Livro Claudio

Claude Kipnis and Claudio Landim
Scaling Limit of Interacting

Particle Systems
Springer-Verlag
Berlin Heidelberg NewYork
London Paris Tokyo
Hong Kong Barcelona
Budapest
A minha lha Anna.
Aos meus pais Raul e Regina.
VI
Preface
The idea to write up a book on the hydrodynamic behavior of interacting particle

systems grew up after a series of lectures Claude gave at the University of Paris
7 in the spring of 1988. Claude wrote by this time some notes in French that
covered Chapters 1 and 4 and parts of Chapters 2, 5 and Appendix 1 of this book.
His intention was to prepare a text, as much as possible self-contained. It would
include, for instance, all tools from the Markov process theory (cf. Appendix 1,
Chaps. 2 and 4) necessary to enable mathematicians and mathematical-physicists
with some knowledge in probability, at the level of Chung (1974), to understand
the techniques of the theory of hydrodynamic limit of interacting particle systems.
In the fall of 1991 Claude invited me to complete with him his notes and
transform them into a book that would present to a large audience the latest
development of the theory in a simple and accessible form. To concentrate on the
main ideas and to avoid unnecessary technical difficulties, we decided to consider
systems evolving in finite lattice spaces and for which the equilibrium states are
product measures. To illustrate the techniques we chose two well known particle
systems, the generalized exclusion processes and the zero range processes. We also
conceived the book in such a manner that most chapters can be read independently
from the others.
Here are some comments that might help the reader to find his way. Let me
just mention that at the end of each chapter we give a list of references, we discuss
possible or proved extensions and we present a brief survey on some related topics
that could not be covered here.
To present the main concepts and the principal goals of the theory in the
simplest possible context, in Chapter 1 we prove the hydrodynamic behavior of
a system with no interaction among particles, more precisely for a superposition
of independent random walks. We recommend this chapter for someone who has
never studied this subject before. We explain there the need for a microscopic and
a macroscopic time scale, the natural time renormalization in order to observe non
trivial hydrodynamic phenomena for symmetric and asymmetric processes and the
concepts of local equilibrium, conservation of local equilibrium, hydrodynamic
behavior of interacting particle systems and hydrodynamic equations.
The purpose of Chapters 2 and 3 is to review the main properties of gener-
alized exclusion processes and zero range processes and to introduce some weak
VIII Preface
formulations of local equilibrium. These two chapters can be skipped by someone

familiar with interacting particle systems.
We start in Chapter 4 to present the main steps and the main tools in a general
proof of the hydrodynamic behavior of interacting particle systems. To illustrate
the approach we consider symmetric simple exclusion processes. The proof of the
hydrodynamic behavior of these systems turns out to be very simple because for
symmetric simple exclusion processes the density field is closed under the action
of the generator. We discuss there the natural topological spaces to consider and
we state some tightness results that are used throughout the book.
In Chapter 5 we explain the entropy method in the context of zero range pro-
cesses. Here the density field is no longer closed under the action of the generator
and one needs to project local fields on the density field. This is made possible
by the one and two blocks estimates that rely on bounds on the entropy and the
Dirichlet form obtained through the examination of the time evolution of the en-
tropy. The one and two blocks estimates are one of the most important results of
this book. Most of the following chapters depend on these estimates.
In Chapter 6 we present the relative entropy method, an alternative approach to
prove the hydrodynamic behavior of systems whose hydrodynamic equation admits
smooth solutions. This is a quite simple method to derive the hydrodynamic limit
of an interacting particle system. Its development led to the investigation of the
Navier–Stokes corrections to the hydrodynamic equations. This chapter can be
read independently from the others, we just need the one block estimate.
In Chapter 7 we extend to reversible nongradient systems, i.e., to processes in
which the instantaneous current cannot be written as the difference of a continuous
function and its translation, the entropy method presented in Chapter 5. The proof
of the hydrodynamic behavior of nongradient systems relies on a sharp estimate
for the spectral gap of the generator of the process restricted to finite cubes and
d
on the characterization of closed and exact forms of N Z . The need of a sharp
estimate on the spectral gap should be emphasized because it is completely hidden
in the proof and because one of the main open questions in the field consists in
obtaining a proof of the hydrodynamic behavior of nongradient systems that does
not use any information of the size of the gap. The sharp estimate on the spectral
d
gap and the characterization of closed and exact forms of N Z are contained in
Appendix 3. We also prove there estimates on the largest eigenvalue of small
perturbations of Markov generators.
In Chapter 8 we prove the hydrodynamic behavior of asymmetric attractive
processes in arbitrary dimensions. This chapter can also be read independently
from the others, we just use the one block estimate. The proof relies on the theory
of measure valued entropy solutions of hyperbolic equations and on the entropy
inequalities presented in Appendix 2.
In Chapter 9 we show how to derive the conservation of local equilibrium
from a law of large numbers for the local fields in the case of attractive systems.
This result together with the methods presented in Chapter 8 for asymmetric pro-
cesses and in Chapters 5, 6 and 7 for symmetric processes permits to prove the
Preface IX
conservation of local equilibrium for attractive systems. This chapter can be read
independently from the previous ones.
In Chapter 10 we prove a large deviation principle for the density field by
showing that the one and two blocks estimate are in fact superexponential. In
Chapter 11 we investigate the equilibrium fluctuations of the density field. The
nonequilibrium fluctuations remain a mainly open problem, though there exist
some one-dimensional results.
In Appendix 1 we present all the tools of Markov processes used in the book.
In Appendix 2 we derive, from the local central limit theorem for independent
and identically distributed random variables, estimates on the distance between
the finite marginals of canonical and grand canonical equilibrium measures. We
also prove some results on large deviations needed in Chapter 10, fix the termi-
nology and recall some well known results about weak solutions of hyperbolic
and parabolic quasi-linear partial differential equations.
In September 1993, when Claude died, Chapters 1, 2, 4, 5 and parts of Ap-
pendix 1 were ready and I decided to conclude alone our original project. I just
made one important modification : the language. All those who met Claude, re-
member certainly his incredible facility to learn languages. Even though he could
speak and write in perfect English, the natural language to write a scientific book
by the end of this century, Claude insisted in writing this book in French, in
some sense our native language. When he died in 1993, to increase the number
of potential readers I decided to switch to English. I hope Noemi will forgive me.
I take this opportunity to express my thanks to all our collaborators and col-
leagues from whom we learned most of the techniques presented here and all
those who read parts of the book for their comments and their encouragements.
I particularly acknowledge the influence on us of Pablo Ferrari, Antonio Galves,
Gianni Jona-Lasinio, Tom Liggett, Stefano Olla, Errico Presutti, Raghu Varadhan
and Horng-Tzer Yau. I would like also to thank Ellen Saada for her careful reading
of most of the chapters and for her uncountable suggestions and Olivier Benois
for his helpful support on all the subtleties of TEX.
Campos dos Goytacazes, April 1998. Claudio Landim

X Preface
Contents
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1. An introductory example : independent random walks . . . . . . . . . . 7
1.1 Equilibrium states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Local equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3 Hydrodynamic equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.4 Equivalence of ensembles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.5 Comments and references . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2. Some interacting particle systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.1 Some remarks on the topology of N Z and M1 (N Z ) . . . . . . . . .
d d
21
2.2 Simple exclusion processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3 Zero range processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4 Generalized exclusion processes . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.5 Attractive systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.6 Zero range processes in infinite volume . . . . . . . . . . . . . . . . . . . . 39
3. Weak formulations of local equilibrium . . . . . . . . . . . . . . . . . . . . . . . 43

4. Hydrodynamic equation of symmetric simple exclusion processes . 49
4.1 Topology and compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.2 The hydrodynamic equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5. An example of reversible gradient system : symmetric zero range

processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.1 The law of large numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.2 Entropy production . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.3 Proof of the replacement lemma . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.4 The one block estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.5 The two blocks estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.6 A L2 estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.7 An energy estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
XII Contents
6. The relative entropy method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

6.1 Weak conservation of local equilibrium . . . . . . . . . . . . . . . . . . . . . 122
7. Hydrodynamic limit of reversible nongradient systems . . . . . . . . . . 147

7.1 Replacing currents by gradients . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
7.2 An integration by parts formula . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.3 Nongradient large deviation estimates . . . . . . . . . . . . . . . . . . . . . . 161
7.4 Central limit theorem variances . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
7.5 The diffusion coefficient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
7.6 Compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
8. Hydrodynamic limit of asymmetric attractive processes . . . . . . . . . 197

8.1 Young measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
8.2 An entropy inequality at microscopic level . . . . . . . . . . . . . . . . . . 214
8.3 Law of large numbers for the empirical measure . . . . . . . . . . . . . 224
9. Conservation of local equilibrium for attractive systems . . . . . . . . . 237

9.1 Replacement lemma for attractive processes . . . . . . . . . . . . . . . . . 239
9.2 One block estimate without time average . . . . . . . . . . . . . . . . . . . 242
9.3 Conservation of local equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . 253
10. Large deviations from the hydrodynamic limit . . . . . . . . . . . . . . . . . 263

10.1 The rate function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
10.2 Weakly asymmetric simple exclusion processes . . . . . . . . . . . . . . 270
10.3 A superexponential estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
10.4 Large deviations upper bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
10.5 Large deviations lower bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
11. Equilibrium Fluctuations of Reversible Dynamics . . . . . . . . . . . . . . 293

11.1 The Boltzmann–Gibbs principle . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
11.2 The martingale problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
11.3 Tightness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
11.4 Generalized Ornstein–Uhlenbeck processes . . . . . . . . . . . . . . . . . . 314
Appendices
1. Markov chains on a countable space . . . . . . . . . . . . . . . . . . . . . . . . . . 319

Contents XIII
1.1 Discrete time Markov chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320

1.2 Continuous time Markov chains . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
1.3 Kolmogorov’s equations, generators . . . . . . . . . . . . . . . . . . . . . . . 330
1.4 Invariant measures, reversibility and adjoint processes . . . . . . . . . 334
1.5 Some martingales in the context of Markov processes . . . . . . . . . 338
1.6 Estimates on the variance of additive functionals of Markov
processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
1.7 The Feynman–Kac formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
1.8 Relative entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
1.9 Entropy and Markov processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
1.10 Dirichlet form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
1.11 A maximal inequality for reversible Markov processes . . . . . . . . 355
2. The equivalence of ensembles, large deviation tools and weak

solutions of quasi–linear differential equations . . . . . . . . . . . . . . . . . 357
2.1 Local central limit theorem and equivalence of ensembles . . . . . 357
2.2 On the local central limit theorem . . . . . . . . . . . . . . . . . . . . . . . . . 365
2.3 Remarks on large deviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
2.4 Weak solutions of nonlinear parabolic equations . . . . . . . . . . . . . 373
2.5 Entropy solutions of quasi–linear hyperbolic equations . . . . . . . . 378
3. Nongradient tools : spectral gap and closed forms . . . . . . . . . . . . . . 383

3.1 On the spectrum of reversible Markov processes . . . . . . . . . . . . . 385
3.2 Spectral gap for generalized exclusion processes . . . . . . . . . . . . . 388
3.3 Spectral gap in dimension d 2 . . . . . . . . . . . . . . . . . . . . . . . . . . 405
3.4 Closed and exact forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
XIV Contents
Frequently Used Notation
M+ (Td R+ ) 200 Tz ;z 392

(a b) 63, 144
0 1
C0 144, 151
< ; > 152 Cm 23
< ; >`;K 152 Cm;b 23
< ; > 63 F 152
< > 144 F` 152
< >;K 151 H1 norm 332
< (t; u); () > 196 H 145, 171
< ; g > 313 H 1 norm 332
< ;NG > 47 I 40
< d; G > 43
T
M+ () 43
Z
C (N d ) 21 M1 (f0; 1g N )
d
Z
26
d
C (2T d) 49 M1 (N ) 22
C m;n(T ) 55 M+;1
o 261
C m;n ([0; T ] d Td) 58 S 40
CK (R+ T ) 196 N 31, 49
Cb (E ) d 313 g 144
Z
Cb (N ) 21 23
D([0; T ]; M+ ) 48 ` 85
D(f ) 345 151
D`;K 160
k () 163
E 9 () 30, 36
Hx;y(A30C) 338 b 153
Im 210 144, 398
LC0 144, 172 1 404
L(2x; y) 322 N;K
27
L1() 327 ;K 350
L1 ([0; T ]) 200 d 12
Lsym([0; T ]; M+ (T R+ )) 200
N ;K 378
LN 80 k k 13
L` 143 S̄tN 238
MGtG 58, 67, 191 ¯' 350
NNt 58, 67 ¯';g 29
QN;` 195 ¯' 36
Q 1fAg 8
QN;`
198
() 144
N 198 u 43
QNN 195 `f 145
RN() 29, 36 `f 159
S N (t) 26 (x) 8
St 73 23
XVI Frequently Used Notation
`x;y(x) 77 x 42
26 x 13
^b 167 ~( ) 34, 42

" 69, 77 0 214

; 171
' 28
162 a b 44

g; h ;0 161
d(x; ) 163
g ;j 161
di;j () 150
24
ej 104
r 144 g 28
r 15 s 151
r 1b ;b g 143 sf0 145
N () 42 w ( ) 50
2
`;j 88
0
w ( ) 51
26, 30 E 9
N;K 27 PNN 69
;K 350 PN 141
`;K 159 P 9
N() 9 TN 7
N () 12 M+ 47
d 47
TdN;K 11 M+ (T )
0
@uNi 142, 300

M1 () 13
Ak 144, 161
NN;`(; du) 43 D(f ) 343
200 D(;K ; ) 152
j ¯ 238 Db (;K ; ) 152
tNx;y(du) 47 F 176
401 dx 106
x 281 [r] 12
Introduction
The problem we address in this book is to justify rigorously a method often used by
physicists to establish the partial differential equations that describe the evolution
of the thermodynamic characteristics of a fluid.
Suppose that we are interested in examining the evolution of a system con-
stituted of a large number of components or degrees of freedom, e. g., a fluid
or a gas. Since the total number of components is very large (typically of the
order of 1023 ), we are not interested nor able to give a precise description of the
microscopic state of the system. Hence, following the statistical mechanics ap-
proach introduced by Boltzmann, we first examine the equilibrium states of the
the system and characterize them by a small number of macroscopic quantities
p = (p1 ; : : : ; pa ), called the thermodynamic characteristics, e. g. the temperature,
the density, the pressure (which are often related through the equation of state).
Once the equilibrium states are known, we may investigate the evolution of
a system out of equilibrium. Assume that the gas is confined in a volume V .
For each point u in V , denote by Vu a small neighborhood of u, that is, small if
compared to the total volume V , but large enough if compared to the intermolecular
distance, in order to assume that each neighborhood contains an infinite number
of particles. Due to the “strong" interaction among the molecules, it is natural to
believe that the system reaches immediately a local equilibrium state, i.e., that in
each neighborhood Vu the system is close to an equilibrium state characterized by
the parameter p(u), that may depend on u.
As the system evolves, we expect this local equilibrium picture to evolve in a
smooth way. More precisely, one would expect to observe at a later time t in a
small neighborhood Vu a state close to a new equilibrium state characterized by
a parameter p(t; u), which now depends on space and time. Moreover, it seems
reasonable to believe that p(t; u) evolves smoothly in time according to a partial
differential equation, called the hydrodynamic equation.
Despite many efforts, this program has not been completely achieved for
Hamiltonian systems where particles evolve deterministically according to New-
ton’s equations, mainly due to the lack of good ergodic properties of the system.
Two simplifications have traditionally then been made. Either to assume the evo-
lution of the microscopic system to be stochastic or to consider systems with
low density of particles in such a way that the total number of collisions of each
particles remains finite in finite intervals of time in order to avoid correlations.
2 Introduction
Our main purpose in this book is to present general methods that permit to de-
duce the hydrodynamic equations of the thermodynamic characteristics of infinite
systems assuming that the underlying microscopic dynamics is stochastic, i.e., to
deduce the macroscopic behavior of the system from the microscopic interaction
among particles. In all cases the microscopic dynamics will consist of random
walks on the lattice submitted to some local interaction, the so-called interacting
particle systems introduced by Spitzer (1970).
For simplicity, we shall assume that there is a unique conserved quantity :
the total number of particles. Moreover, to highlight the main ideas and to avoid
unnecessary technical difficulties, we shall present all results for processes whose
equilibrium states are product measures. It should be emphasized, however, that
all methods presented here depend on the explicit knowledge of the equilibrium
states.
To illustrate the type of results we are going to prove, we conclude this intro-
duction deriving the hydrodynamic equation of an interacting particle system under
general assumptions on the microscopic evolution. To fix ideas, consider a system
evolving on a one-dimensional torus, denoted by T = [0; 1). Fix a positive integer
N , that represents the inverse of the distance between particles and that eventually
will increase to infinity. Denote by (x) the total number of particles at time 0
in the interval [x=N; (x + 1)=N ). The configuration = f (0); : : :; (N 1)g is
therefore an element of N TN , provided TN stands for the discrete one-dimensional
torus with N points : TN = f0; : : : ; N 1g. Two space scales appear naturally.
The macroscopic scale T and the microscopic scale TN . Points of the macroscopic
scale T are denoted by the letters u, v and correspond to the sites [uN ], [vN ] in
the microscopic scale. Here, for a real number a, [a] stands for the integer part
of a. Sites in the microscopic scale TN are denoted by the letters x, y , z and
correspond to the points x=N , y=N , z=N in the macroscopic scale T.
The time evolution of the system can be informally described as follows. We
fix cylinder functions fc(x; 1; ); x 2 TN g. If the state of the process is , at
rate c(x; 1; ) a particle jumps from x to x 1. This dynamics corresponds to a
Markov process whose generator LN is given by
X
(LN f )( ) = c(x; y; )[f (x;x+y) f ()] :
x2TN
jyj=1
In this formula x;x+y is the configuration obtained from by letting a particle
jump from x to x + y :
8 (z ) z 6= x, x + y;
< if
x;x+y (z ) = : ((xx)+ y)1+ 1 if z = x;
if z = x + y:
We shall assume that the process is translation invariant in the sense that the rate
c(x; y; ) is equal to c(y; x ) for two cylinder functions c(1; ). Here x stands
for the configuration translated by x so that (x )(y ) = (x + y ). This group of
translations is naturally extended to functions by the identity (x f )( ) = f (x ).
Introduction 3
To avoid degeneracies, we assume that c(1; ) > 0 if (0) > 0 so that a

jump from x to x 1 is possible whenever there is at least one particle at x. Of
course, when there are no particles at x a jump is impossible so that c(1; ) = 0
if (0) = 0. P
Denote by K the total number of particles at time 0 : K = x2TN (x).
Notice that the total number of particles is conserved by the dynamics and that
the process is irreducible for each fixed number of total particles. Hence, for each
fixed K , ft ; t 0g is an irreducible finite state Markov process. In particular,
there exists a unique invariant measure, denoted by N;K . Since the process is
translation invariant, so is the measure N;K . Therefore, since N;K is concentrated
on configurations with K particles, EN;K [ (x)] = K=N for all x in TN . Here,
for a probability measure , E [] stands for the expectation with respect to .
We shall assume that these invariant measures have nice local properties as
N " 1. More precisely, that there exists product probability measures f ;
0g on N Z such that the finite marginals of N;K , as N " 1 and K=N ! ,
converge to the finite marginals of :
lim E [f ] = E [f ]
N !1 N;K
K=N !
for every bounded cylinder function f . The measures inherit all properties of
the measure N;K . For each fixed 0, is translation invariant, invariant for
the infinite volume dynamics and parametrized by the density : E [ (0)] = .
The requirement that the measures are product is not essential but simplifies
the exposition and will systematically be satisfied by the examples considered here.
This property permits to introduce in a simple way the idea of local equilibrium.
Fix a continuous function : T ! R+ . Denote by N() the product measure on
N TN with marginals given by
N() f; (x) = kg = (x=N ) f; (0) = kg :

Thus, at site x (that corresponds to the macroscopic location x=N ) we place
particles according to the invariant distribution with density (x=N ). It is easy
to check that locally around the macroscopic point u 2 T the measure N() is
close to the invariant measure with density (u). Indeed, since the measure N()
is product, for every positive integer ` and every sequence (a1 ; : : : ; a` ),
h Ỳ i h Ỳ i
EN 1f (x + [uN ]) = ax g E u 1f (x) = ax g :
N !1
lim = ( )
x=1 x=1
( )
We shall now deduce the macroscopic evolution of the density profile ()
assuming that the local equilibrium is conserved. More precisely, assume that at
each macroscopic time t, that will correspond to a microscopic time t(N ) for
some scaling (N ), the state of the process is still in local equilibrium, i.e., that
around each u and at each time t the process is close to some equilibrium state.
4 Introduction
Denote by (t; ) the density profile at time t, so that at time t and in a small
neighborhood of u the state of the process is close to (t;u) .
Denote by Wx;x+1 the instantaneous current between x and x + 1, i.e, the rate
at which a particle jumps from x to x + 1 minus the rate at which a particle jumps
from x + 1 to x. Since the process is translation invariant, Wx;x+1 = x W0;1 and
W0;1 = c(1; ) c( 1; 1). There are two cases to be considered : Either in the
stationary regime the current has mean zero with respect to all invariant measures
or not. Assume first that E [W0;1 ] does not vanish uniformly in .
Fix a smooth function H : T ! R and consider the martingale MtH defined by
X X
MtH = N 1
H (x=N )t(x) N 1
H (x=N )(x)
x2TN x2TN
Zt X
ds N 1
H (x=N )LN s (x) :
0 x2TN
By definition of the current, LN (x) = Wx 1;x Wx;x+1. Therefore, after a sum-
mation by parts, the martingale MtH becomes
X X
N 1
H (x=N )t(x) N 1
H (x=N )(x)
x2TN x2TN
Zt X
ds N 2
(@ N H )(x=N )Wx;x+1(s) :
0 x2TN
In this formula (@ N H ) stands for the discrete derivative of H : (@ N H )(x=N )
= N [H ((x + 1)=N ) H (x=N )]. If we now rescale time by N , i.e., if we consider
times of order N and change variables in the time integral, the martingale MNt H
turns out to be equal to
X X
N 1
H (x=N )tN (x) N 1
H (x=N )(x)
x2TN x2TN
Zt X (0:1)
ds N 1
(@ N H )(x=N )Wx;x+1(sN ) :
0 x2TN
Since MtH is a martingale vanishing at time 0, its expectation is equal to 0
uniformly in time. Since we are assuming that the system is in local equilibrium at
the macroscopic time t, which corresponds in this set up to the microscopic time
tN , the expectation of tN (x) is close to the expectation of (0) with respect to
(t;x=N ) because the site x in the microscopic scale corresponds to the point x=N
in the macroscopic scale. By the same reasons, the expectation of Wx;x+1 (sN )
is close to the expectation of W0;1 with respect to (s;x=N ) . Therefore, taking
expectation in (0.1), up to lower order terms, we have that
Introduction 5
X X
N 1
H (x=N )(t; x=N ) N 1
H (x=N )(x=N )
x2TN x2TN
Zt X
= ds N 1
(@ N H )(x=N )W
~ ((s; x=N )) ;
0 x2TN
provided W ~ () stands for the expectation of the current W0;1 with respect to the
invariant measure with density : W ~0;1 () = E [W0;1 ( )]. Since this expectation
does not vanishes uniformly, integrating by parts, we obtain that the density (t; u)
is a weak solution of the first order partial differential equation
(
@t + @u W~ () = 0 ;
(0:2)
(0; ) = 0 () :
In the case where the current has mean zero with respect to all equilibrium
measures, assume that it can be written as the difference of a continuous function
and its translation : W0;1 = h 1 h. Interacting particle systems satisfying this
assumption are called gradient processes. We discuss this property in Chapters
5 and 7. This assumption permits a second summation by parts in (0.1). The
martingale MtH is now equal to
X X
N 1
H (x=N )t (x) N 1
H (x=N )(x)
x2TN x2TN
Zt X
ds N 3
[@ N (@ N H )]((x 1)=N )x h(s ) :
0 x2TN
Rescaling now time by N 2 , by the assumption of conservation of local equilibrium
H 2 vanishes, up to lower order terms, we have that
and since the expectation of MtN
X X
N 1
H (x=N )(t; x=N ) N 1
H (x=N )(x=N )
x2TN x2TN
Zt X
= ds N 1
[@ N (@ N H )]((x 1)=N )h~((s; x=N )) ;
0 x2TN
where h~() is the expectation of the continuous function h with respect to the
invariant measure with density : h~() = E [h( )]. Therefore, (t; u) is a weak
solution of the parabolic equation
(
@t = @u2 h~() ;
(0:3)
(0; ) = 0 () :
It is therefore easy to derive the macroscopic evolution of the density profile
() under the assumption of local equilibrium. The purpose of this book is to
present general methods that permit to prove the convergence in probability of the
density field
6 Introduction
X
N 1
H (x=N )t(N )(x)
x2TN
R
to the integral T du H (u)(t; u), where is the weak solution of a partial dif-
ferential equation of type (0.2) or (0.3) and (N ) a rescaling constant. In some
cases we will be able to deduce from this convergence the conservation of local
equilibrium (cf. Chap 9).
1. An Introductory Example : Independent Random
Walks
The main purpose of this book is to present general methods that permit to deduce
the hydrodynamic equations of interacting particle systems from the underlying
stochastic dynamics, i.e., to deduce the macroscopic behavior of the system from
the microscopic interaction among particles.
In order to present all main concepts that will appear throughout the book in a
very simple setting, we consider in this chapter a system in which particles evolve
according to independent continuous time random walks, a special case where
there is no interaction among particles. In section 1 we describe all equilibrium
states. In our stochastic context this corresponds to the specification of all invariant
measures. We show further that these equilibrium states can be parametrized by the
density of particles, which is also the unique quantity conserved by the stochastic
dynamics. We then introduce in section 2 the concept of local equilibrium in our
mathematical model. In section 3 we prove that the local equilibrium picture is
conserved by the time evolution and deduce the (linear) hydrodynamic equations
that describe the macroscopic evolution of the density. In section 4 we further
discuss some properties of the equilibrium states.
1. Equilibrium states
In this chapter we investigate in detail the case of indistinguishable particles mov-

ing as independent random walks. Denote by Z the set of integers and by Zd the
d-dimensional lattice. For a positive integer N , denote by TN the torus with N
points : TN = Z=N Z and let TdN = (TN )d . Here N represents the inverse of the
distance between molecules. The points of TdN , called sites, are represented by the
last characters of the alphabet (x, y and z ).
To describe the evolution of the system, we begin by distinguishing all par-
ticles. Let K denote the total number of particles at time 0 and let x1 ; : : : ; xK
denote their initial positions. Particles evolve as independent translation invariant
continuous time random walks on the torus. To be rigorous, fix a translation in-
variant transition probability p(x; y ) on Zd : p(x; y ) = p(0; y x) =: p(y x) for
some probability p() on Zd, called the elementary transition probability of the
system.
8 1. An Introductory Example : Independent Random Walks
Let pt (x; y ) represent the probability of being at time t on site y for a

continuous time random walk with elementary transition probability p() start-
ing from x. pt ( ; ) inherits the translation invariance property from p( ; ) :
pt (x; y) = pt (0; y x) =: pt (y x). Moreover, by the Markov property, pt ( ; )
is the unique solution of the linear differential equations
8 X h i
>
< @t pt(x; y) = p(x; z) pt(z; y) pt(x; y) ;
: p (x; y) = 1zf2xZ= yg :
> d
0
In this formula @t pt stands for the time derivative of pt and 1fx = y g for a
function which is equal to 1 if x = y and 0 otherwise.
We are now in a position to describe the motion of each particle. Denote by
fZti ; 1 i K g K independent copies of a continuous time random walk with
elementary transition probability p() and initially at the origin. For 1 i K ,
let Xti represent the position at time t of the i-th particle. We set
Xti = xi + Zti mod N:
Since particles are considered indistinguishable, we are not interested in the
individual position of each particle but only in the total number of particles at each
site. In particular the state space of the system, also called configuration space,
d
is N TN . The configurations will be denoted by Greek letters , and . In this
way, for a site x of TdN , (x) will represent the number of particles at site x for
the configuration . Therefore, if the initial positions are x1 ; : : : ; xK , for every
x 2 TdN :
X
K
(x) = 1fxi = xg :
i=1
Inversely, given f (x); x 2 TdN g, to define the evolution of the system, we
can first label all particles and then let them evolve according to the stochastic
dynamics described above.
Clearly, if we denote by t the configuration at time t, we have
X
K
t (x) = 1fXti = xg :
i=1
The process (t )t0 inherits the Markov property from the random walks fXti ; 1
i K g because all particles have the same elementary transition probability and
they do not interact with each other.
The first question raised in the study of Markov processes is the characteriza-
tion of all invariant measures. Since the state space is finite and since the total
number of particles is the unique quantity conserved by the dynamics, for every
positive integer K representing the total number of particles, there is only one
invariant measure, as long as the support of the elementary transition probability
p( ) generates Zd. The Poisson measures will, however, play a central role.
1. Equilibrium states 9
Recall that a Poisson distribution of parameter 0 is the probability measure

fp;k = pk ; k 1g on N given by
k
pk = e k! ; k 2 N ;
and its Laplace transform is equal to
X
1 k
e e k k! = e ee
= exp (e 1)
k=0
for all positive .
For a fixed positive function : TdN ! R+ , we call Poisson measure on TdN
d
associated to the function a probability on the configuration space N TN , denoted
N N
by () , having the following two properties. Under () , the random variables
f(x); x 2 TdN g representing the number of particles at each site are independent
and, for every fixed site x 2 TdN , (x) is distributed according to a Poisson
distribution of parameter (x). In the case where the function is constant equal
to , we denote N() just by N . Throughout this book, expectation with respect
to a measure will be denoted by E .
The measure N() is characterized by its multidimensional Laplace transform :
2 3
n X o Y
EN 4exp (x) (x) 5 = exp (x) e (x) 1
x2TdN x2TdN
( )
X
= exp (x) e (x) 1
x2TdN
for all positive sequences f(x); x 2 TdN g (cf. Feller (1966), Chap. VII).
The first result consists in proving that the Poisson measures associated to
constant functions are invariant for a system of independent random walks.
Proposition 1.1 If particles are initially distributed according to a Poisson mea-

sure associated to a constant function equal to then the distribution at time t is
exactly the same Poisson measure.
d
Proof. Denote by PN the probability measure on the space D(R+ ; N TN ) induced
by the independent random walks dynamics and the initial measure N . Expecta-
tion with respect to PN is denoted by E N . Notice the difference between EN
and E N . The first expectation is an expectation with respect to the probability
d
measure N defined on N TN , while the second is an expectation with respect to
d
the probability measure PN defined on the path space D(R+ ; N TN ). In particular,
d
on N TN .
E N [F (0 )] = EN [F ( )] for all bounded continuous function F
d
Since a probability measure on N TN is characterized by its multidimensional
Laplace transform, we are naturally led to compute the expectation
2 3
X
E N 4exp (x) t (x)5
x2TdN
for all positive sequences f(x); x 2 TdN g. For a site y in TdN , we will denote
by Xty;k the position at time t of the k -th particle initially at y . In this way, the
number of particles at site x at time t is equal to
X Xy 0( )
t (x) = 1fXty;k = xg :
y2TdN k=1
From this formula and inverting the order of summations we obtain the identity
X X Xy 0( )
(x) t (x) = (Xty;k ) :
x2TdN y2TdN k=1
Since each particle evolves independently and the total number of particles at
each site at time 0 is distributed according to a Poisson distribution of parameter
, 0 1
X
E N @exp (x) t (x)A
x2TdN
Y " ( Xy 0( )
)#
= E N exp (Xty;k )
y2TdN k=1
YZ h i y)
N (d) E exp (Xty; )
0(
1
=
y2TdN
Y y Xt
= exp E e 1 ; ( + )
y2TdN
where Xt is the position at time t of a random walk on the torus TdN starting from
t ( ) defined by
the origin and with transition probability pN
X
pNt (x; y) := pt (x; y + Nz )
z2Zd
for x and y in TdN . Since
X
E e (y+Xt ) = pNt (x y)e (x);
x2TdN
inverting the order of summation, we obtain that
2. Local Equilibrium 11
0 1 8 9
X <X = :
E N @exp (x) t (x)A (x)
x2TdN
= exp
:x2Td e 1
;
N
P
Remark 1.2 Since the total number of particles x2Td (x) is conserved by the
N
stochastic dynamics it might seem more natural to consider as reference probability
measures the extremal invariant measures that are concentrated on the “hyper–
planes" of all configurations with a fixed total number of particles. These measures
are given by
X
TdN;K : = N (x) = K :
x2TdN
Besides the fact that they enable easier computations, the Poisson distributions
present other intrinsic advantages that will be seen in the forthcoming sections.
We shall return to this discussion on extremal invariant measures in section 4.
Notice that only one quantity is conserved by the dynamics : the total number
of particles. On the other hand, Poisson distributions are such that their expectation
X
is equal to
k
e k! k = :
k 0
The Poisson measures are in this way naturally parametrized by the density of
particles. Furthermore, by the weak law of large numbers, if the number of sites
of the set TdN is denoted by jTdN j,
X
1
N !1 jTdN j
lim (x) =
x2TdN
in probability with respect to N . The parameter describes therefore the “mean"
density of particles in a “large" box.
In conclusion, we obtained above in Proposition 1.1 a one–parameter family
of invariant and translation invariant measures indexed by the density of particles,
which is the unique quantity conserved by the time evolution.
2. Local Equilibrium
We announced that the passage from microscopic to macroscopic would be done

performing a limit in which the distance between particles converges to zero. This
point does not present any difficulty in formalization. We just have to consider
the torus TdN as embedded in the d-dimensional torus Td = [0; 1)d, that is, taking
the lattice Td with “vertices" x=N , x 2 TdN . In this way the distances between
molecules is 1=N and tends to zero as N " 1.
We shall refer to Td as the macroscopic space and to TdN as the microscopic
space. In this way each macroscopic point u in Td is associated to a microscopic
site x = [uN ] in TdN and, reciprocally, each site x is associated to a macroscopic
point x=N in Td . Here and below, for a d-dimensional real r = (r1 ; : : : ; rd ), [r]
denotes the integer part of r : [r] = ([r1 ]; : : : ; [rd ]).
On the other hand, since we have a one–parameter family of invariant mea-
sures, one way to describe a local equilibrium with density profile 0 : Td ! R+ is
the following. We distribute particles according to a Poisson measure with slowly
varying parameter on TdN , that is, for each positive N we fix the parameter of
the Poisson distribution at site x to be equal to 0 (x=N ). Since this type of mea-
sure will appear frequently in the following chapters, we introduce the following
terminology.
Definition 2.1 (Product measure with slowly varying parameter associated to

a profile 0 : Td ! R+ ). For each smooth function 0 : Td ! R+ , we represent
d
by N0 () the measure on the state space Td = N TN having the following two
N
properties. Under N0 () the variables f (x); x 2 TdN g are independent and, for a
site x 2 TdN , (x) is distributed according to a Poisson distribution of parameter
0 (x=N ) :
N0 () f; (x) = kg = N0 (x=N ) f; (0) = kg
for all x in TdN and k in N .
We have thus associated to each profile 0 : Td ! R+ and each positive integer
N a Poisson measure on the torus TdN .
As the parameter N increases to infinity, the discrete torus TdN tends to the full
lattice Zd . We can also define a Poisson measure on the space of configurations
over Zd. For each 0 we will denote by the probability on N Z that makes
d
the variables f (x); x 2 Zdg independent and under which, for every x in Zd,
(x) is distributed according to a Poisson distribution of parameter .
With the definition we have given of N0 () , and since 0 : Td ! R+ is assumed
to be smooth, as N " 1 and we look “close" to a point u 2 Td – that is “around"
x = [Nu] – we observe a Poisson measure of parameter (almost) constant equal
to 0 (u). In fact, since the function 0 () is smooth, for every positive integer `
and for every positive family of parameters f(x); jxj `g,
P P
EN e jxj` (x)([uN ]+x) E e jxj` (x)(x) : (2:1)
N !1
lim = u
0( )
0( )
In this formula and throughout this book, for u = (u1 ; : : : ; ud ) in Rd , kuk denotes
the Euclidean norm of u and juj the max norm :
X
kuk2 = u2i ; juj = max ju j :
id i
id
1
1
3. Hydrodynamic equation 13
In this sense the sequence N0 () describes an example of local equilibrium.
This definition of product measure with slowly varying parameter is of course too
restrictive. To generalize it we introduce a concept of convergence.
d
In the configuration space N TN , endowed with its natural discrete topology,
we denote by fx ; x 2 TN g the group of translations. Thus, for a site x, x is
d
the configuration that, at site y , has (x + y ) particles :
(x )(y ) = (y + x); y 2 TdN :

The action of the translation group extends in a natural way to the space of
d
functions and to the space of probability measures on N TN . In fact, for a site x
and a probability measure , (x ) is the measure such that
Z Z
f ()(x)(d) = f (x )(d);
for every bounded continuous function f .
d d
To perform the limit N " 1 we embed the space N TN in N Z identifying a
configuration on the torus to a periodic configuration on the full lattice. We will
d
endow the configuration space N Z with its natural topology, the product topology.
By M1 (N Z ) or simply by M1 , we represent the space of probability measures
d
d
on N Z endowed with the weak topology.
In this topological setting, formula (2.1) establishes that for all points u of Td
the sequence [uN ] N0 () converges weakly to the measure 0 (u) . It is then natural
to introduce the following definition.
Definition 2.2 (Local equilibrium) A sequence of probability measures (N )N 1

d
on N TN is a local equilibrium of profile 0 : Td ! R+ if
lim
N !1 [uN ]
N = 0 (u)
for all continuity points u of 0 ().
3. Hydrodynamic equation
We turn now to the study of the distribution of particles at a later time t starting
from a product measure with slowly varying parameter. Repeating the compu-
tations we did to prove Proposition 1.1 we see that if we start from a Poisson
measure with slowly varying parameter then
0 1
X
E N @exp (x) t (x)A
x2TN
0( )
d
X X N
= exp 0 (x=N ) pt (y x) e (y) 1
x2TdN y2TdN
X X pN (y x) (x=N )
= exp e (y) 1 t 0
y2TdN x2TdN
X (y ):
=: exp e (y) 1 N;t
y2TdN
Therefore, at time t, we still have a Poisson measure with slowly varying param-
eter, which is now N;t () instead of 0 (=N ). Up to this point we have not used
the particular form of pt () besides the fact that it makes pt (; ) translation invari-
P
ant and thus bistochastic : x pt (x; y ) = 1 for every y . We shall now see what
happens when t is fixed and N increases to infinity. In this case pt () is a function
with essentially finite support, that is, for all " > 0, there exists A = A(t; ") > 0
so that X
pt (x) 1 " :
jxjA
From the explicit form of N;t , we have that for every continuity point u of 0 ,
([uN ]) = 0 (u) :
N !1 N;t
lim
The profile has remained unchanged. The system did not have time to evolve
and this reflects the fact that at the macroscopic scale particles did not move. In-
deed, consider a test particle initially at the origin. Since it evolves as a continuous
time random walk, if Xt denotes its position at time t, for every " > 0, there exists
A = A(t; ") > 0 such that P [ jXtj > A] ". Therefore, with probability close
to 1, in the macroscopic scale, the test particle at time t is at distance of order
N 1 from the origin. In a fluid, however, a “test" particle traverses a macroscopic
distance in a macroscopic time, as do, for instance, particles in suspension.
We solve this problem distinguishing between two time scales (as we have two
space scales : Td and N 1 TdN ) : a microscopic time t and a macroscopic time
which is infinitely large with respect to t.
To introduce the macroscopic time scale, notice that the transition probabilities
pt ( ) are equal to
X
1 k
pt (x) = e t kt ! pk (x) ;
k=0
where pk stands for the k -th convolution power of the elementary transition
probability of each particle.
Assume that the elementary transition probability p( ) has finite expectation :
P
m := x p(x) 2 Rd . We say that the random walk is asymmetric if m 6= 0,
3. Hydrodynamic equation 15
that it is mean–zero asymmetric if p() is not symmetric but m = 0 and that it is

symmetric if p() is symmetric. Recall that Xt stands for the position at time t of
a continuous time random walk with transition probability p() and initially at the
origin. By the law of large numbers for random walks, for all " > 0,
X h XtN i
lim
N !1 x; jx=N mtj"
ptN (x) = lim
N !1
P N mt " = 1:
In particular, from the explicit expression for N;tN and since we assumed the
initial profile to be smooth, we have that
([uN ]) = 0 (u mt) = : (t; u)

N !1 N;tN
lim
for every u in Td .
We obtained in this way a new time scale, tN , in which we observe a new
macroscopic profile : the original one translated by mt. More precisely, in this
macroscopic scale tN we observe a local equilibrium profile that has been trans-
lated by mt since N;tN is itself slowly varying in the macroscopic scale.
Of course, the profile (t; u) satisfies the partial differential equation
@t + m r = 0 (3:1)
if r denotes the gradient of : r = (@u ; : : : ; @ud ).

1
In conclusion, if we restrict ourselves to a particular class of initial measures,
we have established the existence of a time and space scales in which the particles
density evolves according to the linear partial differential equation (3.1). We have
thus derived from the microscopic stochastic dynamics a macroscopic deterministic
evolution for the unique conserved quantity.
An interacting particle system for which there exists a time and space macro-
scopic scales in which the conserved quantities evolve according to some partial
differential equation is said to have a hydrodynamic description. Moreover, the
P.D.E. is called the hydrodynamic equation associated to the system.
We summarize this result in the following proposition.
Proposition 3.1 A system of particles evolving as independent asymmetric ran-

dom walks with finite first moment on a d-dimensional torus has a hydrodynamic
description. The evolution of the density profile is described by the solution of the
differential equation
@t + m r = 0 :
When the expectation m vanishes, the solution of this differential equation is

constant, which means that the profile didn’t change in the time scale tN . Nothing
imposed, however, the choice of Nt as macroscopic time scale. In fact, when the
mean displacement m vanishes, to observe an interesting time evolution, we need
to consider a larger time scale, times of order N 2 .
Assume that the elementary transition probability that describes the displace-
ment of each particle has a second moment. Let = (i;j )1i;j d be the covariance
matrix of this distribution :
X
i;j = x i x j p ( x) ; 1 i; j d:
x2TdN
By the central limit theorem for random walks, we see that
X
lim 2 ([Nu]) =
N !1 N;N t
lim
N !1
pNt ([Nu] x)0 (x=N )
x2TdN
h i Z
= lim
N !1
E 0 (u N 1 XtN ) 2 = 0 ()Gt (u ) d ;
Rd
where 0 : Rd ! R+ is the periodic function, with period Td and equal to 0 on
the torus Td and Gt is the density of the Gaussian distribution with covariance
matrix t .
Since the Gaussian distribution is the fundamental solution of the heat equation
(which can be checked by a simple computation) we obtain the following result.
Proposition 3.2 A system of particles evolving as independent mean–zero asym-

metric random walks with finite second moment on a d-dimensional torus has a
hydrodynamic description. The evolution of the density profile is described by the
solution of the differential equation
8 X
>
< @t = i;j @ui ;uj 2
> i;j d
: (0; u) = (u) :
1
Let fS N (t); t 0g be the semigroup on M1 associated to the Markov pro-

cess (t )t0 . In Propositions 3.1 and 3.2, we have proved that there is a time
renormalization N such that
lim
N !1
S N (tN ) [uN ] N() 0
= (t;u) ;
for all t 0 and all continuity points u of (t; ). It is therefore natural to introduce
the following definition.
Definition 3.3 (Conservation of local equilibrium) The local equilibrium (N )N 1

of profile 0 () is conserved by the time renormalization N if there exists a func-
tion : R+ Td ! R+ such that
lim
N !1
S N (tN ) [uN ] N = (t;u) ;
for all t 0 and all continuity points u of (t; ).
4. Equivalence of ensembles 17
Usually (t; ) is the solution of a Cauchy problem with initial condition 0 ().
As we said earlier, this differential equation is called the hydrodynamic equation of
the interacting particle system. In this section we took advantage of several special
features of the evolution of independent random walks to obtain an explicit formula
for the profile (t; ). The type of result, however, is characteristic of the subject.
We have proved :
1. conservation of the local equilibrium in time evolution.
2. characterization at a later time of the new parameters describing the local
equilibrium and derivation of a partial differential equation that determines how
the parameters evolve in time.
The aim of the following chapters is to prove a weak version of the conser-
vation of local equilibrium for a class of interacting particle systems. We would
like in fact to prove a more general result, that is, one for initial states that are not
product measures with slowly varying parameter – thus without assuming a strong
form of local equilibrium at time 0 – but for initial states having a density profile
and imposing that it is not too far, in a sense to be defined later, from a local
equilibrium; the process establishing by itself a local equilibrium at later times.
4. Equivalence of ensembles
In Section 2 we chose a class of invariant measures to describe the equilib-

rium states (the Poisson measures) when others would seem more appropriate.
Indeed, since the “total number of particles" is conserved by the evolution it
P
would have seemed more natural to choose the so-called canonical measures
N j x2TdN (x) = K (that do not depend on ).
However, since we want to describe the equilibrium state associated to a given
density on the torus Td , we would be led to study the behavior, as N " 1 and
is kept fixed, of 0 1
X
N @ (x) = N d A :
x2TdN
A simple computation presented below shows that for each fixed positive integer
r, and for all sequences (k1 ; : : : ; kr ) in N r and (x1 ; : : : ; xr ) in Zr,
0 1
X
lim N @ (x ) = k ; : : : ; (xr ) = kr (x) = N d A
N !1 1 1
x2TdN
= (x ) = k ; : : : ; (x ) = k :

1 1 r r
Indeed, the addition of independent Poisson distributions is still a Poisson dis-
tribution of parameter equal to the sum of the parameters, as we can see computing
the Laplace transforms. Therefore, the left hand side of the above identity is equal
to
" #
k ++kr e N d [(N d r)]N d (k ++kr ) e N d (N d )N d
1
1 1
k1 ! : : : k r ! [N d (k1 + + kr )]! (N d )!
(N d r)N (k ++kr ) d
d
d d
k1 ! kr !(N d)N d (N )(N 1) [N (k1 + + kr ) + 1]
1
=
that, as N " 1, converges to
r k1 ++kr
k1 ! : : : kr ! e :
1
It is in this sense, known to the physicists as the “equivalence of ensembles",

that the Poisson measures are indeed “natural" in our problem. One advantage
that they present among others is that computations are much easier and that the
definition of local equilibrium is expressed in a very simple and elegant way in
terms of these measures.
P
In Appendix 2 we present in more detail and in a wide context some results
connecting the canonical measures N j x2TdN (x) = K to the grand
canonical measures (). We obtain, for instance, some estimates of the total
variation distance between these two measures.
5. Comments and References
The mathematical formulation of the derivation of macroscopic evolution equa-

tions from microscopic interactions goes back to Morrey (1955). The first rigor-
ous results are due to Dobrushin and Siegmund–Schultze (1982), Rost (1981) and
Galves, Kipnis, Marchioro and Presutti (1981). Dobrushin and Siegmund–Schultze
(1982), that we essentially followed, consider the evolution of independent copies
of stochastic processes on Rd and derive linear first order hydrodynamic equations.
Rost (1981) obtains a first order quasi–linear hyperbolic equation for the one di-
mensional totally asymmetric simple exclusion process. Galves, Kipnis, Marchioro
and Presutti (1981) deduce a non linear heat equation that describes the macro-
scopic evolution of a one dimensional non conservative spin system. De Masi,
Ianiro and Presutti (1982) and Ferrari, Presutti and Vares (1987) prove the con-
servation of local equilibrium for nearest neighbor one dimensional symmetric
exclusion processes and zero range processes with jump rate g (k ) = 1fk 1g.
Kipnis, Marchioro and Presutti (1982) consider a one dimensional system of
harmonic oscillators in contact with reservoirs at different temperature. They obtain
the stationary measure, the temperature profile and prove the local convergence to
the Gibbs measure. De Masi, Ferrari, Ianiro and Presutti (1982) prove the same
statement in the context of symmetric simple exclusion processes in contact with
stochastic reservoirs at different temperature.
5. Comments and References 19
Presutti and Spohn (1983) derive the hydrodynamic behavior of the voter
model. Malyshev, Manita, Petrova and Scacciatelli (1995) extend the result to
weak perturbations of the voter model. They obtain equations of type
X
n
@t = m r C + dj j + e
0
j =1
X
d X
n
dj j + e :
0
and @t = Bi;j @u2 i ;uj C +

i;j =1 j =1
Greven (1984) derives the hydrodynamic limit of a branching interacting particle
system.
Conservation of local equilibrium for continuous spin systems or interacting
diffusions were also considered. Rost (1984) derives the hydrodynamic behavior of
Brownian hard spheres moving on R. Fritz (1987a,b), (1989) proves the hydrody-
namic limit of lattice Ginzburg–Landau processes by the method of resolvent. This
method reduces the proof of conservation of local equilibrium to the verification
of certain smoothness properties of the evolution as function of the initial condi-
tion. Funaki (1989a,b) extends the method to one dimensional Ginzburg–Landau
models of diffusion processes evolving on the line.
Though we shall not consider this problem here, there exist an extended liter-
ature on the hydrodynamic behavior of deterministic or mechanical systems. The
first results in this direction were obtained by Boldrighini, Dobrushin and Suhov
(1983), Dobrushin, Pellegrinotti, Suhov and Triolo (1986), (1988), Dobrushin, Pel-
legrinotti and Suhov (1990) and Fritz (1982), (1985). We refer to Spohn (1991)
for a clear and complete presentation of the subject. De Masi and Presutti (1991)
is another excellent reference on hydrodynamic behavior of interacting particle
systems. Some reviews have also been published or physical discussions on the
derivation of the equations of motion. We mention Lebowitz and Spohn (1983),
Presutti (1987), Boldrighini, De Masi, Pellegrinotti and Presutti (1987), Lebowitz,
Presutti and Spohn (1988), Presutti (1997), Jensen and Yau (1997).
Applications. The methods and the results of the theory of hydrodynamic limit
of interacting particle systems have been used to solve several different types of
problems. Here are some examples.
Ulam’s problem. For a positive integer N , denote by LN the length of the largest
increasing subsequence of a random permutation of f1; : : : ; N g. Consider a Pois-
p by RN the maximal number of points in an
son point process on R2 and denote
increasing path contained in [0; N )2 . Hammersley (1972) pointed out the connec-
tion between the distribution of RN and LN . Proving the hydrodynamic behavior
p
of a one dimensional continuous spin asymmetric process, Aldous and Diaconis
(1995) deduce that (1= N )E [LN ] converges to 2 (cf. also Seppäläinen (1996a)
for another proof through hydrodynamic limit). Seppäläinen (1997b) solves the
same problem on a planar lattice.
Stationary measures. Relying on the hydrodynamic behavior of a superposition of

a speeded up exclusion process with a Glauber dynamics, derived by De Masi,
Ferrari and Lebowitz (1986), Noble (1992) and Durrett and Neuhauser (1994)
prove the existence of a non trivial stationary measure for a class of non conser-
vative processes in the case where the stirring rate is large enough. Maes (1990)
investigates the decay of correlations for the stationary measures of anisotropic
lattice gases on Z2 through the behavior of the equilibrium fluctuations around the
deterministic macroscopic evolution.
Occupation time large deviations. Landim (1992) and Benois (1996) deduce the
occupation time large deviations for one dimensional symmetric exclusion pro-
cesses and for independent one dimensional symmetric random walks from the
large deviations from the hydrodynamic limit. Similar results were obtained previ-
ously by Cox and Griffeath (1984) for independent random walks and by Bramson,
Cox and Griffeath (1988) for the voter model.
Random domino tillings. Jockush, Propp and Shor (1995) deduce from the hydro-
dynamic behavior of the totally asymmetric simple exclusion process obtained by
Rost (1981), the asymptotic shape of a subregion in a random domino tillings of
a square.
2. Some Interacting Particle Systems
We introduce in this chapter the interacting particle systems we consider through-

out the book and present their main features. We shall refer constantly to Liggett
(1985) for some proofs and some extensions of the results presented here.
d
In section 1 we review the basic properties of the product topology of N Z . In
section 2 we introduce the simple exclusion processes, in section 3 the zero range
processes and in section 4 the generalized exclusion processes. In section 5 we
present the main tools in the investigation of attractive systems and in section 6
we characterize the set of invariant and translation invariant measures of attractive
d
zero range processes evolving on N Z .
1. Some remarks on the topology of N Z and

d
M (N1
Zd )
d
We briefly discuss in this section some aspects of the topology of N Z and of
the weak convergence of probability measures. Recall that we denote the config-
d
urations of N Z by the Greek letters , and so that, for each x in Zd, (x)
stands for the total number of particles at x for the configuration . We endow the
d
space N Z with the product topology which is metrizable : Define, for instance,
the distance d(; ) on N Z by
d
X 1 j (x) (x)j
d(; )
d 2 1 + j (x) (x)j
= jxj
x2Z
It is straightforward to check that d is a distance compatible with the product
d
topology. Moreover, with this distance N Z is a complete separable metric space.
d
Since N is not a compact set, N Z is not itself compact. Nevertheless, the
d
compact subsets of N Z are easy to describe. We leave to the reader to check
d
that a subset K of N Z is compact if and only if K is closed and there exists a
collection of positive numbers fnx ; x 2 Zdg such that (x) nx for all in K .
d d
Denote by C (N Z ) (resp. Cb (N Z )) the space of continuous (resp. bounded
d
continuous) functions on N Z and by C (resp. Cb ) the space of cylinder (resp.
bounded cylinder) functions, i.e., functions that depend on the configurations only
through a finite set of coordinates. To clarify ideas and illustrate that continuous
22 2. Some Interacting Particle Systems
functions can not be uniformly approximated by cylinder functions, we present an

example of a continuous function that is not uniformly continuous.
In dimension 1 and for k 1, denote by k the set f k; : : : ; k g and by Ak
the set of configurations with k particles on each site of k : Ak = f ; (x) =
k ; x 2 k g. Define f : NZ ! R by f () = 1 if belongs to [k0 Ak and f () = 0
otherwise. We leave to the reader to check that f is continuous but not uniformly
continuous. In particular, f can not be approximated by cylinder or uniformly
continuous functions in the uniform topology.
Denote by M1 (N Z ) the space of probability measures on N Z endowed with
d d
the weak topology defined as follows :

d
Definition 1.1 A sequence of probability measures k on N Z converges weakly
to a probability measure if Ek [f ] converges to E [f ] for every bounded
continuous function f .
To prove the weak convergence of probability measures, we need only to

investigate the limit behavior of expectations of cylinder functions :
d
Lemma 1.2 A sequence of probability measures k on N Z converges weakly to a
probability measure if and only if Ek [ ] converges to E [ ] for every bounded
cylinder function .
Proof. Consider a sequence of probability measures k such that Ek [ ] con-

verges to E [ ] for every bounded cylinder function and fix " > 0 and
a bounded continuous function f . It follows from the convergence of expecta-
tions of bounded cylinder functions that for every positive integer `, there exists
A` = A` (") > 0 such that
n X o
k ; (x) > A` 2"`
x2`
for every 1 k 1, where to keep notation simple we denoted by 1 . In
this formula ` stands for f `; : : :; `gd . In particular,
n X o
k ; (x) A` for all ` 1 1 "
x2`
P
for every 1 k 1. The set K" of configurations such that x2` (x) A`
for all ` 1 is compact. On a compact set, a continuous function is uniformly
continuous and may therefore be uniformly approximated by a cylinder function :
there exists a bounded cylinder function " such that k " k1 kf k1 and
sup j " () f ()j " :
2K"
In particular, for each fixed 1 k 1, jEk [f ] Ek [ " ]j is bounded above
by
Some remarks on the topology 23
Ek [f 1fK"cg]

Ek [ " 1fK"cg] + Ek jf " j1fK"g 2kf k1" + " :
+

The convergence of Ek [f ] to E [f ] follows therefore from the convergence of
expectations of bounded cylinder functions.
d
Lemma 1.3 A sequence of probability measures k on N Z converges weakly
to a probability measure if and only if for every finite subset of Zd and every
sequence fax ; x 2 g, k f; (x) = ax ; x 2 g converges to f; (x) = ax ; x 2
g.
Proof. By Lemma 1.2, we just need to show that Ek [ ] converges to E [ ]
for every bounded cylinder function . So fix a bounded cylinder function and
" > 0. Denote by the support of the cylinder function . There exists B > 0
such that n X o
; (x) > B " :
x2
Denote by the configurations of N . With this notation we may rewrite Ek [ ]
h i
as
X X
( )k f; (y ) = (y ); y 2 g + Ek ( )1f (y ) > B g :
P
; x B
( ) y2
By assumption, as k " 1, the first term converges to
X
P ( )f; (y) = (y); y 2 g ; (1:1)
; x B
( )
while the absolute value of the second is bounded above by

X h X i
k k1k f; (y ) > B g = k k1 1 k f; (y ) B g
y2 y2
that converges, as k " 1, to
h X i
k k1 1 f; (y) B g "k k1
y2
by definition of B . Since the absolute value of the difference between (1.1) and
E [ ] is also bounded by "k k1 , we proved that Ek [ ] converges to E [ ].

Introduce on N Z the natural partial order : if and only if (x) (x)
d
for every x in Zd. Denote by Cm (resp. Cm;b ) the space of monotone (resp.
bounded monotone) cylinder functions in the sense that f ( ) f ( ) for all .
d
The partial order extends to the space of probability measures over N Z in a natural
way :
Z Z
1 2 provided f d1 f d2 (1:2)
for all functions f in Cb;m .

d
Theorem 1.4 Let 1 and 2 be two probability measures on N Z . The following
two statements are equivalent :
(a) 1 2 .
(b) There exists a probability measure ¯ on N Z N Z such that the first (resp.
d d
second) marginal is equal to 1 (resp. 2 ) and ¯ is concentrated on configura-

tions “above the diagonal" :
¯ f(; ); g = 1:
The reader will have no difficulty in adapting the proof of Theorem II.2.4 in
Liggett (1985) to the present setting. One first prove the result on N S N S , where
S is a finite set, and then use Kolmogorov theorem and a sequence of finite set
Sk increasing to Zd to conclude the proof. ¯ is called the coupling measure of 1
and 2 .
d
Lemma 1.5 A sequence of probability measures k on N Z converges weakly to a
probability measure if and only if Ek [ ] converges to E [ ] for every bounded
monotone cylinder function .
Q
Proof. By Lemma 1.3, we just need to show the convergence of the expected
value of cylinder functions of type x2 1f (x) = ax g for finite sets . The result
follows therefore from the representation of 1f (x) = ax g as the difference of the
two bounded monotone cylinder functions 1f (x) ax g and 1f (x) ax + 1g.

Remark 1.6 In Chapter 1, 3 and 9 we face the problem of proving the conver-
gence of probability measures defined in different spaces. More precisely, for large
d
positive integers N , we consider probability measures N defined on N TN and
wish to prove that this sequence converges weakly to some probability measure
defined on N Zd. It is easy to build a mathematical framework to render d
the
N
argument rigorous. For each N , extend the measure to the space N in theZ
d
natural way : denote by ~N the periodic measure on N Z with period TdN defined
by the following two properties. We first require the projection of ~N on TdN to
be equal to N :
n o
~N ; (x) = ax ; x 2 f [(N 1)=2]; : : :; [N=2]gd
o
= N f; (x) = ax ; x 2 TdN
and then impose the measure to be periodic :

2. Simple exclusion processes (hard core interaction) 25
n o
~N ; (x) = (x + Ny); y 2 Zd = 1:
In the first formula, [r] stands for the integer part of r. We have now a sequence
d
of probability measures ~N defined on N Z and we may investigate the weak
N
convergence of ~ to . It corresponds to the convergence of E~N [ ] (which
is equal to EN [ ] provided N is large enough for f [N=2]; : : :; [N=2]gd to
contain the support of ) to E [ ], for every bounded cylinder function .
Remark 1.7 All analysis developed above extends in a straightforward man-

ner to the configuration space f0; : : : ; gZ and to probability measures de-
d
fined on this space. The situation is in fact even simpler because the space
f0; : : : ; gZd endowed with the product topology (and therefore the measure space
M1(f0; : : :; gZd) endowed with the weak topology) is compact.
2. Simple exclusion processes (hard core interaction)
Among the simplest and most widely studied interacting particle systems is the
exclusion process. In contrast with superpositions of random walks presented in
Chapter 1, the exclusion process allows at most one particle per site. The state
space is therefore f0; 1gTN .
d
To prevent the occurrence of more than one particle per site we introduce
an exclusion rule that suppresses each jump to an already occupied site. In fact,
we shall focus only on the simplest class of exclusion processes : systems where
particles jump, whenever the jump is allowed, independently of the others and
according to the same translation invariant elementary transition probability.
Definition 2.1 Let p be a finite range, translation invariant, irreducible transition

probability on Zd :
p(x; y) = p(0; y x) =: p(y x)

for all pair (x; y ) of d-dimensional integers and for some finite range probability
measure p() on Zd :
X
p( z ) = 1 and p(x) = 0 for jxj large enough :
z2Zd
We shall refer to p() as the elementary jump probability. The generator
X X
(LN f )( ) := (x)[1 (x + z )] pN (z ) [f (x;x+z) f ()] ;
x2TdN z2TdN
where x;y is the configuration obtained from letting a particle jump from x to
y:
8 (z ) z == x; y ;
< if
X
x;y (z ) = : ((xy)) + 11 if z = x; and pN (z ) := p(z + yN ) (2:1)
if z=y y2Zd
defines a Markov process called simple exclusion process with elementary jump
probability p(). In the particular case where p(z ) = p( z ) we say that it is a
symmetric simple exclusion process.
The interpretation is clear. Between 0 and dt each particle tries, independently

from the others, to jump from x to x + z with rate pN (z ). The jump is suppressed
if it leads to an already occupied site.
Note. Irreducibility of the transition probability p() means that the set fx ; p(x) >
0g generates Zd, i.e., that for any pair of sites x, y in Zd, there exists M 1
and a sequence x = x0 ; : : : ; xM = y such that p(xi ; xi+1 ) > 0 for 0 i M 1.
P
We usually renormalize the transition probability in a slightly different way (we
assume x p(x) = d instead of 1) in order to get the heat equation as hydrodynamic
equation for the symmetric simple exclusion process. Finally, since the transition
probability is assumed to be of finite range, there exists A0 in N such that p(z ) = 0
for all sites z outside the cube [ A0 ; A0 ]d . In particular, pN () and p() coincide
provided N A0 . For this reason, from now on we omit the superscript N in the
elementary jump probability.
We denote by fS N (t); t 0g the semigroup of the Markov process with
generator LN . We use the same notation for semigroups acting on continuous
functions or on the space M1 (f0; 1gTN ) of probability measures on f0; 1gTN .
d d
It might be worthwhile to justify the terminology. The rule that forbids jumps
to occupied sites explains the term exclusion. Notice, on the other hand, that the
rate at which a particle jumps from x to y depends on the configuration only
through the occupation variables (x) and (y ). This last dependence on (x) and
(y) reflects the exclusion rule. To distinguish this class from processes where the
jump rate depends in a more complicated way on the configuration, we call the
first family simple exclusion processes. Finally, notice that the total number of
particles is conserved by the dynamics.
For 0 1, we denote by = N the Bernoulli product measure of
parameter , that is, the product and translation invariant measure on f0; 1gTN with
d
density . In particular, under , the variables f (x); x 2 TdN g are independent

with marginals given by
f(x) = 1g = = 1 f(x) = 0g :
In Chapter 1, we have used the same notation to designate
the family
of invariant measures for superpositions of independent random walks. In fact,
throughout this book we will represent the equilibrium states of various processes
by the symbol . The context will clarify to which measure we are referring to.
Proposition 2.2 The Bernoulli measures f ; 0 1g are invariant for simple

exclusion processes. In addition, with respect to each , exclusion processes with
2. Simple exclusion processes (hard core interaction) 27
elementary jump probability p̌(z ) := p( z ) are adjoint to processes with elementary

jump probability p(z ). In particular symmetric simple exclusion processes are self-
adjoint with respect to each .
Proof. Notice that by a simple change of variables

Z Z
f (0;z ) g() (0)[1 (z )] (d) = f () g(z;0) (z )[1 (0)] (d) :
P
This identity, the fact that 1 = z2Zd p(z ) =
P
z2Zd p( z ) and a change in the
order of summation prove the proposition.
The family of invariant measures is parametrized by the density, for
E [(0)] = f(0) = 1g = :
Terminology 2.3 We shall say that an interacting particle system is symmet-
ric if the transition probability p() is symmetric (p(x) = p( x) for x in Zd),
P P
that it is mean-zero asymmetric if the transition probability is not symmetric but
x xp(x) = 0 and that it is asymmetric if x xp(x) 6= 0.
Remark 2.4 Since the total number of particles is conserved by the dynamics
the measures 0 1
X
N;K ( ) := @ (x) = K A
x2TdN
are invariant and it could have seemed more natural to consider them instead of the
Bernoulli product measures . Nevertheless, a simple computation on binomials
shows that for all finite subsets E of Zd, for all sequences f"x ; x 2 E g with
values in f0; 1g and for all 0 1,
n X o
N !1
lim (x) = "x ; x 2 E (y) = [0 N d ]
y2TdN

= (x) = "x ; x 2 E

0
uniformly in 0 . Therefore, the Bernoulli product measures are obtained as limits

of the invariant measures N;K , as the total number of sites increases to infinity.
For each 0 K N d , denote by N;K

1
the “hyperplanes" of all configura-
tions with K particles :
n X o
N;K
1
= 2 f0; 1gTdN ; ( x) = K :
x2TdN
Invariant measures of density preserving particle systems that are concentrated on
hyperplanes with a fixed total number of particles are called canonical measures
(the family fN;K ; 0 K N d g in the context of simple exclusion processes,

for instance). In contrast, the measures obtained as weak limits of the canonical
measures, as the number of sites increases to infinity, are called the grand canonical
measures (here the Bernoulli measures as we have just seen in Remark 2.4). We
return to this point in Appendix 2, where we investigate in a general context the
asymptotic behavior of canonical measures and deduce some estimates on the total
variation distance between the canonical measures and the grand canonical ones.
3. Zero range processes
This time we consider evolutions without restrictions on the total number of parti-
d
cles per site. The state space will therefore be N TN . The process is defined through
a function g : N ! R+ vanishing at 0, which represents the rate at which one par-
ticle leaves a site, and a translation invariant transition probability p(; ) on Zd. It
can be described as follows. If there are k particles at a site x, independently of
the number of particles on other sites, at rate g (k )p(x; y ) one of the particles at x
jumps to y . In this way particles only interact with particles sitting on the same
site. For this reason these processes are called zero range processes.
Definition 3.1 Let g : N ! R+ be a function vanishing at 0 and p(; ) be a finite

range, irreducible, translation invariant transition probability. We assume that g is
strictly positive on the set of positive integers and that it has bounded variation in
the following sense :
g := sup jg (k + 1) g (k )j < 1 : (3:1)
k 0
Denote by ' the radius of convergence of the partition function Z : R+ ! R+
defined by
X 'k
Z ( ')
k g (k )!
=
In this last formula g (k )! stands for

Q g(j ) and by convention g(0)! = 1.
0
jk
Z is analytic and strictly increasing on [0; ' ). Assume that Z ( )
1
Notice that
increases to 1 as ' converges to ' :
lim
'"'
Z (') = 1: (3:2)
The generator
X X
LN f () = pN (z ) g((x)) [f (x;x+z) f ()]
x2TdN z2TdN
d
defines a Markov process on N TN called zero range process with parameters
(g; p). Here, as in the previous section, x;y represents the configuration where
3. Zero range processes 29
one particle jumped from x to y and pN () represents the transition probability
translated to the origin and restricted to the torus :
X
pN (z ) := pN (0; z ) = p(0; z + yN )
y2Zd
for every d-dimensional integer z .
The assumption (3.2) is not necessary to define the process but will always be
required to prove the hydrodynamic behavior of zero range processes. We therefore
preferred to include it in the definition.
In these processes each particle jumps, independently of particles sitting at
other sites, from x to y at a rate pN (y x)g ( (x))( (x)) 1. In particular, if g (k ) = k
for every k 0, we obtain the superposition of independent random walks studied
in Chapter 1. On the other hand, the case g (k ) = 1fk 1g models a system of
queues with mean-one exponential random times of service.
We now describe some invariant measures of the process. For each 0 ' <
N denote the product measure on N TdN with marginals given by
' , let ¯';g = ¯';g
1 'k
¯';g f; (x) = kg =
Z (') g(k)! (3:3)
for each k 0 and x in TdN .

Proposition 3.2 For each 0 ' < ' the product measure ¯';g is invariant
for the zero range process with parameters (g; p). Moreover, the adjoint process
with respect to any of the measures ¯';g is the zero range process with parameters
(g; p̌). In particular, if p is symmetric the process is self-adjoint.
Proof. The proof relies on the same computations that were made for the simple
exclusion process and on the following identity
k j k 'j+1
g(k) g'(k)! g'(j )! g(j + 1) g('k
1
1)! g (j + 1)!
=
Since the function g () will always be fixed, to keep notation as simple as
possible, we hide the dependence on g of the measure ¯';g and denote it simply
by ¯' . Let R(') denote the expected value of the occupation variable under ¯' :
X 'k
R(') = E¯' (0) =
1
Z (') k g(k)! (3:4)
k 0
From the last equality we obtain a relation that will be often used in the sequel :
R(') Z 0 (')' '@' log Z (') : (3:5)

=
Z ( ') =
Computing the first derivative of R() we show that it is strictly increasing. In

fact, it is easy to show that log Z (e ) is strictly convex in .
Since we want to parametrize the invariant measures by the conserved quantity,
which is here the density of particles, we change variables in the definition of the
invariant measures ¯' as follows. For 0, define the product measure by
() = ¯() () : (3:6)
In this formula () stands for the inverse function of R() defined in (3.4). In
the next lemma we show that assumption (3.2) guarantees that the range of the
function R() is all R+ . We obtained in this way a family f ; 0g of invariant
measures parametrized by the density since the expected value of the occupation
variables (x) under is equal to :
E [(x)] = (3:7)
for every 0. Moreover, a simple computation shows that the function () is

the expected value of the jump rate g ( (0)) under the measure :
() = E [g((0))] : (3:8)
Lemma 3.3 Recall that we denoted by ' the radius of convergence of the partition
function Z .
lim R(') = 1 :
'"'
Furthermore, for each 0 ' < ' the measure ¯' has a finite exponential mo-
ment : there exists (') > 0 such that

E¯' e(0) < 1 : (3:9)
Proof. We consider separately two different cases. Assume first that Z is defined
for all positive reals or, equivalently, that the radius of convergence ' is infinite.
Suppose, by contradiction, that the function R is bounded by some constant C0 .
From identity (3.5) we obtain that
@' log Z (') C0 ' 1 :

Hence, integrating over ' we get that for every ' > 1,
Z (') Z (1)'C : 0
But this is in contradiction with the fact that Z (') 'k [g (k )!] 1
for every integer
k by the definition of Z .
Assume now that the radius of convergence is finite. Fix some positive '0 <
' . Since Z () is a smooth increasing function, for ' '0 ,
Z'
log Z (') log Z ('0 ) +
1
'0 ' @ log Z ( ) d :
0
Since, on the other hand, by relation (3.5)
R(') = '@' log Z (') ;

we obtain that
Z'
'0 log ZZ(('')) R( ) d :
0 ' 0
Since the left hand side of this inequality, by assumption (3.2), increases to 1 as
' " ' , it follows that Z'
lim
'"' '0
R ( )d = 1:
Since the function R is increasing the first statement of the lemma is proved.
Notice that E¯ ' [expf (0)g] is equal to Z ('e )=Z ('). Thus (3.9) follows from
assumption (3.2).
Before proceeding, we present an example of zero range dynamics that does

not possess an invariant product measure for each density 0. In virtue of
Lemma 3.3, the partition function Z () can not satisfy assumption (3.2).
Example 3.4 Consider a one-dimensional, nearest neighbor, symmetric zero range

process with jump rate g (k ) = (1 + k 1)3 for k 1. Then, ' = 1 and the partition
function is
X 'k
Z ( ')
k1 (k + 1)
= 1 + 3
so that
X
lim
'!1
Z (') = 1 +
1
(k + 1)3
< 1:
k 1
R LConsider a product invariant measure . Since is invariant, we have that
N(x)d = 0 for every x. Denote by ' the expectation of g((x)) under :
x
'x = E [g((x))]. Since LN (x) = (1=2)fg((x + 1)) + g((x 1)) 2g((x))g, the
previous identity gives that (N ')x = 0 if N stands for the discrete Laplacian.
This identity forces 'x to be constant, equal, say, to '
R.
On the other hand, for every x in TdN and a > 0, LN 1f (x) = ag (d ) = 0.
Since
LN 1f(x) = ag = g(a)1f(x) = ag

+ (1=2)1f (x) = a 1g g ( (x + 1)) + g ( (x 1)) ;
since the measure is assumed to be product and since E [g ( (x))] = ' is
constant, we have that
g(a) f; (x) = ag = ' f; (x) = a 1g :
In particular, an invariant product measure must be of the form (3.3). In this

example, since g (k ) = (1 + k 1 )3 ,
X k
R(') = k (k '+ 1)3
k 1
so that
X k
R(') R < 1 :
k1 (k + 1)
lim = =
'!1 3
Thus for > R there is no invariant product measure with density .

We now present a few results concerning the family of invariant measures
that will be needed later. These technical lemmas might be skipped in a first
reading. We start by showing that the family of invariant measures f ; 0g
is an increasing continuous sequence of measures for the order defined in section
1 of this chapter. Continuous family means that k converges to for every
0 and every k that approaches as k " 1 : lim0 ! E0 [ ] = E [ ] for
every bounded continuous function .
Lemma 3.5 Suppose that 1 2 . Then . 1 2
Proof. Since these measures are product measures, it is enough to prove that the
marginals are ordered. Define therefore, for each 0 ' < ' , a measure m' on
N by
k
m'(k) = Z (1') g'(k)!
We need to show that the family fm' ; 0 ' < ' g is an increasing set of
measures. In order to do it, we have to prove that for every A 1, the function
FA : [0; ' ) ! [0; 1] defined by
F A ( ') :=

m' k; k A

is increasing. A simple computation shows that the derivative of FA is equal to
8 9
< X 'k X 'k =
1
'Z (') :kA k g(k)! R(') g(k)! ;
kA
We denote by RA (') the expression inside parentheses. To conclude the proof of
the lemma it is enough to show that RA (') is positive. We prove it by induction
on A. Fix 0 < ' < ' . Since R1 (') is equal to R(') it is positive. On the other
'A R(') A :
hand,
RA+1 (') RA (') =
g(A)!
Therefore, for each ', R (') is increasing in the set f1; : : : ; [R(')] + 1g and
decreasing in the complementary set. In particular,
n o
RA (') min R1 ('); Alim R
!1 A (')
:
Since
lim R (') = 0 ;
A!1 A
we proved that RA (') is nonnegative for every A and '.
Corollary 3.6 Recall the definition of g given in (3.1). The function : R+ !
[0; ' ) appearing in (3.8) is uniformly Lipschitz on R+ with constant g :

( ) ( ) = E [g((0))] E [g((0))] gj j
2 1 2 1
2 1
for all non negative reals 1 and 2 .
Proof. Fix 1 2 . Recall from (3.8) that () is the expected value of g ( (0))
under the measure :
() = E [g((0))] :
Denote by 1 ;2 a measure on N TN N TN whose first marginal is equal to 1 ,
d d
whose second marginal is equal to 2 and which is concentrated on configurations

(1 ; 2 ) such that 1 2 . The existence of this measure is guaranteed by Theorem
1.4 because 1 2 . Hence, we have
h i
( ) ( ) E ; g( (0)) g( (0))
h i
2 1 2 1
gE ; (0) (0) :

1 2
1 2
2 1
Since the measure 1 ;2 is concentrated on configurations 1 2 we can remove

the absolute value in the last expression and obtain that it is equal to g (2 1 ).

The cylinder function g ( (0)) does not play any particular role in Corollary
3.6. The statement applies to a broad class of cylinder functions that we now
introduce. A cylinder function is called Lipschitz if there exists a constant C0
and finite subset of Zd such that
X
() () C (x) (x) 0 (3:10)
x2
for every configurations and . If is Lipschitz, then, taking in the last formula
to be the configuration with no particles, we obtain that there are finite constants
C0 and C1 and a finite subset of Zd such that
X
( ) C 1 + C0 (x) (3:11)
x 2
for every configuration .
A cylinder function is said to have sublinear growth if there exists a finite
subset of Zd such that for each > 0 there exists a finite constant C ( ) such
that
() C () + X (x)
x2
for every configuration . Every bounded cylinder function has a sublinear growth.
For a cylinder function satisfying (3.11) we denote by ~ : R+ ! R the
function whose value at some density 0 is equal to the expectation of with
respect to the invariant measure :

~() : = E () : (3:12)
With the notation just introduced, we have the identity g~ = .

The proof of Corollary 3.6 shows that ~ is a uniform Lipschitz function pro-
vided is a Lipschitz cylinder function :
Corollary 3.7 Let be a Lipschitz cylinder function. The function ~: R+ ! R

defined by (3.12) is uniformly Lipschitz on R+ : there exists a finite constant C0 =
C0 ( ) such that

~(2 ) ~(1 ) C0 j2 1 j
for all 1 , 2 in R+ .
Lemma 3.8 f ; 0g is a continuous family : ~() = E [ ] is a continuous

bounded real function for every bounded cylinder function .
Proof. Fix 0 in R+ and a sequence fk ; k 1g converging to 0 . By Lemma

1.3, we just need to show that for every finite subset of Zd and every sequence
fax; x 2 Ag, k f; (x) = ax ; x 2 g converges to 0 f; (x) = ax ; x 2 g.
This is obvious because the measures are product and is Lipschitz.
d
We conclude this section showing that the set of probability measures on N TN
bounded by some product measure is weakly relatively compact.
Lemma 3.9 For every 0, the set of probability measures A = f; g

is weakly relatively compact.
Proof. By Prohorov theorem, we just need to show that the collection of prob-
ability measures f; g is tight. Fix a positive " > 0. For each positive
integer `, there exists A` such that
n X o n X o
sup ; (x) > A` ; (x) > A` 2"`
2A x2` x2`
Therefore,
n X o
inf
2A
; (x) A` for every ` 1 " ;
x2`
4. Generalized exclusion processes 35
P
what concludes the proof of the lemma because \`1 f; x2` (x) A` g is a
compact set (cf. characterization of compact sets given at the beginning of section
1).
4. Generalized exclusion processes
In this section we consider a third type of interacting particle systems that will ap-
pear in the next chapters : a mixture of zero range and simple exclusion processes.
We admit this time no more than particles per site for some positive integer .
Like in the simple exclusion process the jump rate of a particle from x to y depends
exclusively on the occupancy at x and y and on a translation invariant transition
probability. More precisely, we are given a function r : f0; : : :; g2 ! R+ so
that r(a; b) represents the rate at which a particle jumps from a site occupied by
a particles to a site occupied by b particles. Since no particles may jump from a
site x if there are no particles at x, r(0; ) 0. Our exclusion rule also imposes
that r(; ) 0 to avoid sites with more than particles. In conclusion, a particle
jumps from site x to site y at rate r( (x); (y ))p(x; y ) independently of the number
of particles at other sites.
Definition 4.1 Let p be a finite range, translation invariant, irreducible transition

probability on Zd. The generator
X
(LN f )( ) = pN (y)r((x); (x + y))[f (x;x+y) f ()] ;
x;y2TdN
where x;x+y is the configuration obtained from by letting a particle jump from
x to x + y, defines a Markov process on f0; : : : ; gTdN called the generalized
exclusion process with elementary jump probability p(). In the particular case
where p(z ) = p( z ) we say that it is a symmetric generalized exclusion process.
As the reader will notice in next sections all proofs presented in this book of
hydrodynamic behavior of interacting particle systems rely on the explicit knowl-
edge of the invariant measures. Moreover, we shall only consider processes with
product invariant measures. In order to guarantee the existence of product invariant
measures for generalized exclusion processes we need to impose some restrictions
on the jump rate r(; ). We shall restrict our analysis to the special case where
particles jump whenever a jump is allowed and the transition probability is sym-
metric : r(a; b) = 1fa > 0; b < g, p(x) = p( x). In this case the generator is
given by
X
(LN f )( ) = p(y)1f(x) > 0; (x + y) < g[f (x;x+y) f ()] : (4:1)
x;y2TdN
As in the previous section, for ' 0, consider the product measure ¯' = ¯'N
on f0; : : : ; gTN with marginals given by
d
¯' f; (x) = kg 'k

=
Z (')
for 0 k , where Z (') = P0j 'j is the normalizing constant.
Proposition 4.2 Assume the elementary jump probability p() to be symmetric.
Then the Markov process with generator given in (4.1) is self–adjoint with respect
to each product measure ¯' .
The proof is identical to the proof of Propositions 2.2 or 3.2 and is therefore
omitted. To parametrize the invariant measures by the density observe that the
density of particles R(') for the measure ¯' , which is equal to
X

R(') = E¯' [(0)] =
1
Z ( ') k'k ; (4:2)
k=0
is strictly increasing. Denote by : [0; 1] ! [0; ] the inverse function of R()
and, for in [0; ], define by
:= ¯() :
We have now a one-parameter family of invariant measures indexed by the density
since, by definition, E [ (0)] = .
5. Attractive systems
In sake of completeness, in this section we briefly define and state the main prop-
erties of attractive interacting particle systems. To fix ideas, we shall consider zero
range processes but every definition or statement can be translated to generalized
exclusion processes. On the other hand, since we have nothing to add to section
II.2 of Liggett (1985), we refer the reader to this book for a complete discussion on
attractiveness and the corresponding coupling features of some particle systems.
Recall the partial order on M1 (N Z ) introduced in (1.2). A similar order can be
d
d
defined on M1 (N TN ).
Definition 5.1 An interacting particle system is said to be attractive if its semi-

group preserves the partial order :
1 2 ) S N (t)1 S N (t)2
for all t 0.
5. Attractive systems 37
Theorem 5.2 below is the main result of this section. It gives a sufficient
(and in fact necessary) condition for zero range processes to be attractive. It also
shows how closely related are the partial order and the coupling of two copies of
the process. This coupling technique is the main tool to prove the hydrodynamic
behavior of asymmetric interacting particle systems beyond the appearance of the
first shock.
Theorem 5.2 A zero range process with parameters (g; p) is attractive provided
the jump rate g () is a non decreasing function.
d
Proof. Fix two probability measures 1 and 2 on N TN such that 1 2 . The
idea is to construct two copies of zero range processes on the same probability
space whose evolution preserves the order. Denote therefore by ¯ the coupling
measure given by Theorem 1.4 and consider the Markov process (t ; t ) on N TN
d
d
N TN starting from ¯ and with generator L̄N given by

L̄N f (; )
X
=
n o
= p(y) min g((x)); g( (x)) f (x;x+y ; x;x+y ) f (; )
x;y2TdN
X n o (5:1)
p(y) g((x)) g( (x)) f (x;x+y ; ) f (; )
+
+
x;y2TdN
X n o
p(y) g( (x)) g((x)) f (; x;x+y ) f (; ) :
+
+
x;y2TdN
Notice that for the dynamics generated by L̄N , the -particles and the -
particles try to jump as much as possible together. On the other hand, both coor-
dinates evolve as zero range processes with parameters (g; p). Indeed, consider a
function f : N TN N TN depending, say, on the first coordinate only. It is simple
d d
to check that L̄N f is also a function that depends only on the first coordinate.
Moreover, (L̄N f )(; ) = (LN f0 )( ) if f0 denotes the restriction of f to the first
coordinate and LN the generator of a zero range process with parameters (g; p).
For the coupling measure ¯ on N TN N TN , denote by P̄¯ = P̄N;¯ the measure
d d
on the path space D([0; 1); NTN N TN ) corresponding to the Markov process of
d d
generator L̄N starting from ¯ . Expectation with respect to P̄¯ is denoted by Ē ¯ .

