Out of equilibrium dynamics of complex

systems
Leticia F. Cugliandolo
Universite Pierre et Marie Curie - Paris VI
Laboratoire de Physique Theorique et Hautes Energies
February 28, 2013

Contents
1 Introduction
1.1 Falling out of equilibrium . . . . .
1.2 Nucleation . . . . . . . . . . . . . .
1.3 Phase ordering kinetics . . . . . . .
1.4 Critical dynamics . . . . . . . . . .
1.5 Structural disorder: glassy physics
1.6 Quenched disorder: still glassiness
1.7 Random manifolds . . . . . . . . .
1.8 Aging . . . . . . . . . . . . . . . .
1.9 Driven systems . . . . . . . . . . .
1.10 Interdisciplinary aspects . . . . . .
1.10.1 Optimization problems . . .
1.10.2 Biological applications . . .
1.11 Summary and interesting questions

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

7
7
9
10
11
11
20
21
22
23
24
24
27
30

2 Modeling
2.1 Fluctuations . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 The classical reduced partition function . . . . . . . . .
2.3 The Langevin equation . . . . . . . . . . . . . . . . . . .
2.3.1 Langevins Langevin equation . . . . . . . . . . .
2.3.2 Derivation of the Langevin equation . . . . . . .
2.3.3 Irreversibility and dissipation. . . . . . . . . . . .
2.3.4 Smoluchowski (overdamped) limit . . . . . . . .
2.3.5 Discretization of stochastic differential equations
2.3.6 Markov character . . . . . . . . . . . . . . . . . .
2.3.7 Generation of memory . . . . . . . . . . . . . . .
2.4 The basic processes . . . . . . . . . . . . . . . . . . . . .
2.4.1 A constant force . . . . . . . . . . . . . . . . . .
2.4.2 Relaxation in a quadratic potential . . . . . . . .
2.4.3 Thermally activated processes . . . . . . . . . . .
2.5 Driven systems . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

32
32
32
34
34
35
39
40
40
41
41
41
41
49
52
54

3 Dynamics at or through a phase transition

3.1 Snapshots . . . . . . . . . . . . . . . . . . .
3.2 Relaxation, equilibration and reversal times
3.2.1 Quench from T Tc to T > Tc . . .
3.2.2 Quench from T Tc to T Tc . . .
3.2.3 Summary . . . . . . . . . . . . . . .
3.3 Growing length and dynamic scaling . . . .
3.4 Critical coarsening . . . . . . . . . . . . . .
3.5 Sub-critical coarsening . . . . . . . . . . . .
3.5.1 Dynamic scaling hypothesis . . . . .

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

56
57
59
60
61
62
63
65
67
67

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

3.5.2

R(t) in clean one dimensional cases with non-conserved order

parameter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5.3 R(t) in non-conserved order parameter curvature dynamics in
d>2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5.4
R(t) in conserved order parameter dynamics and the role of
bulk diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5.5 Crossover between critical and sub-critical coarsening . . . . .
3.5.6 Role of weak disorder: thermal activation . . . . . . . . . . . .
3.5.7 Temperature-dependent effective exponents . . . . . . . . . . .
3.6 Time-dependent Ginzburg-Landau description . . . . . . . . . . . . . .
3.6.1 Short-time dynamics . . . . . . . . . . . . . . . . . . . . . . . .
3.6.2 R(t) in non-conserved order parameter curvature dynamics in
d>2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.6.3 Scaling functions for subcritical coarsening . . . . . . . . . . .
3.6.4 Breakdown of dynamic scaling . . . . . . . . . . . . . . . . . .
3.6.5 Annealing: crossover from critical to subcritical coarsening . .
3.7 The large N approximation . . . . . . . . . . . . . . . . . . . . . . . .
3.8 The 2d xy model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.10 Nucleation and growth . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.11 Elastic manifold dynamics . . . . . . . . . . . . . . . . . . . . . . . . .
3.11.1 Scale invariance . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.11.2 Solid-on-solid models of surface growth . . . . . . . . . . . . .
3.11.3 Continuous models . . . . . . . . . . . . . . . . . . . . . . . . .
3.11.4 Dynamic scaling at work . . . . . . . . . . . . . . . . . . . . . .
3.11.5 Scale invariance . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.11.6 Universality classes . . . . . . . . . . . . . . . . . . . . . . . . .
3.11.7 Roughening transition . . . . . . . . . . . . . . . . . . . . . . .
4 Disordered systems: statics
4.1 Quenched and annealed disorder . . . . . . . . . . .
4.2 Bond disorder: the case of spin-glasses . . . . . . . .
4.2.1 Lack of homogeneity . . . . . . . . . . . . . .
4.2.2 Frustration . . . . . . . . . . . . . . . . . . .
4.2.3 Gauge invariance . . . . . . . . . . . . . . . .
4.2.4 Self-averageness . . . . . . . . . . . . . . . . .
4.3 Models with quenched disorder . . . . . . . . . . . .
4.3.1 Spin-glass models . . . . . . . . . . . . . . . .
4.3.2 Random ferromagnets . . . . . . . . . . . . .
4.3.3 Random manifolds . . . . . . . . . . . . . . .
4.4 The spin-glass transition . . . . . . . . . . . . . . . .
4.4.1 The simplest order parameter . . . . . . . . .
4.4.2 Pure states and more subtle order parameters
4.4.3 Pinning fields . . . . . . . . . . . . . . . . . .
3

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

70
70
70
71
71
73
74
76
76
78
78
78
78
82
84
86
87
88
88
89
92
92
93
95
96
96
96
97
97
98
98
100
100
102
103
104
104
106
108

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

109
110
110
111
113
114
116
117
117
124
127
128
128
130
130
131
137
137
137
141

5 Formalism: dynamic generating functional and symmetries

5.1 Classical dynamics: generating functional . . . . . . . . . . . .
5.2 Generic correlation and response. . . . . . . . . . . . . . . . . .
5.2.1 The linear response as a variable-noise correlation . . .
5.3 Time-reversal . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4 An equilibrium symmetry . . . . . . . . . . . . . . . . . . . . .
5.4.1 Invariance of the measure . . . . . . . . . . . . . . . . .
5.4.2 Invariance of the integration domain . . . . . . . . . . .
5.4.3 Invariance of the action functional . . . . . . . . . . . .
5.4.4 Invariance of the Jacobian (Grassmann variables) . . . .
5.5 Consequences of the transformation . . . . . . . . . . . . . . .
5.5.1 The fluctuation-dissipation theorem . . . . . . . . . . .
5.5.2 Fluctuation theorems . . . . . . . . . . . . . . . . . . .
5.6 Equations on correlations and linear responses . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

143
143
145
146
147
147
147
148
148
149
149
149
150
153

6 Dynamic equations
6.1 Connection with the replica formalism . . . . . .
6.2 Average over disorder . . . . . . . . . . . . . . .
6.3 The equations . . . . . . . . . . . . . . . . . . . .
6.3.1 Supersymmetry and saddle-points . . . .
6.3.2 Field equations . . . . . . . . . . . . . . .
6.3.3 The thermodynamic limit and time-scales
6.3.4 Single spin equation . . . . . . . . . . . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

154
154
154
155
155
160
161
162

4.5

4.6

4.7
4.8

4.9

4.4.4 Divergent susceptibility . . . . . . . . . . . .

4.4.5 Calorimetry . . . . . . . . . . . . . . . . . . .
4.4.6 Critical scaling . . . . . . . . . . . . . . . . .
The TAP approach . . . . . . . . . . . . . . . . . . .
4.5.1 The complexity or configurational entropy . .
4.5.2 Weighted averages . . . . . . . . . . . . . . .
Metastable states in two families of models . . . . .
4.6.1 High temperatures . . . . . . . . . . . . . . .
4.6.2 Low temperatures . . . . . . . . . . . . . . .
The replica method . . . . . . . . . . . . . . . . . . .
Saddle-point evaluation . . . . . . . . . . . . . . . .
4.8.1 Replica symmetry (RS) . . . . . . . . . . . .
4.8.2 One step replica symmetry breaking (1RSB)
4.8.3 k-step replica symmetry breaking (kRSB) . .
4.8.4 Full replica symmetry breaking . . . . . . . .
4.8.5 Interpretation of replica results . . . . . . . .
Finite dimensional systems . . . . . . . . . . . . . .
4.9.1 The Griffiths phase . . . . . . . . . . . . . . .
4.9.2 Droplets and domain-wall stiffness . . . . . .
4.9.3 The droplet theory . . . . . . . . . . . . . . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

162
164
165
167

7 Glassy dynamics: Generic results

7.1 The weak-ergodicity breaking scenario . . . . . . . .
7.2 The weak long-term memory scenario . . . . . . . .
7.3 Slow time-reparametrization invariant dynamics . . .
7.4 Correlation scales . . . . . . . . . . . . . . . . . . . .
7.4.1 Properties . . . . . . . . . . . . . . . . . . . .
7.4.2 Definition of a characteristic time . . . . . . .
7.5 Modifications of fdt . . . . . . . . . . . . . . . . . .
7.5.1 Time domain . . . . . . . . . . . . . . . . . .
7.5.2 Frequency domain . . . . . . . . . . . . . . .
7.5.3 Time-reparametrization invariant formulation
7.5.4 fdt part . . . . . . . . . . . . . . . . . . . .
7.5.5 Diffusion . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

169
169
170
172
172
173
177
178
178
179
180
182
182

8 Solution to mean-field models

8.1 Numerical solution . . . . . . . . . . . . .
8.2 Solution at high temperatures . . . . . . .
8.3 Solution at low-T . . . . . . . . . . . . . .
8.3.1 The Lagrange multiplier . . . . . .
8.3.2 The stationary regime . . . . . . .
8.3.3 The aging regime . . . . . . . . . .
8.3.4 The Edwards-Anderson parameter
8.3.5 Fluctuation - dissipation relation .
8.3.6 Discontinuous classical transition .
8.3.7 The classical threshold level . . . .
8.3.8 Two p models . . . . . . . . . . . .
8.3.9 sk model and similar . . . . . . . .
8.3.10 Mode dependence . . . . . . . . .
8.3.11 Quantum fluctuations . . . . . . .
8.3.12 Driven dynamics . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

184
184
185
187
187
188
189
190
191
192
192
192
194
194
194
195

9 Effective temperature measurements

9.1 Diffusion . . . . . . . . . . . . . . . . . . .
9.2 Coarsening . . . . . . . . . . . . . . . . .
9.3 Critical dynamics . . . . . . . . . . . . . .
9.4 Quenches to the lower critical dimension .
9.5 Relaxation in structural glasses . . . . . .
9.5.1 Simulations of microscopic models
9.5.2 Kinetically constrained models . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

197
197
198
199
201
202
202
203

6.4
6.5
6.6

6.3.5 Diagrammatic techniques . . . . .

The mode coupling approximation (mca)
mca and disordered models . . . . . . . .
mca for super-cooled liquids and glasses .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

9.5.3 Experiments . . . . . . . . . . . .
Relaxation in frustrated magnetic systems
9.6.1 Remarks on model systems . . . .
9.6.2 Simulations . . . . . . . . . . . . .
9.6.3 Experiments . . . . . . . . . . . .
9.7 Driven liquids and glasses . . . . . . . . .
9.8 Granular matter . . . . . . . . . . . . . .
9.9 Activated dynamics . . . . . . . . . . . .
9.10 Biological systems . . . . . . . . . . . . .
9.6

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

203
205
205
206
207
207
208
209
209

10 Conclusions

211

A Conventions
A.1 Fourier transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.2 Commutation relations . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.3 Time ordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

216
216
216
216

B Classical statics: the reduced partition function

217

C The instanton calculation

218

Discrete MSRJD for additive noise: Stratonovich prescription

Mid-point discretization
219
D.1 Construction of the MSRJD action . . . . . . . . . . . . . . . . . . . . 220
D.2 Evaluation of the Jacobian . . . . . . . . . . . . . . . . . . . . . . . . . 221
D.3 Markovian case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
D.4 Non Markovian case . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
D.5 Discrete MSRJD for multiplicative noise: Stratonovich prescription
Mid-point discretization . . . . . . . . . . . . . . . . . . . . . . . . . . 223

E Mean-field theory for ferromagnets

224

F Grassmann variables and supersymmetry

229

Introduction

1.1

Falling out of equilibrium

In standard condensed matter or statistical physics focus is set on equilibrium

systems. Microcanonical, canonical or grand canonical ensembles are used depending
on the conditions one is interested in. The relaxation of a tiny perturbation away from
equilibrium is also sometimes described in textbooks and undergraduate courses.
More recently, attention has turned to the study of the evolution of similar macroscopic systems in far from equilibrium conditions. These can be achieved by
changing the properties of the environment (e.g. the temperature) in a canonical setting or by changing a parameter in the systems Hamiltonian in a microcanonical one.
The procedure of rapidly (ideally instantaneously) changing a parameter is called a
quench. Right after both types of quenches the initial configuration is not one of
equilibrium at the new conditions and the systems subsequently evolve in an out of
equilibrium fashion. The relaxation towards the new equilibrium (if possible) could
be fast (and not interesting for our purposes) or it could be very slow (and thus the
object of our study). There are plenty of examples of the latter. Dissipative ones
include systems quenched through a phase transition and later undergoing domain
growth, and problems with competing interactions that behave as glasses. Energy
conserving ones are of great interest at present due to the rapid growth of activity in
cold-atom systems.
Out of equilibrium situations can also be established by driving a system, that
otherwise would reach equilibrium in observable time-scales, with an external perturbation. In the context of macroscopic systems an interesting example is the one of
sheared complex liquids. Yet another interesting case is the one of powders that stay
in static metastable states unless externally perturbed by tapping, vibration or shear
that drives them out of equilibrium and makes them slowly evolve towards more compact configurations. Such situations are usually called non-equilibrium steady states
(NESS). Small systems can also be driven out of equilibrium with external perturbations. Transport in nano-structures is the quantum (small) counterpart phenomenon
of these cases, also of special interest at present.
Our interest is, therefore, in macroscopic complex1 systems:
With out of equilibrium initial condition. These include
open dissipative systems;
closed systems with energy conserving dynamics.
with external driving forces.
1 Complex

simply means here not easy to understand.

A number of questions one would like to give an answer to naturally arise. Among
these are:
Is the (instantaneous) structure out of equilibrium similar to the one in equilibrium (at some temperature, pressure, etc.)?
What microscopic/mesoscopic relaxation mechanism takes place after
the quench?
Does the system quickly settle into a stationary state? In more technical terms,
is there a finite relaxation time to reach a steady state and which are the
properties of the system on which the steady state depends?
What is the microscopic/mesoscopic dynamics in non-equilibrium steady
states when these are reached?
Can one describe the states of the system sometime after the quench with some
kind of effective equilibrium-like measure?
Are there thermodynamic concepts, such as temperature, entropy, freeenergy, playing a r
ole in the non-equilibrium relaxation? Under which conditions?
One notices that some of these questions apply to the free as well as to the driven
dynamics.

In the last 20 years or so a rather complete theory of the dynamics of classical

macroscopic systems evolving slowly in a small entropy production limit
(asymptotic regime after a quench, small drives), that encompasses the situations
described above has been developed [1, 2]. This is a mean-field theory type in
the sense that it applies strictly to models with long-range interactions or in the
infinite dimensional limit. It is, however, expected that some aspects of it also apply
to systems with short-range interactions although with some caveats. A number of
finite dimensional problems have been solved demonstrating this fact.
ole. For
In several cases of practical interest, quantum effects play an important r
instance, glassy phases at very low temperatures have been identified in a large variety
of materials (spin-glass like systems, interacting electrons with disorder, materials
undergoing super-conductor transitions, metallic glasses, etc.). Clearly, the driven
case is also very important in systems with quantum fluctuations. Take for instance
a molecule or an interacting electronic system driven by an external current applied
via the coupling to leads at different chemical potential. It is then necessary to settle
whether the approach developed and the results obtained for the classical dynamics
in a limit of small entropy production carry through when quantum fluctuations are
included.
In these notes we start by exposing some examples of the phenomenology of
out of equilibrium dynamics we are interested in. We focus on classical problems and
their precise setting. We introduce nucleation [3], phase ordering kinetics [4], critical
dynamics [5] structural glasses [6] and disordered systems [7, 8]. We also discuss
8

some interdisciplinary problems that have many points in common with glassy physics
including optimization problems [9], neural networks [10] and active matter [11].
Next we go into the formalism used to deal with these problems. The basic
techniques used to study classical glassy models with or without disorder are relatively
well documented in the literature (the replica trick, scaling arguments and droplet
theories, the dynamic functional method used to derive macroscopic equations from
the microscopic Langevin dynamics, functional renormalization, Monte Carlo and
molecular dynamic numerical methods). On the contrary, the techniques needed
to deal with the statics and dynamics of quantum macroscopic systems are much
less known in general. I will briefly discuss the role played by the environment in a
quantum system and introduce and compare the equilibrium and dynamic approaches.
Concretely, we recall some features of the Langevin formalism and its generating
function. We dwell initially with some emblematic aspects of classical macroscopic
systems slowly evolving out of equilibrium: concerning models. We focus on two, that
are intimately related: the O(N ) model in the large N limit that is used to describe
coarsening phenomena, and the random manifold, that finds applications to
many physical problems like charge density waves, high-Tc superconductors, etc. Both
problems are of field-theoretical type and can be treated both classically and
quantum mechanically. These two models are ideal for the purpose of introducing
and discussing formalism and some basic ideas we would wish to convey in these
lectures. Before entering the technical part we explain the two-fold meaning of the
word disorder by introducing the glass problem and some of the numerous questions
it raises.

1.2

Nucleation

When a system with a first order phase transition is taken to a region in the
phase diagram in which it is still locally stable but metastable with respect to the
new absolute minimum of the free-energy, its evolution towards the new equilibrium
state occurs by nucleation of the stable phase. The theory of simple nucleation [3] is
well established and the time needed for one bubble of the stable state to conquer the
sample grows as an exponential of the free-energy difference between the metastable
and the stable states over the thermal energy available, kB T . Once the bubble has
reached a critical size that also depends on this free-energy difference it very rapidly
conquers the full sample and the system reaches the stable state. The textbook
example is the one of a magnetic system, e.g. an Ising model, in equilibrium under a
magnetic field that is suddenly reversed. The sample has to reverse its magnetization
but this involves a nucleation process of the kind just explained. Simple nucleation is
therefore not very interesting to us but as soon as multiple nucleation and competition
between different states intervenes the problem becomes rapidly hard to describe
quantitatively and very relevant to the mean-field theory of fragile structural glasses
that we will discuss.

1.3

Phase ordering kinetics

Figure 1: Four images after a quench of a two species mixture (of glasses!) that tends
to demix under the new working conditions. Images courtesy of E. Gouillart (St.
Gobain), D. Bouttes and D. Vandembroucq (ESPCI).
Choose a system with a well-understood equilibrium phase transition and take
it across the critical point (second order phase transition) very quickly by tuning a
control parameter. If the system is taken from its disordered (mixed) phase to its
ordered (demixed) phase the sample will tend to phase separate in the course of time
to approach the ideal equilibrium configuration under the new conditions. Such an
example of phase ordering kinetics [4], i.e. phase separation, is shown in Fig. 1.
None of the two species disappears, they just separate. This is such a slow process
that the time needed to fully separate the mixture diverges with the size of the sample,
as we will see later on.
Another example of phase ordering kinetics is given by the crystal grain growth
sketched in the left-most panel in Fig. 2. Grains are formed by pieces of the lattice
with the same orientation. Boundaries between these grains are drawn with lines in
the figure. The other panels show snapshots of a 2d isotropic ferromagnetic Potts
model
X
HJ [{si }] = J
(1.1)
J >0,
si sj
hiji

with si = 1, . . . , q = 8 quenched below its first order phase transition at the initial
time t = 0 from a configuration in equilibrium at infinite temperature. The quench is
done well below the region of metastability and the dynamics are the ones of domain
growth. Indeed, domains of neighboring spin ordered in the same direction grow in
the course of time. This is clear from the subsequent snapshots taken at t = 128 MCs
and t = 1024 MCs. This model has been used to mimic this kind of physical process
when the number of spin components becomes very large, q 1. Note that the
number of spins of each kind is not conserved along the systems evolution.
These problems are simple in that the systems try to order in configurations that
are easy to visualize and to characterize. It is also quite clear from the figures that two
kinds of processes coexist: what happens within the domains, far from the interfaces,
10

Figure 2: Grain boudaries in crystal growth. Three snapshots of the 2d ferromagnetic

Potts model with q = 8 quenched below its (first order) phase transition to T = Tc /2.
The times at which the images were taken are t = 0, 128, 1024 MCs. Data from M.
P. Loureiro, J. J. Arenzon (Porto Alegre, Brazil, and LFC.
and what the interfaces do. We will come back to this very important issue. To
conclude phase ordering kinetics are rather well understood qualitatively although a
full quantitative description is hard to develop as the problem is set into the form of
a non-linear field theory with no small parameter.

1.4

Critical dynamics

In critical quenches [5], patches with equilibrium critical fluctuations grow in

time but their linear extent never reaches the equilibrium correlation length that
diverges. Clusters of neighboring spins pointing in the same direction of many sizes
are visible in the figures and the structure is quite intricate with clusters within
clusters and so on and so forth. The interfaces look pretty rough too. A comparison
between critical and sub-critical coarsening are shown in Figs. 3 and 4.
Critical slowing down implies that the equilibrium relaxation time diverges
close to the phase transition as a power law of the distance to criticality
(T Tc )z

(1.2)

with the exponent that controls the divergence of the correlation length and z the
dynamic critical exponent. This problem has, though, a small parameter (the distance
to criticality) that allows for a perturbative treatment and a dynamic renormalisation
group approach that is very successful.

1.5

Structural disorder: glassy physics

While the understanding of equilibrium phases, the existence of phase transitions

as well as the characterization of critical phenomena are well understood in clean
systems, as soon as competing interactions or geometric frustration are in-

11

200

200

data

200

data

150

150

150

100

100

100

50

50

50

50

100

150

200

50

100

150

200

data

50

100

150

200

Figure 3: Monte Carlo simulations of a 2d Ising model. Three snapshots at t =

1, 3 105 , 3 106 MCs after a quench to Tc . Data from T. Blanchard, LFC and M.
Picco.
200

200

data

200

data

150

150

150

100

100

100

50

50

50

50

100

150

200

50

100

150

200

data

50

100

150

200

Figure 4: Monte Carlo simulations of a 2d Ising model. Three snapshots at t =

1, 3 105 , 3 106 MCs after a quench to 0.5 Tc . Thermal fluctuations within the
domains are visible. Data from T. Blanchard, LFC and M. Picco.
cluded one faces the possibility of destroying this simple picture by giving way to
novel phenomena like glassy behavior [6].
Glassy systems are usually dissipative, that is to say in contact with a much
larger environment, that has a well defined temperature and with which the systems
in question can exchange heat. We deal with open dissipative systems here.
Competing interactions in physical systems can be dynamic, also called annealed,
or quenched. A simple example illustrates the former: the Lennard-Jones potential2 ,
V (r) = V0 [(r0 /r)a (r0 /r)b ]

(1.3)

with usually, a = 12 and b = 6 (see Fig. 5-left) that gives an effective interaction
2 The first term is chosen to take care of a quantum effect due to Pauli repulsion in a phenomenological way, the asymptotically leading attractive term is the van der Waals contribution when b = 6.

12

0.75

LJ potential

0.25

-0.25

-0.75
0

0.5

1.5

2.5

Figure 5: Left: The Lennard-Jones potential. Right: the Edwards-Anderson 3d spinglass.

between soft3 particles in a liquid, has a repulsive and an attractive part, depending
on the distance between the particles, a set of dynamic variables. In this example,
the interactions depend on the positions of the particles and evolve with them.
When competing interactions are present the low-temperature configurations may
look disordered but still have macroscopic properties of a kind of crystalline state.
Again, cooling down a liquid to obtain a glass is helpful to exemplify what we mean
here: the liquid cannot support stress and flows while the glass has solid-like properties
as crystals, it can support stress and does not easily flow in reasonable time-scales
(this is why glasses can be made of glass!). However, when looked at a microscopic
scale, one does not identify any important structural difference between the liquid
and the glass: no simple long-range structural order has been identified for glasses.
Moreover, there is no clear evidence for a phase transition between the liquid and the
glass. At present one can only talk about a dynamic crossover. The glassy regime is,
however, usually called a glassy phase and it is sometimes said to be a disordered
phase due to the lack of a clear structural order this does not mean that there is
no order whatsoever (see Fig. 7 for an example of a system with a liquid, a crystal
and a glassy phase). Lennard-Jones binary mixtures are prototypical examples of
systems that undergo a glass transition (or crossover) when cooled across the glass
temperature Tg or when compressed across a density ng [6].
There are many types of glasses and they occur over an astounding range of scales
from macroscopic to microscopic. See Fig. 8 for some images. Macroscopic examples
include granular media like sand and powders. Unless fluidized by shaking or during flow these quickly settle into jammed, amorphous configurations. Jamming can
also be caused by applying stress, in response to which the material may effectively
convert from a fluid to a solid, refusing further flow. Temperature (and of course
quantum fluctuations as well) is totally irrelevant for these systems since the grains
3 Soft means that the particles can overlap at the price of an energy cost. In the case this is
forbidden one works with hard particles.

13

Figure 6: A crystal in a 2d colloidal suspension of hard spheres

Figure 7: A liquid or a glass in a 2d colloidal suspension of hard spheres.

are typically big, say, of 1mm radius. Colloidal suspensions contain smaller (typically micrometre-sized) particles suspended in a liquid and form the basis of many
paints and coatings. Again, at high density such materials tend to become glassy
unless crystallization is specifically encouraged (and can even form arrested gels at
low densities if attractive forces are also present). On smaller scales still, there are
atomic and molecular glasses: window glass is formed by quick cooling of a silica
melt, and of obvious everyday importance. The plastics in drink bottles and the like
are also glasses produced by cooling, the constituent particles being long polymer
molecules. Critical temperatures are of the order of 80C for, say, PVC and these
systems are glassy at room temperature. Finally, on the nanoscale, glasses are also
formed by vortex lines in type-II superconductors. Atomic glasses with very low
critical temperature, of the order of 10 mK, have also been studied in great detail.
A set of experiments explore the macroscopic macroscopic properties of glass
formers. In a series of usual measurements one estimates de entropy of the sample by
using calorimetric measurements and the thermodynamic relation
S(T2 ) S(T1 ) =

14

T2

dT
T1

Cp (T )
.
T

(1.4)

Figure 8: Several kinds of glasses. A colloidal suspension observed with confocal

microscopy. A polymer melt configuration obtained with molecular dynamics. A
simulation box of a Lennard-Jones mixture. A series of photograph of granular matter.
In some cases the specific volume of the sample is shown as a function of temperature.
In numerical simulations the potential energy density can be equally used. Figure 9
shows the entropy of the equilibrium liquid, S(T ) cT and the jump to the entropy of
the equilibrium crystal at the melting temperature Tm , a first order phase transition.
The figure also shows that when the cooling rate is sufficiently fast, and how fast is
fast depends on the sample, the entropy follows the curve of the liquid below Tm ,
entering a metastable phase that is called a super-cooled liquid. The curves obtained
with different cooling rates are reproducible in this range of temperatures. However,
below a characteristic temperature Tg the curves start to deviate from the liquid-like
behavior, they become flatter and, moreover, they depend on the cooling rate (red,
orange and yellow curves in the figure). The slower the cooling rate the lower the
entropy and the closer it comes to the one of the crystal. Typical cooling rates used
in the laboratory are 0.1 100 K/min. Within these experiments Tg is defined as the
15

temperature at which the shoulder appears.

Figure 9: The typical plot showing the four phases observed in a cooling experiment:
liquid, supercooled liquid, glass and crystal. The characteristic temperatures Tm (a
first order phase transition), Tg and the Kauzmann temperature TK are shown as well
as the typical relaxation times in the liquid and super-cooled liquid phases.
The extrapolation of the entropy of the liquid below Tg crosses the entropy of the
crystal at a value of the temperature that was conjectured by Kauzmann to correspond
to an actual phase transition. Indeed, at TK the entropy of the glass is no longer
larger than the one of the crystal and the system undergoes an entropy crisis. Of
course experiments cannot be performed in equilibrium below Tg and, in principle,
the extrapolation is just a theoretical construction. Having said this, the mean-field
models we will discuss later on realize this feature explicitly and put this hypothesis
on a firmer analytic ground. If TK represents a thermodynamic transition it should
be reachable in the limit of infinitely slow cooling rate.
Rheological measurements show that the viscosity of a super-cooled liquid, or the
resistance of the fluid to being deformed by either shear or tensile stress, also increases
by many orders of magnitude when approaching the glass transition. One finds or
alternatively defines Tg as the temperature at which the viscosity reaches = 102
Pa s [Pascal s = k m/s2 s/m2 = kg/(m s)]. At this temperature a peak in the specific
heat at constant pressure is also observed, but no divergence is measured.
Bulk relaxation times are also given in the figure in units of seconds. In the super16

cooled liquid phase the relaxation time varies by 10 orders of magnitude, from
1013 at the melting point to 103 at the glassy arrest. The interval of variation
of the temperature is much narrower; it depends on the sample at hand but one can
say that it is of the order of 50 K. We note that the relaxation times remain finite all
along the super-cooled liquid phase and do not show an explicit divergence within the
temperature range in which equilibrium can be ensured. We discuss below how these
relaxation times are estimated and the two classes, i.e. temperature dependences,
that are found.
The values of Tg depend on the sample. In polymer glasses one finds a variation
from, say, 70 C in rubber to 145 C in polycarbonate passing by 80 C in the ubiquitous
PVC.
There are many different routes to the glassy state. In the examples above we
described cooling experiments but one can also use crunches in which the system is
set under increasing pressure or other.
The structure and dynamics of liquids and glasses can be studied by investigating the two-time dependent density-density correlation:
g(r; t, tw )

h (~x, t)(~y , tw ) i

N 2

N
N X
X
i=1 j=1

with

r = |~x ~y |

h(~x ~ri (t))(~y ~rj (tw ))i

where we ignored linear and constant terms. is the density variation with respect to the mean N/V . The average over different dynamical histories (simulation/experiment) h. . .i implies isotropy (all directions are equivalent) and invariance
under translations of the reference point ~y . Its Fourier transform is
F (q; t, tw ) = N 1

N
X

i,j=1

h ei~q(~ri (t)~rj (tw )) i

(1.5)

The incoherent intermediate or self correlation:

Fs (q; t, tw ) = N 1

N
X
i=1

h ei~q(~ri (t)~ri (tw )) i

(1.6)

can be accessed with (neutron or other) diffraction experiments.

In the main panel of Fig. 10-left the equal-time two-point correlation function of
a Lennard-Jones mixture at different times after an infinite rapid quench below the
glassy crossover temperature Tg is shown. The data vary very little although a wide
range of time-scales is explored. In the inset a zoom over the first peak taken at the
same time for different final temperatures, three of them below Tg the reference one
at the numerically determined Tg . Again, there is little variation in these curves. One
concludes that the structure of the sample in all these cases is roughly the same.
The change is much more pronounced when one studies the dynamics of the sample, that is to say, when one compares the configuration of the system at different
17

1.0

Tf=0.1
Tf=0.3
Tf=0.4

4.0

q=7.23
Fs(q;t,tw)

gAA(r)

gAA(r;t,t)

6.0

4.0
3.0

2.0

2.0

Tf=0.435
0.0
0.9

1.0

1.1

0.0

t=10

1.0

2.0

tw=63100

0.6

tw=10

0.4
1.2

1.3 r 1.4

1.0
t=0

0.8

0.2
Tf=0.4

0.0

3.0

r 4.0

tw=0
Tf=0.4

10

10

10

10

10

10

10

ttw

Figure 10: Structure and dynamics of a binary Lennard-Jones mixture. Left: the
two-point correlation function of the A atoms at different times (main panel) and
at different temperatures (inset). Right: the decay of the Fourier transform of the
correlation function at the wave-vector associated to the first peak in gAA (r). Data
from Kob & J-L Barrat.
times. The curves on the right panel display the relaxation of the correlation function
at different temperatures, all above Tg . The relaxation is stationary in all cases, i.e.
a function of t tw only, but it becomes much slower when the working temperature
approaches Tg .
In a family of glass formers called fragile, in double logarithmic scale used in the
plot, a clear plateau develops for decreasing T and may seem to diverge in the T Tg
limit. In another family of glass formers called strong no plateau is seen.
From the analysis of the temperature dependence of the relaxation time, say the
time needed for the correlation to decay to half its value at zero time delay4 one finds
two kinds of fitting laws:

Vogel-Fulcher-Tamann
0 eA/(T T0 )
=
(1.7)
0 eA/T
Arrhenius
In fits T0 is usually very close to TK . The former class of systems are the fragile
ones while the latter are the strong ones. Note that the first form yields a divergence
at a finite TK while the second one yields a divergence at T = 0. Silica belongs to
the second class while most polymer glasses belong to the first one. This relaxation
time is usually called the alpha or structural relaxation time. Recall that in a
usual second order phase transition (as realized in an Ising model, for instance) the
divergence of the relaxation time close to the critical point is of power law type.
A real space analysis of the motion of the particles in atomic, molecules in
molecualr, or strings in polymeric glasses (and granular matter as well) demonstrates
that the elements move, over short time scales, in cages formed by their neighbors.
During this short time span the correlation function decays to the plateau and the
mean-square displacement reaches a plateau (in a double logarithmic scale). Note,
4 This

is a very naive definition of , others much more precise are used in the literature.

18

however, that the particles displacement is much smaller than the particle radius
meaning that the displacement is indeed tiny during this time regime. the second
structural relaxation is the one that take the correlation (displacement) below (above)
the plateau.

500 nm

Figure 11: Colloidal suspension (data from E. Weeks group) and granular matter
(data from O. Pouliquens group).
Very recently stress has been put on the analysis of the motion of the elements
over longer time-scales. Dynamic heterogeneities [12] were thus uncovered. Dynamic
regions with high mobility immersed in larger regions with little mobility were identified. Sometimes stringly motion of particles following each other in a periodic path
were also observed in confocal microscopy measurements or in molecular dynamics
simulations. The length of these strings seems to increase when approaching the
crossover temperature Tg . Moreover, dynamic heterogeneities, and a growing length
associated to it, were quantified from the analysis of a four-point correlation function.
This function takes different forms depending on the problem at hand but basically
searches for spatial correlations in the displacement of particles between on time intervals. Calling (~r, t) = (~r, t) 0 with 0 = N/V ,
C4 (r; t, tw )

h(~x, tw )(~x, t)(~y , tw )(~y , t)i

h(~x, tw )(~x, t)ih(~y , tw )(~y , t)i .

(1.8)

Terms involving one position only can be extracted from the average since they
do not contain information about the spatial correlation. The idea is, roughly, to
consider that (~x, t)(~x, tw ) is the order parameter. The double spatial integral
of this quantity defines a generalized susceptibility 4 (t, tw ) that has been studied
in many numerical and laboratory experiments. It shows a peak at the time-delay
t tw that coincides with the relaxation time . Assuming a usual kind of scaling
with a typical growing length for the four point correlation the characteristics of the
appearance of the peak should yield the length of these dynamic heterogeneities. The
data can be interpreted as leading to a divergence of the growing length at some
temperature but the actual values found are very small, of the order of a few interparticle distances in the sample.
The defining features of glasses, i.e., the characterization of their out of equilibrium relaxation and aging phenomena [13], will be discussed below.
19

A summary of the liquid-super-cooled liquid-glass behavior is given in the table

below.

Crytallization at Tm is avoided by cooling fast enough.

Liquid
Supercooled liquid
Glass
|
{z
}
Exponential relax
Non-exponential relax
Equilibrium
Metastable equilibrium
Non-equilibrium
|
{z
}
Separation of time-scales &
An exponential number of metastable states!
{z
}
|
Stationary
Aging
Aging means that

correlations and reponses depend on t and tw

ac susceptibilities depend on and tw

There might be an equilibrium transition to an ideal glass at Ts .

1.6

Quenched disorder: still glassiness

In the paragraphs above we characterized the low temperature regime of certain

particle models and claimed that their structure is disordered (at least at first sight).
Another sense in which the word disorder is used is to characterize the interactions. Quenched interactions are due to a very sharp separation of time-scales. The
traditional example is the one of spin-glasses in which the characteristic time for
diffusion of magnetic impurities in an inert host is much longer than the characteristic
time for magnetic moment change:
d exp 0 .

(1.9)

The position of the magnetic moments are decided at the preparation of the sample.
These position are then random and they do not change during experimental times.
The interactions between pairs of spins depend on the distance between the magnetic
moments via the RKKY formula
VRKKY (rij ) = J

cos(2kF rij )
si sj .
3
rij

(1.10)

Therefore quenched competing interactions are fixed in the observational time-scale

and they transmit contradictory messages. Typical examples are systems with ferromagnetic and/or antiferromagnetic exchanges that are not organized in a simple way

20

with respect to the geometry and connectivity of the lattice such as spin-glasses [7]
(see Fig. 5-right).
Theoretically, this is modeled by random interactions drawn from a probability
distribution. For simplicity the spins (magentic moments) are placed on the vertices of a finite dimensional lattice, typically a cubic one. The Edwards-Anderson
Hamiltonian then reads
X
HJ [{si }] =
with
Jij
taken from
P (Jij )
(1.11)
Jij si sj
hiji












$JLQJ&RUUHODWLRQ

&RUUHODWLRQ&W
W&WW

Annealed interactions may have a slow time-dependence. Both lead to disorder. These can be realized by coupling strengths as in the magnetic example in
Fig. 5, but also by magnetic fields, pinning centers, potential energies, etc. Disordered interactions usually lead to low-temperature behavior that is similar to the one
observed in systems with dynamic competing interactions.
Data showing the cusp in the susceptibility of a spin-glass sample are shown in
Fig. 12.




(

(













7LPHGLIIHUHQFHW WW
V












$JLQJ5HOD[DWLRQ

5HOD[DWLRQVW
WVWW






(

(











7LPHGLIIHUHQFHW WW
V

Figure 12: Spin-glasses: Susceptibility data (Mydosh). Aging phenomena (Herisson

and Ocio).

1.7

Random manifolds

A problem that finds applications in many areas of physics is the dynamics of elastic manifolds under the effect (or not) of quenched random potentials, with (KardarParisi-Zhang) or without (Edwards-Wilkinson, Mullins-Herring) non-linear interactions, with short-range or long-range elastic terms [8, 14].
21

Under certain circumstances the interfaces roughen, that is to say, their asymptotic averaged width depends on their linear size. Take for instance, the local height
h(~r, t) of a d dimensional surface (with no overhangs). Its time-dependent width is
defined as
Z
d
WL (t) = L
(1.12)
dd r [h(~r, t) hh(~r, t)i]2

R
where h. . .i = Ld dd r . . .. This quantity verifies the so-called Family-Vicsek scaling. In its simplest form, in which all dependences are power laws, it first increases
as a function of time, WL (t) t2 and independently of L. At a crossover time
tx Lz it crosses over to saturation at a level that grows as L2 . is the growth
exponent, z is the dynamic exponent and is the roughness exponent. Consistency
implies that they are related by z = . The values of the exponents are known in
a number of cases. For the Edwards-Wilkinson surface one has = (2 d)/4, z = 2
and = (2 d)/2 for d 2. For the non-linear KPZ line = 1/3, z = 3/2 and
= 1/2.
In the presence of quenched disorder the dependence of the asymptotic roughness
with the length of the line undergoes a crossover. For lines that are shorter than a
temperature and disorder strength dependent value LT the behavior is controlled by
thermal fluctuations and relation as the one above holds with = T , the thermal
roughness exponent. This exponent is the one corresponding to the EW equation.
In this thermally dominated scale, the dynamics is expected to be normal in the
sense that lengths and times should be thus related by power laws of types with the
exponents discussed above. For surfaces such that L > LT one finds that the same
kind of scaling holds but with a roughness exponent that takes a different value. The
time dependence and cross-over time are expected, though, not to be power laws and
we will discuss them later.
The relaxation dynamics of such elastic manifolds in the very large limit presents
many other interesting phenomena that resemble features observed in more complex
glassy systems. Moreover, such elastic surfaces appear in the nucleation and growth
kinetics problems discussed above as the interfaces between equilibrium (sometimes
metastable) states.

1.8

Aging

In practice a further complication appears [13]. Usually, disordered phases are

prepared with a relatively rapid quench from the high temperature phase. When
approaching a characteristic temperature the systems cannot follow the pace of evolution dictated by the environment and fall out of equilibrium [2]. Indeed, their key
feature is that below some characteristic temperature Tg , or above a critical density
g , the relaxation time goes beyond the experimentally accessible time-scales and the
system is next bound to evolve out of equilibrium. Although the mechanism leading
to such a slow relaxation is unknown and might be different in different cases the
out of equilibrium relaxation presents very similar properties. The left panel in Fig. 13
shows one aspect of glassy dynamics, aging, as shown by the two-time relaxation of
22

the self-correlation of a colloidal suspension, that is remarkably similar to the decay

of the magnetic correlation in the Ising model shown in the right panel and in Fig. 31.

a)

tw

Various shear histories

b)

C(t,tw)

0.14

|g1(tw,t)|

0.12
0.10
0.08

tw=2
4
8
16
32
64
128
256
512

0.1

0.06
0.04
0.02
0.01

0.1

10

100

1000

t (sec)

10

100

1000

t-tw

Figure 13: Left: two-time evolution of the self-correlation in a colloidal suspension

initialized by applying a shearing rate (data from Viasnoff and Lequeux) The longer
the waiting time the slower the decay. Right: two-time evolution in the bi-dimensional
Ising model quenched below its phase transition at Tc . A two-scale relaxation with
a clear plateau at a special value of the correlation is seen in the double logarithmic
scale. Data from Sicilia et al. We will discuss this feature at length in the lectures.
A purely static description, based on the use of the canonical (or grand-canonical)
partition function is then not sufficient. One is forced to include the time evolution
of the individual agents (spins, particles, molecules) and from it derive the macroscopic time-dependent properties of the full system. The microscopic time-evolution is
given by a stochastic process. The macroscopic evolution is usually very slow and, in
probability terms, it is not a small perturbation around the Gibbs-Boltzmann distribution function but rather something quite different. This gives rise to new interesting
phenomena.

1.9

Driven systems

An out of equilibrium situation can be externally maintained by applying forces

and thus injecting energy into the system and driving it. There are several ways to
do this and we explain below two quite typical ones that serve also as theoretical
traditional examples.
Rheological measurements are common in soft condensed matter; they consist
in driving the systems out of equilibrium by applying an external force that does not
derive from a potential (e.g. shear, shaking, etc.). The dynamics of the system under
the effect of such a strong perturbation is then monitored.
The effect of shear on domain growth is one of great technological and theoretical importance. The growth of domains is anisotropic and there might be differ23

ent growing lengths in different directions. Moreover, it is not clear whether shear
might interrupt growth altogether giving rise to a non-equilibrium stationary state or
whether coarsening might continue for ever. Shear is also commonly used to study
the mechanical properties of diverse glasses.
Another setting is to couple the system to different external reservoirs all in
equilibrium but at different temperature or chemical potential thus inducing a heat or
a particle current through the system. This set-up is relevant to quantum situations
in which one can couple a system to, say, a number of leads at different chemical
potential. The heat transport problem in classical physics also belongs to this class.
A pinned interface at zero temperature can be depinned by pulling it with an
external force. The depinning problem that is to say the analysis of the dynamics
close to the critical force needed to depin the manifold, and the creep dynamics at
non-vanishing temperature have also been the subject of much analysis.
v

Moving
Phase

T>0

T=0

Creep

Depinning
Fc

Pe
Newtonian

Shear thinning

Shear thickening

Figure 14: Left: Creep and depinning of elastic objects under quenched randomness.
Right: Rheology of complex fluids. Shear thinning: decreases or thickening
increases

1.10

Interdisciplinary aspects

The theory of disordered systems has become quite interdisciplinary in the sense
that problems in computer science, biology or even sociology and finance have disorder
aspects and can be mimicked with similar models and solved with similar methods to
the ones we will discuss here.
1.10.1

Optimization problems

The most convenient area of application is, most probably, the one of combinatorial optimization in computer science [9]. These problems can usually be stated
in a form that corresponds to minimizing a cost (energy) function over a large set of
variables. Typically these cost functions have a very large number of local minima
an exponential function of the number of variables separated by barriers that scale
with N and finding the truly absolute minimum is hardly non-trivial. Many interesting optimization problems have the great advantage of being defined on random
24

graphs and are then mean-field in nature. The mean-field machinery that we will
discuss at length is then applicable to these problems with minor (or not so minor)
modifications due to the finite connectivity of the networks.

Figure 15: Graph partitioning.

Let us illustrate this kind of problems with two examples. The graph partitioning problem consists in, given a graph G(N, E) with N vertices and E edges, to
partition it into smaller components with given properties. In its simplest realization
the uniform graph partitioning problem is how to partition, in the optimal way, a
graph with N vertices and E links between them in two (or k) groups of equal size
N/2 (or N/k) and the minimal the number of edges between them. Many other variations are possible. This problem is encountered, for example, in computer design
where one wishes to partition the circuits of a computer between two chips. More
recent applications include the identification of clustering and detection of cliques in
social, pathological and biological networks.
Another example is k satisfiability (k-SAT). This is the computer science problem of determining whether the variables of a given Boolean formula can be assigned
in such a way as to make the formula evaluate to true. Equally important is to
determine whether no such assignments exist, which would imply that the function
expressed by the formula is identically false for all possible variable assignments.
In this latter case, we would say that the function is unsatisfiable; otherwise it is
satisfiable. For example, the formula C1 : x1 OR x2 made by a single clause C1 is
satisfiable because one can find the values x1 = true (and x2 free) or x2 = true (and
x1 free), which make C1 : x1 OR x2 true. This example belongs to the k = 2 class
of satisfiability problems since the clause is made by two literals (involving different
variables) only. Harder to decide formul are made of M clauses involving k literals
required to take the true value (x) or the false value (x) each, these taken from a pool

25

of N variables. An example in 3-SAT is

C1 : x1 OR

C2 : x5 OR
F=
C3 : x1 OR

C4 : x2 OR

x2
x7
x4
x5

OR
OR
OR
OR

x3
x9
x7
x8

(1.13)

All clauses have to be satisfied simultaneously so the formula has to be read F : C1

AND C2 AND C3 AND C4 . It is not hard to believe that when M/N > c the
problems typically become unsolvable while one or more solutions exist on the other
side of the phase transition. In random k-SAT an instance of the problem, i.e. a
formula, is chosen at random with the following procedure: first one takes k variables
out of the N available ones. Seconf one decides to require xi or xi for each of them
with probability one half. Third one creates a clause taking the OR of these k literals.
Forth one returns the variables to the pool and the outlined three steps are repeated
M times. The M resulting clauses form the final formula.
The Boolean character of the variables in the k-SAT problem suggests to transform
them into Ising spins, i.e. xi evaluated to true (false) will correspond to si = 1 (1)
. The requirement that a formula be evaluated true by an assignment of variables
(i.e. a configuration of spins) will correspond to the ground state of an adequately
chosen energy function. In the simplest setting, each clause will contribute zero
(when satisfied) or one (when unsatisfied) to this cost function. There are several
equivalent ways to reach this goal. For instance C1 above can be represented by a
term (1s1 )(1+s2 )(1s3 )/8. The fact that the variables are linked together through
the clauses suggests to define k-uplet interactions between them. We then choose the
interaction matrix to be

if neither xi nor xi Ca
0
1
if
x i Ca
(1.14)
Jai =

1
if
xi Ca
and the energy function as

HJ [{si }] =

M
X

a=1

N
X
i=1

Jaj si , k)

(1.15)

where (x, y) is a Kronecker-delta. This cost function is easy to understand. The

PN
Kronecker delta contributes one to the sum only if all terms in the sum i=1 Jai si
are equal 1. This can happen when Jai = 1 and si = 1 or when Jai = 1 and
si = 1. In both cases the condition on the variable xi is not satisfied. Since this is
required from all the variables in the clause, the clause itself and hence the formula
are not satisfied.
These problems are solved numerically, with algorithms that do not necessarily
respect physical rules. Thus, one can use non-local moves in which several variables
are updated at once as in cluster algorithms of the Swendsen-Wang type used to beat
26

critical slowing down close to phase transitions or one can introduce a temperature to
go beyond cost-function barriers and use dynamic local moves that do not, however,
satisfy a detail balance. The problem is that with hard instances of the optimization
problem none of these strategies is successful. Indeed, one can expect that glassy
aspects, as the proliferation of metastable states separated by barriers that grow very
fast with the number of variables, can hinder the resolutions of these problems in
polynomial time for any algorithm.
Complexity theory in computer science, and the classification of optimization
problems in classes of complexity P for problems solved with algorithms that use a
number of operations that grows as a polynomial of the number of variables, e.g. as
N 2 or even N 100 , NP for problems for which no polynomial algorithm is known and
one needs a number of operations that grow exponentially with N , etc. applies to
the worst instance of a problem. Worst instance, in the graph-partitioning example,
means the worst possible realization of the connections between the nodes. Knowing
which one this is is already a very hard problem!
But one can try to study optimization problems on average, meaning that the
question is to characterize the typical and not the worst realization of a problem.
The use of techniques developed in the field of disordered physical systems, notably
spin-glasses, have proven extremely useful to tackle typical single randomly generated
instances of hard optimization problems.
Note that in statistical mechanics information about averaged macroscopic quantities is most often sufficiently satisfactory to consider a problem solved. In the optimization context one seeks for exact microscopic configurations that correspond to the
exact ground state and averaged information is not enough. Nevertheless, knowledge
about the averaged behavior can give us qualitative information about the problem
that might be helpful to design powerful algorithms to attack single instances.

1.10.2

Biological applications

In the biological context disordered models have been used to describe neural
networks, i.e. an ensemble of many neurons (typically N 109 in the human brain)
with a very elevated connectivity. Indeed, each neuron is connected to 104 other
neurons and receiving and sending messages via their axons. Moreover, there is no
clear-cut notion of distance in the sense that axons can be very long and connections
between neurons that are far away have been detected. Hebb proposed that the
memory lies in the connections and the peculiarity of neural networks is that the
connectivity must then change in time to incorporate the process of learning.
The simplest neural network models [10] represent neurons with Boolean variables
or spins, that either fire or are quiescent. The interactions link pairs of neurons and
they are assumed to be symmetric (which is definitely not true). The state of a neuron

27

is decided by an activity function f ,

i (t + 1) = f (

Jij j (t)) ,

(1.16)

j(6=i)

that in its simplest form is just a theta-function leading to simply two-valued neurons.
Memory of an object, action, etc. is associated to a certain pattern of neuronal
activity. It is then represented by an N -component vector in which each component
corresponds to the activity of each neuron. Finally, sums over products of these
patterns constitute the interactions. As in optimization problems, one can study the
particular case associated to a number of chosen specific patterns to be stored and
later recalled by the network, or one can try to answer questions on average, as how
many typical patterns can a network of N neurons store. The models then become
fully-connected or dilute models of spins with quenched disorder. The microscopic
dynamics cannot be chosen at will in this problem and, in general, will not be as
simple as the single spin flip ones used in more conventional physical problems. Still,
if the disordered modeling is correct, glassy aspects can render recall very slow due
to the presence of metastable states for certain values of the parameters.
Another field of application of disordered system techniques is the description
of hetero-polymers and, most importantly, protein folding. The question is how
to describe the folding of a linear primary structure (just the sequence of different
amino-acids along the main backbone chain) into an (almost) unique compact native
structure whose shape is intimately related to the biological function of the protein.
In modeling these very complex systems one proposes that the non-random, selected
through evolution, macromolecules may be mimicked by random polymers. This
assumption is based on the fact that amino-acids along the chain are indeed very
different. One then uses monomer-monomer and/or monomer-solvent interactions
that are drawn from some probability distribution and are fixed in time (quenched
disorder). Still, a long bridge between the theoretical physicists and the biologists
approaches remain to be crossed. Some of the important missing links are: proteins
are mesoscopic objects with of the order of 100 monomers thus far from the thermodynamic limit; interest is in the particular, and not averaged, case in biology, in other
words, one would really like to know what is the secondary structure of a particular
primary sequence; etc. In the protein folding problem it is clear that the time needed
to reach the secondary structure from an initially stretched configuration depends
strongly on the existence of metastable states that could trap the (hetero) polymer.
Glassy aspects have been conjectured to appear in this context too.
The constituents of active matter, be them particles, lines or other, absorb
energy from their environment or internal fuel tanks and use it to carry out motion. In
this new type of soft condensed matter energy is partially transformed into mechanical
work and partially dissipated in the form of heat [11]. The units interact directly or
through disturbances propagated in the medium. In systems of biological interest,
conservative forces (and thermal fluctuations) are complemented by non-conservative
forces. Realizations of active matter in biology are thus manifold and exist at different
28

Figure 16: Active matter.

scales. Some of them are: bacterial suspensions, the cytoskeleton in living cells, or
even swarms of different animals. Clearly enough, active matter is far from equilibrium
and typically kept in a non-equilibrium steady state. The difference between active
matter and other driven systems, such as sheared fluids, vibrated granular matter
and driven vortex lattices is that the energy input is located on internal units (e.g.
motors) and therefore homogeneously distributed in the sample. In the other driven
systems mentioned above, the energy input occurs on the boundaries of the sample.
Moreover, the effect of the motors can be dictated by the state of the particle and/or
its immediate neighborhood and it is not necessarily fixed by an external field.
The dynamics of active matter presents a number of interesting features that
are worth mentioning here. Active matter displays out of equilibrium phase transitions that may be absent in their passive counterparts. The dynamic states display
large scale spatio-temporal dynamical patterns and depend upon the energy flux and
the interactions between their constituents. Active matter often exhibits unusual
mechanical properties, very large responses to small perturbations, and very large
fluctuations not consistent with the central limit theorem. Much theoretical effort
has been recently devoted to the description of different aspects of these systems,
such as self-organization of living microorganisms, the identification and analysis of
states with spatial structure, such as bundles, vortices and asters, the study of the
rheological properties of active particle suspensions with the aim of grasping which
are the mechanical consequences of biological activity. A rather surprisingly result
was obtained with a variational solution to the many-body master equation of the
motorized version of the standard hard sphere fluid often used to model colloids:
instead of stirring and thus destabilize ordered structures, the motors do, in some
circumstances enlarge the range of stability of crystalline and amorphous structures
29

relative to the ones with purely thermal motion.

Figure 17: Left: random graph with finite connectivity

1.11

Summary and interesting questions

The main steps in the development and application of Statistical Mechanics ideas
to macroscopic cooperative systems have been
The development of the basic ideas (Boltzmann-Gibbs).
The recognition of collective phenomena and the identification and mean-field
description of phase transitions (Curie-Weiss).
The correct description of critical phenomena with scaling theories and the
renormalization group (Kadanoff, Widom, M. Fisher, Wilson) and more recently
the development of conformal field theories for two-dimensional systems.
The study of stochastic processes and time-dependent properties (Langevin,
Fokker-Planck, Glauber, etc.).
To describe the problems introduced above the same route has been followed.
There is no doubt that Equilibrium Statistical Mechanics yields the static properties
of these systems. In the case of coarsening problems one understands very well the
static phases and phase transitions. In the case of glassy systems this is not so clear.
In the case of active matter or other driven systems there are equilibrium phases in the
vanishing drive limit only but one can also study the dynamic phase transitions
with a critical phenomena perspective.
Although the study of equilibrium phases might be a little irrelevant from the
practical point of view since most glassy systems are out of equilibrium in laboratory time-scales, it is certainly a necessary step on which one can try to build a truly
dynamic theory. The mean-field study the second step in the list above of the equilibrium properties of disordered systems, in particular those with quenched disorder,
has revealed an incredibly rich theoretical structure. We still do not know whether it
carries through to finite dimensional cases. Even though, it is definitely interesting
per se and it finds a very promising field of application in combinatorial optimization
problems that are defined on random networks, see Fig. 17, with mean-field character.
Scaling arguments have been applied to describe finite dimensional disordered systems
but they remain as their parent ones for clean systems quite phenomenological
30

and difficult to put to sufficiently restrictive numerical or experimental test. The extension of renormalisation group methods to systems with quenched disorder is also
under development and still needs quite a lot of work the third step. As for the out
of equilibrium dynamics of these systems, again, it has been solved at the mean-field
level but little is known in finite dimensions apart from numerical simulations or
the solution to toy models. As in its static counterpart, the results from the study of
dynamic mean-field models have been very rich and they have suggested a number of
new phenomena later searched for in numerical simulations and experiments of finite
dimensional systems. In this sense, these solutions have been a very important source
of inspiration.
Disordered systems (in both senses) are usually in contact with external reservoirs
at fixed temperature; their description is done in the canonical (or grand-canonical in
particle systems with the possibility of particle exchange with the environment) ensemble. In these lectures we will only deal with a canonical setting, the microcanonical
one being more relevant to quantum systems.
Some questions that arise in the non-equilibrium context are
How to characterize the non-equilibrium dynamics of glassy systems phenomenologically.
Which are the minimal models that reproduce the phenomenology.
Which is the relation between the behavior of these and other non-equilibrium
systems, in particular, those kept away from equilibrium by external forces,
currents, etc.
Which features are generic to all systems with slow dynamics.
Could one extend equilibrium statistical mechanics ideas out of equilibrium?
e.g., could one use temperature, entropy and other thermodynamic concepts
out of equilibrium?
Related to the previous item, could one construct a non-equilibrium measure
that would substitute the Gibbs-Boltzmann one in certain cases.

31

Modeling

In this section I will revisit certain aspects of statistical physics that are not
commonly discussed and that become important for our purposes.

2.1

Fluctuations

There are several possible sources of fluctuations:

Thermal: the system is coupled to a classical environment that ensures fluctuations (noise) and dissipation (the fact that the total energy is not conserved).
E.g. coarsening, classical glasses, spin-glasses.
Quantum: the system is coupled to a quantum environment that ensures fluctuations (noise) and dissipation. The temperature of the bath can be zero or
not. E.g. quantum coarsening and glasses, quantum spin-glasses.
Stochastic motors: forces that act on the systems particles stochastically.
The energy injected in the sample is partially dissipated to the bath and partially
used as work. As the system is also coupled to a bath there are also thermal
fluctuations in it. E.g. active matter.
Classical and quantum environments are usually modeled as large ensembles of
non-interacting variables (oscillators [16], spins [17], fermions) with chosen distributions of coupling constants and energies.

