic boundary

80
Energy relative to ground state

70

60

50

40

30

20

10

0
0 50 time t 100 150

Figure 1.5 Illustration of a transition between negative (white) and positive

(black) magnetisation, for a ferromagnetic Ising model in d = 2, in a square lattice
with L = 10 and N = 100, at temperature T = 0.25 under periodic boundary
conditions. Intermediate configurations have m ∼ 0; in these configurations, two
black and white regions are separated by two interfaces.
26 Infinite-Dimensional Models in Statistical Physics

conditions, the best possible arrangement is to have half of the system in the meq
state and the other half in the −meq state, the two parts being separated by two
interfaces, as illustrated in Figure 1.5. The free energy is then F = v(meq )Ld
+ 2σ (β)Ld−1 , and the free energy excess over equilibrium is δF = 2σ (β)Ld−1 .
Because the system relaxes to a partially equilibrated state around magnetisation
m(t) ∼ −meq in a finite time, one can use the theory of equilibrium fluctua-
tions [223] to estimate the probability of observing such a fluctuation starting from
m = −meq . The resulting escape probability is pe ∼ exp(−βδF ), from which it
follows that the time needed to observe such fluctuation is τ ∼ 1/pe and, hence,
Eq. (1.53).

1.5 Metastable States

The aim of this section is to introduce metastable states, which are crucial in the
study of glasses. In this section, we present the essential ideas while still focusing
on the ferromagnetic state, in which the system is translationally invariant, and
metastable states respect this symmetry. The equivalent concept for disordered
systems will be discussed in Chapter 4. Detailed reviews on metastability can be
found in [52, 120], and a short pedagogical introduction can be found in [55].
In Section 1.4, we first discussed the structure of the equilibrium states in finite
d (Section 1.4.1) and then the large d approximation (Section 1.4.2). This was
possible because phase coexistence in finite d is mathematically well understood.
The theory of metastability in finite d is, by contrast, not well established. It is
therefore convenient to discuss first the large d behaviour and then its finite d
extension.

1.5.1 Local Minima of the Free Energy in Large d

Consider the free energy of the Ising ferromagnet for a uniform magnetisation
mi = m,∀i and in the limit d → ∞, given in Eq. (1.26). Its shape for T < Tc and
small enough B is shown in Figure 1.4. As in the case B = 0, the potential v(m)
that corresponds to uniform magnetisation is non-convex. It has two local minima at
positive and negative magnetisation, separated by a local maximum situated around
m = 0. The term −Bm in the free energy has the effect of lowering the free energy
of the minimum that is magnetised parallel to the field, i.e., such that Bm > 0.
Hence, for B = 0, the function v(m) has a unique global minimum. This result is
coherent with the discussion of Section 1.4.1, which stated that for B = 0, at all
temperatures, there is a single translationally invariant pure state. This state is the
global minimum of v(m).
As in the case B = 0, in the region where v(m) is not convex, one can examine
what happens if the magnetisation is allowed to be non-uniform. Consider a global
 1.5 Metastable States 27

magnetisation m. One can always choose two arbitrary values of magnetisation

m± , such that m ∈ [m−,m+ ], and separate the system in two regions, in such a
way that a fraction ρ of the system has magnetisation m− and a fraction 1 − ρ has
m+ , with ρ fixed by m = ρm− + (1 − ρ)m+ + O(1/L). If the two regions are
separated by a flat interface of size Ld−1 , Section 1.4.1, the free energy per spin is
f(m) = ρv(m− )+(1−ρ)v(m+ )+σ (β,B)/L, where σ (β,B) is the surface tension.
In the thermodynamic limit, one then obtains

m = ρm− + (1 − ρ)m+, m ∈ [m−,m+ ],

(1.54)
f(m) = ρv(m− ) + (1 − ρ)v(m+ ).

Therefore, the plot of f(m) is a straight line connecting the points {m−,v(m− )}
and {m+,v(m+ )}. Geometrically it is clear that for all m, f(m) is minimised by
choosing m− and m+ such that f(m) is the convex envelope of v(m), as illustrated in
Figure 1.4. From this geometrical construction, which is nothing but the Maxwell
construction [52], one can see that for any B = 0, there is a compact interval
[m− (B),m+ (B)] (the phase coexistence region) over which v(m) is not convex.
For m ∈ / [m− (B),m+ (B)], one has f(m) = v(m), while for m ∈ [m− (B),m+ (B)],
f(m) is the convex envelope of v(m), according to Eq. (1.54).
Without loss of generality, one can restrict the discussion to the case B > 0;
the other case is symmetric under m → −m. The equilibrium magnetisation
meq (B) > 0, which corresponds to the global minimum of v(m), falls outside
of the phase coexistence region, consistently with the equilibrium state being
translationally invariant. Instead, the local minimum with mm (B) < 0 is very close
to m− (B) but falls within the phase coexistence region. The aim of the rest of
this section is to show that the local minimum can be interpreted as a metastable
state. A metastable state is similar to a thermodynamic state, around which the
system, if it is initialised with a sufficiently negative m, spends a lot of time before
eventually reaching the equilibrium state. This discussion therefore requires a
careful investigation of the dynamics. Beforehand, it is useful to stress three points.
1. For large enough B, the secondary minimum disappears. The value of B at
which this minimum disappears defines a ‘spinodal point’ [52]. Beyond the
spinodal, no metastability exists.
2. The picture described in this section is derived explicitly in d → ∞ and remains
valid in any order in the 1/d expansion. However, the dynamics in the strict limit
d → ∞ is very different from that in any finite d, even when d is very large.
This will be discussed in Section 1.5.3.
3. Because mm is a local minimum of v(m), all the thermodynamic relations, such
as Eqs. (1.8) and (1.9), remain valid. The usual thermodynamic relations hold in
a metastable state during the time the system remains confined within it.
28 Infinite-Dimensional Models in Statistical Physics

1.5.2 Metastable Dynamics in Infinite Dimensions

In this section, in order to discuss metastable dynamics in infinite dimensions,
we consider a simple example for which the dynamics and the role of the local
minima of v(m) can be discussed exactly. This is a microscopic realisation of the
ferromagnetic Curie–Weiss model, already introduced in Section 1.3.1. Note that
we here follow closely the discussion of [311].

Dynamics of an Ising Spin System

Consider a system of N spins, interacting with each other via the Ising Hamiltonian
in Eq. (1.1), and undergoing a continuous time Markovian dynamics entirely char-
acterised by the rates wS;S  of jumping from configuration S  to S. The evolution
of the probability distribution of spin configurations pt (S) can be described by a
master equation [159]:
 
∂t pt (S) = wS;S  pt (S  ) − wS  ;S pt (S) . (1.55)
S

The first term describes jumps from any S  into S, while the second term described
jumps from S to any S  . We denote by S i = {S1, . . . , − Si , . . . ,SN } the config-
uration that differs from S by flipping spin i, and specialise to dynamics that can
only flip spins sequentially. In other words, we assume that wS;S  vanishes unless
S  = S i for some i. The energy change under one spin flip is then24

E = H (S) − H (S i ) = −2ĥi Si , ĥi = Bi + Jij Sj , (1.56)
j (=i)

and transition rates are assumed to depend only on E. In this case, one has
wS;S i = w(E) = w(−2ĥi Si ); hence, we can write the master equation as
 
∂t pt (S) = w(−2ĥi Si )pt (S i ) − w(2ĥi Si )pt (S) . (1.57)
i

The first term describes the probability of flipping spin i so that the system goes
into state S from S i . The second term describes all the events by which the system
leaves S by a single-spin flip. For simplicity, we choose
w(E) = e−βE/2, (1.58)
which is consistent with the detailed balance condition
w(E) = e−βE w(−E), (1.59)

24 The ‘microscopic’ field ĥ bears a hat to emphasise that it depends on the configuration S and, thus, to
i
distinguish it from the average field hi given by Eq. (1.24).
 1.5 Metastable States 29

but any other choice of rate that satisfies Eq. (1.59) guarantees that for t → ∞,
pt (S) converges to the Gibbs–Boltzmann equilibrium distribution with Hamilto-
nian given by Eq. (1.1). This point will be discussed in more detail in Chapter 3.

The Curie–Weiss Model

The study of Ising spin dynamics of d-dimensional systems, even in the limit
d → ∞, is fairly involved. The mean field Curie–Weiss model, which corresponds
to Eq. (1.1) with Jij = 1/(2N) for any pair ij and constant magnetic field Bi = B,
provides instead an easily tractable case. The Hamiltonian then depends only on the
global magnetisation,
1 
H =− M 2 − BM, M= Si . (1.60)
2N i
N
By observing that the number of configurations with magnetisation M is (M+N )/2
,
one can define a free energy at constant magnetisation:
⎡ ⎤

F (M) = −T log ⎣ e−βH (S) ⎦

S| Si =M
i (1.61)
  
N 2
= −T log eβM /2+βBM ≈ N v(m),
(M + N)/2 N→∞

where m = M/N , v(m) is given in Eq. (1.26) (and illustrated in Figure 1.4), and the
limit N → ∞ is easily obtained by Stirling’s formula. The free energy per spin of
the Curie–Weiss model in the thermodynamic limit thus coincides with Eq. (1.26),
which is the leading order for d → ∞ of the d-dimensional Ising ferromagnet with
an appropriate choice of the scale of couplings, as discussed in Section 1.2.3. It is
interesting to note, however, that the Curie–Weiss model defined by Eq. (1.60) is
well defined for any finite N . It essentially describes the behaviour of a system of
N = Ld spins when L is finite and d is very large or, more precisely, the limit of
the d-dimensional model when d → ∞ before L does. Note that even L = 2,
the minimal linear system size, is enough to obtain a thermodynamically large
system of N = 2d spins when d → ∞. For N → ∞, minimising v(m) with
respect to m gives the thermodynamic free energy of the Curie–Weiss model and,
as discussed earlier, two states with positive and negative magnetisation are present
at low enough temperatures.
Because the Hamiltonian depends only on M, it follows that at any time t, pt (S)
also only depends on M (provided that this is true at t = 0). We define pt (M) as
 −1
N
pt (S) = pt (M) . (1.62)
(N + M)/2
30 Infinite-Dimensional Models in Statistical Physics

The master Eq. (1.57) acting on the 2N spin configurations can thus be reduced to a
simpler master equation, defined on the much smaller space of the N + 1 possible
values of the magnetisation M ∈ {−N, − N + 2 . . . N − 2,N}. For simplicity,
we focus on the case B = 0, although the discussion can easily be generalised to
B = 0. Injecting Eq. (1.62) into Eq. (1.57), we get
∂t pt (M) = w+ (M − 2)pt (M − 2) + w− (M + 2)pt (M + 2)
(1.63)
− [w− (M) + w+ (M)]pt (M),
with
 
N −M M +1 N − M β(M+1)/N
w+ (M) = w −2 = e ,
2 N 2
  (1.64)
N +M M −1 N + M −β(M−1)/N
w− (M) = w 2 = e ,
2 N 2
which has the form of a one-dimensional birth-death process [159, section 7.1]. By
construction, at long times, pt (M) converges to its equilibrium form
 
N
eβM /2 ∼ e−βNv(m) .
2
p∞ (M) ∝ (1.65)
(M + N)/2
Therefore, in the high-temperature phase, T > 1, the magnetisation converges with
very high probability to m = 0, while in the low-temperature phase, T < 1, it
converges with very high probability to one of the values corresponding to the
minima of v(m) – i.e., ±meq . More precisely, in the low-temperature phase, the
magnetisation first relaxes on a finite25 time scale τrel to the value ±meq that is clos-
est to the initial condition, as it can be checked explicitly by analysing Eq. (1.63)
(see [311, section 6] for details). Without loss of generality, suppose that the system
is initialised in Mstart < 0. The magnetisation then converges in a finite time to
−Nmeq so that if N → ∞ and times are kept finite, the system remains stuck
forever in the negatively magnetised equilibrium state.

Metastability and Transition Rate

However, we know that the system must eventually sample the two states ±Nmeq
with equal frequency. We are thus interested in the transition rate between these two
states. Transition rates for Eq. (1.63) can be calculated using standard techniques
for estimating mean first-passage times in discrete, one-dimensional birth-death
processes [159, section 7.4]. Consider a system starting in Mstart < 0 in t = 0
and call τ the time at which the system first reaches a state Mend > Mstart . We

25 The time is finite if the initial condition has m = 0. If the system instead starts at m = 0, then it takes a time
∼ log N to reach one of the two values ±meq .
 1.5 Metastable States 31

define the mean first passage time τ (Mstart → Mend ) = τ , where the average is
over the dynamics described by Eq. (1.63). The system is confined by a reflecting
barrier in M = −N , because the magnetisation cannot be smaller than this value.
Using this boundary condition and [159, eq. (7.4.12)], we get (K,K ,M denote
magnetisations and therefore increase in steps of two)

Mend 
K
1
τ (Mstart → Mend ) = φ(K) , (1.66)
φ(K  )w + (K
)
K=Mstart K  =−N

with

M
w− (K)
φ(M) = . (1.67)
w (K)
K=−N +2 +

Eq. (1.66) can be computed numerically for any finite N , in a time that grows only
polynomially in N . One can thus compute the mean first passage time in Mend /N
= mend > 0 of a system that starts in Mstart /N = mstart < 0 at time t = 0, for
B = 0 and T < Tc . An asymptotic analysis shows that the result does not depend
on the start and end points in the thermodynamic limit. It then reads

π 1
τ (mstart → mend ) = eβN [v(0)−v(meq )] . (1.68)
β [1 − β(1 − m2eq )](β − 1)

In this expression, we recognise the Arrhenius law, which states that the leading
order in the scaling of the characteristic time τ with N is given by

τ ∼ eβN ,  = v(0) − v(meq ), (1.69)

where the prefactor  is the free energy barrier that separates the two equilibrium
states. This result is easily generalised to the case B = 0, where the two minima
are non-degenerate,
τ (−meq → meq ) ∼ eβN [v(mmax )−v(−meq )],
(1.70)
τ (meq → −meq ) ∼ eβN [v(mmax )−v(meq )],
where mmax denotes the magnetisation at which v(m) has a local maximum
in [−meq,meq ].
Note that this barrier scaling can also be obtained by the following simple argu-
ment, along the lines of that presented at the end of Section 1.4.3. By continuity,
the probability to jump from positive to negative magnetisation is bounded from
above by the probability of reaching any intermediate value of m. Assuming that the
system first equilibrates in the metastable state, this probability is given, according
to the theory of equilibrium fluctuations [223], by e−βN[v(m)−v(meq )] . The best upper
32 Infinite-Dimensional Models in Statistical Physics

bound is obtained for m = mmax , and its inverse gives a lower bound on the mean
first passage time.

1.5.3 Escaping the Metastable State through Nucleation

In Section 1.5.2, we studied the metastable dynamics of a system of N spins with
fully connected interactions, which essentially amounts to considering the case in
which the dimension of space goes to infinity before N does. We now study what
happens in the opposite situation, when N → ∞ at fixed d. In this case, the anal-
ysis of the previous section cannot be applied and, in particular, the magnetisation
profile does not remain uniform. There exists a more efficient way to escape from
the metastable state – namely, the inhomogeneous nucleation of a critical droplet.
Consider a system in a cubic box of size L, which is prepared at T < Tc with
negative magnetisation and in presence of a positive magnetic field B > 0. In this
case, most spins take a negative value, with some rare positive exceptions. Due to
fluctuations, an island of positive spins can nevertheless spontaneously form in the
sea of negative spins. On a cubic lattice, the most favourable configuration for this
process is a ‘droplet’ made of a sub-cube of side R of the whole lattice, filled with
mostly positive spins (i.e., the equilibrium state) surrounded on all sides by negative
spins (i.e., the metastable state). Let us denote δv(B,T ) > 0 the difference of poten-
tial between the homogeneous metastable state and the homogeneous equilibrium
state, which is positive by construction. Because of the presence of the more stable
phase within the bulk of the sub-cube, the droplet gains a free energy δv(B,T )R d .
However, the surface of the droplet (with area 2dR d−1 ) then becomes an interface
between the two states and therefore increases the free energy by σ (B,T )2dR d−1 ,
which is the associated surface tension, as discussed in Section 1.4.1. The total free
energy difference between the droplet and the homogeneous initial metastable state
is therefore

δF (R) = −δv(B,T )R d + σ (B,T )2dR d−1 . (1.71)

For small R, δF (R) is positive; hence, the cost of forming an interface dominates
over the bulk free energy gain. For large R, however, δF (R) becomes negative
as the bulk dominates. The two regimes are separated by a point at which
dδF /dR = 0, with corresponding

2(d − 1)σ (B,T )

Rc (B,T ) = . (1.72)
δv(B,T )

A droplet of this size is said to be critical. Once such a critical droplet forms via
spontaneous fluctuations, the system can lower its free energy by increasing the
 1.5 Metastable States 33

droplet size; R grows and the equilibrium phase invades quickly all the volume.26
The associated free energy cost is
δF † (B,T ) = δF (Rc ) = 2σ (B,T )Rc (B,T )d−1
[2σ (B,T )]d (1.73)
= (d − 1)d−1 ;
δv(B,T )d−1
hence, the time needed to observe the spontaneous formation of such a droplet is
† (B,T )
τ ∼ eβδF . (1.74)
From this argument, one concludes that the time to form a critical droplet and
escape from the metastable state remains finite when L → ∞, for all spatial
dimensions d and for all B = 0, at finite T . It only diverges when T → 0 or
B → 0. These results can be confirmed rigorously [246]. Note that this case is very
different from phase coexistence (B = 0), discussed in Section 1.4.3, in which the
time to jump from negative to positive m diverges exponentially in L, and from the
infinite dimensional case of Section 1.5.2 in which it diverges exponentially in N .
A rough estimate of the free energy cost of forming the critical droplet can be
obtained at very low temperatures and small B. For T → 0, the surface tension
is simply the energy difference between up and down spins, σ (B,T → 0) ∼ 2J
= 1/d for J = 1/(2d), and the two states have magnetisation m ≈ ±1. For
small B, their potentials are thus v+ = v(m = 0) − Bm = v(m = 0) − B and v−
= v(m = 0) + B; hence, δv(B → 0,T → 0) ∼ 2B, and
(2/d)d 2e 1
δF † (B → 0,T → 0) = (d − 1)d−1 ≈ . (1.75)
(2|B|) d−1 d→∞ d |B|d−1

When d → ∞, from Eq. (1.72) with σ (B,T ) ∼ 1/d and δv(B,T ) remaining finite,
the size Rc of the critical droplet reaches a finite limit (which can be made arbitrarily
large upon decreasing B). If the system has a finite size L < Rc , the critical droplet
cannot form. In this regime, nucleation is impossible, and the only way to escape
the metastable state is through a homogeneous jump of the magnetisation, following
the mechanism discussed in Section 1.5.2. In that case, the time is exponential in
N = Ld . We conclude that, roughly,
 d
eβL , for L  Rc (B,T ),
τ ∼ βδF † (B,T ) (1.76)
e , for L  Rc (B,T ).

The logarithm of the escape time thus increases proportionally to Ld for small L,
and then saturates at a finite value when L ∼ Rc (B,T ). Note that δF † (B,T ) ∝

26 In an infinite volume, the nucleation argument is more complicated because many critical droplets can form
concurrently at different locations in the system, but the conclusion is very similar.
34 Infinite-Dimensional Models in Statistical Physics

Rc (B,T )d from Eq. (1.73), which ensures the continuity of Eq. (1.76) when
Rc (B,T ) ∼ L.
The limits L → ∞ and d → ∞, therefore, do not always commute. If d goes
to infinity at fixed L  Rc (B,T ), one is then asymptotically in the first case of
Eq. (1.76) and the infinite-d dynamics of Section 1.5.2 is recovered. Note that, as
discussed in Section 1.5.2, this is the typical situation in large d because a lattice
of size L = 2 is enough to obtain a thermodynamically large system, while the
critical droplet size Rc (B,T ) is large at small B. The metastable state then has a
lifetime exponential in N . If instead L → ∞ at fixed d, one is then always in the
second case of Eq. (1.76), and nucleation dominates. The metastable state then has
a finite lifetime, which becomes very long when d is large. In both cases, however,
the lifetime of the metastable state becomes extremely long as d grows.

1.6 Wrap-Up
1.6.1 Summary
In this chapter, we have seen that

• Large-d expansions can be constructed conveniently by defining the free energy

as a function of the appropriate order parameter and then using a high-temperature
expansion (Section 1.2).
• The leading (d = ∞) term of this expansion, which corresponds to the exact
solution of the infinite-dimensional model, coincides with the mean field result
and has the analytical structure of the Landau theory (Section 1.3.1).
• Upon adding more terms to the 1/d expansion, the free energy keeps the form
of a Landau theory, and, therefore, the critical exponents remain unchanged
from the mean field ones. The truncated 1/d expansion thus describes the
critical behaviour only for d ≥ du , the upper critical dimension (Section 1.3.2).
Estimating the critical exponents for d < du requires a resummation of the series
or the use of different techniques, such as the renormalisation group.
• The coefficients of the Landau free energy are themselves power series in 1/d.
Thus, by carefully truncating the 1/d expansion or performing appropriate resum-
mations, one can obtain accurate estimates of some of the physical observables
(Section 1.3.2).
• Correlations, even between  nearest
 neighbouring spins, vanish in the limit
d → ∞. For example, Si Sj ∝ 1/(2d) for nearest neighbours in the ferro-
magnetic Ising model. This explains why in the limit d → ∞, the mean field
approximation becomes exact (Section 1.3.3).
• Yet the number of neighbours grows with d. For example, each site of a cubic
lattice has 2d nearest neighbours; hence, their total contribution to the suscepti-
bility, 2d ×1/(2d), remains finite. The same happens for spins at larger distances.
 1.6 Wrap-Up 35

For this reason, the susceptibility generally remains finite when d → ∞ and
diverges at the critical point. In mean field, one can thus have both uncorrelated
spins and a divergent susceptibility (Section 1.3.3).
• Within mean field, one can define a correlation length ξ2 that diverges  at the
critical point, by considering the leading d → ∞ contributions to Si Sj and
taking the large-distance limit before the limit d → ∞ (Section 1.3.3). This
reproduces the result of Landau theory for the associated critical exponent.
• In the low-temperature phase, T < Tc , and for B = 0, there are two equivalent
equilibrium states. These states are homogeneous and have magnetisation ±meq .
One can form non-homogeneous states characterised by phase coexistence of the
two homogeneous states. These states have m ∈ [−meq,meq ] and the same free
energy per spin as the homogeneous states (Section 1.4.1).
• In d → ∞, the homogeneous states are local minima of v(m), the free energy
corresponding to uniform magnetisation mi = m, which can be computed
in a 1/d expansion. Non-homogeneous states make the function f(m), which
is the free energy for a global magnetisation m, constant over the interval m
∈ [−meq,meq ] (Section 1.4.2).
• Dynamically, below Tc , the system is not ergodic; it takes a time τ ∼ exp(Ld−1 )
to reach positive m for a system that starts at negative m (Section 1.4.3).
• For T < Tc and small enough B = 0, the function v(m) in the limit
d → ∞ displays a single absolute minimum, which is the unique equilibrium
state, and a secondary local minimum. The local minimum is a metastable state.
It behaves thermodynamically as an equilibrium state but on a restricted time
scale (Section 1.5.1).
• In the limit d → ∞, if the system is initialised close to the metastable state,
it remains there for a time τ ∼ exp(N). The metastable state therefore has
an infinite lifetime in the N → ∞ limit and is a real thermodynamic state
(Section 1.5.2).
• For any finite d (even if large), there is a critical droplet size Rc . For a system
size L  Rc , the droplet cannot form, and τ ∼ exp(Ld ). If instead L  Rc , the
system can escape from the metastable state in a finite time through nucleation
and τ saturates to a finite value. In this case, the metastable state can be considered
as a thermodynamic state only if its lifetime is large enough (Section 1.5.3).

Metastable states will play a central role in the analysis of the glass state. These
ideas will be generalised to the glass starting from Chapter 4.

1.6.2 Further Reading

We provide here a list of references that can be consulted to further explore the
subjects discussed in this chapter. These references have been chosen using the
36 Infinite-Dimensional Models in Statistical Physics

following criteria. (1) Whenever available, we privileged books and review arti-
cles, because they are usually more pedagogical. (2) Among research articles, we
selected those that are, in our opinion, the most accessible to the reader. We also
privileged more recent articles, because they usually contain more references to the
previous literature. (3) Whenever possible, we selected articles presenting results
that are directly related to those discussed in the chapter. We would like to stress
that this bibliography is not meant to be exhaustive, and, in particular, it does not
represent the historical development of the field. Hence, the fact that we cite an
article should not be interpreted as an attribution of credit to its authors.
In the rest of this book, we focus only on the infinite-dimensional limit. Most of
the concepts that have been introduced in this chapter are, therefore, only needed
to situate this limit in the general setting of statistical physics. Yet further reading
on how to go beyond the limit d → ∞ and include spatial fluctuations beyond
mean field is certainly useful. Concerning critical fluctuations around second-order
phase transitions, the most important method is the renormalisation group, which
is covered in many introductory textbooks and reviews, such as

• Amit and Martin-Mayor, Field theory, the renormalization group, and critical
phenomena [11]
• Brézin, Introduction to statistical field theory [69]
• Parisi, Statistical field theory [282]
• Delamotte, An introduction to the nonperturbative renormalization group [124]
The last reference [124] focuses specifically on the non-perturbative methods that
are needed in the context of disordered systems and glasses, where perturbative
renormalisation group usually fails.
Good introductions to the theory of phase coexistence in equilibrium, metasta-
bility and nucleation can be found in

• Binder, Theory of first-order phase transitions [52]

• Gallavotti, Statistical mechanics: A short treatise [158]
• Martinelli, Lectures on Glauber dynamics for discrete spin models [246]
A detailed discussion of the dynamics of infinite-dimensional spin models such as
the Curie–Weiss model discussed in Section 1.5.2, with a focus on critical dynamics
and metastability, can be found in

• Ruijgrok and Tjon, Critical slowing down and nonlinear response in an exactly
solvable stochastic model [311]
• Mora, Walczak and Zamponi, Transition path sampling algorithm for discrete
many-body systems [265]
 2
Atomic Liquids in Infinite Dimensions
Thermodynamics

The aim of this chapter is to review the key elements of the thermodynamic the-
ory of liquids for understanding the subsequent chapters. The reader is assumed
to be familiar with the classic book by Hansen and MacDonald [175], but the
initial discussion nonetheless offers a short summary of basic notions borrowed
from that book in order to define the main observables and notations (that are
sometimes slightly different from [175]), to introduce the virial expansion and to
show its equivalence to a large-dimensional expansion. The formal analogy with
the discussion of Chapter 1 for magnetic systems will be highlighted. Then, the
large-dimensional limit will be explicitly constructed for typical liquid potentials,
and the results for the gas and liquid phases will be discussed.

2.1 Thermodynamics of Atomic Systems

Many gases, liquids and solids can be modelled as collections of classical1 point
particles, representing the atoms, interacting via spherically symmetric potentials.
Simple models that contain identical atoms, such as Argon atoms [175, chapter 1],
are said to be ‘monodisperse’. Most amorphous solids, however, are not composed
of identical atoms. Monodisperse systems typically crystallise very easily and can-
not be kept in an amorphous solid state for a macroscopic time. Amorphous solids
are thus usually composed either of mixtures of different atomic species (binary or
polydisperse mixtures), of small molecules or of even more complex molecules
such as long chains – i.e., polymers. Non-spherical interactions are sometimes
present, especially in granular matter.
One could then be worried that in order to study glasses, it is necessary to avoid
simple monodisperse and spherical systems and to study instead more complicated

1 Quantum effects can certainly be important in some regimes, especially at very low temperatures, but they
will not be discussed in this book.

37
38 Atomic Liquids in Infinite Dimensions: Thermodynamics

systems. It turns out, however, that this complication is only necessary in d = 2 and
d = 3. These, of course, are the relevant physical dimensions, but as soon as d > 3,
monodisperse spherical systems remain disordered for extremely long times, and it
is very hard to observe them crystallise spontaneously [330, 350] (we come back
to this point in Chapter 8). Because this book is mostly concerned with the large-
dimensional limit, it is thus sufficient for us to focus on monodisperse spherical
systems. In infinite dimensions, polydispersity is not essential for glass formation,
and a small enough deviation from monodispersity [184] or from sphericity [367]
does not affect the qualitative shape of the phase diagram. The rest of this book
is thus focused on monodisperse spherical systems, although the discussion could
be quite easily generalised to binary, polydisperse or non-spherical systems [49,
102, 184, 185, 367]. As discussed in the Preface, establishing to what extent an
infinite-dimensional monodisperse spherical system is a good model of real glassy
materials goes beyond the scope of this book.
Note that for monodisperse spherical systems, even if crystallisation is kinet-
ically heavily suppressed in large dimensions (in other words, if the system is
prepared in the liquid phase it will remain there forever), it is still highly probable
that the true equilibrium thermodynamic state at large densities might be a crystal
(see Chapter 8). If this is true, then the liquid phase would be but metastable with
respect to the crystal. However, this is not a problem because as we have seen in
Chapter 1, in infinite dimensions, metastable states are minima of a suitable free
energy function. In this chapter, we define the free energy as a functional of the
density profile. In this framework, the liquid state is a minimum of the free energy
with uniform density while the crystal is a minimum with a periodically modulated
density. From a theoretical standpoint, the two situations can then be distinguished
without any ambiguity. In Chapter 4, we thus describe how this approach can be
generalised to the glass phase.

2.1.1 Definitions
Consider a system of N identical atoms, enclosed in a compact region of Rd with
volume V . Atoms are modelled as point particles. Their positions are specified by
a set of d-dimensional vectors X = {xi }i=1,...,N , each xi having components xiμ
for μ = 1, . . . ,d. In most atomic systems, the interaction energy is dominated by
pairwise interactions, and the total interaction energy is

V (X) = v(|xi − xj |). (2.1)
i<j

Interactions involving more than two particles could also be considered, but they
are not essential in most cases, and for simplicity, they will be neglected [175,
 2.1 Thermodynamics of Atomic Systems 39

chapter 1]. Because the particles are pointlike, the interaction between a given pair
ij of them must be spherically symmetric – i.e., it can only depend on their distance
rij = |xi − xj |. Examples of typical interaction potentials v(r) are given in [175,
chapter 1] and will be discussed in Section 2.3.2.
The main thermodynamic quantities are defined following [175, chapter 2], to
which the reader is referred for a detailed discussion. The configurational integral
at temperature2 T = 1/β,
 N
ZN = dX e−βV (X), dX = dxi , (2.2)
i=1

plays a central role in the thermodynamics of particle systems. In fact, for Hamil-
tonian systems, the total energy, kinetic plus potential, is the Hamiltonian function
 p2
H (P ,X) = i 2mi + V (X), and the canonical partition function is

1 ZN
QN = dN dP dX e−βH (P ,X) = dN . (2.3)
h N!  N!
The factor N ! in the denominator takes into account the indistinguishability of the
particles, while the constant h (which must be present for dimensional reasons,
because QN must be adimensional) can be identified with the Planck constant.
Both factors can be deduced from a purely classical treatment [158], or they can be
obtained by taking the semiclassical limit of a quantummechanical treatment [223].
With h̄ = h/(2π), the ‘De Broglie wavelength’  = 2πβ h̄2 /m appears after the
Gaussian integration over the momenta has been performed explicitly. The average
of an observable O[X] in the canonical ensemble is
 
  1 dP dX −βH (P ,X) 1
O[X] = e O[X] = dX e−βV (X) O[X], (2.4)
QN hdN N ! ZN
and the Helmholtz free energy of a system with fixed N,V ,T is
 
ZN
F (N,V ,T ) = −T log QN = dT N log  − T log . (2.5)
N!
The thermodynamic limit is taken by sending N,V → ∞ with constant number
density ρ = N/V and temperature T . For this limit to exist, two conditions on the
pair potential v(r) are sufficient [158, 310].
1. ‘Temperedness’: for r → ∞, |v(r)| ≤ Ar −d−δ with constants A ≥ 0 and δ > 0.
This condition ensures that the interaction energy decays sufficiently rapidly
at large distance. In particular, finite range potentials, such that v(r) = 0 for
r ≥ r0 , trivially satisfy this condition.

2 Recall that k = 1 throughout this book.

B
40 Atomic Liquids in Infinite Dimensions: Thermodynamics

2. ‘Stability’: there exists an N -independent constant B ≥ 0, such that V (X) ≥

−BN for all particle configurations X ∈ RdN and for all N . This condition
ensures that the collapse of infinitely many particles in a bounded region of space
is impossible. There are many possible ways of imposing conditions on the pair
potential to ensure stability; some of them are discussed in [310, chapter 3]. In
particular, potentials with a hard core satisfy this condition.
Potentials that do not satisfy the temperedness condition are deemed to be ‘long
ranged’, while potentials that do not satisfy the stability condition are said to be
‘catastrophic’. In both cases, anomalies can appear in the thermodynamic limit.
Such potentials – e.g., the gravitational potential – are not considered in this book.
In the thermodynamic limit, the equilibrium Helmholtz free energy is a function
of ρ and T , which are the only state variables of the system. Moreover, because
F is extensive (proportional to N ), it is convenient to consider the free energy per
particle, which remains finite:
F (N,V ,T )
f(ρ,T ) = lim
N,V →∞, ρ=N/V N
  (2.6)
1 ZN
= dT log  − T lim log .
N,V →∞, ρ=N/V N N!
From the Helmholtz free energy per particle, one can derive most of the inter-
esting thermodynamic quantities – such as the average energy, the entropy, the
pressure and the specific heat. One can consider different ensembles – for example,
with pressure,
∂F (N,V ,T ) ∂f(ρ,T )
P (ρ,T ) = − lim = ρ2 , (2.7)
N,V →∞, ρ=N/V ∂V ∂ρ
fixed instead of volume V (or chemical potential μ fixed instead of particle
number N) – but because all ensembles are equivalent in the thermodynamic
limit [158, 175], the (ρ,T ) ensemble will most often be considered in this book.
The main objective of the following discussion will thus be to compute f(ρ,T ).
For an ideal gas of non-interacting particles, v(r) = 0, one has ZN = V N and
 N
1 V
f (ρ,T ) = dT log  − T
id
lim log
N,V →∞, ρ=N/V N N! (2.8)
= dT log  − T (1 − log ρ).
It is customary to separate the Helmholtz free energy between the ideal gas and the
excess contributions:
f(ρ,T ) = f id (ρ,T ) + f ex (ρ,T ),
 
1 ZN (2.9)
f (ρ,T ) = −T
ex
lim log .
N,V →∞, ρ=N/V N VN
 2.1 Thermodynamics of Atomic Systems 41

The term f ex (ρ,T ) contains all the non-trivial dependence on the interactions and
depends only on the normalised configurational integral ZN /V N . Note that the term
dT log  in f id (ρ,T ) depends only on constants and on temperature T . Because its
contribution to thermodynamic quantities is trivial, it will often be omitted in the
following, although it can be reinserted at any time.

2.1.2 Local Density and Thermodynamic Observables

The discussion of Section 1.1.2 for magnetic systems can be adapted with only
minor modifications to particle systems. The role of the local magnetisation is then
played by the local density, defined as


N
 
ρ[x;X] = δ(xi − x), ρ(x) = ρ[x;X] . (2.10)
i=1

Here the function ρ[x;X] is the local density at point x for a given configuration X,
which is a sum of delta functions because classical particles are at well-defined
positions. The function ρ(x) is the thermodynamic average over X in the canonical
ensemble defined by Eq. (2.4) or, equivalently, in any other ensemble. This quantity
is usually a smooth function
 because particles move under the action of thermal
fluctuations. Note that dxρ(x) = N , as follows from Eq. (2.10).
Most of the interesting observables of particle systems can be written as linear
combinations of products of local densities. In general, taking into account that
particles are identical, a thermodynamic observable has the form
  
O[X] = O1 (xi ) + O2 (xi ,xj ) + O3 (xi ,xj ,xk ) + · · · , (2.11)
i i=j i=j =k

where O1 is a one-body observable, O2 a two-body observable and so on. The

potential energy in Eq. (2.1) is a typical example of a two-body observable, and
so is pressure [175]. Eq. (2.11) can equivalently be written in terms of the local
density,

O[X] = dxO1 (x)ρ[x;X]
 (2.12)
+ dxdyO2 (x,y){ρ[x;X]ρ[y;X] − ρ[x;X]δ(x − y)} + · · · ,

which is the analog of Eq. (1.11) for particle systems.

The thermodynamic average then has the form
 
 
O[X] = dxO1 (x)ρ(x) + dxdyO2 (x,y)ρ (2) (x,y) + · · · , (2.13)
42 Atomic Liquids in Infinite Dimensions: Thermodynamics

which can be decomposed as a sum of integrals of density-density correlations,

where the one-body density is ρ(x) and the two-body density is
 
ρ (2) (x,y) = ρ[x;X]ρ[y;X] − ρ(x)δ(x − y). (2.14)

Three-body and higher-order correlations are defined similarly [175, chapter 2].
The quantity ρ (2) (x,y)dxdy yields the probability of finding two distinct particles
with coordinates in the volume element dxdy, irrespective of the positions of the
remaining particles and irrespective of momenta. From it, a normalised pair corre-
lation function can be defined,

ρ (2) (x,y)
g(x,y) = . (2.15)
ρ(x)ρ(y)

This quantity is central in the theory of atomic systems. For an ideal gas, cor-
relations between particles are absent, and one has ρ (2) (x,y) = ρ(x)ρ(y) and
g(x,y) = 1. Deviations of g(x,y) from unity thus characterise the local structuring
of the system due to interactions. In the gas and liquid phases, due to translational
and rotational invariance, the local density ρ(x) = ρ is independent of x, and the
density-density correlation is only a function of distance, r = |x − y|. In this case
g(r) = ρ (2) (r)/ρ 2 is known as the ‘radial distribution function’.

2.1.3 Free Energy as a Functional of the Local Density

As in magnetic systems, the thermodynamic phases of the system can be identified
with the minima of the free energy F [ρ(x)], expressed as a functional of the local
density ρ(x). However, computing this quantity directly in the canonical ensem-
ble, where the total density is fixed, is cumbersome. Following the treatment of
Chapter 1 for magnetic systems, it is convenient to introduce a field conjugated to
the local density – i.e., the chemical potential – and compute the free energy as a
functional of the chemical potential in the grand canonical ensemble. One can then
perform a Legendre transformation to recover F [ρ(x)].

Definition of the Free Energy Functional

In the grand canonical ensemble [175, chapter 2], the particle number can fluctuate,
but its average is fixed by a chemical potential μ. To fix the local density, one can
introduce an external local potential φ(x). Defining

eβμ−βφ(x)
z(x) = , (2.16)
d
 2.1 Thermodynamics of Atomic Systems 43

the grand canonical partition function is

∞  
eβμN −βH (P ,X)−β N
[z] = dN N !
dP dX e i=1 φ(xi )

N=0
h
∞  (2.17)
 1 N
−βV (X)
= dX e z(xi ).
N=0
N! i=1

The free energy of the grand canonical ensemble is the ‘grand potential’
∞
  
1
[z] = −T log [z] = −T log dX e−βV (X)+ dxρ[x;X] log z(x), (2.18)
N=0
N!
 
where i z(xi ) = exp[ dxρ[x;X] log z(x)] is written in terms of the local den-
sity. From Eq. (2.18), it follows that the average local density and its correla-
tions can be obtained by differentiating the grand potential with respect to log z(x)
= βμ−βφ(x)−d log . This treatment follows exactly how spin–spin correlations
are obtained as derivatives3 of the free energy with respect to the magnetic field
(Section 1.1.2):
∂  
−β = ρ[x;X] = ρ(x),
∂ log z(x)
2
∂   
−β = ρ[x;X]ρ[y;X] − ρ(x)ρ(y) (2.19)
∂ log z(x)∂ log z(y)
= ρ (2) (x,y) + ρ(x)δ(x − y) − ρ(x)ρ(y) = ρS(x,y).
The second derivative of −β[z] is a positive operator; hence, [z] is a concave
function of z(x).
Then, as it was done in Section 1.1.3 for magnetic systems, it is possible to define
the free energy as a functional of ρ(x) by a Legendre transform,
  
F [ρ] = max [z] + T d
dxρ(x) log[ z(x)] , (2.20)
z(x)

from which it follows that

∂F ∂ 2F
−β = − log[d z(x)], −β = −(ρS)−1 (x,y), (2.21)
∂ρ(x) ∂ρ(x)∂ρ(y)
where (ρS)−1 (x,y) is the operator inverse of ρS(x,y).

3 Here and in the following, we use the same notation for regular and functional derivatives in order to
emphasise this similarity.
44 Atomic Liquids in Infinite Dimensions: Thermodynamics

Uniform Chemical Potential

The external potential φ(x) has been introduced with the only purpose of
constructing the free energy functional of ρ(x). In this book, we are mostly
interested in the homogeneous case where φ(x) = 0 and z(x) = z. The equation
for ρ(x) is then
∂F
= T log(d z) = μ. (2.22)
∂ρ(x)
This equation must be solved to determine the average local density profile that
corresponds to a given μ. Let us denote the solution ρ ∗ (x), and note that the global
density is ρ = N  /V = V −1 dxρ ∗ (x). In the gas and liquid phases, ρ ∗ (x) = ρ
is homogeneous, but in a solid phase, ρ ∗ (x) depends on x. Note that in a solid
phase, translational and rotational symmetries are broken.4 There are then many
pure states, one specified by ρ ∗ (x) and a continuous set of other equivalent states
corresponding to all the independent translations or rotations of ρ ∗ (x). In this case,
phase coexistence of different solids5 is possible, leading to the same phenomenol-
ogy as discussed in Section 1.4.
According to Eq. (2.20), the thermodynamic free energy F [ρ ∗ ] is6

∗ ∗
F [ρ ] = (z[ρ ]) + T dxρ ∗ (x) log(d z[ρ ∗ ]) = (μ) + Nμ, (2.23)

where we have used the fact that z[ρ ∗ ] does not depend on x because ρ ∗ (x) is
the solution of Eq. (2.22). Then, F [ρ ∗ ] is identified with (μ) + Nμ, which is
the Helmholtz free energy. Note that Eq. (2.22) for ρ(x), in which the chemical
potential is such that the average density is ρ, can also be obtained by adding a
Lagrange multiplier μ to enforce the value of the density – i.e., determine ρ(x) by
minimising
 
Fμ [ρ] = F [ρ] − μ dxρ(x) − N . (2.24)

We conclude that F [ρ] gives the Helmholtz free energy associated to the local
density ρ(x) and that, given an expression for the functional
 F [ρ], one has to find
−1
ρ(x) by minimising it under the constraint that V dxρ(x) = ρ is fixed. The
minima define the equilibrium phases of the system.

4 In crystals, these are only partially broken as a set of discrete symmetries remain unbroken. In amorphous
solid phases, both symmetries are fully broken.
5 For crystals, a state characterised by phase coexistence of different solids is called ‘polycrystal’, and the
interfaces between coexisting regions are called ‘grain boundaries’.
6 Independently for each pure state if a symmetry is broken, but in this case, all pure states have the same
free energy.
 2.1 Thermodynamics of Atomic Systems 45

In infinite dimensions, as we have seen in Chapter 1, the local minima define

metastable phases. The gas or liquid phase, which are translationally and rotation-
ally invariant, are minima of F [ρ] with uniform density profile – i.e., ρ(x) = ρ
– and, thus, ρ (2) (x,y) = ρ 2 g(|x − y|). A crystal phase, instead, corresponds to a
periodically modulated ρ(x). By imposing translational invariance, we thus select
the liquid phase and rule out any other phase (crystals or separated phases – e.g.,
gas–liquid coexistence).
We now focus on this choice. Introducing h(r) = g(r) − 1, and using Eq. (2.19),
the second derivative of [z] is proportional to

S(r) = δ(r) + ρh(r), S(q) = dr eiq·r S(r) = 1 + ρh(q), (2.25)

where the ‘static structure factor’ S(q) is the Fourier transform7 of S(r). This
quantity is also very important in liquid theory. It contains a lot of information
about liquid structure and is routinely measured by neutron and radiation scattering
[175, chapter 3]. From Eq. (2.21), the second derivative of F [ρ] is proportional to
S −1 (x − y), the operator inverse of S(x − y). In the translationally invariant case, it
can be obtained in Fourier space as the numerical inverse of S(q):

1 1
(S −1 )(q) = dr eiq·r S −1 (r) = = = 1 − ρc(q), (2.26)
S(q) 1 + ρh(q)
which implies
S −1 (r) = δ(r) − ρc(r). (2.27)
The function c(q) and its inverse Fourier transform c(r) are ‘direct correlation func-
tions’ [175, chapter 3]. In rotationally invariant states, all these functions depend
only on the modulus of their argument.
From the functions g(r) and S(q), one can deduce most of the thermodynamic
observables. For example, using Eq. (2.13), and for pairwise additive interactions

V (X) = 12 i=j v(|xi − xj |), the average potential energy is
 
  ρ2 ρN
V (X) = dxdyg(|x − y|)v(|x − y|) = drg(r)v(r)
2 2
 (2.28)
ρNd ∞
= dr r d−1
g(r)v(r),
2 0

where d = 2π d/2 / (d/2) is the solid angle in d dimensions, with (x) being the
Euler gamma function. Similar expressions can be obtained for other observables

7 In order to lighten the notation, we use the same symbol for a function and its Fourier transform when there is
no ambiguity.
46 Atomic Liquids in Infinite Dimensions: Thermodynamics

[175]. Finally, it is worth noting that, in translationally invariant phases, the radial
distribution function can be written as a derivative of the free energy with respect
to the pair potential, both in the canonical and grand canonical ensembles [175,
section 3.4]:

1 ∂F 1 ∂ ρ
= = g(r). (2.29)
N ∂v(r) N ∂v(r) 2
  
Writing V (X) = 12 i=j v(xi − xj ) = 1
2
drv(r) i=j δ(r − xi + xj ), from
Eq. (2.29), one then also obtains

1 
g(r) = δ(r − xi + xj ) . (2.30)
ρN i=j

2.2 The Virial Expansion

For particle systems, as for spin systems (Chapter 1), it is impossible to obtain a
closed analytical expression of the free energy functional F [ρ]. The only way to
make concrete calculations is to set up perturbative expansions for F [ρ], the most
famous being the virial expansion. It expands around the ideal gas limit, and as
such is only formally reliable at high temperatures and low densities. In the high-
temperature regime, one is above the critical point, and, hence, there is no distinc-
tion between gas and liquid. This is the (supercritical) ‘fluid’ phase. Interestingly,
appropriate resummations of the virial expansion can also be used to describe the
dense liquid phase [175].

2.2.1 The Ideal Gas

In the ideal gas case, V (X) = 0, the grand potential can be easily computed:

∞  N 
1
id [z] = dX z(xi ) = e dx z(x),
N!
N=0
 i=1 (2.31)
 [z] = −T
id
dx z(x).

It follows that the density profile is

∂id
ρ(x) = −β = z(x). (2.32)
∂ log z(x)
 2.2 The Virial Expansion 47

Plugging these results in Eq. (2.20), one obtains


F [ρ] =  [z] + T
id id
dxρ(x) log[d z(x)]
 (2.33)
= dT N log  − T dxρ(x)[1 − log ρ(x)],

which is minimised by a uniform profile ρ(x) = ρ. The Helmholtz free energy of

the ideal gas, Eq. (2.8), is thus recovered.

2.2.2 Second Virial Diagram

The virial expansion is based on the following idea. Considering pairwise interac-
tions with pair potential v(r) as in Eq. (2.1), and introducing the Mayer function
f (r) = e−βv(r) − 1, (2.34)
the potential energy term in the partition function can be expanded in powers
of f (r):
 
e−βV (X) = e−βv(|xi −xj |) = [1 + f (|xi − xj |)]
i<j i<j
 (2.35)
=1+ f (|xi − xj |) + · · ·
i<j

The motivation for this expansion is that, contrarily to the Gibbs–Boltzmann factor,
the Mayer function decays to zero at large distances, r → ∞. However, such an
expansion only makes sense if f (r) can be considered ‘small’ in the full range of
r – which is correct, for instance, if β is small, or if the region where f (r) is non-
zero is small. The virial expansion can thus be considered as a partially resummed
high-temperature expansion. Keeping only the first correction, one has8 :
⎡ ⎤
∞   N 
1
[z] = dX z(xi ) ⎣1 + f (|xi − xj |)⎦
N=0
N ! i=1 i<j

 ∞
 1 N(N − 1)  N−2
=e dx z(x)
+ dx z(x)
N=2
N! 2 (2.36)

× dxdy z(x)z(y)f (|x − y|)

  
1
=e dxz(x)
1+ dxdyz(x)z(y)f (|x − y|) .
2

8 Note that terms containing the Mayer function can be present only if N ≥ 2.
48 Atomic Liquids in Infinite Dimensions: Thermodynamics

Therefore, at the first order in f (r),

 
1
−β[z] = dxz(x) + dxdyz(x)z(y)f (|x − y|). (2.37)
2
This correction to [z] also impacts the ideal gas result ρ(x) = z(x), which
becomes ρ(x) = z(x) + O(z2 ). The O(z2 ) correction to ρ(x), however, does not
enter in F [ρ] at the lowest order, for the same reason as in Section 1.2.1. We can
then simply substitute z(x) = ρ(x) in the correction term of Eq. (2.37) to obtain
the second virial correction to F [ρ]:

−βF [ρ] = −dN log  + dxρ(x)[1 − log ρ(x)]
 (2.38)
1
+ dxdyρ(x)ρ(y)f (|x − y|).
2

2.2.3 Full Virial Expansion

A convenient way to work out the full virial expansion is to represent its different
terms as diagrams. The general formalism behind this approach is discussed in [175,
chapter 3], and detailed derivations can be found in [118, 266]. Here, we summarise
some of the key ideas. The second virial correction can be represented by a diagram
of the form

1
= dxdyρ(x)ρ(y)f (|x − y|). (2.39)
2
In this diagram, each black dot (or vertex) represents a factor of ρ(x) with a corre-
sponding integration over x, and each black line represents a factor f (|x − y|). The
diagram is also multiplied by a factor of 1/S, where S is the ‘symmetry number’ of
that diagram, which is calculated as follows. One first chooses an arbitrary labelling
of the vertices; in the example of Eq. (2.39), one might call x the left vertex and y
the right vertex. Then, S is the number of permutations of the labels that give rise
to exactly the same labelling. In Eq. (2.39), exchanging x and y leads to the exact
same result. Hence, S = 2, which explains the factor 1/2 in front of the integral.
Another example of diagram is

1
= dxdydzρ(x)ρ(y)ρ(z)f (|x − y|)f (|x − z|)f (|y − z|), (2.40)
6
where S = 6 because once an arbitrary labelling is chosen (e.g., x for the top
vertex, y for the bottom right and z for the bottom left), there are three rotations of
the labels and, for each rotation, one exchange of two labels that all give the same
result.
 2.2 The Virial Expansion 49

The virial series for the Helmholtz free energy then has the form

−βF [ρ] = −βF id [ρ] + + + + + + ···

(2.41)
The diagrams that appear in Eq. (2.41) are all the possible ‘one-particle irreducible
diagrams’, which are those that cannot be disconnected in two or more parts by
removing one vertex (i.e., one particle). Note that although the series in Eq. (2.41)
has been derived through an expansion in powers of the Mayer function (the number
of lines in each diagram), the diagrams in Eq. (2.41) are organised according to
the number of vertices they contain, which is an easier and more natural scheme.
Because the Mayer function is small at high temperatures, classifying the dia-
grams by the number of lines they contain essentially leads to a (resummed) high-
temperature expansion, while classifying the diagrams according to the number of
vertices amounts to an expansion in powers of ρ(x) – i.e., a low-density expansion.
In any case, one should keep in mind that a rigorous proof of convergence of the
virial expansion requires both high temperatures and low densities, in addition to
some regularity conditions on the interaction potential [158].

2.2.4 Virial Expansion in the Fluid Phase

In the fluid, gas or liquid phases, the homogeneous density ρ(x) = ρ is a solution
of Eq. (2.22). From Eq. (2.41), the Helmholtz free energy per particle can be written
as a power series of ρ as
∞
−βF [ρ(x) = ρ] Bn n−1
−βf(ρ,T ) = = −βf (ρ,T ) −
id
ρ , (2.42)
N n=2
n−1
∞

2 ∂βf(ρ,T )
βP (ρ,T ) = ρ =ρ+ Bn ρ n = ρ + B2 ρ 2 + B3 ρ 3 + · · · ,
∂ρ n=2

where the pressure P (ρ,T ) is obtained from Eq. (2.7), and Bn are the ‘virial
coefficients’.
Each virial coefficient is a sum of contributions coming from all one-particle
irreducible diagrams Gn of n vertices. The algorithm to obtain the contribution of
each diagram to Bn is best understood by examples. Consider the second virial
coefficient, to which a single diagram contributes, Eq. (2.39) with ρ(x) = ρ.
Comparing Eq. (2.42) with Eq. (2.41), one obtains
 
1 ρ2 1
B2 = − × =− dx1 dx2 f (|x1 − x2 |) = − drf (|r|).
ρN 2ρN 2
(2.43)
50 Atomic Liquids in Infinite Dimensions: Thermodynamics

In the last expression, translational invariance has been used to change variables
from {x1,x2 } to {r = x1 − x2,x2 }, in such a way that the integrand is independent
of x2 . The integration over x2 then gives a factor V that cancels the factors of ρ in
front of the virial coefficient. Similarly, the third virial coefficient is given by

2 1
B3 = − 2 × =− dr1 dr2 f (|r1 |)f (|r2 |)f (|r1 − r2 |), (2.44)
ρ N 3
where r1 = x1 − x3,r2 = x2 − x3 , and the integration over x3 gives a factor V by
translational invariance. The procedure can be iterated to obtain the general relation

Bn = −(n − 1) B(Gn ), (2.45)
Gn

where the sum is over all diagrams Gn having n vertices and a set of edges
E = ij  ∈ Gn . For a given diagram, the value of B(Gn ) is obtained by fixing
one arbitrary vertex in xn = rn = 0, and the other vertices correspond to the points
r1 . . . rn−1 . The diagram is multiplied by its symmetry factor 1/S(Gn ), with an
integrand given by a product of Mayer functions f (E) = f (|ri − rj |) assigned to
each edge,
 
1
B(Gn ) = dr1 . . . drn−1 f (E), with rn = 0. (2.46)
S(Gn )
E∈Gn

Explicit expressions for the first few virial coefficients are known for several
choices of interaction potential.

2.3 Liquids in Large Dimensions

In this section, we investigate the behaviour of the virial expansion in the limit
of large spatial dimensions, d → ∞. This problem was first considered by
Frisch, Klein, Percus, Rivier and Wyler [154, 155, 210, 362]. As we will see, in
large d, the virial expansion remains valid also in the dense fluid regime. For this
reason, in the following, we generically refer to the homogeneous phase as the
‘liquid’ phase.

2.3.1 Virial Expansion for Hard Spheres

One of the most studied pair interactions is the hard-sphere potential, which models
a collection of identical balls that cannot overlap and otherwise do not interact. For
spheres of diameter , the potential is

0 if r > ,
v(r) = (2.47)
∞ if r ≤ ,
 2.3 Liquids in Large Dimensions 51

which satisfies the temperedness and stability conditions discussed in Section 2.1.1,
and, thus, the existence of the thermodynamic limit is guaranteed. The correspond-
ing Mayer function is f (r) = −θ(−r), where θ(x) is the Heaviside step function.
Note that because f (r) does not depend on temperature, the temperature scaling of
free energy and pressure is trivial: βfex and βP are independent of T and only
depend on density ρ. If temperature T is used as the unit of energy (which amounts
to setting T = 1), all thermodynamic quantities are solely functions of ρ, which is
thus the unique state variable [175]. The volume and surface of a ball of unit radius
in d dimensions are, respectively,
π d/2 2π d/2
Vd = , d = dVd = . (2.48)
(1 + d/2) (d/2)
Note that d is also the solid angle in d dimensions, previously introduced in
Eq. (2.28). The volume and surface of a sphere of radius r are then Vd (r) = Vd r d
and d (r) = d r d−1 . For hard spheres, it is customary to define the ‘packing
fraction’,
ϕ = ρVd d /2d , (2.49)
which is the fraction of volume occupied by the spheres. For convenience, in this
section, we also define ϕ = 2d ϕ = ρVd d . Without loss of generality, one can also
choose  as unit of length, which amounts to setting  = 1; this is the choice we
make (unless otherwise specified) in this section to lighten the notation.
For hard spheres, many virial coefficients have been computed to high precision.
The second virial coefficient is
1
B2 = Vd d . (2.50)
2
Explicit expressions of B3 and B4 in terms of special functions for arbitrary dimen-
sion d are given in [196, 236]. Numerical values of Bn for n ≤ 11 and d ≤ 8
are given in [96, 319]. Even higher dimensions have been studied in [373]. In this
section, we analyse the asymptotic behaviour of the virial series for hard spheres in
the large d limit [154, 210, 362]. We will follow closely the treatment of [153]. The
discussion is a bit long and technical, but the essential steps are quite simple. The
reader can skip the proofs and jump straight to the conclusions, if desired.

Lower Bound on the Convergence Radius of the Virial Series

Denoting ϕ conv the convergence radius of the virial series, Lebowitz and Penrose
[226] proved that
(W (e/2) − 1)2
ϕ conv ≥ = 0.144767 · · · = ϕ lconv, (2.51)
W (e/2)
52 Atomic Liquids in Infinite Dimensions: Thermodynamics

where W (x) is the Lambert function defined by W (x)eW (x) = x. The virial series
is thus surely convergent for ϕ ≤ ϕ lconv . In this region, the hard-sphere system has
a unique pure state, the liquid phase, and there can be no phase transitions.

Ring Diagrams Dominate at Each Order

A special role in the series analysis is played by ‘ring diagrams’. At any order in
n ≥ 3, the ring diagram Rn is the diagram with the fewest edges; it has exactly n
edges arranged in a ring. For example, R3 is the triangle drawn in Eq. (2.40), R4
is the square in Eq. (2.41), R5 is a pentagon and so on. At any given finite order n,
adding an edge to a diagram makes the resulting diagram exponentially smaller in
d than the original one. Therefore, at any given order n, the ring diagram dominates
the virial coefficient, and Bn ∼ −(n − 1)B(Rn ).
Proof A proof of this statement can be found in [153]. Here we follow a similar
but slightly different route. By rotational and translational
! invariance, the Mayer
functions in Eq. (2.46) depend only on |ri − rj | = r2i + r2j − 2ri · rj – i.e.,
they depend only on the scalar products qij = ri · rj . One can thus change the
integration variables from ri to qij . The details of this change of variable are given
in Section 2.5, but recall nonetheless that rn = 0; hence, q̂ is a (n − 1) × (n − 1)
symmetric matrix. Following this scheme, Eq. (2.46) then becomes

Cn,d d−n 
B(Gn ) = dq̂ (det q̂) 2 f (E),
S(Gn ) Q (2.52)
E∈Gn

Cn,d = 21−n d d−1 · · · d−n+2 .


Here dq̂ = dqij , and the integration domain is
i≤j
√
Q = {q̂ : qii ≥ 0 and |qij | ≤ qii qjj }. (2.53)

In Eq. (2.52), the arguments of the Mayer functions are qii + qjj − 2qij for all
edges except those involving point n, which has rn = 0, and, therefore, edges
√
E = in have qii .
For d → ∞ and finite n, Eq. (2.52) can be evaluated by the saddle point method.
Neglecting constants and polynomial terms in d, the leading exponential order in d
is given by maximising det q̂ over all matrices q̂ ∈ Q that also satisfy the constraints
imposed by the requirement that the Mayer functions must be non-zero for all E:
d
|B(Gn )| ∼ (d )n−1 M(Gn ) 2 ,
(2.54)
M(Gn ) = max | det q̂|.
q̂∈Q: E∈Gn f (E)>0

The absolute value of B(Gn ) is taken because the Mayer functions are negative, and,
therefore, each diagram has a sign (−1)|Gn | where |Gn | is the number of edges of Gn .
 2.3 Liquids in Large Dimensions 53

Adding edges to a diagram amounts to introducing more constraints, which, from

Eq. (2.54), reduces the maximum value of | det q̂| and thus suppresses exponentially
the diagram. To show this explicitly, one can compute the leading order of the first
diagrams from Eq. (2.54). The second virial is known exactly. The triangle diagram
has a 2 × 2 matrix q̂ with det q̂ = q11 q22 − q12
2
, which has to be maximised under
the constraints given by the three Mayer functions and Eq. (2.53). This results in
0 ≤ q11 ≤ 1, 0 ≤ q22 ≤ 1, 0 ≤ q11 + q22 − 2q12 ≤ 1. The maximising configuration
is q11 = q22 = 1 and q12 = 1/2, which corresponds to a triangle of spheres all at
unit distance from each other. Then
   d/2
d/2 2 3
B ∼ (d ) max[q11 q22 − q12 ]
2 2
= (d ) . (2.55)
4

At fourth order, there are three diagrams. The maximisation is over a 3 × 3

matrix, and it can be done easily with a numerical manipulation software. The ring
diagram has
⎛ ⎞
   d/2 1 2/3 1/3
16
B ∼ (d )3 , q̂ = ⎝2/3 4/3 2/3⎠ . (2.56)
27
1/3 2/3 1

The other two diagrams have an additional constraint, and, as expected, each addi-
tional edge makes the diagram exponentially smaller:
⎛ ⎞
   d/2 1 1/2 1/4
9
B ∼ (d )3 , q̂ = ⎝1/2 1 1/2⎠ , (2.57)
16
1/4 1/2 1

and
⎛ ⎞
   d/2 1 1/2 1/2
3 1
B ∼ (d ) , q̂ = ⎝1/2 1 1/2⎠ . (2.58)
2
1/2 1/2 1

The same reasoning can be applied to any higher-order n with similar results.

Asymptotic Behaviour of the Ring Diagrams

The previous result suggests that in large d, the nth virial coefficient is dominated
by the ring diagram, except for B2 , which is given in Eq. (2.50). Therefore, the
nth-order contribution to the free energy in Eq. (2.42) is
 d
B n − 2 2ϕ = − ϕ2 for n = 2,
fn = −ρ n−1 = (2.59)
n−1 ρ n−1 B(Rn ) for n ≥ 3.
54 Atomic Liquids in Infinite Dimensions: Thermodynamics

At fixed n and d → ∞, at leading exponential order in d, one can show that

 d/2
nn−2
|fn | ∼ ϕ n−1 . (2.60)
(n − 1)n−1
For example, |f3 | ∼ ϕ 2 (3/4)d/2 , as in Eq. (2.55). Because nn−2 /(n − 1)n−1 < 1 for
all n ≥ 3, if ϕ = 2d ϕ is at most polynomial in d when d → ∞, then all the fn are
exponentially smaller than f2 . In this case, one obtains
ϕ
−βf(ρ,T ) = −βf id (ρ,T ) − ,
2
(2.61)
βP (ρ) ϕ
p(ρ) = = 1 + B2 ρ = 1 + ,
ρ 2
where the reduced pressure p(ρ) is derived from Eq. (2.7). If instead the scaled
packing fraction ϕ = eγ d (with γ > 0) grows exponentially with d, then the
condition |f2 |  |fn | is equivalent to
  1 &
(n − 1)n−1 2(n−1) 1 log n
γ
e < n−2
= 1 − e 2(n−1) . (2.62)
n n
For n → ∞, the right-hand side tends to 1, and Eq. (2.62) cannot be satisfied for
any γ > 0. In this case, the virial series is divergent; the largest orders in n then
dominate the series. Hence, the convergence radius ϕ conv of the virial series cannot
grow faster than polynomially in d. The numerical results of [373] further suggest
that the convergence radius ϕ conv → 1 for d → ∞.
Proof The proof of Eq. (2.60) can be obtained either from Eq. (2.54), or from a
Fourier representation. We follow this second route because it will be useful for the
next step. The Fourier transform of the Mayer function is

f (q) = dreiq·r f (r) = −Vd Jd (q),
    d2 (2.63)
d 2
Jd (q) =  +1 J d (q),
2 q 2

where Jν (x) is the Bessel function of order ν. An example of the behaviour of Jd (q)
is given in Figure 2.1. Recalling that S(Rn ) = 2n and using the Fourier transform,
fn can be written (for n ≥ 3) as

ρ n−1
fn = ρ n−1 B(Rn ) = dr1 . . . drn f (r1 − r2 )f (r2 − r3 ) . . . f (rn − r1 )
2nV
 
ρ n−1 dq nϕ
n−1
Vd d ∞
= f (q) = (−1)
n
dq q d−1 Jd (q)n . (2.64)
2n (2π)d 2n (2π)d 0
 2.3 Liquids in Large Dimensions 55

Jd (q)

Figure 2.1 An example of the normalised Fourier transform of the Mayer

function, the function Jd (q) = −f (q)/Vd defined in Eq. (2.63), for d = 5. The
values of q corresponding to the first zero of Jd (q) and its first negative maximum
both scale as q ∼ d/2.

For large d, the integrand in q in Eq. (2.64) has an absolute maximum before
Jd (q) changes sign for the first time, in the region q ∝ d/2. This maximum dom-
inates the integral for large d. One can use the asymptotic relation Jν (ν/ cosh α)
∼ exp[ν(tanh α − α)] (with α ≥ 0), choosing ν = d/2 and q = d/(2 cosh α)
∈ (0,d/2]. Keeping only the leading exponential terms in d, the integral becomes
 n  ∞
ϕ n−1 Vd2 d
fn ∼  +1 2 dn/2
dq q d(1−n/2) J d (q)n
(2π)d 2 0
2
(2.65)
n n
∼ ϕ n−1 (2/e)d(n/2−1) ed maxα [( 2 −1) log(cosh α)+ 2 (tanh α−α)] .

√ over α gives e = n − 1, from which tanh α = 1 − 2/n and

2α
The maximisation
cosh α = n/(2 n − 1). Plugging these results in Eq. (2.65) gives Eq. (2.60).

Resummation of the Ring Diagrams

The Fourier expression in Eq. (2.64) can then be used to compute the sum of all fn .
The resulting free energy is
∞

−βf(ρ,T ) = −βf id (ρ,T ) + f2 + fn
n=3

ϕ 1 Vd d ∞
= −βf (ρ,T ) − −
id
dq q d−1 L3 [ϕJd (q)], (2.66)
2 2ϕ (2π)d 0
∞
xn x2
L3 (x) = (−1)n+1 = log(1 + x) − x + .
n=3
n 2
56 Atomic Liquids in Infinite Dimensions: Thermodynamics

The series expansion of L3 (x) is convergent if |x| < 1; hence, |ϕJd (q)| < 1.
Because |Jd (q)| ≤ 1, as illustrated in Figure 2.1, this condition is satisfied if
|ϕ| < 1. In this region, the series of fn is convergent, supporting the asymptotic
analysis of the virial coefficients in Eq. (2.60), and the numerical results of [373].
In the region ϕ = eγ d with γ < (1 − log 2)/2 = 0.153426 . . ., Eq. (2.66) is well
defined and, in fact, coincides with Eq. (2.61). Hence, even if Eq. (2.60) indicates
that the contribution of the ring diagrams increases exponentially upon increasing
n for γ > 0, ring diagrams have alternating signs and cancel each other.
Proof The singularity that makes the series divergent corresponds to q = 0, where
Jd (q) = 1, and ϕ = −1, so ϕJd (q) = −1 and the term log(1 + x) in L3 (x) is
singular. This singularity happens in the unphysical region of negative densities.
By restricting the expression to positive densities, one can analytically continue
Eq. (2.66) to much larger densities. In fact, Eq. (2.66) remains well defined if
ϕ minq Jd (q) > −1, or, equivalently,

1 ' e ( d2 1−log 2
ϕ < ϕ0 = − ∼ = ed 2 . (2.67)
minq Jd (q) 2
The result in Eq. (2.67) is obtained by noting that, as illustrated in Figure 2.1, the
absolute minimum of Jd (q) is its first minimum. The equation for the stationary
points of Jd (q) is
d d d
0= Jd (q) ∝ J d (q) − J d (q) = −J d +1 (q). (2.68)
dq dq 2 2q 2 2

The first minimum corresponds to the first zero of J d +1 (q), which is asymptoti-
2
cally found at q0 ∼ d/2. Using J d (d/2) ∼ 1, one has asymptotically Jd (d/2)
2
∼ (4/d)d/2 (d/2 + 1) ∼ (2/e)d/2 .
A complete proof that Eq. (2.66) coincides with Eq. (2.61) for ϕ < ϕ 0 is given
in [153]; see also [284] for an alternative strategy. Here we give a simpler proof
restricted to the interval γ < 12 log(4/3) = 0.143831 · · · . First, one writes L3 (x)
= x 3 L3 (x) and notes that L3 (x) = L3 (x)/x 3 > 0 is a positive and decreasing
function of x for all x > −1, with L3 (0) = 1/3. The last term in the free energy in
Eq. (2.66) is then bounded by

1 Vd d ∞
dq q d−1 |L3 [ϕJd (q)]|
2ϕ (2π)d 0

ϕ 2 Vd d ∞
≤ L3 [ϕ min Jd (q)] dq q d−1 |Jd (q)|3 (2.69)
q 2 (2π)d 0
 d/2
2 3
∼ L3 [−ϕ/ϕ 0 ] ϕ  ϕ,
4
 2.3 Liquids in Large Dimensions 57

where the last line is obtained as Eq. (2.60) for n = 3, keeping only the leading
exponential order in d. For γ < 12 log(4/3) and d → ∞, one has ϕ/ϕ 0 → 0 and
the last inequality holds, which implies that Eq. (2.66) asymptotically coincides
with Eq. (2.61).
Summary
In short, we have identified different density regimes of the virial series in d → ∞,
in terms of ϕ = 2d ϕ.

• ϕ < ϕ conv = 0.144767 · · · : the series is proven to be convergent, and it converges

l

to Eq. (2.61). The excess free energy is then given by the second virial term alone.
• ϕ conv ≤ ϕ < 1: the series is conjectured to be convergent to Eq. (2.61). The
l

conjecture is supported by the asymptotic behaviour of the virial coefficients in

Eq. (2.60), by the resummation in Eq. (2.66) and by the numerical results of [373].
• 1 < ϕ < e with γ < (1 − log 2)/2 = 0.153426 · · · : the series is formally
γd

divergent, but it can be resummed9 to obtain Eq. (2.66), which at leading order
also coincides with Eq. (2.61).
We conclude that if either ϕ is not exponentially large in d, or ϕ = eγ d with
γ < (1−log 2)/2, then the virial series in large d is dominated by the ideal gas term
plus the second virial coefficient, corresponding to a direct two-particle interaction.
All the other terms can be discarded. When ϕ = eγ d with γ → (1 − log 2)/2, a
singularity appears in the liquid free energy [153, 284]. This singularity has been
identified with an instability of the liquid phase towards a modulated density profile,
also called ‘Kirkwood instability’ [153, 155]. Note that the Kirkwood instability
can also be detected by looking at the linear stability of the liquid against fluctua-
tions of the density field ρ(x) [242], which depends on the spectrum of the operator
S −1 (x − y) defined in Eq. (2.21). However, as we will see in Chapter 3, the liquid
phase becomes dynamically unstable towards a dynamically arrested glass phase
much before then, when ϕ ∝ d, and the glass phase even ceases to exist when
ϕ ∼ d log d. Hence, what happens formally to the liquid free energy for ϕ ∼ eγ d is
irrelevant for the rest of this book.

Liquid Theory Approximations and Numerical Results

In finite d, the virial series cannot be studied exactly, and even its radius of
convergence is not known. However, by resumming some classes of diagrams, one
can obtain closed integral equations for the radial distribution function g(r). These

9 Note that there are subtleties in the exchange of the limit d → ∞ and the resummation of the series, which
we do not discuss here in detail. For any finite d, in this region, crystallisation and the glass transition may
happen, which should mathematically correspond to true (i.e., non-resummable) singularities in the virial
series. These singularities, however, become weaker and weaker upon increasing d, as discussed in Chapter 1
for the ferromagnetic case, and can likely be neglected for d → ∞.
58 Atomic Liquids in Infinite Dimensions: Thermodynamics

provide approximations for g(r) and for thermodynamic quantities. The most
famous examples are the Percus-Yevick (PY) and hypernetted chain (HNC)
approximations [175, chapter 4] that, in d = 3, give reasonable estimates of
thermodynamic quantities with typical errors of ≈ 10%–15% in the dense liquid
phase. In the case of hard spheres, the PY approximation has been solved in all
odd [308] and even [2] dimensions, and in both cases, when d → ∞, one recovers
the scenario outlined in the preceding summary. For the HNC approximation,
no exact solution is available, but an asymptotic analysis again suggests that it
converges to the result given earlier for d → ∞ [284]. Numerical simulations
further show that the hard-sphere equation of state converges to Eq. (2.61) for all
densities at which the liquid can be equilibrated – i.e., all the way to the glass
transition [86]. Furthermore, the accuracy of the PY and HNC approximations
in describing the radial distribution function obtained by numerical simulations
improves quickly with dimension; see [2, 308] and references therein. All these
results support the conclusions of the preceding summary.

2.3.2 Other Interaction Potentials

The analysis of Section 2.3.1 can be extended to many other interaction potentials.
To illustrate this point, we start from the simplest example, a pure power-law soft-
sphere potential of the form:
 νd

v(r) = ε , ν > 1. (2.70)
r
Here, ε is the interaction energy scale,  sets the interaction range (an effective
particle diameter) and ν characterises the softness of the potential. The requirement
that ν > 1 ensures the temperedness of the potential (see Section 2.1.1). Under
this condition, the potential energy also remains extensive;
∞ because g(r) → 1 for
r → ∞, the integral in Eq. (2.28) is finite only if 0 dr r d−1 v(r) is convergent.
A common choice is ν = 4, which gives v(r) = ε(/r)12 in d = 3, corresponding
to the repulsive part of a Lennard-Jones potential [175].
We have seen that for hard spheres, temperature plays a trivial role, and density
is the only state parameter. The soft-sphere potential also has a special scale invari-
ance property that simplifies its phase diagram: temperature and density are not
independent state variables. The thermodynamic state of the system only depends
on the combination  = ρ/T 1/ν , which follows from the observation that the
energy and length scales are not independent in Eq. (2.70). Changing the energy
scale ε → ελ is indeed equivalent to changing the length scale  → λ1/(νd) .
To investigate the behaviour of the soft-sphere potential in large d, we start
by computing the second virial coefficient according to Eq. (2.43), assuming that
) = βε remains constant for d → ∞. Changing variables, to x = r/, one obtains
β
 2.3 Liquids in Large Dimensions 59
 ∞  ∞
d d d )
dr r d−1 [e−βv(r) − 1] = − dx x d−1 [e−β /x − 1].
νd
B2 = −
2 0 2 0
(2.71)

The integrand is strongly peaked around x = 1. In fact, for x < 1, the Mayer
function converges quickly to −1 while the factor x d−1 converges exponentially
)/x νd ], which also
fast to zero. For x > 1, the integrand is approximated by x d−1 [−β
goes exponentially fast to zero. Changing the variable to h, such that x = 1 + h/d,
and r = (1 + h/d), and taking the limit d → ∞, we get
 ∞
Vd d ) −νh Vd d
B2 = − dh eh [e−β e − 1] = I (β ),ν).
),ν) = B2HS I (β (2.72)
2 −∞ 2

The second virial coefficient is then given by the hard-sphere result, Eq. (2.50),
multiplied by a finite integral I (β ),ν) over the scaled variable h. Defining
ϕ = ρVd  = 2 ϕ as for hard spheres (but with  now being an effective diameter),
d d

the hard-sphere asymptotic result Eq. (2.61) becomes

ϕ
−βf(ρ,T ) = −βf id (ρ,T ) − ),ν),
I (β
2
(2.73)
βP (ρ,T ) ϕ
p(ρ,T ) = = 1 + I (β ),ν).
ρ 2

Note that by changing variables in such a way that β )1/ν e−h = e−z , one can show
)
that I (β,ν) = β) −1/ν
I (1,ν), which confirms that thermodynamic functions depend
)
only on ϕ β −1/ν
∝ ρ/T 1/ν = .
This analysis shows that if the limit d → ∞ is taken at constant β ) = βε, the
packing fraction ϕ = ρVd (/2)d defined in terms of the interaction range, , then
plays the same role as in hard spheres in controlling the behaviour of the virial
expansion. The Mayer function indeed goes exponentially to −1 for r <  and to
0 for r > , and only differs from the hard-sphere Mayer function in a tiny region
of the order of 1/d around r = , where r = (1 + h/d). The leading exponential
scalings in d of hard spheres therefore remain unchanged, and the virial coefficients
are simply multiplied by finite factors that depend on the potential. For example,
Eq. (2.52) remains valid, with the only difference that the Mayer functions do not
impose |ri − rj | <  strictly, but with a tolerance of order 1/d. The conclusion
is unchanged: adding an edge to a diagram imposes an additional constraint that
lowers the value of det q̂ and suppresses exponentially the diagram. The main result
of Section 2.3.1 for d → ∞ thus remains correct. One can discard all the virial
coefficients except B2 , as long as ϕ = edγ with γ < (1 − log 2)/2, and the liquid
free energy is given by the sum of the ideal gas term and the two-body contribution.
60 Atomic Liquids in Infinite Dimensions: Thermodynamics

Similar results hold for any potential (or, equivalently, any Mayer function) that
can be cast, for d → ∞, in the form

v(r) = v̄[d(r/ − 1)], f (r) = f [d(r/ − 1)], (2.74)

with the function v̄(h) going to infinity for h → −∞, and to zero faster than e−h
for h → ∞. Changing variables to h = d(r/ − 1), the second virial coefficient
becomes
  ∞
Vd d ∞
B2 = − dh e f (h) = B2 I (f ),
h HS
I (f ) = − dh eh f (h),
2 −∞ −∞
(2.75)
and all the other virial coefficients are subleading. The free energy is then
ϕ
−βf(ρ,T ) = −βf id (ρ,T ) − I (f ),
2 (2.76)
−βf id (ρ,T ) = −d log(/) + 1 − log(ρd ),
where the arguments of the logarithms in the ideal gas term have been made adi-
mensional by using the reference length scale . Using Eq. (2.29) on the virial series
truncated to second order, one obtains the radial distribution function of the liquid
in d → ∞,

g(r) = e−βv(r) . (2.77)

The liquid energy (per particle) and reduced pressure are then given by

∂(βf) ϕ ∞
e(ρ,T ) = = dh eh−β v̄(h) v̄(h),
∂β 2 −∞
(2.78)
βP (ρ) ϕ
p(ρ,T ) = = 1 + I (f ).
ρ 2
Examples of such sufficiently short-ranged potentials are [175, chapter 1] the
following:

• The pure power-law soft-sphere potential discussed earlier, for which

 νd

v(r) = ε , v̄(h) = εe−νh, (2.79)
r
with ν > 1.
• The Lennard-Jones potential, with
 νd  νd/2
 
v(r) = ε − , v̄(h) = ε(e−νh − e−νh/2 ), (2.80)
r r
 2.3 Liquids in Large Dimensions 61

with the restriction ν > 2 (the usual value in d = 3 is ν = 4). This potential is
usually a good model for simple atoms, especially inert atoms such as Helium,
Neon and Argon.
• The Yukawa potential, with
 
− εr e−λd(r/−1) if r > , −εe−λh if h > 0,
v(r) = v̄(h) = (2.81)
∞ if r ≤ , ∞ if h ≤ 0.
Note that the r in the denominator of v(r) can be approximated with  at the
leading order in d → ∞. Furthermore, λ > 1 is necessary for the convergence
of the second virial coefficient. The Yukawa potential is a good model for the
interaction of colloidal particles and of other screened charged systems.
• A square-well potential, with
⎧ ⎧
⎪
⎨0 if r ≥ (1 + h0 /d), ⎪
⎨0 if h ≥ h0,
v(r) = −ε if  < r < (1 + h0 /d), v̄(h) = −ε if 0 < h < h0,
⎪
⎩ ⎪
⎩
∞ if r ≤ , ∞ if h ≤ 0.
(2.82)
This potential finds many applications in the description of colloidal systems.
The ‘sticky sphere’ limit corresponds to a vanishing range and infinite strength
of the attractive potential, h0 → 0 and ε → ∞, with h0 eβε = eu in such a way
that the integral I (f ) that enters in the second virial coefficient in Eq. (2.75),
 h0
 
I (f ) = 1 − dh eh eβε − 1 = 1 − eu, (2.83)
0

remains finite.
• The soft repulsive sphere potential, with
εd α ' r (α ε α
v(r) = − 1 θ( − r), v̄(h) = h θ(−h). (2.84)
α  α
This potential describes soft spheres that interact repulsively only if they overlap
(r < ),  being their diameter. The choice α = 2 corresponds to the harmonic
soft-sphere potential, while α = 5/2 corresponds to Hertzian soft spheres [195].
Both models are commonly used to describe materials such as pastes, emulsions,
soft colloids, and soft macroscopic particles such as in granular materials [233].
• The Weeks–Chandler–Andersen (WCA) potential, given by
 4d  2d
 
v(r) = ε 1 + −2 θ( − r),
r r (2.85)
v̄(h) = ε 1 + e−4h − 2e−2h θ(−h).
62 Atomic Liquids in Infinite Dimensions: Thermodynamics

This is a Lennard–Jones potential modified to have its minimum in r = , trun-

cated to r ≤  and shifted in such a way that v() = 0. Close to h = 0, it reduces
to the harmonic sphere potential.
In the last two cases, taking the limit ε → ∞ recovers the hard-sphere potential.
Note that an important requirement for any potential in d → ∞ is that I (f ) > 0.
Otherwise, the pressure is a decreasing function of density, which signals a thermo-
dynamic instability. An example is the sticky sphere potential, for which, according
to Eq. (2.83), one must have u < 0. The thermodynamic instability could be
related to a gas–liquid phase separation. A detailed analysis for sticky spheres
can be found in [259], where it is shown that in the infinite dimensional limit
one recovers the standard Landau theory of the gas–liquid critical point. Another
possibility is that the instability is due to a violation of the stability condition
discussed in Section 2.1.1. Indeed, the thermodynamic limit does not exist (the free
energy density diverges when N → ∞ at fixed ρ) if the pair potential is always
attractive. For a potential with an attractive and a repulsive part, the situation is
more complex. Only if the repulsive part is sufficiently large is stability satisfied,
but there are apparently harmless potentials for which the thermodynamic limit
does not exist [310].

2.4 Wrap-Up
2.4.1 Summary
In this chapter, we have seen that

• In particle systems, the space-dependent density field ρ(x) plays the role
of magnetisation in magnetic systems. All the interesting thermodynamic
observables can be written as a sum of integrals involving correlations of ρ(x)
(Section 2.1.2).
• The local density is thus the appropriate order parameter to construct a free energy
functional F [ρ(x)]. The equilibrium phases of the system are minima of F [ρ(x)]
with the constraint dxρ(x) = N (Section 2.1.3). The fluid, gas and liquid phases
are minima corresponding to uniform density, ρ(x) = ρ. For attractive potentials,
there is a unique fluid phase at high temperatures (F [ρ] has a unique minimum),
while at low temperatures, there are two phases (gas and liquid) separated by a
first-order phase transition (in mean field, F [ρ] has two local minima; see the
general discussion of Chapter 1).
• The crystal phase is characterised by a periodically modulated ρ(x) and is usually
separated from the liquid by a first-order phase transition. Therefore, if an expres-
sion for F [ρ(x)] is available, the crystal can be eliminated from the theoretical
analysis by imposing uniformity of ρ.
 2.4 Wrap-Up 63

• An explicit expression for F [ρ(x)] can be obtained by means of the virial expan-
sion (Section 2.2). Within this approach, F [ρ(x)] is given by the ideal gas term
plus a sum of diagrams that capture interactions between two, three and more
particles; see Eq. (2.41).
• The virial expansion is both a high-temperature and a low-density expansion and
can also be interpreted as a large-dimensional expansion (Section 2.3.1). For a
hard-sphere liquid, in the limit d → ∞, it can be truncated to the two-particle
(second virial) term, resulting in Eq. (2.38), provided the scaled packing fraction
ϕ = 2d ϕ < ϕ 0 ∼ ed(1−log 2)/2 ; see Eq. (2.67).
• The same result can be generalised to any potential that can be written for
d → ∞ in the form v(r) = v̄[d(r/ − 1)], where  is an arbitrary length and the
function v̄(h) goes to zero faster than e−h for h → ∞ and diverges for h → −∞
(Section 2.3.2).