We prove now the main feature of the coupled process defined above : starting
from two configurations the dynamics keeps them ordered at later times.
Consider two configurations and with and denote by (;) the Dirac
probability on N TN N TN concentrated on (; ). We claim that
d d
h i
P̄(;) t t for all t 0 = 1:
Since (t ; t ) is right continuous (cf. section 2 of Appendix 1) and F0 =

f(; ); g is closed for the product topology, to prove the last statement it
is enough to show that P̄ ; [t t ] = 1 for all t 0.
( )
Notice that L̄N does not admit a jump out of F0 because the jump rate g ()
is non decreasing. In other words, F0 is an absorbing set : L̄N 1fF0 g 0. This
inequality can be checked by a direct computation on the generator L̄N . It follows
that h i h i
@t Ē (;) 1fF0g(t ; t ) = Ē (;) L̄N 1fF0 g(t; t ) 0 :
Therefore,
h i h i
1 Ē (;) 1 fF0g(t ; t ) Ē (;) 1 fF0g(0; 0) = 1
because . This proves the claim.

In conclusion, we proved that the Markov process with generator L̄N preserves
the partial order in the following sense : t t for all t 0 a.s. as soon as
0 0 . To conclude the proof of the theorem it remains to show that this property
implies that the semigroup preserves the order. Denote by fS̄ N (t); t 0g the
semigroup of the Markov process with generator L̄N .
We showed above that F0 is an absorbing set. Since the coupled measure ¯ is
concentrated on F0 (¯ fF0 g = 1), the measure S̄ N (t)¯ shares the same property.
On the other hand, since both coordinates evolve as zero range processes with
parameters (g; p), the first (resp. second) marginal of S̄ N (t)¯ is equal to S N (t)1
(resp. S N (t)2 ) if S N (t) represents the semigroup of a zero range process with
parameters (g; p). In particular, S̄ N (t)¯ is a coupling measure for 1 S N (t) and
2 S N (t). Therefore, by Theorem 1.4, 1 S N (t) 2 S N (t).
Remark 5.3 Theorem 1.4 extends to the compact space f0; : : :; gTN and the
d
proof of Theorem 5.2 applies to generalized exclusion processes provided the jump
rate r(; ) is non decreasing in the first variable and non increasing in the second.
In particular, the simple exclusion process is attractive.
6. Zero range processes in infinite volume
In Chapter 9, while proving the conservation of local equilibrium for attractive

zero range processes, we shall need some results on the invariant measures for
d
the process on N Z . We briefly summarize these results in this section and refer
to Liggett (1973), (1985) or Andjel (1982) for proofs and further details.
Fix a jump rate g () satisfying assumption (3.1) and a finite range, translation
invariant, irreducible transition probability p(; ) :
p(x; y) = p(0; y x) =: p(y x) 0
for all pair (x; y ) of d-dimensional integers,

X
p( z ) = 1 and p(z ) = 0 for all z A0 :
z2Zd
6. Zero range processes in infinite volume 39
d
We desire to prove the existence of a Markov process on N Z whose generator
acts on cylinder functions as
XX
Lf () = p(z ) g((x)) [f (x;x+z) f ()] : (6:1)
x2Zd z2Zd
The process is well defined for any initial configuration with a finite number of
particles since in this case it is a Markov process on a countable state space
with a bounded transition rate. Thus, for a configuration with a finite number of
d
particles, denote by P the probability on the path space D(R+ ; N Z ) corresponding
to the Markov process with generator (6.1) starting from and by E expectation
with respect to P .
To extend the definition to configurations with infinitely many particles, we
need to impose some conditions to avoid interference coming from infinity. Con-
sider a function l: Zd ! R+ for which there exists a finite constant M such that
X
l(y) Ml(x)
y; jy xjA0
for every x in Zd. l(x) = e jxj is an example of function satisfying the above
d
growth restriction. For such fixed function l, on the configuration space N Z ,
define the norm k kl by
X
k kl = j(x) (x)jl(x) :
x2Zd
Denote by El the space of configurations with finite norm : El = f; k kl < 1g
and by Ll the space of Lipschitz functions for this norm, i.e., the space of functions
f : El ! R for which there exists cf such that jf () f ( )j cf k kl for every
, in El . Denote by `f the smallest constant satisfying this inequality. Andjel
(1982) proved the following existence result :
Theorem 6.1 There exists a semigroup fS (t); t 0g defined on Ll such that

S (t)f () = E [f (t)]
for every f in Ll and finite . Moreover, Ll is closed under S (t) :

S (t)f () S (t)f () `f eg M t k kl
( +2)
and
(Lf )( ) = lim
S (t)f () f ()
t!0 t
for all f in Ll and in El .
The existence of a probability P on D(R+ ; N Z ), for every in El , corre-

d
sponding to a Markov process whose generator acts on functions in Ll as (6.1)

follows from this theorem and general Markov process arguments that can be
found in Chapter IV of Ethier and Kurtz (1986) or in Chapter III of Revuz and
Yor (1991), for instance.
From this point on we assume the process to be attractive, i.e., the jump rate
to be non decreasing :
0 = g(0) < g(k) g(k + 1)

for all k 1. Note that in this case coupling two copies of the process permits
d
to prove the existence of a Markov process on N Z with generator given by (6.1)
and starting from any initial state in El . One just needs to consider a sequence
of states k , k 1, with finite total number of particles and increasing to .
Denote by I the set of invariant probability measures for the zero range
process with generator given by (6.1) and by S the space of translation invariant
probability measures. The next result states that all invariant and translation invari-
ant measures are convex combinations of the product measures f ; 0g. Its
proof can be found in Theorem VIII.3.9 of Liggett (1985) for exclusion processes
and in Theorem 1.9 of Andjel (1982) for zero range processes.
nZ
Theorem 6.2
o
I \S = m(d); where m is a probability measure on R+ :
In particular, if (I \ S )e stands for the extremal points of the set I \ S,
(I \ S )e = f ; 0g :
We conclude this section with a characterization of translation invariant sta-

tionary measures.
Theorem 6.3 A translation invariant probability measure bounded above by

some product measure 0 is invariant if and only if for each density 0 there
exists a measure ¯ on the product space N Z N Z with first marginal equal to
d d

, second marginal equal to and concentrated on ordered configurations :
(i) the first marginal of ¯ is ,
(ii) the second marginal of ¯ is and
(iii) ¯ f(; ); or g = 1.
Proof. One direction follows straightforwardly from last theorem and Theorem
1.4. To prove the oder direction, first observe that the proof of Lemma VIII.3.6
in Liggett (1985) applies to the zero range processes considered here because we
assumed to be bounded above by some measure 0 .The result now follows
from the proof of Theorem VIII.3.9 of Liggett (1985).
3. Weak Formulations of Local Equilibrium
In Chapter 1 we presented a mathematical model to describe the evolution of

a gas. In this respect we introduced the notions of local equilibrium and con-
servation of local equilibrium. Unfortunately, at the present state of knowledge,
there is no satisfactory general proof of conservation of local equilibrium for large
classes of interacting particle systems. All existing approaches to this problem rely
too much on particular features of the model, such as the existence of dual pro-
cesses (for symmetric simple exclusion processes by De Masi, Ianiro, Pellegrinotti
and Presutti (1984), superposed with a Glauber dynamics by De Masi, Ferrari
and Lebowitz (1986) or for the voter model Presutti and Spohn (1983)); explicit
computations (for superpositions of independent random walks Dobrushin and
Siegmund–Schultze (1982)); resolvent equation methods (for Ginzburg–Landau
processes Fritz (1987a,b) and (1989a)); or on the attractiveness (Rost (1981),
Andjel and Kipnis (1984) Benassi and Fouque (1987, 1988), Andjel and Vares
(1987)). Furthermore, we would like to prove the hydrodynamic behavior of sys-
tems starting from initial states more general than local equilibrium measures. This
leads us to weaken the concept of local equilibrium.
The purpose of this chapter is to introduce two alternative versions of local
equilibrium, to briefly study some relations among these versions and to present
an idea of how one may deduce conservation of local equilibrium from a weaker
version.
Throughout this chapter we use the notation introduced in Chapter 1. Thus, for
instance, for 0, = N denotes the translation invariant product measure
d
on N TN whose marginals are Poisson measures with parameter . Of course,
all concepts extend naturally to generalized exclusion processes and zero range
processes.
In Chapter 1 we presented a quite restrictive definition of local equilibrium
associated to a density profile. We required the marginal probabilities of the state
of the system at a macroscopic time t, that is, of S N (tN )N , to converge to one of
the extremal invariant and translation invariant measures of the infinite system at
each continuity point of the profile. Even if this definition is “physically" natural,
it is difficult to prove it. With this in mind, we introduce two weaker notions
of local equilibrium. We first recall the notion of product measures with slowly
varying parameter associated to a profile.
42 3. Weak Formulations of Local Equilibrium
Definition 0.1 (Product measures with slowly varying parameter associated

to a profile). Given a continuous profile 0 : Td ! R+ , we denote by N0 () the
d
product measure on N TN with marginals given by
N() f; (x) = kg

0
= N(x=N ) f; (0) = kg
0
for all x in TdN , k 0. This measure is called the product measure with slowly
varying parameter associated to 0 ().
d
Measures on N Z are characterized by the way they integrate bounded cylinder
functions. In order to keep notation simple, to each bounded cylinder function
in N Z we associate the real bounded function ~ : R+ ! R that at is equal to
d
the expected value of under :

Z
~() : = E [ ] = () (d) : (0:1)
By Lemma 2.3.8 ~ is a bounded continuous function.

In the same way that we have defined in Chapter 1 translations of configura-
tions, for all continuous functions , we denote by x the translation of by x
units :
(x )( ) = (x )
for every configuration .
By Chebychev inequality and the dominated convergence theorem, for every
smooth profile 0 : Td ! R+ and every sequence N0 () of Poisson measures with
slowly varying parameter associated to this profile,
lim
N !12 3
X Z
N 4 N1d G(x=N ) (x )() G(u) ~( (u)) du > 5
x2TdN
0( ) 0 = 0
Td
for all continuous functions G: Td ! R, all bounded cylinder functions and all
strictly positive . We leave to the reader to check this statement. Details can be
found in the proof of a slightly stronger result stated as Proposition 0.4 below.
This last statement asserts that the sequence of measures N0 () integrates the
cylinder function around the macroscopic point u in Td in the same way as
an equilibrium measure of density 0 (u) would do. We do not require anymore
the marginals of the sequence of measures to converge to an extremal equilibrium
measure at every continuity point of the profile 0 . We only impose that its spatial
mean converges to the corresponding spatial mean. This notion is therefore weaker
than the one of local equilibrium introduced in Chapter 1. The main difference is
the spatial average over small macroscopic boxes of volume of order N d that is
implicit in this new concept and absent in the definition of local equilibrium.
3. Weak Formulations of Local Equilibrium 43
Definition 0.2 . (Weak local equilibrium) A sequence (N )N 1 of probability

d
measures on N TN is a weak local equilibrium of profile 0 : Td ! R+ if for every
continuous function G: Td ! R, every bounded cylinder function and every
> 0 we have
2 Z 3
1 X G(x=N ) ()
lim N 4 d G

(u) ~( (u)) du > 5 = 0:
x
N !1 N x2TdN Td
0

The function ( ) = (0) plays a special role in the whole study of hydro-
dynamics being closely related to the conserved quantity. Though it is generally
not bounded (for zero range processes, for instance), the definition that follows is
a particularly weakened notion since it demands only convergence for this cylin-
der function. Since it will appear repeatedly throughout the book, we introduce a
special terminology :
Definition 0.3 (Probability measures associated to a density profile) A se-

d
quence (N )N 1 of probability measures on N TN is associated to a profile
0 : Td ! R+ if for every continuous function G: Td ! R, and for every > 0
2 3
we have
1 X G(x=N )(x) Z
lim N 4 d G (u) (u) du > 5 = 0 :
N !1 N x2TdN Td
0

The quantity just introduced in the definition above can be reformulated in

terms of empirical measures. Let N be the positive measure on the torus Td
obtained by assigning to each particle a mass N d :
X
N (; du) : = N1d (x) x=N (du): (0:2)
x2TNd
In this formula, for a d-dimensional vector u, u represents the Dirac measure

concentrated on u. The measure N (; du) is called the empirical measure asso-
ciated to the configuration . The dependence in will frequently
P be omitted to
keep notation as simple as possible. With this notation N d x G(x=N ) (x) is the
integral of G with respect to the empirical measure N , denoted by < N ; G >.
Let M+ (Td ) be the space of finite positive measures on the torus Td endowed
with the weak topology. Notice that for each N , N is a continuous function
from N TN to M+ . For a probability measure on N TN , throughout this book,
d d
we shall also denote by the measure ( N ) 1 on M+ induced by and N

(( N ) 1 [A] = [ N 2 A]). With this convention, a sequence of probability
d
measures (N )N 1 in N TN is associated to a density profile 0 if the sequence of
random measures N (du) converges in probability to the deterministic measure

0 (u)du.
In the study of the hydrodynamic behavior of interacting particle systems we
will sometimes be forced to reduce our goals and to content ourselves in proving
that starting from a sequence of measures associated to a density profile 0 then, at
a later suitably renormalized time, we obtain a state (S N (tN )N in the notation
of Chapter 1) associated to a new density profile (t; ) which is the solution of
some partial differential equation. More precisely, we want to prove that for all
sequences of probability measures N associated to a profile 0 : Td ! R+ and
not too far, in a sense to be specified later, from a global equilibrium N ,
2 Z 3
1 X G(x=N ) (x)
lim N 4 d G

(u) (t; u) du > 5 = 0
tN
N !1 N x2TdN Td
for a suitable renormalization N , for every continuous function G: Td ! R and
every strictly positive. As in the case of local equilibrium, in this formula (; )
will be the solution of a Cauchy problem with initial condition 0 (). This consti-
tutes the program of the next chapters and deserves therefore a special terminology.
We shall say that we proved the hydrodynamic behavior of an interacting particle
system whenever we are able to achieve the goal just described.
We will now compare the strong notion of local equilibrium introduced in
Chapter 1 with the weaker versions just presented. We start showing that it is
really a weaker notion. In this direction we first present two results that are not
stated in the greatest possible generality.
Proposition 0.4 Let (N )N 1 be a local equilibrium of profile 0 almost surely

continuous. Then the sequence (N ) is a weak local equilibrium of profile 0 .
Proof. Let be a bounded cylinder function. For a positive integer `, we introduce

the real positive function hN;` defined on the torus Td by
X
hN;`(u) = 1fx Nu < x + 1g
x2TdN
2 3
X
EN 4 (2` + 1)d
1
y () ~( (x=N ))5 :
jy xj`
0

Here 1 represents the vector of Rd with all coordinates equal to 1 and “" is the
natural partial order of Rd : a b provided ai bi for 1 i d.
hN;` is bounded everywhere because is bounded. Moreover, when N in-
creases to infinity, hN;` (u) converges to
2 3
X
E u 4 (2` +1 1)d y () ~( (u)) 5 :

0( )
jyj`
0

3. Weak Formulations of Local Equilibrium 45
at all continuity points u of 0 . By the law of large numbers, when ` tends to

infinity, this expression converges to 0. On the other hand, the integral of hN;`
with respect to the Lebesgue measure on the torus is
2 3
X 1 X
EN 4 ( ) ~ (x=N ) 5 :
1
d d
N x2Td (2` + 1) jy xj` y
0
N
The dominated convergence theorem permits to conclude the proof of the propo-
sition.
In the same way a weak assumption of equiintegrability suffices to prove an

analogous result for probability measures associated to density profiles.
Proposition 0.5 Let (N )N 1 be a local equilibrium of profile almost surely con-
tinuous and bounded. Then, the sequence (N ) is associated to the density profile
0 if X h i
M !1 N !1 N
lim lim
1
d EN (x)1f(x) > M g = 0 :
x2TdN
The weak notion of local equilibrium is not however too far from the strong
form as it is shown by the following result.
Proposition 0.6 Let (N )N 1 be a weak local equilibrium of profile 0 . Assume

that there exists a sequence (k )k1 decreasing to 0 such that for each integer
k there exists two sequences of measures (Nk ;+ )N 1 and (Nk ; )N 1 having the
following properties:
(a) (Nk ; )N 1 is a weak local equilibrium of profile k ; ;
(b) x Nk ;+ x k ; for every d-dimensional integer x of absolute value
N N
bounded by k N ;
(c) For every continuity points u of 0 ,
lim sup sup jk ; (v) 0(u)j = 0:

k!1 jv ujk
Then the sequence (N )N 1 is a strong local equilibrium of profile 0 .
Proof. Let be a bounded monotone cylinder function. From hypotheses (a), (b)
and (c) the expectation of under the measure N translated by [uN ] can be
bounded below and above by expressions that converge to ~(0 (u)). Indeed, as
is monotone, by property (b), for every integer k
X
E uN N [ ]
[ ]
1
(2[k N ] + 1)d
Ey Nk ; [ ] :
jy [uN ]j[k N ]
+
By hypothesis (a), as N increases to infinity, the right hand side converges to

Z
1
(2k )d [u k ;u+k ]d

~ k ;+(v) dv :
Notice that in the definition of weak local equilibrium, we assume convergence
only for continuous functions G. It is not difficult, however, to prove conver-
gence for an indicator function 1f[u ; u + ]g, by approximating the indicator
by continuous functions, keeping in mind that is bounded.
Finally, since by Lemma 2.3.8 the function ~ is continuous and bounded, by
hypothesis (c) when k converges to infinity this expression converges to ~(0 (u))
for every continuity points u of 0 .
In this way we have proved that
lim sup [uN ] EN [ ] ~(0(u)) = E u[ ] :

N !1 0( )
The same arguments replacing the sequence N N

k ;+ by the sequence k ; prove
the reverse inequality. It follows from Lemma 2.1.5 that the sequence [uN ] N
converges weakly to 0 (u) for every continuity point u of 0 .
A similar argument to this one will permit to prove conservation of local equi-
librium for attractive processes from a weak local equilibrium statement. Details
are given in Chapter 9.
Remark 0.7 The definitions of product measure with slowly varying parameter
associated to a continuous profile, of local equilibrium in strong and weak sense
and the one of density profile depend uniquely on the existence of a parametrized
family of invariant measures. These definitions and the propositions stated above
extend therefore in a natural way to the generalized exclusion processes and to
the zero range processes defined in the previous chapter.
4. Hydrodynamic Equation of Symmetric Simple
Exclusion Processes
In this chapter we prove the hydrodynamic behavior of nearest neighbor symmetric

simple exclusion processes and show that the hydrodynamic equation is the heat
equation :
@t = (1=2) :
In this formula, stands for the Laplacian of : = 1id @u2 i .
P
We briefly present the strategy of the proof. We shall first prove that the
empirical measure solves the heat equation in a weak sense in an integral form.
More precisely, for a positive measure on Td of finite total mass and for a
continuous function G: Td ! R, denote by < ; G > the integral of G with
respect to : Z
< ; G > = G(u) (du) :
Td
We shall prove that the empirical measure tN associated to the symmetric simple
exclusion process converges, in a way to be specified later, to a measure t
absolutely continuous with respect to the Lebesgue measure and satisfying :
Zt
< t ; G > = < 0 ; G > + (1=2) < s ; G > ds (0:1)
0
for a sufficiently large class of functions G : Td ! R and for every t in an interval

[0; T ] fixed in advance.
Recall from Chapter 3 that we denoted by M+ = M+ (Td ) the space of finite
positive measures on Td endowed with the weak topology. In order to work in
a fixed space as N increases, we consider the time evolution of the empirical
measure tN associated to the particle system defined by :
X
tN (du) = N (t ; du) :=
1
Nd t (x) x=N (du) : (0:2)
x2TdN
Notice that there is a one to one correspondence between configurations and
empirical measures N (; du). In particular, tN inherits the Markov property from
t .
We consider the distribution of the empirical measure as a sequence of prob-
ability measures on a fixed space. Since there are jumps this space must be
48 4. Hydrodynamic Equation of Symmetric Simple Exclusion Processes
D([0; T ]; M+), the space of right continuous functions with left limits taking
values in M+ .
Fix a profile 0 : Td ! [0; 1] and denote by N a sequence of probability
measures associated to 0 . For each N 1, let QN be the probability measure
on D([0; T ]; M+) corresponding to the Markov process tN speeded up by N 2
and starting from N . We speeded up the process by N 2 because we have seen
in Chapter 1 that to obtain a non trivial hydrodynamic evolution for mean-zero
processes we need to consider time scales of order N 2 .
Our goal is to prove that, for each fixed time t, the empirical measure tN
converges in probability to (t; u)du where (t; u) is the solution of the heat
equation with initial condition 0 . We shall proceed in two steps. We first prove that
the process tN converges in distribution to the probability measure concentrated
on the deterministic path f(t; u)du; 0 t T g and then argue that convergence
in distribution to a deterministic weakly continuous trajectory implies convergence
in probability at any fixed time 0 t T .
A deterministic trajectory can be interpreted as the support of a Dirac prob-
ability measure on D([0; T ]; M+) concentrated on this trajectory. The proof of
the hydrodynamic behavior of symmetric simple exclusion processes is therefore
reduced to show the convergence of the sequence of probability measures QN to
the Dirac measure concentrated on the solution of the heat equation.
An indirect standard method to prove the convergence of a sequence is to
show that this sequence is relatively compact and then to show that all converging
subsequences converge to the same limit. To show the relative compactness we
will use Prohorov’s criterion. At this point it will remain the identification of all
limit points of subsequences.
Prohorov’s criterion states that a sequence of probability measures fQN , N
1g in a reasonable topological space ( e.g. a polish space cf. Billingsley (1968))
is weakly relatively compact if and only if for every " > 0 there is a compact set
K" such that for every N ,
QN (K" ) 1 ":
We must therefore examine the compact subsets of D([0; T ]; M+). In this respect,
notice that we have the choice of the topology to be attributed to D([0; T ]; M+).
We just need to be able to consider Dirac measures concentrated on a trajectory.
Our choice under this restriction is guided by the estimates we have and by the
simplicity of the corresponding compact sets. The simplest and strongest topology
is of course the topology of uniform convergence. However, the identification of
compact sets forces us to use the Skorohod topology.
To characterize all limit points of the sequence QN , we have to investigate
how we may use the random evolution to make an equation of type (0.1) appear.
Notice first that, under QN , for every function G: Td ! R, the quantities
X
< tN ; G > = N1d G(x=N ) t (x) (0:3)
x2TdN
1. Topology and compactness 49
verify the identity

Zt
< tN ; G > = < 0N ; G > + N 2LN < sN ; G > ds + MtG;N
0
where MtG;N are martingales with respect to the natural filtration Ft = (s ; s
t). The factor N 2 in front of the generator LN appears because we speeded up
the process by N 2 . In the particular case of nearest neighbor symmetric simple
exclusion processes, the second term on the right hand side may be rewritten as a
function of the empirical measure. Indeed, applying the generator to the function
! (x) we have :
d h
X i
L N ( x) = (1=2) (x + ej ) + (x ej ) 2 ( x) :
j =1
After two summations by parts we obtain that under QN
Zt
< tN ; G > = < 0N ; G > + (1=2) < sN ; N G > ds + MtG;N
0
where N is the discrete Laplacian :

d h
X + G (x e )=N 2G(x=N )i :
(N G)(x=N ) = N2 G (x + ej )=N j
j =1
To conclude the proof of the hydrodynamic behavior of symmetric simple

exclusion processes, it remains to show, on the one hand, an uniqueness theorem
for solutions of equations (0.1); uniqueness theorem that will require to prove
identity (0.1) for a certain class of functions G; and, on the other hand, to prove
that the martingales MtG;N vanish in the limit as N " 1 for this family of functions
G. From these two results it follows that the sequence QN has a unique limit point
Q which is the probability measure concentrated on the unique solution of (0.1).
1. Topology and compactness
We work on M+ (Td ), the space of finite positive measures on Td endowed with

the weak topology. Denote by C (Td ) the space of continuous functions on Td .
Recall that we may define a metric on M+ (Td) by introducing a dense countable
family ffk ; k 1g of continuous functions on Td and by defining the distance
( ; ) by
X
1 1 j < ; f > < ; f > j
k k
(; ) = 2k 1 + j < ; fk > < ; fk > j
(1:1)
k=1
We assume hereafter that f1 = 1. M+ endowed with the weak topology is complete

and a set A M+ is relatively compact if and only if
sup < ; 1 > < +1 ; (1:2)
2A
that is, if the total masses are uniformly bounded.
On the other hand, from the existence of jumps, the natural space to consider
for the evolution of the empirical measure t is D([0; T ]; M+), the set of right
continuous functions with left limits taking values in M+ . To endow this space
with a reasonable topology notice that we cannot demand the uniform convergence
because a small time change in a jump should not modify too much the proximity
of two paths.
The results presented below are a short recapitulation of results from Chapter
3 of Billingsley (1968) or Chapter 3 of Ethier and Kurtz (1986). Since in next
chapters we will sometimes be interested in other jump processes than the empirical
measure and since all results do not depend on the general structure of M+ , we
state them in a general setting.
Consider, therefore, a complete separable metric space E with metric (; ) and
a sequence P N of probability measures on D([0; T ]; E ). Elements of E are denoted
by the Greek letters and . Let be the set of strictly increasing continuous
functions of [0; T ] into itself. We define then
(t) (s)
kk
s=t t s
= sup log
=
and
d(; ) = inf max
2
kk ; sup (t; (t) ) :
tT
0
Proposition 1.1 D([0; T ]; E ) endowed with the metric d is a complete separable

metric space.
Even though this definition of distance is important because it allows to use the
property of completeness to identify compact sets, it is not very useful in practice
because it takes into account all functions of . (Remember that a complete set
A is compact if and only if it is precompact, that is, if for every " > 0 A can be
covered by a finite number of balls of radius less than or equal to ").
The main tool will be a modified uniform modulus of continuity that allows
to extend the Ascoli Theorem to the set D([0; T ]; E ). For this reason we introduce
w0 ( ) : = ft ginf 0max sup (s ; t ) ;
i ir i<r ti s<t<ti
0 +1
where the first infimum is taken over all partitions fti ; 0 i rg of the interval
[0; T ] such that (
0 = t0 < t1 < < tr = T
ti ti 1 > i = 1; : : : ; r :
We can easily verify that the trajectory ft ; 0 t T g belongs to D([0; T ]; E ) if

and only if the modified uniform modulus of continuity w0 ( ) vanishes as # 0.
Moreover, we can characterize the compact sets of D([0; T ]; E ) in terms of the
modified modulus of continuity.
Proposition 1.2 A set A of D([0; T ]; E ) is relatively compact if and only if

(1) ft ; 2 A ; t 2 [0; T ]g is relatively compact on E ,
(2) lim !0 sup2A w0 ( ) = 0:
With this result we obtain a statement of Prohorov’s theorem.
Theorem 1.3 Let P N be a sequence of probability measures on D([0; T ]; E ). The

sequence is relatively compact if and only if
(1) For every t in [0; T ] and every " strictly positive, there is a compact K (t; ") E
such that supN P N [t 62 K (t; ")] " :

(2) For every " strictly positive, lim !0 lim supN !1 P N ; w0 ( ) > " = 0 :

Remark 1.4 The second condition that appears in the previous theorem, that
depends on all path ft ; 0 t T g and not only on the behavior at a fixed time
t, is the most difficult to verify. Nevertheless, for any positive , the modified
uniform modulus of continuity w0 ( ) is naturally smaller than the usual uniform
modulus of continuity w calculated at 2 : w0 ( ) w (2 ), where
w ( ) = sup (s ; t ) :
jt sj
In all cases we investigate in the following chapters, instead of condition (2)

of the previous theorem, we prove :
(2’) For every " strictly positive, lim !0 lim supN !1 P N ; w ( ) > " =
0:
Remark 1.5 It is easy to see that all limit points of a sequence PN satisfying
(2’) are concentrated on continuous paths.
We also have a very useful sufficient condition due to Aldous (1978). To state
this result denote by TT the family of all stopping times bounded by T .
Proposition 1.6 A sequence of probability measures P N on D([0; T ]; E ) satisfies

condition (2) of Theorem 1.3 provided
h i
lim lim sup sup
!0 N !1 2TT
P N ( ; + ) > " = 0 (1:3)

for every " > 0.
Here and in the proof of this result, by convention, all times are assumed
bounded by T so, for instance, + should be read as ( + ) ^ T .
Proof of Proposition 1.6. Fix " > 0. By assumption, there exists 0 > 0 and
N0 such that for N > N0
h i
P N ( + ; ) > " " (1:4)
for each 2 0 and each stopping time . Let M > 2T= 0 be a fixed integer.
By the same reasons, there exists 0 > 0 and N1 N0 such that for N > N1 ,
h i
P N ( + ; ) > " M" (1:5)
for each 20 and each stopping time .

For 0 i M , define recursively the stopping times i by 0 = 0 and
n o
i+1 = min t > i ; (i ; t ) 2" :
By convention if the set on the right hand side is empty, we fix i+1 to be T .
To prove the proposition, it is enough to show that for N N1 ,
P N [M < T ] 16" and
h i (1:6)
P N i+1 < T ^ (i + ) for some 0 i M 1 8" :
Indeed, it follows from the definition of the stopping time fi ; 0 i M g and
(1.6) that h i
P w0 () > 4" 24"
for all N N1 , which is exactly condition (2) of Theorem 1.3.
It remains to prove (1.6). Denote by F the -algebra generated by ft ; 0
t T g. Consider a random variable U distributed uniformly on [0; 2 0] and inde-
pendent of F . We abuse of notation and denote by P N the probability measure on
D([0; T ]; E ) [0; 2 0] corresponding to the variable (; U ). For a fixed trajectory
, and 0 t1 t2 T , assume that
h i
t2 t1 < 0 and P N (tj +U ; tj ) < " F > 3=4
for j = 1, 2. In this formula the integration is made with respect to the variable
U . Then, there exists t in [t1 + 0 ; t1 + 2 0 ] such that
(t ; tj ) < "
for j = 1, 2. In particular, (t ; t ) < 2" and
n 1 2
o
(; t1 ; t2 ); (t ; t ) 2"; t1 t2 < t1 + 0
1 2
i n
[ h i o
P N (tj +U ; tj ) " F 1=4 :
+1
(; t1 ; t2 );
j =i
Thus for 0 i M 1,
N
h i
P (i ; i ) 2" ; i+1 < i + 0
+1
X
i n h i o
; P N (j +U ; j ) "F 1=4 :
+1
N P
j =i
By Chebychev inequality and (1.4), last expression is bounded above by
X
i
N
h
+1 i
4 P (j +U ; j ) " 8" :
j =i
Since, for 0 i M 1 the set fi+1 < T g is contained in f(i ; i ) 2"g
+1
we obtain from the two previous estimates that
h i
P N i+1 < i + 0 ; i+1 < T 8" (1:7)
for 0 i M 1. A similar argument taking advantage of (1.5) instead of (1.4)

shows that h i 8"
P N i+1 < i + 0 ; i+1 < T M
for 0 i M 1, what proves the second statement in (1.6).
To prove the first claim in (1.6), notice that fM < T g fi+1 < T g for
0 i M 1. In particular, by Chebychev inequality and estimate (1.7) we have
h i h i
that

E N i+1 i M < T 0 P i+1 i 0 M < T

n P N [i+1 i < 0 ; M < T ] o n 8" o
= 0 1
P N [M < T ] 0 1
P N [M < T ]
Thus,
h i MX1 h i
T E N M M < T = E N i+1 i M < T
i=0
n 8" o
M 0
P N [M < T ]
1
This proves the first claim of (1.6) because we chose M > 2T= 0 .
In the context of this book, i.e., in the case of M+ endowed with the weak
topology, to prove the relative compactness of a sequence of measures QN defined
on D([0; t]; M+) it is enough to check conditions of Theorem 1.3 or the one of
Remark 1.4 for each real process obtained by “projecting" the empirical measure
tN with functions of a dense countable set of C (Td). More precisely, the fol-
lowing result asserts that to establish the relative compactness of a family QN of
probability measures on D([0; T ]; M+) it is enough to study the same problem for
the processes < tN ; gk > for a dense countable family fgk ; k 1g of functions
of C (Td ).
Proposition 1.7 Let fgk ; k 1g be a dense subfamily of C (Td ) with g1 = 1. A

family of probability measures (QN )N 1 on D([0; T ]; M+) is relatively compact if
for every positive integer k the family QN gk 1 of probabilities on D([0; T ]; R) has
this property. Here, QN gk 1 is the sequence of measures obtained by “projecting"
QN with a function gk and is defined by
QN gk 1[A] =

QN N ; < N ; gk > 2 A :
Proof. We have already observed in (1.2) that a set A M+ is weakly relatively

compact if sup2A < ; 1 >< 1. Fix " > 0 and 0 t T . By assumption,
there exists A > 0 such that
h i
P N j < t ; g1 > j > A "
for every N 1. In particular, since g1 = 1 and since the set f; j < ; g1 > j
Ag is weakly relatively compact, the first condition of Theorem 1.3 is proved.
To prove the second condition, fix " > 0 and > 0. Let k" such that 21 k" ".
It follows from the definition of the distance introduced in (1.1) and of k" that
for each > 0,
k" 1X "
w0 ( ) w0
2k <;gk >
( ) +
2
(1:8)
k=1
By assumption, there exists 0 such that
h 0 i
P N w<;g k > ( ) > "=2 2k
for each k k" , 0 and N 1. Therefore,

hX
k" 1 0 i
PN w
2k <;gk >
( ) > "=2
k=1
for each 0 and N 1. This together with (1.8) shows that
h i
P N w0 ( ) "
for each 0 and N 1.
2. The hydrodynamic equation 55
2. The hydrodynamic equation
We are now in possession of all elements to examine the hydrodynamic behavior

of the symmetric simple exclusion process.
Theorem 2.1 Let 0 : Td ! [0; 1] be an initial density profile and let N be the
sequence of Bernoulli product measures of slowly varying parameter associated to
the profile :
N f; (x) = 1g = 0 (x=N ) ; x 2 TdN :
Then, for every t > 0, the sequence of random measures
X
tN (du) =
N d x2Td t (x) x=N (du)
1
N
converges in probability to the absolutely continuous measure t (du) = (t; u) du
whose density is the solution of the heat equation :
(
@t = (1=2)
(2:1)
(0; ) = 0 ( ) :
Proof. We start fixing a time T > 0 and considering a sequence of probability

measures QN on D([0; T ]; M+) corresponding to the Markov process tN , defined
by (0.2), speeded up by N 2 and starting from N .
First step (Relative compactness). We have seen in section 1 that the first step
in the proof of the hydrodynamic behavior consists in showing that the sequence
QN is relatively compact. Denote by C 2 (Td) the space of twice continuously
differentiable functions G : Td ! R. Of course C 2 (Td ) is dense in C (Td ) for
the uniform topology. By Proposition 1.7 it suffices to check that the sequence of
measures corresponding to the real processes < tN ; G > is relatively compact for
all G in C 2 (Td ).
Fix therefore a function G in C 2 (Td ) and denote by QN;G the probability
measure QN G 1 on D([0; T ]; R). Since < tN ; G > is a real process, we shall
apply Theorem 1.3 and Proposition 1.6 with E = R and the usual distance in R.
Since the total mass of the empirical measure tN is bounded by 1, condition (1)
of Theorem 1.3 is trivially verified. It remains to prove the second condition or,
as we shall do, the one of Proposition 1.6. Recall from (0.3) that under QN we
have the identity :
Zt
< tN ; G > = < 0N ; G > + (1=2) < sN ; N G > ds + MtG;N (2:2)
0
where MtG = MtG;N is a martingale. In particular, in order to prove (1.3)R for

< t ; G >, we just need to prove the same relation for the integral term 0t <
N
sN ; N G > ds and for the martingale term MtG. We start with the former.
Of course, since G is of class C 2 and since the total mass of sN is bounded
by 1, the integral
Z +
< sN ; N G > ds

is bounded above by C (G), whether is a stopping time or not. (From now on,
C (G) represents a finite constant depending only on G and that may change from
line to line). In particular, the integral term on the right hand side of (2.2) satisfies
the condition of Proposition 1.6.
To check condition (1.3) for the martingale MtG;N we show that the expected
value of its square converges to 0. Denote by BtG = BtG;N the process given by
BtG
N 2 LN < tN ; G >2 2N 2 < tN ; G > LN < tN ; G > :
=
By Lemma A1.5.1, the process NtG = NtG;N defined by :
NtG MtG
Z t BG ds :
2
= s
0
is a new martingale. Straightforward computations show that

X
BsG =
1
N 2(d 1)
G(y=N ) G(x=N ) s (x)(1 s (y)) : 2
jx yj=1
Since is a stopping time we have that
"Z #

EQN (M M ) = EQN BsG ds CN(Gd)
2
+
+

Condition (1.3) follows from this estimate and Chebychev inequality. This con-
cludes the proof that the sequence QN is relatively compact.
Second step (Uniqueness of limit points). Now that we know that the sequence
QN is weakly relatively compact, it remains to characterize all limit points of QN .
Let Q be a limit point and let QNk be a sub-sequence converging to Q .
First of all, notice that the application from D([0; T ]; M+) to R that associates
to a trajectory ft ; 0 t T g the number
Zt
< t ; G > < ; G > (1=2) < s ; G > ds
tT
sup 0
0
is continuous as long as G is of class C 2 . We therefore have that for every " > 0
! Zt
lim inf QNk sup < t ; G >< ; G > (1=2) < s ; G > > "
k!1 tT
0
Zt !
0
Q sup < t ; G > < ; G > (1=2) < s ; G > > " :
0
tT 0
since QNk converges weakly to Q and since the above set is open.
At this point the same estimate on the martingale (MtG )2
t G R
0 Bs ds obtained

in the first part of the proof shows that every limit point Q is concentrated on
trajectories such that
Zt
< t ; G > = < 0 ; G > + (1=2) < s ; G > ds (2:3)
0
for all 0 t T . Indeed, by Chebychev and Doob inequality,

" G #
QN

sup Mt " 4" EQN (MTG )
2 2
tT
0
"Z T #
= 4" 2
EQN C (G)T
BsG ds "2 N d
0
To conclude that all limit points are concentrated on trajectories t satisfying

relation (2.3), we just need to point out that the difference between < tN ; N G >
and < tN ; G > is of order oN (1) because the total mass of tN is bounded by
1 and G is of class C 2 .
We now prove that all limit points Q of the sequence QN are concentrated
on absolutely continuous measures with respect to the Lebesgue measure. Notice
that X
0tT

sup < tN ; G > N d x2Td jG(x=N )j
1
N
because there is at most one particle per site. Since, on the other hand, for fixed
continuous functions G, the application that associates to a trajectory the value
sup0tT j < t ; G > j is continuous, by weak convergence, all limit points are
concentrated on trajectories t such that
Z
< t ; G > jG(u)j du
for all continuous function G and for all 0 t T . Thus all limit points are
concentrated on absolutely continuous trajectories with respect to the Lebesgue
measure :
Q [; t (du) = t (u)du] = 1 :
Moreover, all limit points of the sequence QN are concentrated on trajectories
that at time 0 are equal to 0 (u)du . Indeed, by weak convergence, for every " > 0
2 Z 3
X
Q 4 N d
1
G(x=N ) (x) G(u) (u) du > "5
x2TdN 0
0
2 Z 3
X
lim inf QNk 4 d G (x=N ) (x) G(u) (u) du > "5
1
k!1 N d 0
0
2 x2TN Z
3
X
= lim Nk 4 d G(x=N )(x) G(u) (u) du > "5 = 0 :
1
k!1 N x2TdN 0
The two previous results show that every limit point is concentrated on abso-
lutely continuous trajectories t (du) = t (u)du whose density is a weak solution,
in the sense (2.3), of the heat equation. To prove, however, an uniqueness result of
weak solutions for the heat equation, we need to prove relation (2.3) for time de-
pendent functions G. For positive integers m and n, denote by C m;n ([0; T ] Td)
the space of continuous functions with m continuous derivatives in time and n
continuous derivatives in space. For G : [0; T ] Td ! R of class C 1;2 , consider
the martingale MtG = MtG;N given by
Zt
MtG = < tN ; Gt > < 0N ; G0 > (@s + N 2 LN ) < sN ; Gs > ds :
0
By Lemma A1.5.1, the process NtG = NtG;N defined by
NtG MtG
N Z t A (s) ds
2
= G
2
0
with AG (s) = LN < sN ; Gs >2 2 < sN ; Gs > LN < sN ; Gs >
is a martingale with respect to the natural filtration. In particular, as before,
h G i
lim sup QN
N !1
Mt > " = 0
for every " > 0. Since, on the other hand,

Zt
MtG = < tN ; Gt > < 0N ; G0 > < sN ; @s Gs (1=2)N Gs > ds ;
0
all limit points are concentrated on paths ft ; 0 t T g such that

Zt
< t ; Gt > = < 0 ; G0 > + < s ; @s Gs (1=2)Gs > ds ; (2:4)
0
In conclusion, all limit points are concentrated on absolutely continuous trajec-

tories t (du) = t (u)du that are weak solutions of the heat equation in the sense
of (2.4) and whose density at time 0 is 0 ().
Third step (Uniqueness of weak solutions of the heat equation). We turn now
to the question of uniqueness of weak solutions. We first fix the terminology on
weak solutions of partial differential equations. Let L be a second order differential
operator acting only on space variables :
X X
L = @u2 i ;uj Ai;j () + @ui Bi () + C ()
1 i;jd 1 id
for smooth functions Ai;j , Bi and C . In this chapter, for example, L represents
one half of the Laplacian.
Definition 2.2 For a bounded initial profile 0 : Td ! R, a bounded function

: [0; T ] Td ! R is a weak solution of the Cauchy problem
(
@t = L
(0; ) = 0 ()
if for every function G: [0; T ] Td ! R of class C 1;2
Z Z
G(T; u)(T; u) du G(0; u)0(u) du
T d Td
ZT Z n X o
= dt du @t G + Ai;j ()@u2 i ;uj G
T d
i;jd
ZT Z
0
Xn
1
dt du Bi ()@ui G + C ()G :
0 Td
1id
It was proved in the previous section that every limit point of the sequence
QN is concentrated on weak solutions of the heat equation with initial profile 0 .
Therefore, to conclude the proof of the uniqueness of limit points, it remains to
show that there exists only one weak solution of this equation.
There exists many methods. Brezis and Crandall (1979) proved such a result for
a class of quasi-linear second order equations. Their theorem gives us immediately
the result since Definition 2.2 of weak solutions implies their condition (1.23)
and since the limit condition on the boundedness in L1 is automatically satisfied
because the total mass in simple exclusion processes is conserved and is at most
1.
We present in Appendix A2.4 a uniqueness result based on the investiga-
tion of the time evolution of the H 1 norm. This method requires, however,
supplementary properties of weak solutions that are not difficult to check in the
case of symmetric simple exclusion processes.
Finally, since the hydrodynamic equation of symmetric simple exclusion is lin-
ear, the methods developed by Oelschläger (1985) give a third possible approach.
Note that for any bounded profile 0 the heat equation admits a strong solution
given by
Z
(t; u) = ¯0 (v) Gt (u v) dv
if ¯0 : Rd ! R represents the Td -periodic function identically equal to 0
on Td and if Gt (w) is the usual d-dimensional Gaussian kernel : Gt (w) =
(2t) d=2 expf (1=2t)jwj2g. In particular the weak solution is in fact a strong
solution.
In conclusion, with any of these uniqueness results, we proved that the se-
quence QN converges to the Dirac measure concentrated on this strong solution.
Fourth step (Convergence in probability at fixed time). Even if in general it
is false that the application from D([0; T ]; M+) to M+ obtained by taking the
value at time 0 < t < T of the process is continuous, this statement is true if the
process is almost surely continuous at time t for the limiting probability measure.
In the present context, the limiting probability measure is concentrated on weakly
continuous trajectories. Thus tN converges in distribution to the deterministic
measure t (u)du. Since convergence in distribution to a deterministic variable
implies convergence in probability, the theorem is proved.
In the previous proof, the initial state N appeared only in the second step. It
was necessary to show that the limit points Q were concentrated on trajectories
t (du) that at time 0 were given by
0 (du) 0 (u) du :
=
Therefore, the special structure of the measure N did not play any particular role
in the proof and the hypothesis of Theorem 2.1 concerning the initial state can be
considerably relaxed :
Theorem 2.3 For 0 : Td ! [0; 1], consider a sequence of measures (N )N 1 on

f0; 1gTdN associated to the profile 0 :
" X Z #
lim sup N
1 G(x=N )(x) G(u) (u) du > = 0:
N !1 Nd x 0
for every > 0 and every continuous function G: Td ! R. The conclusions of

Theorem 2.1 remain in force.
Thus, under the hypothesis of a weak law of large numbers at time 0 for the
empirical measure N , we have proved a law of large numbers for any later time
t.
Remark 2.4 In the proof of the hydrodynamic behavior, we took advantage of
many of the special features of symmetric simple exclusion processes. We will try
to point out here the elements that were used.
(a) From the conservation of the total mass, condition (1) of the compactness
criterion presented in Theorem 1.3 refers only to the total initial mass. This
condition was therefore easy to check.
R
(b) By a summation by parts we have been able to rewrite the integral term
T 2 N
0 N LN < t ; G > dt as a sum involving the discrete Laplacian of G.
This property is shared by many processes. Indeed, for a nearest neighbor
system conserving the total number of particles we often have (in dimension
one) that
L(x) = Wx 1;x () Wx;x+1()
where Wx;x+1 stands for the instantaneous current of particles from x to x +1.
In few particular cases, like in symmetric simple exclusion processes where
Wx;x+1 () = (x) (x + 1), the current is itself a discrete gradient of another
cylinder function. This condition known as the “gradient condition" is the one
that permitted to perform a second discrete integration by parts to cancel the
factor N 2 that appeared from time renormalization.
R
(c) After these two summations by parts we were able to rewrite the integral
T
term 0 N 2 LN < tN ; G > dt as a function of the empirical measure. We
obtained in this way a closed equation for the empirical measure. This is a
very special feature of symmetric simple exclusion processes. In general, we
obtain a correlation field and the main difficulty in the proof is to close the
equation, that is, to replace the correlation field by a function of the empirical
measure.
In the next two chapters we will examine the hydrodynamic behavior of a

gradient system whose correlation field is not a function of the density field and
in Chapter 7 we investigate a nongradient system.
Since gradient systems appear several times in the next chapters we introduce
the following terminology. For two sites x and y we denote by Wx;y ( ) the
instantaneous algebraic current between sites x and y , that is, the rate at which
a particle jumps from x to y minus the rate at which a particle jumps from y to
x. In the nearest neighbor symmetric simple exclusion process, for example, the
current W0;e1 ( ) is equal to p(e1 )[ (0) (e1 )].
Definition 2.5 Let E denote a subset of N . A translation invariant nearest neigh-

d
bor particle system (t ) on E TN with generator LN is said to be gradient if there
exists a positive integer n0 , cylinder functions fhi;n ; 1 i d; 1 n n0 g
and finite range functions fpi;n ; 1 i d; 1 n n0 g such that
X
n X
0
W0;ei () = pi;n (x)x hi;n () and

n=1 x2TdN
X
pi;n (x) = 0 for every 1 n n0
x2TdN
for every 1 i d.
This definition can of course be extended to processes that have finite range
transition probabilities. Notice that for a function G: Td ! R of class C 2 the
gradient condition allows two discrete integrations by part in the integral term of
the martingale
Zt
MtG;N = < tN ; G > < N ; G >
0 N 2 LN < sN ; G > ds
0
defined in the beginning of this section.
Tracer particles and self diffusion. The diffusion of a tracer particle is a

standard problem in non equilibrium statistical mechanics. Consider a large system
of interacting particles and add one, the tracer particle. Denote by Xt its position
at time t. One expects Xt to behave as a Brownian motion with some diffusion
coefficient on a large space and time scale.
P P
To fix ideas consider a mean-zero, asymmetric simple exclusion process
ft ; t 0g on Zd with transition probability p() ( x p(x) = 1, x xp(x) = 0).
The generator L of this process acts on cylinder functions f as
X
(Lf )( ) = p(y)(x)[1 (x + y)][f (x;x+y) f ()] :
x;y2Zd
We refer to Liggett (1985) for a proof of the existence of this process. We start the
evolution with one tagged particle at the origin and a Bernoulli product measure
with density on Zd f0g. Denote by Xt the position of the tracer particle at
time t. One would like to prove that XtN 2 =N converges, as N " 1, to a Brownian
motion with diffusion matrix D = D() (the so called self diffusion matrix).
To investigate the behavior of the tracer particle it is more convenient to
describe the system in terms of the position of the tagged particle and the config-
uration seen from the tagged particle. It is easy to check that the process as seen
from the tagged particle, i. e., in which the tagged particle is fixed to be at the
origin, is itself Markovian. Its generator has two pieces, the first one corresponds
to the jumps of the tracer particle and the second one corresponds to all other
jumps. The generator of the process becomes
X
(Lf )( ) = p(x)[1 (x)][f (Tx ) f ( )]
x2Zd
X
+ p(y) (x)[1 (x + y)][f ( x;x+y) f ( )] ;
x=6 0
y2Zd
where Tx is the configuration obtained from letting one particle jump from
0 to x and then translating the new configuration by x : Tx = x 0;x . If the
transition probability is symmetric and of finite range (p(x) = p( x), p(x) = 0 for
x large enough), the Bernoulli product measure with density conditioned to have
a particle at the origin is reversible and ergodic for this process. (For zero range
processes with the same assumptions on the transition probability, the probability
measure ^ (d ) = f (0)=g(d ) is ergodic and reversible for the process as
seen from the tagged particle.)
Recall that Xt stands for the position of the tracer particle at time t. For a fixed
P
in Rd , a simple computation shows that L(Xt ) = z p(z )(z )[1 (z )] =
( ). In this formula (u v) stands for the inner product of u and v in Rd .
R
t
It follows from Lemma A1.5.1 that Xt = 0 (s ) ds + Nt , where N is a
martingale. Since there exists a wide class of central limit theorems for martingales
t R
(cf. Helland (1982) or Durrett (1991)), the problem is reduced to the analysis of
the asymptotic behavior of t 1=2 0 (s ) ds. To this end Kipnis and Varadhan
(1986) proved a central limit theorem for additive functionals of reversible Markov
processes :
Theorem 3.1 (Kipnis and Varadhan). Let Yt be a Markov process on a state space
E with generator L. Let be an invariant measure and assume that is reversible
and ergodic. Let V : E ! R be a mean-zero function in L2 ( ) (E [V ] = 0) for
which there exists a finite constant C = C (V ) such that for every f in L2 ( ),
< V; f > C < Lf; f > : (3:1)

Here < ; > stands for the inner product on L2 ( ). Then, there exists a P -square
integrable martingale Mt with stationary increments, measurable with respect to
the natural -algebra and such that M0 = 0 and
Z s
lim p sup V (Yr ) dr Ms
1
= 0
t!1 t st 0 0
in P probability.
d
Since in the symmetric case the Bernoulli measures on f0; 1gZ f0g are ergodic
and reversible, this theorem reduces the proof of the central limit theorem for the
the tagged particle to the verification of property (3.1) for the function ( ). This
follows from an elementary computation involving the generator L.
The proof of Theorem 3.1 relies on spectral analysis. For > 0, consider the
resolvent equation
u Lu = V :
R t V (Y ) ds as
With this notation we may rewrite the integral s
Zt Zt
0
V (Ys ) ds M (t)
u ( Y ) u ( Y ) u (Ys ) ds ;
= + 0 t +
0 0
where M (t) is a martingale with stationary increments. It follows from assumption

R
t
(3.1) that M (t) (resp. 0 u (Ys ) ds and u (Y0 ) u (Yt )) converges in L2 (P ),
as # 0, to some martingale M (t) (resp. 0 and some process t ). To conclude the
proof of the theorem it remains to show that t 1=2 sup0st js j converges to 0
in probability as t " 1. This is more demanding. The argument can be found in

Kipnis and Varadhan (1986).
This is in substance the method that allowed Kipnis and Varadhan to prove a
central limit theorem for the position of a tagged particle for symmetric simple
exclusion processes. It follows from their proof that the self diffusion matrix D is
non degenerated in dimension d 2 and in dimension d = 1 if p(1) + p( 1) < 1.
We discuss below the behavior of the tracer particle in the one-dimensional nearest
neighbor exclusion process.
De Masi, Ferrari, Goldstein and Wick (1985, 1989) extended Kipnis and Varad-
han central limit theorem for additive functionals to antisymmetric random vari-
ables of reversible Markov processes. Goldstein (1995) presents a simpler proof of
this latter result. Varadhan (1995) extended Kipnis and Varadhan’s method to a non
reversible context, the case of mean-zero asymmetric simple exclusion processes.
Spohn (1990) obtains a variational formula for the diffusion matrix D(). (cf.
also Quastel (1992)). This variational formula permits to show that for exclusion
processes where the jump rates depend on the configuration, the so called Kawasaki
dynamics, the diffusion matrix is still non degenerated for all reversible dynamics
that are not nearest neighbor one-dimensional. Varadhan (1994b) proves that the
self diffusion coefficient D() is Lipschitz continuous in dimension d 3 and that
there exist constants 0 < C0 < C1 < 1, such that C0 (1 ) D() C1 (1 )
in the matrix sense in all dimensions. Asselah, Brito and Lebowitz (1997) prove
bounds on D() in terms of the range of the jumps.
Grigorescu (1997a,b) investigates the evolution of one-dimensional indepen-
dent Brownian motions. Particles evolve independently until they touch. At this
point, with a small probability " particles collide and with probability 1 " particles
do not collide. Grigorescu (1997a,b) proves a central limit theorem for a tagged
particle starting from a nonequilibrium state. In this context the evolution of the
tagged particle depends on the evolution of the all macroscopic profile, described
by the hydrodynamic equation.
Siri (1996) considers space inhomogeneous symmetric zero range processes
where a particle at site x jumps at rate g (x=N; (x))p(y ) to site x + y . Here p()
is a finite range symmetric transition probability and g : Rd N ! R+ a jump
rate smooth in the first coordinate. She proves a central limit theorem for the
tracer particle in this context and obtains a space dependent diffusion matrix.
Shiga (1988) investigates a class of particle systems on Z in which the tagged
particle converges to a Gaussian random variable with variance of order t1=a for
1 < a 2. In these models more than one particle may jump simultaneously.
Carlson, Grannan and Swindle (1993) prove a central limit theorem for a tagged
particle in a one-dimensional symmetric exclusion process in which a particle at
x jumps to x + y at rate c(jyj) provided all sites between x and y are occupied.
For the one-dimensional, symmetric, nearest neighbor simple exclusion pro- p
cess Arratia (1983) proved that the right space rescaling is t1=4 instead of t.
Harris (1965) already observed that this subdiffusive behavior is peculiar to one-
dimensional, nearest neighbor processes. Arratia (1983) proved that starting from
a Bernoulli product measure with density and conditioned to have a particle at

the origin, Xt =t1=4 converges to a mean-zero Gaussian distribution with variance
equal to r
21
(3:2)
Rost andp Vares (1985), relating this process to a zero range process, prove that
XtN 2 = N converges in the Skorohod space to a fractional Brownian motion.
Landim, Olla and Volchan (1997), (1998), considered the evolution of an asym-
metric tagged particle, jumping with probability p to the right and q = 1 p to
the left, in an infinite, one-dimensional system of particles evolving according to
nearest neighbor, symmetric simple exclusion process. They proved that the diffu-
sively rescaled position of the asymmetric particle, (XtN 2 X0 )=N , converges in
probability to a deterministic function vt that depends only on the initial profile.
p case where the initial profile is constant equal to , they proved
In the particular
that v (t) = v t, where
r
v 21
p q !0 p q
lim =
This result, which relates the mobility v=(p q ) of an asymmetric particle in a

symmetric environment to the diffusivity (3.2) of a tagged symmetric particle, can
be interpreted as an Einstein relation.
Einstein relations. Consider a tagged particle, whose position at time t is denoted
by Xt , in a symmetric environment. Denote by D its self diffusion coefficient.
When the tracer particle is submitted to an external field E , that does not alter the
evolution of the environment, it starts to move with a drift. Denote by v (E ) the
mean drift when the environment is in equilibrium : v (E ) =< Xt > =t and set
= limE!0 v(E )=E . Einstein relation states that = D, where is the inverse
temperature.
Ferrari, Goldstein and Lebowitz (1985) proved Einstein relation for particle
systems for which there exists a unique invariant measure for the process as
seen from the tagged particle. They considered, for instance, exclusion processes
on cubes with periodic boundary conditions. Buttà (1993) proved the Einstein
relation in the case of an interface dynamics. Lebowitz and Rost (1994) proposed
a general approach to derive the Einstein relation for interacting particle systems
and applied it to several distinct class of models.
We refer to Lebowitz and Spohn (1982a), Spohn (1991) for the physical aspects
of the evolution of tracer particles. Ferrari (1996) presents a review on limit
theorems for tagged particles. We discuss in the last section of Chapter 8 the
behavior of tagged particles in asymmetric interacting particle systems.
Propagation of chaos. Rezakhanlou (1994b) proved the propagation of chaos for
symmetric simple exclusion processes and Quastel, Rezakhanlou and Varadhan
(1997) proved the corresponding large deviations principle in dimension d 3.
Consider a symmetric simple exclusion process with transition probability p()
evolving on TdN . Fix a profile 0 : Td ! [0; 1] and denote by N a sequence of

probability measures associated to 0 in the sense that
h X Z i
E N N1d
N !1
lim H (x=N )(x) H (u)0(u)du = 0
x2TdN
for every continuous function H : Td ! R+ .
Fix a time T > 0R and let L stand for the (random) total number of particles
at time 0 (L=N d ! Td 0 (u)du as N " 1). Tag all particles at time 0. Denote
by xi (t), 1 i L, the position of the i-th particle at time t and by yiN (t) the
rescaled position at time t of the i-th particle : yiN (t) = N 1 xi (tN 2 ). Let RN be
the empirical measure associated to these trajectories up to time T :
n o
N d yN () + + yLN () :
RN 1
= 1
Notice that for each N 1, RN is a measure on D([0; T ]; Td). We may thus

interpret RN as a random variable taking values in M(D([0; T ]; Td)). Denote by
QN the probability measure on M(D([0; T ]; Td)) induced by RN and N .
Denote by R the distribution on C (R+ ; Td ) corresponding to a diffusion
process on Td with generator L given by
0
L =
1
2
r S ((t; u))r + 21 [S ((t; u)) D] r(t;(t;uu) ) r
and initial distribution 0 (u)du. Here (t; u) is the solution of the hydrodynamic
equation @t = (1=2)r Dr, D is the covariance matrix of the transition prob-
ability p() and S () is the self–diffusion coefficient of the symmetric simple
exclusion process with density .
Rezakhanlou (1994b) proved that RN converges in distribution (and thus in
probability) to the (constant) random variable R0 : For every continuous function
J : D(R+ ; Td) ! R,
h 1 X
L Z i
lim EQN J N
(yj ()) JdR = 0:
N !1 L j =1
0
Quastel, Rezakhanlou and Varadhan (1997) proved a large deviations principle

for RN in dimension d 3. More precisely, they proved the existence of a lower
semi–continuous rate function I : M(D([0; T ]; Td)) ! R+ which has compact
level sets such that
lim sup QN [RN 2 F ] minf2F I (m)
N !1
lim inf QN [RN 2 G] inf I (m)
N !1 m2G
for all closed subsets F and open subsets G of M(D([0; T ]; Td)).
5. An Example of Reversible Gradient System :
Symmetric Zero Range Processes
In this chapter we investigate the hydrodynamic behavior of reversible gradient

interacting particle systems. To keep notation simple and to avoid minor technical
difficulties, we consider the simplest prototype of reversible gradient system : the
nearest neighbor symmetric zero range process. The generator of this Markov
process is given by
XX
LN f ( ) = (1=2) g((x)) [f (x;x+z) f ()] (0:1)
x2TdN jzj=1
d
for cylinder functions f : N TN ! R.
Recall from the last chapter the strategy adopted to prove the hydrodynamic
behavior of symmetric simple exclusion processes. The argument relied on the
analysis of some martingales associated to the empirical measure. More precisely,
to each smooth function G : Td ! R of class C 2 (Td ), we considered the martin-
gales MtG;N = MtG and its quadratic variation NtG;N = NtG defined by
Zt
MtG = < tN ; G > < N ; G >
0 N 2 LN < sN ; G > ds
0
NtG = (MtG )2
Z tn o
N 2 LN < sN ; G >2 2 < sN ; G > N 2 LN < sN ; G > ds :
0
In the context of nearest neighbor symmetric zero range processes, these martin-
gales can be rewritten as
X X
MtG =
Nd
1
G(x=N ) t (x) N1d G(x=N ) 0 (x)
x2TN
d x2TN
d
Zt X X h i
1
2N d 2
G(x=N ) g(s (y)) g(s (x)) ds
0
x2TdN jy xj=1
and
Zt X
NtG = (MtG)2 2N 2d
1
2
G(y=N ) G(x=N ) g(s (x)) ds :
2
0 jy xj=1
68 5. Reversible gradient systems
In the same way as we did for symmetric simple exclusion processes, we

perform discrete integrations by parts to rewrite the integral term of the martingale
MtG as : Z t 1 X
d N G(x=N ) g(s (x)) ds ;
0 2N
x2TdN
where N represents the discrete Laplacian defined in the previous chapter. Notice
that it is the gradient condition that allowed us to perform two integrations by parts.
This is easy to see in dimension 1 where the instantaneous current over the bond
f0; 1g, i.e., the rate at which a particle jumps from 0 to 1 minus the rate at which
a particle jumps from 1 to 0, is equal to
n o
1
2
g((0)) g((1)) :
This time, however, it is not the density field itself that appears in the integral
part of the martingale MtG but another local function of the configuration : g ( (0)).
The main problem consists therefore in using the particular characteristics of our
model to replace g ( (x)) by an ad hoc function of the density field in order to
“close" the equation.
Recall that we have parametrized the invariant and translation invariant mea-
sures of the zero range processes by the density :
E (0)
= :
Moreover, we denoted by () the expected value of the jump rate g ( (0)) :

E g((0))
= () :
The function plays a crucial role in the hydrodynamic limit. We shall present
a rigorous proof of the following intuitive argument. Consider a box of length "N
around a macroscopic point u 2 Td . Since the variation of the total number of
particles in this box is due to surface effects, the system quickly reaches equilib-
rium in [(u ")N; (u ")N ]d before the total number of particles has had time to
change in a discernible way. We would like therefore to replace, in the average,
X
1
(2N" + 1)d
g(s (y))
jy xj"N
by the expectation of the function g under the invariant measure of density
X
1
(2"N + 1)d
s (y)
jy xj"N
that is, from the definition of , by
0 1
X
@ 1
(2N" + 1)d
s ( y ) A :
jy xj"N
1. The law of large numbers 69
Such a replacement allows to close the equation since the density of particles
in a box of macroscopic length ("N above) is a function of the empirical measure.
In fact, if " () stands for the approximation of the identity defined by " () =
P
(2") d1f[ "; "]dg(), (2"N + 1) d jyj"N (y ) is equal to the integral of " with
respect to the empirical measure :
X (2N")d
1
(2"N + 1)d
(y ) =
(2"N + 1)d
< N ; " > :
jyj"N
From this heuristic argument we may guess the hydrodynamic equation of
nearest neighbor symmetric zero range processes. The macroscopic behavior of
the system should be described by the non linear heat equation :
(
@t = (1=2) (())
(0; ) = 0 () :
Once we performed the replacement of the local field N d x N G(x=N )
P
g((x)) by the density field, to conclude the proof of the hydrodynamic behavior,
we just need to repeat the argument presented in last chapter. There is just an
additional difficulty: we have to prove that the empirical measure converges to an
absolutely continuous measure with respect to the Lebesgue measure since nothing
a priori prevents particles to accumulate at one site creating a Dirac measure. This
phenomenon could be discarded for exclusion processes because we allowed at
most one particle per site.
Throughout this chapter we repeatedly apply results on the relative entropy
and the Dirichlet form presented in Appendix 1.
1. The law of large numbers
To state the main theorem of this chapter, we lack just a few notations and certain
hypotheses.
At Lemma 1.3, where we prove that all limit probability measures are concen-
trated on the subset of absolutely continuous measures with respect to the Lebesgue
measure, and at Lemma 3.2, where we replace cylinder fields by functions of the
empirical measure, we need the family of invariant measures f ; 0g to have
all exponential moments or, in an equivalent way, the partition function Z () to
be finite on R+ . We therefore suppose throughout this section that
(FEM) The partition function Z () introduced in Definition 2.3.1 is finite on R+ .
d
Throughout this chapter, for a fixed time T and for a probability N in N TN ,
d
we represent by PNN = PN the probability on the path space D([0; T ]; N N )
T
corresponding to the nearest neighbor symmetric zero range process with generator
LN , defined at (0.1) above, accelerated by N 2 and starting from the initial measure
N .
Theorem 1.1 Assume the jump rate g () to increase at least linearly : there exists
a positive constant a0 such that g (k ) a0 k for all k 0. Let 0 : Td ! R+ be an
integrable function with respect to the Lebesgue measure. Let N be a sequence of
d
probability measures on N TN associated to the profile 0 and for which there exist
positive constants K0 , K1 and such that the relative entropy of N with respect
to is bounded by K0 N d :
H (N j ) K0 N d (1:1)
and h X i
lim sup EN N d (x)2 K1 : (1:2)
N !1 x2TdN
Then, for every t T , for every continuous function G: Td ! R and for every
strictly positive,
2 Z 3
X
lim PN 4 d G(x=N )t (x) G (u)(t; u) du > 5 = 0;
1
N !1 N x2TdN Td
where (t; u) is the unique weak solution of the non linear heat equation
(
@t = (1=2) (()) ;
(1:3)
(0; ) = 0 () :
Remark 1.2 In the assumption regarding the relative entropy, the parameter
does not play any special role in the sense that if this condition is verified for ,
we can easily show that it is also satisfied for every other choice of the parameter.
Indeed, for each density > 0, the entropy inequality gives that
Z d
H (N j ) (1 + 1
)H (N j ) +
1
log
d d
for every > 0. Since the measures and are product, the second term of
the right hand side can be explicitly computed. It is not difficult to show that this
term is finite and of order N d for = (; ) small enough because by (2.3.9)
each measure has finite exponential moments.
On the other hand, this assumption on the relative entropy is naturally satisfied
by every product measure with slowly varying parameter associated to a bounded
profile.
Remark 1.3 We have seen in Chapter 4 that the proof of the hydrodynamic
behavior relies on an uniqueness theorem for weak solutions of the partial dif-
ferential equation that describes the macroscopic evolution of the system, equa-
tion (1.3) in the present context. We present such a result in Appendix A2.4 for
weak solutions belonging to L2 ([0; T ] Td ). It is only in the proof that all limit
points of the sequence QN (defined right below) are concentrated on trajectories
(t; du) = (t; u)du whose density is in L2 ([0; T ] Td) that the assumption on
the jump rate (g (k ) a0 k ) and assumption (1.2) are required.
There are, however, stronger uniqueness results. We shall prove in section 7,
for instance, that in dimension 1 there exists a unique solution satisfying the
energy estimate :((t; u)) 1=2 r((t; u)) belongs to L2 ([0; T ] Td). This energy
estimate is proved in section 7 assuming only (FEM) and the bound (1.1). Brezis
and Crandall (1979) present an alternative uniqueness result of weak solutions of
(1.3).
Remark 1.4 Assumption (1.2) is weakened in Remark 6.4.
Proof of Theorem 1.1. We fix once for all a time T > 0. As in the previous
chapter we denote by tN (du) the empirical measure defined in (4.0.2) and by QN N
or simply Q the measure on the trajectories space D([0; T ]; M+) associated to
N
the process tN starting from N .
To enable the reader to discern each step of the proof, we will state them
as separate lemmas. The step that consists in proving the relative compactness
of the sequence of probabilities QN being almost always technical, even if it is
instructive since it gives information about the limit trajectories, will systematically
be postponed to the end of the proof in the following chapters. In this chapter
however, we will start by this point because it provides a first opportunity to show
the power of the entropy inequality.
Lemma 1.5 The sequence of probabilities QN is relatively compact.
Proof. As we have already seen in Chapter 4, we have to prove the compactness

of the marginals at every fixed time t and a condition regarding the oscillations.
In the case of positive measures on a compact set, the condition on the com-
pactness of marginals at fixed times t is reduced to estimates on the total mass.
Since the total mass is conserved by the evolution, it is enough to check it at time
zero. We have therefore to prove that
n X o
lim lim sup N ; d
N x2Td (x) A = 0 : (1:4)
1
A!1 N !1
N
This equality holds because the sequence N is associated to the profile 0 that
we assumed integrable.
In our context the Aldous condition regarding the oscillations leads us to prove
that
lim lim sup sup
!0 N !1 2TT
2
3
X X
PN 4 d G(x=N ) (x) G(x=N ) (x) > 5
1 1
N x2TdN + d
N x2Td
= 0
N
for every function G of class C 2 and every > 0. Recall the definition of the
martingale MtG . To prove last equality it is enough to show that
2 Z 3
1 X
lim lim sup sup PN 4 (x=N )g (s (x)) ds > 5
+
N G
2N x2TdN
d
= 0
! N !1 2TT
0

and lim lim sup sup PN MG MG > = 0 :
!0 N !1 2TT +

Since G is of class C 2 and the function g increases at most linearly (cf.

Assumption (2.3.1)), the absolute value of the integral term is bounded by
Z X
C1 (g; G)
+
N d s (x) ds
x2TdN
that, by conservation of the total number of particles, is equal to
X
C1 (g; G)N d 0 (x) :
x2TdN
The stopping time having disappeared and since converges to 0 we are brought
back to the very same estimate (1.4) necessary to prove the relative compactness
of marginals at fixed times.
For the martingale term we use Chebychev’s inequality and the explicit formula
for the quadratic variation NtG to bound it above by
h i
2
E N MG+ MG 2
2Z 3
1 X
E N 4 [G(y=N ) G(x=N )] g (s (x)) ds5
+
2 2
2N d
= 2 2
jy xj
2 =1
3
X
C (g 2;NGd) EN 4 N1d ( x) 5
2
2
x2TdN
because the total number of particles is conserved by the evolution. From the
presence of an additional factor N d , we have only to show that the expectation
of the total density is bounded by a constant independent of N . The fact that the
sequence N is associated to an integrable or bounded profile cannot guaranty it. It
is in fact the assumption on the entropy that permits to prove that the expectation
of the total mass is bounded. Indeed, by the entropy inequality,
2 3
X h P i H (N j )
EN 4 1
d
N x2Td (x)5
1
N d log E e x x +
N d
( )
(1:5)
N
for every > 0. Since the measure is product and the entropy of N with
respect to was supposed bounded above by K0 N d , the right hand side of the
last expression is bounded by
1
log E
e + K :
(0)
0
By assumption (2.3.2), this expression is finite for all sufficiently small. The
proof of the relative compactness of the sequence QN is thus concluded.
Notice that assumption (FEM) is not necessary in Lemma 1.5. We just need
the existence of some exponential moments and this is guaranteed by hypothesis
(2.3.2).
After having established the relative compactness of the sequence QN it re-
mains to show the existence of at most one limit point. We shall do it by proving
some regularity properties of all possible limit points. We start, for instance, show-
ing that all limit points are concentrated on absolutely continuous measures. This
result is a second application of the entropy inequality.
Recall from Appendix 1 that the entropy with respect to some invariant ref-
erence state decreases in time. In particular, H (N StN j ) K0 N d for each
0 t T if StN stands for the semigroup associated to the generator LN defined
in (0.1) accelerated by N 2 . Next lemma states that any sequence of probability
measures N on M+ with entropy with respect to bounded by K0 N d and
that converges must converge to a probability measure concentrated on absolutely
continuous measures.
Lemma 1.6 Under the hypothesis (FEM) stated in the beginning of this chapter,
d
let N be a sequence of probability measures on N TN with entropy with respect to
bounded by K0 N d :
H (N j ) K0 N d :
d
Recall from (3.0.1) that we denote by N : N TN ! M+ the function that associates
to each configuration the positive measure obtained assigning mass N d to each
particle. Let RN be the probability measure N ( N ) 1 on M+ defined by
RN [A]

N ; N () 2 A
=

for every Borel subset A of M+ . Then, all limit points R of the sequence RN
are concentrated on absolutely continuous measures with respect to the Lebesgue
measure :
R f; (du) = (u)dug = 1 :
Proof. The strategy consists in obtaining a positive lower semi–continuous func-
tional I : M+ ! R+ such that
(a) lim supN !1 ERN [I ( )] < 1.
(b) I ( ) = 1 if (du) is not absolutely continuous with respect to the Lebesgue
measure.
Indeed, we would then have that
ER [I ()] < 1

for every limit point R because I was supposed lower semi–continuous. In par-
ticular by property (b), R is concentrated on absolutely continuous measures.
All the problem is therefore to find such an appropriate functional I . Consider
for a while a continuous and bounded function J : M+ ! R+ . By the entropy
inequality,
h N dJ i
ERN [J ()] N d H (N j ) +
N d log ER
1
e ( )
:
By hypothesis, the first expression on the right hand side of this inequality is
bounded above by K0 . The second one, since J is bounded, by the Laplace–
Varadhan theorem (cf. Theorem A2.3.1), converges to
sup [J ( ) I0 ()]
2M+
if I0 represents the large deviations rate function for the random measure un-
der R . This rate function can be easily computed. In order to define it, let
M : R ! [0; 1] be the Laplace transform of (0) under :
M () E [e(0)] Z (e ( )) : (1:6)
= =
Z (( ))
The function log M is convex as well as its Legendre transform h defined by
h() = sup f log M ()g

= log (()) log
Z (())
Z (( ))
The large deviations rate function I0 : M+ ! [0; 1] is equal to
Z Z
I0 () = sup f (u)(du) log M (f (u)) du :
f 2C (Td) Td Td
This supremum can be computed. Under the assumption (FEM) stated just before
Theorem 1.1, we have that
8Z
< h((u)) du if (du) = (u)du ;
I0 () =
: T1d otherwise.
In particular property (b) is satisfied by the rate function I0 . Moreover, it follows
from the variational formula for I0 that this functional is lower semi–continuous.
We have now to check that our candidate has finite expectation with respect to all
limit points of RN .
The variational representation of I0 shows that I0 is the increasing limit of the

bounded and continuous functionals Jk defined by
Z
Jk () = max
jk
< ; fj > log M (fj (u)) du ^k
1 Td
if ffk ; k 1g stands for a dense sequence in C (Td ) with f1 being the function
identically equal to 0 to ensure the positiveness of Jk .
Using the entropy inequality for each functional Jk , we obtain by the Laplace–
Varadhan theorem that
lim sup ERN [Jk ( )] K0

N !1
because Jk I0 for each k 1. In particular for all limit point R of the sequence
R N ,
ER [Jk ()] K0 :
By the monotone convergence theorem and by the variational representation of
the rate function I0 ,
ER [I0 ()] = lim E [J ()]

k!1 R k
K0 :
Remark 1.7 We have in fact established a slightly stronger result. Under the hy-
potheses of the previous lemma, if I0 represents the large deviations rate function,
all limit points R of the sequence RN are such that
Z
ER [I0 ()] = ER h((u)) du K0 :
T
d
We shall take advantage of this property later to prove that all limit points of the
sequence QN are concentrated on weak solutions of the non–linear heat equation
(1.3).
Remark 1.8 The previous result applied to the marginal at time t of the measure
QN shows that all limit points Q of the sequence (QN ) are such that
EQ [I0 (t )] K0

for all 0 t T . Since the rate function I0 is positive, by Fubini’s lemma,
hZ T i ZT h i
EQ dt I0 (t ) = dt EQ I0 (t ) K0 T :
0 0
In particular, changing if necessary t (du) in a time set of measure 0, all limit

points Q are concentrated on absolutely continuous trajectories :
Q f; t (du) = t (u)du ; 0 t T g = 1:

Remark 1.9 Notice that the symmetry assumption on the transition probabilities
does not play any role in the proof of Lemma 1.6. In particular, under assump-
tion (FEM), in the asymmetric case, limit points are concentrated on absolutely
continuous trajectories provided the entropy at time 0 is bounded by K0 N d .
We have just proved that all limit points are concentrated on absolutely con-
tinuous trajectories with integrable densities. By Theorem A2.4.4, there exists a
unique weak solution of the Cauchy’s problem (1.3) in L2 ([0; T ] Td ). To con-
clude the proof of Theorem 1.1, it remains therefore to show that all limit points
are concentrated on weak solutions of (1.3) in L2 ([0; T ] Td ).
We prove in this section that all limit points of the sequence fQN ; N 1g
are concentrated on weak solutions of (1.3) and in section 6 that all limit points
are concentrated on trajectories (t; du) = (t; u)du whose density belongs to
L2 ([0; T ] Td).
We now return to the strategy adopted to prove the hydrodynamic behavior of
symmetric simple exclusion processes. For each fixed smooth function G: [0; T ]
Td ! R of class C 1;2 , MtG;N = MtG defined by
X X
MtG =
1
Nd G(t; x=N ) t (x) N1d G(0; x=N ) 0(x)
x2TdN x2TdN
Zt Xn o
1
N d x2Td @s G(s; x=N )s (x) (1=2)N G(s; x=N )g (s (x)) ds
0
N
is a martingale with quadratic variation < M G >t equal to
X Z th i
< M G >t = G(s; y=N ) G(s; x=N ) g(s (x)) ds :
1 2
2N 2d 2
jx yj=1 0
Since the jump rate g () is at most linear (g (k ) g k ) and the total number
of particles is conserved, by estimate (1.5), the expected value of the quadratic
variation < M G >t vanishes as N " 1. In particular, by Doob’s inequality, for
each > 0,
N Zt
N !1
lim PN sup
tT
< t ; Gt > < N ; G > ds < sN ; @s Gs >
0 0
Zt
0
1 X
0
ds (1=2) N d N G(s; x=N )g(s(x)) > = 0 :

0
x2TdN
(1:7)
In contrast with symmetric simple exclusion processes, the integral term
Zt X
ds (1=2)N d N G(s; x=N )g(s (x))
0
x2TdN
is not a function of the empirical measure. This time the equation is not closed
anymore and a new argument is required.
The next lemma allows the replacement of the local function g ( (x)) by a
function of the empirical density of particles in a small macroscopic box. More
precisely, it states that the difference
Zt X n X o
ds N1d H (s; x=N ) g(s (x)) 1
(2"N + 1)d
s (y)
0
x2TdN jy xj"N
vanishes in probability as N " 1 and then " # 0 for every continuous function
H . Notice that the argument of () in the previous integral is a function of the
empirical measure. Indeed, for each " > 0, denote by " the approximation of the
identity
" () = (2") d1f[ "; "]dg() : (1:8)
We have that
X (2"N )d
1
(2["N ] + 1)d
(y ) =
(2["N ] + 1)d
< N ; " ( x=N ) >
jy xj"N (1:9)
=: CN;" ( N " )(x=N ) :
To keep notation simple, for each positive integer ` and d-dimensional integer
x, denote by ` (x) the empirical density of particles in a box of length 2` + 1
centered at x :
X
` (x) = (2` +1 1)d (y ) : (1:10)
jy xj`
Lemma 1.10 (Replacement lemma). For every > 0,

hZ T X i
lim sup lim sup PN
1
Nd x V"N (s ) ds = 0:
"!0 N !1 0
x2TdN
1 X
(2` + 1)d g((y)) `(0) :
where,
V` ( ) =
jyj`
To keep notation simple, in the previous statement and hereafter we are writing
"N for ["N ], the integer part of "N . Lemma 1.10 is the main step in the proof
of Theorem 1.1. Its proof is postponed to section 3.
Notice that the constant introduced in (1.9) is equal to 1 + O(N 1 ). Since
by Corollary 2.3.6 is Lipschitz continuous and since by (1.5) the total density
of particles has bounded expectation, in last formula we may replace s"N (x) by
(sN " )(x=N ). Therefore, from Lemma 1.10 and equation (1.7), we have that
" ZT
lim sup lim sup Q
"!0 N !1
N < T ; GT > < 0; G0 > 0 < s ; @sGs > ds
ZT 1 X 3

(1=2)G(s; x=N ) (s " )(x=N ) ds 5 = 0 :
0 N
d
x2TdN
We have also replaced the discrete Laplacian N by the continuous one. This
replacement is allowed because increases at most linearly and the expected
value of the total density of particles is bounded in virtue of (1.5).
By the same reasons and because G is of class C 1;2 , we may replace the sum
X
N d G(s; x=N )((sN " )(x=N ))
x
by the integral Z
du G(s; u)((sN " )(u)) :
Td
By the dominated convergence theorem, for each "> 0, the function that
associates to a trajectory the expression
ZT
< T ; GT > < 0 ; G 0 > < s ; @s Gs > ds
ZT Z 0
ds du (1=2)G(s; u) ((s " )(u))

0 Td
is continuous. In particular, all limit point Q of the sequence QN are such that
" ZT

lim sup Q < T ; GT > < 0 ; G 0 > < s ; @s Gs > ds
"!
0
ZT Z #
0
ds d du (1=2)G(s; u) ((s " )(u)) = 0:

0 T
To show that all limit points Q are concentrated on weak solutions of equation
(1.3), it remains to prove that we may replace the convolution s " by s as
"#0:
lim sup
"!
"0 Z T Z n o #
Q ds Td du G(s; u) ((s ")(u)) ((s; u))
0
= 0
for all > 0.