2.2

The classical reduced partition function

We analyze the statistical static properties of a classical canonical system in

equilibrium at inverse temperature and itself formed by two sub-parts, one that
will be treated as an environment (not necessarily of infinite size) and another one
that will be the (sub-)system of interest. We study the partition function or Gibbs
functional, Ztot :
X
Ztot [] =
exp(Htot x)
(2.1)
conf env
conf syst

where the sum represents an integration over the phase space of the full system, i.e.
the systems and the environmental ones. is a source. We take
env .
Htot = Hsyst + Henv + Hint + Hcounter = Hsyst + H

(2.2)

For simplicity we use a single particle moving in d = 1: Hsyst is the Hamiltonian of

the isolated particle,
p2
+ V (x) ,
(2.3)
Hsyst =
2M
32

with p and x its momentum and position. Henv is the Hamiltonian of a thermal
bath that, for simplicity, we take to be an ensemble of N independent harmonic
oscillators [15, 16] with masses ma and frequencies a , a = 1, . . . , N
Henv =

N
X
ma a2 2
a2
+
qa
2ma
2
a=1

(2.4)

with a and qa their momenta and positions. This is indeed a very usual choice
since it may represent phonons. (These oscillators could be the normal modes of
a still quadratic Hamiltonian written in terms of other pairs of conjugate variables;
the bath could be, for instance, a chain of harmonic oscillators with nearest-neighbor
couplings.) Hint is the coupling between system and environment. We will restrict
the following discussion to a linear interaction in the oscillator coordinates, qa , and
in the particle coordinate,
N
X
c a qa ,
(2.5)
Hint = x
a=1

with ca the coupling constants. The counter-term Hcounter is added to avoid the
generation of a negative harmonic potential on the particle due to the coupling to the
oscillators (that may render the dynamics unstable). We choose it to be
N

Hcounter

1 X c2a
=
x2 .
2 a=1 ma a2

(2.6)

The generalization to more complex systems and/or to more complicated baths

and higher dimensions is straightforward. The calculations can also be easily generalized to an interaction of the oscillator coordinate with a more complicated dependence
on the systems coordinate, V(x), that may be dictated by the symmetries of the system. Non-linear functions of the oscillator coordinates cannot be used since they
render the problem unsolvable analytically.
Having chosen a quadratic bath and a linear coupling, the integration over the oscillators coordinates and momenta can be easily performed. This yields the reduced
Gibbs functional
!#
"
N
X
1 X c2a
2
x
. (2.7)
Zred []
exp Hsyst + Hcounter + x
2 a=1 ma a2
conf syst
The counter-term Hcounter is chosen to cancel the last term in the exponential and
it avoids the renormalization of the particles mass (the coefficient of the quadratic
term in the potential) due to the coupling to the environment that could have even
destabilized the potential by taking negative values. An alternative way of curing
this problem would be to take a vanishingly small coupling to the bath in such a way
that the last term must vanish by itself (say, all ca 0). However, this might be
problematic when dealing with the stochastic dynamics since a very weak coupling
33

to the bath implies also a very slow relaxation. It is then conventional to include the
counter-term to cancel the mass renormalization. One then finds
P
Zred [] conf syst exp [ (Hsyst + x)] = Zsyst [] .
(2.8)

PN
For a non-linear coupling Hint =
a=1 ca qa V(x) the counter-term is Hcounter =
PN
c2a
1
2
2 [V(x)] .
a=1 ma a
2
The interaction with the reservoir does not modify the statistical properties of the
particle since Zred Zsyst , independently of the choices of ca , ma , a and N .
If one is interested in the dynamics of a coupled problem, the characteristics
of the sub-system that will be considered to be the bath have an influence on the
reduced dynamic equations found for the system, that are of generic Langevin kind,
as explained in Sect. 2.3.
Quantum mechanically the reduced partition function depends explicitly on the
properties of the bath. The interaction with quantum harmonic oscillators introduces
non-local interactions (along the Matsubara time direction) and there is no physical
way to introduce a counter-term to correct for this feature.
The dynamics of quantum systems has all these difficulties.

2.3

The Langevin equation

Examples of experimental and theoretical interest in condensed matter and biophysics in which quantum fluctuation can be totally neglected are manifold. In this
context one usually concentrates on systems in contact with an environment: one selects some relevant degrees of freedom and treats the rest as a bath. It is a canonical
view. Among these instances are colloidal suspensions which are particles suspended
in a liquid, typically salted water, a soft condensed matter example; spins in ferromagnets coupled to lattice phonons, a hard condensed matter case; and proteins in
the cell a biophysics instance. These problems are modeled as stochastic processes
with Langevin equations [25, 26, 27, 28, 29], the Kramers-Fokker-Planck formalism
or master equations depending on the continuous or discrete character of the relevant
variables and analytic convenience.
The Langevin equation is a stochastic differential equation that describes phenomenologically a large variety of problems. It models the time evolution of a set of
slow variables coupled to a much larger set of fast variables that are usually (but not
necessarily) assumed to be in thermal equilibrium at a given temperature. We first
introduce it in the context of Brownian motion and we derive it in more generality in
Sect. 2.3.2.
2.3.1

Langevins Langevin equation

The Langevin equation5 for a particle moving in one dimension in contact with a
5 P. Langevin, Sur la th
eorie du mouvement brownien, Comptes-Rendus de lAcad
emie des Sciences 146, 530-532 (1908).

34

white-noise bath reads

mv + 0 v = F + ,

v = x ,

(2.9)

with x and v the particles position and velocity. is a Gaussian white noise with
zero mean and correlation h(t)(t )i = 20 kB T (tt ) that mimics thermal agitation.
0 v is a friction force that opposes the motion of the particle. The force F designates
all external deterministic forces and depends, in the most common cases, on the
position of the particle x only. In cases in which the force derives from a potential,
F = dV /dx. The generalization to higher dimensions is straightforward. Note that
0 is the parameter that controls the strength of the coupling to the bath (it appears
in the friction term as well as in the noise term). In the case 0 = 0 one recovers
Newton equation of motion. The relation between the friction term and thermal
correlation is non-trivial. Langevin fixed it by requiring hv 2 (t)i hv 2 ieq . We will
give a different argument for it in the next section.
2.3.2

Derivation of the Langevin equation

Let us take a system in contact with an environment. The interacting system+environment ensemble is closed while the system is open. The nature of
the environment, e.g. whether it can be modeled by a classical or a quantum formalism, depends on the problem under study. We focus here on the classical problem
defined by Htot . A derivation of a generalized Langevin equation with memory is
very simple starting from Newton dynamics of the full system [15, 18].
The generalization to more complex systems and/or to more complicated baths
and higher dimensions is straightforward. The calculations can also be easily generalized to an interaction of the oscillator coordinate with a more complicated dependence
on the systems coordinate, V(x), that may be dictated by the symmetries of the system, see Ex. 1.
Hamiltons equations for the particle are
x(t)

p(t)
,
m

p(t)
= V [x(t)]

N
X

a=1

ca qa (t)

N
X

c2a
x(t) (2.10)
ma a2
a=1

(the counter-term yields the last term) while the dynamic equations for each member
of the environment read
qa (t) =

a (t)
,
ma

a (t) = ma a2 qa (t) ca x(t) ,

(2.11)

showing that they are all massive harmonic oscillators forced by the chosen particle. These equations are readily solved by
Z t
a (0)
ca
qa (t) = qa (0) cos(a t) +
dt sin[a (t t )]x(t ) (2.12)
sin(a t)
ma a
ma a 0
35

with qa (0) and a (0) the initial coordinate and position at time t = 0 when the
particle is set in contact with the bath. It is convenient to integrate by parts the
last term. The replacement of the resulting expression in the last term in the rhs of
eq. (2.10) yields
p(t)
= V [x(t)] + (t)

Rt
0

dt (t t )x(t
) ,

(2.13)

with the symmetric and stationary kernel given by

(t t ) =

c2a
2
a=1 ma a

PN

cos[a (t t )] ,

(t t ) = (t t), and the time-dependent force given by

h

i

PN
a (0)
ca x(0)
(t) = a=1 ca m
cos(
t)
.
sin(
t)
+
q
(0)
+
a
a
a
ma 2
a a
a

(2.14)

(2.15)

This is the equation of motion of the reduced system. It is still deterministic.

The third term on the rhs of eq. (2.13) represents a rather complicated friction
force. Its value at time t depends explicitly on the history of the particle at times
0 t t and makes the equation non-Markovian. One can rewrite it as an integral
running up to a total time T > max(t, t ) introducing the retarded friction:
(t t ) = (t t )(t t ) .

(2.16)

Until this point the dynamics of the system remain deterministic and are completely determined by its initial conditions as well as those of the reservoir variables.
The statistical element comes into play when one realizes that it is impossible to
know the initial configuration of the large number of oscillators with great precision
and one proposes that the initial coordinates and momenta of the oscillators have
a canonical distribution at an inverse temperature . (Note that one needs to
assume that the oscillators interacted in the past to establish ergodicity and reach
this pdf, though they do not do any longer.) Then, one chooses {a (0), qa (0)} to be
initially distributed according to a canonical phase space distribution:

P ({a (0), qa (0)}, x(0)) = 1/Zenv [x(0)] e Henv [{a (0),qa (0)},x(0)]

env = Henv + Hint + Hcounter , that can be rewritten as

with H
"
#

2
N
2
2
X
c
m

(0)

a
a
a
env =
qa (0) +
H
x(0) + a
.
2
2
m

2ma
a
a
a=1

(2.17)

(2.18)

Again, the presence of Hcounter here is for convenience. The randomness in the initial
conditions gives rise to a random force acting on the reduced system. Indeed, is
now a Gaussian random variable, that is to say a noise, with
h(t)i = 0 ,

h(t)(t )i = kB T (t t ) .
36

(2.19)

One can easily check that higher-order correlations vanish for an odd number of
factors and factorize as products of two time correlations for an even number of
factors. In consequence
has Gaussian statistics. Defining the inverse of over the
Rt
interval [0, t], 0 dt (t t )1 (t t ) = (t t ), one has the Gaussian pdf:
P [] = Z 1 e

2k1

BT

Rt
0

dt

Rt
0

dt (t)1 (tt )(t )

(2.20)

Z is the normalization. A random force with non-vanishing correlations on a finite support is usually called a coloured noise. Equation (2.13) is now a genuine
Langevin equation. A multiplicative retarded noise arises from a model in which one
couples the coordinates of the oscillators to a generic function of the coordinates of
the system, see Ex. 1 and eq. (2.31).
The use of an equilibrium measure for the oscillators implies the relation between the friction kernel and the noise-noise correlation, which are proportional,
with a constant of proportionality of value kB T . This is a generalized form of the
fluctuation-dissipation relation, and it applies to the environment.
Different choices of the environment are possible by selecting different ensembles
of harmonic oscillators. The simplest one, that leads to an approximate Markovian
equation, is to consider
that the oscillators are coupled to the particle via coupling

constants ca = ca / N with ca of order one. One defines

S()

1
N

c2a
a=1 ma a

PN

( a )

(2.21)

a function of , of order one with respect to N , and rewrites the kernel as

(t t ) =
A common choice is
S()

= 20

||

S()

1

cos[(t t )] .

fc

||

(2.22)

(2.23)

The function fc (x) is a high-frequency cut-off of typical width and is usually chosen
to be an exponential. The frequency
is a reference frequency that allows one
to have a coupling strength 0 with the dimensions of viscosity. If = 1, the friction
is said to be Ohmic, S()/ is constant when || as for a white noise. This
name is motivated by the electric circuit analog exposed by the end of this Section.
When > 1 ( < 1) the bath is superOhmic (subOhmic). The exponent is
taken to be > 0 to avoid divergencies at low frequency. For the exponential cut-off
the integral over can be computer for = 1 and 6= 1. In the former Ohmic case
one finds

(t) = 20
,
(2.24)
[1 + (t)2 ]

37

that in the limit becomes a delta-function, (t) 20 (t). In the latter

non-Ohmic case the integral over yields
(t) = 20
+1 E ()

cos[ arctan(t)]
[1 + (t)2 ]/2

(2.25)

with E () the Euler Gamma-function. At long times, for any > 0 and 6= 1, one
has
lim (t)

20
cos(/2)E () (
t) ,

(2.26)

a power law decay.

Time-dependent, f (t), and constant non-potential forces, f np , as the ones applied
to granular matter and in rheological measurements, respectively, are simply included
in the right-hand-side (rhs) as part of the deterministic force. When the force derives
from a potential, F (x, t) = dV /dx.
In so far we have discussed systems with position and momentum degrees of freedom. Other variables might be of interest to describe the dynamics of different kind
of systems. In particular, a continuous Langevin equation for classical spins can also
be used if one replaces the hard Ising constraint, si = 1, by a soft one implemented
with a potential term of the form V (si ) = u(s2i 1)2 with u a coupling strength
(that one eventually takes to infinity to recover a hard constraint). The soft spins
are continuous unbounded variables, si (, ), but the potential energy favors
the configurations with si close to 1. Even simpler models are constructed with
spherical
that are also continuous unbounded variables globally constrained to
Pspins,
N
satisfy i=1 s2i = N . The extension to fields is straightforward and we will discuss
one when dealing with the O(N ) model.
PN
Exercise 1. Prove that for a non-linear coupling Hint = V[x] a=1 ca qa there is a
choice of counter-term for which the Langevin equation reads
Z t
dt (t t )V [x(t )]x(t
)
(2.27)
p(t)
= V [x(t)] + (t)V [x(t)] V [x(t)]
0

with the same as in eq. (2.14) and (t) given by eq. (2.15) with x(0) V[x(0)].
The noise appears now multiplying a function of the particles coordinate.
Exercise 2. Take now a systemPmade
of i = 1, . . . , n variables collected in an nn PN
component vector ~x. Use Hint = i=1 a=1 cai qai xi as the coupling between system
and bath and an ensemble of n independent sets of harmonic oscillators for the bath.
Prove that the stochastic equation is
Z t
V [x(t)]
p i (t) =
dt i (t t )x i (t )
i = 1, . . . , n
(2.28)
+ i (t)
xi (t)
0

38

where there is no sum over repeated indices and i and i are given by
N
X

c2ai

(2.29)
2 cos[ai (t t )] ,
m

ai
ai
a=1




N
X
cai xi (0)
ai (0)
cos(
t)
.(2.30)
sin(ai t) + qai (0) +
ca
i (t) =
ai
2
mai ai
mai ai
a=1

i (t t ) =

The i dependence in i can be ignored if the ensembles of oscillators are equivalent

(i.e. same distribution of parameters). Characterise next the mean hi (t)i and the
correlation hi (t)j (t )i and see under which conditions hi (t)j (t )i = ij (t t ).
Exercise 3. P
Take now a system made of i = 1, . . . , n variables collected in ~x. Use
N
Hint = V[x] a=1 ca qa as the coupling between system and bath, where x is the
modulus of the vector ~x. Prove that the stochastic equation is
p i (t) =

V[x(t)]
V [x(t)]
+ (t)
xi (t)
xi (t)
Z t
n
X
V[x(t)]
V[x(t )]

dt (t t )
x j (t )
)
xi (t) 0
x
(t
j
j=1

(2.31)

with the same and as in eqs. (2.14) and (2.15) with x(0) V[x(0)]. There is
only one noise component and it appears multiplying a function of the particles
coordinate.
The electric analog: take an LRC circuit. The resistance is of the usual Ohmic
type, that is to say, the potential drop, VR , across it is given by VR = IR with I the
current and R the resistance. The potential drop, VL , across the inductor L is Rgiven
by VL = LdI/dt. Finally, the potential drop across the capacitor is VC = C 1 Idt.
The balance between these potentials implies a Langevin type equation for the current
circulating across the circuit:
L

dI
d2 I
+R
+ C 1 I = 0 .
dt2
dt

(2.32)

This analogy justifies the Ohmic name given to a dissipative term proportional to the
velocity in the general presentation.

2.3.3

Irreversibility and dissipation.

The friction force 0 v in eq. (2.9) or its retarded extension in the non-Markovian
case explicitly breaks time-reversal (t t) invariance, a property that has to be
respected by any set of microscopic dynamic equations. Newton equations describing
the whole system, the particle and all the molecules of the fluid, are time reversal
39

invariant. However, time-reversal can be broken in the reduced equation in which

the fluid is treated in an effective statistical form and the fact that it is in equilibrium
is assumed from the start.
Even in the case in which all forces derive from a potential, F = dV /dx, the
energy of the particle, mv 2 /2 + V , is not conserved and, in general, flows to the
bath leading to dissipation. At very long times, however, the particle may reach a
stationary regime in which the particle gives and receives energy from the bath at
equal rate, on average.
Exercise 4. Prove the time-irreversibility of the Langevin equation and the fact that
the symmetry is restored if 0 = 0. Show that dhHsyst i/dt 6= 0 when 0 6= 0.
2.3.4

Smoluchowski (overdamped) limit

In many situations in which friction is very large, the characteristic time for the
relaxation of the velocity degrees of freedom to their Maxwellian distribution, tvr , is
very short (see the examples in Sect. 2.4). In consequence, observation times are
very soon longer than this time-scale, the inertia term mv can be dropped, and the
Langevin equation becomes
0 x = F +
(2.33)
(for simplicity we wrote the white-noise case). Indeed, this overdamped limit is
acceptable whenever the observation times are much longer than the characteristic
time for the velocity relaxation. Inversely, the cases in which the friction coefficient
0 is small are called underdamped.
In the overdamped limit with white-noise the friction coefficient 0 can be absorbed
in a rescaling of time. One defines the new time
t = 0

(2.34)

the new position, x

( ) = x(0 ), and the new noise ( ) = (0 ). In the new
variables the Langevin equation reads x
( ) = F (
x, ) + ( ) with h( )( )i =
2kB T ( ).
2.3.5

Discretization of stochastic differential equations

The way in which the stochastic differential equation (2.33) [with no inertia and
with white noise] is to be discretized is a subtle matter that we will not discuss in
these lectures, unless where it will be absolutely necessary. There are basically two
schemes, called the It
o and Stratonovich calculus, that are well documented in the
literature 6 .
In short, we will use a prescription in which the pair velocity-position of the particle
at time t + , with an infinitesimal time-step, depends on the pair velocity-position
at time t and the value of the noise at time t.
6 A clear and non-technical discussion of the two schemes is given in N. G. van Kampen, It
o versus
Stratonovich, J. Stat. Phys. 24, 175 (1981).

40

2.3.6

Markov character

In the case of a white noise (delta correlated) the full set of equations defines
a Markov process, that is a stochastic process that depends on its history only
through its very last step.
2.3.7

Generation of memory

The Langevin equation (2.9) is actually a set of two first order differential equations. Notice, however, that the pair of first-order differential equations could also be
described by a single second-order differential equation:
m
x + 0 x = F + .

(2.35)

Having replaced the velocity by its definition in terms of the position x(t) depends
now on x(t) and x(t2). This is a very general feature: by integrating away some
degrees of freedom (the velocity in this case) one generates memory in the evolution.
Generalizations of the Langevin equation, such as the one that we have just presented
with colored noise, and the ones that will be generated to describe the slow evolution
of super-cooled liquids and glasses in terms of correlations and linear responses, do
have memory.

2.4

The basic processes

We will discuss the motion of the particle in some 1d representative potentials:

under a constant force, in a harmonic potential, in the flat limit of these two (Fig. 18)
and the escape from a metastable state and the motion in a double well potential
(Fig. 21).

Figure 18: Three representative one-dimensional potentials.

2.4.1

A constant force

41

Let us first consider the case of a constant force, F . The first thing to notice is
that the Maxwell-Boltzmann measure


v2
2m
+V (x)
Pgb (v, x) e
(2.36)

is not normalizable if the size of the line is infinite, due to the exp[V (x)] =
exp(F x) term. Let us then study the evolution of the particles velocity and position
to show how these variables behave and the fact that they do very differently.
The problem to solve is a set of two coupled stochastic first order differential
equations on {v(t), x(t)}, one needs two initial conditions v0 and x0 .
The velocity
The time-dependent velocity follows from the integration of eq. (2.9) over time
0

v(t) = v0 e m t +

1
m

dt e m (tt ) [ F + (t ) ] ,

v0 v(t = 0) .

The velocity is a Gaussian variable that inherits its average and correlations from
the ones of . Using the fact that the noise has zero average
0

hv(t)i = v0 e m t +


0
F 
1 e m t .
0

In the short time limit, t tvr = m/0 , this expression approaches the Newtonian
result (0 = 0) in which the velocity grows linearly in time v(t) = v0 + F/m t. In
the opposite long time limit, t tvr = m/0 , for all initial conditions v0 the averaged
velocity decays exponentially to the constant value F/0 . The saturation when the
bath is active (0 6= 0) is due to the friction term. The relaxation time separating
the two regimes is
tvr = m0 .
(2.37)
The velocity mean-square displacement is
v2 (t) h(v(t) hv(t)i)2 i =


0
kB T 
1 e2 m t
m

(2.38)

independently of F . This is an example of the regression theorem according to

which the fluctuations decay in time following the same law as the average value. The
short and long time limits yield
 20
kB T
t tvr
m t
(2.39)
v2 (t) h(v(t) hv(t)i)2 i
1
t tvr
m
and the two expressions match at t tvr /2. The asymptotic limit is the result expected
from equipartition of the velocity mean-square displacement, h(v(t) hv(t)i)2 i
h(v(t) hvistat )2 istat that implies for the kinetic energy hKistat = kB T /2 only if
42

the velocity is measured with respect to its average. In the heuristic derivation of the
Langevin equation for F = 0 the amplitude of the noise-noise correlation, say A, is not
fixed. The simplest way to determine this parameter is to require that equipartition
for the kinetic energy holds A/(0 m) = T /m and hence A = 0 T . This relation is
known under the name of fluctuationdissipation theorem (fdt) of the second
kind in Kubos nomenclature. It is important to note that this fdt characterizes the
surrounding fluid and not the particle, since it relates the noise-noise correlation to
the friction coefficient. In the case of the Brownian particle this relation ensures that
after a transient of the order of tvr , the bath maintains the fluctuations of the velocity,
v2 , constant and equal to its equilibrium value.

kB T /m
v2

hvi

F/0

v0

(b)

(a)
0

tvr

tvr
t

Figure 19: Results for the constant force problem. (a) Mean velocity as a function of
time. (b) Velocity mean-square displacement as a function of time. In both cases the
linear behavior at short times, t tvr and the saturation values are shown.
The velocity two-time connected correlation reads
h[v(t) hv(t)i][v(t ) hv(t )i]i =

i
0

kB T h 0 |tt |
e m (t+t ) .
e m
m

This is sometimes called the Dirichlet correlator. This and all other higher-order
velocity correlation functions approach a stationary limit when the shortest time
involved is longer than tvr . At t = t on recovers the mean-square displacement
computed in eq. (2.38). When both times are short compared to tvr the two-time
correlator behaves as 2kB T 0 /m2 max(t, t ). When at least one of the two times
is much longer than tvr the second term vanishes and one is left with an exponential
decay as a function of time delay:
c
Cvv
(t, t ) h[v(t) hv(t)i][v(t ) hv(t )ii

kB T 0 |tt |
e m
m

t, t tvr .

(2.40)

The two-time connected correlation falls off to, say, 1/e in a decay time tvd = m/0 .
In this simple case tvr = tvd but this does not happen in more complex cases.
43

More generally one can show that for times t1 t2 . . . tn tvr :

hv(t1 + ) . . . v(tn + )i = hv(t1 ) . . . v(tn )i

(TTI)

(2.41)

for all delays . Time-translation invariance (TTI) or stationarity is one

generic property of equilibrium dynamics. Another way of stating (2.41) is
hv(t1 ) . . . v(tn )i = f (t1 t2 , . . . , tn1 tn ) .

(2.42)

Another interesting object is the linear response of the averaged velocity to a small
perturbation applied to the system in the form of V V f x, i.e. a change in the
slope of the potential in this particular case. One finds

1 0 (tt )
hv(t)if 
=
Rvx (t, t )
e m
(t t )
(2.43)


f (t ) f =0
m
1
h[v(t) hv(t)i][v(t ) hv(t )i]i (t t )
(2.44)

kB T
the last identity being valid in the limit t or t tvr . This is an fdt relation between
c
a linear response, Rvx (t, t ), and a connected correlation, Cvv
(t, t ), that holds for one
of the particle variables, its velocity, when this one reaches the stationary state.
c
(t, t ) (t t )
kB T Rvx (t, t ) = Cvv

(FDT) .

(2.45)

In conclusion, the velocity is a Gaussian variable that after a characteristic time

tvr verifies equilibrium-like properties: its average converges to a constant (determined by F ), its multi-time correlation functions become stationary and a fluctuationdissipation theorem links its linear response to the connected correlation at two times.
More general FDTs are discussed in the exercise proposed below.
The position

x(t)

Rt

dt v(t ) is still a Gaussian random variable:



0
F
F
x0 + v0 tvr + (t tvr ) + tvr
v 0 e m t
0
0
Z t
Z t
0

1
dt
dt e m (t t ) (t ) .
(2.46)
+
m 0
0

The particles position, x(t) = x0 +

Its noise-average behaves as the Newtonian result, ballistic motion,

hx(t)i x0 + v0 t +

F
2m

t2

t tvr

for

(2.47)

at short times and it crossover to

hx(t)i x0 + v0 tvr +

F
0 (t

44

tvr )

for

t tvr

(2.48)

at long times. Note the reduction with respect to ballistic motion (x F t2 ) due to
the friction drag and the fact that this one-time observable does not saturate to a
constant.
An interesting result, that we will use later, is the fact that the coordinate and
the noise have vanishing correlation at equal times: hx(t)(t)i = 0. This can be easily
proven by multiplying the expression for x(t) by (t) and taking the average.
The position mean-square displacement approaches
x2 (t) h(x(t) hx(t)i)2 i 2Dx t with Dx

kB T
0

(Diffusion)

(2.49)

in the usual t tvr limit, that is to say normal diffusion with the diffusion constant Dx . This expression can be computed using x(t) hx(t)i as obtained from
the v(t) hv(t)i above (and it is quite a messy calculation) or one can simply go
to the Smoluchowski limit, taking advantage of the knowledge of what we have just
discussed on the behaviour of velocities, and obtain diffusion in two lines.
Another way of obtaining the searched result is the following. Multiply the
Langevin equation evaluated at t by x evaluated at the same instant and use an
obvious identity to find


d
(2.50)
(xv) v 2 = 0 vx + xF + x
mxv = m
dt
Take now the noise average. Use the fact that the average of x, when the two factors
are evaluated at the same time, vanishes identically, and exchange time-derivative
and noise-average (assuming this operation is permitted). The resulting equation is
d
0
F
hxvi = hvxi + hxi + hv 2 i .
dt
m
m

(2.51)

The last two terms in the right-hand-side are a known time-dependent function, A(t):
A
F
hxi
m

hv 2 i

F
hxi + hv 2 i ,
m




F
F
F
m0 t
v
v
v
x 0 + v 0 tr +
,
(t tr ) + tr
v0 e
m
0
0
 
 2
0
0
0
kB T 
F 
1 e2 m t + v0 e m t +
1 e m t
.
m
0

One can now integrate eq. (2.51) over time

Z t
0
0

hxvi = x0 v0 e m t +
dt e m (tt ) A(t )

(2.52)
(2.53)
(2.54)

(2.55)

to find rather lengthy expression. In the long time limit, t tvr , we drop all exponentially decaying terms to obtain
hxvi

kB T
F v
F2
+
(tr v0 + x0 ) + 2 (t tvr )
0
0
0
45

(2.56)

Now, using hxvi =

1 d
2
2 dt hx i

hx2 i 2

one finally finds

i
kB T
F
F2 h
2
t + 2 (tvr v0 + x0 ) t + 2 (t tvr ) t2r
0
0
0

(2.57)

The last two terms are hxi2 in the same regime of times. Therefore, eq. (9.2) is
recovered.
In contrast to the velocity mean-square displacement this quantity does not saturate at any finite value. Similarly, the particle displacement between two different
times t and t is
xx (t, t ) h[x(t) x(t )]2 i 2Dx |t t | .
(2.58)
It is interesting to note that the force dictates the mean position but it does not
modify the fluctuations about it (similarly to what it did to the velocity). xx is
stationary for time lags longer than tvr .
The two-time position-position connected correlation reads
c
Cxx
(t, t ) = h(x(t) hx(t)i)(x(t ) hx(t )i)i = . . .

(2.59)

Exercise 5: compute this correlation function.

2kB T 01
Cxv

Rxx

01

(a)

(b)
0

tvr
t

tvr
t t

Figure 20: Results for the constant force problem. (a) The correlation between the
position and the velocity of the particle measured at different times. (b) The linear
response of the position to a kick applied linearly to itself at a previous time. In both
cases the linear behavior at short times, t tvr and the saturation values are shown.
Another way to measure the diffusion coefficient directly from the velocity that is
commonly used in the literature is
Dx = lim limt

One can check that it gives the same result.

46

dt hv( + t )v(t )i .

(2.60)

The linear response of the particles position to a kick linearly applied to itself at
a previous time, in the form V V f x at t < t, is

0

hx(t)if 
1

Rxx (t, t )
=
(2.61)
[1 e m (tt ) ] (t t ) ,
f (t ) f =0
0
with the limits

Rxx (t, t )

m1 (t t ) (t t )
01 (t t )

t t tvr ,
t t tvr .

(2.62)

A simple calculation proves that in the short time-differences limit this is the result
for Newton dynamics.
Exercise 6: show the property mentioned above.
The correlation between the position and the velocity reads




0
m
2kB T m
1+
e m t
h(x(t) hx(t)i)(v(t ) hv(t )i)i =
m
0
0
2kB T
(2.63)

0
and it is only a function of t . One notices that in the asymptotic limit in which both
sides of the equation saturate
c
2kB T Rxx (t, t ) = Cxv
(t, t )

for t t tvr and t tvr ,

(2.64)

with a factor of 2 different from the relation in eq. (2.45).

In conclusion, the position is also a Gaussian variable but it is explicitly out of
equilibrium. Its average and variance grow linearly in time, the latter as in normal
diffusion, and the fluctuation-dissipation relation has an additional factor of 1/2 (or
2, depending on on which side of the equality one writes it) with respect to the form
expected in equilibrium.
The energy
The averaged potential energy diverges in the long-time limit since the potential
is unbounded in the x limit: hV (t)i = F hx(t)i F 2 /0 t for t tvr .
Two kinds of variables
This example shows that even in this very simple problem the velocity and position
variables have distinct behavior: the former is in a sense trivial, after the transient tvr
and for longer times, all one-time functions of v F/0 saturate to their equilibriumlike and the correlations are stationary. Instead, the latter remains non-trivial and
evolving out of equilibrium. One can loosely ascribe the different behavior to the
fact that the velocity feels a confining potential K = mv 2 /2 while the position feels
47

an unbounded potential V = F x in the case in which a force is applied, or a flat

potential V = 0 if F is switched off. In none of these cases the potential is able to
take the particle to equilibrium with the bath. The particle slides on the slope and
its excursions forward and backward from the mean get larger and larger as time
increases.
Quite generally, the classical problems we are interested in are such that the
friction coefficient 0 is large and the inertia term can be neglected, in other words,
all times are much longer than the characteristic time tvr . We will do it in the rest of
the lectures.
Ergodicity
The ergodic hypothesis states that, in equilibrium, one can exchange ensemble
averages by time averages and obtain the same results. Out of equilibrium this hypothesis is not expected to hold and one can already see how dangerous it is to take
time-averages in these cases by focusing on the simple velocity variable. Ensemble
and time averages coincide only if the time-averaging is done over a time-window that
lies after tvr but it does not if the integration time-interval goes below tvr .
Tests of equilibration have to be done very carefully in experiments and simulations. One can be simply mislead by, for instance, looking just at the velocities
statistics.
A measure for the time dependent fluctuating position and velocity can be written
down, taking advantage of the fact that both variables are Gaussian:


Z
Z
1
dt dt y t (t)A(t, t )y(t )
(2.65)
P (v, x) exp
2
with the 2 2 matrix A being the inverse of the matrix of correlations, A1 ij (t, t ) =
hyi (t)yj (t )i with i, j = 1, 2, y t (t) = (v(t) x(t)) and v(t) = v(t)hv(t)i (similarly
for x). The correlations are given above so the dynamic pdf can be easily constructed.
There will be elements in the matrix that remain time-dependent for all times.
Exercise 7. Confront
hv m (t)xn (t)xk (t )i

hv m (t)xn (t)kxk1 (t )v(t )i ;

and

(2.66)

conclude.
Effect of a colored bath: anomalous diffusion
The anomalous diffusion of a particle governed by the generalized Langevin
equation, eq. (2.13), with colored noise characterized by power-law correlations as
the ones given in eq. (2.14), a problem also known as fractional Brownian motion,
was studied in detail by N. Pottier [21]. The particles velocity equilibrates with
the environment although it does at a much slower rate than in the Ohmic case:
its average and mean-square displacement decay as a power law - instead of exponentially - to their asymptotic values (still satisfying the regression theorem). The
48

particles mean square displacement is determined by the exponent of the noise-noise

correlation, hx2 (t)i t , i.e. the dynamics is subdiffusive for < 1, diffusive for
= 1 and superdiffusive for > 1. A time-dependent diffusion coefficient verifies Dx (t) 1/2 dhx2 (t)i/dt t1 : it is finite and given by eq. (2.58) for normal
diffusion, it diverges for superdiffusion and it vanishes for subdiffusion. The ratio
between the linear response and the time-derivative of the correlation ratio reads
T Rxx (t, t )/t Cxx (t, t ) = Dx (t t )/[Dx (t t ) + Dx (t )]. It approaches 1/2 for normal diffusion and the two-time dependent function 1/[1 + (t /(t t ))1 ] in other
cases.
Exercise 8. Work out these results.

2.4.2

Relaxation in a quadratic potential

Another relevant example is the relaxation of a particle in a harmonic potential,

with its minimum at x 6= 0:
V (x) =

k
(x x )2 ,
2

(2.67)

in contact with noise that we take to be white as the simpler starting case. The
potential confines the particle and one can then expect the coordinate to reach an
equilibrium distribution.
This problem can be solved exactly keeping inertia for all values of 0 but the
calculation is slightly tedious. The behavior of the particle velocity has already been
clarified in the constant force case. We now focus on the overdamped limit,
0 x = k(x x ) + ,

(2.68)

with k the spring constant of the harmonic well, that can be readily solved,
Z t

k
k
x(t) = x0 e 0 t + 01
dt e 0 (tt ) [(t ) + kx ] ,
x0 = x(0) .

(2.69)

This problem become formally identical to the velocity dependence in the previous
example.
Convergence of one-time quantities
The averaged position is
k

hx(t) x i = (x0 x )e 0 t 0

txr 0 /k

(Convergence)

(2.70)

Of course, one-time quantities should approach a constant asymptotically if the system

equilibrates with its environment.
Two-time quantities
49

The two-time connected correlation (where one extracts, basically, the asymptotic
position x ) reads
h k
i

k
(2.71)
hx(t)x(t )i = kB T k 1 e 0 (t+t ) e2 0 min(t,t ) 1 .

Again, the Dirichlet correlator (x(t) = x(t) hx(t)i). For at least one of the two
times going well beyond the position relaxation time txr = 0 /k the memory of the
initial condition is lost and the connected correlation becomes stationary:

Cc (t, t ) = hx(t)x(t )i kB T k 1 e 0 |tt |

min(t, t ) txr .

(2.72)

For time-differences that are longer than txd = 0 /k the correlation decays to 1/e and
one finds txd = txr . Interestingly enough, the relaxation and decay times diverge when
k 0 and the potential becomes flat.
Note that when the time-difference t t diverges the average of the product
factorizes, in particular, for the correlation one gets
hx(t)x(t )i hx(t)ihx(t )i x hx(t )i

(2.73)

for any t , even finite. We will see this factorization property at work later in more
complicated cases.
Fluctuation-dissipation theorem (FDT)
One can also compute the linear response to an infinitesimal perturbation that
couples linearly to the position changing the energy of the system as H H f x at
a given time t :

hx(t)if 
.
(2.74)
R(t, t ) =
f (t ) f =0
The explicit calculation yields

R(t, t ) = 01 ek0
R(t, t ) =

(tt )

Cc (t,t )
1
kB T
t

(t t )

(t t )

(FDT)

(2.75)

The last equality holds for times that are longer than txr . It expresses the fluctuationdissipation theorem (fdt), a model-independent relation between the two-time
linear response and correlation function. Similar - though more complicated - relations for higher-order responses and correlations also exist in equilibrium. There are
many ways to prove the fdt for stochastic processes. We will discuss one of them
in Sect. 2.3.2 that is especially interesting since it applies easily to problems with
correlated noise.
It is instructive to examine the relation between the linear response and the correlation function in the limit of a flat potential (k 0). The linear response is just
01 (t t ). The Dirichlet correlator approaches the diffusive limit:
hx(t)x(t )i = 201 kB T min(t, t )
50

for

k0

(2.76)

and its derivative reads t hx(t)x(t )i = 201 kB T (t t ). Thus,

1
t hx(t)x(t )i (t t )
2kB T

R(t, t ) =
R(t, t ) =

1
2kB T

t Cc (t, t ) (t t )

(FDR for diffusion)

(2.77)

A factor 1/2 is now present in the relation between R and Cc . It is another signature
of the fact that the coordinate is not in equilibrium with the environment in the
absence of a confining potential.
Exercise 9. Evaluate the two members of the FDT, eq. (2.75), in the case of the
tilted potential V (x) = F x.
Reciprocity or Onsager relations
Let us compare the two correlations hx3 (t)x(t )i and hx3 (t )x(t)i within the harmonic example. One finds hx3 (t)x(t )i =
3hx2 (t)ihx(t)x(t )i and hx3 (t )x(t)i = 3hx2 (t )ihx(t )x(t)i. Given that hx2 (t)i = hx2 (t )i
hx2 ieq and the fact that the two-time self-correlation is symmetric,
hx3 (t)x(t )i = hx3 (t )x(t)i .

(2.78)

With a similar argument one shows that for any functions A and B of x:
hA(t)B(t )i = hA(t )B(t)i
CAB (t, t ) = CAB (t , t)

(Reciprocity)

(2.79)

This equation is known as Onsager relation and applies to A and B that are even
under time-reversal (e.g. they depend on the coordinates but not on the velocities or
they have an even number of verlocities).
All these results remain unaltered if one adds a linear potential F x and works
with connected correlation functions.
Colored noise with exponential correlation
Exercise 10. Solve the stochastic dynamics of a particle in a harmonic potential with
a exponentially decaying memory kernel. Hint: use Laplace transform techniques.
Colored noise with power law correlation
Let us now take a power-law correlated noise. The Langevin equation can be solved
by using the Laplace transform. Correlation and linear responses can be computed.
As the system should equilibrate there is confining potential the FDT holds.
However, the decay of these two functions (and more complex ones involving more
times) are not trivial in the sense that their temporal dependence is not exponential.
Instead, the position correlation function and its linear response are given by the

51

Mittag-Leffer function


1
k|t t |
E,1
,
k
0


k|t t |
1
(t t ) ,
Rxx (t, t ) =
E,
0
0

Cxx (t, t ) =

(2.80)
(2.81)

where 0 is a constant that is proportional to 0 and all other pre-factors in (t t ).

For the Ohmic = 1 case the Mittag-Leffer function becomes an exponential, as
expected. For 6= 1 the decay is algebraic, E,1 (x) x1 that implies Cxx (t t )
|t t | . The ration between linear response and time derivative of the correlation
function is
1


T Rxx (t t )
E,1 (kt / 0 )E, (kt / 0 )
t

1
=
1
+
(2.82)
t Cxx (t t )
t
E, (k(t t ) / 0 )
In the long time limit, t t 1, the second term vanishes as long as k > 0 and one
recovers the equilibrium result.

2.4.3

Thermally activated processes

The phenomenological Arrhenius law7 yields the typical time needed to escape
from a potential well as an exponential of the ratio between the height of the barrier
and the thermal energy scale kB T , (with prefactors that can be calculated explicitly,
see below). This exponential is of crucial importance for understanding slow (glassy)
phenomena, since a mere barrier of 30kB T is enough to transform a microscopic time
of 1012 s into a macroscopic time scale. See Fig. 21-right for a numerical study of
the Coulomb glass that demonstrates the existence of an Arrhenius time-scale in this
problem. In the glassy literature such systems are called strong glass formers as
opposed to weak ones in which the characteristic time-scale depends on temperature
in a different way.
In 1940 Kramers estimated the escape rate from a potential well as the one
shown in Fig. 21-center due to thermal fluctuations that give sufficient energy to the
particle to allow it to surpass the barrier8 . After this seminal paper this problem has
been studied in great detail [19] given that it is of paramount importance in many
areas of physics and chemistry. An example is the problem of the dissociation of a
molecule where x represents an effective one-dimensional reaction coordinate and
the potential energy barrier is, actually, a free-energy barrier.
Kramers assumed that the reaction coordinate is coupled to an equilibrated environment with no memory and used the probability formalism in which the particle
motion is described in terms of the time-dependent probability density P (x, v, t) (that
for such a stochastic process follows the Kramers partial differential equation).
7 S. A. Arrhenius, On the reaction velocity of the inversion of cane sugar by acids, Zeitschrift f
ur
Physikalische Chemie 4, 226 (1889).
8 H. A. Kramers, Brownian motion in a field of force and the diffusion model of chemical reactions,
Physica 7, 284 (1940).

52

Figure 21: Left: sketch of a double-well potential. Center: sketch of a potential with
a local minimum. Right: correlation function decay in a classical model of the 3d
Coulomb glass at nine temperatures ranging from T = 0.1 to T = 0.05 in steps of
0.05 and all above Tg . In the inset the scaling plot C(t) f (t/tA ) with a characteristic
time-scale, tA , that follows the Arrhenius activated law, tA 0.45/T . Figure due to
Kolton, Domnguez and Grempel [22].
If the thermal energy is at least of the order of the barrier height, kB T V ,
the reaction coordinate, x, moves freely from the vicinity of one well to the vicinity
of the other.
The treatment we discuss applies to the opposite weak noise limit in which the
thermal energy is much smaller than the barrier height, kB T V , the random
force acts as a small perturbation, and the particle current over the top of the barrier
is very small. Most of the time x relaxes towards the minimum of the potential well
where it is located. Eventually, the random force drives it over the barrier and it
escapes to infinity if the potential has the form in Fig. 21-center, or it remains in the
neighbourhood of the second well, see Fig. 21-left.
The treatment is simplified if a constant current can be imposed by injecting
particles within the metastable well and removing them somewhere to the right of it.
In these conditions Kramers proposed a very crude approximation whereby P takes
the stationary canonical form
Pst (x, v) = N e

v2
2

V (x)

(2.83)

(m = 1 for simplicity here.) If there is a sink to the right of the maximum, the
normalization constant N is fixed by further assuming that Pst (x, v) 0 for x x
>
xmax . The resulting integral over the coordinate can be computed with a saddle-point
approximation justified in the large limit. After expanding the potential about the
minimum and keeping the quadratic fluctuations one finds
N 1 =

V (xmin )
53

eV (xmin ) .

The escape rate, r, over the top of the barrier can now be readily computed by
calculating the outward flow across the top of the barrier:
p
Z
V (xmin ) (V (xmax )V (xmin ))
1
dv vP (xmax , v) =

e
.
(2.84)
r
tA
2
0
Note that we here assumed that no particle comes back from the right of the barrier.
This assumption is justified if the potential quickly decreases on the right side of the
barrier.
The crudeness of the approximation (2.83) can be grasped by noting that the
equilibrium form is justified only near the bottom of the well. Kramers estimated an
improved Pst (x, v) that leads to

r=

2
4

1/2
+ V (xmax )

p
V (xmax )

V (xmin ) (V (xmax )V (xmin ))

e
.
2

(2.85)

This expression approaches (2.84) when V (xmax ), i.e. close to the underdamped
limit, and
p
V (xmax )V (xmin ) (V (xmax )V (xmin ))
r=
e
(2.86)
2
when V (xmax ), i.e. in the overdamped limit (see Sect. 2.3.4 for the definition
of these limits).
The inverse of (2.85), tA , is called the Arrhenius time needed for thermal
activation over a barrier V V (xmax ) V (xmin ). The prefactor that characterises
the well and barrier in the harmonic approximation is the attempt frequency with
which the particles tend to jump over the barrier. In short,
tA e|V |

(Arrhenius time)

(2.87)

The one-dimensional reaction coordinate can be more or less easily identified in

problems such as the dissociation of a molecule. In contrast, such a single variable
is much harder to visualize in an interacting problem with many degrees of freedom.
The Kramers problem in higher dimensions is highly non-trivial and, in the infinitedimensional phase-space, is completely out of reach.
The Arrhenius time can be derived within the path-integral formalism that we
will discuss later [23, 24] (see App. ).

2.5

Driven systems

In the introduction we mentioned that systems can be externally maintained out of

equilibrium We list here two solvable examples, in the form of exercises, that illustrate
this point.

54

Exercise 11. Study the Langevin equation for a single particle moving in d = 1
under no external potential, in a case in which the friction kernel is 1 (t t ) and the
noise-noise correlation in 2 (t t ).
Exercise 11. Take a harmonic oscillator, in its over-damped limit to make the calculations simpler, and couple it to two external reservoirs, at different temperatures,
T1 and T2 , and different memory kernels, for instance, a delta function (white noise)
and an exponential decay (Ornstein-Uhlenbeck process). The Langevin relaxation of
the particle can be solved exactly and it is quite interesting. The particle inherits the
two time-scales (1 0 and 2 finite) from the baths as can be seen from the decay,
in two steps, of the position correlation function or linear response. The temperatures
of the environments appear in the fluctuation dissipation relation between these two
functions in the corresponding time regimes [31, 32].
Exercise 12. Take now a symmetric two-dimensional harmonic oscillator V (x, y) =
k(x2 + y 2 )/2 and apply the non-potential force f~(x, y) = a(y, x) on it, with a a
parameter. This force makes a particle turn within the potential well. Describe the
trajectories and compute mean-square displacement, correlation function and linear
response. One can check, by direct calculation, that the fluctuation-dissipation theorem does not hold.

The Langevin equation and its relation to the Fokker-Planck formalism have been
described in many textbooks on stochastic processes including Riskens [26], Gardiners [27] and van Kampens [28]. Many applications can be found in Coffrey et
al.s [29]. Another derivation of the Langevin equation uses collision theory and
admits a generalization to relativistic cases [19]. The alternative master equation
description of stochastic processes, more adapted to deal with discrete variables, is
also very powerful but we will not use it is these lectures.

55

Dynamics at or through a phase transition

Take a piece of material in contact with an external reservoir. The material will be
characterized by certain observables, energy density, magnetization density, etc. The
external environment will be characterized by some parameters, like the temperature,
magnetic field, pressure, etc. In principle, one is able to tune the latter and study
the variation of the former. Note that we are using a canonical setting in the sense
that the system under study is not isolated but open.
Sharp changes in the behavior of macroscopic systems at critical points (or lines)
in parameter space have been observed experimentally. These correspond to equilibrium phase transitions, a non-trivial collective phenomenon appearing in the
thermodynamic limit. We will assume that the main features of, and analytic approaches used to study, phase transitions are known.
Imagine now that one changes an external parameter instantaneously or with a
finite rate going from one phase to another in the (equilibrium) phase diagram. The
kind of internal system interactions are not changed. In the statistical physics language the first kind of procedure is called a quench and the second one an annealing
and these terms belong to the metalurgy terminology. We will investigate how the
system evolves by trying to accomodate to the new conditions and equilibrate with
its environment. We will first focus on the dynamics at the critical point or going
through phase transitions between well-known phases (in the sense that one knows
the order parameter, the structure, and all thermodynamic properties on both sides
of the transition). Later we will comment on cases in which one does not know all
characteristics of one of the phases and sometimes one does not even know whether
there is a phase transition.

upper critical
critical
lower critical

upper critical
critical
lower critical

Figure 22: Left: second-order phase transition. Right: first order phase transition.

56

The evolution of the free-energy landscape (as a function of an order parameter)

with the control parameter driving a phase transition is a guideline to grasp the
dynamics following a quench or annealing from, typically, a disordered phase to the
phase transition or into the ordered phase. See Fig. 22 for a sketch. We will discuss
quenches to the phase transition and below it. In the former case, the system can
get to a critical point (Fig. 22-left) in which the free-energy is metastable in the
sense that its second derivative vanishes (second order phase transition cases) or to
a first-order phase transition (Fig. 22-right) in which various minima are degenerate.
In the latter case the initial state becomes unstable, that is to say a maximum,
and the phase transition is of second-order (see Fig. 22-left) or metastable, that
is to say a local minimum, and the phase transition is of first order (see Fig. 22right) in the final externally imposed conditions.9 In the former case the ordering
process occurs throughout the material, and not just at nucleation sites. Two
typical examples are spinodal decomposition, i.e. the method whereby a mixture
of two or more materials can separate into distinct regions with different material
concentrations, or magnetic domain growth in ferromagnetic materials. Instead, in
the latter case, the stable phase conquers the system through the nucleation of a
critical localized bubble via thermal activation and its further growth.
Having described the dependence of the free-energy landscape on the external
parameters we now need to choose the microscopic dynamics of the order parameter.
Typically, one distinguishes two classes: one in which the order parameter is locally
conserved and another one in which it is not. Conserved order parameter dynamics
are found for example in phase separation in magnetic alloys or inmiscible liquids.
Ferromagnetic domain growth is an exemple of the non-conserved case.
The kinetics of systems undergoing critical dynamics or an ordering process is
an important problem for material science but also for our generic understanding of
pattern formation in non-equilibrium systems. The late stage dynamics is believed
to be governed by a few properties of the systems whereas material details should
be irrelevant. Among these relevant properties one may expect to find the number
of degenerate ground states, the nature of the conservation laws and the hardness or
softness of the domain walls that is intimately related to the dimension of the order
parameter. Thus, classes akin to the universality ones of critical phenomena have
been identified. These systems constitute a first example of a problem with slow
dynamics. Whether all systems with slow dynamics, in particular structural and
spin glasses, undergo some kind of simple though slow domain growth is an open
question.

3.1

Snapshots

Take a magnetic system, such as the ubiquitous Ising model with ferromagnetic
uniform interactions, and quench it to its Curie point or into the low temperature
9 Strictly

speaking metastable states with infinite life-time exist only in the mean-field limit.

57

200

200

data

200

data

150

150

150

100

100

100

50

50

50

50

100

150

200

50

100

150

200

data

50

100

150

200

Figure 23: Monte Carlo simulations of a 2d Ising model. Three snapshots at t =

1, 3 105 , 3 106 MCs after a quench to Tc .
200

200

data

200

data

150

150

150

100

100

100

50

50

50

50

100

150

200

50

100

150

200

data

50

100

150

200

Figure 24: Monte Carlo simulations of a 2d Ising model. Three snapshots at t =

1, 3 105 , 3 106 MCs after a quench to 0.5 Tc . Thermal fluctuations within the
domains are visible.
phase starting from a random initial condition. Classically, the spins do not have an
intrinsic dynamics; it is defined via a stochastic rule of Glauber, Metropolis, Langevin
(when soft-spins are used) or similar type with or without locally conserved magnetization. For the purpose of the following discussion it is sufficient to focus on nonconserved local microscopic dynamics. Three snapshots taken after times 1, 3 105
and 3 106 MCs in a critical and two sub-critical quenches are shown in Figs. 23, 24,
and 25, respectively.
Time-dependent macroscopic observables are then expressed in terms of the values
of the spins at each time-step. For instance, the magnetization density and its twotime self correlation function are defined as
m(t) N 1

N
X
i=1

h si (t) i ,

C(t, t ) N 1

N
X
i=1

h si (t)si (t ) i ,

(3.1)

where the angular brackets indicate an average over many independent runs (i.e. ran58

dom numbers) starting from identical initial conditions and/or averages over different
initial configurations.
In critical quenches, patches with equilibrium critical fluctuations grow in time
but their linear extent never reaches the equilibrium correlation length that diverges.
Clusters of neighbouring spins pointing the same direction of many sizes are visible
in the figures and the structure is quite intricate with clusters within clusters and so
on and so forth. The interfaces look pretty rough too.
In quenches into the ordered phase through a second order phase transition the ferromagnetic interactions tend to align the neighbouring spins in parallel
direction and in the course of time domains of the two ordered phases form and grow,
see Fig. 26. At any finite time the configuration is such that both types of domains
exist. If one examines the configurations in more detail one reckons that there are
some spins reversed within the domains. These errors are due to thermal fluctuations and are responsible of the fact that the magnetization of a given configuration
within the domains is smaller than one and close to the equilibrium value at the working temperature (apart from fluctuations due to the finite size of the domains). The
total magnetization, computed over the full system, is zero (up to fluctuating timedependent corrections that scale with the square root of the inverse system size). The
thermal averaged spin, hsi (t)i vanishes for all i and all finite t, see below for a more
detailed discussion of the time-dependence. As time passes the typical size of the
domains increases and the interfaces get flatter in a way that we will also discuss
below.
Quenches across first order phase transitions will be discussed separately
below.
200

200

data

200

data

150

150

150

100

100

100

50

50

50

50

100

150

200

50

100

150

200

data

50

100

150

200

Figure 25: Monte Carlo simulations. Three snapshots at t = 1, 3 105 , 3 106 MCs
after a quench to 0.01 Tc . There is almost perfect order within the domains (meq 1).

3.2

Relaxation, equilibration and reversal times

59

Figure 26: Snapshot of the 2d Ising model at a number of Monte Carlo steps after
a quench from infinite to a subcritical temperature. Left: the up and down spins on
the square lattice are represented with black and white sites. Right: the domain walls
are shown in black.
We wish to distinguish the relaxation time, tr , defined as the time needed for a
given initial condition to reach equilibrium in one of the (possibly many equivalent)
phases, from the decorrelation time, td , defined as the time needed for a given
configuration to decorrelate from itself. To lighten the notation we do not signal out
the variable that we use to study these typical times (as we did with the velocity and
position in the examples of Sect. 2.4). We further define the reversal time, tR , as
the time needed to go from one to another of the equivalent equilibrium phases. We
focus on second-order phase transitions here.
3.2.1

Quench from T Tc to T > Tc

If one quenches the system to T > Tc the relaxation time, tr , needed to reach
configurations sampled by the Boltzmann measure depends on the systems parameters but not on its size. Hence it is finite even for an infinite-size system. Once a
short transient overcome, the average of a local spin approaches the limit given by the
Boltzmann measure, hsi (t)i hsi ieq = m = 0, for all i and all other more complex
observables satisfy equilibrium laws. The relaxation time is estimated to behave as
|T Tc |zeq close to Tc , with the critical exponent characterizing the divergence of
the equilibrium correlation length, eq (T Tc ) , and zeq the equilibrium exponent
that links times and lengths, Rc (t) t1/zeq .
The relaxation of the two-time self-correlation at T > Tc , when the time t is

60

Figure 27: Snapshot of the 2d Potts model with q = 3 at a number of Monte Carlo
steps after a quench from the critical to a subcritical temperature.
chosen to be longer than tr , decays exponentially

lim hsi (t)si (t )i e(tt )/td

t t

(3.2)

with a decorrelation time that increases with decreasing temperature and close to
(but still above) Tc diverges as the power law, td (T Tc )zeq . The divergence of
td is the manifestation of critical slowing down. The asympotic value verifies
lim

tt t tr

hsi (t)si (t )i = lim hsi (t)i lim hsi (t )i = hsi ieq hsi ieq = m2 = 0 ,
ttr

t tr

(3.3)

cfr. eq. (2.73).

3.2.2

Quench from T Tc to T Tc

At or below Tc , coarsening from an initial condition that is not correlated with

the equilibrium state and with no bias field does not take the system to equilibrium
in finite times with respect to a function of the systems linear size, L. More explicitly,
if the growth law is a power law [see eq. (3.22)] one needs times of the order of Lzeq
(critical) or Lzd (subcrititcal) to grow a domain of the size of the system. This gives
a rough idea of the time needed to take the system to one of the two equilibrium
states. For any shorter time, domains of the two types exist and the system is out
of equilibrium.
The self-correlation of such an initial state evolving at T Tc involves power laws
or logarithms and although one cannot associate to it a decay time as one does to an
exponential, one can still define a characteristic time that, quite generally, turns out
to be related to the age of the system, td tw [see eq. (3.21)].
In contrast, the relaxation time of an equilibrium magnetized configuration at
temperature T vanishes since the system is already equilibrated while the decorrelation time td is a finite function of T .

61

The relaxation of the two-time self-correlation at T < Tc , when the time t is

chosen to be longer than tr , that is to say, once the system has thermalized in one of
the two equilibrium states, decays exponentially

hsi (t)si (t )i e(tt )/td

(3.4)

with a decorrelation time that decreases with decreasing temperature and close to Tc
(but below it) also diverges as a power law, td (T Tc )zeq . The asympotic value
verifies
lim

tt t tr

hsi (t)si (t )i = lim hsi (t)i lim hsi (t )i = hsi ieq hsi ieq = m2 0 ,
ttr

t tr

(3.5)

cfr. eqs. (2.73) and (3.3), depending on T = Tc or T > Tc .

3.2.3

Summary

The lesson to learn from this comparison is that the relaxation time and the
decorrelation time not only depend upon the working temperature but they also
depend upon the initial condition. Moreover, in all critical or low-temperature cases
we will study the relaxation time depends on (L, T ) and diverges in the infinite
size limit while the decorrelation time depends on (T, tw ). For a random initial
condition and an infinite system one has

T > Tc ,
finite
>
tr
T Tc ,
|T Tc |zeq

T Tc
while for a finite system

tr

Lzeq
Lz d

T = Tc ,
T < Tc .

Still another time scale is given by the time needed to reverse an equilibrium
configuration in the low-T phase. This one is expected to be given by an Arrhenius
law, with the height of the barrier being determined by the extensive free-energy
barrier between the two minima, i.e. F Ld f , therefore,
tR eL

Reversal time-scale .

(3.6)

The Ginzburg-Landau description allows for a pictorial interpretation of these results. The dynamics of the full system is visualized as the motion of its representative
point in the Ginzburg-Landau potential. At high T the potential is harmonic in the
deterministic Allen-Cahn equation, or the double-well structure in the time-dependent
stochastic Ginzburg-Landau equation is completely ignored. The relaxation is similar
to the one of a particle in a harmonic potential studied in Sect. 2.4.2. At low T , the
62

initial position in the double-well potential depends on the type of initial condition
(~x, 0) = 0 or (~x, 0) 6= 0. In the first case, the point sits on top of the central barrier
and it does not detach from it in finite times with respect to a function of L. In the
second case, the point starts from within one well and it simply rolls to the bottom
of the well. This relaxation is similar to the one in the harmonic case. To reverse the
configuration from, say, positive to negative magnetization the point needs to jump
over the barrier in the double well potential and it does via thermal activation ruled
by the Arrhenius law.
Note however that the phase-space of the system is actually N -dimensional while
the description that is given here is projected onto one single coordinate, the one of
the order-parameter. This reduction might lead to some misunderstandings and one
should be very careful with it.