2.4.2 Further Reading

We provide here a list of references that can be consulted to further explore the
subjects discussed in this chapter, selected according to the criteria discussed in
Section 1.6.2.
The thermodynamic theory of liquids is very well developed beyond the d → ∞
limit. Historically, it was first derived by using the virial expansion as a low-density
expansion and resumming classes of diagrams to obtain closed integro-differential
equations for g(r) in fixed dimension d = 2,3. The limit d → ∞ was considered
much later. Famous examples are the hypernetted chain (HNC) and Percus-Yevick
(PY) approximations, but many other schemes have been developed. While the
reference textbook is Hansen and MacDonald, Theory of Simple Liquids [175],
many other excellent textbooks are available. The reader is encouraged to gain
familiarity with this literature, because similar methods can be applied to glasses in
d = 3 using the replica method (Section 4.5.2). The solution of the HNC and PY
approximations in large dimensions can be found in

• Parisi and Slanina, Toy model for the mean-field theory of hard-sphere liquids
[284]
• Rohrmann, Robles, de Haro, et al., Virial series for fluids of hard hyperspheres in
odd dimensions [308]
• Adda-Bedia, Katzav, and Vella, Solution of the Percus-Yevick equation for hard
hyperspheres in even dimensions [2]
These approximation schemes reproduce correctly the d → ∞ liquid properties.
Molecular dynamics numerical simulations of liquids also provide a lot of informa-
tion and allow one to precisely test the theory. An excellent textbook is Frenkel and
64 Atomic Liquids in Infinite Dimensions: Thermodynamics

Smit, Understanding molecular simulation: From algorithms to applications [152],

but also in this case many others are available.
A model for which the truncation of the virial series at its second term is exact
in all dimensions was investigated in

• Kraichnan, Stochastic models for many-body systems: I. Infinite systems in

thermal equilibrium [213]
• Mari and Kurchan, Dynamical transition of glasses: From exact to approximate
[242]
This model contains a tunable parameter which allows one to interpolate continu-
ously between the regular liquid and the truncated one, which has been proven very
useful to compare systematically the mean field theory with the finite d problem.
It was shown in Section 2.3.2 that the soft-sphere potential has a unique control
parameter,  = ρ/T 1/ν . Hence, points in the plane (T ,ρ) that correspond to the
same  are isomorphic. While this property is not exactly true for other potentials,
it has been shown in a variety of situations that it is approximately correct, with the
correct choice of an effective parameter ν. The d → ∞ solution has provided a
justification for this fact, which is an interesting application of the large d limit in
the liquid phase. These results are discussed in

• Dyre, Simple liquids’ quasiuniversality and the hard-sphere paradigm [134]

• Maimbourg and Kurchan, Approximate scale invariance in particle systems:
A large-dimensional justification [238]
• Costigliola, Schroder, and Dyre, Studies of the Lennard-Jones fluid in 2, 3, and 4
dimensions highlight the need for a liquid-state 1/d expansion [105]
Finally, a review of the statistical mechanics of long-range potentials, which are
not covered in this book, can be found in Campa, Dauxois and Ruffo, Statistical
mechanics and dynamics of solvable models with long-range interactions [74].

2.5 Appendix: Rotationally Invariant Integrals

We consider the integral of a rotationally invariant function F ({ra }), a = 1,
. . . , n − 1, of n − 1 vectors ra having each d dimensions. For simplicity, we restrict
here to the case d > n − 1, because we are mostly interested in the case where
d → ∞ and n is fixed.10 By rotational invariance, F ({ra }) = F (q̂), where q̂ is the
(n − 1) × (n − 1) matrix of scalar products qab = ra · rb . We can write

10 In the opposite case, d ≤ n − 1, the matrix Û Û T has zero modes. The derivation remains correct, provided
regular functions are replaced by distributions.
 2.5 Appendix: Rotationally Invariant Integrals 65
 
dr1 . . . drn−1 F ({ra }) = dq̂J (q̂)F (q̂),
 
1,n−1 (2.86)
J (q̂) = dr1 . . . drn−1 δ(qab − ra · rb ),
a≤b

where J (q̂) is the Jacobian of the change of variable from the original vectors to the

matrix q̂. Here, dq̂ = 1,n−1
a≤b dqab because the matrix q̂ is symmetric. To compute
the Jacobian, we can write
 
1,n−1   
J (q̂) = dr1 . . . drn−1 δ (qab − ra · rb ) = dÛ δ q̂ − Û Û T , (2.87)
a≤b

where the matrix U is a (n − 1) × d dimensional matrix with entries Uaμ = raμ


(i.e., each row of the matrix is a vector ra ), and the measure dÛ = aμ dUaμ

= n−1 a=1 dra .
Let us first suppose that q̂ is diagonal, q̂ = qaa δab . Then we have
  
J (q̂) = dÛ δ q̂ − Û Û T
 
n−1 
1,n−1 (2.88)
= dr1 . . . drn−1 δ qaa − |ra |2 δ(ra · rb ).
a=1 a<b

The first delta function on the right-hand side of this equation fixes the length of
the vectors {ra }, while the second imposes that they must be mutually orthogonal.
In polar coordinates, we have
  ∞  ∞
J (q̂) = dr̂1 . . . dr̂n−1 dr1 r1 · · ·
d−1 d−1
drn−1 rn−1
0 0

n−1 
1,n−1 (2.89)
× δ qaa − ra2 δ(ra rb r̂a · r̂b ),
a=1 a<b

where r̂a = ra /ra are unit vectors. Then

  ∞  ∞
J (q̂) = 21−n
dr̂1 . . . dr̂n−1 d−1
dr1 r1 . . . d−1
drn−1 rn−1
0 0
.n−1 / .1,n−1 / 1,n−1
 1 √  1 
× √ δ qaa − ua √ δ(r̂a · r̂b ) (2.90)
a=1
qaa a<b
qaa qbb a<b

n−1
= Cn,d (d−n)/2
qaa = Cn,d (det q̂)(d−n)/2 .
a=1
66 Atomic Liquids in Infinite Dimensions: Thermodynamics

The constant
 
1,n−1
Cn,d = 2 1−n
dr̂1 . . . dr̂n−1 δ(r̂a · r̂b ) (2.91)
a<b

can be computed recursively. If n = 2, there is just one unit vector and C2,d
= dr̂1 = d , where d is the d-dimensional solid angle. If n = 3, there are
two unit vectors: the first can access the entire solid angle d , and the second can
only access the d − 1 space orthogonal to the first vector, which gives d−1 . Hence,
C3,d = 2−1 d d−1 . Continuing this recursion, we get
Cn,d = 21−n d d−1 . . . d−n+2 . (2.92)
This completes the calculation when q̂ is diagonal.
The generalisation to the non-diagonal case is straightforward. Because the
matrix q̂ is symmetric, it can be diagonalised by an orthogonal transformation:
q̂ =  ˆ
ˆ −1 q̂D , ˆ = 1,
det  det q̂D = det q̂, ˆT =
 ˆ −1, (2.93)
and, thus,
 
   −1 
J (q̂) = dÛ δ q̂ − Û Û T = dÛ δ  ˆ − Û Û T
ˆ q̂D 
 (2.94)
 
= ˆ Û Û T 
dÛ δ q̂D −  ˆ −1 .

ˆ Û , we get
Performing the unitary change of integration variables V̂ = 
 
   
J (q̂) = dÛ δ q̂D −  ˆ Û Û T 
ˆ T = dV̂ δ q̂D − V̂ V̂ T
(2.95)
= J (q̂D ) = Cn,d (det q̂) (d−n)/2
,
recalling that det q̂ = det q̂D . We conclude that for any rotationally invariant func-
tion, we have
 
dr1 · · · drn−1 F ({ra }) = Cn,d dq̂ (det q̂)(d−n)/2 F (q̂). (2.96)
 3
Atomic Liquids in Infinite Dimensions
Equilibrium Dynamics

The aim of this chapter is to review the dynamical properties of liquids. For sim-
plicity, we will restrict ourselves to equilibrium dynamics – i.e., the study of a
system that starts in equilibrium and maintains it at all subsequent times. The
dynamics of low-density liquids is quite well understood through kinetic theory at
the microscopic level and through hydrodynamics at large length and time scales, as
discussed in [175]. Here, we are mostly interested in the properties of dense liquids
in the region close to the glass transition, where kinetic theory does not apply and
the hydrodynamic regime is pushed to extremely large scales [120]. Remember
that, as discussed at the beginning of Chapter 2, in large dimensions, one can focus
on the liquid phase without worrying about crystallisation.
The analytical study of the dynamics – e.g., of the correlation of two observables
at different times – is much more complex than the study of thermodynamic
properties – e.g., the correlation of these same observables at the same time in
equilibrium. We have to include a new dimension, time, to the problem. For
instance, the exact solution of the dynamics of large-dimensional liquids requires
the use of advanced dynamical tools such as path integrals within the Martin-
Siggia-Rose-De Dominicis-Janssen formalism [114, 187, 245]. These tools go
beyond the scope of this book, and we refer to [239] for details of this particular
derivation.
In Chapter 4, we will introduce a formalism that allows one to compute prop-
erties of the long time limit of dynamical observables, based on an extension of
standard thermodynamics. This approach considerably simplifies the calculations,
but understanding the motivations behind it requires some basic knowledge of
dynamics. For this reason, in this chapter, we present some essential general notions
about dynamics, and then we discuss, without a formal derivation, some important
results obtained from the exact solution of the infinite-dimensional dynamics.

67
68 Atomic Liquids in Infinite Dimensions: Equilibrium Dynamics

3.1 Properties of Equilibrium Dynamics

We consider here the same setting as in Chapter 2: N identical particles in d
dimensions, described by a set of points X = {xi }, which interact via a potential
 p2
energy V (X), with a total Hamiltonian H (P ,X) = i 2mi + V (X). The Gibbs–
Boltzmann equilibrium probability distribution at temperature T is proportional to
exp[−βH (P ,X)], but the configurational equilibrium distribution, which is propor-
tional to exp[−βV (X)], suffices to specify the probability distribution of observ-
ables O[X] that depend solely on particle positions, integrating away the momenta.
We consider here dynamical equations such that the initial condition at time
t0 is extracted from the equilibrium probability distribution, and at all subsequent
times, t > t0 , the probability distribution remains equal to the equilibrium one,
Pt (P ,X) = Pt0 (P ,X) ∝ exp[−βH (P ,X)]. Notable simplifications arise in this
case. For example, the average of observables that depend on a single time t –
i.e., of the form O[X(t)] – does not depend on t, and correlation functions of two
observables at distinct times depend on the time difference only:

O1 [X(t)]O2 [X(t  )] = CO1, O2 (t − t  ). (3.1)

This property is called ‘time-translational invariance’. By contrast, if the dynamics

starts from an off-equilibrium configuration – e.g., from a configuration ther-
malised at a temperature T0 = T – the situation is much more complex, because
one-time observables depend on t, and correlations depend separately on t,t  .
Time-translation invariance is only recovered if the system reaches equilibrium at
long times.
In the rest of this section, we provide examples of equilibrium dynamical equa-
tions, of relevant dynamical observables and of additional important properties of
the equilibrium dynamics. Because these topics are already covered by several
excellent books [152, 159, 168, 175, 343, 349], we limit our introduction to the
basic notions needed for the subsequent sections.

3.1.1 Dynamical Equations

Many different dynamical equations preserve the equilibrium probability distribu-
tion. We present three examples in this section. Here, as in the rest of this chapter,
a dot denotes a derivative with respect to time.

• Hamiltonian dynamics – The equations of motion are deterministic and derive

from the Hamiltonian function, as
∂H (P ,X) ∂V (X) ∂H (P ,X) pi
ṗi = − =− , ẋi = = . (3.2)
∂xi ∂xi ∂pi m
 3.1 Properties of Equilibrium Dynamics 69

These are equivalent to the Newtonian dynamics

∂V (X)
mẍi (t) = − = F i (X) (3.3)
∂xi
of an isolated system exclusively composed of interacting particles. In this case,
the Gibbs–Boltzmann probability distribution is left invariant by the time evolu-
tion, because the Hamiltonian is conserved by the dynamics and, according to
Liouville’s theorem, the time evolution also preserves volume elements in phase
space.
• Langevin dynamics – The equations of motion are obtained by adding a friction
term and a stochastic noise term to Eq. (3.3), resulting in

∂V (X)
mẍi (t) + ζ ẋi (t) = − + ξ i (t), (3.4)
∂xi
where the stochastic variables ξiμ (t) are uncorrelated Gaussian white noises with
zero mean and variance

ξiμ (t)ξj ν (t  ) = 2T ζ δij δμν δ(t − t  ). (3.5)

Physically, the Langevin equation describes a system of interacting Brownian

particles, immersed in a solvent that acts as a thermal bath. The macroscopic
interaction with the solvent is responsible for the frictional force −ζ ẋi (t), while
random collisions between the Brownian particles and the solvent particles pro-
vide the random force ξ i (t). These two terms have a common physical origin:
the interaction with the solvent. If the solvent is in equilibrium, the same friction
coefficient ζ enters in both, and the prefactor T in Eq. (3.5) corresponds to the
solvent temperature. Note that in the limit ζ → 0, the friction and noise terms
disappear, and one recovers Newtonian dynamics. One can also consider the
opposite limit in which inertia is negligible with respect to friction – i.e., m → 0.
In this case, one obtains the overdamped Langevin equation, which describes only
the configurational degrees of freedom but not the momenta that are undefined
in that limit. An explicit computation (see Section 3.6) shows that the Gibbs–
Boltzmann probability distribution is left invariant by the Langevin evolution,
provided that the value of β in the Gibbs–Boltzmann distribution is equal to 1/T ,
where T is the temperature that appears in the noise variance, Eq. (3.5).
One can also consider a more general Langevin dynamics that depends on a
kernel K(t):
 t
∂V (X)
mẍi (t) + dt  K(t − t  )ẋi (t  ) = − + ξ i (t), (3.6)
t0 ∂xi
70 Atomic Liquids in Infinite Dimensions: Equilibrium Dynamics

with
ξiμ (t)ξj ν (t  ) = T δij δμν K(t − t  ). (3.7)
In this case, ξiμ (t) are uncorrelated Gaussian coloured noises. Note that the kernel
must be symmetric, K(t) = K(−t), to be consistent with Eq. (3.7). The white
noise case is recovered  t by choosing K(t) = ζ e−|t|/τ /τ → 2ζ δ(t) for τ → 0
and observing that t0 dt  K(t − t  )ẋ(t  ) → ζ ẋ(t) in that same limit. For the
coloured case also, if the same kernel appears both in the friction term and in the
correlation of the coloured noise, the Gibbs–Boltzmann probability distribution
is left invariant by the time evolution.1
• Monte Carlo dynamics – Here time is discrete, and the evolution is a Markov
chain. The probability of going from one configuration at time t to another con-
figuration at time t + 1, P [X(t + 1)|X(t)], does not depend on the previous
history of the system. In the simplest Monte Carlo dynamics, one assumes that
the Wegscheider-Einstein ‘detailed balance’ condition [349] is satisfied:
P [Y |X]e−βV (X) = P [X|Y ]e−βV (Y ) . (3.8)
In this case, it is easy to check that the Gibbs–Boltzmann probability distribution
is left invariant by the evolution, because of the relation
 
−βV (X)
dXP [Y |X]e = dXP [X|Y ]e−βV (Y ) = e−βV (Y ), (3.9)

in which we used that dXP [X|Y ] = 1. Eq. (3.9) implies that if at a given
time step the dynamics is in equilibrium, one additional step preserves that
equilibrium.

3.1.2 Dynamical Observables

Several types of observables are of interest in liquid dynamics. We refer to [152,
168, 175] for an extended list. Here, we restrict the discussion to observables related
to the equilibrium dynamics of an isotropic and homogeneous system – e.g., the
liquid phase.
One-time observables are simply the dynamical version of the thermodynamic
observables discussed in Section 2.1.2, and have the form

1 This result has been known for a long time and is consistent with physical intuition. However, a simple formal
mathematical proof is not easily found in the literature. One possibility is to represent K(t) as a sum of
exponentials and use the Ornstein-Uhlenbeck representation to map the coloured Langevin process into a
white noise Langevin process in an extended space with additional degrees of freedom [159]. The equilibrium
state in the extended space is a Gibbs–Boltzmann distribution, which can be marginalised over the additional
degrees of freedom to recover the Gibbs–Boltzmann distribution for X. Another possibility is to prove that the
correlations and responses of Eq. (3.6), if the initial condition is in equilibrium, satisfy the
fluctuation–dissipation relation (see Section 3.1.3) at all times [18], which (indirectly) implies that the
equilibrium is preserved at all times.
 3.1 Properties of Equilibrium Dynamics 71
 
O[X(t)] = O1 [xi (t)] + O2 [xi (t),xj (t)] + · · · (3.10)
i i=j

These observables can be expressed in terms of a local density field


N
 
ρ[x;X(t)] = δ(xi (t) − x), ρ(x,t) = ρ[x;X(t)] , (3.11)
i=1

and its equal time spatial correlations, as discussed in Section 2.1.2 for the thermo-
dynamic analysis. Note that here and in the following, brackets indicate an average
over the ensemble of dynamical trajectories X(t) generated by the chosen model for
the dynamics. Note also that for notational simplicity, we often use the shorthand
O(t) = O[X(t)] in the following.
One can next consider observables of the form of Eq. (3.10), but with particle
positions evaluated at different times. These observables can be expressed as func-
tions of correlations of ρ[x;X(t)] at different times and positions. We here focus
on a particular class of two-time correlations that depend on both the configuration
at time t and that at time t  , of the form

O(t,t  ) = O2 [xi (t) − xj (t  )]. (3.12)
ij

These correlations can be expressed in terms of the ‘collective van Hove function’
[175],
1 
G(r,t − t  ) = ρ[x + r;X(t)]ρ[x;X(t  )]
ρ
1  
= δ[r − xi (t) + xj (t  )] , (3.13)
N ij

 
O(t,t ) = N drO2 (r)G(r,t − t  ),


which represents equivalently the space–time correlation of the density field, or

the (non-normalised) probability that a particle at time t is found at distance r
from the position of another particle at time t  . To go from the first to the second
line in Eq. (3.13), one makes use of translational invariance [175]. From the van
Hove function, one can construct other observables such as the coherent scattering
function or the dynamical structure factor [175].
In the limit of infinite dimensions, as we have seen in Chapter 1, correlations
become extremely short ranged. As a consequence, all terms with i = j in a
correlation function such as Eq. (3.13) vanish. It then becomes useful to consider
‘self’ correlations, which depend on the time evolution of a single particle. One
72 Atomic Liquids in Infinite Dimensions: Equilibrium Dynamics

therefore introduces the distribution of single particle displacements, also called

the ‘self van Hove function’, defined as

1  
N

Gs (r,t − t ) = δ[r − xi (t) + xi (t  )] . (3.14)
N i=1

The self van Hove function contains only the diagonal terms with i = j of
Eq. (3.13). It represents the (normalised) probability that a single particle moves by
r in time t − t  . From it, one can derive several interesting observables, such as the
‘mean square displacement’,

1  
N
D(t) = |xi (t) − xi (0)|2 = dr|r|2 Gs (r,t), (3.15)
N i=1

and the ‘self intermediate scattering function’,


1   iq·[xi (t)−xi (0)] 
N
Fs (q,t) = e = dr eiq·r Gs (r,t), (3.16)
N i=1

which are the second moment and the Fourier transform, respectively, of the self
Van Hove function. In the rest of this chapter, we will focus particularly on D(t).
Finally, we have seen in Chapter 1 that while correlations between different
particles vanish in the limit d → ∞, their volume integral, which defines a sus-
ceptibility, remains finite and can even diverge at a phase transition. This motivates
the study of dynamical susceptibilities such as
  
1 
N
2 2
χ4 (t) = N D̂ (t) − D̂(t) , D̂(t) = |xi (t) − xi (0)|2 . (3.17)
N i=1

These functions can be written as volume integrals of correlation functions that

depend on two spatial points and two times and are thus called ‘four point corre-
lation functions’. They are extremely useful for characterising the glass transition.
We refer to [37] for a detailed review.
Finally, together with correlation functions, we consider the linear responses of
observables to small perturbations. The perturbations are introduced by modifying
the interaction potential
V [X] = V [X] − (t)O2 [X], (3.18)
and the variation of the average of an observable O1 (t) is considered in their pres-
ence. For small (t), the variation is linear and takes the most general form
 t
O1 (t) = O1 (t)=0 + dt  RO1, O2 (t,t  )(t  ), (3.19)
−∞
 3.1 Properties of Equilibrium Dynamics 73

which is equivalent to evaluating the functional derivative at vanishing perturbation:

0
 δ O1 (t) 00
RO1, O2 (t,t ) = . (3.20)
δ(t  ) 0=0
By causality of the equations of motion, the response at time t cannot depend on
times t  > t, which sets the upper bound of the integration in Eq. (3.19). Equilibrium
response functions are also time-translationally invariant and, hence, only depend
on the time difference t − t  .
Several examples of response functions that are relevant to liquid dynamics
are given in [175, chapter 7]. The simplest example is obtained by choosing
O2 [X] = xiμ , for an arbitrary particle i and spatial component μ, and O1 (t)
= piμ (t)/m = ẋiμ (t). This choice physically corresponds to a force (t) being
applied on a particle and its velocity being recorded. If at time t0 a constant
(t) = θ(t − t0 ) is switched on, one then obtains
 t
 
ẋiμ (t)  =  dt  Rẋiμ,xiμ (t − t  ). (3.21)
t0

Assuming that the response function vanishes fast enough at long times, one then
obtains
   ∞
ẋiμ (t) 
μt = lim lim = dt  Rẋiμ,xiμ (t  ). (3.22)
→0 t→∞  0

Hence, at long times, the average velocity reaches a finite limit proportional to
the applied force , via a ‘transport coefficient’ μt called translational ‘mobility’.
Transport coefficients are generally expressed as integrals of response functions, as
in Eq. (3.22). Other important examples of transport coefficients are the electrical
conductivity and the shear viscosity [175, chapters 7 and 8].

3.1.3 Reversibility and Fluctuation–Dissipation Relation

A very interesting general property of the equilibrium dynamical equations
introduced in Section 3.1.1 is that the correlations are invariant under time
reflections. First we note that, due to time-translation invariance, correlations
satisfy the relation

CO1, O2 (t) = O1 (t)O2 (0) = O1 (0)O2 (−t) = CO2, O1 (−t). (3.23)

If O1 [X] and O2 [X] are invariant under time reversal, reversibility implies the
additional relation

CO1, O2 (t) = CO1, O2 (−t) ⇔ CO1, O2 (t) = CO2, O1 (t), (3.24)

74 Atomic Liquids in Infinite Dimensions: Equilibrium Dynamics

where the equivalence is due to Eq. (3.23). This commutation symmetry of the cor-
relations is a direct consequence of the microscopic reversibility of the dynamical
equations in the case of Hamiltonian dynamics and of the detailed balance condi-
tion (3.8) in the case of Monte Carlo dynamics. It may, however, look surprising
for Langevin dynamics, because the microscopic equations are not time reversible.
They indeed describe the motion of particles in presence of friction and noise. This
microscopic arrow of time nonetheless bears no consequence on the equilibrium
behaviour. A proof of this fact is given in Section 3.6.
Another general property of equilibrium dynamics is the fluctuation–dissipation
relation. This relation is a consequence of the zeroth law of thermodynamics, which
states that two bodies in thermal contact acquire the same temperature in equilib-
rium. In fact, the fluctuation–dissipation relation can be derived from the theory
of harmonic thermometers [110, 111]. A basic principle of thermodynamics is
that, when a thermometer is weakly coupled to a much larger system, it reaches
the temperature of the larger system. If the thermometer is a harmonic oscillator,
one obtains the temperature of the system by measuring the average energy of the
harmonic oscillator, which is equal to its temperature by the equipartition relation.
Consider then a harmonic oscillator x(t) of frequency ω, coupled to an observ-
able O(t) = O[X(t)] of a much larger system, through a total Hamiltonian

p 2 ω2 x 2
Htot (p,x;P ,X) = + − xO(X) + H [P ,X], (3.25)
2 2
with a coupling constant . The equation of motion of the harmonic oscillator is
then

ẍ(t) − ω2 x(t) + O(t) = 0. (3.26)

The oscillator perturbs the system slightly by the coupling term −x(t)O(X).
Assuming that the coupling is small and that for simplicity O(t) = 0 in absence
of coupling, linear response theory (Section 3.1.2) gives
 t
O (t) = Of (t) + dt  RO, O (t − t  )x(t  ). (3.27)
−∞

The first term is the fluctuating part of O(t), which has zero average, and the
second term describes the shift of the average of O(t) due to the coupling with
the oscillator. The dynamical equation for the oscillator then becomes
 t
ẍ(t) − ω x(t) +
2
Of (t) + 2
dt  RO, O (t − t  )x(t  ) = 0. (3.28)
−∞
 3.1 Properties of Equilibrium Dynamics 75

This linear equation is a particular case of Eq. (3.6) for a single variable x(t) with
a harmonic potential and a friction kernel2  2 RO, O (t) = −θ(t)K̇(t). The noise
Of (t) has autocorrelation  2 Of (t)Of (0) =  2 CO, O (t) = T K  (t). Imposing
that the kernels K(t) and K  (t) be the same, as in Eq. (3.6), guarantees that, at
long times, the oscillator reaches equilibrium at temperature T independently of its
frequency. This condition leads to the relation
RO, O (t) = −βθ(t)ĊO, O (t), (3.29)
which is the fluctuation–dissipation relation. Alternatively, Eq. (3.28) can be solved
explicitly to compute the average energy of the oscillator. See [110] for details. The
energy of the oscillator is thus equal to T , independently of its frequency ω, only
if Eq. (3.29) holds. This argument shows that the fluctuation–dissipation relation
is a physical requirement for a system to be in equilibrium. If Eq. (3.29) does not
hold, then a thermometer coupled to the system measures a different temperature
depending on its characteristic frequency [110].
The fluctuation–dissipation relation, Eq. (3.29), can also be mathematically
derived by calculating directly the response function, considering the variation of
the dynamical equations under the perturbation; see, e.g., [175, 282, 343, 349].
This derivation can also be generalised to two different observables, leading to a
fluctuation–dissipation relation of the form
RO1, O2 (t) = −βθ(t)ĊO1, O2 (t). (3.30)
The fluctuation–dissipation relation has many important applications. In partic-
ular, it can be used to express transport coefficients in terms of correlation func-
tions. For example, using the fluctuation–dissipation relation and the reversibility
property3 expressed by Eq. (3.24), one can rewrite Eq. (3.22) as [175, chapter 7]
 ∞
 
μt = β dt ẋiμ (t)ẋiμ (0) = βD, (3.31)
0

where
D(t)
D = lim (3.32)
2dt t→∞

is the ‘diffusion constant’. The mobility is then expressed as the time integral of
the autocorrelation of the particle velocity, which is an example of a ‘Green–Kubo
relation’ [175, 343, 349]. The relation between mobility and the diffusion constant
is known as ‘Einstein relation’.

2 The derivative is there because in Eq. (3.28) the response function is coupled to x(t) instead of ẋ(t).
t
An integration by parts of −∞ dt  K̇(t − t  )x(t  ) allows one to obtain Eq. (3.6).
3 Note that because ẋ is odd under time reversal, there is an additional minus sign in Eq. (3.24).
iμ
76 Atomic Liquids in Infinite Dimensions: Equilibrium Dynamics

3.2 Langevin Dynamics of Liquids in Infinite Dimensions

We now consider a system of particles evolving under the Langevin dynamics
defined by Eqs. (3.4) and (3.5). The aim of this section is to show that the Langevin
equation of N interacting particles, taking the thermodynamic limit N → ∞
followed by the limit d → ∞, can be mapped into the dynamics of a single effective
particle, coupled to a coloured thermal bath that represents the interaction of the
effective particle with all the other particles. The kernel of the effective thermal
bath can then be determined self-consistently. More precisely, any dynamical cor-
relation function of the original many-body Langevin process can be computed as
a dynamical correlation of the effective process. This mapping is a general feature
of mean field systems. An example is the spin glass case discussed in [111, 254].
It is important to stress that while this mapping will be discussed for a Langevin
dynamics with friction and noise, one can also consider the limit ζ → 0 at which
the noise disappears and the Langevin dynamics becomes Newtonian. This limit,
however, should be taken after the limits N → ∞ and d → ∞. In the limit ζ → 0,
the coloured noise in the effective Langevin process does not disappear. Newtonian
dynamics is then also described by a self-consistent effective Langevin dynamics,
for which noise is generated by interactions.
The idea of deriving effective equations for the dynamics of liquids in d → ∞
was first introduced by Kirkpatrick and Wolynes [206]. The exact solution by map-
ping onto an effective process was derived in full detail in [239], but this derivation
requires advanced path integral dynamical tools [79, 111, 114, 187, 200, 245] that
will not be discussed in this book. Here, for simplicity, we will not give a full deriva-
tion but a series of physically reasonable arguments, which were first proposed by
Szamel [339]. The advantage of the simplified derivation is that it gives a clear
understanding of the physical properties that underlie the full derivation, and it also
leads to the final result in the simplest possible way. The disadvantage, however,
is that some steps are not fully justified, and the correctness of the derivation can
only be verified a posteriori by comparing the result with that obtained via the path
integral formalism [3].
For additional simplicity, in the rest of this section, we consider the overdamped
case, m = 0, but the inertial term can be reinserted at any time in the discussion.
We also recall that, as in all this chapter, dynamics starts in equilibrium at the initial
time, chosen here to be t0 = 0.

3.2.1 Single-Particle Effective Process

The Langevin equation for particle i has the form
ζ ẋi (t) = F i (t) + ξ i (t),
∂V [X(t)]  ∂v(|xi (t) − xj (t)|)  (3.33)
F i (t) = − =− = F j →i (t),
∂xi j (=i)
∂xi j (=i)
 3.2 Langevin Dynamics of Liquids in Infinite Dimensions 77

where F j →i (t) is the force that particle j exerts on i at time t. In finite d, because
forces are short ranged, particle i only interacts with a finite number of neighbours.
However, when d → ∞, the number of neighbours also diverges,4 and F i (t) is the
sum of an infinite number of terms. This situation is very similar to what happens
when d is finite, but particle i is much bigger than the other fluid particles and thus
undergoes a Brownian motion. We are going to make use of this analogy to derive,
in d → ∞, an effective equation for particle i that has the form of a generalised
Brownian motion.
First, we consider the case in which particle i is immobile (ẋi = 0). Then,
by isotropy of the liquid phase, the force F i (t) exerted on i by the other particles
has zero average. Furthermore, because in d → ∞ the number of neighbours
also diverges, the forces F j →i (t) can be considered as uncorrelated, exactly as
in the Ising model, see Chapter 1. This is illustrated in Figure 3.1. Because

F i (t) = j (=i) F j →i (t) is then the sum of many weakly correlated terms, its
fluctuations are Gaussian by the central limit theorem. The autocorrelation of one
of its components
    
M(t − t  ) = Fiμ (t)Fiμ (t  ) = Fj →i,μ (t)Fk→i,μ (t  ) , (3.34)
j,k(=i)

does not depend on i nor μ, because all particles are identical and the fluid is
isotropic. The contributions of different particles being uncorrelated, one can fur-
ther simplify this expression by removing the terms with j = k to obtain
  1  
M(t − t  ) = Fj →i,μ (t)Fj →i,μ (t  ) = Fj →i,μ (t)Fj →i,μ (t  ) , (3.35)
j (=i)
N i=j

where in the second equality we used the fact that particles are identical.
Next, we consider a moving particle. As in the case of standard Brownian
motion, to the fluctuating part discussed earlier, we should now add an average
force F av
i (t) that is proportional to the particle velocity, because the rest of the fluid
acts as a frictional medium. In equilibrium, the frictional force must
 t be related to the
noise, as discussed in Section 3.1.1, and, hence, F i (t) = −β 0 dt M(t − t )ẋi (t  ).
av 

We thus obtain that, in the limits N → ∞ and d → ∞, the dynamics of any given
particle i is governed by a single-particle Langevin process with coloured noise

4 This intuitive statement can be supported by two observations. First, the ‘kissing number’, which is the
maximal number of non-overlapping spheres that can be put in contact with a central sphere, diverges with
d [103]. Second, for a finite range potential, the average number of neighbours of any particle in the liquid
phase can be defined as the integral for r ≤  of the radial distribution function given in Eq. (2.77),
     0
z = ρd drg(r) = ρd dre−βv(r) = ϕ dh eh−β v̄(h) .
0 0 −∞

Because the dynamical glass transition happens for ϕ ∝ d (see Section 3.3.3) and the integral over h is finite,
the number of neighbours is proportional to d.
78 Atomic Liquids in Infinite Dimensions: Equilibrium Dynamics

i (t) = ξ i (t) + F fli (t), where the fluctuating force is Gaussian, independent from
ξ i (t), and has a memory kernel M(t − t  ):
 t
ζ ẋi (t) = −β dt  M(t − t  )ẋi (t  ) + i (t),
0 (3.36)
 
  
iμ (t)iν (t ) = δμν [2T ζ δ(t − t ) + M(t − t )].
Eq. (3.36) is certainly not exact in finite d, except in the case where particle i
is much bigger than the other particles. In that case, particle i indeed undergoes
Brownian motion. The Brownian motion description is appropriate for a much
bigger particle because it interacts with many small particles at the same time, and
the interactions can then be described using the central limit theorem, which leads
to Eq. (3.36). But in infinite d, the number of neighbours is very large even for
a particle of the same size as its neighbours, which justifies Eq. (3.36). Note that
Eq. (3.36) depends only on ẋi (t), and, therefore, the initial condition for xi (t) is
completely irrelevant.

3.2.2 Two-Particle Effective Process and Equation for the Memory Kernel
Equation (3.35) expresses the memory kernel as an autocorrelation function of
the interparticle force. Its computation requires writing an effective process for
the dynamics of two particles, as illustrated in Figure 3.1. The original Langevin
equation for particles i and j reads
(j )
ζ ẋi (t) = F j →i (t) + F i (t) + ξ i (t),
(3.37)
ζ ẋj (t) = −F j →i (t) + F (i)
j (t) + ξ j (t),

Figure 3.1 Illustration of the one-particle (left) and two-particles (right) effective
stochastic processes. In the one-particle case, the central particle interacts with
a large number of neighbours, diverging with d, which are assumed to be
uncorrelated for d → ∞. The resulting fluctuating and friction forces provide
an effective thermal bath. In the two-particle case, the pair of particle exchange a
force, and each particle also interacts with a large number of neighbours, which
are assumed to be independent and thus provide two independent effective thermal
baths, each identical to a one-particle process.
 3.2 Langevin Dynamics of Liquids in Infinite Dimensions 79
(j )
where F i (t) denotes the total force on particle i, without the contribution coming
(j )
from particle j . In the limit d → ∞, the number of terms in F i (t) diverges with
d, and removing one contribution is a small correction of order 1/d. We can thus
(j )
apply to F i (t) the same treatment as that applied to the total force in Section 3.2.1,
to obtain
 t
ζ ẋi (t) = −β dt  M(t − t  )ẋi (t  ) + i (t) + F j →i (t),
0
 t
(3.38)
  
ζ ẋj (t) = −β dt M(t − t )ẋj (t ) + j (t) − F j →i (t),
0

where the noises i (t) and j (t) have the same statistics as in Eq. (3.36).
In Eq. (3.38), the terms F j →i (t) are small corrections to the noise of order 1/d,
as discussed earlier, because the force between particle i and j is very weakly
correlated with the positions xi (t) and xj (t) of these particles. Yet the relative
displacement r(t) = xi (t) − xj (t) satisfies the equation obtained by taking the
difference of the first and second lines in Eq. (3.38) and dividing by 2 for later
convenience,

ζ β t 
ṙ(t) = − dt M(t − t  )ṙ(t  ) + (t) + F (r(t)),
2 2 0
  (3.39)
  1
μ (t)ν (t  ) = δμν T ζ δ(t − t  ) + M(t − t  ) ,
2
where
∂v(|r|)
F (r) = −
. (3.40)
∂r
Now, the term F (r(t)) is strongly correlated with r(t) because, for central poten-
tials, the force is parallel to the distance. Then this term cannot be neglected5 in
Eq. (3.39).
The initial condition for r(t) in Eq. (3.39), which we denote by r0 = r(0),
should be taken at random from the equilibrium distribution of interparticle dis-
tances, which is proportional to g(r0 ) = e−βv(|r0 |) in d → ∞, as discussed in
Chapter 2. Indeed, one can check that this distribution is left invariant by Eq. (3.39).
However, because g(r0 ) → 1 for |r0 | → ∞, this distribution is not normalisable.
To determine the correct normalisation, we consider the value of M(t − t  ) at equal
times, t = t  . Using the thermodynamic results of Section 2.1.2, Eq. (3.35) gives

5 When d → ∞, the projection of F (r(t)) on a random direction is much smaller than its projection on r(t),
which explains why it can be neglected in Eq. (3.38) while it must be kept in Eq. (3.39). This statement will
be made precise in Section 3.3.3.
80 Atomic Liquids in Infinite Dimensions: Equilibrium Dynamics

1   ρ2
M(0) = |F j →i (0)|2 = dxdyg(x − y) |F (x − y)|2
Nd i=j Nd
 (3.41)
ρ
= dr0 g(r0 ) |F (r0 )| .
2
d
At different times, by continuity, one therefore obtains

1   ρ
M(t) = F j →i (t) · F j →i (0) = dr0 g(r0 ) F (r(t)) · F (r0 ), (3.42)
Nd i=j d

which reduces to Eq. (3.41) for t = 0. The dynamical average F (r(t)) in

Eq. (3.42) is originally over the many-body dynamics but can be replaced by
an average over the effective process in Eq. (3.39) that encodes the dynamical
evolution of interparticle distances. Therefore, one should consider every possible
initial condition r(0) = r0 , evolve it according to Eq. (3.39) to compute F (r(t))
averaged over the noise (t) and, finally, integrate over r0 to obtain the memory
kernel. Note that because F (r0 ) decays quickly to zero for |r0 | → ∞, the integral
over r0 in Eq. (3.42) is convergent.
Eqs. (3.39) and (3.42) then constitute a closed system of equations for M(t). One
could solve it numerically by starting with a guess for M(t), solving the process in
Eq. (3.39), using the result to compute a new guess for M(t) using Eq. (3.42) and
repeating until convergence.

3.2.3 Mean Square Displacement and the Diffusion Constant

Once the self-consistent equation for M(t) has been solved, one can use the effec-
tive particle dynamics given by Eq. (3.36) to obtain dynamical observables. As
an example, we derive here the equation for the mean square displacement D(t)
introduced in Eq. (3.15). To this end, let us rewrite Eq. (3.36) as
 ∞
dt  (t − t  )ẋ(t  ) = (t),
(3.43)
0
 
  
μ (t)ν (t ) = δμν T [(t − t ) + (t − t)],
where we dropped the suffix i because all particles follow the same identical equa-
tion, and we introduced
(t − t  ) = θ(t − t  )[2ζ δ(t − t  ) + βM(t − t  )]. (3.44)
We then have
 ∞
ẋ(t) = dt   −1 (t − t  )(t  ),
0
 t  t  ∞
(3.45)
    −1   
x(t) − x(0) = dt ẋ(t ) = dt dt  (t − t )(t ),
0 0 0
 3.2 Langevin Dynamics of Liquids in Infinite Dimensions 81

where  −1 (t) is the operator inverse of (t). Hence,

Ḋ(t) = 2 [x(t) − x(0)] · ẋ(t) (3.46)
 t  ∞  ∞
 
=2 dt  dt  dt   −1 (t  − t  ) −1 (t − t  ) (t  ) · (t  ) .
0 0 0

Using Eq. (3.43) for the noise correlation, we have

 t  ∞  ∞
 
Ḋ(t) = 2dT dt dt dt   −1 (t  − t  ) −1 (t − t  )
0 0 0

× [(t − t ) + (t − t  )]
  

 t
  (3.47)
= 2dT dt   −1 (t − t  ) +  −1 (t  − t)
0
 ∞
= 2dT dt   −1 (t − t  ),
0

where in the last step we used that  −1 (t − t  ) vanishes for t  > t, because it is the
inverse of an operator that has the same property.6
Equation (3.47) can be integrated over time using D(0) = 0 to obtain an explicit
expression of D(t) in terms of M(t), but this requires inverting (t). It is more
convenient to derive an equation for D(t) by writing
 ∞  ∞  ∞
    
dt (t − t )Ḋ(t ) = 2dT dt (t − t ) dt   −1 (t  − t  ) = 2dT ,
0 0 0
(3.48)
which can be written more explicitly as
 t
ζ Ḋ(t) = −β dt  M(t − t  )Ḋ(t  ) + 2dT . (3.49)
0

Equation (3.49) is the desired result that

 ∞expresses D(t) as a function of M(t).
) −st
Introducing Laplace transforms, D(s) = 0 dt e D(t), it can also be written as

) 2dT
D(s) = . (3.50)
s 2 [ζ )
+ β M(s)]
It is reasonable to assume on very general grounds that in the liquid, D(t) is
diffusive – i.e., linear at large times, D(t) ∼ 2dDt – and M(t) decays to zero

6 The operators (t − t  ) that vanish for t  > t form a closed algebra with identity, and, therefore, the inverse
 −1 (t  − t) also vanishes for t  > t. These operators are the continuum analog of upper triangular matrices,
which also have the same properties.
82 Atomic Liquids in Infinite Dimensions: Equilibrium Dynamics
t
sufficiently fast when t → ∞. We then have lim 0 dt  M(t − t  )Ḋ(t  ) = 2dD
∞ t→∞

0 dt M(t), and we obtain from Eq. (3.49) an expression for the diffusion
coefficient:
 ∞
T
ζ D = T − βD dt M(t) ⇒ D= ∞ . (3.51)
0 ζ + β 0 dt M(t)
The same expression is equivalently obtained from Eq. (3.50) by observing that for
s → 0, ) ) has a finite limit corresponding to the integral
D(s) ∼ 2dD/s 2 , while M(s)
of M(t). Because the memory kernel M(t) originates from interactions, the absence
of interactions at low density gives M(t) = 0, as it is clear from Eq. (3.42). One
then recovers the Einstein expression of the diffusion coefficient in Eq. (3.31),
D = T /ζ = T μt with mobility μt = 1/ζ , for independent particle dynamics.
Upon increasing density, M(t) increases and the diffusion coefficient decreases.

3.3 Dynamical Glass Transition

The self-consistent expression for the memory kernel M(t) provided by Eqs. (3.39)
and (3.42) predicts that the system dynamically arrests at high density or low tem-
perature. Dynamical arrest corresponds to the memory kernel not decaying to zero
in the limit t → ∞. We can thus write
M(t) = Mf (t) + M∞, lim Mf (t) = 0, (3.52)
t→∞

and assume that Mf (t) decays to zero sufficiently fast to be integrable. If M∞ > 0,
then the integral of M(t) diverges and according to Eq. (3.51) the diffusion constant
vanishes.7 The system is then dynamically arrested. As a result, the mean square
displacement also reaches a plateau8 at long times – i.e., D(t) → D for t → ∞.
Using the decomposition in Eq. (3.52) into Eq. (3.49), and recalling that D(0) = 0,
we obtain
 t
ζ Ḋ(t) = −β dt  Mf (t − t  )Ḋ(t  ) − βM∞ D(t) + 2dT. (3.53)
0

When t → ∞, both Ḋ(t) and Mf (t) quickly decay to zero. For large t, in the
integral 0 dt  Mf (t − t  )Ḋ(t  ), the variable t  can be either finite, in which case
t

7 Note that the diffusion constant can also vanish if M(t) ∼ t −α for t → ∞, with 0 < α < 1. In this case,
)
M(s) ∼ s α−1 for s → 0. From Eq. (3.50), the mean square displacement behaves as ) D(s) ∼ s −1−α ; hence,
D(t) ∼ t α – i.e., it is sub-diffusive. While this situation can happen at some dynamical critical points [168],
we do not consider it in this book.
8 We avoid adding a suffix ∞ to the plateau value, D, to simplify the notation in Chapter 4. We also stress that
D represents the diffusion coefficient, while D represents the plateau of the mean square displacement, which
are two different quantities.
 3.3 Dynamical Glass Transition 83

Mf (t − t  ) → 0, or of order t, in which case Ḋ(t  ) → 0; the integral therefore goes

to zero for t → ∞. Using these results, we obtain from Eq. (3.53) the relation
2d
−βM∞ D +2dT = 0 ⇒ D= , (3.54)
β 2 M∞
between the plateau of the mean square displacement and the plateau of the memory
kernel. The same result can be obtained from Eq. (3.50) by observing that for
)
s → 0, M(s) ∼ M∞ /s and ) D(s) ∼ D /s. In the following, we derive a closed
equation for M∞ and, from it, an equation for the dynamical transition point.

3.3.1 Calculation of the Plateau Value of Correlation Functions

Using the decomposition in Eq. (3.52), we can write Eq. (3.39) as

ζ β t 
ṙ(t) = − dt Mf (t − t  )ṙ(t  ) + f (t)
2 2 0
∂v β
− (|r(t)|) − M∞ [r(t) − r(0)] + ∞,
∂r 2
  (3.55)
  1
f ,μ (t)f ,ν (t  ) = δμν T ζ δ(t − t  ) + Mf (t − t  ) ,
2
  1
∞,μ ∞,ν = δμν M∞ .
2
Note that we decomposed the noise into two independent components: a constant
one, ∞ , whose correlation is related to M∞ , and a fast one, f (t), whose corre-
lation is related to Mf (t) and decays quickly over time. The two variables ∞ and
f (t) are assumed to be independent and Gaussian, with zero average; their sum
(t) = ∞ + f (t) thus has zero average and the correct variance, as requested
by Eq. (3.39).
Eq. (3.55) describes the dynamics of an effective particle, r(t), moving in the
potential
β
w(r) = v(|r|) + M∞ (r − r0 )2 − ∞ · r (3.56)
4
and subjected to a retarded friction and coloured noise with fast-decaying ker-
nels that also satisfy the fluctuation–dissipation relation. Note that, thanks to the
quadratic term proportional to M∞ , this potential is confining around r0 . Therefore,
at long times, the effective particle position equilibrates in the potential w(r), and
its distribution is given by the equilibrium result,
β2
e−βw(r) e−βv(|r|)− 4 M∞ |r| +βs·r
2

P (r|s) =   −βw(r ) =  2
, (3.57)
dr e 
dr e −βv(|r |)− β4 M∞ |r |2 +βs·r
84 Atomic Liquids in Infinite Dimensions: Equilibrium Dynamics

with s = ∞ + βM∞ r0 /2. The average force, F (r(t)), then converges to the
average force over the equilibrium distribution and the noise
 
F (r) = d∞ P (∞ ) drP (r|∞ + βM∞ r0 /2)F (r), (3.58)

|∞ |2
where P (∞ ) = e− M∞ /(πM∞ )d/2 is the distribution of the Gaussian noise ∞ .
Plugging Eq. (3.58) in Eq. (3.42) gives a self-consistent equation for M∞ ,

ρ
M∞ = dr0 g(r0 )F (r0 ) · F (r) . (3.59)
d
Note that M∞ = 0 – i.e., the solution corresponding to the liquid phase – is always
a solution of Eq. (3.59). In fact, for M∞ = 0, one has∞ = 0 and then s = 0.
Eq. (3.58) then becomes F (r) = drF (r)e−βv(|r|) / dre−βv(|r|) . In absence of
the confining term provided by M∞ , the integral in the denominator, dre−βv(|r|) ,
is divergent because e−βv(|r|) → 1 at large |r|. The integral in the numerator is
instead finite, because F (r) is short ranged. It follows that F (r) = 0, which can
be inserted in Eq. (3.59) and self-consistently shows that M∞ = 0 is a solution.
Finding non-zero solutions of Eq. (3.59) is not easy because numerically eval-
uating its right-hand side requires calculating several integrals over d-dimensional
vectors that are not rotationally invariant. It can, however, be simplified by assum-
ing that g(r0 ) = exp[−βv(|r0 |)], which is exact for d → ∞, as discussed in
Chapter 2. After some manipulations9 (details can be found in [339]), one finds
that Eq. (3.59) can be equivalently expressed in terms of D, as
d
= F1 (D ;β),
ρVd
 ∞
∂q(D ,β;r)
F1 (D ;β) = −d D dr r d−1 log[q(D ,β;r)] , (3.60)
0 ∂D
 d|u|2
e− 2 D
q(D ,β;r) = du e−βv(|r+u|) .
(2π D /d)d/2
Eq. (3.60) is much easier than Eq. (3.59) to evaluate numerically, because the
function q(D ,β;r) is the convolution of a Gaussian with a rotationally invariant
function, and, therefore, it is itself rotationally invariant. It can thus be transformed
into a one-dimensional integral either by moving to Fourier space or by using
bipolar coordinates, which gives [292]

9 Note that the equivalence of Eq. (3.60) and Eq. (3.59) only holds when M > 0, which means that the liquid
∞
solution M∞ = 0 is lost in going from one equation to the next. However, this limitation is not severe because
we already know that the solution M∞ = 0 is always present.
 3.3 Dynamical Glass Transition 85
 ' u ( d−1 2 &  
∞ − d(r−u)
−βv(u) 2 e 2D
− dDru d ru d ru
q(D ,β;r) = du e  e 2π I d−2 ,
0 r 2π D /d D 2 D
(3.61)
where In (x) is the modified Bessel function of the first kind. The derivative of
q(D ,β;r) with respect to D is then easily computed, and F1 (D ;β) can be obtained
via another one-dimensional integral. The solution of Eq. (3.60) for D provides
the plateau value of mean square displacement and that of the memory kernel via
Eq. (3.54).

3.3.2 Equation for the Dynamical Glass Transition

A dynamical glass transition can emerge from Eq. (3.60) if the density is high
enough or if the temperature is low enough. Because the function F1 (D ;β) does not
depend on density, which only appears in the left-hand side of the first Eq. (3.60),
the case in which temperature is held constant and density is increased is easier to
discuss graphically. Thus, we here specialise to this case, but similar results can be
obtained at constant density upon lowering the temperature.
One can show that for simple interaction potentials, the function F1 (D ;β) is
positive, vanishes both for D → 0 and D → ∞ and therefore has an absolute
maximum in between.10 Therefore, besides the liquid solution with M∞ = 0 and
D = ∞ that always satisfies Eq. (3.59), no additional solution can be found if
ρ < ρd (β), where
d
= max F1 (D ;β), (3.62)
ρd (β)Vd D

as illustrated in Figure 3.2. This condition allows one to determine the dynamical
transition density ρd (β) for each temperature. For ρ > ρd (β), Eq. (3.60) admits at
least two solutions, which correspond to the intersections of the function F1 (D ;β)
with a horizontal line at level d/(ρVd ). For reasons that will be better understood
from the discussion of Chapter 4, the solution that correctly describes the arrested
glass phase is that with the smaller D.
Eq. (3.62) can be solved in any spatial dimension d to obtain the dynamical tran-
sition, but in low d, it overestimates the glass transition because the rich structure
of the liquid is neglected11 by the approximation g(r) = exp[−βv(r)]. In general,

10 For some particular potentials, there can be multiple local maxima, leading to multiple glass phases and even
to glass–glass transitions [240, 325]. We will not further discuss this possibility in this book.
11 Better approximation schemes that give more reasonable estimates in d = 3 have been discussed in [241].
Eqs. (3.60) and (3.62) also provide a quantitative, although approximate, description of a particular model,
the Mari-Krzakala-Kurchan model [242, 243], where distances between particles are randomly shifted to
induce a mean field like interaction in finite d. See [90] for a detailed discussion.
86 Atomic Liquids in Infinite Dimensions: Equilibrium Dynamics

F1 (Δ)

Figure 3.2 The function F1 () = limd→∞ F1 (D = 2 /d) corresponding to the

hard-sphere potential (for which the dependence on β is absent). The horizontal
dashed line corresponds to 1/) ϕd with )ϕd = 4.8067 . . . For )
ϕ<)ϕd , Eq. (3.66) has
no solution at finite . At )
ϕ = ) ϕ coincides with the maximum of F1 ()
ϕd , 1/)
and a solution appears, indicating dynamical arrest. For ) ϕ > )
ϕd , there are two
solutions of Eq. (3.66). The plateau of the mean square displacement is the smaller
of the two.

one expects that Eq. (3.62) becomes exact only in the limit d → ∞ [239]. We then
consider the limit d → ∞ for the class of potentials introduced in Section 2.3.2,
such that Eq. (2.74) holds – i.e., v(r) = v̄[d(r/ − 1)] with a typical interaction
scale . The numerical solution [241] indicates that the dynamical transition density
and plateau, expressed in terms of adimensional quantities, have the following
scaling when d → ∞:
 d
 d)
ϕ 2 
ϕ = ρVd = d, D= , (3.63)
2 2 d
where )ϕ and  are finite quantities. This scaling can also be checked by proving
that Eq. (3.62) has a finite limit when d → ∞ if and only if Eq. (3.63) is obeyed.
Plugging Eq. (2.74) and Eq. (3.63) in Eq. (3.61) and using the asymptotic properties
of Bessel functions (details can be found in [292]), one obtains that q(D ,β;r) tends,
for h = d(r/ − 1), to
 ∞ (z−h)2
e− 2 −β v̄(z+/2)
q(,β;h) = dz √ e . (3.64)
−∞ 2π
 3.3 Dynamical Glass Transition 87

From this, one can show that F1 (D ;β) tends to

 ∞
∂q(,β;h)
F1 (;β) = − dh eh log[q(,β;h)] , (3.65)
−∞ ∂

and Eq. (3.62) becomes

1
= max F1 (;β). (3.66)
)
ϕd (β) 

For the hard-sphere potential, for which β is irrelevant, this equation gives a
dynamical transition at )ϕd = 4.8067 . . ., above which the dynamics is arrested (see
Figure 3.2). A more detailed discussion and results for other potentials are given in
Chapter 4.
Note that the dynamical transition happens when the scaled packing fraction
ϕ = 2d ϕ = d) ϕ is proportional to d, which according to the results of Chapter 2,
belongs to the region in which the liquid thermodynamics is well defined and given
by the ideal gas result plus the first virial correction.

3.3.3 Large-Dimensional Scaling of the Dynamical Equations

Having determined the correct scaling of the density and of the mean square dis-
placement in the limit d → ∞, in Eq. (3.63), we can write the full dynamical
equations, Eq. (3.39), in this same limit. We begin by representing the distance
vector between two particles in ‘polar coordinates’ as r(t) = r(t)r̂(t), where r(t)
= |r(t)| is the modulus and r̂(t) = r(t)/r(t) is an angular unit vector. One can
then show that in the limit d → ∞, under the appropriate scaling, the angular
vector remains constant on the relevant time scales and that the dynamics can then
be reduced to a one-dimensional equation for the modulus, or more precisely, the
variable h(t) = d[r(t)/ − 1].
The stochastic Eq. (3.39) can be written in the form

ζ
ṙ(t) = G(r(t)) + (t),
2
 (3.67)
β t 
G(r(t)) = F (r(t)) − dt M(t − t  )ṙ(t  ).
2 0

To write equations for r(t) and r̂(t), we use the well-known Itô’s formula (or
lemma), which provides the correct way of performing a change of coordinates
in a stochastic process with white noise [159, section 4.3.3], to change to polar
coordinates. Eq. (3.67) then becomes
88 Atomic Liquids in Infinite Dimensions: Equilibrium Dynamics
ζ (d − 1)T
ṙ(t) = + r̂(t) · [G(r(t)) + (t)], (3.68)
2 r(t)
ζ dr̂(t) (d − 1)T 1
=− 2
r̂(t) + [G(r(t)) + (t)] (3.69)
2 dt r(t) r(t)
1
− r̂(t)r(t) · [G(r(t)) + (t)].
r(t)2

Note that because the norm of r̂(t) is constant, one has r̂(t) · ṙ(t) = ṙ(t).
Furthermore,

∂v(|r|) d
r̂ · F (r) = −r̂ · = −v  (r) = − v̄  (h), (3.70)
∂r 

Using these relations, from Eq. (3.68) we can derive the equation for h(t)
= d[r(t)/ − 1] at leading order in 1/d,
 t
ζ d β d
ḣ(t) = {T − v̄  [h(t)]} − dt  M(t − t  )ḣ(t  ) + (t), (3.71)
2d  2d 0 

having defined (t) = (/d)r̂ · (t), which is the projection of the noise over a
given coordinate (rescaled by d for a reason that will become immediately clear).
Its correlation is deduced from Eq. (3.39):

  2 2
(t)(t  ) = 2T 2 ζ δ(t − t  ) + 2 M(t). (3.72)
2d 2d

Eqs. (3.71) and (3.72) show that in order to obtain a finite limit when d → ∞, one
needs to define rescaled variables,

2 2
)
ζ = ζ, M(t) = M(t). (3.73)
2d 2 2d 2

The scaling of the friction coefficient ζ is needed for the derivative ḣ(t) and force
v̄  (h) terms to be of the same order in Eq. (3.71); the scaling of the memory
function M(t) is needed to keep the two noise terms of the same order in Eq. (3.72).
Eq. (3.71) then becomes
 t
)
ζ ḣ(t) = T − v̄  [h(t)] − β dt  M(t − t  )ḣ(t  ) + (t),
0 (3.74)
 
(t)(t ) = 2T )
 
ζ δ(t − t ) + M(t − t ), 
 3.4 Critical Properties of the Dynamical Glass Transition 89

where all quantities remain finite in the limit d → ∞. Having defined the scaling
of ζ and M(t) for large d, one can also check that Eq. (3.69) gives dr̂(t) dt
∝ d1 , which
implies that r̂(t) is constant in time (for finite t).12
Finally, we can write in this limit the self-consistent equation that determines the
memory kernel, Eq. (3.42). Under the assumption that r̂(t) is constant, the angular
direction of F [r(t)], which is parallel to r̂(t), is also constant. Using Eq. (3.70),
we get

2 ρ
M(t) = 3 dr0 e−βv(|r0 |) F (r(t)) · F (r0 )
2d
 (3.75)
ϕ ∞
) 
h0 −β v̄(h0 ) 
 
= dh0 e v̄ [h(t)] v̄ (h0 ).
2 −∞
The two equations, Eqs. (3.74) and (3.75), provide a self-consistent equation for
M(t) that holds in the limit d → ∞, with the appropriate scaling of all relevant
quantities. Finally, Eq. (3.49), which relates the memory kernel with the mean
square displacement, can be immediately rescaled by defining, consistently with
Eq. (3.63),
 t
2 ˙ ˙  ) + T , (3.76)
D(t) = (t) ⇒ )
ζ (t) = −β dt  M(t − t  )(t
d 0

which allows one to compute (t) from M(t). Note that Eqs. (3.74), (3.75) and
(3.76) have also been derived via an exact solution of the dynamics in [239].

3.4 Critical Properties of the Dynamical Glass Transition

We conclude this chapter with a brief discussion of the critical behaviour of dynam-
ical correlators upon approaching the glass transition from the liquid phase.

3.4.1 Stress Correlations, Viscosity and Stokes–Einstein Relation

The shear viscosity of the liquid can be deduced from the autocorrelation function
of the pressure tensor13 via a Green–Kubo relation; see [83, 186] for details. In the
overdamped m → 0 limit, the pressure tensor (at zero momentum) is defined by

12 The proof goes, schematically, as follows. Assuming dr̂(t) ∝ 1 , then

dt d

dr̂(t) 1
ṙ(t) = r(t) + r̂(t)ṙ(t) ∝ .
dt d
As a consequence of this scaling and of the scaling in Eq. (3.73), G(t) ∝ d and (t) ∝ d. The right-hand side
of Eq. (3.69) is then proportional to d, while the left-hand side, with )
ζ ∝ d 2 , is proportional to d 2 dr̂(t)
dt . This
proves self-consistently the original assumption.
13 The stress tensor, commonly used in mechanics, is the negative of the pressure tensor. In the overdamped
limit, its kinetic contribution can be neglected.
90 Atomic Liquids in Infinite Dimensions: Equilibrium Dynamics
1  rij μ (t)rij ν (t) 
μν (t) = − v (rij (t)), (3.77)
2 i=j rij (t)

where rij = xi − xj is the interparticle distance. Denoting by μ = ν two distinct

(and otherwise arbitrary, by isotropy of the liquid) spatial components of this tensor,
the shear viscosity is then given by

β ∞  
ηs = dt μν (t)μν (0) . (3.78)
V 0

In the limit d → ∞, following similar steps as in Section 3.2.1, one obtains

1 2
 1  rij μ (t)rij ν (t)  rklμ (0)rklν (0) 
μν (t)μν (0) = v (rij (t)) v (rkl (0))
4 i=j,k=l rij (t) rkl (0)
1 2
1  |rij (t) · rij (0)|2  
∼ 2 v (rij (t))v (rij (0)) (3.79)
2d i=j rij (t) rij (0)
 1 2
ρ 2V |r(t) · r0 |2  
= dr0 g(r0 ) v (r(t))v (r0 ) ,
2d 2 r(t) r0

where, in the second line, it has been assumed that only pairs ij  = kl or ij 
= lk are correlated, and an average has been taken over all directions μν (includ-
ing μ = ν, which gives subdominant 1/d corrections), while in the third line, the
average over the many-body Langevin dynamics has been replaced by an average
over the effective process defined in Section 3.2.2.
Finally, in Section 3.3.3, it was shown that in the limit d → ∞, the unit vector
r̂(t) is constant on the relevant time scales, and dynamics happens only in the
direction parallel to it. Also, at leading order, r(t) ∼ , with corrections of order
1/d. As a consequence, |r(t) · r0 |2 /(r(t) r0 ) ∼ 2 , and

1   ρ 2 2  
μν (t)μν (0) ∼ dr0 g(r0 ) v  (r(t))v  (r0 )
V 2d 2
(3.80)
2 M(t)
∼ρ = ρ d M(t),
2d

using the expression for M(t) in Eq. (3.42), under the same constant r̂(t) hypoth-
esis. In the limit d → ∞, the stress correlation thus coincides with the memory
kernel, and
 ∞  ∞
βρ2
ηs = dt M(t) = βρ d dt M(t). (3.81)
2d 0 0
 3.4 Critical Properties of the Dynamical Glass Transition 91

Eq. (3.51) then becomes a relation between the diffusion coefficient and the shear
viscosity,
T T
D= ∞ = . (3.82)
ζ +β 0 dt M(t) ζ+ 2d ηs
ρ2

At low densities, where ηs → 0 and D → T /ζ , one recovers the ideal gas (over-
damped Langevin) kinetics. Upon approaching the glass transition, M(t) develops
an infinite plateau. The viscosity thus diverges, ηs → ∞, the diffusion coefficient
vanishes, D → 0, and one obtains an effective Stokes–Einstein relation [120] of
the form
T 2 ρ T 2d d−2 2π d/2
Dηs = = , ζeff = = . (3.83)
2d ζeff ρ2 )
ϕ (d/2 + 1)
This scaling of the effective Stokes drag ζeff is very close to that obtained via a
hydrodynamic treatment [83] and can be also derived via a proper treatment of the
dynamics of a single Brownian particle in a solvent, in the limit d → ∞ [84].

3.4.2 Dynamical Critical Exponents and Dynamical Susceptibility

The structure of the dynamical equations in d → ∞ is qualitatively similar to
those of the mode-coupling theory (MCT) of glasses, which has long been used to
describe dynamical arrest. We refer to the books [168, 175] for a detailed discussion
of this theory.
Like Eq. (3.49), MCT is a set of equations that relate dynamical correlators to
memory kernels, derived in fixed d via a series of approximations of the exact
dynamics. In MCT, however, the memory kernels are expressed as polynomial func-
tions of the correlators. This differs from the exact d → ∞ solution, in which the
memory kernel is self-consistently determined via an average over the effective pro-
cess, as discussed in Section 3.2.2. Quantitative differences thus arise between the
two theories. For example, the dynamical glass transition of hard spheres predicted
by MCT has an incorrect scaling in the limit d → ∞, as ϕdMCT ∝ d 2 /2d [183, 317].
The scaling of the plateau of the mean square displacement and the Stokes–Einstein
relations obtained from MCT also differ from the exact solution in d → ∞. The
qualitative structure of the two theories is nonetheless similar. In particular, the
exact d → ∞ theory displays precisely the same critical scaling of the dynamical
correlators as MCT in the vicinity of the dynamical transition, which is illustrated
in Figure 3.3. A direct derivation of these critical properties from the dynamical
equations is possible within MCT. The same derivation is technically much more
involved in the case of the d → ∞ equations because the equation that deter-
mines the memory kernel is implicit and involves an average over the effective
process. Regardless of the nature of the memory kernel, however, the existence of a
92 Atomic Liquids in Infinite Dimensions: Equilibrium Dynamics

D(t)
t−a tb

t2 D
t
Figure 3.3 Illustration of the critical behaviour of the mean square displacement
in d → ∞, in the liquid phase very close to the dynamical glass transition.
The figure is in log–log scale. At short times, for Newtonian dynamics, one has
D(t) ∼ t 2 , while at very long times one has diffusive behaviour, D(t) ∼ Dt. At
intermediate times, a plateau emerges at D(t) ∼ D. The approach to the plateau
is a power law, D(t) − D ∝ t −a , and the departure from the plateau is also a
power law, D(t) − D ∝ t b . Upon approaching the dynamical glass transition, the
diffusion constant D vanishes, and the plateau extends to infinite times.

long-time plateau simplifies the dynamical equations in this regime, leading to

schematic MCT equations near the plateau [288]. In the following, we only briefly
review the critical scaling results, without derivation.
As an illustration, we consider the hard-sphere case, for which the control param-
eter is the packing fraction ϕ. For other systems, the dynamical glass transition is
generically also controlled by temperature or other parameters. The same scaling
results then hold by replacing ϕ with the appropriate control parameter. As dis-
cussed in Section 3.3.2, for ϕ ≥ ϕd , dynamical correlations have an infinite plateau
at long times, limt→∞ D(t) = D and limt→∞ M(t) = M∞ . The critical regime
is observed when ϕ → ϕd . For ϕ → ϕd− , a very long plateau emerges at time
t ∼ tp with D(t) ∼ D, followed by a diffusive regime with a very small diffusion
constant D. The plateau value D tends continuously to the true, infinite plateau
solution of Eq. (3.60), when ϕ → ϕd− . The mean square displacement D(t) scales
as power law upon approaching and leaving the plateau:

−t −a for tm  t  tp,
D(t) − D ∝ b (3.84)
t for t  tp,

where tm is a microscopic time scale, given by tm ∼ 2 ζ /T for Langevin dynamics

√
and tm ∼  m/T for Newtonian dynamics. The two dynamical critical exponents
a and b that appear in Eq. (3.84) are related by [168]
(1 − a)2 (1 + b)2
λ= = , (3.85)
(1 − 2a) (1 + 2b)
 3.5 Wrap-Up 93

where (x) is the Gamma function. The parameter λ is not universal. It depends on
the system under investigation and is different within MCT and the d → ∞ solu-
tion. A calculation of λ from thermodynamics is also possible [72, 288]. For hard
spheres in d → ∞, one finds λ = 0.70698 . . ., which implies that a = 0.32402 . . .
and b = 0.62915 . . . [221]. The dynamical equations additionally predict a power-
law divergence of the shear viscosity ηs and vanishing of the diffusion constant D
with the same critical exponent γ (as implied by the Stokes–Einstein relation of
Section 3.4.1):
1 1
ηs ∼ D −1 ∼ |ϕ − ϕd |−γ , ϕ → ϕd−, γ = + . (3.86)
2a 2b
For hard spheres, this relation gives γ = 2.33786 . . . [221]. On the other side
of the dynamical transition, when ϕ → ϕd+ , one has D(t) − D ∝ t −a , with the
same exponent a given by Eq. (3.85), but the exponent b is not defined because the
plateau extends to infinite time.
At a standard second-order phase transition, the correlation function of the order
parameter displays a divergent correlation length, and the susceptibility defined by
its volume integral diverges, as discussed in Chapter 1. By analogy, in the case
of the dynamical transition, one can study the dynamical susceptibility defined in
Eq. (3.17), which is the volume integral of a four-point dynamical correlation [37,
56, 146, 204]. It encodes the fluctuations of dynamical correlators, here represented
by the mean square displacement. In the dynamically arrested phase, the long-time
limit of this susceptibility goes to a constant – i.e., limt→∞ χ4 (t) = χ – that diverges
upon approaching the dynamical transition with ϕ → ϕd+ as χ ∼ |ϕ − ϕd |−1/2 .
Approaching the dynamical point from the liquid side instead, χ4 (t) has a peak at
t ∼ 1/D, with χ = χ4 (1/D) similarly diverging [37, 56, 146, 204]. In addition,
the dynamical correlation length associated with χ4 diverges as ξ4 ∼ |ϕ − ϕd |−1/4
near the transition [56, 145, 149].

3.5 Wrap-Up
3.5.1 Summary
In this chapter, we have seen that

• There exist several models of equilibrium dynamics. Examples are Hamiltonian,

Langevin and Monte Carlo dynamics. For these dynamics, if the initial condition
is described by the equilibrium Gibbs–Boltzmann distribution, the same distri-
bution describes the system at all times. The dynamics then satisfies reversibility
and fluctuation–dissipation relations, and the transport coefficients can be written
in terms of equilibrium correlations via Green–Kubo relations (Section 3.1).
• In the limit d → ∞, each particle performs an independent Langevin dynamics,
in the presence of a thermal bath (encoded by effective noise and friction terms)
94 Atomic Liquids in Infinite Dimensions: Equilibrium Dynamics

that describes the average (mean field) interaction with all the other particles.
The memory kernel M(t) that describes the thermal bath is self-consistently
determined as an autocorrelation of the force, in the effective Langevin dynamics
(Sections 3.2.1 and 3.2.2).
• Once M(t) is determined, all the other dynamical observables, such as the mean
square displacement D(t) and the diffusion constant D, can be expressed in terms
of M(t) (Section 3.2.3).
• The effective d → ∞ dynamical equations predict that at a sufficiently high
density or low temperature, the dynamics of the liquid phase becomes arrested.
The memory kernel does not decay to zero in the arrested phase, and the diffusion
constant vanishes. Explicit equations describe the dynamical transition point and
the plateau value of the memory kernel (Section 3.3.1).
• From these equations, one can identify the proper scaling of all the dynamical
quantities for d → ∞. The most important scaling forms are those of the mean
square displacement, D(t) = 2 (t)/d, and of the packing fraction, ϕ = d) ϕ /2d
(Section 3.3.2).
• Using this scaling, the equations can be simplified in the d → ∞ limit and
reduced to a one-dimensional effective Langevin equation. The final result given
by Eqs. (3.74), (3.75) and (3.76) coincides with what has been obtained via an
exact solution of the dynamics in [239] (Section 3.3.3).
• One can also show that the viscosity diverges at the dynamical transition point,
that an effective Stokes–Einstein relation holds in the vicinity of dynamical
arrest and that the dynamical correlations display power-law scalings close to
the plateau, as in the mode-coupling theory of glasses. A dynamical four-point
susceptibility diverges upon approaching the dynamical transition (Section 3.4).

3.5.2 Further Reading

We provide here a list of references that can be consulted to further explore the
subjects discussed in this chapter, selected according to the criteria discussed in
Section 1.6.2.
Introductory reviews on the phenomenology of liquid dynamics close to the glass
transition are

• Debenedetti, Metastable liquids: Concepts and principles [120]

• Ediger, Spatially heterogeneous dynamics in supercooled liquids [135]
• Donth, The glass transition: Relaxation dynamics in liquids and disordered
materials [130]
• Henkel, Pleimling and Sanctuary (eds), Ageing and the glass transition [176]
• Hunter and Weeks, The physics of the colloidal glass transition [181]
• Binder and Kob, Glassy materials and disordered solids: An introduction to their
statistical mechanics [53]
 3.5 Wrap-Up 95

Mode-coupling theory (MCT) is by far the most accurate microscopic theory to

describe liquid dynamics in d = 3. Good introductory reviews to MCT are

• Reichman and Charbonneau, Mode-coupling theory [301]

• Janssen, Mode-coupling theory of the glass transition: A primer [188]
• Götze, Complex dynamics of glass-forming liquids: A mode-coupling theory
[168]
• Götze, Recent tests of the mode-coupling theory for glassy dynamics [169]
In particular, the third reference [168] provides all the technical details of the
theory, while the fourth [169] reviews the numerical and experimental tests of
MCT. A systematic numerical investigation of the supercooled liquid dynamics in
dimensions d = 3, . . . ,12, which focuses in particular on the comparison with
MCT and on the Stokes–Einstein relation, can be found in

• Charbonneau, Ikeda, Parisi et al., Glass transition and random close packing
above three dimensions [86]
• Charbonneau, Ikeda, Parisi et al., Dimensional study of the caging order param-
eter at the glass transition [85]
• Charbonneau, Charbonneau, Jin et al., Dimensional dependence of the Stokes–
Einstein relation and its violation [83]
The complete derivation of the d → ∞ dynamical solution via path integrals is
based on the Martin–Siggia–Rose–De Dominicis–Janssen formalism. Pedagogical
introductions to this formalism can be found in

• Cugliandolo, Dynamics of glassy systems [111]

• Castellani and Cavagna, Spin-glass theory for pedestrians [79]
• Kamenev, Field theory of non-equilibrium systems [200]
Both MCT and the d → ∞ dynamical equations derived in this chapter fall into
a general class of ‘dynamical mean field equations’. These equations describe the
dynamics of a variety of mean field systems and are discussed in references [79,
111]. A general discussion of the connection between MCT and dynamical mean
field equations and the application of dynamical mean field equations to several
regimes of the out-of-equilibrium dynamics of liquids can be found in

• Cugliandolo and Kurchan, Analytical solution of the off-equilibrium dynamics of

a long-range spin-glass model [109]
• Bouchaud, Cugliandolo, Kurchan, et al., Out of equilibrium dynamics in spin-
glasses and other glassy systems [65]
• Berthier, Barrat and Kurchan, A two-time-scale, two-temperature scenario for
nonlinear rheology [36]
• Berthier and Kurchan, Non-equilibrium glass transitions in driven and active
matter [41]
96 Atomic Liquids in Infinite Dimensions: Equilibrium Dynamics

A general discussion of the analogy between dynamical mean field equations and
replica equations, which is used in the calculation of dynamical critical exponents
presented in Section 3.4.2, can be found in

• Kurchan, Supersymmetry, replica and dynamic treatments of disordered systems:

A parallel presentation [218]
• Parisi and Rizzo, Critical dynamics in glassy systems [288]
All these ingredients are behind the solution of the dynamics of particle systems in
d → ∞. Detailed derivations can be found in

• Maimbourg, Kurchan and Zamponi, Solution of the dynamics of liquids in the

large-dimensional limit [239]
• Agoritsas, Maimbourg and Zamponi, Out-of-equilibrium dynamical equations of
infinite-dimensional particle systems [3]
In particular, the second reference [3] contains an extension of both the approach
presented in this chapter, and the path integral derivation, to the out-of-equilibrium
dynamics.
It is well established that the mean field sharp dynamical transition becomes
(at best) a dynamical crossover in d = 2,3. The disappearance of the dynamical
transition is due to fluctuations of different nature, which are not taken into account
by the mean field theory. Theories of the fluctuations around the dynamical transi-
tions have been proposed in

• Schweizer and Saltzman, Entropic barriers, activated hopping, and the glass
transition in colloidal suspensions [320]
• Bhattacharyya, Bagchi and Wolynes, Facilitation, complexity growth, mode cou-
pling, and activated dynamics in supercooled liquids [47]
• Franz, Parisi, Ricci-Tersenghi et al., Field theory of fluctuations in glasses [145]
• Rizzo and Voigtmann, Qualitative features at the glass crossover [304]
• Janssen and Reichman, Microscopic dynamics of supercooled liquids from first
principles [189]
• Charbonneau, Jin, Parisi et al., Hopping and the Stokes–Einstein relation break-
down in simple glass formers [90]
At temperatures well below the dynamical transition, dynamical relaxation is
believed to be due to activated events, and many glass forming materials show
some sort of Arrhenius (or modified Arrhenius) behaviour of their relaxation time.
The reviews mentioned at the beginning of this section provide a good introduction
to this dynamical regime; additional references are given in Section 4.5.2.
 3.6 Appendix: Reversibility for Langevin Dynamics 97

3.6 Appendix: Reversibility for Langevin Dynamics

We briefly discuss here how to prove reversibility at equilibrium for a Langevin
equation. In order to simplify the notation, we consider a one-dimensional example
(a single particle at position x evolving in a potential V (x)); hence, the overdamped
Langevin equation is
dV (x)
ẋ = − + ξ(t), ξ(t)ξ(t  ) = 2T δ(t − t  ). (3.87)
dx
The probability that the particle be at x at time t, P (x,t), then satisfies the Fokker-
Planck equation [159]
 
d dV (x) dP (x,t) d2 P (x,t)
Ṗ (x,t) = +T . (3.88)
dx dx dx dx 2
If we choose an initial condition in equilibrium with probability

P (x,0) = Peq (x) ∝ e−βV (x), (3.89)

we can check that Ṗ (x,0) = 0, as expected because the equilibrium distribution is a

stationary solution of Eq. (3.88). We could also introduce the transition probability
T (x,y;t), defined by

P (x,t) = dy T (x,y;t)P (y,0), (3.90)

which is the probability of finding the particle at x at time t, provided that it is

at y at time 0. The function T (x,y;t) is not symmetric as a consequence of the
microscopic irreversibility.
To prove
 reversibility at equilibrium, it is convenient to introduce the function
ρ(x) = Peq (x) and the function Q(x,t) defined by

ρ(x)Q(x,t) = P (x,t). (3.91)

From Eq. (3.88), it follows that

Q̇(x,t) = −HQ(x,t),
 2 (3.92)
d2 1 dV (x) 1 d2 V (x)
H = −T 2 + W (x), W (x) = − .
dx 4T dx 2 dx 2
The operator H is self-adjoint; hence, we have

Q(x,t) = dy G(x,y;t)Q(y,0), (3.93)
98 Atomic Liquids in Infinite Dimensions: Equilibrium Dynamics

with a symmetric G(x,y;t) = G(y,x;t). Comparison of Eq. (3.90) and Eq. (3.93)
leads to
1 1
G(x,y;t) = T (x,y;t)ρ(x)/ρ(y) = e− 2 βV (x) T (x,y;t)e 2 βV (y) . (3.94)
We finally arrive to the needed expression of the equilibrium correlations:

O1 (x(t))O2 (x(0)) = dxdyO1 (x)T (x,y;t)O2 (y)Peq (y)
 (3.95)
= dxdyρ(x)O1 (x)G(x,y;t)O2 (y)ρ(y),

which is clearly symmetric under the exchange of O1 and O2 .

1 1
Notice that Eq. (3.94) implies that e− 2 βV (x) T (x,y;t)e 2 βV (y) is a symmetric
function, which is a sort of balance equation. The detailed balance Eq. (3.8) can
also be written as
1 1 1 1
e− 2 βV (Xt+1 ) P (Xt+1 |Xt )e 2 βV (Xt ) = e− 2 βV (Xt ) P (Xt |Xt+1 )e 2 βV (Xt+1 ) . (3.96)
 4
Thermodynamics of Glass States

In Chapter 3, we have seen that simple liquids in the limit d → ∞ generically

undergo a dynamical phase transition to an arrested phase upon lowering tempera-
ture or increasing density. In this phase, the long-time limit of the mean square dis-
placement is finite. In other words, the system is trapped into a restricted portion of
phase space, selected by the initial condition, and is unable to diffuse away from it.
This restricted portion of phase space defines a ‘glass state’. We can then invoke
an ergodic hypothesis, restricted to a single glass state, in order to convert the
dynamical long time average of observables within the glass into an appropriate
thermodynamical average. In this way, the long time behaviour can be described by
a thermodynamic framework called the ‘state following’ formalism. This construc-
tion requires the introduction of replicas of the original system, which brings about
a new symmetry into the problem – namely the permutation symmetry between
replicas, or ‘replica symmetry’. In this chapter, we introduce the state following
construction, we obtain the main formulae that describe the restricted glass thermo-
dynamics, and we derive the phase diagram for standard models of glasses within
the simplest scheme, in which replica symmetry is unbroken.

4.1 Arrested Dynamics and Restricted Thermodynamics

In Chapter 3, the equations that describe the dynamics of simple liquids in high
dimensions have been discussed. In the liquid phase, close to the dynamical
transition, the equilibrium dynamics happens on two well-separated time scales
(Section 3.4.2): at short times, particles vibrate around an amorphous glassy
structure, while at long times, they diffuse away from the initial structure. In the
infinite-dimensional limit, upon approaching the dynamical transition, the diffusion
constant D vanishes as a power law, given in Eq. (3.86). Beyond the dynamical
transition, diffusion is arrested (D = 0), and the mean square displacement reaches
a finite plateau in the long-time limit.

99
100 Thermodynamics of Glass States

In a nutshell, the dynamical arrest corresponds to the formation of a metastable

state [204, 205, 206, 207, 208]. We have discussed metastability for the ferromagnet
in Chapter 1, but the physics is different here. In ferromagnets, there are only
one equilibrium state and one metastable state, while in glasses, there is a quite
large number of states (exponentially large in the number of particles, as we shall
see later). The aim of this chapter is to generalise the discussion of metastability
presented in Chapter 1 in order to recover and extend the results of Chapter 3,
without solving the dynamics of the system but remaining in the framework of
the Gibbs–Boltzmann statistical mechanics, in which no explicit reference to the
dynamics is made.
Studying dynamical arrest without solving the dynamics may look like a con-
tradiction, but it is not. In fact, we have already seen in Section 1.5 that, for fer-
romagnets in the d → ∞ limit, the metastable state with positive magnetic field,
B > 0, and negative magnetisation, m < 0, can be obtained by (1) studying the
system in equilibrium with the standard Gibbs–Boltzmann distribution at B < 0, (2)
computing the observables we are interested in, (3) and then analytically continuing
their values to B > 0. In other words, metastable states can always be reached by
starting from a stable equilibrium state and changing adiabatically (i.e., analyti-
cally continuing on) a control parameter. This scheme thus provides a well-defined
statistical mechanics procedure to compute observables in a metastable state.1
We have thus shown that the construction of statistical mechanics for metastable
states is based on adding an appropriate term in the Hamiltonian that stabilises the
metastable state and then gradually removing it. Of course, different metastable
states could be reached by changing the form of the Hamiltonian. To study
dynamical arrest, we want to study the metastable states in which the system
explores a small region of phase space around an initial equilibrium liquid
configuration [144, 205, 206, 252, 260]. To construct these states, we can add a new
term in the Hamiltonian that increases the energy of configurations that are too far
away from the initial configuration. If this new term is sufficiently large, in the new
equilibrium the system is confined around the initial configuration. We can then
perform bona fide equilibrium statistical mechanics computations. When we
remove the extra term, the system remains trapped in the metastable state defined
by this procedure.
Note that one can choose as a starting point an equilibrium configuration at a
given temperature and repeat the aforementioned procedure at another temperature.

1 One should keep in mind, however, that the analytical continuation is only well defined in d → ∞, while
in finite dimensions, an essential singularity appears at B = 0 [282]. This singularity is weak enough that the
analytical continuation remains quite unambiguous if B remains sufficiently close to zero.
 4.1 Arrested Dynamics and Restricted Thermodynamics 101

This procedure is called ‘state following’ [144, 214]. It allows one to compute
physical observables of a glass phase prepared by cooling or compressing an ini-
tial equilibrium configuration. These physical ideas are quite simple, and we shall
implement them in the rest of the chapter, hoping that the technical difficulties do
not obscure their intrinsic simplicity.
As in Chapters 2 and 3, we consider here a system of N identical particles,
enclosed in a volume V , with number density ρ = N/V , with positions X = {xi },
interacting via a potential V (X) with a typical interaction scale  and strength ε.
We consider the thermodynamic limit N → ∞ and V → ∞ at constant ρ, in
which the boundary conditions become irrelevant, and there are only two adimen-
sional control parameters, the scaled temperature T /ε and the packing fraction
ϕ = ρVd (/2)d defined in terms of the interaction range  (see Section 2.3.2).
In the following, we will be interested in comparing two configurations X and
Y at different density. It will be convenient to take N and V (and thus ρ) to be
the same for the two configurations so that particle positions, xi and yi , can be
directly compared.2 A change in packing fraction, with constant N and V , can
be conveniently achieved by changing the interaction scale , as we will do in the
following, while for temperature, it is simpler to vary T at fixed ε.

4.1.1 Arrested Dynamics in the Glass Phase

Consider the dynamics starting from an equilibrium configuration Y ≡ {yi },
extracted from the Gibbs–Boltzmann distribution, at packing fraction ϕg (with
interaction scale g ) and temperature Tg that fall into the dynamically arrested
region – i.e., Tg < Td (ϕg ). We call X(t) ≡ {xi (t)} the configuration of the system
at time t, such that X(0) = Y . The mean square displacement

1 
N
D(X,Y ) = |xi − yi |2 , (4.1)
N i=1

is a natural measure of similarity between two configurations X,Y .