2. Entropy production 79
By Remark 1.8, s has a density with respect to the Lebesgue measure. There-
fore, Z
(s " )(u) = (2") d (s; v) dv
u ";u+"]d
[
for all 0 s T . Q is concentrated on integrable trajectories . On the other

hand, for each integrable function f : Td ! R, f " converges in L1 (Td) to f as
" # 0. In particular, since is Lipschitz, the random variable
ZT Z
ds du G(s; u) ((s " )(u))
0 Td
converges Q almost surely to
ZT Z
ds du G(s; u) ((s; u)) :
0 Td
Finally, letting # 0, we obtain that Q is concentrated on weak solutions of
the non linear parabolic equation (1.3) :
" ZT
Q < T ; GT > < ; G > < s ; @s Gs > ds
0 0
ZT Z 0
#
ds d du (1=2)G(s; u) ((s; u)) =0 = 1:
0 T
2. Entropy production
Recall from Appendix 1 the definition of the relative entropy and the Dirichlet
form of the state of the process with respect to some reference invariant measure.
Hereafter, to keep terminology simple, we refer to these two concepts as the
entropy and the Dirichlet form.
This section is devoted to the examination of the entropy time evolution. We
prove not only that the entropy decreases in time but that its evolution is closely
connected to the time evolution of the Dirichlet form. Since the entropy decrease
is not a particularity of reversible processes, in this section, we place ourselves in
a more general context to include asymmetric evolutions.
Denote by LN the generator of a zero range process as defined in section 2.3.
Consider a sequence (N )N 1 of probability measures on the configuration space
d
N TN and denote by StN the semigroup associated to the generator LN accelerated
by (N ). Here (N ) = N 2 in case the drift of each elementary particle vanishes
P
( x xp(x) = 0) and N otherwise.
Let ftN be the density of N StN with respect to a reference invariant measure
. ftN is the solution of the forward Kolmogorov equation
(
@t ftN = (N )LN ftN
f0N = dN =d ;
where LN represents the adjoint operator of LN in L2 ( ).
We have seen in Appendix 1 that the entropy of N StN with respect to is
given by Z
H (N StN j ) = ftN log ftN d :
To keep notation simple, we shall abbreviate it by HN (ftN ).
The time derivative of the entropy, known as the entropy production, is there-
fore equal to
Z Z Z
@t ftN log ftN d = (N ) log ftN LN ftN d + (N )
LN ftN d :
(2:1)
Since is invariant, the second term on the right hand side vanishes. Since LN
is the adjoint of LN in L2 ( ), the first term can be rewritten as
Z
(N ) ftN LN log ftN d :
From the elementary inequality
a log (b=a) 2 a ( b a)
p p p
that holds for all positives a, b, we can bound above the last integral by
Zq q Zq q
2 (N ) ftN LN ftN d = 2 (N ) ftN Lsym
N ftN d :
In this formula, Lsym
N stands for the symmetric part of the generator LN :
Lsym
N = 2 1
(LN + LN ) :
Recall the definition of the Dirichlet form introduced in section A1.10. We

have bounded above the entropy production by the Dirichlet form associated to
the symmetric part of generator :
Zq q
@t HN (ftN ) 2(N ) ftN Lsym
N ftN d = 2 (N ) DN (ftN ) :
Integrating in time we get

Zt
HN (ftN ) + 2 (N ) DN (frN ) dr HN (f0N ) :
0
In particular, if the entropy at time 0 is bounded by K0 N d, by convexity of the

entropy and of the Dirichlet form,
3. Proof of the replacement lemma 81
Zt Zt d
HN 1t frN dr K0 N d ; DN 1t frN dr 2Kt 0N
(N )
(2:2)
0 0
To keep terminology simple, up to the end of this section we shall refer to

P
interacting particle systems whose elementary particles evolve, when interaction is
suppressed, according to mean-zero random walks ( x xp(x) = 0) as mean-zero
processes. All other systems are called asymmetric processes. In Chapter 1 we
have seen that in order to observe interesting hydrodynamic behavior of mean-
zero processes we have to rescale time by N 2 and by N otherwise. In particular,
in the mean-zero case we proved a bound on the Dirichlet form of order CN d 2
and of order CN d 1 in the asymmetric case.
The proof of the replacement lemma relies exclusively on the above two es-
timates. In particular, the symmetry of the transition probabilities, assumed to
clarify the exposition, is not required. All proofs extend therefore to mean-zero
asymmetric zero range processes.
The proof of the replacement lemma is divided in two steps. We first prove
that we are allowed to replace large microscopic spatial averages of g (), that is,
averages over cubes of length `, a parameter independent of N and that increases
to infinity after N , by ( ` ()). This statement is just the one of Lemma 1.10
with "N replaced by ` and letting ` " 1 instead of " # 0. This result stated
in Lemma 3.1, called the one block estimate, requires a bound on the Dirichlet
form of order K0 N d 1 . The proof presented in next section applies therefore to
asymmetric processes.
In contrast, the second step which consists in showing that the particles density
over large microscopic boxes is close to the particles density over small macro-
scopic boxes strongly relies on an estimate of order CN d 2 of the Dirichlet form.
It applies therefore only to mean-zero processes rescaled by N 2 .
3. Proof of the replacement lemma
The proof is divided in several steps. We first reduce the dynamical problem
involving the stochastic evolution to a static one by means of the estimates obtained
in last section. Let N (T ) be the Cesaro mean of N StN :
ZT
N (T ) : = T1 N StN dt
0
and f¯TN the Radon–Nikodym density of N (T ) with respect to . In last section

we showed that the entropy of N (T ) with respect to is bounded by K0 N d
and that its Dirichlet form is bounded by K0 N d 2 (2T ) 1. We claim that to prove
the replacement lemma it is enough to show that
Z X
N d x2Td x V"N ()f () (d) (3:1)
1
lim sup lim sup sup 0
"!0 N !1 DN (f )C0 N d 2
HN (f )C0 N d N
for every positive constant C0 . In this formula, for a positive integer `, V` repre-
sents the cylinder function
1 X
V` () : = g ( (y )) (0)
`
(2` + 1)d
jyj`
and x translation by x.
Indeed, by Markov inequality,
2 Z T X 1 3
X
P N 4
N
1
d d g(s (y)) s"N (x) ds 5
d (2N" + 1)
0
x2TN jy xj"N
is less than or equal to
2 Z T X 1 3
X
1
E N 4 1
Nd d g(s (y)) s"N (x) ds5 :
0 d (2N" + 1)
x2TN jy xj"N
With the notation just introduced, we may rewrite this expectation as
Z X
1T 1
Nd x V"N ()f¯TN () (d)
x2TdN
and (3.1) is proved.
The proof of inequality (3.1) is divided in two steps. We first show that we
may replace the spatial average of g over large microscopic boxes, that is of length
` independent of N and that increases to infinity after N , by (` ()). It consists
therefore in proving inequality (3.1) with ` in place of "N . This is the content of
Lemma 3.1.
In the proof of this result we need weaker assumptions than the one formulated
in the beginning of this chapter. Since this replacement of spatial averages of
cylinder functions over large microscopic hypercubes will be required several
times in next chapters, we prove it here in the greatest possible generality. For
this reason, consider the hypothesis of sub–linear growth of the jump rate g () :
(SLG) A zero range process is said to have a sub–linear jump rate if
g (k ) = 0:
k!1 k
lim sup
Lemma 3.1 (One block estimate). Under hypothesis (SLG) or (FEM), for every
finite constant C0 ,
Z X
N d x2Td (x V` )()f () (d) = 0:
1
lim sup lim sup sup
`!1 N !1 DN (f )C0 N d 2
HN (f )C0 N d N
3. Proof of the replacement lemma 83
The second step in the proof of inequality (3.1) consists in showing that the
particles density over large microscopic boxes are close to the particles density
over small macroscopic boxes :
Lemma 3.2 (Two blocks estimate). Under hypothesis (FEM), for every finite
constant C0 ,
lim sup lim sup lim sup sup sup
`!1 "!0 N !1 DN (f )C0 N d 2 jyj"N
HN (f )C0 N d
Z X
N d ` (x + y) N" (x) f () (d) = 0:
x2TdN
Before turning to the proof of Lemmas 3.1 and 3.2, we check that inequality
(3.1) follows from these two statements. Add and subtract the expression
X n X o
1
(2"N + 1)d
1
(2` + 1)d
g((z )) (` (y))
jyjN" jz yj`
inside the absolute value that appears in the definition of V"N . We shall estimate
three terms separately.
The first and also the simplest one is equal to
Z X
N d x2Td x V"N;`()f () (d) ;
1

N
1 o
where
X n X
V"N;` () =
(2"N + 1)d
g((y)) 1
(2` + 1)d
g((z )) :
jyjN" jz yj`
This expression is bounded above by
C1 (d)` Z 1 X g((x))f () (d) ;
N" N d x2Td
N
where C1 (d) is a constant that depends only on dimension. Since the jump rate g
increases at most linearly and HN (f ) C0 N d , by the entropy inequality, the last
integral is bounded above by

gC1 (d) C + log E h exp n( `=N")(0)oi

(3:2)
0
for every > 0 because is a product measure. For N sufficiently large the
expectation that appears in this expression is finite in virtue of (2.1.2). Moreover,
this sum vanishes as N " 1 and then " 1.
The second term is bounded by

Z X 1 X `

f () (d) :
1

N d x2Td x (2` + 1)d jyj` g ( ( y )) ( (0))
N
According to Lemma 3.1, it vanishes as N " 1 and ` " 1.
Finally, the third term is bounded above by
Z X
sup N d (` (x + y)) (N" (x)) f () (d)
jyj"N x2TdN
Z X `
g sup N d (x + y) N" (x) f () (d)
jyj"N x2TdN
because, by Corollary 2.3.6, is Lipschitz. This expression is estimated in Lemma
3.2.
4. The one block estimate
We prove in this section Lemma 3.1. To detach the main ideas, we divide the
proof in several steps. We first assume hypothesis (SLG), the adjustments needed
for the case (FEM) will be presented at the end of the proof.
Step 1 : Cut off of large densities. We start with a technical lemma. It allows to
introduce an indicator function that later will guarantee that some set of probability
measures is compact for the weak topology.
Lemma 4.1 For every positive constant C0 , there exists a finite constant C3 =
C3 (C0 ; ) such that for every integer N ,
Z X
sup
1
Nd (x)f () (d) C3 :
HN (f )C0 N d x2TdN
Proof. By the entropy inequality, since is a translation invariant product

measure, for every positive , the supremum is bounded above by
1 C0 +

E e (0) :
1
log
By (2.1.2), the Laplace transform of (0) under is finite for sufficiently
small.
This result permits to restrict the integral that appears in the statement of
Lemma 3.1 to configurations with bounded particles density over large microscopic
boxes. Indeed, from the previous result, to prove Lemma 3.1, it is enough to show
that
4. The one block estimate 85
lim sup lim sup sup

`!1 N !1 DN (f )C0 N d
1 X
Z
2
d (x V` )( ) a ` (x) f ( ) (d )

N x2Td 0
N
for every a > 0. For each b > 0, hypothesis (SLG) guarantees the existence of a
constant C4 (b) such that
g(k) C4 (b) + bk :
Replacing k by (0) and taking expectations with respect to , we get a bound
for () :
() C4 (b) + b :
In particular, the cylinder function V` is bounded above by
C4 (a=2) + (a=2)`(0) :
Therefore, V` ( ) a ` (0) is negative as soon as ` (0) 2C4 (a=2)a 1 =: C5 (a) =
C5 . We can thus restrict last integral to configurations satisfying the reversed
inequality. In conclusion, to prove the one block estimate, it is enough to show
that
lim sup lim sup sup
`!1 N !1 DN (f )C0 N d 2
XZ (4:1)
(x V` )( )1f ` (x) C5 g f ( ) (d ) 0
1
d
N x2Td
N
for every positive constants C0 and C5 .
Step2 : Reduction to microscopic cubes. Notice that V` ( )1f ` (0) C5 g de-
pends on the configuration only through f (x); jxj `g. The second step in
the proof consists in taking advantage of this fact to project the density f over a
configuration space that does not depend on the scale parameter N .
Since the measure is translation invariant, we can rewrite last sum as
Z X
V` ()1f`(0)) C5 g N d x2Td x f () (d)
1

Z N
= V` ()1f`(0) C5 g f¯() (d) ;
where f¯ stands for the space average of all translations of f :
X
f¯() =
Nd
1
x f () :
x2TdN
Before proceeding, we introduce some notation. For a fixed positive integer `
we represent by ` a cube of length 2` + 1 centered at the origin :
` = f `; : : :; `gd ; (4:2)
by X ` the configuration space N ` , by the Greek letter the configurations of X `

d
and by ` the product measure restricted to X ` . For a density f : N TN !
R+ , f` stands for the conditional expectation of f with respect to the -algebra
generated by f (z ); z 2 ` g, that is, obtained by integrating all coordinates
outside this hypercube :
Z
f`( ) = ` 1( ) 1f ; (z ) = (z ); z 2 `gf () (d) for 2 X` : (4:3)

Since V` ( )1f ` (0) C5 g depends on the configuration only through the
occupation variables f (x); x 2 ` g, in last integral we can replace f¯ by f¯` . In
particular, we may rewrite inequality (4.1) as
Z
lim sup lim sup sup V` ( )1f `(0) C5 g f¯`( )` (d ) 0 :
`!1 N !1 DN (f )C0 N d 2
Step 3 : Estimates on the Dirichlet form of f¯` . The third step consists in obtaining
information concerning the density f¯` from the estimate on the Dirichlet form of
f . For a fixed pair of neighbor sites x, y, denote by Lx;y the piece of the generator
corresponding to jumps across the bond with ends x and y :
(Lx;y f )( ) = (1=2)g ( (x))ff ( x;y) f ( )g + (1=2)g ( (y ))ff ( y;x) f ( )g :

Denote furthermore by Ix;y (f ) the piece of the Dirichlet form corresponding to
jumps over the bond fx; y g :
p p
Ix;y (f ) < Lx;y f; f >
=
Z p p
= (1=2) g ( (x))f f ( x;y ) f ()g2 (d) :
A simple change of variables ( = x;y ) shows that we may interchange x and
y in the previous integral. With this notation, the Dirichlet form DN (f ) may be
written as X
DN ( f ) = Ix;y (f ) ;
jx yj=1
where summation is carried over all non oriented pairs of neighbors. By non
oriented we understand that each bond fx; y g appears only once in the previous
sum.
Keep in mind that the Dirichlet form is translation invariant :
DN (x f ) = DN (f )
for every x in Zd because it is defined through a sum over all bonds of TdN . In
particular, since the Dirichlet form is also convex, we have that
X X
DN (f¯) = DN N d x f N d DN (x f ) = DN (f ) :
x2TdN x2TdN
Taking advantage again of the convexity of the Dirichlet form, we prove now
that the Dirichlet form restricted to bonds in ` of f¯` is bounded by DN (f ).
For a positive integer `, denote by D` the Dirichlet form defined on all densities
h: X ` ! R+ by X
D` (h) = ` (h);
Iy;z
jy zj=1
y;z2`
where, Z hp p i
` (h) = (1=2) g ( (y ))
Iy;z h( y;z ) h( ) ` (d ) :
2
(4:4)
Like before summation is carried here over all pairs of non oriented bonds of ` .
With this notation, since the Dirichlet form is convex and since conditional
expectation is an average,
` (f¯` ) Iy;z (f¯)
Iy;z (4:5)
for every pair of neighbors y , z in ` . This inequality can also be deduced applying
Schwarz inequality to the explicit expression for Iỳ;z (f¯` ) presented in (4.4) and
keeping in mind the definition of f¯` given in (4.3)
By inequality (4.5), we have that
X
D` (f¯` ) Iy;z (f¯):
jz yj=1
z;y2`
On the other hand, by translation invariance of f¯, Iz;y (f¯) = Iz+x;y+x (f¯) for every
d-dimensional integer x. Thus,
X X
d
Iy;z (f¯) = (2` + 1)d 1
(2`) I0;ej (f¯) = (2` + 1)d 1(2`)N dDN (f¯) :
jz yj=1 j =1
z;y2`
Therefore, for every density f with Dirichlet form bounded by C0 N d 2

we have
that
D` (f¯` ) C0 (2` + 1)dN 2 =: C6 (C0 ; `)N 2 : (4:6)
Notice that we bounded the Dirichlet form of f¯` by a constant vanishing as N " 1.
In conclusion, to prove Lemma 3.1, it remains to show that
Z
lim sup lim sup sup V` ( )1f `(0) C5 g f ( )` (d ) 0 ;
`!1 N !1 D` (f )C6 (C0 ;`)N 2
(4:7)
where this time the supremum is carried over all densities with respect to ` .
Notice that the scaling parameter N appears now only on the bound of the Dirichlet
form.
Step 4 : The Limit as N " 1. The fourth step consists to examine the behavior
of last expression as N " 1. Relying on the lower semicontinuity of the Dirichlet
form and on the relative compactness provided by the indicator function, we bound
last expression by another expression in which the supremum over all densities is
replaced by one over all densities with Dirichlet form equal to 0.
From the presence of the indicator function and since V` is positive, we can
restrict last supremum to densities concentrated on the set f ; ` (0) C5 g or on
Z
the set of all densities such that
` (0) f ( )` (d ) C5 :
X`
This subset of M1;+ (X ` ) is compact for the weak topology. This explains the
reason we introduced the indicator function 1f ` (0) C5 g in the beginning of
the proof.
Since this set is compact, for each fixed N , there exists a density f N with
Dirichlet form bounded by C6 N 2 that reaches the supremum. Consider now a
subsequence Nk such that
Z
lim
k!1
V` ( )1f `(0) C5 g f Nk ( )` (d )
Z
= lim sup V` ( )1f `(0) C5 g f N ( )` (d ) :
N !1
To keep notation simple, assume, without loss of generality, that the sequences
Nk and N coincide. By compactness, we can find a convergent subsequence f Nk .
Denote by f 1 the weak limit. Since the Dirichlet form is lower semicontinuous,
D ` (f 1 ) = 0 :
Moreover, by weak continuity,
Z
` (0)f 1( )` (d ) C5 and
Z
k!1
lim V` ( )1f `(0) C5 g f Nk ( )` (d )
Z
= V` ( )1f `(0) C5 g f 1( )` (d ) :
In conclusion, expression (4.7) is bounded above by
Z
lim sup sup V` ( )1f `(0) C5 gf ( )` (d ) :
`!1 D` (f )=0
Step 5 : Decomposition along hyperplanes with a fixed number of particles. A
probability density with Dirichlet form equal to 0 is constant on each hyperplane
with a fixed total number of particles. It is convenient therefore to decompose
each density f along these hyperplanes with particles density bounded by C5 .
For each integer j 0, denote by `;j the measure ` conditioned to the
P
hyperplane f ; z2` (z ) = j g :
X ` "` ( X )# 1
`;j () = ` (z ) = j = () ; (z ) = j

z2` z2`
Notice that this measure does not depend on the parameter .
With this notation we have that
Z [(2`+1)d C5 ] X Z
V` ( )1f `(0) C5 gf ( )` (d ) = Cj (f ) V` ( ) `;j (d ) ;
j =0
where Z X
Cj ( f ) = 1f (z ) = j gf ( )` (d ) :
z2`
P
Since j Cj (f ) = 1, to conclude the proof of the one block estimate we
0
Z
have to show that
lim sup sup V` ( ) `;j (d ) 0 :

`!1 j [(2`+1)d C5 ]
Step 6 : An application of the local central limit theorem. The previous inequal-
ity follows from the equivalence of ensembles presented in Appendix 2. Since the
measure `;j is concentrated on configurations with j particles, the last integral is
equal to
Z 1 X h i `;j
g( (x)) Ej= ` d g( (0)) (d ) :
(2` + 1)d
jxj`
(2 +1)
Fix a positive integer k , that shall increase to infinity after `. Decompose the set
` in cubes of length 2k + 1 : Consider the set A = f(2k + 1)x; x 2 Zdg \ ` k
and enumerate its elements : A = fx1 ; : : : ; xq g in such a way that jxi j jxj j for
i j . For 1 i q, let Bi = xi + k . Notice that Bi \ Bj = for i 6= j and that
[1iq Bi ` . Let B0 = ` [1iq Bi . By construction, jB0j Ck`d 1 for
some universal constant C . The previous integral is bounded above by
Xq
jBi j Z 1 X g( (x)) E h i `;j
jBi j x2B j= ` d g ( (0)) (d ) :
i j` j
(2 +1)
=0 i
Since jB0 j Ck`d 1 , since g (m) g m and since under `;j the variables (x)
have mean j=(2` + 1)d , this sum is equal to
jk j Xq Z 1 X h i `;j

j` j i=1 jk j x2B g( (x)) Ej= ` d g( (0)) (d ) + O(k=`) :
i
(2 +1)
Since the distribution of the vectors f (z ); z 2 Bi g does not depend on i, the

previous sum is equal to
Z 1 X h i
g x E g
`;j (d ) + O k=` :
(2k + 1)d
( ( )) j= ` d ( (0))
jxjk
(2 +1)
By the equivalence of ensembles (cf. Corollary A2.1.7), as ` " 1 and j=(2` +

1)d ! , last integral converges to
Z 1 X h i
g( (x)) E g( (0)) (d ) :
(2k + 1)d
jxjk
This convergence is uniform in on every compact subset of R+ . On the other
hand, as k " 1, by the law of large numbers, this integral converges to 0 uniformly
on every compact subset of R+ . The proof of Lemma 3.1 when the jump rate
satisfies assumption (SLG) is thus concluded.
It remains to consider the case (FEM). Notice in this respect that we used
assumption (SLG) only to justify the introduction of the indicator function in the
beginning of the proof. Next lemma states that this indicator function can also be
introduced under assumption (FEM).
Lemma 4.2 Under hypothesis (FEM),
lim sup lim sup lim sup sup

A!1 `!1 N !1 HN (f )C0 N d
Z 1 X
` `
N d x2Td (x)1f (x) Ag f () (d) = 0 :

N
Proof. By the entropy inequality, the integral in the statement of the lemma is
bounded above by
Z 8 9
< X ` =
C0 exp (x)1f ` (x) Ag (d )
1
: x2Td ;
1
+
N d log
N
for each > 0, because HN (f ) C0 N d by assumption.
With respect to the product measure , the random variables ` (x) and ` (y )
are independent if jx y j > 2`. Recall from (4.2) the definition of the hypercube
` . For each x in ` , denote by
x the set of sites z in TdN equal to x modulo
2` + 1 :
x = fz 2 TdN ; z x 2 (2` + 1)Zdg. With this notation,
X XX
A (`(x)) = A (` (y)) :
x2TdN x2` y2
x
In this formula, to keep notation simple, we abbreviated ` (x)1f ` (x) Ag by
A (` (x)).
Notice that for each fixed x, f A ( ` (y )); y 2

x g are independent random
variables with respect to . Therefore, by Hölder inequality and the translation
invariance of , last integral is bounded above by
Z n o
C0 1
+
1
(2` + 1)d log exp (2` + 1)d` (0)1f`(0) Ag (d) :
By Cauchy–Schwarz inequality and Chebychev exponential inequality, the last
integral is bounded by
P
1 + E 1f`(0) Ag e jxj` (x)
1 +
n ` P
; (0) A E e jxj` x
o =
2 ( ) 1 2
n d o ;
1 + exp (1=2)(2` + 1) A log M (1) log M (2 )
where M is defined in (1.6). Since log(1 + u) u for all u > 0, up to this point
we proved that the supremum appearing in the statement of the lemma is bounded
above by
n A o ;
C0 (1=2)(2` + 1)d log M (1) log M (2 )
1 1
+
(2` + 1)d exp
for every > 0. This expression is finite because by assumption (FEM),
M () < 1 for every in R. Moreover, the right hand term vanishes in the
limit as ` " 1 for all A large enough. In particular, we may choose = (A) so
that limA!1 (A) = 1 and the right hand term vanishes as ` " 1 for every A
(take for instance A=2 = log M (2 )).
Going back to the proof of the one block estimate, the reader should notice
that the bound on the Dirichlet form was used only in (4.6) to estimate D` (f¯` ) by
a constant C (`; N ) that vanishes as N " 1. In particular the estimate DN (f )
C0 N d 2 is not crucial. In fact any bound on the Dirichlet form of order o(N d)
would be enough to prove the lemma. Such a bound on the Dirichlet form, of
order O(N d 1 ), was obtained in section 2 for asymmetric zero range processes.
In particular, we have the following result.
Lemma 4.3 Let LN be the generator of an asymmetric zero range process satis-
fying hypothesis (SLG) or (FEM), let N be a sequence if initial measures with
entropy bounded by C0 N d :
H (N j ) C0 N d
d
and let PN be the probability on the paths space D([0; T ]; NTN ) corresponding
to the Markov process with generator LN accelerated by N and starting from N .
For every > 0 and 0 < t T ,
lim sup lim sup

`!12 N !1 3
Zt X 1 X
PN 4 d
1
(2` + 1)d g(s (y)) s` (x) ds 5 = 0:
x2TdN
N 0 jy xj`
5. The two blocks estimate
The proof of the two blocks estimate follows closely the proof of the one block
estimate. We thus keep all notation introduced in the previous section and leave
some details to the reader. The unique novelty appears in the proof of an estimate
for the restricted Dirichlet form of the conditional expectation of f .
The first step in the proof of the two blocks estimate consists in replacing the
density average over a small macroscopic box by an average of densities average
over large microscopic boxes. More precisely, for every N sufficiently large, the
integral appearing in the statement of Lemma 3.2 is bounded above by
Z X X `
1 1
N d x2Td (2N" + 1)d jzjN" (x + y) ` (x + z ) f () (d)
N
2`<jy z j
` Z 1 X
+ C1 (d) N" N d (x)f () (d) ;
x2TdN
where C1 (d) is a constant that depends only on the dimension. Notice that summa-
tion over z in the first line is carried over all d-dimensional integers at a distance
at least 2` from y . All other terms are included in the second line. In this way
the averages ` (x + y ) and ` (x + z ) are performed over disjoint sets of sites. The
entropy inequality shows that the limit, as N " 1, of the second line vanishes
(cf. (3.2)). It remains therefore to estimate the first line.
The second step of the proof consists in introducing an indicator function to
avoid possible large values of particles density. In fact, by Lemma 4.2, in order
to prove Lemma 3.2, it is enough to show that
lim sup lim sup lim sup sup sup N d
"! N !1
Z`!1
X `
0

`<jyj N"
DN (f )C0 N d 2 2 2
(x) ` (x + y) 1f`(x) _ ` (x + y) Ag f () (d) 0

x2TdN
(5:1)
for every A > 0.
With the notation introduced in the proof of Lemma 3.1, this integral can be
rewritten as
Z
`(0) `(y) 1f`(0) _ `(y) Ag f¯() (d) ;
5. The two blocks estimate 93
where f¯ stands for the average of all space translations of f . ` (0) and ` (y )
depend on the configuration only through the occupation variables (x) in the
set
y;` := f `; : : :; `gd [ y + f `; : : : ; `gd :

We can thus replace in the last integral f¯ by its conditional expectation with
respect to the -algebra generated by f (z ); z 2 y;` g.
Before proceeding we introduce some notation. For a positive integer `, we
represent by X 2;` the configuration space N ` N ` , by the couple = (1 ; 2 )
configurations of X 2;` and by 2;` the product measure restricted to X 2;` . For a
d
density f : N TN ! R+ , fy;` stands for the conditional expectation of f with respect
to the -algebra generated by f (z ); z 2 y;` g, that is, obtained integrating all
coordinates outside y;` :
Z
fy;`(1 ; 2 ) 1f ; (z ) = (z ); z 2 y;`g f () (d)
1
2;` (1 ; 2 )
=
for in X 2;` . fy;` is really a density on X 2;` because y has absolute value larger
than 2`.
With these notations we can rewrite (5.1) as
lim sup lim sup lim sup sup sup
"!0
`!1
Z N!1 DN (f )C0N d 2 2`<jyj2"N
1`(0) 2`(0) 1f1`(0) _ 2`(0) Ag f¯y;`()2;` (d) 0:
Next step consists in obtaining information concerning the density f¯y;` from
the bound on the Dirichlet form of f . It is here that the proofs of Lemma 3.1
and 3.2 differ sensibly. First of all, as we have seen in the proof of the one block
estimate, the Dirichlet form of f¯ is bounded above by the one of f :
DN (f¯) DN (f ) :
Let D2;` be the Dirichlet form defined on positive densities h: X 2;` ! R+ by
X X
D2;` (h) = I0`;0 (h) + 1;`
Ix;z ( h) + 2;`
Ix;z (h) ;
jx zj=1 jx zj=1
x;z2` x;z2`
where
Z q p 2
1;`
Ix;z (h) = (1=2) g( (x)) h(1x;z ; 2 ) h( ) 2;` (d ) ;
Z q p 2
Ix;z;` (h) = (1=2)
2
g( (x)) h(1 ; 2x;z ) h( ) 2;` (d )
Z q q 2
and I0`;0 (h) = (1=2) g(1 (0)) h(10; ; 20;+ ) h( ) 2;` (d ) :
In last formula, for a configuration = (1 ; 2 ) of X 2;` and for i = 1; 2, i0; is

defined as (
i (0) 1 if x = 0
i ; (x) =
0
i (x) otherwise :
This Dirichlet form corresponds to an interacting particle system on `
` where particles evolve according to nearest neighbor symmetric zero range
processes on each coordinate and where particles can jump from the origin of one
of the coordinates to the origin of the other.
d
Recall that for a density f on N TN , we denote by fy;` the conditional expec-
tation of f with respect to the -algebra generated by f (z ); z 2 y;` g. In the
same way as we did in (4.5), we can show that for each pair of neighbor sites x,
1;` 2;`
x0 in ` (resp. z , z 0 in in ` ) the Dirichlet form Ix;x 0 (fy;`) (resp. Iz;z0 (fy;` )) is
bounded above by the Dirichlet form of f :
1;` 2;`
Ix;x 0 (fy;` ) Ix;x0 (f ) and Iz;z0 (fy;` ) Iy+z;y+z0 (f )
for each neighbor sites x, x0 in ` and each neighbor sites z , z 0 in ` . In particular,
X 1;` ¯
X 2;` ¯
Ix;z (fy;` ) + Ix;z (fy;` ) 2C0 (2` + 1)d N 2
jx zj=1 jx zj=1
x;z2` x;z2`
for every density with Dirichlet form DN (f ) bounded by C0 N d 2 .
It remains to show that we can also estimate the Dirichlet form I0`;0 (f¯y;` ) by the
Dirichlet form of f . In order to do it we first derive an equivalent expression for
the Dirichlet forms I0`;0 and Ix;z . A change of variable = d0 and = dx
permits to rewrite these Dirichlet forms as
Z q q 2
I0`;0 (h) = (1=2)() h(10;+; 2 ) h(1 ; 20;+ ) 2;` (d1 ; d2 ) ;
Z p p
Ix;z (f ) = (1=2)() f (x;+) f (z;+) (d)
2
where, as before, for a configuration and an integer x, x;+ represents the con-
figuration obtained from adding a particle to site x :
(
(x) + 1 if z = x
x;+(z ) =
(z ) otherwise :
In the same way as we proved that Ix̀;z (f` ) Ix;z (f ), for each density f with
respect to , I0`;0 (f¯y;` ) is bounded above by
Z q q 2
(1=2)() f¯(0;+) f¯(y;+) (d) : (5:2)
Let (xk )0kjj yjj be a path from the origin to y , that is, a sequence of sites
such that the first one is the origin, the last one is y and the distance between two
consecutive sites is equal to 1 :
x0 = 0; xjjyjj = y and jxk+1 xk j = 1 for every 0 k jjyjj 1:
In these formulas, jj jj represents the sum norm :

X
jj(y1; : : : ; yd )jj = jyi j :
1id
We use the telescopic identity
p p jjX
yjj 1 p p
f (0;+) f (y;+) = f ( xk ;+ ) f (xk ;+ ) +1
k=0
and the Cauchy–Schwarz inequality
0jjyjj 1
@ X ak A jjyjj X ak
1 jjyjj 2
1
2
k=0 k=0
to estimate the integral (5.2) by
yjj 1 Z
jjX q q 2
(1=2)( )jjy jj f¯(xk ;+ ) f¯(xk+1 ;+ ) (d)
k=0
jjX
yjj 1
= jjyjj Ixk ;xk (f¯) :
+1
k=0
Since f¯ is translation invariant, for each k , Ixk ;xk+1 (f¯) = Ixk +z;xk+1 +z (f¯) for all z
in Zd so that Ixk ;xk+1 (f¯) N d DN (f ). In particular,
I0`;0 (f¯y;`) jjyjj2 N dDN (f ) :

Recall that jy j 2N" so that jjy jj djy j 2dN". Since the Dirichlet form is
assumed bounded by C0 N d 2 , we have proved that
I0`;0 (f¯y;`) 4C0 d2 "2 :

In conclusion, for every density f with Dirichlet form bounded by C0 N d 2
and for every d-dimensional integer y with max norm between 2` and 2N",
D2;` (f¯y;`) C7 (C0 ; d; `)"2 :

Notice that a factor "2 that vanishes in the limit as " # 0 appeared. In particular
to conclude the proof it is enough to show that
lim sup lim sup sup

`!1 "!0 D2;` (f )C7 (C0 ;`)"2
Z ` (0)
1 2` (0) 1f1`(0) _ 2` (0) Ag f ( )2;` (d ) 0;
for every A > 0. This time the supremum is taken over all densities with respect
to 2;` .
We may now follow the arguments presented in the proof of Lemma 3.1 to
conclude. The unique worthwhile mentioning slight difference is that every density
f with Dirichlet form equal to 0 (D2;`(f ) = 0) is constant on hyperplanes having
a fixed total number of particles on ` [fy + `g because particles can jump from
the origin of one of the coordinates to the origin of the other.
Remark 5.1 Notice that in the proof of Lemmas 3.1 and 3.2 only the bounds on
the Dirichlet form of order N d 2 and on the entropy of order N d were used. In
particular, by section 2, Lemmas 3.1 and 3.2 apply to mean-zero asymmetric zero
range processes satisfying assumption (FEM) or (SLG).
Remark 5.2 The time integral in the statement of Lemmas 3.1 and 3.2 is crucial.
Indeed, we have no a priori information on the order of magnitude of the Dirichlet
form at a fixed time. We are only able to prove in the mean-zero case a bound
of order O(N d 2 ) for the time integral of the Dirichlet form and this gives no
information on its value at a fixed time.
The cylinder function g ( (0)) does not play any particular role in Lemmas
3.1 and 3.2. The statement applies to a broad class of cylinder functions. Recall
from (2.3.10) the definition of Lipschitz cylinder functions and Lipschitz cylinder
functions with sublinear growth and recall from (2.3.12) the definition of the real
function ~. We proved in Corollary 2.3.7 that ~ is uniformly Lipschitz for every
Lipschitz cylinder function . The next two results follow from the proof of
Lemma 3.1.
Lemma 5.3 Under assumptions of Lemma 3.1, for every cylinder function with
sublinear growth and for every C0 ,
lim sup lim sup sup
`!1 N !1 DN (f )C0 N d 2
HN (f )C0 N d
Z X X

1
d
1
N x2Td (2` + 1)d y () ~(` (x)) f () (d) = 0 :
N jy xj`
Lemma 5.4 Under hypothesis (FEM), the statement of the previous lemma holds
for cylinder Lipschitz functions .
We mentioned above that ~ is a Lipschitz function for all cylinder Lipschitz

functions . In particular, we have the following result.
Lemma 5.5 Under hypothesis (FEM), the statement of Lemma 1.10 remains in
force if g is replaced by a cylinder Lipschitz function and by its homologue ~.
We conclude this section with a remark concerning the hypotheses (FEM)

and (SLG) and few words about the replacement lemma for generalized simple
exclusion processes.
Remark 5.6 Attractive zero range processes, i.e., systems with non decreasing
jump rate g , satisfy either hypothesis (FEM) or (SLG) because if the jump rate
g() is unbounded it fulfills hypothesis (FEM) and if it is bounded it satisfies
assumption (SLG). In particular, the one block estimates can be proved for all
attractive zero range processes.
At last, notice that in the case of generalized exclusion processes, the entropy of
d
any probability measure N on f0; : : :; gTN with respect to an invariant product
N d
state is bounded by C ()N . Indeed, by convexity of the entropy,
N
H N N max

H :
Here stands for the Dirac measure concentrated on the configuration . Since
X
H N = log N ( ) = Z (())N d log () (x)
x2TdN
and since the total number of particles per site is at most , the previous expression
is bounded above by C ()N d for some finite constant C (). Thus, for generalized
exclusion processes the bound on the entropy and on the time integral of the
Dirichlet form proved in section 2 apply to any sequence of initial measures N .
Moreover, in the statement of the replacement lemma the bound on the entropy is
unnecessary :
Lemma 5.7 For generalized exclusion processes, for every cylinder function
and for every C0 ,
lim sup lim sup sup

"!0 N !1 DN (f )C0 N d
Z 1 X
1 X
2

"N
N d x2Td (2"N + 1)d jy xj"N y () ( (x))f () (d) = 0:
~
N
6. A L2 estimate
We prove in this section that all limit points Q of the sequence fQN ; N
1g are concentrated on absolutely continuous measures whose density is in
L2 ([0; T ] Td). We start introducing some notation related to Fourier transforms.
Fix a positive integer N and consider the space L2 (TdN ) of complex functions on
P
TdN endowed with the inner product < f; g >= N d x2Td f (x)g (x) , where a
N
stands for the conjugate of a. For each z in TdN , denote by z = N;z the L2 (TdN )
function defined by
2i n o
z (x) = exp ( z x) ;
N
where zx stands for the usual inner product in Rd . It is easy to check that
f z ; z 2 TN g forms an orthonormal basis of L2 (TdN ). In particular, each function
d
f in L2 (TdN ) can be written as
X
f = < f; z > z :
z2TdN
We shall repeatedly use the following three properties of the Fourier transform.
Since f z ; z 2 TdN g is an orthonormal basis, for f , g in L2 (TdN ),
X
< f; g > = < f; z > < g; z > : (6:1)
z2TdN
P
Denote by N the discrete Laplacian : (N f )(x) = jd ff (x + ej ) + f (x
ej ) 2f (x)g. A double summation by parts gives that
1
d n
X 2zj o
< N 2 N f; z > = 2N 2 1 cos
N < f; z > (6:2)
j =1
P
because N z = 2 1j d f1 cos(2zj =N )g z . For two functions f , g in
L2 (TdN ), denote by f g the convolution of f and g :
X
(f g)(z ) = f (x)g(z x) :
x2TdN
From this definition it is easy to deduce that
< (f g); z > = N d < f; z > < g; z > : (6:3)
We now introduce two additional norms in L2 (TdN ) and investigate the relations
between them. For a function f in L2 (TdN ), define the H1 norm kf k1 of f by
kf k21 = < f; (I N 2N )f > :

By properties (6.1) and (6.2), the H1 norm of f is equal to
6. A L2 estimate 99
X X
< f; z > < (I N 2 N )f; z > = < f; z > 2 aN (z ) ; (6:4)
z2TdN z2TdN
where aN : TdN ! R+ is the positive function given by
X
d
aN (z ) = 1 + 2N 2 f1 cos(2zj N 1
)g :
j =1
Denote by AN : TdN ! R, the inverse Fourier transform of the function aN :
AN (z ) =< aN ; z >. Since AN is the inverse Fourier transform of aN ,
X
AN =
N d z2Td aN (z ) z ;
1
N
what can be confirmed by an elementary computation. Notice that AN is an even
real function because aN is even. We claim that the H1 norm can be rewritten as
X
kf k21 = < f; AN f > = N1d f (z )AN (z y)f (y) : (6:5)
z;y
Indeed, by property (6.1) and (6.3),
X
< f; AN f > = N d < f; z > 2 < AN ; z > :
z2TdN
Since
P
AN = N d y2TdN aN (y) y and < y ; z >= y;z , we have that
< AN ; z >= N daN (z ), what proves (6.5) in view of (6.4) because aN is
a real function.
We now introduce the dual norm of H1 with respect to L2 (TdN ). For each
function f in L2 (TdN ), define
n o
kf k2 1 = sup 2 < f; g > kgk21 ;
g
where the supremum is taken over all functions g in L2(TdN ). We claim that the
H 1 norm is X
kf k2 1 = < f; z > 2 a 1(z )
(6:6)
z2TdN N
Indeed, for fixed real functions f , g , by properties (6.1), (6.4),
< f; g > kgk
X
2
X
2
= 2 < f; z >< g; z >
1

< g; z > 2 aN (z ) : (6:7)
z2TdN z2TdN
P
Take g = z aN (z ) 1 < f; z > z (so that < g; z >= aN (z ) 1 < f; z >) to
obtain that the left hand side of (6.6) is bounded below by the right hand side. To
prove the reverse inequality, notice that (6.7) is bounded above by
X n o
2 < f; z > < g; z > aN (z ) < g; z > 2 :
z2TdN
Since 2ab Rb2 a2 R
P < f; >
a (z) z
1
, this expression is less than or equal to z
2
N 1
, what proves (6.6).
Denote by KN : TdN ! R the inverse Fourier transform of aN1 : KN (z ) =<
aN ; z > or
1
X
KN = N1d aN1 (z ) z : (6:8)
z2TdN
Since aN () is an even function, KN () is an even real function. Repeating the
arguments that lead to (6.5), we deduce that the H 1 norm can be expressed as
X
kf k2 1 = < f; KN f > = N1d f (x)KN (x y)f (y) (6:9)
x;y
because KN is a real function.
Two properties of the kernel KN () will be used in the proof of the L2 estimate
for the density of the empirical measure. On the one hand, it follows from formula
(6.8) and the identity (I N 2 N ) z = aN (z ) z that
(I N 2 N )KN (y) = 0;y ; (6:10)
where x;y stands for the delta of Kronecker. On the other hand, the same formula
(6.8) for the kernel KN () gives that
X
d
N2 KN (0) KN (ej ) 1
2
: (6:11)
j =1
We conclude this preamble with two observations. It follows from formula
(6.6) for the H 1 norm and from (6.2) that kf k2 1 kf k20 because aN () 1.
Furthermore, by Schwarz inequality, for any real functions f , g in L2 (TdN ),
< f; K g > "kf k " 1 kgk2 1 (6:12)
2 N 2
1 +
for every " > 0. Indeed, by (6.1) and (6.3), the left hand side is bounded above
by X
2N d < f; z > < g; z > < KN ; z > :
z2TdN
Since < KN ; z >= N d aN ( z ) 1
and 2ab a2 + b2 , the previous expression is
bounded above by
6. A L2 estimate 101
X X
" < f; z > 2 a 1(z ) + " 1 < g; z > 2 a 1(z )
z2TdN N z2TdN N
for every " > 0. This concludes the proof of (6.12) in view of (6.6).
We are now ready to state the main result of this section.
Proposition 6.1 Fix T > 0 and a sequence of probability measures fN ; N 1g

such that h i
lim sup EN k k2 1 = K2 < 1 : (6:13)
N !1
There exists a finite constant C depending only on K2 , the jump rate g and T such
that
h i hZ t X i
E N k(t)k 2
1 + E N ds 2N1 d g(s(x))fs (x) 1g C
0
x2TdN
for all t T .
Proof. Consider the martingale M (t) defined by

Z t
M (t) = k(t)k2 1 k(0)k2 1 ds N 2 LN k(s)k2 1 :
0
A simple computation relying on the explicit formula (6.9) for the H 1 norm
permits to compute N 2 LN k k2 1 . Since the kernel KN is an even function,
N 2 LN kk2 1 is equal to
X
1
Nd (x)(N 2 N KN )(x y)g((y))
x;y2TdN
d
X X
+ 2N 2 fKN (0) KN (ej )g N1d g((x)) :
j =1 x2TdN
By property (6.11) of the
P
kernel, the second term of the previous expression is
bounded above by N d x2Td g ( (x)). On the other hand, the first term can be
N
rewritten as
X
N d x;y2Td (x)([I N N ]KN )(x y)g((y))
1 2
N
X
+
1
Nd (x)KN (x y)g((y)) :
x;y2TdN
By (6.10), (6.9) and (6.12), this expression is bounded above by
1 X
(x)g((x)) A kk2 1
kg()k2 1
Nd x2TdN
+
2 1 +
2A
for every A > 0. Since the H 1 norm is bounded above by the L2 norm the third
term of this sum is bounded above by (2A) 1 kg ( )k20. Recollect all the previous
estimates. Up to this point we proved that N 2 LN k k2 1 is bounded above by
X A kk2
N d x2Td g((x))f(x) 1g + kg()k20
1 1
2A
1 +
2
N
for every A > 0. Since g (k ) g k , g (k ) 2g (k 1) for k 2. In particular,
g(k)2 is bounded above by g(1)2 + 2gg(k)(k 1). Choosing therefore A = 2g,
we obtain that N 2 LN k k2 1 is less than or equal to
X
2N d
1
g((x))f(x) 1g + g kk2 1 + C (g )
x2TdN
for some finite constant C (g ) that depends only on the jump rate.
Let RN (t) = E N [k (t)k2 1]. Since M (t) is a mean-zero martingale, we just
proved that
hZ t X i
RN (t) + E N ds 2N1 d g(s (x))fs(x) 1g
0
x2TdN
Z t
RN (0) + g ds RN (s) + C (g )T
0
for all t T . We conclude the proof of the proposition applying Gronwall in-
equality.
Corollary 6.2 In the case where the jump rate g is such that g (k ) a0 k for some
positive constant a0 , under the assumptions of the previous proposition, for each
T > 0, there exists a finite constant C depending only on g, K2 and T such that
h i hZ t X i
E N k(t)k 2
1 + E N ds N1d s (x)2 C
0
x2TdN
for all t T .
Corollary 6.3 Under the assumptions of Theorem 1.1, all limit points Q of the
sequence fQN ; N 1g are concentrated on paths (t; du) = (t; u)du such that
Z T Z
dt du (t; u)2 < 1
0 Td
almost surely.
Proof. It follows from the previous corollary and Schwarz inequality that
7. An energy estimate 103
hZ T X 2 i
lim sup lim sup E N ds N1d s"N (x) C
"!0 N !1 0
x2TdN
for some finite constant C . It is now easy to conclude the proof of the corollary.

Remark 6.4 In view of Proposition 6.1, we may replace assumption (1.2) on the
sequence of initial measures fN ; N 1g by the weaker one (6.13)
7. An energy estimate
We prove in this section an energy estimate for the trajectories (t; u). At the
end of the section we present a simple proof of uniqueness of weak solutions of
equation (1.3) in dimension 1 in the class of paths satisfying an energy estimate.
Fix a limit point Q of the sequence QN and assume, without loss of generality,
that the sequence QN converges to Q . The main theorem of this section can be
stated as follows.
Theorem 7.1 The probability measure Q is concentrated on paths (t; u)du

with the property that there exist L1 ([0; T ] Td ) functions denoted by f@uj (
(s; u)); 1 j dg such that
Z T Z Z T Z
ds d
du (@uj G)(s; u)((s; u)) = ds du G(s; u)@uj ((s; u))
0 T 0 Td
for all smooth functions G and all 1 j d. Moreover,
Z T Z
du kr((s; u))k < 1 :
2
ds ((s; u)) (7:1)
0 Td
The proof of this theorem relies on the following estimate.
Lemma 7.2 Recall the definition of the constant K0 introduced in Theorem 1.1.
For 1 j d,
nZ T Z
EQ sup ds du (@uj H )(s; u)((s; u))
H 0 Td
Z T Z o
2 ds du H (s; u) ((s; u))
2
K0 :
0 Td
In this formula the supremum is taken over all functions H in C 0;1 ([0; T ] Td ).
Before proving this lemma, we show how to deduce Theorem 7.1 from this
statement.
Proof of Theorem 7.1. From Lemma 7.2, for Q almost every path , there exists
a finite constant B = B () such that
Z T Z Z T Z
ds du (@uj H )(s; u)((s; u)) du H (s; u)2((s; u)) B2 ds
Td Td
(7:2)
0 0
for every 1 j d and every C 0;1 ([0; T ] Td ) function H . Fix such a path
and consider on C ([0; T ] Td ) the inner product < ; > defined by
Z T Z
< F; G > = ds du H (s; u)G(s; u)((s; u)) :
0 Td
Denote by L2 the Hilbert space induced by C ([0; T ] Td) and this inner product.
For 1 j d, let `j (H ) be the linear functional on C 0;1 ([0; T ] Td ) defined
by
Z T Z
`j (H ) = ds du (@uj H )(s; u)((s; u)) :
0 Td
It follows from estimate (7.2) that
Z T Z
a`j (H ) 2a2 ds du H (s; u)2((s; u)) B
0 Td
for every a in R. Maximizing over a we show that the linear operator `j is
bounded in L2 . In particular, it can be extended to a bounded linear functional
in L2 . By Riesz representation theorem there exists a L2 function, denoted by
@uj log ((s; u)), such that
Z T Z
ds du (@uj G)(s; u)((s; u))
0 Td
Z T Z
= ds du G(s; u)@uj log ((s; u))((s; u))
0 Td
for every smooth function G: [0; T ] Td ! R. Moreover,
d Z
X T Z 2
ds d
du @uj log ((s; u)) ((s; u)) < 1 :
j =1 0 T
To conclude the proof of the theorem, define @uj ((s; u)) as ((s; u))@uj
log ((s; u)). It is straightforward to check the properties of @uj ((s; u)).
The proof of Lemma 7.2 relies on the following estimate. Fix 1 j d and
recall that fej ; 1 j dg stands for the canonical basis of Rd . For a smooth
function H : Td ! R, > 0, " > 0 and a positive integer N , define WN ("; ; H; )
by
WN ("; ; H; )
X n o
= N1 d H (x=N )("N ) 1
(N (x)) (N (x + "Nej ))
x2TdN
X "N
X
2N d H (x=N )2("N ) 1
(N (x + kej )) :
x2TdN k=0
Lemma 7.3 Consider a sequence fH`; ` 1g dense in C 0;1([0; T ] Td). For

every k 1, and every " > 0,
nZ T o
lim sup lim sup E N max
ik
ds WN ("; ; Hi(s; ); s ) K0 :
!0 N !1 1 0
Proof. It follows from the replacement lemma that in order to prove the lemma
we just need to show that
nZ T o
lim sup E N max
ik
ds WN ("; Hi (s; ); s ) K0 ;
N !1 1 0
where
X
WN ("; H; ) = N1 d H (x=N )("N ) 1
g((x)) g((x + "Nej ))
x2TdN
X "N
X
2N d H (x=N )2("N ) 1
g((x + kej )) :
x2TdN k=0
By the entropy inequality and the Jensen inequality, for each fixed N , the
previous expectation is bounded above by
Z T o
H (N jN ) 1
N d ds WN ("; Hi(s; ); s )
n
:
Nd +
N d log E N exp
1ik 0
max
P
Since expfmax1j k aj g 1j k eaj and since lim supN N d logfaN + bN g
is bounded above by maxflim supN N d log aN ; lim supN N d log bN g, the limit,
as N " 1, of the second term of the previous expression is less than or equal to
h n Z T oi
K0 + max lim sup
1ik N !1
1
N d log E N exp Nd ds WN ("; Hi (s; ); s ) :
0
We now prove that for each fixed i the limit of the second term is nonpositive.
Fix 1 i k . By Feynman–Kac formula and the variational formula for the
largest eigenvalue of a symmetric operator, for each fixed N , the second term of
the previous expression is bounded above by
Z T nZ o
ds sup WN ("; Hi (s; ); )f ()N (d) N 2 dDN (f ) : (7:3)
0 f
In this formula the supremum is taken over all densities f with respect to N .
Recall the formula for WN ("; Hi (s; ); ) and that for an integer x, dx stands
for the configuration with no particles but one at x. The change of variables
= dx shows that
Z Z
g((x))f ()N (d) = ( ) f ( + dx )N (d) :
In particular, we have that
Z

H (s; x=N ) g((x)) g((x + "Nej )) f ()N (d)
Z

= ( )H (s; x=N ) f ( + dx ) f ( + dx+"Nej ) N (d)
"N
X1 Z
= ( )H (s; x=N ) f ( + dx+kej ) f ( + dx+(k+1)ej ) N (d)
k=0
p p p
pf ( + df ))( and
Writing + dy ) f ( + dz ) as ( f ( + dy ) f ( + dz ))( f ( + dy ) +
applying the elementary inequality 2ab a2 + 1 b2 that holds
z
for every a, b in R and > 0, we obtain that the previous expression is bounded
above by ( ) times
H (s; x=N )2 "NX1 Z nqf ( + d q

f
o2
( + dx+kej ) N (d )
2
x +(k+1)ej ) +
k=0
x "N
+X 1 Z nq q o2
+
2
f ( + d ) f ( + d ) N (d)
x+(k+1)ej x+kej
y =x
for every > 0. The inequality (a + b)2 2a2 + 2b2 and the change of variables
= + dz for z = x + kej , x + (k + 1)ej permit to show that the first term is
bounded above by
"N Z
2H (s; x=N )2 X
g((x + kej ))N (d) :
k=0
Therefore, setting = N , recalling the definition of the Dirichlet form DN (f ) and
summing over x we obtain that
X Z
N1 d H (s; x=N )("N ) 1
fg((x)) g((x + "Nej ))gf ()N (d)
x2TdN
X "N Z
X
N2d H (s; x=N )2 "N
1
g((x + kej ))f ()N (d)
x2Td N k=0
+ N 2 d DN (f ) :
This proves that (7.3) is nonpositive because the Dirichlet forms cancel and the
second term of WN ("; H; ) is just the first term on the right hand side of the
previous inequality.
Proof of Lemma 7.2. Fix 1 j d. Since QN converges weakly to Q , it
follows from Lemma 7.3 that for every k 1
Z T Z
lim sup EQ max ds du
!n
0 1ik 0 Td

Hi (s; u)" 1
(s (u)) (s (u + "ej ))
Z o
2 Hi (s; u)2" 1
dv (s (v)) K0 ;
[ u;u+"ej ]
where is the approximation of the identity () = (2 ) 11f[ ; ]g().
Letting # 0, changing variables and then letting " # 0, we obtain that
Z T Z
EQ 1max
ik 0
ds d du
T
n o
(@uj Hi )(s; u)((s; u)) 2 Hi (s; u)2((s; u)) K0 :
To conclude the proof it remains to apply the monotone convergence theorem and
recall that fH` ; ` 1g is a dense sequence in C 0;1 ([0; T ] Td ) for the norm
kH k1 + k(@uH )k1 .
We conclude this section with a proof in dimension 1 of the uniqueness of
weak solutions in H1 of the Cauchy problem (1.3). We start introducing some
terminology.
Definition 7.4 A measurable function : [0; T ] T ! R+ is said to be a weak

solution in H1 of (1.3) provided
(a) (t; ) belongs to L1 (Td) for every 0 t T and sup0tT kt kL < 1. 1
(b) There exists a function in L2 ([0; T ] T), denoted by @u ((s; u)), such that
for every smooth function G: [0; T ] T ! R
Z T Z
ds du (@u G)(s; u)((s; u))
0 T
Z T Z
= ds du G(s; u)@u((s; u)) :
0 T
(c) For every smooth function G: R ! R and for every 0 < t T ,
Z Z
du (t; u)G(u) du 0 (u)G(u)
T T
Z t Z
= ds du G0 (u)@u((s; u)) :
0 T
Notice that we require here @u ((s; u)) to belong to L2 ([0; T ] T), while we
proved only Q to be concentrated on paths satisfying (7.1). We need therefore
further estimates on the trajectories in order to show that for zero range processes
all limit points are concentrated on weak solutions in the H1 sense. However, for
other models like generalized exclusion processes or Ginzburg–Landau processes,
the same proof of Lemma 7.2 provides stronger estimates from which it is easy to
deduce that all limit points are concentrated on weak solutions in the H1 sense.
To prove the uniqueness of weak solutions we need to introduce some notation.
On the torus T the kernel of 1 , the inverse of the Laplacian, is given by :
(
u + v(u 1) provided 0 v u 1 ;
K (u; v)
uv provided 0 u v 1 :
=
For a smooth function H : T ! R, denote by K H : T ! R the convolution of

K with H : Z
(K H )(u) = K (u; v)H (v) dv :
T
A simple computation shows that (K H ) = H , what confirms that K is the
kernel of 1 .
On C 2 (T) consider the inner product < H; G >1 defined by
Z
< H; G >1 = H (u)(G)(u) du :
T
Let H1 be the Hilbert space induced by C 2 (T) and the inner product < ; >1 .
The H1 norm is denoted by k k1 .
Denote by H 1 the dual Hilbert space of H1 with respect to L2 (T). Since K
is the kernel of the inverse of the Laplacian, the H 1 norm is given by
Z
kH k 2
1 = H (u)(K H )(u) du :
We are now in a position to investigate the uniqueness of weak solutions of
(1.3).
Theorem 7.5 There is at most one weak solution in H1 of equation (1.3).
Proof. Consider two weak solutions 1 and 2 and denote by ¯ their difference :
¯ = 2 1 . It follows from property (c) of weak solutions that
Z t Z Z
k¯(t; )k2 1 = k¯(0; )k2 1 + 2 ds du ¯(s; u) dv (@v K )(u; v)(@v ¯ s )(v)
0 T T
for every 0 t T . In this formula, ¯ s (v ) stands for (2 (s; v )) (1 (s; v )).
The explicit formula for the kernel K permits to rewrite the right hand side as
Z t Z
k¯(0; )k2
1 2 ds du ¯(s; u)f¯ s(u) ¯ s (0)g :
0 T
R
Since the total mass is conserved (@t T (t; u) du = 0) and 1 , 2 are weak solu-
tions, the second term of the previous formula is equal to
Z t Z
2 ds du ¯(s; u)¯ (s; u) ;
0 T
which is negative because () is an increasing function and
¯ s (v) = (2 (s; v)) (1 (s; v)) :

This computation proves that the H 1 norm of the difference of two weak solutions
with the same total mass does not increase in time. Uniqueness follows.
The entropy method presented in this chapter is due to Guo, Papanicolaou and
Varadhan (1988). It permitted to prove the hydrodynamic behavior of a large
class of interacting particle systems and interacting diffusion processes through
the investigation of the time evolution of the entropy.
Fritz (1990) extended the entropy estimate to infinite volume for Ginzburg–
Landau processes. Yau (1994) proposed an alternative method to estimate the
infinite volume entropy. Landim and Mourragui (1997) adapted Yau’s approach
to zero range processes.
Lu (1995) proved that starting from a class of deterministic configurations,
there is a microscopic time t0 and a finite constant C0 with the property that at
time t0 the entropy of the process with respect to a reference invariant measure
is bounded by C0 N in dimension 1. This estimate permits to extend the entropy
argument to processes starting from deterministic configurations.
Yau (1994) introduced a method, based on the time evolution of the H 1
norm, to prove a law of large numbers for the density field at a finer scale than
the hydrodynamic scale for dissipative systems. It permits to show the existence
of 0 < a < 1 such that
X
N d H (x=N; tM (x))
x2Zd
R
converges to H (u; (t; u)) du, with M = N a . Landim and Vares (1996) applied
this method to the superposition of a speeded up Kawasaki dynamics and a Glauber
dynamics.
Applying the entropy method Funaki, Handa and Uchiyama (1991) prove the
hydrodynamic behavior of a one-dimensional, reversible, gradient, symmetric ex-
clusion process with speed change. Suzuki and Uchiyama (1993) show that the
macroscopic evolution of a gradient reversible and conservative [0; 1)-valued
spin process is described by a nonlinear parabolic equation. In these models the
usual entropy bound can be relaxed and a moment estimate is required. Ekhaus
and Seppäläinen (1996) and Feng, Iscoe and Seppäläinen (1997) consider similar
models giving rise to the porous medium equations @t = , for > 0,
> 1.
Fritz (1989b) consider the hydrodynamic behavior of a one-dimensional
Ginzburg–Landau process in random environment. Quastel (1995b) examines the
same question for exclusion processes. In this case the process turns out to be
nongradient. Koukkous (1997) investigates the hydrodynamic behavior of mean-
zero asymmetric zero range processes in random environment. The model is the
following. Consider a sequence of independent random variables fax ; x 2 Zdg
taking values in some interval [a; b] with a > 0. Denote by m the distribution of
a0 . For a fixed realization of the environment, consider the zero range process t
in which a particle at x jumps to x + y at rate p(y )g ( (x))ax. In this model, for
each x in Zd, the jump rate is speeded up or slowed down by the random factor
ax . The hydrodynamic equation of this model is shown to be
@t = ^() ;
where ^ is defined as follows. Recall the definition of the function R(). Let
R^(') = Em [R('a0 1)]. ^ is the inverse of R^ and is the second order differential
operator defined just before (1.1).
Systems in contact with stochastic reservoirs. The question is to characterize
the density profile in a pipe connecting two infinite reservoirs containing a fluid
with two different densities in a stationary regime.
To fix ideas, consider a simple exclusion process on ZN = f0; 1; : : :; N
1g with symmetric jump rates in the interior of ZN and with jump rates at the
boundary chosen in order to obtain there a priori fixed densities. The generator of
this process is :
X
(LN f )( ) = (x)[1 (y)][f (x;y) f ()] + (L f )() + (L+ f )( ) ;
x;y2ZN
jx yj=1
where L , L+ are the boundary generators given by
(L+ f )( ) = (N 1)[f ( dN 1 ) f ( )]
+ + [1 (N 1)][f ( + dN 1 ) f ( )] ;
(L f )( ) = (0)[f ( d0 ) f ( )] + [1 (0)][f ( + d0 ) f ( )] :
Here and + are two positive constants that stand for the rate at which particles
are created at the boundary. A simple computation shows that the Bernoulli product
measure + =(1++ ) (resp. =(1+ ) ) is reversible for the process with generator
L+ (resp. L ).
Since the process is indecomposable, there exists a unique stationary measure,
denoted by N . Only in very special cases is this measure explicitly computable.
The problem is to investigate the density profile associated to this stationary state.
More precisely, to prove the existence of a profile 0 : [0; 1] ! R+ such that
n Z 1 o
lim sup N < N ; G >

; G(u)0(u) du > = 0
N !1 0
for every continuous function G : [0; 1] ! R and every > 0. One expects 0 to
be the solution of an elliptic equation with boundary conditions :
(
@u (D((u))@u) = 0 ;
(0) = =(1 + ) ; (1) = + =(1 + + ) ;
where D() is the diffusion coefficient defined by the Green–Kubo formula.
Fick’s law of transport for the expected value of the current in the stationary
regime can also be examined. For 0 x N 2, denote by Wx;x+1 the current
over the bond fx; x + 1g, i.e., the rate at which a particle jumps from x to x + 1
minus the rate at which a particle jumps from x + 1 to x. In the example we
introduced above, the current is equal to (x) (x + 1). One would like to prove
that for every u in [0; 1],
h i
E N NW[uN ];[uN ]+1
N !1
lim = D(0 (u))@u0 (u) :
Finally, we may also investigate the relaxation to equilibrium starting from a
state associated to some profile. To illustrate this question, in the example intro-
duced above, fix a profile : [0; 1] ! [0; 1], consider a sequence of probability
measures fN (); N 1g on f0; 1g N associated to the profile and denote by
Z
N
P the probability on the path space corresponding to the Markov process with
generator LN speeded up by N 2 and starting from N () . It is natural to prove
a law of large numbers for the empirical measure under PN . More precisely, to
show that for every t > 0
n Z 1 o
lim sup PN < tN ; G >

G(u)(t; u) du > = 0
N !1 0
for every continuous function G : [0; 1] ! R and every > 0, provided (t; u)
stands for the solution of the nonlinear parabolic equation
8
>
< @t = @u (D((u))@u) ;
>
(0; ) = () ;
:
(; 0) = =(1 + ) ; (; 1) = + =(1 + + ) :
Goldstein, Lebowitz and Presutti (1981) and Goldstein, Kipnis and Ianiro
(1985) investigate the stationary state of N particles moving on a bounded re-
gion of R3 according to a deterministic Hamiltonian equation in which particles
are thermalized at the boundary. They prove the existence of a stationary state,
which is equivalent to the Lebesgue measure and show convergence in variation
norm of any probability measure under the time evolution. Goldstein, Lebowitz
and Ravishankar (1982) and Farmer, Goldstein, Speer (1984) prove the existence
of a nonequilibrium steady state for a one-dimensional infinite system of molecules
confined in a region in interaction with atoms which flow to from two semi-
infinite reservoirs separated by .
Kipnis, Marchioro and Presutti (1982) consider a one-dimensional system of
harmonic oscillators in contact with reservoirs at different temperature. They obtain
the stationary measure, the temperature profile and prove the local convergence to
the Gibbs measure. De Masi, Ferrari, Ianiro and Presutti (1982) proved the same
statement in the context of symmetric simple exclusion processes in contact with
stochastic reservoirs at different temperature.
Ferrari and Goldstein (1988) consider a symmetric simple exclusion process on
Z3 with creation and destruction of particles at the origin. They deduce the density
profile of the nonequilibrium stationary measure, that turns out to be non product,
and compute the decay of the two point correlation function. Lebowitz, Neuhauser
and Ravishankar (1996) deduce asymptotic occupation properties of the stationary
measure of a semi–infinite asymmetric one-dimensional particle system with a
source at the origin, coupled jumps and annihilation. This is a first approximation
of the so-called Toom cellular automaton.
For zero range processes, as noticed by De Masi and Ferrari (1984), the sta-
tionary measure of a system in contact with an infinite reservoir is a product
measure with slowly varying parameter. All computations are thus explicit.
Based on the entropy method introduced by Guo, Papanicolaou and Varadhan
(1988), Eyink, Lebowitz and Spohn (1990, 1991) obtained the macroscopic profile
of the stationary measure and proved the hydrodynamic behavior of the system for
a gradient exclusion process where the jump rates depend locally on the configura-
tion. Kipnis, Landim, Olla (1995) extended this result to a nongradient generalized
exclusion process relying on Varadhan’s nongradient method (Varadhan (1994a),
Quastel (1992)). Systems in contact with stochastic reservoirs have never been
considered in higher dimensions.
Onsager’s reciprocity relations. Consider a zero range process with two types
of particles. For a = 1, 2, fixP
jump rates ga : N
P
N ! R+ and mean-zero transition
probabilities pa : Zd ! R+ ( y pa (y ) = 1, y ypa (y ) = 0). Define the generator
d d
of the Markov process (t ; t ) on N TN N TN by
X
2 X
(LN f )(; ) = pa (y)ga((x); (x))(ax;x+y f )(; ) ;
a=1 x;y2TdN
where
(1x;x+y f )(; ) = [f ( x;x+y ; ) f (; )]

and (2x;x+y f )(; ) = [f (; x;x+y ) f (; )] :
If the jump rates are not degenerated, this process has only two conserved
quantities, the total number of and particles. Moreover, for each = (1 ; 2 ) in
R+ R+ , there exists an invariant measure, denoted by N1 ;2 , with global density of
P
-particles (resp. -particles) equal to 1 (resp. 2 ) : EN ; [N d x (x)] = 1 ,
P
EN ; [N d x (x)] = 2 .
1 2
1 2
Assume that this family of invariant measures has good regularity prop-
erties in order to be able to define local equilibrium R states. For each profile
= (1; 2 ): Td ! (R+ )2 of density = (1 ; 2 ) ( Td a (u)du = a , a = 1, 2),
denote by N 1 ;2 a local equilibrium of profile . Assume that the specific entropy
of N 1 ;2 with respect to N1 ;2 converges, as N " 1, and denote by S (1 ; 2 )
its limit : S (1 ; 2 ) = limN !1 N d H (N 1 ;2 j1 ;2 ). RSuppose that the entropy is
N
written as the integral of a density s() : S (1 ; 2 ) = Td s(1 (u); 2 (u))du.
The macroscopic behavior of a system starting from N 1 ;2 is expected to be
described by a system of diffusion equations : denote by a (t; u) the density of
(; )–particles at the macroscopic point u at time t, (t; u) should evolve according
to the equation 8
> d
X
>
< @t = @ui Di;j ()@uj ;
> i;j =1
>
(0; ) = () :
:
Here stands for the vector (1 ; 2 ) and Di;j () is a two by two matrix for each
1 i; j d.
The Onsager coefficients are defined in this context by
Li;j () = Di;j () R() ;

where the matrix R is determined by the entropy density s((u)) in the following
way
(R
@2
)a;b =
@a(u)@b(u) s((u)) ;
1
which is by definition a symmetric matrix. Onsager’s reciprocity relations (cf.

Onsager (1931a,b) mean that the matrices Li;j = fLa;b i;j ; 1 a; b 2g are such
a;b b;a
that Li;j = Lj;i for 1 i; j d, 1 a; b 2. The exact microscopic conditions
required to prove these relations are far to be understood.
Gabrielli, Jona–Lasinio and Landim (1996) proved Onsager’s reciprocity re-
lations for mean-zero, non reversible zero range processes in the case where the
jump rates are g1 ( (x); (x)) = g ( (x) + (x))f (x)=( (x) + (x))g, g2 ( (x); (x)) =
g((x) + (x))f (x)=((x) + (x))g for some jump rate g: N ! R+ and p1 = p2 . In
this case the invariant measures are product and all computations presented in this
chapter can be done explicitly. Moreover, the Onsager matrices L are diagonal.
Lebowitz and Spohn (1997) observed that the previous models belong in fact to
a larger class that has a mirror type symmetry : the dynamics is invariant by the
exchange of and particles. Gabrielli, Jona–Lasinio and Landim (1998) present
sufficient conditions to guarantee the validity of Onsager’s reciprocity relations
for a general class of interacting particle systems. This question is discussed in
great generality in Eyink, Lebowitz and Spohn (1996)
6. The relative entropy method
In Chapter 1 we introduced in the context of interacting particle systems the phys-

ical concepts of local equilibrium and conservation of local equilibrium and we
proved the persistence of local equilibrium in a model where particles evolve inde-
pendently. Consider a particle system t evolving on the torus TdN and possessing
a family fN ; 0g of product invariant measures indexed by the density. Fix a
profile 0 : Td ! R+ and assume that the process t has a hydrodynamic behavior
described by the solution (t; u) of some partial differential equation with initial
condition 0 . Denote by N a sequence of initial states associated to the profile
0 and by Nt the state at the macroscopic time t of the process that started from
N . The conservation of local equilibrium states that Nt should be close to the
product measure N(t;) with slowly varying parameter associated to (t; ).
In contrast with the entropy method, where the hydrodynamic behavior is
deduced from the investigation of the time evolution of the entropy H (N t j )
N
of the state of the process with respect to a fixed invariant measure, the relative
entropy method examine the time evolution of the entropy H (N t j(t;) ) of the
N
N
state of the process with respect to the product measure (t;) with slowly varying
parameter associated to the solution of the hydrodynamic equation.
The relative entropy method requires some regularity of the solution of the
hydrodynamic equation. Of course, this is not a restriction for systems where
the average displacement of each elementary particle has mean zero since the
macroscopic behavior of these processes are described by second order quasi–
linear equations, whose weak solutions are smooth. This is not the case, however,
for asymmetric processes described by first order hyperbolic equations whose
solutions develop shocks. In this latter case, the relative entropy method allows
to deduce the hydrodynamic limit of the system up to the appearance of the first
shock. In fact, the relative entropy approach is the unique method that derives the
hydrodynamic behavior of non attractive asymmetric processes.
A last remark concerns the assumptions on the solutions of the hydrodynamic
equation. While the entropy method to be implemented requires a theorem assert-
ing the uniqueness of weak solutions and proves the existence of weak solutions,
the relative entropy method requires the existence of a smooth solution and proves
the uniqueness of such smooth solutions.
116 6. The relative entropy method
1. Weak conservation of local equilibrium
To illustrate the relative entropy method, we consider a mean-zero asymmetric

zero range process on the torus TdN . This is the Markov process introduced in
Chapter 2 whose generator is
X
(LN f )( ) = p(y)g((x))ff (x;x+y) f ()g :
x;y2TdN
To avoid minor technical difficulties, we assume the transition probability p() to
be associated to a finite range mean-zero random walk :
(i)
X
xj p(x) = 0 for 1 j d:
x2Zd
(ii) There exists an integerA0 such that p(x) = 0 if jxj A0 .
Notice that we did not assume the matrix p() to be symmetric and thus the process
to be reversible.
In the proof of the hydrodynamic behavior by means of the relative entropy
method, we shall need the one block estimate. This result (Lemma 5.3.1) was
proved in the previous chapter under hypotheses (FEM) or (SLG) on the jump
rate g (). We shall therefore assume either one of these hypotheses throughout this
chapter without mentioning it again.
We now describe the hydrodynamic equation. Denote by = (i;j )1i;j d the
matrix of correlations of the displacement of an elementary particle :
X
i;j = xi xj p(x)
x2Zd
and by the second order differential operator
X
= i;j @ui @uj :
1 i;jd
Let 0 < e< 1 and 0 : Td ! R+ be a profile of class C 2+e (Td). The Cauchy
problem (
@t = ()
(1:1)
(0; ) = 0 ()
admits a classic solution, that we denote by (t; u), twice continuously differen-
tiable in space and once continuously differentiable in time (cf. Oleinik et Kružkov
(1961)). Moreover, for each t 0, the profile (t; ) is of class C 2+e (Td ).
To avoid uninteresting technical difficulties, we assume that the initial profile
0 () is bounded below by a strictly positive constant :
1. Weak conservation of local equilibrium 117
K1 := inf
u2Td
0 (u) > 0 :
By the maximum principle, for every t > 0, (t; ) is bounded below by K1 :
inf inf (t; u) = K1
t0 u2Td
and above by the L1 norm of 0 (), denoted by K2 :
sup sup (t; u) = sup 0 (u) =: K2 < 1 :
t0 u2Td u2T
d
Hereafter, for t 0, we denote by N(t;) the product measure with slowly

varying parameter associated to the profile (t; ) (cf. Definition 3.0.1) :
N(t;) f; (x) = ng = (t;x=N ) f; (0) = ng ; for x 2 TdN and n2N :
We may now state the main result of this chapter.
Theorem 1.1 Under the assumption (FEM) or (SLG), let (N )N 1 be a sequence
d
of probability measures on N TN whose entropy with respect to N(0;) is of order
o(N d) :
H (N jN(0;) ) = o(N d ) :
Then, the relative entropy of the state of the process at the macroscopic time t with
respect to N(t;) is also of order o(N d ) :

H N StN jN(t;) = o(N d) for every t 0 :
In this formula, StN stands for the semi-group associated to the generator LN
speeded up by N 2 .
Remark 1.2 Fix a bounded profile 0 : Td ! R+ . The computation performed in

Remark 5.1.2 shows that every sequence of probability measures N with entropy
H (N jN0 ()) of order o(N d ) is such that
H (N jN ) = O (N d )
for every > 0.
Before proving Theorem 1.1, we deduce the conservation of local equilibrium
in the weak sense, as defined in Chapter 3.
Corollary 1.3 Under the assumptions of the theorem, for every continuous function
H : Td ! R and every bounded cylinder function ,
h X Z i
N StN N d

lim
N !1
E H (x=N )x () d
H (u)E t;u [ ] du
( ) = 0:
x2TdN T
Proof. To concentrate exclusively on the essential problems, we assume that the

cylinder function depends on the configuration only through (0) :
() = ((0)) :
Since both functions H () and (t; ) are continuous and since the cylinder function
is bounded, a summation by parts shows that in order to prove the corollary we
only need to check that
lim sup lim sup
`!1 N !1
h X X i
EN StN N d (2 + 1) ` d ((y)) E t;x=N [ ] 0 :
( )
x2TdN jy xj`
By the entropy inequality, for every > 0, the expectation in the previous formula
is bounded above by
h
N N N
N d H ( St j(t;) ) N d log EN(t;)
1 1
+
n X X oi
exp (2 + 1) ` d ((y)) E t;x=N [ ] :
( )
x2TdN jy xj`
At the end of the proof, we shall choose as a function of `. By Theorem 1.1,
the first term converges to 0 as N " 1. On the otherP hand, since the measure
N(t;)Pis product, the random variables (2` + 1) d jy x1j` ((y)) and (2` +
1) d jy x2 j` ( (y )) are independent as soon as jx1 x2 j > 2`. In particular,
by Hölder inequality, the second term is bounded above by
X h X n oi

log EN(t;) exp ((y)) E t;x=N [ ] :
1 1

N d x2TN
d (2 ` + 1) d
jy xj`
( )
This step is explained in more details in the proof of Lemma 1.8. Since the profile
(t; ) is continuous, as N " 1, this sum converges to
Z h X n oi

d
du (2`1+ 1)d log E t;u ( ) exp ((y)) E t;u [ ] ( ) :
T jyj`
Since the cylinder function is bounded, it follows from the elementary identities
ex 1 + x + 2 1x2 ejxj, log(1 + x) x, that this integral is bounded above by
Z h X n o i

d
du (2`1+ 1)d E t;u ( ) ( ( )) y E t;u [ ]
( )
T jyj`

+ 2 (2` + 1) d k k21 exp 2 (2` + 1)d k k1
2 2
:
To conclude the proof, it remains to choose = (2` + 1) d . In this case, by the
law of large numbers, this expression converges to 0 as ` " 1 and than # 0.
The proof of Theorem 1.1 is divided in several lemmas. We start introducing

some notation used throughout this chapter. For > 0, N stands for a reference
invariant measure and N (t) is the Radon–Nikodym derivative of N(t;) with
respect to N :
dN(t;)
N (t) := dN
A simple computation allows to obtain an explicit formula for N (t) because the
measures N(t;) , N are product and the profile (t; ) is bounded below by a
strictly positive constant uniformly in time :
( )
Xh i
N (t) = exp (x) log ((t; x=N )) log Z ((t; x=N )) :
x
In this formula and Z are given by
( ) = (( )) and Z ( ) = ZZ (((( )))) ;
where () and the partition function Z () have been defined in section 2 of Chapter
3. To keep notation as simple as possible, hereafter we denote by N t the measure
N at macroscopic time t :
Nt := N StN ;
by fN (t) = ftN the Radon–Nikodym derivative of N t with respect to the reference
measure N :
N N N
fN (t) := d t = d St
d N d N

and by HN (t) the relative entropy of t with respect to N(t;) :
N

HN (t) = H Nt jN(t;) :
Since N N
t is absolutely continuous with respect to (t;) , the explicit formula for
the relative entropy presented in Theorem A1.8.3 gives that
Z
H N ( t) dNt dNt d N
=
dN(t;) log
dN(t;) (t;)
Z N
= ftN () log ftN (()) N (d) :
t
We turn now to the proof of Theorem 1.1. The strategy consists in estimating
the relative entropy HN (t) by a term of order o(N d ) and the time integral of the
entropy multiplied by a constant :
Z t
HN (t) o( N d ) + 1
HN (s) ds
0
and apply Gronwall lemma to conclude. The first step stated in Lemma 1.4 below
gives an upper bound for the entropy production.
Lemma 1.4 For every t 0,

Z
1
@t HN (t) N 2 LN tN @t tN ftN dN ;
tN
where LN is the adjoint of LN in L2 (N ).
Proof. We have seen in Chapter 5 that fN (t) is the solution of the Kolmogorov
forward equation
@t fN (t) = N 2 LN fN (t) :
Since the profile (t; ) is smooth, a simple computation shows that
Z N
@t HN (t) = N 2 LN ftN log ftN dN
Z
t
N
N L f f @t t d N :
2 N N
N t t
tN
+
Since LN is the adjoint of LN in L2 ( ),

Z
LN ftN dN = 0:
By the same reason, the first expression on the right hand side may be rewritten
as
Z
f N f N
N N t t
t N LN log N d :
2
t t
The elementary inequality
a [log b log a] (b a)
that holds for positive reals a, b, shows that for every positive function h and for
every generator L of a jump process,
h L(log h) L h :
In particular, the last integral is bounded above by
Z Z
N N ftN d N ftN L N d N :
t LN N
2 2
N N t
tN
=
t
We now estimate the upper bound for the entropy production obtained in
the previous lemma using the explicit formula for tN . First of all, a simple
computation shows that ( tN ) 1 N 2 LN tN is given by

g((x)) p(x y) ((((t;t; x=N
y=N ))
X
N2 ))
1 ;
x;y2Td N
that is well defined because (t; ) is strictly positive. On the other hand, the sum

((t; x=N )) p(x y) ((((t;t; x=N
y=N ))
X
N 2
))
1 (1:2)
x;y2Td N
clearly vanishes. Therefore, if stands for the second order differential operator
defined in the beginning of this chapter, since the solution of (1.1) is of class
C 2+e (Td), Taylor expansion gives that
N N 2 LN tN ()
t ( )
1
X ( )((t; x=N ))

+ o(N d ) :
=
x2TdN
((t; x=N )) g((x)) ((t; x=N ))
We shall see in a while the reason for adding a vanishing term to ( tN ) 1
N 2 LN tN .
The identity
Z 0 (') = R(') (1:3)
Z (') '
proved in section 2.3 and the fact that is the solution of equation (1.1) gives that

( tN ( )) 1
@t tN () = @t log tN ()
X (((t; x=N )))
=
( ( t; x=N ))
0 ((t; x=N )) [(x) (t; x=N )] :
x2Td N
To keep notation as simple as possible, we denote by F (t; ) the function of
class C e (Td ), ((t; )) 1 (((t; ))). Up to this point we proved that
N [N 2 LN@t ] ( tN ())
t ( )
1
X n
= F (t; x=N ) g((x)) ((t; x=N )) (1:4)
x2TdN
o
0 ((t; x=N )) [(x) (t; x=N )] + o(N d ) :
It is important to stress that a microscopic Taylor expansion up to the second

order appeared in this formula since, by local equilibrium, the mean value of
g(()) at the microscopic point x is given by E(t;x=N ) [g((0))] = ((t; x=N )).
This explains why we introduced above the term (1.2).
To fully take advantage of the Taylor expansion that appeared, the next step
consists in applying the one block estimate to replace the cylinder function g ( (x))
by (t̀ (x)), the expected value of g ( (x)) under the invariant measure with density
equal to the empirical density of particles in a microscopic box centered at x. We
shall obtain in this way a second order Taylor expansion of (t̀ (x)) around
((t; x=N )) in formula (1.4).
Lemma 1.5 Under the assumptions (FEM) or (SLG), for every t>0
2 3
Z t 1 X
lim lim E N
`!1 N !1
4
d F (s; x=N ) g(s (x)) (s` (x)) ds5 = 0:
0 N
x2TdN
Here E N stands for the expectation with respect to PN , the probability
d
measure on the path space D([0; T ]; NTN ) induced by the Markov process with
generator LN speeded up by N 2 starting from N and, for a positive integer `,
` (x) stands for the empirical density of particles in a cube of length ` centered
at x : X
` (x) = (2` +1 1)d (y ) :
jy xj`
Lemma 1.5 is proved in section 5.4. It permits to replace in formula (1.4) the
cylinder function g ( (x)) by its mean value ( ` (x)). On the other hand, since
F (t; )0 ((t; )) is a continuous function, a summation by parts permits to replace
(x) in the same formula by ` (x). We may thus rewrite ( tN ) 1 fN 2LN @t g tN
as X n
F (t; x=N ) (` (x)) ((t; x=N ))
x2TdN
0 ((t; x=N )) [`(x) (t; x=N )] + o(N d ) :
Here o(N d ) stands for an expression whose expectation is of order o(N d ) as
N " 1 and ` " 1. In conclusion, it follows from Lemma 1.4, Lemma 1.5 and the
computations just performed that for every t > 0 the entropy HN (t) is bounded
above by
hZ t X i
HN (0) + E N F (s; x=N )M (s`(x); (s; x=N )) ds + o(N d ) ;
0
x2TdN
where
M (a; b) = (a) (b) 0 (b) (a b) :
Besides the expectation, all terms in this expression are of order o(N d ) since
we assumed the initial entropy to be of this order. To conclude the proof of the
theorem it remains to show that the expectation is bounded by the sum of a term
of order o(N d ) and the time integral of the entropy multiplied by a constant. This
estimate is obtained through the entropy inequality. We start rewriting the last
expectation as
Z t hX i
ds ENs F (s; x=N ) M (`(x); (s; x=N )) :
0
x2TdN
By the entropy inequality, for every > 0, this integral is bounded above by
Z t
1
ds HN (s)
0
Z t h n X oi
+ 1
ds log ENs; exp F (s; x=N )M (`(x); (s; x=N )) :
x2TdN
( )
0
The next result concludes the proof of Theorem 1.1.
Proposition 1.6 There exists 0 > 0 such that for all 0 s t

lim sup lim sup
`!1 N !1
h n X oi (1:5)
1
N d log ENs; exp 0 F (s; x=N )M (`(x); (s; x=N )) 0:
x
( )
A rigorous and complete proof of this result is a bit long. The idea is however
simple, and relies on large deviations arguments. Since the measure N(s;) are
product, the random variables ` (x1 ), ` (x2 ) are independent as soon as jx1 x2 j >
2`. In particular, the Laplace–Varadhan theorem and a large deviations principle
for i.i.d. random variables give an upper bound for the left hand side of (1.5) of
the form Z n o
du sup 0 F (s; u)M (; (s; u)) J(s;u) () :
Td
where J () is a rate function strictly convex vanishing at . Since M (; ) also
vanishes at , is quadratic close to and linear at infinity, it will not be difficult
to show that the supremum vanishes for 0 small enough.
We conclude this section with a rigorous proof of Proposition 1.6. Fix a se-
quence of i.i.d. random variables with distribution
P0 [X1 = k] ( )k k2N
g(1) g(k) Z (( )) ; (1:6)
1
=
and recall the large deviations principle :

PN
Lemma 1.7 The sequence N 1
k=1 Xk satisfies a large deviations principle
with rate function given by
8
>
< ( )
Z (()) 0
J1 () =
log
( ) log
Z (( )) for
(1:7)
>
: 1 otherwise :
We refer to Deuschel and Stroock (1989) for a proof of this large deviations
principle. This lemma provides an upper bound for the left hand side of (1.5).
In order to deduce this bound, we need to introduce some notation. Recall that
' (which might be infinite) stands for the radius of convergence of the partition
function Z () defined in section 2.3. Let : Td ! R+ be a continuous function
bounded by K2 . We shall denote by N() the product measure with slowly varying
parameter associated to .
Lemma 1.8 Let G : Td R+ ! R be a continuous function such that

sup jG(u; )j C0 + C1 for all 2 R+ (1:8)
u2T d
where C0 is a finite constant and C1 a constant bounded by log[' =(K2 )] :
C1 < ' :
log
(K2 )
Then,
X
lim sup lim sup
1
N d log EN exp G(x=N; ` (x))
`!1 N !1 ( )
x2TdN
Z

du sup G(u; ) J1(u) () ;
T d 0
where the rate function J1 () is defined in (1.7).
If the partition function Z () is finite on R+ , the assumption in the previous

lemma on C1 requires only C1 to be a finite constant.
Proof. Since ` (0) depends on the variables (x) only for jxj `, ` (x) and
` (y) are independent under N() for jy xj 2` + 1. We shall take advantage of
this property to decompose the expectation in a product of simpler terms. Assume,
Pgenerality, that 2` + 1 divides N .
without loss of
The sum x G(x=N; ` (x)) can be rewritten as

G x + (2N` + 1)y ; ` (x + (2` + 1)y) ;
X X
x2` y;x+(2`+1)y2TdN
where ` is a cube of length 2` + 1 centered at the origin :
` = f `; : : :; `gd :
It is important to remark that the variables f ` (x+(2`+1)y ); y g are independent
under N() for each x fixed. Therefore, by Hölder inequality and by independence,

X
log E N exp
()
G(x=N; ` (x))
x2TdN
is bounded by
X h
1
(2` + 1) d log EN
x2`
( )

G x + (2N` + 1)y ; ` (x + (2` + 1)y)
X i
exp (2` + 1)d
y
X h i
1
E N ` d G(x=N; ` (x)) :
=
(2` + 1)d
log
exp (2 + 1)
x2TdN
( )
For a positive , let P be the probability in (R+ )N corresponding to a se-

quence of i.i.d. random variables with distribution given by (1.6) and denote by
E expectation with respect to P . Since, by assumption, the function G(; ) and
the profile () are continuous and the family faN ; a 0g of product measures
defined in (2.3.6) is weakly continuous in virtue of Lemma 2.3.8; the last line
divided by N d converges, as N " 1, to
Z

du log E(u) exp (2` + 1)d G(u; X̄(2`+1)d ) ; (1:9)
1
(2` + 1)d T d
where, for a positive integer k , X̄k stands for the average of the first k elements :
k
1 X
X̄k =
k j=1 Xj :
If G was bounded, it would follow from the large deviations principle for the
sequence (Xj )j 1 stated in Lemma 1.7 and the Laplace–Varadhan theorem (cf.
Theorem A2.3.1), that the limit, as ` " 1, of the last line is equal to
Z

du supG(u; ) J1(u) () :
Td >0
A cut off argument permits to reduce the general case to the case of bounded
functions. We first compute, for each fixed u, the limit of the expression in (1.9).
We shall than argue to exchange the limit with the integral.
We start with the upper bound. Fix = (u). For A > 0, denote by GA (; )
the function G truncated at the level A :
GA (u; ) = G(u; )1fjj Ag + G(u; A)1fjj > Ag :
By Hölder inequality and by the assumption (1.8),
h i
E ` G(u; X̄`)
exp

E exp ` GA (u; X̄`) + E exp `[C0 + C1X̄` ] 1fX̄` Ag
h n oi
E exp ` GA (u; X̄`)
h n X̀ oi
+ eC0 ` E1=q 1fX̄` Ag E1=p exp pC1 Xk :
k=1
In this last formula, p and q are conjugates : p > 1 and p 1 + q 1 = 1 and

p is chosen such that E [epC1(0) ] = Z (epC1 ( ))=Z (( )) < 1. This choice
is possible because C1 < log[' =(K2 )] and the function is bounded by K2 .
Therefore, since the variables Xk are independent, we have that

X̀

1
` log E exp pC1 Xk = log E epC X 1 1
k=1
is finite. On the other hand, by the large deviations principle for the sequence
(Xk )k1 ,
log P [jX̄` j A] J1 (A) 1:

1
lim lim sup
A!1 `!1 ` lim
A!1
=
The last equality follows from the explicit formula for the rate function J1 .
Finally, GA being bounded, by the large deviations principle for the sequence
(Xk )k1 and by the Laplace–Varadhan theorem, for every A,

E exp f` GA (u; X̄` )g = sup GA (u; ) J1 () :
1
lim
`!1 ` log
>0
We have thus proved that

E exp f` G(u; X̄`)g GA (u; ) J1 () :
1
lim sup
`!1 ` log lim sup
A!1 >0
To derive the upper bound, it remains to show that the term on the right hand
side of the last inequality is equal to

sup G(u; ) J1 () :
>0
It is enough to show that we may restrict the supremum on the right hand side
of the inequality to a compact subset of [0; 1) because for each compact there
exists a real A0 such that G and GA0 coincide in this compact for every A > A0 .
By assumption (1.8), we have that
G(u; ) J1 () C0 + C1 J1 () :

Taking the derivative of the rate function J1 () and keeping in mind identity (1.3)
and that J1 ( ) = 0, we deduce that
Z
J1 () () d :
=

log
( )
From the choice of the constant C1 ,
J 1 ()
1 ' > C :
!1 ( )
lim = log 1
In consequence,
n o
lim sup sup GA (u; ) J () lim C + C J1 ()
!1 0 1
1
= 1
!1 A
and we may restrict the supremum to a compact subset of [0; 1).
A lower bound for the expression in (1.9) can be proved with similar argu-
ments :

log E exp f` G(u; X̄`)g sup fG(u; ) J1 ()g :
1
lim inf
`!1 ` >0
It remains to justify the exchange of the limit and the integral. From assumption
(1.8) and from the definition of the constant C1 , we deduce a bound, uniform over
` and u, of

E(u) exp (2` + 1)d G(u; X̄(2`+1)d ) :
1
(2` + 1)d
log
The theorem of the dominated convergence permits to conclude the proof of the
lemma.
Applying the previous lemma to the function

G(u; ) = F (s; u) () ((s; u)) 0 ((s; u))[ (s; u)]
we conclude the first step of the proof of Proposition 1.6. We summarize the
conclusions in the next corollary. To state it notice that the just defined function
G is such that n o

sup G(u; ) kF k1 2g + sup ( ) + sup 0 ( )
u2Td 2[0;K2 ] 2[0;K2 ]
because (0) = 0 and, by Corollary 2.3.6, 0 ( ) () g ( ) for
so that () g and 0 () g . In this formula, kF k1 stands for the
L1 ([0; t] Td) norm of F :
kF k1 = sup jF (s; u)j :
s;u)2[0;t]Td
(
Corollary 1.9 Recall that K2 stands for the upper bound for the initial profile 0 .