3.3

Growing length and dynamic scaling

In usual coarsening systems the averaged space-time correlation function

P
Ld C(r, t) = ij/|~ri ~rj |=r hsi (t)sj (t)i

allows for the identification of a growing length from, for example,

Z
Z
d
a+1
Ra (T, t) d r r C(r, t)/ dd r ra C(r, t)

(3.7)

(a is a parameter chosen to weight preferentially short or long distances; the timedependence of Ra (t) should not depend on a.) Here and in the following h. . .i
stands for an average over different realizations of thermal histories at heat-bath
temperature T and/or initial conditions. In presence of quenched disorder one adds
an average
over it and denotes it [. . .]. The stochastic time-dependent function
P
N 1 ij/|~ri ~rj |=r si (t)sj (t) after a quench from a random initial condition does not
fluctuate in the thermodynamic limit. Therefore, the averages are not really necessary
but they are usually written down. In spin-glasses and glasses this observable does
not yield information on the existence of any growing length as we will discuss below.
The spherically averaged structure factor S(k, t) the Fourier transform of C(r, t)
can be measured experimentally with small-angle scattering of neutrons, x-rays or
light and from it Ra (T, t) can be extracted.
The ordering process is characterized by the growth of a typical length, R(T, t).
The growth regimes are summarized in the following equation and in Fig. 28:

T > Tc saturation,
Rc (t) eq (T ) < +
Rc (t) eq (T )
T = Tc critical coarsening,
(3.8)

Rc (t) eq (T ) < R(T, t) L T < Tc sub-critical coarsening.

After a quench to the high temperature phase T > Tc the system first grows equilibrium regions until reaching the equilibrium correlation length eq and next relaxes in
equilibrium as explained in the previous section. The correlation length could be very
63

eq

R(T, t)

Tc
Figure 28: Sketch of the growth process in a second-order phase transition. The
thick line is the equilibrium correlation length eq |T Tc | . The thin solid (red)
arrows indicate the growing length Rc in the critical coarsening regime and the dashed
(black) arrow the sub-critical growing length R in the coarsening regime.
short and the transient non-equilibrium regime be quite irrelevant (T Tc ). In the
critical region, instead, the equilibrium correlation length is very long and it becomes
important. In a critical quench the system never orders sufficiently and R(Tc , t) <
for all finite times. Finally, a quench into the subcritical region is characterized by
two growth regimes: a first one in which the critical point dominates and the growth
is as in a critical quench; a second one in which the proper sub-critical ordering is at
work. The time-dependence of the growth law is different in these two regimes as we
will see below. (Note that below Tc the equilibrium correlation length eq does not
measure the size of ordered regions but the typical distance until which a fluctuation
has an effect.)
In the asymptotic time domain, when R(T, t) has grown much larger than any
microscopic length in the system, a dynamic scaling symmetry sets in, similarly
to the usual scaling symmetry observed in equilibrium critical phenomena. According
to this hypothesis, the growth of R(T, t) is the only relevant process and the whole
time-dependence enters only through R(T, t).

64

3.4

Critical coarsening

The scaling behavior of binary systems quenched to the critical point is quite well
understood. It can be addressed with scaling arguments and renormalization group
approaches [5] which give explicit expressions for many of the quantities of interest
up to two loops order. Numerical simulations confirm the analytic results and probe
exponents and scaling functions beyond the available perturbative orders. In this
case the system builds correlated critical Fortuin-Kasteleyn clusters 10 with fractal
dimension DF K = (d + 2 )/2, where is the usual static critical exponent, in
regions growing algebraically as Rc (Tc , t) Rc (t) t1/zeq ; henceforth we simplify
the notation and avoid writing Tc within R. [As an example, for the bidimensional
critical Ising class = 1/4 and DF K = (2 + 2 1/4)/2 = 15/8.]
In the asymptotic time regime the space-time correlation function has the scaling
form
C(t, t) = Cst (r)Cag (r, t)


r
C(r, t) = Cst (r) f
Rc (t)


r
= r2(dDF K ) f
Rc (t)


r
2d
Multiplicative separation.
C(r, t) = r
f
Rc (t)

(3.9)

The pre-factor r2(dDF K ) takes into account that the growing domains have a fractal
nature (hence their density decreases as their size grows) and the fact that the order
parameter vanishes at the second order critical point. The dependence on r/Rc (t)
in f (x) expresses the similarity of configurations at different times once lengths are
measured in units of Rc (t). At distances and times such that r/Rc (t) 1 the
equilibrium power-law decay, Ceq (r) r2d , should be recovered, thus f (x) ct
at x 0. f (x) falls off rapidly for x 1 to ensure that spins are uncorrelated at
distances larger than Rc (t). [More precisely, correlated as in the initial condition that,
in the case of a quench from infinite temperature, means indeed uncorrelated.]
For two-time quantities, when t is sufficiently large one has


Rc (t)

(3.10)
C(t, t ) = Cst (t t ) fc
Rc (t )


Rc (t)
C(t, t ) = Rc (t t )2d fc
Multiplicative separation.
Rc (t )
10 The Fortuin-Kasteleyn clusters are constructed as follows. Starting with a spin domain, one first
draws all bonds linking nearest-neighbor spin on the cluster and then erases bonds with a temperature
dependent probability e J. In such a way, the original bond-cluster typically diminishes in size
and may even get disconnected.

65

100
qea

aging
Cag

101

(b)

(a)
102
101

stationary Cst

101

103

105

t tw

101

101

103

105

t tw

Figure 29: Sketch of the decay of the two-time correlation at Tc (a) and T < Tc (b)
for different values of the waiting-time, increasing from left to right.
Here Cst (t t ) Rc (t t )2(dDF K ) = Rc (t t )2d . The scaling function fc (x)
describes the non-equilibrium behavior. It satisfies fc (1) = 1 and fc (x ) =
0, see the sketch in Fig. 29 (a). The former condition ensures that equilibrium is
established up to the length Rc (t). The latter that decorrelation occurs faster than
what the equilibrium relaxation tells beyond this length. In the scaling forms the
equilibrium and non-equilibrium contributions enter in a multiplicative structure.
Non-equilibrium effects are taken into account by taking ratios between the sizes of
the correlated domains at the observation times t and t in the scaling functions. Note
that the reason why the equilibrium results are recovered for t t is that for very
similar times one does not let the system realize that it is out of equilibrium.
In the case of non-conserved scalar order-parameter dynamics the growing length
behaves as
Rc (t) t1/zeq
(3.11)
with zeq the equilibrium dynamics exponent (note that zeq is different from zd ). We
will not discuss critical dynamics in detail; this problem is treated analytically with
dynamic renormalization group techniques and it is very well discussed in the literature [5]. In short, the exponent zeq is given by [33]


4 1 2
N +2
(3.12)
+ O(3 )
3 ln
zeq = 2 +
(N + 8)2
3 2
where N is the dimension of the possibly vector field, N = 1 for a scalar one, and
= 4 d with d the dimension of space. Note that zeq is larger than 2 for all finite
N and it approaches 2 in the large N limit (at least up to this order in perturbation

66

theory). In particular, one finds

zeq

2.0538
2.0134

d=2
d=3
d=4

(3.13)

for N = 1. d = 4 is the upper critical dimension in this problem. Numerical simulations indicate zeq 2.13 in d = 2. These results are valid for white noise dynamics.
The effect of colored noise is to change the value of the exponent zeq when it is
sufficiently long-range correlated (sub-Ohmic noise with a power-law decay with an
exponent smaller than a critical value that depends on the dimension of space).
The multiplicative scaling is also obtained with the dynamic RG mthod.

Sub-critical coarsening

0.6
C(r,t)

t=4
8
16
32
64
128
256
512
1024

0.8

0.4
0.2

0.6
0.4
0.2

-0.2

-0.2
0

20

40

60

80

t=4
8
16
32
64
128
256
512
1024

0.8

C(r,t)

3.5

100

120

1/2

r/t

Figure 30: The equal-time correlation as a function of distance in the 2dIM quenched
below Tc . Raw (left) and scaled (right) data. These numerical simulations were
performed by A. Sicilia.

3.5.1

Dynamic scaling hypothesis

The dynamic scaling hypothesis states that at late times and in the scaling
limit
r (g) ,
R(g, t) (g) ,
r/R(g, t) arbitrary ,
(3.14)

where r is the distance between two points in the sample, r |~x ~x |, and eq (g)
is the equilibrium correlation length that depends on all parameters (T and possibly
others) collected in g, there exists a single characteristic length, R(g, t), such
that the domain structure is, in statistical sense, independent of time when lengths
are scaled by R(g, t). Time, denoted by t, is typically measured from the instant
when the critical point is crossed. In the following we ease the notation and write
67

only the time-dependence in R. This hypothesis has been proved analytically in very
simple models only, such as the one dimensional Ising chain with Glauber dynamics
or the Langevin dynamics of the d-dimensional O(N ) model in the large N limit (see
Sect. 3.7).
The late stage of phase-ordering in binary systems is characterized by a patchwork
of large domains the interior of which is basically thermalized in one of the two
equilibrium phases while their boundaries are slowly moving. This picture suggests the
splitting of the degrees of freedom (spins) into two categories, providing statistically
independent contributions to observables such as correlation or response functions.
More precisely, a quasi-equilibrium stationary contribution arises as due to bulk spins,
while boundaries account for the non-equilibrium part. Then asymptotically one has
C(r, t) Cst (r) + Cag (r, t)

Additive separation.

(3.15)

The first term describes the equilibrium fluctuations in the low temperature broken
symmetry pure states
 
r
Cst (r) = (1 hsi i2eq ) g
,
(3.16)

where hsi ieq is the equilibrium expectation value of the local spin in one of the two
symmetry breaking states, hsi ieq = m, and g(x) is a function with the limiting values
g(0) = 1, limx g(x) = 0. The second term takes into account the motion of the
domain walls through


r
,
(3.17)
Cag (r, t) = hsi i2eq f
R(t)

with f (1) = 1 and limx f (x) = 0. Both Cst and Cag obey (separately) scaling
forms with respect to the equilibrium and the non-equilibrium lengths , R(t). In
particular, eq. (3.17) expresses the fact that system configurations at different times
are statistically similar provided that lengths are measured in units of R(t), namely
the very essence of dynamical scaling.
Monte Carlo simulations of the Ising model and other systems quenched below
criticality and undergoing domain growth demonstrate that in the long waiting-time
limit t t0 , the spin self-correlation hsi (t)si (t )i separates into two additive terms
C(t, t ) Cst (t t ) + Cag (t, t )

Additive separation

(3.18)

see Fig. 31, with the first one describing equilibrium thermal fluctuations within the
domains,

1 hsi i2eq = 1 m2 ,
t t = 0 ,
Cst (t t ) =
(3.19)
0,
t t ,
and the second one describing the motion of the domain walls
 

R(t)
hsi i2eq ,
t t ,
=
Cag (t, t ) = hsi i2eq fc

0,
t t .
R(t )
68

(3.20)

To ease the notation we have not written the explicit T -dependence in R that, as
we will see below, is less relevant than t. Note that by adding the two contributions
one recovers C(t, t) = 1 as expected and C(t, t ) 0 when t t . The first term
is identical to the one of a system in equilibrium in one of the two ordered states,
see eq. (3.5) for its asymptotic t t t limit; the second one is inherent to the
out of equilibrium situation and existence and motion of domain walls. They vary in
completely different two-time scales. The first one changes when the second one is
fixed to hsi i2eq , at times such that R(t)/R(t ) 1. The second one varies when the
first one decayed to zero. The mere existence of the second term is the essence of
the aging phenomenon with older systems (longer t ) having a slower relaxation than
younger ones (shorter t ). The scaling of the second term as the ratio between two
lengths is a first manifestation of dynamic scaling.
A decorrelation time can also be defined in this case by expandind the argument of the scaling function around t t. Indeed, calling t t t one has
R(t)/R(t ) R(t + t)/R(t ) [R(t ) + R (t )t]/R(t ) 1 + t/[d ln R(t )/dt ]1
and one identifies a t -dependent decorrelation time
td [d ln R(t )/dt ]1

decorrelation time

(3.21)

which is, in general, a growing function of t .

C(t,tw)

C(t,tw)

tw=2
4
8
16
32
64
128
256
512

0.1

tw=2
4
8
16
32
64
128
256
512

0.1

10

100

1000

t-tw

10

100

1000

t-tw

Figure 31: The two-time self-correlation in the 2dIM with non-conserved order parameter dynamics at several waiting-times given in the key at temperature T = 0.5
(left) and T = 2 (right). Data obtained with Monte Carlo simulations. Note that the
plateau is located at a lower level in the figure on the right consistently with the fact
that hieq decreases with increasing temperature. Data from A. Sicilia et al.
In order to fully characterise the correlation functions one then has to determine
the typical growing length, R, and the scaling functions, g, f , fc , etc. It turns out that
the former can be determined with semi-analytic arguments and the predictions are
well verified numerically at least for clean system. The latter, instead, are harder
69

to obtain. We will give a very brief state of the art report in Sect. 3.6.3. For a much
more detailed discussion of these methods see the review articles in [4].
The time-dependent typical domain length, R(t), is determined numerically by
using several indirect criteria or analytically within certain approximations. The
most common ways of measuring R are with numerical simulations of lattice models
or the numerical integration of the continuous partial differential equation for the
evolution of the order parameter. In both cases one
Computes the inverse perimeter density R(t) = hHieq /[hH(t)i hHieq ] with
hH(t)i the time-dependent averaged energy and hHieq the equilibrium energy both
measured at the working temperature T .
Puts the dynamic scaling hypothesis to the test and extracts R from the analysis.

3.5.2

R(t) in clean one dimensional cases with non-conserved order parameter

In one dimension, a space-time graph allows one to view coarsening as the diffusion and annhilitation upon collision of point-like particles that represent the domain
walls. In the Glauber Ising chain with non-conserved dynamics one finds that the
typical domain length grows as t1/2 while in the continuous case the growth is only
logarithmic, ln t.
3.5.3

R(t) in non-conserved order parameter curvature dynamics in d > 2

In this case the growth law scales as (T )t1/2 as we will show in Sec. ?? using the
time-dependent Ginzburg-Landau equation.
There are a number of ways to find the growth law
R(t) = t1/zd

(3.22)

with zd the dynamic exponent, in pure and isotropic systems (see [4]). The
effects of temperature enter only in the parameter and, for clean systems, growth is
slowed down by an increasing temperature since thermal fluctuation tend to roughen
the interfaces thus opposing the curvature driven mechanism. We estimate the T
dependence of in Sect. 3.5.5.
In curvature driven Ising or Potts cases with non-conserved order parameter the
domains are sharp and zd = 2 with a weakly T -dependent coefficient. For systems
with continuous variables such as rotors or XY models and the same type of dynamics,
a number of computer simulations have shown that domain walls are thicker and
zd = 4.
3.5.4

R(t) in conserved order parameter dynamics and the role of bulk

diffusion

70

A different type of dynamics occurs in the case of phase separation (the water
and oil mixture ignoring hydrodynamic interactions or a binary allow). In this case,
the material is locally conserved, i.e. water does not transform into oil but they just
separate. The main mechanism for the evolution is diffusion of material through the
bulk of the opposite phase. After some discussion, it was established, as late as in
the early 90s, that for scalar systems with conserved order parameter zd = 3.
3.5.5

Crossover between critical and sub-critical coarsening

Matching critical coarsening with sub-critical one allows one to find the T -dependent
prefactor [37]. The argument goes as follows. The out of equilibrium growth at
criticality and in the ordered phase are given by
 1/z
at
T = Tc ,
t eq
R(t)
(3.23)
((T )t)1/zd
at
T < Tc .
zeq is the equilibrium dynamic critical exponent and zd the out of equilibrium growth
exponent. Close but below criticality one should have an interpolating expression of
the kind


t
at
T = Tc
(3.24)
R(t) a t1/zd f
zeq
with the T -dependent equilibrium correlation length, eq (T ) (Tc T ) . The
last factor tends to one, f (x ) 1, when R(t) , that is to say when
the argument diverges and the system enters the sub-critical coarsening regime. It is
however non-trivial when R(t) , the argument is finite and critical coarsening must
be described. In particular, we determine its behavior for x = O(1) by requiring that
eq. (3.24) matches the subcritical growing length which is achieved by (i) recovering
the correct t dependence, (ii) cancelling the factor. (i) implies
f (x) x1/zd +1/zeq
Then eq. (3.24) becomes

for

x = O(1) .

(3.25)

R(t) a+zeq /zd 1 t1/zeq

(3.26)

a = zeq /zd 1 .

(3.27)

and to eliminate we need

Comparing now the subcritical growing length and (3.24) in the very long times limit
such that R(t) and f (x ) 1:
[(T )]1/zd a (Tc T )(zeq zd )/zd .
Note that quite generally one finds zeq > zd and (T ) vanishes at Tc .
3.5.6

Role of weak disorder: thermal activation

71

(3.28)

The situation becomes much less clear when there is weak quenched disorder in the
form of non-magnetic impurities in a magnetic sample, lattice dislocations, residual
stress, etc. These are assumed not to modify the nature of the equilibrium states with
respect to the ones of the clean system. Qualitatively, the dynamics are expected to be
slower than in the pure cases since disorder pins the interfaces. In general, based on an
argument due to Larkin (and in different form to Imry-Ma) one expects that in d < 4
the late epochs and large scale evolution is no longer curvature driven but controlled
by disorder. Indeed, within a phase space view disorder generates metastable states
that trap the system and thus slow down the relaxation.
A hand-waving argument to estimate the growth law in dirty systems is the following. Take a system in one equilibrium state with a domain of linear size R of the
opposite equilibrium state within it. This configuration could be the one of an excited state with respect to the fully ordered one with absolute minimum free-energy.
Call F (R) the free-energy barrier between the excited and equilibrium states. The
thermal activation argument (see Sect. 2.3) yields the activation time scale for the
decay of the excited state (i.e. erasing the domain wall)
tA eF (R)/(kB T ) .

(3.29)

For a barrier growing as a power of R, F (R) (T, J)R (where J represents the
disorder) one inverts (3.29) to find the linear size of the domains still existing at time
t, that is to say, the growth law
R(t)

kB T
(T )

ln t

1/

(3.30)

All smaller fluctuation would have disappeared at t while typically one would find
objects of this size. The exponent is expected to depend on the dimensionality of
space but not on temperature. In normal systems it is expected to be just d 1
the surface of the domain but in spin-glass problems, it might be smaller than d 1
due to the presumed fractal nature of the walls. The prefactor is expected to be
weakly temperature dependent.
One assumes that the same argument applies out of equilibrium to the reconformations of a portion of any domain wall or interface where R is the observation
scale.
However, already for the (relatively easy) random ferromagnet there is no consensus about the actual growth law. In these problems there is a competition between the pure part of the Hamiltonian, that tries to minimize the total (d 1)
dimensional area of the domain wall, and the impurity part that makes the wall
deviate from flatness and pass through the locations of lowest local energy (think of
Jij = J + Jij with J and Jij contributing to the pure and impurity parts of the
Hamiltonian, respectively). The activation argument in eq. (3.29) together with the
power-law growth of barriers in F (R) (T, J)R imply a logarithmic growth
of R(t). Simulations, instead, suggest a power law with a temperature dependent
exponent. Whether the latter is a pre-asymptotic result and the trully asymptotic
72

one is hidden by the premature pinning of domain walls or it is a genuine behavior

invalidating F (R) (T, J)R or even eq. (3.29) is still an open problem. See
the discussion below for a plausible explanation of the numerical data that does not
invalidate the theoretical expectations.
In the 3d RFIM the curvature-driven growth mechanism that leads to (3.22) is
impeded by the random field roughening of the domain walls. The dependence on the
parameters T and h has been estimated. In the early stages of growth, one expects the
zero-field result to hold with a reduction in the amplitude R(t) (A Bh2 ) t1/2 . The
time-window over which this law is observed numerically decreases with increasing
field strength. In the late time regime, where pinning is effective Villain deduced a
logarithmic growth R(t) (T /h2 ) ln t/t0 by estimating the maximum barrier height
encountered by the domain wall and using the Arrhenius law to derive the associated
time-scale.
In the case of spin-glasses, if the mean-field picture with a large number of equilibrium states is realized in finite dimensional models, the dynamics would be one in
which all these states grow in competition. If, instead, the phenomenological droplet
model applies, there would be two types of domains growing and R(t) (ln t)1/
with the exponent satisfying 0 d 1. Some refined arguments that we will
not discuss here indicate that the dimension of the bulk of these domains should be
compact but their surface should be rough with fractal dimension Ds > d 1.
3.5.7

Temperature-dependent effective exponents

The fact that numerical simulations of dirty systems tend to indicate that the
growing length is a power law with a T -dependent exponent can be explained as due
to the effect of a T -dependent cross-over length LT . Indeed, if below LT T the
growth process is as in the clean limit while above LT quenched disorder is felt and
the dynamics is thermally activated:
 1/z
t d
for
R(t) LT ,
R(t)
(3.31)
1/
(ln t)
for
R(t) LT .
These growth-laws can be first inverted to get the time needed to grow a given length
and then combined into a single expression that interpolates between the two regimes:

t(R) e(R/LT ) Rzd

(3.32)

where the relevant T -dependent length-scale LT has been introduced. 


R (t)
1
z(T )
one finds z(T ) zd + ln R(t)
that
Now, by simply setting t(R) R
L
T


)
t/z(T )
replacing R t1/z(T ) becomes z(T ) zd + z(T
. Using now z(T ) zd in
ln t
L
T

the correction term and focusing on times such that t/zd / ln t is almost constant and
(T )
/T )
equal to c one finds z(T )zd c zd /L
T . Similarly, by equating t(R) exp(R
one finds that (T ) is a decreasing function of T approaching at high T .
73

3.6

Time-dependent Ginzburg-Landau description

In order to treat phase-transitions and the coarsening process analytically it is

preferable to introduce a coarse-grained description in terms of a continuous coarsegrained field,
1 X
(~x, t)
si (t) ,
(3.33)
V
iV~
x

the fluctuating magnetization density. In a first approximation a Landau-Ginzburg

free-energy functional is introduced
Z
o
nc
[(~x, t)]2 + V [(~x, t)] .
(3.34)
F [] = dd x
2
With the choice of the potential one distinguishes between a second order and a
first order phase transition. In the former case, the typical form is the 4 form:
V () = a4 + b(g)2 .

(3.35)

The first term in eq. (3.34) represents the energy cost to create a domain wall or the
elasticity of an interface. The second term depends on a parameter, g, and changes
sign from positive at g > gc to negative at g < gc . Above the critical point determined
by b(gc ) = 0 it has a single minimum at = 0, at gc it is flat at = 0 and below
gc it has a double well structure with two minima, = [b(g)/(2a)]1/2 = hieq (g),
that correspond to the equilibrium states in the ordered phase. Equation (3.34) is
exact for a fully connected Ising model where V () arises from the multiplicity of
spin configurations that contribute to the same (~x) = m. The order-parameter
dependent free-energy density reads f (m) = Jm2 hm + kB T {(1 + m)/2 ln[(1 +
m)/2] + (1 m)/2 ln[(1 m)/2] that close to the critical point where m 0 becomes
f (m) (kB T 2J)/2 m2 hm + kB T /12 m4 demonstrating the passage from a
harmonic form at kB T > kB Tc = 2J, to a quartic well at T = Tc , and finally to a
double-well structure at T < Tc .
Exercise 13. Prove the above.
With a six-order potential one can mimic the situation in the right panel of Fig. 22.
When discussing dynamics one should write down the stochastic evolution of the
individual spins and compute time-dependent averaged quantities as the ones in (3.1).
This is the procedure used in numerical simulations. Analytically it is more convenient
to work with a field-theory and an evolution equation of Langevin-type. This is
the motivation for the introduction of continuous field equations that regulate the
time-evolution of the coarse-grained order parameter. Ideally these equations should
be derived from the spin stochastic dynamics but in practice they are introduced
phenomenologically. In the magnetic case as well as in many cases of interest, the

domain wall and interface dynamics can be argued to be overdamped (i.e. t tr ).

74

Two very similar approaches are used. Assuming T is only relevant to determine
the equilibrium coarse-grained field one uses the phenomenological zero-temperature
time-dependent Ginzburg-Landau equation or model A in the classification of
Hohenberg-Halperin deterministic equation
(~x, t)
F []
=
t
(~x, t)

(3.36)

(the friction coefficient has been absorbed in a redefinition of time). Initial conditions
are usually chosen to be random with short-range correlations
[ (~x, 0)(~x , 0) ]ic = (~x ~x )

(3.37)

thus mimicking the high-temperature configuration ([. . .]ic represent the average over
its probability distribution). The numeric solution to this equation with the quartic
potential and b < 0 shows that such a random initial condition evolves into a field
configuration with patches of ordered region in which the field takes one of the two
values [b/(2a)]1/2 separated by sharp walls. It ignores temperature fluctuations
within the domains meaning that the field is fully saturated within the domains
and, consequently, one has access to the aging part of the correlations only, see e.g.
eq. (3.15). The phase transition is controlled by the parameter b in the potential.
Another, similar approach, is to add a thermal noise to the former
F []
(~x, t)
=
+ (~x, t) .
t
(~x, t)

(3.38)

This is the field-theoretical extension of the Langevin equation in which the potential
is replaced by the order-parameter-dependent funcitonal free-energy in eq. (3.34) with
a potential form with fixed parameters (independent of T ). is a noise taken to be
Gaussian distributed with zero mean and correlations
h(~x, t)(~x , t )i = 2kB T d (~x ~x )(t t ) .

(3.39)

The friction coefficient has been absorbed in a redefinition of time. For a quartic
potential a dynamic phase transition arises at a critical Tc ; above Tc the system freely
moves above the two minima and basically ignores the double well structure while
below Tc this is important. Within the growing domains the field fluctuates about
its mean also given by [b/(2a)]1/2 and the fluctuations are determined by T . One
can describe the rapid relaxation at ties such that the domain walls do not move
with this approach. This formulation is better suited to treat critical and sub-critical
dynamics in the same field-theoretical framework.
These equations do not conserve the order parameter neither locally nor globally.
Extensions for cases in which it is conserved exist (model B). Cases with vectorial or
even tensorial order parameters can be treated similarly and are also of experimental
relevance, notably for vectorial magnets or liquid crystals.

75

3.6.1

Short-time dynamics

Take an initial configuration (~x, 0) = 0 on average with small fluctuations, as in

equilibrium at very high temperature, and quench the system. At very short time one
can expand the non-linear potential and the Ginzburg-Landau equation (3.36), for the
R
~
Fourier components, (~k, t) = Ld/2 dd x (~x, t)eik~x with ~k = 2/L (n1 , . . . , nd )
and nk integer, reads
(~k, t)
= [k 2 V (0)](~k, t) + (~k, t) .
t

(3.40)

If V (0) > 0 all modes decay exponentially and no order develops. If V (0) < 0
instead modes with k 2 V (0) > 0 are unstable and grow exponentially until a time
t 1/V (0) when the small expansion ceases to be justified. The instability of
the small wave-vector modes indicates that the system tends to order. To go beyond
this analysis one needs to consider the full non-linear equation.
3.6.2

R(t) in non-conserved order parameter curvature dynamics in d > 2

The time-dependent Ginzburg-Landau model allows us to gain some insight on the

mechanism driving the domain growth and the direct computation of the averaged
domain length. In clean systems temperature does not play a very important role in
the domain-growth process, it just adds some thermal fluctuations within the domains,
as long as it is smaller than Tc . In dirty cases instead temperature triggers thermal
activation.
We focus first on the clean cases at T = 0 and only later we discuss thermal effects.
Equation (3.36) for T = 0 is just a gradient descent in the energy landscape F . Two
terms contribute to F : the bulk-energy term that is minimized by = 0 and the
elastic energy ()2 which is minimized by flat walls if present. As a consequence
the minimization process implies that regions of constant field, (~x, t) = 0 , grow
and they separated by flatter and flatter walls.
Take a flat domain wall separating regions where the configuration is the one
of the two equilibrium states, (~x, t) = 0 + (~x, t). Linearizing eq. (3.36) around
0 and looking for static configurations, i.e. (~x, t) = (~x) = (n) where n
2
2
is the distance from the wall along the normal direction one
finds d (n)/dn =
V (0 )(n). This equation has the solution (n) e V (0 )n where n is the
perpendicular distance to the wall. The order parameter approaches 0 on both
sides of the wall very rapidly. This means that the free-energy of a configuration with
an interface (sum of the elastic and potential terms) is concentrated in a very narrow
region close to it. In consequence, the domain-wall curvature is the driving force for
domain growth.
Allen and Cahn showed that the local wall velocity is proportional to the local
curvature working with the Ginzburg-Landau equation at T = 0. The proof goes as
follows. Take the Ginzburg-Landau equation and trasform the derivatives to apply in

76

t1

t1
t2

t2

o< 0

o> 0 o< 0

o> 0

Figure 32: Left: domain wall profile. Right: view from the top. (g is n.)
the direction normal to the wall:



(~x, t) 
(~x, t)  n 
(~x, t)
~
,
(~x, t) =
n
,
=
t
n t t 
n t


2 (~x, t) 
(~x, t)  ~
2 (~x, t) =
n

+
n2 
n 
t

where the subscripts mean that the derivatives are taken at t or fixed. Using now
2 (~
x,t)

n2 |t = V () (note that the derivative is taken at fixed t) in the GL equation

one finds the Allen-Cahn result
~ n
v t n| =

(3.41)

valid in all d with the geodesic curvature.

Equation (3.41) allows one to get an intuition about the typical growth law in such
processes. Take a spherical wall in any dimension. The local curvature is constant
and = (d1)/R where R is the radius of the sphere within the hull. Equation (3.41)
is recast as dR/dt = (d 1)/R that implies R2 (t) = R2 (0) 2(d 1)t.
A closer look at the 2d equation allows one to go beyond and prove, in this case,
that all areas enclosed by domain walls irrespective of their being other structures
within (the so-called hull-enclosed areas) tend to diminish at constant rate dA/dt =
. This, of course, does not mean that all domains reduce their area since a domain
can gain area from the disappearance of an internal domain of the opposite sign,
for instance. The proof is simple and just uses the Gauss-Bonnet theorem: dA
dt =
H
H
~
~v d = vd. The local wall-velocity, ~v , is proportional
to the local geodesic
H
curvature, , and the Gauss-Bonnet theorem implies d = 2 for a planar 2d
manifold with no holes. Therefore, the hull-enclosed area decreases with constant
velocity for any geometry.
Therefore the local velocity points in the direction of the local centre of curvature.
The effect is to reduce the wall roughness by rendering them smoother.
77

3.6.3

Scaling functions for subcritical coarsening

Even though the qualitative behavior of the solution to eq. (3.36) is easy to grasp,
it is still too difficult to solve analytically and not much is known exactly on the
scaling functions. A number of approximations have been developed but none of
them is fully satisfactorily (see [4] for a critical review of this problem).
The super-universality hypothesis states that in cases in which temperature
and quenched disorder are irrelevant in the sense that they do not modify the nature
of the low-temperature phase (i.e. it remains ferromagnetic in the case of ferromagnetic Ising models) the scaling functions are not modified. Only the growing length
changes from the, say, curvature driven t1/2 law to a logarithmic one. This hypothesis has been verified in a number of two and three dimensional models including the
RBIM and the RFIM.
3.6.4

Breakdown of dynamic scaling

Some special cases in which dynamic scaling does not apply have also been exhibited. Their common feature is the existence of two (or more) growing lengths
associated to different ordering mechanisms. An example is given by the Heisenberg
model at T 0 in which the two mechanisms are related to the vectorial ordering
within domains separated by couples of parallel spins that annhilate in a way that is
similar to domain-wall annihilation in the Ising chain.
3.6.5

Annealing: crossover from critical to subcritical coarsening

There has been recent interest in understanding how a finite rate cooling affects
the defect density found right after the quench. A scaling form involving equilibrium
critical exponents was proposed by Zurek following work by Kibble. The interest is
triggered by the similarity with the problem of quantum quenches in atomic gases, for
instance. An interplay between critical coarsening (the dynamics that occurs close in
the critical region) that is usually ignored (!) and sub-critical coarsening (once the
critical region is left) is the mechanism determining the density of defects right after
the end of the cooling procedure.
The growing length satisfies a scaling law
R(t, (t))

f (x)

(t) f [tzeq (t)]


ct
x 1

x
x1

(t) = |T (t) Tc |
Equilibrium at high T
Coarsening at low T

with t measured from the instant when the critical point is crossed and x (1, 1)
is the critical region. A careful analysis of this problem can be found in Biroli, LFC,
Sicilia (2010).

3.7

The large N approximation

78

A very useful approximation is to upgrade the scalar field to a vectorial one with
N components
~ x, t) = (1 (~x, t), . . . , N (~x, t)) ,
(~x, t) (~
(3.42)

and modify the free-energy



Z
1 ~ 2 N 2
1 2 2
d
(3.43)
F = d x () + (0 N ) ,
2
4
PN
with 2 = =1 2 and 0 > 0. Note that the double-well structure of the potential
is assumed from the start.
The Langevin equation then becomes
t (~x, t) = 2 (~x, t) 4 (~x, t) [20 N 1 2 (~x, t)] + (~x, t)

(3.44)

and the initial condition is taken from a Gaussian distribution with zero-mean and
correlations
h (~x, 0) (~x, 0)i = d (~x ~x ) .
(3.45)
The dynamics is clearly isotropic in the N -dimensional space implying
C (~x, t; ~x , t ) = C(~x, t; ~x , t )

(3.46)

In the limit N while keeping the dimension of real space fixed to d, the cubic
term in the right-hand-side can be replaced by
(~x, t)N 1 2 (~x, t) (~x, t)N 1 [ 2 (~x, t) ]ic (~x, t) a
(t)

(3.47)

since N 1 2 (~x, t) does not fluctuate, it is equal to its average over the initial conditions and it is therefore not expected to depend on the spatial position if the initial
conditions are chosen from a distribution that is statistically translational invariant.
For the scalar field theory the replacement (3.47) is just the Hartree approximation. The dynamic equation is now linear in the field (~x, t) that we rename (~x, t)
(and it is now order 1):
t (~x, t) = [2 + a(t)](~x, t) + (~x, t) ,

(3.48)

where the time-dependent harmonic constant a(t) = 20 [2 (~x, t)]ic = 20 [2 (~0, t)]ic
has to be determined self-consistently. Equation (3.48) can be Fourier transformed
t (~k, t) = [k 2 + a(t)](~k, t) + (~k, t) ,

(3.49)

and it takes now the form of almost independent oscillators under different timedependent harmonic potentials coupled only through the self-consistent condition on
a(t). The stability properties of the oscillators depend on the sign of the prefactor in
the rhs. The solution is
Z d
Rt
Rt
k2 (tt )+ dt a(t )
k2 t+
dt a(t )
t
0
dt e
(~k, t )
(3.50)
(~k, t) = (~k, 0) e
+
0

79

and the equation on a(t) reads:

a(t)

20

Rt
0

dt a(t )

2
4t

d/2

+T

dt
0

2
4(t t )

d/2

Rt

dt a(t )

(3.51)
,

where one used [2 (~x, t)]ic = [2 (~0, t)]ic and a delta-correlated Gaussian distribution
of initial conditions with strength . The self-consistency equation is not singular
at t = 0 since there is an underlying cut-off in the integration over k corresponding
to the inverse of the lattice spacing, this implies that times should be translated as
t t + 1/2 with = 1/a the lattice spacing.
Without giving all the details of the calculation, eq. (3.51) can be solved at all
temperatures [42]. One finds that there exists a finite Tc (d) and
Upper-critical quench
2
a(t) eq
<0

(3.52)

with eq the equilibrium correlation length, and the mass (in field theoretical terms)
or the harmonic constant saturates to a finite value: k 2 + a(t) k 2 2 .
Critical quench
a(t)

w
with w = 0 for d > 4 and w = (d 4)/2 for d < 4 .
2t

(3.53)

The dynamics is trivial for d 4 but there is critical coarsening in d < 4. zeq equals
2 in agreement with the result from the expansion once evaluated at N . The
averaged field is
t
h(~k, t)i (~k, 0)ek t(4d)/4
(3.54)
Zero-temperature sub-critical coarsening
The zero-T equation admits a simple solution. In the long times limit in which the
system tends to decrease its elastic and potential energies [ 2 (~x, t) ]ic must converge
Rt
to 20 6= 0 below criticality and this imposes 2 0 dt a(t ) d2 ln(t/t0 ) with t0 =
/2 (/20 )2/d at large times, i.e.
a(t)

d
4t

for

t t0

(3.55)

and the time-dependent contribution to the spring constant vanishes asymptotically.

Knowing the long-time behavior of a(t) implies that each mode [(~k, t)]ic with
~k 6= 0 vanishes exponentially both in critical and sub-critical quenches but the ~k = 0
mode grows as td/4 . The growth of the ~k = 0 reflects the domain growth process
whereby all modulations tend to disappear and the configuration gets more and more
uniform as time passes.

80

We focus now on two interesting cases: quenches to Tc and 0 = T < Tc (in this way
the equilibrium relaxation is set to one in the correlation). The asymptotic behavior
of the space-time correlation function is




(~x ~x )2
,
exp
4(t + t )

(3.56)



(~x ~x )2
[ (~x, t)(~x , t ) ]ic = 20 t1d/2 f (t/t ) exp
,
4(t + t )

(3.57)

[ (~x, t)(~x , t ) ]ic = 20

4tt
(t + t )2

d/4

for t t for a quench to T < Tc and

for a quench to Tc . We focus on d < 4. Note that 0 is still present These expressions
capture the main features of the domain growth process:
In critical quenches the two-time dependent pre-factor is of the form expected
from dynamic scaling. The correlation decays to zero due to the prefactor that
goes as t(2d)/2 and vanishes in all d > 2. The aging curves have an envelope
that approaches zero as a power law. d = 2 is the lower critical dimension in
this problem.
In sub-critical quenches, for any finite and fixed (~x ~x ), in the long times limit
the exponential factor approaches one and one obtaines a function of t /t only.
Due to the exponential factor, for fixed but very
large time t and t the corre
lation falls off to zero over a distance |~x ~x | t + t . This means that, at
time t, the typical size of the equilibrated regions is R(t) t1/2 . This holds
for critical and sub-critical quenches as well and it is a peculiar property of the
large N O(N ) model that has zeq = zd .
For fixed |~x ~x |, the correlation always falls to zero over a time separation
t t which is larger than t . This means that the time it takes to the system to
decorrelate from its configuration at time t is of the order of t itself, td t . The
age of the system is the characteristic time-scale for the dynamical evolution:
the older is the system, the slower is its dynamics. After a time of the order of
the age of the system any point ~x will be swept by different domain walls and
the correlation will be lost.
In a sub-critical quench, for any finite and fixed (~x ~x ), in the long t and t
limit such that t /t 1 the time dependence disappears and the correlation
between two points converges to 20 . This means that, typically, if one looks at
a finite spatial region on a finite time-scale this region will be in one of the two
states 0 , i.e. within a domain.
For T > 0 in a sub-critical quench an addituve stationay contribution has to be
added to the aging one.
Note that we have obtained the field and then computed correlations from the
time-dependent configuration. We have not needed to compute the linear response.
We will see later that in other more complex glassy systems one cannot follow this
81

simple route and one needs to know how the linear response behave. We refer to the
reviews in [50] for detailed accounts on the behavior of the linear response in critical
dynamics.

3.8

The 2d xy model

The xy model in d = 2 is quite special since it is critical at all temperatures below

Tkt . It is then worth analyzing this special case in detail. Moreover, it has topological
defects in the form of vortices and the rate of approach to the equilibrium state is
affected by them.
The model is fully solvable in the spin-wave approximation in which the field is
supposed to vary smoothly in space and, hence, vortices are neglected. The twocomponent spin is parametrized as
~s(~x) = cos (~x) ex + sin (~x) ey

(3.58)

where the modulus has been fixed to one and the angle is measured between the
local spin and a chosen fixed axis. The functional Ginzburg-Landau free-energy is
proposed to be
Z
F =

d2 x ((~x))2 .

(3.59)

The angle correlation functions in equilibrium are

2

h((~r) (0)) i =

kB T
r
ln
J
a

(3.60)

leading to
C(r) = h~s(~x)~s(~y )i =

 a kB T /J

 a (T )

(3.61)
r
r
The equilibrium correlation length is eq (T ) = a/ ln(kB T /J) that tends to zero
only at T and diverges at T 0. Spin-waves are non-local and extensive
while vortices are local and intensive. The latter cannot be eliminated by simple
perturbations but they annihilate when a vortex and an anti-vortex encounter.
The Langevin equation acting on the angle is
t (~x) = 2 (~x, t) + (~x, t)

(3.62)

where a white-noise scalar noise is proposed to act on the angle . This can be readily
solved in Fourier space
Z t
2

2
dt ek (tt ) (~k, t ) .
(3.63)
(~k, t) = (~k, 0)ek t +
0

82

R
The global correlation and linear response, C(t, t ) = V 1 d2 r hvecs(~r, t) ~s(~r, t )i
R
s(~
r ,t)i 
take the following scaling forms in the limit
and R(t, t ) = V 1 d2 r h~

~
h(~
r ,t )
~
h=0

t t 2 :

C(t, t )

R(t, t )



1
Rc (t)

Rc (t )
(t t )(T )/2


1
Rc (t)

Rc (t )
4(T )(t t )1+(T )/2

(3.64)
(3.65)

with a scaling function and Rc (t) the growing correlation length (that should not be
confused with the linear response). The first remarkable property of these functions is
that they are both decomposed in the product of a function of the time-difference tt
and a function of the ratio Rc (t )/Rc (t), like in the general critical coarsening
case. When t t Rc (t ) and 1, the decay is stationary
C(t, t ) (t t )(T )/2 ,

R(t, t ) (t t )1(T )/2

and the fdr equals one. This limit defines a quasi-equilibrium regime. When the time
difference increases and becomes smaller than one the relaxation enters an aging
regime in which the decay of the correlation and response depends on the waiting-time
t . The behavior in the aging regime depends on the initial conditions as discussed
below.
Uniform initial conditions.
The uniform initial condition contains no free vortices and none are generated by
thermal fluctuations at any T < Tkt . The evolution is well captured by the simple
spin-wave approximation and after a simple calculation one finds

Rc (t)
Rc (t )

(1 + )
4

(T )/4

Rc (t) = t1/2 .

(3.66)

Beyond the crossover time tt t , when C(2t , t ) t(T )/2 and becomes smaller
than one, the correlation and response decay to zero as power laws of the waitingtime t . There is no clear-cut separation of time-scales characterised by the correlation
reaching a constant value independently of the waiting-times but only a t dependent
pseudo-plateau where the behavior of the two-time correlation changes. This is to
be confronted to the behavior of ferromagnetic coarsening systems quenched to the
low-temperature phase for which the crossover occurs at C(2t , t ) = m2eq . Above
this plateau, the relaxation corresponds to the equilibrium fluctuations of short wavelength while below the plateau the decorrelation is due to the domain-wall motion
that manifests into a scaling in terms of = t /t only. In the 2d xy case the order
parameter vanishes and there is no plateau at any finite value of C.
In the aging regime the fluctuation dissipation ratio is larger than one. This a
priori surprising result can be understood when interpreted in terms of the effective
83

 temperature. The completely order configuration is the equilibrium state at zero

temperature. The evolution of this initial state at finite temperature can be thought
of as representing a sudden inverse quench of the system from T = 0 to T > 0. If the
fdr is related to a remembrance of the temperature of the initial condition, in this
case this is lower than the working temperature T and thus, the effective temperature
also turns out to be lower than T .
Random initial conditions.
When random initial conditions with only short-ranged spatial correlations are
considered, free vortices and antivortices are present initially. The relaxation occurs
via the annihilation of vortex-antivortex pairs and this coarsening process is much
slower than the relaxation of spin-waves. The simple Gaussian theory is no more
suited to describe this dynamics and a full analytic treatment is too hard to implement. With scaling and numeric analysis the dynamic correlation length has been
estimated to be [4]
Rc (t) (t/ ln t)1/2 .
The numerical simulations of Berthier, Holdsworth and Sellitto have proven that the
two-time correlation and response are correctly described by the scaling form (3.64)
and (3.65) with this length scale and the full decay looks like the one shown in the
sketch above. The fdr is rather different from the one following the evolution of
a uniform initial condition. The non-equilibrium susceptibility is now smaller than
the equilibrium value, and in terms of the effective temperature this means that the
fluctuations of the wave-lengths longer than Rc (t) occur at a Teff > T and hence keep
a memory of the initial temperature T = .

3.9

Summary

In the table below we summarize the results describe above.

In short, critical and sub-critical coarsening occurs in models with conventional
second order phase transitions (or for systems with first order phase transitions when
one quenches well below the region of metastability). Close to the critical point
the dynamics is characterized by critical slowing down with the relaxation time
diverging as a power law of the distance to criticality. Growth of order is characterized
by a growing length that depends on time as a power law at criticality and with
a different power below the transition (in the absence of disorder). The dynamic
mechanisms are well understood but quantitative results are hard to obtain since the
equation to solve are highly non-linear and there is no small parameter to expand
around.
In structural glasses the slowing down is not of power law type so such a simple
coarsening description seems to be excluded for these systems.

84

Order param.

Growing length

Rc (t)

gc

g < gc

6= 0

t1/zeq

clean

1
t2
1
ln 2 tt
0

2d xy

disordered

t1/2
t1/3
R(t) 
 1
ln tt0

sc. NCOP
sc. COP
dis.

DFV < d

DFV = d

DFS < d 1

DFS = d 1

V RDF (t)
S RDF (t)

r
Rc (t)

C(r, t)

r2d f

C(t, t )

Rc2d (t t ) g

Rc (t )
Rc (t)

Cst (r) + Cag

Cst (t t ) + Cag

r
Rc (t)

Rc (t )
Rc (t)

Table 1: This table summarizes the behavior of growing structures and correlation
functions in critical and sub-critical quenches. V and S are the volume and surface
of the equilibrium growing structures (FK clusters and geometric domains in critical
and sub-critical quenches respectively). DFV and DFS are their fractal dimension.
Interesting information is also contained in the behavior of the linear response function
but we will discuss it later.

85

For spin-glasses this modeling has been pushed by Bray, Moore, Fisher and Huse.
It is not clear whether it is correct as no clearcut experimental evidence for the
coarsening type of scaling has been presented yet.

3.10

Nucleation and growth

In a first-order phase transition the equilibrium state of the system changes

abruptly. Right at the transition the free-energies of the two states involved are
identical and the transition is driven by lowering the free-energy as the new phase
forms, see Fig. 22. The original phase remains meta-stable close to the transition. The
nucleation of a sufficiently large bubble of the truly stable phase into the metastable
one needs to be thermally activated to trigger the growth process [3]. The rate of
the process can be very low or very fast depending on the height of the free-energy
barrier between the two states and the ambient temperature.
Two types of nucleation are usually distinguished: homogeneous (occurring at the
bulk of the meta-stable phase) and heterogeneous (driven by impurities or at the
surface). The more intuitive examples of the former, on which we focus here, are the
condensation of liquid droplets from vapor and the crystallization of a solid from the
melt.
The classical theory of nucleation applies to cases in which the identification
of the nucleus is easy. It is based on a number of assumptions that we now list. First,
one associates a number of particles to the nucleus (although in some interesting cases
this is not possible and a different approach is needed). Second, one assumes that
there is no memory for the evolution of the number densities of clusters of a given
size in time (concretely, a Markov master equation is used). Third, one assumes
that clusters grow or shrink by attachment or loss of a single particle, that is to say,
coalescence and fission of clusters are neglected. Thus, the length-scale over which the
slow part of the dynamics takes place is the one of the critical droplet size, the first one
to nucleate. Fourth, the transition rates satisfy detail balance and are independent of
the droplet form. They just depend on the free-energy of the droplet with two terms: a
contribution proportional to the droplet volume and the chemical potential difference
between the stable and the metastable states, f , and a contribution proportional
to the bubble surface that is equal to the surface area times the surface tension, ,
that is assumed to be the one of coexisting states in equilibrium - that is to say the
energy of a flat domain wall induced by twisted boundary conditions. Fift, the bubble
is taken to be spherical and thus dependent of a single parameter, the radius. Thus
F [R] = d1 Rd1 |f | d Rd

(3.67)

for d > 1. d is the volume of the unit sphere in d dimensions. For small radii
the surface term dominates and it is preferable to make the droplet disappear. In
contrast, for large radii the bulk term dominates and the growth of the bubble is
favored by a decreasing free-energy. Thus the free-energy difference has a maximum

86

at
R =

(d 1) d1
|f |1
d d |f |

(3.68)

and the system has to thermally surmount the barrier F F [R ]. The Kramers
escape theory, see Sect. 2.3, implies that the nucleation rate or the average number
of nucleations per unit of volume and time is suppressed by the Arrhenius factor
F
rA = t1
A e

with

F =

(d 1)d1 dd1
d
d1
d
d
|f |d1
d

(3.69)

As expected, F increases with increasing and/or |f |1 and r1 vanishes for

T 0 when thermal agitation is switched off. The implicit assumption is that the
time to create randomly the critical droplet is much longer than the time involved in
the subsequent growth. The relaxation of the entire system is thus expected to be
given by the inverse probability of escape from the metastable well. The determination
of the pre-factor [that is ignored in eq. (3.69)] is a hard task.

3.11

Elastic manifold dynamics

An interface is a frontier separating two regions of space with two phases. It could
be the border between water and oil in a liquid mixture, the border between regions
with positive and negative magnetization in a magnet, the limit of a fluid invading
porous media, etc. The static and dynamic properties of interfaces have many points
in common with the ones of (sometimes directed) manifolds with d internal dimensions
embedded in N + d dimensional spaces with N the dimension of the transverse space.
In this way, one includes cases such as directed lines (d = 1) that mimic vortex lines
in N + d = 3 dimensional high-Tc superconductors, polymers in (N + d = 2 or 3dimensional) random media, etc.
A slightly different situation is the one of growth phenomena, as for instance, the
burning front in a forrest, the advance of a crack in a rock, fluid invasion in porous
media, the growth of a surface on a substrate due to material deposition combined
(or not) with transverse diffusion of the material that reaches the surface, or even
the growth of bacterial colony. The surface is usually defined by a height field h(~x, t)
defined with respect to an origin level zero on the d-dimensional substrate. We will
assume that h is a single-valued function of ~x. N = 1 in the parametrization described
above.
As a physicist one would like to characterize the static and dynamic properties of
these interfaces and surfaces. The analysis os the static properties of domain walls and
interfaces corresponds, typically, to determining their equilibrium conformations (geometric properties, numbers, degeneracies, etc.) The study of the dynamic properties

87

u(x)
defect
(random bond)

random field

Figure 33: Left: an experimental view of a domain wall. Right: a sketch of a domain
wall in a 2d Ising magnet.
of domain walls and interfaces includes the analysis of their relaxation to equilibrium,
response to external driving forces, creep motion and depinning transition.
Domain growth and interface growth in the presence of quenched disorder is sometimes considered to be a baby spin-glass problem due to the presence of frustration
given by the competition between the elastic energy that tends to reduce the deformations and quenched disorder that tends to distort the structure.
3.11.1

Scale invariance

In general, the morphology of an interface depends on the length scale of observation: the Alps look rough on Earth but they look smooth seen from the Moon. However, a number of surfaces called self-similar do not depend on the scale of observation;
they are characterized by the absence of a characteristic scale. Such scale-invariance
is ubiquitous in nature, with the classical example of critical phenomena, and it is
characterized by the existence of power laws that characterize many quantities over
many orders of magnitude.
3.11.2

Solid-on-solid models of surface growth

These models are discrete and microscopic; they represent a truly experimental
situation, such as atom deposition as in film growth by molecular beam epitaxy, and
they are also advantageous to do numerical simulations.
A substrate d dimensional surface of size Ld is divided into cells that can be
occupied by columns of falling particles. Particles fall on this substrate and stick to it
according to different rules that define different models. The height of the surface is a
discrete variable hi where i labels the cell on the substrate. In restricted SOS models

88

a)

*
*

*
*
*

*
l1

b)

*
*
*
*

c)

l2

*
*

Figure 34: Schematic evolution of a pinned object on well separated time scales.
On scale t(1 ), the object reconforms by flipping a small portion of size 1 from one
favourable configuration to another (a b). On a much longer time scale t(2 )
t(1 ), the conformation on scale 2 (dotted lines) has evolved (b c). The dynamics
of the short wavelengths happens on a time scale such that long wavelengths are
effectively frozen.
the height is also constrained to satisfy hi hj | 1 with i and j nearest-neighboring
cells on the substrate.
In ballistic deposition particles are released from a randomly chosen position above
the surface, they follow a vertical trajectory, and they simply stick to the nearest or
next-nearest neighbor encountered, see Figs. 35 and 36.
In random deposition the particle falls vertically until reaching the top of its column.
In random deposition with surface relaxation, after reaching the top of its column
the particle further diffuses along the surface until it finds the position with the lowest
height.
3.11.3

Continuous models

Continuous models often describe the surfaces at larger length-scales. A coarsening

process is employed in such a way that the surface can be described by a continuous
function.
The Edwards-Wilkinson model
The simplest model for the growth and rearrangement of a surface is due to Edwards and Wilkinson who showed that the continuum limit of the process of sedimentation of granular particles under gravity on an initial flat substrate and their further

89

Figure 35: Balistic deposition

diffusion on the surface leads to
h(~x, t)
= 2 h(~x, t) + (~x, t) ,
t

(3.1)

where ~x is a d-dimensional spatial vector denoting position on the substrate, t is time

and h is a scalar function taking real values and measuring the height with respect
to its average value. The last term is a thermal noise, typically chosen to have a
Gaussian probability distribution a zero average and correlations
h (~x, t)(~x , t ) i = 2T (~x ~x )(t t ) .

(3.2)

Equation (3.3) is a stochastic field equation. The first term on the rhs of eqn (3.3)
penalizes rapid variations of the surface in space and is then a measure of the surface
tension. The noise term describes the randomness in the deposition process. Note
that the Edwards-Wilkinson equation can in fact be seen as describing the equilibrium
fluctuations of an interface (for example liquid vapor), with the noise term describing
the thermal fluctuations. The stochastic equation (3.3) is linear and can be easily
solved using a Fourier transform of the space coordinate ~x.
The Kardar-Parisi-Zhang model
More generally, one can expect that the deterministic force acting on the surface
~ and higher power of
may depend on other kinds of local terms, such as H itself, h,
2
this and h. Fortunately, the use of symmetries and conservation laws allows one
to reduce the number of allowed terms to just a few ones.
For example, one expects the dynamics to be translational invariant, meaning that
a symmetry breaking term such as ch, with c a constant, or any other term depending
directly on h, is not allowed. The next term one can consider is one proportional to the
local slope. Now, by symmetry, one expects that positive and negative slopes should
lead to the same result. Thus, this term should appear squared: [h(~x, t)]2 . Indeed,
90

Figure 36: Snow fall

the effect of this term has a simple geometric interpretation sketched in Fig. 37. One
can then assume that higher order derivatives n h are less relevant that the first
two ones, n = 1, 2, and neglect them all (one can actually prove that they do not
change the scaling properties at long times and large length scales). This leads to the
well-studied Kardar-Parisi-Zhang equation
h(~x, t)
~ 2 + (~x, t) ,
= 2 h(~x, t) + [h]
t

(3.3)

that seems to describe most of the random growth problems encountered in nature.
The new term is proportional to the slope of the surface and pushes it to grow in the
directions of large slope (see Fig. 37).

Figure 37: Sketch of the KPZ growth

91

3.11.4

Dynamic scaling at work

For concreteness let us use the continuous notation. The following definitions
carry through to the discrete case with minor modifications.
A simple quantity that describes whether the surface is actually growing is the
averaged height
Z L
1
dd x h(~x, t) .
(3.4)
h h(~x, t) i = d
L 0

Here and in what follows the angular brackets represent a spatial average over the
substrate coordinates or else an average over thermal noise.
The kinetics of the surface is characterized by its width, defined as the meansquared displacement of the total height
W (L, t) h (h(~x, t) h h(~x, t) i)2 i1/2 .

(3.5)

The surface is said to be rough when W (L, t) diverges with the system size L. In
rough surfaces the height is uncorrelated over long distances while in smooth ones the
height is correlated.
In the rough case, asymptotically W (L, t) has the Family-Vicsek scaling limit given
by:
W (L, t) = L F(tLz ) ,
(3.6)

where the roughness exponent , the dynamical exponent z and the scaling function
F are defined by the above equation in the large L limit. The average denotes
either
P
an average over thermal histories or an average over space, h i = A1

with
the
L
~
x
area of the substracte of linear dimension L indicated by AL . The scaling function F
is such that
1/2
,
for short times
t
(3.7)
W (L, t)
t/z ,
for intermediate times

L ,
for very long times

The dynamic exponent describes the evolution of correlated regions with time; initially, different sites on the substrate are not correlated but correlation develop over
time with correlated regions growing as R(t) t1/z . In each correlated region the
width of the surface grows as the observation scale raised to the roughness exponent
. R(t) is time-dependent correlation length.
3.11.5

Scale invariance

The Family-Vicsek scaling is a manifestation of the scale invariance of the growth

process. Indeed, if one assumes that in the very long time limit t under the
rescaling of space
~x b~x ,
(3.8)
92

the field changes in such a way that

h(~x, t ) b h(b~x, t )

(3.9)

one has
W 2 (L, t )

Ld

b2 W 2 (bL, t ) .

dd xh2 (~x, t ) Ld

dd xb2 h2 (b~x, t )
(3.10)

Now, taking b = L1 in such a way that the last factor is just a numerical constant,
one has
W 2 (L, t ) L2
(3.11)
and the roughness exponent is related to the scale invariant properties of the stationary surface, see eq. (3.9). Note that one transforms differently the longitudinal (~x)
and transverse (h) directions.
Moreover, the Family-Vicsek scaling means that the statistical properties of the
roughness are invariant under the changes
~x b~x ,

t bz t ,

h b h .

(3.12)

b is a dilation parameter.
3.11.6

Universality classes

It has been proposed to use the exponents and to classify growing surfaces
in universality classes. Although the microscopic processes leading to the interface
morphology can be very different, one finds that indeed they group in classes, determined by the symmetry properties of the continuous stochastic equations defining
different growth processes.. This is so since the asymptotic (long times, large scales)
morphologies depend only on mesoscopic aspects such as the diffusion process, the
random character of the process, the presence of non-linear terms, etc. Let us give a
few examples below.
A simple random deposition process, h = has = 1/2 and and z are not
defined.
An Edwards-Wilkinson surface has
= (2 d)/2 ,

= (2 d)/4 ,

z = / = 2 ,

d dim EW

(3.13)

(z = 2 shows the the diffusive character of the process in all dimensions) The EW
equation is invariant under
~ ,
~x ~x + x
h h + h ,
h h ,

t t + t ,
~x ~x ,
93

(3.14)

Figure 38: Family-Vicsek scaling.