We first consider the case in which the dynamics is run at the same state point
(ϕg,Tg ) at which the initial condition is prepared, as illustrated in Figure 4.1. In a
dynamically arrested phase, the diffusion coefficient vanishes, and
 
lim D(X(t),Y ) = Dr , (4.2)
t→∞

2 In fact, if the volume is changed, the particle coordinates should be rescaled accordingly, and if the particle
number is changed, no one-to-one correspondence between the xi and the yi is possible.
102 Thermodynamics of Glass States
Mean Square
Phase Space Real Space
Displacement

D(t)
Δ(t)
∼ Dt
Liquid

∼ t2
t
ϕd
Δ(t)
D(t)
Normal Glass

D
∼ t2 D
Δ1

Figure 4.1 At low densities or high temperatures (top row), the phase space is
composed by only one ergodic component, which coincides with the liquid state.
For Newtonian dynamics, the mean square displacement behaves ballistically at
short times and diffusively at long times (here, in log-log scale). Beyond the
dynamical transition (bottom row), the phase space clusters in a set of metastable
glass states, as a consequence of ergodicity breaking. The long time limit of the
mean square displacement is then finite, which signals that particles are caged by
their neighbours.

where Dr is a finite constant,3 and the brackets represent either a dynamical noise
average, for Brownian dynamics, or an average over a large enough time window
around t, for Newtonian dynamics. By contrast to the liquid phase, in which par-
ticles diffuse away from their initial positions and D(X(t),Y ) grows with time,
in the dynamically arrested region, the system remains close, at all times, to its
initial starting point. Furthermore, because the initial condition Y is sampled at
equilibrium, the dynamics is, on average, time-translationally invariant
D(X(t + τ ),X(t)) = D(τ ), ∀t > 0 ; lim D(τ ) = D = Dr . (4.3)
τ →∞

The behaviour of the system at long times is therefore described by a well-defined

quantity, Dr , that represents both the mean square displacement between X(t) and
Y = X(0) at large t and the mean square displacement between two configurations

3 In equilibrium, as it will be shown below, D = D, so one recovers the notation of Chapter 3.

r
 4.1 Arrested Dynamics and Restricted Thermodynamics 103

X(t), X(t + τ ), separated by a long time window τ → ∞. A priori, the value of

Dr depends on the initial configuration Y . However, Dr (Y ) is an intensive quantity,
and its fluctuations obey the central limit theorem4
Dr (Y ) = Dr + O N −1/2 ; (4.4)
hence, the dependence on Y disappears in the thermodynamic limit. A quantity with
this property is said to be ‘self-averaging’ [254]. Note that here and in the rest of
this chapter, an overbar denotes an average over the initial configuration Y , sampled
from a Gibbs–Boltzmann distribution at (ϕg,Tg ).
We now generalise the previous discussion to the case in which the system starts
from an initial configuration Y , prepared in equilibrium at (ϕg,Tg ) as before, but
evolves at a different state point (ϕ,T ) = (ϕg,Tg ). This procedure corresponds
to preparing a glass state at some temperature and packing fraction and then
instantaneously bringing that glass to another temperature and packing fraction,
where it is then kept.5 If the new state point (ϕ,T ) is such that the system remains
in the glass phase, the diffusion constant still vanishes, and Eq. (4.2) still holds.
Under the assumption that there is no phase transition that intervenes during the
process of glass cooling or compression (this condition is not always satisfied,
as will be discussed in Chapter 6), several additional properties are verified. First,
one has
lim D(X(t + τ ),X(t)) = D(τ ) ; lim D(τ ) = D = Dr . (4.5)
t→∞ τ →∞

In other words, in the long time limit, the dynamics at the new temperature and
packing fraction becomes stationary and is described by two quantities Dr and D,
which correspond respectively to the average mean square displacement between
Y = X(0) and X(t) for large t, and between two configurations X at two very
different and both very large times. This situation is illustrated in Figure 4.2.
Second, the fluctuations of both D and Dr over the initial condition Y fol-
low a central limit theorem analogous to Eq. (4.4), and, hence, D and Dr are
self-averaging. Third, the long time properties are independent of how the system
has been brought from (ϕg,Tg ) to (ϕ,T ). We described earlier an instantaneous
change of the control parameters, but any other process (e.g., a slow cooling
from Tg to T ) would give the same results for times much larger than the
preparation time.

4 Unless there are long-range correlations in the system.

5 An instantaneous quench can be realised, in practice, by using Langevin dynamics with noise at temperature
T = Tg , or by using Newtonian dynamics with the initial velocities being rescaled by an appropriate factor to
bring the system from Tg to T . An instantaneous compression is realised by preparing the system with a
potential with length scale g , and running the dynamics with a different potential with  = g .
104 Thermodynamics of Glass States

(ϕg , Tg )

(ϕ, T )
Y
Dr
Y
X(t)
D

Dg

Figure 4.2 A schematic picture of the state following procedure. A glass state is
selected by the configuration Y that is extracted in equilibrium at (ϕg ,Tg ). The
configuration Y can thus be regarded as a pinning field for selecting a particular
metastable state, as discussed in Chapter 1. (Left) If Y is dynamically evolved
at the same state point (ϕg ,Tg ), then the long time limit of the mean square
displacement is Dg . (Right) If the dynamics is instead evolved starting from Y
but at different (ϕ,T ), the state selected by Y at (ϕg ,Tg ) (light grey in the right
panel) evolves as X(t) into a slightly different state (dark grey). The long time
limit of the mean square displacement between Y and X(t) is Dr while the typical
distance between two configurations separated by a long time window is D.

4.1.2 Restricted Thermodynamics of the Glass State

The idea of using a thermodynamic formalism to describe dynamical arrest in
liquids dates back to the work of Kirkpatrick, Thirumalai and Wolynes [204, 205,
206, 329] and is at the root of the random first-order transition approach to the
glass transition [40, 80, 208, 357]. The ‘state following construction’, which we
introduce in this section, is based on this idea and was originally developed by
Franz and Parisi in the context of spin glasses in [28, 144], where the equivalence
with the long time limit of the dynamics was also discussed [27]. It is also called
the ‘Franz–Parisi construction’, and it has been adapted to particles in d → ∞ in
[59, 299].
The method is based on the following idea. In equilibrium statistical physics, the
long time limit of dynamical observables can be computed (in the thermodynamic
limit) using a probabilistic approach that relies on the use of appropriate statistical
ensembles [158, 310]. In equilibrium, entropy is maximised under the constraint
that the total energy of the system takes a prescribed value. This leads to the
Gibbs–Boltzmann distribution of the canonical ensemble [158]. By analogy, the
 4.1 Arrested Dynamics and Restricted Thermodynamics 105

state following construction aims to describe metastable glass states by a probability

distribution that maximises entropy under a minimal set of constraints.
In Section 4.1.1, it was shown that in the dynamically arrested phase, the dynam-
ics starting from an equilibrium configuration Y remains confined to a portion of
phase space around the initial configuration (Figure 4.1). While the configuration
Y selects one of the possible glass states (Figure 4.1), at large times, the config-
uration X(t) ergodically samples the restricted region of phase space defined by
 
D(X,Y ) = Dr . This constraint defines a region of phase space that we call a
‘glass state’. We further assume that the stationary long time average of dynamical
observables coincides with a restricted statistical average over this glass state. The
simplest assumption to construct such a restricted statistical average is that, besides
the energy, the only other constraint to be enforced concerns the mean square
displacement. We thus seek a probability distribution over X that maximises the
entropy under the two constraints of fixed energy and fixed mean square displace-
ment from Y . Because the additional constraint does not affect the kinetic energy,
the distribution of the momenta remains Maxwellian; hence, we omit it in the
following. Focusing on the configurational part, we obtain a generalisation of the
canonical distribution
P (X|ϕ,β;Y,λ) ∝ e−βV [X,]−λ D(X,Y ), (4.6)
in which the dependence of V [X,] on  is indicated explicitly to highlight the
dependence of the probability on the packing fraction. Here the inverse temperature
β is a Lagrange multiplier coupled to the energy, while λ is coupled to the mean
square displacement. Just like β is determined by requiring a given average energy,
λ is determined by requiring that the average of D(X,Y ) coincides with Dr . Equiv-
alently, one can use a different ensemble, in which the mean square displacement
constraint is strictly enforced:
1
P (X|ϕ,β;Y, Dr ) = e−βV [X,] δ(Dr − D(X,Y )),
Z[ϕ,β;Y, Dr ]
 (4.7)
Z[ϕ,β;Y, Dr ] = dXe−βV [X,] δ(Dr − D(X,Y )).

where the factor Z[ϕ,β;Y, Dr ] is a ‘restricted configurational integral’ that gen-

eralises Eq. (2.2) to the restricted equilibrium construction. The two ensembles
defined in Eqs. (4.6) and (4.7) are equivalent in the thermodynamic limit for the
same reason that the canonical and microcanonical ensembles are equivalent [158].
The free energy of the glass state selected by Y and brought to (ϕ,T ) is then
given by
T
f(ϕ,T ;Y, Dr ) = − log Z[ϕ,β;Y, Dr ]. (4.8)
N
106 Thermodynamics of Glass States

Here again, this free energy fluctuates with the configuration Y but is self-averaging
in the thermodynamic limit, in which it does not fluctuate unless the system under-
goes a phase transition [254]. The average of the free energy over Y is thus repre-
sentative of the typical value of f(ϕ,T ;Y, Dr ) and is given by

fg (ϕ,T ;ϕg,Tg, Dr ) = f(ϕ,T ;Y, Dr )

 (4.9)
T dY
=− e−βg V [Y,g ] log Z[ϕ,β;Y, Dr ],
N Z[ϕg,βg ]
where (ϕg,Tg ) are the parameters
 that control the distribution of the initial condi-
tion Y , and Z[ϕg,βg ] = dY e−βg V [Y,g ] is the standard configurational integral at
(ϕg,Tg ). This expression defines the ‘Franz–Parisi potential’, or ‘glass free energy’.
It represents the free energy of a typical glass state prepared at (ϕg,Tg ) and followed
to the state point (ϕ,T ), for a given choice of Dr .
Note that the glass free energy depends explicitly on Dr , which was introduced as
a dynamical quantity in Eq. (4.2). In principle, one should take Dr from the solution
of the dynamics and use it in the calculation of the glass free energy. This approach
is, however, unsatisfying. One would like to have a fully thermodynamical compu-
tation of the glass free energy without having to solve the dynamics first. To resolve
this problem, one can use the general principles of statistical mechanics [223]. The
constrained equilibrium free energy, expressed as a function of Dr , can be thought
of as a large deviation function for the fluctuations of Dr in an unconstrained
ensemble. This free energy should then be minimised over Dr . Because the global
minimum is found at Dr → ∞ and corresponds to the liquid state, to describe the
glass, one should instead choose the local minimum at finite Dr (see Figure 4.3 for
an example). In other words, the proper value of Dr in Eq. (4.2) is a local minimum
of the Franz–Parisi potential at finite Dr , keeping all other parameters fixed.
In fact, as we show explicitly in the rest of this chapter, at high T and low ϕ, the
Franz–Parisi potential has a unique minimum when Dr → ∞. This corresponds
to the liquid phase in which the dynamics is not arrested. When T is low enough
and ϕ is high enough (provided the preparation values ϕg and Tg fall in the dynam-
ically arrested region), a new local minimum of the Franz–Parisi potential appears
at finite Dr . We will show that the value of Dr at this local minimum precisely
corresponds to the value of Dr selected by the dynamics. One thus concludes that
the free energy of a typical glass state prepared at (ϕg,Tg ) and followed to (ϕ,T ) is
given by

fg (ϕ,T ;ϕg,Tg ) = min fg (ϕ,T ;ϕg,Tg, Dr ), (4.10)

Dr :Dr <∞

which provides a fully thermodynamical expression for the glass free energy, with-
out any reference to the dynamics.
 4.1 Arrested Dynamics and Restricted Thermodynamics 107

The structure of Eq. (4.9) is also interesting. One has first to compute the free
energy of the system X, which depends on the configuration Y , and then average
the free energy over Y . One can think to the system X as feeling an external disor-
dered potential, due to the configuration Y , that is fixed in time (or ‘quenched’).
The configuration X evolves in presence of this external potential and samples
its equilibrium distribution. One then averages the free energy over the disorder,
represented by Y . In other words, the configuration Y acts as a quenched disorder
for the system X. This formulation therefore reduces to a standard problem in the
statistical physics of disordered systems [254], and the techniques developed in this
field can then be used in the context of structural glass problems without quenched
disorder, as we now describe.

4.1.3 The Replica Method

It is usually very difficult to average the logarithm of the disorder-dependent free
energy [254]. Disorder indeed explicitly breaks translational invariance, which is a
crucial symmetry for most theoretical physics methods, from mean field methods
to perturbative expansions. In the case of Eq. (4.9), more specifically, one cannot
apply the virial expansion to the Y -dependent free energy f(ϕ,T ;Y, Dr ), because
of the presence of the disordered space-dependent external potential.
The replica method is designed to solve this problem. It makes use of the simple
identity log x = lims→0 ∂s x s , using the shorthand notation ∂s = d/ds. If x is a
random variable, then
log x = lim ∂s x s . (4.11)
s→0

Applying this identity to Eq. (4.9), we get


T dY
fg (ϕ,T ;ϕg,Tg, Dr ) = − lim ∂s e−βg V [Y,g ] Z[ϕ,β;Y, Dr ]s . (4.12)
s→0 N Z[ϕg,βg ]
Computing this expression for a general value of s is as difficult as computing the
logarithm, but a big simplification occurs if s is an integer. We can then write
s+1 

dXa e−βV [X ,] δ Dr − D(X a,Y ) ,
a
Z[ϕ,β;Y, Dr ]s = (4.13)
a=2

and, defining X = Y , we obtain

1

T 1
fg (ϕ,T ;ϕg,Tg, Dr ) = − lim ∂s Zs+1 [ϕa,βa, Dr ], (4.14)
N s→0 Z[ϕg,βg ]
 .
s+1
/ . s+1

/
dX a e−βa V [X ,a ]
a
Zs+1 [ϕa,βa, Dr ] = δ Dr − D(X a,X 1 ) ,
a=1 a=2
108 Thermodynamics of Glass States

where β1 = βg , ϕ1 = ϕg , βa≥2 = β and ϕa≥2 = ϕ. We thus have to compute the

configurational integral of a system of s + 1 replicas of the original system, such
that all replicas a ≥ 2 are coupled to replica 1 by the mean square displacement
constraint. This constraint imposes that particle i in each replica must be close
to particle i in all other replicas, xai ∼ xbi for all a,b. The s + 1 particles xai
for each given i thus form a ‘molecule’, bound by the mean square displacement
constraint, and Zs+1 [ϕa,βa, Dr ] corresponds to the configurational integral of a
‘molecular liquid’.
The fundamental simplification introduced by the replica method for each integer
value of s is that the molecular liquid is translationally and rotationally invariant.
A global translation of all replicas, xai → xai + X, leaves the partition function
invariant,6 as does a global rotation. This symmetry allows us to use the virial
expansion and the methods introduced in Chapter 2 to compute Zs+1 [ϕa,βa, Dr ]
for each integer s. It is important to stress that, in principle, even the molecular
liquid could have glassy or crystalline phases in which translational invariance is
spontaneously broken, but we are not interested in such phases. Averaging over all
equilibrium configurations Y (as illustrated in Figure 4.2) corresponds to sampling
all possible glass states, which restores the translational and rotational invariance of
the system. Therefore, the ensemble of glass states is described (once replicas are
introduced) by a molecular liquid that remains in its translationally and rotationally
invariant – i.e., liquid – phase.
Assuming that one is able to find an expression for Zs+1 [ϕa,βa, Dr ] that can
be analytically continued from integer s to real s, the Franz–Parisi potential can
be extracted as follows. First, we note that Z1 [ϕa,βa, Dr ] = Z[ϕg,βg ]. Then, we
define a replicated free energy7
1
−fs+1 (ϕa,Ta, Dr ) = log Zs+1 [ϕa,βa, Dr ]
N (4.15)
= −βg f(ϕg,Tg ) − sβfg (ϕ,T ;ϕg,Tg, Dr ) + O(s ) , 2

which also implies

βfg (ϕ,T ;ϕg,Tg, Dr ) = lim ∂s fs+1 (ϕa,Ta, Dr ). (4.16)
s→0

From the expansion of the replicated free energy in powers of s, we can thus derive
the glass free energy. In Section 4.2, we discuss how to obtain an exact expression
for the replicated free energy in the infinite dimensional limit.

6 Provided periodic boundary conditions are used for a finite volume V .

7 Note that usually −βf is the logarithm of the partition function. Here, we omitted the factor β because the
replicas are at different temperatures βa . As a consequence, fs+1 (ϕa ,Ta , Dr ) is adimensional. Another
possible convention, which we do not follow here, would be to multiply the partition function by any of the
replica temperatures Ta = 1/βa .
 4.2 Restricted Thermodynamics in Infinite Dimensions 109

4.2 Restricted Thermodynamics in Infinite Dimensions

In Section 4.1, we have seen that in order to describe the thermodynamics of a
glass one needs to introduce identical replicas of the original atomic system. In the
replicated system, the copies of a same particle in different replicas are close to one
another and, thus, form a molecule. In this section, we derive an exact expression
for the glass free energy in infinite dimensions. First, we generalise the discussion
of Section 2.3.1 to a molecular liquid. Then, we use the translational and rotational
invariance of the molecular liquid to express its free energy in terms of the matrix
of scalar products of particle displacements in a molecule. Finally, we consider the
scaling in the limit d → ∞ and we obtain the expression for the Franz–Parisi
potential.
In this section, we provide an exact derivation of the replicated free energy in
d → ∞, adapted from the original papers [88, 220, 221] but using the important
simplifications obtained in [59]. Its content is more technical than the rest of the
book. In a first reading, or if the reader is not interested in the technical details, it is
possible to jump directly to Section 4.2.4, where the final result is given in compact
form. A simpler derivation is also possible by means of a Gaussian ansatz, which,
however, introduces an additional assumption [88]. More details on this simpler
derivation are given in Section 4.6.3.

4.2.1 Virial Expansion for a Molecular System

We begin by considering a generic molecular liquid, in which a basic constituent
is not a simple atom (labelled by its position xi ) but a molecule, composed
of n atoms, and labelled xi = {x1i , . . . ,xni }, where xai is the position of atom
a = 1, . . . ,n in molecule i = 1, . . . ,N. The total configuration of the system is
then X = {xi }i=1,...,N . The interaction potential has the form
 
V (X) = w(xi ) + v(xi − xj ), (4.17)
i i<j

where w(x) is the interaction between atoms inside a same molecule, which is
responsible for its binding, while v(x − y) is the interaction between molecules
x and y. Although we ultimately want to consider one special replica at temperature
Tg and density ϕg , and s identical replicas at different temperature T and density ϕ,
as in Section 4.1, for notational simplicity, we first consider the case in which all
replicas are at the same (ϕ,T ).
Identifying V (X) with Eq. (4.17), all the definitions of Section 2.1.1 remain iden-
tical, except that the total number of degrees of freedom is now ndN instead of dN.
Therefore, one should replace dN → ndN in the denominator of QN in Eq. (2.3).
110 Thermodynamics of Glass States

This modification only affects the ideal gas term of the free energy, in which the
kinetic term dT log  is multiplied by n. The definitions of Section 2.1.2 are also
straightforwardly extended to the molecular liquid by replacing x → x everywhere.
The local density thus becomes a local ‘molecular density’ ρ(x), normalised by
dxρ(x) = N , and is proportional to the probability of finding a molecule with
atom 1 in position x1 , atom 2 in position x2 and so on. Similarly, the definitions and
results of Section 2.1.3 can be straightforwardly extended to the molecular liquid
by replacing x → x and replacing the definition of z(x) in Eq. (2.16) by
eβμ−βw(x)−βφ(x)
z(x) = . (4.18)
ndN
Introducing the molecular Mayer function
f (x − y) = e−βv(x−y) − 1, (4.19)
the virial expansion can be obtained as in Section 2.2 (once again replacing x → x).
At second order in ρ(x), the free energy of the molecular liquid is

−βF [ρ] = −ndN log  + dxρ(x)[1 − log ρ(x)]
 (4.20)
1
+ dxdyρ(x)ρ(y)f (x − y).
2
The third and higher-order virial terms have diagrammatic forms identical to
those of Section 2.2.3, but with integrations over molecular coordinates x, with a
molecular density ρ(x) on each vertex and a molecular Mayer function f (x − y)
on each edge.

4.2.2 The Liquid Phase: Rotational and Translational Invariance

We are interested in the liquid phase of the molecular liquid, which is rotationally
and translationally invariant. We first discuss some important preliminary conse-
quences of these symmetries.

Translational Invariance
Translational invariance implies that the probability of finding a molecule in
x = {x1, . . . ,xn } is the same as that of finding it in a translated configuration8
x − X = {x1 − X, . . . ,xn − X}. Therefore, the density satisfies ρ(x) = ρ(x − X).
A convenient way to take advantage of this property is to introduce a change of
variables that takes atom 1 as reference – i.e., X = x1 – and then defines

8 Note that this expression defines, in our notation, the operation of summing a single vector and a molecular
configuration.
 4.2 Restricted Thermodynamics in Infinite Dimensions 111

ua = xa − x1, u = {0,u2, . . . ,un },


n (4.21)
dx = dX du, du = dua .
a=2

Because ρ(x) is translationally invariant, it does not depend on X; hence,

ρ(x) = ρ(x − X) = ρ(u). We define π(u) = ρ(u)/ρ, where ρ = N/V is
the molecular number density, normalised by
  
1 ρ
1= dxρ(x) = dXduπ(u) = duπ(u), (4.22)
N N

where in the last equality we used that ρ dX = ρV = N . Hence, π(u) can be
interpreted as the probability distribution of displacements u in a molecule, relative
to the first atom. With this change of variable, the second virial term becomes, for
instance,

1
= dxdyρ(x)ρ(y)f (x − y) (4.23)
2
 
ρ2 ρ
= dXdYdudvπ(u)π(v)f (X − Y + u − v) = N dRfeff (R),
2 2
where
 
feff (R) = dudvπ(u)π(v)f (R + u − v) = f (R + u − v) . (4.24)
u,v

In this section, •u denotes an average over the internal fluctuations of a molecule.
The function feff (R) can be interpreted as an effective interaction between the
atoms in replica 1, once internal molecular degrees of freedom associated to the
other replicas have been averaged out.
The second virial coefficient B2 is then given by Eq. (2.43), with the Mayer
function feff (R). Similarly, the third virial coefficient can be written as

1
= dxdydzρ(x)ρ(y)ρ(z)f (x − y)f (x − z)f (y − z)
6
 (4.25)
ρ2 (3)
=N dR1 dR2 feff (R1,R2 ),
6
where we defined R1 = X − Z and R2 = Y − Z with X − Y = R1 − R2 , and

(3)
feff (R1,R2 ) = f (R1 − R2 + u − v)f (R1 + u − w)f (R2 + v − w) .
u,v,w
(4.26)
112 Thermodynamics of Glass States

This function represents an effective three-particle interaction between reference

coordinates. In general, it cannot be factorised as a product of two-particle
interactions.

Rotational Invariance
Rotational invariance implies that the probability distribution function of displace-
ments π(u) is invariant under a global rotation of all vectors ua ; it then becomes a
function of qab = ua · ub only, i.e. π(u) = π(q̂). Rotational invariance also implies
that ua u = 0. One can introduce a change of variables from the displacements u
to the scalar products matrix q̂. Recall that here a,b = 2, . . . ,n and the matrix q̂ is
thus a (n − 1) × (n − 1) matrix. The corresponding Jacobian is
 
2,n
(d−n)/2
J (q̂) = du δ(qab − ua · ub ) = Cn,d det q̂ , (4.27)
a≤b

as proven in Section 2.5. The averages over duπ(u) can then be replaced by
averages over dq̂J (q̂)π(q̂). Note that both distributions are, in fact, correctly
normalised to unity.
Finally, note that rotational invariance implies that feff (R) depends only on the
(3)
modulus R = |R| and that feff (R1,R2 ) depends only on |R1 |, |R2 |, and |R1 − R2 |.
The third virial coefficient B3 is thus given by Eq. (2.44) with the replacement
(3)
f (|r1 |)f (|r2 |)f (|r1 − r2 |) → feff (|R1 |,|R2 |,|R1 − R2 |). This procedure can be
further generalised to higher-order virial coefficients.

4.2.3 Scaling in the Large-Dimensional Limit

We now focus on the limit d → ∞, using a specific scaling that is relevant for our
problem, and is motivated by the discussions of Chapters 2 and 3.

Scaling of the Molecular Liquid in d → ∞

We take the limit d → ∞ of the molecular liquid under two assumptions:
1. The inter-molecular interaction has the form
 n
v(x − y) = v(|xa − ya |), v(r) = v̄[d(r/ − 1)]. (4.28)
a=1

This amounts to assuming that only atoms of the same type a interact directly,
which is the relevant situation for computing the Franz–Parisi potential
(Section 4.1.3) and that the interaction potential satisfies the correct scaling
to obtain a non-trivial liquid phase in d → ∞ (Section 2.3.2). In the following,
for simplicity, we consider that all the replicas have the same potential v̄(h),
 4.2 Restricted Thermodynamics in Infinite Dimensions 113

the same interaction scale , and the same inverse temperature β. We will later
generalise the result to the case of different interaction scales and temperatures,
which is required for computing the Franz–Parisi potential.
2. We define a (n − 1) × (n − 1) adimensional matrix α̂ from9

αab 2
u  = 0,
a
= u · u  = duπ(u)ua · ub,
a b
(4.29)
d
for a,b = 2, . . . ,n. Because of rotational invariance, one has
1 a b αab 2
uaμ ubμ  = u · u  = , (4.30)
d d2
for each component μ = 1, . . . ,d. We assume that α̂ remains finite when
d → ∞, which corresponds to molecules being strongly localised. In fact,
Eq. (4.30) with a = b implies that the projection of ua along an arbitrary unit
vector e is a random variable with zero mean and variance αaa 2 /d. Its typical
√
value thus scales as ua · e ∝ αaa /d. We can also extend the matrix α̂ to a
n×n matrix γ̂ such that γab = αab for a,b = 2, . . . ,n and γa1 = γ1b = 0, which
is consistent with u1 = 0. Then we define the n × n mean square displacement
matrix ˆ by the relation
ab 2  a 
= |u − ub |2 ⇔ ab = γaa + γbb − 2γab . (4.31)
d
The assumption that α̂ remains finite when d → ∞ is equivalent to the assump-
ˆ remains finite – i.e., that the mean square displacement between
tion that 
atoms in the molecule is proportional to 1/d. This assumption is consistent with
the discussion of Section 3.3.2, which shows that the long time limit of the
mean square displacement in the glass phase (here encoded by the displacement
between different replicas) is indeed proportional to 1/d. See, in particular,
Eq. (3.63).

Saddle Point Method

In the limit d → ∞, the averages over the internal displacements of a molecule
can be calculated via the saddle point method. To illustrate this point, consider the
normalisation of π(u) together with Eq. (4.29):
  
d−n
1 = duπ(u) = dq̂J (q̂)π(q̂) = Cn,d dq̂e 2 log det q̂+log π(q̂),
  (4.32)
αab 2 d−n
= qab  = dq̂J (q̂)π(q̂)qab = Cn,d dq̂e 2 log det q̂+log π(q̂)
qab .
d

9 In this section, we drop the suffix u in the averages • to simplify the notation.
u
114 Thermodynamics of Glass States

When d → ∞, the term in the exponent is proportional to d at leading order.10

According to the saddle point method, both integrals in Eq. (4.32) are then domi-
nated by the value of q̂ that maximises this term, i.e.,
 
d
q̂ ∗ = argmax log det q̂ + log π(q̂) . (4.33)
q̂ 2
The first of Eqs. (4.32) also imposes that, at leading exponential order in d,
d−n
log det q̂ ∗ +log π(q̂ ∗ )
1 = Cn,d e 2 ⇒
 
d(n − 1) 2πe d (4.34)
⇒ 0= log + log det q̂ ∗ + log π(q̂ ∗ ),
2 d 2
where the second line is obtained by taking the logarithm and keeping only
terms that are proportional to d, using the asymptotic relation log Cn,d
∼ (n − 1) log d ∼ d(n−1)
2
log(2πe/d). Then, the second of Eqs. (4.32) becomes
αab 2 d−n ∗ ∗
= Cn,d e 2 log det q̂ +log π(q̂ ) qab
∗ ∗
= qab . (4.35)
d
A similar result holds for any function F (q̂) that does not depend exponentially
on d. We thus obtain
 
F (q̂) → F (q̂ ∗ ) = F (α̂2 /d). (4.36)
d→∞

Truncation of the Virial Expansion

We now show that when d → ∞, the virial expansion of the molecular liquid
can be truncated at second order, as for atomic liquids. The center-of-mass Mayer
function, defined in Eq. (4.24), can be written explicitly as
  n 
feff (R) = dudvπ(u)π(v) e−β a=1 v̄[d(|R+u −v |/−1)] − 1 .
a a
(4.37)

When |h| → ∞, one has h = d(R/ − 1) ≈ d(|R + ua − va |/ − 1) ,∀a, and

therefore, v̄[d(|R + ua − va |/ − 1)] ≈ v̄(h). This equivalence holds because the
component of each ua −va along the direction of R is of order 1/d. From Eq. (4.37),
it then follows that for |h| → ∞,

−1 h → −∞,
feff (R) ≈ e−βnv̄(h) − 1 → (4.38)
0 h → ∞.

10 This statement relies on the assumption that log π(q̂) ∝ d, because the two terms must have the same scaling
to obtain a non-trivial equation for q̂. The validity of this assumption is strongly suggested by Eq. (4.34),
which shows that it holds at least for q̂ = q̂ ∗ (with log d corrections). It can also be checked for q̂ = q̂ ∗ ,
see [221] for details.
 4.2 Restricted Thermodynamics in Infinite Dimensions 115

The function feff (R) thus increases from −1 to 0 when h increases from −∞ to ∞
– i.e., over an interval of order 1/d around R = . The same scaling is obtained for
the potentials considered in Section 2.3.2. As a consequence of these hypotheses,
the Mayer function scales in the same way as the potential:

feff (R) = f eff [d(R/ − 1)] = f eff (h) . (4.39)

An explicit expression of f eff (h) is given next. We can then follow the procedure
described in Section 2.3.2 to show that the second virial coefficient has the same
form as in Eq. (2.75),
 ∞
Vd d
B2 = − dh eh f eff (h) = B2HS I (π, v̄) , (4.40)
2 −∞

with I (π, v̄) being a finite integral that depends on the details of the interaction
potential v̄(h) and the molecular distribution π(u). We thus conclude that B2 has
the same leading exponential scaling with d as for hard spheres and all the potentials
of Section 2.3.2.
A similar analysis can be applied to all the other diagrams that contribute to
the higher-order virial coefficients. The expression in Eq. (4.26), together with the
(3)
hypotheses made earlier that lead to Eq. (4.39), imply that feff (|R1 |,|R2 |,|R1 −R2 |)
coincides with a product of three hard-sphere Mayer functions if at least one of the
three arguments differs from  by a quantity of order /d. This implies that the
third virial coefficient of the molecular liquid has the same leading exponential
scaling in d as the third virial coefficient of hard spheres, and the same is true for
higher-order virial coefficients. We conclude that the analysis of Section 2.3.1 holds
for the molecular liquid; the virial series can be truncated at the second order, and
Eq. (4.20) becomes exact in the large-dimensional limit. We now compute the two
terms of Eq. (4.20).

The Ideal Gas Free Energy

Recalling that ρ(x) = ρπ(u) and dx = dXdu, we can write the ideal gas contribu-
tion to the free energy per particle as

1  
−βfid + nd log  = dxρ(x)[1 − log ρ(x)] = 1 − log ρ − log π(u) .
N
(4.41)

The average of log π(u) can be computed by changing variable to q̂ and applying
Eq. (4.36), because we assumed that log π(u) is not exponential in d. At leading
order in d, we get from Eq. (4.34) that
116 Thermodynamics of Glass States
 
  d(n − 1) 2πe d
log π(q̂) = log π(α̂ /d) = −2
log − log det(α̂2 /d),
2 d 2
(4.42)
and therefore, still at leading order in d,
 
d(n − 1) 2πe d
−βf = −nd log  − log ρ +
id
log + log det(α̂2 /d)
2 d 2
 
d(n − 1) 2πe d
= −nd log(/) − log(ρ ) +
d
log 2
+ log det α̂ .
2 d 2
(4.43)

Note that in the second line, all factors have been made adimensional by using the
reference length scale .

The Excess Free Energy

We now derive an expression for the Mayer function – and, therefore, for the
second virial coefficient that determines the excess free energy – as a function of q̂.
Recalling that u1 = v1 = 0, Eq. (4.39) can be written as
 n
feff (R) = e−β v̄[d(R/−1)] e−β a=2 v̄[d(|R+u −v |/−1)] − 1.
a a
(4.44)

In order to evaluate this expression, we need to compute the distribution of random

variables xa = |R + ua − va |, which depend on random variables11 wa = ua − va .
In order to characterise the statistics of xa , it is convenient to introduce yet another
random variable,


d
ya = xa2 − R = |w | + 2R · w =
2 a 2
[(wμa )2 + 2Rμ wμa ].
a
(4.45)
μ=1

Each ya is the sum of d random variables. By the central limit theorem, when
d → ∞, ya then become Gaussian random variables specified by their mean and
covariance, which can be computed as follows. For the mean, we note that wa  = 0
and ua · va  = ua  · va  = 0. We then have
     
ya  = |wa |2 = |ua |2 + |va |2 = 2αaa 2 /d. (4.46)

The term |wa |2 is a sum of d positive terms, each of order 1/d 2 . Its mean is thus of
order 1/d, and its standard deviation is of order 1/d 3/2 . Its fluctuations can then be

11 The distribution of wa can be determined by recalling that u and v are independently and identically
distributed according to π(u), but we do not need here its explicit form.
 4.2 Restricted Thermodynamics in Infinite Dimensions 117
 
neglected, |wa |2 ≈ |wa |2 . To compute the covariance, we thus consider at leading
order that ya − ya  ≈ 2R · wa , and we obtain
  
ya yb  − ya  yb  = 4 Rμ Rν wμa wνb = 8R 2 αab 2 /d 2, (4.47)
μν
     
where we used that wμa wνb = uaμ ubν + vμa vνb = δμν 2αab 2 /d 2 by rotational
invariance and Eq. (4.30).
Equivalently, we can introduce random Gaussian variables z = {za }a=2,...,n with
probability measure
3    4
−1
exp − 12 na,b=2 2α̂ za zb n
Dz = dz  ab
, dz = dza, (4.48)
(2π)n−1 det(2α̂) a=2

such that za  = 0 and za zb  = 2αab , and express ya as

ya = 2αaa 2 /d + 2Rza /d, (4.49)
which ensures that ya are Gaussian variables with the correct mean and covariance.

Eq. (4.49) shows that ya ∼ 1/d. Recalling from Eq. (4.45) that xa = R 2 + ya ,
and defining h = d(R/ − 1), we obtain at leading order in 1/d:
 ' 
 R ya (
d(xa / − 1) = d( R + ya / − 1) ∼ d
2 1+ −1
 2R 2
(4.50)
d ya
∼ h + 2 = h + αaa + za .
2
Eq. (4.44) can then be written as
 n
f eff (h) = e−β v̄(h) e−β a=2 v̄(h+αaa +za ) − 1
z
 (4.51)
 n 
= e−β v̄(h) Dz e−β a=2 v̄(h+αaa +za ) − 1.

Using the general identity in Eq. (4.88), the Gaussian convolution can be rewritten
in a differential form, and Eq. (4.51) becomes
n ∂2  n 00
f eff (h) = e−β v̄(h) e a,b=2 ab ∂ha ∂hb e−β a=2 v̄(ha +αaa ) 0
α
− 1. (4.52)
ha =h

Plugging this result in Eq. (4.40) allows us to write the second virial coefficient and
the excess free energy per particle as
−βfex = −ρB2 (4.53)
 ∞ 5 6
ϕ −β v̄(h)
n ∂2
a,b=2 αab ∂ha ∂hb
 −β n v̄(ha +αaa ) 00
= h
dh e e e e a=2 0 −1 ,
2 −∞ ha =h

where ϕ = ρVd d = 2d ϕ is defined as in Chapter 2.

118 Thermodynamics of Glass States

Total Free Energy

In summary, in d → ∞, the total free energy per particle, f = fid + fex , of a
molecular liquid can be written as
 
d(n − 1) 2πe d
− βf = −nd log(/) − log(ρ ) + d
log 2
+ log det α̂
2 d 2
 ∞ 5 6
ϕ −β v̄(h)
n ∂2
a,b=2 αab ∂ha ∂hb
 −β n v̄(ha +αaa ) 00
+ h
dh e e e e a=2 0 − 1 . (4.54)
2 −∞ ha =h

This expression clearly depends on density, temperature and interaction potential.

It also depends on the molecular distribution π(u), but only through the matrix α̂
defined in Eq. (4.29). This matrix encodes the average scalar products of the internal
displacements of a molecule. Minimising the free energy with respect to π(u) thus
amounts to minimising it with respect to α̂.
Eq. (4.54) has been derived by using replica 1 as reference, which is clearly
arbitrary; an equivalent result would be obtained by choosing any other replica as
reference. To make this symmetry more explicit, we can express Eq. (4.54) as a
function of the scaled mean square displacement matrix  ˆ defined in Eq. (4.31). In
Section 4.6.1, it is shown that

ˆ T ˆ −1
det α̂ = 2 det(−/2)(−1  1), (4.55)

where 1 = {1, . . . ,1} is the vector of all ones. This allows us to express the ideal gas
term as a function of . ˆ For the excess free energy term – introducing, as before, a
n × n matrix γ̂ such that γab = αab for a,b = 2, . . . ,n and γa1 = γ1b = 0 – we can
use the identity in Eq. (4.90) together with Eq. (4.31) to obtain
 
d(n − 1) 2πe
−βf = −nd log(/) − log(ρ ) + d
log
2 d2
d ˆ T ˆ −1
+ log[2 det(−/2)(−1  1)] (4.56)
2
 3 1 n
ϕ ∞ − ∂2  n 4
dh eh e 2 a,b=1 ab ∂ha ∂hb e−β a=1 v̄(ha ) − 1

+ .
2 −∞ ha =h

Both Eqs. (4.54) and (4.56) show that the molecular liquid free energy does not
depend, in the limit d → ∞, on the full shape of the molecular distribution π(u)
but only on its second moments. A Gaussian ansatz for π(u) would thus provide
an equivalent result, as discussed in Section 4.6.3. Minimising the free energy over
π(u) then also becomes equivalent, in the limit d → ∞, to minimising Eq. (4.54)
ˆ
over α̂ or, equivalently, Eq. (4.56) over .
 4.2 Restricted Thermodynamics in Infinite Dimensions 119

4.2.4 Replicated Free Energy in Infinite Dimensions

The result obtained in Section 4.2.3 can be used to compute exactly the replicated
free energy introduced in Eq. (4.15) and the Franz–Parisi potential, with some
minor adaptations. We need to consider n = s + 1 replicas with the following
properties:

• The first replica has density ϕ1 = ϕg and temperature T1 = Tg , while the others
have ϕa = ϕ and Ta = T . The limit d → ∞, as discussed in Chapter 3, should
be taken with a constant scaled packing fraction ) ϕ = 2d ϕ/d. In the general case,
each replica has a different interaction range a , with all )
ϕa finite. We then need
to scale the ranges as a = g (1 + ηa /d), in such a way that

)
ϕa = )
ϕg (1 + ηa /d)d → )
ϕg eηa . (4.57)

Here g = 1 , which corresponds to ) ϕg = ) ϕ1 , is used as reference unit of length.

In order to compute the glass free energy, we are eventually interested in setting
η1 = 0 and ηa = η for a ≥ 2.
• The interaction potential of replica a is, following the analysis of Section 2.3.2,
  
ra
v(ra ) = v̄ d −1
a
   (4.58)
ra
= v̄ d −1 −→ v̄(ha − ηa ),
g (1 + ηa /d) d→∞

where ha = d(ra /g − 1) is the scaled interparticle distance for replica a, using
g as reference unit of length.
• Following the analysis of Section 4.2.3, the mean square a displacement
 between
 scaled as ab = (d/g ) |u − u | = (d/g ) Dab ,
2 b 2 2
replicas a and b should be
where Dab = D(X ,X ) . The constraint in Eq. (4.14) imposes that D1a = Da1
a b

= Dr , ∀a = 1, which implies 1a = a1 = (d/2g ) Dr ≡ r . The first row and

column of ab are therefore fixed by the constraint. The other matrix elements
ab , with a = 1 and b = 1, which encode the distance between the replicas
different from 1, are left free and, according to the analysis of Section 4.2.3, the
free energy should be minimised over them.

To summarise, we can start from Eq. (4.56), with  = g and n = s +1, with replica
a having temperature βa and an interaction potential v̄(ha − ηa ), which corresponds
ϕa = )
to ) ϕg eηa , and impose the constraint that 1a = a1 = r . In Section 4.2.3,
it was assumed for simplicity that all replicas have the same β v̄(ha ), but the reader
can easily check that this assumption was not actually used in the derivation, and
one can therefore simply substitute β v̄(ha ) → βa v̄(ha − ηa ) in Eq. (4.56). Recall
120 Thermodynamics of Glass States

that we omit the factor β in front of the free energy, because the temperatures are
now different, as in Eq. (4.15). We thus obtain
 
ˆ = −(s + 1)d log(/g ) − log(ρg ) +
ds 2πe
−fs+1 ()
ϕa,Ta, ) d
log
2 d2
d d)ϕg
ˆ
+ log[2 det(−/2)(−1 T ˆ −1
 1)] − F(),ˆ (4.59)
2 2
 ∞ 5  2  n 6
ˆ − 12 na,b=1 ab ∂ha∂ ∂h − −η
F() = − h
dh e e b e a=1 βa v̄(ha a )
−1 ,
−∞ ha =h

which should be minimised over all matrix elements ab with a = 1 and b = 1.
Note that computing the restricted thermodynamic glass free energy, according to
Eq. (4.10), requires an additional minimisation over r . The two minimisations can
be performed concurrently. In order to compute Eq. (4.10), one thus has to minimise
ˆ
the free energy over the full matrix .
The submatrix ab for a = 1 and b = 1 is an s × s matrix. In principle, we
should determine it by minimising the free energy for each integer value of s and
then perform an analytical continuation to noninteger s in order to take the s → 0
limit and extract the glass free energy according to Eq. (4.15). In general this task
is very complex, unless we consider a special prescription for  ˆ that permits a
simple analytical continuation. Such a prescription is discussed in [254] and is at
the core of the replica approach to disordered systems. In Section 4.3, we discuss
the simplest example of this construction and its physical consequences.

4.3 Replicated Free Energy and Replica Symmetry

Finding explicitly the matrix  ˆ that minimises the replicated free energy in
Eq. (4.59) is greatly simplified by the use of symmetry. Because we consider a
system prepared at () ϕg,Tg ) and then followed to () ϕ,T ), we need the replicated
free energy in which the first replica is at ()ϕg,Tg ) and all the remaining s = n − 1
replicas are at () ϕ,T ), as discussed in Section 4.1.3. As a consequence, the s
replicas at ()ϕ,T ) are equivalent, and the free energy must be symmetric under their
exchange. This symmetry is manifest in Eq. (4.12), where the s replicas give rise to
the s equivalent factors Z[ϕ,β;Y, Dr ]. This permutation symmetry of the s replicas
a = 2, . . . ,s + 1 is called ‘replica symmetry’ and can be exploited analytically to
obtain constraints on observables derived from the replicated free energy [254].

4.3.1 Replica Symmetric Matrices

The permutations of k objects form a group Pk with a generic element acting as
P(a1, . . . ,ak ) → (aP (1), . . . ,aP (k) ). For the molecular liquid, the function ρ(x)
 4.3 Replicated Free Energy and Replica Symmetry 121

expresses the probability of finding a given molecule (composed of an atom from

each of the n replicas) in position x. Because of the replica symmetry, the free
energy must satisfy

F [ρ(x)] = F [ρ(Px)], Px = {x1,xP (2), . . . ,xP (n) }, P ∈ Pn−1 . (4.60)

The same symmetry is also present when the free energy is expressed in terms
of π(u). The mean square displacement matrix, defined in Eq. (4.31) as ab
∝ |ua − ub |2 , transforms under the action of a permutation of the replicas as
ab = P (a)P (b) . As a consequence, when ) ϕa = )
ϕ and Ta = T , ∀a ≥ 2, the
replicated free energy in Eq. (4.59) satisfies

ϕa,Ta,ab ) = fs+1 ()
fs+1 () ϕa,Ta,P (a)P (b) ), P ∈ Pn−1 . (4.61)

In other words, the free energy takes the same value on all matrices  ˆ that are
obtained by permuting lines and columns in the block of the s replicas with a ≥ 2.
This situation is similar to that discussed in Chapter 1 for the homogeneous
Ising model in absence of an external magnetic field, in which the free energy
is a symmetric function of magnetisation, with f(m) = f(−m). In the param-
agnetic phase, in which there is a single equilibrium state, the free energy has a
unique minimum that is invariant under the symmetry operation and, therefore, has
m = −m = 0. Conversely, in presence of many equilibrium states, the symmetry is
spontaneously broken, and the free energy has multiple minima that transform one
into another under the action of the symmetry group. In the case of the Ising model,
only two minima can be present, one for m = m∗ and the other for m = −m∗ .
For the molecular liquid, when there is a single equilibrium state, the free energy
must have a unique minimum. This corresponds to a matrix  ˆ that is invariant
under the action of the permutation group, ab = P (a)P (b) . Such a matrix is
called ‘replica symmetric’ (RS) and necessarily has the form:

aa = 0, a = 1, . . . ,n,
1a = a1 = r , a = 2, . . . ,n, (4.62)
ab = , a,b = 2, . . . ,n, and a = b.

For example, when n = 4,

⎛ ⎞
0 r r r
⎜ ⎟
ˆ =⎜ ⎟
r 0 
 ⎜ ⎟. (4.63)
⎝r  0 ⎠
r   0
122 Thermodynamics of Glass States

In Section 4.2.3 we also introduced an s × s matrix for a,b = 2, . . . ,n = s + 1:

αab = (d/2g ) ua · ub  = (d/2g ) (xa − x1 ) · (xb − x1 )

(4.64)
= (a1 + 1b − ab )/2.

See Section 4.6.1 for a proof of the last equality. The replica symmetric matrix has
the form αab = r δab + (r − /2)(1 − δab ). For example, when n = 4,
⎛ ⎞
r r − /2 r − /2
α̂ = ⎝r − /2 r r − /2⎠ . (4.65)
r − /2 r − /2 r

In the rest of this chapter, we restrict our analysis to the replica symmetric case. In
Chapter 5, we will introduce the notion of spontaneous replica symmetry breaking
and discuss its consequences.

4.3.2 Replica Symmetric Free Energy

The replicated free energy can be evaluated explicitly if the matrices  ˆ and α̂ have
the replica symmetric form of Eqs. (4.63) and (4.65). The computation is slightly
simpler for the expression as a function of α̂, Eq. (4.54). Here again, we neglect the
kinetic term proportional to log  in the free energy, which comes from integrating
over the momenta, because it is an irrelevant additive constant. Taking into account
the temperature and potential difference for the different replicas – i.e., β1 = βg ,
η1 = 0, βa = β and ηa = η, ∀a ≥ 2 – gives
 
ds 2πe d d)ϕg
ϕa,Ta, α̂) =
− fs+1 () − log(ρdg ) + log 2
+ log det α̂ − F(α̂) ,
2 d 2 2
(4.66)
 3   0 4
∞ n 2  0
e a,b=2 ab ∂ha ∂hb e−β a=2 v̄(ha +αaa −η) 00
∂ n
dh eh e−βg v̄(h)
α
F(α̂) = − −1 .
−∞ ha =h

In order to obtain the replica symmetric free energy, we first need to compute det α̂
for α̂ given in Eq. (4.65). The eigenvectors of α̂ are the vector 1 = {1, . . . ,1}, with
associated eigenvalue λ1 = sr − (s − 1)/2 and all the s − 1 vectors orthogonal
to 1, with degenerate eigenvalues λ2 = /2. Therefore,
  s−1
 
det α̂ = sr − (s − 1) . (4.67)
2 2
 4.3 Replicated Free Energy and Replica Symmetry 123

The function F(α̂) that enters in the excess term, for a replica symmetric matrix,
gives
 ∞ 3 
'
n ∂
(2 n ∂2
−βg v̄(h) (r − 2 ) +
F(α̂) = − h a=2 ∂ha 2 a=2 ∂h2
dh e e e a
−∞ (4.68)
 −β n v̄(ha +r −η) 00 4
× e a=2 0 −1 .
ha =h

In the following, we denote the Gaussian convolution of a generic (integrable)

function r(h) as
 ∞
 d2 dz z2
γ  r(h) = e 2 dh2 r(h) = √ e− 2 r(h − z). (4.69)
−∞ 2π
See Section 4.6.2 for details. From this definition, it follows that
 ∞  ' (2
n
 − ∂
F(α̂) = − dh e e−βg v̄(h) e( r 2 ) a=2 ∂ha
h
−∞
0 ⎫
 n 0 ⎬
0 (4.70)
× gRS (,β;ha + r − η)0 −1 ,
0 ⎭
a=2 ha =h

gRS (,β;h) ≡ γ  e−β v̄(h) .

Applying the generic identity for a function of k variables T (x1, . . . ,xk ),
0

k 0
d ∂ 0
T (t, . . . ,t) = T (x1, . . . ,xk )0 , (4.71)
dt ∂xi 0
i=1 xi =t

to Eq. (4.70) gives

 ∞ 3 2 4
 − d
F(α̂) = − dh eh e−βg v̄(h) e( r 2 ) dh2 gRS (,β;h + r − η)s − 1 . (4.72)
−∞

Plugging these results into Eq. (4.66), one obtains an explicit expression in terms
of s; hence, the analytical continuation to real s is straightforward. Expanding in
powers of s up to linear order, one obtains the following expression for the glass
free energy:
 
d πe d 2r − 
ϕ,T |)
− βfg () ϕg,Tg ) = log 2
+
2 d 2 
 ∞
d)ϕg  − d
2
+ dh eh−βg v̄(h) e( r 2 ) dh2 log gRS (,β;h +r − η). (4.73)
2 −∞
The analytical continuation to s → 0 of the replica symmetric free energy can
thus be performed explicitly, and the glass free energy be written as a function of
124 Thermodynamics of Glass States

 and r . These last two parameters should be determined by minimising the free
energy. It is important to stress, however, that because of the analytical continuation,
the convexity property of the free energy in Eq. (4.73) with respect to  changes
when s < 1. The thermodynamic value of  is the maximum, and not the minimum,
of the free energy. This fact is well known in the context of spin glasses [254] and
we will not discuss it further. Because the free energy must then be maximised
with respect to  and minimised with respect to r , the thermodynamic values of
,r correspond to a saddle point of Eq. (4.73). The values of ,r should thus be
determined by setting to zero the derivatives of fg with respect to these parameters,
and to avoid any confusion, we refer to this process as an ‘extremisation’ rather
than a minimisation of the free energy.

4.3.3 Recipe for State Following within the Replica Symmetric Scheme
In this section, we provide in compact form all the formulae that are needed to
compute the thermodynamic properties of glass states prepared at an initial state
ϕg,Tg ) and followed to a new state ()
() ϕ,T ), for a general interaction potential v̄(h).
The equations presented here can be used in full generality; examples for specific
systems are presented in Section 4.4.
The free energy in Eq. (4.73) can be compactly written as follows. One can first
integrate by parts the differential operator in Eq. (4.73) in such a way that it acts
on the term at its left, then use the identity in Eq. (4.89) with r(h) → e−βg v̄(h) and
 → 2r −  and finally shift the variable h → h − r + /2. By defining the
function
q(,β;h) = gRS (,β;h + /2) = γ  e−β v̄(h+/2), (4.74)
one then obtains
 
d πe d 2r − 
− βfg ()
ϕ,T |)
ϕg,Tg ) = log +
2 d2 2 
 ∞
(4.75)
d)
ϕg
+ dh e q(2r − ,βg ;h) log q(,β;h − η),
h
2 −∞

which is the final expression of the replica symmetric free energy of the glass state
with η = log() ϕg ). As discussed at the end of Section 4.3.2,  and r should
ϕ /)
be determined by extremisation of Eq. (4.75), which results in the two coupled
equations:
 ∞
∂  
2r =  + ) ϕg 2 dh eh q(2r − ,βg ;h) log q(,β;h − η) ,
−∞ ∂
 ∞   (4.76)
2 ∂
= −) ϕg dh e h
q(2r − ,βg ;h) log q(,β;h − η).
 −∞ ∂r
 4.3 Replicated Free Energy and Replica Symmetry 125

In practice, Eqs. (4.76) can be solved efficiently by iteration. Start from a reasonable
guess for  and r , and compute numerically the right-hand side to obtain new
estimates of  and r . Upon iteration, the new estimates typically get closer to the
previous ones; hence, the procedure converges. The derivatives of q(,β;h) with
respect to  are not given here explicitly but can be deduced easily from Eq. (4.74).
Once the equations for  and r are solved, plugging the results in Eq. (4.75)
gives the thermodynamic free energy of the glass. Note that () ϕg,Tg ) specify
the preparation of the glass, through the reference configuration Y that acts
as a quenched disorder, as discussed in Section 4.1, while () ϕ,T ) control the
thermodynamic state of the system. By taking derivatives of the free energy
with respect to () ϕ,T ), one can thus obtain the averages of the thermodynamic
observables.12 For example, the average glass energy is the derivative of the free
energy with respect to the inverse temperature,

∂(βfg ) d)ϕg ∞ ∂
eg = =− dh eh q(2r − ,βg ;h) log q(,β;h − η),
∂β 2 −∞ ∂β
(4.77)
and the entropy is sg = −β(fg − eg ). The pressure can be obtained from Eq. (2.7);
observing that ρ ∝ ) ϕ , and recalling that η = log() ϕg ), the ‘reduced pressure’ or
ϕ /)
‘compressibility factor’ is given by
βP ∂(βf) ∂(βf) ∂(βf)
p= =ρ =)
ϕ = . (4.78)
ρ ∂ρ ∂)
ϕ ∂η
For the glass, one therefore obtains

∂(βfg ) ϕg ∞
d) ∂
pg = =− dh eh q(2r − ,βg ;h) log q(,β;h − η).
∂η 2 −∞ ∂η
(4.79)
The case in which a glass state is both prepared and studied at the same state
ϕ,T ) = ()
point – i.e., () ϕg,Tg ) – is special and displays an additional symme-
try. Replica 1 is then equivalent to all the others and  = r , as discussed in

12 When taking derivatives of the free energy with respect to a thermodynamic control parameter, one should
keep in mind the following structure. The free energy depends on the state point ()
ϕ,T ) but also on ,r
which are determined by setting the derivatives of the free energy to zero and, therefore, depend implicitly
on ()
ϕ,T ). When taking the first derivatives, one has, for example,
dfg ∂fg ∂fg ∂ ∂fg ∂r
= + + .
dT ∂T ∂ ∂T ∂r ∂T
Because the derivatives with respect to ,r are set to zero on the thermodynamic values of these
df ∂f
parameters, only the first term remains and dTg = ∂Tg , which considerably simplifies the calculation. Note,
however, that this is only true for first derivatives. When taking second derivatives (e.g., to compute the
specific heat), the derivatives of ,r with respect to ()ϕ,T ) appear explicitly.
126 Thermodynamics of Glass States

ϕ,T ) = ()
Section 4.1. For () ϕg,Tg ), one can indeed check that choosing  = r ,
both Eqs. (4.76) reduce to the same equation for r , which can be conveniently
written as
 ∞
1 ∂q(r ,βg ;h)
= F1 (r ;βg ) = −r dh eh log[q(r ,βg ;h)] . (4.80)
)
ϕg −∞ ∂r
Eq. (4.80) coincides with Eq. (3.65), derived in Chapter 3. This key result proves
that the hypothesis that the state following construction can reproduce the long
time limit of dynamical correlations in the arrested phase is indeed correct. Further-
more, for () ϕg,Tg ) and  = r the glass reduced pressure in Eq. (4.79)
ϕ,T ) = ()
simplifies to
 0
ϕ ∞
d) ∂ 0
pg = − h
dh e q(,β;h) log q(,β;h − η)00
2 −∞ ∂η η=0
 ∞ (4.81)
d)
ϕ 
=− dh e h−β v̄(h)
β v̄ (h),
2 −∞
which coincides with the reduced pressure in the liquid phase given in Eq. (2.78).
The pressure is thus continuous at the glass transition; the same also holds for the
internal energy.
Examples of the Franz–Parisi potential, Eq. (4.75), are given in Figure 4.3. The
corresponding function F1 (r ), defined in Eq. (4.80), is illustrated in Figure 3.2.
The solution of Eq. (4.80), which can be easily found numerically, can be used as
initial condition for the iteration of Eqs. (4.76) while one changes slowly () ϕ,T )
to follow the evolution of the state [299]. Note that there are usually at least two
solutions of Eq. (4.80), and the correct solution corresponds to a local minimum of
the Franz–Parisi potential fg ; in the example of Figure 4.3, this is the solution with
smaller value of r .
As discussed in Section 3.3.2, one can show that the function F1 (r ;βg ) gen-
erally vanishes both for r → 0 and r → ∞ and has a maximum in between.
Therefore, no solution to Eq. (4.80) can be found if )ϕg < )ϕd (βg ), where
1
= max F1 (r ;βg ). (4.82)
)
ϕd (βg ) r

The absence of a solution means that the Franz–Parisi potential has no local min-
imum at finite r . No stable glass phase then exists, and the system is therefore
a liquid. This condition allows one to determine the dynamical transition den-
sity )
ϕd (βg ) for each temperature, consistently with the dynamical treatment of
Chapter 3. For )ϕg > )ϕd (βg ), stable glass states exist and can be followed at different
state points.
 4.3 Replicated Free Energy and Replica Symmetry 127

ϕ = 4.5
ϕ = 4.81
ϕ = 5.0

βfg /d − log d
–

Figure 4.3 Franz–Parisi potential Eq. (4.75), multiplied by β and scaled to have
a finite limit in d → ∞, for hard spheres in equilibrium, corresponding to )
ϕ=) ϕg
and r = , at various densities, below, close to, and above the dynamical
transition at ) ϕd = 4.8067 . . . The corresponding function F1 (r ), defined by
Eq. (4.80), is illustrated in Figure 3.2.

In addition, the Franz–Parisi potential develops an inflection point at the dynam-

ical transition (Figure 4.3). It can be shown that the presence of this inflection
point gives rise to the dynamical criticality discussed in Section 3.4.2 [146, 204].
A development of the potential around its inflection point then gives access to the
dynamical critical exponents of mode-coupling theory [72, 288] and to the critical
properties of the four-point dynamical correlations [37, 145], which have been
discussed in Section 3.4.2.
Note that in the dynamically arrested phase, the Franz–Parisi potential is not
convex, as shown in Figure 4.3. This non-convexity is a necessary condition to
have a local minimum at finite r , and it is possible because, in the calculation of
the potential, r was assumed to be spatially uniform. The Franz–Parisi potential
is then a ‘potential’ similar to the function v(m) discussed in Section 1.4. In the
non-convex region, phase coexistence would lead to a non-uniform spatial profile
r (x) and restore the convexity of the free energy [73, 257]. We do not further
discuss this issue in this book; see Section 4.5.2 for additional references.
To summarise, the recipe to study the glass states of a given potential v̄(h) is the
following [299]:
1. Solve Eq. (4.82) to determine the dynamical transition line ) ϕd (βg ).
ϕg,Tg ) in the dynamically arrested region )
2. Choose an initial state () ϕg > )
ϕd (βg ).
128 Thermodynamics of Glass States

3. Solve Eq. (4.80) to determine r at the initial state point ()

ϕg,Tg ).
4. Choose a new state point ()ϕ,T ) at which to study the glass state, and iteratively
solve Eqs. (4.76), using as starting point the value of r at the initial state.
5. Once  and r at () ϕ,T ) are determined, one can obtain the free energy from
Eq. (4.75), the energy from Eq. (4.77) and the pressure from Eq. (4.79). Other
thermodynamic quantities can be computed along similar lines.

4.4 Replica Symmetric Phase Diagram of Simple Glasses

We now apply the recipe discussed in Section 4.3.3 to compute and discuss the
phase diagram of two simple glass models: hard spheres [299] and soft repulsive
spheres [315].

4.4.1 Hard Spheres

Hard spheres are the simplest model of glass forming liquid. The scaled infinite-
dimensional potential v̄(h) is deduced from Eq. (2.47). It is infinite for h < 0
and vanishes for h > 0, and thus, the Mayer function f (h) = e−β v̄(h) − 1
= −θ(−h). Because the potential has no energy scale (it is either zero or infinite),
the Mayer function does not depend on β. As a consequence, the only energy scale
in the problem is the temperature, which can thus be eliminated by expressing the
thermodynamic quantities in units of T . The only remaining control parameter is
thus the scaled packing fraction )
ϕ = 2d ϕ/d. Conveniently, an explicit expression
for the function q(,β;h) (which is also independent of β) can be derived:
 ∞  
dz 2
z
− 2 h + /2
qHS (;h) = √ e θ(h − z + /2) =  √ ,
−∞ 2π 2
(4.83)
1 1
(x) = (1 + erf(x)) = erfc(−x) .
2 2
The recipe of Section 4.3.3 can be applied using this analytical expression, but the
rest of the calculation has to be performed numerically [299]. The resulting phase
diagram is reported in Figure 4.4.
According to the analysis of Chapter 2, the liquid equation of state is p = 1
+ d) ϕ /2. The glass region corresponds to finite values of ) ϕ when d → ∞, and
in that case, the liquid reduced pressure is p ∼ d) ϕ /2; one can therefore define a
scaled pressure p ) = p/d, which in the liquid phase is given by p )= ) ϕ /2. As a first
step, one has to determine the dynamical transition point. Because qHS (;h) does
not depend on β, neither does F1 (r ), whose maximum (Figure 3.2) defines ) ϕd
= 4.8067 . . . according to Eq. (4.82). The corresponding pressure is p )d = ) ϕd /2
= 2.4033 . . . If )
ϕ<) ϕd , the liquid phase is ergodic and the corresponding dynamics
is diffusive, as discussed in Chapter 3.
 4.4 Replica Symmetric Phase Diagram of Simple Glasses 129

ϕd

g = 8, 7, 6.667, 6, 5.5, 5.25, 4.9

ϕ
1/p

Figure 4.4 Following hard-sphere glasses in compression and decompression,

within the replica symmetric ansatz [299]. The inverse of the scaled reduced
pressure p) = p/d is plotted versus the scaled packing fraction ) ϕ = 2d ϕ/d. The
)= )
liquid equation of state (full line) is p ϕ /2. The dynamical transition ) ϕd is
marked by a full circle. For ) ϕg > ) ϕd , the liquid is a collection of glasses. The
equations of state of glasses for different ) ϕg (dashed lines) intersect the liquid
equation of state at )
ϕg . Upon compression, for low ) ϕg they end at an unphysical
spinodal point (open circle), while at high )ϕg they end at a jamming point at which
pressure diverges. Upon decompression, the glass pressure falls below that of the
liquid, and reaches a minimum before growing again until a physical spinodal
point (open square) at which it melts into the liquid.

If )
ϕ > ) ϕd , the liquid dynamics is arrested. Glass states appear and each equi-
librium liquid configuration at ) ϕg > ) ϕd selects a glass. For each choice of ) ϕg , we
can follow the recipe of Section 4.3.3 to solve the equations for ,r and compute
the pressure of the corresponding glass. Recall that the energy is always zero for
hard-sphere systems. In Figure 4.4, the equations of state of several glasses, corre-
sponding to different choices of ) ϕg , are plotted. The pressure of the glass coincides
with that of the liquid at )
ϕg , implying that the pressure is continuous for each glass,
but the slope of the glass equation of state at ) ϕg is different from that of the liquid
equation of state. When the system becomes confined in the glass state selected
at )
ϕg , the compressibility thus has a jump. Following glasses in compression, the
pressure increases faster than in the liquid – i.e., the compressibility is smaller.
For larger values of )ϕg , it is found that, upon compression, the pressure increases
and eventually diverges at a finite ‘jamming’ density ϕ )j ()
ϕg ), where the mean square
130 Thermodynamics of Glass States

displacement of the glass  → 0. Jamming thus corresponds to the point at which

the glass ceases to exist; beyond that point, the hard-sphere exclusion constraints
cannot be satisfied while remaining within the same glass state. Because pressure is
infinite and  = 0, at jamming spheres touch their first neighbours. Jamming thus
also corresponds to the ‘close-packed’ limit of a given glass. Jamming points have
several interesting properties that will be carefully studied in Chapter 9. The glasses
prepared at lower ) ϕg , by contrast, show a different behaviour upon compression.
+
Above some density ) ϕsp ϕg ), the solution for ,r disappears, and the glass state
()
cannot be followed anymore. The disappearance of the solution happens through
a bifurcation as in a spinodal point. This spinodal point is, however, unphysical
because it is reasonable to expect that once the hard spheres are confined to a solid
phase, this solid phase can be compressed up to infinite pressure without becoming
unstable. Compression should only stabilise the solid. We will confirm in Chapter 6
that this spinodal point is, in fact, an artefact of the replica symmetric ansatz. It
disappears because replica symmetry is broken in that region.
Finally, a glass prepared at ) ϕg can also be followed in decompression – i.e.,
for )
ϕ <) ϕg . In this case, the glass pressure becomes lower than that of the liquid
until, upon decreasing density, a spinodal point ) ϕg ) is reached, at which the
ϕsp ()
solution for ,r disappears through a bifurcation. This spinodal point is physical
and corresponds to the glass becoming unstable and thus melting into the liquid.
At the spinodal point, due to its bifurcation nature,
 all thermodynamic quantities
display a square-root singularity – e.g.,  ∼ ) ϕsp − )ϕ , and similarly for r and
the pressure. Note also that the region of densities between the pressure minimum
and the spinodal is thermodynamically unstable because the compressibility is then
negative. Such an unstable region is an artefact of the mean field treatment.

4.4.2 Soft Spheres

We now discuss the simplest example of soft-sphere potential [315]: the pure
power-law repulsive potential, v(r) = ε(/r)νd , which corresponds to a scaled
potential v̄(h) = εe−νh (Section 2.3.2). The specific choice ν = 4 corresponds to
the repulsive part of a Lennard-Jones potential in d = 3, generalised to arbitrary
dimension. In Section 2.3.2, it was shown that the thermodynamics of this system
is controlled by a single parameter  = ρ/T 1/4 ; hence, temperature and density are
not independent. We can choose an arbitrary value of density and study the phase
diagram using temperature as a control parameter. The liquid equation of state has
also been studied in Section 2.3.2.
In d → ∞, it is convenient to use )  = ) ϕ (ε/T )1/4 as control parameter. The
dynamical transition point is )d = 4.304 . . . It is then convenient to work at fixed
density ) )
ϕ = d , such that Td /ε = 1. The results are reported in Figure 4.5. The
 4.4 Replica Symmetric Phase Diagram of Simple Glasses 131

Td

e

Tg = 0.1, 0.3, 0.5, 0.7, 0.8, 0.9, 0.95

Figure 4.5 Following soft-sphere glasses with potential v̄(h) = εe−4h , within the
replica symmetric ansatz, at fixed density ) ϕ = 4.304 . . . [315]. The evolution of
the reduced energy ) e = e/d with temperature T under heating and cooling is
reported, both quantities being expressed in units of the interaction strength ε. On
the liquid equation of state (solid line), dynamical arrest happens at the dynamical
transition Td = 1 (solid circle). For Tg < Td , the liquid is a collection of glasses.
The glass equations of state are reported for several choices of Tg (dashed lines)
and intersect the liquid equation of state at Tg . Upon cooling, for high Tg , they
end at an unphysical spinodal point (open circle), while for low Tg , they can be
continued down to zero temperature. Upon heating, the glass energy falls below
that of the liquid, until a physical spinodal point (open square) is reached, at which
the glass melts into the liquid.

phase diagram is very similar to that of hard spheres, identifying temperature with
inverse pressure and inverse density with energy. For T > Td , the liquid dynamics
is diffusive, while for T < Td , it is arrested. Equilibrium liquid configurations at
Tg < Td select a glass state that can be followed to both lower and higher temper-
atures. The specific heat (i.e., the derivative of the energy with respect to tempera-
ture) has a discontinuous jump at Tg , the specific heat of the liquid being larger than
that of the glass. Upon cooling, the glasses with lower Tg can be followed down to
T = 0, while glasses with higher Tg display an unphysical spinodal before T = 0
can be reached. Upon heating, the glass energy remains lower than that of the liquid,
until a physical spinodal point Tsp (Tg ) is reached, at which the glass melts into
the liquid.
132 Thermodynamics of Glass States

4.5 Wrap-Up
4.5.1 Summary
In this chapter, we have seen that

• In the dynamically arrested region, equilibrium configurations remain stuck

inside a glass state. The averages of dynamical observables, in the long time limit,
reach a stationary value (Section 4.1.1). These long-time dynamical averages can
be computed by defining a probability measure over phase space, restricted to
a glass state. This procedure, which goes under the name of the Franz–Parisi
construction (Section 4.1.2), allows one to compute the free energy of a glass
state that was prepared in equilibrium at () ϕg,Tg ) and then followed to ()
ϕ,T ).
• The computation of the Franz–Parisi potential (or glass free energy) requires
the introduction of replicas of the system to average over the quenched disorder
(Section 4.1.3).
• The replicated free energy corresponds to the free energy of a molecular liquid.
The methods of Chapter 2 can also be generalised to a liquid made of molecules,
if the limit d → ∞ is taken by imposing a strong localisation of the molecules,
expressed by Eq. (4.29) with finite α̂ (Section 4.2.1). The order parameter is
the molecular density π(u) and the resulting free energy is the sum of an ideal
(molecular) gas term plus a two-molecule interaction, Eq. (4.20).
• When d → ∞, under the same hypothesis of strong molecule localisation, the
dependence of the molecular free energy Eq. (4.20) on the molecular distribution
π(u) is greatly simplified.
 a In  fact, the free
 energy can
 be expressed as a function
of the matrix αab ∝ u · u or ab ∝ |u − u | only, resulting in Eq. (4.54)
b a b 2

or Eq. (4.56), respectively (Sections 4.2.2 and 4.2.3). The matrices α̂ or  ˆ are
then determined by minimising the free energy. The mean square displacement
matrix  ˆ corresponds to the dynamical mean square displacement in the long
time regime.
• The number of replicas s should be sent to zero to extract the glass free energy
from the replicated free energy. This requires an analytical continuation. To sim-
plify the procedure, the symmetry of the free energy under permutations of the
replicas – i.e., replica symmetry – is used (Section 4.3.1). Assuming that the
replica symmetry is not broken, the matrix  ˆ depends only on two parameters, 
and r , which have a simple physical meaning. The expression of the glass free
energy simplifies considerably, and the analytic continuation to s → 0 can be
performed explicitly (Section 4.3.2). Minimisation of the free energy becomes an
extremisation in the s → 0 limit, so the analytically continued free energy should
then be extremised with respect to  and r .
• The main result of this chapter is the final recipe to compute the thermodynam-
ical properties of the glass states of any potential v̄(h), under the assumption of
unbroken replica symmetry (Section 4.3.3).
 4.5 Wrap-Up 133

• Results are given in Section 4.4 for hard spheres and soft repulsive spheres. The
phase diagrams, with the liquid and glass equations of state, are computed in both
cases.

4.5.2 Further Reading

We provide here a list of references that can be consulted to further explore the
subjects discussed in this chapter, selected according to the criteria discussed in
Section 1.6.2.
The state following construction was applied to structural glasses in d = 3
shortly after it was formulated. The replica-HNC approximation used in early
papers gives good results close to the equilibrium liquid line but unfortunately
fails when glasses are brought to high densities or low temperatures. Better
approximation schemes for replicated liquids in finite dimensions have thus been
subsequently developed. The relevant papers include

• Mézard and Parisi, A tentative replica study of the glass transition [252]
• Cardenas, Franz and Parisi, Constrained Boltzmann–Gibbs measures and
effective potential for glasses in hypernetted chain approximation and numerical
simulations [77]
• Mézard and Parisi, Glasses and replicas [253]
• Parisi and Zamponi, Mean-field theory of hard-sphere glasses and jamming [292]
In d → ∞, the method has been applied to variations of the square-well potential
discussed in Section 2.3.2 that model attractive or patchy colloids. For these models,
the function F1 (;β) defined in Eq. (3.65) can have multiple maxima, which leads
to the Franz–Parisi potential having multiple minima that describe possible values
of the plateau  of the mean square displacement. Details can be found in

• Sellitto and Zamponi, A thermodynamic description of colloidal glasses [325]

• Maimbourg, Sellitto, Semerjian et al., Generating dense packings of hard spheres
by soft interaction design [240]
• Yoshino, Translational and orientational glass transitions in the large-
dimensional limit: A generalized replicated liquid theory and an application
to patchy colloids [367]
Similar phenomena have been previously found within MCT, where the coexistence
of multiple glass solutions leads to non-trivial dynamical effects. An introductory
review can be found in Sciortino and Tartaglia, Glassy colloidal systems [323].
The state following method has also been used in the context of random
optimisation problems, of interest for computer science. The approach then
provides various insights into the structure of their energy landscape and the
performances of algorithms trying to find good minima. A selection of relevant
references includes
134 Thermodynamics of Glass States

• Krzakala and Kurchan, Landscape analysis of constraint satisfaction problems

[215]
• Zdeborová and Krzakala, Generalization of the cavity method for adiabatic
evolution of Gibbs states [371]
• Krzakala and Zdeborová, Performance of simulated annealing in p-spin glasses
[216]
In the context of particle glasses in finite dimensions, as mentioned in Sec-
tion 3.5.2, the dynamical transition is not sharp. The Franz–Parisi construction
described in this chapter provides a simple way of understanding this observation.
In finite dimensions, the Franz–Parisi potential must be convex. As discussed in
Chapter 1, a non-convex region indicates phase coexistence, which is realised via
nucleation of the stable liquid phase into the metastable glass phase. In mean field,
the glass state corresponds to a local minimum of the potential, and its lifetime is
infinite, while in finite dimensions, it is not. Numerical measurements of the Franz–
Parisi potential support this scenario. A selection of relevant references is

• Mézard and Parisi, Statistical physics of structural glasses [257]

• Parisi, Glasses, replicas and all that [286]
• Cammarota, Cavagna, Giardina, et al., Phase-separation perspective on dynamic
heterogeneities in glass-forming liquids [73]
• Berthier, Overlap fluctuations in glass-forming liquids [39]
In addition to the dynamical processes mentioned in Section 3.5.2 that destroy the
sharp dynamical transition observed in mean field, the restricted thermodynamic
analysis developed in this chapter suggests that nucleation could help escape from
metastable glass states. The identification of this process initiated the development
of the random first-order transition (RFOT) theory of the glass transition beyond
the mean field limit. Details on the nucleation processes within the RFOT approach
can be found in

• Kirkpatrick, Thirumalai and Wolynes, Scaling concepts for the dynamics of

viscous liquids near an ideal glassy state [208]
• Cavagna, Supercooled liquids for pedestrians [80]
• Berthier and Biroli, Theoretical perspective on the glass transition and amor-
phous materials [40]
• Wolynes and Lubchenko (eds), Structural glasses and supercooled liquids:
Theory, experiment, and applications [357]
Thanks to the various escape processes mentioned previously, finite-dimensional
systems can leave the metastable glass minimum of the Franz–Parisi potential. The
liquid phase can thus be equilibrated even for Tg < Td (ρg ). Because of the different
 4.6 Appendix 135

microscopic time scales, this process is most efficient in structural glasses, while it
is much less efficient in colloids, emulsions and grains.
Recent algorithmic and experimental developments further accelerate this sam-
pling to efficiently achieve equilibration for Tg < Td (ρg ). In numerical simulations,
‘swap’ algorithms can be used to speed up the equilibration process, while in
experiments, vapour deposition techniques achieve the same goal. Examples of
these recent developments can be found in

• Swallen, Kearns, Mapes, et al., Organic glasses with exceptional thermodynamic

and kinetic stability [338]
• Ninarello, Berthier and Coslovich, Models and algorithms for the next generation
of glass transition studies [273]
• Fullerton and Berthier, Density controls the kinetic stability of ultrastable
glasses [157]

These methods provide very well equilibrated liquid configurations, corresponding

to glass basins that lie very deep in the free energy landscape. One can thus use
these starting configurations to perform simulations or experiments on time scales
much shorter than the relaxation time of the liquid, in such a way that the system
is effectively confined into a glass basin. The equation of state of thermal glasses
(Figure 4.5) can then be directly compared with the results obtained experimentally
in [338, figure 1B]. The sharp peak in the specific heat reported in [338, figure 1A]
during heating indeed mirrors the divergence of the slope of ) e(T ) at the melting
spinodal in Figure 4.5. The equation of state of hard-sphere glasses (Figure 4.4) can
be compared with the numerical results of [157, figure 1]; the mean field melting
spinodal is hidden, in finite dimensions, by nucleation processes.

4.6 Appendix
4.6.1 Determinant of the Scalar Product Matrix
We consider an n × n symmetric matrix γ̂ that satisfies γ1a = γa1 = 0 for
all a = 1, . . . ,n, and a second symmetric matrix  ˆ defined by the relation ab
= γaa + γbb − 2γab . Note that this definition implies aa = 0, ∀a. We want to
express the determinant of the (1,1) cofactor γ̂ (1,1) = α̂ – i.e., the (n − 1) × (n − 1)
matrix obtained by removing from γ̂ the first (vanishing) line and column, as a
function of . ˆ This determinant can be computed using different techniques, and
we present here one method that is particularly simple.
ˆ From the definition
As a first step, it is convenient to express γ̂ as a function of .
ˆ
of , one has a1 = γaa , and therefore, γab = (a1 + 1b − ab )/2, which
holds for all a,b = 1, . . . ,n. Next, we consider a Gaussian integral representation
of det γ (1,1) :
136 Thermodynamics of Glass States

1 dy2 · · · dyn − 1 na,b=2 αab ya yb
 =  e 2
det(γ̂ (1,1) ) (2π)n−1
 (4.84)
dy2 · · · dyn − 1 na,b=2 (a1 +1b −ab )ya yb
=  e 4 .
(2π)n−1

One can then add to the Gaussian integration an additional variable y1 = − na=2 ya
and express the argument in the exponential by means of this additional variable
(recall that 11 = 0) by the identity
 . n /
1 dy1 · · · dyn  1 n
 =  δ ya e 4 a,b=1 ab ya yb
det(γ̂ (1,1) ) (2π)n−1 a=1
 
dλ dy1 · · · dyn iλ na=1 ya + 1 na,b=1 ab ya yb
=  e 4
2π (2π)n−1
 (4.85)
1 dλ 2  ˆ −1
=! √ eλ ab ( )ab
ˆ
det(−/2) 2π

1
=! ,
ˆ
2 det(−/2)(−1T ˆ −1
 1)
where 1 = {1, . . . ,1} is the vector of all ones. We therefore obtain
ˆ
det α̂ = 2 det(−/2)(−1 T ˆ −1
 1). (4.86)

4.6.2 Gaussian Convolutions

We collect here some useful identities involving Gaussian integration and convo-
lution that are used throughout this chapter. We begin by defining the Gaussian
convolution as in Eq. (4.69),
 ∞
 d2 dz z2
γ  r(h) = e 2 dh2
r(h) = √ e− 2 r(h − z), (4.87)
−∞ 2π
where r(h) is such that the expressions in Eq. (4.87) are well defined and otherwise
arbitrary. The identity of the second and third expressions in Eq. (4.87) can be
 d2
proven by expanding e 2 dh2 in powers of  in the second expression and expanding
r(h − z) in powers of z in the third.
Eq. (4.87) can be generalised to the multidimensional case as
k  ∞ k k
1 k ∂2 dza 1 k −1
e− 2 a,b=1 za M̂ab zb
M
e2 a,b=1 ab ∂ha ∂hb
r(ha ) = ! a=1 r(ha − za ),
−∞ k
a=1 (2π) det M̂ a=1

(4.88)
 4.6 Appendix 137

where M̂ is a k×k symmetric and invertible matrix, and r(h) is once again such that
the expressions are well defined and otherwise arbitrary. The proof of Eq. (4.88) can
be obtained similarly to Eq. (4.87), by expanding the exponential of the differential
operator in a power series of M̂ in the left-hand side of the identity and expanding
r(ha − za ) in a power series of za in its right-hand side.
Another useful identity is

 d2   d2
e2 dh2 [eh r(h)] = eh+ 2 e 2 dh2 r(h + ) , (4.89)

which can be proven either by series expansion in  or by an appropriate change

of variable in the integral representation in Eq. (4.87).
Finally, by defining a matrix γ̂ as in Section 4.6.1, we obtain the identity
 5 6
∞
−β v̄(h)
n ∂2  −β
n 00
dh e h
e e a,b=2 αab ∂ha ∂hb
e a=2 v̄(ha +αaa ) 0 −1
−∞ ha =h
 5 6
∞ n ∂2
a,b=1 γab ∂ha ∂hb +
n ∂
a=1 γaa ∂ha
 −β
n 
= dh e h
e e a=1 v̄(ha ) −1 (4.90)
−∞ ha =h
 5 6
∞
− 12
n ∂2
a,b=1 (γaa +γbb −2γab ) ∂ha ∂hb
 −β
n 
= dh e h
e e a=1 v̄(ha ) −1 .
−∞ ha =h

To prove the last equality in Eq. (4.90), one can replace

n ' (
∂2 − 12 n ∂ d
− 12 a,b=1 γaa ∂ha ∂hb a=1 γaa ∂ha
e →e dh
(4.91)

when these operators are acting inside the integral, which is an application of
Eq. (4.71). Then one can expand the exponential in Taylor series and observe that
d k
dh
can be integrated by parts, acting on the term eh , giving rise to a factor (−1)k .
Then one can replace
' (
n n
− 12 ∂
a=1 γaa ∂ha
d 1 ∂
e dh
→ e2 a=1 γaa ∂ha . (4.92)

4.6.3 Gaussian Ansatz for the Glass Free Energy

In this section, we briefly sketch the derivation of the glass free energy based on
a Gaussian ansatz for the molecular density ρ(x). This derivation is simpler than
the one discussed in Section 4.2 but relies on an additional, and a priori unjustified,
assumption.
We start from the free energy of the molecular liquid truncated at the second
order in the virial expansion, as given by Eq. (4.20), but neglecting the kinetic
contribution, which is but an irrelevant constant,
138 Thermodynamics of Glass States
 
1
−βF [ρ] = dxρ(x)[1 − log ρ(x)] + dxdyρ(x)ρ(y)f (x − y),
2

s+1 (4.93)
f (x − y) = e−βg vg (x1 −y1 ) e−βv(xa −ya ) − 1.
a=2

Here, as discussed in Section 4.2.4, the first replica has temperature Tg and a poten-
tial vg (r) which contains an interaction length scale g , while the other replicas
have temperature T and a potential v(r) containing . As in Section 4.2.2, we then
perform the change of variable ua = xa − x1 , as given in Eq. (4.21), and introduce
the normalised distribution π(u) = ρ(u)/ρ. We assume that π(u) is Gaussian, with
average and covariance given by Eq. (4.29) – i.e.,

1 1  s+1 a −1 b 2
π (u) = e− 2 μ a,b=2 uμ Aab uμ ,  = α̂. (4.94)
[(2π)s det Â]d/2 d

Under this assumption, the calculation of the ideal gas term is straightforward.
Starting from Eq. (4.41), we obtain
 
1
−βfid = dxρ(x)[1 − log ρ(x)] = 1 − log ρ − duπ (u) log π (u)
N
d d
= 1 − log ρ + log[(2π)s det Â] + Aab A−1
ab (4.95)
2 2 a,b

d ds
= 1 − log ρ + log det  + log(2πe).
2 2
The excess term is also easily evaluated. Starting from Eq. (4.23), we get

ρ
−βf =ex
dR dudv π (u)π (v) f (u − v + R). (4.96)
2
Given that u and v are independently and identically distributed according to π (u),
their difference w = u − v is also Gaussian, with twice the covariance, such that

ρ
−βf =ex
dR dwπ2Â (w) f (w + R)
2
  
s+1
ρ
= dR e−βg vg (R) dwπ2Â (w) e−βv(wa +R) − 1 (4.97)
2 a=2
⎧ 0 ⎫
 ⎨  s+1 ∂2  0 ⎬
ρ Aab s+1 0
dR e−βg vg (R) e
a ∂wb −β
a=2 v(wa +R) 0
μ a,b=2
= ∂wμ μe −1 ,
2 ⎩ 0 ⎭
w=0
 4.6 Appendix 139

where in the last line we used the identity in Eq. (4.88). The reader can check that
the total free energy f = fid + fex coincides with Eq. (4.54), if the limit d → ∞ is
taken with the appropriate scalings.
The replica symmetric matrix has the form Aab = Dr δab + (Dr − D /2)(1 − δab ).
Plugging this form into the replicated free energy, the determinant and the Gaussian
convolution can then be evaluated explicitly and the limit s → 0 can be taken to
extract the glass free energy, as in Section 4.3.2. For completeness, we give the final
result:
d d 2 Dr − D
−βfg (ϕ,T |ϕg,Tg ) = log(πe D) +
2 2 D
 (4.98)
ρ
+ d dr r d−1
qg (2 Dr − D ,βg ;r) log q(D ,β;r),
2
where q(D ,β;r) is given in Eq. (3.61), and qg (2 Dr − D ,βg ;r) is defined with the
potential vg (r). Details on how to take the limit d → ∞ of this expression can
be found in [292], and the result coincides with Eq. (4.75). Note that the equation
for the mean square displacement plateau (and then for the dynamical transition)
obtained from this expression coincides with Eq. (3.60) in any finite dimension d.
 5
Replica Symmetry Breaking and Hierarchical
Free Energy Landscapes

In Chapter 4, we have illustrated how the replica method can be a natural tool to
compute the properties of the glass states of simple systems, despite the absence
of quenched disorder in the interaction potential. The Franz–Parisi construction
provides a way to average over an ensemble of glass states selected by a reference
equilibrium configuration. The replica symmetric glass phase diagram has then
been computed. Both in the case of hard-sphere and soft-sphere potentials, the
replica symmetric solution predicts some unphysical features – in particular, a
spinodal instability of the glass state upon compression or cooling. Something
might thus be wrong with the assumption of replica symmetry, which motivates
a deeper investigation of its validity.
In this chapter, we clarify the physical meaning of the replica symmetry. In
particular, we discuss how, as for ordinary symmetries, it can be spontaneously
broken at phase transitions and what the ensuing consequences are. Replica sym-
metry breaking was first discovered in the context of spin glasses, and it has since
been applied in physics and in research areas as distant as computer science and
neural networks. Because the subject is already discussed in great detail in several
books [10, 138, 251, 254, 274], we here only provide a compendium of the basic
notions and tools needed for the rest of the book. The reader who is not interested
in the details of replica symmetry breaking can skip this chapter in a first reading.
Note that in this chapter, we provide general results about the replica formalism
while, in Chapter 6, we apply them to the specific case of particle glasses in the
limit d → ∞.