1 = 2kF k1 g log ('K ) :
Let
1 2
Then, for all < 1 and all 0 s t,
lim sup lim sup
`!1 N !1
X
1
N d log ENs; exp F (s; x=N )M (`(x); (s; x=N ))
x2TdN
( )
Z
du sup F (s; u)M (; (s; u)) J s;u) () : 1
>0
(
Td
To conclude the proof of Proposition 1.6, we have to show that the right hand
side of the previous inequality is non positive for all sufficiently small. This
result follows from the next lemma.
Lemma 1.10 For every 0 < K1 < K2 < 1,

C2 := sup
jM (; )j < 1 :
2[K ;K ] J1 ()
0
1 2
Proof. We first choose 0 < < K1 =2. Throughout this proof, K1 and K2 stand
respectively for K1 and K2 + . We decompose the set [K1 ; K2 ] R+ in three
disjoint subsets ( << ; and >> ) and prove the result in each of
these subsets by different ways. We start with the region .
Consider the set

E1 = (; ) 2 R+ [K1; K2] ; K1 K2 :
Let A be the constant defined by
A= sup j00 ()j :
K K
1 2
Taylor expansion permit to bound M on E1 :

jM (; )j A2 ( )2 for (; ) 2 E1 :
On the other hand, a simple computation taking advantage of the relation (2.3.5),
permits to compute the first two derivatives of the rate function J1 :

@ J ()
1 () ; @2 J1 () 0 ()
= log
( ) =
()
In particular, both J1 and its derivative vanish at . Let B be the constant defined
by 0
B = K inf (@ J )( ) = inf
2 1 ( )
:
1
K2 K1 K2 ( )
B is strictly positive because and 0 are smooth functions strictly positive on
(0; 1). J1 and its derivative vanishing at , by Taylor expansion and the definition
of B ,
J 1 ( ) B ( ) 2
2
for (; ) 2 E : 1
In conclusion, for (; ) 2 E1

jM (; )j A C3 < 1 :
J1 () B =:
We turn now to the set

E2 = (; ) 2 R+ [K1; K2] ; K2 :
Notice that on this set + . On the one hand, since ( ) g and
K2 > ,
0 0
jM (; )j () + ( ) + [( ) + ( )]
1 1 1
2 [g + 0 ( )] :
On the other hand, since J1 ( ) = 0, (@ J1 )() = log (()=( )) and + , by
an integration by parts,
Z 0
J ()
1 1
=
1
[ ] (()) d
Z +=2 0
2( + ) () d :
()

This last expression is denoted by C ( ). Thus,
jM (; )j sup 2[g + 0 ( )] =: C < 1
sup
(; )2E J1 ()
2 2[K ;K ] C ( ) 1
4
2
because 0 () et C () are continuous

and positive.
Finally, for the set E3 = (; ) 2 R+ [K1 ; K2 ] ; K1 we proceed in
the following way. Since on this set K1 , by Taylor expansion
jM (; )j 1
( ) 2 sup j00 ()j 12 2 sup j00 ()j :
2 K2
On the other hand, repeating the arguments presented for the set E2,
Z 0 Z 0 () d :
J1 () [ ] (()) d 2
=2 ( )
=

We denote by 2 1 C ( ) this last function. From these two inequalities, we deduce
an estimate on the set E3 :
jM (; )j K22 supK j00 ()j

=: C5 < 1 :
C ( )
2
sup sup
(; )2E
3
J 1 ()
2[K ; K ] 1 2
It remains to set C2 = C3 _ C4 _ C5 to conclude the proof of the lemma.

Corollary 1.11 There exists 0 > 0 such that for all < 0

sup F (s; u) M (; (s; u)) J1(s;u) () 0 :
s;u)2[0;t]Td
(
Proof. Straightforward from Lemma 1.10 because F is bounded in [0; t] Td

and the range of (; ) is contained in [K1 ; K2 ].
Remark 1.12 Without the assumption concerning the order of magnitude of the
entropy at time 0, the same proof gives an upper bound for the entropy production.
Indeed, if for t 0 H (t) stands for the limit of the specific entropy :
H (t) := lim sup N d HN (t) ;
N !1
the arguments presented above show that there exists > 0 such that
Z t
H (t) H (0) +
1
H (s) ds :
0
Remark 1.13 The special form of the hydrodynamic equation (1.1) played no
special role. We just needed the existence of a smooth solution of the hydrody-
namic equation. In particular, the relative entropy method extends to a large class
of interacting particle systems that includes conservative asymmetric dynamics,
described by first order quasi–linear hyperbolic equations, up to the appearance of
the first shock.
Remark 1.14 As noticed in Chapter 5, for attractive processes either one of

the assumptions (SLG) or (FEM) is fulfilled. Theorem 1.1 applies therefore to
attractive zero range processes.
The method presented in this chapter is due to Yau (1991). It was extensively used
to investigate the first order correction to the hydrodynamic equation. This topic
is discussed in the last section of the next chapter.
Euler equations. Olla, Varadhan and Yau (1993) considered a superposition of
an Hamiltonian dynamics with an infinite range stochastic noise on the velocities
that exchanges momenta and preserves the conserved quantities (the density, the
momentum and the energy). Adapting the relative entropy method to this context,
they proved that the conserved quantities evolve according to the Euler equations
8
> X
3
>
>
>
>
@t + @uj fj g = 0 ;
>
> j =1
>
>
>
>
< X
3
>
@t (i ) + @uj fi j + i;j P g = 0 ;
>
>
> j =1
>
>
>
> X
3
>
>
>
: @t (e) + @uj fej j P g = 0 ;
j =1
in the time interval where the solutions of these equations are smooth. In this
formula stands for the density, for the velocity per particle, e for the energy
per particle and P is the pressure, a function of , , e.
One of the main ingredients in this derivation is the proof of the ergodicity of
the dynamics. Liverani and Olla (1996) proved ergodicity for Hamiltonian systems
superposed to finite range stochastic interactions on the velocities : they proved
that translation invariant measures that are stationary for the deterministic Hamil-
tonian dynamics, reversible for the stochastic dynamics and have finite specific
entropy are convex combinations of Gibbs states. Fritz, Liverani and Olla (1997)
removed the requirement of reversibility proving that all translation invariant sta-
tionary states of finite specific entropy are reversible with respect to the stochastic
evolution. For lattice systems this question has been solved by Fritz, Funaki and
Lebowitz (1994) : They proved that all translation invariant stationary states with
finite local entropy are microcanonical Gibbs states in the case of Hamiltonian
systems with a local random perturbation that conserves the energy.
Cahn–Hilliard equations. Bertini, Landim and Olla (1997) deduced the Cahn–
Hilliard equation
@t = (F 0 ())
from a stochastic microscopic Ginzburg–Landau dynamics. Giacomin and Lebo-
witz (1997a) examined an interacting particle system evolving according to a
local mean field Kawasaki dynamics and showed that the hydrodynamic equation
is given by

@t = r A()r F () ;
where A() = (1 ), is the inverse of the temperature and
Z Z Z
F ( ) =
1
Td du s ((u))
1
2 Td
du d dv J (u v)(u)(v) :
T
In this formula s() = log (1 ) log(1 ) and J () is the mean field inter-
action. Giacomin and Lebowitz (1997b) compared the solution of these equations
with the behavior of solutions of Cahn–Hilliard equations.
Reaction–diffusion equations. De Masi, Ferrari and Lebowitz (1986) considered
a superposition of Glauber and speeded up Kawasaki dynamics to obtain reaction–
diffusion equations. In these models at most one particle is allowed per site. To
describe the stochastic evolution, fix a positive cylinder function c( ). For each
site x, at rate N 2 the occupation variables (x) and (x + ei ) are exchanged
and at rate c(x ) the occupation variable (x) is flipped. De Masi, Ferrari and
Lebowitz (1986) proved that the hydrodynamic behavior of the system is given
by the solution of the reaction–diffusion equation
@t = + F () ;
where F () = E [(1 2 (0))c( )] and is the Bernoulli product measure of
density . Mourragui (1996) extended this analysis to zero range processes with
creation and annihilation of particles applying the relative entropy method. Nappo,
Orlandi (1988) and Nappo, Orlandi and Rost (1989) deduce a non linear reaction–
diffusion equation for Brownian particles moving on Rd with an interaction that
kills a particle at some rate which depends on its distance to the others.
Noble (1992) and Durrett and Neuhauser (1994) investigate the behavior of
a superposition of an attractive Glauber dynamics with a speeded up Kawasaki
dynamics. Using the result of De Masi, Ferrari and Lebowitz (1986) they prove
the existence of non trivial stationary states for a wide class of examples provided
the stirring rate is large enough.
Bramson and Lebowitz (1991) consider a system with two types of particles
that evolve according to independent random walks. When two particles of dif-
ferent type meet, they annihilate each other. They investigate the limit density
of each type of particles and they examine the spatial structure of the process.
In this model the critical dimension is 4 and while in dimension d < 4 there is
segregation of types of particles, in dimension d > 4 there is coexistence of the
two types.
Stefan problems. Chayes and Swindle (1996) analyze a one-dimensional ex-
clusion process with two types of particles. The first type of particle evolves as
an usual exclusion process and the other type is kept frozen. Superposed to this
evolution there is an annihilation mechanism that either eliminates one particle of
each type when they meet or that transforms a free particle in a frozen particle
when their distance reaches 1. The hydrodynamic behavior of these systems for a
class of initial states is shown to be described by the solution of a Stefan problem
with one free boundary :
8 8
>
< @t = ; >
< B (0) = B0 ;
(0; u) = 0 (u) for 0 u B0 ; (t; B (t)) = 0 ;
> >
(dB=dt)(t) = (@u )(t; B (t)) :
: :
(t; 0) = a(t) ;
Landim, Olla and Volchan (1997), (1998) considered a nearest neighbor one-
dimensional symmetric simple exclusion process with an asymmetric tagged par-
ticle. The hydrodynamic behavior is given by the solution of the Stefan problem
8
>
>
@t = (1=2) ;
>
< vt = (@u log )(t; vt+) = (@u log )(t; vt ) ;
>
>
>
pf1 (t; vt +)g = qf1 (t; vt )g ;
:
(0; ) = 0 () :
Gravner and Quastel (1998) derived the hydrodynamic equation of a system where
particles are created at a finite number of fixed sites and then perform zero range
random walks. Each particle jumping to a site occupied by less than particles is
kept frozen at this site. They showed that the macroscopic behavior of this process
is described by the solution of a Stefan problem.
Carleman and Broadwell equation. The Carleman equation is a special case of

a discrete Boltzmann equation. It describes the evolution of two types of particles
on R whose density 0 (t; ), 1 (t; ) evolves according to
(
@t a + m @u a = (1 a)2 (a)2 ;
a (0; ) = a;0 () ; a = 0; 1 :
De Masi and Presutti (1991) deduce this equation by the method of truncated
correlation functions from a microscopic model where two type of particles evolve
on the discrete torus TdN according to independent asymmetric random walks
speeded up by N . Particles of type 0 jump only to the right nearest neighbor
while particles of type 1 jump only to the left nearest neighbor. Superposed to this
displacement there is a collision dynamics whose generator is
X
1 X
(Lc f )( ) = (x; a)[(x; a) 1][f ( 2dx;a + 2dx;1 a ) f ()] :
a=0 x2TdN
In this formula, (x; a) stands for the total number of a-particles at site x and dx;a
is the configuration with no particles but one a-particle at site x.
Carleman equation is derived by Caprino, De Masi, Presutti and Pulvirenti
(1989, 1990) from diffusion processes evolving on R. The two-dimensional version
of the system gives rise to the Broadwell equation and is derived in Caprino, De
Masi, Presutti and Pulvirenti (1991) in the time interval where the equation admits
smooth solutions.
Boltzmann equations. Rezakhanlou (1996a) considered an interacting particles
system from which he deduced a discrete Boltzmann equation. Fix a finite set
I of labels. Particles evolve on Z, each one with a label a in I . A particle
with label a evolves independently according to P a random walk with finite range
transition probability pa () with mean drift qa = x xpa (x). Two particles at the
same site with labels a, b collide with probability N 1 , the rescaled interdistance
between sites. If they collide they gain new labels a0 , b0 at rate K (a; b; a0; b0 ). These
rates are chosen symmetric and vanish when the conservation of momentum is
violated : K (a; b; a0 ; b0 ) = K (b; a; a0 ; b0 ) = K (a; b; b0 ; a0 ) and K (a; b; a0 ; b0 ) = 0
when qa + qb 6= qa0 + qb0 or when qa = qb .
Starting from a product measure associated to a bounded integrable profile
0 () = fa0 (); a 2 Ig, Rezakhanlou (1996a) proved that the macroscopic evolu-
tion of the empirical measure is described by the discrete Boltzmann equation
X
@t a + qa @u a = K (c; d; a; b)cd K (a; b; c; d)ab : (2:1)
b;c;d
Rezakhanlou (1996b) proved the propagation of chaos (cf. Chapter 8 for the
terminology) for this model. Rezakhanlou and Tarver (1997) deduced the hydro-
dynamic equation (2.1) for a one-dimensional, continuous version of the previous
model. Here, instead of moving according to random walks, each particle moves
deterministically with a velocity determined by its label. The collision rules are
similar but the assumption on conservation of momentum is dropped. The proof
of the above results in higher dimension remains an open problem.
Rezakhanlou (1997) proved the equilibrium fluctuations in any dimension for
the model introduced by Rezakhanlou and Tarver (1997). He showed that the
rescaled fluctuation field converges to an Ornstein–Uhlenbeck process with a drift
given by the linearized Boltzmann equation.
Caprino and Pulvirenti (1995) consider a system of N identical particles mov-
ing freely on R until they collide. Particles collide independently with probability
". They prove that in the Boltzmann–Grad limit, i.e., as N " 1 and "N ! ,
the density profile converges to the solution of a Boltzmann equation, globally in
time. Caprino and Pulvirenti (1996) consider the same evolution with stochastic
reflection at the boundary of the interval [0; 1]. They show that the density profile
of the unique invariant measure converges in the Boltzmann–Grad limit to the
solution of the Boltzmann stationary equation.
Degenerate diffusions. Rezakhanlou (1990) considers Ginzburg–Landau models
where the equilibrium states are canonical Gibbs measures for a finite range inter-
action. He deduces the hydrodynamic behavior of the system under assumptions
on the interaction that do not exclude the possibility of phase transition, in which
case the diffusion coefficient might vanish. Carmona and Xu (1997) extend this
result to the case of random finite range interactions.
Lebowitz, Orlandi and Presutti (1991) consider a class of one-dimensional,
infinite volume exclusion processes with a small drift toward the region of higher
density. They deduce a non–linear parabolic hydrodynamic equation of type @t =
@u (D()@u ). The diffusion coefficient might be negative on an interval (a; b). In
this case the hydrodynamic equation is proved only for initial data taking values
on [0; a) [ (b; 1]. Giacomin (1991) extends this investigation to reversible models.
Computer simulations suggest that the system undergoes phase segregation on the
scale of the interaction and that the system does not change on the macroscopic
scale, indicating that the diffusion coefficient vanishes in this region. There are,
however, no rigorous results.
Carlson, Grannan, Swindle and Tour (1993) prove the hydrodynamic behavior
of a one-dimensional symmetric exclusion process in which a particle at x jumps
to x + y at rate c(jy j) provided all sites between x and y are occupied. If c() decays
slowly, the diffusion coefficient D() has a singularity at = 1 : lim!1 D() = 1.
Interface motion. In dimension 1 the hydrodynamic behavior of an interacting
particle system can be interpreted as the motion of an interface (cf. for instance De
Masi, Ferrari and Vares (1989)). Indeed, consider to fix ideas a nearest neighbor
symmetric zero range process on f0; : : : ; N g with
P reflexive boundary conditions.
The generator of this process is given by LN = 0xN 1 Lx;x+1 , where
(Lx;x+1f )( ) = g((x))[f (x;x+1) f ()] g((x + 1))[f (x+1;x) f ()] :

+
P
Denote by the “integral" of the configuration : (x) = 0yx (y ). t is
itself a Markov process with one conserved quantity (N ) and with generator
LN = P0xN 1 Lx;x+1, where

(Lx;x+1 f )() = g ((x) (x 1))[f ( dx ) f ()]
+ g((x + 1) (x))[f ( + dx ) f ()] :
In this formula dx stands for a configuration with no particles but one at x.
Denote by N (t; ) : [0; 1] ! R+ the profile N (t; u) = N 1 tN 2 ([uN ]). It
is easy to check that N (t; ) converges to some nondecreasing function h(t; )
if and only if the empirical measure associated to the t process converges to
an absolutely continuous measure with density (t; u) = (@u h)(t; u). In particular,
it follows from the hydrodynamic behavior of the symmetric zero range process
that N (t; ) converges to h(t; ), as N " 1, where h(t; ) is the solution of the
equation 8
< @t h = @u (@u h) ;
>
>
h(0; ) = h0 () ;
:
h(; 0) = 0 ; h(; 1) = a0 :
Marchand and Martin (1986) used similar ideas to investigate the macroscopic
behavior of a droplet evolving according to a Glauber dynamics in Z2.
The whole problem is to extend these ideas to higher dimensions. At the
moment there are very few rigorous results. Naddaf and Spencer (1997) and Funaki
and Spohn (1997) considered d-dimensional, continuous spins Ginzburg–Landau
models on a cube with periodic boundary conditions. The spins ft (x); jxj N g
evolve according to the differential equations
X p
dt (x) = V 0 (t (x) t (y)) dt + 2dWt (x) ;
jx yj=1
where Wx (t) is a collection of independent Brownian motions and V is a strictly
convex, smooth, symmetric potential. Naddaf and Spencer (1997) examined the
fluctuations of density field at a fixed time.
Defining N (t; u) by N (t; u) = N 1 t ([uN ]), Funaki and Spohn (1997)
proved that starting from a measure N associated to some profile h0 : Td ! R,
N (t; ) converges in L2 to the solution of
8
> d
X
>
< @t h = @uj j (rh) ;
j =1 (2:2)
>
>
:
h(0; ) = h0 () ;
where j = @uj and is the surface tension. Giacomin, Olla and Spohn (1998)
derived the equilibrium fluctuations for this model.
A second possible way to derive a macroscopic interface motion from micro-
scopic local dynamics is to examine zero temperature Glauber dynamics. Consider,
for instance, the Ising model starting from a configuration in which the + domain
is separated from the domain by a single contour without self intersections.
The zero temperature Glauber dynamics forbids flips that increase the energy. We
modify slightly this dynamics excluding also flips that create a second contour
in the separation of the + and domains. Such a model has been considered by
Spohn (1993) who proved the hydrodynamic behavior of a two-dimensional sys-
tem for some special initial configurations. The macroscopic evolution is shown to
be described by solutions of the equation (2.2). Landim, Olla and Volchan (1997)
investigated the interface motion obtained by the zero temperature dynamics of a
Potts model.
A third possible approach would consist in studying the evolution of a +
region in a see of for a reversible Ising model without external field at low
enough temperature to have phase coexistence. There are no rigorous results in
this direction and we refer to Spohn (1993) for an overview on the problems and
on the available techniques. Recently Evans and Rezakhanlou (1997) derived the
hydrodynamic equation of a sandpile model.
Motion by mean curvature, Ising models with long range interactions
Kac potential. Fix a smooth potential J : R ! R symmetric and with compact
support, > 0, that will represent the inverse of the range of the interaction,
and an external field h > 0. Define the Kac potential J : Zd Zd ! R+ by
J (x; y) = d J ( ky xk) and the formal energy H : f 1; 1gZd ! R of a spin
configuration by
1 X X
H ( ) =
2
J (x; y)(x)(y) h (x) :
x;y2Zd x2Zd
Fix > 0, the inverse of the temperature, and consider the Glauber dynamics
associated to the energy H at temperature 1 . This is the Markov process on
f 1; 1gZd whose generator L acts on cylinder functions as
X
(L f )( ) = c(x; )[f (x) f ()] :
x2Zd
x
Here is the spin configuration obtained form by flipping the spin at x :

(y ) y 6= x ;
( x )(y ) =
if
(x) if y = x;
and c(x; ) is the jump rate given by
expf (=2)(xH )( )g
c(x; ) =
expf (=2) (x)(xH )( )g + expf(=2) (x)(xH )( )g
;
and (x H )( ) = H ( x ) H ( ). Notice that jump rates fc(x; ); x 2 Zdg
satisfy the detailed balance condition :
cx(x ) e(x H )() :
cx () =
Lebowitz-Penrose limits A basic question in the theory of Ising models with long
range interactions is the investigation of the behavior of the system as the range
of the interaction increases to infinity, i.e., as # 0. This limit is known as the
Lebowitz–Penrose limit (Lebowitz and Penrose (1966)).
De Masi, Orlandi, Presutti and Triolo (1994) considered this Ising model start-
ing from a product measure with marginals given by
E [(x)] = m0 ( x) ; x 2 Zd ;
where m0 is a profile in C 1 (Rd ) with bounded derivatives. They proved that for
each n 1,
n
hY i n
Y

lim sup E
!0
t (xi ) m(t; xi ) = 0;
i=1 i=1
provided m is the unique solution of
(
@t m + m tanhf (J m + h)g = 0 ;
(2:3)
m(0; ) = m0 () :
In the previous formulas, t is the state of the Markov process at time t, the
supremum is taken over all (x1 ; : : : ; xn ) in (Zd)n such that xi 6= xj for i =
6 j and
J m stands for the convolution of J and m :
Z
(J m)(t; u) = dv J (ku vk)m(t; v) :
Rd
Notice that time is not rescaled so that each spin undergoes a finite number of
flips in this regime. The deterministic limit is obtained by means of a law of large
numbers since for small a large numbers of spins feel the effect of the same
potential. This limit is called by the authors a mesoscopic limit to differentiate it
from asymptotic behaviors where time is rescaled.
Motion by mean curvature. Assume now that the potential is nonnegative and
normalized,
R that the external field vanishes and that the temperature is below 1 :
J 0, Rd J (kuk)du = 1, h = 0, > 1. = 1 is the inverse critical temperature
in the Lebowitz–Penrose limit. Denote by m the strictly positive solution of the
equation
m = tanh(m) :
m are the magnetization of the two extremal Gibbs states in the limit ! 0.
To define the macroscopic time and space variables, let
= p 1
; " = = p
:
log 1 log 1
We shall rescale space by " and time by 2 so that the macroscopic space variable
is equal to x" and the macroscopic time variable is equal to t 2 .
Fix a compact domain 0 of Rd whose boundary 0 is a C 1 connected

surface. Let t evolve according to the motion by mean curvature with normal
velocity ( depends only on the potential J and on the instanton solution of
(2.3). Its explicit form is given in De Masi, Orlandi, Presutti and Triolo (1994),
equation (5.2a)). We refer to Evans and Spruck (1991) for the terminology of
motion by mean curvature. Assume that is smooth for 0 and denote
by the compact domain whose boundary is .
d
Let " , be a product measure on f 1; 1gZ such that

m if x 2 "0 ;
E" [(x)] =
m otherwise :
De Masi, Orlandi, Presutti and Triolo (1994) proved that the Ising process t with
space rescaled by " and time rescaled by 2 (t = "t 2 ) converges, as # 0,
to m 1ft g m 1fct g. It is not a problem to generalize this result to the case
of several initial interfaces. Buttà (1994) extended this result to all times in the
two-dimensional case, where the unique singularity is the shrinking of a surface
to a point. Katsoulakis and Souganidis (1995) extend this convergence beyond the
appearance of the first geometric singularity of the interface.
De Masi, Orlandi, Presutti and Triolo (1996a,b) investigate the fluctuations
around the mesoscopic limit showing that they converge to a generalized Ornstein–
Uhlenbeck process. They prove also the existence of a deterministic time scale
(log 1 ) where phase separation occurs for a system starting from a Bernoulli
product measure with zero average.
Interface dynamics and reaction–diffusion equations. Bonaventura (1995) con-
siders a spin system on Zd, where a Glauber dynamics is superposed to a Kawasaki
dynamics. Speeding up the Kawasaki dynamics by " 2 and rescaling space by
", one obtains an interacting particle system whose hydrodynamic equation is a
reaction–diffusion equation of type
@t m = m V 0 (m) :
Assume that V is a symmetric double well potential. Exploiting the connection be-
tween motion by mean curvature and reaction–diffusion equations with vanishing
viscosity (cf. Evans, Soner and Souganidis (1992)) and the scaling properties of
the reaction–diffusion equation (the fact that n(t; u) defined by n(t; u) = m(t; u)
is the solution of @t n = 2 n V 0 (n)), Bonaventura proves the existence of
a > 0 for which the macroscopic behavior of the process when the Kawasaki
dynamics is speeded up by N 2 a , is given by the motion of an interface evolving
by mean curvature up to the appearance of the first singularity. Katsoulakis and
Souganidis (1994) prove the same result for processes evolving on a torus without
the assumption of smoothness of the interface.
Interacting diffusions. Varadhan (1991) investigated the evolution of reversible
and repulsive interacting Brownian motions on the one-dimensional torus. He
proved that starting from an initial state associated to some profile 0 : Td ! R
whose entropy is of order N , the empirical measure converges to a Lebesgue abso-

lutely continuous measure whose density is the solution of the nonlinear parabolic
equation
@t = (1=2)P () ;
where P () is the pressure. Olla and Varadhan (1991) extended this investigation
to interacting Ornstein–Uhlenbeck processes, whose dynamics is not reversible.
Uchiyama (1994) removed the assumption made in Varadhan (1991) concerning
the initial entropy bound.
Weakly interacting diffusions. Dittrich (1987) consider a system of independent
Brownian motions on Rd , d 3, that are replaced after exponential random times
by a random number of particles with finite second moment and first moment equal
to 1. He proves that the hydrodynamic behavior is described by the heat equa-
tion and deduces the non equilibrium fluctuations. Dittrich (1988a,b) proves the
conservation of local equilibrium and non equilibrium fluctuations for independent
Brownian motions evolving in the interval [0; 1] with reflexion at the boundary and
local destruction of particles. The macroscopic behavior of the process is shown
to be governed by reaction–diffusion equations.
Oelschläger (1984) considers a finite number of diffusion processes evolving
on Rd according to a mean–field type interaction and shows that the empirical
distribution converges, as the number of particles increases to infinity, to a de-
terministic measure–valued process. Oelschläger (1985) investigates the evolution
of finitely many weakly interacting diffusions and deduces a non linear parabolic
equation as hydrodynamic limit for the process. The nonequilibrium fluctuations
are studied in Oelschläger (1987) and a reaction–diffusion equation is obtained in
Oelschläger (1989) for a model where, besides the space evolution, particles are
created and destroyed.
7. Hydrodynamic Limit of Reversible Nongradient
Systems
We investigate in this chapter the hydrodynamic behavior of reversible nongradient

systems. To fix ideas we consider one of the simplest examples, the so called
symmetric generalized exclusion process. This is the Markov process introduced
in section 2.4 that describes the evolution of particles on a lattice with an exclusion
rule that allows at most particles per site. Here is a fixed positive integer greater
or equal than 2. The generator of this Markov process acts on cylinder functions
as X
(LN f )( ) = (1=2) r((x); (y))[f (x;y) f ()] ; (0:1)
x;y2TNd
jx yj=1
where r(a; b) = 1fa > 0; b < g and x;y is the configuration obtained from
moving a particle from x to y .
Before proceeding we explain the terminology. For a site x and 1 i d,
denote by Wx;x+ei the instantaneous current from x to x + ei , i.e., the rate at which
a particle jumps from x to x + ei minus the rate at which a particle jumps from
x + ei to x. With this definition, for nearest neighbor interacting particle systems,
we have that
d
X
LN (x) = fWx ei ;x Wx;x+ei g :
i=1
For the generalized exclusion process considered in this chapter the current Wx;x+ei
writes

Wx;x+ei = (1=2) 1f(x) > 0; (x + ei ) < g 1f (x + ei ) > 0; (x) < g :
In contrast with zero range processes, where Wx;x+ei = g ( (x)) g ( (x + ei )),
the current can not be written as a difference x h x+ei h for some cylinder
function h. This characteristic of nongradient systems adds a major difficulty in
the derivation of the hydrodynamic behavior of the process.
d
Fix a sequence of initial probability measures N on f0; : : : ; gTN and denote
d
by PN the probability measure on D(R+ ; f0; : : :; gTN ) induced by the Markov
process with generator LN defined by (0.1) speeded up by N 2 and the measure
N . Hereafter E N stands for expectation with respect to PN .
We have presented in Chapters 4 and 5 a general method to deduce the hydro-
dynamic behavior of interacting particle systems. We started considering a class
142 7. Hydrodynamic Limit of Reversible Nongradient Systems
of martingales associated to the empirical measure : for each smooth function

H : [0; T ] Td ! R, let M H;N (t) = M H (t) be the martingale defined by
Z t
M H (t) = < tN ; Ht > < N ; H0 >
0 (@s + N 2 LN ) < sN ; Hs
> ds :
(0:2)
0
A simple computation shows that its quadratic variation is of order O(N d ) so

that by Doob inequality, for every T > 0 and > 0,
h i
lim PN
N !1
sup jM H (t)j = 0: (0:3)
0tT
P
Since LN (x) = 1id fWx ei ;x Wx;x+ei g, and Wx;x+ei = x W0;ei , a
spatial summation by parts permits to rewrite the martingale M H (t) as
Z t
M H (t) = < tN ; Ht > < N ; H0 >
0 < sN ; @s Hs > ds
0
d Z
X t X (0:4)
N1 d (@uNi H )(s; x=N )x W0;ei (s) ds ;
i=1 0
x2TdN
where @uNi represents the discrete derivative in the i–th direction :
h i
(@uNi H )(x=N ) = N H ((x + ei )=N ) H (x=N ) :
At this point we face the additional difficulty in the proof of the hydrodynamic
behavior of nongradient systems. For gradient systems, like zero range processes,
the current Wx;x+ei is itself equal to the difference x h x+ei h of a cylinder
function and its translation. This property permits a second summation by parts
in the integral term of the martingale. After this second summation by parts, the
integral term becomes
Z t X
N d (H )(s; x=N )x h(s ) ds + O(N 1
):
0
x2TdN
The proof is thus reduced to the replacement of the cylinder function h by a
function of the empirical measure.
In contrast with the gradient case the current W0;ei in the generalized exclusion
process can no longer be written as a difference h ei h for some cylinder function
h so that a second summation by parts is impossible. We have thus not only to
replace a cylinder function but a cylinder function multiplied by N .
The main theorem of this chapter asserts that there exists a collection of con-
tinuously differentiable increasing functions fdi;j : [0; ] ! R+ ; 1 i; j dg
such that
7. Hydrodynamic Limit of Reversible Nongradient Systems 143
Z t X
lim sup lim sup E N

N d H (s; x=N )
"!0 N !1 0
x2TdN
n d h
X io
N Wx;x+ei (s) + di;j (s"N (x + ej )) di;j (s"N (x)) ds = 0
j =1
(0:5)
for all 1 i d, smooth function H and t > 0.
A sketch of the proof of this result is presented at the end of this introduction.
d
This proof relies on the characterization of the closed forms of f0; : : : ; gZ .
This characterization, presented in section A3.4, requires a sharp estimate for the
second eigenvalue of the generator of the process restricted to finite cubes : for
a positive integer ` and 0 K j` j, denote by L` the restriction to `
of the generator of the symmetric generalized exclusion process and by `;K the
largest, strictly negative eigenvalue of the symmetric operator L` on ` ;K . The
proof of the characterization of germs of closed forms requires that j`;K j shrinks
at a rate slower than ` 2 , i.e., that there exists a universal constant C such that
j`;K j C` 2 for all ` 2 and 0 K j`j. This estimate on the spectrum
of L` is proved in section A3.2. Therefore, though hidden in Appendix 3.4,
the proof presented in this chapter of the hydrodynamic behavior of nongradient
reversible systems relies on a sharp estimate of the spectral gap for the generator
of the process restricted to finite cubes.
In possession of (0.5), the conclusion of the proof is rather simple. Statement
(0.5) permits a second summation by parts in the integral term of the martingale.
M H (t) can thus be written as
Z t
M H (t) = < tN ; Ht > < 0N ; H0 > < sN ; @s Hs > ds
0
d Z
X t X
N d (@u2 i ;uj H )(s; x=N )di;j (s"N (x)) ds + oN (1) ;
i;j =1 0
x2TdN
where oN (1) indicates a random variable that converges to 0 in L1 (N ) as N " 1.
It follows from (0.3) that in the limit as N " 1, the empirical measure satisfies
the nonlinear parabolic differential equation :
d
X d
X
@t = @u2 i ;uj di;j () = @ui fDi;j ()@uj g ;
i;j =1 i;j =1
where Di;j () = (d=d)di;j ().
To state the hydrodynamic behavior of the generalized exclusion process it
remains to present an explicit formula for the diffusion coefficient Di;j () intro-
duced above. Hereafter, by an oriented bond b, we understand a pair of nearest
neighbor sites, most of the time denoted by b = (b1 ; b2 ). Notice that with this
convention (b1 ; b2 ) =
= (b2 ; b1). For an oriented bond (b1; b2 ) and a cylinder function
g, define the gradient (rb1 ;b2 g) by
(rb1 ;b2 g )( ) = rb1 ;b2 ( )[g ( b1 ;b2 )

g()] : (0:6)
d
Set = f0; : : : ; gZ . For every cylinder function g : ! R, denote by g ( )
the formal sum X
g = x g
x2Zd
which does not make sense but for which the “gradient"

r g = r0;e g ; : : : ; r0;ed
1 g
is well defined since it involves only a finite number of non zero differences. For
each , the diffusion coefficient of the hydrodynamic equation for the generalized
symmetric exclusion process is the unique symmetric matrix fDi;j (); 1 i; j
dg such that
d X
X d 2
a D()a = 1
2() h2C0
inf j
aj (A )i + r0;ei h (0:7)
i=1 j =1
for every vector a in Rd . In this formula a represents the transposition of a,
stands for the static compressibility, which in our case is equal to
() = < (0)2 > < (0) >2 ;
Aj stands for the vector defined by
(Aj )i ( ) = r0;ej ( )j;i ;
< > for the expectation with respect to and C0 for the space of cylinder
functions on with mean zero with respect to all canonical measures, i.e., the
space of cylinder functions g such that E` ;K [g ] = 0 for all large enough ` so that
` contains the support of g and all 0 K (2` + 1)d. Examples of functions
in C0 are the currents fW0;ei ; 1 i dg, the gradients f (ei ) (0); 1 i dg
and the range of the generator L : LC0 = fLg; g 2 C0 g. Here L sands for the
generator extended to Zd.
The reader can find in Spohn (1991) (Proposition II.2.2) the equivalence of
the variational formula (0.7) for the diffusion coefficient with the Green–Kubo
formula based on the current–current correlation functions :

X
Di;j () =
1
2()
(x ei ) < (x)W0;ej >
x2Zd
Z 1 X

+ dt < etL W0;e j ; Wx;x+ei > :
0 x2Zd
Here (a b) stands for the usual inner product of Rd .
We are now ready to state the hydrodynamic behavior of the symmetric gen-
eralized exclusion process.
7. Hydrodynamic Limit of Reversible Nongradient Systems 145
Theorem 0.1 Assume d = 1. Let 0 : Td ! R+ be a bounded function. Let

fN ; N 1g be a sequence of probabilities on TdN associated to the profile
0 :
h X Z i
N 1 G(u)0(u) du > = 0 ;
lim
N !1 N d x2Td G(x=N )(x)
N
for every continuous function G: Td ! R and for every strictly positive. Then,
for every t 0,
2 3
Z
1 X
lim PN d
N !1
4
N G(x=N )t (x) G(u)(t; u) du > 5

= 0;
x2Td
N
for every continuous function G: Td ! R and for every strictly positive, where
(t; u) is the unique weak solution of the parabolic equation
8 X n o
>
< @t = @ui Di;j ()@uj
i;jd
1 (0:8)
>
:
(0; ) = 0 () :
Remark 0.2 We assumed the dimension to be equal to 1. The reason is that we

know only the diffusion coefficient given by the variational formula (0.7) to be
continuous and there is no uniqueness result for weak solutions of equation (0.8)
in dimension d 2 under such weak assumptions on D(). Proofs are presented in
general dimension to stress that we need nowhere else the dimension to be equal
to 1.
We conclude this section presenting the strategy of the proof of the replacement
of the current by a gradient. Recall that C0 stands for the space of mean-zero
cylinder functions and that L` stands for the restriction of the generator LN to
the cube ` . In section 4 we prove that for each pair f , g of cylinder functions in
C0, and each sequence K` such that K`=(2`)d converges to ,
1 D X X E
j` j x f; ( L` ) 1
y g (0:9)
jxj`f jyj`g ` ;K`
converges as ` " 1. In this formula, for a cylinder function f , `f stands for

` sf , where sf is the smallest integer k such that k contains the support of f .
In this way the support of x f is contained in ` for all jxj `f . The proof of
d
this convergence requires the characterization of all closed forms of f0; : : : ; gZ .
It is here that a sharp estimate on the spectral gap of L` is needed.
The limit of (0.9), denoted by f; g , defines a semi-inner product on the
space C0 . The Hilbert space induced by C0 and this semi-inner product is denoted
by H .
In section 3, with the type of arguments used in the proof of the one block
estimate (the entropy inequality, Feynman–Kac formula, a variational formula for
the largest eigenvalue of a symmetric operator), we prove that for each t > 0,
smooth function H : Td ! R and h in C0 ,
hZ t X i
lim sup E N

N1 d H (s; x=N )(xh)(s ) ds
N !1 0
x2TdN
D X X E
C (t; H ) lim sup sup j1 j x h; ( L` ) 1
y h
`!1 K ` jxj`h jyj`h ` ;K
for some finite constant C (t; H ). This inequality together with (0.9) shows that
for each t > 0, smooth function H : Td ! R and h in C0 ,
hZ t X i
lim sup E N

N1 d H (s; x=N )(xh)(s) ds
N !1 x2TdN (0:10)
0
C (t; H ) sup h; h
0
for some finite constant C (t; H ).
The structure of H is quite simple and is examined in details in section 5. The
gradients f (ei ) (0); 1 i dg are linearly independent vectors orthogonal
to the space LC0 and H is generated by these two subspaces. There exists, in
particular, a matrix fDi;j (); 1 i; j dg, that depends on the density because
the inner product depends on the density, such that
d
X
W0;ei + Di;j ()[(ej ) (0)] 2 LC0
j =1
for 1 i d. The matrix Di;j that appears in the above formula is the diffu-
sion coefficient of the hydrodynamic equation (0.8). Thus the diffusion coefficient
Di;j () is just the coefficient of the projection of the current W0;ei on the gradient
(ej ) (0) in the Hilbert space H . Moreover, for each fixed > 0, there exists
f in C0 such that W0;ei + dj=1 Di;j ()[(ej ) (0)] Lf .
P
It follows from this observation, (0.10) and some two blocks type argument
(to replace Di;j ()[ (ej ) (0)] by Di;j ( "N (0))[ "N (ej ) "N (0)] that
Z t X
inf lim sup lim sup E N
f 2C0 "!0 N !1

N1 d H (s; x=N )
0
x2TdN
n d
X o
x W0;ei (s) + Di;j (s"N (0))[s"N (ej ) s"N (0)] (x Lf )(s ) ds = 0:
j =1
By Taylor’s expansion and the continuity of the diffusion coefficient D(),
statement (0.5) follows from the previous limit and from the easy to prove identity
(cf. the proof of Corollary 1.2)
1. Replacing currents by gradients 147
hZ t X i
lim sup E N

N1 d H (x=N )Lf (s) ds = 0 (0:11)
N !1 0
x2TdN
for each f in C0 .
1. Replacing currents by gradients
We show in this section that the current W0;ei may be decomposed as a linear
combination of the gradients f (ej ) P (0); 1 j dg with a function in the
range of the generator LN : W0;ei + 1j d Di;j ()f (ej ) (0)g = LN f for
some matrix Di;j () that depend on the density and a cylinder function f . The
gradient part in the decomposition permits a second summation by parts, while the
LN f term turns out to be negligible. This is the content of Theorem 1.1 below.
For positive integers `, N , a smooth function H in C 2 (Td ) and a cylinder
function f, let
X
;i (H; ) : = N 1 d
XN;`
f
H (x=N )xVi ;` () f
x2TdN
where,
d
X n o
Vi ;`
f
= W0;ei () + Di;j ` (0) ` (ej ) ` (0) LN f() :
j =1
Theorem 1.1 For every smooth function H in C 1;2 ([0; T ] Td), 1 i d and
T > 0,
" Z #
T
1
log E N exp N d ;i (H ; ) ds = 0 :
XN;"N
f
s s
2C "!0 N !1 N d
inf lim sup lim sup
f 0
The proof of this result is postponed to section 3. We first conclude the proof
of the hydrodynamic behavior of the generalized symmetric exclusion process.
We start showing that the LN f term is negligible. For a positive integer ` and a
smooth function H in C 2 (Td), let
i (H; ) : =
YN;`
X d
X n o
N1 d H (x=N ) Wx;x+ei + Di;j ` (x) ` (x + ej ) ` (x) :
x2TdN j =1
Corollary 1.2 For every smooth function H in C 1;2 ([0; T ] Td), 1 i d and
T > 0,
" Z #
T

lim sup lim sup E N i (Hs ; s ) ds
YN;N" = 0:
"!0 N !1 0
Proof. Fix a smooth function H , T > 0 and 1 i d. For a cylinder function f

write
i (H; ) = X f;i (H; ) + N 1 d X H (x=N )x LN f( ) :
YN;` N;`
x2TdN
To prove the corollary, we have just to show that
" Z #
T
;i (H ; ) ds
XN;N"
f
(1:1)
f
inf lim sup lim sup E N
2C "!0 N !1 s s = 0
0
and that for every cylinder function f,

Z T
X
lim E N
N !1

N1 d H (t; x=N )xLN f(t ) dt = 0: (1:2)
0
x2TdN
To prove (1.1) apply the entropy inequality to obtain that
" Z #
T
;i (H ; ) ds
XN;N"
f
E N s s
0
" Z #
T
N N log E N exp N d ;i (H ; ) ds
N d H ( j ) + XN;"N
1 1 f
N d s s
0
for every positive . Since there are at most particles per site, the entropy
H (N jN ) is bounded by C (; )N d (cf. discussion preceding the statement of
Lemma 5.5.7). In particular, (1.1) follows from Theorem 1.1 and the arbitrariness
of .
We turn now to (1.2). Since time is speeded up by N 2 , for any smooth function
H : [0; T ] Td ! R and any cylinder (and thus bounded) function f,
X
M H; (t)
f
= N (d+1) H (t; x=N )x f(t )
x2TdN
X
N (d+1) H (0; x=N )xf(0 )
x2TdN
Z t X
N (d+1) (@s H )(s; x=N )xf(s ) ds
0
x2TdN
Z t X
N1 d H (s; x=N )xLN f(s ) ds
0
x2TdN
1. Replacing currents by gradients 149
is a martingale. The first three terms on the right hand side are of order N 1
because f is bounded. A simple computation of the quadratic variation of the
martingale shows that E N [M H;f (t)2 ] is bounded above by
Z t o2
N 2d X
E N rx;y (s )
n X
H (s; z=N )[(zf)(sx;y ) (z f)(s )] :
2
jx yj=1 0
z2TdN
Since f is a cylinder function, the difference (z f)( x;y ) (z f)( ) vanishes for all
but a finite number of sites z . This expectation is therefore of order N d , which
proves (1.2).
We have now all elements to prove the hydrodynamic behavior of a nongradient
system.
Proof of Theorem 0.1. Recall that we denote by tN the empirical measure defined
by X
tN () = N d t (x)x=N :
x2TdN
Fix T > 0 and denote by QN the probability measure on the path space
D([0; T ]; M+;(Td)) corresponding to the process tN with generator LN speeded
up by N 2 starting from N . We have already seen in the proof of the hydro-
dynamic behavior of symmetric simple exclusion processes that a law of large
numbers for the empirical measure tN follows from the weak convergence of
the sequence QN to the probability measure concentrated on the deterministic,
absolutely continuous trajectory (t; du) = (t; u)du whose density is the weak
solution of equation (0.8).
In section 6 we prove that the sequence fQN ; N 1g is weakly relatively
compact and that all limit points Q are concentrated on weakly continuous paths
t that are absolutely continuous with respect to the Lebesgue measure with density
bounded by :
(t; du) = (t; u)du and (t; u) for all 0 t T , Q almost surely.
From Brezis and Crandall (1979), in dimension 1, there exists a unique weak
solution of (0.8). Therefore, to conclude the proof of the theorem, it remains to
show that all limit points of the sequence fQN ; N 1g are concentrated on
absolutely continuous trajectories (t; du) = (t; u)du whose density are weak
solutions of equation (0.8).
Fix a smooth function H : [0; T ] Td ! R and recall from (0.2) the definition
of the martingale M H (t). Since E N [(M H (t))2 ] vanishes as N " 1, by Doob’s
inequality, for every > 0,
h i
lim PN sup M H (t) > = 0: (1:3)
N !1 0 tT
Applying Corollary 1.2 to the second integral term in the explicit formula (0.4) of
M H (t), we get that for every > 0,
Z T

lim sup lim sup PN < T ; HT > < 0 ; H0 > < s ; @s Hs > ds
"!0 N !1 0
Xd ZT X
+ N1 d (@uNi H )(s; x=N )x Vi;j;"N (s ) ds

> = 0;
i;j =1 0
x2TdN
where h i
Vi;j;"N () = Di;j "N (0) "N (ej ) "N (0) :
R
Denote by di;j the integral of Di;j : di;j () = 0 Di;j ( )d . Since by Theorem
5.8 Di;j is continuous,
n o
Di;j ("N (0)) "N (ej ) "N (0)
= di;j ( "N (ej )) di;j ("N (0))
N 1 oN (1) ; +
where oN (1) represents a term that converges uniformly to 0 as N " 1. Summing
by parts, for each 1 i; j d, we obtain that
X h i
N1 d (@uNi H )(s; x=N )x Di;j "N (0) "N (ej ) "N (0)
x2TdN
X
= N d (@u2 i ;uj H )(s; x=N )di;j "N (x) + oN (1) :
x2TdN
Therefore,
Z T

lim lim PN < T ; HT > < 0 ; H0 > < s ; @s Hs > ds
"!0 N !1 0
Xd ZT X
N d (@u2 i ;uj H )(s; x=N )di;j (s"N (x)) ds

> = 0:
i;j =1 0
x2TdN
Recall the definition of the approximation of the identity " . By continuity, for
every limit point Q of the sequence QN ,
Z T
lim Q

"!0 < T ; HT > < 0 ; H0 > < s ; @s Hs > ds
0
d Z
X T Z
ds du (@ui;uj )H (s; u)di;j (< s ; " (u ) > >
2
) = 0:
i;j =1 0 Td
Since each limit point Q is concentrated on absolutely continuous paths t =

(t; u)du with density (t; u) bounded by , for each fixed 0 s T , as " # 0,
< s ; " (u ) > converges to (s; u) for almost all u in Td . From this remark
and the continuity of fdi;j ; 1 i; j dg, we obtain that
2. An integration by parts formula 151
Z T
Q < T ; HT > < 0 ; H0 > < s ; @s Hs > ds
0
d Z
X T Z
ds du (@u2 i ;uj )H (s; u)di;j ((s; u)) > = 0
i;j =1 0
for all H in C 1;2 ([0; T ]; Td). In particular, Q is concentrated on weak solution

of (0.8) what concludes the proof of the theorem.
It may seem odd to define ;i as we do since it involves the term
XN;` f
X
N1 d H (x=N )xLN f
x2TdN
which vanishes and gives no contribution to the equation in the limit. As a matter
of fact this term is important in order to establish Theorem 1.1 where the current
is not only multiplied by N but also exponentiated.
2. An integration by parts formula
For a cylinder function , denote by the smallest d-dimensional rectangle that

contains the support of and by s the the smallest positive integer s such that
s . If is the gradient (ei ) (0), for instance, = f0; eig and s = 1.
in this example s and do not coincide. Let C0 be the space of cylinder
functions with mean zero with respect to all canonical invariant measures :
n o
C0 = g 2 C ; < g >g ;K = 0 for all 0 K jg j :
Here, for a finite subset of Zd and 0 K jj, < >;K stands for the
expectation with respect to the canonical measure ;K . Examples of functions in
C0 are L g for finite subsets and cylinder functions g with support contained
in , currents fW0;ei ; 1 i dg and gradients f (ei ) (0); 1 i dg.
For each cylinder g in C0 , < g >;K = 0 for all g , 0 K jj,
X
< g > = 0 and < g; (x) > = 0 for all 0 :
x2Zd
To prove the first assertion, fix a set g and 0 K jj. < g >;K is
equal to h h X ii
E;K E;K g (x) :
x2g
P
The conditional expectation E;K [g j x2g (x)] is equal to the expectation of
g with respect to the canonical measure g ;Px2 (x), that vanishes because g
g
belongs to C0 . The second identity follows immediately from the previous result
and the convergence of the finite marginals of `;K to , as ` " 1 and K=(2`)d !
(cf. Section A2.2). The third identity is also easy to check. First observe that
< g; (x) > does not vanish only for a finite number
P of sites x since g has mean
zero and is a product
P measure. In particular, x g; (x) > is well defined
<
and equal to < g; x2g (x) > . Taking conditional expectation with respect to
P P P
x2g (x), this last expectation writes E [E [g j x2P g (x)] x2g (x)]
that vanishes because the conditional expectation E [ j x2g (x)] reduces
to the expectation with respect to the canonical measure g ;P and g
x2g (x)
belongs to C0 .
For a finite subset of Zd, denote by F the -algebra generated by
f(x); x 2 g :
F = f(x); x 2 g
and abbreviate F` by F` .
For a rectangle and a canonical measure ;K on ;K , denote by <
; >`;K (resp. < ; > ) the inner product in L (`;K ) (resp. L2( )) and by
2
D(;K ; ) the Dirichlet form relative to ;K defined by

D E X
D(;K ; f ) = L f; f ;K = Db (;K ; f ) ;
b 2
where, for each bond b = (b1 ; b2 ),
Z h i2
Db (;K ; f ) = 14 rb ;b () f (b ;b ) f () ;K (d) :
1 2
1 2
In this formula and below summation is carried out over all oriented bonds b in
(a bond b = (b1 ; b2 ) is said to be in if both end points are in ). Notice that
(b1 ; b2 ) =
= (b2; b1 ) so that both (b1 ; b2 ) and (b2 ; b1 ) appear in the above summation.
Consider a cylinder function in C0 . We claim that is in the range of
L . To prove this statement fix 0 K j j and consider the generalized
symmetric exclusion process on ;K . The kernel of L in L2 ( ;K ) has
dimension 1 : assume that L f = 0. Multiply both sides of the identity by f
and integrate with respect to ;K to obtain that f must be constant. Since the
kernel of L in L2 ( ;K ) has dimension 1, the range of L has codimension
1. Since the range is included in the subspace of mean-zero functions that has
codimension 1, both spaces are equal, i.e., all ;K -mean-zero functions are in
the range of L .
We may therefore write the cylinder function as = ( L )( L ) 1 for
some mean-zero function ( L ) 1 , measurable with respect to the variables
f(z ); z 2 g. Fix a rectangle that contains and 0 K jj. Since
the canonical measure ;K is reversible, we have that
D E D E
; h ;K = ( L )( L ) 1 ; h ;K
X D E X D E
= (1=4) rb ( L ) 1 ; rb h ;K = b ; rb h ;K
b2 b 2
2. An integration by parts formula 153
provided we set b = (1=4)rb ( L ) 1 for b in . In this formula summation

is carried out over all oriented bonds b in and rb is defined in (0.6). We may
extend the definition of b for bonds not in setting b = 0 if b 62 . From

this explicit formula for b , it is easy to check that y b = b+yy for all y in Zd.
Moreover, by reversibility, we have that
X D 2 E X D E
b ;K
= (1=16) rb ( L ) 1
; rb ( L ) 1
;K
b2 b2
D E
= (1=4) ; ( L ) 1
= C( )
;K
for some finite constant that depends only on (and not on nor on K ). We
have thus proved the integration by parts formula :
Lemma 2.1 (Integration by parts formula) Let be a cylinder function in C0 .

There exists a family of cylinder functions fb ; b 2 g, measurable with respect
to F , such that D E D E
X
; h ;K = b ; rb h ;K
b2
for all rectangles , 0 K jj and functions h in L2 (;K ). Moreover,
X D 2 E
b ;K
C( )
b2

for some finite constant C ( ) and y b = b+yy for all d-dimensional integer y .
The same result may be restated with canonical measures ;K replaced by grand
canonical measures .
The integration by parts formula assumes a particularly simple form for three
types of functions. First of all, if = L h, for some finite rectangle and some
F -measurable cylinder function h, then b = (1=4)rbh for b 2 . If is a
current, = W(b1 ;b2 ) , we obtain that
(1=4) ifb0 = b = (b1 ; b2 ) ;

b0 (2:1)
(1=4) if b0 = (b2 ; b1 ) :
=
For each 1 k d, an elementary computation shows that L0;e1k [(ek ) (0)] =

[ (ek ) (0)]f (1=3) (0) (ek) + F ( (0) + (ek ))g, where
F ( A) = F (A)

(1=6)(A + 1)(A + 2) for 0 A ;
(1=6)A + [ + (1=2)]A [ + + (1=3)] for A 2 .
= 2 2
Both definitions coincide at A = . Let 0;ek = (1=4)r(0;ek ) L0;e1k [ (ek ) (0)]

and ek ;0 = (1=4)r(ek ;0) L0;e1k [ (ek ) (0)]. From the explicit formula for
L0;e1k [(ek ) (0)], we get that
n o
0;ek =
4
1
r0;ek ( ) 2 (0) (ek ) [ (ek ) (0)] G ( (0) + (ek ))
n o
ek ;0 = 41 r0;ek () 2(0)(ek ) [(ek ) (0)] + G((0) + (ek )) ;
where,

A for 0 A ;
G(A) G (A)
(2 + 1)A 2( + ) for A 2 .
= = 2
We leave the reader to check that 2 0;ek + 2 ek ;0 = [ (ek ) (0)] and that
< (ek ) (0) ; h > = < 0;ek ; r0;ek h > + < ek ;0 ; rek ;0 h > (2:2)
for every 0 and for all cylinder functions h in L2 ( ).
Remark 2.2 A change of variables shows that
< rb ;b f ; rb ;b h > = < rb ;b f ; rb ;b h >

1 2 1 2 2 1 2 1
for every bond (b1 ; b2 ), 0 and cylinder functions f , h in L2 ( ). Since

for in C0 , b = (1=4)rb ( L ) 1 , we have
X d
X X
< b ; rb ;b h > =
1 2 2 < (z;z+ei ) ; rz;z+ei h >
b2 i=1 z; z2
z+ei 2
and the integration by parts formula becomes
d
X X
< ; h > = 2 < (z;z+ei ) ; rz;z+ei h > :
i=1 z; z2
z+ei 2
3. Nongradient large deviations estimates
We prove in this section Theorem 1.1. To detach the main arguments of the proof
we divide it in several steps.
Step 1 : Reduction to an eigenvalue problem. Our purpose in this first step is to
reduce the dynamic problem stated in the theorem to a static problem involving the
largest eigenvalue of a small perturbation of the generator N 2 LN . This reduction
relies mainly on Feynman–Kac formula and on a variational formula for the largest
eigenvalue of a symmetric operator.
Since ejxj ex + e x and since
lim sup N d logfaN + bN g lim sup N d log aN _ lim sup N d log bN ;

N !1 N !1 N !1
3. Nongradient large deviations estimates 155
it is enough to show that

" #
Z T
inf lim sup lim sup d log E N exp N d
1 ;i (H ; ) ds
XN;"N
f

s s
"!0 N !1 N
0
f2C 0
for every smooth function H in C 1;2 ([0; T ] Td ), 1 i d and T > 0.

Fix such smooth function H , 1 i d and T > 0. By Feynman-Kac formula
(A1.7.5),
" ( )# (Z )
Z T T
Nd ;i (H ; ) ds
XN;"N
f
N (s) ds ;
E N exp s s exp
0 0
where
f;i
N (s) is the largest eigenvalue of the symmetric operator N 2 LN +
d
N XN;"N (Hs ; ). From the variational formula for the largest eigenvalue of an
operator in a Hilbert space (A3.1.1),
nD E o
N (s) ;i (H ; )f ( )
N dXN;"N N 2 DN ( f ) ;
f
sup s
f
where the supremum is taken over all densities f with respect to N and < >
denotes the expectation with respect to . In particular,
" (Z )#
T
d log E N exp f;i
N N d XN;"N (Hs ; s ) ds
0
Z T nD E o
;i (H ; )f ( )
ds sup XN;"N f
N 2 d DN ( f ) :
s
0 f
Therefore, to prove Theorem 1.1 we have to show that
nD E o
inf lim sup lim sup sup ;i (H; )f ( )
XN;"N N 2 d DN ( f ) 0
f
f2C "!0 N !1 f
uniformly over the set of continuous functions H in C 2 (Td ) that are bounded as
well as their first and second derivatives by some fixed constant.
The proof of this inequality is divided in three steps. The strategy follows
closely the one adopted in thePproof of the replacement lemma in Chapter 5. Recall
f;i
that XN;"N is equal to N 1 d x H (x=N )x Vif;"N ( ). We first reduce the problem
on a small macroscopic cube to the same problem on a large microscopic block.
In our context this corresponds to replace Vif;"N ( ) by Vif;` ( ) and constitutes the
goal of step 2 below. We will then follow the arguments presented in the proof
f;i
of the one block estimate to bound the largest eigenvalue of N 2 LN + XN;` by
f;`
the largest eigenvalue of L` + N Vi , where, for a finite subset of Zd, L
represents the restriction of the generator LN to and N is a small constant.
In order to estimate the largest eigenvalue of L` + N Vif;` we use a perturbation
method that relies on the existence of a spectral gap for the generator restricted to
finite boxes. This argument provides a bound on the largest eigenvalue in terms of
the variance of Vif;` . To conclude the proof it will remain to compute the variance
and to show that it vanishes for some cylinder function f.
We start localizing the eigenvalue problem, a rather technical step.
Step 2. Reduction to microscopic blocks. Notice that there is no spatial average
f;i
of the current W0;ei in the definition of XN;"N and recall from the proof of the
one block estimate in Chapter 5 that such a spatial average is crucial. It can easily
be inserted because
X n X o
N1 d H (x=N ) x W0;ei () (2`0 + 1) d y W0;ei ()
x2TdN jy xj`0
is of order `2 =N as one can see after performing a summation by parts and
from the presence of a discrete Laplacian. Here and below ` denotes a positive
integer independent of N and " that increases to infinity after N " 1, " # 0 and
`0 = ` 1. We averaged over jy xj `0 so that y W0;ei is measurable with
respect to f (z ); z 2 ` g for y in `0 .
Denote by sf the linear size of the support of the cylinder function f. In the
f;i
definition of XN;` , we may replace LN by L for some cube large enough to
contain sf +1 . Furthermore, since f is a cylinder function, a summation by parts
shows that
X X X
N1 d H (x=N )x(L f)() N1 d H (x=N ) j1 j x+y (Lf)()
x2TdN x2TdN ` jyj`
is of order
P
`2 =N . This justifies the replacement of x LN f by the average
j` j jyj` x+y L f in XN;"N
1 ;i . We have thus averaged in space the terms
f
W0;ei and LN f in the definition of XN;"N;i . We turn now to the substitution of

f
Di;j ("N (0))f"N (ej ) "N (0)g by a local function.

Lemma 3.1 For each 1
i;j ( ) by
i; j d, " > 0 and positive integers N , `, define
V`;N"
i;j ( )
V`;N"
n o n o
= Di;j ( "N (0)) "N (ej ) "N (0) Di;j (`(0)) `0 (ej ) `0 (0) :
For every > 0 and 1 i; j d,
`!1 "!0 N !1 f
D E
N1 d X H (x=N ) x V i;j ( )f ( ) N 2 d DN (f ) 0 :
`;N"
x2TdN
0 0
We averaged the density over jy j `0 = ` 1 for ` (ej ) and ` (0) to be measurable
with respect to f (x); x 2 ` g.
i;j ( ) as
Proof. We first rewrite the difference V`;N"
nh i h 0 io
Di;j ("N (0)) "N (ej ) "N (0) (ej ) `0 (0)
`
n on o (3:1)
+ Di;j ("N (0)) Di;j (` (0)) `0 (ej ) `0 (0)
and consider the two lines separately. A summation by parts shows that the first
one translated by x, multiplied by H (x=N ) and summed over x is equal to
X
bx()[(x + ej ) (x)] ;
x2TdN
where bx ( ) is given by
X
bx () : = j1 j Di;j ("N (y))H (y=N )
"N jy xj"N
X
1
j 0 j Di;j (` (y))H (y=N ) :
` jy xj`0
We shall prove that

`!1 "!0 N !1 f
D E
N1 d X bx ( )[ (x + ej ) (x)]f () N 2 d DN (f ) 0
x2TdN
for each > 0. P
The main difficulty in obtaining an estimate for N 1 d x < bx ( )[ (x +
ej ) (x)]f () > in terms of the Dirichlet form comes from the factor N
that is multiplying the space average (otherwise the problem would reduce to the
replacement lemma of Chapter 5 because D is a continuous function in virtue
of Theorem 5.8). The extra factor N 1 shall be obtained integrating by parts the
function [ (x) (0)] and applying Schwarz inequality.
Recall from (2.2) the definition of the cylinder functions 0;ek , ek ;0 for 1
k d. By formula (2.2) and Remark 2.2, for all cylinder functions h,
< (ek ) (0) ; h > = 2 < 0;ek ; r0;ek h >
P
Applying this identity to h( ) = bx ( )f ( ), we get that x < bx ( )[ (x + ej )
(x)]f () > may be rewritten as
X n o
2 (x 0;ej )( )rx;x+ej ( ) bx ( x;x+ej )f ( x;x+ej ) bx()f () (3:2)
x2TdN
Note that the difference bx ( x;x+ej ) bx() is equal to
X n o
1
j"N j jy H (y=N ) Di;j "N (y) (2"N + 1) d Di;j ("N (y))
xj"N
yj =xj "N
which is of order ("N ) 1 oN (1) because Di;j () is continuous. Here oN (1) stands
for a constant that vanishes in the limit as N " 1. In particular, (3.2) is equal to
X n o
2 (x 0;ej )( )bx ( )rx;x+ej ( ) f ( x;x+ej ) f ( ) + " 1 N d 1oN (1)
x2TdN
To estimate the first term of this expression, rewrite jf (x;x+ej ) f ()j as

p p p p

( f x;x+ej ) f () f (x;x+ej ) + f () :
By the elementary inequality 2cd Ac2 + A 1 d2 that holds for any positive A,
the expectation < (x 0;ej )bx rx;x+ej f > is bounded by
hp p i2 E
1
2A
< (x 0;ej ())2bx ()2 rx;x+ej () f (x;x+ej ) + f ()
A Dr hp
( x;x+ej )
p i2 E
+ x;x +ej ( ) f f ( )
2
for all A > 0. The second term of this expression is just 2AIx;x+ej (f ). Since
(c + d)2 2c2 + 2d2 , since 0;ej is bounded by a finite constant that depends only
on and since by reversibility < rx+ej ;x ( )h( x+ej ;x ) > =< rx;x+ej ( )h( ) >
for every h in L2 ( ), the first term is bounded above by
C () Db ()2 f ()E + C () Db (x+ej ;x)2 f ()E :
A x A x
Since bx ( x+ej ;x
) = bx ( ) + ("N ) oN (1), summing over x and diving by N d 1 ,
1
we obtain that (3.2) divided by N d 1 is bounded above by

C () N 1 d X Db ()2 f ()E + 2AN 1 d D (f ) + " 1 1 + 1 o (1)
A x N "NA N
x2Td N
for every A > 0. Choosing A = =4N , we get that

X D E
sup N1 d [ (x + 1) (x)]bx()f () N 2 d DN (f )
f x2TdN
X D E
sup C 1 N d [bx ( )]2 f ( )

(=2)N 2 dDN (f ) + oN (1) :
f x2TdN
The proof that the limit of this expression as N " 1, " # 0 and ` " 1 is
nonpositive follows from the usual two blocks estimate since Di;j () is continuous
by Theorem 5.8 below.
The second expression in (3.1) is handled in the same way. This concludes the
proof of the lemma.
For every positive integers N , `, 1 i d, smooth function H 2 C 2 (Td) and
cylinder function f, let
X
;i (H; ) : = N 1 d
X~N;`
f
H (x=N )xV~i ;` ()
f
x2TdN
where,
X d
X n o
V~i ;`
f
= (2`0 + 1) d y W0;ei () + Di;j ` (0) `0 (ej ) `0 (0)
jyj`0 j =1
X
(y LN f)( )
1
(2`f + 1) d
y2`f
and `f = ` sf 1 so that y Lf is F` -measurable for every y in `f . Up to this
point we showed that in order to prove Theorem 1.1 it is enough to prove the
f;i f;i
same statement with X~N;` ( ) instead of XN;"N , i.e., to show that for all > 0,
nD E o
;i ( )f ( )
X~N;` f
N 2 d
D (f ) 0 : (3:3)
inf lim sup lim sup sup
f 2C `!1 N !1 f N
Step 3. Estimate on small perturbations of a reversible generator. Since is

f;i
translation invariant, we may rewrite < X~N;` ( )f ( ) > as
X D E
N1 d H (x=N )V~i ;` () xf ()
f
x2TdN
We now repeat the usual procedure of the proof of the one block estimate. We
first project the density x f on the finite hyperplanes with fixed total number of
particles ` ;K . Recall that for 0 K j` j and a density f with respect to
, we denote by `;K the measure conditioned on the hyperplane ` ;K :
X
`;K () = j ( x) = K
x2`
and by f`;K the projection of f on ` ;K :
h i
E f (x) = (x) ; x 2 `
f`;K ( ) = h P i
E f x2` (x) = K
for all configurations of ` ;K . Since V~if;` ( ) depends on only through
f(z ); z 2 ` g, we have that
D E X D E
V~i ;` ()( xf )()
f
= c(x; f; K ) V~i ;`()( x f )`;K () `;K ;
f
K
where c(x; f; K ) is given by
Z X
c(x; f; K ) = 1f ; (x) = K g( xf )()(d)
x2`
and < >`;K stands for expectation with respect to the canonical measure
`;K = ` ;K . Notice that summation over K of C (x; f; K ) is equal to 1 for
all x. Moreover, with respect to `;K , ` (0) is a constant equal to K=(2` + 1)d .
Denote by D`;K the Dirichlet form on ` ;K :
X D hp p i2 E
D`;K (f ) = (1=4) rx;y () f (x;y ) f ( ) `;K

x;y2`
jx yj=1
By convexity of the Dirichlet form, we have that
X X
c(x; f; K )D`;K ( x f )`;K (2` + 1) dDN (f ) :
x2TdN K
In conclusion, the expression inside braces in (3.3) is bounded by
D E
X X
c(x; f; K ) N 1 d H (x=N )V~i ;` ()( xf )`;K () `;K
f
x2TdN K

C` N 2 d (2`) dD`;K (( xf )`;K )
for some constant C` that converges to 1 as ` " 1. Since ( x f )`;K is a density
with respect to `;K , and since summation over K of c(x; f; K ) is equal to 1, this
expression is bounded above by
C` N 2 d(2`) d
X n D E o
sup sup (C` N ) 1
(2`)d H (x=N ) V~if;` ( )h( ) D`;K (h) ;
`;K
x2TdN K h
(3:4)
where the supremum is carried over all densities h with respect to `;K . Let
= (x; H; ; N; `) = (C` N ) 1 (2`)dH (x=N ). The expression
D E
sup V~i ;` ()h() `;K
f
D`;K (h)
h
is a variational formula for the largest eigenvalue of a small perturbation of the
generator LN restricted to a cube of length 2` + 1.
Recall that L` stands for the restriction of the generator LN to ` . Since
the generalized symmetric exclusion process on a finite cube is ergodic, L` has
4. Central limit theorem variances 161
a strictly positive spectral gap ` . Denote by be the largest eigenvalue of

L` + V~if;`(). Theorem A3.1.1 asserts that
2 D
L` ) 1V~i ;` ; V~i ;` `;Kf f
E
1 2kV~i ;` k1 j j`
f
(
uniformly in K . Since vanishes in the limit as N " 1, the right hand side of
the last inequality is bounded by 2 2 < ( L` ) 1 V~i ;` ; V~i ;` >`;K for sufficiently
f f
large N . Therefore, (3.4) is bounded above by
2kH k21 D E
(2`)d ( L` ) 1 V~i ;` ; V~i ;`f
f
K C`
sup
`;K
To conclude the proof of the proposition, it remains to show that
D E
inf lim sup (2`)d (
f 2C `!1 K
L` ) 1 V~i ;` ; V~i ;` `;K
f f
= 0: (3:5)
This follows from Theorem 4.6 and Corollary 5.9 below.
4. Central limit theorem variances
We assume in this section that the reader is acquainted with the concept of closed
and exact forms in the context of interacting particle systems. The main ideas and
all results needed below are presented in section A3.4.
In last section we reduced the proof of the hydrodynamic limit of nongradient
systems to the computation of a central limit theorem variance. The purpose of this
section is to obtain a variational formula for this variance. We start introducing
a semi–norm on C0 closely related to the central limit theorem variance. For
1 k d denote by Ak = (Ak1 ; : : : ; Akd ) the d-dimensional cylinder function
with coordinates defined by
(Ak )i ( ) = i;k r0;ek ( ) for 1 i d : (4:1)
Here i;j stands for the delta of Kronecker, equal to 1 if i = j and 0 otherwise.
For cylinder functions g and h in C0 and 1 j d, let
X D E X D E
g; h ;0 = g; xh and g ;j = xj g; (x) ;
x2Zd x2Zd
where xj stands for the j -th coordinate of x 2 Zd. Both g; h ;0 and
g ;j are well defined because g and h belongs to C0 and therefore all but a
1=2
finite number of terms vanish. For h in C0 , define the semi–norm h by
d
X
h = sup 2 g; h ;0 + 2 ai h ;i
g2C0 i=1
a2Rd
d D X
X 2 E
(1=2) aj (Aj )i + r(0;ei ) g
:
i=1 1jd
P
In this formula g stands for the formal sum x2Zd x g . To keep notation simple,
denote by r g the vector (r(0;e1 ) g ; : : : , r
P (0;ed ) g ) and by kak the Euclidean
norm of a d-dimensional vector a : kak2 = 1id (ai )2 . With this notation, the
semi–norm writes
d
X
h = sup 2 g; h ;0 + 2 ai h ;i
g2C0 i=1
a2Rd
D X E
=
aj Aj + r g :
2
(1 2)
1j d
We investigate in the next section the main properties of the semi–norm

1=2, while in this section we prove that the variance
X X
(2`) d < ( L` ) 1
x ; x >`;K`
jxj` jxj`
of any cylinder function in C0 converges to , as ` " 1 and
K`=(2`)d ! . Here ` stands for ` s so that the support of x is included
in ` for every jxj ` . To prove this result we need an alternative formula for
. By elementary computations relying on an adequate change of variables,
we Pobtain that the quadratic term in last formula writes (1=2)kak2 < r0;e1 >
+2 1j d aj W0;ej ; g ;0 (1=2) < kr g k2 > . The norm may
thus be rewritten as
d
X
h = sup 2 g; h ;0 + 2 ai h ;i
g2C0 i=1
a2Rd
d
X kak2 < r 1D E
+ 2 ai W0;ei ; g ;0 2
0;e1 > 2
kr g k 2

i=1
(4:2)
We are now in a position to state the main result of this section.
Theorem 4.1 Consider a cylinder function in C0 and a sequence of positive

integers K` such that 0 K` (2` + 1)d and lim`!1 K` =(2`)d = . Then,
D X X E
lim (2`) d ( L` ) 1
x ; x `;K = :
`!1 jxj` jxj` `
We first prove that the left hand side of this identity is bounded above by
the right hand side. This is done in two steps that we state as separate lemmas
in sake of clarity. In the first lemma we estimate the variances with respect to
canonical measures by variances with respect to grand canonical measures. We
then recall that the space of germs of closed forms is a direct sum of the germs
fAk ; 1 k dg introduced in the beginning of this section and the gradients
fr g ; g 2 C0g. This permits to bound the variance, with respect to the grand
canonical measure , of a cylinder function in C0 by . The necessity
of a sharp estimate for the spectral gap of the generator restricted to finite cubes
is hidden in this second step, for such an estimate is crucial in the description of
the structure of the space of germs of closed forms (cf. Theorem A3.4.14).
We start with a corollary of the integration by parts formula. For a subset
of Zd, a site x 2 Zd and a positive integer k , denote by d(x; ) the distance from
x to for the norm j j and by
k () the collection of sites at a distance less
than or equal to k from :

k () = fy 2 Zd; d(y; ) kg :
Lemma 4.2 Fix a cylinder function in C0 . There exist constants C1 ( ), C2 ( )
depending only on such that for every positive integer `, 0 K (2` + 1)d ,
subset of ` and F
s () -measurable function h in L2 (`;K ),
DX E X
2 x ; h `;K A1 C1 ( )jj + AC2 ( ) Db (`;K ; h)
x2 b2
s ()
for every A > 0. The statement remains in force if the canonical measure `;K is
replaced by the grand canonical measure .
In this lemma shall be thought as a cube x + k much smaller than ` :

k `. The lemma is just saying that the left hand side can be estimated by jj
and the Dirichlet form of h on a set slightly larger than x + k .
Proof. By the integration by parts formula,

XD E X X D E
2 x ; h `;K = 2 bx ; rb h `;K
x2 x2 b2x+
because x = x + . For each oriented bond b, denote by Ab; the set of sites
x in such that b belongs to x + : Ab; = fx 2 ; b 2 x + g. By Schwarz
inequality 2ad A 1 a2 + Ad2 , the right hand side of the last identity is bounded
above by
1 X X D x 2 E X D 2 E

A x2 b2x+ b `;K
+ A jAb; j rb h `;K
(4:3)
b2
s ()
for all A > 0. The explicit formula for bx derived in the proof of Lemma
P
2.1 shows that b2x+ < (bx )2 >`;K is bounded by a finite constant C1 ( )
that depends only on and not on x because under `;K the distribution of the
collection ( (z ); z 2 x + ) does not depend on x. To conclude the proof of the
lemma it remains to observe that jAb; j C2 ( ).
Lemma 4.3 Under the assumptions of Theorem 4.1,
D X X E
lim sup (2`) d ( L` ) 1
x ;
x `;K
`!1 jxj` jxj` `
D X X E
lim sup (2k) d ( Lk ) 1 x ; x
k!1 jxjk jxjk
Proof. By the variational formula for the variance,

X X
(2`) d < ( L` ) 1
x ; x >`;K`
x x
D X E D E (4:4)
= (2`) d sup 2 x ; h `;K L` h ; h `;K ;
h jxj` ` `
where the supremum is taken over all functions in L2 (`;K` ). By Lemma 4.2 with
= ` , K = K` and A = (1=2)C2( ) 1 , the expression inside braces in (4.4) is
bounded by
(2`)dC ( ) (1=2)D(`;K` ; h)
which is negative if D(`;K` ; h) C ( )(2`)d. Here C ( ) is a constant depending
only on that may change from line to line. On the other hand, since has
mean zero with
P respect to all canonical measures, for a constant function h the
difference 2 x2` < x ; h >`;K` D(`;K` ; h) vanishes. We may therefore
restrict the supremum to functions h with Dirichlet form bounded by C ( )(2`)d.
Fix a positive integer k larger than s + 1 and that shall converge to infinity
after `. Divide the hypercube ` in cubes of length 2k + 1. Denote these subcubes
by Ba , 1 a p = [(2` + 1)=(2k + 1)]d . Here [r] stands for the integer part of r.
Since 2` + 1 might not be divisible by 2k + 1, let Bp+1 denote the set of sites that
do not belong to any of the cubes Ba and notice that jBp+1 j Ck`d 1 for some
universal constant C . Thus
p[
+1
` = Ba and Ba \ Ba
1 2 = for a1 == a2 :
a=1
For each fixed 1 a p, denote by Bao the “interior" of Ba , that is, the set of
points in Ba that are at a distance at least s + 1 of the boundary of Ba and by
o` the union of all interior points of ` :
p
[
Bao = fx 2 Ba ; d(x; Bac ) s + 1g ; o` = Bao and 1` = ` o` ;
a=1
c
where Ba stands for the complement of Ba . Notice that 1` contains Bp+1 and that
j1` j C ( )`dfk` 1 + k 1g.
For each fixed h in L2 (`;K` ) with Dirichlet form bounded by C ( )`d , rewrite
the expression inside braces in (4.4) as
X X D E X D E
2 x ; h `;K + 2 x ; h `;K D(`;K` ; h) : (4:5)
a x2Bao ` `
x21` \`
qterm in this formula is bounded above by C ( ) minA>0

By Lemma 4.2, the second
fA 1j1` j + A`d g = C ( `d j1` j because the supremum is restricted to functions
)
with Dirichlet form bounded by C ( )`d . The second term in (4.5) is thus bounded
above by C ( )`d fk` 1 + k 1 g1=2 since j1` j C ( )`d fk` 1 + k 1 g.
We turn now to the other two terms in (4.5). By construction, for x in Bao ,
x is measurable with respect to f(x); x 2 Ba g. Denote by ha the conditional
expectation of h with respect to this -algebra : ha = E`;K` [h j (x); x 2 Ba ] so
that < x ; h >`;K` =< x ; ha >`;K` for x in Ba . Let Da be the restriction of
the Dirichlet form to Ba :
X
Da (`;K` ; h) = Db (`;K` ; h) :
b2Ba
P convexity, Da (`;K` ; ha ) Da (`;K` ; h). On the other hand, we have that
By
a Da (`;K` ; ) D(`;K` ; ) because in the first Dirichlet form bonds that
links different cubes Ba do not appear. Therefore, expression (4.5) is bounded
above by
r
X X D E
x ; ha `;K Da (`;K` ; ha ) k + 1 : (4:6)
+ C ( )`d
a
2
x2Bao ` ` k
Fix 1 a p and denote by FBa the -algebra generated by f (z );P z 2 Ba g.
The expression inside braces in last formula is bounded above by supg f2 x2B o <
a
x ; g >`;K` Da (`;K` ; g)g where the supremum is taken over all f(z ); z 2
Ba g-measurable functions g in L2 (`;K` ). Since under `;K` the distribution of the
collection ( (z ); z 2 x + ) does not depend on x,
n X D E o
sup 2 x ; g `;K Da (`;K` ; g)
g2FBa x2Bao `
n X D E o
= sup 2 x ; g `;K Dk (`;K` ; g)
g2Fk x2k `
where on the right hand side the supremum is taken over all Fk -measurable func-
tions in L2 (`;K` ) and Dk (`;K` ; ) is the Dirichlet form D(`;K` ; ) restricted to
P
k : Dk (`;K` ; ) = b2k Db (`;K` ; ). In particular, for each h in L2(`;K` ),
formula (4.6) and thus (4.5), is bounded above by
n X D E o
(2` + 1)d(2k ) d sup 2 x ; g `;K Dk (`;K` ; g)
g2Fk x2k `
r
+ C ( )(2` + 1)d
k+1:
` k
Since the expression inside braces in (4.4) is just (4.5), to prove the lemma it is
enough to show that for each fixed k ,
n X D E o
lim sup sup 2 x ; g `;K Dk (`;K` ; g)
`!1 g2Fk x2k `
n X D E o (4:7)
= sup 2 x ; g Dk ( ; g)
g2Fk x2k
By the variational formula for the variance,

P the supremum
P on the left hand
side of this identity is equal to < ( Lk ) 1 x2k x ; x2k x >`;K`
while the right hand side is equal to the variance
X X
< ( Lk ) 1
x ; x > :
x2k x2k
P P
Since both ( Lk ) 1 x2k x and x2k x are cylinder functions, iden-
tity (4.7) follows from the equivalence of ensembles stated in Lemma A2.2.2.
We now conclude the proof of the upper bound for the central limit theorem
variances.
Lemma 4.4 Under the assumptions of Theorem 4.1

D X X E
lim sup (2k ) d ( Lk ) 1
x ; x : (4:8)
k!1 jxjk jxjk
Proof. By the variational formula for the variance, for each fixed k , the expression
on the left hand side of (4.8) is equal to
X D E
(2k ) d sup 2 x ; h Dk ( ; h) :
h x2k
In this formula the supremum is taken over all Fk -measurable functions h in
L2 ( ). Lemma 4.2 and arguments similar to the ones presented at the beginning
of the proof of the previous lemma permit to restrict the supremum to functions
h with Dirichlet form Dk ( ; h) bounded by C ( )kd for some finite constant
C ( ). P
By the integration by parts formula, 2 x2k < x ; h > is equal to
X X D E X D X E
2 bx ; rb h bx ; rb h ;
= 2
x2k b2x+ b2k x2Ab;k
where Ab;k stands for the set of sites x in k such that b 2 x + . For
P x P x
each fixed bond b in k s , x2Ab;k b = x; b2x+ b . Let ^b =
P x
x; b2x+ b . With this notation the last sum writes
X D E X D X E
2 ^b ; rb h + 2 bx ; rb h
b2k b2k k s x2Ab;k
X D E
2 ^b ; rb h
b2k k s
By Schwarz inequality, the a priori bound on the Dirichlet form of h and the
estimates of the L2 norm of bx , the second and third terms are bounded by
C ( )kd (1=2) for some constant C ( ) that depends only on . In particular, the
left hand side of (4.8) is bounded above by

X D E
lim sup (2k ) d sup 2 ^b ; rb h Dk ( ; h)
k!1 h b2k

X D E X (4:9)
= lim (2k ) d 2 ^b ; rb hk D
b k
( ; h )
k!1 b2k b2k
for some sequence of Fk -measurable functions hk in L2 ( ).