The d-dimensional KPZ surface has
= 1/2 ,
= 0.38 ,

= 1/3 ,
= 0.24 ,

z = 3/2 ,
z = 1.58 ,

d=1,
d=2,

(3.15)

note that z = /. The KPZ exponents in generic d are known numerically. Note
that the KPZ suface does not evolve through normal diffusion z < 2). Moreover, the
height reversal invariance, h h, is not preserved by the KPZ equation, reflecting
the fact that growing surfaces in which atoms fall from above should not have this
symmetry.
At nonzero temperature an equilibrium domain wall in a pure system as imposed,
for example, by twisting boundary conditions is rough for d 3 because of thermal
fluctuations. In the presence of weak-disorder, RG arguments suggest that disorder
becomes relevant in d > 5/3. In a random ferromagnet in d = 2 one has = 4/3
(through a connection with the continuum model and then the KPZ-equation, see
below), while in d > 2 = 0.416(4 d).
Recently, there has been growing interest in characterizing the complete dynamic
probability distribution PL (W 2 , t) since it has been suggested that it might be used to
define universality classes for non-equilibrium steady states via the scaling function:


W2
2
2
.
hW iL PL (W , t) =
hW 2 iL
Note the similarity between this problem and the one of the fluctuations of the global
magnetization in the critical 2d XY model that we discussed in Sect. . Indeed, the

94

two problems are intimately connected since the dynamics of the 2d XY model in the
spin-wave approximation is given by eq. (3.3).
The above definition leads to what can be called a global characterization of
the roughness of the interface, averaged over the whole surface. Another, more local,
characterization of the roughness of the surface, is defined as
2

W 2 (~x, ~x , t, t ) h (h(~x, t) h(~x , t ) ) i = |~x ~x | F (|t t ||~x ~x |z ) ,

(3.16)

where we are now looking at the limit 1 |~x ~x | L. Often, the exponents and
the scaling function defined globally over the whole surface [as in eq. (3.6)] coincide
with their local counterpart, but this is not always the case. Note that in (3.16) one
measures at two different times t and t .
3.11.7

Roughening transition

It is a transition between a smooth and a rough phase. It is usually accompanied

by a diverging correlation length in the d-dimensional substrate space. In the smooth
phase the surface is typically rough over this (finite) length-scale while it is smooth
at longer length-scales.
Transitions between phases with different roughness properties are also possible, especially when adding quenched disordered potential. One observes a hightemperature phase where disorder is irrelevant and the roughness properties are those
of the clean limit e.g. the EW thermal exponent appears and a low-temperature
phase where the surface is rougher and its roughness is characterized by a disordercontrolled exponent that takes a large value than the thermal one.

95

Disordered systems: statics

No material is perfectly homogeneous: impurities of different kinds are distributed

randomly throughout the samples.
A natural effect of disorder should be to lower the critical temperature. Much
attention has been payed to the effect of weak disorder on phase transitions, that
is to say, situations in which the nature of the ordered and disordered phases is not
modified by the impurities but the critical phenomenon is. On the one hand, the
critical exponents of second order phase transitions might be modified by disorder,
on the other hand, disorder may smooth out the discontinuities of first order phase
transitions rendering them of second order. Strong disorder instead changes the
nature of the low-temperature phase and before discussing the critical phenomenon
one needs to understand how to characterize the new ordered glassy phase.
In this Section we will discuss several types of quenched disorder and models that
account for it. We will also overview some of the theoretical methods used to deal with
the static properties of models with quenched disorder, namely, scaling arguments and
the droplet theory, mean-field equations, and the replica method.

4.1

Quenched and annealed disorder

First, one has to distinguish between quenched and annealed disorder. Imagine
that one mixes some random impurities in a melt and then very slowly cools it down
in such a way that the impurities and the host remain in thermal equilibrium. If
one wants to study the statistical properties of the full system one has to compute
the full partition function, summing over all configurations of original components
and impurities. This is called annealed disorder. In the opposite case in which upon
cooling the host and impurities do not equilibrate but the impurities remain blocked in
random fixed positions, one talks about quenched disorder. Basically, the relaxation
time associated with the diffusion of the impurities in the sample is so long that these
remain trapped:
o 1012 1015 sec tobs 104 sec tdiff ,

(4.1)

where o is the microscopic time associated to the typical scale needed to reverse a
spin.
The former case is easier to treat analytically but is less physically relevant. The
latter is the one that leads to new phenomena and ideas that we will discuss next.

4.2

Bond disorder: the case of spin-glasses

Spin-glasses are alloys in which magnetic impurities substitute the original atoms
in positions randomly selected during the chemical preparation of the sample. The
96

interactions between the impurities are of RKKY type:

Vrkky = J

cos(2kF rij )
si sj
3
rij

(4.2)

with rij = |~ri ~rj | the distance between them and si a spin variable that represents
their magnetic moment. Clearly, the location of the impurities varies from sample to
sample. The time-scale for diffusion of the magnetic impurities is much longer than
the time-scale for spin flips. Thus, for all practical purposes the positions ~ri can be
associated to quenched random variables distributed according to a uniform probability distribution that in turn implies a probability distribution of the exchanges. This
is quenched disorder.
4.2.1

Lack of homogeneity

It is clear that the presence of quench disorder, in the form of random interactions, fields, dilution, etc. breaks spatial homogeneity and renders single samples
heterogenous. Homogeneity is recovered though, if one performs an average over
all possible realizations of disorder, each weighted with its own probability.
4.2.2

Frustration

Depending on the value of the distance rij the numerator in eq. (4.2) can be positive or negative implying that both ferromagnetic and antiferromagnetic interactions
exist. This leads to frustration, which means that some two-body interactions cannot be satisfied by any spin configuration. An example with four sites and four links
is shown in Fig. 39-left, where we took three positive exchanges and one negative one
all, for simplicity, with the same absolute value, J. Four configurations minimize the
energy, Ef = 2J, but none of them satisfies the lower link. One can easily check
that any closed loop such that the product of the interactions takes a negative sign
is frustrated. Frustration naturally leads to a higher energy and a larger degeneracy of the number of ground states. This is again easy to grasp by comparing the
number of ground states of the frustrated plaquette in Fig. 39-left to its unfrustrated
counterpart shown on the central panel. Indeed, the energy and degeneracy of the
ground state of the unfrustrated plaquette are Eu = 4J and nu = 2, respectively.
Frustration may also be due to pure geometrical constraints. The canonical example is an anti-ferromagnet on a triangular lattice in which each plaquette is frustrated,
see Fig. 39-right. This is generically called geometric frustration.
In short, frustration arises when the geometry of the lattice and/or the nature of
the interactions make impossible to simultaneously minimize the energy of all pair
couplings between the spins. Any loop of connected spins is said to be frustrated
if the product of the signs of connecting bonds is negative. In general, energy and
entropy of the ground states increase due to frustration.

97

+
+

+
+

Figure 39: A frustrated (left) and an unfrustrated (center) square plaquette. A

frustrated triangular plaquette (right).
4.2.3

Gauge invariance

The gauge transformation

si = i si ,

Jij
= i Jij j ,

with

i = 1

(4.3)

leaves the energy and the partition function of an Ising spin model with two-body
interactions invariant:
EJ [{s}] = EJ [{s }]

ZJ = ZJ .

(4.4)

This invariance means that all thermodynamic quantities are independent of the particular choice of the quenched disordered interactions.
Whenever it exists a set of i s such that frustration can be eliminated from all
loops in the model, the effects of disorder are less strong than in trully frustrated
cases, see the example of the Mattis model in Sect. .
4.2.4

Self-averageness

If each sample is characterized by its own realization of the exchanges, should one
expect a totally different behavior from sample to sample? Fortunately, many generic
static and dynamic properties of spin-glasses (and other systems with quenched disorder) do not depend on the specific realization of the random couplings and are
self-averaging. This means that the typical value is equal to the average over the
disorder:
Atyp
(4.5)
J = [ AJ ]
in the thermodynamic limit. More precisely, in self-averaging quantities sample-tosample fluctuations with respect to the mean value are expected to be O(N a ) with
a > 0. Roughly, observables that involve summing over the entire volume of the
system are expected to be self-averaging. In particular, the free-energy density of
models with short-ranged interactions is expected to be self-averaging in this limit.
An example: the disordered Ising chain
The meaning of this property can be grasped from
P the solution of the random
bond Ising chain defined by the energy function E = i Ji si si+1 with spin variables
98

si = , for i = 1, . . . , N and random bonds Ji independently taken from a probability

distribution P (Ji ). For simplicity, we consider periodic boundary conditions. The
disorder-dependent partition function reads
X P J s s
i i i+1
i
ZJ =
(4.6)
e
{si =1}

and this can be readily computed introducing the change of variables i si si+1 .
One finds.
Y
X
ZJ =
2 cosh(Ji )

FJ =
ln cosh(Ji ) + N ln 2 .
(4.7)
i

The partition function is a product of i.i.d. random variables and it is itself a random
variable with a log-normal distribution. The free-energy density instead is a sum of
i.i.d. random variables and, using the central limit theorem, in the large N limit
becomes a Gaussian random variable narrowly peaked at its maximum. The typical
value, given by the maximum of the Gaussian distribution, coincides with the average,
limN fJtyp [ fJ ] = 0.
General argument
A simple argument justifies the self-averageness of the free-energy density in
generic finite dimensional systems with short-range interactions. Let us divide a,
say, cubic system of volume V = Ld in n subsystems, say also cubes, of volume
v = d with V = nv. If the interactions are short-ranged, the total free-energy is the
sum of two terms, a contribution from the bulk of the subsystems
Pand a contribution
from the interfaces between the subsystems: FJ = ln ZJ = ln conf eEJ (conf ) =
P
P
P
ln conf eEJ (bulk)EJ (surf ) ln bulk eEJ (bulk) +ln surf eEJ (surf ) = FJbulk

FJsurf (we neglected the contributions from the interaction between surface and
bulk). If the interaction extends over a short distance and the linear size of the boxes
is P
, the surface energy is negligible with respect to the bulk one and FJ
ln bulk eEJ (bulk) . In the thermodynamic limit, the disorder dependent free-energy
is then a sum of n = (L/)d random numbers, each one
the disorder dependent
Pnbeing P
free-energy of the bulk of each subsystem: FJ k=1 ln bulkk eEJ (bulkk ) . In
the limit of a very large number of subsystems (L or n 1) the central limit
theorem implies that the total free-energy is Gaussian distributed with the maximum reached at a value FJtyp that coincides with the average over all realizations of
the randomness [ FJ ]. Morever,the dispersion about the typical value vanishes in
the large n limit, FJ /[ FJ ] n/n = n1/2 0 in the large n limit. Similarly,
fJ /[fJ ] O(n1/2 ) where fJ = FJ /N is the intensive free-energy. In a sufficiently
large system the typical FJ is then very close to the averaged [ FJ ] and one can
compute the latter to understand the static properties of typical systems.
Lack of self-averageness in the correlation functions

99

Once one has [FJ ], one derives all disordered average thermal averages by taking
derivatives of the disordered averaged free-energy with respect to sources introduced
in the partition function. For example,

[ FJ ] 
[ h si i ] =
,
(4.8)
hi hi =0

[ FJ ] 
,
(4.9)
[ h si sj i h si ih sj i ] = T
hi hj 
hi =0

with E E i hi si . Connected correlation functions, though, are not self-averaging

quantities. This can be seen, again, studying the random bond Ising chain:
h si sj iJ h si iJ h sj iJ = ZJ1

...
ZJ = tanh(Ji ) . . . tanh(Jj ) , (4.10)
Ji
Jj

where we used h si i = 0 (valid for a distribution of random bonds with zero mean)
and the dots indicate all sites on the chain between the ending points i and j. The
last expression is a product of random variables and it is not equal to its average (4.9)
not even in the large separation limit |~ri ~rj | .
Quenched vs. annealed averages
Take a case in which the partition function equals
 N
e
p = 1/N
Z=
e2N p = (1 1/N )

(4.11)

The annealed free-energy density, fa is

fa = (N )1 ln[ Z ] = (N )1 ln N 1 eN + (1 N 1 )e2N
= lim (N )1 ln N 1 eN = 1
N 1

The quenched free-energy density, fq , is

fq = (N )1 [ ln Z ] = (N )1 N 1 (N ) + (1 N 1 )(2N )
= lim (N )1 (2N ) = 2
N 1

As the logarithm is a concave function, [ ln Z ] < ln[ Z ] and then fa < fq .

4.3
4.3.1

Models with quenched disorder

Spin-glass models

In the early 70s Edwards and Anderson proposed a rather simple model that
should capture the main features of spin-glasses. The interactions (4.2) decay with a
100

cubic power of the distance and hence they are relatively short-ranged. This suggests
to put the spins on a regular cubic lattice model and to trade the randomness in the
positions into random nearest neighbor exchanges taken from a Gaussian probability
distribution:
Eea =

Jij si sj

J2
ij

P (Jij ) = (2 2 ) 2 e 22 .

with

(4.12)

hiji

The precise form of the probability distribution of the exchanges is suppose not to
be important, though some authors claim that there might be non-universality with
respect to it.
A natural extension of the EA model in which all spins interact has been proposed
by Sherrington and Kirkpatrick
X
X
E=
Jij si sj
hi s i
(4.13)
i

i6=j

and it is called the SK model. The interaction strengths Jij are taken from a
Gaussian pdf and they scale with N in such a way that the thermodynamic is nontrivial:
2
1

2 2
P (Jij ) = (2N
)
e

J
ij
2 2
N

2
N
= 2 N .

(4.14)

2
The first two-moments of the exchange distribution are [Jij ] = 0 and [Jij
] = J 2 /(2N ).
This is a case for which a mean-field theory is expected to be exact.
A further extension of the EA model is called the p spin model
X
X
E=
Ji1 ...ip si1 . . . sip
hi s i
(4.15)
i1 <...<ip

with p 3. The sum can also be written as i1 <i2 <...<ip = 1/p!

exchanges are now taken from a Gaussian probability distribution
2
P (Jij ) = (2N
)

21

J2
ij
2 2
N

2
N
= J 2 p!/(2N p1 ) .

i1 6=i2 6=ip .

The

(4.16)

J p!
with [Ji1 ...ip ] = 0 and [Ji21 ...ip ] = 2N
p1 . Indeed, an extensive free-energy is achieved
(p1)/2
. This scaling can be justified as follows. The
by scaling Ji1 ...ip with N P
local field hi = 1/(p 1)! ii2 6=ip Jii2 ...ip mi2 . . . mip should be of order one. At low
temperatures the mi s take plus and minus signs. In particular, we estimate the order
of magnitude of this term by working at T = 0 and taking mi = 1 with probability
1
2 . In order to keep the discussion simple, let us take p = 2. In this case, if the
strengths Jij , are of order one, hi is a sum of N i.i.d. random variables, with zero
mean and unit variance11 , and hi has zero mean and variance equal to N . Therefore,
P
P
11 The calculation goes as follow: h F i =
J J h mj mk i =
Jij h mj i = 0 and h Fi2 i =
i
jk ij ik
j
P 2
j

Jij

101


one can
argue that hi is of order N . To make it finite we then chose Jij to be of
2
order 1/ N or, in other words, we impose [ Jij
] = J 2 /(2N ). The generalization to
p 2 is straightforward.
Cases that find an application in computer science are defined on random graphs
with fixed or fluctuating finite connectivity. In the latter case one places the spins on
the vertices of a graph with links between couples or groups of p spins chosen with a
probability c. These are called dilute spin-glasses.
Exercise 14. Study the statics of the fully connected p-spin ferromagnet in which
all coupling exchanges are equal to J. Distinguish the cases p = 2 from p > 2. What
are the order of the phase transitions?
4.3.2

Random ferromagnets

Let us now discuss some, a priori simpler cases. An example is the Mattis random
magnet in which the interaction strengths are given by
j = with p = 1/2 .

with

Ji1 ...ip = i1 . . . ip

(4.17)

In this case a simple gauge transformation, i i si , allows one to transform the

disordered model in a ferromagnet, showing that there was no true frustration in the
system.
Random bond ferromagnets (RBFMs) are systems in which the strengths of
the interactions are not all identical but their sign is always positive. One can imagine
such a exchange as the sum of two terms:
Jij = J + Jij ,

with

Jij

small and random .

There is no frustration in these systems either.

Models with site or link dilution are also interesting:
P
P
Esite dil = J hiji si sj i j ,
Elink dil = J hiji si sj ij ,

(4.18)

(4.19)

with P (i = 0, 1) = p, 1 p in the first case and P (ij = 0, 1) = p, 1 p in the second.

Link randomness is not the only type of disorder encountered experimentally.
Random fields, that couple linearly to the magnetic moments, are also quite common;
the classical model is the ferromagnetic random field Ising model (RFIM):
Erfim = J

X
hiji

si sj

s i hi

with

h2
i

P (hi ) = (2 2 ) 2 e 22 .

(4.20)

The dilute antiferromagnet in a uniform magnetic field is believed to behave similarly to the ferromagnetic random field Ising model. Experimental realizations of
the former are common and measurements have been performed in samples like
Rb2 Co0.7 Mg0.3 F4 .
102

Note that the up-down Ising symmetry is preserved in models in which the impurities (the Jij s) couple to the local energy (and there is no applied external field)
while it is not in models in which they couple to the local order parameter (as the
RFIM).
The random fields give rise to many metastable states that modify the equilibrium
and non-equilibrium behavior of the RFIM. In one dimension the RFIM does not
order at all, in d = 2 there is strong evidence that the model is disordered even
at zero temperature, in d = 3 it there is a finite temperature transition towards a
ferromagnetic state. Whether there is a glassy phase near zero temperture and close
to the critical point is still and open problem.
The RFIM at zero temperature has been proposed to yield a generic description of
material cracking through a series of avalaches. In this problem one cracking domain
triggers others, of which size, depends on the quenched disorder in the samples. In a
random magnetic system this phenomenon corresponds to the variation of the magnetization in discrete steps as the external field is adiabatically increased (the time
scale for an avalanche to take place is much shorter than the time-scale to modify the
field) and it is accessed using Barkhausen noise experiments. Disorder is responsible
for the jerky motion of the domain walls. The distribution of sizes and duration of
the avalanches is found to decay with a power law tail cut-off at a given size. The
value of the cut-off size depends on the strength of the random field and it moves to
infinity at the critical point.
4.3.3

Random manifolds

Once again, disorder is not only present in magnetic systems. An example that has
received much attention is the so-called random manifold. This is a d dimensional
directed elastic manifold moving in an embedding N + d dimensional space under
the effect of a quenched random potential. The simplest case with d = 0 corresponds
to a particle moving in an embedding space with N dimensions. If, for instance
N = 1, the particle moves on a line, if N = 2 it moves on a plane and so on and
so forth. If d = 1 one has a line that can represent a domain wall, a polymer, a
vortex line, etc. The fact that the line is directed means it has a preferred direction,
in particular, it does not have overhangs. If the line moves in a plane, the embedding
space has (N = 1) + (d = 1) dimensions. One usually describes the system with an
~ that locates in the transverse space each point on the
N -dimensional coordinate, ,
manifold, represented by the internal
coordinate ~x,
p
R d-dimensional
The elastic energy is Eelas = dd x 1 + ((~x))2 with the deformation cost
of a unit surface. Assuming the deformaiton isRsmall one can linearize this expression
and get, upto an additive constant, Eelas = 2 dd x ((~x))2 .
Disorder is introduced in the form of a random potential energy at each point
~ x), ~x)
in the N + d dimensional space. The manifold feels, then a potential V ((~
characterized by its pdf. If the random potential is the result of a large number of
impurities, the central limit theorem implies that its probability density is Gaussian.
103

Just by shifting the energy scale one can set its average to zero, [ V ] = 0. As for its
correlations, one typically assumes, for simplicity, that they exist in the transverse
direction only:
~ x), ~x)V (
~ (~x ), ~x ) ] = d (~x ~x )V(,
~
~) .
[ V ((~

(4.21)

If one further assumes that there is a statistical isotropy and translational invariance
~
~ ) = W/2 V(|
~
~ |/) with a correlation length and
of the correlations, V(,
d2 1/2
(W )
the strength of the disorder. The disorder can now be of two types:
short-ranged if V falls to zero at infinity sufficiently rapidly and long-range if it either
grows with distance or has a slow decay to zero. An example involving both cases is
given by the power law V(z) = ( + z) where is a short distance cut-off and
controls the range of the correlations with > 1 being short-ranged and < 1 being
long-ranged.
The random manifold model is then
Z
h
i
~ x), ~x) .
H = dd x
(4.22)
((~x))2 + V ((~
2
This model also describes directed domain walls in random systems. One can
derive it in the long length-scales limit by taking the continuum limit of the pure Ising
part (that leads to the elastic term) and the random part (that leads to the second
disordered potential). In the pure Ising model the second term is a constant that
can be set to zero while the first one implies that the ground state is a perfectly flat
wall, as expected. In cases with quenched disorder, the long-ranged and short-ranged
random potentials mimic cases in which the interfaces are attracted by pinning centers
(random field type) or the phases are attracted by disorder (random bond type),
respectively. For instance, random bond disorder is typically described by a Gaussian
~ x), ~x), V (
~ (~x ), ~x )] = W d2 d (~x
pdf with zero mean and delta-correlated [V ((~

~
~
~x )( ).

4.4

The spin-glass transition

Let us now discuss a problem in which disorder is so strong as to modify the nature
of the low temperature phase. If this is so, one needs to define a new order parameter,
capable of identifying order in this phase.
4.4.1

The simplest order parameter

The spin-glass equilibrium phase is one in which spins freeze in randomly-looking

configurations. In finite dimensions these configurations are spatially irregular. A
snapshot looks statistical identical to a high temperature paramagnetic configuration
in which spins point in both directions. However, while at high temperatures the spins
flip rapidly and another snapshot taken immediately after would look completely

104

different from the previous one, at low temperatures two snapshots taken at close
times are highly correlated.
In a spin-glass state the local magnetization is expected to take a non-zero value,
mi = h s i i =
6 0, where the average is interpreted in the restricted sense introduced in
the discussion of ferromagnets, that we will call here within a pure state (the notion
of a pure state will be made more precise below). Instead, the total magnetization
PN
density, m = N 1 i=1 mi , vanishes since one expects to have as many averaged
local magnetization pointing up (mi > 0) as spins pointing down (mi < 0) with each
possible value of |mi |. Thus, the total magnetization, m, of a spin-glass vanishes at
all temperatures and it is not a good order parameter.
The spin-glass transition is characterized by a finite peak in the linear magnetic
susceptibility and a diverging non-linear magnetic susceptibility. Let us discuss the
former first and show how it yields evidence for the freezing of the local magnetic
moments. For a generic magnetic model
such that the magnetic field couples linP
early to the Ising spin, E E i hi si , the linear susceptibility is related, via
the static fluctuation-dissipation theorem to the correlations of the fluctuations of the
magnetization:

h si ih 
= h (si h si i)(sj h sj i) i .
(4.23)
ij
hj h=0

The averages in the rhs are taken without perturbing field. This relation is proven
by using the definition of h si ih and simply computing the derivative with respect to
hj . In particular,


(4.24)
ii = h (si h si i)2 i = 1 m2i 0 ,
P
1
with mi = h si i. The total susceptibility measured experimentally is N
ij ij .
On the experimental side we do not expect to see O(1) sample-to-sample fluctuations
in this global quantity. On the analytical side one can use a similar argument to
the one presented in Sect. to argue that should be self-averaging (it is a sum
over the entire volume of site-dependent terms). Thus, the experimentally observed
susceptibility of sufficiently large samples should be given by
X
X
X


= [ ] = N 1
[ ij ] N 1
[ ii ] = N 1
1 [ m2i ] ,
(4.25)
ij

since we can expect that cross-terms will be subleading in the large N limit under the
disorder average (note that ij can take negative values). The fall of at low temperatures with respect to its value at Tc , i.e. the cusp observed experimentally, signals
the freezing of the local magnetizations, mi , in the non-zero values that are more
favourable thermodynamically. Note that this argument is based on the assumption
that the measurement is done in equilibrium.
Thus, the natural and simpler global order parameter that characterizes the
spin-glass transition is
X
q N 1
[ m2i ]
(4.26)
i

105

as proposed in the seminal 1975 Edwards-Anderson paper. q vanishes in the high

temperature phase since all mi are zero but it does not in the low temperature phase
since the square power takes care of the different signs. Averaging over disorder
eliminates the site dependence. Thus, q is also given by
q = [ m2i ] .

(4.27)

These definitions, natural as they seem at a first glance, hide a subtle distinction that
we discuss below.
4.4.2

Pure states and more subtle order parameters

Let us keep disorder fixed and imagine that there remain more than two pure or
equilibrium states in the selected sample. A factor of two takes into account the spin
reversal symmetry. Later we will consider half the phase space, getting rid of this
trivial symmetry. Consider the disorder-dependent quantity
X
qJ = N 1
m2i
(4.28)
i

where the mi depend upon the realization of the exchanges. Then, two possibilities
for the statistical average in mi = hsi i have to be distinguished:
Restricted averages
If we interpret the statistical average in the same restricted sense as the one
discussed in the paramagnetic - ferromagnetic transition of the usual Ising model, i.e.
under a pinning field that selects one chosen pure state, in (4.28) we define a disorder
and pure state dependent Edwards-Anderson parameter,
qJ ea = N 1

N
X

2
(m
i ) ,

(4.29)

where we label the selected pure state. Although qJ ea could depend on it turns
out that in all known cases it does not and the label is superfluos.
In addition, qJ ea could fluctuate from sample to sample since the individual mi
do. It turns out that in the thermodynamic limit qJ ea does not fluctuate. With this
in mind we will later use

qea = qea
= qJ
(4.30)
ea
for the intra-state average. This is the interpretation of the order parameter proposed by Edwards-Anderson who did not take into account the possibility that is
discussed next.
In the clean Ising model, had we taken into account all the phase space, = 1, 2
(1)
(2)
and m
= mi = m > 0. If we kept only half phase space = 1
i = hsi i with mi
and mi = m > 0, say. The dependence on J does not exist in this case from the
106

very definition of the model. In the RBIM the J and i-dependences remain but there
are still only two states to be considered. In the Potts model with q equilibrium FM
states there are q possible values for and consequently m
i .
Full statistical averages
If, instead, the statistical average in mi runs over all possible equilibrium states
(and we imagine there are more than one even if we consider only half the phase
space, that is to say, if we eliminate spin-reversal) the quantity (4.28) has non-trivial
contributions from overlaps between different states.
Imagine each state has a probability weight wJ . Then
X
mi =
wJ m
(4.31)
i

and
N
1 X X J
1 X 2
mi =
w m
qJ =
N i
N i=1 i

!2

wJ wJ

N
1 X
m m .
N i=1 i i

(4.32)

A number of examples will clarify what we mean here. In the ferromagnetic phase
of the clean or dirty Ising models one has only two pure states with w1 = w2 = 1/2
and the fully averaged mi is mi = 1/2 hsi i1 + 1/2 hsi i2 = 0. If one considers half the
phase space, where spin-reversal is not allowed, then w1 = 1, mi = hsi i and one does
not see any difference between the intra-state and the fully averaged mi . As for the q
(1)
(2)
parameters, considering the full phase space qea = qea = m2 , while q = 0. Instead,
taking into account half the phase space q = qea = m2 , and qea and q are identical
order parameters in this case. In the Potts model with more than two equilibrium
states the intra-state and fully averaged local magnetizations are not identical.
In the ferromagnetic modes discussed in the previous paragraph the EdwardsAnderson order parameter takes the same value in each equilibrium state. This is
also the case in spin-glass models, qJ
ea independently of there being only two (as in
the usual ferromagnetic phase) or more (as we will see appearing in fully-connected
spin-glass models). Therefore it does not allow us to distinguish between the two-state
and the many-state scenarii. Instead, qJ does.
Having defined a disorder-dependent order parameter, qJ , and its disorder average,
q, that explains the decay of the susceptibility below Tc , we still have to study whether
this order parameter characterises the low temperature phase completely. It will turn
out that the knowledge of q is not enough, at least in fully-connected and dilute
spin-glass models. Indeed, one needs to consider the probability distribution of the
fluctuating qJ quantity, P (qJ ). The more pertinent definition of an order parameter
as being given by such a probability distribution allows one to distinguish between
the simple, two-state, and the many-state scenarii.
In practice, a way to compute the probability distribution of the order parameter is by using an overlap or correlation between two spin configurations,
107

say {si } and {i }, defined as

J
qs
= N 1

X
i

h si i i

(4.33)

J
where h . . . i is an unrestricted thermal average. qs
takes values between 1 and 1.
It equals one if {si } and {i } differ in a number of spins that is smaller than O(N ),
it equals 1 when the two configurations are totally anticorrelated with the same
proviso concerning a number of spins that is not O(N ) and it equals zero when {si }
and {i } are completely uncorrelated. Note that the self-overlap of a configuration
with itself is identically one for Ising spins. Other values are also possible. A related
definition is the one of the Hamming distance:

dJs = N 1

N
X
i=1

J
h (si i )2 i = 2(1 qs
).

(4.34)

The overlap can be computed by running a Monte Carlo simulation, equilibrating

a sample and recording many equilibrium configurations. With them one computes
the overlap and should find a histogram with two peaks at qea and qea (the values
of the overlap when the two configurations fall in the same pure state or in the sign
reversed ones) and, in cases with many different pure states, other peaks at other
J
J
values of qs
. This is observed in the SK model. Note that qs
is related to the q
definition above.

Figure 40: The overlap between two equilibrium configurations in a FM system.

4.4.3

Pinning fields

In the discussion of the ferromagnetic phase transition one establishes that one of
the two equilibrium states, related by spin reversal symmetry, is chosen by a small
pinning field that is taken to zero after the thermodynamic limit, limh0 limN .
In a problem with quenched disorder it is no longer feasible to choose and apply
a magnetic field that is correlated to the statistical averaged local magnetization in
a single pure state since this configuration is not known! Moreover, the remanent
108

magnetic field that might be left in any experience will certainly not be correlated
with any special pure state of the system at hand.
Which is then the statistical average relevant to describe experiments? We will
come back to this point below.
4.4.4

Divergent susceptibility

In a pure magnetic system with a second-order phase transition the susceptibility

of the order parameter to a field that couples linearly to it diverges when approaching
the transition from both sides. In a paramagnet, one induces a local magnetization
with a local field
N
X
ij hj
(4.35)
mi = h s i i =
j=1

P
with ij the linear susceptibilities, the magnetic energy given by E = E0 i si hi ,
and the field set to zero at the end of the calculation. Using this expression, the order
parameter in the high temperature phase becomes
q = qea =

N
N N N
1 XXX
1 X
[ ij ik hj hk ]
[ m2i ] =
N i=1
N i=1 j=1

(4.36)

k=1

If the applied fields are random and taken from a probability distribution such that
hj hk = 2 jk one can replace hj hk by 2 jk and obtain
q=

N
N N
1 XX 2
1 X
[ ] 2 SG 2 .
[ m2i ] =
N i=1
N i=1 j=1 ij

(4.37)

2 acts as a field conjugated to the order parameter. (One can also argue that a
uniform field looks random to a spin-glass sample and therefore the same result holds.
It is more natural though to use a trully random field since a uniform one induces a
net magnetization in the sample.) The spin-glass susceptibility is then defined as
SG

2 X
2 X
1 X 2
[ ij ] =
[ (h si sj i h si ih sj i)2 ] =
[ h s i s j i2 ]
N ij
N ij
N ij

in the high T phase and one finds that it diverges as T Tc+ as expected in a secondorder phase transition. (Note that there is no cancelation of crossed terms because
of the square.) Indeed, the divergence of SG is related to the divergence of the
non-linear magnetic susceptibility that is measurable experimentally and numerically.
An expansion
P of the total mangnetization in powers of a uniform field h acting as
E E h i si is
(3) 3
Mh = h
h + ... ,
(4.38)
6
109

and the first non-linear susceptibility is then given by

(3)

!4 +
*


4
X

3 Mh 
3N
1 ln Zh 

si
=
=
h3 h=0
h4 h=0
3
i

(4.39)
c

with the subindex c indicating that the quartic correlation function is connected.
Above Tc , mi = 0 at zero field,

3
X
X

(3) = 3
(h si sj sk sl i 3h si sj ih sk sl i) =
3 4N 6
h s i s j i2 ,
N
ij
ijkl

and one can identify SG when i = k and j = l plus many other terms that we assume
are finite. Then,
2
(4.40)
(3) = (SG 2 ) .
3
This quantity can be accessed experimentally. A careful experimental measurement
of (3) , (5) and (7) done by L. Levy demonstrated that all these susceptibilities
diverge at Tc .
4.4.5

Calorimetry

No cusp in the specific heat of spin-glasses is seen experimentally. Since one

expects a second order phase transition this means that the divergence of this quantity
must be very weak.
4.4.6

Critical scaling

Having identified an order parameter, the linear and the non-linear susceptibility
one can now check whether there is a static phase transition and, if it is of second
order, whether the usual scaling laws apply. Many experiments have been devoted to
this task. It is by now quite accepted that Ising spin-glasses in 3d have a conventional
second order phase transition. Still, the exponents are difficult to obtain and there is
no real consensus about their values. There are two main reasons for this: one is that
as Tc is approached the dynamics becomes so slow that equilibrium measurements
cannot really be done. Critical data are thus restricted to T > Tc . The other reason
is that the actual value of Tc is difficult to determine and the value used has an
important influence on the critical exponents. Possibly, the most used technique to
determine the exponents is via the scaling relation for the non-linear susceptibility:
 2 
h
(4.41)
nl = t f +
t
110

1
0.5

1
4

2
9

-1

1
2

Table 2: Critical exponents in the Ising spin-glass transitions.

with t = |T Tc |/Tc and one finds, approximately, the values given in Table 2.

4.5

The TAP approach

Disordered models have quenched random interactions. Due to the fluctuating

values of the exchanges, one expects that the equilibrium configurations be such that
in each equilibrium state the spins freeze in different directions. The local averaged
magnetizations need not be identical, on the contrary one expects h si i = mi and, if
many states exist, each of them can be identified by the vector (m1 , . . . , mN ).
One may try to use the naive mean-field equations (E.2) to characterize the low
temperature properties of these models at fixed quenched disorder and determine then
the different (m1 , . . . , mN ) values. It has been shown by Thouless-Anderson-Palmer
(TAP) that these equations are not completely correct even in the fully-connected
disordered case: a term which is called the Onsager reaction term is missing. This
term represents the reaction of the spin i: the magnetization of the spin i produces
a field hj = Jji mi = Jij mi on spin j; this field induces a magnetization mj =
jj hj = jj Jij mi on the spin j and this in turn produces a mean-field hi = Jij mj =
2
Jij jj Jij mi = jj Jij
mi on the site i. The equilibrium fluctuatio-dissipation relation
between susceptibilities and connected correlations implies jj = h (sj h sj i)2 i =
2
(1 m2j ) and one then has hi = (1 m2j )Jij
mi . The idea of Onsager or cavity
method is that one has to study the ordering of the spin i in the absence of its own
effect on the rest of the system. Thus, the field hi has to be subtracted
from the
P
mean-field created by the other spins in the sample, i.e. hcorr
= j Jij mj + hi
i
P 2
mi j Jij
(1 m2j ) where hi is the external field.
The generalization of this argument to p spin interactions is not so straightforward.
The TAP equations for p-spin fully connected models read

X
1
Jii ...i mi . . . mip
mi = tanh
(p 1)! 2 p 2
i2 6=...6=ip
i
.
+mi Jii2 2 ...ip (1 m2i2 ) . . . (1 m2ip ) + hi
(4.42)

the first contribution to the internal field is proportional to Ji12 ...ip N (p1)/2
111

and once the p 1 sums performed it is of order one. The reaction term instead is
proportional to Jii2 2 ...ip and, again, a simple power counting shows that it is O(1).
Thus, In disordered systems the reaction term is of the same order of the usual meanfield; a correct mean-field description has to include it. In the ferromagnetic case this
term can be neglected since it is subleading in N . Using the fact that there is a sum
over a very large number of elements, Ji21 ...ip can be replaced by its site-independent
variance [Ji21 ...ip ] = p!J 2 /(2N p1 ) in the last term in (4.42). Introducing the EdwardsP
Anderson parameter qea = N 1 i=1 m2i (note that we study the system in one pure
state) the TAP equations follow:

2 2
X

J p
mi = tanh
mi (1 qea )p1 .
Jii2 ...ip mi2 . . . mip + hi
(p 1)!
2
i2 6=...6=ip

(4.43)
The argument leading to the Onsager reaction term can be generalized to include the
combined effect of the magnetization of spin i on a sequence of spins in the sample,
i.e. the effect on i on j and then on k that comes back to i. These higher order terms
are indeed negligible only if the series of all higher
effects does not diverge. The


P order
ensuing condition is 1 > 2 1 2qea + N 1 i m4i .
The importance of the reaction term becomes clear from the analysis of the linearized equations, expected to describe the second order critical behavior for the
SK modelP(p = 2) in the absence of an applied field. The TAP equations become
mi ( j Jij mj J 2 mi + hi ). A change of basis to the one in which the Jij
matrix is diagonal leads to m ( J 2 )m + h . The staggered susceptibility
then reads


1
m 
= 1 2J + (J)2
.
(4.44)


h h=0
The staggered susceptibility for the largest eigenvalue of an interaction matrix in the
Gaussian ensemble, Jmax = 2J, diverges at c J = 1. Note that without the reaction
term the divergence appears at the inexact value T = 2Tc (see Sect. for the replica
solution of the SK model).
The TAP equations are the extremization conditions on the TAP free-energy density:
1 X
Ji1 ...ip mi1 . . . mip
f ({mi }) =
p!
i1 6=...6=ip

4p

X
i

i1 6=...6=ip

Ji21 ...ip (1 m2i1 ) . . . (1 m2ip )

h i mi + T

N 
X
1 + mi
i=1

1 + mi
1 mi 1 mi
ln
+
ln
2
2
2

(4.45)


The free-energy density as a function of the local magnetizations mi defines what is

usually called the free-energy landscape. Note that this function depends on N 1
112

variables, mi , and these are not necessarily identical in the disordered case in which the
interactions between different groups of spins are different. The stability properties
of each extreme {ml } are given by the eigenvalues of the Hessian matrix

f ({mk }) 
.
(4.46)
Hij
mi mj {m }
l

The number of positive, negative and vanishing eigenvalues determine then the number of directions in which the extreme is a minimum, a maximum or marginal. The
sets {ml } for which f ({ml }) is the absolute minima yield a first definition of equilibrium or pure states.
The TAP equations apply to {mi } and not to the configurations {si }. The values
of the {mi } are determined as minima of the TAP free-energy density, f ({mi }), and
they not need to be the same as those of the energy, H({si }), a confusion sometimes
encountered in the glassy literature. The coincidence of the two can only occur at
T 0.
4.5.1

The complexity or configurational entropy

There are a number of interesting questions about the extreme of the TAP freeenergy landscape, or even its simpler version in which the Onsager term is neglected,
that help us understanding the static behavior of disordered systems:
For a given temperature, T , how many solutions to the mean-field equations
exist? The number of solutions can be calculated using


YZ 1
YZ 1

eqi 
dmi (eqi ) det
dmi (mi mi ) =
NJ =
.
(4.47)
mj 
1
1
i
i
{mi } are the solutions to the TAP equations that we write as {eqi = 0}. The
last factor is the normalization of the delta function after the change of variables,
it ensures that we count one each time the integration variables touch a solution
to the TAP equations independently of its stability properties.
We define the complexity or the configurational entropy as the logarithm of the
number of solutions at temperature T divided by N :
J (T ) N 1 ln NJ (T ) .

(4.48)

The normalization with N suggests that the number of solutions is actually an

exponential of N . We will come back to this very important point below.
Does NJ (T ) dependend on T and does it change abruptly at particular values
of T that may or may not coincide with static and dynamic phase transitions?
One can define a free-energy level dependent complexity
J (f, T ) N 1 ln NJ (f, T )
113

(4.49)

where NJ (f, T ) is the number solutions in the interval [f, f + df ] at temperature

T.
From these solutions, one can identify the minima as well as all saddles of different type, i.e. with different indeces K. These are different kinds to metastable
states. Geometry constrains
the number of metastable states to satisfy Morse
PN
theorem that states l=1 (1)l = 1, where l is the number of negative eigenvalues of the Hessian evaluated at the solution l, for any continuous and wellbehaved function diverging at infinity in all directions.
One can then count the number of solutions to the TAP equations of each index,
NJ (K, T ), and define the corresponding complexity
J (K, T ) N 1 ln NJ (K, T ) ,

(4.50)

or even work at fixed free-energy density

J (K, f, T ) N 1 ln NJ (K, f, T ) .

(4.51)

Even more interestingly, one can analyse how are the free-energy densities of
different saddles are organized. For instance, one can check whether all maxima
are much higher in free-energy density than minima, etc.
What is the barrier, f = f1 f0 , between ground states and first excited
states? How does this barrier scale with the actual free-energy difference, f
between these states? The answer to this question is necessary to estimate
the nucleation radius for the reversal of a droplet under an applied field, for
instance.
The definitions of complexity given above are disorder-dependent. One might then
expect that the complexity will show sample-to-sample fluctuations and be characterized by a probability distribution. The quenched complexity, quenched , is then
the most likely value of J , it is defined through max P (J ) = P (quenched ). In
practice, this is very difficult to compute. Most analytic results concern the annealed
complexity
(4.52)
ann N 1 ln [ NJ ] = N 1 ln[ eN J ] .
One can show that the annealed complexity is smaller or equal than the quenched
one.

4.5.2

Weighted averages

Having identified many solutions to the TAP equations in the low-T phase one
needs to determine now how to compute statistical averages. A natural proposal is
to give a probability weight to each solution, w , and to use it to average the value
the observable of interest:
X
hOi =
wJ O
(4.53)

114

where labels the TAP solutions, O = O({m

i }) is the value that the observable
J
are
their
statistical weights, satisfying the
O takes in the TAP solution
,
and
w

P
normalization condition wJ = 1. Two examples can illustrate the meaning of this
average. In a spin-glass problem, if O = si , then O = m
i . In an Ising model in its
=
m
and
w = 1/2. Within the
ferromagnetic phase, if O = si , then O = m
i
TAP approach one proposes
J
eF
J
w = P F J
(4.54)

with FJ the total free-energy of the -solution to the TAP equations. The discrete
sum can be transformed into an integral over free-energy densities, introducing the
degeneracy of solutions quantified by the free-energy density dependent complexity:
Z
Z
1
1
hOi =
df eN f NJ (f, T )O(f ) =
df eN (f J (f,T )) O(f ) .
(4.55)
Z
Z

The normalization is the partition function

Z
Z
N f
Z = df e
NJ (f, T ) = df eN (f J (f,T )) .

(4.56)

We assumed that the labelling by can be traded by a labelling by f that implies

that at the same free-energy density level f the observable O takes the same value.
In the N limit these integrals can be evaluated by saddle-point, provided the
parenthesis is positive. The disorder-dependent complexity is generally approximated
with the annealed value introduced in eq. (4.52).
Freeenergy
Diverging barriers with N

Finite barriers

Threshold level
Metastable states
Equilibrium
Order parameters ( N dim. )

Figure 41: Sketch of the free-energy landscape in the p-spin model.

The equilibrium free-energy

115

The total equilibrium free-energy density, using the saddle-point method to evaluate the partition function Z in eq. (4.56), reads
feq = N 1 ln Z = min [f T J (f, T )] min J (f, T )
f

(4.57)

It is clear that J (f, T ) is the Landau free-energy density of the problem with f
playing the r
ole of the energy and J of the entropy. If we use f = (E T S)/N =
e T s with E the actual energy and S the microscopic entropy one has
J (f, T ) = e T (s + J (f, T )) .

(4.58)

Thus, J is an extra contribution to the total entropy that is due to the exponentially large number of metastable states. Note that we do not distninguish here their
stability.
Note that J is subtracted from the total free-energy. Thus, it is possible that in
some cases states with a higher free-energy density but very numerous have a lower
total free-energy density than lower lying states that are less numerous. Collectively,
higher states dominate the equilibrium measure in these cases.
The order parameter
The Edwards-Anderson parameter is understood as a property of a single state.
Within the TAP formalism on then has
X
1
2
qJ
(m
(4.59)
ea = N
i ) .
i

2
An average over pure states yields wJ (m
i ) .
P
Instead, the statistical equilibrium magnetization, mi = hsi i = wJ m
i , squared
is
!2
X
X

2
2
J
q J h s i i = mi =
w mi
=
wJ wJ m
(4.60)
i mi .

If there are multiple phases, the latter sum has crossed contributions from terms with
6= . These sums, as in a usual paramagnetic-ferromagnetic transition have to
be taken over half space-space, otherwise global up-down reversal would imply the
cancellation of all cross-terms.
This discussion is totally equivalent to the one we developed when we introduced
qJ and qea .

4.6

Metastable states in two families of models

In this subsection we summarize the structure of metastable states found in two

families of models, the ones in the SK and the ones in the p-spin class.

116

4.6.1

High temperatures

For all models, at high temperatures f (mi ) is characterized by a single stable

absolute minimum in which all local magnetizations vanish, as expected; this is the
paramagnetic state.
In models of the p-spin class the mi = 0 for all i minimum continues to exist at all
temperatures. However, even if it is still a minimum of f the TAP free-energy density,
f ({mi }),it becomes unstable thermodynamically at a given temperature, and it is
substituted as the equilibrium state, by other non-trivial configurations with mi 6= 0
that are the absolute minima of . Note the difference with the ferromagnetic problem
for which the paramagnetic solution is no longer a minimum below Tc .
4.6.2

Low temperatures

At low temperature many equilibrium states appear (and not just two as in an
Ising ferromagnetic model) and they are not related by symmetry (as spin reversal
in the Ising ferromagnet or a rotational symmetry in the Heisenberg ferromagnet).
These are characterized by non-zero values of the local magnetizations m
i that are
different in different states.
At low-temperatures both the naive mean-field equations and the TAP equations
have an exponential in N number of solutions and still an exponential in N number of them correspond to absolute minima of the mi -dependent free-energy density.
This means that J (T ) and even J (0, f0 , T ) are quantities O(1). These minima can
be identified as different states that could be accessed by applying the corresponding
site-dependent pinning fields.
The derivation and understanding of the structure of the TAP free-energy landscape is quite subtle and goes beyond the scope of these Lectures. Still, we will
briefly present their structure for the SK and p-spin models to give a flavor of their
complexity.
The SK model
The first calculation of the complexity in the SK model appeared in 1980. After
25 years of research the structure of the free-energy landscape in this system is still
a matter of discussion. At present, the picture that emerges is the following. The
temperature-dependent annealed complexity is a decreasing function of temperature
that vanishes only at Tc but takes a very small value already at 0.6Tc . Surprisingly
enough, at finite but large N the TAP solutions come in pairs of minima and saddles
of type one, that is to say, extrema with only one unstable direction. These states
are connected by a mode that is softer the larger the number of spins: they coalesce
and become marginally stable in the limit N . Numerical simulations show that
starting from the saddle-point and following the left direction along the soft mode one
falls into the minimum; instead, following the right direction along the same mode
one falls into the paramagnetic solution. See Fig. 42 for a sketch of these results. The
free-energy difference between the minimum and saddle decreases for increasing N and

117

1/N < ln N >

0.2

0.1
0
0
N

0.6

T/Tc

1/2

PM

Figure 42: Left: sketch of the temperature dependent complexity, (T ), of the SK. It
actually vanishes only at Tc but it takes a very small value already at 0.6Tc . Right:
pairs of extrema in the SK model with N large and N limit.
one finds, numerically, an averaged f N 1.4 . The extensive complexity of minima
and type-one saddles is identical in the large N limit, (0, T ) = (1, T ) + O(N 1 )
[Aspelmeier, Bray, Moore (06)] in such a way that the Morse theorem is respected.
The free-energy dependent annealed complexity is a smooth function of f with support
on a finite interval [f0 , f1 ] and maximum at fmax . The Bray and Moore annealed
calculation (with supersymmetry breaking) yields fmax = 0.654, max = 0.052,
(fmax ) = 8.9. The probability of finding a solution with free-energy density f can
be expressed as
r
N (fmax ) N | (fmax )|(f fmax )2
eN (f,T )
N (f, T )
, (4.61)
=

e 2
p(f, T ) =
N (T )
N (T )
2
R
where we evaluated the total number of solutions, N (T ) = df eN (f,T ) , by steepest
descent The complexity vanishes linearly close to f0 : (f, T ) (f f0 ) with < .
Only the lowest lying TAP solutions contribute to the statistical weight. The
complexity does not contribute to in the large N limit:
=

f ann (f, T ) f (f f0 )
>0

iff

>

(4.62)

and min fmin = f0 . See Fig. 43. The total free-energy density in the expo118

f0

fmax

Figure 43: The complexity as a function of f for the SK model.

nential is just the free-energy density of each low-lying solution.
The (spherical) p-spin model
The number and structure of saddle-points is particularly interesting in the p
3 cases and it is indeed the reason why these models, with a random first order
transition, have been proposed to mimic the structural glass arrest. The p 3 model
has been studied in great detail in the spherical case,
is to say, when spins are
Pthat
N
not Ising variables but satisfy the global constraint, i=1 s2i = N .
Although in general the minima of the mean-field free energy do not coincide with
the minima of the Hamiltonian, they do in the spherical p-spin model. Their positions
in the phase space does not depend on temperature, while their self-overlap does. At
T = 0 a state (stable or metastable) is just a minimum (absolute or local) of the
energy. For increasing T energy minima get dressed up by thermal fluctuations, and
become states. Thus, the states can be labeled by their zero-temperature energy E 0 .
The complexity is given by


pz 2
p1 2
2
2p
1
ln
,
(4.63)
+
z 2 2+
(E) =
2
2
2
p z
p
where z is

z = E
The complexity vanishes at

E02

Ec2

/(p 1) .

E 0 = Emin = f (p) ,

119

(4.64)

(4.65)

Free energy

(3)
(5)
(2)
(4)

(1)

Ts

T*

Tmax
(Ec)

T
tap

Figure 44: The TAP free-energy as a function of T in the spherical p-spin model.
(1) : free energy of the paramagnetic solution for T > T , Ftot for T < T ; (2) :
free energy of the lowest TAP states, with zero temperature energy Emin ; (3) : free
energy of the highest TAP states, corresponding to Ec ; (4) : an intermediate value of
E0 leads to an intermediate value of f at any temperature; (5) : feq (T ); the difference
between curves (5) and (1) gives the complexity T Sc (feq (T ), T ).

120

the ground state of the system, and it is real for zero-temperature energies E < Eth
with
s
2(p 1)
Eth =
.
(4.66)
p
Emin is the zero-T energy one finds with the replica calculation using a 1-step RSB
Ansatz as we will see below. The finite-T energy of a state is
p

E = q2 E0


1 
(p 1)qp pqp1 + 1 .
2T

(4.67)

This means that:

There can be only a finite number of states with E < E0 .
It can be shown that below Eth minima dominate on average.
Above Eth one can show that there are states but these are unstable.
Each zero-temperature state is characterized
by unit N -vector s
i and it gives rise
p

to a finite-T state characterized by mi = q(E, T )s

i with q(E, T ) given by
p
2 )2
(E + E 2 Eth
.
(4.68)
q p2 (1 q)2 = T 2
2
(p 1)
(q(E, T = 0) = 1 and at finite T the solution with q closest to 1 has to be chosen.)
The self-overlap at the threshold energy, E Eth , is then
p2
qth
(1 qth )2 = T 2

2
.
p(p 1)

(4.69)

Another way for the q equation to stop having solution, is by increasing the temperature, T > Tmax (E 0 ), at fixed bare energy E 0 . This means that, even though
minima of the energy do not depend on the temperature, states, i.e. minima of the
free energy do. When the temperature becomes too large, the paramagnetic states
becomes the only pure ergodic states, even though the energy landscape is broken up
in many basins of the energy minima. This is just one particularly evident demonstration of the fundamental different between pure states and energy minima. Tmax (E 0 )
is obtained as the limiting temperature for which eq. (4.68) admits a solution. It is
given by

 p2
 

2
p1
2
p2
q
.
(4.70)
Tmax (E 0 ) =
p
p
2
E 0 E 0 2 Eth

Tmax is a decreasing function of E 0 . The last states to disappear then are the ones
with minimum energu Emin , ceasing to exist at TT AP = Tmax (Emin ).
Below a temperature Td , an exponential (in N ) number of metastable states contribute to the thermodynamics in such a non-trivial way that their combined contribution to the observables makes them those of a paramagnet. Even if each of these
121

states is non-trivial (the mi s are different from zero) the statistical average over all
of them yields results that are identical to those of a paramagnet, that is to say, the
free-energy density is 1/(4T ) as in the mi = 0 paramagnetic solution. One finds
s
p(p 2)p2
Td =
.
(4.71)
2(p 1)p1
At a lower temperature Ts (Ts < Td ) there is an entropy crisis, less than an
exponential number of metastable states survive, and there is a static phase transition
to a glassy state.
In the p-spin models there is a range of temperatures in which high lying states
dominate this sum since they are sufficiently numerous so as to have a complexity that
renders the cimbined term f J (f, T ) smaller (in actual calculations the disorder
dependent complexity is approximated by its annealed value). In short:
Above Td the (unique) paramagnetic solution dominates, q = 0 and = f =
1/(4T ).
In the interval T [Ts , Td ] an exponentially large number of states (with q 6= 0
given by the solution to pq p2 (1 q) = 2T 2 ) dominate the partition sum.
= 1/(4T ) appearing as the continuation of the paramagnetic solution.
At T < Ts the lowest TAP states with E 0 = Emin control the partition sum.
Their total free-energy is different from 1/(4T ).
This picture is confirmed with other studies that include the use of pinning fields
adapted to the disordered situation, the effective portential for two coupled real replicas, and the dynamic approach.
Low temperatures, entropy crisis
The interval of definition of (E, T ) is the same as (E), that is E [Emin : Eth ].
Assuming that at a given temperature T the energy Eeq (T ) minimizing lies in this
interval, what happens if we lower the temperature? Remember that the complexity
is an increasing function of E, as of course is f (E, T ). When T decreases we favor states with lower free energy and lower complexity, and therefore Eeq decreases.
As a result, it must exist a temperature Ts , such that, Eeq (Ts ) = Emin and thus,
(Eeq (T )) = (Emin ) = 0. Below Ts the bare energy Eeq cannot decrease any further: there are no other states below the ground states Emin . Thus, Eeq (T ) = Emin
for each temperature T Ts . As a result, if we plot the complexity of equilibrium
states (Eeq (T )) as a function of the temperature, we find a discontinuity of the
first derivative at Ts , where the complexity vanishes. A thermodynamic transition
takes place at Ts : below this temperature equilibrium is no longer dominated by
metastable states, but by the lowest lying states, which have zero complexity and
lowest free energy density.
We will show that Ts is the transition temperature found with a replica calculation.
The temperature where equilibrium is given for the first time by the lowest energy
122

states, is equal to the static transition temperature. Above Ts the partition function is
dominated by an exponentially large number of states, each with high free energy and
thus low statistical weight, such that they are not captured by the overlap distribution
P (q). At Ts the number of these states becomes sub-exponential and their weight
nonzero, such that the P (q) develops a secondary peak at qs 6= 0.
The threshold
The stability analysis of the TAP solutions on the threshold level demonstrates
that these are only marginally stable, with a large number of flat directions.
Hierarchy of metastable states
The ordering of TAP solutions with different stability properties according to their
free-energy density has also been studied in great detail. The exact scaling with N
of the height of the barriers separating these solutions is harder to obtain. All these
are accepted to be exponential on the number of spins in the sample.
very
warm
liquid

1/Tcage
warm
liquid

1/Td

viscous
liquid

1/Tg

?
"conguration space"

glass

1/T

Figure 45: Sketch for the Random first order transition (RFPT) scenario. Left: the
experimental observation. Right: its interpretation in terms of metastable states.
Finite dimensions
In finite-dimensional systems, only equilibrium states can break the ergodicity, i.e.
states with the lowest free energy density. In other words, the system cannot remain
trapped for an infinite time in a metastable state, because in finite dimension free
energy barriers surrounding metastable states are always finite.
The extra free energy of a droplet of size r of equilibrium phase in a background
metastable phase has a positive interface contribution which grows as rd1 , and a

123

negative volume contribution which grows as rd ,

f = rd1 f rd

(4.72)

where here is the surface tension and f is the bulk free energy difference between the
two phases. This function has always a maximum, whose finite height gives the free
energy barrier to nucleation of the equilibrium phase (note that at coexistence f = 0
and the barrier is infinite). Therefore, if initially in a metastable states the system
will, sooner or later, collapse in the stable state with lower free energy density. For
this reason, in finite dimension we cannot decompose the Gibbs measure in metastable
components. When this is done, it is always understood that the decomposition is
only valid for finite times, i.e times much smaller than the time needed for the stable
equilibrium state to take over. On the other hand, in mean-field systems (infinite
dimension), barriers between metastable states may be infinite in the thermodynamic
limit, and it is therefore possible to call pure states also metastable states, and to
assign them a Gibbs weight wJ . We will analyze a mean-field spin-glass model, so that
we will be allowed to perform the decomposition above even for metastable states.
Comments
There is a close relationship between the topological properties of the model and
its dynamical behavior. In particular, the slowing down of the dynamics above but
close to Td is connected to the presence of saddles, whose instability decreases with
decreasing energy. In fact, we have seen that the threshold energy level Eth separating saddles from minima, can be associated to the temperature Tth = Td , marking
the passage from ergodicity to ergodicity breaking. In this context the dynamical
transition can be seen as a topological transition. The plateau of the dynamical correlation function, which has an interpretation in terms of cage effect in liquids, may
be reinterpreted as a pseudo-thermalization inside a saddle with a very small number
of unstable modes.

4.7

The replica method

A picture that is consistent with the one arising from the naive mean-field approximation but contradicts the initial assumption of the droplet model arises from the
exact solution of fully-connected spin-glass models. These results are obtained using
a method which is called the replica trick and that we will briefly present below.
In Sect. we argued that the typical properties of a disordered system can be
computed from the disorder averaged free-energy
Z
[ FJ ] dJP (J)FJ .
(4.73)
One then needs to average the logarithm of the partition funtion. In the annealed
approximation one exchanges the ln with the average over disorder and, basically,
124

considers the interactions equilibrated at the same temperature T as the spins:

[ ln ZJ ] ln[ ZJ ] .