5.1 An Introduction to Replica Symmetry Breaking

Replica symmetry may seem different from ordinary symmetries of physical
systems. For example, in ferromagnetic systems without an external magnetic
field (see Chapter 1), the spin inversion symmetry Si → −Si is a symmetry of

140
 5.1 An Introduction to Replica Symmetry Breaking 141

the Hamiltonian. It then plays a crucial role in understanding the thermodynamic

properties of the model – e.g., its thermodynamic pure states in the low-temperature
phase, as discussed in Chapter 1. By contrast, in glassy systems, the Hamiltonian
does not display any symmetry beyond the trivial translational and rotational
symmetries. In particular, once the system is frozen into a glass state, particles
interact in a random environment fixed by the initial configuration which breaks
translational and rotational invariance, and therefore, the Hamiltonian is not invari-
ant under any symmetry group. Replica symmetry only emerges as a consequence
of the average over all possible glass states, which is encoded by the average over
the initial configuration. One may therefore wonder if replica symmetry has any
physical meaning or if it is just a technical artefact. Before going into the details
of replica symmetry breaking, we thus provide here a short discussion of the main
physical properties behind this phenomenon.

5.1.1 Trivial Replica Symmetry Breaking

We have seen in Chapter 1 that in the ferromagnetic Ising model, at low temper-
atures and zero magnetic field, there are two translationally invariant equilibrium
states, corresponding to magnetisation ±meq . There may also be non-translationally
invariant equilibrium states, with a free energy that is higher by a factor proportional
to Ld−1 , where L is the linear size of the system.
The existence of more than one equilibrium state is common in physics. For
instance, it occurs everywhere along the gas–liquid coexistence line. In order to
generalise this notion, let us summarise the procedure to monitor the presence of
two phases first described in Chapter 1. We consider a spatially dependent observ-
able O(x) that takes a different value in the two phases, such as the magnetisation
m(x) in spin systems and the local density ρ(x) near the gas–liquid transition, and
add to the Hamiltonian a small field  conjugated to this observable:

H = − dxO(x). (5.1)

Denoting • the Gibbs–Boltzmann average in the thermodynamic limit with the
additional field, we can define •± = lim→0± • . Phase coexistence is observed
if O(x)− = O(x)+ .
We next consider the slightly more complex problem of a liquid that crystallises.
In this case, we have an infinite number of equilibrium states, which differ from
each other by a global translation. If one does not know the structure of the crystal,
it is difficult to identify an observable O(x) that takes a different value in each of
the many possible phases, and the previous construction is not viable. A possible
way out of this problem, introduced by Edwards and Anderson [136], consists in
142 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

considering two replicas (or clones) X and Y of the same system and writing the
total Hamiltonian as
H (X,Y ) = H (X) + H (Y ) + N D(X,Y ), (5.2)
where
1  N
D(X,Y ) = min |xi − yP (i) |2 (5.3)
N P ∈PN i=1

is the mean square displacement between the two configurations. It is important to

note that, with respect to the definition given in Eq. (4.1), we added here a min-
imisation over all possible permutations P ∈ PN of the N particles. In Chapter 4,
we were interested in using Y as a reference configuration to constrain X into a
restricted equilibrium. Here, by contrast, we want to treat X and Y on equal footing
in a full equilibrium, and therefore, a minimisation over particle permutations is
required to define a meaningful distance. The Hamiltonian in Eq. (5.2) is invariant
under the Z2 group, whose elements exchange the two replicas and which coincides
with replica symmetry in this context.
In a crystal phase, the limits  → 0± produce different results for many observ-
ables that involve two replicas, while single-replica observables have the same value
in the two limits. In particular, for  > 0, the two configurations are in the same
crystal, and the value of the mean square displacement D() = D(X,Y )  is as
small as possible; it would be D() = 0 at zero temperature. By contrast, for
 < 0, one configuration is at the maximum distance (of the order of half the
lattice spacing) from the other, and D() takes the largest possible value. The point
 = 0 is singular, thus indicating a phase transition.
Note that at  = 0, one recovers the original Hamiltonian in which the two
replicas are independent, but the structure of equilibrium states indicates that replica
symmetry is then spontaneously broken. In order to show this, we focus, for sim-
plicity, on the case of a simple cubic lattice with lattice spacing equal to unity. Each
replica then occupies a cubic lattice, and the two lattices are shifted by an arbi-
trary vector a ∈ [−1/2,1/2]d , because there is no coupling between the replicas.
The corresponding value of mean square displacement is D(X,Y ) = |a|2 , and its
probability distribution is

P (D) = da δ(|a|2 − D). (5.4)
[−1/2,1/2]d

A distribution P (D) different from a delta function then signals the presence of
multiple states and replica symmetry breaking. Note, however, that in this case, as
soon as  = 0, replica symmetry breaking disappears and P (D) becomes a delta
function.
 5.1 An Introduction to Replica Symmetry Breaking 143

5.1.2 Non-Trivial Replica Symmetry Breaking

In the example of crystallisation discussed in Section 5.1.1, replica symmetry
breaking is a consequence of an underlying spontaneous breaking of an internal
symmetry – i.e., translations. In this case, the effect disappears if we change the
definition of the distance between two configurations in a way that respects the
internal symmetry,

1  N
D(X,Y ) = min min |xi − yP (i) + a|2, (5.5)
N a∈Rd P ∈PN i=1

that is, if we define the distance as the minimum with respect to all possible trans-
lations of one configuration with respect to the other. With this new definition,
P (D) is a delta function both in the liquid and in the crystal phases, and no replica
symmetry breaking occurs. This treatment is easily generalised to any situation in
which an internal symmetry is spontaneously broken.
Phase coexistence, which is the essence of replica symmetry breaking, is present
at first-order transition points, such as the gas–liquid transition. In this case, only
two phases are present. Three phases are present at the triple point: gas, liquid and
solid [175]. The Gibbs phase rule states that in order to have coexistence of K + 1
phases, we need to tune K control parameters. To prove this, let us label phases by
an index α. The probability wα that the system is in phase α at equilibrium in a
finite volume is given by
e−βFα  
wα = , Z= e−βFα , wα = 1, (5.6)
Z α α

where Fα is the total free energy of phase α. Then, consider adding a perturbation
H to the Hamiltonian, which perturbs the free energies as
Fα → Fα + H α , (5.7)
where •α denotes the equilibrium average restricted to phase α (as discussed
in Chapter 1). Because the different phases have different structures (consider,
e.g., gas, liquid and solid phases), it is very unlikely that the H α are equal.
The perturbed free energies are then also different. Because the difference is of
order N , only one phase survives in the thermodynamic limit. The only way to
remain at coexistence is to impose that all the H α be equal, which imposes K
conditions on the control parameters. This proves the Gibbs phase rule. In other
words, having K + 1 equal free energies is an unlikely event that is destroyed by a
small perturbation.
The preceding argument, however, fails in disordered systems, because one can
then construct many distinct phases that have statistically identical properties. One
144 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

can think, for instance, of the many glasses that can be formed by a same system
of interacting particles. Suppose that at a given state point, in the thermodynamic
limit, there is an infinite number of coexisting states with free energy Fα , and
a perturbation is added to the Hamiltonian. Because the states have statistically
independent properties, H α is then self-averaging and thus independent of α.
The shift of the extensive free energy, Fα , is therefore the same for all states.
More concretely, consider adding a small perturbation v(r) to the pair interaction
potential of a particle system. According to Eq. (2.29), the variation of the free
energy of state α is

Nρ
Fα = drgα (r)v(r). (5.8)
2
If the coexisting phases are structurally different, such as a gas and a crystal,
then their gα (r) are different, which results in different Fα . By contrast, if the
coexisting phases are microscopically different glasses, gα (r) is the same, as is well
known numerically and experimentally [40, 80, 120], resulting in identical Fα .
Thanks to disorder, infinitely many phases might thus coexist in a whole region of
parameter space.
Note that the coexistence of infinitely many phases could also be described
within a fully probabilistic framework, known as the ‘cavity method’ [254], thus
avoiding the use of the replica formalism. We do not make this choice in this book
for three reasons:

• The probabilistic arguments used in the cavity method are subtle and not easy
to follow for a reader who is not already familiar with the structure of a replica
symmetry broken phase.
• The cavity method is much easier to develop for lattice models. The absence of
an underlying lattice for particle systems complicates the derivation.
• The replica approach is computationally much easier to handle, and new appli-
cations are often first developed using this method. This is certainly the case for
particles in d → ∞.
Therefore, here we use the very compact algebraic replica formalism, even if this
may sometimes hide the underlying physical interpretation. Hopefully, the main
elements of its physical interpretation provided in this chapter somewhat compen-
sate for this choice.

5.1.3 Stability of the Replica Symmetric Solution

It has been shown in Chapter 4 that the thermodynamic properties of glass states
can be derived, within the Franz–Parisi construction, by minimising the replicated
free energy fs+1 () ˆ defined by Eq. (4.59), with respect to the n × n mean
ϕa,Ta, ),
 5.1 An Introduction to Replica Symmetry Breaking 145

square displacement matrix . ˆ Then, one should take the analytic continuation of
the minimum to real s = n − 1 and take the limit s → 0. In order to perform this
operation, in Chapter 4 it was assumed that the matrix  ˆ is invariant under replica
permutations – i.e., it has the ‘replica symmetric’ form,  ˆ RS , given by Eq. (4.62)
and illustrated by Eq. (4.63). The assumption of replica symmetry allows one
to perform the analytic continuation straightforwardly. However, the replica
symmetric form was taken as an assumption, and we did not check that the
minimum of the free energy is really assumed on matrices of this form.
For the ferromagnetic model discussed in Chapter 1 the paramagnetic minimum
at m = 0 can become unstable for two reasons. Either the curvature of the free
energy f (m = 0) vanishes, leading to a linear instability of the minimum, identified
with a second-order phase transition. Or another minimum appears at m = 0, and
its free energy becomes smaller than f(m = 0), thus leading to a discontinuous
(first-order) transition.
By analogy, a minimal way to support the replica symmetric assumption is to
check that  ˆ RS is a stable local minimum of the replicated free energy.1 This check
amounts to excluding the possibility of a second-order phase transition. Although
this does not exclude the existence of another free energy minimum characterised
by replica symmetry breaking – i.e., a first-order transition – it at least excludes the
possibility that the replica symmetric solution is unstable to small fluctuations. To
perform this check, we need to compute the stability operator (or Hessian) of the
free energy, defined as
0
∂ 2 fs+1 00
Hab;cd = , a < b, c < d. (5.9)
∂ab ∂cd 0= ˆ 
ˆ RS

We define the replica symmetric solution to be locally stable if all the eigenvalues
of H are positive. Recall that the matrix  ˆ is by definition symmetric and has
aa = 0. Hence, only elements with a < b can be varied independently, and the
Hessian is a n(n − 1)/2 × n(n − 1)/2 symmetric matrix. Note that a pair a < b is
here considered as a single index taking n(n − 1)/2 values.
The positivity of the Hessian matrix given by Eq. (5.9) is a natural condition.
The lowest-order corrections to mean field theory indeed contain terms that involve
the matrix
(H−1 )ab;cd = χab;cd = ab cd  − ab  cd  . (5.10)
The matrix χab;cd is positive definite by construction, analogously to the mag-
netic susceptibility defined in Eq. (1.9). Because it is the inverse of the Hessian, it

1 Recall that in the analytical continuation to s < 1 a minimum may become a saddle point. The correct
procedure is thus to check the stability for s > 1 and then analytically continue the eigenvalues to s < 1.
146 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

diverges when the Hessian develops a zero eigenvalue and becomes (unphysically)
negative in the region where the Hessian has negative eigenvalues. The divergence
of χab;cd thus signals a phase transition, and the presence of negative eigenvalues in
a positive definite matrix signals that the replica symmetric computation is incon-
sistent, because a phase transition has been ignored.
Replica symmetry strongly constrains the form of the Hessian. Its eigenvalues
can thus be computed explicitly. We do not report this calculation here, but it can
be found in several classic references [113, 254]. In many models, this computation
reveals that at some point in parameter space, one of the eigenvalues of the stability
operator vanishes, signalling the breaking of replica symmetry. The set of degener-
ate eigenvectors φab that correspond to a vanishing eigenvalue are called ‘replicon’

modes and satisfy the condition b φab = 0. This phase transition was first discov-
ered by de Almeida and Thouless in the context of the equilibrium thermodynamics
of the Sherrington–Kirkpatrick model [113], a mean field spin glass model. In that
case, the phase transition separates the high-temperature paramagnetic phase from
the low-temperature spin glass phase. In other models, the same transition was also
shown to happen inside a glass state, by Gross, Kanter and Sompolinsky [171]
and by Gardner [160], who first computed exactly the transition point in a given
model. In this context, the phase transition is known as a Gardner transition. We
will describe more precisely this transition in Chapter 6.
In all these models, it has been shown that the correct structure of the solution in
the replica symmetry broken phase is given by ‘hierarchical replica matrices’ [254].
The structure of these matrices allows one to perform a simple analytic continuation
to real s and then take the limit s → 0, as in the replica symmetric case. It
has been proven mathematically that this solution gives the correct free energy
and that it correctly describes the underlying ultrametric structure of pure states,
which we discuss in Section 5.3.3. In the context of infinite dimensional structural
glass models, there is no rigorous proof that the hierarchical construction gives
the correct free energy. Nevertheless, we will make this assumption because it
provides physically consistent results. The consequences of this assumption will
be discussed in Chapter 6.
In the rest of this chapter, we show how to compute the transition point based on
the hierarchical matrix construction, without explicitly obtaining the eigenvalues
of the stability operator defined by Eq. (5.9). We also show how to compute the
properties of the low-temperature phase, first in an expansion around the transition
point and then by solving the replica symmetry broken equations.

5.2 The Algebra of Hierarchical Matrices

In this section, we discuss some of the mathematical preliminaries needed for the
replica formalism. In particular, because in all known cases the replica symmetry
 5.2 The Algebra of Hierarchical Matrices 147

broken phase displays a hierarchical structure in replica space, we introduce and

discuss the algebraic properties of this structure. Note that within the Franz–Parisi
construction, one has n = 1 + s replicas. The first one is special, and the other s are
all equivalent. Considerations of replica symmetry are thus restricted to the s × s
sub-block of the matrices with indices a,b = 2, . . . , s + 1. Here we consider a
simplified setting, in which there are only s equivalent replicas – i.e., we momen-
tarily discard the special replica. To generalise the discussion to the Franz–Parisi
construction, it will suffice to add back the special replica with a = 1 at the end
(see Section 5.3).

5.2.1 k-Step Replica Symmetry Breaking

In the following, we construct the algebra of hierarchical matrices step by step. We
consider s × s real symmetric matrices q̂, whose elements are denoted by qab , and
for notational simplicity, we use indices a,b = 1, . . . , s.
Replica Symmetric (RS) Algebra
The replica symmetric algebra is composed of matrices of the form

qaa = qd , qa=b = q. (5.11)

It is easy to check that these matrices form a closed commutative algebra, in the
sense that the product of two RS matrices is still a RS matrix, and the order of the
matrices in the product is irrelevant.

One-Step Replica Symmetry Breaking (1RSB) Algebra

Consider an integer m0 , such that s is divisible by m0 – i.e., s mod m0 = 0. We
divide the replica indices in s/m0 groups, and define ceil(x) = x, a function
that returns the smallest integer larger<than= or equal to x. If a ∈ {1, . . . ,s}, then
replica a belongs to the group la = m0 , with la(1) ∈ [1, . . . ,s/m0 ]. The 1RSB
(1) a

parametrisation of q̂ is then given by

⎧
⎨qd if a = b,
⎪
qab = q1 if a = b but la(1) = lb(1), (5.12)
⎪
⎩ (1)
q0 if la = lb .
(1)

In other words, the diagonal is set to qd , and the off-diagonal elements are equal to
q1 if a,b are in the same group and to q0 if they are in a different group. A graphical
representation of this construction is given in Figure 5.1.

Two-Step Replica Symmetry Breaking (2RSB) Algebra

The 2RSB parametrisation can be obtained starting from the 1RSB one. Consider
an integer m1 , such that m0 mod m1 = 0. We divide each group la(1) = 1, . . . ,s/m0
148 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

m0

q1
qd
q1

q0

q0
q1
qd
q1

Figure 5.1 Illustration of a matrix q̂ of the 1RSB form.

> ?
in m0 /m1 subgroups of replicas, labelled by la(2) = a
m1
, each group containing m1
replicas. If two replicas a and b belong to different groups, la(1) = lb(1) , the matrix
element qab = q0 is unchanged. Let us now consider the case in which a and b
belong to the same group, la(1) = lb(1) . On the diagonal, we still have qaa = qd . If a
and b belong to the same subgroup, la(2) = lb(2) , the matrix element is qab = q2 , while
it remains qab = q1 for la(2) = lb(2) . The graphical representation of this construction
is given in Figure 5.2.

From kRSB to (k + 1)RSB: Iterative Construction

This construction can be iterated an arbitrary number of times, provided s is large
enough to find integers that divide the matrix in subgroups. Let us suppose that a
kRSB parametrisation has been constructed. We denote its parameters

{s = m−1,m0, . . . ,mk−1,mk = 1} → {q0, . . . ,qk ;qd }, (5.13)

where we defined mk = 1 and m−1 = s for later convenience. We also have

s ≥ m0 ≥ . . . ≥ mk−1 ≥ mk = 1. The RS case corresponds to k = 0, and we
have already discussed the cases k = 1,2.
To move from kRSB to (k + 1)RSB, we change the value of mk = 1 to a number
mk > 1, such that mk−1 mod mk = 0, and we add a new mk+1 = 1. Then, we
consider the innermost blocks of size mk−1 and divide the replicas that belong to
each of these blocks in mk−1 /mk groups of mk replicas. We keep all matrix elements
as before, except if a pair of replicas a and b belong to the same subgroup (but are
 5.2 The Algebra of Hierarchical Matrices 149
m0

q1
qd
q1

q0

m1
q0
q2
q1 qd
qd q2
q1

q1

m0

q1
q2
qd
q2

Figure 5.2 Illustration of a matrix q̂ of the 2RSB form.

not identical), in which case, we replace the matrix element qab = qk with a new
qab = qk+1 .
Properties of Hierarchical Matrices
An important property of hierarchical matrices is that each line is a permutation of
the other lines. Hence, for any function f (x),
 
f (qab ) = f (qcb ), ∀a,c. (5.14)
b b

In other words, hierarchical matrices break the replica symmetry in a way that does
not differentiate the single replicas but only induces correlations among replicas.
One-replica observables thus remain symmetric under the action of the permutation
 
group. This property also implies that the limit lims→0 1s ab f (qab ) = b f (qab ),
which appears in the computation of the replicated free energy, is finite. Other
schemes of replica symmetry breaking, however, do not guarantee the existence
of this limit.
Another important property is that, at any level k of RSB, the hierarchical matri-
ces with fixed {mi } form a closed commutative algebra, because for any two kRSB
matrices  and B̂, their product is a kRSB matrix Ĉ = ÂB̂ = B̂ Â. If the matrices
 and B̂ are parametrised by
{s = m−1,m0, . . . ,mk−1,mk = 1} → {A0, . . . ,Ak ;Ad },
(5.15)
{s = m−1,m0, . . . ,mk−1,mk = 1} → {B0, . . . ,Bk ;Bd },
150 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

their product Ĉ is given by

{s = m−1,m0, . . . ,mk−1,mk = 1} → {C0, . . . ,Ck ;Cd }, (5.16)
where the parameters Cd and Ci are given by

k
Cd = Ad Bd + (mi−1 − mi )Ai Bi ,
i=0
. / . /

k 
k
Ci = Ad + Al (ml−1 − ml ) Bi + Bd + Bl (ml−1 − ml ) Ai (5.17)
l=0 l=0

i
+ (ml−1 − ml )(Al − Ai )(Bl − Bi ).
l=0

Using these equations, one can compute, for instance, the inverse of a matrix Â,
by imposing that Cd = 1 and Ci = 0 for all i = 0, . . . ,k. This property thus
guarantees that, once a RSB structure with a given k is chosen, all the terms that
appear in the free energy preserve this same structure. We come back to this point
in Section 5.2.2, after having discussed the analytic continuation to s → 0.

5.2.2 Full Replica Symmetry Breaking

The kRSB matrices, as defined in Section 5.2.1, can only exist for large enough
values of s. The smallest size corresponds to choosing only two groups of replicas at
each step – i.e., s = 2k+1 . Because s cannot be smaller than 2k+1 , one might wonder
how to perform the analytic continuation to the non-integer and small values of s
and, in particular, to the limit s → 0 in which we are ultimately interested. The
continuation to real values of s is possible, however, because a kRSB matrix is
fully parametrised by {mi }, {qi }, as in Eq. (5.13), and specifying a proper analytic
continuation of these parameters is indeed straightforward. Here, we are specifically
interested in the case where s is continued to a real positive value2 s ≥ 0.
Let us first consider the analytic continuation of the parameters mi . The simplest
case of 1RSB has s ≥ m0 ≥ m1 = 1. If m0 = s and m0 = 1, the 1RSB
matrix reduces to a RS matrix. In particular, when s = 1, one necessarily has
s = m0 = m1 = 1, and the matrix q̂ then becomes a single number qd . Because m0
is bounded between 1 and s and is necessarily equal to 1 for s = 1, any meaningful
continuation to real values of s should be such that for s < 1 the inequality is

2 Negative values of s have also been considered in order to compute large deviations of the free energy.
See for example [131].
 5.2 The Algebra of Hierarchical Matrices 151
m

m−1 = s

m0

m1 = 1

0 1 2 3 4 s

Figure 5.3 Analytic continuation of a 1RSB matrix to s < 1. The inequality

s = m−1 ≥ m0 ≥ m1 = 1 is naturally reversed to s ≤ m0 ≤ 1 in the analytical
continuation to s < 1.

reversed. One then has s ≤ m0 ≤ 1, as illustrated in Figure 5.3. Generalising this

reasoning to kRSB, all the inequalities between the {mi } are reversed:

s ≥ m0 ≥ m 1 . . . ≥ m k = 1 → s ≤ m0 ≤ m1 . . . ≤ mk = 1. (5.18)

One can then collect the {mi } and the off-diagonal parameters {qi } into a piecewise
constant function, as illustrated in Figure 5.4. Once represented this way, a kRSB
matrix can be analytically continued to any value of s. The product of two kRSB
matrices can also be written as in Eq. (5.17), which remains well defined even for
non-integer {mi } and s. As we discuss in more detail in Section 5.4, for generic
forms of the free energy, if one considers only kRSB matrices, then the free energy
can be analytically expressed as a function of {mi } and {qi }, i.e., the piecewise
constant function q(x). It can then also be continued to real s. In summary, the
algebra of kRSB matrices remains formally well defined for any continuous value
of mi for s ≥ 0, provided Eq. (5.18) is respected. Inserting this structure in the free
energy gives a function of {mi } and {qi }, which upon extremisation determines the
{mi } and {qi }.
Full replica symmetry breaking (fullRSB) is reached when the number k of
replica symmetry breaking steps is sent to infinity. If the kRSB algebra is rep-
resented by a piecewise function q(x), then the limit k → ∞ is well defined,
provided q(x) converges to a continuous function. This process, which is illustrated
152 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes
q(x)
q(x)

qM

qk
qk−1

q2

q1 qm
q0
x
s m0 m1 m2 mk−2 mk−1 mk = 1 0 s xm xM 1 x

Figure 5.4 Parametrisation of a kRSB (left) and of a fullRSB (right) matrix for
s < 1 via a piecewise constant function q(x), which encodes the off-diagonal
matrix elements of q̂. The function q(x) is always monotonous; depending on the
definition of qab , it can either be an increasing or a decreasing function of the
parameter x.

in Figure 5.4, can only happen if {mi } converges to a dense set on the interval
[0,1]. The difference between two successive mi must thus become infinitesimal,
i.e., mi+1 − mi = dx. In this continuum limit, sums over replica indices can be
transformed into integrals. An explicit and useful example is (recall the convention
m−1 = s and mk = 1)


1,s 
k   1 
p p p p
qab =s qd + (mi−1 − mi )qi = s qd − p
dxq(x) . (5.19)
a,b i=0 0

5.2.3 The Algebra of fullRSB Matrices

We now give explicit expressions for the product of two fullRSB matrices, and
from it, we derive an expression for the inverse and for the determinant of a fullRSB
matrix, following [255]. These results are especially useful because these quantities
usually appear in the replicated free energy of interesting models.

Product
A fullRSB version of Eq. (5.17) can be obtained in the continuum limit. Given two
fullRSB matrices Â, B̂ parametrised by

 → {A(x);Ad }, B̂ → {B(x);Bd }, (5.20)

their product

Ĉ = ÂB̂ = B̂ Â → {C(x);Cd } (5.21)

 5.2 The Algebra of Hierarchical Matrices 153

is expressed by taking the continuum limit of Eq. (5.17) as

 1
Cd = Ad Bb − dxA(x)B(x),
s
C(x) = A(x) (Bd − B) + B(x) (Ad − A) (5.22)
 x
− dy (A(y) − A(x)) (B(y) − B(x)) ,
s
where we have defined
 1
A = dxA(x) . (5.23)
s
Hence, fullRSB matrices also form a closed commutative algebra. Remarkably,
the fullRSB algebra contains any finite-kRSB algebra as a particular case. Insert-
ing piecewise constant functions A(x) and B(x) in Eq. (5.22), as illustrated in
Figure 5.4, one indeed recovers Eq. (5.17).

Fourier Transform
For simplicity we now restrict to the case s → 0, which is the interesting limit
in most applications, and discuss some additional properties of fullRSB matrices.
To each hierarchical matrix  → {A(x);Ad }, we can associate its ‘Fourier trans-
form’ [106, 116, 289] È → {Ã(x); à }, defined by the linear transformation
d
 1  1
Ã(x) = Ad − xA(x) − dyA(y), Ãd = Ad − dxA(x). (5.24)
x 0

The inverse Fourier transform is then given by3

 1
dy dÃ(y)
A(x) = A(1) − , Ad = Ã(1) + A(1). (5.25)
x y dy
These relations can also be expressed in a differential form,
dÃ(x) dA(x)
= −x , Ã(1) = Ad − A(1), (5.26)
dx dx
which is equivalent to the integral relations in Eqs. (5.24) and (5.25).
The reader can verify that the multiplication of two matrices, as defined by
Eq. (5.22), becomes a simple multiplication in terms of the Fourier transformed
matrices:
C̃(x) = Ã(x)B̃(x), C̃d = Ãd B̃d . (5.27)

3 Note that the transformation {A(x);A } → {A(x) + C;A + C} leaves the Fourier transform invariant for any
d d
constant C. As a consequence, the inversion of the Fourier transform leaves an undetermined constant in
{A(x);Ad }. This is why A(1) appears in the right-hand side of Eqs. (5.25).
154 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

In other words, the Fourier transform reduces the matrix multiplication to a simple
product and diagonalises the algebra of fullRSB matrices.4

Determinant and Inverse

The expressions for the determinant and the inverse of a fullRSB matrix
q̂ → {q(x);qd } can be derived in compact form in terms of the Fourier trans-
form [255]. The standard notation for the Fourier transform of q̂ is λ̂,
 1
λ(x) = qd − xq(x) − dyq(y), λ̇(x) = −x q̇(x), (5.28)
x

where here and in the rest of this chapter, the dot denotes a derivative with respect
to the argument of a fullRSB function. One then obtains [255]
 1  
d q(0) dx λ(x)
lim log det q̂ = log (λ(0)) + − 2
log
s→0 ds λ(0) 0 x λ(0)
 1 (5.29)
q(0) q̇(x)
= log (λ(1)) + + dx .
λ(0) 0 λ(x)
The inverse of q̂ can be parametrised by q̂ −1 → {[q −1 ](x);[q −1 ]d }, and in the
s → 0 limit, one obtains
  1 
−1 1 dy λ(0) − λ(y) q(0)
[q ]d = 1− 2
− ,
λ(0) 0 y λ(y) λ(0)
  x  (5.30)
−1 1 q(0) λ(0) − λ(x) dy λ(0) − λ(y)
[q ](x) = − + + 2
.
λ(0) λ(0) xλ(x) 0 y λ(y)
Note that from Eq. (5.30) one also obtains a useful relation,
 1
1  −1
1,s
−1 1
lim (q )ab = [q ]d − dx[q −1 ](x) = . (5.31)
s→0 s 0 λ(0)
a,b

5.3 Probability Distribution of the Mean Square Displacement

We now examine a first physical consequence of the hierarchical structure of the
replica matrices by considering the probability distribution of the mean square
displacement (MSD) between replicas. In Section 5.2.1, we discussed the algebra
of hierarchical matrices in a general setting, but here we revisit the setting of
Chapter 4, in which one follows a glass state prepared at a given initial temper-
ature and density to a different state point. The free energy, in the Franz–Parisi
construction, is then expressed in terms of the MSD matrix D̂ of 1 + s replicas,

4 This result is not surprising. It is well known by the Gelfand representation theorem that any closed
commutative algebra can be diagonalised [104].
 5.3 Probability Distribution of the Mean Square Displacement 155

where the first replica encodes the initial state, and the block of the remaining s
replicas is symmetric under permutations. We assume, in the following, that this
block is described by a hierarchical matrix of the type discussed in Section 5.2.1.

5.3.1 Replicas and Hierarchical Matrices

As discussed in Section 4.1.2, in the glass phase the long time dynamics of
the system is described by a restricted Gibbs–Boltzmann measure, given by
Eq. (4.7). Consider two particle configurations X1 and X 2 , independently extracted
according to Eq. (4.7) with identical parameters. Their relative distance, or MSD,
D = D X 1,X 2 , is a random variable5 with probability distribution

PY, Dr (D) = dX1 dX 2 P (X1,ϕ,β|Y, Dr )P (X2,ϕ,β|Y, Dr )
(5.32)
× δ D − D(X 1,X 2 ) .
Because Y is a random configuration extracted from the Gibbs–Boltzmann equilib-
rium measure, PY, Dr (D) is itself a random object and can be averaged over Y :

dY dX 1 dX2 −βg V [Y,g ]
PY, Dr (D) = e
Z[ϕg,βg ] (5.33)
× P (X1,ϕ,β|Y, Dr )P (X2,ϕ,β|Y, Dr )δ D − D(X 1,X 2 ) .
Following the reasoning and notations of Section 4.1.3, one can introduce replicas
and write Eq. (5.33) as6
 .
s+1
/
1
dX a e−βa V [X ,a ]
a
PY, Dr (D) = lim
s→0 Zs+1 [ϕa ,βa , Dr ]
a=1
. s+1 /⎡ ⎤
 2 
2,s+1
× δ Dr − D(X a,X 1 ) ⎣ δ D − D(X a,X b ) ⎦
a=2
s(s − 1) a=b

2 
2,s+1
= δ D − D(X a,X b ) ≡ P(D), (5.34)
s(s − 1) a=b

where the average is over the replicated liquid defined in Section 4.1.3.

5 Note that in Chapter 4, D denoted the average of this random variable. The motivation for this change of
notation will be given shortly.
6 In short, the proof is as follows. By replica symmetry, one can eliminate the average over pairs a = b in the
last term of the second line of Eq. (5.34) and choose two arbitrary replicas of the s block, say a = 2 and b = 3.
The other s − 2 replicas can then be integrated, giving rise to a factor Z[ϕ,β|Y, Dr ]s−2 . In the limit s → 0,
this factor provides the denominator of the two identical copies of P (X,ϕ,β|Y, Dr ) that appear in Eq. (5.33).
The result follows after recalling that Zs+1 [ϕa ,βa , Dr ] → Z[ϕg ,βg ] when s → 0.
156 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

In the d → ∞ limit, the free energy can be expressed as a function of the

rescaled MSD matrix  ˆ = (d/2g )D̂, via the closed Eq. (4.59), as discussed in
Section 4.2.4. It follows from the general theory of thermodynamic fluctuations
ˆ is given by7
that the probability distribution of 
ˆ
ˆ ϕa,Ta ) ∝ e−Nfs+1 ()
P (|) ϕa ,Ta , )
. (5.35)
ˆ becomes
Therefore, in the thermodynamic limit, the probability distribution of 
extremely sharply peaked around the value that minimises the free energy. Rescal-
ing the argument of P(D) as  = (d/2g ) D, the average in Eq. (5.34) simplifies
to [254]

1 
2,s+1
P() = lim δ ( − ab ) , (5.36)
s→0 s(s − 1)
a=b

where ab are the elements of the matrix  ˆ that minimises (for s > 1) or extrem-
ises (for s < 1) the free energy fs+1 () ˆ Eq. (5.36) is a general relation
ϕa,Ta, ).
between this matrix and the equilibrium probability distribution P() of the MSD
between two identical copies, extracted from the distribution of the same glass
state – i.e., generated by the same reference configuration Y – and followed to a
target state point.
The replica symmetric structure for ab , as discussed in Chapter 4, Eq. (4.62),
reads8
aa = d = 0, a = 1, . . . ,n,
1a = a1 = r , a = 2, . . . ,n, (5.37)
ab = 0, a,b = 2, . . . ,n, and a = b,
where 0 and r are the solutions of Eqs. (4.76). Plugging this structure in
Eq. (5.36), we get
P() = δ( − 0 ). (5.38)
Therefore, within the replica symmetric assumption, the average probability distri-
bution of the mean square displacement of two copies of the system sampled from
the same glass state is peaked around the mean value 0 . In phase space, each glass
state thus corresponds to a restricted portion of the set of configurations, which is
sampled ergodically by the dynamics.

7 Remember that f
s+1 () ˆ incorporates a factor β, and it is therefore adimensional. See Section 4.2.4.
ϕa ,Ta , )
8 Note that while in Chapter 4,  denoted the off-diagonal elements of  , here we denote them by 
ab 0
to be consistent with the notation of Section 5.2.1. Instead,  denotes the argument of the probability
distribution P().
 5.3 Probability Distribution of the Mean Square Displacement 157

5.3.2 The Replica Symmetry Broken Phase

We now show that Eq. (5.36) is the basis for a physical interpretation of replica
symmetry breaking. We consider a RSB parametrisation for the matrix  ˆ in the
block of the s identical replicas labelled by a = 2, . . . ,n, and we derive the average
distribution of the MSD between two replicas that belong to a same glass state. Let
us consider first a 1RSB structure – that is,
aa = 0, a = 1, . . . ,n,
1a = a1 = r , a = 2, . . . ,n,
 (5.39)
1 if a,b = 2, . . . ,n in same subgroup and a = b,
ab =
0 if a,b = 2, . . . ,n in different subgroups,

where m0 is the size of the 1RSB subgroups, as illustrated in Figure 5.5. The MSD
probability distribution, from Eq. (5.36), is then

P() = m0 δ( − 0 ) + (1 − m0 )δ( − 1 ) . (5.40)

In the 1RSB case, two identical glass configurations X1 and X 2 can typically be
found either at a distance 0 , with probability m0 , or at distance 1 , with probabil-
ity 1−m0 . This result can be interpreted in terms of the phase space structure. In the
replica symmetric case, a glass state is a unique cluster of particle configurations.
Two typical configurations visited by the dynamics, separated by a long time, are
always at a distance 0 . Within the 1RSB ansatz, by contrast, what was a glass state

0 Δr
Δ1
0 m0
Δ1

Δ0

n
Δr
s

Δ0

Δ1
0
Δ1

Figure 5.5 Illustration of the 1RSB parametrisation for the state following mean
square displacement matrix, within the Franz–Parisi construction.
158 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

in the RS construction becomes a ‘metabasin’ that contains several glass substates.

Configurations that belong to the same substate are typically at distance 1 apart,
while configurations that belong to different substates within the same metabasin
are typically at distance 0 apart. For this reason, one physically expects that
0 > 1 . This reasoning is a first illustration of the physical consequences of
the hierarchical structure of the MSD matrix in replica space.
Consider next a kRSB or fullRSB parametrisation of the form
aa = 0, a = 1, . . . ,n,
1a = a1 = r , a = 2, . . . ,n, (5.41)
ab |a,b=2,...,n, a=b → {(x);0},
where the function (x) encodes the off-diagonal elements of the block of the
s identical replicas. As in the 1RSB case, values of (x) for smaller x encode
replicas that are farther away, and therefore, we expect (x) to be a monotonically
decreasing function of x, as illustrated9 in Figure 5.6. From Eq. (5.36), one can
show that
0 0
0 dx() 0
P() = 0 0 0, (5.42)
d 0
where x() is the inverse function of (x) over the interval x ∈ [0,1]. This
inverse is well defined and monotonically decreasing, because (x) is assumed
to be monotonically decreasing. The RSB profile (x), therefore, defines the aver-
age probability distribution P() of the MSD between two typical configurations
that belong to a same metabasin. Note that Eq. (5.42) can also be applied to a
piecewise constant function. For example, in the RS case where (x) = 0 , the
inverse function is formally x() = θ(0 − ), which is consistent with P()
= δ( − 0 ). The 1RSB result in Eq. (5.40) is also reproduced by this general
expression.
A typical profile for (x) is shown in Figure 5.6. It displays two flat parts,
(x) = m for x ∈ [0,xm ] and (x) = M for x ∈ [xM ,1], with a continuously
changing part c (x) in between. Using Eq. (5.42) one obtains that the average
probability distribution of the overlap is given by
P() = xM δ( − M ) + xm δ( − m ) + Pc (), (5.43)
where Pc () denotes the regular part of the probability distribution.

9 In the spin glass and structural glass literatures, an overlap q ∈ [0,1] is often used as a measure of distance,
ab
such that qab = 1 corresponds to identical configurations, and decreasing qab corresponds to increasingly
different configurations. In that context q(x) is an increasing function of x, as illustrated in Figure 5.4.
 5.3 Probability Distribution of the Mean Square Displacement 159

P(Δ)
Δ(x)
Δm

ΔM

x Δ
xm xM 1 ΔM Δm

Figure 5.6 (Left) Illustration of a typical profile of the function (x). The two
points xm and xM separate the flat and continuous regions of (x). These points
are often called ‘breaking points’. (Right) Illustration of the corresponding P().
The two flat parts of (x) are responsible for the two Dirac delta functions
appearing at M and m . Their weights are respectively given by xM and xm .

5.3.3 The Ultrametric Structure of States

The replica symmetric assumption implies that a glass state is a restricted portion
of phase space ergodically sampled by the dynamics. Two typical configurations
that belong to this portion of phase space have a MSD that satisfies the replica
symmetric saddle point equations. At this level, a glass is a simple thermo-
dynamic state.
At the 1RSB level, a glass state is not a simple state anymore. It is a metabasin
that contains many individual substates: a pair of typical configurations within this
metabasin can have two different MSDs depending on whether they belong to the
same substate or not.
Upon iterating the RSB construction – i.e., moving to higher number of replica
symmetry breaking steps – the glass metabasin splits into sub-basins, each con-
taining other sub-basins, and so on. This hierarchy terminates when one obtains
individual states. As a consequence, the MSD of two typical configurations that
belong to the same glass metabasin can take a set of k values, corresponding to
the configurations being separated at different levels in the hierarchical structure
of substates. This set becomes continuous for k → ∞ and is thus encoded in the
profile (x). For completeness, in this section, we discuss briefly a few additional
properties of the organisation of glass states in phase space that are important in
some applications. We refer the reader to [254, chapter IV] for a more detailed
discussion.
160 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

Triplet Distribution and Ultrametricity

The hierarchical states description can be refined by computing additional geomet-
rical properties of phase space. Consider, for instance, three configurations X1, X2
and X 3 , with mutual MSDs10

d
˘ ab =
 D X a,X b , a,b = 1, . . . ,3. (5.44)
2g

The average probability distribution of the three MSDs can be computed using an
approach along the same lines as Section 5.3.1,
1
P( ˘ 13, 
˘ 12,  ˘ 23 ) = lim
s→0 s(s − 1)(s − 2)


2,s+1 (5.45)
× ˘ 12 − ab )δ(
δ( ˘ 13 − ac )δ(
˘ 23 − bc ) .
a=b=c(=a)

Assuming a hierarchical structure for ab , one then obtains [254]

˘ 12, 
P( ˘ 23 ) = 1 P(
˘ 13,  ˘ 12 )x( ˘ 12 )δ( ˘ 12 −  ˘ 13 )δ( ˘ 12 −  ˘ 23 )
2
1
+ P( ˘ 12 )P( ˘ 13 )θ( ˘ 13 −  ˘ 12 )δ( ˘ 13 −  ˘ 23 )
2 (5.46)
+ P( ˘ 13 )P( ˘ 23 )θ( ˘ 23 −  ˘ 13 )δ( ˘ 23 −  ˘ 12 )

+ P( ˘ 23 )P( ˘ 12 )θ( ˘ 12 −  ˘ 23 )δ( ˘ 12 −  ˘ 13 ) .

Therefore, this probability is non-vanishing only if two MSDs are equal and the
third is equal or smaller. In other words, for every triplet of configurations extracted
according to the Boltzmann–Gibbs measure restricted to one glass metabasin, the
triangle formed by their mutual distances is either equilateral or isosceles, with the
two equal edges longer than the third. This structure, which follows directly from
the hierarchical structure of ,ˆ is called ‘ultrametric’ in mathematics [300] and is
also common in taxonomy; see, e.g., [17]. In the framework of disordered systems,
it was first discovered in spin glasses [254].
The replica symmetry breaking structure thus corresponds to a hierarchical
organisation of states in phase space. The states form an ultrametric tree, as
illustrated in Figure 5.7 for a 3RSB landscape. We refer the reader to [254, chapter
IV and reprint 16] and [300] for more details.

˘ ab the MSDs that are used as arguments of probability distributions to

10 In this section, we denote by 
distinguish them from the thermodynamic matrix ab that extremises the free energy.
 5.3 Probability Distribution of the Mean Square Displacement 161
Y k = 0 ≡ RS level

Δ0

1RSB

Δ1
2RSB

Δ2
3RSB

Pure states Sub-basins Metabasin

Figure 5.7 An example of a 3RSB ultrametric structure of states. Individual states

are the leaves of the tree. They contain configurations whose typical MSD is 3 .
The states are organised in metabasins with a hierarchical structure: configurations
belonging to states at distance 1 in the tree have typical MSD 2 , those at distance
2 have 1 , and those at distance 3 have 0 .

Fluctuations of the MSD Distribution

Another important property that follows from the hierarchical matrix structure can
be obtained from the fluctuations of PY, Dr (D) from metabasin to metabasin. This
probability density, defined by Eq. (5.32), is a random object because it depends
on the configuration Y that selects one of the possible glass metabasins. One can
thus study its probability distribution over the realisations of Y . Consider four con-
˘ ab defined by Eq. (5.44). Following
figurations {X i }i=1,...,4 with mutual distances 
the same strategy as in Section 5.3.1, the probability distribution of  ˘ 34
˘ 12 and 
is [254]

˘ 34 ) = 1 PY (
Y Y 2 Y Y
˘ 12, 
PY ( ˘ 12 ) δ  ˘ 34 + PY (
˘ 12 −  ˘ 12 ) PY (
˘ 34 ) . (5.47)
3 3
This result shows that
Y Y Y
˘ 12, 
PY ( ˘ 34 ) = PY (
˘ 12 ) PY (
˘ 34 ) (5.48)
and, thus, that PY () is not self-averaging. This quantity does not converge in
probability to P(), and finite fluctuations persist even in the thermodynamic limit.
The structure of the fluctuations of PY () is also entirely fixed by the hierarchical
ansatz for the MSD matrix. We refer the reader to [254, reprint 16] for more details.

Generation of the Ultrametric Tree of States

We conclude this section by discussing the statistical properties of the free
energies of glass metabasins, refining the analysis of Section 5.1. Recall that
the total partition function of a given metabasin, selected by a configuration Y , is
given by Eq. (4.7),
162 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

Z[ϕ,β|Y, Dr ] = dXe−βV [X,] δ(Dr − D(X,Y )). (5.49)

When replica symmetry is broken, the metabasin is a collection of sub-basins,

each broken into sub-basins, and so on, down to the lowest level in the hierarchy
of Figure 5.7 that corresponds to pure states. Each pure state is a minimum of
the free energy and corresponds to a cluster of configurations. These clusters are
disjoint, and the probability that the system is found far away from one of them
is exponentially small in N . The partition function is therefore the sum of the
partition functions of the pure states.11 For notational simplicity, here we consider
the parameters ϕ,β, Dr as fixed and focus on the dependence on Y . The partition
function can then be written as
  T
Z(Y ) = Zα (Y ) = e−βNfα (Y ), fα (Y ) = − log Zα (Y ), (5.50)
α α
N
where α labels the pure states that compose the metabasin selected by Y , and fα (Y )
is the free energy of α. The weight of state α is therefore
e−βNfα (Y ) 
wα (Y ) =  −βNfγ (Y ) , wα (Y ) = 1, (5.51)
γ e α

and given an observable O, one can write its equilibrium average as


OY = wα (Y )Oα, (5.52)
α

where Oα is the average of O restricted to state α.

All the pure states that contribute to the equilibrium measure must have the
same (and lowest possible) free energy per particle,12 fα = fmin , as discussed
in Section 5.1. States with fα > fmin are unstable to nucleation, as discussed in
Section 1.5.3. We therefore assume that, if equilibrium can be attained within the
glass metabasin selected by Y , the relevant equilibrium pure states have N fα =
Nfmin + δfα , where δfα is a finite correction in the thermodynamic limit.13 Their
weight is then finite and given by

11 In the case of a ferromagnet, this statement can be verified based on the discussion of Chapter 1. For T < T
c
and zero external field, the two minima of the free energy correspond to positive and negative magnetisation
m = ±meq . Excluding the possibility of phase coexistence – i.e., focusing on homogeneous magnetisation
profiles – the probability that the system has magnetisation m = meq is exponentially small in N , and the
fluctuations around ±meq within each state are of the order of N −1/2 . The partition function is therefore
dominated by configurations with m = ±meq + O(N −1/2 ) and can be written as the sum of two partition
functions, corresponding to disjoint sets of configurations with positive or negative magnetisation.
12 We will discuss in Chapter 7 how to extend this reasoning to take into account metastable states.
13 This is certainly the case for all states in fully connected models such as the Curie–Weiss model introduced in
Section 1.5.2, for which mean field theory is exact. In this case, the free energy can be expressed as a series in
1/N , as can be shown from the high-temperature expansion discussed in Chapter 1. In finite dimensions, this
is only true for pure states. For instance, in presence of interfaces, we have δfα ∝ Ld−1 = N (d−1)/d , where
L is the linear size of the system.
 5.3 Probability Distribution of the Mean Square Displacement 163

e−βδfα (Y )
wα (Y ) =  −βδfγ (Y ) . (5.53)
γ e

We wish to characterise the statistical properties of δfα , in order to provide a statis-

tical procedure to generate the ultrametric tree of states.
We first discuss these statistical properties for a 1RSB hierarchical structure, in
which the glass metabasin is partitioned into independent pure states. Because of
the randomness of Y , the pure states, their partition functions and their weights are
themselves random variables. It turns out that the random free energies δfα have
an extremely simple distribution [254]. They are extracted independently from a
Poisson point process with intensity ρm0 (δf) = eβm0 δf , where m0 is the point at
which (x) is discontinuous. The number of states with free energy δfα ∈ [δf,δf
+ dδf] is then given by

dρm0 (δf) = eβm0 δf dδf. (5.54)

We now consider the case in which the glass metabasin has a 2RSB structure.
Individual pure states are then organised into sub-basins. We assign to each
sub-basin a label α (0) , and we label (α (0),α (1) ) the individual states within this
basin. As for individual states, we can assign to each sub-basin a free energy,
which is the logarithm of the sum of the partition functions of the individual states
that belong to that basin. We denote N fmin + δfα(0) the reference free energy for
the sub-basins, and N fmin + δfα(0) + δfα(0),α(1) the free energy of the individual
states. It turns out that the free energies of the sub-basins are here again extracted
according to a Poisson point process. The number of free energies δfα(0) in the
interval [δf0,δf0 + dδf0 ] is thus

dρm0 (δf0 ) = eβm0 δf0 dδf0 . (5.55)

The free energies of individual states within a sub-basin of free energy δfα(0) are
extracted according to another point process, for which the density of free energies
in the interval [δf1,δf1 + dδf1 ] is given by

dρm1 (δf1 ) = eβm1 δf1 dδf1 . (5.56)

This construction can be repeated an arbitrary number of times in order to generate

the free energies of states and basins on arbitrary ultrametric trees. This construction
is particularly helpful to describe the evolution of the ultrametric tree under an
external perturbation in order to characterise, for example, the equilibrium statis-
tical distribution of avalanches [148, 225]. We refer the reader to [254, reprint 18]
for more details.
It is interesting to discuss the behaviour of the weights when a random pertur-
bation is added to the original Hamiltonian, as discussed in Section 5.1. The free
energies are modified, and the new ones, denoted N fα , are
164 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

N fα = N fmin + δfα, fmin = fmin + f, δfα = δfα + α, (5.57)
where the shift of the extensive part of the free energy, f, is the same for all
states, as discussed in Section 5.1. We assume once again that the perturbations
α are finite in the thermodynamic limit; otherwise, one state dominates as soon
as the perturbation is added. Let us also assume that α are random Gaussian
variables with√zero mean and variance σ 2 , i.e., with a probability distribution γσ 2 ()
= e− /(2σ ) / 2πσ 2 . The new density of free energy states is then
2 2


ρ (δf ) = dγσ 2 ()ρ(δf − ).
 
(5.58)

If the moments of δf are well defined, we have

[δf2 ]c = [δf2 ]c + σ 2, [δf2 ]c = [δf2 ] − [δf]2, (5.59)
where [•] here denotes an average over ρ(δf). The distribution ρ(δf) thus widens
when the perturbation is added. The situation is completely different if ρ(δf) = ekδf .
We then have
 σ2
ρ  (δf ) = ek(δf −c) = ρ(δf − c), . c=− (5.60)
2k
In this case, the shape of the distribution of weights is left invariant by the perturba-
tion. This invariance property of the exponential weight distribution characterises
ultrametric trees, as given by Eq. (5.55).
Let us stress one final point. Although in the fullRSB limit, the ultrametric tree
has an infinite number of branching points, and at each branching point, the number
of branches is infinite, the distribution of weights is such that only states with
positive weight contribute.14 More precisely, if we restrict our analysis to states
such that wα > , we have for small 

wα θ(wα − ) = 1 − O( ω ), (5.61)
α

with a positive exponent ω = 1 − xM [254].

5.4 Replicated Free Energies and Hierarchical Matrices

The replicated free energies of many mean field models have a common structure,
and their evaluation on hierarchical matrices can be done using the same approach.
In this section, we consider the free energy f(q̂) to be a function of a replica matrix
q̂, of a generic form, and we provide some mathematical results that are useful to

14 A contrarian could argue that a very large number of states with vanishing weight gives a finite contribution.
This possibility is excluded by Eq. (5.61).
 5.4 Replicated Free Energies and Hierarchical Matrices 165

evaluate this free energy on a hierarchical RSB matrix q̂. The expressions we obtain
can be applied to many different problems, and in Chapter 6, we consider the case
of particle systems in the limit d → ∞ more specifically.
Consider a generic free energy that depends on a s × s matrix, q̂, written as the
sum of two terms,

f(q̂) = fa (q̂) + fd (q̂), (5.62)

the first term being a simple ‘algebraic’ function of q̂ and the second being a
‘differential’ term to be specified later. Under the assumption that q̂ → {q(x);qd }
is a hierarchical matrix, our aim is to compute the limit s → 0 of the free energy
d a,d
fa,d [q(x);qd ] = lim f (q̂). (5.63)
s→0 ds
The term fa (q̂) is assumed to be a simple explicit function of q̂, for which the
evaluation on the hierarchical matrix and the limit s → 0 can be taken easily. Two
common examples are the sum of an arbitrary power p of matrix elements,

1,s  1
p p
f (q̂) =
a
qab ⇒ f [q(x);qd ] =
a
qd − dxq(x)p, (5.64)
a,b 0

where we used Eq. (5.19), and fa (q̂) = log det q̂, which – using Eq. (5.29) – leads
to
 1
q(0) q̇(x)
f [q(x);qd ] = log (λ(1)) +
a
+ dx . (5.65)
λ(0) 0 λ(x)
The term fd (q̂) is expressed in terms of a differential operator and a function
r(h) as
0
1 s 
s 0
q ∂2 0
fd (q̂) = e 2 a,b=1 ab ∂ha ∂hb r(hc )0 . (5.66)
0
c=1 {hc =H }

The aim of this section is to show how the differential term can be treated when q̂
has a hierarchical structure and the limit s → 0 is taken. We will give a general
recipe for computing this term, and deduce equations for qd ,q(x) from extremising
the free energy.
In the context of spin glasses, for instance, the free energy of the Sherrington–
Kirkpatrick model, as well as many of its generalisations, has this form [254]. In the
context of glasses, q̂ corresponds to the MSD matrix, and the replicated free energy
given in Eq. (4.59) is written in a similar form, with some small complications
that will be discussed in Chapter 6. Hence, the results of this section are broadly
applicable to mean field models.
166 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

5.4.1 The Differential Free Energy Term

We now describe how to compute the differential term fd (q̂) using the iterative
algorithm developed in [133]. In the rest of this section, to simplify the notation,
we do not explicitly indicate the ranges of indices a,b = 1, . . . ,s in the sums
and products. For the new indices we introduce, the range is specified only at the
moment they are defined.

The Replica Symmetric Case

Let us start by the simplest case, in which q̂ has a replica symmetric structure, i.e.,
q(x) = q0 . This case has already been discussed in Section 4.3.2 in the specific
setting of particles in the limit d → ∞. The differential operator appearing in
Eq. (5.66) can then be written as
. /2
1 ∂2 qd − q0  ∂ 2 q0  ∂
qab = 2
+ . (5.67)
2 a,b ∂ha ∂hb 2 a
∂ha 2 a
∂ha

Applying the identity in Eq. (4.69) to Eq. (5.67) gives

 ' (2 ' (2
qd −q0 ∂2
+
q0 ∂  q0 ∂ 
r(hc ) = e 2
2 a ∂h2 2 a ∂ha a ∂ha
e a γqd −q0  r(hc ). (5.68)
c c

Next, we can apply the identity in Eq. (4.71) to express the derivatives with
respect to the variables {ha } evaluated at a point where they all take the same value
{ha = H },
0 0
' (2
 0  s 0
q0 ∂
0 q0 ∂ 2
e 2 a ∂ha
γqd −q0  r(hc )0 =e 2 ∂h2
γqd −q0  r(h) 00
0 h=H (5.69)
c {hc =H }
 s
= γq0  γqd −q0  r(H ) .

Taking the s → 0 limit, we then obtain

d  s  
fd [q(x);qd ] = lim γq0  γqd −q0  r(H ) = γq0  log γqd −q0  r(H ) . (5.70)
s→0 ds

The 1RSB Case

We now repeat the same calculation within a 1RSB parametrisation. The function
q(x) is now a step function, q(x) = q0 for x < m0 and q(x) = q1 for x > m0 . For
finite s, this corresponds to dividing the s replicas in s/m0 blocks that we label with
 5.4 Replicated Free Energies and Hierarchical Matrices 167

an index l ∈ {1, . . . ,s/m0 }. We write a ∈ l if replica a is one of the m0 replicas

that belongs to block l. The differential operator can then be written as
. /2
1 ∂2 qd − q1  ∂ 2 q1 − q 0   ∂
qab = +
2 a,b ∂ha ∂hb 2 a
∂h2a 2 l a∈l
∂ha
. /2 (5.71)
q0  ∂
+ .
2 a
∂ha

The action of this operator on the product a r(ha ) can be evaluated in three steps:
 ∂2
1. The operator qd −q
2
1
a ∂h2a acts independently on each replica, via Eq. (4.69),
and thus leads to the replacement

r(ha ) → γqd −q1  r(ha ) = g(1,ha ), (5.72)

for each replica. From this point on, there is no need to consider different fields
ha in replicas of the same block because the remaining operators act on each
group of replicas via Eq. (4.71). Therefore, we can consider that ha = hl , if
a ∈ l, and
 
g(1,ha ) → g(1,hl )m0 . (5.73)
a l

   ∂2
2. Replacing the operator q1 −q
2
0
l
∂ 2
a∈l ∂ha with q1 −q
2
0
l ∂h2 via Eq. (4.71)
l
makes it act independently on each block of replicas, leading to the replacement

g(1,hl )m0 → γq1 −q0  g(1,hl )m0 = g(m0,hl ) (5.74)

in each block. From this point on, there is no need to consider different fields
hl in different blocks because the remaining operator acts on all replicas via
Eq. (4.71). Therefore, we can consider that hl = h, and

g(m0,hl ) → g(m0,h)s/m0 . (5.75)
l

q0  ∂ 2 q0 ∂ 2
3. Replacing the operator 2 a ∂ha with 2 ∂h2
via Eq. (4.71) gives

q0 ∂ 2
fd (q̂) = e 2 ∂h2 g(m0,h)s/m0 = γq0  g(m0,h)s/m0 , (5.76)

which should be computed in h = H .

168 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

Collecting the results, we obtain

  m s/m0
fd (q̂) = γq0  γq1 −q0  γqd −q1  r(H ) 0 , (5.77)

and
d d 1   m 
fd [q(x);qd ] = lim f (q̂) = γq0  log γq1 −q0  γqd −q1  r(H ) 0 . (5.78)
s→0 ds m0

The kRSB Case

The 1RSB construction we just discussed can be immediately generalised to a
kRSB structure with finite k. Using the conventions

q−1 = 0, qk+1 = qd , m−1 = s, mk = 1, (5.79)

we can introduce an index, l (i) ∈ {1, . . . ,s/mi }, to label the blocks at level i of
the hierarchical structure. Note that l (k) ∈ {1, . . . ,s} coincides with the original
replica index a, while l (−1) ∈ {1} can take a single value corresponding to the
block of all replicas. As for the 1RSB case, we write a ∈ l (i) if replica a belongs
to block l (i) . With those conventions, the differential operator in Eq. (5.66) can be
written as
⎛ ⎞2
1  (qi+1 − qi )  ⎝  ∂ ⎠
k
∂2
qab = . (5.80)
2 a,b ∂ha ∂hb i=−1
2 (i) (i)
∂ha
l a∈l

As in the 1RSB case, the action of this operator on the product a r(ha ) can then
be evaluated in k + 2 steps. Because they are very similar to the 1RSB case, we
describe them succinctly, the main purpose being the introduction of the notational
convention.
 ∂2
1. The operator qd −q
2
k
a ∂h2 acts independently on each replica, leading to
a

r(h) → γqd −qk  r(h) = g(mk,h),

  (5.81)
r(ha ) → g(mk,hl (k−1) )mk−1 /mk ,
a l (k−1)

recalling that mk = 1 and that in each of the l (k−1) blocks there are mk−1
replicas.

2. The operator qk −q2 k−1 l (k−1) ∂h2∂
2
acts independently on each l (k−1) block of
l (k−1)
replicas, leading to the replacement

g(mk,h)mk−1 /mk → γqk −qk−1  g(mk,h)mk−1 /mk = g(mk−1,h) (5.82)

 5.4 Replicated Free Energies and Hierarchical Matrices 169

in each block. All the g(mk−1,h) that pertain to the same l (k−2) block can now
be grouped together. There are mk−2 /mk−1 sub-blocks in the same block, hence
 
g(mk,hl (k−1) )mk−1 /mk → g(mk−1,hl (k−2) )mk−2 /mk−1 . (5.83)
l (k−1) l (k−2)

3. The process is iterated for all i, leading to the recursion relation

qi −qi−1 ∂ 2
g(mi−1,h) = e 2 ∂h2 g(mi ,h)mi−1 /mi = γqi −qi−1  g(mi ,h)mi−1 /mi . (5.84)

4. After all steps are completed, one obtains

q0 ∂ 2
f d (q̂) = e 2 ∂h2 g(m0,h)s/m0 = γq0  g(m0,h)s/m0 , (5.85)

which should be evaluated in h = H .

In summary, one obtains

g(mk,h) = γqd −qk  r(h),

g(mi−1,h) = γqi −qi−1  g(mi ,h)mi−1 /mi , i = k, . . . ,1,
(5.86)
d 1
f d [q(x);qd ] = lim f d (q̂) = γq  log g(m0,H ).
s→0 ds m0 0

The fullRSB Case

In the limit k → ∞, assuming that q(x) has the form illustrated in Figure 5.4, one
can take a continuum limit of Eq. (5.86). At mk = 1, we have g(1,h) = γqd −q(1) 
r(h). First, we note that for x ∈ [xM ,1], q(x) is constant, q(x) = q(1) = qM . Any
discretisation of x would lead, using Eq. (5.84), to

g(x,h) = g(1,h)x , x ∈ [xM ,1]. (5.87)

We then consider the generic i-iteration step given by Eq. (5.84). In the region
x ∈ [xm,xM ], q(x) has a continuum limit with q̇(x) = 0; one thus has

qi − qi−1 ∼ q̇(x)dx, (5.88)

where x = mi , dx = mi − mi−1 , and the iteration Eq. (5.84) becomes

1 ∂2
q̇(x)dx
g(x − dx,h) = e 2 ∂h2 g(x,h)1−dx/x . (5.89)

By developing at linear order in dx we get the differential equation

∂g(x,h) 1 ∂ 2 g(x,h) 1
= − q̇(x) + g(x,h) log g(x,h), (5.90)
∂x 2 ∂h2 x
170 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

and defining
1
f (x,h) = log g(x,h) (5.91)
x
gives a differential equation for f (x,h),
 2
∂f (x,h) 1 ∂ 2 f (x,h) ∂f (x,h)
= − q̇(x) +x . (5.92)
∂x 2 ∂h2 ∂h
Note that both Eq. (5.87) and Eq. (5.92) imply that when q̇(x) = 0 – i.e., q(x) is
constant – the function f (x,h) is independent of x. The initial condition for f (x,h)
can then be equivalently specified in x = 1 or in x = xM and is given by
f (1,h) = f (xM ,h) = log g(1,h) = log γqd −q(1)  r(h). (5.93)
This initial condition should be evolved down to x = xm using Eq. (5.92).
According to Eq. (5.86), the final value of the free energy is given by fd [q(x);qd ]
= γq(0)  f (m0,H ). The specific value of m0 is irrelevant as long as m0 < xm ,
because f (x,h) is independent of x for x < xm . One can thus choose m0 = 0. In
summary, the free energy is given by the solution of
f (1,h) = log γqd −q(1)  r(h),
1
f˙(x,h) = − q̇(x) f  (x,h) + xf  (x,h)2 , 0 < x < 1, (5.94)
2
d
fd [q(x);qd ] = lim fd (q̂) = γq(0)  f (0,H ) ,
s→0 ds

recalling that all the functions are constant in the regions where q̇(x) = 0, and
using primes to denote derivatives with respect to h. The differential term defined
by Eq. (5.66) can thus be computed implicitly by solving the partial differential
Eq. (5.92). Note that Eqs. (5.94), which were first introduced in [285], are very
general. They only depend on the differential structure of the term fd (q̂) and on the
hierarchical structure of the matrix q̂.

5.4.2 Variational Equations

We found that, for the fullRSB case, a generic replicated free energy, in the limit
s → 0, is given by
f[q(x);qd ] = fa [q(x);qd ] + γq(0)  f (0,H ), (5.95)
where fa [q(x);qd ] is a simple functional of q(x) and qd , and the second term
depends on the function f (x,h) that satisfies Eq. (5.94). The values of qd and
q(x) have to be obtained as stationary points of the replicated free energy. This
operation is difficult because unlike fa [q(x);qd ], for which the explicit dependence
 5.4 Replicated Free Energies and Hierarchical Matrices 171

on q(x) and qd (and therefore the derivatives) are known, the term fd [q(x);qd ] is
given implicitly in terms of the solution of Eq. (5.94). A standard way to solve this
problem is to introduce Lagrange multipliers that enforce both the partial differ-
ential equation in (5.92) and its initial condition [333]. One can therefore define a
variational replicated free energy as
fL [q(x);qd ] = fa [q(x);qd ] + γq(0)  f (0,H )
 ∞
 
− dhP (1,h) f (1,h) − log γqd −q(1)  r(h) (5.96)
−∞
 1  ∞  
1
+ dx dhP (x,h) f˙(x,h)+ q̇(x) f (x,h) + xf (x,h) .
  2
0 −∞ 2
The Lagrange multipliers P (x,h) and P (1,h) actually have a physical meaning.
For example, in the Sherrington–Kirkpatrick spin glass model, they encode the
distribution of local magnetic fields within metabasins at level x in the hierarchi-
cal structure of states [254]. Having introduced the Lagrange multipliers P (x,h)
and P (1,h), the variations must be taken in the enlarged space of all independent
functions. These are the two functions P (x,h) and f (x,h), their initial conditions
P (1,h) and f (1,h), the function q(x) and qd . Note that the initial condition f (1,h)
can also be replaced by the final condition f (0,h) because the equation for f (x,h)
can be solved in both directions in x. This substitution has the advantage that f (0,h)
appears explicitly in the first line of Eq. (5.96). Differentiating the variational free
energy with respect to P (1,h), P (x,h), f (x,h) and f (0,h) gives
f (1,h) = log γqd −q(1)  r(h),
1
f˙(x,h) = − q̇(x) f  (x,h) + xf  (x,h)2 ,
2
1   (5.97)
Ṗ (x,h) = q̇(x) P  (x,h) − 2x P (x,h)f  (x,h) ,
2
1 (h−H )2
P (0,h) = √ e− 2q(0) .
2πq(0)
The last two equations are obtained by integrating by parts the last line of Eq. (5.96)
in order to make explicit the dependence on f (x,h) and f (0,h). Note that these
equations are very general because they follow only by the form of the differential
term fd [q(x);qd ] and its expression in terms of Eq. (5.94).
Finally, taking the derivatives with respect to qd and q(x) gives
dfa [q(x);qd ] 1 γqd −q(1)  r  (h)
=− ,
dqd 2 γqd −q(1)  r(h)
 (5.98)
δfa [q(x);qd ] 1
= dhP (x,h)f  (x,h)2 .
δq(x) 2
172 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

Eqs. (5.97) and (5.98) form a complete set of equations that can be solved numer-
ically to fully determine f (x,h), P (x,h), q(x) and qd . Once this is done, the free
energy is obtained from Eq. (5.95).

5.5 De Almeida–Thouless Transition and Marginal Stability

The fullRSB variational equations obtained in Section 5.4.2 contain the kRSB
ansatz, for any finite k, as a particular case. Because the kRSB matrices form a
closed algebra, a kRSB ansatz for q(x) always provides a consistent solution of
the variational equations. This is true, in particular, for the case k = 0, which
corresponds to a constant q(x) = q0 , and, hence, a RS solution always exists.
One can ask, however, whether this solution is the correct one. As discussed in
Section 5.1.3, the RS solution sometimes has (at least) one negative eigenvalue
of the stability matrix, and it then corresponds to a saddle point in the space of
q(x), which makes it linearly unstable towards RSB. In this section, we discuss,
using Eqs. (5.97) and (5.98), how this instability can occur via a continuous phase
transition called the de Almeida–Thouless (dAT) transition [113, 254]. In order to
discuss the stability of the RS solution, instead of computing the whole spectrum
of the stability matrix, we take a shortcut due to Sommers [331, 332].
Our derivation relies on the assumption that the derivatives of fa [q(x);qd ] have
a simple structure. An example, which we use in this section, is

d δfa [q(x);qd ]
= f(2)
a [q(x);qd ] q̇(x),
dx δq(x)
(5.99)
d (n)
f [q(x);qd ] = f(n+1) [q(x);qd ] q̇(x), n ≥ 2.
dx a a

One can check that Eq. (5.99) is true for the examples of fa [q(x);qd ] discussed
in Section 5.4. In other examples – in particular, the one discussed in Chapter 6 –
the term q̇(x) in Eq. (5.99) is replaced by the Fourier transform, λ̇(x) = −x q̇(x),
introduced in Eq. (5.28). This change only affects certain details of the formulae
we derive in this section, but the overall procedure is generic.

5.5.1 Instability of the Replica Symmetric Solution

The second of Eqs. (5.98) holds for all x ∈ [0,1]. Its derivative with respect to x,
under the assumption of Eq. (5.99), gives

1 d
f(2)
a [q(x);qd ] q̇(x) = dhP (x,h)f  (x,h)2 . (5.100)
2 dx
 5.5 De Almeida–Thouless Transition and Marginal Stability 173

Using Eqs. (5.97), one can show15 that

 
d
dhP (x,h)f (x,h) = q̇(x) dhP (x,h)f  (x,h)2,
 2
(5.101)
dx
and, therefore, Eq. (5.100) becomes

q̇(x)
a [q(x);qd ] q̇(x) =
f(2) dhP (x,h)f  (x,h)2 . (5.102)
2
If there exist an interval for x ∈ [xm,xM ] ⊆ [0,1] over which q̇(x) = 0, we can
simplify the factor q̇(x) in Eq. (5.102) and obtain the relation

1
fa [q(x);qd ] =
(2)
dhP (x,h)f  (x,h)2, x ∈ [xm,xM ]. (5.103)
2
Let us consider a model for which replica symmetry is continuously broken at a
dAT phase transition. In the replica symmetry broken region, close to the phase
transition, we have a finite region x ∈ [xm,xM ] in which Eq. (5.103) holds. Upon
approaching the phase transition point, xM − xm → 0 so that the replica symmetric
solution is recovered, and q(x) becomes constant.16 Yet, by continuity, at the tran-
sition point, Eq. (5.103) is satisfied at x = xm = xM . Because q(x) = q0 is constant
at the transition point, we have, in particular, q(xm ) = q(xM ) = q0 , and P (x,h)
and f (x,h) are independent of x. We then obtain the condition

1
fa [q(x);qd ]|q(x)=q0 =
(2)
dhP (0,h)f  (1,h)2, (5.104)
2
where f (1,h) and P (0,h) are given by Eqs. (5.97) with q(0) = q(1) = q0 .
Starting from the fullRSB equations, one can therefore obtain a simple condition
for the instability of the replica symmetric solution. Eq. (5.104) is equivalent to the
vanishing of one of the eigenvalues of the stability matrix.

5.5.2 Perturbative Expansion Close to the dAT Transition

In Section 5.5.1, we have shown that the instability of the RS solution is obtained
from Eq. (5.104). Next, we can compute the properties of the function q(x) close to
that instability, in order to understand what happens when replica symmetry breaks
down. In this section, we first derive an equation for the ‘breaking point’ x ∗ where
the solution q(x) develops a deviation from the constant profile – i.e., q̇ (x ∗ ) = 0 –
and we then derive an equation for the value of q̇ (x ∗ ). As we show next, these two
numbers characterise the RSB solution in the vicinity of the dAT instability.

15 The proof is based on using Eqs. (5.97) to replace Ṗ (x,h) and f˙(x,h) by derivatives over h and then
integrating by parts on h. Note that the boundary terms at h → ±∞ in the integrations by parts vanish
because of the asymptotic properties of P (x,h) and f (x,h).
16 A concrete example will be given in Figure 6.2 in a slightly different context.
174 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

The Breaking Point

Taking the derivative of Eq. (5.103) with respect to x, using Eqs. (5.99) and the
relation

d
dhP (x,h)f  (x,h)2 =
dx
   (5.105)
 
= q̇(x) dhP (x,h)f (x,h) − 2x dhP (x,h)f (x,h) ,
2 3

which follows from the variational Eqs. (5.97), one obtains

 
2fa [q(x);qd ] = dhP (x,h)f (x,h) − 2x dhP (x,h)f  (x,h)3 .
(3)  2
(5.106)

This equation can be inverted to obtain


dhP (x,h)f  (x,h)2 − 2f(3)
a [q(x);qd ]
x=  . (5.107)

2 dhP (x,h)f (x,h)3
Evaluating this last equation on the replica symmetric solution, at the dAT phase
transition when the marginal stability condition in Eq (5.104) is satisfied, one
obtains an expression for the breaking point x ∗ . There are then two possibilities:
1. x ∗ ∈ [0,1]: the instability of the replica symmetric phase gives rise to a contin-
uous transition to a genuine RSB phase. The broken phase could be fullRSB or
kRSB with finite k.
2. x ∗ ∈
/ [0,1]: the dAT transition is unphysical and must be preceded by another
instability, typically a discontinuous RSB transition.

The Nature of the Broken Phase

We now assume that x ∗ ∈ [0,1] and that q(x) is analytic in the vicinity of the
transition point and of x = x ∗ . Taking the derivative of Eq. (5.106) with respect
to x, and using Eqs. (5.99) and the relation
  
d  
dhP (x,h)f (x,h) − 2x dhP (x,h)f (x,h) =
2 3
dx
 
= q̇(x) dhP (x,h)q4 (x,h) − 2 dhP (x,h)f  (x,h)3, (5.108)

q4 (x,h) = f  (x,h)2 − 12xf  (x,h)f  (x,h)2 + 6x 2 f  (x,h)4,

which again follows from Eqs. (5.97), we obtain

2 dhP (x,h)f  (x,h)3
q̇(x) =  . (5.109)
dhP (x,h)q4 (x,h) − 2f(4)
a [q(x);qd ]
 5.5 De Almeida–Thouless Transition and Marginal Stability 175

Evaluating this equation on the replica symmetric solution in x ∗ has two possible
outcomes:
1. q̇(x ∗ ) > 0: the solution is of the continuous fullRSB type, and the profile q(x)
is continuous in the vicinity of the transition point.
2. q̇(x ∗ ) < 0: the fullRSB solution cannot be accepted, because the term
fd [q(x);qd ] is well defined only for monotonously increasing functions, as
follows from Eq. (5.86), which is not well defined if qi < qi−1 . Usually, the
transition point is then a continuous transition towards a 1RSB phase, in which
the profile q(x) has a discontinuous jump at x = x ∗ .

Going on with the Expansion

When the dAT instability is a transition towards a continuously broken phase, one
can iterate this approach to compute higher derivatives of q(x) at the breaking
point x ∗ . In this way, one can construct a series expansion of q(x) around the
replica symmetric solution near the phase transition point. This strategy was carried
out extensively in the study of the Sherrington–Kirkpatrick model close to the
transition temperature [107]. This analysis shows that the transition is continuous
and thermodynamically of third order. In other words, the derivatives of the free
energy up to the second order are continuous, but from the third order on, they
display a jump with no divergence.

5.5.3 Marginal Stability

The existence of multiple statistically equivalent equilibrium states is the distinctive
hallmark of replica symmetry breaking. The way in which these states are dis-
tributed in phase space defines two very different scenarios.
1. States are scattered in phase space in a nearly random way and stay at a minimum
distance from one another, and the free energy barriers separating them are very
high. In this situation, if the system is inside one state, it does not feel the
existence of the other equilibrium states. This scenario physically corresponds
to stable glasses and mathematically corresponds to a finite (and typically small)
number of replica symmetry breaking steps.
2. States are not randomly distributed in phase space, but each state is surrounded
by a large number of other states that are arbitrarily close. The barriers that
separate one state from another might then be very small. This situation
mathematically corresponds to full replica symmetry breaking. Physically, if
the system is inside a state, it can move in phase space in many directions
(typically corresponding to the directions of nearby states) without the free
energy increasing too much. In other words, there are nearly flat directions in
176 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

the free energy function, exactly as at a second-order phase transition point.

The spectrum of small oscillations within an equilibrium state then displays
an excess of low-frequency modes that in some cases dominate over those
coming from more conventional sources (e.g., phonons or magnons) [150, 254].
In the limit s → 0, the permutation symmetry indeed becomes similar to a
continuous symmetry, and one can thus derive Ward identities and the existence
of Goldstone modes [117]. The precise form of this anomalous low-frequency
spectrum and the localisation properties of the eigenvalues depend on the details
of the model, but marginally stable glasses of this kind are all self-organised
critical systems.
The two scenarios described previously differ in many physically observable
ways:

• Correlation functions of non-conserved local quantities, in both space and

time, decay exponentially in stable glasses, while they decay as power laws
in marginally stable glasses. This is a consequence of the stability matrix in
Eq. (5.9) having zero modes.
• In marginally stable glasses, there is a divergent linear susceptibility, given by a
four-point correlation χab;cd = qab qcd  − qab  qcd , which is the inverse of the
stability matrix. The physical meaning of χab;cd depends on the model, but it is
generally the response of qab  to a field conjugated to qab . In spin glass models,
this can be written as the so-called spin glass susceptibility [254]. In addition,
non-linear susceptibilities are finite in stable glasses, while they are often infinite
in marginally stable glasses [64].
• If we consider a marginally stable glass on which an infinitesimal perturbation
is applied, typically, the system changes state, leading to a rearrangement (an
‘avalanche’) of the order of the size of the system. These avalanches are power
law distributed at zero temperature, with an exponent related to the shape of
the function x(q) [148, 225]. For both stable and marginally stable glasses, the
application of a finite external perturbation may force the system to jump from
one state to another, quite different, state. This behaviour is said to be ‘chaotic’
because the set of equilibrium states has a chaotic dependence on the parameters
of the Hamiltonian [68].

5.6 Wrap-Up
5.6.1 Summary
In this chapter, we have provided a compendium of mathematical and physical
results on replica symmetry breaking. In particular, we have seen that
 5.6 Wrap-Up 177

• Replica symmetry breaking corresponds to the coexistence of a large number of

equivalent thermodynamic states. In disordered systems, the Hamiltonian has no
explicit symmetry, and the coexistence of many states results from the disorder.
The states are statistically equivalent and can thus coexist over a full region of the
phase diagram (Section 5.1).
• In all known mean field models, replica symmetry breaking is mathematically
described by a particular, hierarchical, structure of the mean square displacement
matrix ab (or overlap matrix qab ). Hierarchical matrices are defined, for finite
number s of replicas, by a block structure with k levels, called a kRSB structure.
These matrices have many interesting properties: (1) at fixed k, they form a
closed commutative algebra; (2) they break the replica symmetry by inducing
correlations between replicas but preserving the statistical equivalence of each
replica; (3) a kRSB hierarchical matrix q̂ is fully specified by a finite set of 2k + 2
parameters {mi },{qi },qd for any s and, thus, an explicit analytical continuation to
any real s and in particular to s → 0 is possible (Section 5.2.1).
• When the number of RSB steps k goes to infinity, the {mi } accumulate and
become a continuous parameter x ∈ [0,1] (Section 5.2.2). The resulting fullRSB
matrices are then parametrised by {q(x);qd }, where qd is the diagonal element
and q(x) is a continuous function of x ∈ [0,1], which encodes the off-diagonal
elements. FullRSB matrices also form a closed commutative algebra and allow
an explicit analytical continuation to real s. The algebra of fullRSB matrices has
interesting mathematical properties. In particular, it is diagonalised by a simple
Fourier transform, which permits the easy computation of the inverse and deter-
minant of such matrices (Section 5.2.3).
• In a replica symmetry broken phase, the coexisting thermodynamic states have
a non-trivial distribution of mean square displacements, P(), which can be
expressed in terms of the matrix ab (Section 5.3.1). For the replica symmet-
ric case (k = 0), P() is a delta function; for kRSB, P() has k + 1 delta
peaks; for the fullRSB limit, P() has a finite support on an interval [M ,m ]
(Section 5.3.2). From the distribution of distances between three states, one can
further infer that the states are organised in an ultrametric tree, which can be
generated by a simple Poisson process (Section 5.3.3).
• In many mean field models, the replicated free energy is expressed as a function
of the overlap matrix q̂, by the sum of two terms: a simple ‘algebraic’ term
and a ‘differential’ term of a particular form. Thanks to this structure, if q̂ is a
hierarchical matrix, the free energy can be compactly written in the limit s → 0 as
a functional of {q(x);qd }, in terms of a differential equation for a function f (x,h).
In addition, the variational equations for {q(x);qd } can be written explicitly by
introducing a Lagrange multiplier P (x,h) conjugated to f (x,h). One ends up
178 Replica Symmetry Breaking and Hierarchical Free Energy Landscapes

with two differential equations that must be solved to obtain the thermodynamic
free energy (Section 5.4).
• From the differential equations, one can investigate the stability of the replica
symmetric solution via a simple argument and derive a condition for the insta-
bility point, which defines the de Almeida–Thouless (dAT) phase transition (Sec-
tion 5.5.1). A perturbative expansion for the non-constant part of q(x) can be
developed around the dAT transition in order to investigate the nature of the
replica symmetry broken phase in the vicinity of the transition (Section 5.5.2).
A fullRSB phase is characterised by marginal stability, which leads to peculiar
mathematical and physical properties (Section 5.5.3).

5.6.2 Further Reading

We provide here a list of references that can be consulted to further explore the
subjects discussed in this chapter, selected according to the criteria discussed in
Section 1.6.2.
As mentioned in the introduction to this chapter, replica symmetry breaking has
been applied to a variety of problems, giving rise to a very large literature. Several
books and reviews provide solid access points to this literature:

• Binder and Young, Spin glasses: Experimental facts, theoretical concepts, and
open questions [51]
• Mézard, Parisi and Virasoro, Spin glass theory and beyond [254]
• Fischer and Hertz, Spin Glasses [140]
• Nishimori, Statistical physics of spin glasses and information processing: An
introduction [274]
• Mézard and Montanari, Information, Physics and Computation [251]
• Talagrand, Spin glasses: A challenge for mathematicians; Cavity and mean field
models [340]
• Talagrand, Mean field models for spin glasses: Volume I; Basic examples [341]
• Panchenko, The Sherrington–Kirkpatrick model [281]
The reader can consult the first four references [51, 140, 254, 274] for a basic
introduction to the subject, including applications to spin glasses, neural networks,
and information theory. The fifth one [251] provides a more recent and complete
introduction to the application to information theory, including mathematically rig-
orous results. The last three [281, 340, 341] provide introductions to the subject for
more probabilistically oriented readers.
The existence of replica symmetry broken phases is well established in mean
field models. Their existence in finite dimensional models is, however, the subject
of a hot debate. A variety of renormalisation group methods have been developed to
investigate this question. From the field theory point of view, an introductory book
and a few more recent papers are
 5.6 Wrap-Up 179

• De Dominicis and Giardina, Random fields and spin glasses: A field theory
approach [115]
• Moore and Bray, Disappearance of the de Almeida–Thouless line in six dimen-
sions [264]
• Parisi and Temesvári, Replica symmetry breaking in and around six dimensions
[290]
• Castellana and Parisi, Non-perturbative effects in spin glasses [78]
• Charbonneau, Hu, Raju et al., Morphology of renormalization-group flow for the
de Almeida–Thouless–Gardner universality class [95]
Real space scaling arguments and renormalisation group techniques have also been
used, a few examples being

• Fisher and Huse, Equilibrium behavior of the spin-glass ordered phase [141]
• Yeo and Moore, Renormalization group analysis of the M − p-spin glass model
with p = 3 and M = 3 [364]
• Angelini, Parisi and Ricci-Tersenghi, Ensemble renormalization group for disor-
dered systems [15]
• Angelini and Biroli, Spin glass in a field: A new zero-temperature fixed point in
finite dimensions [14]
Detailed numerical simulations have also been carried out in spin glass models,
and their accuracy improves in step with the quick growth of standard computer
performances. The use of dedicated supercomputers within the Janus collaboration
has now matched the numerically and experimentally investigated time scales and
has greatly improved the system sizes studied in equilibrium. A sample of results
can be found in

• Marinari, Parisi, Ricci-Tersenghi et al., Replica symmetry breaking in short-range

spin glasses: Theoretical foundations and numerical evidences [244]
• Leuzzi, Parisi, Ricci-Tersenghi et al., Dilute one-dimensional spin glasses with
power law decaying interactions [229]
• Larson, Katzgraber, Moore et al., Spin glasses in a field: Three and four dimen-
sions as seen from one space dimension [224]
• Wang, Machta, Munoz-Bauza et al., Number of thermodynamic states in the
three-dimensional Edwards-Anderson spin glass [355]
• Janus collaboration, An in-depth view of the microscopic dynamics of Ising spin
glasses at fixed temperature [32]
• Janus collaboration, Nature of the spin-glass phase at experimental length
scales [26]
• Janus collaboration, Critical parameters of the three-dimensional Ising spin
glass [23]
 6
The Gardner Transition

In Chapter 4, we discussed the Franz–Parisi, or state following, construction and

presented results for two typical potentials within the replica symmetric ansatz. In
Chapter 5, we then saw that this replica symmetric ansatz can become unstable,
which leads to a phase transition characterised by spontaneous replica symmetry
breaking. In this chapter, we apply the general tools presented in Chapter 5 to the
state following construction of Chapter 4. We show that deep into the glass a phase
transition, at which replica symmetry is spontaneously broken, can happen. The
resulting transition is known as a Gardner transition. We then discuss the impact
of this phase transition on the phase diagram of hard and soft spheres, as well as
general properties of the replica symmetry broken phase.