Since by Lemma 2.1 y b = b+yy , ^b is translation covariant in the sense that
y ^b = ^b+y . Therefore, recalling that b = (b1; b2 ), that is translation invariant
and that y rb = rb+y y
D E D E D E
^b ; rb h = b ^b ; b rb h = ^(0;b b ) ; r(0;b b ) b h
1 1 2 1 2 1 1
On the other hand, for every oriented bond b = (b1 ; b2 ), a change of vari-
ables shows that < r(b ;b ) f; r(b ;b ) g > is equal to < r(b ;b ) f; r(b ;b ) g > .
In particular, since b = (1=4)rb( L ) 1 , for 1 i d, we have that
1 2 1 2 2 1 2 1
< ^(ei ;0); r(ei ;0)h > =< ^(0;ei ) ; r(0;ei )h > . Thus,
X D E
2(2k ) d ^b ; rb hk
b2k
d D
X X E
= 4 ^(0;ei ) ; r(0;ei )(2k) d x hk
i=1 x; x2k
x+ei 2k
On the other hand, by Schwarz inequality and the bound on the Dirichlet form
of h,
d D
X X 2 E
r(0;ei ) (2k) d x hk
i=1 x; x2
x+ei 2kk
d
X X D 2 E
(2k ) d rx;x+ei hk + O (k 1
)
i=1 x; x2k
x+ei 2k
X
= 2(2k ) d Db ( ; hk ) + O (k 1
):
b2k
The remainder O(1=k ) appeared because last summation is carried over (2k +
1)d 1 (2k ) bonds, while we are dividing only by (2k )d and the Dirichlet form is
bounded by k d .
Therefore, if we denote by Rik the cylinder function
X
r(0;ei )(2k) d x hk ;
x; x;x+ei 2k
for 1 i d, the right hand side of identity (4.9) is bounded above by
d
X
D E D E
lim
k!1 i=1
4 ^(0;ei ) ; Rik (1=2) (Rik )2

The upper bound on the Dirichlet form of hk implies that the sequence of
vectors fRik ; k 1g is bounded in L2 ( ). There exists therefore a weakly con-
verging subsequence. Denote by Ri a weak limit and assume, without loss of
generality, that the sequence Rk converges weakly to R = (R1 ; : : : ; Rd ). Since the
L2 norm may only decrease along weakly converging subsequences, the limit of
last sum, as k " 1, is bounded by
d D
X E D E
Ri (1=2) (Ri )2
4 ^(0;ei ) ;
i=1
It is not difficult to check that Ri is a germ of a closed form in the terminol-
ogy of Definition A3.4.12. Therefore, according to Theorem A3.4.14, R can be
decomposed asPa sum of the germs fAj ; 1 j dg defined in (4.1) and of a
gradient : R = 1id aj Aj + r g for some cylinder function g in C0 . Therefore,
the right hand side of (4.9) is bounded above by
d D
X X E
sup 4 ^(0;ei ) ; aj (Aj )i + r(0;ei ) g
a2Rd i=1 1j d
g2C0
D X E
=
aj Aj + r g :
2
(1 2)
1j d
To conclude the proof of the lemma it remains to check that

d D
X E X D E
4 ^(0;ei ) ; (Aj )i = 2 xj ; (x) = 2 ;j
i=1 x2Zd
for 1 j d and
d D
X E X D E
4 ^(0;ei ) ; r(0;ei ) g = 2 ; x g = 2 ; g ;0
i=1 x2Zd
for each cylinder function g in C0 . The first identity relies on the integration by
parts formula for the current : by definition of the germ Aj , summation over i
is equal to 4 < ^(0;ej ) >. By definition of ^b and a change of variables, this
expectation is equal to
X D E
r0;ej x ( L ) 1

x; 02x+
ej 2x+
X D E
= 2 ( L ) 1
; W x; x+ej
x; 02x+
ej 2x+
P P
A simple computation shows that L x2 xj (x) = z;z+ej 2 Wz;z+ej from
what the first identity follows. The second identity is elementary to check. One
has just to replace ^(0;ei ) by its value and use the integration by parts formula
after performing change of variables of type = x .
To conclude the proof of Theorem 4.1 , it remains to obtain a lower bound for
the variance. This is much easier since the variational formula for the variance is
expressed as a supremum.
Lemma 4.5 Under the assumptions of Theorem 4.1 ,

D X X E
lim (2`) d ( L` ) 1
x ; x `;K :
`!1 jxj` jxj` `
Proof. We need to obtain a lower bound for the variance given by the variational
formula
P (4.4). Consider
P a cylinderP
function g . For ` sg + 1, take h in (4.4) as
h = jxj`g x g + 1id ai x2` xi (x).
On the one hand, for 1 i d, the equivalence of ensembles gives that
D X X E
lim (2`) d x ; y g `;K = ; g ;0 ;
`!1 jxj` jyj`g `
D X X E
lim (2`) d x ; yi (y) `;K = ;i
`!1 jxj` y2` `
because belongs to C0 and, by assumption, K` =(2`)d ! .

On the otherPhand, to compute
P the limit P as ` " 1 of the Dirichlet form
(2`) dD(`;K` ; jxj`g x g + 1id ai x2` xi (x)), recall the elementary
P P
identity L` x2` xi (x) = x;x+ei 2` Wx;x+ei . We may therefore decompose
the Dirichlet form as a sum of terms of three kinds :
D X X E
(2`) d L` x g ; y g `;K ;
jxj`g jyj`g `
D X X E
2(2`) d Wx;x+ei ; y g `;K
x;x+ei 2` jyj`g `
D X X E
and (2`) d Wx;x+ei ; yj (y) `;K
x;x+ei 2` y2` `
By the equivalence of ensembles, as ` " 1, the first term converges to

Lg; g ;0 that may be rewritten as (1=2)< kr g k2 > . By similar reasons
the second Pexpression converges to 2 W0;ei ; g ;0 and the third one to
< W0;ei ; y yj (y) >= i;j < W0;ei ; (ei ) > = (1=2)i;j < r0;ei > .
In conclusion, the limit, as ` " 1, of (4.4) is bounded below by
d
X d
X
2 ; g ;0 +2 ai ;i +2 ai W0;ei ; g ;0
i=1 i=1
kak2 < r 1D
kr k
E

2
;ei >
0
2
g 2

To conclude the proof of the lemma it remains to take a supremum over a 2 Rd
and g in C0 . and recall formula (4.2)
We conclude this section proving that for each in C0 the function
: [0; ] ! R+ that associates to each density the value is continuous
and that the convergence of the finite volume variances to is uniform
on [0; ]. For each ` in N Pand 0 K (2` + 1)d, denote by V` (K=(2` + 1)d )
the variance of (2` + 1) d
jxj` x with respect to `;K :
D X X E
V` (K=(2` + 1)d) = (2`) d ( L` ) 1
x ; x `;K
jxj` jxj`
We may interpolate linearly to extend the definition of V` to the all interval [0; ].
With this definition V` is continuous. Theorem 4.1 asserts that V` (K` =(2` + 1)d)
converges, as ` " 1, to , for any sequence K` such that K` =(2`+1)d ! .
In particular, lim`!1 V` (` ) = for any sequence ` ! . This implies
that is continuous and that V` ( ) converges uniformly to as
` " 1. We have thus proved the following theorem.
5. The diffusion coefficient 171
Theorem 4.6 For each fixed h in C0 , h is continuous as a function of the

density on [0; ]. Moreover, the variance
X X
(2`) d < ( L` ) 1
x h ; x h >`;K
jxj`h jxj`h
converges uniformly to h as ` " 1 and K` =(2`)d ! . In particular,
D X X E
lim sup (2`) d ( L` ) 1
x h ; x h `;K
`!1 0K (2`+1)d jxj`h jxj`h
= sup h :

0
5. The diffusion coefficient
We investigate here the main properties of the semi norm introduced

in the previous section and of the diffusion coefficient defined in the beginning
of the chapter. We first show that we may define from a semi–inner
product on C0 through polarization.
Lemma 5.1 For every g , h in C0 and 2 R

(a) g 0,
(b) g = 2 g and
(c) (parallelogram identity) g + h + g h = 2f g + h
g.
The proof of this lemma is elementary. On C0 C0 let ; be defined
by
1n o
g; h = 4
g + h g h : (5:1)
It is easy to check that (5.1) defines a semi–inner product on C0 : for all g1 , g2 , h

in C0 and 2 R we have that
(a) (symmetry) g1 ; g2 = g2 ; g1 ,
(b) (linearity) g1 + g2 ; h = g1 ; h + g2 ; h and
(c) (positiveness) h 0.
Linearity is a simple consequence of the parallelogram identity in Lemma 5.1 and
its proof can be found in any standard text on functional analysis.
1=2
Denote by N the kernel of the semi–norm on C0 . Since ;
is a semi–inner product on C0 , the completion of C0 jN , denoted by H , is a Hilbert
space.
Recall that the linear space generated by the currents fW0;ei ; 1 i dg and
LC0 = fLg; g 2 C0 g are subsets of C0 . The first main result of this section consists
in showing that H is the completion of LC0 jN + fW0;ei ;P1 i dg, in other
words, that all elements of H can be approximated by 1id ai W0;ei + Lg
for some a in Rd and g in C0 . To prove this result we derive two elementary
identities :
h; Lg = h; g ;0 and h; W0;ei = h ;i
(5:2)
for all h, g in C0 and 1 i d.
These identities are easily explained. By Theorem 4.1 and (5.1), thePsemi–inner
product h; g is the limit of the covariance (2`) d < ( L` ) 1 jxj`g x g ,
P
jxj`h x h >`;K` , as ` " 1 and K`=(2`) ! . In particular, if g = Lg0, for
d
some cylinder function g0 , since x Lg0 = LP` x g0 for jxj `P
g0 , we have that
h; Lg = lim`!1 (2`) d < ( L` ) 1 jxj`g0 L` x g0 ; jxj`h x h >`;K`
for some sequence K` such that K` =(2`)d ! . The inverse of the generator
cancels with the generator. Therefore, h; Lg is equal to
X X
lim (2`) d < x g; x h >`;K` = g; h ;0 :
`!1 jxj`g jxj`h
The second identity
P is proved inP
a similar way, we just need to recall the elementary
relation L` x2` xj (x) = x; x;x+ej 2` Wx;x+ej . It is also possible to prove
both identities directly from the definition of the semi–norm through
=
1 2
the variational formula. We leave the second proof to the reader as an exercise.
It follows from the first identity that the gradients f (ei ) (0); 1 i dg
are orthogonal to the space LC0 , while the second identity permits to compute
inner product of cylinder functions with the current :
(ei ) (0); Lh = 0 for all 1 i d and all h in C0.
(ei ) (0); W0;ej = ()i;j (5:3)
and W0;ei ; W0;ej = (1=2) < r0;e > i;j 1
for 1 i; j d. In this formula () stands for the static compressibility and is

equal to < (0)2 > < (0) >2 . Furthermore,
X d Dn X
X o2 E
aj W0;ej + Lg = (1=2) aj (Aj )i + r(0;ei ) g
(5:4)
jd
1 i=1 1jd
for a in Rd and g in C0 . In particular, by (5.2), the variational formula for h
writes
h

=
X X
sup 2 h; aj W0;ej + Lg aj W0;ej + Lg :
a2Rd jd jd
g2C0 1 1
(5:5)
We may now prove that in H a function can be approximated by

X
aj W0;ej + Lg
1 jd
for some a in Rd and g in C0 .
Proposition 5.2 Recall that we denote by LC0 the space fLg ; g 2 C0 g. For each
0 ,
H = LC0N fW0;ei ; 1 i dg :
P
Proof. Let us first show that the sum is direct. Suppose 1j d aj W0;ej belongs
to LC0 for some vector a. There
P exists, therefore, a sequence of functions gk in C0
such that Lgk converges to 1j d aj W0;ej . Take the inner product with respect
to (ei ) (0). By (5.3),
X
()ai = aj W0;ej ; (ei ) (0)
jd
1
= lim
k!1
Lgk ; (ei ) (0) = 0 :
Thus aj = 0 for 1 j d proving that the sum is direct.
We now turn to the proof that H is generated by LC0 and the currents.
Since fW0;ei ; 1 i dg and LC0 are contained in C0 , by definition, H
contains the right hand space. To prove the converse inclusion, let h 2 C0 so that
h; W0;ei = 0 for 1 i d and h; Lg = 0 for g in C0. From (5.5) it
follows that h = 0. Thus, C0 jN (LC0 + fW0;ei ; 1 i dg)jN .
Corollary 5.3 For each g in C0, there exists a unique vector a in Rd such that
d
X
g aj W0;ej 2 LC0 in H :
j =1
From (5.3) we know that the space generated by f (ei ) (0); 1 i dg

is orthogonal to LC0 . Thus H is a Hilbert space with a quite simple structure. It
is generated by LC0 and the current fW0;ei ; 1 i dg and has inner product
defined by (5.1). Moreover, f (ei ) (0); 1 i dg and LC0 are orthogonal in
H .
We shall now start to describe the diffusion coefficient D of the hydrodynamic
equation. From Corollary 5.3, there exists a matrix fQi;j ; 1 i; j dg such that
d
X
(ei ) (0) + Qi;j W0;ej 2 LC0 in H : (5:6)
j =1
We replaced the minus sign by a plus for the matrix Q to be positive as we shall
see in the next lemma. Notice that the matrix Q = Q() depends on the density
because the inner product depends on . We claim that Q is symmetric, strictly
positive and has all eigenvalues bounded below by a finite constant.
Lemma 5.4 Consider the matrix Q defined by (5.6). Recall that () denotes the
static compressibility and is equal to < (0)2 > < (0) >2 . Q is a symmetric,
strictly positive matrix with eigenvalues bounded below by f2()= < r0;ei > g.
Proof. The proof of this lemma is quite simple. Since the vectors f(ei)
(0); 1 i dg are orthogonal to LC0,
d
X
(ei ) (0); (ek ) (0) = Qi;j W0;ej ; (ek ) (0)
j =1
= ()Qi;k
(5:7)
because, by (5.3), W0;ej ; (ek ) (0) = ()j;k . Q is therefore symmetric
and strictly positive. It remains to show that all eigenvalues of Q are bounded
below by f2()= < r0;ei > g.
To keep notation simple, assume that there exist functions Hi in C0 , 1 i
d, so that (ei ) (0) + dj=1 Qi;j W0;ej = LHi . Otherwise, we approximate
P
(ei ) (0) + dj=1 Qi;j W0;ej by a strongly convergent sequence fLHin; n 1g

P
with Hin in C0 for each 1 i d and n 1.

Taking inner product on both sides with respect to W0;ek , by (5.3), we obtain
that ()i;k + (1=2)< r0;e > Qi;k = LHi ; W0;ek . Denote by M the
matrix with entries Mi;k = LHi ; W0;ek . Thus,
1
()I + (1=2) < r0;e > Q = M ;

1
if I stands for the identity. Taking now innerPproduct with respect to LHk , since
(ei ) (0) is orthogonal to LC0, we get that 1jd Qi;j W0;ej ; LHk =
LHi ; LHk . Notice that the matrix with entries LHi ; LHk is positive
definite. Therefore, QM 0 in the matrix sense, if M denotes the adjoint of
M . Since, by the first part of the proof, M = ()I + (1=2)< r0;e > Q, 1
n o
Q ()I (1=2) < r0;e1 > Q 0
in the sense of matrices. Let be an eigenvalue of Q and denote by v an associated

eigenvector. is positive because Q is a positive definite matrix. The previous
inequality asserts that
D E
0 [()I (1=2) < r0;e1 > Q]v; Qv
n o
= () (1=2) < r0;e > kv k2 :
1
Thus, (() (1=2) < r0;e1 > ) 0. Since is positive, is bounded below
by f2()= < r0;ei > g.
Denote by D = D() the inverse of Q. We shall see below that D() is
the diffusion coefficient of the hydrodynamic equation (0.8). From the previous
lemma, D is symmetric, positive definite, with eigenvalues bounded above by
f< r0;ei > =2()g :
D < 2r0;e(i>
)
I (5:8)
in the sense of matrices. Our purpose now is to obtain an explicit formula for
D and then prove that D is continuous on [0; ] and nonlinear. Since D is the
inverse of Q, we have that
d
X
W0;ei + Di;j [(ej ) (0)] 2 LC0 in H
j =1
for 1 i d. This relation provides a variational characterization of the diffusion
coefficient D. Indeed, for all vectors a in Rd ,
n d
X d
X o
inf
g2C0
ai W0;ei + ai Di;j [(ej ) (0)] Lg = 0:
i=1 i;j =1
Since gradients are orthogonal to the space LC0, since
(ej ) (0); W0;ei = ()i;j
and since, by (5.7),
(ej ) (0); (ek ) (0) = ()Qj;k = ()[D 1 ]j;k ;

the last identity reduces to
n d
X o
inf
g2C0
()a Da + ai W0;ei Lg = 0;
i=1
where a stands for the transposition of a. We have thus obtained a variational
formula for D().
Theorem 5.5 The diffusion coefficient D() is such that

d
X
a D()a =
1
() ginf
2C0
aj W0;ej Lg
j =1
d
X D Xd 2 E
=
1
2() g2C0
inf aj (Aj )i r(0;ei ) g
i=1 j =1
for all a in Rd .
The second identity follows from equation (5.4). Moreover, this formula de-
termines the matrix D since D is symmetric by Lemma 5.4.
We now prove that the diffusion coefficient D is continuous. The proof is
divided in three steps. We first show that D is continuous on the open interval
(0; ). Then, taking advantage of the integration by parts formula for (ei ) (0),
we prove a lower bound for D. This lower bound in addition to the upper bound
(5.8) shall prove that D is continuous at the boundary of [0; ].
The following functional space plays a key role in the proof of the continuity of
the diffusion coefficient. Denote by F the space of functions f : [0; ] Z d ! R
such that
(i) For each 2 [0; ], f(; ) is a mean-zero cylinder function with uniform
support : there exist a finite set Zd that contains the support of each
f(; ) and the expected value of f(; ) with respect to all canonical measures
;K vanishes :
E;K [f(; )] = 0 for all 0 K jj :
(ii) For each configuration , f(; ) is a smooth function of class C 2 ([0; ]).
Theorem 5.6 The diffusion coefficient Di;j ( ) is continuous on (0; ).
Pd
Proof. Fix " > 0 and in [0; ]. Since W0;ei + j =1 Di;j ()[ (ej ) (0)] belongs
to LC0 , there exists a cylinder function Hi (; ) in C0 such that
d
X
W0;ei Di;j ()[(ej ) (0)] LHi (; ) " :
j =1
Since by Theorem 4.6 h is continuous in for all h in C0 , for each
P0 in [0; ], there exists a neighborhood O0 of 0 such that W0;ei
d D ( )[ (e ) (0)] LH ( ; ) 2" for in O . The family
j =1 i;j 0 j i 0 0
fO ; 2 [0; ]g forms an open covering of the compact set [0; ]. There exists
therefore a finite subcovering fOk ; 1 k ng.
From fDi;j (k ); 1 k ng and fH (k ; ); 1 k ng it is possible to
" : [0; ] ! R, 1 j d, and a
define by interpolation continuous functions Di;j
function H (; ) in F so that
d
X
W0;ei + " ()[ (ej ) (0)] LH " (; ) 4" :
Di;j
sup i
0 j =1
We now prove that the continuous functions Di;j" uniformly approximate Di;j
P
on compact sets of (0; ). On the one hand, by Schwarz inequality, j [Di;j
"
Di;j ][(ej ) (0)] LfHi (; ) Hi (; )g is bounded above by 2 W0;ei
"
P P
j Di;j ()[ (ej ) (0)] LHi (; ) +2 W0;ei
" "
j Di;j ()[ (ej )
(0)] LHi(; ) . By the previous estimate, this last sum is bounded above by
10". On the other hand,
P
since the vectors f (ej ) (0); 1 j dg are orthogonal
to the space LC0 , j [D Pi;j "Di;j ][ (ej ) (0)] L[Hi (; ) Hi (; )]
" "
is bounded below by j [Di;j Di;j ][ (ej ) (0)] . In conclusion, we
have that
X d
[Di;j " Di;j ][ (ej ) (0)] 10" :
j =1
Recall the definition of the matrix Q defined in (5.6) and keep in mind that
Q()Pis the inverse of the diffusion coefficient D. By (5.7), last sum thus writes
() j;k Bi;j" Qj;k B " . Here, to keep notation simple, we denoted the difference
" i;k
" . Since
Di;j Di;j by Bi;j Q is bounded below by 2()= < r0;e > , We obtain1
that
2()2 X h i
" Di;j 2 10" :
< r0;e > 1jd D i;j
1
This proves that D is uniformly approximated by continuous functions on any

compact set of (0; ). In particular, D is continuous on (0; ).
We are now ready to prove a lower bound for the diffusion coefficient.
Lemma 5.7 For every a in Rd , we have that
a Da 8 < 2(r) > kak2 :

0;e 0;e1 1
Here 0;ei is the cylinder function defined in (2.2) and related to the gradients by
the integration by parts formula.
Proof. By definition of the static compressibility () = (1=2) < (ei )

(0); (ei) (0) >. Since r0;ei [(ei ) (0)] = 2r0;ei , by the integration by
parts formula (2.2) and Remark 2.2, for all a in Rd
d
X D E
kak2() = 2 a2i 0;ei ; r0;ei
i=1
Up to the end of this proof we denote by the inner product on Rd . Recall the
definition of the germs fAi ; 1 i dg, as vectors of Rd . We may rewrite this
last identity as
d
Dh X d
i hX iE
(1=2)kak2() = ai 0;ei Ai ai Ai
i=1 i=1
In contrast, since gradients are orthogonal to the space LC0, by the integration by
parts formula, for every g in C0 ,
d
X
0 = (1=2) ai [(ei ) (0)] ; Lg
i=1
d
DX E
= (1=2) ai [(ei) (0)] ; g
i=1
d
Dh X i E
= ai 0;ei Ai r g
i=1
Adding the two previous identities and applying Schwarz inequality, we obtain
that
Dh d
X d
i hX iE 2
(1=4)kak () =
4 2
ai 0;ei Ai
ai Ai + r g
i=1 i=1
D Xd 2 E D X d 2 E
ai 0;ei Ai aj Aj + r g
i=1 j =1
for every g in C0 . Minimizing over all g in C0 , by the variational characterization
of the diffusion coefficient presented in Theorem 5.5, we obtain that
a Da ()kak4
D P 2 E (5:9)
d
8 i=1 ai 0;ei Ai
Since the vectors Aj are orthogonal, the denominator is equal to 8kak2 <
02;e1 r0;e1 > .
It is now easy to prove that the diffusion coefficient is continuous at the
boundary of [0; ]. By duality among particles and holes we need only to check the
continuity at one of the boundary points, say the origin. From the explicit formulas
for (), < r0;e1 > and < 0;e1 r0;e1 > , we have that (R(')) = ' + O('2 ),
< r0;e1 >R(')= ' + O('2 ) and < 02;e1 r0;e1 >R(') = (1=4)' + O('2 ). Thus, by the
previous lemma, a D(R('))a (1=2)kak2 + O('). In contrast, by the lower bound
(5.8) for the diffusion coefficient, a D(R('))a (1=2)kak2 + O('). Therefore,
Theorem 5.8 The diffusion coefficient D() is continuous on [0; ]. Moreover it

converges to (1=2)I as # 0 or " .
From the continuity of the diffusion coefficient and the proof of Theorem 5.6
we have
Corollary 5.9 Let D be the matrix defined in Theorem 5.5. Then, for each 1
i d,
d
X
inf sup
2C0 0
W0;ei + Di;j ()[(ej ) (0)] Lf() = 0 :
f
j =1
This result together with (3.5), the definition of V~if;` and Theorem 4.6 con-
cludes the proof of Theorem 1.1.
Proof of Corollary 5.9. Fix 1 i d and " > 0. From the proof of Theorem 5.6,
there exists H (; ) in F such that
d
X
sup W0;ei + " ()[ (ej ) (0)] LH (; ) " :
Di;j
0 j =1
Fix a positive integer ` and set f( ) = H ( ` (0); ). Notice that, for sufficiently
large `, f belongs to C0 because H is in F. By the triangle inequality,
d
X
sup W0;ei + " ()[ (ej ) (0)] Lf( )
Di;j
0 j =1 (5:10)
2 sup Lf LH (; ) + 2" :
0
By identity (5.4) with a = 0,
d
1 X Dn X h io2 E
Lff H (; )g = 2
r (0;ei ) x H (`(0); ) H (; )

i=1 x2Zd
Since r(0;ei ) x = x r( x; x+ei ) and since is translation invariant, the previous
expression is equal to
d
1 X Dn X h io2 E
r `
(x;x+ei ) H ( (0); ) H (; )
2
i=1 x2Zd
Since H belongs to F, there exists a cube such that
X n o
r(x;x+ei ) H (` (0); ) H (; )
x2Zd n o (5:11)
X
= r(x;x+ei ) H (` (0); ) H (; ) + O (` 1
):
x2
The second term on the right hand side comes from a jump of a particle from `
to c` or from a jump in the opposite direction. For ` large enough the contribution
of this jump is H ( ` (0) (2` + 1) d ; ) H ( ` (0); ). Since H belongs to F, this
difference is bounded by C (H )` d. Summing over all sites at the boundary of ` ,
we obtain (5.11).
From identity (5.11) and since for every bond b and every L2 ( ) function h,
< (rb h)2 > 4 < h2 > , we obtain that the first term in (5.10) is bounded
above by
n Dn o2 E o
sup C (H ) H (` (0); ) H (; )
+ O (` 2
)

that vanishes as ` " 1 by the law of large numbers. This concludes the proof of
the corollary.
We conclude this section proving that the diffusion coefficient is nonlinear.
Proposition 5.10 D() is a nonlinear function and

()kak4 < r0;ei > kak2
2 E a D ( ) a
2()
D P
d
8 i=1 ai 0;ei Ai

for every a in Rd .
Proof. The inequalities were proved in (5.8) and (5.9) so that D(0) = D() =
(1=2)I . By the upper bound for D, that may also be obtained setting g = 1 in
Theorem 5.5, we have
a D(R('))a 12 f1 2' + O('2 )gkak2
for ' small. This proves that D is nonlinear.
6. Compactness
We prove in this section that the sequence QN is compact.
Theorem 6.1 The sequence of probability measures QN is relatively compact.

Moreover, every limit point Q is concentrated on absolutely continuous paths
(t; du) with density bounded above by : (t; du) = (t; u)du, (t; u) .
The proof of this theorem relies on the following exponential estimate.
Lemma 6.2 For any smooth function H : Td ! [0; 1], 1 i d and s < t :
n Z t o
X
E N exp d N N1 d H (x=N )Wx;x+ei (r)dr
s x2TdN
n X o
2 exp (1=2)(t s) H (x=N )2
x2TdN
6. Compactness 181
Proof. Since ejxj ex + e x, it is enough to prove that

n Z t o
X
E N exp N d N 1 d H (x=N )Wx;x+ei (r)dr
s x2TdN
n X o
exp (1=2)(t s) H (x=N )2
x2TdN
for every smooth function H : Td ! [0; 1], 1 i d and s < t.
Fix a smooth function H : Td ! R+ . By Feynman–Kac formula (A1.7.5) and
by stationarity of N ,
nZ t o
X
E N exp N H (x=N )Wx;x+ei (r)dr
s x2TdN
nZ t s o
X
= E N exp N H (x=N )Wx;x+ei (r)dr e(t s)N (H ) ;
0
x2TdN
P
where N (H ) is the largest eigenvalue of N 2 LN + N x2TdN H (x=N )Wx;x+ei and
has the variational expression :
n X o
N ( H ) = sup N H (x=N ) < Wx;x+ei f > (1=2)N 2DN (f ) :
f x2TdN
By the integration by parts formula (2.1) for the current Wx;x+ei and a computation
similar to the one performed just after (3.2), H (x=N ) < Wx;x+ei f > is bounded
above by (A=8)H (x=N )2 < (rx;x+ei + rx+ei ;x )f > +(2=A)Dx;x+ei (f ) for every
A > 0. Since the jump rate is bounded by 1, setting A = (2=N ), we obtain that
the expression inside braces in last formula is bounded above by
X
(1=2) H (x=N )2 ;
x2TdN
what concludes the proof of the lemma.
Corollary 6.3 Assume that T > 1=2. For any smooth function H : Td ! R+ and
1 i d, there is a constant C = C (H; T ) depending only on H and T such that
for any N and any small enough :
Z t X p
E N sup

N1 d H (x=N )Wx;x+ei (r)dr C log 1 :
jt sj< s x2TdN
s<tT
0
Rt P
Proof. Denote the time integral 0 N 1 d x2TdN H (x=N )Wx;x+ei (r)dr by g (t).
By the entropy inequality, E N [sup jg (t) g(s)j] is bounded above by

C1 () H (N j N ) + C1 () log E N exp nN d C () 1 sup jg(t) g(s)jo (6:1)
Nd Nd 1
where C1 ( ) is a function of to be determined later. Since there are at most

particles per site, the relative entropy H (N jN ) is bounded by C (; )N d .
Recall the Garsia–Rodemich–Rumsey inequality (cf. Stroock and Varadhan
(1979)) that can be stated as follows. Let g (t) a continuous function and (u), p(u)
strictly increasing functions such that (0) = 0, p(0) = 0 and limu!1 (u) = 1,
and define B as Z T Z T
ds dt jgp((tj)t gs(js))j :

B =
0 0
Then, Z 4 B p(du) :
sup jg(t) g(s)j 8 1
u2
jt sj< 0
0t;sT
p
Set p(u) = u and (u) = expfN d ug 1 so that 1
(u) = N d log(1 + u).
Integrating by parts we get that
Z n 4B o 1 p n 4B o p
0
log 1 +
u2 pu du 2 log 1+
2 + 8
because B=(4B + u2 ) p1=4. For < e 2 , the right hand side of the last inequality
is bounded above by 8 log p1 f1 + log+ (4B + 2 )g, where log+ u = (log u) _ 0.
Therefore, choosing C1 ( ) = 32 log 1 , we get that
N d C1 () 1
sup jg (t) g(s)j 1 + log+ (4B + 2 ) :
In particular, the second term of (6.1) is bounded above by

p
log 1 n1 + log(4E N [B ] + 2 + 1)o :
32
Nd
Recalling the definitions of B and g (t), for 2 < 4T 2 1, which is possible because
we assumed T > 1=2, by Lemma 6.2, 4E N [B ] + 2 + 1 is bounded above by
Z T Z T Z t
X
4 dt ds E N
exp N j t sj 1 2 = H (x=N )Wx;x+ei (r)dr

0 0 s x2TdN
Z T Z T n X o
8 dt ds exp (1=2) H (x=N )2
0 0
x2TdN
n X o
= 8T 2 exp (1=2) H (x=N )2 :
x2TdN
This estimate together with the bound on the relative entropy obtained earlier in
the proof shows that (6.1) is bounded above by
6. Compactness 183
p n X o
32 log 1
C (; ) + N d 1 + log(8T 2) + (1=2) H (x=N )2 :
x2TdN
This concludes the proof of the corollary.
An estimate of the modulus of continuity of the trajectories follows immedi-
ately :
Corollary 6.4 For any smooth function H : Td ! R+ , there exists a constant

C = C (H; T ) depending only H and T such that
p
lim sup E N sup

< H; t > < H; s >
C (H; T )d log 1 :
N !1 jt sj<
0t;sT
Proof. Fix a smooth function H and consider the martingale M H (t) = M H;N (t)
< H; 0N > 0t N 2LN < H; sN > ds.
R
defined by M H (t) =< H; tN >
A simple computation shows that the quadratic variation of this martingale is
bounded by C (d; H )N d . In particular, by Doob’s inequality,
h i
E N sup

MtH MsH 2E N sup MtH

C (d; H )N d=2 :
jt sj< tT
0
0t;sT
On the other hand,

Z t Z t X
N 2 LN < H; sN > ds = N 1 d (@uNi H )(x=N )Wx;x+ei (r)dr :
0 0 i;x
To prove the corollary, it is therefore enough to show for each 1 i d that
Z t
X
lim supE N sup

N 1 d (@uNi H )(x=N )Wx;x+ei (r)dr

N !1 jt sj< s x2TdN
0s<tT
p
C (T; H ) log 1
which is the content of the previous corollary.

It is now easy to conclude the proof of Theorem 6.1.
Proof of Theorem 6.1. Since there are at most particles per site,
h i
PN sup < t ; 1 > A = 0
t0
for all A > . By Theorem 4.1.3, Remark 4.1.4 and Proposition 4.1.7, the tightness
of the sequence QN follows from this identity and the previous corollary.
Furthermore, since there are at most particles per site, for any
P continuous
function H : Td ! RR, j < tN ; H > j is bounded above by N d x jH (x=N )j
that converges to jH (u)jdu as N " 1. All limit points Q of the sequence
QRN are thus concentrated on paths (t; du) such that supt0 j < t ; H > j
jH (u)jdu. The trajectories are therefore absolutely continuous with density
bounded by Q -almost surely.
The nongradient method just presented is due to Varadhan (1994a) and Quastel
(1992). It permitted to extend to reversible nongradient systems the entropy method
presented in Chapter 5, provided the generator of the system restricted to a cube o
linear size ` has a spectral gap that shrinks as ` 2 . The integration by parts formula
for the current W0;ei is presented in Varadhan (1994a) and Quastel (1992). It was
extended to mean-zero functions by Esposito, Marra and Yau (1994). The proof
proposed here is taken from this latter article, as well as the one of Theorem 4.1.
Section 5 is a mixture of Esposito, Marra and Yau (1994) and Landim, Olla and
Yau (1997). The proof of the continuity of the diffusion coefficient, Theorem 5.6,
is taken from Landim, Olla and Yau (1997) while the proof of Corollary 5.9 is
taken from Funaki, Uchiyama and Yau (1995). The continuity of the diffusion
coefficient was already present in Varadhan (1994a) and Quastel (1992).
Wick (1989) proved the hydrodynamic behavior of a one-dimensional non-
gradient model in which the current can be written as the sum of a gradient
h h and a term of type Lf , for cylinder functions h and f . Kipnis, Landim
and Olla (1994) applied Varadhan (1994a) and Quastel (1992) ideas to derive
the hydrodynamic behavior of the symmetric generalized exclusion process. Xu
(1993) extended the nongradient approach to the non reversible setting by consid-
ering mean-zero asymmetric simple exclusion processes. Spohn and Yau (1995),
based on the variational formula presented in Theorem 5.5 for the diffusion co-
efficient, obtained a lower and an upper bound for the diffusion matrix of lat-
tice gases that are valid close to the critical temperature. They showed that
d (1 )() 1 D() d+ (1 )() 1, where d , d+ are univer-
sal constants and () is the static compressibility. Funaki, Uchiyama and Yau
(1995), assuming that the diffusion coefficient is smooth, applied the relative en-
tropy method to derive the hydrodynamic equation of nongradient lattice gases
that are reversible with respect to Bernoulli product measures. Komoriya (1997)
extended these ideas to asymmetric mean-zero exclusion processes with speed
change. Varadhan and Yau (1997) prove the hydrodynamic limit of Kawasaki
dynamics satisfying mixing conditions.
In the nongradient context, there are two problems that deserve to be stud-
ied. The first one consists in proving the hydrodynamic behavior of a nongradient
system without using any information on the size of the spectral gap. In another
direction, it would be interesting to derive the hydrodynamic behavior of nongra-

dient interacting Brownian particles.
Navier–Stokes equations. A fundamental question in mathematical physics
is the derivation and the interpretation of the Navier–Stokes equations. One of
difficulties in the interpretation of this equation is that it is not scaling invariant
and thus cannot be obtained by a scaling limit. Although this problem is still out
of reach for Hamiltonian systems, important progress has been made recently in
the context of interacting particle systems.
To fix ideas consider an asymmetric zero range process evolving on the lattice
TdN . The macroscopic evolution of the process under Euler scaling is described
by the first order quasi–linear hyperbolic equation
@t r() = 0 ;
+ (7:1)
P
where stands for the mean drift : = x2Zd xp(x) and () for the expected
value of the jump rate g ( (0)) under the invariant measure with density . Assume
that the system starts from a product measure with slowly varying parameter
associated to a profile 0 : Td ! R+ . We shall see in Chapter 8 that under Euler
scaling (times of order tN ) the density has still a slowly varying profile q N (t; u) =
E N [tN ([uN ])] that converges weakly (in fact pointwisely at every continuity
0 ()
point according to Theorem 9.0.2) to the entropy solution of equation (7.1) with
initial data 0 .
In the context of interacting particle systems with one conserved quantity the
Navier–Stokes equations takes the form
d
X
@t N + r(N ) = N 1
@ui Di;j (N )@uj N ; (7:2)
i;j =1
where D is a diffusion matrix. Three different interpretations have been proposed
for the Navier–Stokes corrections :
(a) The incompressible limit : Consider a small perturbation of a constant profile
: 0N (u) = + N 1 a0 (u). Assuming that this form persists at latter times
(N (t; u) = + N 1 a(t; u)) we obtain from (7.2) the following equation for
aN (t; u) = a(tN; u)
@t aN + N0 () raN + (1=2)00 () ra2N
d
X
= Di;j ()@u2 i ;uj aN + O (N 1
):
i;j =1
A Galilean transformation mN (t; u) = aN (t; u + Nt0 () ) permits to remove
the diverging term of the last differential equation and to get a limit equation
for m = limN !1 mN :
d
X
@t m + (1=2)00 () rm2 = Di;j ()@u2 i ;uj m : (7:3)
i;j =1
(b) First order correction to the hydrodynamic equation : Fix a smooth profile
0 : Td ! R+ and consider a process starting from a product measure with
slowly varying parameter associated to the profile 0 (). We have seen that
under Euler scaling the expected density q N (t; u) = E N [tN ([uN ])] has
0 ()
still a slowly varying profile that converges weakly to the entropy solution of
equation (7.1) with initial data 0 . This second interpretation asserts that the
solution of equation (7.2) with initial profile 0 approximates q N up to the
order N 1 :
lim N q N N = 0
N !1
in a weak sense as N " 1.
(c) Long time behavior : The third interpretation relies on the following obser-
vation. Denote by (t; u) the solution of the hyperbolic equation (7.1). In the
interacting case in finite volume with periodic boundary conditions, asymptot-
ically as t " 1, the entropy solution (t; u) converges to a stationary solution
which is constant along the drift :
Z 1
lim (t; ) = 1 (u) = 0 (u + r ) dr ;
t!1 0
provided 0 stands for the initial data. In particular, if we consider the asymp-
totic process under diffusive scaling, we expect it to become immediately
constant along the drift direction :
r Nlim E N [ ([uN ])]

!1 tN0( )
2 = 0
for every t > 0 and for any initial profile. In contrast, on the hyperplane
orthogonal to the drift the profile should evolve smoothly in time according
to a parabolic equation.
The third interpretation consists therefore in analyzing the behavior of the
solution of equation (7.2) in time scales of order tN on the hyperplane orthogonal
to the drift direction. Let bN (t; u) = (tN; u). From (7.2) we obtain the following
equation for bN :
d
X
@t bN + N r(bN ) = @ui Di;j (bN )@uj bN :
i;j =1
To eliminate the diverging term N r(bN ), assume that the initial data (and
therefore the solution at any fixed time) is constant along the drift direction :
r0 = 0. In this case we get the parabolic equation
d
X
@t b = @ui Di;j (b)@uj b (7:4)
i;j =1
that describes the evolution of the system in the hyperplane orthogonal to the drift.
Notice that while the first and the third interpretation concern the behavior of
the system under diffusive scaling, the second one is a statement on the process
under Euler scaling.
Dobrushin (1989) was the first to investigate the corrections to the hydrody-
namic equations. He considered the evolution of independent Markov processes
and proposed a systematic approach to deduce the corrections of all orders to the
hydrodynamics equations of interacting particle systems. The method has been
successfully applied to harmonic random oscillators in Dobrushin, Pellegrinotti,
Suhov and Triolo (1988) and in Dobrushin, Pellegrinotti and Suhov (1990).
Esposito and Marra (1994) deduced formally the Navier–Stokes equations
(
div A = 0
(7:5)
@t A + K0 A rA = rP + K1 r DrA
from Hamiltonian dynamics. Here P stands for the pressure and D for the diffusion
matrix. In the sequel Esposito, Marra and Yau (1994) proved the incompressible
limit for asymmetric simple exclusion processes in dimension d 3 : They con-
sidered an asymmetric simple exclusion process evolving on the torus TdN starting
form a product measure with slowly varying parameter associated to a profile
N0 (u) = + N 1a0 (u), where is a fixed parameter in (0; 1). Recall that in the
context of exclusion processes () = (1 ). Denote by tN (du) the corrected
empirical measure defined by
X
tN =
Nd
1
1
tN (x) x=N :
2
x2TdN
Notice the diffusive scaling of time and that the sum is divided by N d 1
instead
of N d . Esposito, Marra and Yau (1994) proved the following result :
Theorem 7.1 In dimension d 3, as N " 1, [tN0 () ] tN converges weakly

in probability to an absolutely continuous measure whose density is the solution
of the equation (7.3) with initial data a0 . Moreover, the diffusion coefficient D is
given by a variational formula.
Landim, Olla and Yau (1997) examined the question of the Navier–Stokes
equations from the point of view of the first order corrections. They considered
an asymmetric simple exclusion process evolving on the torus TdN starting from a
product measure with slowly varying parameter associated to a profile 0 . Denote
by q N (t; u) the expected density of particles at time t around u : q N (t; u) =
E N () [tN ([uN ])] and denote by N the solution of equation (7.2) with () =
(1 ). Landim, Olla and Yau (1997) proved
0
Theorem 7.2 For the asymmetric simple exclusion process in dimension d 3,

as N " 1, N (q N N ) converges weakly to 0 in some appropriate H 1 space.
Landim, Olla and Yau (1996) proved regularity properties of the diffusion
matrix D of the Navier–Stokes equation (7.2), (7.3). Landim and Yau (1997)
filled a gap left in the previous works showing that in the exclusion models,
each cylinder function h such that E [h] = (d=d)E [h] = 0 for all can be
approximate in some H 1 space by functions in the range of the generator.
Benois, Koukkous and Landim P studied an asymmetric zero range process
evolving on TdN with drift = x xp(x) along the first direction : = ce1
for some constant c = 6 0. Fix a profile 0: Td ! R+ constant along the drift direc-
tion : @u1 0 = 0 and consider as initial state a product measure with slowly varying
parameter associated to a profile 0 . Benois, Koukkous and Landim (1997) proved
that in dimension d 2 the empirical measure diffusively rescaled converges to
an absolutely continuous measure whose density is the solution of
d
X
@t = i;j @ui 0 ()@uj
i;j =1
with initial condition 0 . In this formulaPi;j stands for the covariance matrix of
the transition probability p() : i;j = x xi xj p(x). Janvresse (1997) obtained
the first order corrections to the parabolic hydrodynamic equations of Bernoulli
reversible speed change exclusion processes in dimension d 3.
The proofs of Theorems 7.1 and 7.2 rely on the relative entropy method present
in Chapter 6 and require a logarithmic Sobolev inequality, which at the moment
where this book has been concluded has been proved only for reversible general-
ized exclusion processes and lattice gases with mixing conditions in Yau (1996),
(1997).
More recently exclusion processes in which particles have velocities have been
considered. The dynamics can be briefly described as follows. Fix a finite set
Q Rd of possible velocities. Each particle has a velocity q and evolves on the
lattice TdN . An exclusion rule forbids the presence of two particles with the same
d
velocity at some site. The state space is thus ff0; 1gQgTN .
The evolution can be decomposed in two pieces. Particles evolve according
to random walks on TdN obeying the exclusion rule described above. Thus a
particle with velocity q at site x, independently from the others, waits a mean-one
exponential time at the end of which it jumps to x + y with probability p(y; q ).
If the chosen site is already occupied by a particle with velocity q the jump is
suppressed.
P Here the transition probabilities
p(; q) are such that their mean drift
are q : x xp(x; q ) = q .
Superposed to this jump dynamics there is a collision process that exchanges
the velocity of a pair of particles sitting on the same site. More precisely, if two
particles with velocities q1 and q2 are at x and there are no particles on this site
with velocities q10 , q20 , then simultaneously at rate 1 the particle with velocity qi
assumes velocity qi0 , i = 1, 2, provided q1 + q2 = q10 + q20 . This latter assumption is

imposed to guarantee the conservation of momentum.
Notice that both the total number of particles and the total momentum are
conserved by the dynamics. For x in TdN , denote by (x; q ) the total number of
particles with velocity q at site x, by I` (x; ) the total
P momentum at site x for
the configuration in the `-th direction : I` (x; ) = q2Q (q e` ) (x; q ) and by
tN;`(du) the corrected empirical measure associated to the moment in the `-th
direction :
X
tN;` = N d1 1 I` (x; t )x=N :
x2TNd
Starting from a product measure N with slowly varying parameter associated

to the profile (0 ; A10 ; : : : ; Ad0 ) with density 0 (u) = r + N 2 a0 (u) and momentum
in the `-th direction A`0 (u) = N 1 a0` (u), Esposito, Marra and Yau (1996) proved,
under some assumptions on the set Q, that the corrected empirical measure tN;`
diffusively scaled converges in probability to a Navier–Stokes equation of type
(7.5), provided the solution of this equation is smooth, and obtained a variational
formula for the diffusion matrix D.
Smooth solutions of the incompressible Navier–Stokes equations are only
known to exist for short period of times. To avoid this problem, Quastel and
Yau (1997) proved that the sequence of probability measures on the path space
induced by the process (tN;1 ; : : : ; tN;d ) and the initial state N is tight and
that all its limit points are concentrated on weak solutions of the incompressible
Navier–Stokes equation (7.5).
8. Hydrodynamic Limit of Asymmetric Attractive
Processes
We examine in this chapter an alternative method to prove the hydrodynamic be-

havior of asymmetric interacting particle systems. This approach has the advantage
over the one presented in Chapter 6 that it does not require the solution of the
hydrodynamic equation to be smooth. On the other hand its main inconvenience is
that it assumes the process to be attractive to permit the use of coupling arguments
and the initial state to be a product measure. To illustrate this approach we con-
sider an asymmetric attractive zero range process on the discrete d–dimensional
torus TdN . The generator of this Markov process, denoted by LN , is given by
X h i
(LN f )( ) = p(y)g((x)) f (x;x+y ) f () ; (0:1)
x2TdN
y2Zd
where p() is a finite range irreducible transition probability on Zd. Irreducible
means here that for every z in Zd, there exists a positive integer m and a sequence
0 = x0 ; x1 ; : : : ; xm = z such that p(xi+1 xi ) + p(xi xi+1 ) > 0 for 0 i m 1.
Throughout this chapter we assume the process to be attractive. We have seen
in Chapter 2 that this hypothesis corresponds to assume that the rate at which a
particle leaves a site is a non decreasing function of the total number of particles
at that site :
(H) g() is a non decreasing function.
To avoid minor technical difficulties we shall assume also that g () is bounded.
This hypothesis is not a crucial one and can be removed without problems.
We have seen in Chapter 1 that to study the hydrodynamic behavior of asym-
metric processes we have to speed up the process by a factor N . We denote
d
therefore by t the Markov process on N TN associated to the generator LN de-
fined by (0.1) speeded up by N .
A formal argument, assuming conservation of local equilibrium, permits to
foresee the hydrodynamic equation of asymmetric zero range processes. Indeed,
assume the initial state N to be a local equilibrium of profile 0 . For a smooth
function H : Td ! R, MtH defined by
192 8. Hydrodynamic Limit of Asymmetric Attractive Processes
X X
MtH = N d H (x=N )t(x) N d H (x=N )0 (x)
x2TdN x2TdN
Z t X
N d H (x=N )NLN s (x) ds (0:2)
0
x2TdN
is a martingale vanishing at 0. The additional factor N in the integral term comes
from the time renormalization. A discrete integration by parts shows that the
integral term is equal to
d Z
X t X X
N d yj p(y) (@uj H )(x=N )g(s(x)) ds + O (N 1
):
j =1 0
x2TdN y
For 1 j d, denote by j the drift in the j -th direction of the evolution of each
elementary particle : X
j = yj p(y) :
y
Since the martingale vanishes at 0, if N (t; x=N ) denotes the expected value of
the number of particles at time t and at site x,
X X
N d H (x=N )N (t; x=N ) N d H (x=N )N (0; x=N )
x2TdN x2TdN
X d X Z t h i
= N d j (@uj H )(x=N )E N g(s (x)) ds :
j =1 x2TdN 0
Recall from Chapter 2 that we denoted by () the expected value of the jump
rate g ( (0)) with respect to the invariant measure N . If there is conservation of
local equilibrium, around site x at time s the state of the process should be close
to some equilibrium. Since the equilibrium states are parametrized by the density
of particles and since we denote by N (s; x=N ) the density at time s at site x, the
state of the process should be close to NN (s;x=N ) . In consequence the expected
value of g (s (x)) should be close to (N (s; x=N )). From these considerations
we see that N (t; x=N ) should be the solution of
X X
N d H (x=N )N (t; x=N ) N d H (x=N )N (0; x=N )
x2TdN x2TdN
X d X Z t
= N d j (@uj H )(x=N )(N (s; x=N )) ds :
j =1 x2TdN 0
Thus, we expect the macroscopic behavior of asymmetric zero range processes to

be described by solutions of the first order quasi–linear hyperbolic partial differ-
ential equation
8. Hydrodynamic Limit of Asymmetric Attractive Processes 193
8
>
>
d
X
< @t + i @ui () = 0;
i=1 (0:3)
>
>
:
(0; ) = 0 () :
This equation is well known by now. We refer to Smoller (1983) for a back-
ground on properties of solutions of this equation. We just point out here and
review in Appendix 2 some of the main features.
First of all, even if the initial profile 0 () is smooth, solutions of equation (0.3)
may develop shocks. Therefore, there may not exist classical solutions for this
equation and we are forced to consider weak solutions in the sense of Definition
4.2.2. Enlarging, however, the set of admissible solutions in this way we loose
uniqueness and we have to introduce a criterion to pick among all weak solutions
the physically significant one. We shall consider the entropy condition proposed
by Kružkov.
Definition 0.1 (Entropy solutions) A bounded function : R+ Td ! R is an

entropy solution of equation (0.3) if :
(a) (Entropy inequality) For every c 2 R,
d
X
@t c + i @ui () (c) 0
i=1
in the sense of distributions on (0; 1) Td, that is, if for every smooth function
H : (0; 1) Td ! R with compact support,
Z n d
X o
dt du (@t H ) c + i (@ui H )() (c) 0 :
(0;1)T d i=1
(b) (L1 (Td) convergence at t = 0) (t; ) converges in L1 (Td ) to 0 () as t decreases
to 0 : Z
lim
t!0 Td
j(t; u) 0(u)jdu = 0 :
Notice that the initial data does not play any role in part (a) of this definition
since the smooth functions are assumed to have compact support on (0; 1) Td .
It is condition (b) that connects the solution to the initial data.
Kružkov (1970) proved the existence of a unique entropy solution of equation
(0.3) if the initial data is bounded :
Theorem 0.2 Assume that 0 2 L1 (Td). There exists a unique entropy solution
(t; u) of equation (0.3).
In dimension 1, if () is concave and = 1 > 0, the entropy inequality is
equivalent to require the solution not to have decreasing jumps :
lim inf (t; u)

u#u0
lim sup (t; u) for every u0 2 Td :
u"u0
In view of this discussion, our natural goal in this chapter is to prove that the
empirical measure converges in probability to the absolutely continuous measure
whose density is the entropy solution of the first order quasi-linear equation (0.3).
This is the content of the main theorem of this chapter. To state this result we
have to describe the initial measures considered. Let (N )N 1 be a sequence of
d
probability measures on N TN . We shall assume that
(M1) N is a product measure.
(M2) There exists an equilibrium state N that bounds above N :
N N for all N 1:
0 : T ! R+ as the profile associated to
(M3) Define N d N : on an hypercube of
N
length 1=N centered at x=N , 0 is equal to the expected value of the total
number of particles for N :
X
0N (u) = EN [(x)]1fu 2 N (x)g ;
x2TdN
where N (x) is a hypercube of length 1=N centered at x:
n o
N (x) = y 2 Rd ; xi =N 1=2N yi < xi =N + 1=2N for 1 i d :
We assume that N
0 converges in L1 (Td) to the initial profile 0 :
Z
lim
N !1 Td
jN0 (u) 0 (u)jdu = 0:
Of course such initial measures exist. Fix for example 0 in L1 (Td ) and define
N0 as the profile given by
Z
0N (u) = Nd 0 (v) dv if u 2 N (x)
N (x)
and N as the product measure with marginals
N f; (x) = kg = NN (x=N ) f; (0) = kg
0
for x in TdN .
It is easy to check that this sequence of initial states satisfies hypotheses (M1)–
(M3).
We are now ready to state the main theorem of this chapter.
Theorem 0.3 Fix a bounded profile 0 : Td ! R+ . Let (N )N 1 be a sequence of

d
probability measures on N TN satisfying assumptions (M1)–(M3). For each t 0
the empirical measure tN converges in probability to the absolutely continuous
measure (t; du) = (t; u)du whose density (t; u) is the entropy solution of equa-
tion (0.3).
We have presented in Chapter 4 an approach to prove this type of result

in which the first step is to show that the empirical measure N converges in
distribution to a Dirac measure. Thus, for a fixed time T > 0 and for a sequence
d
of probabilities (N )N 1 on N TN denote by QN N the probability measure on the
path space D([0; T ]; M+(T )) corresponding to the process tN with generator
d
(0.1) speeded up by N and starting from N . To keep notation simple we will
often omit the subscript of QN N .
We would like to show that QN converges to a Dirac measure concentrated on
the path (t; du) whose density is the entropy solution of (0.3). In view of Chapter
5 the first natural idea is to repeat in this context the entropy approach. Consider,
therefore, for a smooth function H : Td ! R, the martingale MtH defined by (0.2).
We showed that after a summation by parts this martingale is equal to
X X
N d H (x=N )t (x) N d H (x=N )0 (x)
x2TdN x2TdN
X d Zt X
N d j (@uj H )(x=N )g(s(x)) ds + O (N 1
):
j =1 0
x2TdN
At this point to close the equation we would like to replace the expression
X
N d (@uj H )(x=N )g (s (x))
x2TdN
P
by N d x2Td (@uj H )(x=N )(sN" (x)). Notice that once the replacement is per-
N
formed the equation is indeed closed since N" (x) is a function of the empirical
measure. In fact,
N" (x) = < N ; " (x=N ) >
if " = ";N is the approximation of the identity defined by
" ( u) N d 1f[ "; "]g(u) :

(2["N ] + 1)d
=
Once the equation is closed we would proceed as in the proof of the hydrodynamic
equation for symmetric simple exclusion processes to show that QN converges to
a Dirac measure.
There are however two serious problems in this approach. First of all the
generator LN is speeded up by N and not by N 2 . This factor N 2 was crucial in
the proof of the two blocks estimate. Thus the replacement lemma allows only
to replace g ( (x)) by ( ` (x)) where ` is an integer independent of N and that
increases to infinity after N . But ` (x) is not a function of the empirical measure
and hence the equation is not closed.
The second problem comes from the previously mentioned existence of several
weak solution of the hyperbolic equation (0.3). It is therefore not enough to close
the equation and to show that all limit point of the sequence QN are concentrated
on weak solutions of (0.3), we still have to guaranty that they are concentrated on
entropy solutions.
The second problem is solved proving an entropy inequality at a microscopic
level. This result is stated at Corollary 2.2 of the next section and relies on coupling
techniques allowed by the attractiveness assumption. It consists in showing that
for every c 2 R,
d
X
@t t` () c + i @ui (t` ()) (c) 0 (0:4)
i=1
in the sense of distributions on (0; 1) Td and for N " 1 and then ` " 1.
Notice that here also we need a two blocks estimate to replace t̀ () by tN" () in
order to obtain functions of the empirical measure. If the replacement was possible
it would follow from this entropy inequality that all limit points of the sequence
QN are concentrated on entropy solutions of (0.3) and to conclude the proof of
Theorem 0.3 it would remain to show the relative compactness and some L1 (Td )
continuity at t = 0 to ensure condition (b) of Definition 0.1.
Unfortunately there is a proof of the two blocks estimate in the asymmetric
case only for attractive nearest neighbor processes in dimension 1 (cf. Rezakhanlou
(1991)). A new approach is thus needed to close the equation. The idea here is
to introduce the measure valued solutions of equation (0.3). For a background on
this subject we refer to Di Perna (1985).
Definition 0.4 (Measure valued solutions) Denote by P (R+ ) the set of probability
measures on R+ . A measurable map : (0; 1) Td ! P (R+ ) is a measure valued
solution of (0.3) if
d
X
@t < (t; u); > + i @ui < (t; u); () > = 0
i=1
in the sense of distributions, that is, if for every smooth function H in CK1;1 (R+
Td ),
Z Z n d
X o
dt du @t H (t; u) < (t; u); > + i (@ui H )(t; u) < (t; u); () >
Z
i=1
+ H (0; u)0(u) du = 0:
Here and below, for a continuous function : R+ ! R+ , < (t; u); () >
stands for the expected value of with respect to the probability (t; u) and
CKm;n (R+ Td) for the space of continuous functions with compact support and
with m continuous derivatives in time and n continuous derivatives in space.
This concept is clearly weaker than the one of weak solution since to each
weak solutions (t; u) we may associate the measure valued solution defined by
(t; u)(d) = (t;u) (d). Since this concept is weaker it will be easier to prove
existence of measure valued solution of (0.3). In fact such solutions can easily be
obtained by the viscosity method which consists in adding a small diffusion term
to the hyperbolic equation (0.3) and in obtaining a measure valued solution as a
weak limit of classical solutions for the second order equation as the diffusion
coefficient vanishes.
The real problem is therefore to prove uniqueness of measure valued solutions.
To consider this problem we need some terminology. A measure valued solution
(t; u)(d) is said to be a Dirac solution if there exists a bounded measurable
function (t; u) such that (t; u)(d) = (t;u) (d).
From the previous discussion we know that to guarantee uniqueness we have
at least to impose an entropy condition and some L1 (Td ) continuity at t = 0.
Definition 0.5 (Entropy measure valued solutions) A measure valued solution

(t; u)(d) is said to be an entropy measure valued solution if
(a) (Entropy inequality) For every c2R
d
X
@t < (t; u); j cj > + i @ui < (t; u); j() (c)j > 0
i=1
in the sense of distributions on (0; 1) Td .
(b) (L1 (Td ) continuity at t = 0)
Z
lim
t!0 Td
du < (t; u); j 0 (u)j > = 0 :
All entropy measure valued solutions (t; u)(d) that are Dirac solutions
(t;u) (d) are such that the associated profile (t; u) is a weak entropy solution
of (0.3). Therefore the proof of uniqueness of entropy measure valued solutions is
reduced to the proof that all entropy measure valued solutions are Dirac solutions.
Moreover, condition (b) in Definitions 0.1 and 0.5 can be relaxed. It is enough to
prove that
1 t
Z Z
t!0 t 0
lim ds du < (s; u); j 0 (u)j > = 0 : (0:5)
Td
The following result solves the question of uniqueness (cf. Di Perna (1985)).
Theorem 0.6 Assume that the initial profile 0 is in L1 (Td ). An entropy measure
valued solution satisfying (0.5) instead of condition (b) is a Dirac solution if there
exists 0 such that
sup (t; u) [ 0 ; 0 ]c = 0 :
t;u
In view of this discussion on measure valued solutions of quasi-linear hyper-

bolic equations, for positive integers N and `, associate to each configuration
the Young measure
X
N;`() = N;` (du; d) = N d x=N (du) `(x)(d) :
x2TdN
denote by tN;` the Young measure at time t : tN;` = N;` (t ).
Notice first that for a continuous function H : Td ! R and a smooth function
: R+ ! R, the expression
X
N d H (x=N ) (t`(x))
x2TdN
is a function of the Young measure since it is equal to
Z
H (u) ()tN;`(du; d) =: < tN;` ; H (u) () >
In particular, for the martingale MtH defined in (0.2), the one block estimate closes
the equation
P in terms of the Young measure.
P More precisely, after replacement of
N d x H (x=N )g(s (x)) by N d x H (x=N )(s̀ (x)), we obtain that MtH is
equal to
< tN;`; H > < 0N;` ; H >

d
X Z t (0:6)
i ds < sN;`; (@ui H )() > + oN (1) :
i=1 0
Here we performed a discrete integration by parts and used the smoothness of H

to get that
X X
N d H (x=N )t(x) = N d H (x=N )t` (x) + O(`=N 2 )
x2TdN x2TdN
= < tN;` ; H > + O(`=N 2 ) :
Notice that the expression (0.6) is closed in N;` . Thus the introduction of the
Young measure N;` solves the first objection raised in the beginning of this
section. To prove that the Young measures converges in probability to the en-
tropy measure valued solution of (0.3), we first show that it converges in distri-
bution. Thus for integers N and ` and a sequence N of probability measures
on N TN , denote by QN;`
d N;` the probability measure on the path space
N = Q
D([0; T ]; M+(Td R+ )) corresponding to the process tN;` = N;`(t ) with gen-
erator (0.1) speeded up by N and starting from N .
To prove that all limit points are concentrated on measure valued entropy
solutions we need to prove that they are concentrated on absolutely continuous
measures t (du; d) = t (u; d)du that satisfy the conditions of Definition 0.5 and
Theorem 0.6. It is easy to prove that the measures are absolutely continuous in
the first coordinate (cf. section 3). To show that the limit points are concentrated
on measures that satisfy the entropy inequality, notice that inequality (0.4) may
be rewritten in terms of the Young measure as
d
X
@t < tN;`; j cj > + i @ui < tN;`; j() (c)j > 0 :
i=1
From this result it will be easy to prove that all limit points of QN;` are concen-
trated on measure valued solutions satisfying condition (a) of Definition 0.5. The
proof of condition (b) relies on coupling techniques and is the unique point in the
proof where we need the initial measures to be product. At this point, we are left
to show that the condition of Theorem 0.6 is satisfied and that the sequence is
relatively compact. These problems are solved in the third section of this chapter.
From this result we obtain the convergence in probability by standard arguments
already used in Chapter 4 for symmetric simple exclusion process.
Finally, we claim that the convergence in probability of the empirical measure
to an absolutely continuous measure whose density is the entropy solution of (0.3)
follows from the convergence of the Young measure N;` (du; d) to the entropy
measure valued solution t (u; d)du of (0.3). Indeed, for every smooth function
H , we have that
X
< tN ; H > = N d H (x=N )t(x)
x2TdN
X
= N d H (x=N )t` (x) + O(`=N 2 ) :
x2TdN
Since, by assumption, the Young measure N;` (du; d) converges in probability,
the first term on the right hand side, equal to < tN;` ; H >, converges to
Z
< t (u; d); > H (u) du :
Since t (u; d) is the measure valued entropy solution, t (u; d) = (t;u) (d) and
last integral is equal to Z
H (u)(t; u) du :
This shows that tN converges in probability to the entropy solution of (0.3).
1. Young Measures
We prove in this section, under restrictive assumptions regarding the initial profile
0 and the initial state N , that the Young measures N;` converges in distribution
to the entropy measure valued solution of equation (0.3).
Recall from section 1 that for positive integers N and `, to each configuration
we associate the Young measure N;`(du; d) on Td R+ defined by
X
N;`() = N;` (du; d) = N d x=N (du) `(x)(d) :
x2TdN
Thus, if H : Td R+ ! R is a smooth function, the integral of H with respect to
N;`(du; d), denoted by < N;`; H >, is equal to
X
< N;` ; H > = N d H (x=N; ` (x)) :
x2TdN
Denote by M+ (Td R+ ) the space of positive Radon measures on Td R+
endowed with the weak topology and fix a time T 0. Denote by L1 ([0; T ])
the space of bounded functions from [0; T ] to R endowed with the weak topology.
This topology is generated by the metric
R
T
dt hk (t)g(t) 0T dt hk (t)f (t)
R
X 1 0
d(f; g) =
2k 1 + R T
;
dt hk (t)g(t) 0T dt hk (t)f (t)
R
k 1 0
where fhk ; k 1g is a dense sequence of functions in L1 ([0; T ]). Denote,

furthermore, by L1 ([0; T ]; M+(Td R+ )) the space of bounded functions
: [0; T ] ! M+ (Td R+ ) such that for each function F in CK (Td R+ ), < ; F >
belongs to L1 ([0; T ]). This space is endowed with the weak topology generated
by the metric
dMV (; ~)

X 1 d(< ; Fk >; < ;
~ Fk >)
;
2k 1 + d(< ; Fk >; < ;
~ F k >)
=
k 1
where fFk ; k 1g is a dense, with respect to the uniform topology, sequence of
functions in CK (Td R+ ). In particular, a sequence n converges to some measure
valued function if < n ; F > converges weakly in L1 ([0; T ]) to < ; F > for
d
every F in CK (Td R+ ). For a sequence N of probability measures on N TN
N;`
satisfying assumptions of Theorem 0.3, denote by QN;` = QN the measure on
the space L1 ([0; T ]; M+(Td R+ )) induced by the process tN;` := N;` (t ) with
generator (0.1) speeded up by N and starting from N . Recall also that PNN = P
N
d
stands for the probability on D([0; T ]; NTN ) corresponding to the process t with
generator (0.1) speeded up by N and starting from N . Expectation with respect
to PN is denoted by E N .
1. Young Measures 201
The following theorem is the main result of this section.
Theorem 1.1 Let 0 : Td ! R+ be a Lipschitz continuous profile. For each N 1,

define N as the product measure with marginals given by
N f; (x) = kg = N(x=N )f; (0) = kg

0
x 2 TdN ; k 2 N :
Then, the sequence QN;` converges, as N " 1 and then ` " 1, to the probability
measure concentrated on the entropy measure valued solution of equation (0.3).
We have seen in Chapter 4 that the proof of the hydrodynamic behavior of

interacting particle systems comprehend essentially two steps. We first show that
the sequence of probability measures is weakly relatively compact and then prove
uniqueness of limit points, that is, that all converging subsequences converge to
the same limit. We start with the tightness.
Lemma 1.2 The sequence QN;` is weakly relatively compact.
Proof. It is well known that the unit ball of L1 ([0; T ]) for the strong topology is
compact for the weak topology. In particular, for each b > 0, the sets Ab = ff 2
L1 ([0; T ]); kf k bg are weakly compact.
Fix a countable, dense subset fGk ; k 1g of CK (Td R+ ). It is not difficult
to check that for any sequence of finite numbers fCk ; k 1g, the set
n o
; k < ; Gk > k1 Ck for all k 1
is compact in L1 ([0; T ]; M+(Td R+ )). In particular, the tightness of the sequence
QN;` follows from the existence of finite constants fC (F ); F 2 CK (Td R+ )g
such that h i
sup EQN;` < ; F > 1 C (F )
N;`

every F . This estimate in our case is trivial because, by definition, < ; F >
for
is trivially bounded by kF k1 .
1
It remains to investigate the uniqueness of limit points.
Proof of Theorem 1.1. In view of Definition 0.5 and of Theorem 0.6 of section
1, to prove that a probability Q is concentrated on the entropy measure valued
solution of equation (0.3) we have to show that Q almost surely
(i) For every 0 t T , t is absolutely continuous on Td :
t (du; d) = t (u; d) du :

(ii) For every 0 t T , t is concentrated on a compact set : there exists
0 > 0 such that
t (u; [0; 0]c ) = 0

for all (t; u) in [0; T ] Td .
(iii) Entropy inequality : for every c2R
d
X
@t < t ; j cj > + i @ui < t ; j() (c)j > 0
i=1
on (0; 1) Td in the sense of distributions.
(iv) Convergence in L1 (Td ) to the initial data:
Z t Z
lim inf
t!0
1
t ds du < s ; j 0 (u)j > = 0 :
0
For each fixed integer `, denote by Q;` the set of probability measures on
1
L ([0; T ]; M+(Td R+ )) that are limit points of the sequence QN;` with N " 1.
Denote also by Q the set of probability measures that are limit points of sequences
fQ;`; ` 1g such that, Q;` belongs to Q;` for each `. Fix Q in Q . We shall
prove that Q is concentrated on paths satisfying (i)-(iv).
We start with property (i). Fix G: R+ ! R and H : Td !R two bounded
continuous functions. From the definition of tN;` ,
X

< tN;`; G()H (u) > = N d

H (x=N )G(t`(x))
x2TdN
X
kGk1N d jH (x=N )j :
x2TdN
R
Since H is continuous, the right hand side converges to kGk1 jH (u)j du, as
N " 1. In particular,
h Z i

lim
N !1
QN;` sup < t ; G()H (u) > kGk1 jH (u)j du = 1:
tT
0
Since k k1 is lower semicontinuous for the weak topology of L1 ([0; T ]), the
above set is closed. Therefore, Q a.s.
Z

sup < t ; G()H (u) > kGk1 jH (u)j du :
tT0
This shows that Q a.s. the first coordinate of t is absolutely continuous :
t (du; d) = t (u; d)du :

Statement (ii) is a simple consequence of attractiveness and the boundness of
the initial profile. Indeed, for every A > 0,
hZ T i hZ T X i
EQN;` < t ; 1fjj Ag > dt = EN N d 1ft` (x) Ag :
0 0
x2TdN
To avoid minor technical difficulties we assumed in this chapter the jump rate to
be bounded. It follows from the boundness of g that the equilibrium measures
do not have all exponential moments finite ( EN [expf (0)g] = 1 for >
logfsupk g (k )g log ()). This characteristic of bounded jump rate zero range
processes prevent us in using the entropy inequality in order to show that the
last expectation is small for A large. Coupling arguments, available due to the
attractiveness assumption, replace the entropy inequality.
Recall that by assumption N N . Denote by StN the semigroup of the
Markov process (t ) with generator LN speeded up by N . By the attractiveness
assumption and since N is invariant N StN N . Since 1f ` (x) Ag is an
increasing function last expectation is bounded above by
T N [`(0) A] :
For A > , by the law of large numbers, this expression converges to 0 as `
increases to 1. We proved therefore that for A sufficiently large,
hZ T i
lim sup lim sup EQN;` < t ; 1fjj Ag > dt = 0:
`!1 N !1 0
! 0T < t ; 1fjj > Ag > dt is lower semicontinuous,

R
Since the application
hZ T i
EQ < t ; 1fjj > Ag > dt = 0
0
for A sufficiently large. Thus

Z T Z
dt du t (u; [0; A]c) = 0
0
Q a.s. Redefining, if necessary, t (u; d) in a subset of measure 0 of [0; T ] Td

we conclude the proof of (ii).
We now turn to property (iii). In the next section we prove the following
entropy inequality at the microscopic level :
Theorem 1.3 Let N be a sequence of probability measures bounded by some

equilibrium state N : N N . For every smooth positive function H with
compact support in (0; 1) Td , every constant c 2 R and every " > 0,
"Z
1 X

lim lim PN dt N d @t H (t; x=N )t` (x) c
`!1 N !1 N 0
x2TdN
#
d
X

+ i @ui H (t; x=N
) t` (x) c " = 1:
i=1
Notice that in this theorem, which is the main step toward the proof of Theorem
1.1 we only require the initial state to be bounded above.
We claim that (iii) follows from this result. Indeed, in terms of the Young
measure this result can be restated as
Z T D E
lim lim
`!1 N !1
QN;` dt t ; (@t H )(t; u)j cj
0
d
X D E
+ i t ; (@ui H )(t; u)j() (c)j " = 1
i=1
for all positive smooth functions H : (0; T ) Td ! R+ with compact support, all
c 2 R and all " > 0. Since Q is a weak limit point concentrated on absolutely
continuous measures, from this result and property (ii) already proved, we obtain
that
hZ T Z n
Q dt d
du (@t H )(t; u) < t (u; d); j cj >
0 T
d
X o i
+ i (@ui H )(t; u) < t (u; d); j() (c)j > " = 1:
i=1
Letting " # 0 and since the statement is valid for every smooth function H we
have that Q a.s.
d
X
@t < t ; j cj > + i @ui < t ; j() (c)j > 0
i=1
on (0; T ) Td in the sense of distributions for every c 2 R.
It remains to prove (iv). Here is the unique point in the proof of Theorem 1.1
where we need the initial measure N to be product. We claim that in order to
prove (iv) it is enough to show that
h i
lim lim sup lim sup EQN;`
t!0 `!1 N !1
< t ; j 0 (u)j > = 0: (1:1)
Indeed, it follows from (1.1) that

h1 Z t i
lim lim sup lim sup EQN;`
t!0 `!1 N !1 t 0 = 0:ds < s ; j 0 (u)j >
Since Q is concentrated on paths t such that t (u; [0; 0 ]c ) = 0 for large enough
0 and 0 () is a bounded profile, by weak convergence, for every t > 0,
hZ t i
EQ ds < s ; j 0 (u)j >
0
hZ t i
lim sup lim sup EQN;` ds < s ; j 0 (u)j > :
`!1 N !1 0
In particular,
h Z t i
lim Q ds < s ; j 0 (u)j >
1
t!0 t 0
= 0
and, by the dominated convergence theorem,

h Z t i
Q lim sup
1
t ds < s ; j 0 (u)j > = 0:
t!0 0
We now turn to the proof of (1.1). Recall that we denoted by 0N (x=N ) the
expected value of particles at site x for the measure N : 0N (x=N ) = EN [ (x)].
From the definition of the empirical measure N;` , we have that
h i h X i
EQN;` < t ; j 0 (u)j > = EN N d jt` (x) 0(x=N )j : (1:2)
x2TdN
We shall need to couple two and three coordinate processes starting from
product measures. We first define the initial measures and the evolution. For each
fixed N and sites y1 and y2 of TdN , consider the product measure ¯ N N
y1 (resp. ¯ y1 ;y2 )
d 2 d 3
on N TN (resp. N TN ) with first marginal equal to N , second marginal
equal to N0 (y1 =N ) (resp. third marginal equal to N0 (y2 =N ) ) and ordered at each site
z:
0 (z=N ) 0 (y1 =N ) ;
1 if N
¯ Ny1 f(; y1 ); (z ) y1 (z )g =
0 otherwise ;

0 (z=N ) 0 (y1 =N ) ;
if N
¯ Ny ;y f(; y ; y ); (z ) y (z )g
1
1 2
1 2 1
=
0 otherwise ;
an analogous inequality with y2 replacing y1 and

¯ Ny ;y f(; y ; y ); y y g = 1 if 0 (y1 =N ) 0 (y2=N ) ;
1 2
1 2 1 2
¯ Ny ;y f(; y ; y ); y y g = 1 otherwise :
1 2
1 2 1 2
In these last formulas and below y stands for the second coordinate and y
1
for 2
the third one in the three coordinates coupled process.