(4.74)

This approximation turns out to be correct at high temperatures but incorrect at low
ones.
The replica method allows one to compute [ FJ ] for fully-connected models. It is
based on the smart use of the identity
ZJn 1
.
(4.75)
n0
n
The idea is to compute the right-hand-side for finite and integer n = 1, 2, . . . and then
perform the analytic continuation to n 0. Care should be taken in this step: for
some models the analytic continuation may be not unique. It turns out that this is
indeed the case for the emblematic Sherrington-Kirkpatrick model, as discussed by
Palmer and van Hemmen in 1979 though it has also been recently shown that the
free-energy f (T ) obtained by Parisi with the replica trick is exact!
The disorder averaged free-energy is given by
Z

Z
1
n
[ FJ ] = dJP (J) ln ZJ = lim
dJP (J)ZJ 1 ,
(4.76)
n0 n
ln ZJ = lim

where we have exchanged the limit n 0 with the integration over the exchanges.
For integer n the replicated partition function, ZJn , reads
X
1
n
ZJn =
e[EJ ({si })+...+EJ ({si }] .
(4.77)
}
{sa
i

Here {sa } {s1 =1} . . . {sn =1} . ZJn corresponds to n identical copies of the
i
i
i
original system, that is to say, all of them with the same realization of the disorder.
Each copy is characterized by an ensemble of N spins, {sai }. We label the copies with
a replica index a = 1, . . . , n. For p-spin disordered spin models ZJn takes the form
i
P hP
P
a
a
X na=1
Ji1 ...ip sa
...s
+
h
s
i
i
i
i
p
1
i1 6=...6=ip
i
ZJn =
e
.
(4.78)
{sa
}
i

The average over disorder amounts to computing a Gaussian integral for each set of
spin indices i1 , . . . ip . One finds
P a a 2 P P
P
J2
X
X 2p1
a
hi sa
s ...sip ) +
(
i
n
2N
a
i
a i1
i1 6=...6=ip

eF ({si }) . (4.79)
[ ZJ ] =
e
{sa
}
i

{sa
}
i

The function F ({sai }) is not random. It depends on the spin variables only but it
includes terms that couple different replica indices:

!p
2 2 X
X
XX
1
N

sai sbi
+ n
hi sai .
(4.80)
F ({sai })
2
N i
a
i
a6=b

125

In writing the last expression we have dropped terms that are subleading in N (in
complete analogy with what we have done for the pure p spin ferromagnet). The
constant term N n 2 J 2 /2 originates in the terms with a = b, for which (sai )2 = 1.
To summarize, we started with an interacting spin model. Next, we enlarged the
number of variables from N spins to N n replicated spins by introducing n noninteracting copies of the system. By integrating out the disorder we decoupled the
sites but we payed the price of coupling the replicas. Hitherto the replica indices
act as a formal tool introduced to compute the average over the bond distribution.
Nothing distinguishes one replica from another and, in consequence, the free-energy
F ({sai }) is invariant under permutations of the replica indices.
The next step to follow is to identify the order parameters and transform the freeenergy into an order-parameter dependent expression to be rendered extremal at their
equilibrium values. In a spin-glass problem we already know that the order parameter
is not the global magnetization as in a pure magnetic system but the parameter q or
more generally the overlap between states. Within the replica calculation an overlap
between replicas
X
qab N 1
sai sbi
(4.81)
i

naturally appeared in eq. (4.80). The idea is to write the free-energy density as a
function of the order parameter qab and look for their extreme in complete analogy
with what has been done for the fully-connected ferromagnet. This is, of course, a
tricky business, since the order parameter is here a matrix with number of elements n
going to zero! A recipe for identifying the form of the order parameter (or the correct
saddle-point solution) has been proposed by G. Parisi in the late 70s and early 80s.
This solution has been recently proven to be exact for mean-field models by two
mathematical physics, F. Guerra and M. Talagrand. Whether the very rich physical
structure that derives from this rather formal solution survives in finite dimensional
systems remains a subject of debate.
Introducing the Gaussian integral
Z
P a b 2
P a b N 2
N
1
(4.82)
dqab eJqab i si si 2 qab = e 2 ( N J i si si )

for each pair of replica indices a 6= b, one decouples the site indeces, i, and the
averaged replicated partition function can be rewritten as
Z Y
[ZJn ] =
dqab eF (qab )
(4.83)
a6=b

and
F (qab )

(qab )

N 2J 2 X p

(4.84)
qab + n N ln (qab ) ,

2
a6=b
X
X
X
eH(qab ,sa ) ,
H(qab , sa ) = J
qab sa sb h
sa(4.85)
sa

ab

126

where for simplicity we set hi = h. The factor N in front of ln comes from the
decoupling of the site indeces. Note that the transformation (4.82) serves to uncouple
the sites and to obtain then the very useful factor N in front of the exponential. The
partition function Z(qab ) is the one of a fully-connected Ising model with interaction
matrix qab .
We next summarize the saddle-point method used to evaluate the partition sum
and free-energy and we enumerate and briefly discuss the Ansatze used to characterize
the Q matrix and the phases they represent.

4.8

Saddle-point evaluation

Having extracted a factor N in the exponential suggests to evaluate the integral

over qab with the saddle-point method. This, of course, involves the a priori dangerous
exchange of limits N and n 0. The replica theory relies on this assumption.
One then writes
1
sp
)
(4.86)
lim [ fJ ] lim f (qab
n0 n
N
and searches for the solutions to the n(n 1)/2 extremization equations

f (qab ) 
=0.
(4.87)
qcd qsp
ef

In usual saddle-point evaluations the saddle-point one should use is (are) the one(s)
that correspond to absolute minima of the free-energy density. In the replica calculation the number of variables is n(n 1)/2 that becomes negative! when n < 1
and makes the saddle-point evaluation tricky. In order to avoid unphysical complex
results one needs to focus on the saddle-points with positive (or at least semi-positive)
definite Hessian

f (qab ) 
H
,
(4.88)
qcd qef qsp
ab

and these sometimes corresponds to maxima (instead of minima) of the free-energy

density.
The saddle-point equations are also self-consistency equations
sp
qab
= hsa sb iH(qab ,{sa }) = [ hsa sb i ]

(4.89)

where the second member means that the average is performed with the single site
Hamiltonian H(qab , sa ) and the third member is just one of the averages we would
like to compute.
The partition function in eq. (4.85) cannot be computed for generic qab since
there is no large n limit to exploit on the contrary, n 0. Thus, one usually looks
for solutions to eqs. (4.87) within a certain family of matrices qab . We discuss below
the relevant parametrizations.
127

4.8.1

Replica symmetry (RS)

In principle, nothing distinguishes one replica from another one. This is the reason
why Sherrington and Kirkpatrick looked for solutions that preserve replica symmetry:
for all a 6= b .

qab = q ,

Inserting this Ansatz in (4.84) and (4.85) and taking n 0 one finds
!
r
Z
dz z2 /2
pq p1
2

q=
e
tanh
z + h .
2
2

(4.90)

(4.91)

This equation resembles strongly the one for the magnetization density of the p-spin
ferromagnet, eq. (E.7).
Let us first discuss the case p = 2, i.e. the SK model. In the absence of a magnetic
field, one finds a second order phase transition at Ts = J from a paramagnetic (q = 0)
to a spin-glass phase with q 6= 0. In the presence of a field there is no phase transition.
SK soon realized though that there is something wrong with this solution: the entropy
at zero temperature is negative, S(0) = 1/(2), and this is impossible for a model
with discrete spins, for which S is strictly positive. de Almeida and Thouless later
showed that the reason for this failure is that the replica symmetric saddle-point is not
stable, since the Hessian (4.88) is not positive definite and has negative eigenvalues.
The eigenvalue responsible for the instability of the replica symmetric solution is
called the replicon.
Comparison with the TAP equations
P shows that the RS Ansatz corresponds to the
assumption that the local fields hi = ii ...ii Ji1 ...ip mi1 . . . mip + h are independent
1

and have a Gaussian distribution with average h and variance 2 = J 2 q p1 . Numerical

simulations clearly show that this assumption is invalid.
Interestingly enough, the numerical values for several physical quantities obtained
with the replica symmetric solution do not disagree much with numerical results. For
instance, the ground state zero-temperature energy density is E 0 = 0.798 while with
numerical simulations one finds E 0 0.76.
For the p > 2 model one finds that the replica symmetric solution is stable at all
temperatures. However, the problem of the negative entropy remains and should be
solved by another solution. The transition must then have aspects of a first-order
one, with another solution appearing at low temperatures and becoming the most
convenient one at the transition.

4.8.2

One step replica symmetry breaking (1RSB)

The next challenge is to device a replica symmetry breaking Ansatz, in the form
of a matrix qab that is not invariant under permutations of rows or columns. There is
128

q1

q1

2
q

q1

o
q

2
q

q1

q1
o

q1
o
q

o
o

q1

o
q

2
q

q1

q1
q

Figure 46: Left: a one-step replica symmetry breaking (1RSB) Ansatz. Right: a twostep replica symmetry breaking Ansatz. The elements on the main diagonal vanish
identically. In the 1RSB case the diagonal blocks have size m m. In the 2RSB the
proceudre is repeated and one has blocks of size m1 m1 with smaller diagonal blocks
of size m2 m2 .
no first principles way of doing this, instead, the structure of the Ansatz is the result
of trial and error. Indeed, a kind of minimal way to break the replica symmetry is to
propose a structure in blocks as the one shown in Fig. 46-left. The diagonal elements
are set to zero as in the RS case. Square blocks of linear size m close to the main
diagonal are filled with a paramater q1 . The elements in the rest of the matrix take
a different value q0 and one takes 0 q0 q1 . The matrix qab depends on three
parameters q0 , q1 , m and one has to find the values such that the free-energy density
is maximized! The conditions for a extreme are
f (q0 , q1 , m)
f (q0 , q1 , m)
f (q0 , q1 , m)
=
=
=0.
q0
q1
m

(4.92)

In the SK model (p = 2) the 1RSB Ansatz yields a second order phase transition
(q0 = q1 = 0 and m = 1 at criticality) at a critical temperature Ts = J, that remains
unchanged with respect to the one predicted by the RS Ansatz. The 1RSB solution
is still unstable below Ts and in all the low temperature phase. One notices, however,
that the zero temperature entropy, even if still negative and incorrect, takes a value
that is closer to zero, S(T = 0) 0.01, the ground state energy is closer to the
value obtained numerically, and the replicon eigenvalue even if still negative has an
absolute value that is closer to zero. All this suggest that the 1RSB Ansatz is closer
to the exact solution.
Instead, in all cases with p 3 the 1RSB Ansatz is stable below the static critical
temperature Ts and all the way up to a new characteristic temperature 0 < Tf < Ts .
Moreover, one can prove that in this range of temperatures the model is solved exactly
by this Ansatz. The critical behavior is quite peculiar: while the order parameters q0
and q1 jump at the transition from a vanishing value in the paramagnetic phase to
a non-zero value right below Ts , all thermodynamic quantities are continuous since
129

m = 1 at Ts and all q0 and q1 dependent terms appear multiplied by 1 m. This

is a mixed type of transition that has been baptized random first-order. Note that
disorder weakens the critical behavior in the p 3-spin models. In the limit p
the solutions become m = T /Tc , q0 = 0 and q = 1.
4.8.3

k-step replica symmetry breaking (kRSB)

The natural way to generalize the 1RSB Ansatz is to propose a k-step one. In
each step the off-diagonal blocks are left unchanged while the diagonal ones of size
mk are broken as in the first step thus generating smaller square blocks of size mk+1 ,
close to the diagonal. At a generic k-step RSB scheme one has
0 q0 q1 . . . qk1 qk 1 ,
n = m0 m1 . . . mk mk+1 ,

(4.93)
(4.94)

parameters. In the n 0 limit the ordering of the parameters m is reversed

0 = m0 m1 . . . mk mk+1 .

(4.95)

In the SK model one finds that any finite k-step RSB Ansatz remains unstable.
However, increasing the number of breaking levels the results continue to improve
with, in particular, the zero temperature entropy getting closer to zero. In the p 3
case instead one finds that the 2RSB Ansatz has, as unique solution to the saddlepoint equations, one that boils down to the 1RSB case. This suggests that the 1RSB
Ansatz is stable as can also be checked with the analysis of the Hessian eigenvalues:
the replicon is stricly positive for all p 3.
4.8.4

Full replica symmetry breaking

In order to construct the full RSB solution the breaking procedure is iterated an
infinite number of times. The full RSB Ansatz thus obtained generalizes the block
structure to an infinite sequence by introducing a function
q(x) = qi ,

mi+1 < x < mi

(4.96)

with 0 x 1. Introducing q(x) sums over replicas are traded by integrals over x;
for instance
Z 1
1X l
dx q l (x) .
(4.97)
qab =
n
0
a6=b

The free-energy density becomes a functional of the function q(x). The extremization
condition is then a hard functional equation. A Landau expansion expected to

130

be valid close to the assumed second order phase transition simplifies the task of
solving it. For the SK model one finds
 x
0 x x1 = 2q(1) ,
2 ,
q(x) =
(4.98)
qea qmax = q(1) ,
x1 = 2q(1) x 1 ,
at first order in |T Tc |, with q(1) = |T Tc |/Tc and x1 = 2q(1). The stability
analysis yields a vanishing replicon eigenvalue signalling that the full RSB solution is
marginally stable.
One can also recover the particular case of the 1RSB using a q(x) with two
plateaux, at q0 and q1 and the breaking point at x = m.

RS

q
0

Full RSB

qea

qea

qea

1RSB

0
0

x1

x1

Figure 47: The function q(x) for a replica symmetric (left), one step replica symmetry
breaking (center) and full replica symmetry breaking Ans
atze.
Marginality condition
In the discussion above we chose the extreme that maximize the free-energy density since we were interested in studying equilibrium properties. We could, instead,
use a different prescription, though a priori not justified, and select other solutions.
For example, we can impose that the solution is marginally stable by requiring that
the replicon eigenvalue vanishes. In the p = 2 this leads to identical results to the
ones obtained with the usual prescription since the full-RSB Ansatz is in any case
marginally stable. In the p-spin models with p 3 instead it turns out that the
averaged properties obtained in this way correspond to the asymptotic values derived
with the stochastic dynamics starting from random initial conditions. This is quite a
remarkable result.

4.8.5

Interpretation of replica results

Let us now discuss the implications of the solution to fully-connected disordered

models obtained with the, for the moment, rather abstract replica formalism.
The interpretation uses heavily the identification of pure states. Their definition
is a tricky matter that we will not discuss in detail here. We will just assume it can
131

be done and use the analogy with the ferromagnetic system and its two pure states
and the TAP results at fixed disorder. As we already know, which are the pure
states, its properties, number, etc. can depend on the quenched disorder realization
and fluctuate from sample to sample. We will keep this in mind in the rest of our
discussion.
Let us then distinguish the averages computed within a pure state and over all
configuration space. In a ferromagnet with no applied magnetic field this is simple
to grasp: at high temperatures there is just one state, the paramagnet, while at low
temperatures there are two, the states with positive and negative magnetization. If
one computes the averaged magnetization restricted to the state of positive (negative)
magnetization one finds meq > 0 (meq < 0); instead, summing over all configurations
meq = 0 even at low temperatures. Now, if one considers systems with more than just
two pure states, and one labels them with Greeks indices, averages within such states
are denoted hOi while averages taken with the full Gibbs measure are expressed as
X
hOi =
wJ h O i .
(4.99)

wJ is the probability of the state given by

J

wJ

eF
=
,
ZJ

with

ZJ =

eF

(4.100)

P
and thus satisfying the normalization condition wJ = 1. FJ can be interpreted as
the total free-energy of the state . These probabilities, as well as the state dependent
averages, will show sample-to-sample fluctuations.
One can then define an overlap between states:
X
X

qJ N 1
hsi i hsi i = N 1
m
(4.101)
i mi
i

and assume rename the self-overlap the Edwards-Anderson parameter

X
qJ N 1
hsi i hsi i qJ ea .

(4.102)

The statistics of possible overlaps is then characterized by a probability function

X
PJ (q)
wJ wJ (q q ) ,
(4.103)

where we included a subindex J to stress the fact that this is a strongly sampledependent quantity. Again, a ferromagnetic model serves to illustrate the meaning of
PJ (q). First, there is no disorder in this case so the J label is irrelevant. Second, the
high-T equilibrium phase is paramagnetic, with q = 0. P (q) is then a delta function
with weight 1 (see the left panel in Fig. 48). In the low-T phase there are only two
132

pure states with identical statistical properties and qea = m2 . Thus, P (q) is just the
sum of two delta functions with weight 1/2 (central panel in Fig. 48).
Next, one can consider averages over quenched disorder and study
Z
X
[ PJ (q) ] dJ P (J)
wJ wJ (q q ) .
(4.104)

How can one access PJ (q) or [ PJ (q) ]? It is natural to reckon that

!
X
X
2
EJ () EJ (s)
1
PJ (q) = Z
e
e
N
i si q
s

(4.105)

that is to say, PJ (q) is the probability of finding an overlap q between two real replicas
of the system with identical disordered interactions in equilibrium at temperature T .
This identitiy gives a way to compute PJ (q) and its average in a numerical simulation: one just has to simulate two independent systems with identical disorder in
equilibrium and calculate the overlap.
But there is also, as suggested by the notation, a way to relate the pure state
structure to the replica matrix qab . Let us consider the simple case

n1 X
X
1
Z
si eEJ ({si }) = J n
[ mi ] = ZJ1
s1i eEJ ({si })
ZJ
{si }
{s1i }

1 X 1 Pn EJ ({sai })
a=1
(4.106)
s e
= n
ZJ a i
{si }

where we singled out the replica index of the spin to average. This relation is valid
for all n, in particular for n 0. In this limit the denominator approaches one and
the average over disorder can be simply evaluated
X
ef f
a
[ mi ] =
s1i eE ({si })
(4.107)
}
{sa
i

and introducing back the normalization factor Z n = 1 =

Pn
a
P
ef f
a
= [ {sa } e a=1 EJ ({si }) ] = eE ({si }) we have
i

[ mi ] = h sai iEef f

{sa
}
i

Pn

a=1

EJ ({sa
i })

(4.108)

with a any replica index. The average is taken over the Gibbs measure of a system with
effective Hamiltonian Eef f . In a replica symmetric problem in which all replicas are
identical this result should be independent of the label a. Instead, in a problem with
replica symmetry breaking the averages on the right-hand-side need not be identical
for all a. This could occur in a normal vectorial theory with dimension n in which
133

not all components take the same expected value. It is reasonable to assume that the
full thermodynamic average is achieved by the sum over all these cases,
n

1X a
h si i .
n0 n
a=1

[ mi ] = lim

(4.109)

Let us now take a less trivial observable and study the spin-glass order parameter
q
q

[ h si i2 ] = ZJ1

si eEJ ({si }) ZJ1

{si }

{i }

i eEJ ({i })

n2
X
Z

si i eEJ ({si })EJ ({i })

Zn
{si },{i }

Pn
X
a
1

E
({s
})
J
i

a=1
s1 s2 e
ZJn a i i

(4.110)

{si }

In the n 0 limit the denominator is equal to one and one can then perform the
average over disorder. Introducing back the normalization one then has
q = h sai sbi iEef f ({sai })

(4.111)

for any arbitrary pair of replicas a 6= b (since h sai sai i = 1 for Ising spins). The average
is done with an effective theory of n interacting replicas characterized by Eef f ({sai }).
Again, if there is replica symmetry breaking the actual thermal average is the sum
over all possible pairs of replicas:
X
1
q ab .
n0 n(n 1)

q = lim

(4.112)

a6=b

A similar argument allows one to write

X
1
k
qab
.
n0 n(n 1)

q (k) = [ h si1 . . . sik i2 ] = lim

(4.113)

a6=b

One can also generalize this argument to obtain

X
1
(q q ab )
n0 n(n 1)

P (q) = [ PJ (q) ] = lim

(4.114)

a6=b

Thus, the replica matrix qab can be ascribed to the overlap between pure states.
Note that a small applied field, though uncorrelated with a particular pure state,
is necessary to have non-zero local magnetizations and then non-zero q values.
134

RS

high T

-1

0
q

-1

-m2

p-spin

0
q

m2

m2

Full RSB

-1 -qEA

0
q

qEA

-1

-m2

0
q

Figure 48: [ PJ (q) ] in a paramagnet (top left), in a ferromagnet or a replica symmetric

system (top right), in a model with a one step replica symmetry breaking solution
(bottom left) and for system with full RSB (bottom right).
The function P (q) then extends the concept of order parameter to a function. In
zero field the symmetry with respect to simultaneous reversal of all spins translates
into the fact that PJ (q) must be symmetric with respect to q = 0. [ PJ (q) ] can
be used to distinguish between the droplet picture prediction for finite dimensional
spin-glasses two pure states that simply corresponds to
[ PJ (q) ] =

1
1
(q qea ) + (q + qea )
2
2

(4.115)

(see the central panel in Fig. 48) and a more complicated situation in which [ PJ (q) ]
has the two delta functions at qea plus non-zero values on a finite support (right
panel in Fig. 48) as found in mean-field spin-glass models.
The linear susceptibility
Taking into account the multiplicity of pure states, the magnetic susceptibility,
eq. (4.25), and using (4.99) becomes
Z
X
1 X
2
J J
T = T [ J ] = 1
[ h si i ] = 1
[ w w ] q = dq (1 q) P (q) . (4.116)
N i

135

There are then several possible results for the susceptibiliy depending on the level of
replica symmetry breaking in the system:
In a replica symmetric problem or, equivalently, in the droplet model,
= (1 qea ) .

(4.117)

This is also the susceptibility within a pure state of a system with a higher level
of RSB.
At the one step RSB level, this becomes
= [1 (1 m)qea ] .

(4.118)

For systems with full RSB one needs to know the complete P (q) to compute ,
as in (4.116).
Note that in systems with RSB (one step or full) the susceptibility is larger than
(1 qea ).
A system with qea = 1 in the full low-temperature phase (as the REM model
or p limit of the p spin model, see below) has just one configuration in each
state. Systems with qea < 1 below Tc have states formed by a number of different
configurations that is exponentially large in N . (Note that qea < 1 means that the two
configurations differ in a number of spins that is proportional to N .) The logarithm
of this number is usually called the intra-state entropy.
Even if the number of pure states can be very large (exponential in N ) only a
fraction ofPthem can have a non-negligible weight. This is the case if one finds, for
example, w2 < +
Symmetry and ergodicity breaking

In all p 2 spin models there is a phase transition at a finite Ts at which the rather
abstract replica symmetry is broken. This symmetry breaking is accompanied by
ergodicity breaking as in the usual case. Many pure states appear at low temperatures,
each one has its reversed si si counterpart, but not all of them are related by
real-space symmetry properties.
The one-step RSB scenario
In this case the transition has first-order and second-order aspects. The order
parameters q0 and q1 jump at the critical point as in a first-order transition but the
thermodynamic quantities are continuous.
The full RSB scenario
Right below Tc an exponential in N number of equilibrium states appear. The
transition is continuous, the order parameter approaches zero right below Tc . Lowering further the temperature each ergodic component breaks in many other ones. In
this sense, the full spin-glass phase, T < Tc , is critical and not only the single point
Tc .

136

Tdis
0

Tpure
Tloc

Figure 49: A sketch of critical temperatures.

4.9

Finite dimensional systems

We start now the discussion on the statics of spin-glass models by describing briefly
scaling arguments and the droplet theory. Similar arguments can be used to study
other models with strong disorder, as a manifold in a random potential.
4.9.1

The Griffiths phase

The effects of quenched disorder show up already in the paramagnetic phase of

finite dimensional systems. Below the critical point of the pure case (no disorder)
finite regions of the system can order due to fluctuations in the couplings. Take the
case of random ferromagnetic interactions. Fluctuations in bonds can be such that in
a given region they take higher values than on average. In practice, at the working
temperature T that is higher than the transition temperature of the full system, Tcdis ,
a particular region can behave as if it had have an effective Tcloc that is actually higher
than Tc , see Fig. 49. Similarly, fluctuations can make a region more paramagnetic
than the average if the Jij s take smaller values [ Jij ]. (Note that Tc is typically
proportional to J, the strength of the ferromagnetic couplings. In the disordered case
we normalize the Jij s in such a way that [ Jij ] = Jpure . We can then compare the
disordered and the pure problems.)
These properties manifest in non-analyticities of the free-energy that appear in
a full interval of temperatures above (and below) the critical temperature of the
disordered model, as shown by Griffiths. For instance, deviations from Curie-Weiss
( = 1/T ) behavior appear below the Neel temperature of dilute antiferromagnets
in a uniform field. These are sometimes described with a Lorentzian distribution of
local temperatures with the corresponding Curie-Weiss law at each T . It is clear that
Griffiths effects will also affect the relaxation of disordered systems above freezing.
We will not discuss these features in detail here.
4.9.2

Droplets and domain-wall stiffness

Let us now just discuss one simple argument that is at the basis of what is needed
to derive the results of the droplet theory without entering into the complications of
the calculations.
It is clear the structure of droplets, meaning patches in which the spins point in the
direction of the opposite state, plays an important role in the thermodynamic behavior
of systems undergoing a phase transition. At criticality one observes ordered domains
137

of the two equilibrium states at all length scales with fractal properties. Right above
Tc finite patches of the system are indeed ordered but these do not include a finite
fraction of the spins in the sample and the magnetization density vanishes. However,
these patches are enough to generate non-trivial thermodynamic properties very close
to Tc and the richness of critical phenomena. M. Fisher and others developed a droplet
phenomenological theory for critical phenomena in clean systems. Later D. S. Fisher
and D. Huse extended these arguments to describe the effects of quenched disorder
in spin-glasses and other random systems; this is the so-called droplet model.
Critical droplet in a ferromagnet
Let us study the stability properties of an equilibrium ferromagnetic phase under
an applied external field that tends to destabilize it. If we set T = 0 the free-energy is
just the energy. In the ferromagnetic case the free-energy cost of a spherical droplet
of radius R of the equilibrium phase parallel to the applied field embedded in the
dominant one (see Fig. 50-left) is
F (R) = 2d Rd hmeq + d1 Rd1 0

(4.119)

where 0 is the interfacial free-energy density (the energy cost of the domain wall)
and d is the volume of a d-dimensional unit sphere. We assume here that the droplet
has a regular surface and volume such that they are proportional to Rd1 and Rd ,
respectively. The excess free-energy reaches a maximum

d1
d1
d dd1
Fc =
0d
(4.120)
d dd
2dhmeq
at the critical radius
Rc =

(d 1)d1 0
,
2dd hmeq

(4.121)

see Fig. 50 (h > 0 and m > 0 here, the signs have already been taken into account).
The free-energy difference vanishes at
F (R0 ) = 0

R0 =

d1 0
.
2d hmeq

(4.122)

Several features are to be stressed:

The barrier vanishes in d = 1; indeed, the free-energy is a linear function of R
in this case.
Both Rc and R0 have the same dependence on hmeq : they monotonically
decrease with increasing hmeq vanishing for hmeq and diverging for
hmeq 0.
In dynamic terms that we will discuss later, the passage above the barrier is
done via thermal activation; as soon as the system has reached the height of
the barrier it rolls on the right side of potential F and the favorable phase
nucleates.
138

droplet free-energy

fc

Rc
R

Figure 50: Left: the droplet. Right: the free-energy density f (R) of a spherical
droplet with radius R.
As long as the critical size Rc is not reached the droplet is not favorable and
the system remains positively magnetized.
In this example the field drives the system from one state to the other. In studies of
phase transitions at zero external field, temperature generates fluctuations of different
size and the question is whether these are favourable or not. The study of droplet
fluctuations is useful to establish whether an ordered phase can exist at low (but finite)
temperatures. One then studies the free-energy cost for creating large droplets with
thermal fluctuations that may destabilize the ordered phase, in the way we have done
with the simple Ising chain. Indeed, a fundamental difference between an ordered and
a disordered phase is their stiffness (or rigidity). In an ordered phase the free-energy
cost for changing one part of the system with respect to the other part far away is of
the order kB T and usually diverges as a power law of the system size. In a disordered
phase the information about the reversed part propagates only a finite distance (of
the order of the correlation length, see below) and the stiffness vanishes.
The calculation of the stiffness is usually done as follows. Antiparallel configurations (or more generally the two ground states) are imposed at the opposite boundaries
of the sample. A domain wall is then generated somewhere in the bulk. Its free-energy
cost, i.e. the difference between the free-energies of the modified configuration and
the equilibrium one, is measured and one tests when creating a wall is favourable.
The Imry-Ma argument for the random field Ising model
Take a ferromagnetic Ising model in a random field, defined in eq. (4.20). In zero
applied field and low enough temperature, if d > 1 there is phase transition between a
paramagnetic and a ferromagnetic phase. Under the effect of a random field with very
strong typical strength, the spins align with the local external fields and the system
139

is paramagnetic. It is, however, non-trivial to determine the effect of a relatively

weak random field on the ferromagnetic phase at sufficiently low temperature. The
long-range ferromagnetic order could be preserved or else the field could be enough
to break up the system into large but finite domains of the two ferromagnetic phases.
A qualitative argument to estimate whether the ferromagnetic phase survives or
not in presence of the external random field due to Imry and Ma. Let us fix T = 0
and switch on a random field. If a domain D of the opposite order (say down) is
created within the bulk of the ordered state (say up) the system pays an energy due
to the unsatisfied links lying on the boundary that is
Eborder 2JLd1

(4.123)

where L is the linear length of the border and d 1 is the dimension of the border of
a domain embedded in d a dimensional volume, assuming it is compact. By creating
a domain boundary the system can also gain a magnetic energy in the interior of the
domain due to the external field:
Erf hLd/2

(4.124)

since there are N = Ld spins

domain of linear length L and, using the
P inside the
central limit theorem, h jD si h N = hLd/2 . The comparison between
these two energy scales yields
2
  d2
h
L0
J

d/2

JL0d1 hL0

(4.125)

In the large L limit E diverges to + with increasing L in d > 2. The marginal

case d = 2 is more subtle and we do not discuss it in detail here. One can also search
for an extreme in E(L) finding
Lc

4J(d 1)
hd

2

(4.126)

Several comments are in order:

In d = 1 the energy difference is a monotonically decreasing function of L thus
suggesting that the creation of droplets is very favorable and there is no barrier
to cross to do it.
In d > 2 the energy difference first decreases from E(L = 0) = 0 to reach
a nagative minimum at Lc , and then increases back to pass through zero at
L0 and diverge at infinity. This indicates that the creation of domains at zero
temperature is not favorable in d > 2. Just domains of finite length, upto L0
can be created. Note that L0 increases with h/J in d > 2 and thus a higher
field tends to generate larger droplets and thus disorder the sample.

140

With this argument one cannot show the existence of a phase transition at hc nor the
nature of it. The argument is such that it suggests that order can be supported by
the system at zero temperature and small fields.
An elastic line in a random potential
Let us take an interface model of the type defined in eq. (4.22) with N = 1.
If one assumes that the interfaces makes an excursion of longitudinal length L and
transverse length the leastic energy cost is
Z
c
dd x ((~x))2

Eelast cLd (L1 )2 = cLd2 2 (4.127)

Eelast =
2
If the excursion is sufficiently large, the interface meets Ld /d+1 impurities (that
is to say the volume of the displacement over the typical volume between impurities
given by the correlation length of disorder to the power given by the number
of
p
dimensions). Each impurity applies a pinning force of the order of dV /d W/d
and then the energy gain due to the random potential is
q
(4.128)
Erandom W/d .
The balance between the cost of elastic energy and the gain in random energy leads
to
(L/)(4d)/3
(4.129)

where = (c2 4 /W )1/(4d) is the Larkin length and = (4 d)/3 is the Flory
exponent for the roughness of the surface. One then concludes that for d > 4 disorder
is irrelevant and the interface is flat ( 0 when L ). Since the linearization of
the elastic energy [see the discussion leading to eq. (4.22)] holds only if /L 1, the
result (4.129) may hold only for d > 1 where < 1.
The 3d Edwards-Anderson model in a uniform magnetic field
A very similar reasoning is used to argue that there cannot be spin-glass order
in an Edwards-Anderson model in an external field. The only difference is that the
domain wall energy is here assumed to be proportional to Ly with an a priori unknown
d-dependent exponent y that is related to the geometry of the domains.
Comments
These arguments are easy to implement when one knows the equilibrium states.
They cannot be used in models in which the energy is not a slowly varying function
of the domain wall position.

4.9.3

The droplet theory

The droplet theory is a phenomenological model that assumes that the low temperature phase of a spin-glass model has only two equilibrium states related by an
141

overall spin flip. It is then rather similar to a ferromagnet, only that the nature of
the order in the two equilibrium states is not easy to see, it is not just most spins
pointing up or most spins pointing down with some thermal fluctuations within. At
a glance, one sees a disordered paramagnetic like configuration and a more elaborate
order parameter has to be measured to observe the order. The spin-glass phase is then
called a disguised ferromagnet and a usual spontaneous symmetry breaking (between
the two equilibrium states related spin reversal symmetry) leading to usual ergodicity
breaking is supposed to take place at Tc .
Once this assumption has been done, renormalization group arguments are used to
describe the scaling behavior of several thermodynamic quantities. The results found
are then quantitatively different from the ones for a ferromagnet but no novelties
appear.

142

Formalism: dynamic generating functional and

symmetries

In this Section we discuss some static and dynamic aspects of classical statistical
systems in the canonical ensemble. In this chapter we introduce the classical path
integral formalism. The symmetry arguments follow closely the discussion in [36].

5.1

Classical dynamics: generating functional

In Sect. 2.3 we showed a proof of the (generally non-Markov) Langevin equation

based on the integration over a large ensemble of harmonic oscillators that act as a
bath with which the system is set in contact.
Observables which are functions of the solution to the Langevin equation can also
be computed using a dynamic generating functional that reads [57]
R T R T 1 R T
Z
2k1 T
dt (t ) (t t )(t )+
dt (t )x (t )
dt
B
t0
t0
t0
Zd [] D dP (t0 ) e
x (t) is the solution to the Langevin equation with initial condition x0 = x(t0 ), x 0 =
x(t
0 ) at the initial time t0 . The factor dP (t0 ) is the measure of the initial condition, dP (t0 ) dx0 dx 0 Pi [x0 , x 0 ]. The Gaussian factor is proportional to P [] the
QN
functional probability measure of the noise. The measure is D k=0 d(tk ) with
k = 0, . . . , N , tk = t0 + k(T t t0 )/N and N while (T t0 ) remains finite. The
RT
inverse kernel 1 is defined within the interval [t0 , T ]: t0 dt (t t )1 (t t ) =
(t t ).
A very useful expression for Zd [], usually called the Martin-Siggia-Rose generating functional (actually derived by Janssen [58]), is obtained by introducing the
identity
Z T
dt (t t )x(t
) + V [x(t)] = (t)
(5.1)
Eq[x(t)] m
x(t) +
t0

valid at each time t, with the delta function

Z
1 = Dx [Eq[x(t)] (t)]





det Eq[x(t)]  ,


x(t ) 

(5.2)

QN
with Dx k=1 dx(tk ). The factor |det . . . | is the determinant of the operator
(t t ){mt2 + V [x(t)]} + (t t )t and ensures that the integral equals one.12
R 12 Its

origin isR in the change of variables. In the same way as in the one dimensional integral,
dx [g(x)] = dz 1/|g [g 1 (z)]| (z) = 1/|g [g 1 (0)]|, to get 1 as a result one includes the inverse

of the Jacobian within the integral:

dx [g(x)] |g (x)| = 1.

143

The delta function can be exponentiated with an auxiliary field i

x (using the Fourier
QN 1
representation of the -function). Di
x = k=1 di
x(tk ). The determinant can be
exponentiated with time-dependent anticommunting variables opening the way to
the use of super-symmetry [48], a subject that we will not touch in these notes.
However, since it does not yield a relevant contribution to the kind of dynamics we
are interested in, we forget it (one can show that the determinant is a constant for
Langevin processes with coloured noise and/or inertia and that the discretization of
an over-damped Langevin equation with white-noise can also be chosen to set it to
one It
o convention, see App. ). Zd reads
Z
Zd [, ] DDxDi
x dP (t0 )
h
i
R
R
e

t0

2k1

dt i
x(t ) m
x(t )+

BT

RT

t0

dt

RT

t0

t0

dt (t t )x(t
)+V [x(t )](t )

dt (t )1 (t t )(t )+

RT

t0

(t )i
x(t )]
dt [(t )x(t )+

where we have introduced a new source (t), coupled to the auxiliary field i
x(t). The
integration over the noise (t) is Gaussian and it can be readily done; it yields
kB T
+
2

dt
t0

T
t0

dt i
x(t ) (t t ) i
x(t )

(5.3)

and, for a coloured bath, the environment generates a retarded interaction in the
RT
effective action. In the usual white noise case, this term reduces to, kB T 0 t0 dt [i
x(t )]2 ,
a local expression. In the end, the generating function and resulting Martin-SiggiaRose-Jaenssen-deDominicis (MSRJD) action reads
Z
Zd [, ] DxDi
x dP (t0 ) eS[x,ix,,]


Z
Z
S[x, i
x, , ] = dt i
x(t ) m
x(t ) + dt (t t )x(t
) + V [x(t )]
Z
Z
kB T
dt dt i
x(t )(t t )i
x(t ) + sources .
(5.4)
+
2
All integrals runs over [t0 , T ]. Causality in the integral over t is ensured by the fact
that is proportional to .
The MSRJD action has two kinds of contributions: the ones that depend on the
characteristics of the bath (through ) and the ones that do not. The latter also
exist in a functional representation of Newton dynamics and we call them S det (for
deterministic) while the former contain all information abour thermal fluctuations
and dissipation and we call it S diss (for dissipation):
S[x, i
x, , i
] = S diss [x, i
x; ] + S det [x, i
x, , i
] .

144

(5.5)

If the distribution of the initial condition were to be included in the action as an

additional term, ln Pi [x0 , i
x0 ], 6t would be part of S det .
Interestingly enough, the dynamic generating functional at zero sources is
identical to one for any model:
Zd [ = 0, = 0] = 1

(5.6)

as can be concluded from its very first definition. In particular, it does not depend
on the coupling constants of the chosen model. This property will be utilized in
disordered systems to render the dynamic calculations relatively easier than the static
ones.

5.2

Generic correlation and response.

The mean value at time t of a generic observable A is

Z
hA(t)i = DxDi
x dP (t0 ) A[x(t)] eS[x,ix] ,

(5.7)

where S[x, i
x] is ashort-hand notation for S[x, i
x, = 0, = 0]. The self-correlation
and linear response of x are given by


2 Zd [, ] 
2 Zd [, ] 
1

=
(5.8)
C(t, t ) = hx(t)x(t )i =
Zd [, ] (t)(t ) ==0
(t)(t ) ==0

hx(t)i 
S[x, i
x; h]
R(t, t ) =
i|h=0 = hx(t)i
x(t )i
= hx(t)
h(t ) h=0
h(t )


2 Zd [, ] 
2 Zd [, ] 
1
=
(5.9)
=
Zd [, ] (t) (t ) ==0
(t) (t ) ==0

with h(t ) a small field applied at time t that modifies the potential energy according
to V V h(t )x(t ). The i
x auxiliary function is sometimes called the response
field since it allows one to compute the linear response by taking its correlations with
x. Similarly, we define the two-time correlation between two generic observables A
and B,
Z
CAB (t, t ) DxDi
xdP (t0 ) A[x(t)]B[x(t )] eS[x,ix] = hA[x(t)]B[x(t )]i
(5.10)
and the linear response of A at time t to an infinitesimal perturbation linearly applied
to B at time t < t,

hA[x(t)]ifB 
,
(5.11)
RAB (t, t )
fB (t ) fB =0

with V (x) 7 V (x) fB B(x). The function B(x) depends only on x (or on an even
number of time derivatives, that is to say, it is even with respect to t t). By
145

plugging eq. (5.7) in this definition we get the classical Kubo formula for generic
observables:

S[x, i
x; fB ] 
RAB (t, t ) = hA[x(t)]
i
= hA[x(t)]i
x(t )B [x(t )]i
(5.12)

fB (t )
fB =0

with B [x(t )] = x B[x(t )]. This relation is also causal and hence proportional to
(tt ); it is valid in and out of equilibrium. For B[x] = x it reduces to the correlation
between x and i
x.
If the system has quenched random exchanges or any kind of disorder, one
may be interested in calculating the averaged correlations and responses over different
realizations of disorder. Surprisingly, this average is very easy to perform in a dynamic
calculation [59]. The normalization factors 1/Zd [, ] in (5.8) and (5.9) have to be
evaluated at zero external sources in which they are trivially independent of the
random interactions. Hence, it is sufficient to average Zd [, ] over disorder and
then take variations with respect to the sources to derive the thermal and disorder
averaged two-point functions. This property contrasts with
P an equilibrium calculation
where the expectation values are given by [hAi] = [1/Z conf A exp(H)], with []
denoting the disorder average. In this case, the partition function Z depends explicitly
on the random exchanges and one has to introduce replicas [49] to deal with the
normalization factor and do the averaging.
Having assumed the initial equilibration of the environment ensures that a normal
system will eventually equilibrate with it. The interaction with the bath allows the
system to dissipate its energy and to relax until thermalization is reached. However,
in some interesting cases, as the dyamics across phase transitions and glassy models,
the time needed to equilibrate is a fast growing function of N , the number of dynamic degrees of freedom. Thus, the evolution of the system in the thermodynamic
limit occurs out of equilibrium. In real systems, a similar situation occurs when the
equilibration time crosses over the observation time and falls out of the experimental
time-window.
A final interesting remark on the relevance of quenched disorder is the following.
When a system with quenched disorder evolves out of equilibrium at finite temperature, the correlation function and the response function do not depend on the realization of disorder if the size of the system is large enough (the realization of disorder
has to be a typical one). These quantities are self-averaging. This statement is
easily checked in a simulation. When times get sufficiently long as to start seeing the
approach to equilibrium, dependencies on the realizations of disorder appear.
5.2.1

The linear response as a variable-noise correlation

The correlation between coordinate and a generic colored noise can be obtained
from the variation with respect to (t, t ) of the generating functional once the source
Z
dt dt (t , t ) x(t )(t )
(5.13)
146

has been added. Integrating over the noise and keeping only the linear terms in in
the effective action since all others will vanish when setting = 0
Z
kB T
Linear terms =
dt1 dt2 dt3 dt4 [(t1 , t2 )x(t1 )(t2 , t3 )i
x(t4 )(t4 t3 )
2
+i
x(t1 )(t1 t2 )(t2 , t3 )(t4 , t3 )x(t4 )] .
(5.14)
R
The variation with respect to (t, t ) yields (kB T )/2 dt [(t , t ) + (t , t )]
hx(t)i
x(t )i = hx(t)(t )i.

5.3

Time-reversal

Since it will be used in the rest of this chapter, we introduce the time-reversed
variable x
by x
(t) x(t) for all t. The time-reversed observable is defined as
Ar ([x], t) A([
x], t).

(5.15)

It has the effect of changing the sign of all odd time-derivatives in the expression of
local observables, e.g. if A[x(t)] = t x(t) then Ar [x(t)] = t x(t). As an example
for non-local observables, the time-reversed Langevin equation reads
Eqr ([x], t)

m
x(t) Fr ([x], t)

T
T

du (u t)x(u)

(5.16)

Notice the change of sign in front of the friction term that is no longer dissipative in
this new equation.

5.4

An equilibrium symmetry

If the initial time t0 is set to t0 = T and the system is in equilibrium at this

instant, PT is given by the Maxwell-Boltzmann measure. One can then check that
the zero-source action, S[x, i
x], is fully invariant under the field transformation Tc
defined as

xu 7 xu ,
Tc
i
xu 7 i
xu + du xu .
We introduced the notation xt = x(t) so as to make the notation more compact. Note
that du xu = du xu . This transformation does not involve the kernel and it includes a time-reversal. The invariance is achieved independently by the deterministic
(S det ) and dissipative (S diss ) terms in the action. The former includes the contribution from the initial condition, ln PT . Moreover, the path-integral measure is
invariant since the Jacobian associated to this transformation is also block triangular
with ones on the diagonal. The proof goes as follows.
5.4.1

Invariance of the measure

147

The Jacobian Jc of the transformation Tc is the determinant of a triangular matrix:

"
#
xu

2
0
(x, x
)
1
xv
Jc det
= detuv
=1
= det1
x
u
x
u
uv [u+v ]
(Tc x, Tc x
)
xv
x
v

and it is thus identical to one.

5.4.2

Invariance of the integration domain

Before and after the transformation, the functional integration on the field x is
performed for values of xt on the real axis. However, the new domain of integration
for the field x
is complex. For all times, x
t is now integrated over the complex line
with a constant imaginary part it xt . One can return to an integration over the
real axis by closing the contour at both infinities. Indeed the integrand, eS , goes
to zero sufficiently fast at xt for neglecting the vertical ends of the contour
thanks to the term 1 0 (i
xt )2 (in the white noise limit or the correspondong ones in
colored noise cases) in the action. Furthermore the new field is also integrated with
the boundary conditions x
(T ) = x
(T ) = 0.
5.4.3

Invariance of the action functional

The deterministic contribution satisfies

Z


S det [Tc x, Tc x
] = ln Pi (xT , x T ) [i
xu + u xu ] mu2 xu + V (xu )
u
Z
Z

xu + V (xu )]
xu [m
xu + V (xu )] + x u [m
= ln Pi (xT , x T ) i
u
u
Z
Z
= ln Pi (xT , x T ) i
xu [m
xu + V (xu )] + u ln Pi (xu , x u )
u

=S

det

[x, x
] ,

where we used the initial equilibrium measure ln Pi (x, x)
= 21 mx 2 + V (x) ln Z.
In the first line we just applied the transformation, in the second line we made the
substitution u 7 u, in the third line we wrote the last integrand as a total derivative
the integral of which cancels the first term and recreates the initial measure at T .
Secondly, we show that the dissipative contribution is also invariant under Tc . We
have
Z
Z
S diss [Tc x, Tc x
] =
[i
xu + u xu ] 1 uv i
xv
u
v
Z
Z
xv
[i
xu x u ] vu 1 i
=
v

S diss [x, x
] .
148

In the first line we just applied the transformation, in the second line we made the
substitution u 7 u and v 7 v and in the last step we exchanged u and v.
5.4.4

Invariance of the Jacobian (Grassmann variables)

Finally we show that the Jacobian term in the action is invariant once it is expressed in terms of a Gaussian integral over conjugate Grassmann fields (c and c )
and provided that the transformation Tc is extended to act on these as follows13

xu 7 xu ,
cu 7 cu ,
Tc
(5.17)
i
xu 7 i
xu + u xu ,
cu 7 cu .
We start from
J

S [c, c , x] =

Z Z
u

cu

and we have

mu2 uv

+ u uv cv +

cu V (xu )cu

(5.18)

S J (Tc c, Tc c , Tc x)
Z
Z Z
 2

cu mu uv + u uv cv + cu [V (xu )] cu
=
u
u v
Z Z
Z
 2


=
cv mu vu u vu cu + cu V (xu )cu
u

= S J (c, c , x) .

In the first line we just applied the transformation, in the second line we exchanged
the anti-commuting Grassmann variables and made the substitutions u 7 u and
v 7 v, finally in the last step we used v vu = v uv and we exchanged u and
v. Finally the set of boundary conditions [ c(T ) = c(T

) = c (T ) = c (T )] is left
invariant.

5.5

Consequences of the transformation

We now use the transformation Tc to derive a number of exact results.

5.5.1

The fluctuation-dissipation theorem

This symmetry implies

hxt i
xt iS[x,ix]
13 More

=
=

hTc xt Tc i
xt iS[Tc x,Tc ix]
hxt i
xt iS[x,ix] + hxt dt xt iS[x,ix]

(5.19)

generally, the transformation on c and c is cu 7 cu and cu 7 1 cu with C .

149

that, using time-translational invariance and t t , becomes

R( ) R( )

d C( ) = d C( ) .

(5.20)

For generic observables one can similarly apply the Tc transformation to expression (5.12) of the linear response
RAB ( ) RAr Br ( )

d CAB ( ) = d CAB ( ) .

(5.21)

where we defined Ar and Br as

Ar ([x], t) A([
x], t) .
(5.22)
R
R
Take
for instance a function A[x(t), t] = duf (x(u))(u t) +

t) +
R
R duf (x(u))(u
duf
(
x
(u))(u

t)
+
.
.
.
then
A
[x(t),
t]
=
A[x(t),
t]
=
duf
(x(u))(u
+
t) +
r
R
R
duf (x(u))(u

+ t) + duf (
x(u))(u + t) + . . ..
Relations between higher order correlation functions evaluated at different times
t1 , t2 , . . . tn are easy to obtain within this formalism.
5.5.2

Fluctuation theorems

Let us assume that the system is initially prepared in thermal equilibrium with
respect to the potential V (x, T )14 . The expression for the deterministic part of the
MSRJD action functional is
S det [x, x
; , f]

H([xT ], T ) ln Z(T )
Z
i
xu [m
xu + V (xu , u ) fu [x]] ,
u

where H([xt ], t ) 21 mx 2t + V (xt , t ) and f is a non-conservative force applied on

the system. The external work done on the system along a given trajectory between
times T and T is given by
Z
Z u
Z u
Z u
T
T
T
u u V + x u x V
du V =
dE =
W [x; , f]
=

uT
uT

uT

uT

uT

uTT u V (xu , u ) +

uT

x u fu [x]

(5.23)

uT

where we take into account the time variation of the parameter .

Fluctuation Theorem 1.
14 This is in fact a restriction on the initial velocities, x
T , that are to be taken from the Maxwell
distribution with temperature 1 , independently of the positions xT . These latter can be chosen
from a generic distribution since the initial potential can be tailored at will through the dependence
of V .

150

The transformation Tc does not leave S det invariant but yields

Tc
fr ] F W [x; ,
fr ]
S det [x, x
; ,
S det [x, x
; , f] 7

(5.24)

fr ] is the MSRJD action of the system that is prepared (in equiwhere S det [x, x
; ,

librium) and evolves under the time-reversed protocol ((u)

(u)) and external
forces (fr ([x], u) f([
x], u)). F is the change in free energy: F = ln Z((T ))
ln Z((T )) between the initial and the final virtual equilibrium states. W is defined
above. The dissipative part of the action, S diss does not involve and it is still invariant under Tc . This means that, contrary to the external forces, the interaction with
the bath is not time-reversed: the friction is still dissipative after the transformation.
This immediately yields

hA[x, x
]iSc [x,x;,f] = eF hA[Tc x, Tc x
]eW [x;,fr ] iSc [x,x;,f
r]

(5.25)

for any functional A of x and x

. In particular for a local functional of the field,
A[x(t)], it leads to the Crooks relation

hA[x(t)]iSc [x,x;,f] = eF hAr [x(t)]eW [x;,fr ] iSc [x,x;,f

r] ,

(5.26)

or also
hA[x(t)]B[x(t )]iSc [x,x;,f]

= eF hAr [x(t)]Br [x(t )]eW [x;,fr ] iSc [x,x;,f

r].

(5.27)

Setting A[x, x
] = 1, we obtain the Jarzynski equality
1 = eF heW [x;,f] iSc [x,x;,f] .

(5.28)

Setting A[x, x
] = (W W [x; , f]) we obtain the transient fluctuation theorem
P (W ) = Pr (W ) e(W F ) ,

(5.29)

where P (W ) is the probability for the external work done between T and T to be
W given the protocol (t) and the non-conservative force f([x], t). Pr (W ) is the same
and time-reversed force fr .
probability, given the time-reversed protocol
Fluctuation Theorem 2.
The result we prove in the following lines is not restricted to Langevin processes
with an equilibrium dissipative bath. It applies to generic classical equations of
motion, with or without stochastic noise. In short, the proof consists in applying
time-reversal on the system and yields an equality between observables and their
time-reversed counterparts in a so-called backward (B) process in which the system
is prepared in equilibrium with respect to the final conditions of the forward process

151

and is evolved according to the time-reversed equations of motions and protocol. Let
us rewrite the action as
Z
Sc [x, x
, ] = H(xT , x T , T ) i
xu Eq([xu ], u )
u
Z Z
1
+
i
xu 1 uv i
xv ln Z(T ) ,
2 u v
and apply the following time-reversal of the fields

xu 7 xu ,
Ttr
i
xu 7 i
xu .

(5.30)

This yields
Sc [x, x
, ]

Z
T ) i
u)
H([xT ],
xu Eqr ([xu ],
u
Z Z
1
i
xu 1 uv i
xv ln Z(T )
+
2 u v

or, by introducing zeroes:

Z
T ) i
u)
Wr H([xT ],
xu Eqr ([xu ],
u
Z Z
1
T ) ,
+
i
xu 1 uv i
xv F ln Z(
2 u v

(5.31)

where F F(T ) F(T ) is the free-energy difference between the two virtual
T ) H([xT ],
T )
equilibrium states corresponding to T and T . Wr H([xT ],
is the work applied on the system that evolves with the time-reversed equation of
In particular and contrary to
motion Eqr and under the time-reversed protocol .
the previous paragraph, the friction is no longer dissipative after the transformation.
This defines the backward (B) process. Finally, for any observable A[x, x
] we get the
relation
]eWr iB .
hA[x, x
]iF = eF hA[
x, x

(5.32)

In particular, for two-time correlations, it reads

hA[x(t)]B[x(t )]iF = eF hAr [x(t)]Br [x(t )]eWr iB .

(5.33)

Setting A[x, x
] = (W W [x; , f ]) we obtain the transient fluctuation theorem
PF (W ) = PB (W ) e(W F ) ,

(5.34)

where PF (W ) is the probability for the external work done between T and T to be
W in the forward process. PB (W ) is the same probability in the backward process.
152

5.6

Equations on correlations and linear responses

Take any Langevin process in the MSRJD path-integral formalism. From the
following four identities
 


 


x(t)
x(t)
i
x(t)
i
x(t)

=
=
(t

t
)
,
=
=0,
(5.35)
i
x(t )
x(t )
i
x(t )
x(t )
where the angular brackets indicate an average with the MSRJD weight, after an
integration by parts, one derives four equations




S
S

x(t)
= (t t ) ,
i
x(t)
= (t t ) ,
(5.36)
x(t )
i
x(t )




S
S
x(t)
=0,
i
x(t)
=0.
(5.37)

i
x(t )
x(t )
The second and third one read



Z
i
x(t) m
x(t ) + dt (t t ) x(t
) + V [x(t )]
Z
+kB T
dt (t t ) hi
x(t)i
x(t )i = (t t ) ,



Z

x(t) m
x(t ) + dt (t t ) x(t
) + V [x(t )]
Z
+kB T
dt (t t ) hx(t)i
x(t )i = 0 ,

(5.38)

while the other ones, once causality is used (basically hx(t )i

x(t)i = 0 for t > t and

hi
x(t)i
x(t )i = 0) do not yield further information. All terms are easily identified with
the four types of two-time correlations apart from the ones that involve the potential
and are not necessarily quadratic in the fields. The linear terms in two-time functions
can be put together after identifying the free-operator

G1
0 (t , t ) = (t t )m

d2

2 + (t t ) t

dt

(5.39)

The non-linear terms can be approximated in a number of ways: perturbation theory

in a small parameter, Gaussian approximation of the MSRJD action, self-consistent
approximations, etc. The choice should be dictated by some knowledge on the systems behavior one wishes to reproduce. In short then
Z
Z
1

0 = dt G0 (t , t )C(t , t) + hx(t)V [x(t )]i + kB T dt (t t )R(t, t ) ,

Z

x(t)V [x(t )]i .

(5.40)
(t t ) = dt G1
0 (t , t )R(t , t) + hi
153

Dynamic equations

The generating functionals, with their effective actions, are the adequate starting
point to apply perturbation theory (when it is accepted), self-consistent approximations such as the mode-coupling approach, or even more sophisticated techniques as
the functional renormalization group. In this section we discuss a number of properties of the generating functional and the ensuing dynamic equations, that we derive
in a different ways.

6.1

Connection with the replica formalism

The effective action in a supersymmetric formulation of the generating function has

a kinetic minus a potential term V []. When applying the replica trick to compute the
free-energy a replicated effective potential V [a ] appears. A connection between the
two formalism, that is based on the similarity between the zero-dimensional replica
space and the susy one, has been P
exploited. Roughly speaking, many properties
N
of the replica overlap Qab N 1 i=1 hsai sbi i finds a counteraprt in the
Pndynamic
susy correlator Q(a, b). For instance, a summation over a replica
index,
a=1 when
R
n 0, translates into an integration over the supercoordinate da. For the moment,
though, the connection is empirical and a formalization of the relation between the
two approaches would be welcome.

6.2

Average over disorder

In general one is interested in the evolution of a model in which the configuration

of disorder is typical. One could either attempt to solve the dynamics for one such
disorder realization or one can assume that the behavior of a typical system is described by the averaged behavior over all systems, each weighted with its probability.
Since the former procedure is more difficult than the latter one usually studies the
dynamics averaged over disorder and computes:
[hA(t)i] =

eSeff [,i]
dJP (J) DDi A[, i]
.
R
R
R

dJP (J) D Di eSeff [,i]

(6.1)

J represents here the random exchanges. Similarly, one can perform an average over
a random potential.
One of the advantages of using a dynamic formalism is that when the initial
conditions are uncorrelated with disorder there is no need to use the replica trick to
average over disorder [59]. Indeed, the classical generating functional is constructed
from a path integral that is identical to 1 (and hence independent of disorder) in the
absence of sources. The same holds for the quantum Schwinger-Keldysh generating

154

functional, Tr
red (0) = 1, since we have chosen a diagonal density matrix as the
initial condition for the system. Thus,
Z
Z

eSeff [,i]
[hA(t)i] = dJP (J) DDi A[, i]
(6.2)
and these averages can be simply computed from [ZJ ].
If the initial condition is correlated with the random exchanges or the random
potential, the situation is different. One such example is the study of the equilibrium
dynamics of a disordered model, i.e. the study of the evolution of initial conditions
taken from Pgb . In this case, the use of replicas to average ln Z is unavoidable and
one is forced to treat replicated dynamic correlators. The initial density operator is
a Boltzmann factor that is represented with the Matsubara formalism while the realtime dynamics is written with the Schwinger-Keldysh approach. Mixed correlators
and responses intervene in the dynamic equations.