6.1 State Following in the Replica Symmetry Broken Phase

We begin by recalling some results of Chapter 4 and deriving from them the expres-
sion of the glass free energy in the replica symmetry broken (RSB) phase. As
discussed in Section 4.1, we are interested in the average free energy fg (ϕ,T |ϕg,Tg )
of glass states prepared at some initial equilibrium state point (ϕg,Tg ) and then
followed adiabatically to a new state point (ϕ,T ). As shown in Section 4.2, this
free energy is expressed, in the limit d → ∞, as the extremum of a function of
a s × s matrix α̂ or, equivalently, of a n × n (with n = s + 1) matrix  ˆ in the
limit s → 0. The matrices α̂ and  ˆ encode the phase space distances between
the n replicas introduced to average over glass states. Replica 1 is special because
it encodes the initial state at (ϕg,Tg ), but all the other replicas, a = 2, . . . ,n are
equivalent. As a consequence, the mean square displacement, a1 = 1a = r , is
independent of a > 1 (recall that 11 = 0). For a,b = 2, . . . ,n we have, following
Eq. (4.64):

αab = (d/2g ) (xa − x1 ) · (xb − x1 ) = r − ab /2, (6.1)

180
 6.1 State Following in the Replica Symmetry Broken Phase 181

where g is the typical interaction scale of the potential in the initial state. One
should keep in mind that replica symmetry only holds for the s replicas – i.e.,
a = 2, . . . ,n – and, therefore, spontaneous RSB can only occur in that same sector.
Replica 1 always remains distinct. In Section 4.3, we used the simplest assumption
of replica symmetry for this sector and derived from it the replica symmetric (RS)
expression of the glass free energy. From this, in Section 4.4, we obtained the
corresponding phase diagram of glassy states. We now discuss the stability of this
ansatz against RSB.

6.1.1 The Glass Free Energy

A convenient starting point is the expression of the free energy as a function of α̂,
given by Eq. (4.54). This result should be adapted to the state following formalism,
as discussed in Section 4.2.4. In particular, we consider n = s + 1 replicas, omit the
factor β and the kinetic part of the ideal gas term of the free energy and keep replica
ϕg,Tg ) while the others are in state point ()
1 in state point () ϕ,T ), with )
ϕ=)
ϕg eη , as
defined in Eq. (4.57). We then obtain
 
ds 2πe d
− fs+1 (α̂) = − log(ρg ) +
d
log 2
+ log det α̂
2 d 2
 ∞
d)ϕg @ A
+ dh eh e−βg v̄(h) fd (α̂,h) − 1 , (6.2)
2 −∞
n ∂2  n 00
fd (α̂,h) = e a,b=2 ab ∂ha ∂hb e−β a=2 v̄(ha −η+αaa ) 0
α
.
ha =h

After extremising over α̂, the glass free energy is recovered taking the derivative
with respect to s and setting s = 0, as in Eq. (4.16).
We now assume that the matrix  ˆ has a hierarchical structure, as discussed in
Chapter 5, and follow the procedure detailed in Section 5.4 to obtain the fullRSB
expression of the free energy. Because the diagonal term aa = 0, the hierarchi-
cal matrix is fully encoded by a function (x). Correspondingly, the matrix α̂ is
encoded, according to Eq. (6.1), by

α̂ → {α(x) = r − (x)/2;αd = r }. (6.3)

Our aim is to express the free energy as a function of r and (x). The determinant
of α̂ can then be obtained by substituting Eq. (6.3) into the general expression given
in Eq. (5.29):
   1
d λ(1) 2r − (0) ˙
(x)
lim log det α̂ = log + − dx , (6.4)
s→0 ds 2 λ(0) 0 λ(x)
182 The Gardner Transition

where1
 1
λ(x) = x(x) + dy(y),
x
˙
λ̇(x) = x (x), λ(1) = (1), (6.5)
 1
(x) = λ(1) + dy λ̇(y)/y.
x

The differential term fd (α̂,h) that appears in Eq. (6.2) has the same structure as
that discussed in Section 5.4.1. Similarly to Eq. (5.80), but taking into account that
αd = r , d = 0 and αi = r − i /2, we thus have (all the sums are over
a,b = 2, . . . ,n)
   . /2
∂2 0 ∂
αab = r −
a,b
∂ha ∂hb 2 a
∂ha
⎛ ⎞2 (6.6)
  
(i − i+1 )  ⎝ ∂ ⎠ k  ∂ 2
k−1
+ + .
i=0
2 (i) (i)
∂ha 2 a ∂ha
l a∈l

Following the same steps as in Section 5.4.1, we thus obtain

g(mk,h) = γk  e−β v̄(h−η+r ),
g(mi−1,h) = γi−1 −i  g(mi ,h)mi−1 /mi , i = k, . . . ,1, (6.7)
fd (α̂,h) = γ2r −0  g(m0,h)s/m0 .
Plugging these results into Eq. (6.2), taking the derivative with respect to s and the
limit s → 0, one obtains in the continuous RSB limit
    1 
d πeλ(1) 2r − (0) ˙
(x)
−βfg ()
ϕ,T |)
ϕg,Tg ) = log + − dx
2 d2 λ(0) 0 λ(x)
 ∞ (6.8)
d)ϕg
+ dh e h−βg v̄(h)
γ2r −(0)  f (0,h),
2 −∞
with
f (1,h) = log γ(1)  e−β v̄(h−η+r ),
1 (6.9)
f˙(x,h) = (x)
˙ f  (x,h) + xf  (x,h)2 , 0 < x < 1.
2
One can check that in the RS case, where (x) = λ(x) =  are constant and
f (x,h) is independent of x, the results of Chapter 4, Eq. (4.73), are recovered.

1 Note that in order to compensate for the factor 1/2 appearing in Eq. (6.3), λ(x) has here been multiplied by 2
with respect to the definition given in Eq. (5.28).
 6.1 State Following in the Replica Symmetry Broken Phase 183

6.1.2 Variational Equations

The variational equations for r and (x) can be obtained following the approach
described in Section 5.4.2. Adding to the free energy the Lagrange multipliers
P (x,h) and P (1,h) gives
    1 
d πeλ(1) 2r − (0) ˙
(x)
−βf =L
log + − dx
2 d2 λ(0) 0 λ(x)
 ∞
d)
ϕg
+ dh eh−βg v̄(h) γ2r −(0)  f (0,h) (6.10)
2 −∞

d ∞  
− dhP (1,h) f (1,h) − log γ(1)  e−β v̄(h−η+r )
2 −∞
  ∞  
d 1 ˙
(x)
+ dx ˙
dhP (x,h) f (x,h) −  
f (x,h) + xf (x,h) ,
2
2 0 −∞ 2
and differentiating with respect to f (x,h) and f (0,h) leads to the differential
equation

P (0,h) = )
ϕg γ2r −(0)  eh−βg v̄(h), (6.11)
1   
˙
Ṗ (x,h) = − (x) P (x,h) − 2x P (x,h)f  (x,h) , 0 < x < 1.
2
Note that the expression for P (0,h) can also be rewritten using the identity
in Eq. (4.89). Equations for r and (x) are then obtained by differentiating
Eq. (6.10) with respect to these two quantities. Using the identity
 ∞
d
dh P (x,h)f  (x,h) = 0, (6.12)
dx −∞
which can be proven using Eqs. (6.9) and (6.11) and a few additional manipulations,
one obtains

1 1 ∞
=− dh P (0,h)[f  (0,h) + f  (0,h)],
λ(0) 2 −∞
 x ˙  (6.13)
2r − (0) (y) 1 ∞ 
− dy = dhP (x,h)f (x,h) .2
λ(0)2 0 λ(y)2 2 −∞
Eqs. (6.9), (6.11) and (6.13) constitute a set of closed equations for f (x,h), P (x,h),
(x) and r . In solving these equations, it is convenient to use λ(x) instead of
(x); once λ(x) is determined, one can obtain (x) using Eq. (6.5). The solution
can then be obtained iteratively using the following numerical procedure.
1. Start from a reasonable guess for r and λ(x).
2. From λ(x), compute (x) by Eq. (6.5).
184 The Gardner Transition

3. Solve Eq. (6.9), starting from the initial condition in x = 1 and evolving towards
x = 0 to obtain f (x,h) for all x.
4. Solve Eq. (6.11), starting from the initial condition in x = 0 and evolving
towards x = 1 to obtain P (x,h) for all x.
1 ∞ 
5. Compute K(x) = 2 −∞ dhP (x,h)f (x,h)2 .
6. Obtain a new estimate of λ(x) by first computing a new estimate of λ(0) from
the first Eq. (6.13) and then integrating the relation2
 x
λ̇(x) 1 1
K̇(x) = − ⇒ = + dy y K̇(y). (6.14)
xλ(x)2 λ(x) λ(0) 0

7. Compute a new estimate of r by using the second Eq. (6.13) computed in

x = 0,
(0) + λ(0)2 K(0)
r = . (6.15)
2
8. Repeat steps 2–7 until convergence of r and λ(x) to a fixed point is reached.

6.1.3 Observables
Taking derivatives of the free energy in Eq. (6.10), we can obtain expressions for
several interesting observables. For example, the energy is

∂(βfg ) d ∞ ∂
eg = =− dhP (1,h) f (1,h)
∂β 2 −∞ ∂β
 ∞ (6.16)
d γ(1)  [e−β v̄(h−η+r ) v̄(h − η + r )]
= dhP (1,h) ,
2 −∞ γ(1)  e−β v̄(h−η+r )
and the reduced pressure is

∂(βfg ) d ∞
pg = = dhP (1,h)f  (1,h)
∂η 2 −∞
 ∞ (6.17)
d γ(1)  [e−β v̄(h−η+r ) β v̄  (h − η + r )]
=− dhP (1,h) .
2 −∞ γ(1)  e−β v̄(h−η+r )
More generally, Eq. (2.29) can be expressed in terms of v̄(h) to define a distribution
function g(h) of interparticle gaps h = d(r/g − 1). Starting from Eq. (2.29), using
rotational invariance and changing variables from r to h, the functional derivative
of the free energy with respect to the potential is

2 Eq. (6.14) is obtained by taking a derivative with respect to x of the second Eq. (6.13) and recalling that
˙
λ̇(x) = x (x).
 6.2 Gardner Transition and Replica Symmetry Breaking 185
  ∞
ρ d)
ϕg
δfg = drg(r)δv(r) = dheh g(h)δ v̄(h)
2 2 −∞
(6.18)
∂fg d)
ϕg h
⇒ = e g(h).
∂ v̄(h) 2
Hence, the radial distribution function of the glass, g(h), can be computed by
taking the variation of the free energy in Eq. (6.10) with respect to the potential.
It is important to stress, however, that in the state following construction, all the
thermodynamic and structural properties of the glass are expressed in terms of the
s replicas with a ≥ 2, while the replica 1 is used only to select one among all
possible glasses. Therefore, one should formally introduce two distinct potentials,
v̄g (h) for replica a = 1 and v̄(h) for those with a ≥ 2, and the variation in Eq. (6.18)
should be taken only with respect to v̄(h) at constant v̄g (h). From Eq. (6.10), one
then obtains3
 
−β v̄(h) P (1,h + η − r )
)
ϕg e g(h) = e
h
γ(1)  . (6.19)
γ(1)  e−β v̄(h)
This result can be used to express the average of any two-body observable that
1 
depends on the gaps, O[X] = 2N i=j O[d(|ri − rj |/g − 1)], as
 ∞
d)ϕg
O = dheh g(h)O(h)
2 −∞
 (6.20)
d ∞ γ(1)  [e−β v̄(h−η+r ) O(h − η + r )]
= dhP (1,h) .
2 −∞ γ(1)  e−β v̄(h−η+r )
Note that a particular case of this relation is the energy expression in Eq. (6.16).

6.2 Gardner Transition and Replica Symmetry Breaking

In this section, we obtain explicit expressions for the quantities that characterise the
instability of the RS solution at the Gardner transition and then describe a general
recipe to study RSB effects within the state following construction.

6.2.1 Instability of the Replica Symmetric Solution

Following the same procedure as in Section 5.5, we now write the equations that
characterise the instability of the replica symmetric solution at the Gardner transi-
tion. Taking the derivative with respect to x of the second Eq. (6.13) gives that for
˙
(x) = 0,

3 Because v̄ (h) (corresponding to replica a = 1) appears in the second line of Eq. (6.10), while v̄(h)
g
(corresponding to replicas a ≥ 2) appears in the third line, to compute g(h), one should take the variation
of the third line of Eq. (6.10) only.
186 The Gardner Transition
 ∞
1 1
2
= dhP (x,h)f  (x,h)2 . (6.21)
λ(x) 2 −∞

This condition must be satisfied on the continuous part of the fullRSB solution.
Taking the RS limit, in which λ(x) = , we then obtain4 that

2 ∞
λR = 1 − dhP (0,h)f  (1,h)2 (6.22)
2 −∞
  2 2
ϕg 2 ∞
) ∂
=1−  dh e q(2r − ,βg ;h)
h
log q(,β;h − η)
2 −∞ ∂h2
must vanish at the point where the replica symmetric solution becomes unstable.
Hence, λR is proportional to the replicon eigenvalue.
Following Section 5.5, we take one additional derivative of Eq. (6.21) with
respect to x using Eq. (5.105) and obtain another condition,
∞
dhP (x,h)f  (x,h)2
x = 4 −∞  ∞ , (6.23)
λ(x) 3 + 2 −∞ dhP (x,h)f  (x,h)3

which holds in the continuous region of the fullRSB solution. Evaluating this equa-
tion on the RS solution at the instability point gives the breaking point
∞
dhP (0,h)f  (1,h)2
x = 4 −∞ ∞
∗

3
+ 2 −∞ dhP (0,h)f  (1,h)3
∞ ' 3 (2 (6.24)
)
ϕg −∞ dh eh q(2r − ,βg ;h) ∂h ∂
3 log q(,β;h − η)
= ∞ ' (3 .
∂2
4
3
+ 2)
ϕg −∞ dh e q(2r − ,βg ;h) ∂h2 log q(,β;h − η)
h

Taking another derivative with respect to x using Eq. (5.108), one further obtains
∞ 
3 + 2 −∞ dhP (x,h)f (x,h)
4 3
˙ λ(x)
(x) = 12x 2  ∞ ,
λ(x)4
− −∞ dhP (x,h)A(x,h) (6.25)
A(x,h) = f  (x,h)2 − 12xf  (x,h)f  (x,h)2 + 6x 2 f  (x,h)4 .
Evaluating this expression at the instability point of the RS solution formally
amounts (after some manipulation) to replacing, as before, λ(x) → , x → x ∗ ,
P (x,h) → )ϕg eh q(2r − ,βg ;h), f (x,h) → log q(,β;h − η).
˙ ∗ ), we can study the stability of the RS solution following
Given λR , x ∗ and (x
the approach presented in Section 5.5.

4 Note that the second line of Eq. (6.22) is obtained from the first by using the RS expressions of P (0,h) and
f (1,h), applying the identity in Eq. (4.89) and using the definition of q(,β;h) in Eq. (4.74).
 6.3 Gardner Transition of Simple Glasses 187

6.2.2 Recipe for Replica Symmetry Breaking in State Following

Based on the results of Sections 5.5 and 6.1, we can outline a recipe for studying
RSB effects in the state following construction for particle glasses in d → ∞.
ϕg,Tg ) that falls into the dynamically arrested region (see
For a given initial state ()
Section 4.3.3), the procedure is as follows.
1. Use the recipe of Section 4.3.3 to follow the evolution of the initial state at any
state point ()ϕ,T ) within the RS ansatz. The RS values of  and r are then
obtained.
2. At each ()ϕ,T ), check the stability of the RS ansatz by computing the replicon
λR according to Eq. (6.22). If λR ≥ 0, the RS ansatz is stable. It is generically
found that the RS ansatz is stable on the equilibrium line, i.e., for () ϕ,T )
= () ϕg,Tg ) [303]. Usually, upon heating or decompressing, the RS ansatz
remains stable, while upon cooling or compressing, it may become unstable.
3. Identify the Gardner transition point by the condition λR = 0. Beyond this point,
the RS solution is unstable. Note that the replicon vanishes at the dynamical
transition; therefore, a glass state prepared on the dynamical transition line is
immediately unstable and cannot be followed at the RS level (as discussed in
Chapter 4).
4. Compute the breaking point x ∗ using the RS solution at the Gardner point,
according to Eq. (6.24). If x ∗ ∈ [0,1], then the transition is continuous. Oth-
erwise, there is an inconsistency, which signals that the RS ansatz has already
become unstable via another mechanism (usually a discontinuous transition).
5. Compute the slope (x ˙ ∗ ) at the Gardner point according to Eq. (6.25). If
˙ ∗
(x ) < 0, a perturbative fullRSB solution exists around the Gardner point,
which suggests that the solution is fullRSB throughout the unstable phase. If
˙ ∗ ) > 0, a 1RSB solution describes the unstable phase in the vicinity
instead (x
of the instability.
6. Once the nature of the solution has been determined, solve numerically the
RSB Eqs. (6.9), (6.11) and (6.13), as described in Section 6.1.2 (see also Sec-
tion 6.6). From this solution, compute the observables (energy, pressure, etc.),
as described in Section 6.1.3.
This procedure can be applied to any interaction potential. Specific examples are
considered in Section 6.3.

6.3 Gardner Transition of Simple Glasses

In this section, we apply the procedure outlined in Section 6.2.2 to two represen-
tative potentials, hard spheres and the inverse power-law potential (soft spheres),
already studied at the RS level in Chapter 4.
188 The Gardner Transition

RS glass ϕd

g = 8, 7, 6, 5.5, 5
ϕ
1/p

Equilibrium liquid

FullRSB glass

ϕ Jamming line

Figure 6.1 Following hard-sphere glasses in compression and decompression,

within the full replica symmetric breaking ansatz [298]. The inverse of the reduced
pressure p ) = p/d is plotted versus packing fraction ) ϕ = 2d ϕ/d. As in Figure 4.4,
the liquid equation of state is plotted as a full line, and the dynamical transition )
ϕd
is marked by a full circle. The equations of state of glasses prepared at ) ϕg > ) ϕd
are reported here as full lines in the region where the replica symmetric ansatz is
stable. Upon compression, all glasses undergo a Gardner transition at a density
)
ϕG ()ϕg ), marked by a triangle. The ‘Gardner line’ obtained by connecting the
Gardner transitions of different glasses is plotted as a dotted line. Only data for
ϕg > )
) ϕg† are reported, for which glasses are described by a fullRSB ansatz
beyond the Gardner transition (dashed lines). The glass equations of state end
at the jamming point, at which the pressure diverges5 .

6.3.1 Hard Spheres

The replica symmetric phase diagram of hard spheres has been discussed in
Section 4.4.1. Replica symmetry breaking effects for this model have been studied
in [88, 89, 221], and, in particular, state following results have been obtained
in [298, 299]. See [93] for a review. Results obtained5 by solving the kRSB
equations with k = 99, following the procedure discussed in Section 6.6, are
reported in Figure 6.1 and Table 6.1. These results remain stable upon further
increasing k [298]. Upon decompression, the RS ansatz always remains stable, and

5 Some of the results reported in Figure 6.1 and in Table 6.1 are taken from [298], but the numerical solution
has been improved and new results have been produced for this book. Despite these improvements, for
ϕg = 5, p
() )  5.85), ()ϕg = 5.5, p
)  21.4) and () ϕg = 6, p
)  324), the code is unstable, and convergence
could not be reached. We believe that this is just a numerical instability of the code, and in Figure 6.1, the
ϕg ≥ 7, the
equations of state have thus been extended to infinite pressure by a linear fit. By contrast, for )
code converges up to infinite pressure.
 6.3 Gardner Transition of Simple Glasses 189
Table 6.1 Gardner transition density ) ϕG , order parameters  and r , breaking
∗
point x , slope at the breaking point (x˙ ∗ ) and dynamical critical exponents γ
and a for several initial densities )
ϕg in hard spheres. The first line corresponds to
ϕg = )
) ϕd and is reported with higher precision because the calculation is
numerically simpler.

)
ϕg )
ϕG r  x∗ ˙ ∗)
(x γ = 1/a
4.80677 4.80677 1.15336 1.15336 0.70698 19.357 3.08627
5 5.64 0.599 0.436 0.512 −6.950 2.547
5.5 6.61 0.333 0.169 0.397 −1.929 2.363
6 7.33 0.088 0.228 0.340 −0.990 2.295
7 8.54 0.133 0.033 0.280 −0.402 2.228
8 9.63 0.0888 0.0156 0.248 −0.204 2.194
9 10.67 0.0643 0.00847 0.228 −0.119 2.176

the results of Chapter 4 are unchanged. The glass states then melt into the liquid
at a spinodal point. Upon compression, by contrast, there is always a Gardner
transition, reached at a finite pressure, before jamming. For low-density states,
this Gardner transition occurs before the unphysical RS spinodal (see Figure 4.4)
is reached. Beyond the Gardner instability, a RSB solution appears and remains
stable until reaching jamming, at infinite pressure.5 Hence, no spinodal is observed
upon compression once RSB is taken into account.
The values of the Gardner transition density are given in Table 6.1 for several
values of initial density ) ϕg . By joining all the Gardner transitions corresponding
to different ) ϕg one obtains a ‘Gardner transition line’ in the () ϕ,1/)p ) phase
diagram (Figure 6.1). When the initial density tends to the dynamical transition
point – i.e., for )ϕg → ) ϕd+ – the Gardner transition approaches ) ϕg , and ultimately,
)
ϕG ()ϕd ) = ) ϕd . This feature is generic of mean field models because it can be
shown that the replicon eigenvalue vanishes at ) ϕd . As a consequence, glass
states prepared at ) ϕg = ) ϕd are always marginally stable. The Gardner line then
originates from the dynamical transition and moves to higher pressures upon
increasing )ϕg .
The breaking point x ∗ and the corresponding slope (x ˙ ∗ ) are also given in
Table 6.1. Note that x ∗ always belong to the interval [0,1], which confirms that the
transition is continuous. The inverse slope 1/(x˙ ∗ ) vanishes linearly at ) ϕg† ≈ 4.85 –
i.e., the slope changes sign at ) ϕg† . For )
ϕg > )ϕg† , the slope is negative, which is
consistent with a continuous transition towards a fullRSB solution. In Figure 6.2,
we report explicit examples of (x) for ) ϕg = 7, in the vicinity of the Gardner
transition. The fullRSB part of (x) emerges continuously around the breaking
point with the correct slope. In these cases, the solution remains fullRSB up to
190 The Gardner Transition

η = 0.198, 0.2, 0.21, 0.22, 0.23, 0.24

Δ

Figure 6.2 The function (x) for initial density ) ϕg = 7 and for several values
of η = log() ϕg ). The perturbative prediction for (x) at the instability, (x)
ϕ /)
∼  + (x˙ ∗ )(x − x ∗ ), with parameters , x ∗ and (x
˙ ∗ ) given in Table 6.1, is
reported as a dashed line. Once replica symmetry is broken, upon compression,
the difference between the two plateaus (0) and (1) increases, and the shape
of the continuous part becomes increasingly non-linear. The function of (x) at
larger values of η, close to jamming, is reported in Figure 9.6.

jamming.6 Instead, in the region of densities )

ϕd ≤ )ϕg  )ϕg† , the slope is positive,
and the Gardner transition is a continuous transition towards a 1RSB phase. This
transition is followed by a second continuous transition from 1RSB to a fullRSB
solution, which then persists up to jamming.7 The properties of the fullRSB phase
that emerges beyond the Gardner transition are further discussed in Chapter 9.

6.3.2 Soft Spheres

The soft sphere potential v̄(h) = εe−4h , which is the infinite-dimensional limit
of a pure inverse power-law potential, v(r) = ε(/r)4d , was studied at the RS
level in [315], as discussed in Section 4.4.2. In the following, we fix ε = 1 for
simplicity. Because for this potential the unique control parameter is )
 = )
ϕ T 1/4

6 The numerical solution of the fullRSB equations indicates that for low )
ϕg  6, at some intermediate density,
there might be a phase transition from a phase with continuous (x) to another phase in which (x) develops
a discontinuity. This transition has been discussed in [108] for a spin glass model but has not been investigated
systematically for hard spheres in d → ∞.
7 This second transition has not yet been studied in detail, and RSB results for )
ϕd ≤ )ϕg  )ϕg† are thus not
reported in Figure 6.1. The solution is expected to display a discontinuity and a continuous part, as discussed
in [108].
 6.3 Gardner Transition of Simple Glasses 191

Td RS glass

FullRSB glass
e

Equilibrium
liquid

Tg = 0.1, 0.3, 0.5, 0.7, 0.8, 0.9

Figure 6.3 Following soft-sphere glasses with potential v̄(h) = εe−4h , under
heating and cooling, within the full replica symmetric breaking ansatz [315, 316].
The reduced energy ) e = e/d is plotted versus temperature T , both expressed in
units of the interaction strength ε, at fixed density )ϕ = 4.304. As in Figure 4.5, the
liquid equation of state is plotted as a full line, and the dynamical transition Td = 1
is marked by a full circle. The equations of state of glasses prepared at Tg < Td
are reported as full lines in the region where the replica symmetric ansatz is stable.
Upon cooling, higher-energy glasses undergo a Gardner transition at a temperature
TG (Tg ), marked by a triangle. The ‘Gardner line’ obtained by connecting the
Gardner transitions of different glasses is plotted as a dotted line. Only data for
Tg < Tg† are reported, for which glasses are described by a fullRSB ansatz below
the Gardner transition (dashed lines). Lower-energy glasses, by contrast, display
no Gardner transition. Note that once RSB is taken into account, no unphysical
spinodal is present upon cooling8 (unlike in the RS case of Figure 4.5), and all
glasses can thus be followed down to T = 0.

(see Section 4.4.2), one can fix the density to )

ϕ = 4.304, corresponding to Td = 1,
and study the phase diagram as function of temperature [315]. The results includ-
ing replica symmetry breaking effects8 are reported in Figure 6.3 and Table 6.2.
Additional results on the Gardner transition in soft harmonic spheres can be found
in [58, 59, 316].

8 Most of these results are taken from [315, 316], but as in the case of hard spheres, the numerical solution of the
fullRSB equations has been improved, and new results have been produced for this book. Despite these
improvements, for (Tg = 0.9,T  0.14) and (Tg = 0.8,T  0.03), the code is unstable, and convergence
could not be reached. As for hard spheres, we believe that this is just a numerical instability of the code, and in
Figure 6.3, the equations of state have thus been extended to zero temperature by a quadratic fit. By contrast,
for Tg = 0.7, the code converges down to zero temperature.
192 The Gardner Transition
Table 6.2 Gardner transition temperature TG , order parameters  and r ,
breaking point x ∗ , slope at the breaking point (x ˙ ∗ ) and dynamical critical
exponents γ and a for several Tg of soft spheres at constant density ) ϕ = 4.304.
The first line corresponds to Tg = Td and is reported with higher precision because
the calculation is numerically simpler. Below Tg = 0.58, no Gardner transition is
observed.

Tg TG r  x∗ ˙ ∗)
(x γ = 1/a
1 1 1.31228 1.31228 0.66128 13.0198 2.92034
0.98 0.730 1.05 0.924 0.554 22.7 2.637
0.95 0.570 0.912 0.722 0.488 41.7 2.508
0.90 0.414 0.767 0.529 0.407 −124 2.382
0.85 0.306 0.667 0.399 0.344 −33.2 2.300
0.80 0.216 0.586 0.292 0.283 −23.7 2.231
0.75 0.150 0.518 0.209 0.225 −18.5 2.173
0.70 0.098 0.459 0.141 0.168 −16.5 2.123
0.65 0.0585 0.406 0.0871 0.112 −16.7 2.078
0.60 0.018 0.359 0.0282 0.0417 −33.8 2.027
0.58 0 0.338 0 0 ∞ 2

As for hard spheres, we find that the RS solution for states prepared at Tg < Td
and heated to T > Tg remains stable up to the spinodal point where the glass melts
into the liquid. The results of Section 4.4.2 are therefore unchanged upon heating.
Upon cooling, the RS solution for the lowest-energy states (corresponding to the
lowest Tg ) remains stable down to T = 0. These states, therefore, do not display any
Gardner transition (see Figure 6.3), and also in this case, the results of Section 4.4.2
remain unchanged. A Gardner transition emerges continuously from T = 0 when
Tg  0.58, and for Tg ∈ [0.58,1], the states display a Gardner transition at a
finite TG (Tg ) upon cooling. The values are reported in Table 6.2. In the vicinity of
Tg = 0.58, when TG → 0, one can show analytically that  → 0, x ∗ ∼ 1.59 → 0
˙ ∗ ) ∼ −0.76/ → −∞. Hence, the breaking point tends to zero, and the
and (x
slope tends to negative infinity when TG = 0. For Tg > 0.58, x ∗ always remains
between 0 and 1, signalling a continuous transition, as for hard spheres. The slope
˙ ∗ ) changes sign around Tg† ≈ 0.92. For Tg† < Tg < Td , the slope is positive,
(x
indicating a continuous transition towards a 1RSB phase.9

9 It is currently not known whether, during the cooling process, this 1RSB phase undergoes another transition to

fullRSB at a lower temperature. As for hard spheres, RSB results for Tg† < Tg < Td are thus not reported in
Figure 6.3. Furthermore, as in the case of hard spheres, the numerical solution of the fullRSB equation
suggests the existence of a phase in which (x) has a discontinuity, on top of a continuous part [108],
for relatively high Tg and low T .
 6.4 Critical Properties of the Gardner Transition 193

6.4 Critical Properties of the Gardner Transition

To conclude the discussion of the Gardner transition, we discuss here (shortly and
without proof) some of its critical properties. The Gardner transition is a continu-
ous phase transition that takes place within one of the glass basins that appeared
discontinuously at the dynamical transition.
In order to better illustrate this idea, the organisation of phase space in the
Gardner phase is sketched in Figure 6.4. As discussed in Chapter 4, glass states
emerge discontinuously at the dynamical transition, where diffusion stops. In phase
space, this means that glass basins form somewhat before the dynamical arrest
but remain connected by relatively narrow phase space bottlenecks that must be
dynamically explored in order for the liquid to relax. A plateau in the mean square
displacement is correspondingly observed at intermediate times (Figure 3.3). At the
dynamical transition, however, these bottlenecks close; hence, the relaxation time
becomes infinite, and the long-time limit of the MSD discontinuously jumps from
infinity to a finite value. Equilibrium configurations at (ϕg,Tg ) in the dynamically
arrested phase thus select one of the stable glass states that are uniformly scattered

(ϕg , Tg )

(ϕ, T )
Dr Y
X5 X2
Y

X3
Dg X4

Figure 6.4 A schematic picture of the organisation of phase space in the Gardner
phase. (Left) As in Figure 4.2, a glass state is selected by the configuration Y
extracted in equilibrium at (ϕg ,Tg ), with associated mean square displacement Dg .
Glass states are uniformly scattered in phase space. The connections between them
close discontinuously at the dynamical transition. (Right) Once followed at a state
point (ϕ,T ) within the Gardner phase, the state selected by Y at (ϕg ,Tg ) becomes
a metabasin, hierarchically organised in sub-states, as discussed in Chapter 5. The
structure of sub-states emerges continuously at the Gardner transition. The replicas
X2, . . . Xs+1 explore this structure, giving rise to a distribution of mean square
displacements Dab . The displacement between Y and any of the Xa , however,
remains Dr .
194 The Gardner Transition

across phase space. The Gardner transition, by contrast, is a continuous transition

that happens independently within each glass basin, when the basin is followed
adiabatically from (ϕg,Tg ) to (ϕ,T ). It is this glass basin that becomes critical.
Its bottom becomes flat and then fractures into many sub-basins, similarly to what
happens to the paramagnetic state at the transition to ferromagnetism (Figure 1.4).
In the following, we denote by  the distance from the Gardner transition point
along a state following path, with  > 0 in the RS phase and  < 0 in the RSB
phase. For example, in the hard-sphere case, we have  = ) ϕg ) − )
ϕG () ϕ for a given
ϕg , while in soft spheres, we have  = T − TG (Tg ) for a given Tg . Note that for
)
generic interaction potentials, both density and temperature are control parameters,
and one can move away from the Gardner transition point in both directions within
a given state [316].

6.4.1 Thermodynamic Properties

The thermodynamic properties of the Gardner transition are quite unusual. In the
Ehrenfest classification, the transition is of third order – i.e., all the derivatives of
the free energy (with respect to density, temperature, shear strain, . . . ) up to second
order included are continuous at the Gardner line, while higher order derivatives
are discontinuous but not divergent.
As an example, consider the equations of state of the glasses reported in
Figures 6.1 and 6.3. The pressure and energy of the glass are first derivatives
of its free energy, with respect to density and temperature, respectively. They are
always continuous functions of the control parameters. The derivative of the energy
with respect to temperature is the specific heat, the derivative of the pressure with
respect to density is related to the compressibility, and both are second derivatives
of the free energy. While they display a jump at the glass transition Tg , they remain
continuous at the Gardner point; the slope of the equations of state is indeed
continuous in Figures 6.1 and 6.3. Only the third derivatives of the free energy,
such as the derivative of the specific heat with respect to temperature, show a
discontinuous jump at the Gardner transition.
A possible order parameter for this transition, as shown in Figure 6.2, is the
difference δ = (0) − (1), which grows linearly, δ ∝ ||, on the RSB side of
the transition.

6.4.2 Dynamical Properties

The Gardner transition displays a dynamical critical behaviour similar to a standard
second-order phase transition [288]. This dynamical criticality is also similar to that
of the dynamical transition discussed in Section 3.4, with the important difference
that, in this case, the system is confined into a glass state, and the dynamical
 6.5 Wrap-Up 195

mean square displacement (t) has a plateau at long times, limt→∞ (t) = ;
hence, dynamics is arrested, and caging is permanent (Figure 4.1). At the Gardner
transition, one has for large t

(t) −  ∝ −t −a, (6.26)

where the exponent a is expressed in terms of the breaking point by a relation

similar to Eq. (3.85):

 2 (1 − a)
x∗ = . (6.27)
(1 − 2a)
Upon approaching the transition from the RS phase, one has at large times

(t) −  = −f (t/τ ), τ ∝  −γ , γ = 1/a, (6.28)

which reduces to Eq. (6.26) at the Gardner transition when  → 0, under the
assumption that f (x → ∞) ∼ x −a . Eq. (6.28) indicates that the relaxation to
the long time (restricted equilibrium) limit happens on a time scale τ that diverges
at the Gardner transition. The critical exponent a is reported in Table 6.1 for hard
spheres and in Table 6.2 for soft spheres.
Beyond the Gardner transition, in the RSB phase, the dynamics becomes
extremely slow. Relaxation to the restricted equilibrium limit then takes an infinite
time. If the system is prepared in the RS phase and suddenly cooled or compressed
into the RSB phase, a stationary state is never observed, and the system ages
forever. The properties of this ageing dynamics are particularly complex and will
not be reported here. Yet, even though the system is unable to equilibrate within the
glass basin in the RSB phase, the restricted equilibrium RSB calculation provides
a good approximation of its physical properties.

6.5 Wrap-Up
6.5.1 Summary
In this chapter:

• We have applied the general RSB framework discussed in Chapter 5 to the

calculation of the glass free energy in the Franz–Parisi framework. We have
derived RSB expressions for the glass free energy; the corresponding variational
equations; and the most important observables such as energy, pressure and radial
distribution function (Section 6.1).
• We have obtained stability criteria for the RS solution, as well as the perturbative
expression of the RSB solution around the RS instability. Using these criteria, we
196 The Gardner Transition

have described a recipe for studying RSB effects in the Franz–Parisi framework
(Section 6.2).
• We have applied this recipe to two representative potentials, hard spheres and
soft spheres, thus obtaining phase diagrams that include the Gardner transition
and RSB equations of state for the glass (Section 6.3).
• We have described the most important static and dynamical critical properties of
the Gardner transition and provided values for the dynamical critical exponents
(Section 6.4).

6.5.2 Further Reading

We provide here a list of references that can be consulted to further explore the
subjects discussed in this chapter, selected according to the criteria discussed in
Section 1.6.2.
The idea that glass basins are not simple minima but complex metabasins con-
taining a hierarchical structure of sub-basins has a long history in the literature
on the potential energy landscape of glasses. Some early papers and introductory
reviews are

• Goldstein, Viscous liquids and the glass transition: A potential energy barrier
picture [166]
• Stillinger and Weber, Dynamics of structural transitions in liquids [335]
• Debenedetti and Stillinger, Supercooled liquids and the glass transition [119]
• Wales, Energy landscapes: Applications to clusters, biomolecules and glasses
[354]
• Heuer, Exploring the potential energy landscape of glass-forming systems: From
inherent structures via metabasins to macroscopic transport [177]

Since the existence of a Gardner transition has been proposed in hard spheres in
the limit d → ∞, signatures of this transition in finite dimensional glasses have
been searched for in many numerical and experimental studies. Because the field is
moving rapidly, we only provide here a complete list of papers appeared until the
completion of this book:

• Hicks, Wheatley, Godfrey, et al. Gardner transition in physical dimensions [178]

• Berthier, Charbonneau, Jin et al., Growing timescales and lengthscales charac-
terizing vibrations of amorphous solids [44]
• Seguin and Dauchot, Experimental evidence of the Gardner phase in a granular
glass [324]
• Jin and Yoshino, Exploring the complex free-energy landscape of the simplest
glass by rheology [192]
 6.6 Appendix: Numerical Resolution of the RSB Equations 197

• Seoane and Zamponi, Spin-glass-like aging in colloidal and granular glasses

[326]
• Liao and Berthier, Hierarchical landscape of hard disk glasses [230]
• Jin, Urbani, Zamponi et al., A stability-reversibility map unifies elasticity, plastic-
ity, yielding and jamming in hard sphere glasses [193]
• Scalliet, Berthier and Zamponi, Absence of marginal stability in a structural
glass [315]
• Seoane, Reid, de Pablo et al., Low-temperature anomalies of a vapor deposited
glass [327]
• Geirhos, Lunkenheimer and Loidl, Johari-Goldstein relaxation far below Tg :
Experimental evidence for the Gardner transition in structural glasses? [162]
• Charbonneau, Corwin, Fu et al., Glassy, Gardner-like phenomenology in mini-
mally polydisperse crystalline systems [94]

The first paper, which reports a study of a one-dimensional system of hard spheres,
identified sub-basins associated to localised defects but pointed out that these
defects could misleadingly give rise to Gardner-like physics in certain observables.
The next six papers considered hard-sphere glasses in d = 2 and d = 3,
both numerically and experimentally. These studies observed several anomalies
potentially associated with a Gardner transition but could not establish whether
these effects were associated to a true phase transition or to a sharp crossover.
The defects associated with sub-basins are, however, certainly extended over large
regions in space and are, therefore, of collective origin. The next three papers
considered soft-potential models of structural glasses. Only localised defects have
been identified, and no Gardner transition is detected numerically. Experimental
studies remain inconclusive. The last paper reports the existence of Gardner-like
anomalies in complex crystals of hard particles.

6.6 Appendix: Numerical Resolution of the RSB Equations

The fullRSB Eqs. (6.9), (6.11) and (6.13) constitute a set of closed equations for
f (x,h), P (x,h), (x) and r , which must be solved numerically. One possibil-
ity is to follow the procedure outlined in Section 6.1.2 and simply discretise the
differential equations Eqs. (6.9) and (6.11) on a grid in (x,h).
Another possibility is to write explicitly the kRSB equations for finite k and then
increase k until the results converge to the desired precision. For completeness,
we report here the kRSB equations. These can be obtained from Eqs. (6.9), (6.11)
and (6.13) by inserting a piecewise constant form of (x), following the convention
of Figure 5.4, with s = m−1 = 0. Also, we recall that while in Figure 5.4 the
function q(x) increases monotonically, the function (x) decreases monotonically.
198 The Gardner Transition

The procedure to solve the kRSB equations is as follows. Keeping the breaking
points {mi } fixed (recalling that m−1 = 0 and mk = 1), one starts by guessing the
values of {i } and r . Then, the following steps are iterated until convergence.
1. For a piecewise constant (x), introducing the function q(,β;h) defined in
Eq. (4.74), Eq. (6.9) takes the form:
f (mk,h) = log q(k,β;h − η + r − k /2), (6.29)
1
f (mi ,h) = log γi −i+1  emi f (mi+1,h), i = k − 1, . . . ,0.
mi
Similarly, applying the identity in Eq. (4.89), Eq. (6.11) takes the form:
ϕg eh+r −0 /2 q(2r − 0,βg ;h + r − 0 /2),
P (m0,h) = )
P (mi ,h) = emi−1 f (mi ,h) γi−1 −i  [P (mi−1,h)e−mi−1 f (mi−1,h) ], (6.30)
i = 1, . . . ,k.
Solve these two equations in the order in which they are written to obtain
f (mi ,h) and P (mi ,h).
2. Compute

1
Ki = dh P (mi ,h)f  (mi ,h)2, i = 0, . . . ,k, (6.31)
2
and from it, compute a new estimate of λ(x) from a discrete version of
Eq. (6.14):

1 1
=− dh P (m0,h)[f  (m0,h) + f  (m0,h)],
λ0 2
(6.32)
1 1
= + mi−1 (Ki − Ki−1 ), i = 1, . . . ,k.
λi λi−1
3. Obtain a new estimate of (x) and r using the relation
k = λk,
1
i = i+1 − (λi+1 − λi ), i = k − 1, . . . ,0, (6.33)
mi
1
r = (0 + K0 λ20 ).
2
Once convergence is reached, the observables can be computed using the results of
Section 6.1.3.
 7
Counting Glass States
The Complexity

In Chapters 4 and 6, we studied the properties of individual glass states. In this

chapter, we introduce the ‘complexity’, which counts the number of distinct glass
states that can be found in a given sample. This function is similar to the standard
entropy, which counts the number of typical microscopic configurations that con-
tribute to a macroscopic state. We first introduce the equilibrium complexity, which
is simply the difference between the liquid entropy and the internal entropy of
typical glass states. The complexity vanishes at the Kauzmann transition, which sets
a limit on the existence of the liquid state. At the Kauzmann point, a phase transition
towards an ideal glass phase takes place. Next, we describe how to compute the
out-of-equilibrium complexity and the properties of the ideal glass through the
Monasson method. We provide the expression of the Monasson free energy at the
replica symmetric level, and, as an example, we give results for the phase diagram
of hard spheres. Finally, we briefly discuss spontaneous replica symmetry breaking
effects in this context.

7.1 Equilibrium Complexity and the Kauzmann Temperature

Consider a simple liquid with potential v̄(h) of the class discussed in Section 2.3.2,
in the limit d → ∞. Its dynamically arrested phase, which was investigated in
Chapter 4, is found in a regime with finite temperature T = 1/β and scaled packing
fraction )
ϕ = 2d ϕ/d. As discussed in Chapter 3, when T < Td () ϕ ), with Td ()
ϕ ) given
by Eq. (3.66), the liquid dynamics starting from an equilibrium configuration is
fully arrested, and the system is confined to a glass state. In this section, we intro-
duce the ‘equilibrium complexity’, which counts the number of such glass states.

7.1.1 Equilibrium Liquid and Glass Free Energy

Before proceeding, it is useful to recall some of our previous results. Recall that
ϕ = 2d ϕ/d = ρd d /d 2 , where ρ is the number density, d is the solid angle and
)

199
200 Counting Glass States: The Complexity

 is the typical interaction scale of the potential. At leading order in d → ∞, with

fixed )ϕ , we then have
 
)
ϕ d2 d 2πe
ρ =
d
⇒ 1 − log(ρ ) ∼ log d ∼ log
d
, (7.1)
d 2 d
with corrections growing slower than d. In Chapter 2, it was shown that the liquid
free energy per particle is given by Eq. (2.76) – which, using Eq. (7.1), becomes1
  
d 2πe ϕ ∞
d)
−βfliq ()
ϕ,T ) ∼ log + dh eh [e−β v̄(h) − 1]. (7.2)
2 d 2 −∞
In Chapter 4, we computed the free energy of a typical glass state prepared in
ϕg,Tg ) and followed to a different state point ()
equilibrium at () ϕ,T ) by using the
replica method within the Franz–Parisi construction. It was further shown in Chap-
ter 6 that when () ϕg,Tg ) = ()
ϕ,T ), the replica symmetric calculation is correct.
Specialising the replica symmetric result of Section 4.3.3 given in Eq. (4.75) to the
ϕg,Tg ) = ()
case () ϕ,T ), with r = , we obtain for the typical glass free energy2
  
d πe d d)ϕ ∞
−βfg ()
ϕ,T ) = log + + dh eh q(,β;h) log q(,β;h),
2 d2 2 2 −∞
(7.3)
where  is the solution of Eq. (4.80). It was further shown in Chapter 4 (see
Figures 4.4 and 4.5) that the equilibrium liquid pressure and energy coincide with
those of the glass,3 as given by Eqs. (4.77) and (4.79), respectively, computed at
ϕg,Tg ) = ()
() ϕ,T ). Note that, as discussed in Chapter 4, in equilibrium, the liquid
and glass phases are not distinct, coexisting phases. The glass phase corresponds to
the exploration of a restricted portion of phase space, in which the system is trapped
for an infinite time in the limit N → ∞ and d → ∞. The liquid phase corresponds
to the collection of all glasses, which is the total accessible phase space at this state
point and is explored on time scales that diverge exponentially in N or d.

7.1.2 Equilibrium Complexity

Because the energies of the liquid and glass phases are identical in equilibrium,
their difference in free energy, f = e − T s, must originate from an entropy

1 In this chapter, we omit the irrelevant kinetic free energy term dT log(/) from the ideal gas contribution.
Because the most important terms of the free energy are proportional to d, we also neglect all terms that
diverge slower than d.
2 In this chapter, we drop the suffixes ‘g’ from ()
ϕg ,Tg ) and ‘r’ from r to lighten the notation. Because we do
not follow glasses at different state points, this does not create any ambiguity.
3 Note that the glass pressure and energy are not obtained by differentiating Eq. (7.3). One should instead
differentiate Eq. (4.75) with respect to ) ϕg ,Tg ) and compute the result in
ϕ or T , respectively, at fixed ()
ϕg ,Tg ) = ()
() ϕ,T ).
 7.1 Equilibrium Complexity and the Kauzmann Temperature 201
50

d = 50

40

30
Σeq

20

10

0
4 ϕd 5 6 7 ϕK 8

Figure 7.1 Equilibrium complexity for hard spheres given by Eq. (7.4), as a
function of packing fraction ) ϕ . Because different terms in the complexity have
different scaling with d, the curve is plotted for d = 50, neglecting all terms that
diverge slower than d. The complexity jumps to a non-zero value at the dynamical
transition )
ϕd and then decreases continuously until it vanishes at the Kauzmann
transition )
ϕK .

difference4 [80, 119, 120, 201]. We thus define an equilibrium ‘configurational

entropy’, or ‘complexity’, as
eq ()
ϕ,T ) = −β[fliq ()
ϕ,T ) − fg ()
ϕ,T )] = sliq ()
ϕ,T ) − sg ()
ϕ,T )
 
d d d 
= log d − − log (7.4)
2 2 2 2
 ∞
d)
ϕ
+ dh eh [e−β v̄(h) − 1 − q(,β;h) log q(,β;h)].
2 −∞
ϕ,T ) are kept finite (so that  is also finite) and d → ∞, the first term
If ()
dominates. Hence, eq () ϕ,T ) is always positive and diverges as d2 log d in this
regime. In other words, the liquid entropy,
ϕ,T ) = sg ()
sliq () ϕ,T ) + eq ()
ϕ,T ), (7.5)
is the sum of the entropy of a typical glass and of a positive contribution, eq ()
ϕ,T ),
that captures the multiplicity of possible glasses selected by distinct equilibrium

4 This can also be checked by explicitly computing the liquid and glass entropies.
202 Counting Glass States: The Complexity

liquid configurations, as discussed in Chapter 4. The equilibrium complexity thus

counts the number Neq () ϕ,T ) of distinct glass states that compose the equilibrium
liquid, or, more precisely,

1
eq ()
ϕ,T ) = lim log Neq ()
ϕ,T ). (7.6)
N→∞ N

Note that when the liquid is not dynamically arrested, for T > Td () ϕ ), glass states
do not exist in equilibrium because there is no solution for the parameter . In
this case, the complexity is not defined by Eq. (7.4) but vanishes because the liquid
ϕ ), the complexity thus jumps from zero to a
is the only equilibrium state. At Td ()
positive value. Upon decreasing temperature or increasing density, the complexity
decreases, as illustrated in Figure 7.1. The density (or temperature) at which the
complexity vanishes is called the Kauzmann density (or temperature).

7.1.3 The Kauzmann Transition

The extrapolation of the complexity to densities above or temperatures below the
Kauzmann point is formally negative. As noted already by Kauzmann [201], this
result is physically inconsistent. According to Eq. (7.6), the complexity is the log-
arithm of the number of glass states, which is certainly Neq () ϕ,T ) ≥ 1, and hence,
ϕ,T ) ≥ 0. A formally negative complexity implies that the number of glass
eq ()
states is exponentially small in N – i.e., there are no glass states at all. Because in
this region the liquid state is a superposition of glass states, the liquid phase thus
ceases to exist at the Kauzmann point. A phase transition to another phase must
then happen. Before investigating this transition in more detail (Section 7.2), it is
illustrative to compute the Kauzmann point for some model potentials.
Because of the stronger divergence of the d2 log d term, the complexity remains
positive for any finite )
ϕ , as d → ∞. The density ) ϕ should thus itself diverge when
d → ∞ for the complexity to vanish. According to Eq. (4.80),  decreases with
increasing ) ϕ . When  → 0, one can show that

q(,β;h) ∼ e−β v̄(h) [1 +  q1 (β;h) + O(2 )]. (7.7)

ϕ −1 =  A(β) + O(2 ),
Plugging this expansion in Eq. (4.80), we conclude that )
where A(β) is some constant. We want to show that ) ϕ ∝ log d and  ∝ 1/ log d.
Under this assumption, plugging the leading order  = A(β)/) ϕ and Eq. (7.7) in
Eq. (7.4), we obtain at leading order
 ∞
d d)
ϕ
ϕ,T ) ≈ log d +
eq () dh eh [e−β v̄(h) (1 + β v̄(h)) − 1]. (7.8)
2 2 −∞
 7.1 Equilibrium Complexity and the Kauzmann Temperature 203
8
Soft spheres
WCA

6
TK

0
0.0 0.5 1.0 1.5 2.0 2.5
ϕ/ log d

Figure 7.2 The Kauzmann temperature TK plotted as a function of scaled density

ϕ / log d for the soft-sphere potential v̄(h) = εe−4h and the WCA potential
)
v̄(h) = ε(1 + e−4h − 2e−2h )θ (−h). In both cases, the unit of energy is set to
ε = 1. For the soft-sphere potential, TK = 0.713 ()
ϕ / log d)4 because T and )
ϕ are
not independent. For the WCA potential, the Kauzmann temperature vanishes at
)
ϕ / log d = 1, which corresponds to the hard-sphere Kauzmann density.

From this result, it is straightforward to obtain an expression for the Kauzmann

density by the condition eq () ϕ,T ) = 0,

 ∞
log d
)
ϕK = , gK (β) = − dh eh [e−β v̄(h) (1 + β v̄(h)) − 1]. (7.9)
gK (β) −∞

Note that for the hard-sphere potential, gK (β) = 1, and, thus, ) ϕK = log d. The
Kauzmann transition line in the () ϕ,T ) plane for two representative potentials is
given in Figure 7.2.
It is important to stress that the Kauzmann transition is not always present in
mean field models. Consider, for instance, the class of finite range potentials that
vanish outside their interaction distance ; hence, v̄(h) = 0 for h > 0. These
potentials reduce to the hard-sphere potential when T → 0, and in that case, the
complexity is positive for )ϕd = 4.8067 . . . < ) ϕ <)ϕK = log d. Therefore, for all
these potentials the Kauzmann transition line must vanish at ) ϕ = log d, and if one
studies them as a function of T for 4.8067 . . . < )
ϕ < log d, the complexity remains
strictly positive down to zero temperature.
204 Counting Glass States: The Complexity

7.2 Out-of-Equilibrium Complexity: The Monasson Method

In this section, we discuss how to obtain a better understanding of the complexity
function and describe a concrete method to compute it, from the decomposition of
the partition function into glass states discussed in Chapters 4 and 5. We consider
not only the equilibrium states that dominate the Gibbs–Boltzmann measure but
also other states that dominate the out-of-equilibrium behaviour, using a method
due to Monasson [260].

7.2.1 The Liquid Partition Function as a Sum over Glass States

We begin by discussing again the equilibrium liquid at a state point (ϕ,T ). Each
typical equilibrium liquid configuration Y falls into a glass state that can be defined
by the Franz–Parisi construction of Chapter 4. According to the discussion of
Section 7.1.2, there is an exponential number of such states. The partition function
of each state, labelled by an index α, can then be written as

Zα = e−βNfα . (7.10)

As was shown in Chapter 4, distinct states are essentially disjoint. The probability
that a configuration Y cannot be unambiguously attributed to a single state is expo-
nentially small in N . Each configuration Y thus defines a state, with a typical free
energy, fα = fg (ϕ,T ), as computed in Chapter 4.
At a given state point (ϕ,T ), there are, however, also atypical configurations
Y that belong to atypical states, which have fα = fg (ϕ,T ). By analogy with the
typical glass states, we assume that the total number of glass states with free
energy f, for a given state point (ϕ,T ), is exponentially large in N in an interval
[fmin (ϕ,T ),fth (ϕ,T )]:

N (f;ϕ,T ) = δ(f − fα ) ∼ eN (f;ϕ,T ), f ∈ [fmin,fth ]. (7.11)
α

Note that the interval [fmin,fth ] is defined by the condition that the complexity
(f;ϕ,T ) ≥ 0. If (f;ϕ,T ) < 0, then the number of states is, on average, expo-
nentially small in N , which means that, with probability 1 for N → ∞, there are
no states. An illustration of the complexity (f;ϕ,T ) for hard spheres is given in
Figure 7.5. In most cases, the complexity increases with increasing f. States with
high energy, low entropy (small size) or both are more numerous than states with
low energy and high entropy.
Under these assumptions, we can write the total contribution of glass states to
the equilibrium (configurational) partition function as
 7.2 Out-of-Equilibrium Complexity: The Monasson Method 205
  
Zeq = e−βNfeq (ϕ,T ) ∼ e−βNfα = df δ(f − fα )e−βNf
α α
  (7.12)
fth (ϕ,T )
−βNf
= dfN (f;ϕ,T )e = df e N[ (f;ϕ,T )−βf]
.
fmin (ϕ,T )

For N → ∞, evaluating the last integral by the maximum of the integrand follow-
ing Laplace’s method gives

−βfeq (ϕ,T ) = max { (f;ϕ,T ) − βf}

f∈[fmin (ϕ,T ),fth (ϕ,T )]
(7.13)
= (fg (ϕ,T );ϕ,T ) − βfg (ϕ,T ),
where fg ∈ [fmin,fth ] is such that f−T (f;ϕ,T ) is minimal – i.e., it is the solution of
d 1
= , (7.14)
df T
provided that it belongs to the interval [fmin,fth ]. Eq. (7.13) shows that the equi-
librium value of the free energy fg depends on a trade-off between lowering the
individual free energy fg of the glass, which comes at the price of having fewer
states (lower complexity), and having more states that contribute (higher complex-
ity) at the price of increasing their individual free energy. This trade-off is very
similar to what happens in a standard thermodynamic equilibrium, except that the
energy is here replaced by the free energy of individual states. Note that while
Eq. (7.13) looks like a standard Legendre transform of the glass free energy with
parameter β, the complexity (f;ϕ,T ) itself depends explicitly on temperature,
which complicates the inversion of the Legendre transform. The inversion would
be greatly simplified if the Legendre parameter could be varied independently of
the temperature that appears in the complexity. This is precisely the aim of the
Monasson construction, as we discuss in Section 7.2.2.
For consistency, the maximum in Eq. (7.13) must be identified with the typical
value of the free energy of glass states fg (ϕ,T ) computed via the Franz–Parisi
method, given by Eq. (7.3). Therefore, the equilibrium complexity that we com-
puted explicitly in d → ∞ in Section 7.1.2 should be identified with

eq (ϕ,T ) = (fg (ϕ,T );ϕ,T ). (7.15)

We have seen, however, that eq (ϕ,T ) is only well defined for ϕd < ϕ < ϕK ,
for hard spheres and more generally for Td (ϕ) > T > TK (ϕ). In the rest of this
section, we will show that, generally speaking, in Eq. (7.13), one can encounter
three distinct situations, depending on the specific state point (ϕ,T ) [65, 79, 217,
250, 253].
206 Counting Glass States: The Complexity

• At high temperatures or low densities – or, more precisely, for T > Td (ϕ) – either
there are no states at all and thus (f;ϕ,T ) is always formally negative, or there is
a regime within which (f;ϕ,T ) > 0 and the solution of the maximum condition
in Eq. (7.13) is formally given by f = fth . In the former case, feq is not defined,
while in the latter case, it is usually found that feq = fth − T (fth ;ϕ,T ) > fliq ,
where fliq is the liquid free energy computed in Chapter 2 and given by Eq. (7.2)
in d → ∞. Therefore, even if glass states exist, they are subdominant in the
equilibrium measure, which is given by a single pure state – i.e., the liquid.
• At intermediate temperatures or densities – or, more precisely, for Td (ϕ) ≥ T
≥ TK (ϕ) – a value fg ∈ [fmin,fth ] can be found, which is a solution of Eq. (7.14)
such that
fliq (ϕ,T ) = feq (ϕ,T ) = fg − T (fg ;ϕ,T ). (7.16)
In other words, the total free energy of all glass states coincides with the analytic
continuation of the free energy of the liquid state below Td . We already know from
the study of the dynamics in Chapter 3 that in this regime the liquid dynamics
is fully arrested. The interpretation is therefore that the liquid state has become
a superposition of an exponential number of glass states of higher individual
free energy density fg . The Gibbs–Boltzmann measure is split on this exponen-
tial number of contributions. Yet no equilibrium phase transition happens at Td
because Eq. (7.16) guarantees that the free energy is analytic upon crossing Td .
Note that Eq. (7.16) is non-trivial because fliq (ϕ,T ) is the result of a simple liquid
virial calculation, while feq (ϕ,T ) follows from a complicated computation of the
complexity of glass states. It will be justified a posteriori in Section 7.2.3.
• At low temperatures or high densities – or, more precisely, for T < TK (ϕ) – the
partition function is dominated by the lowest free energy states, fg = fmin , with
(fmin ) = 0. In this case, the total glass free energy is feq (T ) = fmin −T (fmin ) =
fmin . At TK , a phase transition takes place. The free energy and its first derivatives
are continuous, but the second derivative of feq with respect to T (the specific
heat) has a jump, as will be shown in Section 7.2.3.
Note that such an ‘entropy crisis’ scenario [65, 79, 80], where the number of glass
states vanishes at a critical temperature TK , is also realised in a class of completely
solvable spin glass models, the simplest of which being the random energy model
(REM) [125].

7.2.2 Computing the Complexity Curve

In Section 7.2.1, we introduced the complexity function (f;ϕ,T ), which gives the
entropy of glass states of free energy f at a state point (ϕ,T ), and we have shown that
this function encodes a lot of information about the glass phase. For instance, the
 7.2 Out-of-Equilibrium Complexity: The Monasson Method 207

typical free energy of glass states, the equilibrium complexity and both the dynam-
ical and Kauzmann temperatures can be derived from the complexity. However, we
did not yet provide a method to compute this quantity.
In order to do so, Monasson suggested [260] to consider m identical replicas
of the original system, which (1) are constrained to be in the same glass state and
(2) are uncorrelated within this glass state. We will discuss in Section 7.3 how to
implement these two requirements in practice. For now, assuming that they can be
implemented, we observe that the free energy of m copies inside a single glass state
is just m times fα , because these copies independently sample the state. Then, at
low enough temperatures, the partition function of the replicated system is the sum
over all states of the contribution of each state,
  fth
∗ ∗
Zm ∼ e−βN mfα = df eN[ (f;ϕ,T )−βmf] ∼ eN[ (f ;ϕ,T )−βmf ], (7.17)
α fmin

where now f∗ (m;ϕ,T ) is such that mf − T (f;ϕ,T ) is minimal. It thus satisfies the
equation
d m
= . (7.18)
df T
The introduction of the m identical replicas thus adds a weight m to the term −βf
in Eq. (7.17). This weight can then be tuned, in order to extract the complexity
function from a Legendre transform of the replicated free energy. Defining
1
−β!(m;ϕ,T ) = log Zm = max { (f;ϕ,T ) − βmf}
N f∈[fmin (ϕ,T ),fth (ϕ,T )] (7.19)
∗ ∗
= (f (m;ϕ,T );ϕ,T ) − βmf (m;ϕ,T ),
one indeed has [253, 260]
∂ !(m;ϕ,T )
f∗ (m;ϕ,T ) = ,
∂m
(7.20)
∗ [m−1 β!(m;ϕ,T )]
2∂
(m;ϕ,T ) = (f (m;ϕ,T );ϕ,T ) = m .
∂m
The function (f;ϕ,T ), for a given state point (ϕ,T ), can then be reconstructed
from the parametric plot of (m;ϕ,T ) versus f∗ (m;ϕ,T ) varying m.
Before proceeding, let us recapitulate the two main assumptions of the Monasson
method:
1. Phase space can be partitioned in glass states labelled by α, with free energies fα .
2. m replicas can be confined within the same glass state in such a way that their
free energy is the Legendre transform of the complexity function, as expressed
by Eqs. (7.19) and (7.20).
208 Counting Glass States: The Complexity

The method thus gives the complexity (f;ϕ,T ) at any given state point, provided
we are able to compute the free energy of m copies of the original system, con-
strained to be in the same glass state, and to perform the analytical continuation to
real m, in such a way that we can take the derivatives in Eq. (7.20). This method
was developed and tested in the context of spin glasses, in which a direct compar-
ison with other methods to compute the complexity is possible, e.g., via the TAP
equations [79, 250]. Such a comparison is unfortunately not possible in the context
of particle systems, because there is no straightforward way of individually defining
the atypical glass states and counting them.5 The Monasson method is therefore the
only method thus far available to compute the complexity in particle systems.

7.2.3 Properties of the Replicated Free Energy

To conclude the presentation of the Monasson method, we now discuss a few
important properties of the replicated free energy !(m;ϕ,T ) defined in Eq. (7.19).

The Equilibrium Line

The equilibrium line corresponds to m = 1, on which the partition function Zm
defined in Eq. (7.17) reduces to the equilibrium partition function in Eq. (7.12). We
can therefore identify
f∗ (1;ϕ,T ) = fg (ϕ,T ),
(7.21)
(1;ϕ,T ) = eq (ϕ,T ),

and, therefore,
!(1;ϕ,T ) = fg (ϕ,T ) − T eq (ϕ,T ). (7.22)
On the other hand, because in the Monasson construction !(m;ϕ,T ) is the free
energy of a system of m constrained replicas, for m = 1, it reduces to the free
energy of a single replica, which is then equal to that of the equilibrium liquid,
!(1;ϕ,T ) = fliq (ϕ,T ). (7.23)
Combining Eqs. (7.22) and (7.23) thus provides a proof of Eq. (7.16), which implies
that the free energy is analytic at the dynamical transition. Note that the Monasson
construction makes sense only if the complexity (m;ϕ,T ) is positive. A for-
mally negative complexity indicates a wrong choice of maximum in Eq. (7.19); the
correct one must lie in the region [fmin (ϕ,T ),fth (ϕ,T )], in which the complexity is
positive. In particular, for m = 1, this implies that Eq. (7.22) can only hold when
eq (ϕ,T ) ≥ 0.

5 In principle, one should use density functional theory [206, 329] to do so, but this approach is technically
difficult to implement. An exact treatment in d → ∞ has therefore not yet been found.
 7.2 Out-of-Equilibrium Complexity: The Monasson Method 209

The Kauzmann Transition

In order to lighten the notation, we do not consider here the dependence of the
complexity on both temperature and density but focus instead on temperature.
In the following, ‘lower temperature’ can, however, always be replaced by ‘higher
density’.
The complexity decreases with decreasing f. Usually,6 upon lowering temper-
ature, the value of fg (T ) decreases towards the minimum fmin (T ). One can then
expand the complexity around fmin (T ) as
1
(f;T ) = σ1 (T )(f − fmin (T )) + σ2 (T )(f − fmin (T ))2 + · · · (7.24)
2
Because fg (T ) is the solution of Eq. (7.14), it coincides with fmin (T ) at a tempera-
ture TK , such that
0
d 00 1
0 = σ1 (TK ) = . (7.25)
df f=fmin (TK ) T K

In equilibrium (m = 1) the total free energy of the system is then given by


fg (T ) − T (fg (T );T ) = fliq (T ), T > TK,
feq (T ) = (7.26)
fmin (T ), T < TK .
A perturbative calculation for T  TK , using Eq. (7.24), under the assumption that
σ1 (T ) and σ2 (T ) are analytic functions in the vicinity of the Kauzmann transition
defined by TK σ1 (TK ) = 1, shows that
1 − T σ1 (T )
fg (T ) − fmin (T ) = + O((T − TK )2 ),
T σ2 (T )
(7.27)
(1 − T σ1 (T ))2
feq (T ) − fmin (T ) = + O((T − TK )3 ).
2T σ2 (T )
Hence, the total free energy is given by fmin (T ) for T < TK , while for T > TK ,
it is given by fmin (T ) + O((T − TK )2 ). The entropy seq = −dfeq /dT is therefore
continuous at TK , but the specific heat cV = T dseq /dT has a jump:
TK2 σ1 (TK )2
cV = cV (TK+ ) − cV (TK− ) = − . (7.28)
σ2 (TK )
Note that because usually σ2 (TK ) < 0, the specific heat drops upon lowering the
temperature across TK .

6 With notable exceptions. For example, in hard spheres with an attractive short range interaction, an ‘inverse
freezing’ transition can take place [321].
210 Counting Glass States: The Complexity

The Ideal Glass Phase

The equilibrium Kauzmann transition can be extended into a whole line mK (T ) or
TK (m) in the (m,T ) plane, defined by the condition of vanishing complexity:
mK (T )
(mK (T );T ) = 0 ⇔ σ1 (T ) = , (7.29)
T
where the second condition is derived from Eq. (7.18). Because the temperature
dependence of σ1 (T ) is usually mild, the Kauzmann transition shifts to lower tem-
peratures (higher densities) for m < 1 and to higher temperatures (lower densities)
for m > 1. See Figure 7.3 for an illustration.
For m < mK (T ) – or equivalently T > TK (m) – the system is in the ‘high-
temperature’ phase, where f∗ > fmin and > 0. By contrast, for m > mK (T ) or
T < TK (m), it is in the ‘low-temperature’ phase, where f∗ = fmin and = 0. Note
that, because a phase transition happens on the line mK (T ), calculations based
on a high-temperature (low-density) expansion such as those of Chapters 2 and
4 only give access to the free energy !(m;T ) in the high temperature region m
≤ mK (T ) [250, 253, 260]. In the low-temperature phase, because one has
!(m;T ) = mfmin (T ), the free energy can instead be reconstructed from the
value of fmin (T ), obtained using the relation
!(mK (T );T )
fmin (T ) = = f∗ (mK (T );T ). (7.30)
mK (T )
This equation is evaluated on the transition line, hence one can use the high-
temperature expression of !(m;T ). We will better illustrate this point with a
concrete example in Section 7.3. The low-temperature phase at T < TK , where the
complexity is identically zero and the free energy is given by Eq. (7.30), is called
the ‘ideal glass’ or ‘condensed’ phase.

7.3 The Monasson Construction in Infinite Dimensions

In this section, we discuss the practical implementation of the Monasson construc-
tion in d → ∞, and we provide results for the free energy and for thermodynamic
observables. To compute the free energy of m identical replicas of the original
system, we use the methods of Chapter 4 – in particular, those of Section 4.2.4.
Because these results are based on the truncation of the low-density virial expan-
sion, they only hold in the equilibrium liquid phase of the system of m replicas for
T ≥ TK (m), where the complexity is positive, as discussed in Section 7.2.3.

7.3.1 Free Energy

ˆ
Eq. (4.56) provides the free energy of a replicated liquid in terms of the matrix 
of mean square displacements between replicas. This result can be used to compute
 7.3 The Monasson Construction in Infinite Dimensions 211

the complexity by straightforwardly renaming the number of replicas n → m and

using again Eq. (7.1) to keep only the leading order in large d of the ideal gas term.
One then has
   
d 2πe d (m − 1) 2πe
−β!(m;) ˆ
ϕ,T , ) = log + log
2 d 2 d2
d d)
ϕ
+ ˆ
log[2 det(−/2)(−1 T ˆ −1
 1)] − F(), ˆ (7.31)
2 2
 ∞ 5   m 6
∂2
ˆ =− −1 m
dh eh e 2 a,b=1 ab ∂ha ∂hb e− a=1 β v̄(ha ) − 1

F() ,
−∞ ha =h

where 1 = {1, . . . ,1}. Only terms proportional to d log(d) and to d have been kept;
subleading terms have been discarded. As discussed in Section 7.2.2, the replicas
must satisfy two requirements:
1. They must be in the same glass state; hence, ab should remain finite for all
pairs ab.
2. They should be independently equilibrated within that glass state; hence,
ab should be given by the unconstrained mean square displacement of the
glass state.
Therefore, ˆ should be set to a local extremum7 of the free energy in Eq. (7.31),
with finite matrix elements [250, 253, 260, 292], similarly to the Franz–Parisi
construction discussed in Chapter 4. Because m should be continued to real values,
we need to make an ansatz for the matrix  ˆ that allows for such an analytic contin-
uation. As for the Franz–Parisi construction, we assume that the correct ansatz has
the hierarchical structure introduced in Chapter 5.

7.3.2 Replica Symmetric Ansatz

We now specialise the calculation to the replica symmetric ansatz. Because all the
replicas are equivalent, and aa = 0, we have ab = (1 − δab ). The eigenvectors
of ˆ are the vector 1, with eigenvalue λ1 = (m − 1), and any vector orthogonal
to 1, with (m − 1) degenerate eigenvalues λ2 = −. Therefore, we have
. /  
ˆ (1 − m)  m−1 m
det − = , ˆ −1 1 =
− 1T  . (7.32)
2 2 2 (1 − m)

ˆ
Applying the same steps as in Sections 4.3.2 and 4.3.3 to the calculation of F(),
we obtain

7 A local minimum for m > 1, which is analytically continued to a local maximum for m < 1.
212 Counting Glass States: The Complexity
 ∞ 5 6
d2
ˆ =− −
F() dh e h
e 2 dh2 gRS (,β;h) − 1
m
−∞
 (7.33)
∞ @ A
=− dh e h
q(,β;h)m − 1 ,
−∞

where the function gRS (,β;h) is defined in Eq. (4.70) and q(,β;h) is defined
in Eq. (4.74). Plugging these results in Eq. (7.31), we obtain the replica symmetric
expression of the Monasson replicated free energy,
   
d 2πe d (m − 1) πe
−β!(m;) ϕ,T ,) = log + log
2 d 2 d2
 (7.34)
d ϕ ∞
d) @ A
+ log m + dh e q(,β;h) − 1 .
h m
2 2 −∞
The equation for  is obtained by imposing a vanishing first derivative of
!(m;)ϕ,T ,) with respect to . The result is very similar to Eq. (4.80):
 ∞
1 m ∂q(,β;h)
= Fm (;β) = − dh eh q(,β;h)m−1 . (7.35)
)
ϕ m − 1 −∞ ∂
Taking derivatives with respect to m according to Eq. (7.20) and keeping in mind
that the derivative with respect to  should not be taken because  is set variation-
ally, we obtain
 
d πe d
−βf∗ (m;) ϕ,T ) = log +
2 d2 2m

ϕ ∞
d)
+ dh eh q(,β;h)m log q(,β;h),
2 −∞
 
d d d 
(m;)ϕ,T ) = log d − − log (7.36)
2 2 2 2m

ϕ ∞
d)
+ dh eh [q(,β;h)m − 1 − mq(,β;h)m log q(,β;h)].
2 −∞
One can now check explicitly that when m → 1, which corresponds to equilibrium
sampling of the glass states, the following properties hold:

• The replicated free energy Eq. (7.34) reduces to the liquid free energy Eq. (7.2) –
ϕ,T ) = fliq ()
i.e., !(1;) ϕ,T ).
• The equation for , Eq. (7.35), reduces to the equilibrium equation for  of the
Franz–Parisi construction, Eq. (4.80), which itself coincides with the equation for
the long time limit of the mean square displacement obtained from the equilib-
rium dynamical equations in the arrested phase in Section 3.3.2.
 7.3 The Monasson Construction in Infinite Dimensions 213

• The glass free energy in Eq. (7.36) coincides with that obtained from the Franz–
Parisi construction given by Eq. (7.3) – i.e., f∗ (1;)
ϕ,T ) = fg ()
ϕ,T ).
• The complexity in Eq. (7.36) coincides with the equilibrium complexity given in
Eq. (7.4) – i.e., (1;)ϕ,T ) = eq ()ϕ,T ).
Some of these results have already been discussed in the general case in
Section 7.2.3. They show that the Monasson method is fully consistent with the
Franz–Parisi method and with equilibrium dynamics. In particular, in equilibrium,
a finite solution for  of Eq. (7.35) appears at the dynamical transition density
ϕd (β). For lower densities or higher temperatures, the only solution is  = ∞,
)
which implies that replicas cannot be confined in a same glass state – i.e., there are
no such states. For higher densities or lower temperatures, the Monasson method
predicts the existence of glass states with free energy f∗ (1;) ϕ,T ) = fg () ϕ,T ),
consistently with the Franz–Parisi method, and complexity (1;) ϕ,T ) = eq () ϕ,T ).
Furthermore, because

ϕ,T ) = f∗ (1;)
!(1;) ϕ,T ) − T ϕ,T ) = fliq ()
(1;) ϕ,T ), (7.37)

the total free energy is predicted by the Monasson method to be analytic at the
dynamical transition )ϕd (β), as discussed in Section 7.2.3.
The results in Eqs. (7.34) and (7.36) hold only when the complexity (m;) ϕ,T )
is positive. In the equilibrium case of m = 1, as discussed in Section 7.1.3, the
complexity remains positive up to ) ϕK ∼ log d. The case of hard spheres for d = 50,
for which )ϕK ∼ 8, is illustrated in Figure 7.1.

7.3.3 Thermodynamic Observables

For fixed m and () ϕ,T ), the partition function is dominated by a set of glass states
that have an identical free energy f∗ (m;) ϕ,T ) as well as identical values of all
intensive observables – e.g., internal entropy, pressure or energy. Therefore, to
each state point (m;) ϕ,T ), one can equivalently associate a state point using other
thermodynamic variables. The pressure, internal entropy and energy cannot be com-
puted by taking derivatives of f∗ (m;) ϕ,T ), because this free energy corresponds to
different glass states at each state point. In order to bypass this difficulty, we express
these observables in terms of correlation functions, as discussed in Section 2.1.2.
For instance, the radial distribution function g(r|m;) ϕ,T ) of the glass states selected
at the state point (m;)ϕ,T ) gives the average potential energy per particle, according
to Eq. (2.28),

ρd ∞
e(m;)ϕ,T ) = dr r d−1 v(r)g(r|m;) ϕ,T )
2 0
 (7.38)
ϕ ∞
d)
= h
dh e v̄(h) ḡ(h|m;)ϕ,T ),
2 −∞
214 Counting Glass States: The Complexity

where in the second equality we changed variables to h = d(r/−1) and introduced

ḡ(h|m;) ϕ,T ) = g(r|m;) ϕ,T ). The pressure can be similarly written through the
virial theorem [175, section 2.5],

βρVd ∞
p =1− dr r d v  (r)g(r|m;)
ϕ,T )
2 0
 (7.39)
βd)ϕ ∞ h 
=1− dh e v̄ (h)ḡ(h|m;) ϕ,T ),
2 −∞
which, for hard spheres, gives [175, section 2.5]

2d ϕ d)
ϕ
p =1+ g(+ |m;)
ϕ) = 1 + ḡ(0+ |m;)
ϕ ). (7.40)
2 2
The problem is then reduced to that of computing the radial distribution function.
In the liquid phase, Eq. (2.29) expresses g(r) as a functional derivative of the liquid
free energy with respect to the pair potential. This result can be generalised to a
system of m replicas. The radial distribution function of any of the replica can be
written as the functional derivative of the replicated free energy with respect to the
pair potential of that replica. This amounts to replacing, either in Eq. (7.31) or in
its replica symmetric version (7.34), the pair potential v̄(ha ) by a replica-dependent
potential v̄a (ha ), taking the derivative with respect to one of these potentials, and
then setting v̄a (ha ) = v̄(ha ). Because all replicas are equivalent, we can also use a
functional analog of Eq. (4.71) to express the average of the partial derivatives with
respect to v̄a (ha ) in terms of the total derivative with respect to v̄(h). Changing
variables from r to h, Eq. (2.29) generalises to
∂!(m;) ϕ,T ) d)
ϕ
= m eh ḡ(h|m;)
ϕ,T ). (7.41)
∂ v̄(h) 2
Using the replica symmetric expression (7.34) for !(m;) ϕ,T ), we then obtain
 ∞
ḡ(h|m;)ϕ,T ) = e−β v̄(h) dz ez−h q(,β;z)m−1 γ (z − h + /2). (7.42)
−∞

Plugging this result in Eqs. (7.38) and (7.39) provides the energy and pressure of
the glass states as a function of (m;)
ϕ,T ).

7.4 The Phase Diagram of Hard Spheres

We now turn to the study of the phase diagram, specialising to the case of hard
spheres, for which the function q(;h) does not depend on β and is given in
Eq. (4.83).
 7.4 The Phase Diagram of Hard Spheres 215
10
d = 50 md
mK

6
Condensed (Σ = 0)
m

Σ>0
2 Liquid
d
ϕ K
ϕ
ϕth GCP
ϕ
0
4 6 8 10

Figure 7.3 Replica symmetric phase diagram of hard spheres, in the plane (m,) ϕ ),
for d = 50. The long-dashed line is the equilibrium line m = 1. The solid line is
the dynamical line md ()
ϕ ), above which a finite solution for  exists and replicas
can be constrained to the same state. The short-dashed line is the Kauzmann line
mK ()ϕ ) above which the m replicas are stuck into the lowest free energy state.
Because on this scale the Kauzmann line moves to infinite m for d → ∞, it is
plotted here for d = 50, but the other lines are independent of d.

7.4.1 Phase Diagram in the (m,)

ϕ ) Plane
To compute the complexity as given by Eq. (7.36), one should first solve Eq. (7.35)
to obtain . A solution exists only for m > md () ϕ ), as given in Figure 7.3. In the
region where a solution for  exists, the complexity (m;) ϕ ) can be computed
numerically, and one can obtain the Kauzmann line defined by (mK () ϕ ) = 0.
ϕ );)
This line is also given in Figure 7.3 for d = 50. The two lines mK () ϕ ) and md ()ϕ)
intersect, defining three distinct regions in Figure 7.3:

• A region where no solution for  exists and, hence, replicas cannot be kept in a
same state. Here the replicated system is liquid, and the complexity is not defined.
• A region where > 0, and glass states thus exist. In this region, the low-density
results are correct, and one can use Eqs. (7.36) to compute parametrically the
curve (f;) ϕ ). Examples for some )ϕ are given in Figure 7.5.
• A region where, formally, < 0 if one uses Eq. (7.34) but, in reality, the
system is condensed in the lowest free energy state, with = 0. In this
regime, Eqs. (7.34) and (7.36) are invalid because the low-density expansion
216 Counting Glass States: The Complexity

has broken down at the Kauzmann transition. The free energy is instead given by
!(m;) ϕ ) = m fmin ()
ϕ ). One can compute fmin ()ϕ ) by using the continuity of the
free energy on the Kauzmann line:
 
!(m;)ϕ) !(mK ()ϕ );)
ϕ)
ϕ) =
fmin () = , (7.43)
m ∀m≥mK ()
ϕ) mK ()
ϕ)
as discussed in Section 7.2.3. On the Kauzmann line, one can use Eqs. (7.34)
and (7.36).
On the equilibrium line, which corresponds to m = 1 in Figure 7.3, these three
regions correspond to the regime )
ϕ<) ϕd (liquid phase), )
ϕd < )
ϕ<)
ϕK (dynamically
arrested liquid phase) and )ϕ > )ϕK (thermodynamically stable ideal glass phase),
respectively.

7.4.2 Phase Diagram in the () ϕ ) Plane

p,)
The parameter m can be used to bias the Gibbs–Boltzmann measure towards
atypical states. Its physical meaning is, however, not fully transparent. One
can interpret the quantity T /m as an ‘effective temperature’, conjugated to the
internal free energy of the states, exactly as temperature is conjugated to energy
[65, 109, 110]. The study of the off-equilibrium dynamics has indeed shown that in
some regimes, this effective temperature controls the equilibration of the slowest
degrees of freedom in the system [109]. Alternatively, one can convert m into
another more directly measurable physical quantity. As an example, we focus on
reduced pressure. To each state point (m,) ϕ ), we can associate a scaled pressure
)(m,)
p ϕ ) = p(m,)ϕ )/d. Within the replica symmetric ansatz, substituting Eq. (7.42)
into Eq. (7.40), we obtain at leading order

ϕ ∞
)
)(m,)
p ϕ) = dh eh q(,β;h)m−1 γ (h + /2). (7.44)
2 −∞
This result holds, as for Eq. (7.34), in the region m < mK () ϕ ). Using Eq. (7.44), we
can compute the pressure in each point (m,) ϕ ), and then replace m by p ) as the state
variable.
The resulting phase diagram in the plane () p,)ϕ ) is given in Figure 7.4. Note that
for m = 1 and ) ϕ≤) ϕK , we get p)(1,)ϕ) = )ϕ /2 = p )liq ()
ϕ ), which coincides with the
equilibrium liquid equation of state (EOS), obtained in Chapter 2. This confirms
that m = 1 corresponds to the equilibrium case. The dynamical and Kauzmann
lines have similar shapes in the () p,)ϕ ) plane. In this case, they also delimit three
regions, but two of them are inaccessible. More precisely:

• In the region = 0 of the (m,) ϕ ) plane, the m replicas are independent, and  is
formally infinite. In this situation, each replica is an independent liquid, and they
 7.4 The Phase Diagram of Hard Spheres 217

1.0
d = 50 pd
pK
0.8
Liquid EOS

0.6
No states
1/
p

0.4
Σ>0
d
ϕ
K
ϕ
0.2 No states

th
ϕ GCP
ϕ
0.0
3 4 5 6 7 8 9 10
ϕ Jamming line

Figure 7.4 Replica symmetric phase diagram of hard spheres, in the plane
(1/)p,)ϕ ), for d = 50. The long-dashed line is the equilibrium line p )liq ()
ϕ) = )
ϕ /2.
)d ()
The full line is the dynamical line p ϕ ), above which a finite solution for  exists
and replicas can be constrained to be in the same state. The short-dashed line is the
Kauzmann line p )K ()
ϕ ) corresponding to the lowest free energy states. Above this
line there are no states. Because on this scale the Kauzmann line moves to infinite
1/)p for d → ∞, it is plotted here for d = 50, but the other lines are independent
of d.

all have the same pressure p )liq ()

ϕ ). Hence, in the () ϕ ) plane, this whole region
p,)
collapses onto the equilibrium liquid EOS, p ϕ) = )
)liq () ϕ /2.
• In the region > 0, there are exponentially many glass states with given density
and pressure ()p,)ϕ ). The equilibrium states are still those with p )= p )liq . The other
state points with p) = p)liq are dominated by atypical states that can be visited if
the system falls out of equilibrium.
• In the ‘condensed’ region of the (m,) ϕ ) plane, the system is stuck in the lowest
free energy state. The pressure of this state gives the Kauzmann (ideal glass) line
)K . The whole negative complexity region of the (m,)
p ϕ ) plane thus collapses onto
a single line in the () ϕ ) plane.
p,)

In summary, in equilibrium the system follows the liquid line p ϕ ) for )

)liq () ϕ <)
ϕK
and the Kauzmann (ideal glass) line p )K ()
ϕ ) for )
ϕ > ) ϕK . In all other state points
()
p,)ϕ ) for which > 0, there exists an exponential number of glass states that
are never sampled in equilibrium but can be accessed if the system falls out of
equilibrium. In the rest of the () ϕ ) plane, no glass states are present.
p,)
218 Counting Glass States: The Complexity
60
d = 50

50

40
Σ

30

20

10 =5
ϕ
=6
ϕ
=7
ϕ
0
-400 -350 -300 -250 -200 -150
s= βf

Figure 7.5 Examples of the complexity (f;) ϕ ), plotted as a function of the

entropy −βf = s (because, for hard spheres, the energy is identically zero), for
three representative values of )ϕ . Because, according to Eq. (7.36), the complexity
diverges when d → ∞, it is plotted here for d = 50. For ) ϕ = 7, the entropy s
can take values going from s = −∞ (jammed states) to a maximal value where
= 0. For )ϕ = 5 and ) ϕ = 6, the entropy can take values over a finite interval.
The dotted line has slope −1, corresponding to the equilibrium condition given by
Eq. (7.14). For all curves, the black dot denotes the equilibrium value of s and .