We let processes (t ; ty1 ; ty2 ) and (t ; ty1 ) evolve according to the basic cou-
pling defined in section 2.5 speeded up by N . Though we defined it only for two
coordinate models, it is straightforward to define a generator for a three coordinates
processes that preserve the order. Denote by P¯ N y1 (resp. P¯ N
y1 ;y2 ) the probability
measure induced by the basic coupling speeded up by N and the probability mea-
sure ¯ N N
y1 (resp. ¯ y1 ;y2 ). Expectation with respect to P¯ Ny1 (resp. P¯ Ny1 ;y2 ) is denote
by E ¯ N
y1 (resp. E ¯ N
y1 ;y2 ).
Expectation (1.2) is bounded above by
X h i X h i
N d E ¯ N
x jt` (x) tx;`(x)j + N d E ¯ N
x jtx;` (x) 0(x=N )j : (1:3)
x2TdN x2TdN
P
In this formula, x;` (y ) stands for (2` + 1) d jz yj` x (z ). It is easy to show
that the second sum vanishes in the limit, as N " 1 and then ` " 1, since the
second coordinate starts from the equilibrium state N0 (x=N ) . More precisely, recall
that the initial state N is bounded above by N . In particular 0 () is bounded
above by . Since x starts from the equilibrium state N0 (x=N ) , the second sum
is equal to
X h i h i
N d EN x=N j` (x) 0 (x=N )j sup

E N j ` (0) j :
x2TdN
0( )

Schwarz inequality shows that the right hand side is bounded above by
n o1=2
` 1=2 sup EN ((0) )2

and thus converges to 0 as ` increases to 1.
On the other hand, the first term of (1.3) is bounded by
X X h i
N d (2` + 1) d E ¯ N
x jt (z ) tx (z )j :
x2TdN jz xj`
Using now the three coordinates coupling we may bound this expression by
X h i
N d E ¯ N
x jt (x) tx (x)j
x2TdN
X X h i (1:4)
+ N d (2` + 1) d E ¯ N
x;z jtx (z ) tz (z )j :
x2TdN jz xj`
By construction of the coupling measure ¯ Nx;z , either 0 0 or 0 0 . Since
x z x z
the order is preserved by the dynamics, we may move the absolute value to outside
the expectation. Since both marginal start from equilibrium, this term is equal to
X X
N d (2` + 1) d j0(x=N ) 0(z=N )j :
x2TdN jz xj`
Since the initial profile 0 is Lipschitz continuous, this expression converges to 0
as N " 1.
It remains to prove that the first term of (1.4) converges to 0. It is equal to
X h i X hZ t i
d x d ds N L̄N js (x) sx(x)j :
N x j0 (x) 0 (x)j +
E ¯ N N E ¯ N
x
x2TdN x2TdN 0
(1:5)
Here L̄N stands for the generator of the basic coupling introduced in section
2.5. Since at each site the measure ¯ N
x is ordered, in the first expression we
may exchange the absolute value and the expectation to obtain that this expected
value is identically 0. On the other hand, since by a straightforward computation

involving the generator of the coupled process, we have that
L̄N j(x) x (x)j

X
jg((x)) g( x(x))j + p( y)jg((x + y)) g( x(x + y))j ;
y
the second expression is bounded above by the time integral of
X X h i
N1 d p( y)E ¯ Nx jg(s (x + y)) g(sx(x + y))j
x2TdN y
X h i (1:6)
N1 d E ¯ N
x jg(s (x)) g(sx(x))j :
x2TdN
Using once more the three coordinates coupled process, the first line is bounded
above by
X X h i
N1 d p( y)E ¯ Nx;x y jg(sx(x + y)) g(sx+y (x + y))j
x2TdN y
+
X h i
+ N1 d E ¯ N
x jg(s (x)) g(sx(x))j :
x2TdN
The second line is compensated by the negative part of (1.6). Therefore, (1.5) is
bounded above by
Z t X X h i
ds N 1 d p( y)E ¯ Nx;x y jg(sx(x + y)) g(sx+y (x + y))j :
x2TdN y
+
0
Since the jump rate g is an increasing function and x and x+y are ordered, as we
did before, we may move the absolute value outside the expectation and obtain,
since both marginals start from equilibrium, that this integral is equal to
X X
tN d p( y)N j(0((x + y)=N )) (0 (x=N ))j :
x2TdN y
Since the transition probability p() is of finite range and 0 is Lipschitz continuous,
this expression is bounded above by tC (0 ) and converges to 0 as t decreases to
0. This proves (iv).
Remark 1.4 It is known that the entropy solution starting from a smooth initial
profile remains smooth in a finite time interval. One may therefore use the relative
entropy method, presented in Chapter 6, to prove statement (iv) for processes
starting from smooth initial profiles and avoid all computations performed above.
2. An entropy inequality at microscopic level
Recall from Chapter 2 that we denote by LN the generator of the basic coupling
of two copies of zero range processes. This section is devoted to the proof of
an entropy inequality at a microscopic level. In order to state this result for a
d d N
probability N on the configuration space N TN N TN = XN2 denote by PN the
probability measure on the path space D([0; 1); XN2 ) corresponding to the Markov
process (t ; t ) evolving according to the generator LN defined above speeded up
N N
by N and starting from N . Expectation with respect to PN is denoted by E N .
Furthermore, for a measure N on the product space XN2 , we denote by N i its
i–th marginal.
Theorem 2.1 Let N be a measure with both marginals bounded by an invariant
product measure N0 :
Ni N0 (2:3)
for i = 1, 2 and some density 0 . Recall the definition of ` (x) given in (5.1.10).
1;1
For every smooth positive function H in CK ((0; 1) Td ) and every positive ",
"Z
N 1 X

lim lim P N
`!1 N !1
dt N d @t H (t; x=N )t` (x) t` (x)
0
x2TdN
#
d
X
+ i (@ui H )(t; x=N
) t` (x) t` (x) " = 1:
i=1
Theorem 1.3 follows straightforwardly from this result :
Proof of Theorem 1.3. Fix a constant c 0. Define the measure N

c on the product
space XN2 by
Nc = N
cN :
Since cN is a translation invariant stationary measure, for every positive t0
Z t0
X
E N dt N d `
x
( t ( )) c
( ) = t0 N ( ` (0))
c c
( ) :
0
x2TdN
This last expression converges to 0 as ` increases to infinity by the law of large
numbers because () is Lipschitz continuous in virtue of Corollary 2.3.6. By the
Rt P
same reasons E N [ 0 0 dt N d x2Td jt̀ (x) cj] vanishes as N " 1 and ` " 1.
N

The proof of Theorem 2.1 is divided in several lemmas. We first prove that
in the limit as N " 1 the configurations and are ordered. Before stating
2. An entropy inequality at microscopic level 209
this result, notice that assumption (2.3) on the initial measure N implies that the
expected value of the mean density of particles is bounded :
2 3
X n o
lim sup E N 4N d 0 (x) + 0 (x) 5 < 1: (2:1)
N !1 x2TdN
Next lemma requires only this weaker assumption on the sequence of initial mea-
sures.
Lemma 2.2 Assume that the sequence of initial measure N satisfies hypothesis
(2.1). Then, for every positive time T and every d–dimensional integer y ,
"Z #
N T X
lim E N dt N d Gx;x+y (t ; t ) = 0;
N !1 0
x2TdN
where, for two sites x and y in Zd, Gx;y (; ) is an indicator function equal to one
if the configurations and are not ordered at sites x and y :
Gx;y (; ) = 1f (x) < (x) ; (y ) > (y )g + 1f (x) > (x) ; (y ) < (y )g :
Proof. The proof is divided in 3 steps.

Step 1. We first derive from the fact that the total number of uncoupled particles
decreases in time that
"Z
N T
lim E N
N !1
dt
0
#

N d X p( y g x y
) ( t ( + ))

g x y Gx;x+y (t ; t )
( t ( + )) = 0:
x;y
Consider the martingale M (t) vanishing at 0 given by
X X
M (t) = N d t( ) x t (x) N d 0( ) x 0 (x)
x2TdN x2TdN
Z t X
ds N d NLN s (x) s (x) :
0
x2TdN
Since the total number of uncoupled particles decreases in time, the time integral
of the last formula has to be negative. In fact, a careful computation relying on
the explicit form of the generator NLN shows that the integrand is equal to
X
N (d 1)
( t ( )) g x g(t(x))
x2TdN
X
N (d 1)
p( y)Gx;x+y(t ; t )g(t (x + y)) g(t (x + y))
x;y
X h i
+ N (d 1)
p( y) 1 Gx;x+y (t ; t ) g(t(x + y)) g(t(x + y))
x;y
X
(d 1)
= 2N p( y)Gx;x+y(t ; t )g(t(x + y)) g(t(x + y)) :
x;y
Since the total number of particles is conserved by the dynamics, by assumption
(2.1) on the initial measure, the expected value of the martingale M (t) is well
defined. Therefore, since M (t) vanishes at time 0, we have that
Z T X
EN
N dt N d p( y)Gx;x+y(t ; t )g(t (x + y)) g(t (x + y))
0 x;y

X
N 1EN N d ( ) x (x) :
x2TdN
A factor N 1 coming from the time renormalization of the dynamics appeared on
the right hand side of the last inequality. This factor and assumption (2.1) permits
to conclude the first step of the proof.
Step 2. We deduce from the previous step that the configurations and are
ordered on sites x and x + y such that p(y ) + p( y ) > 0.
From Step 1 and since by assumption g () is bounded below by a strictly
positive constant (g (k ) g (1) > 0), it follows that
"Z #
T n o
N
lim E N dt N d X p( y) I0x;y (t ; t ) + J0x;y (t ; t ) = 0;
N !1 0 x;y
where, for a positive integer m,
Imx;y (; ) = 1fm = (x + y) < (x + y); (x) > (x)g
Jmx;y (; ) = 1f(x + y) > (x + y) = m; (x) < (x)g :
We now remove the condition t (x + y ) ^ t (x + y ) = 0 in the left hand side of
the last equality. For a positive integer m, let
X
Im (; ) = N d p( y)1fm = (x + y) < (x + y) ; (x) > (x)g :
x;y
We shall prove by induction that
N hZ T i
lim E N
N !1
dt Im (t ; t ) = 0
0
for every positive m. We just proved this equality for m = 0. Consider now for two
fixed d–dimensional integers x and y the mean-zero martingale Mm x;y (t) defined
by
Z t
Mmx;y (t) = Imx;y (t ; t ) Imx;y (0 ; 0 ) NLN Imx;y (s ; s ) ds :
0
If we compute carefully the expression LN Im x;y we obtain two different kinds of

terms. The first ones are negative and correspond to jumps of Im x;y (t ; t ) from
1 to 0 and the second ones are positive and correspond to jumps from 0 to 1.
Notice that while the negative terms are equal to Im x;y LN I x;y , the positive terms
x;y x;y m
are given by (1 Im )LN Im . Among the positive terms there is one which
comes from a jump of two coupled particles from site x + y to another site, when
site x + y is occupied by m + 1 –particles and by at least m + 2 –particles and
site x is occupied by more –particles than –particles. This jump happens at rate
g(m + 1)N . Therefore
(1 Imx;y )LN Imx;y g(m + 1)1fm + 1 = (x + y) < (x + y) ; (x) > (x)g
x;y :
= g (m + 1)Im +1
In particular, since Mm x;y (t) is a mean-zero martingale,

hZ T i
g(m + 1)N E NN dt Imx;y+1 (t ; t )
0
Z T (2:2)
N h x;y i h i
E N Im (T ; T ) N dt E NN Imx;y NLN Imx;y (t ; t ) :
0
A simple computation shows that Imx;y LN Imx;y (; ) is equal to Imx;y (; ) times
h i
g ( m) + g(m + 1) g(m) 1f (x + y) = m + 1g
h i
+ g((x)) g( (x)) 1f(x) = (x) + 1g
X h i
+ p(x z ) g( (z )) g((z )) 1f(z ) < (z ); (x) = (x) + 1g
z
X
+ p(x + y z )g((z ))
z
Therefore, multiplying inequality (2.2) by p( y )N (d+1), summing over all x and
y and keeping in mind that g() is bounded by g(1), we obtain that
Z T h i
g(m + 1) dt E NN Im+1 (t ; t )
0
Z T
N h i h i
N 1
E N Im (T ; T ) + 4g (1) dt E NN Im (t ; t ) :
0
Since Im (; ) is bounded by 1 it follows by induction that

Z T h i
N !1
lim dt E NN Im (t ; t ) = 0 (2:3)
0
for every positive integer m. We prove In the same way this result with the roles
of and interchanged.
We are now ready to prove that
Z T h X i
lim
N !1
dt E NN N d Gx;x+y (t ; t ) = 0
0
x2TdN
for every y such that p(y )+ p( y ) > 0. Fix such integer y and assume without loss
of generality that p(y ) > 0. The time integral of last formula is bounded above by
Z T h X i
p(y) 1
dt E NN N d p(z )Gx;x+z(t ; t )
0 x;z
Z T h X i
= p( y ) 1
dt E NN N d p( z )Gx;x+z(t ; t ) :
0 x;z
m, N dP
For every integer x;z p( z )Gx;x+z(; ) is bounded above by
m n
X o n o
Im (; ) + Im (; )
1
N d X p( z ) ( x + z ) + ( x + z ) :
n=0
+
m x;z
The claim follows therefore from (2.3), from conservation of total number of
particles and from assumption (2.1) on the initial measure.
Step 3 It remains now to remove the assumption p(y ) + p( y ) > 0. Fix z in Zd.
Since p() is irreducible, there exists m 1 and a sequence 0 = x0 ; x1 ; : : : ; xm = z
such that p(xi+1 xi ) + p(xi xi+1 ) > 0 for 0 i m 1. Denote by M (z ) the
length of the smallest path linking the origin to z . We shall prove that
N hZ T X i
lim E N
N !1
dt N d Gx;x+z (t ; t ) = 0 (2:4)
0
x2TdN
by induction in M . In Step 2 we proved the above statement for paths of length
m = 1. Assume that this statement is true for all n m. Fix z in Zd such that
M (z ) = m + 1 and a path 0 = x0 ; x1 ; : : : ; xm+1 = z from the origin to z of length
m + 1. Assume, without loss of generality, that p(xm z ) > 0. If p(xm z ) = 0,
we would consider the path xm z; 0; x1 ; : : : ; xm and replace in the proof below
xm by x0 , z by xm z and x0 by xm .
By the martingale argument presented in steps 1 and 2,
h X i h X i
N N
E N N (d+1) I x;xm (T ; T ) E N N (d+1) I x;xm (0 ; 0 )
x2TdN x2TdN
N h T
Z X i
= E N dt N d LN I x;xm (t ; t ) ;
0
x2TdN
where
I x;y (t ; t ) = 1f (x) < (x); (x + y ) > (x + y )g :
The left hand side of the previous identity vanishes as N " 1 and therefore the
right hand side. We may decompose LN I x;xm in positive and negative terms. The
positive terms are (1 I x;xm )LN I x;xm and the negative terms are I x;xm LN I x;xm .
A simple computation similar to the one performed in the second step shows that
the negative terms are bounded in absolute value by a constant depending only on
g multiplied by I x;xm . In particular, by the induction hypothesis, the expectation
of the negative terms vanishes as N " 1. Therefore,
N hZ T X i
lim E N
N !1
dt N d [1 I x;xm (t ; t )]LN I x;xm (t ; t ) = 0:
0
x2TdN
Since p(xm z ) is positive, the positive expression [1 I x;xm ]LN I x;xm is
bounded below by p(xm z )[g ( (x + z )) g( (x + z ))]1f (x) < (x); (x + xm ) =
(x + xm ); (x + z ) > (x + z )g. This indicator function can be written as
1f (x) < (x); (x + z ) > (x + z )g
1f (x) < (x); (x + xm ) > (x + xm ); (x + z ) > (x + z )g
1f (x) < (x); (x + xm ) < (x + xm ); (x + z ) > (x + z )g :
By the induction assumption, the expectation of the average of the second and
third terms vanishes as N " 1. Therefore, since g (k ) g (1) > 0 for k 1,
N hZ T X i
lim E N
N !1
dt N d 1ft (x) < t (x); t (x + z ) > t (x + z ) = 0g = 0:
0
x2TdN
We now repeat the arguments presented in the previous step to remove the con-
dition (x + z ) = 0. We may of course repeat the proof with the roles of and
interchanged to conclude the proof of the lemma.
This ordering of coordinates made by the process permits to replace averages
of absolute values of differences of monotone functions by absolute values of
averages. This statement is made clear in the next lemma.
Recall from Chapter 2 that a cylinder function is said to be Lipschitz if there
exists a finite subset of Zd and a constant C ( ) such that
X

( ) ( ) C ( ) x
( ) (x)
x2
for all configurations and . Notice that for all Lipschitz cylinder functions
there exists a constant C 0 ( ) and a finite subset of Zd such that
X

C 0 ( ) 1

( ) + ( x)
x2
for all configurations . On the other hand, every monotone cylinder function
is bounded below by (0), if 0 represents the configuration without particle :
0(x) = 0 for every site x.
Lemma 2.3 Let N be a sequence of measures satisfying the assumption stated

in Theorem 2.1. Let be a monotone Lipschitz function. Then, for every positive
integer ` and every positive time T ,
"Z #
N T X
lim E N dt N d x V (t ; t ) = 0;
N !1 0
x2TdN
where V (; ) is the positive function defined by
X
V (; ) = (2` + 1) d y () y ( )
jyj`
X

`
(2 + 1)
d y () y ( ) :
jyj`
Proof. In the case where is a bounded function this result is an immediate

consequence of the previous lemma. Indeed, for a finite subset of Zd, let D;`
be the subset of Zd consisting of all integers at a distance smaller than ` from
: D;` = fy 2 Zd; 9 x 2 ; jx yj `g. Define G;` (; ) as the indicator
function equal to 1 if and are not ordered at D;` :
Y
G;` (; ) = 1 1 Gx;y (; ) :
x;y2D;`
If the configurations and are ordered on the set D;` translated by x x V (; )
vanishes because is monotone. Therefore to prove the lemma for bounded
functions it is enough to show that
"Z #
N T X
lim E N dt N d x G;` (t ; t )
N !1
= 0
0
x2TdN
and this follows from the previous lemma since G;` (; ) is bounded by
X
Gx;y (; ) :
x;y2D;`
We now turn to the general case. Recall that a monotone Lipschitz function is
bounded below (by (0)). The idea is to reduce the general case to the bounded
case by means of a cutoff. Thus, for a real positive A, let A be the cutoff of
at level A :
A () = () ^ A :
The expected value which appears in the statement of the lemma is bounded above
by
"Z #
N T X
E N dt N d x V A (t ; t )
0
x2TdN
"Z #
T X + +
N d
+ 2E N dt N x (t ) A + x (t ) A :
0
x2TdN
For every A the first term converges to 0 as N increases to 1 by the first part of
the proof. On the other hand, since both marginal of the initial

measureN are
bounded by the translation invariant measure N0 and since ( ) A
+
is an
increasing function, the second term is bounded by
" #
X + h i
4E0 N d x () A 4 E ( ) ( ) ^ A :
0
x2TdN
This expected value converges to 0 as A increases to 1 by the dominated con-
vergence theorem since is Lipschitz and therefore 0 –integrable.
The third result towards the proof of Theorem 2.1 is a one block estimate for
the uncoupled process. This result is stated in section 5.4. Notice that all zero
range processes considered in this chapter satisfy assumption (SLG) because g ()
is bounded.
We are now ready to prove the entropy inequality at the microscopic level.
Proof of Theorem 2.1. Fix a smooth positive function H with compact support in
(0; 1) Td . Let MtH be the martingale vanishing at 0 defined by
X
MtH = N d H (t; x=N )t (x) t (x)
x2TdN
Z t X (2:5)
@s + NLN N d
H (s; x=N s (x) x ds :
)

s ( )
0
x2TdN
We will show in Lemma 2.4 below that under assumption (2.1) the expected value
of the square of the martingale converges to 0 :
2
N
lim E N
N !1
MtH = 0:
From Chebychev inequality we obtain that for every t 0 and every positive ",

N h H
i
lim PN
N !1
M t >" = 0: (2:6)
On the other hand, since H has compact support, for sufficiently large t, the
martingale MtH is equal to
Z 1 X
@s + NLN N d H (s; x=N )s (x) s (x) ds :
0
x2TdN
A straightforward computation shows that this integral is equal to
Z 1 X

N d @s H (s; x=N )s (x) s (x)
0
x2TdN
d
X

+ i @ui H (s; x=N g s (x)) g x
) ( ( s ( )) ds
i=1
+ R(H; ; ) + O(1=N ) :
Here R is a positive term and a remainder of order O(1=N ) appeared when we
replaced the discrete partial derivative by the usual one. Therefore, for sufficiently
large t, the martingale MtH is bounded below by
Z 1 X

N d @s H (s; x=N )s (x) s (x)
0
x2TdN
Xd
+ i @ui H (s; x=N )g(s (x)) g x
( s ( )) ds O(1=N ) :
i=1
By assumption (2.1), for every continuous function G: R+ Td ! R with
compact support,
"Z
N 1 X h X i
lim E N
N !1
dt N d G(t; x=N ) (2` + 1) d G(t; y=N )
0
x2TdN jy xj`
#

x x t( ) g x
t ( ) + ( t ( )) g x
( t ( )) = 0 :
Applying this result to the functions @t H and @ui H and making a discrete inte-
gration by parts we obtain that for t sufficiently large, the martingale is bounded
below by
Z 1 X

X
N d @s H (s; x=N )(2` + 1) d
s (y) s (y)
0
x2TdN jy xj`
d
X X
(s; x=N )(2` + 1) d

+ i @ui H g s (y)) g y
( ( s ( )) ds
i=1 jy xj`
oN (1) :
Therefore, from (2.6) and Lemma 2.3, we obtain that for every " > 0,
"Z
N 1 X X
lim PN
N !1
N d @s H (s; x=N )(2` + 1) d s (y) s (y)
0
x2TdN jy xj`
d #
X
+ i @ui H (s; x=N )xV` (s ) ds < " = 0;
i=1
where X
V` () = (2` + 1) d g(s (y)) g(s (y)) :
jyj`
Finally applying the one block estimate to the function g ( (0)) and recalling
that the expectation of this function with respect to the product measure N is
(), we conclude the proof of the theorem.
Lemma 2.4 For a smooth positive function H : (0; 1) Td ! R, let MtH be the
martingale defined by (2.5). Then,
h 2 i
N
lim E N
N !1
MtH = 0:
Proof. From Lemma A1.5.1 N H (t) given by

Z t 2
N H (t) = (M H (t))2 ds NLN AH (s) 2AH (s)NLN AH (s)
0
is a martingale provided
X
AH (t) = N d H (t; x=N )t (x) t (x) :
x2TdN
A straightforward computation shows that the expression inside braces is equal to
X
N 1 2d jg(s (x)) g(s (x))jp(y) Gx;x+y (s ; s )[r+N;y Hs (x=N )]2
x;y
X h i
+ N 1 2d jg(s (x)) g(s (x))jp(y) 1 Gx;x+y [rN;y Hs (x=N )]2 ;
x;y
where rN;y Hs (x=N ) = H (s; (x + y )=N ) H (s; x=N ). The second line is of order
O(N 1 ) and therefore converges to 0 as N increases to 1. The first line is of
d
order O(N 1 d ). Its expected value integrated in time converges to 0 in virtue of
Lemma 2.2.
3. Law of large numbers for the empirical measure
In this section we prove Theorem 0.3. The proof relies on the law of large numbers
for the Young measure proved in section 1.
We first show that if the initial profile 0 () is Lipschitz continuous and the
initial measure is product with marginals given by
N f; (x) = kg = N(x=N ) f; (0) = kg

0
for x 2 TdN ; k 2 N ;
then a law of large numbers for the empirical measure follows from Theorem 1.1.
For each N 1, denote by QN = QN N the probability measure on
D([0; T ]; M+(T )) induced by the empirical measure tN speeded up by N and
d
the sequence of initial measures N .
Lemma 3.1 The sequence QN is tight. Moreover, all limit points are concentrated
on weakly continuous paths (t; du) that are absolutely continuous with respect to
the Lebesgue measure : (t; du) = (t; u)du.
This lemma is proved essentially in the same way as the tightness is proved
in the diffusive case presented in Chapter 5. We leave the details to the reader.
We now claim that it follows from Theorem 1.1 that all limit points of the
sequence QN are concentrated on entropy solutions of the hyperbolic equation
(0.3). Indeed, denote by (t; u) the entropy solution of (0.3), fix a smooth function
H : [0; T ] Td ! R of class C 1;2 ([0; T ] Td) and " > 0. By definition of the
empirical measure,
Z T Z T Z
QN
dt < t ; Ht > dt du H (t; u)(t; u) > " =
0 0
Z T Z T Z
X

PN dt N d H (t; x=N )t(x) dt
du H (t; u)(t; u)du > " :

0
x2TdN 0
By a discrete integration by parts,

X X
N d H (t; x=N )t (x) = N d H (t; x=N )t`(x) + rN;` (t )
x2TdN x2TdN
where the remainder rN;` (t ) has absolute value bounded above by
`2 C (H )N d X (x) :
N2 t
x2TdN
Since the total number of particles is conserved, by Chebychev inequality,
h 2` i
C (H )
2
PN jrN;" j > "=2 "N 2
3. Law of large numbers for the empirical measure 219
if is an upper bound for the initial profile 0 . Therefore,

Z T Z T Z
lim sup QN

dt < t ; Ht > dt du H (u)(t; u > "
)
N !1 0 0
Z T X
lim sup PN

dt N d H (t; x=N )t`(x) (3:1)
N !1 0
x2TdN
Z T Z
dt du H (t; u)(t; u) > "=2 :
0
for every ` in N . Notice that the expressions inside the absolute value in
the last probability may respectively be rewritten as the time integral of <
tN;`; H (t; u) > and the time integral of < ~t ; H (t; u) >, if ~t = (t;u) (d)du
stands for the entropy measure valued solution of equation (0.3). Thus, with the
notation introduced in section 1, last probability may be rewritten as
Z T Z T
QN;`
dt < t ; H (t; u) > dt < ~t ; H (t; u) > > "=2 :

0 0
By Theorem 1.1, this expression converges to 0 as N " 1 and then ` " 1. In

particular, we have from (3.1) that
Z T Z T Z
lim sup QN

dt < t ; Ht > dt du H (u)(t; u) > " = 0;
N !1 0 0
what concludes the proof of the claim.

Finally, since by Lemma 3.1, all limit points are concentrated on weakly con-
tinuous paths, we have that for all 0 t T , all continuous functions H : Td ! R
and all " > 0,
Z
lim sup QN

< t ; H > du H (u)(t; u > " ) = 0:
N !1
To prove Theorem 0.3, it remains to extend this result to profiles in L1 (Td).
Fix 0 2 L1 (Td ) and N a sequence of product measures satisfying assumptions
of Theorem 0.3.
Consider a sequence " : Td ! R+ of Lipschitz continuous functions con-
verging in L1 (Td) to 0 . Denote by " (t; u) the entropy solution of equation (0.3)
with initial data " . It is known (cf. Theorem A2.5.11) that the L1 (Td) norm of
the difference of two entropy solutions of equation (0.3) decreases in time :
Z Z
d
du j" (t; u) (t; u)j du j" (u) 0 (u)j :
T Td
To each profile " associate the product measure ";N with marginals given by
";N f; (x) = kg = N" (x=N ) f; (0) = kg ;

for x in TdN
and k in N .
d d
For each " > 0, define the measure ¯ ";N on N TN N TN with first marginal
equal to N , second marginal equal to ";N and such that

0 (x=N ) " (x=N ) ;
1 if N
¯ ";N f(; ); (x) (x)g =
0 otherwise :
Let (t ; t ) evolve according to the basic coupling defined in section 2.5 and denote
by P̄N;" = P̄¯ ";N the probability on the path space induced by the Markov process
with generator N L̄N and by the initial measure ¯ ";N . Fix > 0 and a continuous
function H : Td ! R. From the definition of the empirical measure, we have
Z
QN
< t ; H > H (u)(t; u)du >

Z
X

= P̄N;" N d H (x=N )t (x)
H (u)(t; u)du > :

x2TdN
R
For " such that Td du j" (u) 0 (u)j < (3kH k1) 1 , the last probability is
bounded above by
X
P̄N;" N d

H (x=N )[t(x) t (x)] > =3
x2TdN
Z
X

+ P̄N;" N d H (x=N )t(x)
H (u)"(t; u)du > =3

x2TdN
because the L1 (Td ) norm of the difference of two entropy solutions decreases in
time. By the first claim of this section and since " (u) is Lipschitz continuous,
the second term converges to 0 as N increases to 1 for each " > 0. On the other
hand, since the total number of uncoupled particles decreases in time, the first
term is bounded above by
X
3kH k1 1
N d E¯ N;" [j(x) (x)j] :
x2TdN
Since at each site the measure ¯ N;" is ordered, we may move the absolute value
outside the expectation and obtain that this last sum is equal to
X
3kH k1 1
N d jEN [(x)] " (x=N )j :
x2TdN
Recall from section 1 that we denoted by N : Td ! R+ the profile associated
N
to the sequence :
0
0N (x=N ) = EN [(x)]

for x in TdN . With this notation the last line is bounded above by
Z Z
3kH k1 1
du jN0 (u) 0 (u)j + du j0 (u) " (u)j
X Z
+ 3kH k1 1
N d " (x=N ) Nd du N d" (u) :
x2TdN N (x)
Here for a site x in TdN , N (x) stands for an hypercube of length N 1

centered
at x :
N (x) = fy 2 Rd ; xi =N yi < xi =N + 1=2N for 1 i dg :

1 =2 N
By assumption, the first integral converges to 0 as N " 1. Since for each " > 0
" is Lipschitz continuous, the third term converges to 0 as N " 1. At last,
by construction of the initial profiles " (u), the second term converges to 0 as "
decreases to 0. In conclusion, we proved that
Z
lim sup QN < t ; H

> H (u)(t; u)du > = 0
N !1
for every continuous function H and every > 0. This concludes the proof of
Theorem 0.3.
The first proof of the hydrodynamic behavior of asymmetric interacting particle

systems was given by Rost (1981). He proved the conservation of local equilibrium
for a one-dimensional nearest neighbor totally asymmetric (particles jumps only
to the right) simple exclusion process starting from the configuration with all
sites at the left of the origin occupied and all others empty. This result was
extended by Andjel and Kipnis (1984) for a one-dimensional nearest neighbor
totally asymmetric zero range process with rate g (k ) = 1fk 1g starting from a
product measure associated to initial profiles of type 1fu < 0g + + 1fu 0g.
They also proved conservation of local equilibrium for general decreasing profiles
in the case where the drift is toward the right (in this case the entropy solution
does not present shocks). Benassi and Fouque (1987) proved conservation of local
equilibrium for the one-dimensional nearest neighbor asymmetric simple exclusion
process starting from a product measure with profile 1fu > 0g and extended it
in Benassi and Fouque (1988) to one-dimensional attractive zero range processes
starting from the same type of initial profile. By the same time and independently,
Andjel and Vares (1987) prove the same result for zero range processes with
initial profiles of type 1fu < 0g + + 1fu 0g and general decreasing profiles
if the drift is to the right. Landim (1991a,b) extended the conservation of local
equilibrium for attractive zero range processes in any dimension with initial profile
given by 1fC g + 1fC c g, where C is a cone. The arguments presented in this
chapter follow Rezakhanlou (1991) who proved a law of large numbers for the
empirical measure and the weak conservation of local equilibrium, as stated in

Chapter 3, for attractive particle systems in any dimension starting from bounded
initial profiles. The spaces L1 ([0; T ]), L1 ([0; T ]; M+(Td R+ )) were introduced
by Bahadoran (1996a).
Extensions.
(a) Infinite volume. The techniques presented in this section permit to prove the
hydrodynamic limit for systems evolving on the infinite lattice Zd. In fact the
original proof of Rezakhanlou (1991) is in infinite volume.
(b) Conservation of local equilibrium. Benassi, Fouque, Saada and Vares (1991)
proved the conservation of local equilibrium for one-dimensional attractive
processes starting from general monotone initial profiles. Fouque and Saada
(1994) considered one-dimensional exclusion processes starting from product
initial measures associated to bumps : 0 = + 1f[u0; u1 ]g. Landim (1993)
deduced the conservation of local equilibrium from the weak local equilib-
rium for attractive systems. Together with Rezakhanlou (1991), this proves
conservation of local equilibrium for initial measures associated to bounded
initial profiles. This argument is explained in details in Chapter 9.
(c) Non product initial states. We already pointed out in Chapter 6 that the rel-
ative entropy method introduced by Yau (1991) permits to prove the hy-
drodynamic limit of asymmetric processes up to the appearance of the first
shock. With coupling arguments, Venkatraman (1994) proved the law of large
numbers for the empirical measure for one-dimensional totally asymmetric
nearest neighbor exclusion processes and zero range processes with jump
rate g (k ) = 1fk 1g starting from deterministic initial configurations as-
sociated to bounded profiles. Independently, Seppäläinen (1996b) proved the
same result for totally asymmetric processes, through the explicit Lax–Oleinik
formula for the entropy solution for one-dimensional hyperbolic equations.
Bahadoran (1997) extended the previous results to one-dimensional nearest
neighbor misanthrope processes.
(d) Spatially inhomogeneous processes. Landim (1996) examined the hydro-
dynamic behavior of a one-dimensional totally asymmetric zero range pro-
cess with bounded jump rates where the rates at a finite number of sites
are slowed down. This result is further discussed below under large devia-
tions. Bahadoran (1996a,b) and Covert and Rezakhanlou (1996) considered
independently the case of asymmetric processes where the jump rate varies
smoothly in the macroscopic scale. Bahadoran (1996a) considered an attrac-
tive zero range process where a particle at site x jumps to site x + y at rate
g(x=N; (x))p(y) for a irreducible transition probability p() and a rate g(; )
such that g (u; ) is an nondecreasing function for every u in R and g (; n) is a
twice continuously differentiable function for every n 1. With further mild
technical assumptions on the jump rate, the hydrodynamic equation is shown
to be @t + ru g^(u; (t; u)) = 0, where g^ is a smooth function depending on
the jump rate and the invariant measures of the process. Bahadoran (1996b)
extended the result to inhomogeneous exclusion processes where a particle at

x jumps to x + y at rate (x=N )p(y)(x)[1 (x + y)] for a irreducible tran-
sition probability p() and a twice continuously differentiable function ().
The same result was obtained independently and at the same time by Covert
and Rezakhanlou (1997) for misanthrope processes. In both cases the hydro-
dynamic equation is of type @t + ru f(u)h^((t; u))g = 0. Moreover, in
these last two models, the invariant measures are not known explicitly and
the method presented in this chapter can not be applied in a straightforward
manner.
(e) Continuous spin systems. Aldous and Diaconis (1995) and Rost (private com-
munication) considered a version of a one-dimensional continuous spin at-
tractive zero range process, called the Hammersley or stick process. The
state space is R+Z and the dynamics may be described as follows. At rate
(x) a random piece U(x) is taken out from site x and transferred to site
x + 1, where U is a random variable with uniform distribution in the interval
[0; 1]. A simple computation shows that the product measure with marginals
distributed according to exponential variables are invariant. Aldous and Dia-
conis (1995) and Rost derived the hydrodynamic behavior of this model for
initial profiles of type 1fu < 0g. Seppäläinen (1996a) extended this result to
processes starting from general initial states that include deterministic ones.
(f) Random rates. Benjamini, Ferrari and Landim (1996) proved the hydro-
dynamic limit of two types of asymmetric zero range processes with random
rates. In the first model jumps are speeded up by random variables : let
fx; x 2 Zdg be i.i.d. random variables taking finitely many positive val-
ues. For a fixed realization, particles at site x jump to site x + y at rate
x g((x))p(y) for some nondecreasing bounded jump rate g. The second
model is one-dimensional. Fix 1=2 < c < 1 and consider i.i.d. random vari-
ables fpx ; x 2 Zg taking finitely many values in the interval [c; 1]. For a
fixed realization, at rate px 1f (x) 1g (resp. [1 px 1 ]1f (x) 1g) a
particle at x jumps to the right (resp. left).
Long time behavior of weakly asymmetric processes. The one dimensional
nearest neighbor weakly asymmetric simple exclusion process is the Markov pro-
d
cess on f0; 1gZ whose generator LN writes as Ls + N 1 La , where Ls (resp. La )
is the generator of the one-dimensional nearest neighbor symmetric (resp. totally
asymmetric) exclusion process. To fix ideas assume that La permits only jumps to
the right. De Masi, Presutti and Scacciatelli (1989) proved that the hydrodynamic
behavior of this process is described by the solution of the Burgers equation with
viscosity :
@t = (1=2) @u f(1 )g (4:1)
and deduced the nonequilibrium fluctuations. Gärtner (1988) presented an alterna-
tive proof of the hydrodynamic behavior of this process and Dittrich and Gärtner
(1991) obtained the nonequilibrium fluctuations of this model. Fritz and Maes
(1988) derived a similar equation with a non linear second order term as equation
of motion for a continuous spin Ginzburg–Landau model in presence of a small

external field. Due to the factor N 1 in front of the asymmetric part, it is easy to
show that N 2 is the correct time renormalization one needs to perform in order to
investigate the hydrodynamic behavior of the process.
Dittrich (1990) studies the evolution of this process in time scales of order
N 3 . He proves that in this scale the behavior is entirely determined by the ini-
tial configuration. From this result Dittrich (1992) shows that the hydrodynamic
behavior of this process starting from an initial configuration is described in
macroscopic times of order N (thus microscopic times of order N 3 ) by the solu-
tion of the discrete version of the Burgers equation with viscosity (4.1) and initial
data . More precisely, he shows that for every continuous function with compact
support H : R ! R and every sequence fuN ; n 1g
1 X
N x2Zd H (x=N uN ) tN (x) N (tN; x=N )
3
converges to 0 in probability as N " 1, provided N is the solution of the discrete

partial differential equation
(
@t N = (1=2)N N rN [N (t; (x + 1)=N )f1 N (t; x=N )g] ;
N (0; x=N ) = (x) :
In this formula, r N and N stand respectively for the discrete space deriva-
tive and the discrete Laplacian : (r N f )(x=N ) = N [f ((x 1)=N ) f (x=N )],
(N f )(x=N ) = N 2 ff ((x + 1)=N ) + f ((x 1)=N ) 2f (x=N )g.
Dittrich (1992) proves also that on the space scale N 2 and time scale N 3 the
process evolves according to the Burgers equation without viscosity. Fix a piece-
wise continuous profile 0 : R ! [0; 1] and denote by N the Bernoulli product
measure associated to 0 on the scale N 2 : E N [ (x)] = 0 (x=N 2 ). Then, for
every continuous function H with compact support,
1 X
N 2 x2Zd H (x=N )tN (x)
2
3
R
converges in probability, as N " 1, to R H (u)(t; u)du, where is now the
entropy solution of the inviscid Burgers equation
@t + @u f(1 )g = 0:
Tracer particles. In contrast with the investigation of the hydrodynamic limit,

where we are interested in the behavior of the whole system, we now tag a single
particle and examine its evolution. To fix ideas, consider a zero range process t
with jump rate g () and transition probability p() satisfying the assumptions of
this chapter. Denote by At the position at time t of the tagged particle and assume
without loss of generality that A0 = 0. Let t stand for the process as seen from
the tagged particle : t = At t . By Harris (1967) and Port and Stone (1973), for
each 0, the Palm measure t , which is product with marginals given by

f; (x) = kg ifx == 0 and k 0 ;
t f; (x) = kg = k 1 ()k
Z (()) g(k)! for x = 0 and k 1
is invariant for the process t .
The mean displacement of the tagged particle is LA = fg ( (A))= (A)g ,
provided L stand for P the generator of the process (At ; t ) and for the mean drift
of the particles : = z zp(z ). Assume the local equilibrium assumption and that
the density around the tagged particle is . In this case the expectation of the mean
displacement is Et [g ( (0))= (0)] = h() , where h() = ()= if > 0 and
h(0) = 0 (0). Denote by aN (t) the macroscopic position of the tagged particle at
the macroscopic time t : aN (t) = N 1 A(tN ). If at the macroscopic time t the
tagged particle is at the macroscopic point u, where, by the hydrodynamic limit,
the density is (t; u), we expect daN (t)=dt to be close to h((t; u)) . Therefore,
in the limit as N " 1, a(t) should be the solution of
8
< da(t) = H (t; a(t))
:
dt (4:2)
a(0) = 0 ;
where H (t; u) = h((t; u)). Since the entropy solution (t; u) might be discontin-
uous the same lack of smoothness is inherited by H (t; u) and, following Reza-
khanlou (1994a), we shall interpret equation (4.2) in the Filippov (1960) sense :
a Lipschitz continuous function a: R+ ! Rd is a solution of (4.2) in the Filippov
sense if for almost all t, da=dt belongs to the interval
h i
ess lim inf H (t; a); ess lim sup H (t; a) :
Rezakhanlou (1994a) proved the existence and the uniqueness of a solution in the
Filippov sense of equation (4.2).
For one-dimensional nearest neighbor simple exclusion processes, the tagged
particle may not jump over the other particles. In particular, the mass at the left of
the tagged particle is constant in time. Therefore, starting from a product measure
with slowly varying parameter associated to an integrable profile, a law of large
numbers for the tagged particle follows from the hydrodynamic limit provided
the tagged particle never hits a hole (in which case its position may not be well
defined). More generally we have
Theorem 4.1 (Rezakhanlou (1994a)) Consider a one-dimensional nearest neigh-

bor zero range process with bounded jump rate g (). Let N be a sequence of prob-
ability measures satisfying assumptions (M1)–(M3) associated to an integrable
profile 0 such that
Z 2
0 (u) du > 0
1
for all 1 0 2 , 2 1 > 0. Add a particle at the origin and denote by aN (t)
its macroscopic position. Then, for each fixed t 0, aN (t) converges in L1 to a(t),
the solution of (4.2).
Rezakhanlou proved this theorem for misanthrope processes, a class that in-
cludes at the same time zero range processes and simple exclusion processes (cf.
Cocozza (1985)). The behavior of a single tagged particle starting from a sequence
of product measures associated to a general profile in higher dimension or in di-
mension 1 with general finite range transition probability p() is still open. There
is a partial answer due to Rezakhanlou (1994a) that shows that in the average, par-
ticles follow equation (4.2). This result, called propagation of chaos, is discussed
below.
The question of a law of large numbers for a tagged particle was already
present in Spitzer (1970). Kipnis (1986) proved a strong law of large numbers in
the case of a nearest neighbor one-dimensional exclusion process starting from
an equilibrium measure conditioned on the presence of a particle at the origin
The asymptotic velocity of the particle was shown to be (1 )(p q ). This
result was extended by Saada (1987a,b) to asymmetric exclusion processes in any
dimension and asymmetric zero range processes with jump rate g (k ) = 1fk 1g.
Seppäläinen (1996b) gives an alternative proof of Theorem 4.1 in the case of
totally asymmetric one-dimensional exclusion processes and zero range processes
with jump rate g (k ) = 1fk 1g. We refer the reader to Kipnis (1985) and Ferrari
(1996) for clear reviews of the subject.
Seppäläinen (1997a) proves a law of large numbers for a tagged particle in the
one-dimensional totally asymmetric simple exclusion process in a scale different
from the Euler scale. Bramson et al. (1986) investigate a one-dimensional nearest
neighbor symmetric exclusion process speeded up by N 2 in which particles create
at rate 1=2 new particles at the neighboring sites. They prove the existence of a
stationary measure for the process as seen from the rightmost particle and compute
the asymptotic velocity, as N " 1, of the rightmost particle. The velocity is related
to the velocity of the traveling wave solution of the reaction–diffusion equation
with a heaviside function as initial data.
Central limit theorem for a tagged particle. Ferrari and Fontes (1996) obtained
a sharp estimate for the position of a tagged particle. Fix > 0 and consider a one-
dimensional nearest neighbor simple exclusion process jumping with probability
p to the right, q = 1 p to the left and starting from the equilibrium measure .
To fix ideas set p > q . Add, if necessary, a tagged particle at the origin or simply
tag the particle already there. Denote by At the position of the tagged particle at
time t. Ferrari and Fontes (1994a) proved the existence of a Poisson point process
of rate (p q )(1 ) and a stationary process St such that Xt = Nt St + S0 and
h i
sup E eSt <1
t
for some > 0. It follows from this result that
lim
XtN (p
p q)(1 )tN =
p
(p q)(1 )Wt
N !1 N
in distribution, where W is a Brownian motion. Moreover, Ferrari (1992) proved
that the fluctuations are due to fluctuations of the initial measure showing that
h 2 i
t!1
lim
1
t E Xt 1 M0 (0 ; (p q)t) = 0 (4:3)
number of holes in the interval f0; : : : ; k g for the con-

if M0 (; k ) stands for theP
figuration : M0 (; k ) = 0xk f1 (x)g.
The central limit theorem for the tagged particle just stated is due to Kipnis
(1986). He obtained a lower bound for the diffusion coefficient. De Masi and
Ferrari (1985) showed that the diffusion coefficient is (p q )(1 ). The fact
that the fluctuations of the tagged particle are due to fluctuations of the initial
configuration was first observed by Gärtner and Presutti (1990).
In nonequilibrium, Wick (1985) proved a central limit theorem for the tagged
particle in a totally asymmetric simple exclusion process starting from a product
measure 0; with density at the right of the origin and 0 at the left and showed
that the diffusion coefficient is (1 ). This result was extended to nearest neigh-
bor one-dimensional exclusion processes by De Masi, Kipnis, Presutti and Saada
(1989). In the same context, Ferrari (1992) proved (4.3), giving an alternative
proof of the central limit theorem in the case where the initial measure is 0;
since M0 (; k ) is a sum of i.i.d. random variables.
Propagation of chaos. To fix ideas, consider a zero range process with jump
rate g () and transition probability p() satisfying the assumptions of this chapter.
Instead of proving that the asymptotic behavior of a single particle is described by
the solutions of equation (4.2), Rezakhanlou (1994a) proved that in the average
particles behave according to (4.2).
We fix our attention on initial configurations with a finite number of par-
ticles and label all particles. The state space is thus E = [L0 (Zd)L , where
L indicates the total number of particles. Configurations are denoted by A =
(A1 ; : : : ; AL ). The microscopic dynamics is clear. Each particle at site y waits a
rate p(z )g ( (A; y ))= (A; y ) exponential random time at the end of which it jumps
to site y +Pz . Here, (A; y ) stands for the total number of particles at site y :
(A; y) = i 1fAi = yg. Denote by Ai (t) the position of the i-th particle at time
t. To write the generator of this Markov process, for each positive integer i and
each site z , denote by Ti;z : E ! E the transformation that moves the i-th particle
by z :
Aj ifj == i ;
(Ti;z A)j =
Aj + z if j = i :
A(t) is a Markov process with generator given by
(Lf )(A) =
X g((A; y)) p(z )1fA = ygf (T A) f (A) :
(A; y) i i;z
y;z2Zd
i1
Fix a bounded integrable profile 0 : Rd ! R+ of bounded variation. Denote

by N a sequence of product measures on E such that
(i) (symmetry) N fA; A = (z1 ; : : : ; zL )g = N fA; A = (z1 ; : : : ; zL )g for ev-
ery L 1 and every permutation ,
(ii) it is associated to the profile 0 in the sense of assumption (M3) : the sequence
N0 : Rd ! R+ that on N (x) is equal to EN [(A; x)] converges to 0 in
L1 (Rd ).
For each v in Rd , denote by av (t) the solution of equation (4.2) with initial
condition v instead of 0. Denote by RN the probability measure on D(R+ ; E )
induced by N and the process aN (t) = (aN N
1 (t); : : : ; aL (t)) = N
1
A(tN ).
Theorem 4.2 (Rezakhanlou (1994a)) For every continuous function J : D( R+ ; Rd )

! R, L Z
h X i

lim
N !1
E R N L 1
J (ak ) J dR = 0;
k=1
where R is the probability measure on D(R+ ; Rd ) concentrated on the solutions of
(4.2) and such that R
(u)du
Rfa; a(0) 2 g = R 0 (u)du
Rd 0
It follows from this theorem that for every n 1 and every family of contin-
uous functions Ji : D(R+ ; Rd ) ! R, 1 i n,
n
hY i n
Y
E
N !1 RN
lim Jk (ak ) = ER [Jk (a)] :
k=1 k=1
This last identity is called propagation of chaos.
Large deviations. This is one of the main open questions in the theory of hydro-
dynamic limits. There are only two results in this direction. Kipnis and Léonard
(1995) proved a large deviation principle for the empirical measure in the non inter-
acting case (particles move according to asymmetric independent random walks).
In this case the probability of observing a large deviation decays as expf CN d+1 g,
instead of expf CN dg as in the diffusive case considered in Chapter 10. In the
interacting case the picture is expected to be completely different. Landim (1996)
proved the hydrodynamic behavior of a one-dimensional totally asymmetric at-
tractive zero range process with bounded jump rate g (), where the jumps from
the origin or the jumps of a tagged particle are slowed down by a fixed factor.
In the first case, the hydrodynamic behavior is given by the entropy solution of
an hyperbolic equation with a boundary condition at the origin. This boundary
condition allows the appearance of a Dirac mass at the origin. In the second case
the hydrodynamic limit is described by a non entropy solution of the hyperbolic
equation. From these results he deduced that in dimension 1, the probability of
large deviations are of order bounded below by expf CN g, in contrast with the
non interacting case. This suggests, as pointed out by Varadhan, that the large
deviations for the asymmetric exclusion process should be of order expf CN g
and given by non entropy solutions of the Burger’s equation.
Fluctuations of the empirical measure. This is another mainly open question,
even in equilibrium. Ferrari and Fontes (1994b) proved the convergence of the
finite dimensional distribution of the fluctuation density field in the case of a one-
dimensional nearest neighbor exclusion process starting from the product measure
; with density at the left of the origin and at the right. In the totally
asymmetric case with = 0, the convergence away from the shock was obtained
by Benassi and Fouque (1991).
9. Conservation of Local Equilibrium for Attractive
Systems
In Chapter 1 we introduced the concept of local equilibrium and proved the con-
servation of local equilibrium for a superposition of independent random walks.
Then, from Chapter 4 to Chapter 8, we proved a weaker version of local equilib-
rium for a large class of interacting particle systems : we showed that the empirical
measure tN converges in probability to an absolutely continuous measure whose
density is the solution of some partial differential equation. The purpose of this
chapter is to to show that in the case of attractive processes, the conservation of
local equilibrium may be deduced from a law of large numbers for local fields,
i.e., from the convergence in probability of the averages
X Z
N d H (x=N )x (t ) to H (u) ~((t; u)) du
x Td
for every t 0, every continuous function H and every bounded cylinder function
. Here (t; u) is the solution of the hydrodynamic equation. This statement is
slightly stronger than the convergence of the empirical measures since it involves
all local fields.
d
To fix ideas, we consider in this chapter a zero range process on N TN with
generator given by
X X
(LN f )( ) = g((x))p(y)[f (x;x+y) f ()] ; (0:1)
x2TdN y2Zd
where p() is a finite range irreducible transition
P probability : there exists R0
such that p(x) = 0 for all x not in R0 , x p(x) = 1 and for every x, y in Zd,
there exists M 1 and x = x0 ; : : : ; xM = y such that ps (xi+1 xi ) > 0 for
0 i M 1, where ps (y ) = p(y ) + p( y ). Denote by StN the semigroup
associated to the generator LN defined in (0.1). Notice that in this chapter StN is
the semigroup associated to the generator LN , which has not been speeded up.
We consider two cases : asymmetric processes where the mean displacement of
each elementary particle does not vanish
X
m = ( m1 ; : : : ; m d ) : = xp(x) 6= 0 (0:2)
x
and mean-zero asymmetric processes, where the mean displacement vanishes. In
the second case, we denote by the covariance matrix defined by
232 9. Conservation of Local Equilibrium for Attractive Systems
X
i;j = xi xj p(x) for 1 i; j d: (0:3)
x2Zd
= fi;j ; 1 i; j dg is a symmetric non–negative definite matrix that we shall
assume to be positive definite to avoid degeneracy of the hydrodynamic equation :
there exist a > 0 such that,
X X
vi i;j vj a vi2
i;j i
for every v in Rd .
For a continuous function 0 : Td ! R+ , denote by N0 () the product measure
with slowly varying parameter associated to 0 , this is the product measure on
d
N TN with marginals given by
N() f; (x) = kg 1 (0 (x=N ))k (0:4)

0
=
Z ((0 (x=N ))) g(k)!
for k 0 so that EN [(x)] = 0 (x=N ) for all x in TdN .
0( )
Theorem 0.1 Assume that m = 0. For every t 0,

lim S N N = (t;u) ;
N !1 tN 2 [uN ] 0 ()
where (t; u) is the unique weak solution of the nonlinear heat equation
8 X
>
< @t = i;j @u2 i ;uj ()
1 i;jd (0:5)
>
:
(0; ) = 0 ()
and is the covariance matrix defined in (0.3).
Theorem 0.2 Assume that the mean displacement m defined in (0.2) does not
vanish and that is strictly concave or convex in the range of 0 (). For every
t 0 and for every continuity point u of (t; ),
lim S N N = (t;u) ;
N !1 tN [uN ] 0 ()
where (t; u) is the unique entropy weak solution of
8
> d
X
>
< @t + mj @uj () = 0
j =1 (0:6)
>
>
:
(0; ) = 0 ()
and m = (m1 ; ; md ) is the mean drift of each elementary particle defined in
(0.2).
1. Replacement lemma for attractive processes 233
1. Replacement lemma for attractive processes
We consider in this section a mean-zero asymmetric zero range process with

generator given by (0.1). We have seen in Chapter 5 that the main step in the
proof of the hydrodynamic behavior of these processes is the replacement lemma
that permits to replace average of cylinder functions by a function of the empirical
measure. This theorem was proved under the assumption that the partition function
is finite on R+ excluding, for instance, the case where the rate function g () is
bounded. We prove in this section a replacement lemma in the attractive case
for sequences of measures bounded above by a product invariant measure. The
approach illustrates how coupling arguments may replace entropy estimates.
d
For a fixed time T > 0 and for a probability measure N on N TN , denote by
d
PN T
N = PN the probability measure on the path space D([0; T ]; N N ) correspond-
ing to the mean-zero asymmetric zero range process with generator LN , defined
in (0.1), speeded up by N 2 and starting from the initial measure N , and by E N
expectation with respect to PN .
Theorem 1.1 Consider a sequence of probability measures fN ; N 1g bounded

by some invariant measure N0 . Assume that the entropy of N with respect to N0
is bounded by K0 N d for some universal constant K0 : H (N j N0 ) K0 N d . For
every cylinder Lipschitz function ,
lim sup lim sup

"!0 N !1
hZ T X X i
E N dt N1d
1
d y (t ) ~(t"N (x)) = 0:
0 x (2"N + 1) jy xj"N
Remark 1.2 In the previous theorem we assumed for simplicity that the sequence
N is bounded above and has finite entropy with respect to the same invariant
measure N0 . The fact that in both assumptions appears the same reference measure
N0 is not restrictive : by Remark 5.1.2 a sequence N which has relative entropy
with respect to some invariant measure N2 bounded by K0 N d , has entropy with
respect to N1 bounded by K1 N d for some K1 = K1 (K0 ; 1 ; 2 ).
Proof. Denote by ftN the Radon–Nikodym derivative of StN N 2 N with respect to

R T
N0 . Set f¯tN = T 1 0 ftN dt. By attractiveness, the probability measures ftN dN0
are bounded by N0 and so is f¯tN dN0 . Moreover, by the entropy computation
performed in section 5.2, the Dirichlet form of f¯TN is bounded by C (K0 ; T )N d 2.
With the notation just introduced, we may write the expectation appearing in
the statement of the theorem as
Z X X
T f¯TN () N1d
1
(2"N + 1)d
y () ~("N (x))N (d) :
0
x jy xj"N
By Corollary 2.3.7, ~ is a uniformly Lipschitz continuous function. In particular,

the arguments presented in the beginning of section 5.3 show that the proof of
the replacement lemma may be reduced to the proof of the one and two blocks
estimates :
Lemma 1.3 (One block estimate) Under the hypotheses of Theorem 1.1,
lim sup lim sup

`!1
Z
N !1
1 X X
Nd x
1
(2` + 1)d
y () ~(` (x))f¯TN () N (d)
0
= 0:
jy xj`
Lemma 1.4 (Two blocks estimate) Under the hypotheses of Theorem 1.1,
lim sup lim sup lim sup

`!1 "!Z0 N !1
X
sup N d ` ( x + y ) N" (x) f¯TN ()N (d) = 0:
jyj"N x
0
We already pointed out in Remark 6.1.14 that in the attractive case either one of
the assumptions (FEM) or (SLG) is fulfilled. The proof of the one block estimate
presented in Chapter 5 applies therefore to the present context. Nevertheless, to
illustrate how coupling arguments may replace entropy estimates, we present below
a proof of the one block estimate in the attractive set–up.
Proof of Lemma 1.3. In the proof of the one block estimate, the assumptions
(FEM) or (SLG) were only used to introduce the indicator 1f ` (x) Ag inside
the expectation to avoid large densities. In the case of attractive processes, this
P by a simple coupling argument : since is
indicator function can be introduced
a Lipschitz function, j(2` + 1) d jyj` y ( ) ~( ` (0))j is bounded above by
C ( )`+s (0), where s stands for the linear size of the support of . In particular,
Z
1 X X
Nd
1
(2` + 1)d
y () ~(` (x))1f`(x) Agf¯TN ()N (d) 0
x jy xj`
Z
1 X
`+s ` ¯N N
C ( ) N d x (x)1f (x) AgfT () (d) : 0
Since `+s (0)1f `(x) > Ag is an increasing function and f¯TN ( ) N (d ) is
bounded by the product measure N , this expression is less than or equal to
0
0
Z Z
` +s ` N C ( )
C ( ) (0)1f (0) Ag (d) A (0)2N (d) :
0 0
that converges to 0 as A " 1.

2. One block estimate without time average 235
Therefore, in order to prove the one block estimate, it suffices to show that
Z
1 X ` ¯N N
lim sup lim sup
`!1 N !1 N x d (x V ;` )( )1f (x) AgfT ( ) 0 (d ) = 0
P
for every A > 0, where V ;` ( ) = j(2` + 1) d jyj` y ( ) ~( ` (0))j. This
follows from the proof of the one block estimate presented after formula (5.4.1)
and the estimate on the Dirichlet form of f¯TN obtained in the beginning of the
proof of Theorem 1.1.
Proof of Lemma 1.4. In the same way, in the proof of the two blocks estimate
given in section 5.3, the assumption (FEM) was invoked only at formula (5.5.1)
to justify the introduction of the indicator function 1f ` (x) _ ` (x + y ) Ag. The
arguments presented in the proof of Lemma 1.3 above show that in the attractive
case such indicator function may also be introduced.
Since the proof of the hydrodynamic behavior of mean-zero asymmetric zero
range processes presented in Chapter 5 relies almost exclusively on the replacement
lemma and we just proved this result in the context of attractive systems, we have
Theorem 1.5 Consider a mean-zero attractive zero range process with generator
given by (0.1) speeded up by N 2 and starting from a measure N satisfying the
following three assumptions : N is bounded above by an invariant measure N0 ,
it has relative entropy with respect to N0 bounded by K0 N d and
h Z i
E
N d X H (x=N ) (x) H (u)0(u) du
lim
N !1 N = 0
x
for every continuous function H and some continuous initial profile 0 : Td ! R+ .
Then, for every t > 0, the empirical measure tN defined in (4.0.2) converges in
probability to the absolutely continuous measure (t; u)du whose density (t; u) is
the solution of the nonlinear heat equation (0.5).
2. One block estimate without time average
The first step toward the proof that for attractive processes the conservation of
local equilibrium follows from a law of large numbers for the empirical measure
is a one block estimate without time average. The proof of this estimate relies on
the fact that under some assumptions on the sequence of initial measures N all
limit points of the sequence
X
N d x StN(N )N
x2TdN
are invariant. Here (N ) stands for the hydrodynamic time renormalization :
(N ) = N in the asymmetric case and (N ) = N 2 is the mean-zero case. The
condition on the initial measure is simple to state and relies on the behavior
of the microscopic density field : we shall assume that for all uniformly Lips-
chitz continuous function F : R+ ! R (there exists a constant C (F ) such that
jF (u) F (v)j C (F )ju vj for all u, v in R+ )
" #
Z T Z T
(MF) lim lim ds ESsN N N N d X F ( ` (x)) = A(F; s) ds :
`!1 N !1 0
( )
x 0
for some continuous function A(F; ).

Before stating the main theorem of this section, we shall discuss assumption
(MF). In the asymmetric case, consider a sequence N = N0 () of product measures
with slowly varying parameter associated to a continuous profile 0 : Td ! R+ . It
follows from the law of large numbers for the Young measure proved in Chapter
8 that the left hand side of (MF) is equal to
Z T Z
ds du F ((s; u)) ;
0 Td
where (s; u) is the entropy solution of the hyperbolic equation (0.6). Since by
Theorem A2.5.11 the entropy solution is L1 (Td ) continuous in the sense that (t; )
converges to (s; ) in RL1 (Td ) as t approaches s, for every uniformly Lipschitz
continuous function F , du F ((s; u)) is time continuous, what proves assumption
(MF) for attractive asymmetric zero range processes.
We consider now the case of mean-zero attractive asymmetric zero range pro-
cesses. Fix a sequence of measures N satisfying the assumptions of Theorem
1.5. It follows from the two blocks estimate proved in section 1 that assumption
(MF) is equivalent to the statement that for every uniformly Lipschitz continuous
function F
"Z #
T X Z T
lim lim E N ds N d F (s"N (x)) = A(F; s) ds
"!0 N !1 0 x 0
for some continuous function A(F; ). By the hydrodynamic behavior of attractive
mean-zero asymmetric zero range processes stated in Theorem 1.5 above, the left
hand side is equal to
Z T Z
ds du F ((s; u)) :
0 Td
Assumption (MF) is therefore satisfied for attractive mean-zero asymmetric zero
range processes because by Theorem A2.4.3 the solution (t; u) is uniformly
Hölder continuous on each compact set of (0; 1) Td .
Theorem 2.1 Let N be a sequence of probability measures bounded by N0 and

satisfying assumption (MF). For every bounded cylinder function and for every
t > 0,
lim sup lim sup

`!1 N !1 2 3
X X
EStN N N 4N d
1
y () ~( ` ( ))5 = 0
x :
d
x2Td (2` + 1) jy xj`
( )
Proof. Fix t > 0. Let N (t) be the spatial average of StN(N )N :
X
N (t) =
Nd
1
x StN(N )N :
x2TdN
With this notation, we may rewrite last expectation as
h X i
EN (t)

1
(2` + 1)d
x ( ) ~(`(0)) :
jxj`
By attractiveness, N (t) is bounded above by N0 , for N0 is invariant and
translation invariant. This bound on N (t) in turn permits to show that the sequence
fN (t); N 1g is weakly relatively compact (cf. Lemma 2.3.9 for a similar
statement). Denote by A the set of limit points of this sequence. Since is a
bounded cylinder function, in order to prove the theorem, it is enough to show
that h i
X
lim sup sup E
1
x ~( ` (0)) = 0 :

(2` + 1)d
( )
`!1 2A jxj`
Due to the presence of the space average, all measures in A are translation
invariant. Moreover, all elements of A are bounded above by 0 because by
d
Lemma 2.3.9 the set of probability measures f 2 M1 (N Z ); 0 g is weakly
relatively compact.
Let us assume for the moment that all limit points of the sequence fN (t); N
1g are invariant : A I . This is the content of Proposition 2.2 below. A
is therefore contained in the set of invariant and translation invariant measures
bounded by 0 . By Theorem 2.6.2, the convex hull of the compact convex set
f 2 I \S ; 0 g is equal to the set of product translation invariant measures
, 0 0 . In particular,

X
sup E

1
(2` + 1)d
x () ~(` (0))
2A jxj`

X
sup E

1
(2` + 1) d x () ~(`(0))
2I\S jxj`
0
X
= sup E

1
(2` + 1) d x () ~(` (0)) :
0 jxj`
Since is a bounded cylinder function,

X
lim sup sup E
1
(2` + 1) d x E [ ] = 0:
`!1 0 jxj`
On the other hand, for a fixed density 1 > 0 ,
h i h i
~( ` (0)) k k1E 1f`(0) 1 g
~( )
E 2
h i
+ sup j ~0 (u)j E j ` (0) j
u1
and the right hand side converges to 0 as ` " 1 uniformly in [0; 0].
An elementary coupling argument permits to extend Theorem 2.1 to cylinder
Lipschitz functions .
We now complete the proof of the previous theorem showing that all limit
points of the sequence fN (t); N 1g are invariant measures.
Proposition 2.2 Under the assumptions of Theorem 2.1, A I .
Proof. The proof of this result is divided in several lemmas. Consider a limit
point and assume without loss of generality that the sequence N (t) converges
to . By Theorem 2.6.3, the translation invariant measure is invariant if for
each density 0 there exists a translation invariant measure ¯ on the product
d d
space N Z N Z with first marginal equal to , second marginal equal to and
concentrated on ordered configurations :
(i) the first marginal of ¯ is ,
(ii) the second marginal of ¯ is and
(iii) ¯ f(; ); or g = 1.
Fix 0. To obtain a measure ¯ satisfying assumptions (i)–(iii), we shall
consider a coupled process (t ; t ) evolving according to the generator L̄N defined
= . d
in (2.5.1) starting from ¯ N N N
d
For a measure ¯ on the product space N Z N Z and j = 1, 2, let j ¯ stand for
the j -th marginal of ¯ . Denote by S̄tN the semigroup associated to the generator
L̄N . Since each marginal evolves according to the original dynamics and N is an
invariant measure, at each time the first marginal of the coupled process is StN N
and the second is N :
1 S̄tN ¯ N
StN N ; 2 S̄tN ¯ N = N :
=
Fix 1 > _ 0 . Denote by ¯ N N N

(t) the space average of the measure ¯ S̄t(N ) :
X
¯ N (t) =
1
Nd x S̄tN(N )¯ N
x2TdN
and by ¯N1 the measure on the product space concentrated on the diagonal and with
both marginals equal to N1 : ¯N1 f(; ); = g = 1, j ¯N1 = N1 for j = 1, 2. Since
the measures N and N are bounded above by N1 , since the translation invariant
measure ¯N1 is invariant with respect to the coupled process and since the coupled
dynamics is itself attractive, a simple argument shows that S̄t N ¯ N (and therefore
(N )
¯ (t)) is bounded above by ¯1 . In particular, the sequence f¯ N (t); N 1g is
N N
weakly relatively compact. Denote by ¯ a limit point and assume, without loss
of generality, that the sequence ¯ N (t) converges to ¯ . It is clear that ¯ satisfies
requirements (i) and (ii) of Theorem 2.6.3. The third property follows from the
next four lemmas and and is closely related to the proof of the entropy inequality
at the microscopic level for the Young measures presented in section 8.2 (cf. proof
of Lemma 8.2.2).
Lemma 2.3 Fix t0 > 0. For every x such that p(x) + p( x) > 0,
n o
¯ (t0) (; ); 0 = (0) < (0); (x) < (x) = 0 ;
¯ (t0 )f(; ); 0 = (0) < (0); (x) < (x)g = 0 :
Proof. Denote by dN (t) the density of uncoupled particles at time t:

" #
X h i
d N ( t) = ES̄tN N ¯ N N d (z ) z
( ) = E¯ N (t) (0) (0) :
z
( )
For the dynamics defined by the generator L̄N , the number of uncoupled particles
may only decrease in time. Indeed, a simple computation taking advantage of the
translation invariance of the measures ¯ N
(t) shows that the time derivative of dN
is given by
hX i
d0N (t) = 2(N )E¯ N
(t) p( x)G0;x(; )jg((x)) g( (x))j ;
x
where, for two distinct sites x, y , Gx;y (; ) is the cylinder functions equal to 1
if the configurations and are not ordered at sites x, y and 0 otherwise :
Gx;y (; ) = 1f (x) < (x); (y ) > (y )g + 1f (x) > (x); (y ) < (y )g :
Denote by (N )fN (t) the time derivative of dN (t). In Lemma 2.6 below,
we prove that the sequence of continuous monotone functions fdN (); N 1g
converges uniformly on each compact set of R+ . It is in the proof of this assertion
that assumption (MF) is required.
From the definition of the sequence fN , we have that
Z t+(N )
dN t + (N )
1
dN (t) = (N ) fN (s) ds

t
for every > 0. By Taylor expansion,
Z t+(N ) 1
2 f 0 (s (t));
(N ) fN (s) ds = fN (t) +
2 (N ) N
N
t
for some sN (t) in the interval [t; t + (N ) 1
]. Therefore,
fN (t) dN (t + (N ) 1
) dN (t) f 0 (s (t)) :
=
2 (N ) N
N
A simple but rather long computation shows that the derivative of fN is of
order (N ). Since by Lemma 2.6 below the sequence dN converges uniformly on
each compact set, for every t1 > 0,
lim sup f (t) = 0 : (2:1)

N !1 tt1 N
Recall the definition of fN (t). Since by assumption the sequence ¯ N
(t0 ) converges
weakly to ¯ (t0 ), it follows from (2.1) that
hX i
E¯ (t ) 0 p( z )G0;z (; )g((z )) g( (z )) = 0:
z
Since the jump rate is nondecreasing and vanishes at 0, the left hand side is
bounded below by
X h
g(1) p( z )E¯ (t ) 1f(0) < (0); 0 = (z ) < (z )g
0
z i
+ 1f (0) < (0); 0 = (z ) < (z )g :
This proves the lemma because the measure ¯ (t0 ) is translation invariant.
Lemma 2.4 Fix t0 > 0. For all x in Zd such that p(x) + p( x) > 0,
¯ (t0 )f(; ); (0) < (0); (x) < (x)g = 0 ;
¯ (t0 )f(; ); (0) < (0); (x) < (x)g = 0 :
Proof. We prove the first identity and leave the second one to the reader. Fix x in
Zd so that p(x) + p( x) > 0. For a positive integer m, denote by Im the indicator
function of the set
f(; ); m = (0) < (0); (x) < (x)g

m (t) its expectation with respect to the measure ¯ N (t) :
and by fN
fNm (t) = E¯ N (t) [Im ] :
We shall prove that for all m 0, the sequence ffNm; N 1g
converges to
0 uniformly over all compact set of R+ . In the previous lemma we proved this
statement for m = 0 because
fN0 (t) 2p(x)g(1)fN (t) :

We shall proceed by induction. Fix m > 0 and assume that fN n converges uniformly
to 0 on all compact subsets of R+ for all 0 n m. The time derivative of fN m (t),
m
denoted by (N )hN (t), is given by
(fNm )0 (t) = (N )E N [L̄N Im ] = : (N )hm (t) :

¯ (t) N
Computing L̄N Im we obtain positive terms that correspond to jumps of
Im (t ; t ) from 0 to 1 and negative terms that correspond to jumps of Im (t ; t )
from 1 to 0. Since Im takes only the values 0 and 1, the positive terms, whose ex-
pectation with respect to ¯ N m;+
(t) is denoted by hN (t), are given by (1 Im )L̄N Im :
hNm;+ (t) = E¯ N (t) [(1 Im )L̄N Im ] :
In the same way, the negative terms of L̄N Im are given by Im L̄N Im and their
expectation with respect to ¯ N m;
(t) is denoted by hN (t) :
hm;
N (t) = E¯ N (t) [Im L̄N Im ] :
Among the positive terms there is one which corresponds to a jump of two
coupled particles from site 0 to some site x when site 0 is occupied by m + 1
-particles and more that m + 2 -particles and the site x is occupied by more
-particles than -particles. Since the rate of such jump is g(m + 1) this term is
equal to
p(x)g(m + 1)1f(; ); m + 1 = (0) < (0); (x) < (x)g

= p(x)g (m + 1)Im+1 :
In particular,
hm; m+1
N (t) p(x)g (m + 1)E¯ N (t) [Im+1 ] = p(x)g (m + 1)fN (t)
+
because all other terms of (1 Im )L̄N Im are positives. Therefore, to show that
the sequence ffN m+1; N 1g converges to 0 uniformly over all compact sets of
R+ , it is enough to prove the same result for hm; m;+
N (). By definition of hN and
+
m;
hN we have that
Z t+(N ) 1
fNm (t + (N ) 1) fNm(t) = (N ) m;+ (r) + hm; (r)] dr

[hN :
N
t
By Taylor expansion,
Z t+(N ) 1
2 (hm;+)0 (s (t))
(N ) hNm;+ (r) dr = hm;
N (t)
+
2 (N ) N
+ N
t
for some sN (t) in [t; t + (N ) 1 ]. In particular, since hN m;+ is positive,
m fNm (t)
hNm;+(t) = fN (t + (N) (hm; + 0
1
N ) (sN (t))
)
2 (N )
0
(2:2)
(N ) Z t+(N ) hm; (r) dr :
1
t N
By the induction assumption the sequence (fN m )N 1 converges to 0 uniformly over
all compact sets of R+ . On the other hand, an elementary computation gives that
the time derivative of hm;
M is of order (N ). Therefore the first two terms on the
+
right hand side of the equality converge uniformly to 0 as N " 1 and # 0. It

remains to show that the third term converges to 0. It is not difficult to bound the
absolute value of hm;
N (t) by
X h i
(g (m) + 2g )E¯ N
(t) [Im ] + g p( z )E¯ N (t) f(z ) + (z + x)gIm ;
z
where g = supn0 fg (n + 1) g (n)g. In this last computation we used the trivial
bound g (n) g n. Since both marginals of ¯ N N
(t) are bounded above by and
E [(0) ] is finite, by Schwarz inequality, there exists a constant C0 (1 ; g ; m)
1
2
1
such that q
hm;
N (t) C0 E¯ N (t) [Im ] = C0 fNm (t)1=2 :
In particular, by the induction assumption, the third term of the right hand side of
(2.2) converges to 0 uniformly over compact sets of R+ .
To conclude the proof of the lemma, it remains to notice that
¯ (t0 )f(; ); (0) < (0); (x) < (x)g
n
X
¯ (t0 )f(; ); m = (0) < (0); (x) < (x)g + ¯ (t0 )f (0) > ng
m=0
n

X
lim fNm (t0) + 1
m=0 N !1 n
because by assumption ¯ N (t0 ) converges to ¯ (t0 ) and both marginals are bounded
by N1 . We just proved that the first term on the right hand side vanishes for all
m. The second vanishes as n " 1.
Lemma 2.5 For every t0 > 0,
¯ (t0)f(; ); or g = 1:
Proof. Since ¯ (t0 ) is translation invariant, we just need to prove that
¯ (t0 )f(; ); (0) < (0); (x) < (x)g = 0

for all x in Zd and a similar identity with and interchanged.
Fix x in Zd and denote by axN (t) the function
h i
axN (t) = E¯ N (t) 1f(0) < (0); (x) < (x)g :
We shall prove that for each x in Zd axN converges to 0 uniformly on each compact
interval of R+ , as N " 1 :
lim ax = 0 locally uniformly : (2:3)

N !1 N
Fix z in Zd. Since the transition probability is irreducible, there exists a positive
integer m and a path 0 = x0 ; x1 ; : : : ; xm = z such that p(xi xi+1 ) p(xi+1 xi ) >
0. Denote by M (z ) the length of the smallest path. We shall prove (2.3) by
induction on M (z ). In Lemma 2.4 above we proved (2.3) for all z such that
M (z ) = 1. Indeed, with the notation introduced in the previous lemma,
n
X 1
axN (t) fNm(t) +
n
m=0
for every n 1 and fN converges to 0 uniformly on each compact subset of R+
m
for each m 0.
Fix m 1 and assume that (2.3) has been proved for each z such that M (z )
m. Fix z in Zd such that M (z ) = m + 1 and a path 0 = x0 ; x1 ; : : : ; xm+1 = z . To
fix ideas, assume that p(xm z ) > 0. If p(xm z ) = 0, in the proof below, we
just need to consider the path xm z; 0; x1 ; : : : ; xm and let xm z , 0 and xm
play the roles of z , xm and 0, respectively.
Denote by I x the indicator function of the set f (0) < (0); (x) < (x)g.
We have that Z t
axm (t) = (N ) fb+ (s) + b (s)g ds ;
N N N
0
where b+N = (1 I xm )L̄N I xm , bN = I xm L̄N I xm . Notice that while b+N is positive,

bN is negative.
By Taylor expansion, for each t 0, > 0, we have
b+N (t) axNm (t + (N ) 1) axNm (t)

0 =

Z t+(N )
bN (s) ds 2(N ) (b+N )0 (sN (t)) ;
1
(N )
t
where sN (t) belongs to the interval [t; t + (N ) 1 ]. By the induction assumption,
the first term on the right hand side converges to 0 uniformly on each compact
subset of R+ . A simple computation shows that (b+N )0 is of order (N ) so that
p by . Finally, as in the proof of the previous lemma,
the third term is bounded
bN (t) is bounded by C axNm (t) for some constant C = C (1 ; g). In particular,
the second term is bounded by . This proves that b+N converges to 0 uniformly
on each compact subset of R+ .
Since p(xm z ) > 0 among the positive terms of b+N (t) there is one that
corresponds to a jump of a -particle from z to xm when (0) < (0), (xm ) =
(xm ) and (z ) > (z ). This jump happens with rate p(xm z )[g((z )) g( (z))].
Therefore, since g (k ) g (1) > 0,
h i
0 E¯ N (t) 1f(0) < (0); (xm ) = (xm ); (z ) > (z ) = 0g
(2:4)
fp(xm z )g(1)g 1b+N (t) :
The indicator function of the previous equation can be rewritten as
1f (0) < (0); (z ) > (z ) = 0g

1f (0) < (0); (xm ) > (xm ); (z ) > (z ) = 0g
1f (0) < (0); (xm ) < (xm ); (z ) > (z ) = 0g :
Since by the induction assumption E¯ N (t) [1f(0) < (0); (xm ) > (xm )g] and
E¯ N (t) [1f(xm) < (xm ); (z ) > (z )g]
(t) [1f (0) < (0); (z xm ) > (z xm )g]
= E¯ N
converge to 0 uniformly on each compact subset of R+ , the same holds for

E¯ N (t) [1f(0) < (0); (z ) > (z ) = 0g] in virtue of (2.4).
It remains now to repeat the proof of Lemma 2.4 to remove the restriction
(z ) = 0. This concludes the proof of the lemma.
We turn now to the proof that the sequence dN converges uniformly to 0 on
each compact set of R+ .
Lemma 2.6 The sequence dN converges uniformly on each compact set of R+ .
Proof. Fix an interval [0; u]. To prove that a sequence of continuous functions
converges uniformly on the compact set [0; u], we may proceed as follows : prove
first that the limit is unique and then show that all sequences admit a converging
subsequence.
To show that the limit is unique, consider a subsequence dNj that converges
uniformly on [0; u] to d1 . To keep notation simple assume that the sequence dN
converges. The sequence is bounded since
h X i
dN (t) = ES̄tN N ¯ N N d (x) (x)
x
( )
h i h i
EStN N N N d X (x) + X
EStN N N N d (x) 0 + :
x x
( ) ( )
By the dominated convergence theorem, for every 0 t u,

Z t Z t
lim
N !1
dN (s) ds = d1 (s) ds :
0 0
On the other hand, recalling the definition of dN (t),

Z t Z t h X i
dN (s) ds = ds ES̄sN N ¯ N N d (x) (x)
x
( )
0 0
Z t h i
= ds ESsN N N N d X ` (x)
x
( )
0
Z t X ih
+ ds ES̄sN N ¯ N N d x V` (; ) ;
x
( )
0
where,
X `
V` (; ) =
1
(2` + 1) d
(y ) (y )
(0) :
jyj`
By assumption (MF), the first term on the right hand side of the last identity
converges to Z t
A(F ; s) ds
0
where, for a real positive , the function F is equal to F (u) = ju j. The
second term is bounded above by
Z t h X X ` i
dsES̄sN N ¯ N N d (2` +1 1)d ( ) y (y ) x
( ) ` (x)
x jy xj`
( )
0
Z t X h i
+ ds ESsN N N N d s` (x) :
x
( )
0
By Lemma 8.2.3, the first term vanishes as N " 1 and then ` " 1 (this result was
proved for asymmetric systems but the arguments apply to mean-zero processes
speeded up by N 2 ). On the other hand, since the measure N is invariant, the
second expectation is equal to
h
t EN ` (0)
i t
E N
h
(0)
2 i1=2
(2` + 1)d=2
that vanishes as ` " 1. In conclusion, we proved that the time integral of any
converging subsequence is equal to the time integral of A(F ; ). In particular,
the unique possible limit point of the sequence dN is the continuous function
A(F ; ).
To prove that every sequence admits a converging subsequence, fix a subse-
quence Nj , that we denote by N to keep notation simple and consider a countable
and dense subset D of [0; u]. Since the functions dN are uniformly bounded, by
Cantor diagonal procedure, we may obtain a subsequence Nj so that dNj con-
verges at each point of D. Denote by d1 the function defined on D thus obtained.
It is a nonincreasing function since, for each N , dN is a nonincreasing func-
tion. We may therefore extend the definition of d1 to [0; u] taking right limits :
d~1 (r) = supfd1 (s); s > r; s 2 Dg. Notice that on points r of D, d~1 (r) d1 (r)
and there might be strict inequality. The function thus obtained, denoted from now
on by d1 to keep notation simple, is right continuous, nonincreasing and has there-

fore at most a countable number of discontinuities. Moreover, it is not difficult to
see that the subsequence dNj converges pointwisely to d1 at all continuity points
of d1 . In particular, dNj converges almost surely to d1 . Since the sequence is
uniformly bounded, by the dominated convergence theorem,
Z t Z t
lim
j !1
dNj (s) ds = d1 (s) ds
0 0
for all t u. We have seen in the first part of the proof that
Z t Z t
lim
j !1
dNj (s) ds = A(F ; s) ds :
0 0
Therefore, d1 is almost surely equal to the continuous function A(F ; ). Since
d1 is nonincreasing, the two functions are equal. We have thus a sequence of
nonincreasing continuous functions dNj that converges pointwisely to a continuous
function A(F ; ). This forces the sequence to converge uniformly on [0; u].
3. Conservation of local equilibrium
Recall from Appendix 2 the terminology of weak solutions of the nonlinear partial
differential equation X
@t = i;j @u2 i ;uj () (3:1)
1i;jd
and Theorem A2.4.3 and Theorem A2.4.5 therein that shall be invoked throughout
this section.
We start with a weak version of local equilibrium for attractive systems. This
question was already investigated in Chapter 6 in the case of mean-zero processes
and in Chapter 8 in the case of asymmetric processes. The next result follows
from Corollary 6.1.3 and some elementary coupling arguments.
Theorem 3.1 Fix a continuous profile 0 : Td ! R+ and recall that we denote by

N0 () the product measure with slowly varying parameter associated to 0 . Consider
an attractive mean-zero asymmetric zero range process with generator given by
(0.1). For every t > 0, every bounded cylinder function and every continuous
function H : Td ! R,
h i Z
lim E N N N d X H (x=N )(x )( ) = du H (u) ~((t; u)) ;
N !1 StN 2 0 () x
where (t; u) is the unique weak solution of (3.1) with initial data 0 .
Proof. This result was proved in Corollary 6.1.3 for smooth enough strictly
positive profiles 0 . The strategy is thus to approximate 0 by a sequence "0
3. Conservation of local equilibrium 247
of such smooth profiles and couple two copies of the process starting from an
initial measure with first marginal equal to N0 () and second marginal equal to
N"0 (). To use then coupling arguments to show that EStN N 2 N () [x ] is close to
EStN N 2 N" [x ] ~(" (t; x=N )) for " small and the continuous dependence on
0
0 ()
the initial data of solutions of the equation (3.1) to prove that ~(" (t; x=N )) is
close to ~((t; x=N )).
Fix a continuous function 0 : Td ! R+ and consider a sequence "0 : Td ! R+
of strictly positive profiles of class C 1 (Td ) converging to 0 uniformly on Td
and bounded below by 0 : 0 0" . For each " > 0, denote by " (t; u) (resp.
(t; u)) the unique weak solution of equation (3.1) with initial data 0" (resp. 0 ).
By Theorem A2.4.5, for each t 0, " (t; ) converges uniformly to (t; ).
Assume for a while that
h i
E N N N d X H (x=N )(x )( )
N !1 StN
lim
x
2 0( )
h i (3:2)
E N N N d X H (x=N )(x )( ) :
"!0 N !1 StN "
= lim lim
x
2 ( )
0
In this case, by Corollary 6.1.3 and Remark 6.1.14, the right hand side is equal to
Z Z
lim
"!0
du H (u) ~(" (t; u)) = du H (u) ~((t; u))
because " (t; ) converges uniformly to (t; ) and ~() is a continuous bounded
function. Therefore, to prove the theorem we just need to justify identity (3.2)
For each " > 0, denote by ¯ N " the probability measure on the product space
d d
N N
Z Z whose first marginal is equal to N0 () , second marginal is equal to
"0 () and concentrated on configurations (; ) such that . This is possible
N
because N0 () and N" () are product measures and 0 "0 .
Recall that we denote by S̄tN the semigroup measure associated to coupling
0
generator L̄N defined in (2.5.1). With this notation we may write

h X i
N N" N d
EStN H (x=N )(x )( )

xh
2 ( )
0
X i
N N N d
EStN H (x=N )(x )()
x
2 0( )
h i
ES̄tN d X H (x=N )x f ( ) ( )g :
= " N
N ¯ N
x
2
Since is a bounded cylinder

P function, there exists a constant C = C ( ) such
that j ( ) ( )j C jxjs j (x) (x)j for any two configurations , .
In particular, the absolute value of the right hand side of the previous equality is
bounded above by
h i
C ( ; H )ES̄tN d X j (x) (x)j :
" N
N ¯ N
x
2
Since at time 0 the configurations are ordered (0 0 ) and since the dynamics
preserves the order, t t for all t 0. In particular, in the previous formula,
j (x) (x)j = (x) (x). Since, on the other hand, the total number of particles
is conserved, last expectation is equal to
n h X i h X io
C ( ; H ) EN" N d (x) EN N d (x)
x 0( )
x
( )
0
that converges to
nZ Z o
C ( ; H ) du 0" (u) du 0 (u)
as N " 1. This expression vanish as " # 0 because "0 converges to 0 uniformly
on Td , what proves (3.2).
In the asymmetric context the same statement follows from the law of large
numbers for the Young measure proved in Chapter 8.
Theorem 3.2 Fix a continuous profile 0 : Td ! R+ and recall that we denote

by N0 () the product measures with slowly varying parameter associated to 0 .
Consider an attractive asymmetric zero range process with generator given by
(0.1). For every t > 0, every bounded cylinder function and every continuous
function H : Td ! R,
h i Z
lim E N N N d X H (x=N )(x )( ) = du H (u) ~((t; u)) ;
N !1 StN 0 () x
where (t; u) is the unique entropy solution of the differential equation (0.6).
Proof. Due to Theorem 2.1, we just need to show that

h i Z
lim lim E N N N d X H (x=N ) ~( ` (0)) = du H (u) ~((t; u))
`!1 N !1 StN 0 () x
t
(3:3)
for every t > 0, every bounded cylinder function and every continuous function
H . Recall the definition of the Young measure tN;`(dq; du) introduced in Chapter
8. With this notation, we may rewrite last expectation as
hZ Z i
EStN
N N H (u) ~(q)N;`(dq; du)
0( ) Td R+
R
that converges, by Theorem 8.1.1, to du H (u) ~((t; u)) as N " 1 and then
` " 1 provided 0 is Lipschitz continuous. A simple coupling argument, very
similar to the one presented in the proof of Theorem 3.1, permits to extend (3.3)
to continuous profiles.
3. Conservation of local equilibrium 249
We have now all tools to prove the conservation of local equilibrium.
Proof of Theorem 0.1. Fix a continuous function 0 : Td ! R+ and denote by

(t; u) the solution of (3.1) with initial condition .
: Td ! R+ by 0
For each " > 0, define ";0
";0 + (u) = sup 0 (v) ; ";0 (u) = inf

v2B (u;")
0 (v) ;
v2B (u;")
where B (u; ") is a ball centered at u of radius " for the max norm : B (u; ") =
fv; max1id jvi ui j "g. Notice that for each " > 0 0"; is a continuous
function and that
0"; (v1 ) 0 (u) ";+ (v2 ) (3:4)
for each v1 , v2 in B (u; ").
Denote by "; (t; u) the solution of (3.1) with initial condition ";
0 . By The-
orem A2.4.5,
lim "; (t; u) = (t; u) (3:5)
"!0
for every t 0 and u in T . By property (3.4), for all u1 , u2 in B (0; "),
d
[u1 N ]N";0 () N0 () [u2 N ]N";0 + () :

Therefore,
N [v N ] N"; N N N N
StN 2
() StN [uN ] () StN [v N ] "; ()
1
0
2
0
(3:6) 2 2
0
+
for each v1 , v2 in B (u; ") because the order is preserved by attractive processes.
Fix t > 0, u in Td , " > 0 and consider a bounded monotone cylinder function
. We claim that
lim sup ES N uN N [ ] E t;u [ ] = ~((t; u)) : (3:7)
N !1 tN ( )
2 [ ]
0( )
Indeed, for 0 < < ", consider a sequence of continuous approximations H;k :
Rd ! R+ of the function H = d 1fB (0; =2)g with support contained in
[ ; ]d :
H = d 1fB (0; =2)g H;k d 1fB (0; )g

and lim H;k = H
k!1
pointwisely. By inequality (3.6) and by definition of H;k we have that
h X i
N uN N [ ] ES N N";
EStN 2 [ ]
tN
N d H;k (x=N )[uN ]+x ()
x
+
0( ) 2 ( )
0
because < ". By Theorem 3.1, as N " 1, the right hand side converges to
Z
dv H;k (v) ~(";+(t; u + v)) :
Since ~ is a bounded function and the sequence H;k converges pointwisely to

the function H , as k " 1, this expression converges to
Z
dv H (v) ~(";+ (t; u + v)) : (3:8)
Since H is an approximation of the identity, as decreases to 0, the integral

converges to
~(";+(t; u))
because the solution of the parabolic equation (3.1) is continuous and ~() is a
continuous bounded function. Finally, by (3.5), letting " # 0 we obtain (3.7).
The same argument with an approximation from below of the function
d 1fB (0; =2)g and with "; replacing ";+ shows that
lim inf ES N 2 [uN ] N () [ ]
N !1 tN 0
~((t; u)) :
This concludes the proof of the theorem since, by Lemma 2.1.5, to prove that a
sequence of probability measures N converges weakly to a probability measure
, it is enough to show that the expected value of all bounded monotone cylinder
functions converges.
The proof of the conservation of local equilibrium for asymmetric processes is
exactly the same and relies on the following result on the continuous dependence
on initial data of entropy solutions of equation (0.6). Recall from the proof of
Theorem 0.1 the definition of the profiles 0"; . For each " > 0, denote by "; (t; u)
the entropy solution of equation (0.6) with initial condition 0"; .
Lemma 3.3 Fix t > 0 and recall the notation introduced in the proof of Theorem
0.1. For every continuity point u of (t; ),
lim sup
"!0 jv uj"
j"; (t; v) (t; u)j = 0:
Proof. The proof of this lemma relies on a result of P. Lax (1957), stated here as
Theorem A2.5.10, that asserts that the entropy solution depends continuously on
the initial data for-one dimensional equations.
A change of coordinates permits to rewrite equation (0.6) as
@t + @u ()
1 = 0:
It is therefore enough to prove the lemma for this equation, which is a one-
dimensional equation and for which the variables u2 ; : : : ; ud may be interpreted
as parameters.
Fix t > 0 and a continuity point u of (t; ). In order to prove the lemma,
it is enough to show that for every sequence fa" ; " > 0g such that ja" j ",
"; (t; u+a" ) converges to (t; u). Fix such sequence fa" ; " > 0g. By definition of
";0 , ";0 ( + a" ) converges uniformly to 0 () because 0 is uniformly continuous.