6.3

The equations

We present three derivations of the dynamic equations for the macroscopic order parameters that use the classical or quantum dynamic generating functionals as
starting points. Each method is better adapted for different kinds of models.
6.3.1

Supersymmetry and saddle-points

In the white noise limit Z can be written in a much more compact form if one introduces the super-field formulation of stochastic processes. One first enlarges (space)time to include two Grassmann coordinates and , i.e. t a = (t, , ). The
dynamic variable x(t) and the auxiliary variable i
x(t) together with the fermionic
ones (t) and (t), used to express the Jacobian, are encoded in a super-field,
(a) = x(t) + (t) + (t) + i
x(t) .

(6.3)

With these definitions,


 Z
Z
Z
Z
1
da (a)Da(2) (a) da V [(a)] + da(a)(a)
Z[] = d exp
2
(2)

with a = (t, , ), da = dtdd, and the dynamic operator Da

Da(2) = 2kB T

defined as

3
2
2
M
.
+ 2
t
t
t2

(2)

(2)

(6.4)

If the model is spherically constrained, Da Da (a) with (a) a super

Lagrange multiplier introduced to enforce
R the constraint. The delta function (a b)
is defined in Appendix and it satisfies db(a b)f (b) = f (a). The super-symmetric
notation allows one to encode in the single super correlator Q(a, b) h(a)(b)i all
155

correlators and responses. The generalization to a system with N degrees of freedom

is immediate.
Even though the use of the susy notation is not necessary to derive the dynamic
equations, it is very useful in several aspects. Firstly, it allows to establish contact
with the replicated version of the static partition function and the further study of
this quantity; secondly, it is very useful as a bookkeeping tool; thirdly, it allows us to
develop more sophisticated techniques amenable to derive the dynamic equations of
models without fully connected interactions.
Since for classical models the use of white noises is rather generally justified we
will stick to this case. Moreover, we will drop the inertial contribution to further sim~ = (1 , . . . , N ).
plify the presentation. We analyze here models with N variables
In susy notation we then set the sources to zero and we add a Lagrange multiplier
(2)
(2)
Da into Da s (a) with a a super-field to impose the spherical constraint
(s (a) = (t) + fermionic +
(t),
(t) is a Lagrange multiplier that fixes the measure of integration and (t) enters the Langevin equation). Soft spins with their
corresponding potential energy can be studied in a similar way though their treatment is slightly more complicated. The potential energy of a rather generic fully
connected disordered model can be expressed as a series expansion of the form
V [] = g

Fr

Ji1 ...ir+1 i1 . . . ir+1 ,

(6.5)

i1 <...<ir+1

r0

For each r the sum is taken over all possible groups of r +1 spins. The fully-connected
character of the model implies that there is no notion of distance or geometry. Ji1 ...ir+1
are random interactions taken from a Gaussian distribution with zero mean and variance [Ji21 ...ir+1 ] = (r + 1)!/(2N r ). Thus (6.5) is a Gaussian random potential with
~
~
[V ((a))V
((b))]
= N g2

X
r0

Fr2

~
~
(a)
(b)
N

!r+1

=NV

~
~
(a)
(b)
N

. (6.6)

P
~
~
The scalar product in the second member is defined as (a)
(b)
= i i (a)i (b).
The bullet means that the powers are taken locally in the super-coordinates a and b
and they do not involve an operational product. The term r = 0 corresponds to a
random field linearly coupled to the spin, the term r = 1 is quadratic in the fields
while for r 2 we obtain higher order interactions. If Fr = Fp 6= 0, p 2 and all
other Fr = 0 one recovers a spherical p spin model. If two parameters are non-zero
one obtains a model with two p spin terms. The model of a particle in an infinite
dimensional spherical random environment correlated also falls in this category if one
can expand the correlator in a power series.
The disordered averaged generating functional reads


R
R
P
Z
~
~
(a)
(b)
(2)
dadb V
i (a) (Da
s (a)) i (a)+ N
da 12
2
N
i
.
(6.7)
[Z] = D e
156

PN
Introducing the order parameter Q(a, b) = N 1 i=1 i (a)i (b) through
Z


R
PN
1

Q(a,b)
i (a)i (b)
e 2 da db N iQ(a,b)Q(a,b)i
i=1
1 DQDiQ

(6.8)

yields
[Z] =


Z

exp 1
b)Q(a, b) N V (Q(a, b))
DDQDiQ
dadb N iQ(a,
2
!#
N
X
(2)
b)) i (b)
i (a) (D s (a)) (a b) iQ(a,
.
(6.9)

i=1

(Again we omit irrelevant normalization constants.) Note that all terms in the exponent are order N if the integrals yield finite contributions. We call the models for
which this is true mean-field since the saddle-point evaluation of the integral when
N is exact without including fluctuations.
The saddle-point values for the Landau fields Q are simply related to correlations
of the original spins. Indeed, evaluating the generating function in Eq. (6.9) with a
saddle-point approximation

N
X
S 
hi (a)i (b)iZ[Q]
,

N
Q
(a,
b)
=
0=

sp
b) 
iQ(a,
Qsp
i=1

(6.10)

where the average on the rhs is taken with the generating functional
Z
 PN

R
(2)
dadb 12
i (a)(Da
s (a))(ab)i (b)+N V (Q(a,b))

i=1
Z[Q]

DDQ e
.
Opening up the susy notation Eq. (6.10) implies, as expected,
N Csp (t1 , t2 ) =

N
X
i=1

N Rsp (t1 , t2 ) =

sp (t1 , t2 ) =
hqi (t1 )qi (t2 )iZ[Q]
, NQ

N
X
i=1

N
X
i=1

hqi (t1 )i
qi (t2 )iZ[Q]
, N Rsp
(t1 , t2 ) =

hi
qi (t1 )i
qi (t2 )iZ[Q]
,

N
X
hi
qi (t1 )qi (t2 )iZ[Q]
.

i=1

b),
Q (Da(2) (t)) (ab)iQ(a,
Going back to Eq. (6.9) we can now shift iQ,
and integrate over i
Z
R
(2)

N
[Z] = DQDQ e 2 dadb [Q(a,b)Q(a,b)+(Da (t))(ab) Q(a,b)V (Q(a,b))]
e 2 Tr Ln Q .
N

157

Using
a saddle-point evaluation, we eliminate Q, and we obtain [Z]
R
DQ exp[N Seff (Q)] with
Z
h
i
2Seff (Q) =
dadb [Da(2) s (a)](a b) Q(a, b) V (Q(a, b))
TrLn Q .

(6.11)

The saddle-point equation over Q, Seff /Q = 0, yields the dynamic equation

(Da(2) + (t))(a b) + Q1 (a, b) + V (Q(a, b)) = 0 ,
that takes a more convenient form after multiplying operationally by Q:
Z
(Da(2) + (t))Q(a, b) + (a b) + da (a, a )Q(a , b) = 0 ,

(6.12)

(6.13)

with the self-energy defined as

(a, b) V (Q(a, b)) = g 2

Fr2 (r + 1)Q(a, b) r .

(6.14)

r0

We have recasted the saddle-point dynamic equation in the form of a Schwinger-Dyson

equation. The dynamic field is here a susy correlator that encodes the usual correlation function, the advance and retarded linear responses and the fourth correlator
(that vanishes for causal problems):
Z
1
t ) + dt [(t, t )R(t , t )
Go (t)R(t, t ) = (t t ) + 2 Q(t,
t, t )] ,
+D(t, t )Q(t

1
Go (t)C(t, t )

= 2kB T R(t , t) +

dt (t, t )C(t , t )

dt + D(t, t )R(t , t ) ,
Z
1

Go (t)R (t, t ) = (t t ) +
dt (t , t)R(t , t )
0
Z
t )C(t , t ) + 2
+ dt (t,
(t)C(t, t ) ,
Z
Z
1
t ) = dt (t, t )Q(t
, t ) + dt (t,
t )R(t , t )
Go (t)Q(t,
+

+2
(t)R(t, t ) ,

with Go (t) M t2 + t + (t), Go (t) M t2 t + (t), (t, t ) = (t , t) and

X
(t, t ) = g 2
Fr2 (r + 1)r C r1 (t, t )R(t, t )
(6.15)
r0

158

D(t, t )

g2

Fr2 (r + 1) C r (t, t )

(6.16)

t ) .
Fr2 (r + 1)r C r1 (t, t )Q(t,

(6.17)

r0

t )
(t,

g2

X
r0

We set to zero all fermionic correlators. We call the above integro-differential equa respectively.
tions the Schwinger-Dyson equations for R, C R and Q,
Causality can be used to simplify the four Schwinger-Dyson equations considerably. For t > t one has R(t, t ) = 0 while for t > t one has R(t , t) = 0. Rewriting the
t ) = 0
equations for R and R with these two choices of times one easily sees that Q(t,

is symmetric in t and t ) and 0 (t) = 0 for all t. Thus,

for all t and t (note that Q
vanishes identically when causality holds. In the following we
the equation for Q
search for causal solutions and we work with their simplified version. We loose in
this way the possibility of finding solutions that break causality which are related to
instantons. We will come back to this point later. If we focus on the case t > t the
dynamic equations x(
1

Go (t)R(t, t ) =
1
Go (t)C(t, t )

dt (t, t )R(t , t ) ,

(6.18)

dt D(t, t )R(t , t ) +
0

dt (t, t )C(t , t ) . (6.19)

In their integrated form they read

R(t, t ) = Go (t, t ) +
C(t, t ) =

t
0

dt

dt
t

dt Go (t, t ) (t , t ) R(t , t ) , (6.20)

dt R(t, t ) D(t , t ) R(t , t ) ,

(6.21)

0
1

with the propagator given by G1

o (t, t ) (t t )Go (t).
The equation for (t) can be derived from the Schwinger-Dyson equation by imposing the spherical constraint through the evaluation at t = t . Multiplying operationally by G1
o one obtains

(t)

dt [(t, t )C(t, t ) + D(t, t )R(t, t )]

0
Z t
Z t
dt (t R(t, t )) D(t , t ) (t R(t, t ))
(6.22)
dt
+M
0
0


2
2
+ M 2 t R(t, s)st
C(s, t) st
R(t, s)t C(s, t ) s0 tt .

The last two terms are a consequence of having a kinetic term with second derivatives.
It can be easily identified with minus the second-derivative of the correlation at equal

159

times by taking the limit t t in Eq. (6.19). Thus


Z t

2


dt [(t, t )C(t, t ) + D(t, t )R(t, t )] M 2 C(t, t )
(t) =
.
t
0
t t

(6.23)

One way of deriving the equation for (t) for a Langevin process with white noise
and no inertia goes as follows. Considering t > t in the complete Schwinger-Dyson
equation for C and taking t t , and considering t < t in the same equation and
taking t t+ , one finds
lim t C(t, t ) = lim+ t C(t, t ) 2kB T

t t

t t

(6.24)

where we used R(t, t t ) = 1/. The derivative of C has a cusp at t =

t . The symmetry of the correlation function about t = t implies C(t + , t ) =
C(t , t ) and an expansion up to first order in implies limt t t C(t, t ) =
limt t+ t C(t , t). From Eq. (6.24) one has limt t t C(t, t ) = kB T . Now,
one rewrites the complete equation for C exchanging t and t and adds this equation to the same equation in the limit t t : limt t [t C(t, t )+ t C(t, t )] =
2(t) + limt t [rhs eq. for C+rhs eq for C(t t)]. From the discussion above the
lhs vanishes and the rhs implies
Z
dt [(t, t )C(t, t ) + D(t, t )R(t, t )] .
(6.25)
(t) = kB T +
0

For the spherical p spin model (t) is simply related to the energy density E(t).
Indeed, take the Langevin equation evaluated at time t, multiply it by si (t ), sum over
all sites, average over the noise and take the limit t t. Repeat this procedure with
the Langevin equation evaluated at t and multiplying by si (t). Adding the resulting
PN

equations and using N 1

DP i=1 hsi (t)i (t )iE = 2kB T R(t, t ) (see Appendix 5.2.1) we
HJ (~
s(t))

have (t) = limt t

i
si (t) si (t ) + kB T that for the spherical p spin model
becomes
(t) = pE(t) + kB T .

(6.26)

Thanks to the mean-field character of the model the action is proportional to

N and the saddle-point evaluation is exact when N . For the fully connected
models considered in this Section the self-energy is given by a rather simple function of
the interactions. For finite dimensional problems none of these procedures are exact.
An effective action in terms of local order parameters Qi (a, b) can be written but the
evaluation of the generating functional by saddle-point has to include fluctuations.

6.3.2

Field equations

160

Once we have written the dynamic action in terms of i and ii the field equations follow from exact properties of the functional integration. Indeed,
Z
R
i (t ))
]+

i (t )i
dt (t )i (t )+
S [ ,i
0 = DDi
e eff i i C ( i
ii (t)
#
"
Z
R
i )
i (t ))
)+
S
(
,
i

i (t )i
dt (t )i (t )+
S ( ,i
eff
i
.
+ i (t) e eff i i C ( i
= DDi
ii (t)
The subindex C in the integrals stands for time contour and it can describe the
usual integration from the initial time to infinity for classical models or the close time
path for quantum ones. Taking now the variation with respect to the source i
j (t )
and evaluating at = i
= 0 for all times and components we find
+
*

Seff (i , ii )

(6.27)
0 = (t t )ij ij (t )
ii (t)
where the brackets denote an average with the measure weighted by the dynamic
action Seff . If, instead one takes the variation with respect to j (t ) and later evaluates
at = i
= 0 one obtains:


S
ii (t)
=0.
(6.28)
j (t )

A way to derive dynamic equations for the two-point correlators amounts to use
Wicks theorem and rewrite these averages as a sum over all possible factorizations in
products of two point-functions. This is of course exact if the action is quadratic but it
is only a Gaussian approximation for more general models. This kind of derivation has
been mainly used in the study of the dynamics of manifolds in random potentials [41].
6.3.3

The thermodynamic limit and time-scales

It is very important to stress that the dynamic equations derived with the saddlepoint approximation hold only when N before any long-time limit is taken.
They describe the dynamics in finite time-scales with respect to N and they cannot
capture the crossover from the non-equilibrium relaxation to the equilibrium dynamics
reached in time scales that diverge with N [remember that teq (N )].
Old attempts to study the dynamics of disordered glassy systems assumed that
these same equations hold for the equilibrium dynamics when N is finite and timescales diverge with N . This assumption is wrong as shown by several inconsistencies
found in the solution at low temperatures: (i) the asymptotic values of one timequantities do not necessarily coincide with the values calculated with the equilibrium
distribution. (ii) the solution exhibited violates the fluctuation - dissipation theorem.
These two results are not compatible with equilibrium.
In order to study the equilibrium dynamics of these models one should (i) start
from random initial conditions but reach times that grow with N or (ii) impose
equilibrium initial conditions. The second route has been implemented though
161

without solving the full dynamic problem by Houghton, Jain and Young. They
showed that in this case one is forced to introduce the replica trick to average over
disorder.
The dynamic equations here derived are correct when N at the outset. Since
times are always finite with respect to N , when teq (N ) diverges with N the dynamics
is not forced to reach equilibrium and there is no contradiction if the solution violates
the equilibrium theorems.
6.3.4

Single spin equation

In the limit N one can also write the full action Seff in terms of a single
variable. This is at the expense of modifying the thermal kernel and the interaction
term in a self-consistent way, through the introduction of terms arising from the nonlinear interactions (the vertex and self-energy, respectively). For a classical model
with white external noise the single variable equation reads
Z t

dt (t, t ) i (t ) + i (t) + i (t) .

(6.29)
M i (t) + i (t) + (t) i (t) =
0

Its generalisation is straightforward. There are two noise sources in this equation:
i (t) is the original white noise while i (t) is an effective (Gaussian) noise with zero
mean and correlations self-consistently given by hi (t)j (t )i = ij D(t, t ). The vertex
D(t, t ) plays the r
ole of the colored noise correlation in a usual Langevin equation.
The self-energy (t, t ) appears here in the place of an integrated friction. A solution of the problem can be attempted numerically using this equation and the
self-consistent definitions of and D.
This procedure is not particular useful for the analysis of polynomial models
since the transformation into a Q dependent effective action can be done exactly. It
does however become useful for dealing with models whose single-spin effective action
has higher order interaction terms. An example is the quantum sk model.
Interestingly enough a rather flat harmonic oscillator coupled to a bath made of a
white and a coloured part at different temperatures acquires two time-scales controlled
by the two temperatures involved. We see that a similar structure might appear for
the glassy system if the self-energy and vertex self-consistently arrange to act on each
degree of freedom as the friction and noise-noise correlator of a complex bath. We
will see that this is indeed what happens to mean-field models.
6.3.5

Diagrammatic techniques

In this Section we first describe the perturbative solution to the Langevin process
and how it is used to construct series expansions for the correlations and responses.
Self-consistent approximations, such as the mode coupling or the self-consistent screening, correspond to a selection of a subset of diagrams from the full series. The connection with disordered models is demonstrated. An extension to quantum problems
is possible using the generating functional formalism.
162

Let us focus on a single scalar degree of freedom, q, with potential energy

V (x) =

(t) 2
g
x + x3 ,
2
3!

(6.30)

and dynamics given by the Langevin in the white noise limit. We take the initial
condition x(t = 0) = 0. (t) is a time-dependent function that we fix at the end of
the derivation by requiring C(t, t) = 1. In vector models it is the Lagrange multiplier that self-consistently imposes a spherical constraint. Note that this potential is
not bounded from below. Setting Go (t, t ) = [(t) + t + M t2 ]1 , a perturbative
expansion for x(t) in powers of the noise is easily written as
g
(Go [Go Go ]) (t) + ...
(6.31)
2
Rt
where means a time convolution, (Go f )(t) = 0 dt Go (t, t )f (t ), and is a
simple product at equal times. This notation is equivalent to the one used in the
susy formalism, see Appendix . Causality implies
Go (t, t ) (t t ). If inertia

R
t
can be neglected Go (t, t ) = exp t d ( ) (t t ). If one keeps the second-time
x(t) = (Go )(t)

derivative Go (t, t ) takes a more complicated form. Equation (6.31) can be graphically
represented as in Fig. 51. Crosses indicate noise and oriented lines indicate the bare
propagator Go . Each vertex carries a factor g/2. Note that the unknown q is evaluated
at time t while the noises are evaluated at all previous times.
x
x

x +

x
+ ...

Figure 51: Terms O(g 0 ), O(g 1 ) and O(g 2 ) in the perturbative solution to the Langevin
equation.
The expansion for q leads to two expansions for the correlation and response. In
simple words, the former corresponds to sandwiching, i.e. averaging over the noise,
the usual product of two series as the one in Fig. 51 evaluated at different times t
and t . Due to the average over the Gaussian noise noise factors have to be taken by
pairs. Let us illustrate this with a few examples.
The first term in the expansion is the result of averaging two O(g 0 ) terms (first
term in Fig. 51):

Co (t, t ) = h(Go )(t) (Go )(t )i = 2kB T

dt Go (t, t )Go (t , t ) ,

t t . We depict this term and its contributions to more complicated diagrams with
a single crossed line, see the first graph in Fig. 52. The term O(g), as well as all
terms which are odd powers of g, vanishes. There are two contributions to the term
163

O(g 2 ). One is the result of multiplying a term O(g 2 ) with a term O(g 0 ) and it is a
tadpole, see the second graph in Fig. 52; we assume this term and all its corrections
are included in the contributions from the time-dependent mass and we henceforth
ignore them. The other comes from multiplying two O(g) terms, see the third graph
in Fig. 52.
Higher order terms are of two types: they either dress the propagators or they
dress the vertices, see the last two diagrams in Fig. 52. These two terms are order
O(g 4 ). The first one follows from averaging two O(g 2 ) contributions while the second
one is the result of averaging an O(g 3 ) and an O(g) term. The full series yields the
exact perturbative expansion for C.

x
x
x

x
x
x
x
x
x
x
x
Figure 52: From left to right: O(g 0 ), two O(g 2 ) and two O(g 4 ) terms in the series
for C. The next to last diagram dresses the propagator and the last term dresses the
vertex. The former is kept in the mca while the latter is neglected.
The series expansion for the response follows from the relation (??) in the white
noise limit. In graphical terms we obtain it by multiplying the series in Eq (6.31) and
Fig. 51 evaluated at time t by a noise evaluated at time t and taking the average.

6.4

The mode coupling approximation (mca)

The diagrammatic expansions for C and R can be represented analytically by

introducing the kernels (t, t ) and D(t, t ) through the Schwinger-Dyson equations
(6.20) and (6.21) in their integral form. Each of them is a compact notation for a
series of diagrams. These equations are exact perturbatively. However, for a generic
model one cannot compute the kernels and D exactly.
The mode coupling approach amounts to approximating the kernels (t, t ) and
D(t, t ) in the following way. One takes their values at O(g 2 ) and substitutes in them
the bare propagator Go and the bare correlation C0 by their dressed values, i.e. by
R and C themselves. For the model defined in Eq. (6.30) this yields
(t, t )
D(t, t )

g 2 C(t, t ) R(t, t ) ,

(6.32)

g2 2
C (t, t ) .
2kB T (t t ) +
2

(6.33)

This approximation neglects vertex renormalization in the sense that all diagrams
correcting the values of the lines are taken into account while all diagrams correcting
the vertices are neglected. For instance, one keeps the fourth diagram in Fig. 52 that
represents a line correction, while leaving aside the fifth diagram drawn in the same
figure that represents a vertex correction.
164

The same procedure can be implemented using the susy representation of the
dynamics. Each line represents the superfield and the super-correlator follows from
the sandwich of two series for the super-field evaluated at different super-coordinates
a and b.
The Schwinger-Dyson equations can be recast, after multiplying by G1
o , into the
form (6.18) and (6.19) for a random potential (6.5) with only one term r = p = 3.
Applying the mca to the trivial (and ill-defined) model (6.30) we derived the dynamic
equations for the p = 3 spin spherical model! On the one hand, this result is worrying
since it shows that the mca can be rather uncontrolled and it can generate glassy
behavior by itself. On the other hand, since the same equations hold in the mca of a
model of interacting particles with realistic interactions, this calculation allows one to
understand why the dynamic equations of the mct for super-cooled liquids coincide
with the ones of disordered spin models above Td . In the next Subsection we show
how the diagrams neglected in the mca vanish in a disordered model with a large
number of components.

6.5

mca and disordered models

The first to notice that the mca for a quadratic dynamic equation corresponds
to the exact dynamic equation of a disordered problem with a large number of components was Kraichnan in the context of the Navier-Stokes equation. More recently,
Franz and Hertz showed that the schematic mct equations of the Fp group for
super-cooled liquids are identical to those arising from a spin model with pseudorandom interactions between groups of three spins. (The schematic mct focus on a
chosen wavevector.)
Indeed, for the example chosen in this Section, one easily demonstrates that the
diagrams retained by the mca are precisely those which survive if one modifies the
initial model (6.30) and considers instead the following disordered problem. First, let
us upgrade q to a vector with N components or colors i , where i = 1, 2, ..., N .
Second, let us modify the potential energy (6.30) into
X
~ =g
V ()
Jijk i j k
(6.34)
i<j<k

with couplings Jijk that are independent quenched Gaussian random variables of zero
2
mean and variance [Jijk
]J = 1/N p1 = 1/N 2 . (p is the number of spins in each term
in V .) In the large N limit, the noise and disorder averaged correlation and response
of this modified model obey Eqs. (6.18) and (6.19) with and D given by Eqs. (6.32)
and (6.33), respectively. The fact that these equations are recovered can be seen
either directly on the perturbation theory, or using the functional methods given in
Section . Since we want to stress that the diagrams neglected in the mca vanish
exactly for this model we use here the first approach.
The bare propagator is diagonal in the color indices, Go ij = Go ij . The vertex
is now proportional to the random exchanges Jijk . The perturbative solution to the
165

Langevin equation reads

i (t) = (Go i )(t) Jijk Go (Go j Go k ) (t) + . . . .

(6.35)

One is interested
PN in computing the self-correlation averaged over the noise and disorder, N 1 i=1 [hi (t)i (t )]. The latter average eliminates all terms with an odd
number of couplings. Similarly, since Jijk 6= 0 only if all indices i, j, k are different,
tadpole contributions as the one in the second graph in Fig. 52 vanish (the noise-noise
correlation enforces that two indices in the random exchange must coincide). Finally,
one can check that due to the scaling with N of the variance of the disordered interactions, vertex corrections as the one in the last graph in Fig. 52 are sub-leading and
vanish when N . Instead, all line corrections remain finite in the thermodynamic
limit. We can check this statement in the two examples shown in Fig. 52 extended to
include color indices. The vertex correction has four random exchanges that due to
the averaging over the noise are forced to match as, e.g. Jijk Jjlm Jmni Jkln leaving 6
free-indices. Averaging over disorder one identifies the indices of two pairs of J s, e.g.
i = l and k = m, this yields a factor (1/N 2 )2 and, at most, it leaves 4 color indices
over which we have to sum from 1 to N (i, j, k, n). We have then an overall factor
1/N 4 N 4 = 1 and this term vanishes when one normalises the correlation by N .
Instead, in the line correction, after averaging over the noise, we are left with 6 free
indices, e.g. Jikj Jklm Jlmn Jinj , the average over the noise only imposes k = n in its
most convenient contribution, and the overall factor is 1/N 4 N 5 = N . This term
contributes to the normalisaed global correlation.
Interestingly enough, the equivalence between the mca and a disordered system
extends P
to an arbitrary non-linear coupling F (q). Expanding F in a power series

F (q) = r=2 Fr!r q r the mca leads to

(t, t )

g2

X
r=2

D(t, t )

Fr2
C r1 (t, t ) R(t, t ) ,
(r 1)!

2kB T (t t ) + g 2

X
F2
r

r=2

r!

C r (t, t ) .

(6.36)
(6.37)

[Note that for r odd, there appears an additional tadpole contribution in Eq. (6.36),
which we have assumed again that it has been re-absorbed into the mass term (t).]
The dynamic equations can also be obtained as the exact solution of the Langevin
dynamics of N continuous spins i interacting through the potential
~ =g
VJ []

X
r2

Fr

Ji1 ...ir+1 i1 . . . ir+1

(6.38)

i1 <...<ir+1

where J1 ,..r+1 are quenched independent Gaussian variables with zero mean and
[(J1 ,..r+1 )2 ] N r . Therefore the mc equations for a single dynamic variable in
contact with a heat reservoir and under an arbitrary nonlinear potential F (q) describe
166

exactly a fully-connected spin-glass problem with arbitrary multi-spin interactions

or a particle evolving in an N dimensional space in a quenched random potential
~ with a Gaussian distribution with zero mean and variance (6.6). Let us note
V []
that in order to be well defined, the model given by V must be supplemented by a
constraint preventing the field i from exploding in an unstable direction set by the
coupling tensor Ji1 ...ir+1 . This problem is cured by imposing the spherical constraint
PN 2
i=1 i (t) = N C(t, t) N .
The extension of the mapping to a space dependent (~x, t) (or to a multicomponent
field) is straightforward. Several interesting physical examples involve an equation of
the type:
~k, t)
(
t

~k, t)
= (k 2 + )(
+(~k, t)

X
X
Fr
~k1 , t)....(
~kr , t)
Lr (~k|~k1 , .....~kr )(
r!
r=2
~
k1 ,..~
kr

~k, t) is the Fourier transform of (~x, t), and (~k, t) a Gaussian noise such that
where (
h(~k, t)(~k , t )i = 2kB T (~k + ~k )(t t ). The Kardar-Parisi-Zhang (kpz) equation
corresponds to r = 2, L2 (~k|~k1 , ~k2 ) = [~k1 ~k2 ] (~k1 + ~k2 + ~k), while domain coarsening
in the 4 theory corresponds to r = 3, L3 (~k|~k1 , ~k2 , ~k3 ) = (~k1 + ~k2 + ~k3 + ~k), with
a negative . The Navier-Stokes equation is similar to the kpz case with, however,
an extra tensorial structure due to the vector character of the velocity field. The
correlation and response functions now become ~k dependent, d (~k + ~k )C(~k, t, t ) =
~k, t)(
~k , t )i and d (~k + ~k )R(~k, t, t ) = h (
~k, t)/(~k , t )i. The generalized mc
h(
equations then read (assuming that the structure factors are invariant under the
permutation of ~k1 , ..., ~kr ):
(~k, t, t )

g2

X
r=2

D(~k, t, t )

X
Fr2
Lr (~k|~k1 , .....~kr )Lr (~kr |~k1 , .....~k)
(r 1)!
~
k1 ,..~
kr

C(~k1 , t, t )...C(~kr1 , t, t )R(~kr , t, t )

(6.39)


2 X 
X
2
Fr
2kB T (t t ) + g 2
Lr (~k|~k1 , .....~kr )
r!
r=2
~
k1 ,..~
kr

C(~k1 , t, t )...C(~kr , t, t )

(6.40)

where (~k, t, t ) and D(~k, t, t ) are defined in analogy with Eqs. (6.36) and (6.37).

6.6

mca for super-cooled liquids and glasses

In the last 20 years the mca has been much used in the study of super-cooled
liquids. Starting from the realistic interactions between the constituents of a liquid,
Gotze et al used the mca together with an assumption of equilibrium to derive a dynamic equation for the density-density correlator. This analysis lead to the schematic
167

mode coupling theory (mct) of super-cooled liquids and generalizations (with no reference to wave-vector dependence) and to more sophisticated versions that include a
dependence on space. The difference between these models lies on the form of the
kernels and D. Kirkpatrick, Thirumalai and Wolynes realized in the late 80s that
the schematic mode coupling equation is identical to the dynamic equation for the
spin-spin correlator in the disordered Potts or p spin model, building a bridge between
the study of structural and spin glasses. Why these models and not sk? This will
become clear when we present their dynamic and static behavior.
In this Section we explained why the dynamic equation of a disordered model and
the one stemming from a mca of a model with more realistic interactions concide:
the terms neglected in the latter vanish exactly in the former. The example studied
here serves also to signal the danger in using a mca. One could conclude that a trivial
model has a highly non-trivial dynamics, this being generated by the approximation
itself.
In the derivation of the dynamic equations presented in this Section no assumption
of equilibrium was used. Therefore, these equations hold also in the low temperature
phase where equilibrium is lost. It is then natural to propose that the dynamics of the
p spin spherical model below Td schematically describes the dynamics of glasses just as
its dynamics above Td yields the schematic mct of super-cooled liquids. To go beyond
the schematic theory while still keeping a single mode description one simply has to
consider p1 + p2 spherical disordered models. Moreover, the dynamics of a manifold
in a random potential is described by dynamic equations with a ~k dependence that
goes beyond the single mode mct.

168

Glassy dynamics: Generic results

Before presenting the explicit solution to the mean-field models we state some
generic features of the low-T dynamics that we believe hold in general.
1

qea (T )

0.1

d
qea

0 (T1 )
0.01

10

0 (T2 )
100

0.1

1000

0 (t01 )
10000

100000

1e+06

0.01

1e+07

10

0 (t02 )
100

t0

1000

10000

100000

1e+06

1e+07

t0

Figure 53: Left: Sketch of the decay of the stationary correlations in the high T phase
close to Td , T1 > T2 > . . .. Right: Sketch of the decay of the aging correlations in the
low T phase, at fixed T < Td , t1 < t2 < . . .

7.1

The weak-ergodicity breaking scenario

Figure 53-right shows a sketch of the decay of the correlation as obtained from the
numerical solution to the dynamic equations for the mean-field models (see Section ).
It develops a separation of time scales in the long t limit. It first approaches a
plateau at qea in a stationary manner and it then decays below this value with an
explicit waiting-time dependent form. For each waiting-time there is a sufficiently
long t such that the correlation decays to zero. These properties are included in the
weak-ergodicity breaking (web) scenario that states that, for t t , C decays in such
a way that
lim C(t, t )

lim

tt

Cst (t t ) = 0
lim C(t, t )

qea + Cst (t t )

lim

(7.1)

lim C(t, t ) = qea

tt t

at fixed t .

(7.2)
(7.3)

Equation (7.2) defines the Edwards-Anderson order parameter, qea . For finite t
there is a crossover between two time-scales controlled by a waiting-time dependent
characteristic time 0 (t ) that is a growing function of t whose precise form depends
on the model. For large t t such that t t is small with respect to 0 (t ), the
correlation function first decays from 1 to qea in a tti manner. At longer t t it goes
169

further below qea to eventually reach 0 in a manner that depends both upon t and t
(the aging effect). This behavior suggests the presence of at least two time-sectors in
which the dynamics is stationary and non-stationary, respectively. We will see that
the number of time-scales, or more precisely correlation scales, depends on the model.
We write C as the sum of a stationary and an aging contribution:
C(t, t )

Cst (t t ) + Cag (t, t ) .

(7.4)

The matching conditions at equal times between Cst and Cag are C(t, t) = 1 implying
Cst (0) + Cag (t, t) = 1 with Cst (0) = 1 qea and Cag (t, t) = qea . Together with Eq.
(7.2) they ensure that in the two-time sector in which Cst decays from 1 qea to 0,
Cag is just a constant qea . Instead, in the two-time sector in which Cag decays from
qea to 0, Cst vanishes identically.
The name web reflects the fact that for short time-differences the system behaves
as if it were trapped in some region of phase space of size qea suggesting ergodicity
breaking. However, it is always able to escape this region in a time-scale 0 (t ) that
depends upon its age t . Hence, trapping is gradual and ergodicity breaking is weak.
qea depends on temperature when T < Td .
We have already described, phenomenologically, such a separation of time-scales
in the decay of correlation functions when we discussed the domain growth problem.
The first term describes in this case the fast fluctuations within domains while the
second term characterises the domain growth itself. A second example where such
a separation of time-scales occurs are the trap models in phase space. The first
term corresponds to the dynamics within the traps while the second describes the
wandering of the system from trap to trap. In glasses, the first term corresponds to
the rapid rattling of each particle within its cage while the second one describes the
destruction of the cages and hence the structural relaxation.
In driven models rendered stationary by a weak perturbation we also find a separation of time-scales with 0 increasing with weaker strengths of the perturbation.
We can also propose that C and R separate in two terms, both being stationary but
evolving in different time-scales.
In classical purely relaxational models governed by a Langevin equation with no
inertia the correlation functions are monotonic with respect to both times t and t , as
it is easily checked numerically. Inertia introduces oscillations and the decay can be
non-monotonic. The magnitude of the oscillations depends upon the relative value of
the mass M with respect to the other parameters in the problem. However, for a wide
choice of parameters the oscillations appear only in the stationary regime, the aging
dynamics having a monotonic decay towards zero. This is relevant since it allows one
to use the general properties of monotonic correlation functions to find the two-time
scaling of Cag (t, t ).

7.2

The weak long-term memory scenario

Regarding the response function, we propose a similar separation in two terms:

R(t, t ) = Rst (t t ) + Rag (t, t )
170

(7.5)

with Rst (t t ) limt R(t, t ). The matching conditions close to equal times
are different for a model with or without inertia. In the former case, R(t, t) = 0,
R(t, t ) = 1/M while in the latter, using the Ito convention, R(t, t) = 0 , R(t, t ) =
1/. In both cases the equal-times condition implies Rst (0) = 0, Rag (t, t) = 0 while
the next-to-main diagonal values yield Rst () = 1/M, Rag (t, t ) = 0 and Rst () =
1/ , Rag (t, t ) = 0, respectively.
The response tends to zero when times get far apart, and its integral over a finite
time-interval as well:
Z t

dt R(t, t ) = 0
fixed t .
(7.6)
lim R(t, t ) = 0 ,
lim
t

lim

These properties imply

lim R(t, t ) = 0

lim

tt t

tt

Rst (t t ) = 0 , lim Rag (t, t ) = 0 . (7.7)

t

However, the contribution of the response to the dynamic equations and to other
measurable quantities is not trivial. Examining the integral of the response function
over a growing time interval one finds that even if the response vanishes, it yields a
contribution to the integration. Figure 56-left shows the integrated linear response
(??). Using (7.5)

(t, t ) =

dt [Rst (t t ) + Rag (t, t )] = st (t t ) + ag (t, t ) .

(7.8)

If, for long enough t , the contribution of the second term in (7.8) were negligible,
(t, t ) should be a stationary quantity. Instead, for all t s studied and for t long
enough one clearly sees a waiting-time dependence that can only come from the
integration of the second term. This is a weak long-term memory (wltm), the system
has an averaged memory of its past.
When a system is in equilibrium, the response is simply related to the correlation
via fdt. We then assume (and test on the dynamic equations) that the dynamics in
the stationary regime satisfies fdt:
Rst ( )

Rst ()

1 dCst ( )
0,
kB T
d


Z
1
2
h
d
Cst ( )
lim+
tanh
h 0

2 + i
2

(7.9)

in a classical and quantum problem, respectively. One can formally prove that fdt
has to hold for any generic relaxing model fro short time-differences. For longer
time-differences, when Cag and Rag vary in time while Cst and Rst have decayed to
zero, one cannot assume the validity of fdt and, as we will see, the equations have a
solution that explicitly modifies fdt.

171

7.3

Slow time-reparametrization invariant dynamics

We have already mentioned that the correlations decay monotonically (only below
qea if M 6= 0). The final insight coming from the numerical solution to the full
equations is that the dynamics becomes slower and slower for fixed waiting-time and
as tt increases. In the stationary regime t2 [C(t, t ), R(t, t )] and t2 [C(t, t ), R(t, t )]
are not negligible with respect to the terms in the rhs of Eqs. (6.18) and (6.19). On
the contrary, in the second decay below qea , C and R decay in a much slower manner
such that, t C(t, t ) (t)C(t, t ) and t2 C(t, t ) (t)C(t, t ) (similarly for R),
and the time-derivatives can be neglected.
We choose the following strategy to solve the equations in the long t limit where
a sharp separation of time-scales can be safely assumed. First, we take advantage
of the fact that one-time quantities approach a limit, as one can verify numerically,
and write the asymptotic form of Eq. (6.25) for . The integrals on the rhs are
approximated using the separation of C and R in two terms that vary in different
time-scales that we assume are well-separated. We detail this calculation below. As
regards to the equations for C and R, we proceed in two steps. On the one hand, we
choose t t short in such a way that C > qea and we write the dynamic equations for
Cst and Rst . On the other hand, we take t and t widely separated so as C < qea and
we write the dynamic equations for Cag and Rag . In this way we double the number
of unknown functions and equations but we simplify the problem enough as to make
it solvable.
Once the time-derivatives are neglected and the integrals are approximated as we
explain in Section the aging equations become invariant under reparametrizations of
time t h(t) that transform the two-point functions as
Cag (t, t ) Cag (h(t), h(t )) ,

Rag (t, t ) [dt h(t )] Rag (h(t), h(t )) .

(7.10)

This is not an exact invariance of the dynamic equations. It is only generated when
dropping the time-derivatives. This invariance was first noticed by Sompolinsky in his
study of the equilibrium dynamics (see also [?] and it later appeared in the nonequilibrium dynamics. We will see that this approximation forbids us to solve completely
the dynamic equations, in particular, to fix the time scaling (select h(t)).

7.4

Correlation scales

Take three ordered times t3 t2 t1 . The correlations are C(ti , tj ) =

P
k hsk (ti )sk (tj )i cos ji . The monotonicity of the decay of the correlations
with respect to the longer time (keeping the shorter time fixed) and the shorter time
(keeping the longer time fixed) allows us to derive general properties that strongly
constrain the possible scaling forms. Indeed, one can relate any three correlation
functions via triangle relations constructed as follows. Using the fact that the decay
is monotonic, one can invert the relation between correlation and times to write, for
example, t2 = g(C(t2 , t1 ), t1 ) with g : [0, 1][0, ] [0, ]. This allows us to rewrite
1
N

172

C(t3 , t1 ) as
C(t3 , t1 ) = C(g(C(t3 , t2 ), t2 ), t1 ) = C(g(C(t3 , t2 ), g(C(t2 , t1 ), t1 ), t1 ) .

(7.11)

We now define a real function f (x, y), f : [0, 1] [0, 1] [0, 1], by taking the limit
t1 while keeping the intermediate correlations fixed
lim
C(t3 , t1 )
t1
C(t2 , t1 ) and C(t3 , t2 ) fixed

f (C(t3 , t2 ), C(t2 , t1 )) .

The fact that the limit exists is a reasonable working assumption. This function completely characterizes the correlations and their scales in the asymptotic limit. (Note
that we defined f using the correlation between the longest time and the intermediate
as the first argument.)
7.4.1

Properties

The definition of the function f , as well as the properties shown in this Subsection,
are model independent. The form taken by f for each model is determined by the
dynamic equations.
Time reparametrization invariance The function f is invariant under reparametrizations of time that satisfy (7.10).
Associativity Take now four times t4 t3 t2 t1 . The correlation between t4 and
t1 can be written in two ways
C(t4 , t1 )

f (C(t4 , t2 ), C(t2 , t1 )) = f (f (C(t4 , t3 ), C(t3 , t2 )), C(t2 , t1 )) ,

C(t4 , t1 )

f (C(t4 , t3 ), C(t3 , t1 )) = f (C(t4 , t3 ), f (C(t3 , t2 ), C(t2 , t1 ))) .

Thus f satisfies f (f (x, y), z) = f (x, f (y, z)), i.e. it is an associative function.
Identity. If one takes t1 = t2
C(t3 , t1 ) = f (C(t3 , t2 ), C(t2 , t1 )) = f (C(t3 , t1 ), C(t1 , t1 )) = f (C(t3 , t1 ), 1) ,

(7.12)

for all C(t3 , t1 ) [0, 1]. Equivalently, if one takes t2 = t3

C(t3 , t1 ) = f (C(t3 , t2 ), C(t2 , t1 )) = f (C(t3 , t3 ), C(t3 , t1 )) = f (1, C(t3 , t1 )) ,

(7.13)

for all C(t3 , t1 ) [0, 1]. The correlation at equal times acts as the identity since
x = f (x, 1) and y = f (1, y) for all x, y [0, 1].
Zero. Taking t3 and t2 and much larger than t1 in such a way that C(t2 , t1 ) 0 and
C(t3 , t1 ) 0 while C(t3 , t2 ) > 0,
0 C(t3 , t1 ) = f (C(t3 , t2 ), C(t2 , t1 )) f (C(t3 , t2 ), 0) .
173

(7.14)

Equivalently, taking t3 t2 and t1 , then C(t3 , t2 ) 0 and C(t3 , t1 ) 0 while

C(t2 , t1 ) > 0 and one has
0 C(t3 , t1 ) = f (C(t3 , t2 ), C(t2 , t1 )) f (0, C(t2 , t1 )) .

(7.15)

The minimum correlation acts as a zero of f (x, y) since 0 = f (x, 0) and 0 = f (0, y) for
all x, y [0, 1]. (This property can be easily generalised if the correlation approaches
a non-zero limit.)
Bound. Given that we assume that the system drifts away in phase space, C(t2 , t1 )
decays as a function of t2 for t1 fixed, and C(t2 , t1 ) increases as a function of t1 for
t2 fixed. This property implies
y = f (1, y) f (x, y) y, x < 1 ,

x = f (x, 1) f (x, y) x, y < 1 .

(7.16)

Therefore f (x, y) min(x, y).

Forms for f In [?] we proved that
f (x, y)

1 ((x)(y))

Isomorphic to the product

(7.17)

f (x, y)

min(x, y)

Ultrametricity

(7.18)

are the only possible forms that satisfy the properties of f shown above. Note that
for equal to the identity the first type of function becomes simply f (x, y) = xy,
hence the name. It is also possible to prove that the first kind of function (7.17) is
only compatible with the time scaling


h(t2 )
C(t2 , t1 ) = 1
(7.19)
h(t1 )
with h(t) a monotonically growing function. The actual correlation can have a piecewise form. Here, we explain these statements reviewing the scaling forms found for
some physical systems and in the analytic solution to mean-field models.
Examples: domain growth
The correlation decays in two steps, see the right panel in Fig. 53 and for C >
qea = m2eq the decay is stationary:
C21 C(t2 , t1 ) = qea + Cst (t2 t1 ) ,

(7.20)

and it can be put in the form (7.19) using h(t) = exp(ln t)) and 1 (x) = qea + Cst (x).
1
(C31 qea ) = t3 t2 +
Any three correlation satisfying (7.20) also verify t3 t1 = Cst
1
1
t2 t1 = Cst (C32 qea ) + Cst (C21 qea ) that implies
1
1
C31 = Cst [Cst
(C32 qea ) + Cst
(C21 qea )] + qea .

(7.21)

This equation is equivalent to (7.17). This means that any three correlations above
qea can be related with an f that is isomorphic to the product, see (7.17), with
1
1
st (x) = Cst (ln x) + qea and st (x) = exp(Cst (x qea ).
174

When the times are such that the domain walls move, the self-correlation decays
below qea in an aging manner, with


R(t2 )
1
C21 C(t2 , t1 ) = Cag (t2 , t1 ) = ag
,
(7.22)
R(t1 )
1
1
ag (1) = qea and ag (0) = 0. It is obvious that any three correlations below qea also
satisfy (7.17)
Take now t3 = t2 + 32 with 32 < 0 (t2 ) and C32 > qea , and t3 and t2 sufficiently
larger than t1 (t3 = t1 + 31 with 31 > 0 (t1 ) and t2 = t1 + 21 with 21 > 0 (t1 ))
such that C31 < qea and C32 < qea . One has





R(t3 )
R(t3 )
R(t2 )
1
1
C31 = 1
=

)
ag
ag
ag
ag
R(t1 )
R(t2 )
R(t1 )


R(t
)
3
= 1
ag (C21 ) = C21 .
ag
R(t2 )

The last idendity is a consequence of R(t3 )/R(t2 ) 1 since for a sufficiently small
32 , R (t2 )32 /R(t2 ) 1.
Thus, when the times are such that two correlations, say with values a and b, are
both greater than qea one explores the dynamics in the stationary regime and f (a, b)
is isomorphic to the product. When they are both smaller that qea one explores the
dynamics in the aging coarsening regime and again f (a, b) is isomorphic to the product
though with a different function . Finally, if a > qea and b < qea , f (a, b) = min(a, b)
and one finds dynamic ultrametricity.
The structure discussed in the context of the domain growth problem is indeed
generic. Some special values of the correlation act as fixed points of f (a, a),
f (a, a) = a. A correlation scale spans the values of correlations comprised between two subsequent fixed points. Within a correlation scale f is isomorphic to the
product. Any two correlations falling into different correlation scales are related by
an ultrametric f . In the domain growth example 1, qea and 0 are fixed points that
are simple to visualize physically. In more abstract models as the sk spin-glass the
form of f is more involved, with a stationary scale between 1 and qea and a dense set
fixed points, hence correlation scales, that fill the interval [0, qea ].
Scaling functions
Most solvable models, numerical data and experimental results can be described
with only two correlation scales, a stationary and a slow one. Several scaling functions h(t) for the slow decay have been proposed in the literature. In the following
we summarize and discuss the main ones. In Fig. 54 we compare the decay of the
correlation from qea for three of the four laws discussed below.
Power law: h(t) = at . This is the simplest scaling also called simple aging. Ferromagnetic domain growth realizes this form with = 1/2 for non conserved dynamics
and = 1/3 for conserved dynamics. Several solvable model have simple aging, an
175

t=t0 tw 1

ln(2t=t0 ) tw 1

eln

0.1

tw 1
tw 2
tw 3

qea

(t=t0 )

t=t0

ln(t=t0 )
2 (t=t )
0

eln
0.01

10

100

1000

10000 100000 1e+06 1e+07 1e+08

0.1

t=t

Figure 54: Comparison between three h(t)s, power law, enhanced power law and
logarithm. Plot of C(t, t ) = (1 qea ) exp((t t )) + qea h(t)/h(t ) against the timedifference t t (on the left) and against the ratio of times t/t (on the right) for three
waiting times. Note the drift of the curves in the right panel. For the logarithmic
law (sub-aging) the curves drift towards the left for increasing waiting-time. Instead,
for the enhanced power law (super-aging) the curves drift to the right for increasing
waiting-time. For the power law (simple aging) the scaling is perfect. In real systems
the decay of the stationary part towards qea is much slower than exponential (typically
power law with a small exponent) and the separation of time-scales is not so neat.
example being the classical spherical p = 2 model. In it was conjectured that a power
law also characterized the aging dynamics of the fully connected p spin-model with
p 3. This was later confirmed with the algorithm of Kim and Latz that allows one
to reach much longer times. Aging below Tc in the simplest trap model also scales
with this law. The molecular dynamic simulations of Lennard-Jones mixtures show
this type of scaling too. Note that for all , C scales as a function of t2 /t1 .
Enhanced power law: h(t) = exp(ln (t/t0 )) This law yields the most accurate description of spin-glass experimental data. The exponent typically takes a possibly
T -dependent value about 2.
Stretched exponential: h(t) = exp[(t/t0 ) ] This law has been proposed to describe
the slowing down of the full correlation above the critical temperature. As far as we
know, no aging model that satisfies a scaling (7.19) with a stretched exponential has
been found yet.
Logarithm: h(t) = ln (t/t0 ) In the Fisher and Huse droplet model for spin-glasses,
activated dynamics is assumed and the domains are found to grow as R(t) ln(t/t0 ).
This leads to C(t2 , t1 ) g(ln(t2 /t0 )/ ln(t1 /t0 )). However, this law does not fit the
aging experimental data.
Dynamic ultrametricity: Even though it seems mysterious at first sight there is a
simple graphical construction that allows one to test it. Take two times t3 > t1
such that C(t3 , t1 ) equals some prescribed value, say C(t3 , t1 ) = 0.3 = C31 . Plot
now C(t3 , t2 ) against C(t2 , t1 ) using t2 , t1 t2 t3 , as a parameter. Depending on
176

the value of C31 with respect to qea we find two possible plots. If C(t3 , t1 ) > qea ,
for long enough t1 , the function f becomes isomorphic to the product. Plotting
then C(t3 , t2 ) for longer and longer t1 , the construction approaches a limit in which
C(t3 , t2 ) = 1 ((C31 )/(C(t2 , t1 ))). If, instead, C31 < qea , in the long t1 limit the
construction approaches a different curve. We sketch in Fig. 55 two possible outcomes
of this construction. On the right, we represent a model with two correlation scales,
ultrametricity holds between them and within each of them f is isomorphic to the
product. On the left instead we represent a model such that dynamic ultrametricity
holds for all correlations below qea . The construction approaches, in the long t1 limit,
the broken curve depicted in the sketch.
The sk spin-glass and the dynamics of manifolds in an infinite dimensional embedding space in the presence of a random potential with long-range correlations have
ultrametric decays everywhere within the aging regime. This scaling is also found in
the trap model at the critical temperature. Dynamic ultrametricity in finite dimensional systems has been search numerically. There is some evidence for it in the 4dEA
model. In 3d instead the numerical data does not support this scaling. Whether this
is due to the short times involved or if the scaling asymptotic is different in 3d is still
an open question.

C (t3 ; t2 )

C (t3 ; t2 )

0.8

0.6

0.8

0.6

qea

qea
0.4

0.4

C (t3 ; t1 )

C (t3 ; t1 )

0.2

0.2

0
0

0.2

0.4

C (t3 ; t1 )

0.6

qea

0.8

C (t2 ; t1 )

0.2

0.4

C (t3 ; t1 )

0.6

qea

0.8

C (t2 ; t1 )

Figure 55: Sketch of a check of ultrametricity using the parametric plot C(t3 , t2 )
against C(t2 , t1 ) for C(t3 , t1 ) = 0.3 < qea fixed. On the left an ultrametric model, on
the right a model with two correlation scales.

7.4.2

Definition of a characteristic time

177

Expanding the argument in (7.19) for t2 = t1 + with t1 one finds, to leading

order,
1

h(t1 )

h (t1 )
=1
tc (t1 )
,
(7.23)
h(t2 )
tc (t1 )
h(t1 )

 
2
with O 2 h (t1 )/h2 (t1 ) + h (t1 )/h(t1 ) corrections. The characteristic time tc (t1 )
is given by

t1 /
Power law

t1 /[ ln1 (t1 /t0 )]

Enhanced power law
tc (t1 ) =

t1 (t0 /t1 )
Stretched exponential

t1 ln(t1 /t0 )
Logarithm
Note that tc (t1 ) is defined close to the limit of equal times and (7.23) does not make
sense for large . Rather often in the literature, the scaling variable x = /ta1 has
been used even for large values of . This scaling is incompatible with the general
properties of the triangular relations recalled in Section if the exponent a is larger
than 1. See the right panel in Fig. 54 to see the different trends of these scalings when
plotted as functions of t/t .
For the power law tc (t1 ) scales just as t1 . In the cases of the stretched exponential
and the enhanced power law, tc (t1 ) has a slower growth than the linear dependence
iff > 0 in the first case and > 1 in the second. This behavior has been called
sub-aging. For the logarithm tc (t2 ) grows faster than linearly. This function belongs
to a different class that we called super aging.

7.5

Modifications of fdt

One of the most important outcomes of the analytic solution to the mean-field
glassy models is the need to modify the fluctuationdissipation relations between
linear responses, R(t, tw ), and their partner correlations between spontaneous fluctuations, C(t, tw ), when T < Td . In this Subsection we discuss different ways of
presenting the modification of fdt expected in rather generic systems with slow dynamics.
7.5.1

Time domain

The fdt is a linear relation between (t, tw ) and C(t, tw ) for any pair of times
(t, tw ). In early simulations of the 3dea model as well as in the analytic solution to
fully-connected disordered models a modification of this relation below Td appeared.
Plotting kB T (t, tw ) and 1 C(t, tw ) for tw fixed as a function of ln(t tw ) one
typically obtains the pair of curves schematically shown on the left panel of Fig. 56.
The two functions go together until t tw reaches a characteristic time 0 (tw ) and
they then depart demonstrating that fdt does not hold beyond this time-scale. The
characteristic time 0 (tw ) is of the order of the time needed to reach the plateau in

178

0.75

tw 1

tw 2
1 qea

0.5
1 C(t, tw )
kB T (t, tw )
1 C(t, tw )
kB T (t, tw )

0.25
0
100

kB T (,
tw )/( C(,
tw ))

102

104
106
t tw

108 101 0

T /T

1
1
2
3
100

102

104

106

tw

Figure 56: Left: sketch of the modification of fdt in the time-domain. Right: sketch
of the modification of fdt in the frequency domain for a glassy system, 1 > 2 > 3 .
the correlation function (this holds for mean-field models but it is not certain in finite
dimensional systems). Summarizing
t tw < 0 (tw )
t tw > 0 (tw )

fdt holds in the fast scale ,

fdt is modified in the slow scale ,

(7.24)
(7.25)

with 0 (tw ) an increasing function of tw that depends on the system considered (see
Fig. 53).
7.5.2

Frequency domain

Taking a Fourier transform with respect to the time-difference while keeping tw

fixed allows one to work in a mixed frequency-time domain. Since many experimental
set-ups are prepared to apply ac-fields it is particularly important to predict the aspect
fdt modification have when using these parameters. The condition t tw < 0 (tw )
to explore the fast relaxation roughly translates into 1 < 0 (tw ), i.e. for a fixed
waiting-time high frequencies are required. The longer the waiting time the lower
the frequency one has to use to see this scale since 0 (tw ) increases with tw . Instead,
when t tw > 0 (tw ) one has 1 > 0 (tw ), and very low frequencies are needed to
explore the slow scale. These conditions imply
0 (tw ) > 1
0 (tw ) < 1

fdt holds in the fast scale ,

fdt does not hold in the slow scale .

(7.26)

Reversing the argument above, if one weakly perturbs the sample with an ac-field of
a fixed frequency 1 at a chosen time tw , one can follow the deviation from fdt using
tw as the control parameter. This procedure yields the solid line on the right panel of
179

Fig. 56. Choosing now a lower frequency 2 (< 1 ) the crossover from the slow to the
fast regime occurs at a larger value of tw . One obtains then the dotted curve on the
right panel of Fig. 56. So on and so forth, the smaller the frequency of the applied
ac-field the longer the slow regime lasts and the longer one sees deviations from fdt.
(Note that the probe does not modify the dynamics.) In the Figure we chose to sketch
the behavior of a system with only two-time scales, in which the fdt ratio takes two
constant values separated at single breaking point in which the correlation reaches
the plateau value qea . This procedure is commonly employed experimentally where
we discuss the measurements of Grigera and Israeloff for glycerol.
7.5.3

Time-reparametrization invariant formulation

A more interesting way of displaying the modification of the fdt has been suggested by the analytic solution to the mean-field models discussed in Section . One
of its advantages is that it allows one to classify the systems into sort of universality
classes according to the form the fdt modification takes.
The analytic solution is such that, in the asymptotic limit in which the waitingtime tw diverges after N , the integrated linear response approaches the limit
lim
(t, tw ) = (C)
tw
C(t, tw ) = C

(7.27)

when tw and t diverge while keeping the correlation between them fixed to C. Deriving
this relation with respect to the waiting time tw , one finds that the opposite of the
inverse of the slope of the curve (C) is a parameter that replaces temperature in the
differential form of the fdt. Thus, using Eq. (7.27) one defines
kB Teff (C) ( (C))1 ,

(7.28)

that can be a function of the correlation. Under certain circumstances one can show
that this quantity has the properties of a temperature in the sense to be described
later.
One of the advantages of this formulation is that, just as in the construction of
triangle relations, times have been divided away and the relation (7.27) is invariant
under the reparametrizations of time defined in Eq. (7.10).
Equation (7.27) is easy to understand graphically. Let us take a waiting time tw ,
say equal to 10 time units after the preparation of the system (by this we mean that
the temperature of the environment has been set to T at the initial time) and trace
(t, tw ) against C(t, tw ) using t as a parameter (t varies between tw and infinity). If
we choose to work with a correlation that is normalized to one at equal times, the
parametric curve starts at the point (C(tw , tw ) = 1, (tw , tw ) = 0) and it arrives at
the point (C(t , tw ) C, (t , tw ) = ). Without loss of generality we can
assume that the correlation decays to zero, C = 0. This first curve is traced in red
in Figs. 57. Now, let us choose a longer waiting time, say tw = 100 time units, and
180

reproduce this construction. One finds the green curves in Figs. 57. Equation (7.27)
states that if one repeats this construction for a sufficiently long waiting time, the
parametric curve approaches a limit (C), represented by the blue curves.

t =100 t w>>teq
w

1/T

1/Teff (C)

=10

1/T

tw < tw < tw

0
0

Figure 57: The asymptotic behavior of the integrated linear response against the
correlation in a parametric plot, for fixed waiting time and using t as a parameter.
Left: behavior in equilibrium. Right: behavior in a slowly relaxing system out of
equilibrium. See text for an explanation.
When the system equilibrates with its environment, the construction approaches
a straight line with slope 1/(kB T ) as predicted by the fdt. This is the result shown
in the left panel of Fig. 57. Instead, for non-equilibrium systems evolving slowly the
asymptotic limit is different, it is given by a curve (C). For solvable fully-connected
models one distinguishes three families, as drawn in the right panel of Fig. 57. They
correspond to some systems undergoing domain growth (e.g. the O(N ) model in d = 3
when N ), systems behaving like structural glasses (e.g. the p-spin model) and
spin-glasses (e.g. the sk model). Several numerical studies in more realistic models of
the three cases tend to confirm this classification. However, two provisos are in order.
First, one has to be very cautious about the numerical results given the very short
time scales and rather small system sizes accessible in simulations. Second, as shown
in Section ??, at least one system that undergoes domain growth, the ferromagnetic
chain, has a non-trivial (C) like the one found for the sk model.
We have already found these asymptotic (C) curves when we discussed the dynamics of a (flat) harmonic oscillator in contact with a complex bath made of subsystems with different characteristic times and temperatures. Here we claim that the
same structure arises in a glassy model coupled to a white-bath. Different values of
the effective temperature are self-generated in the system.
This plot is invariant under reparametrisations of time t h(t) acting on the twopoint functions as in Eqs.(7.10). A different choice of the functions h only changes
the speed with which the (C) curve is traced but not its form.