7.4.3 The Edwards Ensemble for Hard Spheres

A particularly interesting region of the () p,)ϕ ) plane is the ‘jamming line’ cor-
responding to p ) = ∞ (or 1/) p = 0) in Figure 7.4. On this line, hard-sphere
configurations contain a network of particles in direct contact – i.e., separated by a
distance identical to their diameter. This network of mechanical hard core contacts
is able to sustain the infinite external pressure. Such configurations are often called
‘random close packings’ [33, 34] or ‘mechanically stable configurations’ [6, 232].
From Eq. (7.44), it is possible to show that achieving infinite pressure requires
both  = 0 and m = 0. More precisely, one should take the joint limit m → 0
and  → 0 with /m = ϑ [292]. In this ‘jamming’ limit, one can show from
Eq. (7.44) that p) → ∞. The first condition,  = 0, is natural. Because pressure
is infinite and particles touch each other, no vibrations are possible. The second
condition is less trivial but has an interesting interpretation. When m = 0, the
modified partition function defined in Eq. (7.17) becomes a uniform sum over
all glass states. At infinite pressure, these states coincide with the mechanically
 7.4 The Phase Diagram of Hard Spheres 219

stable configurations. In other words, at infinite pressure, the Monasson ensemble

reduces to the so-called “Edwards ensemble” [30], developed in the context of
granular systems, which expresses the partition function as a uniform sum over
all mechanically stable hard-sphere configurations.
The equation for ϑ in the jamming limit is obtained from Eq. (7.35) and reads
 ∞ 2
1 h2 h
= F0 (ϑ) = dh e−h− 2ϑ . (7.45)
)
ϕ 0 2ϑ
As for any finite m, this equation admits a solution only when )ϕ is large enough,
)
ϕ>) ϕth = 1/[maxϑ F0 (ϑ)] = 6.25812 . . ., for ϑ = 0.60487 . . . For )
ϕ>) ϕth , one
can solve for ϑ and obtain the corresponding value of the complexity by taking the
m → 0,  = mϑ limit of Eq. (7.36):
d d d ϑ
Ed
j ()
ϕ) = log d − − log
2 2 2 2
  ∞   (7.46)
d)ϕ h2
−h− 2ϑ h2
+ −1 + dh e 1+ .
2 0 2ϑ
This expression gives the complexity of mechanically stable configurations as a
function of density along the jamming line – i.e., the ‘Edwards complexity’ [30].
The qualitative behaviour of the Edwards complexity is similar to that of the equi-
librium complexity; see Figure 7.6. It jumps discontinuously from zero to a positive
ϕth and then decreases smoothly until it vanishes linearly at a density )
value at ) ϕGCP ,
called ‘glass close packing’ (GCP). This density represents the densest possible
state that can be achieved by hard-sphere glasses. As for the Kauzmann transition
(see Section 7.1.3), the GCP density scales as ) ϕGCP ∼ log d at the leading order,
while the difference )ϕGCP − )
ϕK remains finite when d → ∞.

7.4.4 Atypical Glass States and Protocol Dependence

It is interesting to compare the phase diagram in the () ϕ ) plane obtained via state
p,)
following, Figure 6.1, with that obtained via the Monasson construction, Figure 7.4.
In both cases, glass states exist in a region delimited by the jamming line at infinite
pressure, by the Kauzmann line on the right side8 and by a ‘dynamical line’ on the
left side.9 While the qualitative shape of the two phase diagrams is very similar,
their physical interpretation is quite different. In the Monasson construction, for

8 The equivalent of the Kauzmann line was not included in Figure 6.1 because the Kauzmann transition had not
yet been introduced. In the state following construction, it corresponds to preparing an equilibrium state at )
ϕK
and following it in compression. Note that this line does not coincide with the Kauzmann line of Figure 7.4,
which instead corresponds to the densest glass state for any fixed pressure. See [214] for details.
9 In the state following construction, the dynamical line corresponds to following the states prepared at )
ϕd in
compression and to the envelope of the spinodal points of the different glass states in decompression.
220 Counting Glass States: The Complexity

d = 50 Σeq
60
ΣEd
j
ΣSF
j

40
Σ

20

K
ϕ GCP
ϕ
0
4 ϕd 6 ϕth ϕj,d 8 ϕj,K 10
ϕ

Figure 7.6 Equilibrium and jamming complexities for hard spheres, as a function
of packing fraction )ϕ . Because different terms in the complexity have different
scaling with d, the curves are plotted for d = 50, neglecting all terms that diverge
slower than d. The solid line eq ()ϕ ) is the equilibrium complexity (m = 1), as in
Ed
Figure 7.1. The dot-dashed line j () ϕ ) is the Edwards complexity of the jammed
states (m = 0), given by Eq. (7.46) under the RS approximation. It jumps to a
non-zero value at the threshold )ϕth and vanishes at the glass close packing point
)
ϕGCP . Note, however, that this whole curve is unstable towards replica symmetry
breaking. Finally, the dashed line jSF ()ϕ ) is the equilibrium complexity,12 plotted
as a function of the jamming density, as defined in Eq. (7.47). It represents
(neglecting RSB effects) the complexity of the jammed states obtained via state
following.

each point () ϕ ), one considers the ‘typical’ states for that point. These are the
p,)
states that dominate the modified partition function defined in Eq. (7.17). In the state
following construction, one instead prepares states that are typical of the equilib-
rium measure (i.e., at m = 1) and then follows them adiabatically in compression or
decompression. Obviously, the two constructions coincide on the equilibrium line
(m = 1). They thus provide the same predictions for the dynamical transition ) ϕd , the
Kauzmann transition ) ϕK and the equilibrium complexity. However, as soon as one
leaves the equilibrium line, the states that were typical in equilibrium become atyp-
ical, and the Monasson construction is dominated by a distinct set of states. This
leads to quantitative differences in the location of the lines that delimit the region of
existence of glass states. In particular, the region obtained via state following must
be strictly contained within that obtained by the Monasson construction, because
 7.4 The Phase Diagram of Hard Spheres 221

the latter considers all possible states while the former considers only a subset of
all possible states, namely those that can be obtained by adiabatically following an
equilibrium state.
This is an example of a more general property of the () ϕ ) phase diagram.
p,)
Because of the proliferation of glass states with distinct properties, different out-
of-equilibrium protocols are likely to visit different sets of states [65, 215, 292].
The state following construction corresponds to one specific protocol: prepare a
state in equilibrium, and follow it adiabatically. The Monasson construction corre-
sponds to another specific protocol: sample all glass states at given () ϕ ) according
p,)
to the modified measure. Many other protocols can be devised, and each would
typically lead to distinct results as soon as one leaves the equilibrium line.10 This
specific example reflects a general fact: properties of out-of-equilibrium systems
are strongly protocol dependent.
Another observable that depends sensitively on the protocol is the jamming com-
plexity. The complexity of the packings obtained via state following can be defined
by observing that the state following procedure maps each equilibrium density
)
ϕg to a corresponding jamming point ϕ )j ()
ϕg ) [86, 128, 292, 330]; see Figure 6.1.
Therefore, under the assumption11 that no states appear or disappear during the
compression procedure from ) ϕg to ϕ)j ()
ϕg ), one can define a jamming complexity of
state following as
SF
j ()
ϕj ) = eq [)
ϕg ()
ϕj )], (7.47)
where ) ϕj ) is the inverse function of ϕ
ϕg () )j ()
ϕg ), which maps each jamming density
onto its corresponding equilibrium density. In other words, Eq. (7.47) indicates
that in order to count the number of packings that can be constructed by state
following at density ϕ )j , one can just count how many equilibrium states at density
) ϕj ) exist, under the assumption that each of them would generate a single cor-
ϕg ()
responding jammed state. It is interesting to observe that the Edwards complexity
is distinct from the state following complexity of jammed packings, as shown12
in Figure 7.6. While the two functions are numerically close in the region where
they are both non-zero, they are clearly distinct. In addition, jEd () ϕ ) is non-zero for
)
ϕ ∈ [) ϕth,)ϕGCP ], while j () SF
ϕ ) is non zero-over a smaller interval, )
ϕ ∈ [) ϕj,K ],
ϕj,d,)

10 For example, one can combine the Monasson and state following constructions to first prepare a typical state
at a point () ϕ ) out of the equilibrium line and then follow it adiabatically in compression or decompression.
p,)
11 This assumption is correct only in the p-spin model [65], while it fails in all other known models, even in a
RS phase. It is even more likely to be incorrect in presence of a RSB phase [230]. Yet it often provides a good
approximation of the correct calculation.
12 To construct the state following complexity of jammed packings in Figure 7.6, the function ϕ
)j ()
ϕg ) has been
obtained by fitting the fullRSB numerical data discussed in Section 6.3, which gives
ϕ ϕg ) ≈ 3.324 + 0.730)
)j () ϕg + 0.0127) ϕg2 . Note that these data are not very precise in the low )
ϕg region, and
ϕg ∼ )
especially for ) ϕd , for the reasons discussed in Section 6.3.
222 Counting Glass States: The Complexity

where ) ϕj,d = ϕ
)j ()
ϕd ) and )
ϕj,K = ϕ
)j ()
ϕK ). This is another manifestation of the fact
that the packings that can be reached by compressing an equilibrium state are only
a subset of the full Edwards ensemble.

7.5 Replica Symmetry Breaking Instability in the Monasson Construction

We now discuss how the replica symmetry breaking (RSB) effects discussed in
Chapter 5 modify the replica symmetric (RS) picture that has been discussed ear-
lier. Unfortunately, the role of RSB in the Monasson construction has not been
extensively studied. Because many problems remain open, in this section, we only
briefly review the topic without entering into the full technical details.

7.5.1 Replica Symmetry Breaking Equations

For completeness, we provide here the basic equations needed to discuss RSB
effects within the Monasson construction. The procedure to derive these equations
is similar to that described in Chapter 6, so we do not repeat here the details of
that derivation. Technically, the main difference with respect to Chapter 6 is that,
in the Monasson construction, the number of replicas m is kept finite. The function
(x) is therefore defined over the interval x ∈ [m,1], while in state following, the
number of replicas goes to zero and the function (x) is defined in x ∈ [0,1].

Free Energy and Variational Equations

The expression for the free energy within the fullRSB ansatz can be derived starting
from Eq. (7.31) and then inserting the hierarchical fullRSB ansatz for , ˆ as dis-
ˆ
cussed in Chapter 5. In particular, the determinant of  is given by Eq. (5.29), and
ˆ −1 1 is given by Eq. (5.31), while the differential term can be treated
the term 1T 
as in Section 5.4.1. The details are very similar to Section 6.1. The result is then
  ' πe ( d
d 2πe d (m − 1)
−β!(m;) ϕ,T ) = log + log 2 + log m
2 d 2 d 2
 1  ∞ (7.48)
dm dx d)ϕ @ A
− log λ(x) + dh eh emf (m,h) − 1 ,
2 m x2 2 −∞
where
 1
λ(x) = x(x) + dy(y), (7.49)
x

and
f (1,h) = log γ(1)  e−β v̄(h+(m)/2),
1˙ (7.50)
f˙(x,h) = (x) f  (x,h) + xf  (x,h)2 , m < x < 1.
2
 7.5 Replica Symmetry Breaking Instability in the Monasson Construction 223

The free energy should be optimised with respect to (x), which can be done by
introducing a Lagrange multiplier P (x,h) conjugated to f (x,h) as in Section 6.1.2.
The resulting equations are

P (m,h) = )ϕ eh+mf (m,h), (7.51)

1  

˙
Ṗ (x,h) = − (x) P  (x,h) − 2x P (x,h)f  (x,h) , 0 < x < 1,
2
and
 ∞
1 1
=− dhP (x,h)[f  (x,h) + f  (x,h)]. (7.52)
λ(x) 2 −∞

Eqs. (7.49), (7.50), (7.51) and (7.52) constitute a closed set that can be solved
iteratively to obtain (x). The solution should be inserted in Eq. (7.48) to obtain
the Monasson free energy. The energy, pressure and other observables can be
obtained along similar lines.

Instability of the RS Solution

One can follow the procedure outlined in Section 5.5 to study the linear instabil-
ity of the RS solution and compute the RSB solution perturbatively around that
instability. Note that, according to the definition in Eq. (4.74), the RS solution has

f (1,h) = log q(,β;h), P (1,h) = )

ϕ eh q(,β;h)m . (7.53)

From this, one obtains expressions for the replicon mode,


ϕ 2 ∞
)
λR = 1 −  dheh q(,β;h)m f  (1,h)2, (7.54)
2 −∞

for the breaking point x ∗ ,

∞
∗
)
ϕ −∞dh eh q(,β;h)m f  (1,h)2
x = ∞ , (7.55)
4
3
+ 2)
ϕ −∞ dh eh q(,β;h)m f  (1,h)3

˙ ∗ ) at the breaking point,

and for the slope (x
∞
4
3 + 2) ϕ −∞ dh eh q(,β;h)m f  (1,h)3
˙
(x ∗
) = 12(x ∗ )2

∞ ,
4
−)
ϕ −∞ dh eh q(,β;h)m ARS (h) (7.56)
ARS (h) = f  (1,h)2 − 12x ∗ f  (1,h)f  (1,h)2 + 6(x ∗ )2 f  (1,h)4,

all evaluated on the RS solution. From these expressions, one can study RSB effects
following the same recipe as in Section 6.2.2.
224 Counting Glass States: The Complexity

d = 50 Liquid EOS
pRS
d
d
ϕ pK
0.4
pG
(ϕ† , p† )
ϕ1RSB
th

K
ϕ
1/
p

RS No states
0.2

No states

RSB GCP
ϕ
0.0
5 6 ϕRS
th 7 8 9 10
ϕ

Figure 7.7 Full replica symmetry breaking phase diagram of hard spheres, in the
plane (1/)p,)ϕ ), for d = 50 [88, 221]. For better visualisation, this plot zooms
(with respect to figure 7.4) on the region where fullRSB is observed. The long-
dashed line is the equilibrium line p )liq ()
ϕ) = ) ϕ /2. The dotted line is the RS
dynamical line p ϕ ). The portion of this line that connects )
)d () ϕd with )RS is unstable
ϕth
because of RSB. The short-dashed line is the Kauzmann line p )K ()
ϕ ). The dot-
dashed line is the Gardner line p )G ()ϕ ), where the replicon eigenvalue vanishes.
The black diamond marks the point () )† ), above which a fullRSB solution can
ϕ †, p
be constructed perturbatively around the Gardner line. The black square marks the
location of the 1RSB threshold ) 1RSB .
ϕth

7.5.2 Replica Symmetry Breaking Phase Diagram

In Figure 7.7, we report the phase diagram obtained within the fullRSB ansatz
[88, 221]. The RS solution becomes unstable on a Gardner transition line p )G ()
ϕ ),
defined by λR = 0, which originates from the equilibrium dynamical transition,
and extends all the way to the Kauzmann line, crossing it before GCP is reached.
On this Gardner line, the breaking point x ∗ is always in [0,1], but the function
(x) is defined over [m,1]. To obtain a consistent perturbative expansion around
the Gardner line, one therefore has to require x ∗ ∈ [m,1]. The values of x ∗ and
m along the Gardner line are reported in Figure 7.8. The condition x ∗ > m is
only satisfied for )
ϕ > ) ϕ † = 5.823 (which corresponds to m > m† = 0.4214
and p)> p )† = 4.838) [88]. For ) ϕ<)
ϕ † , no RSB solution thus exist perturbatively
around the Gardner line [303], and one concludes that the Gardner line is the limit of
existence of the glass states. For )
ϕ>)
ϕ † , instead, a RSB solution can be constructed
 7.5 Replica Symmetry Breaking Instability in the Monasson Construction 225
1.0
m
0.9 x∗
† , m† )
(ϕ
0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0.0
4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 9.0 9.5
ϕ

Figure 7.8 The parameter m() ϕ ) and the breaking point x ∗ ()

ϕ ) along the Gardner
transition line of Figure 7.7. The function (x) is defined, at any ) ϕ , for x ∈ [m,1];
hence, it can be constructed perturbatively only when x ∗ > m. The point where
m()ϕ ) = x ∗ ()
ϕ ) is the lower boundary where this condition is satisfied. The crossing
gives )ϕ † = 5.823, which corresponds to m† = 0.4214 and p )† = 4.838.

perturbatively around the Gardner line. Because the slope (x ˙ ∗ ) < 0 in this region,
the solution is described by a fullRSB ansatz in the whole region denoted by ‘RSB’
in Figure 7.7.
The crossing point between the Gardner transition line and the Kauzmann line
defines an equilibrium Gardner transition [160, 171]. Above this density, the equi-
librium glass phase is described by a fullRSB ansatz. Note that, as found in the state
following framework, the whole jamming line (1/) p = 0) falls into the fullRSB
region. The consequences of the fullRSB structure of jamming will be discussed
in Chapter 9. For now, we note that the fullRSB region in Figure 7.7 is delimited
by the Gardner line on the top, by the Kauzmann line on the right and by the jam-
ming line on the bottom. On the left, there must exist a ‘dynamical fullRSB line’,
where the fullRSB solution becomes unstable and disappears, similarly to what
happens to the RS solution on the dynamical RS line. Unfortunately, the dynamical
fullRSB line has only been systematically studied for spin glass models [303], not
for particle models. For hard spheres in d → ∞, the only available calculation
is an approximate one at the 1RSB level, on the jamming line p ) = ∞ [88].
This calculation provides a 1RSB approximation for the threshold density, ) ϕth1RSB
= 6.870 [88]. Joining the point () ϕ †,1/) ϕth1RSB,0) by a straight
p † ) with the point ()
line in Figure 7.7 provides a rough approximation of the dynamical fullRSB line.
226 Counting Glass States: The Complexity

It is conjectured [303] that in the phase diagram of Figure 7.7 no states exist to
the left of the Gardner line for 1/)p ≥ 1/) p† (because the RSB solution cannot be
constructed perturbatively in this regime) and to the left of the dynamical fullRSB
line (which, however, remains to be precisely located) for 1/) p < 1/)p† .

7.5.3 On the Structure of Glass States

The interpretation of the Monasson construction when replica symmetry is broken
is not straightforward. The method relies on the two assumptions formulated in
Section 7.2.2:
1. Phase space can be partitioned in glass states.
2. m replicas can be confined within the same glass state.
When replica symmetry is broken, both assumptions must be revised [262, 263,
303]. While the partitioning in states is always possible [254], states are now organ-
ised in a hierarchical structure of sub-basins, as discussed in Section 5.3.3. In this
situation, a state is fully identified by a series of indices {α1,α2, . . . ,αk }: a first
index α1 labelling the largest metabasins, a second index α2 labelling the distinct
sub-basins inside α1 and so on, down to the last index αk labelling individual states
(see Figure 5.7). To each level of the hierarchy, one can associate a corresponding
complexity: the complexity k (f) counts the number of individual states with free
energy f (labelled by αk ), the complexity k−1 (f) counts the number of sub-basins
that group individual states and so on, up to the complexity 1 (f) that counts the
largest metabasins [262].
The second assumption must also be revised. In the simplest Monasson construc-
tion introduced in Section 7.2, we assumed that m replicas could be confined within
a same state. However, in practical computations, it is hard to implement this con-
straint if replica symmetry is spontaneously broken. It is certainly straightforward to
confine replicas into the same metabasin α1 because escaping from such metabasin
implies diffusion – hence, an infinite . As long as (x) remains finite for all x,
the m replicas thus belong to a same metabasin. But if (x) acquires a fullRSB
structure, the replicas inside this metabasin then have a non-trivial distribution of
mutual overlaps, and different glass states inside this metabasin get visited.
The correct interpretation of the Monasson construction in this case is that repli-
cas are free to explore, independently in equilibrium, a given metabasin (labelled by
α1 ). Their free energy is therefore mfα1 , where fα1 is the equilibrium free energy of
this metabasin. This free energy takes into account the contribution of the individual
glass states within the metabasin and their complexity. Then, the partition function
of the m replicas is given by Eq. (7.17), where the sum is over α1 , and the complex-
ity 1 (f) counts the number of metabasins. Hence, the Legendre transform with
 7.6 Wrap-Up 227

respect to m gives the complexity of the biggest metabasins when replica symmetry
is spontaneously broken [262]. How to compute the complexities at lower levels
in the hierarchy – i.e., j (f) with j = 2, . . . ,k – remains an open problem. In
spin glass models, the calculation of k (f) has been attempted either by counting
directly the solution of the TAP equations [22, 81] or by using a more complicated
RSB ansatz [287], but these methods are more involved than the Monasson method,
and they have not yet been adapted to particle systems in the limit d → ∞.

7.6 Wrap-Up
7.6.1 Summary
In this chapter, we have seen that

• A notion of equilibrium complexity eq () ϕ,T ) can be introduced as the difference

between the liquid entropy and the internal entropy of typical glass states on
the equilibrium line, in the dynamically arrested phase (Section 7.1). This quan-
tity counts the number of typical glass states that contribute to the equilibrium
partition function.
• For some models, the equilibrium complexity vanishes at high density or low
temperatures, defining a Kauzmann transition line TK () ϕ ). For T < TK () ϕ ), the
liquid state does not exist anymore, even as a dynamically arrested phase. A phase
transition towards an ideal glass phase thus takes place at TK () ϕ ) (Section 7.1.3).
• The Monasson method (Section 7.2) makes use of m coupled replicas to compute
the complexity (f;) ϕ,T ) of glass states of any free energy f at a given state point
ϕ,T ). The equilibrium case corresponds to a specific free energy f = fg ()
() ϕ,T )
dominating the equilibrium partition function.
• The lowest free energy states, fmin ()ϕ,T ) such that (f;) ϕ,T ) = 0, dominate the
partition function beyond the Kauzmann transition. The ideal glass free energy,
ϕ,T ), can also be computed from the Monasson method (Section 7.2.3).
fmin ()
• Explicit expressions of the Monasson free energy can be obtained in the limit
d → ∞ (Section 7.3).
• The Monasson phase diagram of hard spheres in d → ∞ displays a region
in which glass states exist. This region is delimited by the Kauzmann line, the
dynamical line and the jamming line (Section 7.4). On the jamming line, where
pressure diverges, glass states reduce to mechanically stable states, and the
Monasson ensemble reduces to the Edwards ensemble.
• Spontaneous replica symmetry breaking is observed in the Monasson construc-
tion in part of the phase diagram. For hard spheres, it surrounds the jamming
line. The interpretation of replica symmetry breaking in this context is not fully
developed (Section 7.5).
228 Counting Glass States: The Complexity

7.6.2 Further Reading

We provide here a list of references that can be consulted to further explore the
subjects discussed in this chapter, selected according to the criteria discussed in
Section 1.6.2.
The existence of a Kauzmann transition and of an ideal glass phase in finite
dimensions has been the subject of a long debate in the glass literature, and the
equilibrium complexity has been measured in a wide variety of glass formers. While
at the mean field level, dynamical arrest happens at Td > TK , in finite dimension, the
dynamics is likely activated and thus very slow for T < Td (see Sections 3.5.2 and
4.5.2 for references). Equilibrating in the vicinity of TK thus requires times growing
exponentially in 1/|T − TK |, at variance with standard phase transitions for which
equilibration times grow as power laws in the distance from the transition. As a
consequence, state of the art methods can only measure the configurational entropy
and the associated length scale quite far from TK . Collections of experimentally
available data can be found in

• Angell, Entropy and fragility in supercooling liquids [16]

• Richert and Angell, Dynamics of glass-forming liquids. V. On the link between
molecular dynamics and configurational entropy [302]
• Capaccioli, Ruocco and Zamponi, Dynamically correlated regions and configu-
rational entropy in supercooled liquids [75]
Note that in experiments, the complexity is usually estimated by the difference
between the liquid and crystal entropies, under the assumption that the crystal and
glass entropies are similar. This assumption has been questioned; see, e.g.,

• Johari, A resolution for the enigma of a liquids configurational entropy-molecular

kinetics relation [194]
• Stillinger, Debenedetti and Truskett, The Kauzmann paradox revisited [336]
for a discussion. In numerical simulations of model systems, the glass entropy can
be computed by several methods. The equilibrium complexity can then be directly
estimated and compared with theoretical predictions. Early attempts are reported in

• Coluzzi, Mézard, Parisi et al., Thermodynamics of binary mixture glasses [102]

• Sciortino, Kob and Tartaglia, Inherent structure entropy of supercooled liq-
uids [322]
• Sastry, Evaluation of the configurational entropy of a model liquid from computer
simulations [313]
• Angelani and Foffi, Configurational entropy of hard spheres [13]
but were severely limited by the impossibility of equilibrating the liquid at
temperatures below the dynamical arrest using standard molecular dynamics. More
 7.6 Wrap-Up 229

recently, thanks to the smart ‘swap’ algorithms already mentioned in Section 4.5.2,
supercooled liquids have been equilibrated in computer simulations down to the
laboratory glass transition and beyond (but still far from TK ). Thanks to these
methods, the configurational entropy has been measured down to much lower
temperatures, both in d = 2 and d = 3; see

• Berthier, Charbonneau, Ninarello et al., Zero-temperature glass transition in two

dimensions [46]
• Berthier, Charbonneau, Coslovich et al., Configurational entropy measurements
in extremely supercooled liquids that break the glass ceiling [45]
Extrapolation of the data in d = 2 suggests that TK = 0, while in d = 3, it
is consistent with a finite TK . While the debate remains open and interesting, it
is important to keep in mind that even at the mean field level, the existence of a
Kauzmann transition is model dependent, and that, in any case, laboratory glasses
are always trapped in free energy states much higher than fmin . The properties of
the Kauzmann transition and of the ideal glass phase have, therefore, little practical
relevance.
It has been theoretically established that the Kauzmann transition should
necessarily be associated to a diverging thermodynamic length scale. This was
later confirmed within the random first-order transition (RFOT) approach, by an
analysis of the excitations around energy minima, by the study of Kac models, by a
rigorous analysis, and by numerical simulations. Within the RFOT framework, this
phenomenon is intimately connected with the non-convexity of the Franz–Parisi
potential discussed in Section 4.5.2. Besides the references already mentioned in
Section 4.5.2 on the non-convexity of the Franz–Parisi potential and nucleation in
the RFOT approach, additional relevant references are

• Adam and Gibbs, On the temperature dependence of cooperative relaxation prop-

erties in glass-forming liquids [1]
• Bouchaud and Biroli, On the Adam–Gibbs–Kirkpatrick–Thirumalai–Wolynes
scenario for the viscosity increase in glasses [63]
• Stillinger, Supercooled liquids, glass transitions, and the Kauzmann paradox
[334]
• Montanari and Semerjian, Rigorous inequalities between length and time scales
in glassy systems [261]
• Franz and Montanari, Analytic determination of dynamical and mosaic length
scales in a Kac glass model [143]
• Biroli, Bouchaud, Cavagna et al., Thermodynamic signature of growing amor-
phous order in glass-forming liquids [57]
• Yaida, Berthier, Charbonneau et al., Point-to-set lengths, local structure, and
glassiness [363]
230 Counting Glass States: The Complexity

Measuring the complexity out of equilibrium is much harder than in equilibrium

because of the absence of a reference liquid state. Efforts have thus exclusively
focused on the simplest case of jammed packings to test the validity of the Edwards
ensemble. The literature on this problem is extremely large; see, e.g.,

• Mehta (ed.), Granular matter: An interdisciplinary approach [249]

• Barrat, Kurchan, Loreto et al., Edwards measures: A thermodynamic construction
for dense granular media and glasses [29]
• Chakraborty, Statistical ensemble approach to stress transmission in granular
packings [82]
• Bowles and Ashwin, Edwards entropy and compactivity in a model of granular
matter [66]
• Asenjo, Paillusson and Frenkel, Numerical calculation of granular entropy [20]
• Martiniani, Schrenk, Stevenson et al., Turning intractable counting into sam-
pling: computing the configurational entropy of three-dimensional jammed pack-
ings [247]
for introductory reviews and a sample of numerical results.
 8
Packing Spheres in Large Dimensions

The sphere packing problem consists of finding the densest arrangement of equal-
sized spheres in Rd , i.e., the infinite d-dimensional Euclidean space. This geometri-
cal problem is simply stated, which in part explains why it has attracted the attention
of mathematicians since the ancient times. But it also has connections to other areas
of mathematics, natural sciences and engineering [98, 103, 292, 344], which makes
it far from being an abstract problem. In physics, this problem is also connected to
the existence and stability of crystalline phases. Since Shannon’s pioneering work,
the large d limit of this problem is known to be connected to the practical problem
of designing error-correcting codes in communication technology [103]. In this
chapter, we review some of the known results and discuss how the results of the pre-
vious chapters can provide additional insight on this problem in the limit d → ∞.

8.1 Statement of the Problem

8.1.1 Close Packing Density
The sphere close packing density of d-dimensional Euclidean space, denoted θ(d),
is the supremum of the packing density over all equal-sized sphere packings P of
Rd . More precisely:1

• A sphere packing P of R is a countably infinite collection of points X(P)

d

= {xi }i∈Z , such that |xi − xj | ≥  for all pairs i,j of points. Here  is the sphere
diameter.
• Consider a continuous sequence of regions V(L) ⊂ R parametrised by a char-
d

acteristic linear size L and of volume V (L). An example is a cubic volume

1 The conventional notations for the sphere packing problem differ a lot in the mathematics and physics
communities; here we mostly follow the physics conventions, with some exceptions, and we provide both
notations when needed to avoid ambiguities.

231
232 Packing Spheres in Large Dimensions

V(L) = [−L/2,L/2]d or a sphere of radius L centered at the origin. The packing

density of P in this region is then
Vc (L,P)
ϕ(L,P) = , (8.1)
V (L)
where Vc (L,P) is the volume of V(L) covered by spheres of diameter  centered
at the points of P (note that spheres can cross the boundary of V).
• The close packing density in d dimensions is defined as
θ(d) = sup lim sup ϕ(L,P). (8.2)
P L→∞

In order to solve the sphere packing problem in dimension d, one would like to
know both the value of θ(d) and at least one packing P ∗ (d) that has a density
asymptotically equal to θ(d) when L → ∞.

8.1.2 Periodic versus Non-periodic Packings

Periodicity is an important property of packings. In fact, all known very dense
packings are periodic; see Figure 8.1. The distinction between periodic and
non-periodic packings is a recurrent theme in the literature. In physics, this
corresponds to the distinction between amorphous (or ‘disordered’) and crystalline
(or ‘ordered’) phases of matter. Disorder in this context is often intended as
‘absence of order’, and in this sense, any non-periodic packing can be considered
as disordered. This grouping is fairly crude, but defining and quantifying disorder
more precisely is particularly difficult [344]. This problem will not be discussed
here, but instead, we list a few classes of packings that can be explicitly constructed,
from the ‘most ordered’ to the ‘most disordered’ ones.
2
• Bravais lattices – The simplest class of periodic packings are ‘Bravais lattices’,
which represent crystals with a single–particle unit cell. They are defined by a
set of d primitive vectors a 1, . . . ,a d , with a i ∈ Rd or, equivalently, by a d × d
matrix  = {a 1, · · · ,a d } having the primitive vectors as columns.3 The points of
the packings are obtained as
x(n) = n1 a 1 + · · · + nd a d = Ân, n ∈ Zd , (8.3)
where n is an arbitrary d-dimensional vector of integers. Because x(0) = 0 is
part of the packing, and by periodicity, the maximum allowed sphere diameter
corresponds to the closest point to the origin, i.e.,

2 In the mathematical literature, they are simply called ‘lattice packings’.

3 Many different matrices  can characterise the same lattice, but there is a well-defined region in the space of
matrices that has a one-to-one correspondence with Bravais lattices.
 8.1 Statement of the Problem 233

(Â)2 = min |Ân|2, (8.4)

n

while the volume of a unit cell is | det Â|. Because there is one sphere per unit
cell, the packing fraction of a Bravais lattice is then simply

Vd (Â)d
ϕ(Â) = . (8.5)
2d | det Â|

The calculation of (Â) for a generic matrix Â, however, is an algorithmically

hard problem when dimension increases [98]. Note that because of periodicity,
the structure factor S(q) of Bravais lattices exhibits Bragg peaks. These are delta
peaks at values of q that belong to the reciprocal Bravais lattice, which is defined
by q · x(n) = 2πp, for some p ∈ Z.
• Non-Bravais lattices – Lattices obtained by considering a unit cell containing
k points in positions {b1 = 0,b2, . . . ,bk }, and repeating it periodically in space
along a set of primitive vectors, are non-Bravais lattices. The lattice points then
have the form

xα (n) = bα + n1 a 1 + · · · + nd a d = bα + Ân, (8.6)

with n ∈ Zd and α = 1, . . . ,k. The unit cell vectors have to be chosen in such a
way that bα = Ân for all n. Non-Bravais lattices describe crystals with complex
unit cells [21], and their packing fraction is easily computed by a generalisation
of Eq. (8.5). In this case, periodicity also induces Bragg peaks in the structure
factor, arranged in periodic positions in Fourier space.
• Quasiperiodic packings – Quasiperiodic packings, or ‘quasicrystals’, are
non-periodic packings which still exhibit Bragg peaks in their structure factors
[126, 190]. A simple way of constructing d-dimensional quasicrystals is to
consider a lattice in some dimension d  > d and project it on a plane of dimension
d that is not one of the lattice planes. Despite the absence of periodicity,
quasicrystals present patterns that repeat often, in such a way that the number of
packings that can be created by a given rule is not exponential in the volume of
the system [219].
• Disordered packings – Several proposals have been made to define disordered
packings in terms of pattern repetitions [219, 309] or local order metrics [344],
but these constructions are difficult to apply in practice. To bypass the problem
of giving a general definition of disordered packings, one can instead define
concretely a class of disordered packings by considering the Gibbs–Boltzmann
measure of hard spheres, which is uniform over all possible packings of a given
packing fraction ϕ. Whenever the system can be equilibrated in the fluid phase,
typical equilibrium configurations of hard spheres are disordered sphere packings.
234 Packing Spheres in Large Dimensions

To obtain higher-density packings, one can compress the system adiabatically

from the liquid phase, as discussed in Chapter 4, up to the jamming density. This
particular class of disordered packings is discussed in more detail in the rest of
this chapter.
An important remark is that the Gibbs–Boltzmann measure is usually defined by
considering a cubic volume V(L) = [−L/2,L/2]d ⊂ Rd with periodic boundary
conditions and then giving uniform probability to all configurations of N points in
V(L) that satisfy the hard-sphere constraint. Therefore, all the configurations that
belong to the Gibbs–Boltzmann measure can be considered, by periodic extension
of the box V(L), as packings over the infinite Euclidean space Rd . For finite L, one
thus obtains non-Bravais lattices with N particles in a unit cell defined by the box
V(L). Only in the limit L → ∞, provided one is still in the liquid phase, does one
obtain fully disordered packings over the infinite space. For the same reason, any
disordered packing can be approximated up to a desired precision by a non-Bravais
lattice with a large enough unit cell.

8.1.3 Constructive versus Non-constructive Approaches

As stated in Section 8.1.1, one would like to know the close packing density θ(d)
and a packing that achieves it. Several strategies have been adopted to reach, at least
partially, this goal:

• Bounds – One approach to the problem is to construct upper and lower bounds to
θ(d), even if non-constructive, in the sense that they do not provide information
about the best packing configuration, but only about the packing density. A review
of available bounds will be given in Section 8.2. The strategies that have been
followed to obtain them are diverse, spanning several fields of mathematics.
• Deterministic packing construction – Another approach is to directly exhibit a
single (typically periodic) packing P of the infinite Euclidean space. The density
ϕ(P) is then a lower bound for θ(d). If ϕ(P) happens to coincide with the best
upper bound, then one has found the best packing. Sometimes packings can be
constructed rather easily (e.g., in the case of Bravais lattices, it is enough to
specify the primitive vectors), but the procedure to construct the packing can
also be rather complex [268]. It is then also interesting to consider how many
operations are needed to obtain the result.
• Stochastic packing construction – One can propose a stochastic procedure to
construct an ensemble of packings. For example, one could sample from the
equilibrium Gibbs–Boltzmann distribution at some constant density (if possible).
In this case, one also encounters the algorithmic problem of the number of oper-
ations needed to achieve a proper sampling.
Examples of these approaches are given in the following sections.
 8.2 Review of Rigorous Results 235

8.2 Review of Rigorous Results

We present in this section a brief review of known rigorous results on the sphere
packing problem. More detailed reviews can be found in [98, 103].

8.2.1 Best Known Packings

Currently, the value of θ(d) is rigorously known only for d = 1,2,3,8,24.

• The case d = 1 is trivial. The densest packing is a one-dimensional lattice, with

particles in positions xn =  n, n ∈ Z, and θ(1) = 1.
• The case d = 2 is elementary
√ but not trivial. The densest packing is the hexagonal
lattice and θ(2) = π 3/6 = 0.9069 . . . The first proof is attributed to Thue in
1892 [98, page 5]. √
• For d = 3, the close packing density θ(3) = π/(3 2) = 0.74048 . . . is achieved
by any sequence of staggered hexagonal planes, including the face-centred cubic
(FCC) and hexagonal close-packed (HCP) lattices. After centuries of efforts,
a proof was finally achieved by Hales. The proof, which is very complex and
heavily based on computer assistance, was first announced in 1998 and then
went through a long review process to be finally published in 2005 [174]. It was
subsequently verified at the level of formal logic [173].
• The cases d = 8 and d = 24 were proven in 2016, the former by Viazovska [353]
(d = 8, the densest packing being a lattice called E8 ) and the latter by Cohn,
Kumar, Miller, Radchenko and Viazovska [101] (d = 24, the densest packing
being the Leech lattice). The proof, which is much simpler than for d = 3, is
based on deriving a series of upper bounds depending on auxiliary functions
and then finding the (non-trivial) function such that the upper bound precisely
coincides with the packing density of the candidate best lattice. A pedagogical
introduction can be found in [99].
It is interesting to note that in all these cases, the best packing is a Bravais lattice –
i.e., a crystal with a single-particle unit cell – but the optimal packing also depends
sensitively on d. A list of the best known packings is available in [103], and more
updated lists are also available online [271]. The density of the best known packings
is reported as a function of d in Figure 8.1. The list in [271] and the figure clearly
show that every dimension has its own geometric oddities. As a result, what is
known about the densest packings in a certain dimension d is generally useless for
other dimensions, even those very close to d. For instance, stacking d-dimensional
good packings generally produces poor packings in d + 1.
Let us define a packing as ‘saturated’ when there is no room for adding an extra
sphere. Clearly, saturated packings exist: they can be constructed by starting from
a non-saturated packing, and adding spheres to fill the holes until the packing is
saturated. An interesting observation is that a cubic lattice is not saturated for d ≥ 4.
236 Packing Spheres in Large Dimensions
106
1019 Minkowski LB Minkowski LB

1017 Venkatesh LB 5 Densest Bravais

10
1015 Mostrou LB Densest non-Bravais

13 KL UB 104 Cohn-Elkies UB
10
Densest Bravais
1011
2d ϕ

2d ϕ
GCP 103
109
107 102
105
103 10
10
1
10−1
1 10 102 103 104 5 10 15 20 25 30 35
d d

Figure 8.1 (Left) Scaled packing fraction 2d ϕ of the best known Bravais lattice
packings as a function of dimension d [271]. For comparison, the best upper
bound (UB) by Kabatiansky and Levensthein [197]; the lower bounds (LB)
by Minkowski, Venkatesh [352] and Mostrou [268]; and the glass close pack-
ing (GCP) density are also shown. Because these are only asymptotic results
for d → ∞, they are only shown for d > 100 for illustration. (Right) Scaled
packing fraction 2d ϕ of the best known packings as a function of d, zooming on
the region of small d. Bravais lattices are shown as crosses joined by a full line;
non-Bravais lattices are shown as full squares for a set of dimensions d where
they are better than Bravais lattices [271]. Also shown is the best upper bound
by Cohn and Elkies [100, table 3], which is saturated to very high precision in
d = 1,2,3,8,24.

√
Indeed, the center of each cubic cell in the lattice is at distance  d/2 from any
point of the lattice, where  is the lattice size. Therefore, a new sphere can be
added there if d ≥ 4. In sufficiently large d, no saturated Bravais lattice has been
found, leading to the conjecture that no such lattice exists in large enough d [98].
If the conjecture is true, then dense periodic lattices in large d must have more than
one particle per unit cell. For instance, the best known packing in d = 10 has 40
particles in the unit cell and is 8% denser than any known Bravais lattice. Non-
Bravais lattices are also the densest known packings in several d ≥ 10 (Figure 8.1).
The best packings in asymptotically large d might thus have large unit cells or even
be disordered [344].

8.2.2 Lower Bounds

When d grows, the unusual geometrical features of high-dimensional spaces
become more pronounced. For instance, the volume of the unit diameter sphere
vanishes exponentially with respect to the volume of the unit hypercube. In this
asymptotic limit, only upper and lower bounds are known (see [98, 103, 344]
for a more detailed review). We present here, in chronological order, a list of
lower bounds.
 8.2 Review of Rigorous Results 237

• Minkowski proved in 1905 that the volume occupied by a saturated packing

must satisfy ϕ ≥ 2−d ; hence, because saturated packings certainly exist, we
have θ(d) ≥ 2−d . The proof is actually elementary. By definition, if we double the
radius of each sphere, then space must be covered completely; otherwise, a new
sphere would fit in the original packing. Doubling the radius multiplies volume
by 2d , and so the original packing must cover at least a 2−d fraction of space.
• This lower bound was improved by a linear d factor by Rogers in 1947, who
proved that θ(d) ≥ d 2−d [306, 307] (see also [283]).
• Lebowitz and Penrose in 1964 [226] proved that the virial series expansion of the
entropy is convergent for ϕ < 0.14467 . . . 2−d ; hence, θ(d) ≥ 0.14467 . . . 2−d .
Additionally, the proof provides an exact expression for the excess equilibrium
entropy per particle, as discussed in Chapter 2:
2d ϕ
s ex (ϕ) = − . (8.7)
2
−d
• Vance’s 2011 bound is θ(d) ≥ (6/e)d 2 (for d divisible by 4) [351].
• In 2012, Venkatesh dramatically stepped up this effort by proving that
θ(d) ≥ 65,963 d 2−d and that θ(d) ≥ 0.5 log(log d)d 2−d for infinitely many
(but not all) dimensions [352].
• While most of these lower bounds are non-constructive, in 2017, Moustrou
proved that, in infinitely many dimensions, packings with density
ϕ ≥ 0.89 log(log d) d 2−d can be constructed with exp(1.5 d log d) binary
operations [268].
• The result of Lebowitz and Penrose was extended to higher densities in 2018
by Jenssen,√Joos and Perkins [191]. They proved that at packing fraction
ϕ = log(2/ 3) d 2−d , the entropy density of the Gibbs measure is s ex ≥ −2d ϕ,
which also proves that there exist packings at this density. Note that this bound
is compatible with the validity of Eq. (8.7). The missing factor 2 is likely due to
technical details of the proof.
All the bounds mentioned have their own interest. Currently, the best lower bounds
are those of Venkatesh and Moustrou, which give ) ϕ ≥ 0.89 log(log d) in infinitely
many (but not all) dimensions, and ) ϕ ≥ 65,963 in all other sufficiently large
dimensions.

8.2.3 Upper Bounds

Proving upper bounds to θ(d) in the limit d → ∞ is also quite challenging.
−d/2
• In 1929, Blichfeldt proved that θ(d) ≤ (d/2)2 [60]. This result was slightly
−d/2
improved by Rogers to θ(d) ≤ (d/e)2 [306].
−0.5990...d
• In 1978, Kabatiansky and Levensthein proved that θ(d) ≤ 2 [197].
238 Packing Spheres in Large Dimensions

To date, these are essentially the only two significant improvements that have been
made to upper bounds in the asymptotic limit d → ∞. However, in 2003, Cohn and
Elkies [100] proved a very important related theorem. It can be stated roughly as
follows. Let f (r) be a function that is smooth enough and decays fast enough, and
let f)(q) = dr eiq·r f (r) be its Fourier transform. If one can find a (non-identically
zero) function such that
f (r) ≤ 0 for |r| ≥ , f)(q) ≥ 0, ∀q, (8.8)
then for spheres of diameter  one has4
 d
 f (0) π d/2
θ(d) ≤ ϕCE [f ] = Vd , Vd = . (8.9)
2 f)(0) (d/2 + 1)
The idea of the proof is very simple, and it can be found in [98]. Note that the bound
obtained from Eqs. (8.8) and (8.9) is invariant under multiplication of f (x) by an
arbitrary constant. The best upper bound from Eq. (8.9) can then be obtained by
fixing, for example, a normalisation f)(0) = 1 and then minimising f (0) subject to
the linear constraints in Eq. (8.8). This problem can be solved by efficient ‘linear
programming’ algorithms. This theorem is at the basis of the proof of the optimal
packing in d = 8 and d = 24 [99]. It has also provided many other interesting
results in finite d, but thus far, for d → ∞, it could only reproduce the asymptotic
exponential scaling of the Kabatiansky–Levensthein 1978 bound, without improv-
ing it [98].
In a follow-up work, Cohn [97] made another interesting observation. Consider
a function5 g(r) that satisfies, for a fixed ρ, the linear constraints
g(r) = 0 for |r| < , g(r) ≥ 0 for |r| ≥ ,
(8.10)
S(q) = 1 + ρ[)
g (q) − (2π)d δ(q)] ≥ 0, ∀q.
Then, for any function satisfying the constraints in Eq. (8.8) one has
 
dq )
f (0) ≥ drf (r)[δ(r) + ρg(r)] = f (q)[1 + ρ) g (q)] ≥ ρ f)(0), (8.11)
(2π)d
and as a consequence,
 d  d
 f (0) 
ϕCE [f ] = Vd ≥ Vd ρ = ϕ. (8.12)
2 )
f (0) 2

4 Note that according to Eq. (8.8), f)(q) ≥ 0 and is not identically zero. It follows that f (0) =
 dq )
f (q) > 0.
(2π )d
If f)(0) = 0, then the bound is useless, as it gives θ (d) ≤ ∞.
5 More precisely, a distribution. See [97] for details.
 8.2 Review of Rigorous Results 239

Note the following:

• The linear constraints in Eq. (8.10) are the minimal constraints one should impose
on the radial distribution function of a system of hard spheres of diameter ,
number density ρ and packing fraction ϕ. The radial distribution function g(r)
should indeed vanish inside the hard core, be positive everywhere and have a
positive structure factor S(q) (see [175] and Chapter 2).
• Finding a solution to Eq. (8.10) at some density ρ does not imply that hard-sphere
configurations exist at that density. Many other constraints have to be satisfied by
pair and higher-order correlations [344].
• Finding a solution to Eq. (8.10) at some density ρ with associated packing fraction
ϕ implies, according to Eq. (8.12), that the Cohn–Elkies method cannot provide
a better upper bound than θ(d) ≤ ϕ. In this sense, the linear problem of finding
the lowest ϕCE [f ] under the linear constraints in Eq. (8.8) and that of finding the
highest ρ under the linear constraints in Eq. (8.10) are thus mathematically dual.
Finding a solution to one implies an upper or lower bound on the solution of the
other.6
• Parisi and Slanina [284] developed a procedure to obtain a solution to the lin-
ear problem in Eq. (8.12) up to a maximal value of packing fraction scaling
asymptotically as 2−0.7786...d . This approach was simplified by Torquato and Still-
inger [344], who proposed the simple test function g(r) = θ(|r|−A)+Bδ(|r|−),
with adjustable parameters A ≥  and B ≥ 0, and found that this test function is a
solution to Eq. (8.12) up to the same packing fraction. These works show that the
Cohn-Elkies method cannot provide a better upper bound than 2−0.7786...d ; hence,
the gap with the best lower bound cannot be closed this way. Interestingly, the
packing fraction 2−0.7786...d is also where the resummation of ring diagrams in the
virial expansion ceases to be valid, as discussed in Chapter 2.
In summary, while the best lower bound on θ(d) scales asymptotically as 2−d , the
best upper bound scales as 2−0.5990... d , and therefore, its ratio with the best lower
bound grows exponentially in d (Figure 8.1). Moreover, there are currently no good
ideas on how to improve the exponential scaling 2−d of the lower bound, while the
most effective method to prove upper bounds due to Cohn and Elkies surely cannot
achieve anything better than 2−0.7786...d (and it is still far from achieving it). This
leaves a huge uncertainty in the asymptotic scaling of θ(d) for large d. In addition,
it is by no means obvious that for large d there exist ‘universal features’ of optimal
packings or that they should necessarily be periodic [98, 344].

6 It is conjectured that there might be no gap between the best solutions of the two problems [97].
240 Packing Spheres in Large Dimensions
0.8
Liquid EOS
d=3 FCC crystal EOS
Coexistence
0.6
1/(pϕ)

0.4

0.2

0.0
0.3 0.4 0.5 0.6 0.7 θ(3)
ϕ

Figure 8.2 Equilibrium equation of state (EOS) of hard spheres in d = 3, showing

the liquid and crystal phases and their coexistence. The inverse scaled pressure
ϕp = ϕβP /ρ = (π/6)βP 3 is shown as a function of the packing fraction ϕ. For
the liquid, the Carnahan-Starling EOS p = (1 + ϕ + ϕ 2 − ϕ 3 )/(1 − ϕ)3 has been
used [175]. For the crystal, the Speedy equation of state p = 3/(1 − z) − a(z − b)/
(z − c) has been used, with z = ϕ/θ (3) and a = 0.620735, b = 0.708194,
c = 0.591663 [25]. The coexistence pressure βPco 3 = 11.5727 is taken
from [139].

8.3 Review of Non-rigorous Results

We now describe some of the results on packing problems in large dimension that
have been found within the physics literature and currently lack a rigorous proof.
We mostly focus on results obtained within statistical mechanics. More detailed
reviews can be found in [30, 292, 344].

8.3.1 Crystals in High Dimensions

In physics, periodic packings correspond to crystals, while disordered packings
correspond to liquids or glasses. It was first suggested by Kirkwood and Monroe in
1940 [209] and then shown numerically by Alder and Wainwright in 1957 [5, 175]
that, upon increasing density or pressure, a hard-sphere liquid in d = 3 transforms
into a crystal7 via a first-order phase transition, as illustrated in Figure 8.2.

7 Remarkably, the structure of the equilibrium hard-sphere crystal in d = 3 has been the subject of a debate
because the free energy difference between the FCC and HCP crystals is extremely small [61, 358].
State-of-the-art numerical results suggest that the FCC lattice is slightly more stable [61, 211, 275]. Because
the FCC and HCP lattices have the same close-packing density, however, which one of the two is more stable
is irrelevant for the present discussion.
 8.3 Review of Non-rigorous Results 241

The crystal then becomes one of the close-packed structures in d = 3 (see

Section 8.2.1) upon further compression to infinite pressure. This observation
suggests that one can start by an ideal gas of hard spheres (i.e., throwing points
uniformly at random) and then slowly compress the system by increasing the size
of the spheres [235]. If the compression rate is slow enough, the system crystallises
√
and eventually reaches the close packing density at ϕ = θ(3) = π/(3 2). The
same strategy would obviously work in d = 1, where there is no phase transition
up to close packing, and in d = 2, where two phase transitions take place but
crystallisation into a hexagonal lattice is nonetheless easily achieved [35]. Slow
compression from the liquid is thus a simple way to obtain the densest packing
in d ≤ 3.
The equilibrium phase diagram has been studied in d = 4,5,6 as well [350].
The result is similar to that of d = 3: a first-order phase transition separates
the low-density liquid phase from a high-density crystal phase whose lattice
structure coincides with the one of the best packing. Crystallisation during a slow
compression becomes, however, increasingly hard when dimension is increased.
Already in d = 4, with current computational power, it is very hard to observe
spontaneous crystallisation during a compression, and the situation is worse in
d = 5 and d = 6, where no hint of crystallisation has been detected [330, 350].
This marked suppression of crystallisation is due to the growth of the surface
tension between the two phases, which is related to the fact that the local geometry
of the fluid becomes increasingly different from that of the crystal [350]. Instead
of crystallising, upon compression, the system follows the metastable liquid
equation of state and ultimately forms a glass, as it has been tested up to d = 12
[86, 330].
Conceptually, the fact that the barrier to crystallise increases with d implies that
the liquid can be compressed at pressures above the coexistence pressure Pco , while
remaining stable for very long times. The liquid at P > Pco (see Figure 8.2) is
usually deemed8 ‘supercooled’ [120] and is metastable with respect to the crystal.
For d ≥ 4, the lifetime of the metastable supercooled liquid becomes extremely
large and grows with d. The metastable liquid phase is therefore better and better
defined upon increasing d. In the limit d → ∞, as discussed in Chapter 1, the
lifetime of the metastable state becomes infinite. In this limit, the liquid and the
crystal are well defined and well separated minima of the free energy F [ρ(x)]
defined in Chapter 2. Within the theory, one can therefore very easily restrict the
study to the liquid phase by assuming that the density profile is homogeneous
and isotropic.

8 Although when the control parameter is the density, ‘overcompressed’ would be a better nomenclature.
242 Packing Spheres in Large Dimensions

8.3.2 Algorithms to Construct Dense Packings

The difficulty to crystallise a system of hard spheres in d ≥ 4 has an important
algorithmic implication. Namely, it is impossible to construct good crystals by
simple compression of low-density liquid configurations [330]. We now discuss
a simple example to illustrate this effect.
Let us assume that the relaxation time of the equilibrium liquid is finite at the
coexistence pressure, which is the case in all d ≤ 6 where the problem has been
studied [139, 350]. The density at which the liquid becomes metastable with respect
to the crystal is then smaller than the dynamical glass transition density ϕd . Sam-
pling equilibrium liquid configurations of N particles slightly below coexistence
therefore requires a time of the order of Nd, the number of degrees of freedom in
the system, because each degree of freedom must be updated a finite number of
times (either by Monte Carlo or molecular dynamics) before equilibrium can be
reached. If the system of N particles is confined in a box of linear size L and
volume V = Ld with periodic boundary conditions, the resulting packing can
be periodically extended to the infinite Euclidean space, for which it would be a
periodic packing of period L with N particles in the unit cell. For this procedure
to be well defined, however, the linear size L of the system must be at least L/ =
1 + O(1/d), in such a way that, for instance, a particle located in the origin can fit
in the box without overlapping with itself. The periodic extension would otherwise
not produce a packing.9 In the following, to obtain a rough estimate of the scalings
with dimension, we neglect all sub-exponential corrections in d. Recalling that the
natural scale of packing fraction in the liquid phase is ϕ ∼ 2−d , and using Stirling’s
approximation for the volume Vd ∼ exp[ d2 log(2πe/d)], the minimal number of
particles must thus be of the order of
Ld ϕ 2d × 2−d d d
N≈ ≈ ≈ e 2 log 2πe ≈ d d/2 . (8.13)
Vd (/2)d Vd
Note that this value of N diverges with d. Because finite size corrections vanish as
1/N in the liquid phase [86], in the large d limit, the resulting system is effectively
in the thermodynamic limit despite the fact that its linear size L is finite.10 The
efficient sampling of liquid configurations then requires a time that scales propor-
tionally to dN; hence,
τliq ≈ dN ≈ d d/2 (8.14)

9 The stricter requirement that a particle cannot collide with itself due to the periodic boundary conditions
would lead to L/ = 2 + O(1/d), which does not change the leading order in Eq. (8.13).
√
10 Note, however, that a d-dimensional cube of side L = 2 has 2d−1 diagonals of length 2 d. Its linear size
along the diagonals
√ is thus very large when d → ∞. Similarly, two random points in this cube have a mean
distance ∝ d, and the distribution of distances is strongly concentrated around this mean.
 8.4 Liquid, Glass and Packings in Infinite Dimensions 243

at leading order in large d. One concludes that sampling configurations of a

sufficiently large system to define a liquid phase requires a time scaling as d d/2 ,
which is why current studies are limited to d ≤ 12 [86]. This scaling sets a ‘natural’
time scale for the problem of constructing large packings that is intrinsically hard
to beat.
In addition, as discussed in Section 1.5.3, the crystallisation time is the exponen-
tial of a free energy barrier that scales proportionally either to N or to d, depending
on the scaling of the size of the system with d. Because the relevant scaling here
is d → ∞ at fixed, small size L – in which case, the barrier is proportional to
N = Ld – one can estimate
d/2
τcryst ≈ τliq eN ≈ ed , (8.15)

which is much worse than τliq . Although many other out-of-equilibrium compres-
sion algorithms have been proposed – see [30, 292, 344] for example – none can
construct dense lattice packings in d ≥ 4 much faster than slow compression.
An interesting alternative idea is to restrict the exploration of configurations to
the space of Bravais lattices, which are fully specified by a d ×d matrix. One can try
to sample the space of these matrices looking for very dense lattices [12, 198, 199],
but the number of possible lattices grows extremely fast with d. In addition, their
sampling is particularly demanding because even obtaining the packing fraction of a
given lattice is a computationally difficult problem. As a result, current studies only
managed to reproduce the densest known packings for d ≤ 20 and were unable to
go beyond that limit. For a related statistical mechanics approach to lattice packings
in large d, see [283].

8.4 Liquid, Glass and Packings in Infinite Dimensions

In this section, we show how the theory of liquids and glasses in d → ∞
described in this book can provide some insight into the problem of packing spheres
in d → ∞.

8.4.1 Dynamical Glass Transition

In Chapter 3, it has been shown that the hard-sphere liquid dynamics has a finite
relaxation time for ) ϕ<) ϕd = 4.8067 . . . One can therefore start from an ideal gas
configuration and compress the system [235] in order to achieve equilibration at
any target density ) ϕ<) ϕd . As discussed in Section 8.3.2, this requires a number of
particles N that scales as in Eq. (8.13) and a time scale (or inverse compression rate)
that grows as Eq. (8.14). The solution of infinite dimensional liquids thus suggests
a first interesting result:
244 Packing Spheres in Large Dimensions

R1 In d → ∞, up to a packing fraction ϕ = d2−d × 4.8067 . . ., there exist

packings of hard spheres that can be constructed in a time τliq ≈ d d/2 by
simple adiabatic compression of the ideal gas. Their excess entropy at packing
d
fraction ϕ is given by s ex = − 2 2ϕ .
R1 is very similar to the rigorous result of Moustrou [268] discussed in
Section 8.2.2. Moustrou proved that packings with density ϕ ≥ 0.89 log(log d)d 2−d
can be constructed with exp(1.5 d log d) binary operations – i.e., with the same
scaling with d but with a prefactor of 1.5 instead of 0.5 in the leading exponential
term. While Moustrou’s result has a better asymptotic scaling of the density for
d → ∞, i.e., log(log d) instead of constant, R1 is actually better as long as
4.8/0.89
0.89 log(log d) < 4.8 ⇔ d < ee ≈ 1095 . (8.16)
Because in practical applications one would never reach such high dimensions,
R1 thus remains of comparable interest to Moustrou’s result. Also, this procedure
can construct an ensemble of packings with the liquid entropy at ϕd . The resulting
entropy has the same scaling with d as the rigorous one of [191], with better
prefactors.
The solution for the dynamical glass transition also suggests a way to slightly
improve the prefactor in R1. Consider a liquid with interaction potential

∞ h < 0,
v̄(h) = (8.17)
v̄+ (h) h ≥ 0,
where v̄+ (h) is a finite function that goes to zero for h → ∞ fast enough that the
hypotheses of Section 2.3.2 are satisfied. Any liquid configuration of the potential
in Eq. (8.17) is also a valid hard sphere packing because particles satisfy the hard
core constraint.11 One can therefore choose an appropriate potential v̄+ (h) in such
a way to push the dynamical glass transition to higher packing fractions. It was
shown in [325] that it suffices to add a short range attractive potential

−1/T) 0 ≤ h ≤ ) σ,
β v̄+ (h) = (8.18)
0 otherwise,

to push the dynamical transition up to ) ϕd ≈ 6.5 (for ) σ = 0.06 and T) = 0.48).

Adding a sequence of attractive and repulsive steps and optimising the parameters
to maximise )ϕd even gives ) ϕd = 6.966 . . . [240]. Because the optimisation proce-
dure of [240] converges upon increasing the number of steps, it is reasonable to
conjecture that it achieves the maximum over all possible functions v̄+ (h) so that no

11 Sampling from the liquid defined by Eq. (8.17), however, does not produce equilibrium hard-sphere
configurations because the Gibbs–Boltzmann measure is biased by the potential v̄+ (h).
 8.4 Liquid, Glass and Packings in Infinite Dimensions 245

additional improvement can be obtained by this strategy. Sampling configurations

of the liquid with optimised potential takes a time τliq ≈ d d/2 , as in R1, and the
resulting liquid has finite entropy, thus leading to the improved result:
R1+ In d → ∞, up to a packing fraction ϕ = d2−d × 6.966 . . ., there exist pack-
ings of hard spheres that can be constructed in a time τliq ≈ d d/2 by simple
adiabatic compression of an ideal gas of hard spheres with an additional finite
potential v̄+ (h). These packings have a finite excess entropy [240].

8.4.2 Adiabatic State Following

As discussed in Chapter 4, during a sufficiently slow compression, a liquid can be
kept in equilibrium up to ϕd . More precisely, the time scale of the compression (the
inverse of the compression rate) must be proportional to τliq ∼ Nd, as given in
Eq. (8.14). In order to reach equilibrium beyond ϕd , a much smaller compression
rate, scaling as exp(−N), must be employed, as we discuss in Section 8.4.3. A slow
compression on a time scale proportional to τliq thus falls out of equilibrium at ϕd .
Beyond that point, the pressure increases faster than that of the equilibrium liquid
and diverges at a jamming point ϕj (ϕd ). As discussed in Chapter 4, the out-of-
equilibrium glass prepared at ϕd can be followed adiabatically through the state
following construction in the limit d → ∞. However, as shown in Chapter 6,
this glass is always in the replica symmetry broken phase. It would then take
a time exp(N) for the system to follow adiabatically the evolution of the state.
If one is interested in a finite compression rate, the state following construction
might then seem irrelevant. Yet experience with spin glass models shows that, for
a continuous transition, the replica symmetry broken equation of state provides
a good approximation of the actual non-equilibrium equation of state. A liquid
of hard spheres compressed on a time scale given by Eq. (8.14) thus falls out of
equilibrium around ) ϕd = 4.8067 and then remains close to the equation of state
of the corresponding glass, which ends at a jamming point ) ϕd ). The precise
ϕj ()
jamming density is not known, but it is likely that )
12
ϕd ) > 7. This analysis
ϕj ()
leads to a further improvement of the prefactor in R1 and R1+ :
R2 In d → ∞, up to a packing fraction ϕ ≈ d2−d ) ϕd ), there exist packings
ϕj ()
of hard spheres that can be constructed in a time τliq ≈ d d/2 by simple non-
equilibrium compression of an ideal gas of hard spheres up to its jamming
point.

12 As already discussed in Section 7.4.4, the numerical results for ϕ

)j ()
ϕg ) reported in Section 6.3 can be well
approximated by ϕ)j ()
ϕg ) ≈ 3.324 + 0.730) ϕg + 0.0127)ϕg2 , which gives ) ϕd ) ≈ 7.1.
ϕj ()
246 Packing Spheres in Large Dimensions

Because the complexity of the packings obtained at ) ϕd ) is positive, as discussed

ϕj ()
in Chapter 7, an exponential number in N of distinct packings can be constructed
via this procedure.
It is important to stress that R2 is much harder to formalise than R1 and R1+ .
Proving R1 ‘only’ requires one to solve the equilibrium dynamics of the liquid,
which has been done at the theoretical physics level. Its rigorous mathematical
justification mostly requires proving that the truncation of the virial expansion is
correct (which is, however, likely to be extremely difficult). By contrast, R2 is not
even precise from the theoretical physics point of view, because the state follow-
ing construction only provides an approximation to the non-equilibrium dynamics
when replica symmetry is broken. A precise computation of the jamming point cor-
responding to slow compression would require the solution of the non-equilibrium
dynamical equations in d → ∞ derived in [3], which is very challenging already at
the theoretical physics level. Transforming R2 in a rigorous mathematical statement
would therefore be considerably harder. Note that R2 could perhaps be slightly
improved by considering the non-equilibrium compression of a system with an
additional potential v̄+ (h), as discussed in Section 8.4.1, but this possibility has
not yet been explored.

8.4.3 Glass Close Packing

To conclude the discussion, we consider the behaviour of the liquid when it is
compressed on a time scale (inverse compression rate) that scales as exp(N) –
which means, according to the discussion of Section 8.3.2, a scaling with d of
the form
d/2
τact ≈ eN ≈ ed , (8.19)

using again the minimal scaling of N with d discussed in Section 8.3.2. This time
scale is slow enough that the system can jump over the barriers that separate the
glass basins and thus also equilibrate in the dynamically arrested phase, ) ϕ>) ϕd .
Because τact has the same scaling as the crystallisation time τcryst given by
Eq. (8.15), it is possible that, at some point, the system could simply crystallise.
Whether this would take place or not depends on the existence of a dense enough
crystal phase and on the prefactors in the scaling of τact and τcryst [80]. If there is no
crystal phase denser than the liquid, or if the prefactors are such that τcryst  τact at
all densities, then crystallisation would not be observed. Because we have no such
information about crystals in large d, we leave this possibility as open and do not
discuss it further.
Instead, we focus on what happens if the system stays in the amorphous phase.
In this case, the system remains in equilibrium in the liquid phase at ) ϕ > ) ϕd
 8.4 Liquid, Glass and Packings in Infinite Dimensions 247

and samples the exponential number of glass basins in which the liquid is decom-
posed, as described in Chapter 7. Upon increasing density, the complexity of
basins decreases up to the Kauzmann transition at ) ϕK = log d, and, hence, ϕK =
d log(d) 2−d . Beyond that density, the number of glass basins is no more exponen-
tial and the system transitions to an ideal glass phase. If one keeps compressing the
system at rate 1/τact , equilibrium can be maintained in the ideal glass phase and the
system then follows the equilibrium ideal glass pressure p )K , which can be computed
using the Monasson method described in Chapter 7; see Figure 7.7. The equilibrium
compression in the ideal glass terminates at the glass close packing (GCP) point,
whose density ) ϕGCP = log d has the same scaling as the Kauzmann density at
leading order. Because the number of ideal glass basins is sub-exponential in N,
the number of distinct packings obtained at ϕGCP is also sub-exponential.
The solution of the infinite dimensional ideal glass of hard spheres discussed in
Chapter 7 thus suggests the following result:
R3 In d → ∞, up to a packing fraction ϕGCP = d log(d) 2−d , there exist
d/2
packings of hard spheres that can be constructed in a time τact ≈ ed by
very slow adiabatic compression of an ideal gas of hard spheres up to its ideal
glass phase. The number of distinct packings that can be constructed in this
way is sub-exponential in N .
Note that this result provides quite a substantial improvement over the best rigorous
lower bound, which predicts a scaling of ) ϕ ∼ 0.89 log(log d). R3 improves this
bound to ) ϕ ∼ log d. Yet the time that would be needed to construct these packings
has an extremely poor scaling with d, which makes a practical implementation
impossible already in low d. R3 is therefore mostly an existence result. Inter-
estingly, the leading exponential scaling of the packing fraction, 2−d , cannot be
improved by this approach. Improving it likely requires considering entirely dif-
ferent packings that have nothing to do with compressing ideal gas configurations.
These packings are thus unlikely to be fully disordered [98, 344].
Finally, note that R3 cannot be improved by considering an additional potential
v̄+ (h) as in Eq. (8.17). The expression of )
ϕK for the potential in Eq. (8.17) is indeed
deduced from Eq. (7.9):
 ∞
log d
ϕK =
) , gK = 1 + dh eh [1 − e−β v̄+ (h) (1 + β v̄+ (h))]. (8.20)
gK 0

Because the integrand function 1 − e−x (1 + x) ≥ 0 for all x, one has gK ≥ 1 (with
gK = 1 corresponding to the hard-sphere case v̄+ (h) = 0), and the leading-order
scaling of the Kauzmann density can only be decreased by adding an additional
potential. One can also check that the expression for the Edwards complexity,
discussed in Section 7.4.3, is independent of v̄+ (h). The presence of an additional
248 Packing Spheres in Large Dimensions

finite potential thus cannot change the structure of jammed packings at infinite
pressure, as expected. As a consequence, the expression of ϕGCP is independent
of v̄+ (h). This result is physically intuitive. The glass close packing is the highest
possible density of a well-defined class of amorphous packings, and as such, only
depends on the geometric properties of the hard-core potential.

8.5 Wrap-Up
8.5.1 Summary
In this chapter, we have seen that

• The sphere packing problem of finding the highest packing fraction θ(d) of iden-
tical spheres in Rd is very difficult. Its solution is only known in a small set of
low dimensions d = 1,2,3,8,24. The structure of the large d Euclidean space is
such that it is also very hard to find regularities in the solution. As a consequence,
no results for good packings are known in large dimensions (Section 8.2.1).
−d
• The best lower bound on θ(d) scales as 2 , and decades of work were only able
to improve it by polynomial or logarithmic factors in d (Section 8.2.2). The best
upper bound instead scales as 2−0.5990d , which leaves an exponentially large range
within which θ(d) can be located (Section 8.2.3). It can be proven that the best
current techniques to obtain upper bounds (i.e., the linear programming method
of Cohn and Elkies) cannot achieve anything better than 2−0.7786d .
• Spontaneous crystallisation upon compression from the liquid phase becomes
increasingly difficult upon increasing d. Current computers cannot observe it for
d > 4. It is therefore unknown whether a stable thermodynamic crystal phase
exists for large enough d. For the same reason, the metastable liquid acquires a
very long lifetime when d increases. It can thus be treated as a stable thermody-
namic phase in computations (Section 8.3.1).
• Algorithmically, the sphere packing problem is complex. Constructing the liquid
phase requires at least N ∼ d d/2 particles, and as a consequence, the natural
equilibration time for the liquid is τliq ∼ N ∼ d d/2 . The crystallisation time
d/2
(provided a stable crystal exists) is likely to scale as τcryst ∼ eN ∼ ed in
large d. Even sampling the restricted class of Bravais lattices in high d has a
poor scaling with d. As a result, Bravais lattices can only be sampled up to
d ∼ 20. This approach thus provides no improvement over the best known
packings (Section 8.3.2).
• The exact solution of the liquid dynamics in d → ∞ discussed in this book
leads to the conjecture that sphere packings exist and can be constructed in
time τliq ∼ d d/2 by slow equilibrium compression of the liquid, up to at least
ϕ=) ϕd d 2−d . The natural value of )ϕd = 4.8067 for hard spheres can be improved
 8.5 Wrap-Up 249

up to )
ϕd = 6.966 by optimising the potential (Section 8.4.1) and up to a higher
(but not precisely known) value, ) ϕd ), by considering out of equilibrium
ϕj ()
compressions via the state following technique (Section 8.4.2).
• The exact calculation of the complexity and of the ideal glass phase thermody-
namics leads to the conjecture that amorphous sphere packings exist up to ϕGCP =
d log d 2−d . However, their construction requires a time τact ∼ ed , which is
d/2

practically impossible to achieve even in low d. The density ϕGCP thus provides a
limit of existence of packings that are structurally similar to a liquid. Beyond
this density, if packings still exist, they likely have a very different structure
(Section 8.4.3).

8.5.2 Further Reading

We provide here a list of references that can be consulted to further explore the
subjects discussed in this chapter, selected according to the criteria discussed in
Section 1.6.2.
The literature on the sphere packing problem is extremely vast and scattered
across many disciplines. The reader interested in the problem can start from reviews
of mathematical results in

• Rogers, Packing and covering [306]

• Conway and Sloane, Sphere packings, lattices and groups [103]
• Cohn, Packing, coding, and ground states [98]
while the reviews

• Torquato and Stillinger, Jammed hard-particle packings: From Kepler to Bernal

and beyond [344]
• Baule, Morone, Herrmann et al., Edwards statistical mechanics for jammed gran-
ular matter [30]
attempt to connect results from physics to those from the mathematical literature.
A natural extension of the sphere packing problem is the problem of packing
identical non-spherical particles and mixtures of different particles. The literature
on this problem is also vast and partially covered in the previously cited reviews.
Some additional results can be found in

• Yerazunis, Cornell and Wintner, Dense random packing of binary mixtures of

spheres [365]
• Pinson, Zu, You et al., Coordination number of binary mixtures of spheres [295]
• Hopkins, Stillinger and Torquato, Disordered strictly jammed binary sphere pack-
ings attain an anomalously large range of densities [179]
250 Packing Spheres in Large Dimensions

• Donev, Cisse, Sachs et al., Improving the density of jammed disordered packings
using ellipsoids [127]
• Torquato and Jiao, Dense packings of the Platonic and Archimedean solids [345]
• Haji-Akbari, Engel, Keys et al., Disordered, quasicrystalline and crystalline
phases of densely packed tetrahedra [172]
All these problems can be, in principle, tackled by the formalism developed in this
book, as in

• Biazzo, Caltagirone, Parisi et al., Theory of amorphous packings of binary mix-

tures of hard spheres [49]
• Ikeda, Miyazaki, Yoshino et al., Decoupling phenomena and replica symmetry
breaking of binary mixtures [184]
 9
The Jamming Transition

In Chapters 4 and 6, we have shown that a glass of hard spheres, prepared at

equilibrium at scaled packing fraction ) ϕg , can be compressed up to a point ϕ)j ()
ϕg ),
at which pressure diverges and spheres touch their neighbours, thus forming a rigid
contact network. The values of ϕ )j ()
ϕg ) span an interval [) ϕj,K ] upon varying
ϕj,d,)
ϕg ∈ [)
) ϕK ]. In Chapter 7, we have seen that stable amorphous packings can
ϕd,)
be found over an even larger interval, which goes from the threshold ) ϕth up to
the glass close packing density ) ϕGCP (see Figure 7.6). In Chapter 8, we have seen
that periodic and quasi-periodic packings might exist in the same density range as
amorphous packings and even possibly at higher densities. In the limit d → ∞,
however, the crystallisation time diverges, and the amorphous branch of the phase
diagram (liquids and glasses) is then disconnected from the crystalline branch.
Amorphous packings can thus be studied without worrying about crystallisation
or other structural transitions, as long as the amorphous state is locally stable.
The purpose of this chapter is to describe in more detail the properties of such
jammed amorphous packings. We show that if the hard-sphere constraint is relaxed
slightly, thus allowing spheres to overlap, jammed configurations separate mechan-
ically floppy structures from mechanically rigid ones, and the associated jamming
transition is critical. We discuss several protocols that are used to obtain jammed
packings, focusing on the physical observables that characterise their properties.
We show that the Gardner transition and full replica symmetry breaking induce
a scaling behaviour close to the jamming transition and, thus, provide a set of
associated universal critical exponents.

9.1 The Jamming Transition as a Satisfiability Threshold

The packing problem introduced in Chapter 8 consists in finding a configuration
of N spheres of diameter  with centers xi ∈ V, where V is a region of the

251
252 The Jamming Transition

d-dimensional Euclidean space with volume V , such that the spheres do not
overlap.1 Defining the (scaled) gap between two spheres, i and j , as
 
|xi − xj |
hij = d −1 , (9.1)

the packing problem consists in finding a configuration X = {xi }i=1,...,N , such that
hij ≥ 0, ∀ i = j, (9.2)
which provides a set of M = N(N − 1)/2 constraints2 on the variables X (because
hij = hj i ). The packing problem therefore consists in finding a configuration of N
variables that simultaneously satisfies M constraints. Such problems are known as
‘satisfiability’ problems. See [19, 50, 251] for a historical introduction, many exam-
ples and a very complete discussion of the theory and algorithms for satisfiability.
Interestingly, most of the satisfiability problems to which statistical mechanics tools
have been applied involve discrete variables, but the packing problem involves
continuous variables, which brings the criticality of the satisfiability transition
[147, 151].
The simplest class of algorithms to solve satisfiability problems are ‘local search
algorithms’. These algorithms start from a random assignment of the variables.
If this assignment satisfies all constraints, then the algorithm stops. Otherwise,
each variable is updated according to its ‘local’ environment – i.e., the set of vari-
ables involved in the same constraints. While local search algorithms are extremely
simple to implement, they may sometimes fail to find solutions even when such
solutions do exist. In the rest of this section, we give some examples of local search
algorithms for the packing problem and discuss their properties.

9.1.1 Gradient Descent

A classical strategy to solve satisfiability problems is to transform them in opti-
misation problems by introducing a ‘cost function’ associated with the unsatisfied
constraints. In the sphere packing problem, the cost function is simply the interac-
tion energy defined in Eq. (2.1) with a particular choice of pair potential, such that
if and only if two spheres overlap, then the system pays a positive energy cost – i.e.,

 v̄(h) > 0 for h < 0,
V (X) = v̄ hij , with (9.3)
i<j
v̄(h) = 0 for h ≥ 0.

1 In Chapter 8, we discussed the problem when V coincides with the infinite space Rd , but in this chapter, we
also discuss finite size effects.
2 Even if the total number of constraints is M = N (N − 1)/2, in any finite d, the number of neighbours
surrounding each sphere is finite. The effective number of relevant constraints is thus proportional to N .
 9.1 The Jamming Transition as a Satisfiability Threshold 253

Configurations X with V (X) = 0 are solutions to the packing problem, while

configurations with V (X) > 0 contain at least one pair of overlapping spheres.
Having introduced the cost function in Eq. (9.3), finding solutions to the packing
problem amounts to minimising V (X), hoping for a zero-energy outcome. A typical
choice of cost function is the soft repulsive potential
ε
v̄(h) = hα θ(−h). (9.4)
α
This potential is not only convenient for the numerical minimisation of Eq. (9.3)
but also a good model of real materials (see Section 2.3.2).
The simplest initialisation for the system is a random Poisson point process [276,
277], which amounts to choosing xi independently and uniformly at random over V.
Each point xi is then associated to a sphere of diameter , which gives a packing
fraction ϕ = Vd (/2)d N/V . In order to minimise the energy, the configuration is
typically evolved according to a gradient descent dynamics
dxi ∂V  ∂
ζ =− =− v̄ hij . (9.5)
dt ∂xi j (=i)
∂xi

Eq. (9.5) is the zero-noise limit of the overdamped (m = 0) Langevin dynamics of

Eq. (3.4), where the parameter ζ defines the unit of time. This algorithm defines a
local dynamics, because the net force acting on particle i depends only on neigh-
bouring particles – i.e., particles that overlap with particle i. Note that v̄(h) should
be differentiable at least once – i.e., α ≥ 1 in Eq. (9.4) – for Eq. (9.5) to be well
defined. Note also that, although the dynamics in Eq. (9.5) is deterministic, the
result of each energy minimisation is stochastic because it depends on the initial
configuration, which is randomly generated.
Final configurations reached by the gradient descent dynamics for t → ∞ have
either positive or zero energy. The former are deemed ‘overjammed’, while the
latter are ‘unjammed’, in reference to their athermal mechanical behaviour. For a
given packing fraction ϕ, from Ns random initial conditions, one can measure the
number Nj ≤ Ns of overjammed final configurations, and for Ns → ∞, one has
Nj /Ns → fj (ϕ;N), which provides the fraction of overjammed configurations.
A schematic plot of fj (ϕ;N) is presented3 in Figure 9.1. For finite N , fj (ϕ;N)
increases smoothly from zero to one upon increasing ϕ. Upon increasing N , the
curve becomes steeper, and for N → ∞, it converges to a step function. The
condition fj (ϕ;N) = 1/2 defines a finite-size jamming transition ϕj (N), which

√
3 Figure 9.1 is obtained by assuming f (ϕ;N ) = [(ϕ − ϕ (N ))/( 2w(N ))], with (x) given in Eq. (4.83),
j j
w(N ) = 0.05N −1/2 and ϕj (N ) = 0.64 − 0.1N −1/2 , for N = 2n and n = 4,5, . . . ,11. The approximate
parameter values are taken from [24, 277].
254 The Jamming Transition

0.8

fj (ϕ; N )
0.6

0.4

Increasing N
0.2

0
0.57 0.58 0.59 0.6 0.61 0.62 0.63 0.64 0.65 0.66
ϕ

Figure 9.1 A schematic plot3 of the fraction fj (ϕ;N ) of overjammed configura-

tions obtained after energy minimisation from a random initial configuration, as
a function of packing fraction ϕ. The figure is roughly consistent with numerical
data in d = 3 [24, 277]. Qualitatively similar results are obtained in d = 2 [347].
Increasing the system size N , the curves become steeper and approach a step
function. The jamming transition ϕj∞ can be defined as the limit for N → ∞
of the ϕ such that fj (ϕ;N ) = 1/2. Above (below) ϕj∞ , energy minimisation
from an infinite size random initial configuration leads to overjammed (unjammed)
configurations.

converges in the thermodynamic limit to the thermodynamic jamming transition

at packing fraction
ϕj∞ = lim ϕj (N). (9.6)
N→∞

Note that the suffix in ϕj∞ does not here refer to the thermodynamic limit taken in
Eq. (9.6) but to the fact that initial configurations are prepared uniformly at random,
which corresponds to an infinite temperature initial configuration, as further dis-
cussed in Section 9.1.2. Note also that ϕj∞ then depends mostly on dimension d, but
the details of the algorithm – e.g., the value of the exponent α in the potential and
the choice of minimisation dynamics – can play a small role. Numerical simulations
give ϕj∞ (d = 2) 0.84 [347], ϕj∞ (d = 3) 0.64 [276, 277], and ϕj∞ (d) has
been measured up to d = 13 [87]. Correspondingly, the energy density e(ϕ) of the
final configurations, averaged over the initial configurations (here denoted by an
overline),
1
e(ϕ) = lim lim
V (X(t)), (9.7)
N→∞ t→∞ N

is such that e(ϕ) = 0 for ϕ ≤ ϕj∞ and e(ϕ) > 0 for ϕ > ϕj∞ .
 9.1 The Jamming Transition as a Satisfiability Threshold 255

9.1.2 Simulated Annealing

The pure gradient descent algorithm discussed in Section 9.1.1 may get trapped in
local minima with finite energy even when plenty of zero-energy configurations
exist. To find higher-density packings, we thus consider a generalisation of the
gradient descent algorithm [9] similar to the simulated annealing protocol [202].
In this protocol, a white noise term at temperature T (t) is added to Eq. (9.5),
thus recovering the full (overdamped, with m = 0) Langevin dynamics given by
Eq. (3.4), with a time-dependent temperature. The function T (t) can then be opti-
mised to improve the performance of the algorithm. For concreteness, we consider
here the simplest possible choice.
1. The system is kept at finite temperature for a long enough time to reach equilib-
rium at state point (ϕ,T ). The initial configuration for the subsequent dynamics
is thus an equilibrium configuration at (ϕ,T ).
2. From time t = 0, temperature decreases with annealing rate r, such that T (t) =
T (1 − rt), while keeping ϕ constant.
3. When t = 1/r, temperature reaches zero, and for t ≥ 1/r, the dynamics follows
a pure gradient descent as in Eq. (9.5).
Fluctuations due to thermal noise can then overcome local energy barriers and find
configurations with a lower energy [9, 202].
We begin the analysis of this more general protocol by considering the par-
ticular case r = ∞, which corresponds to a gradient descent from an equili-
brated configuration at (ϕ,T ). In Section 9.1.1, we have considered the case of
a uniformly random initial configuration, which corresponds to T = ∞. A sharp
jamming transition at ϕj∞ then separates unjammed and overjammed configurations
for N → ∞. The same happens at any finite temperature [347], thus defining a
jamming transition line ϕj (T ), and its inverse Tj (ϕ), in the (ϕ,T ) phase diagram.
The line Tj (ϕ) could, in principle, be computed from the exact solution of the
out-of-equilibrium dynamics obtained in [3], but such a study has not yet been
attempted. The result for the line Tj (ϕ) is thus only schematically illustrated4 in
Figure 9.2 for harmonic soft spheres in d → ∞. Above the line, configurations

4 The schematic line T ()

j ϕ ) in Figure 9.2 has been obtained by fitting the part of the line obtained via RS state
following in [316] to a second-order polynomial,

ϕ ) = A()
1/Tj () ϕj∞ ) + B()
ϕ−) ϕj∞ )2 .
ϕ−)

ϕj∞ ≈ 5.8 is consistent with a linear extrapolation of 2d ϕj∞ (d)/d to d → ∞ in the soft
The resulting )
harmonic sphere Mari-Kurchan model (unpublished data). The value of 2d ϕj∞ (d)/d for the regular soft
harmonic sphere model [87, 267] can also be extrapolated to the same value when d → ∞ but with much
larger finite d corrections.
256 The Jamming Transition
102
Td
Tj , RS-SF
10 Tj , fit

Unjammed-diffusive
Overjammed-diffusive
1
T

10−1 Overjammed-arrested

10−2

Unjammed-arrested
−3
10
5 10 15 20 25 30

Figure 9.2 Phase diagram of harmonic soft spheres in d → ∞. The dynamical

transition line Td ()ϕ ) (solid line) that separates the diffusive and dynamically
arrested regimes is computed exactly by solving Eq. (3.66) [59, 316]. Energy
minimisation from states above (below) the Tj () ϕ ) line leads to overjammed
(unjammed) configurations. The dashed line is a rough estimate of Tj () ϕ ) from
a RS state following calculation [316], which disappears before crossing Td () ϕ)
because of an unphysical spinodal (Section 6.3). The dotted line extrapolates by
fitting these results4 . An example of state following starting from the black dot is
given in Figure 9.3.

reached by gradient descent from (ϕ,T ) are overjammed, while below the line,
they are unjammed.
ϕ ) crosses the dynamical transition line Td ()
The line Tj () ϕ ), computed for soft
harmonic spheres by solving Eq. (3.66) [59, 316]. Note that Td () ϕ ) → 0 for )ϕ →
ϕd = 4.8067 . . ., where )
) HS HS
ϕd is the dynamical transition point of hard spheres, as
discussed in Section 4.4.1. The intersection of the two lines defines four regions
in Figure 9.2. In the unjammed-diffusive and overjammed-diffusive regions, con-
figurations prepared in equilibrium at ()
ϕ,T ) have diffusive equilibrium dynamics
and are, respectively, unjammed or overjammed upon gradient descent minimi-
sation. In the unjammed-arrested and overjammed-arrested regions, by contrast,
configurations prepared in equilibrium at ()
ϕ,T ) are located inside a glass state that
traps the equilibrium dynamics (see Chapter 4). If the glass state is a simple har-
monic energy basin with a single energy minimum,5 then the annealing dynamics

5 This is the case in the spherical p-spin glass model [65], but essentially all other models have a more
complicated energy landscape.
 9.1 The Jamming Transition as a Satisfiability Threshold 257

starting from () ϕ,T ) should eventually reach the unique energy minimum at T = 0,
independently of the annealing rate r. Under this assumption, in the arrested region
ϕ ) should be independent of the annealing rate and can
of Figure 9.2, the line Tj ()
thus be computed by considering a slow annealing r → 0, which corresponds to
following the glass from an initial state () ϕ,T ) to ()
ϕ,0) within the RS construction
introduced in Chapter 4. This calculation has been performed in [316] and the
result is reported in Figure 9.2. In the vicinity of the dynamical transition, however,
the RS solution undergoes an unphysical spinodal instability (see Chapter 4) and
disappears, impeding the calculation of Tj () ϕ ) within the RS ansatz.
It is important to stress that states prepared in the vicinity of the putative RS
ϕ ) actually undergo a Gardner transition upon cooling [316], beyond which
line Tj ()
the structure of energy basins is complex. As an example, the RS state following
phase diagram of a glass prepared in equilibrium at a state point () ϕg,Tg ) located
inside the overjammed-arrested region is reported in Figure 9.3. Other examples
corresponding to different () ϕg,Tg ) can be found in [316]. This glass state can be
adiabatically followed (r → 0) upon cooling or (de)compression. The endpoint of
a constant-density annealing is an overjammed configuration, with energy density
eg ()
ϕ,0|) ϕg,Tg ) > 0, given by Eq. (4.77). Once this zero-temperature overjammed

0.5 (ϕ g , Tg ) TG

0.4

0.3 RS
T

0.2

0.1
RSB
ϕ j
0.0
0 5 10 15 20 25 30 35 40

Figure 9.3 Adiabatic evolution of the glass prepared at the state point () ϕg ,Tg )
marked by a black point in Figure 9.2, under cooling and (de)compression at
ϕ,T ) [316]. The energy at zero temperature is positive above ϕ
() )j ()
ϕg ,Tg ), and
vanishes below it. Below the Gardner line TG ()ϕ ), the RS solution is unstable.
The arrows mark an adiabatic cooling at constant density, which leads to an over-
jammed final state, followed by decompression towards the jamming transition.
258 The Jamming Transition

configuration is reached, decompressing the system leads to unjamming at a sharp

transition density ϕ)j ()
ϕg,Tg ), which depends on the initial point ()
ϕg,Tg ). However,
a Gardner transition surrounds the jamming point in Figure 9.3 and is generically
found for all initial states [59, 316], indicating that the RS calculation is unstable
around jamming. Crossing the Gardner transition makes the annealing dynamics
ϕ ) in Figure 9.2 then depends on the anneal-
more complex. In particular, the line Tj ()
ing rate even in the arrested region. A more detailed investigation of the RSB phase,
however, remains to be performed. See [59, 316] for additional details.
In summary:

• In the diffusive regime of Figure 9.2, a sharp jamming transition line Tj () ϕ ),

originating from the infinite-temperature jamming transition ) ϕj∞ , separates initial
states ()ϕ,T ) that lead to overjammed configurations from those that lead to
unjammed configurations upon gradient descent energy minimisation (r = ∞).
ϕ ) strongly depends on the annealing rate r. This line
In this regime, the line Tj ()
and its dependence on r should be computed via dynamical methods [3], which
have not yet been attempted.
• In the dynamically arrested regime of Figure 9.2, under the assumption of a
simple harmonic glass basin, the line Tj () ϕ ) does not depend on r and can then
be computed through the RS state following method. However, the RS solution is
unstable in the vicinity of jamming because the energy landscape is then more
complex. A weak r dependence of the line Tj () ϕ ) is expected in the fullRSB
regime, and its computation requires either dynamical methods or solving the
fullRSB equations.
To conclude, we recall that while the dynamical transition is sharp in d → ∞,
ϕ ) is then
in any finite dimension, glass states have a finite lifetime. The line Tj ()
expected to always depend on r and exhibit only a crossover in the vicinity of the
avoided dynamical transition.