Since u is a continuity point of (t; ), u1 is a continuity point of (t; ; u2 ; : : : ; ud ).
Therefore, by Theorem A2.5.10, (t; u + a" ) converges to (t; u) as " # 0, what
concludes the proof of the lemma.
Proof of Theorem 0.2. In the proof of Theorem 0.1, we did not used any property
of the solutions of equation (0.5) until formula (3.8). In the hyperbolic setup, we
may bound the integral (3.8) by
~((t; u)) + sup j ~0 (a)j sup j";+ (t; v) (t; u)j
ak0 k1 jv uj"
because ". In the previous lemma we showed that the second term vanishes
as " # 0. This concludes the proof of the conservation of local equilibrium for
asymmetric processes.
The proof of the one block estimate without time average is due to Rezakhan-
lou (1991). It relies partially on Proposition 2.2 whose proof is very similar to
the one of Proposition 5.1 of Andjel (1982) and Theorem VIII.3.9 (a) of Liggett
(1985). The idea to deduce the conservation of local equilibrium from a law of
large numbers for the local fields is taken from Landim (1991b), Landim (1993).
We review here some aspects of the macroscopic motion of asymmetric pro-
cesses.
Microscopic structure of the shock. We have seen in this chapter that in the
asymmetric case the empirical measure converges in probability to the entropy
solution of a first order hyperbolic equation. These equations have the peculiarity
that the solutions develop shocks even when the initial data is smooth. An interest-
ing question at the physical level concerns the microscopic structure of the shock.
It consists in determining whether there exists a finer scale than the hydrodynamic
scale where the density profile becomes smooth or whether such an intermediary
scale does not exists and the shock is sharp.
To examine this question consider a nearest neighbor asymmetric simple ex-
clusion process on Z with probability p to jump to the right and probability
q = 1 p < p to jump to the left. Fix < and let ; be the inhomoge-
neous Bernoulli measure with density at the left of the origin and at the
right. In this case the entropy solution of the hydrodynamic equation is a traveling
wave : (t; u) = 1fu < vtg + 1fu > vtg, where the velocity v of the shock is
equal to (p q )(1 ).
Add a particle at the origin if this site is empty, tag this particle and let it
evolve as a second class particle. This means that at rate p (resp. q ) the particle
will attempt to jump to the right (resp. to the left). If the site chosen is empty
the particle jumps, otherwise nothing happens. In addition, if another particle tries
to jump to the site occupied by the tagged particle they exchange position. In
particular, if the site on the left (resp. right) of the tagged particle is occupied,
at rate p (resp. q ) the tagged particle jumps to the left (resp. right). The tagged
particle is called a second class particle because the other particles have priority
over it and jump to a site even when it is occupied by the tagged particle.
Denote by Zt the position of the tagged particle at time t. Ferrari (1992)
proved that the process as seen from the tagged particle (i.e. Zt t ) converges
weakly to an invariant measure ; with asymptotic distribution and :
limx!1 x ; = , limx! 1 x ; = . He proved moreover that for each
fixed time t the distribution of Zt t has the same asymptotics. This proves the
sharpness of the shock since there is a random position from which microscopically
to the left (resp. right) we see the invariant measure with density (resp. ).
The question of the microscopic structure of the shock was examined by several
authors. Ferrari (1986) investigated the invariant measures of exclusion processes
as seen from a first class tagged particle. Wick (1985) considered the structure of
the shock for the totally asymmetric case (p = 1) with no particles at the left of the
origin ( = 0). De Masi, Kipnis, Presutti and Saada (1989) extended the result to
the case p < 1. In both previous situations the position of the shock is determined
by the leftmost particle and the invariant measure for the process as seen from this
particle is explicitly known and its asymptotics converge exponentially fast to .
In the case > 0, however, there is no such natural choice. Nevertheless, Ferrari,
Kipnis and Saada (1991) proved in this case the existence of a random position
Xt from which to the left (resp. right) we see the invariant measure with density
(resp. ). Ferrari (1992) showed that the position of a second class particle has
this property.
In higher dimension nothing is known. Alexander et al. (1992) present simu-
lations of the shock in a two-dimensional asymmetric exclusion process.
Closely related to the problem of the microscopic structure of the shock is the
question of the position of a second class particle in asymmetric processes.
Asymptotics of a second class particle. Consider the nearest neighbor asym-
metric simple exclusion process on Z with probability p to jump to the right and
q = 1 p < p to jump to the left. Fix < and recall from the previous subsec-
tion the definition of the product measure ; and the definition of the evolution
of a second class particle. Denote by Zt the position of a second class particle
initially at the origin. Ferrari and Fontes (1994b) proved that
lim
ZtNp vtN = Wt ;
N !1 N
weakly in the sense of finite dimensional distributions, provided Wt stands for a
Brownian motion with diffusion coefficient
D = (p q) (1 ) + (1 ) (4:1)
and v stands for the velocity of the shock which is equal to (p q)(1 ).
Wick (1985) proved this central limit theorem in the context of totally asym-
metric simple exclusion processes with = 0. De Masi, Kipnis, Presutti and Saada
(1989) extended this result to the case 1=2 < p < 1. Gärtner and Presutti (1990)
proved in the case p = 1, = 0 that the fluctuations of Zt arise from the fluc-
tuations of the initial state. Ferrari (1992) extended this result to the case p < 1.
Ferrari and Fontes (1994b) proved the result that we just described and extended
Gärtner and Presutti result to the case 0 < < , 1=2 < p < 1. Ferrari and Fontes
(1996) proved a central limit theorem for the current over macroscopic regions for
one-dimensional nearest neighbor asymmetric simple exclusion processes starting
from an invariant state .
From the central limit theorem for the second class particle and the microscopic
structure of the shock we may deduce the behavior of the state of the process at the
discontinuity points of the solution of the hydrodynamic equation. This question
is further discussed in the next subsection.
Rezakhanlou (1995) considered the behavior of a second class particle in one-
dimensional asymmetric misanthrope processes starting from product initial mea-
sures. To fix ideas we shall present the results in the context of exclusion processes
with jump rate p(). Fix a profile 0 : R+ ! [0; 1] in L1 (R) and denote by N a
product measure associated to 0 in the sense that
Z
lim
N !1 jujA
du EN [([uN ])] 0 (u) = 0
for every A > 0. Fix a0 in R and denote by ZtN the position at time t of a second
class particle initially at [a0 N ].
Let (t; u) stand for the uniqueP entropy solution of the hyperbolic equation
@t + @u([1 ]) = 0, where = x xp(x). Recall from section 8.4 the definition
of a Filippov solution of the equation
8
< da(t) = f1 2(t; a(t))g ;
:
dt
a(0) = a0 :
Rezakhanlou (1995) proved that there exists at most one solution in the Filippov
sense of the previous equation provided
Z a0 + " Z a0
lim inf
"#0
1
" a 0 (u) du
" a " 0 (u) du :
lim sup
1
0 "#0 0
N =N converges in probability to a(t) provided

Furthermore, he proved that ZtN
Z a0 + Z a0
[1 0 (u)] du 0 (u) du 6= 0
a0 a0
for every positive .
All the results cited in this subsection show that the second class particles in
asymmetric processes either follow the characteristics or the shocks of the hydro-
dynamic equation. Ferrari and Kipnis (1995) consider the case of a rarefaction
fan, where more than one characteristic starts from the origin. They considered
the nearest neighbor asymmetric simple exclusion process with probability p to
jump to the right and q = 1 p < p to jump to the left starting from a product
measure ; with density at the left of the origin and < at the right. They
proved that the second class particle chooses with uniform distribution one of the
characteristics and then sticks to it. More precisely, denote by Zt the position at
time t of a second class particle sitting at the origin at time 0. They proved that, as
N " 1, N 1 ZtN converges to Ut in distribution, where Ut is a random variable
with uniform distribution over the interval [(1 2)t; (1 2 )t]. Furthermore,
they showed that for 0 < s < t, ZtN =tN ZsN =sN converges to 0 in probability
as N " 1.
Behavior at discontinuity points of the profile or dynamical phase transition.
We proved in this section that the sequence N St(N ) [uN ] converges to (t;u)
at all continuity points u of (t; ). Nothing is said at the discontinuity points. For
diffusive systems this remark is irrelevant since the solutions are Hölder contin-
uous by Nash’s theorem. However, for asymmetric processes, where the entropy
solution may develop shocks even if the initial profile is smooth, the behavior at
the shocks must be examined by different means.
Since the article of Liggett (1975) it is conjectured that the sequence N St(N )
[uN ] converges at a shock to a mixture of extremal invariant measures. The
first result in this direction was proved by Wick (1985) who considered the one-
dimensional totally asymmetric zero range process on Z moving to the right with
rate g (k ) = 1fk 1g and starting from a product measure 0; with density 0 at
the left of the origin and density at the right. In this case, if 0 (u) = 1fu > 0g
stands for the initial profile, the entropy solution (t; u) is a traveling wave equal
to 0 (u vt), where v = v ( ) = (1 + ) 1 is the speed of the shock. Wick (1985)
proved that for any sequence TN that increases to 1 as N " 1,
Z tN +TN
N !1 TN tN
lim
1
drN Sr [v()tN +upN ] = f1 m(t; u)g + m(t; u)0 ;
where m is the solution of the parabolic equation
(
@t m = (1=2)D( )m
m(0; u) = 1fu < 0g
and D( ) = (1 + ) 1 . Wick called this phenomena a dynamical phase transition.
It is closely related to the microscopic structure of the shock and to the asymptotic
behavior of a second class particle.
This result was successively improved for one-dimensional nearest neighbor
asymmetric exclusion processes by Andjel (1986), Andjel, Bramson and Liggett
(1988), De Masi, Kipnis, Presutti and Saada (1989) and Ferrari and Fontes (1994b).
Bramson (1988) presents a short review of some of the previous results. We now
describe Ferrari and Fontes (1994b). Consider an asymmetric simple exclusion
process that attempts to jump with rate p to the right and rate q to the left. For ,
> 0, denote by ; the product measure on N Z that has density at the left
of the origin and at the right and by 0 the associates profile : 0 = 1fu <
0g + 1fu 0g. The entropy solution (t; u) is a traveling wave given by (t; u) =
1fu < vtg+ 1fu > vtg, where the speed of the shock v = v(; ) is equal to (p
q)(1 ). Recall the definition of the diffusion coefficient D given in (4.1) and
denote by w(t; u) the probability of a Brownian motion with R diffusion coefficient
D to be less than u at time t : w(t; u) = (2tD) 1=2 u1 da expf a2=2tDg.
Ferrari and Fontes (1994b) proved that
lim S p = (1 w(t; a)) w(t; a) :

N !1 ; tN [vtN +a N ]
+
By the time this note was written, Ferrari, Fontes and Vares (private communi-
cation) obtained the behavior at the shock for an increasing piecewise constant
initial profile and with a finite number of discontinuities.
Stationary measures of asymmetric systems. Derrida, Domany and Mukamel
(1992) consider a one-dimensional simple exclusion process on f0; : : : ; N g, where
particles jump only to the right. Particles are create with intensity > 0 at 0 and
are destroyed at N with intensity > 0. The generator of this process is therefore
NX1
(LN f )( ) = (x)[1 (x + 1)][f (x;x+1) f ()] + (L f )() + (L+ f )( ) ;
x=0
where L , L+ are the boundary generators given by
(L+ f )( ) = (N 1)[f ( dN ) f ( )] ;
(L f )( ) = [1 (0)][f ( + d0 ) f ( )] :
Derrida, Domany and Mukamel (1992) present a general method to derive an
explicit formula for the stationary measure and implement it in the case = = 1.
The computations were extended by Schütz and Domany (1993) to the general
case ; > 0. Derrida, Evans and Mallick (1995) compute the fluctuations of the
current for this model.
The method introduced by Derrida, Domany and Mukamel (1992) was ex-
tended by Derrida et al. (1993) for asymmetric exclusion processes with first and
second class particles. Derrida et al. (1993) compute the stationary measure of a
totally asymmetric simple exclusion process evolving on TN with first and second–
class particles. They deduce from this result the profile of the first class particles as
seen from a second class particle in the stationary regime. This profile determines
the microscopic shape of the shock linking two different densities. Speer (1994)
extend this result to the infinite volume case. Ferrari, Fontes and Kohayakawa
(1994) propose an alternative method to describe the invariant measure of the
exclusion process in infinite volume with first and second–class particles.
Foster and Godrèche (1994) and Evans et al. (1995) examine the stationary
measures of exclusion processes with two types of species. Janowsky and Lebowitz
(1994) investigate the stationary measures of an inhomogeneous totally asymmetric
exclusion process. Here at all sites but one particles jump to the right at rate one.
In the remaining site, particles jump with rate 0 < r < 1. Schütz (1993) obtains
the stationary measure of a deterministic model with a stochastic defect.
10. Large Deviations from the Hydrodynamic Limit
In Chapters 4 and 5 we proved a law of large numbers for the empirical density
of reversible interacting particle systems. A natural development of the theory is
to investigate the large deviations from the hydrodynamic limit.
To avoid technical problems related to the lack of regularity of the rate function,
we concentrate on symmetric simple exclusion processes. Moreover, for historical
reasons, we decided to consider the process starting from an equilibrium product
state : N , for some density 0 < < 1. In fact, the same approach applies to a
process starting from any sequence N of deterministic configurations associated
to a profile : Td ! [0; 1] (cf. Remark 1.2).
In the case where the process starts from N , two distinct types of large devia-
tions of the same order arise. The first one corresponds to large deviations from the
initial state. It is very simple since it reduces to large deviations of i.i.d. random
variables. The second one comes from the stochastic character of the evolution.
Since we are mainly interested in the latter, we ignore in this introduction the
static large deviations.
We claim that in order to prove an upper bound large deviations, we just need
to find a family of mean-one positive martingales that can be expressed as function
of the empirical measure. Indeed, denote by expfCN J (N )g such a martingale
indexed by in A and fix a compact set K on the path space D([0; T ]; M+). We
have
h i
QN [N 2 K] = EQN e CN J (N )eCN J (N ) 1fN 2 Kg
n o h N i
exp CN inf J ( ) EQN eCN J ( ) 1f N 2 Kg
n
2K o
exp CN inf J ( ) :
2K
The last inequality follows from the fact that expfCN J ( )g is a mean-one pos-
itive martingale. Therefore, minimizing over in A, we have that
log QN [ N 2 K] sup inf J ( ) : (0:1)

1
C
lim sup
N !1 N 2A 2K
To conclude the proof of the upper bound, it remains to justify the exchange
between the supremum and the infimum. This is done through a minimax theorem
relying on some regularity of J . The upper bound rate function obtained in this
258 10. Large Deviations from the Hydrodynamic Limit
way is equal to sup 2A J ( ). Of course the upper bound may be bad if we
considered to few positive martingales or not the relevant ones.
This argument shows that we have to build positive martingales. Following
Donsker and Varadhan (1975a,b), (1976), in the context of Markov processes, the
relevant positive martingales are obtained as small Markovian perturbations of the
original process.
To clarify this general philosophy we return to the case of symmetric simple
exclusion processes. For each H in C 1;2 ([0; T ] Td ) consider the time inhomo-
geneous Markov process with generator at time t given by
(LH
N;t f )( ) =
X
(1=2)N 2 (x)[1 (y)]eH (t;y=N ) H (t;x=N ) [f (x;y ) f ()] :
jx yj=1
This is a small perturbation of the original process in the following sense. At time
t instead of jumping from x to x ei with rate 1=2, a particle jumps with rate
(1=2)f1 N 1 (@ui H )(t; x=N )g. We introduced therefore a small (of order N 1 )
space and time dependent asymmetry in the jump rate.
For each H in C 1;2 ([0; T ] Td ), denote by PH N the probability measure on
D([0; T ]; f0; 1g ) corresponding to the inhomogeneous Markov process t with
Td
N
generator LH N
N . WhenN H =N0, we denote PH simply by P .
N
Denote by (dPH =dP )(t) the Radon–Nikodym derivative of PH N with re-
spect to P restricted to the -algebra generated by fs ; 0 s tg. Of
N
course, (dPH N =dP N )(t) is a mean-one positive martingale. The explicit formula
for the Radon–Nikodym derivative of a Markov process with respect to another
one (cf. Proposition A1.7.3) and a simple computation (cf. section 2) shows that
(dPHN =dP N )(t) is equal to
Z t
exp N d < tN ; Ht > < N ; H0 >
0 < sN ; @s Hs + (1=2)Hs > ds
0
Xd X Z t
(1=2) N d (@ui H (s; x=N ))2 s (x)(1 s (x + ei )) ds
i=1 x2TdN 0

+ OH (N 1
)
where OH (N 1 ) is a constant bounded in absolute value by C (H )N 1 for some

finite constant C (H ).
Unfortunately, (dPHN =dP N )(t) is not a function of the empirical measure due
to the second integral term. This is the main difficulty in the proof of a large
deviations principle : we have to show that the integral term can be rewritten as
a function of the empirical measure.
In view of the replacement lemma proved in Chapter 5, the idea is clear.
Denote by F () the polynomial (1 ). We would like to show that
10. Large Deviations from the Hydrodynamic Limit 259
X Z s n o
N d G(s; x=N )) s (x)(1 s (x + ei )) F (s"N (x)) ds
x2TdN 0
is small as N " 1 and than " # 0. However, this time, since we are interested in
large deviations events with probability of order expf CN d g, we need to show
that this difference is superexponentially small, i.e., that for any > 0
" Z #
T
N
1:
N d log P VN;" (s; s ) ds >
1
lim sup lim sup =
"!0 N !1 0
where
X n o
VN;" (t; ) = N d G(t; x=N ) (x)(1 (x + ei )) F ("N (x)) :
x2TdN
This is the content of section 3. In section 4 we prove the upper bound fol-
lowing the strategy presented above. In order to prove that we may exchange the
infimum with the maximum on the right hand side of (0.1), we rely on Lemma
A2.3.3 that allows such replacement provided K is compact and each J is lower
semicontinuous. In possession of the upper bound for compact sets, the passage
to closed sets is standard and presented with all details at the end of section 4.
The strategy of the proof of the lower bound is also easy to understand. We
start proving a law of large numbers for the empirical measure evolving according
to the perturbations considered in the proof of the upper bound. More precisely,
denote by QN H the probability on the path space D([0; T ]; M+) corresponding to
the inhomogeneous Markov process tN with generator LH N . We show that for each
H in C 1;2 ([0; T ] Td), QNH converges weakly to the measure QH concentrated
on an absolutely continuous deterministic path H (t; du) whose density is the
solution of a differential equation involving H (cf. Proposition 5.1).
Denote by I the large deviations rate function obtained in the proof of the
upper bound. The second step consists in proving that the entropy of PH N with
N d
respect to P divided by N converges to I ( ) : H

N d H P N P N I ( H ) : (0:2)
lim
N !1 H =
At this point it is not difficult to obtain a lower bound large deviations. Consider
an open set O. Fix H in C 1;2 ([0; T ]Td) and recall that we denote by H (t; du) the
hydrodynamic limit of the empirical measure evolving according to the generator
LHN . For each H such that H (t; du) belongs to O, we have
N
N d log QN [O] N d log EPHN dP
=
dPHN 1f 2 Og :
Since O contains H (t; du), under PHN the probability of the event f 2 Og is
close to 1. We may therefore remove the restriction f 2 Og in the last expectation.
Moreover, by Jensen inequality, the right hand side is bounded below by
N
EPHN N d log dP

dP N = N dH PHN P N :
H
By (0.2) the right hand side converges to I ( H ). We have thus proved that
lim inf N d log QN [O] inf I ( H ) :

N !1 H 2C 1;2 ([0;T ]Td); H 2O
To conclude the proof of the lower bound we are left to show that
inf
H 2C 1;2 ([0;T ]Td); H 2O
I ( H ) = inf I ( )
2O
or, equivalently, that each path with finite rate function (I ( ) < 1) can be ap-
proximated by a sequence Hn , Hn in C 1;2 ([0; T ] Td), such that limn!1 I ( Hn )
= I ( ).
The arguments presented above for the proof of the upper bound explain in part
why it is sometimes easier to prove a large deviation principle at a higher level. It
may happen that the natural positive martingales to consider in the problem cannot
be expressed as functions of the process under investigation but only as functions
of higher level processes.
For instance, one might be interested in studying equilibrium
Rt
occupation time
large deviations, i.e., large deviations for the functional 0 ds s (0). Unfortunately,
Rt
no natural martingale is function of 0 ds s (0) only. There is therefore no direct
way to prove an upper bound large deviations for this functional. If instead we
investigate higher level processes like s or sN , a large deviations principle
can be proved through the method presented above. To Rt
get from these results a
large deviations principle for the original functional 0 ds s (0), we just apply
a contraction principle (cf. Benois (1994) for one-dimensional superposition of
independent random walks and Landim (1992) for symmetric simple exclusion
processes in dimension d = = 2.)
1. The rate function
Fix once for all a density in (0; 1). We state in this section the large deviations
principle for the empirical measure. We start introducing all apparatus required to
define the rate function.
Recall from Chapter 4 that we denote by M+ = M+ (Td) the space of all
positive measures on Td with finite total mass. Denote by ! the elements of M+
and by M+;1 = M+;1 (Td ) the closed subset of M+ of all positive measures on
Td with total mass bounded by 1. Hereafter, absolutely continuous measures with
respect to the Lebesgue measure are called absolutely continuous measures. In the
case where ! 2 M+ or 2 D([0; T ]; M+) are absolutely continuous, we denote
respectively by and their density : ! (du) = (u)du, (t; du) = (t; u)du.
1. The rate function 261
We mentioned in the introduction that there are two distinct types of large
deviations arising in this problem. Static large deviations from the initial product
measure and dynamics large deviations due to the stochastic character of the
evolution. This translates in the decomposition of the rate function in two pieces.
We define now the first one which is associated to the static large deviations. This
one is very simple because it reduces to large deviations of independent Bernoulli
random variables.
For each continuous function : Td ! (0; 1), define h : M+ ! R and
h : M+ ! R̄+ by
h (!) = < !; log (1(1 )) > + < ; log 11 > ;
(1:1)
h(!) = sup h (!)

In this formula and below stands for the Lebesgue measure on Td and < !; f >
for the integral of f with respect to ! . Since for each , h () is linear, h is convex
and lower semicontinuous. h is the piece of the rate function associated to large
deviations from the initial measure.
Recall that in simple exclusion processes there is at most one particle per
site. In particular, in the hydrodynamic limit, as the scale parameter N " 1, all
trajectories are absolutely continuous with density bounded by 1. This space plays
therefore a particular role and deserves a special notation.
Denote by Mo+;1 the subset of M+;1 of all absolutely continuous measures
with density bounded by 1 :
n o
M+o;1 = ! 2 M+;1; !(du) = (u)du and 0 (u) 1 a.e. :
M+o;1 is a closed subset of M+ endowed with the weak topology. This property is
inherited by D([0; T ]; Mo+;1) : D([0; T ]; Mo+;1) is a closed subset of D([0; T ]; M+)
for the Skorohod topology.
In section 5 we obtain an explicit formula for h and we show that h is infinite
outside Mo+;1 . We now turn to the piece of the rate function associated to the
dynamics.
For each smooth function H in C 1;2 ([0; T ] Td ), define the functionals JH ,
`H : D([0; T ]; Mo+;1) ! R by
Z T Z
JH () = `H () (1=2) dt d du k(rH )(t; u)k2F ((t; u)) ;
0 T
`H () = < T ; HT > < 0 ; H0 > (1:2)
Z T
dt < t ; @t Ht + (1=2)Ht > :
0
In this formula rH stands for the gradient of H : rH =P(@u H; : : : ; @ud H ),

krH k for the Euclidean norm of the gradient : krH k2 = 1id (@ui H )2 and
1
F : [0; 1] ! R+ for the polynomial F () = (1 ).

Since F is concave and `H linear, JH is convex for each H in C 1;2 ([0; T ]Td).
Moreover, we claim that JH is lower semicontinuous. Since `H is linear, we just
need to show that the integral part is upper semicontinuous.
Let n be a sequence in D([0; T ]; Mo+;1) converging to some . In particular,
n (t; ) converges to (t; ) for almost all 0 t T . For each such t, since
belongs to D([0; T ]; Mo+;1), n (t; [a; b]) converges to (t; [a; b]) for all closed
hypercubes [a; b]. Recall from (5.1.8) that " stands for an approximation of the
identity. Denote by " the convolution of the density with " and keep in
mind that ( " )(t; u) = (2") d (t; [u "; u + "]). Since " converges, as " # 0,
to in L1 ([0; T ] Td ),
Z T Z
dt du k(rH )(t; u)k2F ((t; u))
0 Td
Z T Z
= lim
"!0
dt du k(rH )(t; u)k2F (( " )(t; u)) :
0 Td
Since n (t; [u "; u + "]) converges to (t; [u "; u + "]) for almost all (t; u) and
since F is concave, the right hand side is equal to
Z T Z Z
lim lim dt du k(rH )(t; u)k2F (2") d n (t; v) dv
"!0 n!1 0 Td u ";u+"]
[ d
Z T Z Z
"lim lim
!0 n!1 dt du k(rH )(t; u)k2(2") d dvF (n (t; v)) :
0 Td [ u ";u+"]d
Now, since n is uniformly bounded by 1 and since H is smooth, integrating
by parts the space variable it is easy to see that we may interchange limits. This
shows that the last expression is equal to
Z T Z
n!1
lim dt du k(rH )(t; u)k2F (n(t; u)) :
0 Td
Therefore the integral term is upper semicontinuous.
Since in the proof of the hydrodynamic behavior we considered the empirical
measure N as evolving on D([0; T ]; M+), we need to extend the definition of JH
to this larger space. The natural way is to set JH to be 1 outside D([0; T ]; Mo+;1) :
JH () = 1 if 62 D([0; T ]; M+o;1) : (1:3)
Notice that since D([0; T ]; Mo+;1) is closed, JH is still a convex lower semicon-
tinuous function on D([0; T ]; M+).
We are now ready to define the large deviations rate function. Let I , I0 :
D([0; T ]; M+) ! R̄+ be defined as
I0 ( ) = sup JH () ; I () = I0 () + h(0 ) : (1:4)
; H 2C 1 2 ([0;T ]T )
d
In this formula 0 stands for trajectory at time 0. Of course I is identically

equal to 1 outside D([0; T ]; Mo+;1) :
2. Weakly asymmetric simple exclusion processes 263
I ( ) = 1 62 D([0; T ]; M+o;1) :
if (1:5)
In section 5 we present an explicit formula for rate function I0 .

Recall that we consider for simplicity nearest neighbor symmetric simple ex-
clusion processes, i.e., a simple exclusion process with transition probability p
such that p(x) = 0 if jxj > 1 and p(ej ) = p( ej ) = 1=2 for 1 j d, where
ej stands for the j th vector of the canonical basis of Rd . Denote by QN (resp.
P N ) the probability on the path space D([0; T ]; M+) (resp. D([0; T ]; f0; 1gTdN ))
corresponding to the Markov process tN (resp. t ) with generator LN introduced
in Definition 2.2.1 accelerated by N 2 and starting from the invariant state N .
Expectation with respect to P N is denote by E N .
We have now all elements to state the large deviations principle for the em-
pirical measure.
Theorem 1.1 For each closed set C and each open set O of D([0; T ]; M+),
lim sup N d log QN [C ] inf

2C () ;
I
N !1
lim inf N d log QN [O] inf I ( ) :
N !1 2O
Remark 1.2 We prove in this chapter large deviations starting from the equilib-
rium measure N for historical reasons. It is straightforward to extend this proof
to processes starting from a product measure with slowly varying parameter or to
a process starting from a sequence of deterministic configurations N in f0; 1gTN
d
associated to some profile : Td ! [0; 1] in the following sense :

X Z
lim N d H (x=N )N (x) = H (u)(u) du
N !1
x2TdN
for all continuous functions H.
2. Weakly asymmetric simple exclusion processes
The investigation of large deviations from the initial measure is quite simple since
the occupation variables f (x); x 2 TdN g under N are independent Bernoulli
random variables. We are reduced therefore to study large deviations of i.i.d.
random variables : for each continuous function : Td ! (0; 1), let N() denote
the product measure on f0; 1gTN with marginals given by :
d
n o
N() ; (x) = 1 = (x=N ) (2:1)
for all x in TdN . An elementary computation shows that d N()=dN is equal to

d N() Y ( x=N ) (x) 1 (x=N ) 1 (x)

dN =
x2TdN
1

(x) log (x=N 1 (x=N )
X
+ (1 (x)) log :
)
= exp
x2TNd 1
(2:2)
In order to keep notation simple, for a continuous function : Td ! (0; 1),
define hN : M+ ! R as
hN (!) = < !; log (1(1 )) > + < N ; log 11 > : (2:3)
In this formula N stands for thePdiscrete approximation of the Lebesgue measure

and is defined by N = N d x x=N . With this notation we may write the
Radon–Nikodym derivative d N() =dN as
d N() n o
= exp N d hN (0 )
dN
Notice that d N() =dN is a continuous function of the empirical measure. More-
over, for each continuous : Td ! (0; 1), hN converges uniformly to h in
M+ .
We turn now to the large deviations coming from the dynamics. We have
seen in the beginning of this chapter that in order to prove a large deviations
principle, we need to construct mean-one positive martingales. For Markov pro-
cesses, this is done by considering small perturbations of the original dynamics
and taking Radon–Nikodym derivatives. In the context of symmetric simple ex-
clusion processes, the relevant perturbations to introduce are small time and space
dependent asymmetries in the jump rate. More precisely, for each function H
in C 1;2 ([0; T ] Td ), consider the time inhomogeneous Markov process whose
generator at time t is given by
X
(LH
N;tf )( ) = (1=2) (x)[1 (y)]eH (t;y=N ) H (t;x=N ) [f (x;y ) f ()] :
jx yj=1
(2:4)
The interpretation is simple, at time t and site x, instead of a symmetric jump
rate equal to 1=2 to each neighbor, the jump rate from x to x ei is equal to
(1=2)f1 N 1 (@ui H )(t; x=N )g.
For each continuous function : Td ! (0; 1) and H in C 1;2 ([0; T ]
d
T ), denote by P ;HN (resp. QN ) the probability measure on the path space
;H
D([0; T ]; f0; 1gTdN ) (resp. D([0; T ]; M+)) corresponding to the inhomogeneous
Markov process t (resp. tN ) with generator LH N defined in (2.4) accelerated by
N 2 and starting from N(). In the case where is constant equal to , we denote
N simply by P N .
P ;H H
2. Weakly asymmetric simple exclusion processes 265
N =dP N . It is
In section A1.7, we compute the Radon–Nikodym derivative dPH
equal to

exp N d < TN ; HT > < 0N ; H0 >
Z T n o n o
N d dt exp N d < tN ; Ht > (@t + L) exp N d < tN ; Ht > :
0
Taylor’s expansion up to the second order and the elementary inequality jeu
1 u (1=2)u2j (1=6)juj3ejuj permit to rewrite the Radon–Nikodym derivative
dPHN =dP N as
Z T
exp N d < TN ; HT > < 0N ; H0 > < sN ; @s Hs + (1=2)Hs > ds
0
d
X X Z T
(1=2) N d (@ui H (t; x=N ))2 t (x)(1 t (x + ei )) dt
i=1 x2TdN 0

+ OH (N 1
) ;
where OH (N 1 ) is a constant bounded in absolute value by C (H )N 1 .
In order to write this Radon–Nikodym derivative in a simple form, we intro-
duce some notation. For each smooth function H in C 1;2 ([0; T ] Td ), recall the
linear functional `H introduced in (1.2). We consider from now on `H as defined
on D([0; T ]; M+). Thus `H : D([0; T ]; M+) ! R is given by
Z T
`H () = < T ; HT > < 0 ; H0 >
dt < t ; @t Ht + (1=2)Ht > :
(2:5)
0
With this notation we may write the Radon–Nikodym derivative dP ;H N =dP N ,

which is equal to d N() =dN dPH
N =dP N , as
N
dP ;H
N d `H ( N ) + hN ( N ) + OH (N 1 )
dP N = exp 0
d
X X Z T
(1=2) N d (@ui H (t; x=N ))2 t (x)(1 t (x + ei )) dt :
i=1 x2TdN 0
(2:6)
We have seen in the introduction that prove upper bound large deviations we
need the mean-one positive martingales to be function of the process. Unfortu-
N =dP N is not a function of the empirical measure. The first step in the
nately, dP ;H
N =dP N is super-
proof of a large deviations principle is therefore to show that dP ;H
exponentially close to a function of the empirical measure. Here superexponentially
means that the difference between the Radon–Nikodym derivative dP ;H N =dP N
and a function of the empirical measure has expectation of order smaller than
expf CN d g for all C > 0. We need such a small order because we are interested
in large deviations events that have probability of order expf CN d g.
In view of the replacement lemma of Chapter 5, there is at least one possi-
ble approach. We proved there that for each continuous function G and cylinder
function ,
lim sup lim sup
"!0 N !1
Z T h i
X
P N dt N d G(t; x=N ) x () ~("N (x)) > = 0
0
x2TdN
for all > 0. In this formula ~() = E [ ] and that "N () is a function
of the empirical measure. One could hope to prove that this probability is in
fact superexponentially small and obtain in consequence that the Radon–Nikodym
N =dP N is superexponentially close to a function of the empirical
derivative dP ;H
measure. This is the content of the main theorem of next section.
3. A superexponential estimate
We prove in this section the main ingredient required in the investigation of large
deviations of the empirical measure : a superexponential estimate that allows the
replacement of cylinder functions by functions of the density field. Such replace-
ment lemma was already proved in Chapter 5 to close the differential equation
satisfied by the empirical measure. We shall prove below that the probability in-
volved is superexponentially small, that is, of order smaller that expf CN d g for
all C > 0.
Recall from Chapter 5 that for any cylinder function we defined ~ : [0; 1] !
R as ~() = E [ ].
Theorem 3.1 For each G 2 C ([0; T ] Td), each cylinder function and each
" > 0, let
X h i
G; (t; )
VN;" = VN;" (t; ) = N d G(t; x=N ) x () ~("N (x)) :
x2TdN
Then, for any >0
" Z #
T
log P N 1:

lim sup lim sup
Nd
1
VN;" (t; t ) dt >
= (3:1)
"!0 N !1 0
Proof. Since for any positive sequences aN and bN

3. A superexponential estimate 267
lim sup N d log(aN + bN )

N !1 n o (3:2)
max lim sup N d log aN ; lim sup N d log bN ;
G; and V G; .
it is enough to prove (3.1) without the absolute value for VN;" N;"
For each positive a, by Chebychev exponential inequality, the probability on
the left hand side of (3.1) without the absolute value is bounded above by
" #
nZ T o
expf aN dgE N exp aN d VN;"(t; t ) dt :
0
To conclude the proof of the theorem it is therefore enough to show that

" #
nZ T o
N aN dVN;" (t; t ) dt
d log E exp (3:3)
1
lim sup lim sup
"!0 N !1 N 0
0
for every positive a, because in this case we would have proved that the left hand
side of (3.1) is bounded above by a for every positive a and it would remain
to let a increase to 1.
Denote by Sta;V the semigroup associated to the inhomogeneous Markov pro-
cess with generator N 2 LN + aN d VN;" (t; ) and notice that this generator is sym-
metric with respect to the product measure N . By Feynman–Kac formula, the ex-
pectation in (3.3) is equal to < STa;V 1; 1 > (cf. section A1.7) provided < ; >
stands for the inner product in L2 (N ) and 1 for the function on the configuration
space f0; 1gTN identically equal to 1.
d
Denote by a;V (t) the largest eigenvalue of the symmetric operator N 2 LN +

aN VN;" (t; ) on L2 (N ). By the spectral
d representation (Lemma A1.7.2), <
STa;V 1; 1 > is less than or equal to expf 0T a;V (t) dtg. Moreover, by (A3.1.1)
R
the largest eigenvalue a;V (t) is equal to the variational formula

nZ o
sup aN dVN;" (t; )f ()N (d) N 2 DN ( f ) :
f
In this formula, the supremum is carried overR all densities with respect to N (that
is all integrable positive functions f with fdN = 1) and DN is the convex
p p functional defined in Corollary A1.10.3 and given by
and lower semicontinuous
DN (f ) =< LN f; f >.
Up to this point we proved that the expression in (3.3) is bounded above by
Z T nZ o
dt sup aVN;" (t; )f ()N (d) N 2 d DN ( f ) :
0 f
Since is a cylinder function and G is continuous, both are bounded. In partic-
ular, the supremum is bounded by a finite constant C (a; G; ). Therefore, by the
dominated convergence theorem, in order to prove (3.3) we have to show that
nZ o
lim sup lim sup sup aVN;" (t; )f ()N (d) N 2 d DN ( f ) 0
"!0 N !1 f
for every positive a and 0 t T .
Recall that VN;" (t; ) is bounded in absolute value by a constant C ( ; H )
depending only on and H . In particular, in last formula the expression inside
braces is negative whenever the density f is such that DN (f ) is greater than
aC (H; )N d 2. We may therefore restrict the supremum over all densities with
functional DN (f ) bounded by aC (H; )N d 2 . To conclude the proof it remains
therefore to show that
Z
lim sup lim sup sup aVN;"(t; )f ()N (d) 0
"!0 N !1 f ; DN (f )CN d 2
for every 0 t T , a > 0 and positive constant C . This is exactly the content
of the replacement lemma stated in Chapter 5 (cf. Lemma 5.5.7).
Corollary 3.2 The statement of Theorem 3.1 remains in force if the probability
N replaces P N .
P ;H
Proof. By the explicit formula (2.6) for the Radon–Nikodym derivative and since
for simple exclusion processes there is at most one particle per site,
N =dP N k1 expfC ( ; H; T )N dg :
kdP ;H (3:4)
On the other hand, we have

" #
nZ T o
N
E ;H
1

VN;" (t; t ) dt >

0
" #
N n Z T
dP ;H o
= EN

dP N 1 V N;" ( t; t ) dt > :
0
N =dP N by its L1 norm and to apply Theorem 3.1.

It remains to bound dP ;H
4. Large Deviations Upper Bound
In the previous two sections we essentially proved that the martingales dP ;HN =dP N
are functions of the empirical measure. We are therefore ready to carry out the
strategy presented in the introduction of this chapter to prove the upper bound
large deviations.
We shall prove first an upper bound for compact sets. Recall the definition of
the linear functional `H given in (2.5). For a function H in C 1;2 ([0; T ] Td ) and
, " > 0, let BH;;" denote the set of trajectories (t )0tT defined by
4. Large Deviations Upper Bound 269
d Z
X T
BH;;" 2 D([0; T ]; f0; 1gTdN ); H;i (t; ) ;
dt VN;" (4:1)
= t
i=1 0
where
X n o
H;i (t; )
VN;" = N d (@ui H (t; x=N ))2 (x)(1 (x + ei )) F ("N (x)) :
x2TdN
For each positive integer N and " > 0, denote by ";N the approximation of
the identity defined by
";N (u) d 1f[ "; "]g(u) :

1
(2" + N
= 1)
With this notation t"N (x) = (tN ";N )(x=N ). In particular,

X
N d (@ui H (t; x=N ))2 F (t"N (x))
x2TdN
X
= N d (@ui H (t; x=N ))2 F ((tN ";N )(x=N )) :
x2TdN
Notice moreover that tN ";N belongs to Mo+;1 because there is at most one
particle per site. In view of this identity and (2.6), on BH;;" the Radon–Nikodym
derivative dP ;HN =dP N can be written as a function of the empirical measure
modulo some small errors.
On the other hand, by Theorem 3.1, the set BH;;" has probability superexpo-
nentially close to 1 : for each > 0 and H in C 1;2 ([0; T ] Td ),
N c 1:
N d log P [BH;;"] (4:2)
1
lim sup lim sup =
"!0 N !1
These two observations will provide the large deviations upper bound for compact
sets.
Let O denote an open set of D([0; T ]; M+) and fix > 0, H in C 1;2 ([0; T ]
Td ) and a continuous function : Td ! (0; 1). By (3.2), for each " > 0,
log QN [O]
1
N !1 d
N
lim sup
(4:3)
log P N [ 2 O ; BH;;" ] ; U (H; ; ") ;
1
max lim sup
N !1 Nd
where
U (H; ; ") N c
N d log P [BH;;"] :
1
= lim sup
N !1
Keep in mind that by (4.2), lim"!0 U (H; ; ") = 1 for each fixed H and .
We may rewrite the probability on the right hand side of (4.3) as

N n
N dP 1 2 O ; BH;;" :
o
E ;H N
dP ;H
As observed above, on BH;;" the derivative dP ;H N =dP N can be written as a
function of the empirical density modulo small errors. More precisely, on BH;;" ,
N =dP N is equal to
dP ;H

exp N d `H (N ) + hN (0N ) + O() + OH (N 1)
Z T
exp N d (1=2) < N ; krHt k2F ((tN ";N )()) > dt
(4:4)
0
Because `H () is linear, H belongs to C 1;2 ([0; T ] Td ) and there is at most

one particle per site, `H ( N ) = `H ( N ";N ) + oH ("), where oH (") is a constant
depending on H that vanishes in the limit as " # 0. On the other hand, due to the
definition of ";N and the exclusion rule, N ";N belongs to D([0; T ]; Mo+;1).
Furthermore,
< N ; krHt k2F ((tN ";N )()) >
Z
= du k(rH )(t; u)k2F ((tN ";N )(u)) + OH ( N 1
);
d T
where OH (N 1 ) is bounded by a function of H that vanishes as N " 1. In
particular, from (1.2) and (1.3)
Z T
`H (N ) (1=2) < N ; krHtk2 F ((tN ";N )()) > dt
0
JH (N ";N ) + OH (N 1 ) + oH (") :
=
For similar reasons hN N N

(0 ) = h (0 ) + O (N ) where O (N ) is bounded
1 1
by a function of that vanishes as N " 1. In conclusion, on BH;;" , dP ;H

N =dP N
is equal to

exp N d JH (N ";N ) + h (0N ) + O() + OH (N 1) + O (N 1) + oH (")
In particular, replacing in last expectation the Radon–Nikodym derivative by the
previous expression, taking logarithms and dividing by N d , we obtain that
N
N d log P [ 2 O ; BH;;" ] O() + OH (N ) + O (N ) + oH (")
1 1 1
+ sup JH ( ";N ) h (0 ) :

2O
Letting N " 1, recalling (4.3) and minimizing over H , , and ",
4. Large Deviations Upper Bound 271
N
N d log Q [O] H; ;;"
inf sup JH; ;;" ( ) ;
1
lim sup
N !1 2O
where

JH; ;;"() = max JH ( " ) h (0 ) + O() + oH (") ; U (H; ; ")
and " is the approximation of the identity defined in (5.1.8). We have thus proved
an upper bound large deviations for every open set O of D([0; T ]; M+). Since
for each H in C 1;2 ([0; T ] Td ), continuous function : Td ! (0; 1), > 0
and " > 0, JH; ;;" : D([0; T ]; M+) ! R̄ is upper semicontinuous, by Lemma
A2.3.3, for every compact set K,
N
N d log Q [K] sup inf JH; ;;" ( ) :
1
lim sup
N !1 2K H; ;;"
By (4.2) and the definition of JH; ;;" , for each fixed , H , and
lim J ( ) JH ( ) h (0 ) + O()

"!0 H; ;;"
because JH is lower semicontinuous and " converges to as " # 0. Letting
# 0, by definition of the rate function I given in (1.4), we get
n o
N
N d log Q [K] JH () h (0 ) inf I ( )
1
lim sup sup inf + =
N !1 2K H; 2K
for every compact set K of D([0; T ]; M+).
We conclude this section extending the large deviations upper bound to closed
sets. By remark (3.2), it is enough to find a sequence of compact sets Kn such
that
lim sup d log QN [Knc ] n : (4:5)
1
N !1 N
This is the so called exponential tightness of the sequence QN . To construct
such sequence, we proceed in two steps. We first show that for every continuous
function H : Td ! R and " > 0,

log QN 1:

lim lim sup
!0 N !1 Nd
1
sup < t ; H > < s ; H > > " =
jt sj
(4:6)
From this estimate we prove the exponential tightness.
First of all, for N sufficiently large, we have that
n o
sup < tN ; H > < sN ; H > > "

jt sj
T[ 1 ] n
[ o
sup < tN ; H > < k

N ; H > > "=4 :
k=0 kt<(k+1)
We have here "=4 instead of "=3 due to the presence of jumps. Since we start
from equilibrium, by remark (3.2), in order to prove (4.6), it is enough to show
that

log QN 1

lim lim sup
!0 N !1
1
Nd sup < t ; H > < 0 ; H > > " =
t<
0
(4:7)
for every " > 0 and H in C 1;2 ([0; T ] Td ).
Fix a constant a that will increase to 1 after N " 1 and # 0. Denote by
P Rt
Aa;H
t the integral (1=2)N jx yj=1 0 (expfa[H (y=N )
2
H (x=N )]g 1)s(x)[1
s (y)]ds. We introduced Aa;H a;H
t because Mt = expfN a < t ; H > N a <
d N d
a;H
0 ; H > At g is a positive martingale equal to 1 at time 0.
N
In order to prove (4.7), it is enough to prove the same statement (with a"
instead of ") for N d log Mta;H and for N d Aa;Ht . On the one hand, a simple
computation, similar to the one performed to express the Radon–Nikodym deriva-
tive dPH N =dP N as a function of the empirical measure, shows that N d Aa;H is
t
bounded by C (a; H )t because there is at most one particle per site. In particular,
because t , for small enough the probability in (4.7) with N d Aa;H
t vanishes.
On the other hand, in order to prove (4.7) for N d log Mta;H , we first observe
that we can neglect the absolute value. Without the absolute value, by Chebychev
exponential inequality, the probability is bounded above by expf a"N dg because
Mta;H is a positive martingale equal to 1 at time 0. This concludes the proof of
(4.6).
It is now a simple game to obtain a sequence of compact sets satisfying (4.5).
Consider a sequence H` of C 2 (Td ) functions dense in C (Td ) for the uniform
topology. For each > 0 and " > 0, denote by C`;;" the set of all paths t such
that
n o
C`;;" = 2 D([0; T ]; M+);
sup < t ; H` > < s ; H` > " :
jt sj
We just proved that

N 62 C`;;" 1
N d log Q
1
lim lim sup =
!0 N !1
for each ` 1 and " > 0. In particular, for each positive integers `, m and n,
there exists = (`; m; n) such that

QN 62 C`;;1=m expf N dnm`g
for all N large enough. We may extend this inequality to all positive integers N
by modifying if necessary. Consider the set Kno defined by
\
Kno = C`;(`;m;n);1=m :
`1; m1
5. Large Deviations Lower Bound 273
It is quite simple to check that Kn = Kno \ D([0; T ]; M+;1) is a compact set

for each n 1. On the other hand, since there is at most one particle per site,
QN [Kn ] = QN [Kno ]. Furthermore, by construction,

X
QN 62 Kno expf N dnm`g C expf N dng
`1; m1
for some universal constant C . In particular,

log QN 62 Kno n:
1
lim sup
N !1 Nd
This concludes the proof of the large deviations upper bound for closed sets.
5. Large Deviations Lower Bound
We start this section proving the hydrodynamic behavior of the inhomogeneous

Markov process t with generator LH N;t defined in (2.4).
Proposition 5.1 Fix a continuous function : Td ! (0; 1) and H in C 1;2 ([0; T ]
Td ). The sequence of probabilities QN
;H converges in distribution to the probability
measure concentrated on the absolutely continuous path t (du) = (t; u)du whose
density (t; u) is the unique weak solution of the partial differential equation
8
>
>
d
X
< @t = (1=2) @ui (1 )@ui H ;
i=1 (5:1)
>
>
:
(0; ) = ( ) :
The proof presented in Chapter 5 for symmetric zero range processes can easily
be adapted to the case of inhomogeneous weakly asymmetric simple exclusion
processes described above.
We now obtain an explicit form for the large deviations rate function I .
Lemma 5.2 Recall the definition of h and h introduced in (1.1). Then

Z
h(!) = du (u) log (u) + [1 (u)] log 1 1 (u)
if ! (du) = (u)du for some 0 (u) 1. h(! ) = 1 otherwise. Moreover, if
!(du) = (u)du for some continuous function : Td ! (0; 1),
h(!) = h (!) : (5:2)
The proof is elementary and left to the reader.

To obtain an explicit formula for I0 , for each absolutely continuous path in
D([0; T ]; Mo+;1) consider the inner product < ; > on C 1;2 ([0; T ] Td) defined
by
Z T Z n o
< G; H > = dt du (rH )(t; u) (rG)(t; u) F ((t; u)) :
0 T d
Denote by N ( ) the kernel of this inner product
and by H1() the Hilbert space
obtained by completing C ([0; T ] T )
1;2 d
.
N ()
Lemma 5.3 Assume that I0 ( ) < 1 and denote by the density of : (t; du) =
(t; u)du. There exists H in H1 () such that
Z T Z
I0 () =
1
dt du k(rH )(t; u)k2F ((t; u)) :
2 0 Td
Proof. Assume that I0 ( ) < 1 and recall the definition of the linear functional
`G given in (2.5). By definition of I0 (), for every G in C 1;2 ([0; T ] Td),
Z T Z
`G() (1=2) dt du k(rG)(t; u)k2F ((t; u)) I0 () :
0 Td
Fix a real a. Considering aG instead of G in the previous formula and minimizing
over a, we obtain that
Z T Z 1=2
p
`
G( ) 2I0 ( ) dt du k(rG)(t; u)k F ((t; u))
2

0 Td
Notice that ` ( ) is a linear functional on C 1;2 ([0; T ] Td ). We just proved
that it is bounded in H1 ( ). We may therefore extend ` ( ) to H1 ( ). In this case,
by Riesz’ representation theorem, there exists H in H1 ( ) such that
Z T Z

`G () = dt du rG rH (t; u)F ((t; u))
0 T d
for each G in H1 ( ).
It is now easy to derive the explicit formula for the rate function I0 . By
definition of I0 and by the above representation of the linear functional ` ( ),
I0 () =
n Z T Z o
sup `G ( ) (1=2) dt du k(rG)(t; u)k2F ((t; u))
G2C 1;2([0;T ]Td) 0 T
d
nZ T Z

= sup dt du rG rH (t; u)F ((t; u))
G2C 1;2 ([0;T ]Td) 0 T
d
Z T Z o
(1=2) dt du k(rG)(t; u)k2F ((t; u))
0 T
d
RT R
Adding and subtracting (1=2) 0
dt du k(rH )(t; u)k2F ((t; u)), we rewrite last
supremum as
Z T Z
(1=2) dt du k(rH )(t; u)k2F ((t; u))
0 Td
nZ T Z o
(1=2) sup dt du kr(G H )(t; u)k2F ((t; u))
G2C 1;2([0;T ]Td) 0 T d
Since C 1;2 ([0; T ] Td ) is dense in H1 ( ), this last expression is equal to

Z T Z
(1=2) dt du k(rH )(t; u)k2F ((t; u))
0 Td
and the proof is concluded.
For each continuous : Td ! (0; 1) and H in C 1;2 ([0; T ] Td ), denote by
the weak solution of (5.1). Let ;H stand for the path on D([0; T ]; M+)
;H
with density ;H : ;H (t; du) = ;H (t; u)du. Multiplying both sides of equation
(5.1) by H and integrating by parts we obtain that
Z T Z
`H ( ;H ) = dt du k(rH )(t; u)k2F ( ;H
t ):
0 Td
In particular, from last lemma,
Z T Z
I0 ( ;H ) = (1=2) dt du k(rH )(t; u)k2F ( ;H (t; u))
0 Td
Z T Z
= `H ( ;H ) (1=2) dt du k(rH )(t; u)k2F ( ;H (t; u))
0 Td
= JH ( ;H ) :
This last equality and (5.2) show that
I ( ;H ) = h (0 ;H ) + JH ( ;H ) : (5:3)
The previous two lemmas permit to interpret the large deviations rate function
as an entropy. This statement is clarified in next lemma.
Lemma 5.4 Recall from section A1.8 the definition of the entropy. For each con-
tinuous function : Td ! (0; 1) and smooth function H in C 1;2 ([0; T ] Td ),

N d H P N P N I ( ;H ) :
lim
N !1 ;H =
Proof. Recall the definition of the set BH;;" given in (4.1). By Corollary 3.2, the
c
probability of BH;;" N is superexponentially small. In other
with respect to P ;H
words, statement (4.2) remains correct if we replace P N by P ;H N .
By the explicit formula for the entropy,

N
N log dP ;H :
h i
N P N
N dH P ;H N dE ;H
=
dP N
Since by (3.4) N d log(dP ;HN =dP N ) is bounded by C ( ; H; T ) and since by (4.2)
N instead of P N ) the probability of B c
(with P ;H H;;" is superexponentially small,
the right hand side of last identity is equal to
N
N log dP ;H 1fBH;;" g
h i
N d E ;H dP N + oN (1)
for all > 0 and each " small enough (" < "()). By (4.4), on BH;;" ,
N d log(dP ;H
N =dP N ) is equal to
Z T
`H (N ) + hN (0N ) (1=2) < N ; krHt k2 F ((tN ";N )()) > dt
0
+ O ( ) + OH (N 1
)
c
Since this expression is bounded and the probability of BH;;" with respect to
N vanishes as N " 1, last expectation (and therefore the entropy) is equal to
P ;H
h Z T i
N `H ( ) + h (0 )
E ;H (1=2) < N ; krHt k2F ((t ";N )()) > dt
0
+ O ( ) + oN (1)
for all positive and all " small enough. Notice also that we replaced hN by h
N
because h converges uniformly to h . We may also replace the discrete Lebesgue
measure N by . The details were given in the proof of the upper bound.
All expression inside the expectation are continuous with respect to the Sko-
rohod topology. By Proposition 5.1 the sequence QN ;H converges weakly to the
probability concentrated on the weak solution of (5.1). In particular, as N " 1,
the previous expectation converges to
`H ( ;H ) + h (0 ;H )
Z T Z
(1=2) dt du krH (t; u)k2F ((t ;H " )(u)) + O ( ) :
0 Td
It remains to let " # 0, then # 0 and recall identity (5.3).
Denote by Do ([0; T ]; M+) the subset of D([0; T ]; M+) consisting of all paths
;H associated to some continuous function : Td ! (0; 1) and some H in
C 1;2 ([0; T ] Td). That is, the set of all trajectories of D([0; T ]; M+) that are
absolutely continuous ( (t; du) = (t; u)du) and for which there exists a continuous
: Td ! (0; 1) and H in C 1;2 ([0; T ] Td) so that is solution of (5.1).
We are now ready to prove the lower bound large deviations.
Proof of the lower bound. Let O be an open set of D([0; T ]; M+). We shall first
prove that
log QN [O] I ( ) (5:4)
1
N !1 N d
lim inf
for all paths in O \ Do ([0; T ]; M+). Since belongs to Do ([0; T ]; M+), there
exists a continuous : Td ! (0; 1) and H in C 1;2 ([0; T ] Td) such that = ;H .
O the probability on D([0; T ]; f0; 1gTN ) given by
N d
Denote by P ;H;
N [A] N N
P ;H; N [N 2 O] P ;H [A; 2 O]
1
O =
P ;H
A of D([0; T ]; f0; 1gTdN ). By Jensen’s inequality,
for all measurable set
N i
d log dP
h
N d log QN [O] E ;H;
N
O N dP ;H
d N
N + N log Q ;H [O] :
By Proposition 5.1, since O is a neighborhood that contains ;H , the second

expression on the right hand side converges to 0 as N " 1. The first one is equal
to
N
N N d log dP 1f N 2 Oc g :
h i
1 N P N
N dH P ;H E ;H
QN ;H [O] dP N
;H
;H [O] converges to 1 as N " 1. Since by (3.4)
Once again, by Proposition 5.1, QN
the expression N d log(dP N =dP ;H
N ) is bounded, the second term inside braces
vanishes as N " 1. Therefore,

lim inf QN [O] Nlim N dH P ;H
N PN :
N !1 !1
Recall the statement of Lemma 5.4 to conclude the proof of (5.4) for paths in
O \ Do ([0; T ]; M+).
To conclude the proof of the lower bound it remains to show that all paths
with finite rate function (I () < 1) can be approximated by a sequence n
in Do ([0; T ]; M+) such that limn!1 I (n ) = I ( ). This is the content of next
lemma.
Lemma 5.5 For each with finite rate function (I ( ) < 1), there exist a sequence
n in Do ([0; T ]; M+) converging to in D([0; T ]; M+) and such that
lim
n!1
I (n ) = I ( )
Proof. In this proof we adopt the following terminology. A sequence of paths n is

said to approximate if n converges to in D([0; T ]; M+) and limn!1 I (n ) =
I ().
The proof is divided in three steps. We first show that we can approximate
each path with finite rate function by absolutely continuous paths with density
bounded below by a strictly positive constant and bounded above by a constant
strictly smaller than 1. In this case, we say that the density is bounded away from
0 and 1.
Let in D([0; T ]; M+) with I ( ) < 1. In particular, by (1.5), is absolutely
continuous : (t; du) = (t; u)du with density bounded below by 0 and above
by 1. Denote by e1 (resp. e0 ) the constant path with density equal to 1 (0) :
e (t; du) = du for 0 t T (resp. e0 (t; du) = 0 for 0 t T ). Notice that
1
both e0 and e1 belong to Do ([0; T ]; M+) and I ( e1 ) = log < 1 (I (e0 ) =
log(1 ) < 1).
For 0 " < 1, define " by " = (1 ") + ("=2) e0. Denote by "
e1 + ("=2)
" " "
the density of : (t; du) = (t; u)du. By construction,
"=2 " (t; u) 1 ("=2) : (5:5)
Moreover, it is clear that " converges to as " # 0. In particular, by lower

semicontinuity of the rate function, I ( ) lim inf"!0 I ( " ). On the other
hand, by convexity, I ( " ) (1 ")I ( ) + ("=2)I (
e0) + ("=2)I (
e1). Therefore,
lim sup"!0 I ( ) I ( ).
"
In conclusion, for each path with finite rate function we constructed an approx-
imating sequence " of absolutely continuous trajectories with density bounded
away from 0 and 1.
It remains to show that we can approximate an absolutely continuous path
(t; du) = (t; u)du with density bounded away from 0 and 1 by a sequence n
in Do ([0; T ]; M+).
Consider such a path in D([0; T ]; Mo+;1) For each " > 0, denote by " a
smooth approximation of the identity :
Z
" 0 ; supp " [ "; "]d ; " (u) du = 1 :
Rd
R
Let " denote the spatial convolution of with " : " (t; u) = (t; u v )" (v )dv
and " (t; du) = " (t; u)du. It is easy to check that " converges to in
D([0; T ]; M+). In particular, by lower semicontinuity of the rate function, I ()
lim inf"!0 I ( " ). On the other hand, by convexity and translation invariance of I ,
Z
I ( " ) I (Tu )" (u) du = I ( ) :
In this formula Tu stands for the translation of by u : (Tu )(t; dv ) = (t; v
u)dv. Therefore lim sup"!0 I (" ) I (). Notice that the density " is bounded
below and above by the same values that bound .
To conclude the proof of the lemma we have to show that we can approximate
every absolutely continuous path with density bounded away from 0 and 1 and
which is smooth in the space variable by a sequence in Do ([0; T ]; M+).
Fix such a path (t; du) = (t; u)du. For each 0 t 1, extend the definition
of to [T; T + 1] by setting (T + t; u) = e(t; u), where e(t; u) is the solution of
the heat equation with initial condition (T; ) :
(
@t e = (1=2)e
e(0; ) = (T; ) :
For 0 t < 1, denote by t the time translation of : (t )(s; u) = (t + s; u)
for (s; u) in [0; T ] Td . By extension we define (t )(s; du) = (t )(s; u)du. We
leave to the reader, as an exercise, to check that I0 (t ) I0 ( ). This result
follows from the variational formula for the rate function I0 and Lemma 5.3. It is
a quite natural result since we extended following the hydrodynamic equation.
For each 0 < " < 1, Rdenote by " a smooth one-dimensional approximation
of the identity : " 0, " (s)ds = 1, supp " [0; "]. Notice that the support
of " is contained in R R+ .
Let " (t; u) = " (s)(s ) ds and " (t; du) = " (t; u)du. Clearly, " con-
verges to as " # 0. In particular, by lower semicontinuity,
I () lim inf I ( " ) :

"!0
R
In contrast, by convexity, I ( " ) " (s)I (s ) ds. We left to the reader to show
that I0 (s ) I0 ( ). On the other hand, since h((s )0 ) = h(s ) and since, by the
explicit formula for h, lims!0 h(s ) = h(0 ), we have lims!0 h((s )0 ) = h(0 ).
In conclusion, lim sup"!0 I ( " ) I ( ).
Finally, we claim that " belongs to Do ([0; T ]; M+) for every " > 0. Since
"
has a smooth density bounded away from 0 and 1, we may solve equation
(5.1) in H and obtain that H is of class C 1;2 ([0; T ] Td ). Since = (0; ), is
a continuous function bounded away from 0 and 1, which proves that " belongs
to Do ([0; T ]; M+).
We presented in this chapter the ideas of Kipnis, Olla and Varadhan (1989) and
Donsker and Varadhan (1989) who were the first to investigate the large deviations
of the empirical measure from the hydrodynamic limit.
Extensions.
(a) Infinite volume. A large deviations principle for the empirical measure for
symmetric simple exclusion processes and mean-zero asymmetric zero range
processes on the lattice Zd starting from an equilibrium state were obtained by
Landim (1992) and Benois, Kipnis and Landim (1995). Based on the investi-
gation of the time evolution of the H 1 norm and on estimates derived in Yau
(1994) this result was extended by Landim and Yau (1995) to one-dimensional
Ginzburg–Landau processes in infinite volume starting from a large class of

non equilibrium states, including deterministic initial configurations. They also
proved a similar result for attractive zero range processes through coupling
techniques.
(b) Nonconservative systems. Large deviations from the hydrodynamic limit for
reaction–diffusion models were considered by Jona–Lasinio, Landim and Vares
(1993) and Landim (1991c). In contrast with conservative systems described
by parabolic equations, the rate function involves exponential terms and cannot
be interpreted as due to a simple stochastic perturbation of the hydrodynamic
equation. Jona–Lasinio (1991), (1992) discusses further this issue.
(c) Nongradient systems. Quastel (1995a) proved a large deviations principle for
the empirical measure in the case of a one-dimensional nongradient Ginzburg–
Landau model. His proof applies in any dimension provided the diffusion coef-
ficient is Lipschitz continuous. Quastel and Yau (1997) derived weak solutions
of the incompressible Navier Stokes equations from an interacting particle sys-
tem where particles have velocity. The generator has two part : The first one
corresponds to displacements of particles and the second to collisions. Both
density and momenta are conserved by the dynamics. They deduce further-
more the large deviations from the hydrodynamic limit. Two different large
deviations rates arise in this model : The probability to violate the divergence
free condition decays at rate at least expf CN 1 d g, while the probability to
violate the momentum conservation decays at rate expf CN 2 d g.
Onsager–Machlup time–reversal relation. Consider, to fix ideas, the symmetric
simple exclusion process and recall the notation introduced in this section. Fix
a density . For a profile : Td ! [0; 1], the specific entropy N d H ( N() jN )
converges, as N " 1, to the entropy functional S ( ) that may be written as the
integral of a density s ( ) :
Z
S ( ) = du s( (u)) ;
Td
where s (a) = a logfa=g + (1 a) logf(1 a)=(1 )g.
Denote by : Td ! [0; 1] the constant profile with density : (u) =
for u in Td . is an equilibrium state of the heat equation (4.2.1) and its basin
of attraction consists of all profiles
R with total mass equal to : for every profile
: Td ! [0; 1] with density ( Td du (u) = ), the solution of the hydrodynamic
equation with initial profile relaxes, as t " 1, to .
Denote by L1 d
+;1; (T ) the space of positive, measurable functions on T
d
bounded by 1 and with density . For T1 < T2 , denote by I[T1 ;T2 ] () the large
deviations rate functional for the empirical measure on the time interval [T1 ; T2 ].
Define the quasi–potential V : L+1;1; (Td ) ! R+ by
V ( ) = inf I( 1;0] () ;

where the infimum is taken over all trajectories : ( 1; 0] Td ! [0; 1] such

that (0; ) = , limt! 1 (t; ) = . The quasi–potential V ( ) represents the
price to create a fluctuation from to .
The Onsager–Machlup principle (Onsager and Machlup (1953)) states that for
reversible systems the infimum is attained by the reversed trajectory : V ( ) =
I( 1;0] ( ), where is the time reversed trajectory : (t; u) = ()( t; u) if
() stands for the solution of the heat equation with initial data . Furthermore,
by the Boltzmann–Einstein relation, V and S coincide up to an additive constant.
The Onsager–Machlup time–reversal relation has been conjectured for a wide
class of systems by Eyink (1990). It has been proved by Gabrielli, Jona–Lasinio,
Landim (1996) for a particular class of non reversible, conservative interacting
particle systems showing that reversibility is not a necessary condition for the
Onsager–Machlup principle. Gabrielli et al. (1997) proved the Onsager–Machlup
principle for non reversible and non conservative systems obtained superpos-
ing speeded up symmetric simple exclusion processes with Glauber dynamics.
The evolution of these later processes at the macroscopic level are described by
reaction–diffusion equations (De Masi, Ferrari and Lebowitz (1986)).
The Onsager–Machlup time–reversal relation is further discussed in Eyink,
Lebowitz, Spohn (1996) in a general context. The question remains open for a
general non reversible system associated to a reaction–diffusion equation.
Metastability. Consider a reaction–diffusion model obtained by the superposition
of a speeded up symmetric simple exclusion process with a Glauber dynamics.
d
This is a Markov process on f0; 1gTN whose generator is
X
(LN f )( ) = N2 (x)[1 (y)][f (x;y) f ()]
jx yj=1
X (6:1)
+ (x c)( )[f (x ) f ()] :
x2TdN
In this formula c( ) is a cylinder function and x stands for the configuration
with the spin at x flipped :

1 ( x) if y = x ;
(x )(y ) =
(y ) otherwise :
The hydrodynamic behavior (De Masi, Ferrari and Lebowitz (1986)) of this
system is described by a reaction–diffusion equation :
@t = (1=2) +
F () ; (6:2)
where F () is a polynomial given by F () = E [f1 2 (0)gc( )], provided

stands for the Bernoulli product measure with density . To fix ideas, assume that
F = W 0 is a double well potential and denote by (resp. + ) the local (resp.
global) minima of the potential W . Assume that the constant profiles , + are
stable equilibrium points of the differential equation (6.2). Denote by B , B+
(resp. @B , @B+ ) the basins of attraction (resp. the boundary of the basins of

attraction for the weak topology) of these equilibrium points.
Fix a profile 0 in B and consider a system starting from N0 () . Since
is a stable equilibrium point of the hydrodynamic equation, the empirical
measure converges in probability, as N " 1 and t " 1 to (u)du. The typical
behavior of the empirical measure is therefore to relax towards the equilibrium
and to fluctuate around it due to the stochastic character of the dynamics.
Since is a stable equilibrium point of the hydrodynamic equation, we expect
the system to remains a long time in a stationary situation around (u)du. In
particular, two systems starting from two different profiles 1 and 2 in the basin
of attraction of will rapidly be very close one to the other. In this sense the
process loses memory of its starting point. Eventually one large fluctuation will
drive the empirical measure out of the basin of attraction B to the basin of
attraction of + , where it stays for a much larger time because + is the global
minimum of the potential W . Denote by TN the first time the process leaves B :
TN = infft 0; tN 62 B g. Due to the loss of memory of the system, it is
natural to conjecture that TN correctly renormalized converges in some sense to
an exponential distribution :
TN ! exp(1) : (6:3)
E N [TN ]
0
Cassandro, Galves, Olivieri and Vares (1984) proposed a mathematical for-

mulation to describe this phenomenon. For fixed N in N and R > 0, denote by
ANR (s) the time average
Z s+ R
ANR (s) R s tN dt ;
=
1
where stands for the Dirac mass at . Cassandro, Galves, Olivieri and Vares
suggested to characterize the meta–stable behavior of the process by proving the
existence of a sequence RN " 1 such that AN RN (sE N0 () [TN ]) converges in dis-
tribution to a jump process A(s) given by

(u)du for s<T;
A(s)
sT;
=
(u)du
+
for
where T is a mean-one exponential random variable. They proved this result as
well as (6.3) for the total magnetization in a Curie–Weiss model.
The meta–stable behavior of interacting particle systems has been widely stud-
ied. We refer to Penrose and Lebowitz (1987) and to the recent book by Olivieri
and Vares (1997) for clear and detailed presentations of the subject.
Exit points from a basin of attraction. We keep here the notation introduced
in the previous subsection. Recall that V stands for the quasi–potential corre-
sponding to the stable equilibrium . Denote by A the subset of the boundary
@B where the quasi–potential assumes its minimum :
n o
A = 2 @B ; V ( ) = 0 2inf V ( 0 ) :
@B
Since the quasi–potential represents the price needed to create a fluctuation from
the equilibrium point , it is natural to expect that the system leaves B
through A : for every > 0,
h i
lim P N
N !1 0 ()
d(TN ; A ) > = 0: (6:4)
This statement has been proved by Comets (1987) for a mean field, non conserva-
tive Glauber dynamics under some technical assumptions on the quasi–potential.
For local dynamics this question and the meta–stable behavior of the empirical
measure remain open problems.
Escape from unstable equilibrium points. To examine the escape from an unsta-
ble equilibrium point, consider the same reaction–diffusion process with generator
given by (6.1). Recall that the hydrodynamic behavior is described by the reaction–
diffusion equation (6.2), where F () = E [f1 2 (0)gc( )]. Denote by W () the
potential associated to the reaction part : W 0 () = F () and fix an unstable equi-
librium point : W 0 ( ) = 0, W 00 ( ) < 0. Consider a system starting from the
Bernoulli product measure with density . By the conservation of local equilib-
rium, proved for these systems by De Masi, Ferrari, Lebowitz (1986), for a fixed
macroscopic time t, the system remains close to : limN !1 StN [uN ] N =
for every t > 0, u in Td , where fStN ; t 0g stands for the semigroup of the
Markov process with generator LN defined in (6.1). However, since is an
unstable equilibrium point, one might believe that in a longer time scale the pro-
cess escapes from the unstable equilibrium point and converges to some stable
equilibrium point.
Since the behavior depends drastically on the shape of the potential W close
to the saddle point, we introduce two distinct examples. Following the review of
Vares (1991), let the jump rate c( ) in (6.1) be equal to
(a) c( ) = 1 (0)[(1) + ( 1)] + 2 (1)( 1) ;
(b) c() = 1
n (0)(1) on1 (0)( 1) o1 c (0)(2)(3)(4) :
2 2
In this formula, (x) stands for 2 (x) 1 and 1=2 < 1, 1=4 < c 1. The
upper bounds are required to ensure that the jump rates are positive and the lower
bounds to guarantee that 1=2 is an unstable equilibrium point.
In example (a) the potential is equal to W () = 4 1 (1 2 )(2 1)2 +
8 (2 1)4 and is quadratic in a neighborhood of the unstable equilibrium
1 2
point = 1=2. In example (b) the potential is W () = 8 1 (4 1 c)(2 1)4 +

2 4 c(2 1)6 and is quartic close to the unstable equilibrium point = 1=2.
De Masi, Presutti, Vares (1986), De Masi, Pellegrinotti, Presutti and Vares
(1994) proved for example (a) that

1=2 t < (2A) 1 ;
SN N
if
(6:5)
N !1 t log N (1=2)f + + g
lim =
if t > (2A) 1
and that Z 1
SN N
N !1 (1=2A) log N +t
lim = t (d) ;
0
where A = 2(2 1), convergence is meant by weak convergence, t is a Lebesgue
absolutely continuous probability measure that converges, as t " 1, to (1=2)f +
+ g and , + are the stable equilibrium points of the potential W .
Notice in this example that, although being produced by a stochastic fluctua-
tion, the escape from the unstable equilibrium point occurs at a deterministic time
in the macroscopic scale log N .
For example (b) Calderoni, Pellegrinotti, Presutti, Vares (1989) proved that
lim S Np N = a(t) + [1 a(t)](1=2)f + + g ;

N !1 t N
where a(t) = P [S > t] and S is the explosion time of the diffusion dZt =
AZt3 dt + dWt with A = 2[c (1=4)]; Z0 = 0 and Wt a standard Brownian motion.
Therefore, in the case where the potential is quartic around the saddle point, the
escape time is a random variable. This is the expected behavior for every higher
degree of degeneracy.
De Masi and Presutti (1991) present a proof of (6.5). An extension to infinite
volume is contained in De Masi, Pellegrinotti, Presutti, Vares (1994). Giacomin
(1994), (1995) investigates the problem in higher dimension.
Rare events. Related to the question of metastability is the investigation of
occurrence times of rare events. To fix ideas, consider an interacting particle system
t evolving on the lattice Zd with invariant measure , not necessarily unique. Let
fAN ; N 1g be a sequence of events such that limN !1 (AN ) = 0 and define
the sequence of hitting times TN = infft 0; t 2 AN g. There are two natural
questions regarding the asymptotic behavior of TN : its magnitude and the limit
distribution of TN correctly renormalized.
This type of question has been analyzed for Markov processes with good re-
currence properties, by Bellman and Harris (1951) and Harris (1952) for recurrent,
discrete time Markov chains on a countable set; by Aldous (1982, 1989) and by
Aldous and Brown (1992, 1993) for finite state Markov chains and by Korolyuk
and Sil’vestrov (1984) and Cogburn (1985) in the case of Harris recurrent chains.
In the context of interacting particle systems, the hitting times of rare events
were first examined by Lebowitz and Schonmann (1987). They considered the first
time the density in a cube N = f N; : : : ; N gd performs a large fluctuation in
the case of attractive stochastic spin systems starting from equilibrium. Assuming
rapid convergence to equilibrium they deduced estimates on the order of magnitude
of the hitting times TN and they proved that TN =N , for some suitably chosen
sequence N , converges in distribution to a mean-one exponential random variable.
These results were extended to the one-dimensional supercritical contact process
by Galves, Martinelli and Olivieri (1989).
As in the question of metastability, the main difficulty in the investigation of
hitting times of rare events is to show that the process loses memory in the scale
where the phenomenon occurs. Typically, nonconservative systems lose memory

much faster than conservative ones. There are, however, results for conserva-
tive dynamics. For the one-dimensional, nearest neighbor, totally asymmetric zero
range process with jump rate g (k ) = 1fk 1g, Ferrari, Galves and Landim (1994)
proved the estimate
h i

sup P¯ ' '(1 ')N TN > t e t (N + 3)(1 ')N=2
t0
if TN stands for the first time the sites f1; : : : ; N g become empty and ¯' for
the product, translation invariant, stationary measure with marginals given by
¯' f; (x) = kg = (1 ')'k . Ferrari, Galves and Liggett (1995) studied the
same problem for the one-dimensional nearest neighbor symmetric simple exclu-
sion process. They proved that
h i

sup P N N TN > t e t C0 C N
1
t0
for some sequence 0 < N 0 < 1. In this formula stands for the
Bernoulli product measure with density and C0 , C1 for two finite constants.
Asselah and Dai Pra (1997) extends this result to higher dimension for the first
time the empirical density in the cube f1; : : : ; N gd reaches a level 0 > .
Large deviations of asymmetric models. This issue, which remains one of the
main open questions in the theory of hydrodynamic behavior of interacting particle
systems, is discussed in Chapter 8.
11. Equilibrium Fluctuations of Reversible Dynamics
In Chapters 4 to 7 we examined the hydrodynamic behavior of several mean-zero

interacting particle systems and proved a law of large numbers under diffusive
rescaling for the empirical measure. We now investigate the fluctuations of the
empirical measure around the hydrodynamic limit starting from an equilibrium
state. To fix ideas, we consider the nearest neighbor symmetric zero range process.
The reader shall notice, however, that the approach presented below applies to a
large class of reversible models including nongradient systems. The generator of
this process is
X
(LN f )( ) = p(y)g((x))[f (x;x+y) f ()] ; (0:1)
x;y2TdN
where p(y ) = 1=2 if jy j = 1 and 0 otherwise and g is a rate function satisfying the
assumptions of Definition 2.3.1.
We proved in Chapters 5 and 6 that for a class of zero range processes starting
from a sequence of probability measures fN ; N 1g associated to a profile
0 : Td ! R+ , the empirical measure N converges in probability to an absolutely
continuous measure whose density is the weak solution of the non linear heat
equation 8
> d
X
>
< @t = (1=2) @uj 0 ()@uj
j =1 (0:2)
>
>
:
(0; ) = 0 () :
To investigate the equilibrium fluctuations of N , we fix once for all a density
> 0 and we denote by YN the density fluctuation field that acts on smooth
functions H as
X
YtN (H ) = N d=2 H (x=N )(tN (x) ) :
2 (0:3)
x2TdN
Notice the diffusive rescaling of time on the right hand side of this identity. The
aim of this chapter is to prove that YN converges to a stationary Gaussian process
with a given space–time correlations.
To state the main theorem of this section we need to introduce some notation.
Consider the lattice Zd endowed with the lexicographical order. Let h0 1 and
288 11. Equilibrium Fluctuations of Reversible Dynamics
p
for each z > 0p(resp. z < 0), define hz : Td ! R by hz (u) = 2 cos(2z u)
(resp. hz (u) = 2 sin(2z u)). Here denotes the inner product of Rd . It is well
known that the set fhz ; z 2 Zdg is an orthonormal basis of L2 (Td ) : each function
f in L2(Td) can be written as
X
f = < f; hz > hz :
z2Zd
In this formula and below < ; > stands for the inner product of L2 (Td ).
Consider on L2 (Td ) the positive, symmetric linear operator L = (1 ). A
simple computation shows that the functions hz are eigenvectors :
Lhz z hz ;
=
where z = 1 + 4 2 kz k2. For a positive integer k , denote by Hk the Hilbert space
obtained as the completion of C 1 (Td ) endowed with the inner product < ; >k
defined by
< f; g >k = < f; Lk g > :
It is easy to check that Hk is the subspace of L2 (Td ) consisting of all functions
f such that X
< f; hz >2 zk < 1 :
z2Zd
In particular, if we denote L2 (Td) by H0 ,
H0 H1 H2 (0:4)
Moreover, on Hk the inner product < ; >k can be expressed as
X
< f; g >k = < f; hz >< g; hz > zk :
z2Zd
For each positive integer k , denote by H k the dual of Hk relatively to the
inner product < ; >. H k can be obtained as the completion of L2 (Td) with
respect to the inner product obtained from the quadratic form < f; f > k defined
by n o
< f; f > k = sup 2 < f; g > < g; g >k :
g2Hk
It is again easy to check that H k consists of all sequences f< f; hz >; z 2 Zdg
such that X
< f; hz >2 z k < 1
z2Zd
and that the inner product < f; g > k of two functions f , g in H k can be written
as X
< f; g > k = < f; hz >< g; hz > z k :
z2Zd
It follows also from the explicit characterization of H k and from (0.4) that
11. Equilibrium Fluctuations of Reversible Dynamics 289
H2 H1 H0 H 1 H 2
We shall consider the density fluctuation field YtN as taking values in the
Sobolev space H k for some large enough k . Fix a time T > 0, a positive integer
k0 and denote by D([0; T ]; H k0 ) (resp. C ([0; T ]; H k0 )) the space of H k0 valued
functions, that are right continuous with left limits (resp. continuous), endowed
with the uniform weak topology : a sequence fYj ; j 1g converges to a path Y
if Yj (t) converges weakly to Y (t) uniformly in time, i.e., if for all f in Hk0 ,

lim
j !1 0tT
sup < Yj (t) ; f > < Y (t) ; f > = 0:
Denote by QN the probability measure on D([0; T ]; H k0 ) induced by the density

fluctuation field Y N introduced in (0.3) and the product measure N , by PN the
d
probability measure on D([0; T ]; NTN ) induced by the probability measure N
and the Markov process t speeded up by N 2 and denote by E N expectation with
respect to PN .
Theorem 0.1 Fix a positive integer k0 > 2 + (d=2). Let Q be the probability mea-
sure concentrated on C ([0; T ]; H k0 ) corresponding to the stationary generalized
Ornstein–Uhlenbeck process with mean 0 and covariance
h i
EQ Yt (H )Ys (G) =
() Z Z n (u v )2 o (0:5)
(2 (t s)0 ())d=2 Rd
du dv H̄ ( u ) exp
2(t s)0 ()
Ḡ ( v )
Rd
for every 0 s t and H , G in Hk . Here () stands for the static compressibility
given by () = Var( ; (0)) and H̄ , Ḡ: Rd ! R are periodic functions with
0
period Td and equal to H , G on Td . Then, the sequence QN converges weakly to

the probability measure Q.
Theorem 0.1 relates the covariance of the equilibrium density fluctuation to the
diffusion coefficient of the hydrodynamic equation (0.2), a parameter determined
by the non equilibrium evolution. In the mathematical physics literature this result
is called a fluctuation–dissipation theorem since it connects the non equilibrium
dissipative feature of the system to its equilibrium fluctuations.
The proof of Theorem 0.1 relies on Holley and Stroock’s theory of generalized
Ornstein–Uhlenbeck processes that we now explain. Denote by A the nonnegative
self adjoint operator (1=2)0() defined on a domain of L2 (Td ), by fTt ; t 0g
the semigroup associated to A and by B the linear operator ()r. For t 0, let
Ft be the –algebra on D([0; T ]; H k0 ) generated by Ys (H ) for s t and H in
C 1 (Td) and set F = ([t0 Ft).
Theorem 0.2 Fix a positive integer k1 2. Let Q be a probability measure on
the space fC ([0; T ]; H k1 ); Fg. Assume that for each H in C 1 (Td ),
Z t
MtA;H = Yt (H ) Y0 ( H ) Ys (AH ) ds
0 (0:6)
and (MtA;H )2
kBH k t 2
2
are L1 (Q) Ft –martingales. Then, for all 0 s < t, H in C 1 (Td ) and subsets A
of Rd ,
h i
Q Yt (H ) 2 A Fs =
Z n (y Ys (Tt s H ))2 o
s kBT H k2 dr dy Q a.s. :
1
q Rt exp Rt
s kBT
A 2 0 r H k22 dr 2 0 r 2
(0:7)
In particular, condition (0.6) and the knowledge of the restriction of Q to F0
uniquely determines Q on fC ([0; T ]; H k1 ); Fg.
The proof of this theorem is postponed to section 4. It states that for each distri-
bution q on H k1 , there exists a unique probability measure Q on C ([0; T ]; H k1 )
that solves the martingale problem (0.6) and such that Q restricted to F0 is equal
to q . In our setting, for any fixed time t0 , the limit distribution of YtN is easy to
deduce : we shall prove at the beginning of section 2 that YtN
0
0
converges in law
to a mean-zero Gaussian field with covariance given by
EQ [Y (H )Y (G)] = () < H; G > (0:8)
for each smooth function G, H in Hk . In particular, Theorem 0.2 reduces the

proof of Theorem 0.1 to the verification that the sequence QN converges to a
0
probability measure Q that solves the martingale problem (0.6).
Relation (0.6) and the equal time covariances EQ [Yt (G)Yt (H )] given by (0.8)
permit to deduce the space time covariances EQ [Ys (G)Yt (H )] : an expansion
argument gives that for 0 s < t
h i
EQ Ys (G)Yt (H ) = () < Tt s G; H > ;
which is precisely the right hand side of (0.5). Moreover, by (0.6), for each H
in Hk1 , WtH = kBH k2 1 MtA;H is a martingale with quadratic variation equal to
t. Therefore, by the martingale characterization of Brownian motion due to Levy,
WtH is a Brownian motion and we may rewrite (0.6) as
Z t
Yt ( H ) = Y0 (H ) + Ys (AH ) ds + kBH k2WtH ; (0:9)
0
where Wt is a generalized Brownian motions with covariance