181

7.5.4

fdt part

The Fokker-Planck formalism used to derive the fdt can be also be used to obtain
a bound on fdt violations. Indeed, one bounds the difference between response and
variation of the correlation with the Cauchy-Schwartz inequality leading to
p



kB T R( + tw , tw ) s C( + tw , s)|
(7.29)
s=tw c dtw H(tw )
R
where c is a constant and H(tw ) dqP (q, tw )(E(q) kB T ln P (q, tw )) is a positive
definite function that monotonically decreases towards the free-energy when the system eventually equilibrates. One finds a similar bound for Kramers processes and a
generalization that includes the power input when time-dependent or non-potential
forces are applied. For systems such that dtw H(tw ) 0 sufficiently fast when tw
the bound implies that the lhs vanishes in this limit. This can be achieved in two
ways: either each term is finite and the difference between them vanishes or each
term tends to zero independently. The former possibility is what happens in the fast
regime where fdt holds. The latter holds in the slow regime where both the response
and the variation of the correlation are very small but the relation between them does
not follow fdt. One derives a more useful bound by integrating (7.29) over time:
Z +tw
p
dt dt H(t ) . (7.30)
|kB T ( + tw , tw ) C( + tw , tw ) + C(tw , tw )| c
tw

The terms in the lhs are now always finite while the value of the rhs depends on
the relation between the time-difference and the waiting-time tw . For sufficiently
short such that the rhs vanishes fdt has to be satisfied in its integrated form. This
result explains the existence of a common straight-line with slope 1/(kB T ) in the
nonequilibrium curves in Fig. 57. For sufficiently long such that the rhs takes a
finite value fdt can be violated. In this second scale a departure from the straight
line of slope 1/(kB T ) can occur and it is indeed what happens in systems with slow
non-equilibrium dynamics, see the right panel in Fig. 57. One sees how a separation
of time-scales in the dynamics influences the fdt violations.
In driven systems the bound depends on the power input and only vanishes in
the limit of vanishing applied forces. The fdt is not even enforced in the fast scale
and deviations start as soon as C decays from 1. However, as we will see below, the
modification of fdt follow a very similar pattern to the one shown in Fig. 57 with
the strength of the applied force playing a similar role to the one of the waiting-time
here.
7.5.5

Diffusion

In these Lectures we focus on models with a bounded self-correlation for an observable with zero average that is normalised at equal times. If the averaged observable does not vanish but the equal-time correlation reaches a time-independent
182

limit one can still use the simple self-correlation in the generalisations of fdt. However, in more general diffusive model with an unbounded time-dependent equaltime correlator it is more natural to compare the behaviour of the displacement
(t, t ) C(t, t)+C(t , t )2C(t, t ) (that vanishes by definition at equal times) to the
linear response. In normal diffusion these are linked by R(t, t ) = 1/(2kB T )(t, t ).
In glassy models like the massless manifold in a random potential and others this
relation is modified.

183

Solution to mean-field models

In this Section we turn our attention to the solution to the Schwinger-Dyson

equations derived in previous Sections. We start by describing the simplest numerical
algorithm that solves these equations and we next briefly discuss the asymptotic
analytic solution at high temperatures. Next we describe in quite detail the solution
at low T .

8.1

Numerical solution

One can attempt a numerical solution to the set of causal integro-differential

equations (6.18), (6.19) together with the equation for the Lagrange multiplier (t).
One of the questions we would like to explore is whether they encode a non-equilibrium
evolution as the one we have already described.
The correlation C(t, t ) and response R(t, t ) are two-time quantities, that is, they
depend on t (which physically corresponds to the time of observation) and t (which
corresponds to the age of the system). In the simplest algorithm one discretises the
two-times plane with a uniform grid, t = j and t = i. The correlation and response
on the diagonal and the next-to-main diagonal of the two-times plane (i, j) are known
for all times.
The time-derivatives t2 C(t, t ) and t2 R(t, t ) in their discretized form are used
to update the two-point functions. Due to causality, to advance one time step, the
integrals only need values of C and R that are already known. This algorithm is
simple and efficient but it is severely limited by the computer storage capacity. Since
one has to store C and R for all previous time steps, the memory used grows as i2
and this number becomes rather quickly prohibitive. In standard pcs one can use
imax 104 , get an acceptable precision for 0.1 and reach total times of the order
of 103 .
In the quantum case the presence of non local kernels and , that appear convoluted with C and R, renders the numerical solution harder. The larger the cut-off
, the smaller the iteration step we need to compute these integrals with a good
precision. The maximum total time is of the order of 102 in this case.
A different starting point for a numerical solution is the single variable equation
(6.29). This route was followed by Eissfeller and Opper for spin-glasses and it is
usually used in the so-called dynamic mean-field theory. Again, this method is limited
by the storage capacity.
The knowledge of the qualitative features of the solution helps one devising a more
performant algorithm with a variable two-time grid. As we will see from the analytic
solution, C and R vary rapidly when times are near the diagonal j = i and decay
very slowly far from it. Kim and Latz have exploited this property and wrote such
an algorithm for the spherical p spin model reaching total times of the order of 108 .

184

Finally, one can think of an iterative search where one starts from a trial form of
C and R and uses the dynamic equations to extract the new form. One can expect to
obtain the solution by repeating this procedure until the iteration converges to a fixed
point. This method would allow one to look for solutions of the full set of Schwinger
- Dyson equations that break causality.
The numerical solution for the causal problem, found with the simple uniform grid,
has been of great help in deriving the asymptotic analytic solution. In the following
we describe how this solution builds up.

8.2

Solution at high temperatures

At high temperature the system equilibrates with its environment since

teq (N , T ) = finite .

(8.1)

The mere existence of an asymptotic limit implies that one-time quantities as, e.g., the
energy density, E(t), or the Lagrange multiplier, (t), have to approach an asymptotic
limit, limt E(t) = E and limt (t) = . In equilibrium E = Eeq and
similarly for all one-time quantities. Two time-quantities, as C and R, depend on
times but only through time differences.
To solve the high T dynamics one first assumes that after a transient equilibrium
is reached and a solution of the form (t) ,
C(t, t ) Cst (t t ) ,

R(t, t ) Rst (t t )

(8.2)

with Rst and C|scst related by fdt, for long waiting-times t and all time-differences
t t , exists. These properties also apply to D and that behave as a correlation
and a response, respectively. This Ansatz should solve Eqs. (6.18) and (6.19) when
T > Td , with Td the dynamic critical temperature. In order to prove it we take t
long and we assume that we can separate the integrals in Eqs. (6.19) and (6.18) in a
preasymptotic and an asymptotic contribution,
Z
Z teq
Z

dt .
(8.3)
dt +
dt
0

teq

Next, we assume that the two-point functions

decay as fast as to ensure that all
Rt
preasymptotic contributions vanish, e.g. 0 eq dt A(t, t )B(t , t ) 0 when t and
t t are in the asymptotic regime. Using the Ansatz (8.2) and this assumption the
integrals in the rhs of Eq. (6.19), for a classical problem, read
t

Z t

1 Cst (t t )
1 Dst (t t )
dt
+
Cst (|t t |)
dt Dst (t t )
kB T
t
kB T
t
teq
teq
Z t
Z t

1
1

Dst (t t )

[D
(t

t
)C
(t

t
)]
+
Cst (t t )
dt
dt
=
st
st
kB T teq
t
k B T t
t
Z

185

The first integral in the rhs is a total derivative and it can be readily evaluated, it
yields Dst (t t )Cst (0) Dst (t teq )Cst (t teq ) Dst ( ) where we assumed that t
and t are well in the asymptotic regime in such a way that Cst (t teq ) 0, and we
defined t t . Integrating by parts the last integral in the rhs one finally obtains
the high T equation for the correlation
Z
1
1
1
Go ( )Cst ( ) =
d Dst ( )d Cst ( )
(8.4)
Dst (0)Cst ( )
kB T
kB T 0
with G1
o ( ) = M d 2 + d + . One can check that Eq. (6.18) coincides with
Eq. (8.4) under the same assumptions. To prove this statement one has to integrate
Eq. (6.18) with respect to t from teq to t taking care of the fact that t appears in
the lower limit of the integral.
Equation (8.4) for the spherical p spin model coincides with the schematic mc
equation. This equation has a decaying solution above a sharp critical temperature
that we call Tmct = Td where the assumptions of tti and fdt are justified. After
a short transient (eliminated by the limit t teq ) the system equilibrates with its
environment even if the thermodynamic limit has already been taken. At very high T
the decay to zero is very fast and typical of, say, a high-T liquid. Closer to Td , however,
a very interesting structure appears. The solution takes the form sketched in the left
panel in Fig. 53. In a logarithmic scale one sees a two step relaxation develop with a
first relatively quick decay towards a plateau at a value that we call qea and next a
slower relaxation towards zero. The length of the plateau increases when temperature
d
,
approaches Td from above and it diverges at Td . At Td the height of the plateau, qea
d
follows from the asymptotic analysis of Eq. (8.4). If one loosely considers qea to be
d
an order parameter, the high temperature analysis yields qea
> 0 [see Eq. (8.30)] and
the transition is discontinuous. It is important to stress that, as we will see below,
this does not mean that the model has a first order thermodynamic transition. All
d
susceptibilities are continuous when going across Td even though qea
> 0. In the
mode-coupling literature these transitions are called type B.
The details of the asymptotic analysis of the schematic mc equation and its relation
with the behavior of real systems has been discussed at length in the literature.
We will not develop it here. With the purpose of future comparison with the lowT solution we just recall that the approach and departure from the plateau (beta
relaxation) occurs with two power laws:
d
Cst ( ) qea
+ ca a + . . .

given by

d
Cst ( ) qea
cb b + . . .

(8.5)

d
)
1 2 (1 a)
1 V (qea
1 2 (1 + b)
.
=
=

d
kB Td (1 + 2b)
kB Td (1 2a)
2 (V (qea ))3/2

(8.6)

A similar analysis can be done for a quantum model.

186

8.3

Solution at low-T

Three families of mean-field models have been found so far. In this Section we
present the solution to the spherical mean-field descriptions of ferromagnetic domain
growth and structural glasses in some detail. We use a generic notation that allows
us to treat the classical and quantum problem simultaneously. The presentation
follows. By the end of this Subsection we discuss the generalisation of these results to
models of spin-glass type, models with spatial dependence and the effect of different
microscopic dynamics.
The numerical solution to the dynamic equations at low T shows no evidence for
an arrest in the waiting-time dependence of the decay of C and R. In this regime of
temperatures,
teq (N, T < Td )
(8.7)

and the equations do not admit the choice of a t > teq . In order to consider the
crossover towards the equilibration regime one should revisit the derivation of the
dynamic equations allowing for N finite. This program has not been pursued in the
literature and it remains one of the most interesting open problems in the field.
8.3.1

The Lagrange multiplier

We approximate the integral in Eq. (6.23) by separating its support in three intervals
t : 0 0 , t : 0 t , t : t t .
(8.8)

The first time-interval contains only finite times t . Hence, all correlations and responses of the form C(t, t ) and R(t, t ) vanish due to Eqs. (7.3) and (7.7). In the last
time-interval t is close to t in the sense that correlations of the kind C(t, t ) are of
the form Cst (t t ) + qea and similarly for the responses. Finally, in the intermediate
time-interval the C and R vary in the aging regime. Of course, we are sloppy in that
we do not precise what are the values of 0 and t . The definitions of correlation
scales given in Section correct this imprecision exchanging the time limts by limits
in the correlation. Within these assumptions the asymptotic value of (t) is given by
Z
Z
q
d st ( ) + D
d Rst ( )
= A + qea
ea
0
0
Z
d [ st ( )Cst ( ) + Dst ( )Rst ( ) ] + Last
(8.9)
+
0

and D are made of two terms, one contribution from the bath and one
 contribution
from the interactions. We called Last a term that equals M 2 Cst ( ) 0 in a model
with inertia (classical or quantum) and simply kB T in classical models without inertia.
A is the aging contribution:
Z t
dt [ ag (t, t )Cag (t, t ) + Dag (t, t )Rag (t, t ) ] . (8.10)
A = lim
t

187

The bath does not contribute to the integrals in A when the kernels and decay
sufficiently fast to zero as to yield vanishing integrals. This is trivially true for a white
noise. It can be a working assumption for colored noises based on a weak limit of the
strength of the
R tcoupling to the noise. More precisely, we are neglecting terms of the
form limt 0 dt A(t t )B(t, t ) where A is either or and B is either Cag or
Rag . In this case
Z t
h
i
ag (t, t )Cag (t, t ) + D
ag (t, t )Rag (t, t ) . (8.11)
dt
A = lim
t

The second and third terms in Eq. (8.9) come from the constant (non-zero) limit of the
q
first decay of the correlation qea limtt limt C(t, t ) and the vertex D
ea
p p1

limtt limt D(t, t ). For the classical spherical p spin model Dqea = 2 qea and
this equation also holds for its quantum extension if we use lim Rst ( ) qea , a
property of the wltm scenario. The integral over the stationary parts can be simply
performed using fdt for classical problems but they cannot in quantum problems.
8.3.2

The stationary regime

If (t, t ) are such that C(t, t ) > qea , C(t, t ) = qea + Cst (t t ) and R(t t ) =
Rst (t t ). The Schwinger-Dyson equation for R in this time sector reads
Z


d st ( )Rst ( )
(8.12)
M 2 + Rst ( ) = ( ) +
0

and it keeps the same form as in the high-temperature phase, apart from the fact
that the constant has contributions from the aging regime. The Schwinger-Dyson
equation for C reads
Z
Z

d st ( ) + Dqea
d Rst ( )
M + (qea + Cst ( )) = A + qea
0
0
Z
d [st ( + )Cst ( ) + Dst ( + )Rst ( )] .
(8.13)
+

One can now Fourier-transform both equations

Rst ()

M 2 + Cst () + qea ()

M 2

1
,
+ st ()


q Rst () ()
A + qea st () + D
ea

+st ()Cst () + Dst ()Rst () .

The formal solution to the equation for Cst is



q Rst () ()Rst () + Dst ()|Rst ()|2 .
Cst () = qea + A + qea st () + D
ea
188

The first term on the rhs has an imaginary and a real part. The imaginary part
vanishes identically since, due to fdt, both ImRst () and Imst () are proportional
to tanh (
h/2) which is zero at = 0 for classical and quantum problems. Concerning the real part of this first term, as we have assumed that Cst ( ) goes to zero
for , we need to impose the self-consistent condition
q Rst ( = 0) = 0 .
qea + A + qea st ( = 0) + D
ea

(8.14)

This is the condition that fixes qea . We will find it again in the next section as the
matching condition between the stationary and aging regimes. The final equation for
Cst () is
Cst () = Dst ()|Rst ()|2 .
(8.15)
One can check that these calculations are consistent with the results from . Actually, the integrals in equation for (t) involving the stationary parts can be evaluated
with the help of the equations for Rst and Cst , Eqs. (8.14) and (8.15), and yield once
again Eq. (8.14).
st and
Similarly to the high-temperature case one can now show that fdt for
st implies fdt for Rst and Cst . The remainder of the proof, i.e. to show that fdt
D
st and D
st depends only upon the form
between Rst and Cst implies fdt between
st and D
st as functions of Rst and Cst and is not modified from the one discussed
of
in Section .
8.3.3

The aging regime

If we now choose the times t, t to be well-separated so as to have C(t, t ) =

Cag (t, t ) qea and R(t, t ) = Rag (t, t ), the web and wltm hypotheses allow us to
throw the second time derivatives on the lhs. We assume that their contribution
is much weaker than the one of each of the integral terms on the rhs. This is an
assumption that we have to verify at the end of the calculation, once the solution for
Cag and Rag is known. It corresponds to the over-damped limit.
Using a separation of time-scales in the integrals the equation for R in the aging
regime becomes
Z
Z
ag (t, t )
d st ( )
d Rst ( ) + Rag (t, t )
Rag (t, t ) =
0
0
Z t
ag (t, t )Rag (t , t )
dt
(8.16)
+
t

and we call it the Rag -eq. Similarly, the equation for C becomes
Z
Z
ag (t, t )
d Rst ( )
d st ( ) + D
Cag (t, t ) = Cag (t, t )
0

ag (t, t )Cag (t , t ) +
dt

189

ag (t, t )Rag (t , t ) (8.17)

dt D

and we call it the Cag -eq. In all integrals over the slow regime we neglected the
ag (t, t )
contributions of the noise kernels and and we approximated ag (t, t )

and Dag (t, t ) Dag (t, t ).

8.3.4

The Edwards-Anderson parameter

The Edwards-Anderson parameter, qea , is determined self-consistently from the

matching of limt Cag (t, t) = limtt limt C(t, t ) = qea . Taking the limit
t t in the Rag -eq and Cag -eq one obtains
Z
Z
ag (t, t)
d st ( ) ,
(8.18)
d Rst ( ) + Rag (t, t)
Rag (t, t) =
0
0
Z
Z
ag (t, t)
d Rst ( ) .
(8.19)
d st ( ) + D
qea = A + qea
0

ag (t, t) is proportional
The first equation admits the solution Rag (t, t) = 0 since
to Rag (t, t). This corresponds to the high-temperature solution where there is no aging
regime. Here we concentrate on the other possibility. The response becomes smaller
and smaller as time passes though its integral over an infinite interval gives a finite
contribution. If we neglect all terms that are proportional to Rag (t, t) with respect
to terms that are proportional to qea , only the first term in the power expansions of
and D
survive and

/R

lim

qea

ag (t, t)

Rag (t, t)

ag (t, t)
q lim D
D
ea

(8.20)

q = p q p1 ,
D
ea
2 ea

(8.21)

that for the p spin model become



/R

qea

p(p 1) p2
qea
2

in accord with the large limit of the stationary values (see Section ). Equations
(8.18) and (8.19) become
Z

 Z

d st ( ) ,
(8.22)
=
d Rst ( ) +
/R
qea 0
0
Z
Z
q
d st ( ) + D
d Rst ( ) .
(8.23)
qea = A + qea
ea
0

The second equation is the same as the one arising from the end of the stationary
regime, Eq. (8.14).
From Eqs. (8.14) and (8.15) one derives
Z
1
d Rst ( ) = Rst ( = 0) =
,
(8.24)

st ( = 0)
0
190

and
1 =

/R

qea

2
Rst
( = 0) .

(8.25)

2
We remind that the factor Rst
( = 0) can be written in terms of the stationary
correlation function using fdt; therefore this is a closed equation for the correlation
that determines qea . In the case of the p-spin model it reads


2

Z
d
p(p 1) p2 1
h

Cst ( )
1 =
.
(8.26)
tanh
qea
P

2
h
2

In the classical case, the integral can be readily computed and the final equation for
qea is
p(p 1) p2
qea (1 qea )2 = (kB T )2 ,
(8.27)
2
that coincides with the result for the purely relaxational dynamics. For p 3 fixed, qea
is a function of temperature. Equation (8.27) can be solved graphically. The lhs has
max
= (p 2)/p.
a bell shape. It vanishes at qea = 0, 1 and it reaches a maximum at qea
The equation has two solutions for all temperatures (kB T )2 < (kB T max )2 = p(p
1)/2 [(p 2)/p]p2 (2/p)2 , these merge ar T max and disappear for higher T s. The
physical solution corresponds to the branch on the right of the maximum, the one
that continues the solution qea = 1 at T = 0. The minimum value of qea is reached
d
max
at the dynamic critical temperature Td (< T max ), where qea
qea (Td ) > qea
.
8.3.5

Fluctuation - dissipation relation

In order to advance further we have to relate the response to the correlation. If

we assume that
1 Cag (t, t )
Rag (t, t ) =
,
(8.28)
kB T
t
with T the value of an effective temperature (see Section ??) that is determined by
ea
q Rst ( = 0). Using Eq. (8.28) and the
+D
Eqs. (8.23) and (8.24) 0 = A Rst q(=0)
ea

ag ,
equivalent relation
between

and
D
we obtain A
=
ag


1
1

(kB T ) limt Dag (t, t)Cag (t, t) = (kB T ) qea Dq and

ea

1
(p 2)
=
Rst ( = 0) =
kB T
qea

2(p 2)2 p/2

qea .
p(p 1)

(8.29)

In the classical limit T /T = (p 2)(1 qea )/qea . Note that both in the classical
and quantum case, T if p = 2. Since the case p = 2 is formally connected to
ferromagnetic domain growth in d = 3 (in the mean-field approximation) there is no
memory neither in the classical nor in the quantum domain growth.
The Ansatz in Eq. (8.28) solves classical and quantum aging equations. The modification of the fdt in this regime became thus classical even when quantum fluctuations
191

exist. This is an interesting sort of decoherent effect that will become clearer when
we will discuss the interpretation of this results in terms of effective temperatures.
Using Eq. (8.28) for all values of C below qea we assumed that there is only one
aging correlation scale in the problem. Interestingly enough, one do a more general
analysis using the formalism described in Section and find that the dynamic equations
force the solution to have only one aging correlation scale.
8.3.6

Discontinuous classical transition

The classical dynamic critical point (Td , h = 0) can arise either when qea 0 or
when T T . For the p spin model, using Eqs. (8.27) and (8.29) the latter holds
and
p (p 2)p2
p2
d
(kB Td )2 =
qea
=
.
(8.30)
2 (p 1)p1
p1
The transition is discontinuous since the order parameter qea jumps at Td . However, it is still of second order thermodynamically since there are no thermodynamic
discontinuities, all susceptibilities being continuous across Td . For instance,
lim (t, tw ) =

ttw

1
1
1
(1 qea ) +
qea
when T T at Td .

kB T
kB T
kB T

(8.31)

The dynamic transition occurs at a value Td that is higher than the static transition temperature Ts . The latter is fixed as the temperature where replica symmetry
breaking occurs (using the standard prescription to fix the parameters in the Parisi
Ansatz to compute the free-energy density). This feature is an explicit realisation of
the discussion on Tg and T0 . They are sharp in this model.
8.3.7

The classical threshold level

R
The asymptotic energy density reads E = p1 0 dt [(t, t )C(t, t )+ D(t, t )R(t, t )]
where we used Eq. (6.26). Replacing the solution found above we obtain


1
1
1
p
p
(1 qea ) +
q
Eth .
(8.32)
E =
2 kB T
kB T ea
If one compares this expression with the equilibrium energy density, found studying
the partition function, one discovers that
E = Eth > Eeq .

(8.33)

Thus, the non-equilibrium dynamics does not approach the equilibrium level asymptotically but it reaches a threshold level that is extensively higher than equilibrium
(note that the inequality (8.33) holds for the energy density). The name threshold is
motivated by a similarity with percolation (in phase space).
8.3.8

Two p models

192


In Section we took the limit t t , or equivalently, Cag qea
in the equations
for the slow part of the response and the correlation and this lead us to Eqs. (8.25)
and (8.29) for qea and T . Let us now take subsequent variations of this equation
with respect to the correlation and evaluate them in the same limit. It is easy to see
that if we neglect the contributions from the integral between t and t, assuming that
the integrands are analytic in this limit, we get new equations linking T and qea
that, for generic models, are not compatible. Indeed, as we will see below, the pure
spherical p spin model is the only one for which the solution is given by an analytic
function 1 (x) when x 1 .
The way out of this contradiction is to propose that the correlation approaches
the plateau at qea with a power law decay and that it departs from it with another
non-trivial power law:

(i)

Cst (t t )

Cag (t, t )

a
2a
(1 qea ) + c(1)
+ c(2)
+ ......
a (t t )
a (t t )




b
2b
h(t )
h(t )
(1)
(2)
1
qea cb
1
cb
+ ...
h(t)
h(t)

(8.34)
(8.35)

(i)

with ca and cb constants. If the exponent b is smaller than one, the integrals
generated by taking derivatives with respect to Cag do not vanish when t t . The
expansion of the stationary and aging equations around qea fix the exponents a and
b. One finds
1 ((1 a))2
1 V (qea )
1 ((1 + b))2
=
=

kB T (1 + 2b)
kB T (1 2a)
2 (V (qea ))3/2

(8.36)

that are to be confronted to Eqs. (8.5) and (8.6) for the high T behavior. We recall that
V(C) is the correlation of the random potential. Importantly enough, the exponents
a and b are now T -dependent and they are related via an equation in which T enters.
Classical spherical p spin model
Since V(C) = C p /2 using Eqs. (8.25) and (8.29) to fix T and qea one finds
((1 + b))2 /(1 + 2b) = 1/2 and b = 1 for all T < Td . The exponent a interpolates
between a = 1/2 at T 0 and a = 1 at T Td since ((1a))2 /(12a) = T /(2T ).
Classical mixed p1 + p2 spherical spin model
For adequate choices of the coefficients in V(C) = a1 /2C p1 + a2 /2C p2 (see below)
one finds T -dependent exponents a(T ) and b(T ).
Ultrametric limit
It is interesting to notice that ((1 + b))2 /(1 + 2b) is bounded by one. Thus,
Eq. (8.36) constrains the random potentials for which a solution with only two correlation scales exists. For a particle in a power-law correlated random potential one
sees the transition towards an ultrametric-like solution arrives when the potential
goes from short-range to long-range correlated. To our knowledge this has not been
193

found in a static calculation. An interpretation of the exponents a and b, and this

consequence, in terms of the properties of the tap free-energy landscape is not known
either.
8.3.9

sk model and similar

A different family of models, to which the sk model belongs, are solved by an

ultrametric Ansatz, C31 = f (C32 , C21 ), for all correlations below qea . The (C) plot
yields a non-trivial curve (instead of a straight line) for C [0, qea ]. The transition
d
is continuous qea
= 0. These models are called type A in the mct literature.
Indeed, for a generic disordered model with random potential correlated as in
Eq. (6.6), one finds that the solution is ultrametric if and only if

3/2
V (C) V (qea )
<1.
(8.37)
V (qea ) V (C)
This bound constrains, for instance, the values of the coefficients in a polynomial
random potential for which the dynamic solution is ultrametric. The fdt is modified
with a C dependent factor given by
s
T
V (qea )
.
(8.38)
= qea V (C)
Teff (C)
4(V (C))3/2
8.3.10

Mode dependence

The models we solved so far have no spatial dependence. The manifold problem
has an internal structure that leads to a mode-dependence. This model has been
solved for generic potential correlations. We summarize the outcome without presenting its detailed derivation. All modes are slaved to one in the sense that one has
to solve for the dynamics of one of them and the mode-dependence follows from an
algebraic equation. The value of the effective temperature does not depend on the
mode. The mathematical reason for this is the slaved structure of the equations. The
physical reason is that all interacting observables evolving in the same time-scale have
to partially equilibrate and acquire the same effective temperature. The height of the
plateau, qea , is a k dependent quantity. The approach to it and departure from it
also depends on k but only via the prefactors; the exponents a and b, see Eqs. (8.34)
and (8.35, are the same for all modes.
Mode-couling equations including a wave-vector dependence have been derived
by Latz using the Mori-Zwanzig formalism; the structure of the solution to these
equations shares the properties just described.
8.3.11

Quantum fluctuations

The simplest effect of quantum fluctuations is to introduce oscillations in the first

step of relaxation. These disappear at long enough time-differences and they are
totally suppressed from the second decay, that superficially looks classical.
194

The Edwards-Anderson parameter qea depends upon T and h. As expected, quantum fluctuations introduce further fluctuations in the stationary regime and they
decrease the value of qea , qea (T, h 6= 0) < qea (T, h 0).
The modification of fdt in the quantum model is of a rather simple kind: Rag
and Cag are related as in the classical limit. For the quantum extension of the p spin
model there are two correlation scales, one with the temperature of the environment,
T , the other with another value of the effective temperature, T , that depends on T ,
h and the characteristics of the environment. This is a kind of decoherent effect.

As regards to the transition from the glassy to the liquid or paramagnetic phase,
an interesting effect appears. Keeping all other parameters fixed, the plane (T,
h2 /(JM )) is separated in these two phases by a line that joins the classical dynamic

critical point (Td , = 0) and the quantum dynamic critical point (T = 0, d ). Close
to the classical dynamic critical point the transition is discontinuous but of second
order thermodynamically until it reaches a tricritical point where it changes character
to being of first order. This behavior is reminiscent of what has been reported for the
quantum spin-glass by Aeppli, Rosenbaum and collaborators.
A still more dramatic effect of quantum mechanics is related to the very strong role
played by the quantum environment on the dynamics of a quantum system. Indeed,
the location of the transition line depends very strongly on the type of quantum bath
considered and on the strength of the coupling between system and environment.
8.3.12

Driven dynamics

The effect of non potential forces can be mimicked with a non-symmetric force
with strength playing an analogue role to the shear stress . For strengths that are
not too strong, the dynamics presents a separation of time scales with a fast approach
to the plateau and a slow escape from it that is now, however, also stationary. Indeed,
after a characteristic time tsh the full dynamics becomes stationary though the system
is still far from equilibrium. One defines a structural relaxation time, , as the
time needed to reach, say, a correlation
equal to a half. One relates the structural
R
relaxation to the viscosity via dt C(t). The scaling of with the shear rate
/ has been successfully confronted to the behavior in rheological experiments
in super-cooled liquids and glasses. In terms of the general scalings discussed in
Section , the correlations are characterised by two different functions , one for the
fast decay towards the plateau and another for the slow decay from the plateau, while
the functions h(t) are simple exponentials.
Interestingly enough, from the study of fdt violations above (though close to)
and below Td , when the forcing is weak, one extracts a still well-defined slope of the
(C) plot when C evolves in the slow scale. This means that an effective temperature
can also be identified for these systems kept explicitly out of equilibrium.
Oscillatory forces, as the ones used to perturb granular matter, have a different
effect. Aging is not stopped in a finite region of the phase diagram (T -strength of
the force-frequency of the force). An effective temperature can still be defined as the

195

slope of the (C) plot drawn using stroboscopic time, with a point per oscillatory
cycle.

196

Effective temperature measurements

In this Section we discuss measurements of FDT violations and tests of the effective
temperature notion in a variety of physical systems out of equilibrium. Since we
cannot make the description exhaustive we simply select a number of representative
cases that we hope will give a correct idea of the level of development reached in the
field. This Section follows closely where references to the original papers are given.

9.1

Diffusion

The dynamics of a particle in a potential and subject to a complex environment

(colored noise or baths with several time-scales and temperatures) has a pedagogical
interest but also admits an experimental realization in the form of Brownian particles
immersed in, e.g., colloidal suspensions and controlled by optical tweezers.
A particle coupled to a bath in equilibrium at temperature T with noise-noise
correlations of type h(t)(t )i (tt )a1 , 0 < a < 2, and under no external forces,
performs normal or anomalous diffusion depending on a. The fluctuation-dissipation
ratio, eq. (??), for t t is
Xxx (t, t ) =

D(t t )
T Rxx (t, t )
=
,

t Cxx (t, t )
D(t t ) + D(t )

(9.1)

with the diffusion coefficient D(t) 1/2 dhx2 (t)i/dt ta for a 6= 1 and D(t) = ct
for a = 1. In such colored noise cases Xxx is a non-trivial function of times and it
does not seem to admit a thermodynamic interpretation. Still, for later reference we
consider the long times limit:

a<1
subOhmic,
0
1/2
a=1
Ohmic,
X = lim lim Xxx (t, t )
t t

1
a>1
superOhmic.

Another illustrative example is the non-Markovian diffusion of a particle in a

harmonic potential and subject to different external baths. As already explained in
Sect. ?? this simple system allows one to show how different environments can impose
their temperatures on different dynamic regimes felt by the particle. Tests of other
definitions of out of equilibrium temperatures in this simple case confirmed that the
definition that appears to have the most sensible behaviour is the one stemming from
the long-time limit of the relations between induced and spontaneous fluctuations.
All other definitions yield results that are more difficult to rationalize: in most cases
one simply finds the temperature of the fast bath and in some cases, as with a static
limit, one incorrectly mixes different time regimes even when their time-scales are
well separated.

197

9.2

Coarsening

When a system is taken across a second order phase transition into an ordered
phase with, say, two equilibrium states related by symmetry, it tends to order locally
in each of the two but, globally, it remains disordered. As time elapses the ordered
regions grow and the system reaches a scaling regime in which time-dependencies
enter only through a typical growing length, L(t). Finite dimensional coarsening
systems have been studied in great detail from the effective temperature perspective.
In this context, it is imperative to distinguish cases with a finite temperature phase
transition and spontaneous symmetry breaking from those with ordered equilibrium
at T = 0 only. Some representative examples of the former are the clean or dirty 2d
Ising model with conserved and non-conserved order parameter. An instance of the
latter is the Glauber Ising chain and we postpone its discussion to Sect. .
Let us focus on scalar systems with discrete broken symmetry. When timedifferences are short with respect to a function typically algebraic of the typical
growing length L(tw ), domain walls remain basically static and the only variation is
due to thermal fluctuations on the walls and, more importantly, within the domains.
This regime is stationary, and induced and spontaneous fluctuations are linked by the
FDT. At longer time-differences domain walls move and observables display the out
of equilibrium character of the system.
The motion of the domain walls in the presence of an external perturbing random
field introduced to measure the staggered response is due to two competing factors:
on the one hand, the system tends to diminish the curvature of the interfaces due to
surface tension, on the other hand the random field tends to pin the domain walls in
convenient places.
The correlation and total susceptibility in the tw limit separate in two
contributions C(t, tw ) = C st (ttw )+C (1) (t, tw ) and (t, tw ) = st (ttw )+(1) (t, tw ).
Numerical studies of Teff focused on the parametric construction (C, tw ) at fixed
and finite tw where the chosen observable is the spin itself. The resulting plot has
a linear piece with slope 1/T , as in eq. (??), that goes below C = qea = m2 and,
consistently, beyond = [1 m2 ]/T . The additional equilibrium contribution is due
to the equilibrium response of the domain walls that exist with finite density at any
finite tw . In the truly asymptotic limit their density vanishes and their contribution
disappears. Consequently, limtw (C, tw ) = C st qea satisfies FDT and it is
entirely due to fluctuations within the domains. In cases with L(t) t1/zd , the slow
terms take the scaling forms

f (t/tw ) .
(1) (t, tw ) ta
w

C (1) (t, tw ) fC (t/tw ) ,

(9.2)

It would be natural to assume that (1) (t, tw ) is proportional to the density of defects
d (t) L(t)n tn/zd with n = 1 for scalar and n = 2 for vector order parameter.

198

Although this seems plausible a is instead d-dependent.

n (d dL )/(dU dL )
n (with ln corrections)
zd a =

Another conjecture is
d < dU ,
d = dU ,
d > dU .

(9.3)

dL is the dimension at which a vanishes and may coincide with the lower critical dimension. One finds dL = 1 in the Ising model, dL = 1 in the Gaussian approximation
of Ohta, Jasnow and Kawasaki, and dL = 2 in the O(N ) model in the large N limit.
dU is the dimension at which a becomes d-independent and it does not necessarily
coincide with the upper critical dimension. One finds dU = 3 in the Ising model,
dU = 2 in the Gaussian approximation, and dU = 4 in the large N O(N ) model.
It was then suggested that dU might be the highest d at which interfaces roughen.
In all cases in which a > 0, Teff . This result was confirmed by studies of
second order FDRs in the 2d Ising model that showed the existence of stationary contributions verifying the non-linear equilibrium relation and aging terms that satisfy
scaling and yield Teff as in the linear case. The approach by Henkel et al. based
on the conjecture that the response function transforms covariantly under the group
of local scale transformations, fixes the form of the scaling function f but not the
exponent a and does not make predictions on Teff . The coincidence between statics
and dynamics, see Sect. ??, holds in these cases.
Noise induced spatial fluctuations in the effective temperature of clean coarsening
systems were analyzed in the large N O(N ) model with d > 2 and with numerical simulations. The first study shows that time-reparametrization invariance is
not realized and that Teff is trivially non-fluctuating in this quasi-quadratic model.
The second analysis presents a conjecture on the behaviour of the average over local
(coarse-grained) susceptibility at fixed local (coarse-grained) correlation that consistently vanishes in coarsening (but is more interesting in critical dynamics as we will
discuss in Sect. ).
The results gathered so far and summarized in the conjecture (9.3) imply that
the FD ratio vanishes and thus Teff diverges in quenches into the ordered phase of
systems above their lower critical dimension.

9.3

Critical dynamics

The non-equilibrium dynamics following a quench from the disordered state to the
critical point consists in the growth of the dynamical correlation length, (t) t1/zeq .
This length does not characterize the size of well defined domains but the size of a
self-similar structure of domains within domains, typical of equilibrium at the critical
point. A continuum of finite time-scales associated to different wave-vectors, (k)
k zeq , exists with only the k 0 diverging. At any finite time t, critical fluctuations
of large wave-vectors, k(t) 1, are in almost equilibrium, while those with small
wave-vectors, k(t) 1, retain the non-equilibrium character of the initial condition.
This finite-time separation, and the fact that the order parameter vanishes, leads to

199

the multiplicative scaling forms

C(t, tw ) (t tw )d+2 fC [(t)/(tw ), 0 /(tw )] ,
(t, tw ) (t tw )d+2 f [(t)/(tw ), 0 /(tw )] ,
with the microscopic length 0 ensuring the normalization of the correlation and the
fact that vanishes at equal times. These forms imply that beyond the initial equilibrium part, the (C) plot assumes a non-trivial shape that, however, progressively
disappears and approaches the equilibrium linear form at all C > 0. The limit C = 0
is distinct and the limiting parameter X should be non-trivial and it was conjectured to be universal in the sense of the renormalization group. Whether this one
can be interpreted as a temperature is a different issue that has been only partially
discussed. For this reason, we keep the notation X (instead of Teff ) in most of this
section.
The correct estimation of X has to take into account that the number of out of
equilibrium modes decreases in the course of time (contrary to what happens in the
random manifold problem in the large N limit, for example). The best determination
of X is achieved by selecting the k 0 mode. A thorough review of the properties of
X found with the perturbative field-theoretical approach and some exact solutions
to simple models, as well as the comparison to numerical estimates, is given in. At
the Gaussian level the X of local operators (e.g. powers of the field, first derivatives
of the field, etc.) is independent of the chosen pair but recalls certain features of
the initial condition and the correlations of the environment. This is not the case for
non-local operators as, e.g. the energy or the tensor. Moreover, when fluctuations
are taken into account with, e.g. a two-loop or first order in = 4 d expansion, the
X of local operators is found to depend upon the observables.
In the scalar model, using the field itself as the observable, one finds the diffusive
results, eq. (9.2), at the Gaussian level and corrections when higher orders are taken
into account. For example X = 0.30(5) in d = 2, X = 0.429(6) in d = 3 for a
quench from a disordered state, white noise and up to second order in 4 d. The
trend of X increasing with d was found in other models too. Instead, X 0.78
(d = 3) and X = 0.75 (d = 2) if the initial state is magnetized. A larger X

= Tc /X and the comparison between these values conforms to

implies a lower Teff
the intuitive idea that an ordered initial state leads to a lower effective temperature
than a disordered one. X was found to increase with N in vector models.
A different type of critical phenomena (infinite order) arises in the 2d XY model.
The magnetic order parameter vanishes at all T but there is a low-T critical phase
with quasi long-range order (power-law decaying spatial correlations) that is destroyed
at Tkt where vortices proliferate and restore a finite correlation length. Out of equilibrium the critical scaling forms apply although with a temperature-dependent exponent, (T ), and a growing length scale (t) (t/ ln t)1/2 (the logarithm is due to
the effect of vortices). The r
ole of the EA order parameter is played by the asymptotically vanishing function (tw / ln tw )(T )/2 and the crossover between equilibrium
and out of equilibrium regimes takes place at a tw -dependent value of the correla200

tion. The (C, tw ) plot at finite tw is curved, it does not reach a non-trivial master
curve for tw , but Teff (t, tw ) = fX [(t)/(tw )]. Quenches from the disordered
phase, T0 > Tkt and heating from a T0 = 0 ground state to T < Tkt demonstrate
that the slow modes Teff depends on the initial state and it is higher (lower) when
T0 > T (T0 < T ). We allow ourselves to use the name Teff in this case since these
results point in the direction of justifying its thermodynamic meaning. Similar results
were obtained for 1 + 1 elastic manifolds with and without quenched disorder. As
the dynamic-static link is concerned, Berthier et al. evinced that the extension to
finite-times finite-sizes works, at least at not too high T s where free vortices inherited
from the initial condition are still present.
The exact calculation of the joint probability distribution of the finite-size correlation and linear response in the spherical ferromagnet quenched to its critical temperature was given by Annibale and Sollich. The results prove that these fluctuations
are not linked in a manner akin to the relation between the averaged quantities,
as proposed by Chamon et al., see Sect. ??, for glassy dynamics. The analysis of
correlation-susceptibility fluctuations in non-disordered finite-dimensional ferromagnets quenched to the critical point showed that the restricted average of the susceptibility, at fixed value of the two-time overlap between system configurations, obeys
a scaling form. Within the numerical accuracy of the study the slope of the scaling
function yields, in the asymptotic limit of mostly separated times, the value X .
The first experiments testing fluctuation dissipation deviations in a liquid crystal
quenched to its critical point appeared recently and the results are consistent with
what has been discussed above.
The coexistence of a single time scale in the aging regime together with a smooth
and time-dependent (C, tw ) plot arises naturally in a critical regime and it is due to
the lack of sharp time-scale separation.
Although many evaluations of X in a myriad of models tend to confirm that
it mostly behaves as a critical property, the thermodynamic nature of this parameter has not been explored in full extent yet. Measurements with thermometers and
connections to microcanonical definitions have not been performed at critical points.

9.4

Quenches to the lower critical dimension

The kinetic Glauber-Ising spin chain is the prototype of a dynamic model at its
lower critical dimension. Taking advantage of the fact that this is one of the very
few exactly solvable models of non-equilibrium statistical mechanics, several issues
concerning the effective temperature interpretation have been addressed in this case,
notably the observable dependencies.
After a quench from T0 to T = 0 the correlation and response vary in a
single time-scale with a simple aging scaling (they are both functions of t/tw ) and
the (C) relation is a continuous function. The factor Xs (t, t ), associated to the spin
correlation and susceptibility, is smaller than or equal to one and its value Xs in
the limit Cs 0 evolves smoothly from 1/2 (as in models characterized by simple
diffusion such as the random walk or the Gaussian model) to 1 (equilibrium) as t/eq
201

grows from 0 to [1/eq = 1 tanh(2J/T ) is the smallest eigenvalue of the master

equation operator]. Moreover, Xs is an exclusive function of the auto-correlation Cs
as in more complex instances of glassy behaviour.

The value for the long-wavelength analogue, the fluctuating magnetization, Xm

, is

identical to the local value Xs . The physical origin of the local-global correspondence,
which can also be obtained by field-theoretic arguments, is that the long wavelength
Fourier components dominate the long-time behaviour of both quantities. In contrast,
observables that are sensitive to the domain wall motion have Xd = 0, the difference
residing on the interplay between criticality and coarsening, a peculiar feature of
models with Tc = 0.
The dependence on the initial condition is also interesting. A non-zero initial
magnetization does not change the value of Xs at T = 0. Instead, demagnetized
initial conditions with strong correlations between spins so that only a finite number
of domain walls exist in the system, yield Xs = 0 (the same result is found in the
spherical ferromagnet). The deviations from non-linear FDTs have not been fully
analyzed yet.
The static-dynamics connection sketched in Sect. ?? does not hold in the 1d Ising
chain and the non-trivial (C) cannot be used to infer the properties of the equilibrium state. Indeed, the aging part of the response is finite asymptotically while
the equilibrium P (q) has a double-delta (RS) structure as in higher dimensions. The
reason for the failure is that the hypotheses used to derive the connection are not
fulfilled.
The large N O(N ) model in d = 2 shares many common features with the phenomenology described above although it has not been studied in as much detail.
To sum up, a quench to T = 0 at the lower critical dimension does not seem to be
the dimensional continuation of a line of critical quenches in the (T, d) plane (as often
implicitly assumed), but the continuation of a line of T = 0 quenches: the system
behaves as in the coarsening regime, although X 6= 0 for observables that do not
focus on the domain wall dynamics.

9.5

Relaxation in structural glasses

In particle glassy systems a separation of time-scales exists although it is not as

sharp as in mean-field models or coarsening systems, at least within simulational and
experimental time-scales. In atomic glasses the existence of an FDT part implies that
the rapid particle vibrations within the cages occur in equilibrium while the structural
relaxation is of a different out of equilibrium kind, and it is not necessarily ruled by the
temperature of the bath. Tests of the thermodynamic origin of fluctuation-dissipation
violations in the aging regime of these systems were carried through in much greater
detail and we summarize them below.
9.5.1

Simulations of microscopic models

202

Mono-atomic and binary Lennard-Jones mixtures, soft sphere systems, and the
BKS potential for silica are standard models for glass forming liquids. Both Monte
Carlo and molecular dynamics simulations suggest that the three first cases belong to
the RFOT class of systems defined in Sect. ?? with Teff = T (1) constant in the aging
regime. T (1) depends weakly on the bath temperature and systems parameters but it
does not on the preparation protocol as demonstrated by measurements after quenches
and crunches or the microscopic dynamics. Tests of partial equilibration between
fluctuations at different wave-vector gave positive results. Importantly enough, these
models have a well defined equilibrium behaviour and their energy density is naturally
bounded. Of special interest is the numerical method devised to compute linear
responses in molecular systems with high precision that allowed Berthier to resolve
the paradoxical behavior previously reported for silica.
Numerical evidence for a slow decrease in time of the configurational temperature,
as defined in eq. (??), although with the inherent structure complexity, is in agreement
with the idea of the systems representative point penetrating below the threshold in
the (free)-energy landscape.
The ratchet effect of an asymmetric intruder in an aging glass was studied numerically by Gradenigo et al. The energy flowing from slow to fast modes is rectified
to produce directed motion. The (sub) velocity of the intruder grows monotonically
with Teff /T and this current could be used to measure Teff .
9.5.2

Kinetically constrained models

Kinetically constrained models are toy models of the glassy phenomenon. Their
equilibrium measure is just the Boltzmann factor of independent variables and correlations only reflect the hard core constraint. Still, many dynamic properties of glass
forming liquids and glasses are captured by these models, due to the sluggishness introduced by the constrained dynamic rules. The literature on kinetically constrained
models is vast; a recent review with tests of Teff was written by Leonard et al. In
short, non-monotonic low-temperature response functions were initially taken as evidence against the existence of effective temperatures in these systems. The confusion
arosed from the incorrect construction of the (C) plot by using tw instead of t fixed
(see Sect. ??) that led to the incorrect treatment of the transient regime. Still, even
this taken into account, a large number of observables have negative fluctuationdissipation ratios; this might be related to the fact that these models do not have a
proper thermodynamics.
9.5.3

Experiments

Grigera and Israeloff were the first to measure FDT violations in glasses by comparing dielectric susceptibility and polarization noise in glycerol at T = 179.8K, i.e.
relatively close to Tg 196K. At fixed measuring frequency 8Hz, they found
an effective temperature that slowly diminishes from Teff 185K to roughly 180K
in 105 sec, that is to say in the order of days! This pioneering experiment in such a

203

traditional glass former has not had a sequel yet.

198




7*



186















VF)&

Teff
T = 179 K
T = 196 K

190

FF)&

Teff(K )

194

182

178
1000

10000

t (s)

100000

1e+06


















&&W
W


Figure 58: Left: the waiting-time evolution of the effective temperature of glycerol
by Grigera and Israeloff. Right: the parametric (C) plot for thiospinel, an insulator
spin-glass by Herisson and Ocio.
Particle tracking experiments in a colloidal suspension of PMMA particles revealed
an effective temperature of the order of double the ambient one from the mobilitydiffusivity relation.
In the soft matter realm a favorite is an aqueous suspension of clay, Laponite
RG, in its colloidal glass phase. During aging, because of electrostatic attraction and
repulsion, Laponite particles form a house-of-cards-like structure. After a number
of rather confusing reports the status of Teff in this system can be summarized as
follows. The surprisingly high Teff found with dielectric spectroscopy combined with
spontaneous polarization noise measurements was later ascribed to violent and intermittent events possibly linked to the presence of ions in the solution which may be
the actual source of FDT violation. For the moment dielectric degrees of freedom
are invalidated as a good test ground for Teff in this sample. Using other methods
several groups found that Teff detaches from the bath temperature. Strachan et al.
measured the diffusion of immersed probe particles of different sizes via dynamic light
scattering and simultaneous rheological experiments and found a slightly higher Teff
than T . With micro-rheology Abou and Gallet observed that Teff increases in time
from T to a maximum and then decreases back to T . Using a passive micro-rheology
technique and extracting Teff from the energy of the probe particle via equipartition
Greinert et al. also observed that Teff increases in time. In parallel, a series of global
mechanical tests, and passive and active micro-rheological measurements that monitor the displacement and mobility of probe Brownian particles were performed by
Cilibertos and D. Bonns groups, both finding no violation of FDT over a relatively
wide frequency range. In a very detailed article Jop et al. explain many subtleties
in the experimental techniques employed and, especially, the data analysis used to
extract Teff that could have biased the results quoted above. A plausible reason for
204

the lack of out of equilibrium signal in some experiments using Laponite as well as
other colloidal glasses is that the range of frequency-time explored may not enter
the aging regime. Moreover, none of these works studied the degrees of freedom of
the Laponite disks themselves but, instead, the properties of the solvent molecules
or probe particles. More recently, Maggi et al. combined dynamic light scattering
measurements of the correlation function of the colloid rotations with those of the refringence response and a (C, tw ) plot that is rather constant as a function of C and
slowly recovers the equilibrium form as the arrested phase is approached (tw ranges
from 90 to 1200 min and the violations are observed for time differences between 0.1
and 1 ms, i.e. frequencies between 10 and 1 kHz). Teff is at most a factor of 5 larger
than T . The actual behaviour of Laponite remains mysterious and not only in what
Teff is concerned!
Oukris and Israeloff measured local dielectric response and polarization noise in
polyvinyl-acetate with electric-force-microscopy. They probed long-lived nano-scale
fluctuations just below Tg , achieved a good signal-to-noise ratio down to very low
frequencies, constructed a parametric plot by keeping tw fixed and found a nontrivial asymptotic form with no tw dependence within the available accuracy. The
data combine into the parametric plot Teff (C) T C 0.57 in the aging regime.

9.6

Relaxation in frustrated magnetic systems

Disordered and frustrated magnets behave collectively at low temperatures and

developed ordered phases that although not fully understood are accepted to exist. As macroscopic glassy systems they present a separation of time-scales in their
low-temperature dynamics and are good candidates to admit a thermodynamic interpretation of the FDT violations.
9.6.1

Remarks on model systems

The physics of spin-glasses is a controversial subject. Some authors push an Ising

domain-growth interpretation of their dynamics slowed down by domain wall pinning
by disorder a.k.a. the droplet picture. If the scheme discussed in Sect. were
reproduced under strong disorder, the asymptotic (C) plot would have a linear
piece of slope 1/T and a sharp transition at qea to a flat aging piece. The domaingrowth interpretation is not accepted by other authors and more complex scenarii
based on the static and dynamic solution to the SK model are envisaged, with a nontrivial (C) as a result. Much effort has been put in trying to interpret numerical
and experimental data as validating one description at the expense of the other.
Unfortunately, it is very difficult to distinguish between the two. A third possibility
is that, in a loose sense, the spin-glass be like the low-T phase in the 2d XY model,
with quasi long-range order. Yet another proposal is that actual spin-glass samples are
of Heisenberg-type and that chilarity might be decoupled from spin with a chiral-glass
order arriving at a higher critical temperature than the spin-glass ordering.
205

The trap model was devised to describe slow dynamics in systems with weak
ergodicity breaking and it was applied, notably, to describe experiments in spinglasses. The model shows a glass transition at a Tg below which an equilibrium
Boltzmann state cannot exist. The (C) has a slope that varies continuously even
though there is a single scaling of relaxation times with age, it depends non-trivially
on the observable and one cannot use it to define a meaningful Teff . The reason for
this failure seems to be the unbounded nature of the energy and the fact that an
equilibrium distribution does not exist below Tg .
9.6.2

Simulations

Monte Carlo simulations of the 3d Edwards-Anderson (EA) model were carried

out by several groups. One of the hallmarks of the dynamics of the SK model,
dynamic ultrametricity, is absent from all numerical and experimental data analyzed
so far. Magnetic correlation and susceptibility relax in two scales, the by now usual
stationary one for finite time-differences and an aging one in which the data are well
described by a simple t/tw scaling. This aging scaling does not conform with the
droplet picture either, which predicts an asymptotic ln t/ ln tw form. In all studies
so far, the parametric plot was constructed by keeping tw fixed and the curves drift
towards increasing values of for longer tw s as in a transient or critical system. In
simple coarsening problems the drift with increasing tw goes in the opposite direction
of rendering the aging part of the curves flatter; this remark suggests to discard a
simple droplet picture. The outcome (C) found for the longest tw reached was
interpreted as being non-constant as in the SK model although this is, in our
opinion, not that clear from the data that could be described by a straight line. The
simultaneous t/tw scaling, the lack of unambiguous evidence for a stable plateau at
qea , and a curved (C) in the aging regime is not what would be expected from an
analogy with the SK model. Instead, it would be consistent with critical dynamics
and the 2d XY model similitude. A number of caveats on the numerical analysis
should, however, be lifted before reaching a firm conclusion.
The finite-time finite-length relation between static (C, (tw )) and long-time out
of equilibrium dynamic (C, tw ) was put to the test in the 2d and 3d EA models at
finite T . The notable coincidence of the two functions found in the 2d case, in which
there is no complex equilibrium structure, suggests that the claimed coincidence of
(C) and (C) in 3d might also be valid just in the transient regime.
Simulations of the 3d Heisenberg spin-glass model with weak anisotropy suggest
that Teff associated to the spin degrees of freedom is constant and about twice the
critical temperature for spin-glass ordering. As far as we know, chiral degrees of
freedom have not been used to estimate Teff .
As regards fluctuations, the two kinds were measured in the 3d EA spin-glass.
Disordered induced ones, in which one computes strictly local noise-averaged correlations and linear responses, demonstrate the existence of two types of spins in each
sample: rapid paramagnetic-like ones and slow ones. The former satisfy FDT while
the latter evolve in two time-regimes with a fast one satisfying FDT and a slow one in

206

which i (Ci ) looks quite flat as in coarsening systems. The simulation suggests that
the two ensembles behave independently of each other and are strongly correlated
with the backbone of the ground state configurations. The average over all sites (at
finite tw ) gives rise to a curve with non-constant slope. These results suggest a still
different picture for the spin-glass dynamics in which a rather compact set of spins
undergoes coarsening of the backbone equilibrium configurations while the other ones
behave paramagnetically. This intriguing idea needs to be put to further test.
The analysis of noise induced fluctuations suggests that eq. (??) is valid although
better numerical data would be needed to have definitive evidence for this statement.
A more detailed discussion can be found in the review by Chamon and Cugliandolo.
Very recent studies of non-linear fluctuations that take advantage of FDRs to compute
higher order responses point in the direction of the TRI scenario with a finite Teff .
9.6.3

Experiments

On the experimental side the first attempt to quantify FDT violations in spinglasses was indirect. Simultaneous measurements of global magnetic noise and susceptibility in the thiospinel insulating spin-glass were later performed by Herisson
and Ocio. The data confirm deviations from the FDT with a (C, tw ) plot of relatively curved form although still evolving during the experimental time window.
The authors interpreted it as evidence for the full RSB scenario, via the association
(C) (C). However, as with numerical data, dynamic ultrametricity fails to show
off, the asymptotic limit of the parametric construction is still far, and a clear-cut
distinction between a curved and a linear (C) is hard to assess.
More recent experiments exploit two novel techniques, Hall-sensor based magnetometer and giant magnetoresistance technology to detect signals from very small
samples. The use of these probes opens the way to perform a systematic study of
FDT violations in magnetic systems of different kind (spin-glasses, super-spin glasses,
disordered ferromagnets...). The first of these measurements appeared recently in a
super-spin glass, a system of magnetic nanoparticles suspended in fluid glycerol with
a single-domain magnetic structure that behaves as one large spin, the orientation
of which is the only degree of freedom. The large magnetic moment facilitates the
observation of magnetic noise. For aging times of the order of 1 h, the ratio of Teff
to the bath temperature T grows from 1 to 6.5 when T is lowered from Tg to 0.3 Tg ,
regardless of the noise frequency.
Artificial spin ice is yet another material in which the Teff notion has been tested.

9.7

Driven liquids and glasses

Berthier and Barrat used molecular dynamics of a binary Lennard-Jones mixture

under a steady and homogeneous shear flow. The deviation from FDT is similar to
the one found analytically in disordered spin models of RFOT type with asymmetric
couplings that mimic non-conservative forces. Moreover, it does not depend on the
observable. The tracer particle experiment was also realized. When the tracers

207

Einstein frequency is smaller than the inverse relaxation time of the fluid, a nonequilibrium equipartition theorem holds with mtr vz2 = Teff , where vz is the velocity
in the direction transverse to the flow. For increasing mtr the effective temperature
very slowly crosses over from T to the slow modes value, in perfect agreement with
the notion of a temperature measured by a thermometer sensible to the scale. Teff
also captures the essential phenomenological idea that when a system is sheared more
vigorously its effective temperature increases.
OHern et al. also studied fluctuation-dissipation relations in shear fluids. This
group defined an effective temperature through the static limit limt (ttw )/C(t, t),
a kind of average of the slope of the (C) plot over the full range of C(t tw ) that
mixes different time scales (in particular, the high and low frequency ones). A more
thorough discussion of the comparison between this definition and the one described
in this review was given by Ilg and Barrat within a fully solvable model that demonstrates the importance of not mixing time-scales to get physically sensible results.
A first study of the fluctuations of entropy production in a Lennard-Jones fluid
above and below Tg under a shear flow appeared in and the need to take into account
Teff , as obtained from the modification of the FDT below Tg , was signaled in this
paper. A more detailed analysis of the time-scale dependent effective temperature
would be needed to fully test the proposal in Bonnetto et al.
Another prominent example is the current driven motion of vortices in type II
superconductors. Disorder reduces dissipation, is responsible for non-equilibrium
transport and magnetic properties. The external force induces two dynamic phase
transitions separating plastic flow, smectic flow and a frozen transverse solid. A lowfrequencies Teff that decreases with increasing driving force and reaches the equilibrium melting temperature when the dynamic transverse freezing occurs was computed
from the transverse motion in the fluid moving phase.