9.1.3 Compression of Hard Spheres

The local search algorithms described in Sections 9.1.1 and 9.1.2 seek solutions
of the packing problem by minimising a cost function at constant density. An
alternative approach consists in compressing pure hard spheres – i.e., imposing that
constraints be satisfied at all times. Recall that in the hard-sphere case, the energy is
either zero (for valid configurations) or infinity (in presence of overlaps) and, there-
fore, only the space of zero-energy configurations is accessible. No energy barrier
is then present, but at high density, ‘entropic barriers’ – i.e., narrow bottlenecks in
phase space – can prevent the system from exploring the full phase space, trapping
it into configurations with suboptimal density [215]. Thermal agitation then plays
a similar role as in harmonic soft spheres. It provides the system with a mechanism
to go through entropic bottlenecks and explore wider regions of phase space.
 9.1 The Jamming Transition as a Satisfiability Threshold 259

Hard-sphere configurations are easily constructed in the ideal gas limit

ϕ → 0. In finite-dimensional numerical simulations, the compression algorithm
can be conveniently implemented by progressively increasing the sphere diameter
(t) under periodic boundary conditions [235], which provides a homogeneous
compression mechanism and avoids boundary effects that would be present
if the box volume were reduced instead. The resulting packing fraction, ϕ(t),
and the reduced pressure, p(t), then evolve along the compression, resulting in a
line in the plane (1/p,ϕ) of Figures 6.1 and 7.7.
A protocol similar to that of Section 9.1.2 consists in compressing adiabat-
ically an ideal gas up to density ϕ, thus producing an equilibrium hard-sphere
configuration at ϕ and, from there, compressing at a fast rate6 [278, 279]. The
compression then ends at a jamming density ϕj (ϕ). The value of ϕj (ϕ → 0) is close,
but not identical, to ϕj∞ obtained by gradient descent from random configurations.
A similar protocol consists in using a constant compression rate from ϕ = 0. In
this case, the final jamming density ϕj (r) increases continuously upon decreasing r
[86, 128, 292, 330].
The analysis of this algorithm in d → ∞ follows that of Section 9.1.2.

• In the simplest case, the compression starts from an equilibrium dynamically

arrested configuration at ) ϕg > )ϕd . Then, if the glass basin is simple, a replica
symmetric state following calculation provides the endpoint ϕ )j ()
ϕg ) of the com-
pression, as in Figure 4.4.
• As shown in Figure 6.1, a Gardner transition always happens before jamming
is actually reached. A fullRSB state following calculation then provides the
endpoint ϕ )j ()
ϕg ) of a slow enough compression, such that equilibration inside the
glass basin can be achieved. This requires a rate that vanishes with increasing
system size N . A compression with finite rate would instead fall out of
equilibrium around the Gardner transition and end at a slightly lower jamming
density.
• If the compression is initiated from an equilibrium liquid configuration at ) ϕ<) ϕd ,
or if it is initiated at )ϕ = 0 with finite compression rate, then the analysis of
the algorithm requires solving the out-of-equilibrium dynamical equations [3].
It is not then obvious to compare the results with state following calculations
[214, 216, 263].

9.1.4 Jamming as a Satisfiability Transition with Disorder

The analysis of local search algorithms shows that the jamming transition is char-
acterised by the following properties.

6 The compression rate should remain finite, because hard spheres otherwise get stuck in extremely sub-optimal
configurations. In practice, a wide range of rates produce the same result [278].
260 The Jamming Transition

• It takes place when a local search algorithm ceases to find solutions to the pack-
ing problem. It is therefore an algorithmic satisfiability transition. The jamming
density, however, depends sensitively on the details of the algorithm (cooling or
compression rate, choice of the potential, starting point, etc.) even in the thermo-
dynamic limit.
• It separates unjammed configurations with zero energy from overjammed config-
urations with positive energy.
• Configurations at jamming are disordered because the initial configurations are
disordered, and local search algorithms are typically unable to crystallise the
system. This is precisely the case when d → ∞ and approximately true7 in
finite d, as discussed in Chapter 8. The phase space regions that correspond to
disordered and crystalline configurations are well separated in d → ∞ so that
one can focus on amorphous packings as if none other existed.
• We stress that, as discussed in Chapter 8, disordered packings exist up to
ϕGCP = log d, which is therefore the thermodynamic satisfiability transition.
)
Beyond this point, no solutions exist. However, in general, local search algorithms
)j  log d. There is, therefore, a large range of densities
jam at a finite density ϕ
at which a large number of disordered packings exists, but no local search can
find them. The existence of such an ‘algorithmically hard’ region is well known
in random satisfiability problems [50, 370].
In summary, the jamming point can be seen as the satisfiability threshold of a
constraint satisfaction problem that involves disorder and continuous variables.
This analogy can be further exploited by constructing more general constraint sat-
isfaction problems of the same kind, a research program that was started in [147].
Interested readers can find more details in [138, 150, 151].

9.2 Criticality of Jamming

It has been established that the jamming transition is critical, in the sense that
several physical observables display a power-law singular behaviour close to the
transition [93, 232, 234, 269, 277], and associated critical exponents satisfy scaling
relations [167, 361]. In this section, we describe the critical properties of jam-
ming when it is approached both from the overjammed (positive-energy) and the
unjammed (zero-energy) sides.

9.2.1 Energy, Pressure and Coordination

Recall that physical observables associated to the jamming protocols described in
Section 9.1 are fluctuating quantities because they depend on the initial starting

7 In low dimensions, binary or polydisperse mixtures must sometimes be used to prevent crystallisation.
 9.2 Criticality of Jamming 261

configuration. As discussed in Chapter 4, a double average is then needed, first

over the thermal noise (if present) and then over the starting configuration, which
can be either fully random or thermalised at a given state point. Unless otherwise
specified, to simplify the notation, in the rest of this chapter, physical quantities are
assumed to be averaged over both sources of fluctuations.

Energy, Pressure and Entropy

A first indication of the critical nature of jamming is obtained by looking at the
average energy e(ϕ) of the final configurations reached by jamming algorithms. By
definition, this energy is zero for ϕ ≤ ϕj and positive for ϕ > ϕj ; hence, the function
e(ϕ) must be singular at ϕj . In fact, for the soft harmonic potential in Eq. (9.4) and
for ϕ → ϕj+ , the energy scales as [277]
α
e(ϕ) ∼ ϕ − ϕj . (9.8)

Another interesting observable is pressure. In zero-temperature overjammed con-

figurations, particles are mechanically in contact and thus exert a force on the
boundary of the confining volume. Pressure is thus finite and scales as
de de α−1
P (ϕ) = ρ 2 ∝ ϕ2 ∼ ϕ − ϕj , (9.9)
dρ dϕ
where the last equality holds for ϕ → ϕj+ . In particular, for α = 2, pressure is linear
in the distance from jamming and, thus, provides a convenient control parameter in
the overjammed phase. Unjammed configurations, by contrast, have zero energy
and pressure. Because they are valid hard-sphere configurations, however, adding
thermal noise leads to collisions and, as in hard spheres, makes pressure propor-
tional to temperature [42]. The reduced pressure is then the appropriate observable,
and, for ϕ → ϕj− and T → 0, one finds [128, 292]

βP (ϕ)
p(ϕ) = ∼ (ϕj − ϕ)−1 . (9.10)
ρ
Eqs. (9.9) and (9.10) can be combined into a single scaling relation for the reduced
pressure, which holds over the whole (ϕ,T ) plane in the vicinity of jamming. For
ϕ ∼ ϕj and T ∼ 0 [43, 123, 182],
 
p(ϕ,T ) = T −1/α P T −1/α (ϕ − ϕj ) . (9.11)

In order to match Eqs. (9.9) and (9.10), the asymptotic behaviour of the scaling
function must be

|x|−1 for x → −∞,
P(x) ∼ (9.12)
x α−1 for x → ∞.
262 The Jamming Transition

Similar scaling relations can be obtained for other observables, such as the energy
[43, 123, 182].
Note that on the unjammed (hard-sphere) side of the transition, the reduced
pressure coincides with the derivative of the entropy with respect to density [292].
Hence, for T → 0 and ϕ < ϕj ,
ds
p(ϕ) = −ϕ ⇒ s(ϕ) ∼ log(ϕj − ϕ), ϕ → ϕj− . (9.13)
dϕ
Because the internal entropy is the logarithm of the phase space volume of the
solutions to the packing problem, its divergence towards minus infinity indicates
that the phase space volume shrinks continuously to zero [161]. This critical scaling
of the entropy is a consequence of the continuous nature of the variables involved
in the packing problem – i.e., the sphere positions. It is very different from what
is observed around the satisfiability threshold of discrete constraint satisfaction
problems [370]. These continuous variables are thus responsible for the appearance
of critical scaling near jamming [147].

Coordination
The real-space structure of the configurations around jamming also displays an
interesting criticality. Consider approaching the transition from the overjammed
phase. Some spheres then overlap, and the energy density is positive. Having
defined the gap variables as in Eq. (9.1), contacts correspond to negative gaps. The
number zi of spheres in contact with sphere i is then

zi = θ −hij . (9.14)
j (=i)

Because each contact is shared by two spheres, the total number of contacts is

1
N
Nc = zi . (9.15)
2 i=1

If zi ≥ d + 1 and if the zi neighbouring spheres are not all on the same side
of any hyperplane going through xi (a very unlikely event in random packings),
then sphere i is blocked and belongs to the rigid contact network. Conversely, if
zi ≤ d, the sphere is not blocked and can thus rattle inside the cage formed by
its neighbours. It is then called a ‘rattler’ [128, 344]. The number Nr of rattlers
depends on the packing fraction [279], the protocol used to create jammed pack-
ings and, most strongly, dimension [87]. Numerically, using gradient descent from
random configurations, it is found that the number of rattlers at jamming scales
roughly as Nr /N ∼ exp(−d/dr ) with dr ≈ 2 and, thus, vanishes exponentially for
d → ∞ [87].
 9.2 Criticality of Jamming 263

Being disconnected from the rigid contact network, rattlers are not relevant for
determining the mechanical properties of jammed packings [227, 344] and are thus
typically omitted from the analysis of packings. The contact network is then defined
as the graph having as vertices the spheres with degree zi ≥ d + 1 and as links the
contacts between two such spheres. The number of spheres belonging to the contact
network is Ncn = N − Nr , and the average degree of the contact network is

1 
Ncn
2Nc
z= zi = , (9.16)
Ncn i=1 Ncn

where Nc only counts the particles in the contact network – i.e., rattlers are assigned
zi = 0. Note that z is self-averaging over the ensemble of packings generated
according to a given protocol (see Section 9.1) – i.e., it is equal to its average value
with probability going to one in the thermodynamic limit Ncn → ∞. At jamming,
it is found that z = 2d [128, 234, 277, 344], and in the overjammed phase, z is a
function of the packing fraction. Upon approaching jamming, it scales as [277]
νz
z − 2d ∝ (ϕ − ϕj )νz ∝ P (α−1) , (9.17)
with a non-universal exponent νz that will be further discussed in section 9.2.3.
In the unjammed phase, by contrast, the spheres do not touch. The gaps are all
positive and z = 0. Hence, z jumps from zero to 2d at jamming and then increases
as a power law in the overjammed phase.

9.2.2 Isostaticity
The property that z = 2d at jamming is called ‘isostaticity’. In this section, we
explain the origin of this terminology and discuss some theoretical arguments that
highlight the special role played by isostaticity in determining the properties of
jamming.

Isostaticity, Hypostaticity and Pre-stress

We begin by reproducing an argument originally developed in [359, 360] to explain
why sphere packings are isostatic. The argument generalises a simple counting
argument due to Maxwell [248]. We focus for simplicity on a harmonic potential –
i.e., we fix α = 2 in Eq. (9.4) – but the argument can be easily extended to more
general potentials [91, 122]. We denote C = {ij  : hij < 0} the set of contacts,
whose cardinality |C| = Nc is the total number of negative gaps. Defining a contact
index c = ij  ∈ C, the energy is then
ε
V (X) = hc (X)2 . (9.18)
2
c∈C
264 The Jamming Transition

The argument discussed in this section applies to a generic configuration X

with Ndof degrees of freedom, and for arbitrary gap functions hc (X). The particular
case of spheres is recovered when Ndof = dNcn , and the gap is given in Eq. (9.1),
but the argument also applies to particles of arbitrary shape (e.g., ellipsoids)
[70, 318].
Let us consider an overjammed configuration X at ϕ > ϕj . By definition, such a
configuration is a local minimum of the potential energy, hence ∂V /∂X = 0, with
positive energy, V (X) > 0. If the configuration is perturbed, X → X + δX, the
energy variation is then controlled by the Ndof × Ndof Hessian matrix,

∂ 2 V (X)
H= . (9.19)
∂X ∂X

At order |δX|2 , the energy variation can conveniently be written as

1
δV (X) = V (X + δX) − V (X) ≈ δX · H · δX (9.20)
2
 
ε ∂ 2 hc ε ∂hc 2
=− |hc | δX · · δX + δX · .
2 ∂X∂X 2 ∂X
c∈C c∈C
B CD E B CD E
pre-stress harmonic

Following [359, 360], the energy variation has been split in two contributions.

• The harmonic term is a semidefinite positive matrix, which is non-zero only if

at least a contact c is such that δX · ∂h∂X
c
= 0. Each contact thus stabilises the
direction parallel to its gradient, by providing a positive energy variation in that
direction. Note that if there are Nc contacts, under the assumption that the vectors
∂hc
∂X
are linearly independent, the harmonic term has max [Ndof − Nc,0] directions
δX along which it vanishes – i.e., zero modes.
• The pre-stress term is a sum of terms proportional to |hc | over the negative gaps,
which vanish at jamming. Note that for ϕ → ϕj+ , energy and pressure scale as

1  2 de 1 
e(ϕ) ∝ hc , P (ϕ) ∝ ∝ |hc |. (9.21)
Nc d Nc
c∈C c∈C

In other words, energy is proportional to the average squared negative gap, while
pressure is proportional to the average negative gap. Upon approaching jamming,
the pre-stress term vanishes proportionally to pressure, while the harmonic term
stays finite.

The pre-stress term determines the properties of the jamming transition [359, 360].
There are two possible situations.
 9.2 Criticality of Jamming 265
∂2h
1. If the matrix ∂X∂X c
is positive definite for each contact c ∈ C, then the pre-
stress term is negative definite. By definition, the configuration X is an energy
minimum in the overjammed phase, and the total Hessian matrix must then be
positive definite. Suppose now that Nc < Ndof . The harmonic term then has
Ndof − Nc > 0 zero modes, and the pre-stress term is generically negative
on these modes. The total Hessian matrix is then negative along these modes,
which contradicts the initial assumption that X is a local energy minimum. The
hypothesis Nc < Ndof must thus be rejected. One concludes that, if the pre-
stress term is negative definite, one must necessarily have Nc ≥ Ndof in the
overjammed phase. By continuity, this condition must also hold at jamming.
Note that in this case, all the eigenvalues of the Hessian matrix remain finite
at jamming. If the bound is saturated, Nc = Ndof , such a system is said to be
isostatic.
∂ 2 hc
2. If instead the matrix ∂X∂X is not positive definite, then the pre-stress term can
stabilise some of the zero modes of the harmonic term. In this case, some eigen-
values of the Hessian matrix (those associated with the harmonic term) remain
finite at jamming, while others vanish proportionally to P . The number of con-
tacts is then unconstrained, and, in particular, it can be Nc < Ndof . Its precise
value is system dependent; it is fixed by the properties of the pre-stress matrix.
Such a system is said to be hypostatic.
The critical properties of jamming are very different in the isostatic and hypostatic
case. Spheres belong to the first category, as we show next, but for more general
potentials, jamming can be hypostatic. A notable example of the latter case are
ellipsoids [70, 127, 129, 318]. In the following, we restrict our discussion to the
isostatic case.

Application to Soft Harmonic Spheres

We here show that for soft harmonic spheres the pre-stress term is always nega-
tive [359, 360]. Suppose for simplicity that the spheres are within a d-dimensional
cubic volume with periodic boundary conditions.8 Recall that ∂i denotes the set of
spheres in contact with i and rattlers are excluded from the analysis. The number
of degrees of freedom is then Ndof = dNcn . Each contact particle exerts a force on
particle i,
 ∂ v̄(hij )  d dε
Fi = − = fij nj →i , fij = − v̄  (hij ) = |hij |, (9.22)
j ∈∂i
∂xi j ∈∂i
 

8 The general case, in which walls and external forces are present, is discussed in [91].
266 The Jamming Transition

where fij is the modulus of the force exchanged by spheres i and j , and
xi − xj
nj →i = , (9.23)
|xi − xj |
is a d-dimensional unit vector pointing from the center of sphere j towards the
center of sphere i.
In a local minimum of the potential energy, forces vanish, F i = 0, ∀i =
1, . . . ,Ncn . Writing explicitly the second-order variation of the energy, Eq. (9.20),
in terms of particle displacements, one obtains [359, 360]
εd  |δrij |2 − (δrij · nj →i )2 εd 2 
δV (X) = − |hij | + 2 (δrij · nj →i )2
2 2rij 2
ij ∈C ij ∈C

εd  |δr⊥
ij |
2
εd 2 
=− |hij | + 2 (δrij · nj →i )2 , (9.24)
2 2rij 2
ij ∈C ij ∈C
B CD E B CD E
pre−stress harmonic

where rij = xi − xj , rij = |rij |, δrij = δxi − δxj and δr⊥

ij is the projection of δrij
on the plane orthogonal to nj →i . For soft harmonic spheres, the pre-stress term is
thus always negative, and, hence, jamming is always isostatic.
It is important to emphasise that, because the system is translationally invariant
(it is confined in a cubic box with periodic boundary conditions), any uniform
translation, δxi = a,∀i, is a zero mode of the Hessian. This is manifest in Eq. (9.24)
because δrij = 0, and then δV (X) = 0 for such a mode. Once the d zero modes
associated to translations are excluded, the effective number of degrees of freedom
is reduced by d, and the Hessian is effectively a d(Ncn − 1) × d(Ncn − 1) matrix.
The stability condition then becomes Nc ≥ d(Ncn − 1). Using Eq. (9.16), in the
thermodynamic limit Ncn → ∞, this condition corresponds to z ≥ 2d, as numeri-
cally observed [277, 359, 360]. In the next section, we discuss more precisely the
isostatic value of Nc for a finite system.

Determination of the Contact Forces at Isostatic Jamming

For a given contact network entirely specified by the contacts ij  and the contact
vectors nj →i , the force balance equations read

fij nj →i = 0, (9.25)
j ∈∂i

where 0 is the null vector. Eq. (9.25) can be thought of as a homogeneous linear
system for the forces fij . Obviously, if the potential v̄(hij ) is known, the contact
forces are then functions of the particle positions as in Eq. (9.22), but one can
 9.2 Criticality of Jamming 267

wonder if Eq. (9.25) alone suffices to determine these forces. It turns out that if
jamming is isostatic, then the contact forces fij are entirely determined by the
contact network, independently of the interaction potential.
The system in Eq. (9.25) contains dNcn equations – i.e., one for each spatial
coordinate of each particle in the contact network – and Nc unknowns – i.e., one
force per contact. Because fij = fj i , and ni→j = −nj →i , one has


Ncn 
fij nj →i = 0, (9.26)
i=1 j ∈∂i

because the terms ij and j i cancel. This global force balance follows from the
invariance of the potential energy under a global translation of the system. One of
the vectorial equations in Eq. (9.25), corresponding to d scalar equations, is thus
linearly dependent on the others. In presence of additional symmetries – e.g., in
crystals – there might be additional linear dependencies, but for disordered
configurations, it is reasonable to assume that the remaining d(Ncn − 1) equations
are linearly independent. The linear system in Eq. (9.25) thus admits
max[0,Nc − d(Ncn − 1)] linearly independent solutions. Note that for Nc
= d(Ncn − 1), which is the minimal number of contacts allowed by the stability
condition on the Hessian, the force equation has no solution. Thus, this value must
be excluded. The minimal value of Nc for a stable finite system is then [91, 227]

Nc = d(Ncn − 1) + 1 = Niso, (9.27)

which defines the isostaticity condition in finite systems. This condition still cor-
responds to z = 2d in the thermodynamic limit. At isostaticity, Eq. (9.25) has a
unique solution for which the forces are not identically zero; the forces are therefore
determined by this solution. If the system is hyperstatic, Nc > Niso , then Eq. (9.25)
has multiple solutions and the forces have to be determined from the interaction
potential.
Note that Eq. (9.25) does not fix the overall normalisation of contact forces.
If an isostatic jammed configuration is reached from above using a soft harmonic
potential, then fij = d|hij |/, and the average force is proportional to pressure
P , which vanishes at jamming. By contrast, if jamming is reached from below by
compressing hard spheres, then the contact forces can be defined as the average
momentum transfer due to collisions between particles ij [128], and the average
force diverges proportionally to the reduced pressure p at jamming. If forces are
scaled by requiring that their average is unity,
1 
fij = 1, (9.28)
Nc
ij ∈C
268 The Jamming Transition

then one concludes that the forces fij associated with an isostatic jammed configu-
ration are independent of the potential and the protocol that were used to prepare it.
They are then given by the unique solution of Eq. (9.25) normalised as in Eq. (9.28)
and depend only on the geometry of the contact network.

9.2.3 Critical Exponents of Isostatic Jamming

The jamming transition, of both isostatic and hypostatic nature, is characterised by a
critical scaling of the energy, pressure and average coordination (Section 9.2.1). The
associated critical exponents for pressure and energy are simple functions of the
exponent α of the interaction potential. The exponent for coordination is νz = 1/2
(isostatic) [277, 359, 360] or νz = 1 (hypostatic) [70]. Isostatic jammed packings
display additional universal critical exponents that are not simple rational numbers.
We now describe these additional exponents.

Contact Force Distribution

The probability density of contact forces is

1 
P (f ) = δ f − fij (9.29)
Nc
ij ∈C

and, following Eq. (9.28), is normalised by

 ∞  ∞
df P (f ) = 1, df f P (f ) = 1. (9.30)
0 0

This distribution can be numerically measured either by constructing the contact

network at jamming and then solving the linear system in Eq. (9.25) or by directly
measuring the forces at ϕ > ϕj and then taking the limit ϕ → ϕj+ (the former strat-
egy is more numerically accurate) [91]. It turns out that at the jamming transition
[91, 227]

P (f ) ∼ f θ , f → 0+, (9.31)

which defines a critical exponent θ . This ‘pseudo-gap’ in the force distribution

affects broadly the mechanical stability of packings at jamming, which is controlled
by sphere contacts carrying an extremely small force [269, 361]. If the packing is
perturbed, these small-force contacts are indeed likely destabilised.

Gap Distribution
The radial distribution function of a jammed packing can be computed from
Eq. (2.30),
 9.2 Criticality of Jamming 269

1 
g(r) = δ r − xi + xj . (9.32)
ρNcn i=j

Recall that rattlers are excluded from the analysis [87, 128], so the sums over i = j
in Eq. (9.32) only run over the spheres that belong to the contact network. The
cumulative structure function,
 r
Z(r) = ρd ds s d−1 g(s), (9.33)
0

then gives the number of particle centers contained within a ball of radius r centered
on a reference particle. At jamming, the gap constraint implies Z(r) = 0 for r < .
In the jamming limit, isostaticity also implies that

lim Z(r) = 2d . (9.34)

r→+

Hence, at jamming, Z(r) jumps from 0 to 2d in r = , and g(r) has a delta peak in
r = . It is also found that [87, 128, 361]

g(r) ∼ (r − )−γ , Z(r) − 2d ∼ (r − )1−γ , r → +, (9.35)

which defines a new critical exponent γ . Therefore, for h = d(r/ − 1) → 0+ ,

the gap distribution g(h) = g[(1 + h/d)] diverges as g(h) ∼ h−γ . Isostatic
jammed configurations thus display a large number of very small gaps. If the pack-
ing is perturbed, new contacts between almost touching particles are likely to form
[269, 361].

Mean Square Displacement of the Hard-Sphere Glass

In a hard-sphere glass, the long time limit of the MSD, , is finite and continuously
decreases upon compression, as discussed in Section 4.4.1. In a replica symmetric
glass phase,  has a unique value, while in a RSB phase, a function (x) describes
the different MSD plateaus corresponding to the exploration of different levels in
the hierarchy of sub-basins (see Chapter 5). The smallest of them, M = (1),
corresponds to the MSD plateau inside an individual glass state. In the jamming
limit, each glass state becomes a jammed configuration; hence, M must vanish
when the reduced pressure diverges, p → ∞. It is found that [89, 93, 122, 269]

M ∼ p−κ , p → ∞. (9.36)

In other words, upon approaching jamming, the cage size in which hard spheres are
trapped shrinks to zero as a power law of reduced pressure, which defines a third
critical exponent κ.
270 The Jamming Transition

Critical Exponents in Finite Dimensions

The critical exponents γ and κ have been measured in numerical simulations and
appear to be independent of dimension, within numerical accuracy [87, 89, 128,
330]. The exponent θ , however, depends on dimension. The underlying reason has
been clarified in [91, 227]. In finite dimension, there is a finite probability that a
sphere has zi = d + 1 contacts, and there is also a finite probability that, among
the d + 1 forces that stabilise such particle, d are almost coplanar. The extra force
can then be anomalously small, thus allowing these particles to easily buckle in
and out of this plane [91, 227]. This buckling mechanism produces an anomalous
abundance of small forces. For simplicity, all particles having zi = d + 1 are called
‘bucklers’, and the probability distribution of contact forces is then separated in the
contribution coming from bucklers and from the rest of the system [91],
Pbucklers (f ) ∼ f θl , f → 0+,
(9.37)
Pnon−bucklers (f ) ∼ f θe , f → 0+ .
Denoting the fraction of bucklers as fb = Nb /N, the total probability distribution
of all forces is

P (f ) = fb Pbucklers (f ) + (1 − fb )Pnon−bucklers (f ). (9.38)

Numerical simulations in finite dimensions found that θl < θe [91, 227]. Hence,
bucklers dominate and P (f ) ∼ f θl , as long as
  θ −θ
1
fb e l
fb f  (1 − fb )f
θl θe
⇔ f  . (9.39)
1 − fb
Because fb is small, numerical simulations may observe a value of θ intermediate
between θl and θe , unless small enough values of f can be probed. And because the
smallest observable force decreases upon increasing the system size, large systems
are needed to reach the asymptotic regime.
For large d, the density of bucklers goes to zero exponentially – i.e., fb ∼
exp (−d/db ) with db ∼ 2.0 – as does the density of rattlers [91]. In the d → ∞
limit, the probability distribution of zi in fact strongly concentrates around the
average value, z = 2d, and the probability of observing any deviation from this
value goes to zero exponentially [87]. The contribution of bucklers to the total force
distribution hence vanishes, and

P (f ) = Pnon−bucklers (f ) ∼ f θe , f → 0+, d → ∞. (9.40)

In infinite dimensions, therefore, θ = θe . Numerical measurements further sug-

gest, within current numerical precision, that both θl and θe are independent of
dimension [91].
 9.2 Criticality of Jamming 271

i j fij
fij fij

Figure 9.4 A schematic illustration of a dipolar force applied to two spheres i and
j in a jammed packing.

9.2.4 Marginal Stability and Scaling Relations

For a system with negative definite pre-stress, such as soft harmonic spheres, iso-
staticity at jamming implies that the system is close to a mechanical instability.
To analyse this instability, we reproduce here briefly an argument originally devel-
oped by Wyart [227, 361]. For simplicity, we assume9 that bucklers are absent and
θ = θe , which is correct in d → ∞. The argument is based on

• Identifying the elementary excitations of the system

• Selecting those with smallest energy cost
• Perturbing the system along those excitations and looking for stabilisation
mechanisms
This procedure is standard to assess the stability of many-body low-temperature
systems in condensed matter. Similar approaches have been successfully applied
to a variety of systems ranging from spin glasses to electron glasses and low-
temperature particle glasses [269]. When applied to isostatic jammed packings, it
provides a scaling relation between the critical exponents γ and θ .
Let us consider an isostatic jammed packing with Nc = (Ncn − 1)d + 1 = Niso
total contacts. As discussed in Section 9.2.2, this is the minimal requirement for
mechanical stability in isostatic systems, in which the pre-stress term is negative
definite. Imagine applying a dipolar force on a pair of spheres in contact, as illus-
trated in Figure 9.4. It can be shown that if Nc ≥ Niso , then this dipolar force
induces a well-defined response δxi of the particles in the packing, which can be
calculated in terms of the Hessian matrix [361] (see [91] for a detailed discussion).
Wyart’s argument assumes that these are the elementary excitations of a jammed
packing [361].

9 The argument can be extended to take into account bucklers, which then give rise to localised excitations [227].
272 The Jamming Transition

We then study the stability of the lowest-energy dipolar excitations. Consider a

system kept at constant pressure P . The work done to open a contact carrying a
force f by a distance s > 0 is
P δV (s) f s − ANs 2, (9.41)
where δV (s) is the variation of the volume V due to this perturbation, and the
constant A is proportional to the average force f  [361]. This work is minimal
when the contact carries the smallest force. In a packing of size N , according to
extreme value statistics,10 if P (f ) ∼ f θ , then the smallest force scales as
1
fmin ∼ N − 1+θ . (9.42)
With this choice of f , δV (s) becomes negative for s > smin , with
2+θ
smin ∼ N − 1+θ . (9.43)
The packing is then destabilised in favour of a denser configuration. However,
in the process of opening this contact, a new contact could be formed elsewhere
in the system. If this happens before s reaches smin , then the packing cannot be
destabilised. Because a new contact is likely formed by closing a very small gap,
we consider the scaling of the smallest gap between two spheres,
1
hmin ∼ N − 1−γ . (9.44)
If we assume that contacts close proportionally to s when the packing is perturbed,
then the stability condition becomes [361]
1
hmin ≤ smin ⇔ γ ≥ . (9.45)
2+θ
Numerical simulations show that the bound in Eq. (9.45) is saturated [91, 227],
which provides a scaling relation γ = 1/(2+θ) between γ and θ , and suggests that
jammed amorphous packings are indeed very close to a mechanical instability – i.e.,
they are marginally stable. Other scaling relations between the critical exponents
can be obtained with similar arguments [122, 269]. In particular,
2
κ =2− ; (9.46)
3+θ
hence, a single exponent – e.g., θ – remains undetermined.


10 The probability to extract from P (f ) a force smaller than f is f df  P (f  ) ∼ f 1+θ . If N forces are
0
extracted independently, this probability is ≈ Nf 1+θ , which is of order unity if Eq. (9.42) holds.
 9.3 The Unjammed Phase: Hard Spheres 273

9.3 The Unjammed Phase: Hard Spheres

When hard spheres are compressed towards the jamming point, they undergo a
Gardner transition. At this point, the glass basin continuously fractures into a large
number of hierarchically organised sub-basins (Chapter 6). The fullRSB phase
that appears beyond the Gardner transition is characterised by critical long-range
correlations and marginal stability, in the sense of a diverging susceptibility (Sec-
tion 5.5.3). The equation of state of these fullRSB hard-sphere glasses terminates
at the jamming transition, which is also a critical point characterised by mechan-
ical marginal stability (Section 9.2.4). The relation between these two notions of
marginal stability is not a priori obvious. In this section, we show that at least in
d → ∞, they are, in fact, deeply related. The criticality of the jamming transi-
tion is indeed described by the fullRSB equations, which provide exact analytical
results for the critical exponents. Isostaticity and the scaling relations in Eqs. (9.45)
and (9.46) then emerge as direct consequences of the marginal stability of the
fullRSB solution in the jamming limit.

9.3.1 FullRSB Equations for Hard Spheres

We begin by considering hard-sphere glasses that are adiabatically compressed
towards jamming, as in Chapter 6. The fullRSB function (x) is constant outside
the interval defined by the two breaking points, xm and xM , as shown in Figure 6.2.
The plateau value, M , for x > xM represents the long time limit of the MSD
within an individual glass state, while the function (x) for x < xM describes the
MSD associated to the exploration of different states in the free energy landscape,
as discussed in Chapter 5. The plateau value m for x < xm encodes the largest
possible MSD within a same glass metabasin and, thus, captures the width of that
metabasin. At jamming, the cage of individual glass states vanishes – i.e., M → 0.
Recall that, as described in Chapter 6, a slow compression (at finite rate) of
a glass state selected by an equilibrium liquid configuration at ) ϕg is unable to
follow the glass adiabatically up to the Gardner transition and beyond because
the equilibration time scale of the metabasin diverges at ) ϕG ()ϕg ). Any finite-rate
compression would then explore the glass metabasin out of equilibrium. The value
of M ()ϕ ) computed using the state following method is thus only an approximation
of what would be obtained during a finite-rate compression. The critical exponents
obtained for jamming nonetheless compare very well with those observed by the
protocols described in Section 9.1, which justifies the approximation a posteriori.
In order to characterise the behaviour of the fullRSB state following solution
in the limit M → 0, it is convenient to write the fullRSB equations in a
different form. For x ∈ [xm,xM ], (x) is monotonously decreasing and can be
274 The Jamming Transition

inverted. Its inverse function x() is defined over the interval  ∈ [M ,m ]; see
Section 5.3.2 and Figure 5.6. Expressing λ(x) as a function of , λ() = λ(x()),
gives
 
λ() = M + d x( ), x() = λ̇(), (9.47)
M

where the dot here denotes a derivative with respect to . Introducing11

f (,h) = f (x(),h + η − r ), P (,h) = P (x(),h + η − r ), (9.48)
the set of fullRSB Eqs. (6.9), (6.11) and (6.13) derived in Chapter 6 can be
rewritten as

f (M ,h) = log γM  θ(h),
f˙(,h) = 12 f  (,h) + x()f  (,h)2 ,
⎧ (9.49)
⎨P (m,h) = )
ϕg γ2r −m  eh+η−r θ(h + η − r ),
 
⎩Ṗ (,h) 
= − 12 P  (,h) − 2x() P (,h)f  (,h) ,

where the differential equations are defined over the interval  ∈ [M ,m ], and

1 1 ∞
=− dh P (m,h)[f  (m,h) + f  (m,h)],
λ(m ) 2 −∞
   (9.50)
2r − m  1 1 ∞ 
− d = dhP (,h)f (,h) . 2
λ(m )2 m λ( )2 2 −∞
Restricting the analysis to hard spheres entails replacing e−β v̄(h) = θ(h) in the
expressions for f (M ,h) and P (m,h) in Eq. (9.49).
The jamming limit corresponds to M → 0, while r and the function (x)
for x < xM remain finite and positive. In the rest of this section, the scaling of
Eqs. (9.49) in this limit is derived and compared to a direct numerical resolution of
the fullRSB equations, following the procedure of Section 6.6 with a finite number
k of RSB steps [298]. It has been checked [88, 298] that upon increasing k, the
curves become smoother and converge towards the continuum fullRSB limit.

9.3.2 Jamming Scaling Ansatz for Hard Spheres

The fullRSB equations depend on )ϕ – or, equivalently, on η = log() ϕg ) – as a con-
ϕ /)
)j or ηj is then the point at which M ()
trol parameter. The jamming density ϕ ϕ ) → 0.

11 For later convenience, we have shifted the variable h → h + η −  . Note that this shift only explicitly enters
r
in the form of the boundary functions f (M ,h) and P (m,h).
 9.3 The Unjammed Phase: Hard Spheres 275

ΔM
1/p

η p

Figure 9.5 State following results for a hard-sphere glass prepared at ) ϕg = 7

obtained by numerical integration of the kRSB equations with k = 99.
(Left) Inverse reduced pressure 1/) p = d/p as a function of η = log() ϕ /)
ϕg ).
p = A(ηj − η) with
Circles are numerical results; the dashed line is a fit to 1/)
ηj = 0.2565 and A = 1.0767. (Right) Mean square displacement plateau M
as a function of p ). Circles are numerical results, the dashed line is a fit to
M = j /) p κ with the exponent κ = 1.41574 fixed to the analytically predicted
value (Section 9.3.3), and j = 1.3766.

Recall that all the results also depend on )ϕg , as discussed in Chapter 6, but for the
rest of this chapter, we will not indicate this dependence explicitly. The reduced
pressure can be computed from Eq. (6.17), and, as shown in Figure 9.5, its inverse
vanishes linearly when η → ηj− , consistently with numerical simulations in finite
dimensions [86, 128, 330]. It is then convenient to use 1/) p = d/p as a control
parameter because it encodes the linear distance from jamming. Note, however, that
the scaling 1/)p ∝ ηj − η cannot be proven analytically from the scaling analysis. It
has to be inferred from the numerical solution of the fullRSB equations (Figure 9.5).
For the scaling analysis, it is convenient to introduce

y() = x())
p,
)
λ() = λ())
p,
(9.51)
f)(,h) = f (,h)/)
p,
λ()f) (,h) = λ()f  (,h).
m(,h) = )
Numerical inspection of the fullRSB solution shows that these functions develop a
scaling regime close to jamming [88, 151, 298]. In Figure 9.5, M shows a power-
law behaviour as a function of p), with exponent κ as in Eq. (9.36). If the breaking
−κ
point xM remains finite, then yM = xM p) diverges linearly, while M ∼ p )−κ ∼ yM .
This observation suggests a power-law scaling of y(),

y() ∼ yj −1/κ ,  → 0+, (9.52)

276 The Jamming Transition

η = 0.253, 0.254, 0.255, 0.256, 0.2561, η = 0.253, 0.254, 0.255, 0.256, 0.2561,
0.2562, 0.2563, 0.2564 0.25562, 0.2563, 0.2564
Δ

Δ
x xp

Figure 9.6 State following results for a hard-sphere glass prepared at )ϕg = 7
obtained by numerically integrating the kRSB equations with k = 99. The small
discontinuities disappear in the limit k → ∞ [88]. (Left) The function (x)
for several values of η = log ) ϕg . See Figure 6.2 for lower values of η.
ϕ /)
(Right) Same results, but as a function of y = x p ). In this case the collapse
of (y) to a master curve for finite y is clearly observed, and the cutoff on y
). The dashed line gives the asymptotic behaviour of
diverges proportionally to p
(y) ∼ y −κ at large y.

with some constant yj , as confirmed numerically in Figure 9.6. In the jamming

limit, f (,h) and P (,h) also develop a scaling regime for  → 0. The following
scaling is then conjectured [88, 151, 298]:
⎧ 1−κ ' 1−κ (
⎪
κ−1
⎪  κ p− h κ for h ∼ − κ ,
⎨ ' (
P (,h) ∼ − κ p0 h− 2
a 1 1
for |h| ∼  2 ,
⎪
⎪
⎩ 1 (9.53)
p+ (h) for h   2 ,
1
' 1
(
m(,h) = − 2 M h− 2 .

The scaling functions p− (t), p0 (t), p+ (t) and M(t) are yet to be determined, and
a is a new critical exponent. Rather than deriving this scaling ansatz, we show
here that it provides a consistent scaling description of the fullRSB solution and
agrees with numerical results. More technical details on the derivation can be found
in [88, 151].
The asymptotic behaviour of f (M ,h) for M → 0 is
h2
f (M ,h) = log γM  θ(h) ∼ − θ(−h), M → 0, (9.54)
2M
and then, recalling that λ(M ) = M ,

m(M ,h) = λ(M )f  (M ,h) ∼ −hθ(−h), M → 0. (9.55)

 9.3 The Unjammed Phase: Hard Spheres 277

The differential equation for f (,h) implies that a similar equation holds for
m(,h),
1 y()  
ṁ(,h) = m (,h) + m(,h) 1 + m (,h) . (9.56)
2 )
λ()
This equation admits a solution that scales asymptotically as m(,h) ∼ −hθ(−h)
for h → ±∞ for any . Consistency with Eq. (9.53) therefore requires the bound-
ary conditions for the scaling function M(t) to be
M(t → ∞) = 0, M(t → −∞) ∼ t. (9.57)
The scaling functions p− (t) and p+ (t) that appear in Eq. (9.53) for P (,h) can
be understood as follows. For h → ∞, because m(,h → ∞) → 0, the differ-
ential equation for P (,h) reduces to the heat equation Ṗ (,h) = − 12 P  (,h).
The initial condition P (m,h) is a regular, finite function for h > 0. Then, P (,h)
is also expected to be finite for large enough h > 0, which justifies p+ (h). The
behaviour for h → −∞ is slightly more complicated. In this limit,
  
P (,h) ∝ A() exp −A()h2 , (9.58)
which is compatible with the initial condition P (m,h). Plugging this Gaussian
ansatz within the differential equation for P (,h), in the limit h → −∞, one
obtains a differential equation for A(),
y()
Ȧ() = 2A()2 − 2A() . (9.59)
)λ()
)−κ and y() ∼ yj −1/κ , one has
For  → 0, using M ∼ j p
 
)
λ() = p
)M + d y( )
M
. 1−1/κ / (9.60)
yj κj yj κ 1−1/κ
∼ j − )1−κ +
p  .
κ −1 κ −1

)−κ , the second term in Eq. (9.60) dominates,

In the regime where   M ∼ p
and, thus,
y() κ −1 1
∼ ,  → 0 and   M . (9.61)
)
λ() κ 
This scaling behaviour is confirmed by the numerical solution of the fullRSB
equations, as shown in Figure 9.7. Substituting Eq. (9.61) into Eq. (9.59), the
term 2A()2 is negligible provided κ < 2 (which will be verified a posteriori),
leading to
A() ∝ 2(1−κ)/κ ,  → 0 and κ < 2. (9.62)
278 The Jamming Transition

η = 0.253, 0.254, 0.255, 0.256, 0.2561,

0.2562, 0.2563, 0.2564

xΔ/λ

Figure 9.7 The ratio x()/λ() = y()/) λ() for a hard-sphere glass
prepared at )ϕg = 7, obtained numerically integrating the kRSB equations with
k = 99. The dashed horizontal line corresponds to (κ − 1)/κ, and the convergence
to Eq. (9.61) in the region M    m is visible.

Plugging this result into Eq. (9.58) implies that for h → −∞ and  → 0,
P (,h) ∼ (1−κ)/κ p− (h(1−κ)/κ ), with p− (t → −∞) ∝ e−t , which justifies
2

the scaling form of P (,h) for h →√ −∞ in Eq. (9.53).

The intermediate regime |h| ∼  is needed in Eq. (9.53) to match the scaling
behaviour at positive and negative h. The scaling variable in this matching regime
is the same as in the scaling of m(,h); it is the only possible choice that leads to
a non-trivial scaling equation for p0 (t). Matching p+ (t → 0+ ) and p0 (t → ∞)
then gives

2a
p+ (t → 0+ ) ∼ t −γ , p0 (t → ∞) ∼ t −γ ⇒ γ = , (9.63)
κ

while matching p− (t → 0− ) and p0 (t → −∞) gives

1−κ +a
p− (t → 0− ) ∼ |t|θ , p0 (t → −∞) ∼ |t|θ ⇒ θ= . (9.64)
κ/2 − 1

The scaling in Eq. (9.53) is thus compatible with the asymptotic behaviour of all
the functions in the scaling regime, but the exponents a and κ as well as the scaling
functions remain undetermined at this stage.
 9.3 The Unjammed Phase: Hard Spheres 279

9.3.3 Analytical Computation of the Jamming Exponents

Plugging the scaling ansatz of Eq. (9.53) in Eq. (9.56), using Eq. (9.61), one obtains
an equation for M(t), with boundary conditions given in Eq. (9.57),
  
M(t) − tM (t) = M (t) + 2 1−κ κ
M(t) 1 − M (t) ,
(9.65)
M(t → ∞) = 0, M(t → −∞) ∼ t.

For each value of κ, there is a unique solution of this equation that satisfies the
correct boundary conditions.
Plugging the scaling form in Eq. (9.53) inside the differential equation for
P (,h), Eq. (9.49), is not sufficient to obtain closed equations for p− (t) and
p+ (t). These functions are non-universal and depend on the details of the problem
– e.g., the value of )
ϕg [89]. However, the differential equation for P (,h), together
with Eq. (9.61), gives a closed scaling equation for p0 (t),
⎧  
⎨ a p0 (t) + 1 tp (t) = 1 p (t) + κ−1 p0 (t)M(t)  ,
κ 2 0 2 0 κ
(9.66)
⎩p (t → ∞) = t −2a/κ , p0 (t → −∞) = |t|(1−κ+a)/(κ/2−1) .
0

Universality then appears only in the matching regime of Eq. (9.53) for P (,h).
Note that Eq. (9.66) depends on both κ and a, but there is a unique value a(κ)
such that p0 (t) satisfies the boundary conditions at t → ±∞. Hence, for a given κ,
Eqs. (9.65) and (9.66) fix the exponent a(κ) and the scaling functions M(t), p0 (t).
Only the exponent κ thus remains undetermined at this stage.
The additional condition needed to determine the scaling solution of the fullRSB
equations is provided by Eq. (6.23). Using Eq. (6.21), this equation can be
written as
∞  2
x() 1 −∞ dhP (,h)m (,h)
= ∞ . (9.67)
λ() 2 −∞ dhP (,h)m (,h)2 [1 + m (,h)]

Plugging the scaling ansatz Eq. (9.53) and Eq. (9.61) into Eq. (9.67) gives
∞  2
κ −1 1 −∞ dt p0 (t)M (t)
= ∞ , (9.68)
κ 2 −∞ dt p0 (t)M (t)2 [1 − M (t)]

which provides a self-consistent condition for κ. One can then fix κ, solve
Eqs. (9.65) and (9.66) to obtain M(t) and p0 (t) and use Eq. (9.68) to obtain
a new estimate of κ, repeating until convergence with arbitrary precision. This
procedure gives [88]
a = 0.29213 . . . , κ = 1.41574 . . . ,
(9.69)
γ = 0.41269 . . . , θ = 0.42311 . . . .
280 The Jamming Transition

Note that within the reported numerical precision, the results of Eq. (9.69) satisfy
the relation
a = 1 − κ/2. (9.70)
Inserting Eq. (9.70) into Eq. (9.63) and Eq. (9.64), one obtains the scaling relations
between κ, γ and θ given in Eq. (9.45) and Eq. (9.46). While it must be possible
to prove this relation directly from the system of equations that define a and κ –
i.e., Eqs. (9.65), (9.66) and Eq. (9.68) – such a proof has not yet been achieved.
A different argument supporting the validity of Eq. (9.70) has, however, been
obtained by investigating the jamming transition in the perceptron model [151].
We have thus far shown that the scaling of the fullRSB equations upon
approaching jamming reproduces the observed scaling of the reduced pressure,
) ∼ 1/()
p ϕj − )ϕ ), which also implies that the hard-sphere glass entropy diverges
as s ∼ log() ϕj − )ϕ ). The scaling of the mean square displacement, M ∼ p )−κ , is
also reproduced, as shown in Figure 9.5, and the exact expression of κ is obtained
analytically. Using Eq. (9.48), we can furthermore rewrite12 the gap distribution
function given by Eq. (6.19) in terms of P (,h) and f (,h),
 
ϕg eh g(h) = θ(h) γM  e−f (M ,h) P (M ,h) → θ(h) p+ (h).
) (9.71)
M →0

Therefore, Eq. (9.63) implies that g(h) ∼ h for h → 0+ , with the exponent γ
−γ

given by Eq. (9.69), and the fullRSB scaling also reproduces the scaling of the
gap distribution, providing an analytical expression of the associated exponent.
To complete the discussion, the behaviour of the force distribution remains to be
investigated. We refer the reader to [88] for the proof that P (f ) ∼ f θ for hard
spheres. In Section 9.4.2, we provide a much simpler derivation of this result using
soft harmonic spheres.

9.4 The Overjammed Phase: Soft Harmonic Spheres

In this section, we consider the behaviour of soft harmonic spheres, with poten-
tial v̄(h) = (ε/2)h2 θ(−h), in the overjammed phase. This analysis is interesting
because it reproduces the scaling of pressure, energy and coordination discussed in
Section 9.2.1. It is also a much easier model with which to prove isostaticity and
compute P (f ) because the average coordination z is obtained by counting negative
gaps, and, according to Eq. (9.22), forces are simply proportional to the modulus of
negative gaps.

12 To prove Eq. (9.71), it suffices to observe that for h > 0 and  → 0, one has f ( ,h) → 0 according to
M M
Eq. (9.54) and P (M ,h) → p+ (h) according to Eq. (9.53). The convolution with γM then disappears
because M → 0.
 9.4 The Overjammed Phase: Soft Harmonic Spheres 281

The scaled energy )eliq = eliq /d of a liquid of soft harmonic spheres in d → ∞

is given by Eq. (2.78),

)
ϕ 0 ε 2ε
) ϕ,T ) =
eliq () dh eh−β 2 h h2, (9.72)
2 −∞ 2
which vanishes for T → 0 at any finite density ) ϕ . The equilibrium zero-temperature
liquid is therefore unjammed at all ) ϕ , and coincides with an equilibrium liquid of
hard spheres13 at the same ) ϕ . We can thus (1) prepare an initial equilibrium configu-
ration of a soft-harmonic-sphere liquid at () ϕg,Tg = 0), and (2) adiabatically follow
the corresponding glass at state point () ϕ,T ), computing in particular its energy
) ϕ,T |)
eg () ϕg,0) and reduced pressure p )g ()ϕ,T |)
ϕg,0). According to the discussion of
Section 9.2.1, as long as for T → 0 the glass energy vanishes and the reduced
pressure is finite, the glass is unjammed. The soft-sphere glass then coincides with
the hard-sphere glass, and upon compression, the jamming point14 is reached for
)
ϕ→ϕ )j − , at which the glass reduced pressure p )g ()ϕ,0|) ϕg,0) ∼ () ϕ )−1 diverges.
ϕj − )
The corresponding jamming scaling ansatz was studied in Section 9.3. Upon further
compression, for ) ϕ>ϕ )j , the glass enters its overjammed phase, where energy and
pressure are both positive, ) eg () ϕg,0) ∼ ()
ϕ,0|) ϕ−ϕ )j )2 and Pg () ϕg,0) ∼ ()
ϕ,0|) ϕ−ϕ )j ).
In this section, we write the fullRSB equations that describe the overjammed phase
and investigate their scaling behaviour when ) ϕ→ϕ )j + .

9.4.1 Zero-Temperature Limit of the fullRSB Equations

The overjammed phase can be studied by the formalism of Chapter 6, using the soft-
harmonic-sphere potential and setting Tg = 0 and T = 0, with ) ϕ >) ϕj . The first
step consists in setting Tg = 0, which is easily done by replacing e−βg v̄(h) → θ(h)
in Eq. (6.11). Using the modifications discussed in Section 9.3.1, one arrives at the
same Eqs. (9.49) and (9.50), with the only difference that the initial condition for
f (,h) now reads
ε 2
f (M ,h) = log γM  e−β 2 h θ(−h)
, (9.73)
which corresponds to following the glass at state point () ϕ,T ). In the unjammed
phase, )
ϕ<ϕ )j , we know from Section 9.3.1 that M remains finite when T → 0.
ε 2
Hence, we can simply replace e−β 2 h θ(−h) → θ(h) in Eq. (9.73) and the hard-sphere
case is recovered. In the overjammed phase, ) ϕ>ϕ )j , by contrast, the mean square
displacement vanishes – i.e., M → 0 when T → 0 – because the system is in a

13 Recall that a hard-sphere liquid exists up to the Kauzmann density )

ϕK ∼ log d, and a soft-harmonic-sphere
liquid exists at densities below the Kauzmann transition line given by Eq. (7.9).
14 Recall that the value of ϕ
)j depends on )
ϕg , but for notational convenience, we do not here indicate this
dependence explicitly.
282 The Jamming Transition

local minimum of the potential energy. The correct scaling solution is then obtained
by assuming that

M = χ T , T → 0 and ) )j,
ϕ>ϕ (9.74)

which can be justified by harmonic analysis, as discussed in Section 9.4.3.

Because one should take the limit T → 0 and M = χ T → 0 simultaneously,
the function f (M ,h) does not converge to the hard-sphere limit. Similarly to
Eq. (9.51) for hard spheres, for overjammed soft harmonic spheres, it is convenient
to define
y() = x()/M ,
 
)
λ() = λ()/M = 1 + d y( ),
M (9.75)
f)(,h) = M f (,h),
λ()f) (,h) = λ()f  (,h).
m(,h) = )

In the zero temperature limit, M → 0 and

f)(0,h) = lim f)(M = χ T ,h)

T →0
ε 2 χ ε h2 θ(−h) (9.76)
= lim χ T log γχ T  e−β 2 h θ(−h)
=− .
T →0 1 + χε 2
Note that the hard-sphere result, Eq. (9.54), is recovered for χ → ∞. We can then
write Eq. (9.49) in terms of these scaled variables,
⎧
⎨f)(0,h) = − χ ε h2 θ(−h) ,
1+χ ε 2
⎩f)˙ (,h) = 1 f) (,h) + y()f) (,h)2 ,
2
⎧ (9.77)
⎨P (m,h) = ) ϕg γ2r −m  eh+η−r θ(h + η − r ),
 
⎩Ṗ (,h) 
= − 12 P  (,h) − 2y() P (,h)f) (,h) ,

with  ∈ [0,m ]. Eq. (9.50) also becomes


1 1 ∞
=− dh P (m,h)[f) (m,h) + f) (m,h)],
)
λ(m ) 2 −∞
   (9.78)
2r − m 1 1 ∞
− d 
= dhP (,h)f) (,h)2,
)
λ(m )2 m )
λ( )2 2 −∞
which provides a system of equations that describe the zero-temperature over-
jammed phase.
 9.4 The Overjammed Phase: Soft Harmonic Spheres 283

In the limit  → 0, the solution for y() is singular,

y() ∼ yχ −1/2,  → 0, (9.79)
where yχ is a constant that depends on χ . We refer the reader to [151] for a detailed
discussion. Similarly,
  √
)
λ() = 1 + d y( ) ∼ 1 + 2yχ ,  → 0. (9.80)
0

Note that the same scaling is found in some spin glass models [254]. Correspond-
ingly, P ( → 0,h) is finite and smooth around h = 0 [151]. Finally, the parameter
χ is determined by P (0,h) through Eq. (6.21),
   2
1 ∞ 1 0 χε
1= dhP (,h)f) (,h)2 → dhP (0,h) . (9.81)
2 −∞ →0 2 −∞ 1 + χε

9.4.2 Approaching Jamming from the Overjammed Phase

Taking the zero-temperature limit of Eq. (6.19) gives

P (0,h(1 + χε))(1 + χ ε) for h < 0,
ϕg e g(h) =
) h
(9.82)
P (0,h) for h > 0.
Similarly, from Eqs. (6.16) and (6.17), the scaled glass energy ) eg = eg /d and
)g = Pg /(dρ) become, for T → 0,
pressure P
 0
eg ()
ϕ,0|)ϕg,0) ε
)
eg ()ϕ) = = dh P (0,h) h2,
d 4(1 + χ ε)2 −∞
 0 (9.83)
)g () P g ()
ϕ ,0|)
ϕ g ,0) ε
P ϕ) = = dh P (0,h) |h|.
ρd 2(1 + χ ε) −∞
These expressions suggest that upon approaching the jamming transition from the
overjammed side, ) ϕ→ϕ )j + , one has χ → ∞ because energy and pressure must
then vanish. Consistently, the observation that M is finite for )
ϕ<ϕ )j requires that
χ diverges to match the hard-sphere and the soft-harmonic-sphere scalings.
A careful analysis [151] of the fullRSB equations derived in Section 9.4.1 shows
that, for χ → ∞, one has

yχ −1/2,   ∗,
y() ∼ (9.84)
yj −1/κ ,   ∗ .
The crossover point ∗ then scales as
2κ
∗ ∼ χ − κ−2a , yχ ∼ (∗ )−1/2, y(∗ ) ∼ 1/∗ . (9.85)
284 The Jamming Transition

In the jamming limit, ∗ → 0, and, hence, we recover the jamming critical

behaviour of hard spheres, y() ∼ yj −1/κ , over the full range  ∈ [0,m ]. We
refer the reader to [151] for a derivation of these results.
Using the scaling relation a = 1 − κ/2, Eq. (9.70), one obtains χ ∼ (∗ )(1−κ)/κ ,
and observing that in the jamming limit ∗ → 0, one concludes that P (0,h) ∼
P (∗,h) follows the hard-sphere scaling in Eq. (9.53), because ∗ is the boundary
of the hard-sphere regime [151]. Hence,
⎧
⎪
⎪ for h ∼ −χ −1,
(hχ)
⎨χp−1−κ/2 ' κ
( κ
P (0,h) ∼ χ − 1−κ p0 hχ − 2(1−κ) for |h| ∼ χ 2(1−κ) , (9.86)
⎪
⎪
⎩p (h) κ
for |h|  χ 2(1−κ) .
+

Eqs. (9.82) and (9.86) describe the critical behaviour of soft harmonic spheres for
χ → ∞. The following properties are then indeed obtained.
1. Because P (0,h) ∼ χp− (hχ) for h < 0, Eq. (9.83) gives

) ϕ ) ∼ χ −4,
eg () )g ()
P ϕ ) ∼ χ −2, (9.87)

which suggests that

χ ∼ () )j )−1/2,
ϕ−ϕ (9.88)

in order to reproduce the scaling of energy and pressure in Eqs. (9.8) and (9.9),
respectively. Unfortunately, Eq. (9.88) cannot be proven from the scaling analy-
sis. It must instead be obtained from a numerical solution of the fullRSB equa-
tions (similarly to the relation between reduced pressure and density in the case
of hard spheres).
2. Because for h < 0 and χ → ∞, one has

ϕg eh g(h) = P (0,h(1 + χε))(1 + χ ε) = Aχ p− (Aχ h),

) (9.89)

with Aχ = χ (1 + χ ε), the average contact number is

   0  0
z = ρd drg(r) = d) ϕg dh e g(h) = d
h
dh p− (h). (9.90)
0 −∞ −∞

In the limit χ → ∞, Eq. (9.81) becomes

 0
2= dh p− (h), (9.91)
−∞

which implies z = 2d. The solution in the limit d → ∞ thus predicts that
jammed sphere packings are isostatic. It is also possible to show [151] that
 9.4 The Overjammed Phase: Soft Harmonic Spheres 285
! 
z − 2d ∼ 1/χ ∼ )g ∼
P )
ϕ−ϕ
)j, (9.92)

which reproduces Eq. (9.17).

3. For h > 0,

g(h) ∼ P (0,h) ∼ p+ (h) ∼ + h−γ ; (9.93)

h→0

hence, the positive gaps are characterised by the same power-law divergence as
those of hard spheres. The positive part of g(h) is indeed continuous at jamming.
4. For h < 0, Eq. (9.89) provides the scaling of the negative gaps, which are
proportional to the contact forces. Because f ∝ |h|, one has
P (f ) = Bp− (−Bf )/2, which is correctly normalised as a consequence of
Eq. (9.91). Normalising the average force to unity, as in Eq. (9.30), fixes the
constant B and gives
 0
B 1
P (f ) = p− (−Bf ), B= dh p− (h)|h|. (9.94)
2 2 −∞

From these results, it follows that P (f ) ∼ f θ for small forces, with the exponent
θ computed in Section 9.3.3 from the scaling of the fullRSB solution. Note that
because at jamming the forces are entirely determined by the contact network,
this result also holds upon approaching jamming from the hard-sphere side [88].

9.4.3 The Density of Harmonic Vibrations

In the overjammed phase at zero temperature, soft harmonic spheres are in a min-
imum of the potential energy, which makes a harmonic analysis possible. By def-
inition, in an energy minimum, the eigenvalues λi of the Hessian matrix defined
in Eq. (9.19) are all positive, and the frequency associated to each eigenvalue is
√
ωi = λi . Because each rattler gives d trivial zero eigenvalues, rattlers are once
again removed from the analysis. One can then define the density of vibrational
states as

1  cndN
D(ω) = δ(ω − ωi ), (9.95)
dNcn i=1

where the average is over an ensemble of configurations generated by the same

protocol. While an exact calculation of D(ω) has not yet been achieved for soft
harmonic spheres in d → ∞, this quantity has been studied analytically in simpli-
fied infinite-dimensional models such as the perceptron [150] and in overjammed
286 The Jamming Transition

sphere packings using an (approximate) effective medium theory [121]. This anal-
ysis shows that, if phonons are neglected,15 D(ω) has the following general form
⎧
⎪
⎨0 ω∈/ [ω0,ωmax ],
D(ω) ∼ (ω/ω∗ ) ω0  ω  ω∗,
2 (9.96)
⎪
⎩
const. ω∗  ω  ωmax,
which depends on three characteristic frequencies ω0 , ω∗ and ωmax . The two fre-
quencies ω0 and ωmax are the two edges of the vibrational spectrum, while ω∗
is an intermediate frequency scale. Within mean field theory [150], in the over-
jammed replica symmetric phase, ω0 > 0; hence, the density of states is gapped.
No vibrational excitations of arbitrarily small frequency – i.e., soft modes – are then
present. In the overjammed fullRSB phase, instead, ω0 = 0, and D(ω) ∼ (ω/ω∗ )2
down to zero frequency. Upon approaching the jamming point, the intermediate
frequency ω∗ ∼ () ϕ −ϕ)j )1/2 also goes to zero. It follows that D(ω) remains constant
down to zero frequency. The jamming point is then marginally stable also from the
vibrational point of view, and a lot of vibrational modes with small frequency are
observed. The scaling in Eq. (9.96) is also found in numerical simulations [92, 182,
228, 258, 277, 360], which additionally identified a class of localised vibrational
modes that contribute a term Dloc (ω) ∼ (ω/ω∗ )4 to the density of states [228, 258].
These localised modes likely disappear upon increasing dimension, similarly to
rattlers and bucklers. They are therefore unlikely to be present within mean field
theory, but their disappearance has not yet been systematically studied.
To conclude, note that a single harmonic oscillator x(t) of frequency ω has an
equilibrium mean square displacement
 
 = lim (x(t) − x(0))2 ∝ T /ω2 .
t→∞

Harmonic analysis around a glass minimum provides dNcn independent oscillators

with frequencies distributed according to the density of states D(ω), hence [182]
 ∞
 = χT , χ∝ dω D(ω)/ω2, (9.97)
0

which provides a physical argument in support of Eq. (9.74). Note that setting
ω0 = 0 in Eq. (9.96) gives D(ω) = D(ω/ω∗ ), with a scaling function D(x) ∼ x 2
for x  1 and constant otherwise, which implies χ ∼ 1/ω∗ ∼ () ϕ−ϕ )j )−1/2 ,
consistently with Eq. (9.88).

15 Phonons provide a contribution that scales as ωd−1 and can therefore be neglected at low ω in large enough
dimension.
 9.5 Wrap-Up 287

9.5 Wrap-Up
9.5.1 Summary
In this chapter, we have seen that

• The jamming transition ϕj is the satisfiability transition of the packing problem.

It separates unjammed configurations at ϕ < ϕj , which satisfy all the hard-core
constraints, from overjammed configurations at ϕ > ϕj , which contain overlap-
ping spheres (Section 9.1).
• The glass close packing density ϕGCP corresponds to the thermodynamic jam-
ming transition; amorphous unjammed configurations do not exist for ϕ > ϕGCP .
Local search algorithms that look for unjammed configurations, such as gradient
descent from random initial conditions (Section 9.1.1), annealing soft harmonic
spheres (Section 9.1.2) or compressing hard spheres (Section 9.1.3) generically
get stuck at a lower density ϕj < ϕGCP . The jamming point ϕj thus strongly
depends on the details of the preparation algorithm.
• Pressure, energy and coordination are always singular at ϕj (Section 9.2.1), but
the structure of the Hessian matrix determines additional critical properties of
jamming. If the pre-stress term is negative definite, then jamming is isostatic –
e.g., for spherical particles – while otherwise, jamming is hypostatic – e.g., for
ellipsoids (Section 9.2.2).
• Isostatic jammed sphere packings have universal, protocol-independent proper-
ties. They display a power-law distribution of small gaps, g(r) ∼ (r − )−γ ,
and small forces, P (f ) ∼ f θ . Their mean square displacement vanishes as
M ∼ p −κ √ in the unjammed phase, and their average coordination grows as
z − 2d ∼ P in the overjammed phase (Section 9.2.3). A mechanical marginal
stability argument provides scaling relations between the exponents γ , θ , κ
(Section 9.2.4).
• The fullRSB equations that describe adiabatic compression of a hard-sphere
glass (Section 9.3.1) develop a scaling regime close to jamming, ϕ → ϕj−
(Section 9.3.2). An analysis of these equations leads to exact analytical results
for the exponents γ , θ , κ (Section 9.3.3). Within this framework, isostaticity is
a consequence of the marginality of the fullRSB solution, which also fixes the
value of the exponents.
• The fullRSB equations that describe the overjammed phase of a soft-harmonic-
sphere glass (Section 9.4.1) also develop a scaling regime close to jamming,
ϕ → ϕj+ (Section 9.4.2). The analysis of soft harmonic spheres simplifies the
derivation of isostaticity and of the force distribution, P (f ). Moreover, in the
overjammed phase, the density of vibrational states D(ω) can be defined. It is
288 The Jamming Transition

gapless in the fullRSB phase, where it behaves as D(ω) ∼ (ω/ω∗ )2 for ω  ω∗

and it is flat for ω  ω∗ . At jamming, ω∗ vanishes, and D(ω) is finite at zero
frequency. An anomalous abundance of zero-frequency modes is then observed
(Section 9.4.3).

9.5.2 Further Reading

We provide here a list of references that can be consulted to further explore the
subjects discussed in this chapter, selected according to the criteria discussed in
Section 1.6.2.
The jamming transition has been observed in a variety of materials, ranging from
colloids to foams, granulars and emulsions. Complete reviews on the application of
jamming ideas to soft condensed matter can be found in

• Van Hecke, Jamming of soft particles: Geometry, mechanics, scaling and iso-
staticity [348]
• Liu and Nagel, The jamming transition and the marginally jammed solid [234]
• Liu, Nagel, Van Saarloos, et al., The jamming scenario: An introduction and
outlook [232]
A complete review of the marginal stability ideas described in Section 9.2.4 can
be found in Müller and Wyart, Marginal stability in structural, spin, and electron
glasses [269].
A general introduction to satisfiability problems can be found in

• Percus, Istrate and Moore, Computational complexity and statistical physics [294]
• Arora and Barak, Computational complexity: A modern approach [19]
• Biere, Heule and van Maaren (eds), Handbook of satisfiability [50]
In particular, the first reference [294] contains a contribution from Cocco, Monas-
son, Montanari et al., Approximate analysis of search algorithms with ‘physical’
methods, specifically focused on the analysis of search algorithms. The last
reference [50] contains a contribution from Altarelli, Monasson, Semerjian et al.,
A review of the statistical mechanics approach to random optimization problems,
which reviews advanced statistical mechanics methods to compute the satisfiability
threshold in these problems. A discussion on the extension of these results to
continuous satisfiability problems, in relation to jamming, can be found in Franz,
Parisi, Sevelev et al., Universality of the SAT-UNSAT (Jamming) Threshold in
Non-convex continuous constraint satisfaction problems [151].
A very important practical problem is that of understanding the role of friction in
jamming. Friction indeed plays a very important role in selecting the jammed states
of granular materials. Introductory reviews to this topic are
 9.5 Wrap-Up 289

• Liu and Nagel, Jamming and rheology: Constrained dynamics on microscopic

and macroscopic scales [233]
• Bi, Henkes, Daniels et al., The statistical physics of athermal materials [48]
• Behringer and Chakraborty, The physics of jamming for granular materials:
A review [31]
The study of frictional jammed packings requires the introduction of new ideas,
because the Franz–Parisi and Monasson constructions discussed in this book are
based on a Gibbs–Boltzmann approach, which is not appropriate for dissipative sys-
tems. The out-of-equilibrium dynamical equations for infinite-dimensional spheres
could possibly be generalised to frictional systems.
 10
Rheology of the Glass

In Chapters 4, 6 and 9, we investigated the behaviour of glasses adiabatically

followed under (de)compression, cooling or heating. This chapter extends the state
following construction to the response of an amorphous solid to an applied shear
strain, which leads to a mean field theory of the rheology of these materials. Several
physical phenomena ranging from elasticity to dilatancy, yielding and jamming by
shear are described by this approach.
An elastic response to small applied deformations is the hallmark of solidity.
While for perfect crystals, perturbative calculations are possible, for amorphous
solids, the theory is much less developed. We discuss here the static elastic response
and how the solid breaks when subjected to larger deformations. Because a liquid
can adapt to a change of shape of its container, a response to strain is only observed
in the dynamically arrested phase unless dynamics is considered.
Because the topic of this chapter is rapidly evolving, we limit ourselves to a
description of the formalism and of a few selected results. The first steps towards a
theory of the rheology of amorphous solids using the techniques we have described
in this book were taken in [366, 368], while the exact solution for particles in the
limit d → ∞ was derived in [298, 299, 369].

10.1 Perturbing the Glass by a Shear Strain

Within the state following approach described in Chapter 4, a reference replica Y is
prepared in equilibrium at a state point (ϕg,Tg ) in the dynamically arrested region,
which selects a glass, and a second replica X in constrained equilibrium within
this glass is compressed or cooled to a state point (ϕ,T ); see Figure 4.2 for an
illustration. In this section, we generalise this construction to the case in which a
shear strain is also applied to X. If the system is dynamically arrested, it behaves as
an amorphous solid and thus reacts elastically to a small applied strain. The force
(per unit surface) that the system exerts against the deformation is the shear stress

290
 10.1 Perturbing the Glass by a Shear Strain 291

γL

Figure 10.1 A two-dimensional illustration of a shear deformation, with shear

strain γ .

(or simply the stress) [6, 175, 222]. One of the goals of this chapter is to compute
the stress as a function of the strain; the resulting stress–strain curves describe the
rheology of that solid.

10.1.1 State Following with Shear Strain

Following [366, 368], we consider particles confined in a cubic box of linear size
L and volume Ld , and we denote as x = {xμ }μ=1,...,d a Cartesian system of coordi-
nates in the laboratory frame, such that the box sides are parallel to the coordinate
axes. In this reference frame, the box is subjected to a shear strain γ , as illustrated
in Figure 10.1. Note that because the reference replica Y is used to select the glass
in equilibrium, it should not be strained, while the constrained replica X should
be strained. Hence, in the frame x , replica Y is in the original cubic box, while
replica X is in the strained box. Because in the state following construction, it is
not practical to compare replicas in different boxes, for convenience, we introduce
a ‘strained’ frame with coordinates x = {xμ }μ=1,...,d , to bring back the strained box
to a cubic shape. If the box is strained along the first direction μ = 1, then the
transformation is
x1 = x1 + γ x2, xμ = xμ, ∀μ = 2, . . . ,d, (10.1)
and its inverse is
x1 = x1 − γ x2 , xμ = xμ , ∀μ = 2, . . . ,d. (10.2)

To limit the use of indices, we introduce a strain matrix S(γ ) = 1̂ + γ x̂1 x̂T2 , where
1̂ is the d × d identity matrix and x̂μ is the unit vector parallel to the coordinate axis
μ, such that
x = S(γ )x = x + γ x̂1 x2, x = S(−γ )x . (10.3)
292 Rheology of the Glass

Note that S(γ )S(−γ ) = 1̂, which simply means that reversing the applied strain
brings the system back to the unstrained state. The particles of replica X, in the
laboratory frame x , interact via the normal interaction potential v(|x − y |). Hence,
in the strained frame, their pair interaction is v(|S(γ )(x − y)|) [366, 368], which
results in a total potential energy for replica X

V [X,,γ ] = v(|S(γ )(xi − xj )|). (10.4)
i<j

The average free energy of a glass prepared in equilibrium at (ϕg,Tg ) and adi-
abatically followed to (ϕ,T ,γ ) is then simply obtained from Eq. (4.9) by using
Eq. (10.4) as potential for replica X,

T dY
fg (ϕ,T ,γ ;ϕg,Tg, Dr ) = − e−βg V [Y,g ] log Z[ϕ,β,γ ;Y, Dr ],
N Z[ϕg,βg ]

Z[ϕ,β,γ ;Y, Dr ] = dXe−βV [X,,γ ] δ(Dr − D(X,Y )). (10.5)

The average of the logarithm can then be computed by introducing additional repli-
cas, as in Section 4.1.3, where now replica 1 has no shear strain, γ1 = 0, while
replicas a = 2, · · · ,s + 1 have shear strain γa = γ . Note that because the mean
square displacement D(X,Y ) is computed in the strained frame, the ‘affine part’ of
the displacement – i.e., the linear part corresponding to the straining of the box –
is removed, and D(X,Y ) only measures the ‘non-affine’ contribution to the mean
square displacement. Note also that in full thermodynamic equilibrium, the free
energy does not depend on the shape of the box, even for a solid [314]. Hence, the
free energy in Eq. (10.5) only depends on γ because replica X is in a constrained,
metastable equilibrium within the glass state selected by replica Y .

10.1.2 Replicated Free Energy in Infinite Dimensions

In the limit d → ∞, the derivation of Section 4.2 can be followed identically1 up
to Eq. (4.44), which is now replaced by2
 n 
feff (R) = e−βg v̄[d(R/g −1)] e−β a=2 v̄[d(xa /−1)] − 1, (10.6)
with
xa = |S(γa ) R + wa |, wa = u a − va . (10.7)

1 In Section 4.2, rotational invariance was used to obtain, for example, Eq. (4.30). In the presence of a shear
strain, the coordinates μ = 1,2 are special, and rotational invariance only holds in the subspace of d − 2
coordinates μ = 3, · · · ,d. When d → ∞, however, one can show that this anisotropy can be neglected [59].
2 By contrast to Section 4.2, we here explicitly take into account that replica 1 is at state point (ϕ ,T ), a priori
g g
different from the state point of the other replicas.
 10.1 Perturbing the Glass by a Shear Strain 293

From this point on, however, the derivation of Section 4.2 should be adapted
to take the shear strain into account [59]. For d → ∞, one can show that the
term S(γa )wa gives subleading contributions in 1/d and can be neglected. Then
xa ∼ |R + wa + γa x̂1 R2 |, and one has

ya = xa2 − R 2 ∼ |wa | + 2R · wa + 2γa R1 R2 + γa2 R22, (10.8)

where the term 2γa w1a R2 has been neglected because it also gives subleading con-
tributions. The first two terms in Eq. (10.8) can be analysed as in Section 4.2.3,
resulting in Eq. (4.49) being modified by the addition of the last two terms in
Eq. (10.8),

ya = 2αaa 2 /d + 2Rza /d + 2γa R1 R2 + γa2 R22 . (10.9)

Following the same reasoning as in Section 4.2.3, we expect that ya ∝ 1/d.

Expanding at leading order in 1/d, with R = g (1 + h/d), one obtains a
modification of Eq. (4.50),
2
d ya R1 R2 2 R2
d(xa / − 1) ∼ h − η + = h − η + α aa + z a + dγa + dγa .
22 R2 R2
(10.10)

Here, the variables za are distributed according to Eq. (4.48) and are independent
of R. Eq. (10.6) has to be integrated over R within the second virial coefficient,
which amounts to integrating over R = |R| and over the unit vector R̂ = R/R. It
can be shown that, when d → ∞, the integration
√ over the components of R̂ can be
replaced by an average, R̂μ = Rμ /R → gμ / d, where gμ are independent random
Gaussian variables with zero mean and unit variance. Hence, Eq. (10.10) becomes

d(xa / − 1) ∼ h − η + αaa + za + γa g1 g2 + γa2 g22 . (10.11)

Eq. (4.52) is then replaced by

n n ' (
∂2
−βg v̄(h) a,b=2 αab ∂ha ∂hb + γa g1 g2 + 12 γa2 g22 ∂h∂ a
f eff (h,g1,g2 ) = e e a=2

n 0 (10.12)
× e− a=2 β v̄(ha −η) 0 − 1,
ha =h

and the excess free energy becomes

  ∞
d)ϕg
−βfex = Dg1 Dg2 dh eh f eff (h,g1,g2 ), (10.13)
2 −∞
√
where Dg = dg e−g /2 / 2π denotes the integration over a Gaussian variable
2

of zero mean and unit variance [59, 299]. By integrating by parts and expressing
294 Rheology of the Glass

the free energy in terms of , ˆ one of the two Gaussian integrations can be elimi-
nated, and
  ∞  
d)
ϕg
−βf =ex
Dg ˆ
dh eh fd (,h,g) −1 , (10.14)
2 −∞
   2  0
− 12 na,b=1 ab − g2 (γa −γb )2 ∂ha∂ ∂h − na=1 βa v̄(ha −ηa ) 0
2
d ˆ
f (,h,g) = e b e 0 ,
ha =h

where γ1 = 0, η1 = 0, and γa = γ , ηa = η, ∀a ≥ 2, as in Chapter 4. The addition

of a shear strain γ thus only modifies the interaction term by an additional Gaussian
integration.
As discussed in Chapter 4, the matrix  ˆ should be determined by extremising
ˆ
the free energy. Once the solution for  is found, the physical observables can thus
be obtained by taking the appropriate derivatives of the free energy. For example,
the energy and pressure are still given by the derivatives with respect to tempera-
ture and density, respectively. The shear stress is defined through response theory
[366, 368] as
∂fg (ϕ,T ,γ ;ϕg,Tg )
σ (ϕ,T ,γ ;ϕg,Tg ) = =dT)
σ (ϕ,T ,γ ;ϕg,Tg ). (10.15)
∂γ
Because the glass free energy is proportional to d, so is the stress; hence, as for
pressure, we define an adimensional scaled stress ) σ = βσ/d that remains finite
when d → ∞. Moreover, because, for hard spheres, βfg is independent of temper-
ature, so is )σ . Eq. (10.15) thus allows one to obtain the stress–strain curve of the
glass within the state following approach.
Note that if all the replicas, including the reference replica a = 1, are subjected
to the same shear strain, γa = γ , ∀a, then the free energy in Eq. (10.14) becomes
independent of γ . This result is correct, because a molecular liquid is globally a
liquid and its free energy should not depend on the shape of its container. Consis-
tently, according to Eq. (10.15), the shear stress then vanishes – i.e., σ = 0. Recall
that the state following construction keeps the reference replica unstrained, γ1 = 0,
and applies the strain on the constrained replicas, γa = 0. This produces a non-
trivial result because replicas a > 1 are not liquid but, rather, constrained to remain
within the glass state selected by the first replica. Note that within the Monasson
construction discussed in Chapter 7, in which all replicas are equivalent, extracting
the rheological properties of the glass requires a slightly different procedure, as
discussed in [366, 368].
As in Chapter 4,  ˆ is assumed to be a hierarchical matrix, the simplest choice
being the replica symmetric ansatz. Its stability can be checked along the same lines
as in Chapter 6 [59, 298, 299]. Using the replica symmetric ansatz in Eq. (4.62) for
the matrix ,ˆ the RS free energy is
 10.2 Linear Response 295
 
d πe d 2r − 
−βfg = log 2
+
2 d 2 
 (10.16)
d)ϕg
+ dh eh qγ (2r − ,βg ;h) log q (,β,h − η) ,
2
where

qγ (,β;h) = Dg q( + γ 2 g 2,β;h). (10.17)

Note that for γ = 0, qγ (,β;h) = q(,β;h) and the replica symmetric expression
without shear, Eq. (4.75), is recovered. The parameters  and r are fixed by
extremising Eq. (10.16), and the resulting equations are identical to Eq. (4.76), with
the replacement q(2r − ,βg ;h) → qγ (2r − ,βg ;h). The solution to these
equations can thus be found by starting from equilibrium at β = βg and η = γ = 0,
where  = r , and then following the solution upon changing β, η and γ .

10.2 Linear Response

In the rest of this chapter, we discuss a selection of results for the hard-sphere poten-
tial. Before discussing the full stress–strain curves, we focus on the linear response
regime in the small strain limit. We consider two observables: the shear modulus,
which characterises the rigidity of glasses, and the dilatancy, which describes the
propensity of glasses to expand upon constant-pressure straining.

10.2.1 Shear Modulus

Because the free energy, Eq. (10.14), is an even function of γ , the stress vanishes
at zero strain. Physically, this is obvious because the glass is then in mechanical
equilibrium. At very small strain, elasticity theory assumes that the stress increases
linearly with γ , with a proportionality constant given by the shear modulus,
0 0
dσ 00 σ 00
d)
μ= = dT )
μ, )
μ= . (10.18)
dγ 0γ =0 dγ 0γ =0
The higher the shear modulus is, the more rigid the glass is, because its elastic
response is stronger. Note that it is convenient to define a scaled modulus ) μ that
remains finite for d → ∞ and is independent of temperature for hard spheres. In
the small strain limit, within the replica symmetric solution, one has [299, 369]
1
)
μ= , (10.19)

where  is evaluated at γ = 0. The scaled shear modulus is then simply the inverse
of the mean square displacement plateau. This linear response result connects the
296 Rheology of the Glass

g )
(ϕ
μ

g )
 ϕ
 μ(
g )
(ϕ

 ϕ)/
g
ϕ
μ

μ(
g = 8, 7, 6.667, 6, 5.25, 4.9
ϕ

ϕg

Figure 10.2 Shear modulus of the hard-sphere glass [299]. (Left) Shear modulus
)
μ()ϕg ) on the equilibrium line. For )ϕg > )
ϕd , the shear modulus is finite. It displays
a square root singularity before jumping to zero for ) ϕ → ) ϕd+ , Eq. (10.20),
as better shown in the inset. (Right) Shear modulus for glasses prepared at
different )
ϕg , as a function of )
ϕ . Upon decompression, the shear modulus jumps to
zero at melting (open squares, as in Figure 4.4), with a square root singularity
inherited from the square root singularity of the mean square displacement.
Upon compression, the RS solution becomes unstable at the Gardner transition
(triangles, as in Figure 6.1), beyond which a fullRSB description is needed. The
unstable continuation of the RS shear modulus diverges at jamming (dashed line,
shown only for ) ϕg = 8).

linear rheology of the glass to its unstrained properties. It is consistent, moreover,

with physical intuition: a glass with a smaller cage is more rigid.
We have seen in Chapter 4 that, at the dynamical transition,  jumps from
infinity in the liquid phase to a finite value in the dynamically arrested phase.
Particle cages are then infinitely long lived, and the system suddenly becomes rigid
for an infinite time. Consequently, the shear modulus also jumps from zero (in
the liquid) to a finite value at the dynamical transition, as shown in Figure 10.2.
This observation has a precise dynamical interpretation. Upon approaching the
dynamical glass transition from the liquid phase, ) ϕ → ) ϕd− , the time scale tp –
at which the dynamical mean square displacement shows a plateau before reaching
the diffusive regime – diverges as a power law; see Section 3.4.2. On times t  tp ,
the system is effectively frozen into an amorphous state; it is then able to sustain
a shear strain on this time scale. In other words, if a shear strain is applied at
t = 0, the shear stress σ (t) would only decay to zero at times t  tp . Only at
the dynamical glass transition can the system sustain a finite stress for arbitrarily
long times. While rigidity at infinite times emerges discontinuously, the time over
which the system behaves as a solid diverges continuously.3 Upon approaching

3 We note that this effect is a d → ∞ artefact. In any finite d, the dynamical glass transition becomes a smooth
crossover, and a finite shear stress over an infinite time can only be sustained when the system is frozen in the
ideal glass state – i.e., beyond the Kauzmann point. Yet, because experiments are always performed over finite
time scales, the system becomes effectively solid when tp is larger than the experimentally accessible time
scales [80, 120].
 10.2 Linear Response 297

the dynamical transition from the dynamically arrested phase, the shear modulus
displays a square root singularity before jumping to zero,

) μd + C()
μ∼) ϕ−)
ϕd )1/2, ) ϕd+,
ϕ→) (10.20)

where )μd = 1/() ϕd ) is the shear modulus at the dynamical glass transition.
One can also consider the shear modulus of a glass prepared at ) ϕg and followed
adiabatically at )
ϕ . Upon decompression, the shear modulus decreases and also
displays a square root singularity before jumping to zero at the melting spinodal of
the glass. Upon compression, the shear modulus increases. It diverges at the jam-
ming transition, where  → 0, because a hard-sphere system forms an infinitely
rigid contact network and cannot be deformed anymore. The replica symmetric
approximation gives  ∼ p )−1 , and )
μ ∼ p ) thus diverges upon approaching the
jamming point, but we will see in Section 10.2.2 that this scaling is modified by
fullRSB effects. Results for different ) ϕg are illustrated in Figure 10.2.

10.2.2 Hierarchy of Shear Moduli in the Gardner Phase

We now discuss the behaviour of the shear modulus in the Gardner phase, in
which replica symmetry is broken. Although this chapter only presents results for
hard spheres, the discussion of this section applies to any interaction potential.
As we discussed in Chapter 5, the phase space structure in the Gardner phase
is ultrametrically organised, with individual glass states grouped in sub-basins,
themselves grouped in bigger sub-basins, up to the largest metabasins, as described
by a function (x), or x(). The linear response of the system in equilibrium can
be probed by considering an initially equilibrated system, applying an infinitesimal
shear strain γ at time t = 0 and monitoring the shear stress )
σ (t). This perturbation
takes a long time to relax because, in the Gardner phase, the ultrametric structure
of glass states is associated to an infinite hierarchy of time scales [93, 112].
After a short transient, the system equilibrates in the individual glass state in
which it was initially prepared, the mean square displacement approaches M and,
correspondingly, the stress decays to a value )
σ (t) ∼ )μ(M )γ . The shear modulus
)
μ(M ) corresponding to the intra-state relaxation is then given by
1 1
μ(M ) =
) = , (10.21)
M λ(M )
which generalises Eq. (10.19) to the Gardner phase [369]. Waiting longer times
(that diverge with the system size), the system is able to leave the initial glass state
to find other states, thus exploring a wider portion of the ultrametric tree of states.
Correspondingly, the mean square displacement increases to values  > M . The
value of  and its associated index x() label the level of the hierarchy. One
can associate to it a shear modulus ) μ() = 1/λ(), where λ() is defined in
Eq. (9.47) [369], which gives the stress response on the time scale over which
298 Rheology of the Glass

values of  > M are explored. A full exploration of the glass metabasin then
gives a shear modulus ) μ(m ) = 1/λ(m ) [369]. Note that this idealised situation
corresponds to an equilibrium exploration of the glass metabasin. In reality, the
system explores the metabasin out of equilibrium, and establishing a correspon-
dence between ) μ() and the dynamical time scales is tricky [112, 192, 369]. Note
also that this distribution of shear moduli provides a clear physical meaning to the
function λ(), similarly to the distribution of linear magnetic susceptibilities in
spin glasses [254].
For hard spheres, upon approaching jamming one has

)−κ
M ∼ p ⇒ )
μ(M ) ∼ p
)κ , (10.22)

as determined by the scaling solution discussed in Chapter 9. The divergence of the

intra-state shear modulus is thus characterised by the exponent κ. Conversely, for
any finite  > M , one has λ() = ) λ()/) p, from Eq. (9.51), and the correspond-
ing shear modulus )μ() = p )/)λ() diverges proportionally to the reduced pressure
) [192, 369].
p

10.2.3 Dilatancy
Pressure also exhibits an interesting behaviour upon straining the system. Expand-
ing the RS free energy of the strained glass for γ → 0,

1
fg (η,γ ) fg (η) + μ(η)γ 2 + O(γ 4 ), (10.23)
2
and recalling that βμ(η)/d = 1/(η), Eq. (6.17) gives

∂(βfg (η,γ )) βR(η) 2

pg (η,γ ) = = pg (η,γ = 0) + γ + O(γ 4 ), (10.24)
∂η ρ

where the coefficient

βR(η) d d 1 d 1 d
= =− (10.25)
ρ 2 dη (η) 2 (η)2 dη

is the dilatancy. A positive dilatancy indicates that the system, kept at fixed
pressure, expands under strain. In hard-sphere glasses, (η) decreases upon
increasing η, as shown in Chapter 4; hence, the dilatancy is always positive, as
shown in Figure 10.3. Note that the dilatancy diverges at melting because (η)
has a square root singularity, and at jamming because (η) → 0 while d/dη
remains finite.
 10.3 Stress–Strain Curves 299

ϕg )
g )/(ρd)

R(ϕ)/R(
βR(ϕ


g = 8, 7, 6.667, 6, 5.25, 4.9
ϕ

ϕg

Figure 10.3 Dilatancy of the hard-sphere glass [299]. (Left) Dilatancy R() ϕg ) on
the equilibrium line. For )ϕg > )ϕd , the dilatancy is finite. It scales linearly before
jumping down to zero for ) ϕ → ) ϕd+ . (Right) Dilatancy of glasses prepared at
different )
ϕg , as a function of )
ϕ . Upon decompression, the dilatancy diverges at
melting, because of the square root singularity of . Upon compression, the RS
solution becomes unstable at the Gardner transition (triangles, as in Figure 6.1),
beyond which a fullRSB solution is needed. The unstable continuation of the RS
dilatancy diverges proportionally to 1/2 at jamming (dashed line, shown only
for )
ϕg = 8).

10.3 Stress–Strain Curves

Solving the equations for  and r as a function of γ provides the full stress–strain
curve beyond the small strain, linear response regime. These curves correspond to
an adiabatic (or ‘quasi-static’) following of the glass state under the applied strain.
In this section, we discuss the behaviour of these curves for hard-sphere glasses in
different regimes.

10.3.1 The Yielding Transition

Figure 10.4 shows the stress–strain curves of hard-sphere glasses prepared in equi-
librium at )
ϕ=) ϕg and subjected to a shear strain γ at constant density. In agreement
with Figure 10.2, more stable glasses at higher ) ϕg have a larger shear modulus in the
linear regime, at small γ . Upon increasing γ , all glasses undergo a Gardner tran-
sition, at which replica symmetry spontaneously breaks due to the applied strain.
Recall that beyond the Gardner transition, the replica symmetric solution is only
approximate, and a RSB calculation is needed. See [298] for a preliminary study.
Here, we limit our discussion to the RS results. At high enough shear strain, after the
stress reaches a maximum, a spinodal transition takes place, at which the solution
for r and  is lost. The spinodal point γY () ϕg ) then corresponds, within mean field
theory, to the yielding point of the glass, i.e., where it breaks.
300 Rheology of the Glass

g = 7, 6.5, 6, 5.5, 5
ϕ g = 7, 6.5, 6, 5.5, 5
ϕ
σ

p
γ γ

Figure 10.4 Stress–strain (left) and pressure–strain (right) curves for hard-sphere
glasses prepared in equilibrium at )
ϕg and strained at constant density. Both stress
and pressure increase with strain, and a Gardner transition (triangle) is observed
before the stress and pressure overshoot. At the yielding transition (diamond),
the solution for  and r is lost via a spinodal mechanism in the Franz–Parisi
potential, indicating the breakdown of the glass.

Note that upon increasing the preparation density ) ϕg , the yielding point γY ()
ϕg )
increases and, before yielding, the stress–strain curves display a more pronounced
stress overshoot. In Figure 10.4 the pressure as a function of the strain is also
reported. As predicted by Eq. (10.25), pressure increases quadratically in γ , and
the dilatancy is larger for more stable glasses.
At the yielding transition, both  and r display a square root singularity, indi-
cating that the local minimum of the Franz–Parisi potential becomes an inflection
point and then disappears, as it does upon approaching the dynamical transition.
One can then consider the fluctuations of r , which define a susceptibility
 
χr = N 2r  − r 2 . (10.26)
Because χr is related to the curvature at the local minimum of the Franz–Parisi
potential, it diverges at the yielding point. The yielding transition in large dimension
is thus a critical spinodal with disorder [293, 299].

10.3.2 Stability Map under Shear Strain and Compression

In Section 10.3.1, we discussed the behaviour of the stress–strain curves for glasses
prepared at )ϕg and strained at constant density. More generally, one can consider
following a glass state under a joint applied (de)compression (i.e., a compressive
strain) and shear strain. In other words, one can prepare a glass at () ϕg,γ = 0) and
ϕ,γ ). In the rest of this section, we consider )
then follow it in the plane () ϕg as fixed;
hence, we do not indicate the dependence on ) ϕg explicitly.
 10.3 Stress–Strain Curves 301

p
σ

 = 8.08, 8.41, 8.84, 9.29, 9.48, 9.58, 9.67, 9.77

ϕ  = 8.08, 8.41, 8.84, 9.29, 9.48, 9.58, 9.67, 9.77
ϕ

γ γ

Figure 10.5 Stress–strain (left) and pressure–strain (right) curves for hard-sphere
glasses prepared in equilibrium at ) ϕg = 8, compressed at ) ϕ>) ϕg and strained at
constant density [346]. At low ) ϕ , both stress and pressure increase with strain, and
a Gardner transition (triangle) is observed at γG () ϕ ) before the stress and pressure
overshoot, and the glass yields at γY () ϕ ) (diamond). At higher ) ϕ , a Gardner
transition is observed, but instead of yielding, the stress and pressure diverge at
ϕ ), which indicates shear jamming. At even higher )
a finite γj () ϕ>) ϕG (γ = 0), the
whole curve is unstable towards RSB.

In Figure 10.5, stress–strain and pressure–strain curves for a glass compressed at

)
ϕ>) ϕg , and then strained, are reported. Both the shear modulus and the dilatancy
increase with ) ϕ , resulting in a steeper increase of stress and pressure with γ . The
yielding strain γY () ϕ ), however, decreases upon increasing ) ϕ . For high enough ) ϕ,
a qualitative difference is found in the stress–strain curves: the yielding instability
disappears, and the stress and pressure diverge at a finite strain γj () ϕ ), indicating
that the system jams under strain.
The yielding instability at γY () ϕ ) and the jamming transition at γj ()
ϕ ) thus delimit
a region in which a stable glass exists under strain. Moreover, a Gardner transition
ϕ ) is also observed under strain. These three lines are plotted in Figure 10.6,
line γG ()
resulting in a ‘stability map’ [8, 193, 346] of the glass in the () ϕ,γ ) plane. At zero
strain, the stability region is delimited by the melting point in decompression and by
the jamming point ϕ )j in compression. At non-zero strain, the yielding line extends
from the melting point at low ) ϕ , while the jamming line extends from ϕ )j at high
density. The shear jamming and shear yielding lines meet at a critical point () ϕc,γc ),
at which the system yields at extremely high (divergent) stress.
Note that the Gardner transition line in Figure 10.6 extends from the zero-strain
Gardner transition and crosses the yielding line at a lower density. Hence, the
shear jamming line is fully contained within the RSB region. Shear jamming is
thus described by the same fullRSB scaling solution described in Chapter 9. It is
possible to show that the exponents γ ,θ,κ are universal on the shear jamming
302 Rheology of the Glass


γj (ϕ)

γY (ϕ)

γG (ϕ)

(ϕ c , γc )

ϕ g

ϕ
Figure 10.6 Stability map of a hard-sphere glass prepared at ) ϕg = 8, obtained
by plotting in the () ϕ,γ ) plane the lines γj () ϕ ) and γG ()
ϕ ), γY () ϕ ) defined in
Figure 10.5 [8, 346]. The preparation point () ϕg ,0) is indicated by a dot. At γ = 0,
the glass exists for densities larger than the melting point and smaller than the
jamming point. At high density, close to the jamming transition ϕ )j , the glass jams
under shear due to dilatancy; at low density, the glass yields under shear. The shear
jamming line and part of the yielding line lie within the RSB phase, γ > γG () ϕ ).
Under the RS approximation, the shear jamming and shear yielding lines meet at
a critical point ()
ϕc,γc ) (diamond) where the glass yields at divergent shear stress.

line [298]. The critical point ()

ϕc,γc ) also falls within the RSB region, and because
the Gardner line crosses the yielding line, yielding can be either RS (at low ) ϕ ) or
RSB (at high )ϕ ). Different critical properties around the spinodal are expected in
these two regions [298].

10.4 Wrap-Up
10.4.1 Summary
In this chapter, we have seen that

• The state following formalism can be extended to analyse the behaviour of glasses
subjected to shear strain, leading to a theory of the elasticity and rheology of
amorphous solids (Section 10.1).
• At small shear strains, the glass responds elastically, and the theory predicts
the shear modulus and the dilatancy. These quantities depend on the degree of
annealing of the glass and on the underlying phase space organisation of glass
states. In particular, in the Gardner phase, spontaneous replica symmetry breaking
gives rise to a hierarchy of shear moduli (Section 10.2).
 10.4 Wrap-Up 303

• Beyond the linear response regime, the stress–strain curves can be computed up
to the yielding point, where the glass breaks under the applied strain. Within
mean field theory, yielding is a critical spinodal of the Franz–Parisi potential.
Depending on the glass preparation and on the straining protocol, a Gardner
transition can be present before yielding (Section 10.3.1).
• When a hard-sphere glass prepared at sufficiently high pressure is sheared, it can
undergo a shear jamming transition that prevents yielding. The yielding and shear
jamming lines delimit the stability region of the solid in the (ϕ,γ ) phase diagram.
These lines meet at a critical point, where yielding happens with a divergent stress
(Section 10.3.2).

10.4.2 Further Reading

We provide here a list of references that can be consulted to further explore the
subjects discussed in this chapter, selected according to the criteria discussed in
Section 1.6.2. Because the literature on the rheology of glasses is extremely vast
and scattered among several disciplines (physics, materials science, engineering),
we only provide here a list of references that are closely related to the material
discussed in this chapter.
General introductory reviews to the rheology of amorphous materials are

• Berthier, Yield stress, heterogeneities and activated processes in soft glassy mate-
rials [38]
• Rodney, Tanguy and Vandembroucq, Modeling the mechanics of amorphous
solids at different length scale and time scale [305]
• Bonn, Denn, Berthier et al., Yield stress materials in soft condensed matter [62]
• Nicolas, Ferrero, Martens et al., Deformation and flow of amorphous solids: a
review of mesoscale elastoplastic models [272]
The criticality of yielding has been studied in

• Lin, Lerner, Rosso et al., Scaling description of the yielding transition in soft
amorphous solids at zero temperature [231]
• Parisi, Procaccia, Rainone et al., Shear bands as manifestation of a criticality in
yielding amorphous solids [293]
• Ozawa, Berthier, Biroli et al., Random critical point separates brittle and ductile
yielding transitions in amorphous materials [280]
Beyond yielding, soft glasses such as pastes and emulsions break and start to flow.
A dynamical investigation is then required. Mean field dynamical equations for the
description of this flow regime have been developed in

• Berthier, Barrat and Kurchan, A two-time-scale, two-temperature scenario for

nonlinear rheology [36]
304 Rheology of the Glass

• Fuchs and Cates, Theory of nonlinear rheology and yielding of dense colloidal
suspensions [156]
• Brader, Voigtmann, Fuchs et al., Glass rheology: from mode-coupling theory to a
dynamical yield criterion [67]
The relation between the Gardner transition, plasticity and avalanches has been
discussed in

• Biroli and Urbani, Breakdown of elasticity in amorphous solids [58]

• Nakayama, Yoshino and Zamponi, Protocol-dependent shear modulus of amor-
phous solids [270]
• Franz and Spigler, Mean-field avalanches in jammed spheres [148]
• Jin and Yoshino, Exploring the complex free-energy landscape of the simplest
glass by rheology [192]
• Jin, Urbani, Zamponi et al., A stability-reversibility map unifies elasticity, plastic-
ity, yielding and jamming in hard sphere glasses [193]
An extension of the theory presented here to attractive colloids can be found in
Altieri, Urbani and Zamponi, Microscopic theory of two-step yielding in attractive
colloids [7].
 References

[1] Adam, G., and Gibbs, J. 1965. On the temperature dependence of cooperative
relaxation properties in glass-forming liquids. The Journal of Chemical Physics,
43, 139.
[2] Adda-Bedia, M., Katzav, E., and Vella, D. 2008. Solution of the Percus–Yevick
equation for hard hyperspheres in even dimensions. The Journal of Chemical
Physics, 129, 144506.
[3] Agoritsas, E., Maimbourg, T., and Zamponi, F. 2018. Out-of-equilibrium dynamical
equations of infinite-dimensional particle systems. I. The isotropic case. Journal of
Physics A: Mathematical and Theoretical, 52, 144002.
[4] Aktekin, N. 1997. Simulation of the eight-dimensional Ising model on the Creutz
cellular automaton. International Journal of Modern Physics C, 8, 287.
[5] Alder, B. J., and Wainwright, T. E. 1957. Phase transition for a hard sphere system.
The Journal of Chemical Physics, 27, 1208.
[6] Alexander, S. 1998. Amorphous solids: Their structure, lattice dynamics and
elasticity. Physics Reports, 296, 65.
[7] Altieri, A., Urbani, P., and Zamponi, F. 2018. Microscopic theory of two-step
yielding in attractive colloids. Physical Review Letters, 121, 185503.
[8] Altieri, A., and Zamponi, F. 2019. Mean-field stability map of hard-sphere glasses.
Physical Review E, 100, 032140.
[9] Aluffi-Pentini, F., Parisi, V., and Zirilli, F. 1988. A global optimization algo-
rithm using stochastic differential equations. ACM Transactions on Mathematical
Software, 14, 345.
[10] Amit, D. J. 1992. Modeling brain function: The world of attractor neural networks.
Cambridge University Press.
[11] Amit, D. J., and Martin-Mayor, V. 2005. Field theory, the renormalization group,
and critical phenomena. 3rd edn. World Scientific.
[12] Andreanov, A., and Scardicchio, A. 2012. Random perfect lattices and the sphere
packing problem. Physical Review E, 86, 041117.
[13] Angelani, L., and Foffi, G. 2007. Configurational entropy of hard spheres. Journal
of Physics: Condensed Matter, 19, 256207.
[14] Angelini, M. C., and Biroli, G. 2015. Spin glass in a field: A new zero-temperature
fixed point in finite dimensions. Physical Review Letters, 114, 095701.
[15] Angelini, M. C., Parisi, G., and Ricci-Tersenghi, F. 2013. Ensemble renormalization
group for disordered systems. Physical Review B, 87, 134201.

305
306 References

[16] Angell, C. 1997. Entropy and fragility in supercooling liquids. Journal of research
of the National Institute of Standards and Technology, 102, 171.
[17] Apostolico, A., Comin, M., Dress, A., et al. 2013. Ultrametric networks: A new tool
for phylogenetic analysis. Algorithms for Molecular Biology, 8, 7.
[18] Aron, C., Biroli, G., and Cugliandolo, L. F. 2010. Symmetries of generating
functionals of Langevin processes with colored multiplicative noise. Journal of
Statistical Mechanics: Theory and Experiment, 2010, P11018.
[19] Arora, S., and Barak, B. 2009. Computational complexity: A modern approach.
Cambridge University Press.
[20] Asenjo, D., Paillusson, F., and Frenkel, D. 2014. Numerical calculation of granular
entropy. Physical Review Letters, 112, 098002.
[21] Ashcroft, N. W., and Mermin, N. D. 1976. Solid state physics. Thomson Learning.
[22] Aspelmeier, T., Bray, A., and Moore, M. 2004. Complexity of Ising spin glasses.
Physical Review Letters, 92, 087203.
[23] Baity-Jesi, M., Baños, R., Cruz, A., et al. 2013. Critical parameters of the three-
dimensional Ising spin glass. Physical Review B, 88, 224416.
[24] Baity-Jesi, M., Goodrich, C. P., Liu, A. J., et al. 2017. Emergent SO(3) symmetry of
the frictionless shear jamming transition. Journal of Statistical Physics, 167, 735.
[25] Bannerman, M. N., Lue, L., and Woodcock, L. V. 2010. Thermodynamic pressures
for hard spheres and closed-virial equation-of-state. The Journal of Chemical
Physics, 132, 084507.
[26] Banos, R. A., Cruz, A., Fernandez, L., et al. 2010. Nature of the spin-glass
phase at experimental length scales. Journal of Statistical Mechanics: Theory and
Experiment, 2010, P06026.
[27] Barrat, A., Burioni, R., and Mézard, M. 1996. Dynamics within metastable states in
a mean-field spin glass. Journal of Physics A: Mathematical and General, 29, L81.
[28] Barrat, A., Franz, S., and Parisi, G. 1997. Temperature evolution and bifurcations of
metastable states in mean-field spin glasses, with connections with structural glasses.
Journal of Physics A: Mathematical and General, 30, 5593.
[29] Barrat, A., Kurchan, J., Loreto, V., et al. 2001. Edwards measures: A thermodynamic
construction for dense granular media and glasses. Physical Review E, 63, 051301.
[30] Baule, A., Morone, F., Herrmann, H. J., et al. 2018. Edwards statistical mechanics
for jammed granular matter. Reviews of Modern Physics, 90, 015006.
[31] Behringer, R. P., and Chakraborty, B. 2019. The physics of jamming for granular
materials: A review. Reports on Progress in Physics, 82, 012601.
[32] Belletti, F., Cruz, A., Fernandez, L., et al. 2009. An in-depth view of the microscopic
dynamics of Ising spin glasses at fixed temperature. Journal of Statistical Physics,
135, 1121.
[33] Bernal, J., and Mason, J. 1960. Packing of spheres: Co-ordination of randomly
packed spheres. Nature, 188, 910.
[34] Bernal, J., Mason, J., and Knight, K. 1962. Radial distribution of the random close
packing of equal spheres. Nature, 194, 957.
[35] Bernard, E. P., and Krauth, W. 2011. Two-step melting in two dimensions: First-
order liquid-hexatic transition. Physical Review Letters, 107, 155704.
[36] Berthier, L., Barrat, J., and Kurchan, J. 2000. A two-time-scale, two-temperature
scenario for nonlinear rheology. Physical Review E, 61, 5464.
[37] Berthier, L., Biroli, G., Bouchaud, J.-P., et al. 2011. Dynamical heterogeneities and
glasses. Oxford University Press.
[38] Berthier, L. 2003. Yield stress, heterogeneities and activated processes in soft glassy
materials. Journal of Physics: Condensed Matter, 15, S933.
 References 307

[39] Berthier, L. 2013. Overlap fluctuations in glass-forming liquids. Physical Review E,

88, 022313.
[40] Berthier, L., and Biroli, G. 2011. Theoretical perspective on the glass transition and
amorphous materials. Reviews of Modern Physics, 83, 587.
[41] Berthier, L., and Kurchan, J. 2013. Non-equilibrium glass transitions in driven and
active matter. Nature Physics, 9, 310.
[42] Berthier, L., and Witten, T. A. 2009. Glass transition of dense fluids of hard and
compressible spheres. Physical Review E, 80, 021502.
[43] Berthier, L., Jacquin, H., and Zamponi, F. 2011. Microscopic theory of the jamming
transition of harmonic spheres. Physical Review E, 84, 051103.
[44] Berthier, L., Charbonneau, P., Jin, Y., et al. 2016. Growing timescales and length-
scales characterizing vibrations of amorphous solids. Proceedings of the National
Academy of Sciences, 113, 8397.
[45] Berthier, L., Charbonneau, P., Coslovich, D., et al. 2017. Configurational entropy
measurements in extremely supercooled liquids that break the glass ceiling. Pro-
ceedings of the National Academy of Sciences, 114, 11356.
[46] Berthier, L., Charbonneau, P., Ninarello, A., et al. 2019. Zero-temperature glass
transition in two dimensions. Nature communications, 10, 1508.
[47] Bhattacharyya, S. M., Bagchi, B., and Wolynes, P. G. 2008. Facilitation, complexity
growth, mode coupling, and activated dynamics in supercooled liquids. Proceedings
of the National Academy of Sciences, 105, 16077.
[48] Bi, D., Henkes, S., Daniels, K. E., et al. 2015. The statistical physics of athermal
materials. Annual Reviews of Condensed Matter Physics, 6, 63.
[49] Biazzo, I., Caltagirone, F., Parisi, G., et al. 2009. Theory of amorphous packings of
binary mixtures of hard spheres. Physical Review Letters, 102, 195701.
[50] Biere, A., Heule, M., and van Maaren, H. (eds). 2009. Handbook of satisfiability.
Frontiers in Artificial Intelligence and Applications, vol. 185. IOS Press.
[51] Binder, K., and Young, A. 1986. Spin glasses: Experimental facts, theoretical
concepts, and open questions. Reviews of Modern Physics, 58, 801.
[52] Binder, K. 1987. Theory of first-order phase transitions. Reports on Progress in
Physics, 50, 783.
[53] Binder, K., and Kob, W. 2011. Glassy materials and disordered solids: An introduc-
tion to their statistical mechanics. World Scientific.
[54] Binder, K., and Luijten, E. 2001. Monte Carlo tests of renormalization-group
predictions for critical phenomena in Ising models. Physics Reports, 344, 179.
[55] Binder, K., Block, B. J., Virnau, P., et al. 2012. Beyond the van der Waals loop:
What can be learned from simulating Lennard-Jones fluids inside the region of phase
coexistence. American Journal of Physics, 80, 1099.
[56] Biroli, G., Bouchaud, J., Miyazaki, K., et al. 2006. Inhomogeneous mode-coupling
theory and growing dynamic length in supercooled liquids. Physical Review Letters,
97, 195701.
[57] Biroli, G., Bouchaud, J., Cavagna, A., et al. 2008. Thermodynamic signature of
growing amorphous order in glass-forming liquids. Nature Physics, 4, 771.
[58] Biroli, G., and Urbani, P. 2016. Breakdown of elasticity in amorphous solids. Nature
Physics, 12, 1130.
[59] Biroli, G., and Urbani, P. 2018. Liu-Nagel phase diagrams in infinite dimension.
SciPost Physics, 4, 020.
[60] Blichfeldt, H. F. 1929. The minimum value of quadratic forms, and the closest
packing of spheres. Mathematische Annalen, 101, 605.
308 References

[61] Bolhuis, P. G., Frenkel, D., Mau, S.-C., et al. 1997. Entropy difference between
crystal phases. Nature, 388, 235.
[62] Bonn, D., Denn, M. M., Berthier, L., et al. 2017. Yield stress materials in soft
condensed matter. Reviews of Modern Physics, 89, 035005.
[63] Bouchaud, J.-P., and Biroli, G. 2004. On the Adam-Gibbs-Kirkpatrick-Thirumalai-
Wolynes scenario for the viscosity increase in glasses. The Journal of Chemical
Physics, 121, 7347.
[64] Bouchaud, J.-P., and Biroli, G. 2005. Nonlinear susceptibility in glassy systems:
A probe for cooperative dynamical length scales. Physical Review B, 72, 064204.
[65] Bouchaud, J., Cugliandolo, L., Kurchan, J., et al. 1998. Out of equilibrium dynamics
in spin-glasses and other glassy systems. In Young, A. (ed), Spin glasses and random
fields. World Scientific.
[66] Bowles, R. K., and Ashwin, S. 2011. Edwards entropy and compactivity in a model
of granular matter. Physical Review E, 83, 031302.
[67] Brader, J. M., Voigtmann, T., Fuchs, M., et al. 2009. Glass rheology: From mode-
coupling theory to a dynamical yield criterion. Proceedings of the National Academy
of Sciences, 106, 15186.
[68] Bray, A. J., and Moore, M. A. 1987. Chaotic nature of the spin-glass phase. Physical
Review Letters, 58, 57.
[69] Brézin, E. 2010. Introduction to statistical field theory. Cambridge University Press.
[70] Brito, C., Ikeda, H., Urbani, P., et al. 2018. Universality of jamming of nonspherical
particles. Proceedings of the National Academy of Sciences, 115, 11736.
[71] Butera, P., and Pernici, M. 2012. High-temperature expansions of the higher
susceptibilities for the Ising model in general dimension d. Physical Review E, 86,
011139.
[72] Caltagirone, F., Ferrari, U., Leuzzi, L., et al. 2012. Critical slowing down exponents
of mode coupling theory. Physical Review Letters, 108, 085702.
[73] Cammarota, C., Cavagna, A., Giardina, I., et al. 2010. Phase-separation perspective
on dynamic heterogeneities in glass-forming liquids. Physical Review Letters, 105,
055703.
[74] Campa, A., Dauxois, T., and Ruffo, S. 2009. Statistical mechanics and dynamics of
solvable models with long-range interactions. Physics Reports, 480, 57.
[75] Capaccioli, S., Ruocco, G., and Zamponi, F. 2008. Dynamically correlated regions
and configurational entropy in supercooled liquids. The Journal of Physical Chem-
istry B, 112, 10652.
[76] Cardenas, M., Franz, S., and Parisi, G. 1998. Glass transition and effective potential
in the hypernetted chain approximation. Journal of Physics A: Mathematical and
General, 31, L163.
[77] Cardenas, M., Franz, S., and Parisi, G. 1999. Constrained Boltzmann–Gibbs mea-
sures and effective potential for glasses in hypernetted chain approximation and
numerical simulations. The Journal of Chemical Physics, 110, 1726.
[78] Castellana, M., and Parisi, G. 2015. Non-perturbative effects in spin glasses.
Scientific Reports, 5, 8697.
[79] Castellani, T., and Cavagna, A. 2005. Spin-glass theory for pedestrians. Journal of
Statistical Mechanics: Theory and Experiment, 2005, P05012.
[80] Cavagna, A. 2009. Supercooled liquids for pedestrians. Physics Reports, 476, 51.
[81] Cavagna, A., Giardina, I., and Parisi, G. 1998. Stationary points of the Thouless-
Anderson-Palmer free energy. Physical Review B, 57, 11251.
[82] Chakraborty, B. 2010. Statistical ensemble approach to stress transmission in
granular packings. Soft Matter, 6, 2884.
 References 309

[83] Charbonneau, B., Charbonneau, P., Jin, Y., et al. 2013. Dimensional dependence of
the Stokes–Einstein relation and its violation. The Journal of Chemical Physics, 139,
164502.
[84] Charbonneau, B., Charbonneau, P., and Szamel, G. 2018. A microscopic model
of the Stokes–Einstein relation in arbitrary dimension. The Journal of Chemical
Physics, 148, 224503.
[85] Charbonneau, P., Ikeda, A., Parisi, G., et al. 2012. Dimensional study of the caging
order parameter at the glass transition. Proceedings of the National Academy of
Sciences, 109, 13939.
[86] Charbonneau, P., Ikeda, A., Parisi, G., et al. 2011. Glass transition and random close
packing above three dimensions. Physical Review Letters, 107, 185702.
[87] Charbonneau, P., Corwin, E. I., Parisi, G., et al. 2012. Universal microstructure and
mechanical stability of jammed packings. Physical Review Letters, 109, 205501.
[88] Charbonneau, P., Kurchan, J., Parisi, G., et al. 2014. Exact theory of dense
amorphous hard spheres in high dimension. III. The full replica symmetry breaking
solution. Journal of Statistical Mechanics: Theory and Experiment, 2014, P10009.
[89] Charbonneau, P., Kurchan, J., Parisi, G., et al. 2014. Fractal free energies in structural
glasses. Nature Communications, 5, 3725.
[90] Charbonneau, P., Jin, Y., Parisi, G., et al. 2014. Hopping and the Stokes–Einstein
relation breakdown in simple glass formers. Proceedings of the National Academy
of Sciences, 111, 15025.
[91] Charbonneau, P., Corwin, E. I., Parisi, G., et al. 2015. Jamming criticality revealed
by removing localized buckling excitations. Physical Review Letters, 114, 125504.
[92] Charbonneau, P., Corwin, E. I., Parisi, G., et al. 2016. Universal non-Debye scaling
in the density of states of amorphous solids. Physical Review Letters, 117, 045503.
[93] Charbonneau, P., Kurchan, J., Parisi, G., et al. 2017. Glass and jamming transitions:
From exact results to finite-dimensional descriptions. Annual Review of Condensed
Matter Physics, 8, 265.
[94] Charbonneau, P., Corwin, E. I., Fu, L., et al. 2019. Glassy, Gardner-like phe-
nomenology in minimally polydisperse crystalline systems. Physical Review E, 99,
020901(R).
[95] Charbonneau, P., Hu, Y., Raju, A., et al. 2019. Morphology of renormalization-group
flow for the de Almeida-Thouless-Gardner universality class. Physical Review E, 99,
022132.
[96] Clisby, N., and McCoy, B. M. 2006. Ninth and tenth order virial coefficients for hard
spheres in D dimensions. Journal of Statistical Physics, 122, 15.
[97] Cohn, H. 2002. New upper bounds on sphere packings II. Geometry & Topology,
6, 329.
[98] Cohn, H. 2016. Packing, coding, and ground states. arXiv:1603.05202.
[99] Cohn, H. 2017. A conceptual breakthrough in sphere packing. Notices of the
American Mathematical Society, 64, 102.
[100] Cohn, H., and Elkies, N. 2003. New upper bounds on sphere packings I. Annals of
Mathematics, 157, 689.
[101] Cohn, H., Kumar, A., Miller, S. D., et al. 2017. The sphere packing problem in
dimension 24. Annals of Mathematics, 185, 1017.
[102] Coluzzi, B., Mézard, M., Parisi, G., et al. 1999. Thermodynamics of binary mixture
glasses. The Journal of Chemical Physics, 111, 9039.
[103] Conway, J. H., and Sloane, N. J. A. 1993. Sphere packings, lattices and groups.
Spriger-Verlag.
[104] Conway, J. B. 1990. A course in functional analysis. Springer-Verlag.
310 References

[105] Costigliola, L., Schroder, T. B., and Dyre, J. C. 2016. Studies of the Lennard–Jones
fluid in 2, 3, and 4 dimensions highlight the need for a liquid-state 1/d expansion.
The Journal of Chemical Physics, 144, 231101.
[106] Crisanti, A., and De Dominicis, C. 2015. Replica Fourier transform: Properties and
applications. Nuclear Physics B, 891, 73.
[107] Crisanti, A., and Rizzo, T. 2002. Analysis of the ∞-replica symmetry breaking
solution of the Sherrington-Kirkpatrick model. Physical Review E, 65, 046137.
[108] Crisanti, A., and Leuzzi, L. 2006. Spherical 2 + p spin-glass model: An analytically
solvable model with a glass-to-glass transition. Physical Review B, 73, 014412.
[109] Cugliandolo, L. F., and Kurchan, J. 1993. Analytical solution of the off-equilibrium
dynamics of a long-range spin-glass model. Physical Review Letters, 71, 173.
[110] Cugliandolo, L. F., Kurchan, J., and Peliti, L. 1997. Energy flow, partial equilibra-
tion, and effective temperatures in systems with slow dynamics. Physical Review E,
55, 3898.
[111] Cugliandolo, L. F. 2003. Dynamics of glassy systems. In Barrat, J., Feigelman, M.,
Kurchan, J., et al. (eds), Slow relaxations and nonequilibrium dynamics in condensed
matter. Springer-Verlag.
[112] Cugliandolo, L. F., and Kurchan, J. 1994. On the out-of-equilibrium relaxation of the
Sherrington-Kirkpatrick model. Journal of Physics A: Mathematical and General,
27, 5749.
[113] de Almeida, J., and Thouless, D. 1978. Stability of the Sherrington-Kirkpatrick
solution of a spin glass model. Journal of Physics A: Mathematical and General,
11, 983.
[114] De Dominicis, C. 1978. Dynamics as a substitute for replicas in systems with
quenched random impurities. Physical Review B, 18, 4913.
[115] De Dominicis, C., and Giardina, I. 2006. Random fields and spin glasses: A field
theory approach. Cambridge University Press.
[116] De Dominicis, C., Carlucci, D., and Temesvari, T. 1997. Replica Fourier tansforms
on ultrametric trees, and block-diagonalizing multi-replica matrices. Journal de
Physique I, 7, 105.
[117] De Dominicis, C., Temesvari, T., and Kondor, I. 1998. On Ward-Takahashi identities
for the Parisi spin glass. Journal de Physique IV, 8, Pr6.13.
[118] De Dominicis, C. 1962. Variational formulations of equilibrium statistical mechan-
ics. Journal of Mathematical Physics, 3, 983.
[119] Debenedetti, P. G., and Stillinger, F. H. 2001. Supercooled liquids and the glass
transition. Nature, 410, 259.
[120] Debenedetti, P. 1996. Metastable liquids: Concepts and principles. Princeton
University Press.
[121] DeGiuli, E., Laversanne-Finot, A., Düring, G., et al. 2014. Effects of coordination
and pressure on sound attenuation, boson peak and elasticity in amorphous solids.
Soft Matter, 10, 5628.
[122] DeGiuli, E., Lerner, E., Brito, C., et al. 2014. Force distribution affects vibrational
properties in hard-sphere glasses. Proceedings of the National Academy of Sciences,
111, 17054.
[123] DeGiuli, E., Lerner, E., and Wyart, M. 2015. Theory of the jamming transition at
finite temperature. The Journal of Chemical Physics, 142, 164503.
[124] Delamotte, B. 2012. An introduction to the nonperturbative renormalization group.
In Polonyi, J., and Schwenk, A. (eds), Renormalization group and effective field
theory approaches to many-body systems. Springer-Verlag.
 References 311

[125] Derrida, B. 1981. Random-energy model: An exactly solvable model of disordered

systems. Physical Review B, 24, 2613.
[126] DiVincenzo, D. P., and Steinhardt, P. J. 1999. Quasicrystals: The state of the art.
World Scientific.
[127] Donev, A., Cisse, I., Sachs, D., et al. 2004. Improving the density of jammed
disordered packings using ellipsoids. Science, 303, 990.
[128] Donev, A., Torquato, S., and Stillinger, F. H. 2005. Pair correlation function
characteristics of nearly jammed disordered and ordered hard-sphere packings.
Physical Review E, 71, 011105.
[129] Donev, A., Connelly, R., Stillinger, F. H., et al. 2007. Underconstrained jammed
packings of nonspherical hard particles: Ellipses and ellipsoids. Physical Review E,
75, 051304.
[130] Donth, E. 2001. The glass transition: Relaxation dynamics in liquids and disordered
materials. Springer-Verlag.
[131] Dotsenko, V., Franz, S., and Mézard, M. 1994. Partial annealing and overfrustration
in disordered systems. Journal of Physics A: Mathematical and General, 27, 2351.
[132] Drouffe, J.-M., Parisi, G., and Sourlas, N. 1979. Strong coupling phase in lattice
gauge theories at large dimension. Nuclear Physics B, 161, 397.
[133] Duplantier, B. 1981. Comment on Parisi’s equation for the SK model for spin
glasses. Journal of Physics A: Mathematical and General, 14, 283.
[134] Dyre, J. C. 2016. Simple liquids quasiuniversality and the hard-sphere paradigm.
Journal of Physics: Condensed Matter, 28, 323001.
[135] Ediger, M. D. 2000. Spatially heterogeneous dynamics in supercooled liquids.
Annual Review of Physical Chemistry, 51, 99.
[136] Edwards, S. F., and Anderson, P. W. 1975. Theory of spin glasses. Journal of Physics
F: Metal Physics, 5, 965.
[137] El-Showk, S., Paulos, M. F., Poland, D., et al. 2014. Solving the 3D Ising model with
the conformal bootstrap II. c-minimization and precise critical exponents. Journal of
Statistical Physics, 157, 869.
[138] Engel, A., and Van den Broeck, C. 2001. Statistical mechanics of learning.
Cambridge University Press.
[139] Fernández, L., Martin-Mayor, V., Seoane, B., et al. 2012. Equilibrium fluid-solid
coexistence of hard spheres. Physical Review Letters, 108, 165701.
[140] Fischer, K., and Hertz, J. 1991. Spin glasses. Cambridge University Press.
[141] Fisher, D. S., and Huse, D. A. 1988. Equilibrium behavior of the spin-glass ordered
phase. Physical Review B, 38, 386.
[142] Fisher, M. E., and Gaunt, D. S. 1964. Ising model and self-avoiding walks on
hypercubical lattices and ‘high-density’ expansions. Physical Review, 133, A224.
[143] Franz, S., and Montanari, A. 2007. Analytic determination of dynamical and mosaic
length scales in a Kac glass model. Journal of Physics A: Mathematical and
Theoretical, 40, F251.
[144] Franz, S., and Parisi, G. 1995. Recipes for metastable states in spin glasses. Journal
de Physique I, 5, 1401.
[145] Franz, S., Parisi, G., Ricci-Tersenghi, F., et al. 2011. Field theory of fluctuations in
glasses. The European Physical Journal E, 34, 1.
[146] Franz, S., and Parisi, G. 2000. On non-linear susceptibility in supercooled liquids.
Journal of Physics: Condensed Matter, 12, 6335.
[147] Franz, S., and Parisi, G. 2016. The simplest model of jamming. Journal of Physics
A: Mathematical and Theoretical, 49, 145001.
312 References

[148] Franz, S., and Spigler, S. 2017. Mean-field avalanches in jammed spheres. Physical
Review E, 95, 022139.
[149] Franz, S., Jacquin, H., Parisi, G., et al. 2012. Quantitative field theory of the glass
transition. Proceedings of the National Academy of Sciences, 109, 18725.
[150] Franz, S., Parisi, G., Urbani, P., et al. 2015. Universal spectrum of normal modes
in low-temperature glasses. Proceedings of the National Academy of Sciences, 112,
14539.
[151] Franz, S., Parisi, G., Sevelev, M., et al. 2017. Universality of the SAT-UNSAT
(jamming) threshold in non-convex continuous constraint satisfaction problems.
SciPost Physics, 2, 019.
[152] Frenkel, D., and Smit, B. 2001. Understanding molecular simulation: From algo-
rithms to applications. Elsevier.
[153] Frisch, H. L., and Percus, J. K. 1999. High dimensionality as an organizing device
for classical fluids. Physical Review E, 60, 2942.
[154] Frisch, H. L., Rivier, N., and Wyler, D. 1985. Classical hard-sphere fluid in infinitely
many dimensions. Physical Review Letters, 54, 2061.
[155] Frisch, H., and Percus, J. 1987. Nonuniform classical fluid at high dimensionality.
Physical Review A, 35, 4696.
[156] Fuchs, M., and Cates, M. E. 2002. Theory of nonlinear rheology and yielding of
dense colloidal suspensions. Physical Review Letters, 89, 248304.
[157] Fullerton, C. J., and Berthier, L. 2017. Density controls the kinetic stability of
ultrastable glasses. Europhysics Letters, 119, 36003.
[158] Gallavotti, G. 2000. Statistical mechanics. A short treatise. Springer-Verlag.
[159] Gardiner, C. W. 1985. Handbook of stochastic methods for physics, chemistry and
natural sciences. Springer-Verlag.
[160] Gardner, E. 1985. Spin glasses with p-spin interactions. Nuclear Physics B,
257, 747.
[161] Gardner, E., and Derrida, B. 1988. Optimal storage properties of neural network
models. Journal of Physics A: Mathematical and General, 21, 271.
[162] Geirhos, K., Lunkenheimer, P., and Loidl, A. 2018. Johari-Goldstein relaxation far
below Tg : Experimental evidence for the Gardner transition in structural glasses?
Physical Review Letters, 120, 085705.
[163] Georges, A., and Yedidia, J. S. 1991. How to expand around mean-field theory using
high-temperature expansions. Journal of Physics A: Mathematical and General,
24, 2173.
[164] Georges, A., Kotliar, G., Krauth, W., et al. 1996. Dynamical mean-field theory of
strongly correlated fermion systems and the limit of infinite dimensions. Reviews of
Modern Physics, 68, 13.
[165] Gofman, M., Adler, J., Aharony, A., et al. 1993. Series and Monte Carlo study of
high-dimensional Ising models. Journal of Statistical Physics, 71, 1221.
[166] Goldstein, M. 1969. Viscous liquids and the glass transition: A potential energy
barrier picture. The Journal of Chemical Physics, 51, 3728.
[167] Goodrich, C. P., Liu, A. J., and Sethna, J. P. 2016. Scaling ansatz for the jamming
transition. Proceedings of the National Academy of Sciences, 113, 9745.
[168] Götze, W. 2008. Complex dynamics of glass-forming liquids: A mode-coupling
theory. Oxford University Press.
[169] Götze, W. 1999. Recent tests of the mode-coupling theory for glassy dynamics.
Journal of Physics: Condensed Matter, 11, A1.
[170] Gross, D. J., and Mézard, M. 1984. The simplest spin glass. Nuclear Physics B,
240, 431.
 References 313

[171] Gross, D., Kanter, I., and Sompolinsky, H. 1985. Mean-field theory of the Potts glass.
Physical Review Letters, 55, 304.
[172] Haji-Akbari, A., Engel, M., Keys, A. S., et al. 2009. Disordered, quasicrystalline and
crystalline phases of densely packed tetrahedra. Nature, 462, 773.
[173] Hales, T., Adams, M., Bauer, G., et al. 2017. A formal proof of the Kepler conjecture.
In Forum of Mathematics, Pi, vol. 5. Cambridge University Press.
[174] Hales, T. C. 2005. A proof of the Kepler conjecture. Annals of Mathematics, 162,
1065.
[175] Hansen, J.-P., and McDonald, I. R. 1986. Theory of simple liquids (3rd edition).
Academic Press.
[176] Henkel, M., Pleimling, M., and Sanctuary, R. (eds). 2007. Ageing and the glass
transition. Springer.
[177] Heuer, A. 2008. Exploring the potential energy landscape of glass-forming systems:
From inherent structures via metabasins to macroscopic transport. Journal of
Physics: Condensed Matter, 20, 373101.
[178] Hicks, C. L., Wheatley, M. J., Godfrey, M. J., et al. 2018. Gardner transition in
physical dimensions. Physical Review Letters, 120, 225501.
[179] Hopkins, A. B., Stillinger, F. H., and Torquato, S. 2013. Disordered strictly jammed
binary sphere packings attain an anomalously large range of densities. Physical
Review E, 88, 022205.
[180] Hull, D., and Bacon, D. J. 2011. Introduction to dislocations. Elsevier.
[181] Hunter, G. L., and Weeks, E. R. 2012. The physics of the colloidal glass transition.
Reports on Progress in Physics, 75, 066501.
[182] Ikeda, A., Berthier, L., and Biroli, G. 2013. Dynamic criticality at the jamming
transition. The Journal of Chemical Physics, 138, 12A507.
[183] Ikeda, A., and Miyazaki, K. 2010. Mode-coupling theory as a mean-field description
of the glass transition. Physical Review Letters, 104, 255704.
[184] Ikeda, H., Miyazaki, K., Yoshino, H., et al. 2017. Decoupling phenomena and replica
symmetry breaking of binary mixtures. arXiv:1710.08373.
[185] Ikeda, H., Zamponi, F., and Ikeda, A. 2017. Mean field theory of the swap Monte
Carlo algorithm. The Journal of Chemical Physics, 147, 234506.
[186] Irving, J. H., and Kirkwood, J. G. 1950. The statistical mechanical theory of transport
processes. IV. The equations of hydrodynamics. The Journal of Chemical Physics,
18, 817.
[187] Janssen, H.-K. 1976. On a Lagrangean for classical field dynamics and renormal-
ization group calculations of dynamical critical properties. Zeitschrift für Physik B
Condensed Matter, 23, 377.
[188] Janssen, L. M. C. 2018. Mode-coupling theory of the glass transition: A primer.
Frontiers in Physics, 6, 97.
[189] Janssen, L. M. C., and Reichman, D. R. 2015. Microscopic dynamics of supercooled
liquids from first principles. Physical Review Letters, 115, 205701.
[190] Jaric, M. V. 2012. Introduction to the mathematics of quasicrystals. Elsevier.
[191] Jenssen, M., Joos, F., and Perkins, W. 2019. On the hard sphere model and sphere
packings in high dimensions. Forum of Mathematics, Sigma, 7, E1.
[192] Jin, Y., and Yoshino, H. 2017. Exploring the complex free-energy landscape of the
simplest glass by rheology. Nature Communications, 8, 14935.
[193] Jin, Y., Urbani, P., Zamponi, F., et al. 2018. A stability-reversibility map unifies
elasticity, plasticity, yielding and jamming in hard sphere glasses. Science Advances,
4, eaat6387.
314 References

[194] Johari, G. 2000. A resolution for the enigma of a liquids configurational entropy-
molecular kinetics relation. The Journal of Chemical Physics, 112, 8958.
[195] Johnson, K. L. 1987. Contact mechanics. Cambridge University Press.
[196] Joslin, C. 1982. Third and fourth virial coefficients of hard hyperspheres of arbitrary
dimensionality. The Journal of Chemical Physics, 77, 2701.
[197] Kabatiansky, G. A., and Levensthein, V. I. 1978. Bounds for packings on a sphere
and in space. Problems on Information Transmission, 14, 1.
[198] Kallus, Y. 2013. Statistical mechanics of the lattice sphere packing problem.
Physical Review E, 87, 063307.
[199] Kallus, Y., Marcotte, E., and Torquato, S. 2013. Jammed lattice sphere packings.
Physical Review E, 88, 062151.
[200] Kamenev, A. 2009. Field theory of non-equilibrium systems. Cambridge University
Press.
[201] Kauzmann, W. 1948. The glassy state and the behaviour of liquids at low tempera-
ture. Chemical Reviews, 43, 219.
[202] Kirkpatrick, S., Gelatt, C. D., and Vecchi, M. P. 1983. Optimization by simulated
annealing. Science, 220, 671.
[203] Kirkpatrick, T. R., and Thirumalai, D. 1987. Dynamics of the structural glass
transition and the p-spin-interaction spin-glass model. Physical Review Letters, 58,
2091.
[204] Kirkpatrick, T. R., and Thirumalai, D. 1988. Comparison between dynamical
theories and metastable states in regular and glassy mean-field spin models with
underlying first-order-like phase transitions. Physical Review A, 37, 4439.
[205] Kirkpatrick, T. R., and Thirumalai, D. 1989. Random solutions from a regular
density functional Hamiltonian: A static and dynamical theory for the structural glass
transition. Journal of Physics A: Mathematical and General, 22, L149.
[206] Kirkpatrick, T. R., and Wolynes, P. G. 1987. Connections between some kinetic and
equilibrium theories of the glass transition. Physical Review A, 35, 3072.
[207] Kirkpatrick, T. R., and Wolynes, P. G. 1987. Stable and metastable states in mean-
field Potts and structural glasses. Physical Review B, 36, 8552.
[208] Kirkpatrick, T. R., Thirumalai, D., and Wolynes, P. G. 1989. Scaling concepts for the
dynamics of viscous liquids near an ideal glassy state. Physical Review A, 40, 1045.
[209] Kirkwood, J. G., and Monroe, E. 1940. On the theory of fusion. The Journal of
Chemical Physics, 8, 845.
[210] Klein, W., and Frisch, H. 1986. Instability in the infinite dimensional hard sphere
fluid. The Journal of Chemical Physics, 84, 968.
[211] Koch, H., Radin, C., and Sadun, L. 2005. Most stable structure for hard spheres.
Physical Review E, 72, 016708.
[212] Kossevich, A. M. 1999. The crystal lattice: Phonons, solitons, dislocations. Wiley
Online Library.
[213] Kraichnan, R. H. 1962. Stochastic models for many-body systems. I. Infinite systems
in thermal equilibrium. Journal of Mathematical Physics, 3, 475.
[214] Krzakala, F., and Zdeborová, L. 2010. Following Gibbs states adiabatically–The
energy landscape of mean-field glassy systems. Europhysics Letters, 90, 66002.
[215] Krzakala, F., and Kurchan, J. 2007. Landscape analysis of constraint satisfaction
problems. Physical Review E, 76, 021122.
[216] Krzakala, F., and Zdeborová, L. 2013. Performance of simulated annealing in p-spin
glasses. Journal of Physics: Conference Series, 473, 12022.
 References 315

[217] Krzakala, F., Montanari, A., Ricci-Tersenghi, F., et al. 2007. Gibbs states and the set
of solutions of random constraint satisfaction problems. Proceedings of the National
Academy of Sciences, 104, 10318.
[218] Kurchan, J. 2003. Supersymmetry, replica and dynamic treatments of disordered
systems: A parallel presentation. Markov Processes and Related Fields, 9, 243.
[219] Kurchan, J., and Levine, D. 2010. Order in glassy systems. Journal of Physics A:
Mathematical and Theoretical, 44, 035001.
[220] Kurchan, J., Parisi, G., and Zamponi, F. 2012. Exact theory of dense amorphous
hard spheres in high dimension. I. The free energy. Journal of Statistical Mechanics:
Theory and Experiment, 2012, P10012.
[221] Kurchan, J., Parisi, G., Urbani, P., et al. 2013. Exact theory of dense amorphous hard
spheres in high dimension. II. The high density regime and the Gardner transition.
The Journal of Physical Chemistry B, 117, 12979.
[222] Landau, L. D., and Lifshitz, E. M. 1986. Course of theoretical physics, vol.7: Theory
of elasticity. Butterworth-Heinemann.
[223] Landau, L. D., and Lifshitz, E. M. 1980. Course of theoretical physics, vol.5:
Statistical physics. Butterworth-Heinemann.
[224] Larson, D., Katzgraber, H. G., Moore, M., et al. 2013. Spin glasses in a field: Three
and four dimensions as seen from one space dimension. Physical Review B, 87,
024414.
[225] Le Doussal, P., Müller, M., and Wiese, K. J. 2012. Equilibrium avalanches in spin
glasses. Physical Review B, 85, 214402.
[226] Lebowitz, J., and Penrose, O. 1964. Convergence of virial expansions. Journal of
Mathematical Physics, 5, 841.
[227] Lerner, E., During, G., and Wyart, M. 2013. Low-energy non-linear excitations in
sphere packings. Soft Matter, 9, 8252.
[228] Lerner, E., Düring, G., and Bouchbinder, E. 2016. Statistics and properties of
low-frequency vibrational modes in structural glasses. Physical Review Letters, 117,
035501.
[229] Leuzzi, L., Parisi, G., Ricci-Tersenghi, F., et al. 2008. Dilute one-dimensional spin
glasses with power law decaying interactions. Physical Review Letters, 101, 107203.
[230] Liao, Q., and Berthier, L. 2019. Hierarchical landscape of hard disk glasses. Physical
Review X, 9, 011049.
[231] Lin, J., Lerner, E., Rosso, A., et al. 2014. Scaling description of the yielding
transition in soft amorphous solids at zero temperature. Proceedings of the National
Academy of Sciences, 111, 14382.
[232] Liu, A., Nagel, S., Van Saarloos, W., et al. 2011. The jamming scenario – an
introduction and outlook. In Berthier, L., Biroli, G., Bouchaud, J.-P., et al. (eds),
Dynamical Heterogeneities and Glasses. Oxford University Press.
[233] Liu, A. J., and Nagel, S. R. 2001. Jamming and rheology: Constrained dynamics on
microscopic and macroscopic scales. CRC Press.
[234] Liu, A. J., and Nagel, S. R. 2010. The jamming transition and the marginally jammed
solid. Annual Review of Condensed Matter Physics, 1, 347.
[235] Lubachevsky, B. D., and Stillinger, F. H. 1990. Geometric properties of random disk
packings. Journal of Statistical Physics, 60, 561.
[236] Luban, M., and Baram, A. 1982. Third and fourth virial coefficients of hard
hyperspheres of arbitrary dimensionality. The Journal of Chemical Physics,
76, 3233.
[237] Lundow, P. H., and Markström, K. 2009. Critical behavior of the Ising model on the
four-dimensional cubic lattice. Physical Review E, 80, 031104.
316 References

[238] Maimbourg, T., and Kurchan, J. 2016. Approximate scale invariance in particle
systems: A large-dimensional justification. Europhysics Letters, 114, 60002.
[239] Maimbourg, T., Kurchan, J., and Zamponi, F. 2016. Solution of the dynamics of
liquids in the large-dimensional limit. Physical Review Letters, 116, 015902.
[240] Maimbourg, T., Sellitto, M., Semerjian, G., et al. 2018. Generating dense packings
of hard spheres by soft interaction design. SciPost Physics, 4, 1.
[241] Mangeat, M., and Zamponi, F. 2016. Quantitative approximation schemes for
glasses. Physical Review E, 93, 012609.
[242] Mari, R., and Kurchan, J. 2011. Dynamical transition of glasses: From exact to
approximate. The Journal of Chemical Physics, 135, 124504.
[243] Mari, R., Krzakala, F., and Kurchan, J. 2009. Jamming versus glass transitions.
Physical Review Letters, 103, 025701.
[244] Marinari, E., Parisi, G., Ricci-Tersenghi, F., et al. 2000. Replica symmetry breaking
in short-range spin glasses: Theoretical foundations and numerical evidences.
Journal of Statistical Physics, 98, 973.
[245] Martin, P. C., Siggia, E. D., and Rose, H. A. 1973. Statistical dynamics of classical
systems. Physical Review A, 8, 423.
[246] Martinelli, F. 1999. Lectures on Glauber dynamics for discrete spin models. In
Bernard, P. (ed), Lectures on probability theory and statistics. Springer-Verlag.
[247] Martiniani, S., Schrenk, K. J., Stevenson, J. D., et al. 2016. Turning intractable
counting into sampling: Computing the configurational entropy of three-dimensional
jammed packings. Physical Review E, 93, 012906.
[248] Maxwell, J. C. 1864. On the calculation of the equilibrium and stiffness of
frames. The London, Edinburgh, and Dublin Philosophical Magazine and Journal
of Science, 27, 294.
[249] Mehta, A. (ed). 1994. Granular matter: An interdisciplinary approach. Springer-
Verlag.
[250] Mézard, M. 1999. How to compute the thermodynamics of a glass using a cloned
liquid. Physica A, 265, 352.
[251] Mézard, M., and Montanari, A. 2009. Information, physics and computation. Oxford
University Press.
[252] Mézard, M., and Parisi, G. 1996. A tentative replica study of the glass transition.
Journal of Physics A: Mathematical and General, 29, 6515.
[253] Mézard, M., and Parisi, G. 2012. Glasses and replicas. In Wolynes, P. G., and
Lubchenko, V. (eds), Structural glasses and supercooled liquids: Theory, experiment
and applications. Wiley & Sons.
[254] Mézard, M., Parisi, G., and Virasoro, M. A. 1987. Spin glass theory and beyond.
World Scientific.
[255] Mézard, M., and Parisi, G. 1991. Replica field theory for random manifolds. Journal
de Physique I, 1, 809.
[256] Mézard, M., and Parisi, G. 1999. A first-principle computation of the thermodynam-
ics of glasses. The Journal of Chemical Physics, 111, 1076.
[257] Mézard, M., and Parisi, G. 2000. Statistical physics of structural glasses. Journal of
Physics: Condensed Matter, 12, 6655.
[258] Mizuno, H., Shiba, H., and Ikeda, A. 2017. Continuum limit of the vibrational
properties of amorphous solids. Proceedings of the National Academy of Sciences,
114, E9767.
[259] Mon, K. K., and Percus, J. K. 1999. Virial expansion and liquid–vapor critical points
of high dimension classical fluids. The Journal of Chemical Physics, 110, 2734.
 References 317

[260] Monasson, R. 1995. Structural glass transition and the entropy of the metastable
states. Physical Review Letters, 75, 2847.
[261] Montanari, A., and Semerjian, G. 2006. Rigorous inequalities between length and
time scales in glassy systems. Journal of Statistical Physics, 125, 23.
[262] Montanari, A., and Ricci-Tersenghi, F. 2003. On the nature of the low-temperature
phase in discontinuous mean-field spin glasses. The European Physical Journal B,
33, 339.
[263] Montanari, A., and Ricci-Tersenghi, F. 2004. Cooling-schedule dependence of the
dynamics of mean-field glasses. Physical Review B, 70, 134406.
[264] Moore, M., and Bray, A. J. 2011. Disappearance of the de Almeida–Thouless line in
six dimensions. Physical Review B, 83, 224408.
[265] Mora, T., Walczak, A. M., and Zamponi, F. 2012. Transition path sampling algorithm
for discrete many-body systems. Physical Review E, 85, 036710.
[266] Morita, T., and Hiroike, K. 1961. A new approach to the theory of classical fluids.
III: General treatment of classical systems. Progress of Theoretical Physics, 25, 537.
[267] Morse, P. K., and Corwin, E. I. 2014. Geometric signatures of jamming in the
mechanical vacuum. Physical Review Letters, 112, 115701.
[268] Moustrou, P. 2017. On the density of cyclotomic lattices constructed from codes.
International Journal of Number Theory, 13, 1261.
[269] Müller, M., and Wyart, M. 2015. Marginal stability in structural, spin, and electron
glasses. Annual Review of Condensed Matter Physics, 6, 177.
[270] Nakayama, D., Yoshino, H., and Zamponi, F. 2016. Protocol-dependent shear mod-
ulus of amorphous solids. Journal of Statistical Mechanics: Theory and Experiment,
2016, 104001.
[271] Nebe, G., and Sloane, N. J. A. 2015. Table of densest packings presently known.
www.math.rwth-aachen.de/∼Gabriele.Nebe/LATTICES/density.html. Accessed:
2018-05-15.
[272] Nicolas, A., Ferrero, E. E., Martens, K., et al. 2018. Deformation and flow of
amorphous solids: Insights from elastoplastic models. Reviews of Modern Physics,
90, 045006.
[273] Ninarello, A., Berthier, L., and Coslovich, D. 2017. Models and algorithms for the
next generation of glass transition studies. Physical Review X, 7, 021039.
[274] Nishimori, H. 2001. Statistical physics of spin glasses and information processing:
An introduction. Clarendon Press.
[275] Noya, E. G., and Almarza, N. G. 2015. Entropy of hard spheres in the close-packing
limit. Molecular Physics, 113, 1061.
[276] O’Hern, C. S., Langer, S. A., Liu, A. J., et al. 2002. Random packings of frictionless
particles. Physical Review Letters, 88, 075507.
[277] O’Hern, C. S., Silbert, L. E., Liu, A. J., et al. 2003. Jamming at zero temperature and
zero applied stress: The epitome of disorder. Physical Review E, 68, 011306.
[278] Ozawa, M., Kuroiwa, T., Ikeda, A., et al. 2012. Jamming transition and inherent
structures of hard spheres and disks. Physical Review Letters, 109, 205701.
[279] Ozawa, M., Berthier, L., and Coslovich, D. 2017. Exploring the jamming transition
over a wide range of critical densities. SciPost Physics, 3, 027.
[280] Ozawa, M., Berthier, L., Biroli, G., et al. 2018. Random critical point separates brittle
and ductile yielding transitions in amorphous materials. Proceedings of the National
Academy of Sciences, 115, 6656.
[281] Panchenko, D. 2013. The Sherrington-Kirkpatrick model. Springer Science &
Business Media.
[282] Parisi, G. 1988. Statistical field theory. Addison-Wesley.
318 References

[283] Parisi, G. 2008. On the most compact regular lattices in large dimensions: A
statistical mechanical approach. Journal of Statistical Physics, 132, 207.
[284] Parisi, G., and Slanina, F. 2000. Toy model for the mean-field theory of hard-sphere
liquids. Physical Review E, 62, 6554.
[285] Parisi, G. 1980. A sequence of approximated solutions to the SK model for spin
glasses. Journal of Physics A: Mathematical and General, 13, L115.
[286] Parisi, G. 2003. Glasses, replicas and all that. In Barrat, J.-L., Feigelman, M.,
Kurchan, J., et al. (eds), Slow relaxations and nonequilibrium dynamics in condensed
matter. Springer-Verlag.
[287] Parisi, G., and Potters, M. 1995. On the number of metastable states in spin glasses.
Europhysics Letters, 32, 13.
[288] Parisi, G., and Rizzo, T. 2013. Critical dynamics in glassy systems. Physical Review
E, 87, 012101.
[289] Parisi, G., and Sourlas, N. 2000. P-adic numbers and replica symmetry breaking.
The European Physical Journal B, 14, 535.
[290] Parisi, G., and Temesvári, T. 2012. Replica symmetry breaking in and around six
dimensions. Nuclear Physics B, 858, 293.
[291] Parisi, G., and Zamponi, F. 2006. Amorphous packings of hard spheres for large
space dimension. Journal of Statistical Mechanics: Theory and Experiment, 2006,
P03017.
[292] Parisi, G., and Zamponi, F. 2010. Mean-field theory of hard sphere glasses and
jamming. Reviews of Modern Physics, 82, 789.
[293] Parisi, G., Procaccia, I., Rainone, C., et al. 2017. Shear bands as manifestation of
a criticality in yielding amorphous solids. Proceedings of the National Academy of
Sciences, 114, 5577.
[294] Percus, A., Istrate, G., and Moore, C. (eds). 2006. Computational complexity and
statistical physics. Oxford University Press.
[295] Pinson, D., Zou, R. P., Yu, A. B., et al. 1998. Coordination number of binary mixtures
of spheres. Journal of Physics D: Applied Physics, 31, 457.
[296] Plefka, T. 1982. Convergence condition of the TAP equation for the infinite-ranged
Ising spin glass model. Journal of Physics A: Mathematical and General, 15, 1971.
[297] Presutti, E. 2008. Scaling limits in statistical mechanics and microstructures in
continuum mechanics. Springer Science & Business Media.
[298] Rainone, C., and Urbani, P. 2016. Following the evolution of glassy states under
external perturbations: The full replica symmetry breaking solution. Journal of
Statistical Mechanics: Theory and Experiment, 2016, P053302.
[299] Rainone, C., Urbani, P., Yoshino, H., et al. 2015. Following the evolution of hard
sphere glasses in infinite dimensions under external perturbations: Compression and
shear strain. Physical Review Letters, 114, 015701.
[300] Rammal, R., Toulouse, G., and Virasoro, M. A. 1986. Ultrametricity for physicists.
Reviews of Modern Physics, 58, 765.
[301] Reichman, D. R., and Charbonneau, P. 2005. Mode-coupling theory. Journal of
Statistical Mechanics: Theory and Experiment, 2005, P05013.
[302] Richert, R., and Angell, C. 1998. Dynamics of glass-forming liquids. V. On the link
between molecular dynamics and configurational entropy. The Journal of Chemical
Physics, 108, 9016.
[303] Rizzo, T. 2013. Replica-symmetry-breaking transitions and off-equilibrium dynam-
ics. Physical Review E, 88, 032135.
[304] Rizzo, T., and Voigtmann, T. 2015. Qualitative features at the glass crossover.
Europhysics Letters, 111, 56008.
 References 319

[305] Rodney, D., Tanguy, A., and Vandembroucq, D. 2011. Modeling the mechanics of
amorphous solids at different length scale and time scale. Modelling and Simulation
in Materials Science and Engineering, 19, 083001.
[306] Rogers, C. A. 1964. Packing and covering. Cambridge University Press.
[307] Rogers, C. A. 1947. Existence theorems in the geometry of numbers. Annals of
Mathematics, 48, 994.
[308] Rohrmann, R. D., Robles, M., de Haro, M. L., et al. 2008. Virial series for fluids
of hard hyperspheres in odd dimensions. The Journal of Chemical Physics, 129,
014510.
[309] Ruelle, D. 1982. Do turbulent crystals exist? Physica A, 113, 619.
[310] Ruelle, D. 1999. Statistical mechanics: Rigorous results. World Scientific.
[311] Ruijgrok, T. W., and Tjon, J. 1973. Critical slowing down and nonlinear response in
an exactly solvable stochastic model. Physica, 65, 539.
[312] Russo, L. 2004. The forgotten revolution: How science was born in 300 BC and why
it had to be reborn. Springer-Verlag.
[313] Sastry, S. 2000. Evaluation of the configurational entropy of a model liquid from
computer simulations. Journal of Physics: Condensed Matter, 12, 6515.
[314] Sausset, F., Biroli, G., and Kurchan, J. 2010. Do solids flow? Journal of Statistical
Physics, 140, 718.
[315] Scalliet, C., Berthier, L., and Zamponi, F. 2017. Absence of marginal stability in a
structural glass. Physical Review Letters, 119, 205501.
[316] Scalliet, C., Berthier, L., and Zamponi, F. 2019. Marginally stable phases in mean-
field structural glasses. Physical Review E, 99, 012107.
[317] Schmid, B., and Schilling, R. 2010. Glass transition of hard spheres in high
dimensions. Physical Review E, 81, 041502.
[318] Schreck, C. F., Mailman, M., Chakraborty, B., et al. 2012. Constraints and vibrations
in static packings of ellipsoidal particles. Physical Review E, 85, 061305.
[319] Schultz, A. J., and Kofke, D. A. 2014. Fifth to eleventh virial coefficients of hard
spheres. Physical Review E, 90, 023301.
[320] Schweizer, K. S., and Saltzman, E. J. 2003. Entropic barriers, activated hopping,
and the glass transition in colloidal suspensions. The Journal of Chemical Physics,
119, 1181.
[321] Sciortino, F. 2002. One liquid, two glasses. Nature Materials, 1, 145.
[322] Sciortino, F., Kob, W., and Tartaglia, P. 1999. Inherent structure entropy of super-
cooled liquids. Physical Review Letters, 83, 3214.
[323] Sciortino, F., and Tartaglia, P. 2005. Glassy colloidal systems. Advances in Physics,
54, 471.
[324] Seguin, A., and Dauchot, O. 2016. Experimental evidence of the Gardner phase in a
granular glass. Physical Review Letters, 117, 228001.
[325] Sellitto, M., and Zamponi, F. 2013. A thermodynamic description of colloidal
glasses. Europhysics Letters, 103, 46005.
[326] Seoane, B., and Zamponi, F. 2018. Spin-glass-like aging in colloidal and granular
glasses. Soft Matter, 14, 5222.
[327] Seoane, B., Reid, D. R., de Pablo, J. J., et al. 2018. Low-temperature anomalies of a
vapor deposited glass. Physical Review Materials, 2, 015602.
[328] Simon, B. 2014. The statistical mechanics of lattice gases. Princeton University
Press.
[329] Singh, Y., Stoessel, J. P., and Wolynes, P. G. 1985. Hard-sphere glass and the density-
functional theory of aperiodic crystals. Physical Review Letters, 54, 1059.
320 References

[330] Skoge, M., Donev, A., Stillinger, F. H., et al. 2006. Packing hyperspheres in high-
dimensional Euclidean spaces. Physical Review E, 74, 041127.
[331] Sommers, H.-J. 1983. Properties of Sompolinsky’s mean field theory of spin glasses.
Journal of Physics A: Mathematical and General, 16, 447.
[332] Sommers, H.-J. 1985. Parisi function q(x) for spin glasses near Tc . Journal de
Physique Lettres, 46, 779.
[333] Sommers, H.-J., and Dupont, W. 1984. Distribution of frozen fields in the mean-field
theory of spin glasses. Journal of Physics C: Solid State Physics, 17, 5785.
[334] Stillinger, F. H. 1988. Supercooled liquids, glass transitions, and the Kauzmann
paradox. The Journal of Chemical Physics, 88, 7818.
[335] Stillinger, F. H., and Weber, T. A. 1983. Dynamics of structural transitions in liquids.
Physical Review A, 28, 2408.
[336] Stillinger, F. H., Debenedetti, P. G., and Truskett, T. M. 2001. The Kauzmann
paradox revisited. The Journal of Physical Chemistry B, 105, 11809.
[337] Svidzinsky, A., Scully, M., and Herschbach, D. 2014. Bohrs molecular model, a
century later. Physics Today, 67, 33.
[338] Swallen, S. F., Kearns, K. L., Mapes, M. K., et al. 2007. Organic glasses with
exceptional thermodynamic and kinetic stability. Science, 315, 353.
[339] Szamel, G. 2017. Simple theory for the dynamics of mean-field-like models of glass-
forming fluids. Physical Review Letters, 119, 155502.
[340] Talagrand, M. 2003. Spin glasses: A challenge for mathematicians: Cavity and mean
field models. Springer Science & Business Media.
[341] Talagrand, M. 2010. Mean field models for spin glasses: Volume I: Basic examples.
Springer Science & Business Media.
[342] Thouless, D., Anderson, P., and Palmer, R. 1977. Solution of ‘Solvable model of a
spin glass’. Philosophical Magazine, 35, 593.
[343] Toda, M., Kubo, R., and Saito, N. 1992. Statistical physics I: Equilibrium statistical
mechanics.
[344] Torquato, S., and Stillinger, F. H. 2010. Jammed hard-particle packings: From Kepler
to Bernal and beyond. Reviews of Modern Physics, 82, 2633.
[345] Torquato, S., and Jiao, Y. 2009. Dense packings of the Platonic and Archimedean
solids. Nature, 460, 876.
[346] Urbani, P., and Zamponi, F. 2017. Shear yielding and shear jamming of dense hard
sphere glasses. Physical Review Letters, 118, 038001.
[347] Vagberg, D., Olsson, P., and Teitel, S. 2011. Glassiness, rigidity, and jamming of
frictionless soft core disks. Physical Review E, 83, 031307.
[348] Van Hecke, M. 2010. Jamming of soft particles: Geometry, mechanics, scaling and
isostaticity. Journal of Physics: Condensed Matter, 22, 033101.
[349] Van Kampen, N. G. 1992. Stochastic processes in physics and chemistry. Elsevier.
[350] van Meel, J. A., Charbonneau, B., Fortini, A., et al. 2009. Hard-sphere crystallization
gets rarer with increasing dimension. Physical Review E, 80, 061110.
[351] Vance, S. 2011. Improved sphere packing lower bounds from Hurwitz lattices.
Advances in Mathematics, 227, 2144.
[352] Venkatesh, A. 2013. A note on sphere packings in high dimension. International
Mathematics Research Notices, 2013, 1628.
[353] Viazovska, M. S. 2017. The sphere packing problem in dimension 8. Annals of
Mathematics, 185, 991.
[354] Wales, D. 2003. Energy landscapes: Applications to clusters, biomolecules and
glasses. Cambridge University Press.
 References 321

[355] Wang, W., Machta, J., Munoz-Bauza, H., et al. 2017. Number of thermodynamic
states in the three-dimensional Edwards-Anderson spin glass. Physical Review B,
96, 184417.
[356] Witten, E. 1980. Quarks, atoms, and the 1/N expansion. Physics Today, 33, 38.
[357] Wolynes, P., and Lubchenko, V. (eds). 2012. Structural glasses and supercooled
liquids: Theory, experiment, and applications. Wiley & Sons.
[358] Woodcock, L. 1997. Entropy difference between the face-centred cubic and hexag-
onal close-packed crystal structures. Nature, 385, 141.
[359] Wyart, M., Silbert, L., Nagel, S., et al. 2005. Effects of compression on the
vibrational modes of marginally jammed solids. Physical Review E, 72, 051306.
[360] Wyart, M., Nagel, S., and Witten, T. 2005. Geometric origin of excess low-frequency
vibrational modes in weakly connected amorphous solids. Europhysics Letters, 72,
486.
[361] Wyart, M. 2012. Marginal stability constrains force and pair distributions at random
close packing. Physical Review Letters, 109, 125502.
[362] Wyler, D., Rivier, N., and Frisch, H. L. 1987. Hard-sphere fluid in infinite dimen-
sions. Physical Review A, 36, 2422.
[363] Yaida, S., Berthier, L., Charbonneau, P., et al. 2016. Point-to-set lengths, local
structure, and glassiness. Physical Review E, 94, 032605.
[364] Yeo, J., and Moore, M. A. 2012. Renormalization group analysis of the M-p-spin
glass model with p = 3 and M = 3. Physical Review B, 85, 100405.
[365] Yerazunis, S., Cornell, S., and Wintner, B. 1965. Dense random packing of binary
mixtures of spheres. Nature, 207, 835.
[366] Yoshino, H. 2012. Replica theory of the rigidity of structural glasses. The Journal of
Chemical Physics, 136, 214108.
[367] Yoshino, H. 2018. Translational and orientational glass transitions in the large-
dimensional limit: A generalized replicated liquid theory and an application to
patchy colloids. arXiv:1807.04095.
[368] Yoshino, H., and Mézard, M. 2010. Emergence of rigidity at the structural glass
transition: A first-principles computation. Physical Review Letters, 105, 015504.
[369] Yoshino, H., and Zamponi, F. 2014. Shear modulus of glasses: Results from the full
replica-symmetry-breaking solution. Physical Review E, 90, 022302.
[370] Zdeborová, L., and Krzakala, F. 2007. Phase transitions in the coloring of random
graphs. Physical Review E, 76, 031131.
[371] Zdeborová, L., and Krzakala, F. 2010. Generalization of the cavity method for
adiabatic evolution of Gibbs states. Physical Review B, 81, 224205.
[372] Zdeborová, L., and Krzakala, F. 2016. Statistical physics of inference: Thresholds
and algorithms. Advances in Physics, 65, 453.
[373] Zhang, C., and Pettitt, B. M. 2014. Computation of high-order virial coefficients in
high-dimensional hard-sphere fluids by Mayer sampling. Molecular Physics, 112,
1427.
 Index

avalanche, 163, 176 Einstein relation, 75, 82

elasticity, 290, 295, 302
belief propagation, 9 energy barrier, 255, 258
Bessel function, 54, 85, 86 equilibrium state, 19, 27, 32, 44, 62, 100, 121, 141,
Bethe lattice, 2 162, 175, 202
boundary conditions, 2, 15, 19, 26, 101, 234, extreme value statistics, 272
242, 259, 265, 266
Bragg peaks, 233 first order phase transition, 62, 143, 145, 240, 241
Bravais lattice, 232–236, 243, 248 fluctuation-dissipation relation, 74, 75, 83
Brownian motion, 69, 77, 78, 91, 102 Fokker–Planck equation, 97
buckler, 270, 271, 286 free energy barrier, 31, 175, 241, 243, 246
fully connected model, 2, 32
catastrophic potential, 40
cavity method, 9, 144 Gardner transition, 146, 185, 187, 189–191, 193–195,
central limit theorem, 77, 78, 103, 116 224, 226, 251, 257–259, 273, 297–299, 301, 302
close packing, 231, 232, 234, 241 Gibbs rule, 143
coloured noise, 70, 76, 77, 83 glass close packing, 130, 219, 224, 246, 247, 287
commutative algebra, 147, 149, 153, 177 Goldstone mode, 176
complexity, 199, 201–207, 210, 213, 219–221, gradient descent, 253, 255, 256, 258, 259, 262, 287
226, 227 Green–Kubo relation, 75, 89, 93
configurational entropy, 201
correlation function, 10, 15, 17, 19, 23, 42, 43, 45, Hamiltonian dynamics, 68, 74, 93
67, 68, 71–73, 75–77, 89, 127, 176 hydrodynamics, 67, 91
cost function, 252, 253, 258
critical exponent, 12, 13, 92, 195, 196, 260, 263, 268, ideal glass, 210, 216, 217, 227, 247
269, 271–273, 275, 278–280, 287, 301 interaction graph, 2
crystallisation, 37, 45, 62, 67, 108, 141, 231, 232, 240, interface, 21, 22, 32, 162
241, 243, 246, 248, 251, 260, 267, 290 Ising model, 1, 2, 10, 19, 26, 28, 77, 121, 141
isostaticity, 263, 265–269, 271, 273, 280, 284, 287
de Almeida-Thouless transition, 146, 172, 173,
175, 178 Kauzmann transition, 199, 202, 203, 207, 209, 215,
De Broglie wavelength, 39 216, 219, 224, 225, 227, 247
density of vibrational states, 285–287 kinetic theory, 67
detailed balance, 24, 28, 70, 74, 98 Kirkwood instability, 57
dilatancy, 290, 295, 298, 301, 302
dynamical arrest, 57, 82, 85, 87, 91, 94, 99–101, Landau theory, 12, 34, 62
104, 126, 129, 132, 193, 199, 202, 206, 216, Langevin dynamics, 69, 74, 76, 90, 93, 97, 253, 255
227, 246, 256, 258, 290, 296, 297 large dimensional expansion, 6, 10, 12, 23, 27, 63
lattice, 2, 15, 34, 144, 232, 233, 235, 241, 243
Edwards ensemble, 219, 221, 227, 247 Legendre transform, 5–7, 23, 42, 43, 205, 207, 226
effective temperature, 216 linear programming, 238

322
 Index 323

Liouville theorem, 69 rotational invariance, 42, 44, 64, 77, 84, 108, 109,
local dynamics, 24, 28 112, 113, 117, 141
local search algorithm, 252, 258, 260, 287 roughening transition, 21
long-range potential, 40
satisfiability problem, 252, 260
Markov chain, 70 satisfiability transition, 260, 262, 287
master equation, 28, 30 saturated packing, 235, 237
Maxwell construction, 22, 27 second-order phase transition, 10, 11, 13, 18, 93,
Mayer function, 47, 49, 51, 59, 110, 111, 145, 176, 194
114, 116, 128 simulated annealing, 255, 287
metabasin, 158–161, 163, 226, 273, 297 spinodal point, 27, 130, 131, 140, 189, 192, 257,
metastable state, 26, 27, 31, 38, 100, 241, 297, 299, 303
248, 292 stable potential, 40, 51
mobility, 73, 75, 82 static structure factor, 45, 239
mode-coupling theory, 91, 94, 127 Stokes–Einstein relation, 91, 93, 94
molecular liquid, 108–110, 114, 121, 132 surface tension, 22, 25, 27, 32, 241
Monte Carlo dynamics, 70, 74, 93, 242 susceptibility, 4, 10, 18, 72, 93, 94, 145, 176,
273, 298, 300
Newtonian dynamics, 69, 76, 102
nucleation, 32, 162 tempered potential, 39, 51, 58
time reversal, 73, 75, 97
path integral, 67, 76 time-translational invariance, 68, 73, 102
periodic packing, 232, 234, 239, 242, 251 translational invariance, 2, 9, 11, 16, 23, 24, 26, 42,
phase coexistence, 22, 23, 27, 33, 44, 62, 141, 44, 50, 71, 107, 109, 110, 141, 266
143, 177 transport coefficient, 73, 75
phase diagram, 128, 130, 133, 140, 177, 181, triple point, 143
188, 191, 199, 214, 221, 241
Poisson process, 163, 177, 253 ultrametricity, 146, 159–164, 177, 297
polydispersity, 37, 38
van Hove function, 71, 72
quasicrystal, 233 virial expansion, 46, 50, 57, 63, 87, 107, 110, 114,
quenched disorder, 107, 125, 132, 140 115, 137, 210, 237, 239
virial theorem, 214
radial distribution function, 42, 46, 58, 185, 213, 214,
239, 269 Ward identity, 176
random graph, 2 white noise, 69, 87
rattler, 262, 263, 265, 269, 270, 286
response function, 72 yielding, 290, 299, 301, 303