9.8

Granular matter

Several studies of the effective temperature of granular matter have been pursued
theoretically, numerically and experimentally. In the latter front, DAnna et al. immersed a torsion oscillator in a granular system fluidized by strong high frequency
external vibrations to realize the thermometer experiment. They found Teff 2
with the adimensional measure of vibrational intensity, and quite independently of
. Wang et al. visualized the dynamics of tracer particles embedded in a 3d granular
ensemble slowly sheared by the rotating inner wall of a Couette cell. Teff , as obtained
from the comparison between the tracers diffusion and mobility perpendicular to the
applied rate of strain, is independent of the shear rate used and the tracers properties
but does depend on the packing density of the system. Tests of the thermodynamic
properties of Teff have not been carried through in this system yet. The dependence
on the direction of the applied stress was studied by Twardos and Dennin in a plastic
bead raft close to jamming. As expected, the correlations and linear responses in the
direction of flow do not decay slowly and (C) does not have the same properties as
in the transverse direction. Gei and Behringer stressed the fact that in a granular
208

assembly the outcome of a mobility measurement depends on whether one imposes

the velocity or the external force.
In the powders literature reference is often made to the granular temperature, a
measure of the temperature of the fast modes, as given by the kinetic energy of the
grains TK d2 EK d2 hv 2 i. Importantly enough, TK is a high frequency measure
that does not really access the structural properties of the sample and, in a sense,
plays the role of the environmental temperature in thermal systems. TK is generically
smaller than Teff , as in thermal systems where TK = T , the temperature of the bath.

9.9

Activated dynamics

Activated processes often occur in systems that are out of equilibrium, in the sense
that their response to an external drive is strongly non-linear or that their phase
space distribution is not the Gibbs-Boltzmann one. The question as to whether an
Arrhenius law governs the activation rate, possibly with an effective temperature,
and how the latter compares to the one defined from the deviations from FDT has
been addressed recently. Ilg and Barrat studied the effect of an out of equilibrium
flowing environment, a weakly sheared super-cooled liquid, on the activated dynamics
between the two stable conformations of dumbbell particles. The transition rate is
well described by an Arrhenius law with a temperature that crosses over from the one
of the equilibrium bath to a higher value close to the Teff of the slow modes of the
driven fluid. The crossover roughly occurs at the value of the rate that corresponds
to the inverse of the relaxation time of the fluid.
Three related studies are also worth mentioning. An effective temperature, also
consistent with the one stemming from fluctuation-dissipation measurements, appears
in a phenomenological Arrhenius law that describes transverse jumps between channels in the driven motion of vortex lattices with random pinning. Haxton and Liu
showed that in the shear dominated regime the stress of a 2d sheared fluid follows an
Arrhenius law with the effective temperature. A study of activation and Teff in a 2d
granular system close to jamming was performed by Abate and Durian.

9.10

Biological systems

In biologically inspired problems the relevance of Teff was stressed to reveal the
active process in hair bundles and model cells. Morozov et al. studied a model of the
cytoskeletal network made of semi-flexible polymers subject to thermal and motorinduced fluctuations and found a Teff that exceeds the environmental temperature T
only in the low-frequency domain where motor agitation prevails over thermal fluctuations. Simple gene network models were studied from the Teff perspective in Lu et al.
Fluctuation-dissipation ratios were used to quantify the degree of frustration, due to
the existence of many metastable disordered states, in the formation of viral capsids
and the crystallization of sticky discs, two self-assembly processes. Fluctuations and
responses of blood cell membranes for varying ATP concentration were measured very

209

recently. The measured Teff approaches the bath temperature at high frequencies and
increases at low frequencies reaching 4-10 times the ambient temperature.
Ratchets are simple models of molecular motors, out of equilibrium systems with
directed dissipative transport in the absence of any external bias. Harada and Sasa
proposed to use the violations of FDT in flashing ratchets as a means to measure
the energy input per unit time in molecular motors an otherwise difficult quantity
to access. Kolton showed that the rectified transverse velocity of a driven particle
in a geometric ratchet is equivalent to the response of a 1d flashing ratchet at a
drive-dependent Teff , as defined from the generalized Einstein relation.
Active matter is driven out of equilibrium by internal or external energy sources.
Its constituents absorb energy from their environment or from internal fuel tanks and
dissipate it by carrying out internal movements that lead to translational or rotational
motion. A typical example are self-propelled particle assemblies in bacterial colonies.
The role played by Teff in the stability of dynamic phases of motorized particle
systems was stressed by Shen and Wolynes. Multiple measurements of Teff were
carried out with molecular dynamic simulations of motorized spherical as well as
linear molecules in interaction. All measurements (from fluctuation-dissipation ratio
and using tracers) yield a constant low-frequency Teff > T when the effect of the
motors is not correlated with the structural rearrangements they induce. Instead,
Teff takes a slightly lower value than T when susceptible motors are used. Such an
inversion also occurs in relaxational systems in which the initial configuration is
chosen to be one of equilibrium at a lower T than the working one. In the case of
uncorrelated motors, Teff /T was found to follow the empirical law Teff /T 1 + f 2
with f the active force relative to the mean potential force and 15 a parameter.
Palacci et al.. investigated Teff by following Perrins analysis of the density profile in
the steady state of an active colloidal suspension under gravity. The active particles
used JANUS particles are chemically powered colloids and the suspension was
studied with optical microscopy. The measurements show that the active colloids
are hotter than in the passive limit with a Teff that increases as the square of the
parameter that controls activation, the Peclet number, a dependence that is highly
reminiscent of the f 2 -dependence of the simulations mentioned above.
Joly et al. used numerical techniques to study the non-equilibrium steady state
dynamics of a heated crystalline nanoparticle suspended in a fluid. This problem
models an active colloid that acts as a local heat source and generates a temperature
gradient around it. By comparing the mobility to the velocity correlation function,
they found that the FDT approximately holds at short-time lags with a temperature
value that coincides with the kinetic one. In contrast, at long-time lags data are
compatible with the temperature estimated by using the Einstein relation.
Certainly, many more studies of effective temperatures will appear in this very
active field of research, essentially out of equilibrium, in the near future.

210

10

Conclusions

We discussed the behavior of macroscopic classical systems out of equilibrium.

First, we summarized the dynamics of relatively simple problems, the ones that
undergo coarsening, that is to say, the progressive ordering into patches of the two (or
more) competing equilibrium states. We also recalled, briefly, the nucleation process
relevant to describe the dynamics of systems undergoing first order phase transitions.
Next, we entered the realm of the more interesting, and harder to understand,
glassy problems. We modeled these with models with quenched random interactions which, surprisingly enough, render their solution simpler. We mentioned that
the mean-field solution of these models allows for a classification into models that
mimic spin-glass samples and models that are relevant to describe structural glasses,
although there is no explicit quenched randomness in these.
In particular, we analyzed in detail a family of disordered models that yield a
mean-field description of the glass transition and dynamics of super-cooled liquids and
glasses. The relevance of these models to describe structural glasses was signaled and
explained by Kirkpatrick, Thirumalai and Wolynes in the 80s. Their non-equilibrium
dynamics and hence the connection with other systems far from equilibrium started
to develop more recently.
In short, their behaviour is the following. The dynamic transition arises when
the partition function starts being dominated by an exponentially large number of
metastable states yielding a finite complexity. The static transition instead is due
to an entropy crisis, i.e. it occurs when the complexity vanishes and the number of
states is no longer exponential in N , just as in the Adams-Gibbs-di Marzio scenario.
These transitions mimic, in a mean-field way, the crossover to the glassy phase at Tg
and the putative static transition at TK of fragile glasses.
The equilibrium dynamics close and above Td coincides with the one obtained with
the mode-coupling approach. It describes the relaxation of super-cooled liquids and it
contains its most distinctive feature of having a two step decorrelation. The first step
is ascribed to the motion of particles within the cages made by their neighbors while
the second one is the structural relaxation related to the destruction of the cages.
Below Td the equilibration time diverges with the size of the system and the models
do not equilibrate any longer with their environments (unless one considers times that
grow with the size of the system). This is very similar to the situation encountered
in real systems below Tg . The experimental time-window is restricted and one is
not able to equilibrate the samples any longer below Tg . Aging effects are captured.
The correlations still decay to zero but they do in a waiting-time dependent manner.
Their decay also occurs in two steps separated by a temperature-dependent plateau at
a value related to the size of the cages. One can interpret their stiffness as increasing
with the age of the system given that the beginning of the structural relaxation is
delayed and slowed down for longer waiting-times.

211

The nonequilibrium dynamics below Td approaches a threshold level of flat directions in phase space and it never goes below this level in finite times with respect to
the size of the system. The aging dynamics corresponds to the slow drift of the point
representing the system in the slightly tilted set of channels that form the threshold. The motion that is transverse to the channels is related to thermal fluctuations
and the first stationary step of the relaxation towards qea , that charcaterises then
the transverse size of the channel. The longitudinal motion along the channels
is related to the structural relaxation. The tilt is proportional to the magnitude of
the time-derivatives and these become less and less important as time passes. More
generally one interprets the long but finite time non-equilibrium dynamics following
saddles that are the borders between basins of attraction of more stable states in
phase space.
For times that scale with the size of the system, N , the sharp dynamic transition is
avoided, the system penetrates below the threshold via activation and it approaches
equilibrium in much longer time-scales. Metastable states below the threshold are
typically minima (the fact that they are local minima can be checked studying the
dynamics with initial conditions set to be in one of them). This structure allows
one to describe the cooling rate effects. For large but finite N and sufficiently slow
cooling rate, the system penetrates below the threshold via activation when this is
facilitated by T , i.e. when passing near Td . To which level it manages to arrive
(roughly speaking to which of the curves in the figure) depends on how long it stays
close to Td . The slower the cooling rate the lower level the system reaches with the
ideal equilibrium glass corresponding to an infinitely slow cooling.
The region of phase space reached asymptotically in the thermodynamic limit is
the threshold of flat directions. The replica analysis of the partition function gives
an alternative way of determining its statistical properties. Indeed, by evaluating the
partition function on a marginally stable saddle-point in replica space one selects the
threshold states. Dynamic information such as the value of qea is thus obtained with
a pseudo-static calculation. Other facts as, for instance, the scaling of the correlation
are not accessible in this way.
One of the hallmarks of the glassy non-equilibrium dynamics is the modification
of the relation between correlations and responses, namely, the fluctuation-dissipation
theorem. In mean-field models for structural glasses one finds that the integrated linear response is in linear relation with the associated correlation with a proportionality
constant that takes the equilibrium value 1/(kB T ) when the correlation is above the
plateau and it takes a different value 1/(kB T ) when it goes below the plateau. This
behavior has been found in a number of finite dimensional glassy models numerically.
The behavior just described corresponds to a family of mean-field disordered models to which the p spin models with p 3 and the Potts glass belong. Other two
families exist and they are related to ferromagnetic domain growth and spin-glasses.
Two representative models are the spherical p spin model with p = 2 and the sk
model, respectively. They are characterized by different scalings of the correlations in
the aging regime and by different forms of the modification of fdt. The classification
in families according to the non-equilibrium behavior has a static counterpart given
212

by the structure of replica symmetry breaking in the low-T phase.

The modification of fdt allows one to define an observable and correlation-scale
dependent effective temperature. Fast observables like the kinetic energy are equilibrated with the environment and the effective temperature equals the thermal bath
temperature for them. Other observables though show different values of the effective
temperature depending on the time-scales on which one investigates them. The effective temperature has a thermodynamical meaning even if defined out of equilibrium.
In particular, it can be directly read with a thermometer coupled to the desired observable and a zero-th law holds for interacting observables that evolve in the same
time-scale. As one should have expected the effective temperature shares some of
the qualitative features of the phenomenological fictive temperatures. For instance, a
system that is quenched from high temperatures has effective temperatures that take
higher values than the temperature of the bath, etc. At the mean-field level, when
N , it is history independent but one expects it to depend on the preparation of
the sample for finite size and finite dimensional systems. (This is in close relation to
the discussion above on cooling rate effects.) There is still no precise determination
of which are the necessary conditions a nonequilibrium system has to fulfill to ensure
the existence of well-behaved effective temperatures. A clear condition are the need
to reach a dynamic regime in which the dynamics is slow and heat exchanges are
weak.
Once the effective temperature has been identified one interprets the behavior in
the low T phase as follows: the system adjusts to a situation in which each observable
sees two baths, one is the white external one and the one characterizing the fast motion
of the particles, the other is coloured and at a different temperature T . The latter is
generated by the interactions. In more complex systems as mean-field spin-glasses
the asymptotic regime might be multi-thermalised with several time-scales each with
its own value of the effective temperature. These results, first derived explicitly for
p spin fully-connected models actually hold for any resummation of the perturbative
approach that keeps an infinite subset of diagrams (the mca being one such example).
The structure of time-scales and values of the effective temperature is related to the
breaking of supersymmetry down to a residual group.
The structure of the free-energy landscape can be computed exactly for mean-field
models in general, and for the spherical p spin model in particular. We expect its main
features to be reproduced at least in a smoothen way in real glassy systems. The
free-energy landscape at fixed and low T has a structure as the one roughly sketched
in Fig. 59. A pictorial image of the aging process can be quite helpful to understand
it. Imagine that one fills phase space with water whose level reaches a free-energy
density value, say, f . At high levels of the water, i.e. for high free-energy densities,
the landscape has only some few isolated stationary states. Looking at the landscape
from above one only sees some maxima that are represented as islands in the second
panel in the figure. Lowering the water level the islands grow in size and some of them
merge: land bridges develop. Lowering still the water level, it eventually reaches a
threshold, that corresponds to f = fth , where land percolates. One is left with a
labyrinthic path of water as drawn schematically in the third panel that represents
213

a top view of the landscape. This level is marginal since the bottom of the water
channels is almost completely flat. Draining water from the system the connectivity
of paths is reduced until the water level goes below the threshold, f < fth , where
minima dominate. In the fourth panel we represent them as lakes immersed in land.
Lowering the water level one sees the sizes of the lakes diminish and some of them dry.
These minima exist until the lowest level, f = feq . A gap in free-energy density
separates the threshold and the equilibrium levels.
This picture allows us to give a natural interpretation of the non-equilibrium
dynamics following a quench. Initially, the system is in a configuration typical of
high-T , thus, its initial free-energy density is very high. This corresponds to a
high level of water that fills the landscape. As time passes, water abandons the
landscape in such a way that the quantity of water progressively diminishes lowering
its level. The systems configuration can be associated to a ship and its evolution to
the displacements of the ship sailing on the water. Initially, the water level is very
high and the ship can move very rapidly far away from its initial position. It only
sees some very few isolated islands that it simply avoids along its motion and the
dynamics is very fast. As time passes the water level goes down. Roughly speaking
we can associate the speed of drainage with the magnitude of the rate of change of
the energy-density. When it approaches the threshold the available path becomes a
series of rivers forming a very intricate network. The ship can still follow this network
without remaining trapped in any confining region. Its motion, however, gets slower
and slower. In finite times with respect to N the water level does not go below the
threshold. But for longer times that scale with N it does. When such long times are
attained the ship remains trapped in lakes. For still longer times the higher lakes dry
and, if the ship got trapped in one of them it must be transported through the land to
reach other lakes at lower levels. This action represents an activated process. Part of
this image was introduced by Sibani and Hoffmann phenomenologically. The p spinmodels and the like realize it explicitly. All quantitative features of the landscape
here described with words have been, or in principle can be, calculated analytically.
The value taken by the effective temperature is in direct relationship with the
structure of the free-energy landscape. Indeed, again for p-spin model and the like, it
has been shown analytically that the asymptotic value T reached for long but finite
times with respect to N is given by = (, f )/f |fth , with the complexity. For
even longer times such that the system penetrates below the threshold one expects
the effective temperature to take different values related to the complexity at lower
free-energy density levels. The Edwards-Anderson parameter, qea , also changes since
qea (f ). In the longest time-scale such that equilibrium is reached and qea equals
the equilibrium value also obtained with a replica calculation using the standard
maximization prescription to determine the breaking point parameter x. This result
is intimately related to Edwards flat measure for granular matter and also to the more
recent use of a flat measure over inherent structures to describe the non-equilibrium
dynamics of glasses. Note that the these, being defined using the potential energydensity landscape, are valid only at zero temperature. However, extensive numerical
checks recently performed suggest that the approach, even if not obviously correct at
214

finite T , yields a very good approximation.

Within this picture two distinct regimes would appear in the low-T isothermal
dynamics of real systems: a mean-field-like one when the system approaches a pseudothreshold of flat directions in phase space and a slower activated regime in which the
system jumps over barriers to relax its excess energy density and very slowly progress
towards equilibrium. How and if the aging properties in the first and second regime
resemble is a very interesting open problem.
The existence of a threshold plays a fundamental role in explaining several features of many experimental observations in such diverse systems as driven granular
matter, the rheological properties of complex liquids and glasses, etc. Just to cite
two examples, trapping and Reynolds dilatancy effects in granular matter as well as
the existence of a static yield stress and thixotropic behaviour in some rheological
experiments can be interpreted in terms of threshold and sub-threshold states. These
features support the claim that this free-energy structure exists in real physical systems. Moreover, maybe not surprisingly, this structure also appears in optimization
problems such as xor-sat and k-sat that can be mapped to dilute p-spin models
at zero temperature. In this context the control parameter is the number of requirements over the number variables, , and the static transition, s , is related to the
sat-unsat transition while the dynamic transition, d < s corresponds to the value
where greedy algorithms fail to find the existing solutions.
All these arguments can be adapted to include quantum fluctuations The statics
is studied with the Matsubara replicated partition function, metastability with an extension of the tap approach and the real-time dynamics with the Schwinger-Keldysh
formalism. The picture that arises is very similar to the one above with some intriguing new ingredients as the emergence of truly first order transitions close to the
quantum critical point, highly non-trivial effects due to the quantum environments,
a waiting-time dependent quantum-to-classical crossover in the dynamic scaling, etc.
The models we studied in these notes have quenched random interactions. Real
glassy systems of the structural type do not. One may wonder if this is an important
deficiency of the approach or if similar results can be obtained for models with no
disorder. A large variety of models of mean-field type, or defined on large d spaces,
with no explicit quenched disorder and having the same phenomenology have been
introduced in recent years. Finite d models with similar, eventually interrupted,
dynamic behavior have also been exhibited. Their existence supports the belief that
the scenario here summarized goes beyond simple modelling. Indeed, it is at the basis
of several conjectures for the behavior of other non-equilibrium systems with slow
dynamics that have been later checked numerically. It has also motivated several
experimental investigations in a variety of systems.

215

Above
At threshold
threshold
Below
threshold

Lake

Island

Sea

Land

Figure 59: Left: a 1d simplified sketch of the free-energy density. Three top views of
the free-energy landspcape: above, at and below the threshold.

Conventions

A.1

Fourier transform

The convention for the Fourier transform is

Z
d i
e
f () ,
f ( ) =
2
Z
d e+i f ( ) .
f () =

(A.1)
(A.2)

The Fourier transform of the theta function reads

() = ivp

1
+ () .

(A.3)

The convolution is
[f g]() = f g()

A.2

d
f ( )g( ) .
2

(A.4)

Commutation relations

We defined the commutator and anticommutator: {A, B} = (AB + BA)/2 and

[A, B] = (AB BA)/2.

A.3

Time ordering

We define the time odering operator acting on bosons as

Tx
(t)
x(t )

(t, t )
x(t)
x(t ) + (t , t)
x(t )
x(t) .
216

(A.5)

For fermions, we define the time ordering operator as

Tx
(t)
x(t )

Tx
(t)
x (t )

(t, t )
x(t)
x(t ) (t , t)
x(t )
x(t) ,

(t, t )
x(t)
x (t ) (t , t)
x (t )
x(t) ,

(A.6)
(A.7)

In both cases (t, t ) is the Heaviside-function.

We define the time-ordering operator TC on the Keldysh contour in such a way
that times are ordered along it:
TC x+ (t)x (t ) = x (t )x+ (t)
TC x+ (t)x (t ) = x (t )x+ (t)

TC x (t)x+ (t ) = x (t)x+ (t )
TC x (t)x+ (t ) = x (t)x+ (t )

(A.8)

for all t and t .

Classical statics: the reduced partition function

In order to analyze the statistical static properties of the classical coupled system,
we study the partition function or Gibbs functional, Ztot that reads
X
Ztot [] =
exp(Htot x)
(B.1)
conf osc
conf syst

where the sum represents an integration over the phase space of the full system, the
particles and the oscillators, and is a source. Having chosen a quadratic bath and
a linear coupling, the integration over the oscillators coordinates and momenta can
be easily performed. This yields the reduced Gibbs functional
!#
"
Nb
X
c2a
1X
x2
. (B.2)
Zred []
exp Hsyst + Hcounter + x
2 a=1 ma a2
conf syst

The counterterm Hcounter is chosen to cancel the last term in the exponential and
it avoids the renormalization of the particles mass (the coefficient of the quadratic
term in the potential) due to the coupling to the environment that could have even
destabilize the potential taking negative values. An alternative way of curing this
problem would be to take a vanishingly small coupling to the bath in such a way
that the last term must vanish by itself (say, all ca 0). However, this might be
problematic when dealing with the stochastic dynamics since a very weak coupling
to the bath implies also a very slow relaxation. It is then conventional to include the
counterterm to cancel the mass renormalization. One then finds
X
Zred []
exp [ (Hsyst + x)] = Zsyst [] .
(B.3)
conf syst

The interaction with the reservoir does not modify the statistical properties of the
particle since Zred Zsyst . This doesP
not necessarily happen quantum mechanically. (For a non-linear coupling Hint = c q V(x) the counterterm is Hcounter =
P
c2
1
2
m 2 [V(x)] .)
2

217

The instanton calculation

The path-integral formalism yields an alternative calculation of the Kramers escape time, the Arrhenius exponential law and its prefactor that, in principle, is easier
to generalize to multidimensional cases. For the sake of simplicity let us focus on
the overdamped limit in which we neglect inertia. We first rederive the Arrhenius
exponential using a simplified saddle-point argument, and then show how Kramers
calculation can be recovered by correctly computing the fluctuations around this saddle point. Starting from the following representation of the probability to reach the
top of the barrier from the potential well:
*Z


+
x(t)=xmax


eq[x](t)
 ,
Dx ( eq[x]) det
P (xmax , t|xmin ) =

x(t ) 
x(0)=xmin

and neglecting the determinant (which is justified if one follows the It

o convention),
then, for a Gaussian white noise :
P (xmax , t|xmin ) =

x(t)=xmax
x(0)=xmin

Dx e

4k1

BT

Rt
0

dt (x+
dV
dx )

Expanding the square, we find a total derivative contribution

to the integral equal to
Rt
2[V (xmax ) V (xmin )], plus the sum of two squares: 0 dt [x 2 + (V (x))2 ]. For small T ,
the path, x , contributing most to the transition probability is such that this integral
is minimized. Using standard rules of functional derivation one finds
d2 x
= V (x )V (x )
dt2

x = V (x ).

In order to be compatible with the boundary conditions x (0) = xmin and x(t) = xmax ,
the + solution must be chosen, corresponding to an overdamped motion in the inverted
potential V (x). The action of this trajectory is
Z

t
0



dt x 2 + (V (x ))2 = 2

t
0

dt x V (x ) = 2[V (xmax ) V (xmin )],

that doubles the contribution of the total derivative above. Hence,

P (xmax , t|xmin ) e(V (xmax )V (xmin )) ,
independently of t, as in eq. (2.84). This type of calculation can be readily extended
to cases in which the noise has temporal correlations, or non Gaussian tails, and
to see how these effects change the Arrhenius result. The calculation of the attempt
frequency is done using the standard dilute gas instanton approximation developed
by several authors but we will not discuss it here.
218

The path-integral that we have just computed is a sum over the subset of noise
trajectories that lead from the initial condition to a particular final condition that we
imposed. Imposing a boundary condition in the future destroys the causal character
of the theory.
In a one dimensional problem as the one treated in this Section there is only one
possible reaction path. In a multidimensional problem, instead, a system can transit
from one state to another following different paths that go through different saddlepoints. The lowest saddle-point might not be the most convenient way to go and
which is the most favorable path is, in general, difficult to established.

Discrete MSRJD for additive noise: Stratonovich

prescription Mid-point discretization

The Langevin equation is a stochastic differential equation and one can give a
rigorous meaning to it by specifying a particular discretization scheme. We adopt the
Stratonovitch prescription where the rules of conventional differential calculus can be
used. This correponds to a mid-point discretization scheme and is coherent with the
convention (0) = 1/2 in the continuum limit.
Let us divide the time interval [T, T ] into N + 1 infinitesimal slices of width
2T /(N + 1). The discretized times are tk = T + k with k = 0, ..., N + 1. The
discretized version of x(t) is xk x(tk ). The continuum limit is achieved by sending
N to infinity and keeping (N + 1) = 2T constant. Given some initial conditions xi
and x i , we set x1 = xi and x0 = xi x i meaning that the first two times (t0 and t1 ) are
reserved for the integration over the initial conditions whereas the N following ones
correspond to the stochastic dynamics given by the discretized Langevin equation:
Eqk

xk+2 2xk+1 + xk
Fk+2 (xk+2 , xk+1 , ...)
2
k
X
kl (xl+2 xl+1 ) = k+1 ,
+

(D.1)

l=1

R
defined for k = 0, ..., N 1. The notation kl stands for kl 1 0 du(tk tl +
u) The k (k = 1, ..., N ) are independent Gaussian random variables with variance
hk l i = 1 kl where kl kl + lk . Inspecting the equation above, we notice that
the value of xk depends on the realization of the previous noise realisation k1 so that
there is no need to specify 0 and N . In the Markovian limit, one has kl = 1 0 kl
, hk l i = 20 1 1 kl where is the Kronecker delta, and
Eqk

xk+2 2xk+1 + xk
Fk+2 (xk+2 , xk+1 , ...)
2
xk+2 xk+1
= k+1 .
+0

219

(D.2)

D.1

Construction of the MSRJD action

The probability density P for a complete field history (x0 , x1 , ..., xN +1 ) is set by
the relation
P (x0 , x1 , ..., xN +1 )dx0 dx1 ...dxN +1
= Pi (xi , x i )dxi dx i Pn (1 , 2 , ..., N )d1 d2 ...dN .

(D.3)

Pi is the initial probability distribution of the field. The probability for a given noise
history to occur between times t1 and tN is given by
PN
k 1
l
12
kl
k,l=1
(D.4)
Pn (1 , ..., N ) = M1
e
N
with the normalization MN

(2)N
detkl 1
kl

P (x0 , x1 , ..., xN +1 ) = |JN |Pi (x1 ,

1/2

. From eq. (D.3), one gets

x1 x0
)Pn (Eq0 , ..., EqN 1 ) ,

(D.5)

with the Jacobian

JN det

(xi , x i , Eq0 , . . . , EqN 1 )

(xi , x i , 1 , . . . , N )
= det
,
(x0 , x1 , . . . , xN +1 )
(x0 , x1 , . . . , xN +1 )

(D.6)

that will be discusssed in . The expression (D.4) for the noise history probability
reads, after a Hubbard-Stratonovitch transformation that introduces the auxiliary
variables x
k (k = 1, ..., N ),
Z
P
P
1
1
Pn (1 , ..., N ) = NN1
d
x1 ...d
xN e k ixk k + 2 kl ixk kl ixl
Z
P
P
1
1
x0 ...d
xN +1 (
x0 )(
xN +1 ) e k ixk Eqk1 + 2 kl ixk kl ixl ,
= NN1 d
with NN (2)N . In the last step, we replaced k by Eqk1 and we allowed integrations over x
0 and x
N +1 at the cost of introducing delta functions. Notice that the
Hubbard-Stratonovitch transformation allows for some freedom in the choice of the
sign in front of i
xk in the exponent. Together with eq. (D.5) this gives
Z
P (x0 , x1 , ..., xN +1 ) = NN1 |JN | d
x0 ...d
xN +1
(D.7)
P
P
x1 x0
1
1
e k ixk Eqk1 + 2 kl ixk kl ixl +ln Pi (x1 , )
that in the continuum limit becomes
Z
R
P [x] = N 1 eln Pi +ln |J [x]| D[
x]e

du i
x(u)Eq([x],u)+ 21

220

duvi
x(u) 1 (uv)i
x(v)

with the boundary conditions x

(T ) = x
(T ) = 0 and where all the integrals over time
run fromRT to T . In the following, unless otherwise stated, we will simply denote
them by . The infinite prefactor N lim (2)N can be absorbed in the definition
N

of the measure:

N
+1
Y
1
D[x, x
] = lim
dxk d
xk .
N (2)N

(D.8)

k=0

D.2

Evaluation of the Jacobian

In this section we take the continuum limit of the Jacobian defined in eq. (D.6).
In the additive noise case, we start from
JN

(xi , x i , Eq0 , . . . , EqN 1 )

(x0 , x1 , . . . , xN +1 )

0
1
0...
1/

1/ 0 . . .
Eq

Eq0
Eq0
0

0...
x0

x1
x2
det Eq1

Eq1
Eq1
Eq1
0...
x0

x1
x2
x3

...
0

EqN 1
EqN 1
.
.
.
x0
xN +1
Eq

0
0...
x
1
Eq21
1 NY
Eq1
0...
1
Eqk

x3
det x2
=
.
.
.
0

xk+2

k=0
EqN 1
EqN 1
.
.
.
x2
xN +1
det

(D.9)

We can safely drop the overall 1/ factor since it can be included in the normalization.
Notice that causality manifests itself in the lower triangular structure of the last
matrix involved. In the continuous notation, lim JN reads
N

J [x] = detuv

D.3

Eq([x], u)
x(v)

(D.10)

Markovian case

Let us first consider the Markovian case in which the friction term has no memory
and the force F is a local functional of x which can carry a time-dependence. Defining
F as Fu [xu ]/xv Fu [xu ](u v), the Jacobian reads




J [x] = detuv mu2 + 0 u Fu [xu ] uv .
(D.11)




Now let us write detuv mu2 + 0 u Fu [xu ] uv


Z


Guw Fw [xw ]wv
= detuv (mu2 + 0 u )uv detuv uv
w

221

=
=



detuv (mu2 + 0 u )uv exp Truv ln [uv Guv Fv [xv ]]

X


1
M M ...M
detuv (mu2 + 0 u )uv exp
|
{z
}
n
u
n=1
n times

, (D.12)
uu

where we used the matrix notation Muv Guv Fv [xv ] and product . G is the
retarded Green function solution of
 2

mu + 0 u G(u v) = (u v) ,
(D.13)

which reads

G(t) =

i
1 h
1 e0 t/m (t) .
0

(D.14)

Since (u v)(v u) = 0, u 6= v, the n 2 terms do not contribute to the sum

in eq. (D.12). Furthermore, G(t = 0) = 0 for finite values of m15 , implying that the
n = 1 term is zero as well. Therefore we established


J [x] = detuv mu2 + 0 u (u v) .
(D.15)
This means that the functional determinant is simply a field independent constant.
One can easily generalize this result for time dependent and non potential forces.

D.4

Non Markovian case

Within the Stratonovich prescription, the determinant can be seen as the result
of a Gaussian integration over Grassmannian conjugate fields c and c . Let us first
recall the discretized expression of the Jacobian obtained in eq. (D.9):


1
Eqk
JN = detkl
,
(D.16)

xl+2
where k and l run from 0 to N 1. Introducing ghosts, it can be put in the form
Z
PN 1 PN +1 Eqk
1
c
c
(D.17)
JN =
dc2 dc0 ...dcN +1 dcN 1 e k=0 l=2 k xl l

Z
PN +1 PN +1 Eqk
1
c
c
=
dc0 dc0 ...dcN +1 dcN +1 e k=0 l=0 k xl l c0 c1 cN cN +1 ,

where in the last step, we allowed integration over c0 , c1 , cN and cN +1 at the cost
of introducing delta functions (remember that for a Grassmann number c, the delta
15 If we send m 0 at the end of the calculation, we still get G(0) = 0 and a constant Jacobian.
However, if m is set to 0 from the begining then G(t) = (t)/0 and G(0) = 1/(2
R 0 ) in our
F [xu ]. It
conventions. This leads to the so-called Jacobian extra-term in the action: 1/(20 )
u u
is invariant under time-reversal of the field xu 7 xu as long as F is itself time-reversal invariant.

222

function is achieved by c itself). In the continuum limit, dropping the overall 1/

constant (and infinite) factor, this yields
Z

J [x] = D[c, c ]eK[c,c ,x]

(D.18)
with

K[c, c , x]

dduvc (u)

Eq([x], u)
c(v) ,
x(v)

(D.19)

and with the extra boundary conditions: c(T ) = c(T

) = c (T ) = c (T ) = 0.
Plugging the Langevin equation (2.9), we have
Z
Fu [x]
Equ [x]
2
wv w wv .
= mu uv
+
xv
xv
w
The kinetic term in K[c, c , x] can be re-written
Z
Z Z
1
1
T
T
cu u2 cu + [c c c c]
cu u2 uv cv =
T + 0 [c c]T .
2
2
u
u v
The last two terms in the rhs vanishes by use of the boundary conditions (cT =
cT = cT = cT = 0). The retarded friction can be re-written
Z Z
Z
1
cu u uv cv
c [u+T cT uT cT ] ,
2 u u
u v
where the second line vanishes identically for two reasons: the boundary condition
(cT = 0) kills the first part and the causality of the friction kernel (u = 0 u < 0)
suppresses the second one. Notice that if there is a Dirac contribution to centered
at u = 0 like in the Markovian case, the other boundary condition (cT = 0) finishes
to cancel the second part. Finally we have


Z Z
Z
Fu [x]
cv .
(D.20)
cu u uv
cu u2 cu +
K[c, c x] =
xv
u v
u

D.5

Discrete MSRJD for multiplicative noise: Stratonovich

prescription Mid-point discretization

The discretized Langevin equation reads:

Eqk

xk+2 2xk+1 + xk
Fk+2 (xk+2 , xk+1 , ...)
2
k
X
kl M (xl )(xl+2 xl+1 ) = M (
xk )k+1 , (D.21)
+M (xk )
l=1

223

with the mid-point x

k (xk+1 + xk )/2. The Jacobian is:


1
Eqk
M (
xk )
k+1 l+2 + k l+2
JN = detkl
,

Eqk

xl+2
M (
xk )
2

(D.22)

where k and l run from 0 to N 1. Introducing ghosts, it can be put in the form
Z
JN =
dc0 dc0 ...dcN +1 dcN +1 c0 c1 cN cN +1 eKN ,
with
KN

+1
N
+1 N
X
X

ck

k=0 l=0

N
+1
X
Eqk
M (
xk )
ck+1 + ck
ck
cl
Eqk
.

xl
M (
xk )
2

(D.23)

k=0

In the continuum limit,

K

lim KN =

Z Z
u

cu

Equ [x]
cv
xv

cu

M (xu )
Equ [x] cu , (D.24)
M (xu )

with the boundary conditions c(T ) = c(T

) = 0 and c (T ) = c (T ) = 0.

Mean-field theory for ferromagnets

In spite of their apparent simplicity, the statics of ferromagnetic Ising models has
been solved analytically only in one and two dimensions. The mean-field approximation allows one to solve the Ising model in any spatial dimensionality. Even if the
qualitative results obtained are correct, the quantitative comparison to experimental and numerical data shows that the approximation fails below an upper critical
dimension du in the sense that it does not capture correctly the behavior of the systems close to the critical point. It is however very instructive to see the mean-field
approximation at work.
Naive mean-field approximation
Using the factorization of the joint probability density that defines the mean-field
approximation, one finds
X
X
F ({mi }) =
Ji1 ...ip mi1 . . . mip
h i mi
i

i1 6=...6=ip

+T

N 
X
i=1

1 + mi 1 + mi
1 mi 1 mi
ln
+
ln
2
2
2
2

(E.1)

Note that a Taylor expansion of the entropic contribution around mi = 0 leads to a

polynomial expression that is the starting point in the Landau theory of second order
phase transitions (see Sect. ??).
224

The local magnetizations, mi , are then determined by requiring that they minimize the free-energy density, f ({mj })/mi = 0 and a positive definite Hessian,
2 f ({mj })/mi mj (i.e. with all eigenvalues being positive at the extremal value).
This yields

X
(E.2)
mi = tanh p
Jii2 ...ip mi2 . . . mip + hi
i2 6=...6=ip

p1

) for all p uplets and the applied field is uniform, hi = h,

If Ji1 ...ip = J/(p!N
one can take mi = m for all i and these expressions become (E.4) and (E.7) below,
respectively. The mean-field approximation is exact for the fully-connected pure Ising
ferromagnet, as we will show below. [Note that the fully-connected limit of the model
with pair interactions (p = 2) is correctly attained by taking J J/N and 2d N
leading to Tc = J.]
Exact solution
Let us solve the ferromagnetic model exactly. The sum over spin configurations
PN
in the partition function can be traded for a sum over the variable, x = N 1 i=1 si ,
that takes values x = 1, 1 + 2/N, 1 + 4/N, . . . , 1 4/N, 1 2/N, 1. Neglecting
subdominant terms in N , one then writes
X
Z=
eN f (x)
(E.3)
x

with the x-parameter dependent free-energy density



1+x 1+x 1x 1x
J p
.
ln
+
ln
f (x) = x hx + T
p!
2
2
2
2

(E.4)

The first two terms are the energetic contribution while the third one is of entropic
origin since N !/(N (1 + x)/2)!(N (1 x)/2)! spin configurations have the same magnetization density. The average of the parameter x is simply the averaged magnetization
density:
N
1 X
hxi =
h si i = m
(E.5)
N i=1

In the large N limit, the partition function and all averages of x can be
evaluated in the saddle-point approximation
X

Z
eN f (xsp )
,
(E.6)

where x
sp are the absolute minima of f (x) given by the solutions to f (x)/x|xsp = 0,


J
xsp = tanh
xp1
+
h
,
(E.7)
(p 1)! sp
225

-0.8

0.25

-0.25

p=3
f (m)

0
-1.2

-0.3
-0.25

p=2
-1.6
-1.5

-0.5

0.5

1.5

-0.5
-1.5

-0.5

0.5

1.5

-0.35
-1.5

-0.5

0.5

1.5

Figure 60: The free-energy density f (m) of the p = 2 (left), p = 3 (center) and p = 4
(right) models at three values of the temperature T < Tc (light dashed line), T = Tc
(dark dashed line) and T > Tc (solid line) and with no applied field. (The curves
have been translated vertically.)
together with the conditions d2 f (x)/dx2 |xsp > 0. Note that the contributing saddlepoints should be degenerate, i.e. have the same f (x
sp ) for all , otherwise their
contribution is exponentially suppressed. The sum over then just provides a numerical factor of two in the case h = 0. Now, since
xsp = f (x)/h|xsp = h x i = m ,

(E.8)

as we will show in Eq. (E.9), the solutions to the saddle-point equations determine the
order parameter. We will next describe the phases and phase transition qualitatively
and we will later justify this description analytically.
Model in a finite field
In a finite magnetic field, eq. (E.7) has a unique positive negative solution
for positive negative h at all temperatures. The model is ferromagnetic at all
temperatures and there is no phase transition in this parameter.
2nd order transition for p = 2
In the absence of a magnetic field this model has a paramagnetic-ferromagnetic
phase transition at a finite Tc . The order of the phase transition depends on the
value of p. This can be seen from the temperature dependence of the free-energy
density (E.4). Figure 60 displays f (x) in the absence of a magnetic field at three
values of T for the p = 2 (left), p = 3 (center) and p = 4 (right) models (we call the
independent variable m since the stationary points of f (x) are located at the magnetization density of the equilibrium and metastable states, as we will show below).
At high temperature the unique minimum is m = 0 in all cases. For p = 2, when
one reaches Tc , the m = 0 minimum splits in two that slowly separate and move
towards higher values of |m| when T decreases until reaching |m| = 1 at T = 0 (see
Fig. 60-left). The transition occurs at Tc = J as can be easily seen from a graphical
226

1.5

1.2

1.2

eq(m)

p=2

p=3

p=4

0.5

0.8

0.8

-0.5

0.4

0.4

-1.5
-1.5

0
-0.5

0.5

1.5

0
0

0.5

1.5

0.5

1.5

Figure 61: Graphical solution to the equation fixing the order parameter x for p = 2
(left), p = 3 (center) and p = 4 (right) ferromagnetic models at three values of the
temperature T < T , T = T and T > T and with no applied field. Note that the
rhs of this equation is antisymmetric with respect to m m for odd values of p
while it is symmetric under the same transformation for even values of p. We show
the positive quadrant only to enlarge the figure. T is the temperature at which a
second minimum appears in the cases p = 3 and p = 4.
solution to eq. (E.7), see Fig. 61-left. Close but below Tc , the magnetization increases
1
as m (Tc T ) 2 . The linear magnetic susceptibility has the usual Curie behavior
at very high temperature, , and it diverges as |T Tc |1 on both sides of
the critical point. The order parameter is continuous at Tc and the transition is of
second-order thermodynamically.
1st order transition for p > 2
For p > 2 the situation changes. For even values of p, at T two minima (and
two maxima) at |m| =
6 0 appear. These coexist as metastable states with the stable
minimum at m = 0 until a temperature Tc at which the three free-energy densities
coincide, see Fig. 60-right. Below Tc the m = 0 minimum continues to exist but
the |m| 6= 0 ones are favored since they have a lower free-energy density. For odd
values of p the free-energy density is not symmetric with respect to m = 0. A single
minimum at m > 0 appears at T and at Tc it reaches the free-energy density of the
paramagnetic one, f (m ) = f (0), see Fig. 60-center. Below Tc the equilibrium state is
the ferromagnetic minimum. For all p > 2 the order parameter is discontinuous at Tc ,
it jumps from zero at Tc+ to a finite value at Tc . The linear magnetic susceptibility
also jumps at Tc . While it equals on the paramagnetic side, it takes a finite value
given by eqn. (E.10) evaluated at m on the ferromagnetic one. In consequence, the
transition is of first-order.
Pinning field, broken ergodicity and spontaneous broken symmetry
The saddle-point equation (E.7) for p = 2 admits two equivalent solutions in
no field. What do they correspond to? They are the magnetization density of the
227

equilibrium ferromagnetic
and negative value. At T < Tc if one
PN states with
P positive
N f (x)
e
x summing over the two minima of
computes m = N 1 i=1 h si i =
x
the free-energy density one finds m = 0 as expected by symmetry. Instead, if one
computes the averaged magnetization density with the partition sum restricted to the
configurations with positive (or negative) x one finds m = |msp | (or m = |msp |).
In practice, the restricted sum is performed by applying a small magnetic field,
computing the statistical properties in the N limit, and then setting the field
to zero. In other words,



N
f (xsp ) 
1 X
1 ln Z 
=
= |xsp | .
m
h s i i =

N i=1
N h
h h0
h0

(E.9)

By taking the N limit in a field one selects the positive (or negatively) magnetized states.
For all odd values of p the phase transition is not associated to symmetry breaking, since there is only one non-degenerate minimum of the free-energy density that
corresponds to the equilibrium state at low temperature. The application of a pinning
field is then superfluous.
For any even value of p and at all temperatures the free-energy density in the
absence of the field is symmetric with respect to m m , see the left and right panels
in Fig. 60. The phase transition corresponds to a spontaneous symmetry breaking
between the states of positive and negative magnetization. One can determine the
one that is chosen when going through Tc either by applying a small pinning field
that is taken to zero only after the thermodynamic limit, or by imposing adequate
boundary conditions. Once a system sets into one of the equilibrium states this is
completely stable in the N limit. In pure static terms this means that one can
separate the sum over all spin configurations into independent sums over different
sectors of phase space that correspond to each equilibrium state. In dynamic terms
it means that temporal and statistical averages (taken over all configurations) in an
infinite system do not coincide.
The magnetic linear susceptibility for generic p is given by


m 
xsp 

.
(E.10)
=
=
p1


2
J
J
h h0
h h0
cosh ( (p1)! xsp ) (p2)!
xp2
sp

For p = 2, at T > Tc , xsp = 0 the susceptibility is given by (T J)1 predicting the

second order phase transition with a divergent susceptibility at Tc = J. Approaching
Tc from below the two magnetized states have the same divergent susceptibility,
(Tc T )1 .
For p > 2, at T > Tc , xsp = 0 and the susceptibility takes the Curie form = .
The Curie law, = , jumps to a different value at the critical temperature due to
the fact that xsp jumps.

228

Grassmann variables and supersymmetry

2

+ = 0. The integration rules are

Grassmann
variables
2 = = [, ]
R
R anticommute
R
R
d = d
R = 0 and d = d = 1 while the derivation is such that = d
and = d.
In the supersymmetric formalism used in Section one enlarges the usual bosonic
space to include two conjugate Grassmann variables and : t a = (t, , ). A
superfield and its supercorrelator are then defined as
R

q (t) ,
(a) q(t) + (t) + (t) + i

Q(a, b) h(a)(b)i ,

(F.1)

x(t )i,
b = (t , , ). The latter encodes the usual correlations hx(t)x(t )i, hx(t)i
hi
x(t)x(t )i, hi
x(t)i
x(t )i, as well as fermionic correlators hx(t)(t )i, h(t)i
x(t )i,
h(t)(t )i, etc. The solutions we construct and study are such that all correlators
that involve only one fermionic variable and vanish. We are then left with
the usual four correlators purely bosonic correlators and the fermion bilinears. One
proves that the latter equal the linear response. If, moreover, we only consider causal
t ) hi
solutions, Q(t,
x(t)i
x(t )i = 0 and

Q(a, b) = C(t, t ) ( ) ( R(t, t ) R(t , t)) .

(F.2)

Convolutions, or operational products, and Hadamard, or simple products, are defined

as
Z
Q1 (a, b) Q2 (b, c) =
db Q1 (a, b)Q2 (b, c) ,
Q1 (a, b) Q2 (a, b)

Q1 (a, b)Q2 (a, b) ,

(F.3)

respectively, with db dtdd.

For correlators of the causal form (F.2), the convolution and the Hadamard product respect the structure of the correlator. Indeed, the result of the convolution is
again of the form (F.2) with
Z
Cconv (t, t ) =
dt [C1 (t, t )R2 (t , t ) + R1 (t, t )C2 (t , t )] ,
Z

Rconv (t, t ) =
dt R1 (t, t )R2 (t , t ) ,
(F.4)
and the result of the Hadamard product is also of the form (F.2) with
Chad (t, t )
Rhad (t, t )

=
=

C1 (t, t )C2 (t, t ) ,

C1 (t, t )R2 (t, t ) + C2 (t, t )R1 (t, t ) .

The Dirac delta function is defined as (a b) = (t t )( )( ).

229

(F.5)

References
[1] A. Cavagna, Supercooled liquids for pedestrians, Phys. Rep. 476, 51 (2009). L.
Berthier & G. Biroli, A theoretical perspective on the glass transition and nonequilibrium phenomena in disordered materials, arXiv:1011.2578. Rev. Mod. Phys.
[2] L. F. Cugliandolo, Dynamics of glassy systems, in Les Houuches 2002 (Springer,
2003).
[3] D. W. Oxtoby, Homogeneous nucleation: theory and experiment, J. Phys.: Condens. Matter 4, 7626-7650 (1992). K. Binder, Theory of first-order phase transitions, Rep. Prog. Phys. 50, 783-859 (1987).
[4] A. J. Bray, Thoery of phase ordering kinetics, Adv. Phys. 43, 357 (1994). P.
Sollich, http://www.mth.kcl.ac.uk/ psollich/
[5] P. C. Hohenberg and B. I. Halperin, Rev. Mod. Phys. 49, 435 (1977). H. K.
Janssen, B. Schaub, and B. Schmittman, Z. Phys. B Cond.Mat. 73, 539 (1989).
P. Calabrese and A. Gambassi, J. Phys. A 38, R133 (2005).
[6] P. G. Debenedetti, Metastable liquids (Princeton Univ. Press, 1997). E. J. Donth,
The glass transition: relaxation dynamics in liquids and disordered materials
(Springer, 2001). K. Binder and W. Kob, Glassy Materials and Disordered Solids:
An Introduction to their Statistical Mechanics (World Scientific, 2005).
[7] K. H. Fischer and J. A. Hertz, Spin glasses (Cambridge Univ. Press, 1991). M.
Mezard, G. Parisi, and M. A. Virasoro, Spin glass theory and beyond (World
Scientific, 1986). N. Kawashima and H. Rieger, Recent progress in spin glasses
in Frustrated spin systems, H. T. Diep ed. (World Scientific, 2004). T. Castellani
and A. Cavagna, Spin-glass theory for pedestrians, J. Stat. Mech. (2005) P05012.
F. Zamponi, Mean field theory of spin glasses, arXiv:1008.4844. M. Talagrand,
Spin glasses, a challenge for mathematicians (Springer-Verlag, 2003).
[8] G. Blatter, M. V. Feigelman, V. B. Geshkenbein, A. I. Larkin, and V. M. Vinokur,
Rev. Mod. Phys. 66, 1125 (1994). T. Giamarchi and P. Le Doussal, Statics
and dynamics of disordered elastic systems, arXiv:cond-mat/9705096. T. Nattermann and S. Scheidl, Adv. Phys. 49, 607 (2000). T. Giamarchi, A. B. Kolton,
A. Rosso, Dynamics of disordered elastic systems, arXiv:cond-mat/0503437. P.
Le Doussal, Exact results and open questions in first principle functional RG,
arXiv:0809.1192.
[9] O. C. Martin, R. Monasson, R. Zecchina, Statistical mechanics methods and phase
transitions in optimization problems Theoretical Computer Science 265 (2001)
3-67. M. Mezard and A. Montanari, Information, Physics, and Computation,
(Oxford Graduate Texts, 2009).

230

[10] D. J. Amit, Modeling Brain Function: The World Of Attractor Neural Networks,
(Cambridge Univ. Press, 1992). N. Brunel, Network models of memory, in Les
Houches 2003 (Elsevier, 2004).
[11] D. A. Fletcher and P. L. Geissler, Annu. Rev. Phys. Chem. 60, 469 (2009). G. I.
Menon, arXiv:1003.2032. S. Ramaswamy, arXiv:1004.1933.
[12] Dynamical heterogeneities in glasses, colloids, and granular media, L. Berthier,
G. Biroli, J-P Bouchaud, L. Cipelletti and W. van Saarloos eds. (Oxford University Press, Oxford, 2011).
[13] L. C. E. Struick, Physical Aging in Amorphous Polymers and Other Materials (Elsevier, Amsterdam, 1978). E. Vincent, J. Hammann, M. Ocio, JP Bouchaud, L. F. Cugliandolo, Slow dynamics and aging in spin-glasses,
arXiv:cond-mat/9607224 in Lecture Notes in Physics 492, 184 (1997).
[14] A.-L. Barabasi and H. E. Stanley, Fractal Concepts in Surface Growth, (Cambridge University Press, Cambridge, 1995). T. Halpin-Healey and Y.-C. Zhang,
Phys. Rep. 254, 215 (1995).
[15] R. Zwanzig, J. Stat. Phys. 9, 215 (1973). S. Nordholm and R. Zwanzig, J. Stat.
Phys. 13, 347-371 (1975). K. Kawasaki, J. Phys. A 6, 1289-1295 (1973).
[16] R. P. Feynmann and F. L. Vernon, Jr, Ann. Phys. 24, 114 (1963).
[17] Theory of the spin bath N. Prokofev and P. Stamp Rep. Prog. Phys. 63, 669
(2000).
[18] U. Weiss, Quantum dissipative systems, Series in modern condensed matter
physics vol. 10, World Scientific (1999).
[19] J. Dunkel and P. H
anggi, Phys. Rep. 471, 1-73 (2009).
[20] F. Marchesoni and P. H
anggi, Introduction: 100 years of Brownian motion,
Chaos 15, 026101 (2005).
[21] N. Pottier, Physica A 317, 371 (2003).
[22] A. B. Kolton, D. R. Grempel and D. Domnguez, Phys. Rev. B 71, 024206 (2005).
[23] C. Caroli, R. Combescot, P. Nozi`eres and D. Saint-James, J. Phys. C 5, 21 (1972).
[24] J. S. Langer, Statistical theory of decay of metastable states Ann. of Phys. 54,
258 (1969). C. G. Callan and S. Coleman, Phys. Rev. D 16, 1762 (1977). B.
Caroli, C. Caroli, and B. Roulet, Diffusion in a bistable potential - systematic
WKB treatment J. Stat. Phys. 21, 415 (1979). A. M. Polyakov, Gauge fields and
stings (Harwood Academic Publishers, 1994).

231

[25] S. Chandrasekhar, Stochastic problems in physics and astronomy, Rev. Mod.

Phys. 15, 1 (1943).
[26] H. Risken, The Fokker-Planck equation: methods of solution and applications,
(Springer-Verlag, Berlin, 1996).
[27] C. W. Gardiner, (1985). Handbook of Stochastic Methods for Physics, Chemistry,
and the Natural Sciences, (Springer-Verlag, Berlin, 1985).
[28] N. G. van Kampen, Stochastic Processes in Physics and Chemistry (NorthHolland, Amsterdam, 1992).
[29] W. T. Coffey, Yu. P. Kalmykov, and J. T. Waldron, The Langevin equation,
World Scientific series in contemporary chemical physics, vol. 14, 2nd ed. (World
Scientific, Singapore, 2005).
[30] N. G. van Kampen, It
o vs. Stratonovich, J. Stat. Phys. 24, 175 (1981).
[31] L. F. Cugliandolo and J. Kurchan. J. Phys. Soc. Japan 69, 247 (2000).
[32] P. Ilg and J-L Barrat, J. Phys. Conf. Series 40, 76 (2006).
[33] R. Bausch, H. K. Janssen, and H. Wagner, Z. Phys. B 24, 113 (1976).
[34] J. L. Iguain, S. Bustingorry, A. Kolton, and L. F. Cugliandolo, Growing correlations and aging of an elastic line in a random potential arXiv:0903.4878.
[35] L. F. Cugliandolo, J. Kurchan, and L. Peliti, Phys. Rev. E 55, 3898 (1997).
[36] C. Aron, G. Biroli, and L. F. Cugliandolo, J. Stat. Mech. (2011).
[37] A. Sicilia, J. J. Arenzon, A. J. Bray, and L. F. Cugliandolo, Phys. Rev. E
[38] see, e.g. R. Chitra, T. Giamarchi and P. Le Doussal, Phys. Rev. B 65, 035312
(2002) and references therein. H. Fukuyama and P. Lee, Phys. Rev. B 18, 6245
(1978).
[39] C. Aron, G. Biroli, and L. F. Cugliandolo, Phys. Rev. Lett. 102, 050404 (2009).
[40] M. Mezard and G. Parisi,
[41] L. F. Cugliandolo and P. Le Doussal, Phys. Rev. E L. F. Cugliandolo, J. Kurchan,
and P. Le Doussal, Phys. Rev. Lett.
[42] Phys. Rev. E 65, 046136 (2002).
[43] Grabert Phys. Rep.
[44] Z. Ovadyahu, Phys. Rev. B 73, 214204 (2006).

232

[45] R. MacKenzie, Path integral methods and applications, VIth Vietnam School
of Physics, Vung Tau, Vietnam, 27 Dec. 1999 - 8 Jan. 2000; arXiv:quantumph/0004090.
[46] R. P. Feynman and A. R. Hibbs, Quantum Mechanics and Path Integrals (New
York: McGraw-Hill, 1965). H. Kleinert, Path Integrals in Quantum Mechanics,
Statistics, Polymer Physics, and Financial Markets, 4th edition (World Scientific,
Singapore, 2004) J. Zinn-Justin, Path Integrals in Quantum Mechanics (Oxford
University Press, 2004), R. J. Rivers, Path Integrals Methods in Quantum Field
Theory (Cambridge University Press, 1987).
[47] R. P. Feynman, Statistical mechanics,
[48] J. Kurchan, Supersymmetry, replica and dynamic treatments of disordered systems: a parallel presentation arXiv:cond-mat/0209399.
[49] M. Mezard, G. Parisi, and M. A. Virasoro, Spin glasses and beyond, Fischer and
J. Hertz,
[50] F. Corberi, E. Lippiello and M. Zannetti J. Stat. Mech. (2007) P07002.
[51] Viasnoff and Lequeux
[52] L. F. Cugliandolo and G. Lozano, Phys. Rev. Lett. 80, 4979 (1998) . Phys. Rev.
B59, 915 (1999).
[53] L. F. Cugliandolo, D. R. Grempel, G. Lozano and H. Lozza Phys. Rev. B 70,
024422 (2004). L. F. Cugliandolo, D. R. Grempel, G. S. Lozano, H. Lozza, C. A.
da Silva Santos Phys. Rev. B 66, 014444 (2002).
[54] R. P. Feynmann and A. R. Hibbs, Quantum mechanics and path integrals (Mc
Graw-HIll, New York, 1965).
[55] A. J. Leggett, S. Chakravarty, A. T. Dorsey, M. P. A. Fisher, A. Garg, W.
Zwerner, Rev. Mod. Phys. 59, 1 (1987).
[56] A. J. Bray and M. A. Moore, Phys, Rev. Lett. 49, 1546 (1982).
[57] C. P. Martin, E. Siggia and H. A. Rose, Phys. Rev. A8, 423 (1973), H. K. Janssen,
Z. Phys. B23, 377 (1976) and Dynamics of critical phenomena and related topics,
Lecture notes in physics 104, C. P. Enz ed. (Springer Verlag, Berlin, 1979).
[58] H. K. Janssen, in From Phase Transitions to ChaosTopics in Modern Statistical
Physics, ed. by G Gy
orgyi et al. (World Scientific, Singapore, 1992).
[59] C. de Dominicis, Phys. Rev. B18, 4913 (1978).
[60] J. Schwinger, J. Math. Phys. 2, 407 (1961). L. V. Keldysh, Zh. Eksp. Teor. Fiz.
47, 1515 (1964), Sov. Phys. JETP 20 235 (1965).
233

[61] A. Kamenev,
[62] G. Parisi, Statistical Field Theory, Frontiers in Physics, Lecture Notes Series,
Addison-Wesley (1988).
[63] Cugliandolo, Grempel, Lozano, da Silva Santos

234