Z
h
G H
i
EQ Ws Wt = (s ^ t) d kr rG(u) rH (u) du :
T Gk2 krH k2
11. Equilibrium Fluctuations of Reversible Dynamics 291
To deduce this last relation we used the identity 0 () = ()=() that follows
from the equality 0 (0 ) = [R0 ((0 ))] 1 and a straightforward computation of
R0 (). Equation (0.9) suggests the following formal stochastic differential equation
for Yt : p
dYt = (1=2)0()Yt dt + ()rdWt :
We conclude this section sketching the strategy of the proof of Theorem 0.1.
Theorem 0.2 reduces the proof of Theorem 0.1 to the verification of three proper-
ties : (a) that the sequence of probability measures is tight, (b) that the restriction
to F0 of all limit points Q of the sequence QN are Gaussian fields with covariance
given by (0.8) and (c) that all limit points Q solve the martingale problem (0.6).
The first two properties are straightforward. To check that all limit points solve
the martingale problem, we consider a collection of martingales associated to the
empirical measure. For each smooth function G: Td ! R, denote by MtG and by
NtG the martingales defined by
Z t X
MtG = YtN (G) Y0N (G) N 2 LN N d=2 G(x=N )[s (x) ] ds ;
0
x2TdN
Z tn 2 o
NtG = (MtG )2 N 2 LN YsN (G) 2YsN (G)N 2 LN YsN (G) ds :
0
A simple computation permits to rewrite these martingales as
MtG = YtN (G) Y0N (G)

Z t X
(1=2)N d=2 (N G)(x=N )[g (s(x)) ()] ds ;
0
x2TdN
NtG = (MtG )2
Z t d X
X
(1=2)N d [g (s (x + ej )) + g (s (x))][(@uNj G)(x=N )]2 ds ;
0 j =1 x2TdN
(0:10)
where N stands for the discrete Laplacian and (@uNj G)(x=N ) is equal to
N fG((x P + ej )=N ) G(x=N )g for 1 j d. We took advantage
P here from the
fact that x (N G)(x=N ) = 0 to add the expression () x (N G)(x=N ) to the
martingale MtG in order to obtain the mean-zero cylinder function g ( (x)) ().
To prove that all limit points of the sequence QN solve the martingale problem
(0.6) it remains to close the equations in terms of the fluctuation field YtN . This
is easy for the martingale NtG : an elementary computation shows that for every
continuous function H , the L2 (PN ) norm of
Z t X
N d H (x=N )[g(s(x)) ()] ds
0
x2TdN
is bounded above by t2 N d < H; H > Var( ; g ) because N is a product

invariant measure. In particular, by the definition of the linear operator B, in the
limit N " 1, (MtG )2 kBH k22 t is a martingale.
To close the equation for the martingale MtH , we follow an approach proposed
by Rost (1983). For each Lipschitz cylinder function , denote by YtN; the –
fluctuation field defined by
X
YtN; (H ) = N d=2 H (x=N )[x (t ) ~()] :
x2TdN
Notice that the integral part of the martingale MtG is equal to
Z t
YsN;g ((1=2)N H )ds :
0
Since non conserved quantities fluctuates in a much faster scale than conserved
quantities, in the time scale where the density changes, the non conserved quantities
should average out and only their projection on the density fluctuation field should
persist in the limit. In substance, there should exists a constant C ( ) such that
Z t n o
ds YsN; (H ) C ( )YsN (H )
0
vanishes as N " 1 for every smooth function H . This is the content of the
Boltzmann–Gibbs principle stated in the next section, where we prove conver-
gence to 0 in L2 (PN ) of the above integral term with C ( ) = ~0 (). This con-
vergence and some elementary estimates ensure that in the limit Yt (G) Y0 (G)
Rt
0 Ys (AH )ds is a martingale, concluding the proof of the convergence of the den-
sity fluctuation fields to the stationary generalized Ornstein–Uhlenbeck process
satisfying (0.6).
1. The Boltzmann–Gibbs principle
We show in this section that the martingales MtG introduced just before (0.10) can
be expressed in terms of the fluctuation fields Yt . This replacement of the cylinder
function g ( (0)) () by 0 ()[ (0) ] constitutes the main step toward the
proof of the equilibrium fluctuations.
Theorem 1.1 (Boltzmann–Gibbs principle) For every cylinder Lipschitz function

, every continuous function G on Td and every t > 0,
Z t X 2
lim E N
N !1
dsN d=2 G(x=N )xV (s ) = 0;
0
x2TdN
where
1. The Boltzmann–Gibbs principle 293
V ( ) = () ~() ~0 ()[(0) ] :
Proof. We first localize the problem. Fix a positive integer K that shall increase
to 1 after N . For each N , we subdivide TdN in non overlapping cubes of length
K : let M = [N=K ], where [r] stands for the integer part of r, and denote by
fBj ; 1 j M dg non overlapping cubes of linear size K : for each j
Bj = yj + f1; : : :; K gd for some yj in TdN and Bi \ Bj = if i 6= j :
Denote by B0 the set of sites not included in one of the cubes Bi . By construction
the cardinality of B0 is bounded by dKN d 1 .
Recall from section 7.2 that s is the smallest cube centered at the origin
that contains the support of . Denote by Bio the interior of the cube Bi , i.e., the
sites x in Bi that are at a distance at least s from the boundary :
Bio = fx 2 Bi ; d(x; TdN Bi ) > s g :

We defined the interior Bio so that x is measurable with respect to ( (z ); z 2
Bi ) for all x in Bio . In particular, under N , x and y are independent for
x and y in the interior of distinct cubes. Let B o stand for the set of all interior
points of TdN and B 1 for its complement :
M d
[
Bo = Bio ; B 1 = TdN Bo :
i=1
The cardinality of B 1 is bounded by dN d fC ( )K 1
+ KN 1
g for some finite
constant C ( ) depending only on .
With the notation just introduced, we have that
X X
N d=2 G(x=N )xV () = N d=2 G(x=N )x V ()
x2TdN x2B 1
Md
X X
+ N d=2 [G(x=N ) G(yi =N )]x V ( ) (1:1)
i=1 x2Bio
Md
X X
+ N d= 2
G(yi =N ) x V ( ) :
i=1 x2Bio
We claim that the expected value of the L2 norm of the time integral of the first
two expressions on the right hand side vanishes as N " 1 and then K " 1.
To show that the first expression vanishes in the limit, apply Schwarz inequality
to bound the expected value by
h X 2 i
t2 EN N d=2 G(x=N )x V ()
x2B 1
because N is invariant. Since the cylinder function V has mean zero with respect
to the product measure N , the last expression reduces to
X h i
t2 N d G(x=N )G(y=N )EN x V ()y V ()
x;y2B 1
jx yj2s
that vanishes in the limit as N " 1 and then K " 1 because the cardinality of
B 1 is bounded above by dN d fC ( )K 1 + KN 1g and V belongs to L2(N )
(since is Lipschitz).
For similar reasons and because G is assumed to be continuous, the expectation
of the square of the time integral of the second expression on the right hand side
of (1.1) vanishes in the limit as N " 1.
For each 1 i M d , denote by i the configuration f (x); x 2 Bi g and by
LBi the restriction of the generator LN to the cube Bi :
X
(LBi f )( ) = (1=2) g((x))[f (x;y) f ()] :
x;y2Bi
jx yj=1
Consider a L2 ( ) cylinder function f measurable with respect to ( (x); x 2 B1 )
and denote by fi the translation of f that makes it measurable with respect to
((x); x 2 Bi ). By definition of the generator LBi , LBi fi is also measurable
with respect to the -algebra ( (x); x 2 Bi ).
By Proposition A1.6.1, for every t > 0,
Z t Md
X 2
EN ds N d=2
G(yi =N )LBi fi (i (s))
0 i=1D (1:2)
E
20 t VG;fN ; ( N 2LN ) 1 V N
G;f ;
where
Md
X
N ( ) =
VG;f N d=2
G(yi =N )LBi fi (i ) :
i=1
In this formula and below < ; > stands for the inner product in L2 (N ). By the
variational formula for the H 1 norm, the right hand side of the last expression
is equal to
n Z o
20 t sup 2 N ( )h( ) N (d ) N 2 < h; ( LN )h > ;
VG;f
h
where the supremum is taken over all functions h in L2 (N ). The linear term in
the previous supremum is equal to
Md Z
2N d=2 X G(yi =N ) LBi fi (i )h()N (d) :
i=1
1. The Boltzmann–Gibbs principle 295
R
Integrating by parts the expression LBi fi (i )h( )dN and applying Schwarz
inequality, we obtain that it is bounded above by
1
< ( LBi )fi ; fi > < ( L )h; h >
2
+
2
Bi
for every > 0. The summation over i of the second term is less than or equal to
( =2) < h; ( LN )h > because LBi is the restriction of the generator LN to Bi .
Therefore, taking = N 2+(d=2) jG(yi =N )j 1 , we obtain that the right hand side of
(1.2) is bounded above by
Md
G(yi =N )2 < ( LBi )fi ; fi > tKkGdNk12 < ( LB )f1 ; f1 >
X
d
2
20tN 2
1
i=1
because the dynamics is translation invariant and fi is defined as the translation
of f1 . The last expression vanishes as N " 1.
Up to this point we reduced the proof of the theorem to the proof that
lim inf lim
K !1 f N !1
Z t Md
X n X o2
EN ds N d=2 G(yi =N ) x V (s ) LBi fi (i (s)) =0;
0 i=1 x2Bio
where the infimum is taken over all L2 ( ) functions f measurable with respect
to ( (x); x 2 B1 ) and fi stands for the translation of f that makes it measurable
with respect to ( (x); x 2 Bi ).
By Schwarz inequality, the expectation appearing in the previous expression
is bounded above by
X Md hn X o2
t2 N d G(yi =N )2EN x V () LB f1(1 )
1
i=1 x2B1o
because the product measure N is invariant and translation invariant and because
the support of x V LBi fi , y V Lj fj are disjoints for x in Bio , y in Bjo ,
i =6 j . As N increases to infinity, this expression converges to
hn X o2 i
t2 K dkGk22 EB 1 x V LB f1
1 : (1:3)
x2B1o
Recall that LB1 stands for the restriction to B1 of the generator LN . Denote
by R(LB1 ) the range of the generator LB1 in L2 (B1 ), i.e., the space generated
by LB1 f , f in L2 (B1 ) and by R? (LB1 ) the space orthogonal to R(LB1 ). Fix a
((x); x 2 B1 ) measurable function h in L2 (B1 ). The formula
h i
inf EB fh LB f g2
f 2L2 (B1 )
1 1
corresponds to the projection of h on R? (LB1 ). Denote by B1 ;L the space of

all configurations of N B1 with total number ofPparticles equal to L, by B1 ;L the
restriction of B1 to B1 ;L : B1 ;L () = B1 ( j x2B1 (x) = L) and by B 0
1 ;L
the
space of L (B1 ;L )-mean zero functions. B1 ;L has codimension 1 and R(LB1 ) is
2 0
a subset of B 0
1 ;L
because B1 ;L is invariant for the dynamics generated by LB1 .
On the other hand, the kernel of LB1 in L2 (B1 ;L ) reduces to the constant functions
since LB1 f = 0 implies that < f; LB1 f >= 0 that in turn forces f to be constant. In
particular, the dimension of Ker LB1 (and thus the codimension of R(LB1 )) is equal
to 1. Therefore R(LB1 ) = B 0
1 ;L
because R(LB1 ) B 0
1 ;L
and the codimension
of the latter space is equal to 1. This shows that on L2 (B1 ;L ) R? (LB1 ) is the one-
dimensional space of constant functions. Thus R? (LB1 ) consists of all functions
that depend on the configuration only through its total number of particles. In
particular, the infimum over all f in L2 (B1 ) of the expression (1.3) is equal to
hn h X io2 i
t2 K d kGk22 EB EB 1 1 x V B (y1 )
1
: (1:4)
x2B1o
In this formula B1 (y1 ) stands for the average
P number of particles of the config-
uration on the cube B1 : B1 (y1 ) = K d x2B1 (x). For x in B1o , denote by
~K (B1 (y1)) the conditional expectation of x with respect to B1 (y1) :
h i
~K (B (y1 ))EB x B (y1 )
1
= 1
1
and notice that this expression does not depend on x because is homogeneous.
With this notation, we may rewrite (1.4) as
t2 jBK1dj kGk22 EB
o2 hn o2 i
1 ~K (B (y1)) ~() ~0 ()[B (y1 ) ]
1 1
:
In Corollary A2.1.7 we prove that the absolute value of the difference between
~K () and ~() is bounded above by C ( )K d on all compact sets of R+ . We
may therefore estimate the previous expectation by
hn o2 i
2t2 K d kGk22 E B1 ~K (B (y1 )) ~(B (y1))
1 1

hn o2 i
+ 2t2 K d kGk22 E B1

~(B (y1 )) ~() ~0 ()[B (y1) ]
1 1
:
Since is Lipschitz, j ~K ()j and j ~()j are bounded by C ( )(1 + ). We may
thus introduce the indicator function 1f B1 (y1 ) Ag inside both expectations.
By Corollary A2.1.7, the first one is bounded above by C (A; ; )K 2d . On
the other hand, by Taylor’s expansion up to the second order and since is
a product measure, the second expectation is also bounded by C (A; ; )K 2d .
This concludes the proof of the Boltzmann–Gibbs principle.
2. The martingale problem 297
2. The martingale problem
In section 3 we prove that the sequence of probability measures QN is tight and

that all limit points are concentrated on continuous paths. In view of Theorem
0.2, to conclude the proof of the equilibrium fluctuations, it remains to show that
all limit points Q of the sequence QN solve the martingale problem (0.6) and to
characterize their restriction to F0 . We start with the latter question which is easier.
Fix a limit point Q and assume without loss of generality that QN converges to
Q .
Lemma 2.1 For every continuous function H : Td ! R and every t > 0,
h
lim log E N expfiYt (H )g
i
=
() < H; H > :
N !1 2
Proof. Since N is a product invariant measure,

h i X h i
log E N expfiYt (H )g = log EN exp iN d=2H (x=N )[(0) ] :
x2TdN
(2:1)
Since by assumption (2.3.2), (0) has a finite exponential moment, by Taylor
expansion, last expectation is equal to
() H (x=N )2 O(N 3d=2 ) :
2N d
1 +
In this formula () stands for the static compressibility Var( ; (0)). Therefore,
the right hand side of (2.1) is equal to
() N d X
H (x=N )2 + O(N d=2 ) :
2
x2TdN
As N " 1 this expression converges to (1=2)() < H; H >, what concludes
the proof of the lemma.
Corollary 2.2 Restricted to F0 , Q is a Gaussian field with covariance given by
EQ [Y0 (G)Y0 (H )] = () < H; G > : (2:2)
Proof. Fix a positive integer n, in Rn and H1 ; : : : ; Hn in Hk0 . Since Y0 is linear

and since, by assumption, QN converges weakly to Q , by the previous lemma,
h n
X i h n
X i
log EQ exp i j Y0 (Hj ) = lim E N exp iY0 j Hj
j =1 N !1 j =1
() < n
X n
X
=
2
j Hj ; j Hj >
j =1 j =1
The Q joint distribution of (Y0 (H1 ); : : : ; Y0 (Hn )) is thus Gaussian with covariance
given by (2.2), what concludes the proof of the lemma.
We turn now to the dynamic part of the problem.
Proposition 2.3 Q solves the martingale problem (0.6).

Proof. Fix H in C 2 (Td) and denote by MtA;H , NtA;H the random processes
defined by
Z t
A ;H
Mt = Yt (H ) Y0 (H ) ds Ys (AH )
0
and NtA;H = (MtA;H )2 kBH k22t :
By definition, MtA;H is Ft –measurable. Thus, in order to prove that MtA;H is a
martingale, we just need to check that
h i h i
EQ MtA;H U = EQ MsA;H U
for all 0 s t T and U of the form U = 1fYsi (Hi ) 2 Ai ; 1 i ng,
where n is a positive integer, 0 s1 sn s, Hi are in C 2 (Td ) and Ai
are measurable subsets of Rd for 1 i n.
Recall from the introduction that for each H in C 2 (Td ), the process MtH
defined by
Z t X
MtH = YtN (H ) Y N (H )
0 N 2 LN N d=2 H (x=N )[s(x) ] ds
0
x2TdN
is a martingale so that E N [MtH U ] is equal to E N [MsH U ]. To conclude the proof of
the first statement of the proposition it remains to show that these two expectations
converge respectively to EQ [MtA;H U ] and EQ [MsA;H U ].
By the Boltzmann–Gibbs principle, since U is bounded,
h i
lim E N
N !1
MtH U
hn Z t o i
= lim E N Yt (H ) Y0 (H ) (1=2)0() ds Ys (N H ) U :
N !1 0
A simple computation shows that the square of the expression inside braces on
the right hand side of the equality has uniformly (in N ) bounded expectation.
In particular, since U is bounded and QN converges weakly to Q , which is
concentrated on continuous paths, E N [MtH U ] converges to EQ [MtA;H U ]. Since
the same argument applies to the expectation E N [MsH U ], the first statement of
the proposition is proved.
The argument that shows that the process NtA;H is a martingale is similar
to the one presented above to prove that MtA;H is a martingale. It relies on the
3. Tightness 299
martingale NtH introduced in (0.10), on the Boltzmann–Gibbs principle and on the

fact that the martingale MtH has uniformly bounded fourth moments. We leave
the details to the reader.
3. Tightness
We prove in this section that the sequence of probability measures QN is tight

and that all limit points are concentrated on continuous paths. We first review
some aspects of the uniform weak topology on D([0; T ]; H k ) introduced in the
beginning of the chapter. Throughout this section k stands for a positive integer
larger than 2 + (d=2).
For > 0 and a path Y in D([0; T ]; H k ), define the uniform modulus of
continuity w (Y ) by
w ( Y ) = sup kYt Ys k k :
js tj
0s;tT
The first result provides sufficient conditions for a subset to be weakly relatively
compact.
Lemma 3.1 A subset A of D([0; T ]; H k ) is relatively compact for the uniform

weak topology if
(i) sup sup kYt k k <1
Y 2A 0tT
lim sup w (Y ) = 0 :
!0 Y 2A
(ii)
The proof of this result is standard and left to the reader. From this lemma we
deduce a criterion for tightness of a sequence of probability measures PN defined
on D([0; T ]; H k ).
Lemma 3.2 A sequence fPN ; N 1g of probability measures defined on

D([0; T ]; H k) is tight provided for every 0 t T ,
h i
lim lim sup PN
A!1 N !1
sup kYt k k > A = 0
tT
0
and h i
lim lim sup PN
!0 N !1
w ( Y ) " = 0
for every " > 0.

We have now all elements to prove the tightness of the sequence QN introduced
in the beginning of the chapter.
Proposition 3.3 The sequence of probability measures QN is tight. Moreover, all

limit points are concentrated on continuous paths.
The proof of this proposition is divided in several lemmas. Following Holley

and Stroock (1978), we start with a key estimate. For each z in Zd, denote by
Mtz and Ntz the martingales Mthz and Nthz introduced just before (0.10). To keep
notation simple let
X
z (s) = (1=2)N d=2 (N hz )(x=N )[g (s(x)) ()]
1
x2TdN
d
z (s) = (1=2)N d X X [(@ N hz )(x=N )]2[g (s (x)) + g (s (x + ej ))]
2 uj
j =1 x2TdN
Rt Rt
so that Mtz = YtN (hz ) Y0N (hz ) 0
z (s)ds and N z = (M z )2
1 t t 0
z (s)ds.
2
Lemma 3.4 There exists a finite constant C (; T ) depending only on and T

such that for every z in Zd,
h i
lim sup E N sup j < Yt ; hz > j2
N !1 0tT
n o
C (; T ) < hz ; hz > + < hz ; hz > :
In this formula and below < Yt ; hz > stands for the inner product of Yt 2 H k
and hz 2 Hk .
Rt
Proof. Rewrite < YtN ; hz > as Mtz + Y0N (hz ) + 0 1z (s)ds. A straightforward
computation shows that the limit, as N increases to 1, of E N [j < Y0 ; hz > j2 ]
is equal to () < hz ; hz > because N is a product measure.
On the other hand, since Mz is a martingale, by Doob inequality,
h i h i
EN sup jMtz j2 4 EN jMTz j2 :
0 tT
By definition of the martingale Nz , the right hand side is equal to
hZ T X X i
4 EN ds (1=2)N d [g (s (x)) + g (s (x + ei ))](@uNi hz (x=N ))2 ;
0 id x2TdN
1
where @uNi hz (x=N ) stands for NR[hz ((x + ei )=N ) hz (x=N )]. As N " 1, this
expression converges to 4T() dukrhz (u)k22 , where krhz (u)k2 is the L2 (Td )
norm of the gradient of hz .
Finally, by definition of hz and Schwarz inequality,
3. Tightness 301
h Z t 2 i
EN sup ds 1z (s)
tT
0 0
hZ T X 2 i
T EN ds (1=2)N d=2 (N hz )(x=N )[g (s(x)) ()] :
0
x2TdN
Since the initial measure N is product and invariant, as N " 1, this expression
converges to (1=4)T 2 < hz ; hz > Var( ; g ), what concludes the proof of the
lemma.
Corollary 3.5 For k > 2 + (d=2),
h i
(a) lim sup E N sup kYt k2 k < 1
N !1 tT
0
h X i
(b) lim lim sup E N
n!1 N !1 sup ( < Yt ; hz >)2 z k = 0:
0tT jzjn
Proof. The first expression is bounded above by

X h i
z k E N sup (< Yt ; hz >) 2 :
z2Zd 0tT
By the previous lemma, the limsup as N " 1 of this sum is less than or equal to
X n o
C (; T ) z k 1+ < hz ; hz >
z2Zd
for some finite constant C (; T ) depending only on and T because hz has
L2 (Td) norm equal to 1. By definition of hz and z , this expression is equal to
X 1 + (2 kz k)4 X
C (; T ) C (; T ) 1

z2Zd (1 + (2 kz k) ) kz k)2)k
2 k (1 + (2 2
z2Zd
This estimate proves the first statement. The second claim follows by the same
argument.
It follows from Lemma 3.2 and Corollary 3.5 that in order to prove that the
sequence QN is tight, we only have to show that for every " > 0,
h i
lim lim sup PN
!0 N !1
w ( Y ) > " = 0:
In view of part (b) of the previous corollary, this result follows from the following
lemma :
Lemma 3.6 For every positive integer n and every " > 0,
h X i
lim lim sup PN
!0 N !1
sup ( < Yt Ys ; hz >)2 z k > " = 0:
0js tj jzjn
0s;tT
Proof. To prove this lemma we just have to show that

h i
lim lim sup PN sup ( < Yt Ys ; hz >)2 > " = 0
!0 N !1 0js tj
0s;tT
for every z in Zd and " > 0. Fix z in Zd and Rrecall the definition of the martingale
M z . Since < Yt ; hz >=< Y0 ; hz > +Mtz + 0t 1z (s)ds, the lemma follows from
the next two results.
Lemma 3.7 Fix a function G in C 2 (Td). For every " > 0,
h i
lim lim sup PN
!0 N !1
sup jMtG MsG j > " = 0:
0s;tT
jt sj
Proof. Denote by w0 (M G) the modified modulus of continuity defined as

w0 (M G) = finf max
t g 0i<r
sup jMtG MsGj ;
i ti s<t<ti+1
where the first infimum is taken over all partitions of [0; T ] such that
(
< t1 < < tr = T
0 = t0
ti+1 ti > 0 i < r :
Since supt jMtG MtG j = supt j < Yt ; G > < Yt ; G > j is bounded by
C (G)N (1+(d=2)) and
w (M G) 2w0 (M G) + sup jMtG MtG j ;
t
in order to prove the lemma we just need to show that
h i
lim lim sup PN
!0 N !1
w0 (M G) > " = 0 (3:1)
for every " > 0.

By Proposition 4.1.6, to prove (3.1) it is enough to check that for every " > 0,
h i
lim lim sup sup PN
!0 N !1 2TT
jMG+ MGj > " = 0;
0
where TT stands for all stopping times bounded by T . By Chebychev inequality,

the last probability is less than or equal to
3. Tightness 303
h i h i
E N (MG+ MG)2 E N (MG+ )2 (MG )2
1 1
"2 =
"2
because MtG is a martingale and a bounded stopping time. By (0.10) this ex-
pression is bounded above by
h Z Xd X
1
2"2
E N ds N d [g (s (x + ej )) + g (s (x))]
0 j =1 x2TdN
i
[N fG((x + ej )=N ) G(x=N )g]2
because N is invariant, a stopping time and is bounded above by . The limit
as N " 1 of this last expression is less than or equal to " 2 krGk22 (), what
Rt
We now turn to the additive functional 0 N 2 LN < YsN ; G > ds. From (0.10),
this expression is equal to
Z t X
1
2
ds N d=2 (N G)(x=N )[g (s (x)) ()]
0
x2TdN
Lemma 3.8 Fix a function G in C 2 (Td). For every " > 0,

lim lim sup
!0 N !1
h Z t X i
PN sup

dr N d=2 (N G)(x=N )[g (r (x)) ()] > " = 0:
0s;tT s x2TdN
jt sj
Proof. By Chebychev and Schwarz inequality, the probability is bounded above

by
E h Z T dr N d=2 X ( G)(x=N )[g( (x)) ()]2 i
"2 N 0 x2TdN
N r
because jt sj and s, t T . Since N is an invariant product measure and

g((x)) () has mean zero, this expression is equal to
T N d X [ G(x=N )]2E N hfg((x)) ()g2 i
"2 N
x2Td N
that vanishes as N " 1 and then # 0.
This concludes the proof of the tightness of the sequence QN .
We conclude this section showing that tightness can be proved in a stronger
norm in the case where the jump rate g () is bounded. The argument applies to
reversible nongradient systems and may be skipped without prejudice by those

who are satisfied with convergence in D([0; T ]; H k ) for k > 2 + (d=2).
We shall prove that there exists a finite constant C (; T ) depending only on
and T such that for every z in Zd,
h i
lim sup E N sup j < Yt ; hz > j2
N !1 0 tT (3:2)
n o
C (; T ) < hz ; hz > + < ( )hz ; hz > :
The reader should compare this estimate with Lemma 3.4 and check that Corollary
3.4 with k > 1 + (d=2) instead of k > 2 + (d=2) follows from estimate (3.2). In
particular, by Lemma 3.6, the sequence QN introduced in the beginning of the
chapter is tight (and therefore, by section 2, converges) in D([0; T ]; H k ) for
k > 1 + (d=2).
In view of the proof of Lemma 3.4, in order to deduce estimate (3.2), we just
need to show that
h Z t i
lim sup E N sup
z (s) ds2 C (; T ) < hz ; ( )hz > : (3:3)
N !1 tT
1
0 0
For x in Zd and 1 i d, denote by Wx;x+ei the current over the bond

fx; x + ei g. In the case of the symmetric nearest neighbor zero range process,
the current is Wx;x+ei = (1=2)[g ( (x)) g ( (x + ei ))]. Notice that 1z (s) may be
expressed in terms of the current :
X
z (s) = N d=2 [@uNj hz (x=N )]NWx;x+ej (s) : (3:4)
1
x;j
The proof of (3.3) is divided in two steps. We first obtain an exponential
estimate of the current fluctuation fields and then apply the Garsia–Rodemich–
Rumsey inequality to deduce (3.3).
Lemma 3.9 For every a > 0, continuous functions Gi , i = 1; : : : ; d and 0 s < t,

h n Z t X oi
EN exp

a dr N d=2 NWx;x+ei (r)Gi (x=N )
s x;i
n o
2 exp kgk1(t s)a2N d X Gi (x=N )2
x;i
for all N large enough.
The proof of this lemma follows closely the one of Lemma 7.6.2. In sake of
completeness, we prove it again in the gradient context of zero range processes
with bounded rates.
Proof. Since ejxj ex + e x, it is enough to show that

3. Tightness 305
h n Z t X oi
E N exp a dr N d=2 NWx;x+ei (r)Gi (x=N )
s x;i
n o
exp kgk1(t s)a2N d X Gi (x=N )2
x;i
for all N large enough and all sets of continuous functions Gi , i = 1; : : : ; d. Fix
such functions. By Feynman–Kac formula and by stationarity of ,
n Z t o
X
E N exp a N d=2 Gi (x=N )NWx;x+ei (r)dr
s x;i
n Z t s o
X
= E N exp a N d=2 Gi (x=N )NWx;x+ei (r)dr e(t s)N (G) ;
0 x;i
where P
N (G) is the largest eigenvalue of the symmetric operator N 2 LN +
aN d=2 x;i Gi (x=N )NWx;x+ei . By inequality (A3.1.1),
n X o
N (G) sup aN d=2 Gi (x=N ) < NWx;x+ei f > N 2 DN (f ) ;
f x;i
where the supremum is carried over all densities f and where < > stands for
expectation with respect to N .
Recall that for symmetric nearest neighbor zero range processes, the current
takes the form Wx;x+ei = (1=2)[g ( (x)) g ( (x + ei ))]. A change of variables
permits to rewrite < Wx;x+ei f > as
Z h i
(1=2)() f ( + dx ) f ( + dx+ei ) (d) :
p p p p
Since a b = ( a b)( a + b), by Schwarz inequality, last expression is
bounded above by
() Z p
f ( + dx )
p
f ( + dx+ei ) (d)
2
4
Z
() pf ( + d ) + pf ( + d )2 (d)
+
4
x x+ei
for every > 0. The first term is just a multiple of the piece of the Dirichlet form
corresponding to jumps over the bond fx; x + ei g and is equal to (2 ) 1 Ix;x+ei (f ).
The second one, by a new change of variables, is bounded above by kg k1
because the jump rate g () is boundedPand f is a density with respect to . From
this estimate it follows that aN d=2 x;i Gi (x=N ) < NWx;x+ei f > is bounded
above by
X X
(2 ) 1
Ix;x+ei (f ) + kgk1a2 N 2 d Gi (x=N )2
x;i x;i
for every density f . Choosing = N 2

, we conclude that the largest eigenvalue
N (G) defined above is bounded by
kgk1a2N d X Gi (x=N )2 ;
x;i
what proves the lemma.
Lemma 3.10 There exists a finite constant C (; T ) depending only on and T
such that for each z in Zd,
h Z t i
lim sup E N

sup z (s)ds2 C (; T ) < krhz k22 >
N !1 tT 1
0 0
Proof. Recall from section 7.6 the statement of the Garsia–Rodemich–Rumsey

inequality. Fix a0 > 0 and set
Z t
g ( t) = z (s)ds ; (u) = expfa0 N d ug 1:
1
0
By the just mentioned inequality with p(u) =

pu, we have that
Z T 4B p1 du ;
sup jg(t)j 4 1
u2 u
0 tT 0
where Z T Z T (t) g (s)j

jgp
B ds dt :
j t sj
=
0 0
With our choice of , 1

(u) = (a0 N d ) 1
log(1 + u). Integrating by parts we get
p
that Z T n 4B o 1 n 4B o
0
log 1 +
u2 pu du 2 T log 1 +
T2 + 4
because B=(4B + u2 ) 1=4. Therefore, by the two previous estimates,

na Nd
p sup jg(t)j E N exp log 1 + 4TB2 + 4
h oi h n oi
0
E N exp
8 T 0tT
n o
= e4 1 + 2 E N [B ] :
4
T
Recall the definitionpof B , the explicit formula (3.4) for 1z (r) and apply Lemma
3.9 with a = a0 N = jt sj and Gj = @uNj hz , to obtain that the previous expres-
d
sion is less than or equal to
4. Generalized Ornstein–Uhlenbeck processes 307
n Z T Z T n X o
e4 1 4+
T
8
2
ds dt exp a20 kgk1N d [(@uNj hz )(x=N )]2
0 0 x;j
n X o
8e4 exp a20kgk1N d [(@uNj hz )(x=N )]2 :
x;j
In conclusion, we have proved that
h n a0pN d sup jg(t)j a2 kgk N d X[(@ N h )(x=N )]2oi 8e4 :
E N exp 1 uj z
8 T 0tT
0
x;j
Maximizing over a0 , we get that
2
h n sup0tT jg(t)j oi
E N exp
28 kg k1 TN d P 8e4 :
x;j [(@uNj hz )(x=N )]2
In particular, by Jensen inequality,
h 2 i
EN jg(t)j 28 (4 + log 8)kg k1TN d X[(@ N hz )(x=N )]2 ;
sup uj
0 tT x;j
4. Generalized Ornstein–Uhlenbeck processes
We conclude this chapter proving Theorem 0.2, a particular case of a general result
due to Holley and Stroock (1978). The reader should notice that the proof below
does not rely on the special form of the operators A and B.
By Ito’s formula and (0.6), for each fixed s 0 and H in C 1 (Td),
fXt (H ); t 0g defined by
s
n Z s_t o
Xts (H ) = exp i Yt_s (H ) Ys(H ) Yr (AH )dr + (1=2)kBH k22(t s)+
s
is a martingale (in fact, as already mentioned in the beginning of the chapter,
kBH k2 1MtA;H is a Brownian motion for each H in C 1 (Td)).
We now claim that for each S > 0 and smooth function H in C 1 (Td),
H;S
fZt ; t 0g defined by
n Z t^S o
ZtH;S = exp iYt^S (H(S t) ) + (1=2) + kBHS r k22dr
0
is a martingale. Here, for t > 0, Ht stands for Tt H .

To prove this claim, fix 0 t1 < t2 S and set sn;j = t1 + (j=n)(t2 t1 )
for 0 j n. It is easy to check that the function defined on R2+ that associates
to each (s; t) the value Ys (Ht ) is continuous because fYs ; 0 s T g is weakly

compact on H k and Tt H is uniformly continuous. It follows from this continuity
and from the expansion Ht+" = Ht + "AHt + o(") that
nY1
Xssn;j
n;j (H
+1 S sn;j )
j =0
converges a.s. and in L1 (Q) to ZtH;S =ZtH;S as n " 1.
Let G be a bounded Ft1 –measurable function. Since the convergence takes
2 1
place in L1 (Q),
h Z H;S i h nY1 i
EQ tH;S G2
= lim
n!1 EQ Xssn;j
n;j (H
S sn;j ) G :
Zt 1 j =0
+1
Taking conditional expectation with respect to Fsn;n 1 , we reduce the range of the
product in last expectation to 0 j n 2 because Xs (H ) is a martingale for
each s 0 and smooth H . repeating this argument, we obtain that
h Z H;S i
EQ tH;S G 2
= EQ [G]
Zt 1
what proves that fZtH;S ; t 0g is a L1(Q) martingale.

In particular,
h i n Z t o h i
EQ eiYt (H ) Fs = exp (1=2) kBHt r k22 dr EQ ZtH;t Fs :
0
Since ZH;t is martingale, the conditional expectation on the right hand side of last
formula is equal to ZsH;t . Therefore, a change of variables gives that
h Z t s
i n o
EQ eiYt (H ) Fs = exp iYs (Ht s ) (1=2) kBHr k22 dr :
0
This equation states that conditionally to Fs , Yt (H ) has a Gaussian distribution of

Rt s
mean Ys (Ht s ) and variance 0 kBHr k22 dr. This is precisely (0.7). A standard
Markov argument guarantees then the uniqueness of finite dimension distributions,
which in turn gives the uniqueness of Q.
The first rigorous result in equilibrium fluctuations was obtained by Martin–Löf

(1976) for a superposition of independent Markov processes on Rd . Due to the ab-
sence of interaction between particles, the hydrodynamic equation is linear and the
martingales introduced in (0.10) are functions of the density field. With the present
techniques, the equilibrium fluctuations follows therefore from the Holley–Stroock
theory of generalized Ornstein–Uhlenbeck processes and some compactness argu-
ments.
To prove the equilibrium fluctuations of interacting systems, Rost (1983) intro-
duced the Boltzmann–Gibbs principle described in section 1. Brox and Rost (1984)
proved the validity of the principle for attractive zero range processes : they showed
that for a fixed density , in the Hilbert space generated by theP –mean zero cylin-
der functions and the inner product defined by g; h = x2Zd < x g; h > ,
the semigroup St , as t " 1, acts as a projection on the space generated by the
cylinder function associated to the conserved quantity :
lim
t!1 g; St h =
() g; (0) h; (0)
1
From this result they deduced the Boltzmann–Gibbs principle and, in Rost (1985),
the equilibrium fluctuations as stated in this chapter.
The Boltzmann–Gibbs principle was extended by De Masi, Presutti, Spohn and
Wick (1986) for exclusion processes with speed change (from which they deduced
the equilibrium fluctuations for gradient lattice gases), by Spohn (1985, 1986,
1987) for interacting Brownian motions and by Zhu (1990) for one-dimensional
Ginzburg–Landau lattice models.
Landim and Vares (1994) proposed an alternative proof of the Boltzmann–
Gibbs principle in dimension 1 based on a p superexponential replacement lemma,
at the fluctuations level, for blocks of size " N . The proof we present here of the
Boltzmann–Gibbs principle is due to Chang (1994) and Chang and Yau (1992).
It was extended to nongradient Ginzburg–Landau lattice models by Lu (1994)
and to nongradient generalized exclusion processes by Chang (1995) and Sellami
(1998). Gielis, Koukkous and Landim (1997) proved the equilibrium fluctuations
for symmetric zero range processes in random environment.
Bertini et al. (1994) prove, for one-dimensional stochastic Ising dynamics with
a Kac potential at the critical temperature, that the fluctuation field correctly renor-
malized converges in distribution to the solution of a stochastic partial differential
equation obtained by adding a white noise to a Ginzburg–Landau equation. This
analysis is extended in Fritz and Rüdiger (1995) to infinite volume for a wider
class of initial states and temperatures close to the critical one.
The nonequilibrium fluctuations are a much less understood question and con-
stitutes one of the main open problems in the theory of hydrodynamic limit of
interacting particle systems. Until now only partial results for gradient models
are known. Comets and Eisele (1988) prove the hydrodynamic limit and the
nonequilibrium large deviations for a non conservative mean field stochastic Ising
model. Ravishankar (1992a) proves the nonequilibrium fluctuations for symmet-
ric simple exclusion process in any dimension. De Masi, Presutti and Scacciatelli
(1989), Dittrich and Gärtner (1991) prove nonequilibrium fluctuations for the one-
dimensional nearest neighbor weakly asymmetric simple exclusion process. Rav-
ishankar (1992b) extends this result to 2-dimensional weakly asymmetric simple
exclusion processes. Ferrari, Presutti and Vares (1988) proved a nonequilibrium
version of the Boltzmann–Gibbs principle for symmetric simple exclusion pro-
cesses in dimension 1 and extended the result for the one-dimensional nearest
neighbor symmetric zero range process with jump rate given by g (k ) = 1fk 1g
starting from a local equilibrium. These ideas were applied to a superposition
of Kawasaki and Glauber dynamics by De Masi, Ferrari and Lebowitz (1986)
and to particles systems with unbounded spins associated to reaction–diffusion
equations by Boldrighini, De Masi and Pellegrinotti (1992). Later, Chang and
Yau (1992), using a logarithmic Sobolev inequality extended to one-dimensional
Ginzburg–Landau lattice models with strictly convex potentials Chang’s proof of
the Boltzmann–Gibbs principle to the nonequilibrium setting.
A clear presentation of Brox and Rost proof of the Boltzmann–Gibbs principle
can be found in De Masi, Ianiro, Pellegrinotti and Presutti (1984) or in Spohn
(1991).
The tightness argument and the theory of generalized Ornstein–Uhlenbeck
processes presented in this chapter are taken from Holley and Stroock (1978)
and Chang (1994). These ideas were successfully applied in Holley and Stroock
(1979a,b) to investigate the equilibrium fluctuations of non conservative spin flip
dynamics.
Appendix 1. Markov Chains on a Countable Space
We present in this chapter an overview on continuous time Markov chains on

countable state spaces. We refer the reader to Kemeny, Snell and Knapp (1966),
Breiman (1968), Gikhman et Skorohod (1969) and Ethier and Kurtz (1986) for
a detailed and comprehensive exposition of general properties of discrete and
continuous time jump Markov chains.
The first two sections are devoted to the construction of a continuous time
Markov process on a countable state space and to the investigation of the basic
properties of the underlying discrete time skeleton chain. At the end of the second
section we compute the Radon–Nikodym derivative between two jump Markov
processes. In sections 3 and 4 the basic tools in the theory of Markov processes
are introduced : semigroups, generators, adjoint and reversible processes.
In section 5 we introduce a class of martingales in the context of Markov
processes. With this collection of martingales we derive a bound in section 6
for the variance of additive functionals of Markov processes and we prove the
Feynman–Kac formula in section 7. This Feynman–Kac formula permits to com-
pute explicitly the Radon–Nikodym derivative of a time inhomogeneous Markov
process with respect to another, generalizing the formula obtained in section 2.
In section 8 we review the elementary properties of the relative entropy of a
probability measure with respect to a reference measure. The explicit formula for
the relative entropy permits in section 9 to show that in the context of Markov
processes the entropy of the state of the process with respect to an invariant
measure does not increase in time. In fact, we show that the time derivative of
this entropy is bounded above by the Dirichlet form. This estimate leads us in
section 10 to examine the main properties of the Dirichlet form and in section 11
to prove a maximal inequality for reversible Markov processes.
1. Discrete time Markov chains
Throughout this chapter E stands for a countable state space. The elements of E
are denoted by the last characters of the alphabet.
Let p: E E ! R+ be a transition probability :
X
p(x; y) 0 and p(x; y) = 1
y2E
312 Appendix 1. Markov Chains on a Countable Space
for every x 2 E . Denote by

= E N the path space endowed with the Borel
-algebra B = B(E ) and by ! the elements of
. For n 0 let Xn :
! E be
the state of the chain at time n :
Xn ( ! )
!n =
and let n :
!
be the time translation by n units :

n (!) j = !n+j
for all j 0.
Proposition 1.1 For each x in E , there exists a unique probability measure on
(
; B ), denoted by Px , such that
Px [X0 = x0 ; X1 = x1 ; : : : ; Xn = xn ] = x;x p(x0 ; x1 ) p(xn 1 ; xn )
0
for every n 0. In this formula x;y stands for the delta of Kronecker. Moreover, if
Ex stands for the expectation with respect to Px , for every bounded, B–measurable
function f ,

Ex f n j X0 ; : : : ; Xn = Ex f n j Xn = EXn f :
Proof. The existence of the probability measure Px follows from Kolmogorov’s

theorem (cf. Corollary 2.19
S of Breiman (1968)). We just need to check that this pre-
probability defined on n0 (X0 ; : : : ; Xn ) may be extended to B . The formula
for the conditional expectation follows from an elementary computation.
The first identity of the second statement of the proposition establishes that the
behavior of the Markov chain in the future depends on the past only through the
present or, equivalently, that conditioned on the present, the past and the future
are independent. This result explains the following definition.
Definition 1.2 Let p: N E E ! [0; 1] be a collection of transition probabilities.

A sequence of random variables fXn ; n 0g defined on a probability space
(
; A; P ) and taking values on a countable space E is a Markov chain with
transition probability p if for every n 0,
h i h i
P Xn+1 = y X0 = x0 ; : : : ; Xn = xn = P Xn+1 = y Xn = xn
= p(n; xn ; y)
for every (x0 ; : : : ; xn ; y ) in E n+2 . The Markov chain is said to be homogeneous if
the transition probability p does not depend on n, i.e., if there exists a transition
probability p: E E ! [0; 1] such that
h i
P Xn+1 = y Xn = x = p(x; y)
1. Discrete time Markov chains 313
for every (x; y ) in E E and every n 0.

The second property imposes the process to be time translation invariant in
the following sense. The probability, for a process starting from x at time 0, to be
at state y at time n is equal to the probability, for a process that is at x at time
m, to be at y at time m + n.
If is a probability measure on E , we denote by P the probability measure
on the path space
for a process whose initial position is distributed according
to : X
P [ ] := ( x) P x [ ] :
x2E
Hereafter E stands for expectation with respect to P .
In order to investigate the equilibrium states of a Markov chain, in the space
of bounded measurable functions on E , denoted by Cb (E ), we introduce the
operator P defined by
X
(Pf )(x) = p(x; y) f (y) :
y2E
Notice that for every f in Cb (E ) and every probability measure on E ,
E [f (Xn)] = < ; P nf > ;
where P n stands for the n-th power of P and < ; g > for the integral of a
bounded function g with respect to .
We endow Cb (E ) with its natural topology, the topology of the pointwise
bounded convergence : a sequence (fj )j 1 of bounded functions converges bound-
edly pointwise to f if it converges pointwisely and remains uniformly bounded :
lim f (x) = f (x) for every x 2 E ;

j!1 j

lim sup sup fj (x)

< 1:
j!1 x2E
By duality we may extend the operator P to the space of probability measures
on E , denoted by M1 (E ). In this way, for in M1 (E ), P stands for the
probability measure defined by
X
(P ) (y ) = (x) p(x; y) :
x2E
In this context to find a probability measure invariant under the evolution, i.e., a
probability measure under which X0 and Xn have the same distribution for every
n 0, we need to look for a solution of
< P; f > = < ; f >
for every bounded function f .
Of course, there exists always a solution to this problem in the case where E
is finite. It is enough, for instance, to examine the action of P on the compact and
convex set M1 (E ) or to take any limit point of the sequence
N 1
1 X i
N i=0 P
where is any probability measure.
In the countable case an invariant probability measure may not exist. For
example, in the case of a nearest neighbor symmetric random walk on Z. However,
if there is a state x such that the expectation of the first return to x is finite under
Px , then there exists an invariant probability measure (cf. Definition 7.26 and
Theorem 7.34, Breiman (1968)). Such state x is said to be positive recurrent.
The next natural question in the investigation of a Markov chain is the problem
of the uniqueness of an invariant probability measure. If the Markov chain is
indecomposable, i.e., if there are no two sets A and B with A \ B = and such
that starting from any site x in A (resp. B ) the chain remains in A (resp. B ), there
is at most an invariant probability measure (cf. Theorem 7.16, Breiman (1968)).
2. Continuous time Markov chains
Let E be a countable space, : E ! (0; 1) a bounded function and p a transition

probability on E that vanishes on the diagonal : p(x; x) = 0 for every x in E .
Consider the space
= (E (0; 1))N endowed with the Borel -algebra that
makes the variables (n ; n ) measurable. For each x in E , let Px be the probability
measure under which
(a) n is a Markov chain with transition probability p starting from x,
(b) Given the sequence (n )n0 , the random variables n are independent and
distributed according to an exponential law of parameter (n ).
f 2
Notice that the conditional distribution of the vector (0 ; : : : ; n ) given k ; k
g
N depends only on (0 ; 1 ; : : : ; n ) and therefore that the conditional distribution
of (0 ; : : : ; n ) given (0 ; 1 ; : : : ; n ) is still
n
Y
(i ) e (i )ui 1fui > 0g dui :
i=0
The next result follows easily from this remark.
Proposition 2.1 For n 0, let

T0 = 0 ; Tn+1 = Tn + n :
For each probability Px the sequence (n ; Tn ) is an inhomogeneous Markov chain
on E (0; 1) with transition probability given by
2. Continuous time Markov chains 315

P n+1 = y ; t Tn+1 t + dt n = x ; Tn = s
= p(x; y ) (x) e (x)(t s) 1ft > sg dt :
Proof. It is enough to show that for each integer n 1 and each pair of measurable
bounded functions F , G,
h i
Ex F (n+1 ; Tn+1) G(1 ; T1 ; : : : ; n ; Tn)
h i (2:1)
= Ex E(n ;Tn ) F (1 ; T1 ) G(1 ; T1 ; : : : ; n ; Tn ) ;
where the expectation on the right hand side is defined by

X Z

E(x;t) F (1 ; T1) = p(x; y) ds (x)e (x)(s t) 1fs tgF (y; s) :
y2E R
Identity (2.1) is an elementary consequence of the definition of the probability

Px . Indeed, the left hand side is equal to
X h i
Ex F (xn+1 ; Tn+1) G(x1 ; T1 ; : : : ; xn ; Tn)1 = x1 ; : : : ; n+1 = xn+1
x 2E
1iin+1
h i
Px 1 = x1 ; : : : ; n+1 = xn+1 :
For each x in E , denote by fx the density (x)e (x)s 1fs > 0g. By definition of
Px the previous sum is equal to
2 3 2 3
X Y Z 1 Z 1 Y
4 p(xi ; xi+1 )5 ds1 dsn 4 fxk (sk+1 )5
x1 ;:::;xn 2E in
0 1 0 0 kn
0 1
G(x18
; s1 ; : : : ; xn ; s1 + + sn )
9
<X Z 1 =
: p( xn ; y ) dsn fxn (s)F (y; s1 + + sn + s);
y2E 0
provided x0 stands for x. The notation introduced in the beginning of the proof
permits to rewrite, after a change of variables, the last sum inside braces as

E(xn ;s ++sn ) F (1 ; T1 ) :
1
The next to last expression is thus equal to

h i
Ex E(n ;Tn ) F (1 ; T1 ) G(1 ; T1 ; ; n ; Tn) ;
The next result, whose proof relies on the loss of memory of exponential
random variables, will be used throughout this chapter.
Corollary 2.2 For every bounded, B (

)-measurable function H , on the set fTn
tg,
h i
Ex 1fTn+1 > tgH (f(j ; Tj t); j n + 1g) (n ; Tn)
h i (2:2)
= e (n )(t Tn ) En H (j ; Tj ); j 1 :
Proof. It is enough to prove the corollary for functions that depend only on a
finite number of coordinates. To avoid too long formulas, consider a function
H : (E (0; 1))2 ! R that depends only on two coordinates. By Proposition 2.1,
the left hand side of (2.2) is equal to
h i
E(n ;Tn ) 1fT1 > tgH (1 ; T1 t; 2 ; T2 t)
h i h i
= E(n ;Tn ) H (1 ; T1 t; 2 ; 1 + T1 t) T1 > t P(n ;Tn ) T1 > t :
(2:3)
Since T1 is distributed according to an exponential random variable of parameter
(n ), on the set T1 > t, the variable T1 t is also distributed according to an
exponential variable of parameter (n ). In particular, the right hand side of the
previous identity may be rewritten as
h i
e (n )(t Tn ) En H (1 ; T1; 2 ; 1 + T1 ) :
Identity (2.2) can also be proved recalling the explicit formulas for the transi-
tion probabilities. We may indeed rewrite the left hand side of (2.3) as
X Z Z
p(n ; x1 )p(x1 ; x2 ) ds1 ds2 (n )e (n)(s Tn ) (x1 )e (x )(s s )
1 1 2 1
x1 ;x2 R R
1fs1 > Tn _ tg 1fs2 > s1g H (x1; s1 t; x2; s2 t) :
On the set fTn tg the indicator function of the set fs1 Tn g can be removed.
A change of variables permits now to rewrite this sum as
X Z Z
e (n )(t Tn ) p(n ; x1 )p(x1 ; x2 ) ds1 ds2 (n ) e (n )s 1fs1 > 0g
1
x1 ;x2 R R
(x1) e (x1 )(s2 s1 ) 1fs2 > s1 g H (x1 ; s1 ; x2 ; s2 ) :
This double integral is exactly equal to the expectation of H for the Markov chain
starting from n . This last expression is thus equal to
e (n )(t Tn ) En [H (1 ; T1; 2 ; T2)] :

We now construct a continuous time jump process according to the following

prescription. For each t 2 R+ and ! in
, let
Xt ( ! ) = n (!) if Tn(!) t < Tn+1(!) :
This means that the process visits successively the sites occupied by the discrete
time Markov chain n staying at each site an exponential time of parameter (n ).
Notice that Tn " 1 Px almost surely and that there is no ambiguity in the definition
of Xt because was assumed to be bounded and strictly positive.
For each t 0, denote by t :
!
the time translation by t in the path space
. To define t rigorously we need to introduce some notation. For each t 0,

the random variable nt :
! N indicates the interval [Tn ; Tn+1 ) that contains t :
nt (!) := max fn; Tn(!) tg :
With this notation, t ! is the sequence of elements of E (0; 1) whose n-th
coordinate, denoted by (t ! )n , is equal to

(nt ; 0) for n = 0;
(t ! )n =
(nt +n ; Tnt +n t) for n 1:
A simple computation shows that
Xs (t !) = Xs+t (!)
for every s,t 0. Indeed, by definition,
X
Xs (t !) = j (t!)1fTj (t !) s < Tj+1 (t !)g
j0
X
= nt +j (!)1 (Tnt +j (!) t)+ s < Tnt +j+1 (!) t :
j 0
Since s 0, we may replace (Tnt +j (! ) t)+ by Tnt +j (!) t. A change of variables
gives that
X
Xs ( t ! ) = j (!)1fTj (!) s + t < Tj+1 (!)g
jnt
X
= j (!)1fTj (!) s + t < Tj+1 (!)g
j0
Xs+t (!)
=
because, by definition of nt , Tj (! ) t t + s for j nt .
For t 0, denote by Ft the -algebra all events prior to time t : Ft =
fXs ; 0 s tg and denote by F the -algebra generated by the random
variables fXs ; s 0g.
Proposition 2.3 For each x in E , under Px fX (t); t 0g is Markov process

with a time homogeneous transition probability.
Proof. To prove the Markov property we need to show that for every F–
measurable set A and every Ft -measurable set B ,

Px t 1 (A) \ fXt = yg \ B Px [B \ fXt = yg] Py [A] :
=
In fact we shall prove a stronger property : for every n 0

h i
Px t 1 (A) \ fXt = yg \ B \ fTn t < Tn+1g
h i
= Px B \ fXt = y g \ fTn t < Tn+1 g Py [A] :
By the – class Theorem (cf. Theorem 1.4.3 in Chow and Teicher (1988)),
it is enough to check this identity for sets of the form
k1
\
A = fXsj 2 Fj1 g ;
1 0 s10 < < s1k ;
1
j=0
\k2
B = fXsj 2 Fj2g ;
2 0 s20 < < s2k t ;
2
j=0
where, for i = 1, 2, ki is a positive integer and fFji ; 0 j ki g is a collection
of subsets of E .
On the set fTn t < Tn+1 g, the event B is a function of 0 ; : : : ; n and
T0 ; : : : ; Tn , whereas t 1 (A) can be written in terms of fj ; j ng and of
fTj t; j n + 1g. We may therefore rewrite the left hand side of the previous
identity as
h
Ex F (0 ; T0 ; : : : ; n 1 ; Tn 1; y; Tn) 1fn = yg
i
G(y; f(j ; Tj t); j n + 1g) 1fTn t < Tn+1g :
Taking the conditional expectation with respect to f(j ; Tj ); 0 j ng and
applying the Markov property for the process (j ; Tj ) proved in the previous
proposition, this last expression is equal to

Ex F (0 ; T0; : : : ; n 1 ; Tn 1 ; y; Tn) 1fn = yg 1fTn tg
h i
Ex 1fTn+1 > tg G(y; f(j ; Tj t); j n + 1g) (n ; Tn ) :
By Corollary 2.2, this expectation is equal to
h i
Ex F (0 ; T0 ; : : : ; n 1 ; Tn 1; y; Tn) 1fn = yg 1fTn tg e (n)(t Tn )
h i
Ey G(y; f(j ; Tj ); j 1g) :
By Corollary 2.2 and a simple computation taking advantage of the explicit form
of the set A, this expression may be rewritten as
h i
Ex F (0 ; T0 ; : : : ; n 1 ; Tn 1; y; Tn ) 1fn = yg 1fTn t < Tn+1g Py [A] :
The first term is exactly
h i
Px B \ fXt = yg \ fTn t < Tn+1g ;
what concludes the proof.
Definition 2.4 A collection of variables fXt ; t 0g defined on a probability
space (
; A; P ) and taking values in a countable space E is a homogeneous,
continuous time Markov chain if
(a) (Markov property). For every s, t0
h i h i
P Xs+t = y fXr ; r tg = P Xs+t = y Xt
for every site y of E .
(b) (Homogeneity). For every x in E , let Px be the probability on the path space

defined by
Px := P X0 = x :
Then, for every s, t 0 and y in E ,
h i h i
P Xs+t = y Xt = PXt Xs = y :
(c) (Jump property). There exists a sequence of strictly increasing stopping times
(Tn )n0 such that T0 = 0, Xt is constant on the interval [Tn ; Tn+1 ) and XTn =
=
XTn for every n 0.
Proposition 2.5 . (Converse of Proposition 2.3.) If (Xt )t0 is a homogeneous
continuous time Markov chain, then
(a) The skeleton chain defined by n = XTn for n 0 is a discrete time Markov
chain with transition probability p(x; y ) given by
h i
p(x; y) = P XT 1 = y X0 = x :
(b) Recall the definition of the probability Px introduced in Definition 2.4 (b).
Under Px T1 has an exponential distribution whose parameter is denoted by
(x). Conditionally to the sequence (n )n0 the variables j = Tj+1 Tj are
independent and have exponential distributions of parameter (j ).
(c) (Uniqueness in distribution). To each continuous time homogeneous Markov
chain we just associated a transition probability p( ; ) and a jump rate ().
Two continuous time, homogeneous Markov chains having the same transition
probability p and bounded jump rate have the same distribution.
The assumption on the boundness of the jump rate can be weakened in the
proof of the uniqueness in distribution of the process. Nevertheless, an assump-
tion that guarantees the divergence of the sequence of stopping times Tn must
be imposed. The reader can find in Doob (1953) (pp. 266 ff.) a proof of this
proposition.
The assumption XTn = = XTn in Definition 2.4 guarantees that the transition
probability p of the skeleton chain vanishes on the diagonal.
From now on, we shall abbreviate homogeneous, continuous time Markov
chains as Markov chains. Furthermore, we shall always assume the jump rate to
be positive and bounded and the transition probability p to vanish on the diagonal.
Given two Markov chains on the same countable space E , we may construct
on the same path space two probability measures Px and P x corresponding re-
spectively to the pairs ((x); p(x; y )) and ((x); p(x; y )). If we consider the paths
only up to a certain time t, the probability measures P and P are equivalent if the
allowed jumps are the same, i.e., if for every x in E , the sets fy 2 E; p(x; y ) 6= 0g
and fy 2 E; p(x; y ) 6= 0g are the same.
Proposition 2.6 The Radon–Nikodym derivative dP restricted to Ft is given by

dP
the formula
8 9
dP

<Z t X (Xs ) p(Xs ; Xs ) =
= exp (Xs ) (Xs ) ds log
dP Ft : 0 st (Xs ) p(Xs ; Xs ) ;
Proof. The assumption p(x; x) = 0 ensures that the function p(Xs ; Xs ) vanishes
everywhere but at the jumps. In particular, for almost all realization of the process
the sum reduces to a finite sum of terms.
Fix n 1 and let F : (E (0; 1))n ! R be a bounded measurable function.
The expectation under Px of F (1 ; T1 ; : : : ; n ; Tn ) is equal to
2 3
X Y
Z 1 Z 1
4 p(xi ; xi+1 )5 ds1 dsn
x1 ;:::;xn2E in
0 1 0 0
2 3

Y
4 (xk )e (xk )sk +1 5 F (x1 ; s1 ; : : : ; xn ; s1 + + sn ) :
kn
0 1
In this last formula x0 = x. taking the ratio of the densities, we obtain that this
sum is equal to
2 8 93
Z Tn
Ex 4 F exp
<
(Xs ) (Xs ) ds +
X
log
(Xs ) p(Xs ; X =
s) 5 :
: 0 sTn (Xs ) p(Xs ; X s) ;
3. Kolmogorov’s equations, generators 321
For k 1, consider a sequence of times 0 s1 < < sk t and a bounded

measurable function F : E k ! R. On the set fTn t < Tn+1 g, F (Xs1 ; : : : ; Xsk )
is a function of (1 ; T1 ; : : : ; n ; Tn ). Thus, by Corollary 2.2, we obtain that the
expectation of F (Xs1 ; : : : ; Xsk ) under Px is equal to
X
Ex F~(1 ; T1; : : : ; n ; Tn )1fTn tge (n )(t Tn )
n0
for some function F~. By the first part of the proof, this sum is equal to
"
X
E x F~(1 ; T1; : : : ; n ; Tn )1fTn tge (n )(t Tn )
n0
8 9#
Z Tn
exp
<
(Xs ) (Xs) ds +
X
log
(Xs ) p(Xs ; Xs ) :
=
: 0 sTn (Xs ) p(Xs ; Xs ) ;
By Corollary 2.2, this expression may be rewritten as
" 8 9
Z t
(Xs ) ds + log (Xs ) p(Xs ; Xs ) ;
X < X =
Ex exp (Xs )
n0 : 0 st (Xs ) p(Xs ; Xs )
#
F~(1; T1; : : : ; n ; Tn)1fTn t < Tn+1g
and this sum is equal to
"
E x F ( Xs ; : : : ; X s k )
1
8 9#
Z t
(Xs ) ds + log (Xs ) p(Xs ; Xs ) ; :
< X =
exp (Xs )
: 0 st (Xs ) p(Xs ; Xs )
Since every bounded Ft –measurable function may be approximated by functions
depending only on a finite number of coordinates, the proposition is proved.
3. Kolmogorov’s equations, generators
We now introduce the matrices Pt (x; y ) that represent the probability to be at site
y at time t for the Markov chain that starts from x :
Pt (x; y) := Px [Xt = y] :
It follows from the Markov property of the process that theses matrices form a
semigroup :
X
Pt+s (x; y) = Pt (x; z ) Ps(z; y)
z2E
for every (x; y ) in
E 2 . In terms of operators this identity becomes
Pt+s = Pt Ps :
On the other hand, since is assumed to be bounded, it follows from the previous
identity that the operators Pt are continuously differentiable in time and satisfy
the so called Chapman–Kolmogorov equations :
8
>
< P0 (x; y) = x;y
X (3:1)
: @t Pt (x; y ) =
> (x)p(x; z ) Pt (z; y) Pt (x; y) :
z2E
In this formula x;y is equal to 1 if x = y and 0 otherwise. The integral version of
equation (3.1) may be obtained computing Px [Xt = y ] by decomposing the event
fXt = yg according to the time spent at x before the first jump and the first site
visited after x. At the end of this section we give a sketch of an analytical proof
of this formula.
Introducing the operators

(x) p(x; y) if y == x ;
L(x; y) =
(x) if y = x;
the last formula becomes
@t Pt = L Pt :
In particular,
Pt I L
lim
t!0 t =
so that for t small

Pt (x; y) = x;y + t L(x; y) + o(t) :
In a similar way we can show that the operators fPt ; t 0g satisfy the second
Chapman–Kolmogorov equations :
X
@t Pt (x; y) = Pt (x; z )(z )p(z; y) Pt (x; y)(y) (3:2)
z2E
so that
@t Pt
Pt L : =
Notice that the matrix L is such that for every x in E ,
L(x; y) 0 if y == x L(x; x) < 0 and L1 = 0; (3:3)
where 1 stands for the function on E constant equal to 1.

Like in the discrete case in the space Cb (E ) of bounded measurable functions

we introduce the operators (Pt )t0 and L defined by
X
(Pt f )(x) = Pt (x; y)f (y) ;
y2E
X (3:4)
(Lf )(x) = (x)p(x; y) f (y) f (x) :
y2E
Since the jump rate is assumed to be bounded, Cb (E ) is closed under Pt and
L. The next result is proved at the end of this section.
Lemma 3.1 For every bounded function f , the sequence ft 1
( Pt f f ); t > 0 g
converges boundedly pointwise to Lf as t # 0.
By duality we can extend the operators Pt and L to the space of probability

measures on E : X
(Pt )(x) = (y)Pt (y; x) ;
y2E
X
(L)(x) = (y)L(y; x) :
y2E
Notice that < L; 1 >=< ; L1 >= 0 for every probability measure . In partic-
ular, L is not a probability measure.
Conversely, to each operator L satisfying properties (3.3) is associated a unique
pair of jump rate () and transition probability p(; ) vanishing on the diagonal.
Moreover, solving the differential equation
(
P0 = I
@t Pt = L Pt ;
where I stands for the identity, we obtain a unique solution Pt = etL .
We call the operator L the generator of the semigroup Pt . The Trotter–Kato
formula is easily deduced :
n
Pt t
=
!1 I + n L :
nlim
It is also worthwhile to examine the effect on the generator of a time scale
modification in the semigroup. A simple computation shows that for each c > 0
the semigroup fPtc ; t 0g is associated to the generator cL. Therefore a time
scale modification t ! ct translates into a multiplication by c of the jump rate .
We conclude this section with a proof of Lemma 3.1 and formulas (3.1), (3.2).
They rely on the following estimates that will proved later in this section.

Pt (x; x) 1
x 22 t ;
+ ( ) (3:5)
t

Pt (x; y) (x)p(x; y) 2 t p(x; y) + X p (x; y) (t)n ; (3:6)
t n+2 n!
n0
where stands for the upper bound for the jump rate :
= sup (x) :
x2E
Proof of Lemma 3.1. The estimate (3.6) permits to bound the expression

(
Pt f )(x) f (x) (Lf x ;

)( )
t
which is equal to
i
Xh ih

t Pt (x; y) (x)p(x; y) f (y) f x ;
1
( )
y2E
by
no
pn+2 (x; y) (tn!)
Xn X
2kf k1 t p(x; y) +
2
y2E n0

2kf k1 t 1 + et :
2
This proves that the sequence converges uniformly to Lf as t # 0.

We turn now to the proof of formula (3.1). By the semigroup property of the
operators fPs ; s 0g, we may rewrite h 1 [Pt+h (x; y ) Pt (x; y )] as
LPt (x; y) Ph (x; z ) (x)p(x; z )iP (z; y)

Xh
+
h t
z==x
h P (x; x) 1 i
h + (x) Pt (x; y ) :
+
h
Estimates (3.5), (3.6) and similar arguments to the ones presented in the proof of
Lemma 3.1 show that the last two terms vanish as h # 0, what proves (3.1)
The proof of formula (3.2) is even simpler. By the semigroup property, we
have that the difference h 1 [Pt+h (x; y ) Pt (x; y )] is equal to
Pt (x; z ) Ph (hz; y) Pt (x; y) Ph (y;hy)

X
+
1
:
z==x
Formula (3.2) follows thus from the dominated convergence theorem and from
estimates (3.5), (3.6).
It remains to prove inequalities (3.5), (3.6). We start with (3.5). To compute
the probability to be at x at time t for a Markov chain starting from x, we may
decompose the event fXt = xg according to the number of jumps before time t.
Thus, h i X h i
P x Xt = x = Px n = x; Tn t < Tn+1 :
n0
Keep in mind that under Px T1 is an exponential random variable of parameter
(x) and that the transition probability vanishes on the diagonal. In particular,

Pt (x; x) 1
+ ( ) x
t
1n h io 1X h i
t
1 x P T1 > t x ( ) +
t n2 Px n = x; Tn t < Tn+1
h i
t2 (x) +
t Px T2 t :
1
The random variable T2 is the sum of two independent exponential random vari-
ables of parameter (x) and (1 ). Since the jump rate () is bounded above by
, if U1 , U2 stand for two independent, identically distributed, exponential random
variables of parameter , for every state y of E ,
h i h i
Py T2 t P U1 + U2 t (1=2)2 t ;
what proves inequality (3.5).
We now turn to the estimate (3.6). With the same decomposition of the event
fXt = xg performed in the proof of inequality (3.5), we have that

Pt (x; y) (x)p(x; y)
t h

i
t 1Px 1 = y; T1 t < T2 (x)p(x; y)
X h i
+ t 1
Px n = y; Tn t < Tn+1 :
n2
By Corollary 2.2 and the properties of the skeleton chain (n ) stated in Proposition
2.5, this expression is bounded above by
h i

t 1 Ex 1f1 = yg1fT1 tg e (x)(t T ) (x)p(x; y) 1
X h i
+ t 1 pn (x; y)Px Tn tn = y :
n2
h
Like in the proof of inequality (3.5), we may bound the probability Px Tn
i h i
tn = y by Px U1 + + Un t provided (Uj )j 1 stands for a sequence of
independent, identically distributed, exponential random variables of parameter .
A simple computation shows than that the last expression is bounded above by
n
pn+2(x; y) (n(t+)2)! :
X
2 t p(x; y) +
n0
4. Invariant measures, reversibility and adjoint processes
In the same way that we considered invariant probability measures for discrete
time Markov chains, we may investigate invariant measures for continuous time
processes, i.e, probability measures such that Pt = for every t 0.
Proposition 4.1 A probability measure is invariant if and only if L = 0.
Proof. Of course, if Pt = for every t 0, taking the time derivative at t = 0,

we obtain that L = 0. Conversely, if L = 0 it follows from the Trotter–Kato
formula that Pt = for every t 0.
Corollary 4.2 A probability measure is invariant for Pt if and only if the prob-
ability measure
(x) = P (x) (y(x) ) (y)
y2E
is invariant for the discrete time Markov chain with transition probability p(; ).
Proof. The identity L = 0 can be rewritten as

X
(x) (x) p(x; y) = (y) (y)
x2E
for every y in E .
It follows from the previous result that under the assumption of indecom-
posability, that is in general easy to verify, we have the existence of a unique
invariant probability measure when E is finite. In the countable case it will also
be necessary to check conditions that guaranty the positive recurrence. Most of
the time, however, it will be important to characterize more or less explicitly the
invariant measures. In order to do it we have to solve the jE j linear equations :
X
(x) (x) p(x; y) = (y) (y)
x2E
for every y in E . It is stronger but easier to try to solve the system with jE j2
equations
(x) (x) p(x; y) = (y) (y) p(y; x)
known to the physicists as the “detailed balance" condition.
4. Invariant measures, reversibility and adjoint processes 327
It is clear that a summable solution of these equations is automatically

an invariant probability measure if properly renormalized. In particular, it is the
invariant probability measure if the chain is indecomposable. Furthermore, an in-
variant measure satisfying the detailed balance conditions possesses certain special
properties. To state them we need to introduce some notation.
For an invariant probability measure on E , denote by L2 () the space of
the square integrable functions with respect to . We extend the operators Pt
and L, originally defined on Cb (E ), to the space L2 () : for f in L2 (), we
denote by Pt f , Lf the functions defined by (3.4). An elementary computation,
relying on Schwarz inequality and Proposition 4.1, shows that the operators Pt
are contractions on L2 () and that L is a bounded operator :

< ; Lf 2 > 42 < ; f 2 > :
Moreover, it follows from the estimates (3.5), (3.6) that for every function f in
L2 () t 1 [Pt f f ] converges to Lf in L2 () as t # 0 :
n o2
lim
t!0
< ; t 1 [Pt f f ] Lf > = 0 : (4:1)
Proposition 4.3 The probability measure satisfies the detailed balance condition
if and only if the operators Pt are self–adjoint in L2 (), i.e., if and only if for every
t 0, f , g in L2 (),
X X
(x) f (x) Pt g(x) = (x) g(x) Ptf (x)
x x
or, briefly,
< f; Pt g > = < Pt f; g > : (4:2)
Here < ; > stands for the inner product of L2 ().
Proof. Fix two sites x, y , set f = 1fxg, g = 1fy g and take the time derivative at
t = 0 in identity (4.2) to obtain that
(x) L(x; y) = (y) L(y; x)
which is the detailed balance condition. Inversely, it follows from the previous
identity that
< f; Lg > = < Lf; g >
for every f , g in L2 (). It remains to recall Trotter–Kato formula to conclude the
proof.
A probability measure satisfying the detailed balance conditions is said to
be reversible. The previous result states therefore that a probability measure is
reversible if and only if the generator is self adjoint in L2 (). We shall now look
for conditions that guarantee that the adjoint of the generator L in L2 () is also a
generator.
Proposition 4.4 Let be a probability measure. The adjoint of L in L2 (),

denoted by L , is a generator if and only if is invariant. In this case Pt = etL is
also the adjoint of Pt in L2 () and the semigroup Pt is characterized by the pair
( ; p ) given by
8
< (x) = (x) ; for x 2 E
: p (x; y ) =
(y)(y)p(y; x) ; for x; y 2E:
(x)(x)
Moreover,
< f; Lg > = < L f; g > (4:3)
for every f , g in L2 ().
Proof. A simple computation shows that the adjoint L of L in L2() is given

by the formula
(x) L(x; y) = (y) L(y; x) : (4:4)
In particular, L satisfies always the first two properties of generators :
(
L (x; x) = (x) < 0 ; for x 2 E
L (x; y) 0 ; for x == y :
Therefore, L is a generator if and only if
X
L(x; y) = 0
y2E
for every x in E and hence if and only if is invariant because the explicit
expression (4.4) permits to rewrite the previous sum as
1 X
(x) (y)L(y; x) :
y2E
On the other hand, the Trotter–Kato formula shows that the semigroup Pt
associated to the generator L is the adjoint of the semigroup Pt in L2 ().
Finally, we have already seen in the first part of the proof that the jump rate
and coincide. The explicit formula (4.4) permits than to compute p (x; y).
On the other hand, formula (4.3) follows from the identity (x)(x)p(x; y ) =
(y)(y)p(y; x) and a change of variables :
X X
< f; Lg > = (x)(x)p(x; y)f (x)g(y) (x)(x)f (x)g(x) :
x;y2E x 2E
Since (x)(x)p(x; y ) = (y )(y )p(y; x), the first term on the right hand side can
be written as
X X
(y)(y)p(y; x)f (x)g(y) = (x)(x)p (x; y)f (y)g(x) :
x;y2E x;y2E
4. Invariant measures, reversibility and adjoint processes 329
From the previous two identities it is easy to conclude the proof of the proposition.

The process Pt , that is defined without ambiguity when the original process
admits a unique invariant measure, is called the adjoint process. The adjoint process
is closely connected to the process reversed in time. This is the content of the next
proposition.
Proposition 4.5 If the semigroup Pt is the adjoint of the semigroup Pt with
respect to the invariant probability measure , then for every n > k 0, every
sequence of times 0 t1 < < tn and every sequence of bounded functions
ffj ; 1 j ng,
h i
E f1(Xt ) fn (Xtn )
1
X h i
= (x) fk (x) Ex fk+1 (Xtk tk ) fn (Xtn tk )
+1
x2E h i
Ex fk (Xtk tk 1 ) f1(Xtk t1 ) ;
where Ex stands for the expectation with respect to the Markov chain with transi-
tion probability Pt starting from x.
Proof. The proof is straightforward. We just have to apply successively the identity
(x) Pt (x; y) = (y) Pt(y; x) :
We shall sometimes consider the symmetric part, denote by S , of the generator

L in L2 (). It is given by
S = 2 1
L + L :
A simple computation shows that the symmetric part S is itself a generator char-
acterized by the parameters s and ps given by
s (x) = (x) ; x 2 E

ps (x; y) = 12 p(x; y) + (y)(x(y))p(x(y;) x) :
Furthermore, S satisfies the detailed balance condition :
(x) S (x; y) = (y) S (y; x) :

5. Some martingales in the context of Markov processes
The purpose of this section is to introduce a class of martingales in the context

of Markov processes. Consider a bounded function F : R+ E ! R smooth in
the first coordinate uniformly over the second : for each x in E , F (; x) is twice
continuously differentiable and there exists a finite constant C such that

sup (@sj F )(s; x) C

(5:1)
s;x) (
for j = 1, 2. In this formula (@sj F ) stands for the j -th time derivative of F (; x).
To each function F satisfying assumption (5.1), define M F (t) and N F (t) by
Z t
M F ( t) = F (t; Xt) F (0; X0) ds (@s + L)F (s; Xs) ;
0
Z t n o
N F (t) = (M F (t))2 ds LF (s; Xs)2 2F (s; Xs )LF (s; Xs ) :
0
Lemma 5.1 Denote by fFt ; t 0g the filtration induced by the Markov process :
Ft = (Xs ; s t). The processes M F (t) and N F (t) are Ft -martingales.
Proof. We start showing that M F (t) is a martingale. Fix 0 s < t. We need to
check that Ex [M F (t)jFs ] = M F (s), i. e., that
h i Z t h i
Ex F (t; Xt ) Fs = F (s; Xs ) + Ex (@r + L)F (r; Xr ) Fs dr :
s
For each r 0, denote by Fr : E ! R (resp. by Fr0 : E ! R) the function that at x
takes the value F (r; x) (resp. (@r F )(r; x)). By the Markov property and a change
of variables in the integral, the previous identity is reduced to
Z t sn o
(Pt s Ft )(Xs) = Fs (Xs ) + (Pr Fs0+r )(Xs ) + (Pr LFs+r )(Xs ) dr :
0
Since for t = s this identity is trivially satisfied, we just need to check that the
time derivative of both expressions are equal, i. e., that
@t (Pt s Ft )(x) = (Pt s Ft0 )(x) + (Pt s LFt )(x)

for every x in E and 0 s < t.
To prove this identity we compute the left hand side. Fix h > 0 and rewrite
the difference h 1 f(Pt+h s Ft+h )(x) (Pt s Ft )(x)g as
h i h i
h 1Ex Ft+h (Xt s+h ) Ft(Xt s+h ) + h 1 Ex Ft (Xt s+h) Ft (Xt s ) : (5:2)
The first expression is equal to
5. Some martingales in the context of Markov processes 331
Z t+h h i
1
h t dr Ex Fr0 (Xt s+h) Ft0 (Xt s+h)
h i h i
Ex Ft0 (Xt s+h) Ft0 (Xt s) + Ex Ft0 (Xt s) :
+
Since by assumption (5.1) (@r F )(; x) is Lipschitz continuous uniformly on x, the

first term vanishes as h # 0. The second term also vanishes as h # 0 because the
semigroup Pt is continuous. Therefore, as h # 0, the first term in (5.2) converges
to (Pt s Ft0 )(x).
The second expression in (5.2) is equal to
Z t s+ h h i
h t s
1
dr Ex LFt (Xr )
that converges, as h # 0, to (Pt s LFt )(x). This proves that M F (t) is a martingale.
We show now that N F (t) is a martingale. A simple computation and the first
part of the lemma show that M F (t)2 is equal to
Z t Z t 2
F 2 (t; Xt ) 2F (t; Xt ) ds (@s + L)F (s; Xs) + ds (@s + L)F (s; Xs)
(5:3)
0 0
R
plus a martingale term. Since F 2 (t; Xt ) ds(@s + L)F 2(s; Xs ) is a martingale,
F 2 (t; Xt ) is equal to a martingale added to
Z
ds (@s + L)F 2(s; Xs ) : (5:4)
The second and third expression in (5.3) can be rewritten as

Z t Z t 2
2M0F (t) ds (@s + L)F (s; Xs) ds (@s + L)F (s; Xs) ; (5:5)
0 0
where M0F (t) = M F (t)+F (0; X0). By Ito’s formula, the first term in this expression
is equal to a martingale added to
Z t
2 ds F (s; Xs)(@s + L)F (s; Xs)
0
Z t Z s
+ 2 ds dr (@r + L)F (r; Xr ) (@s + L)F (s; Xs) :
0 0
An integration by parts shows that the second term of this formula cancels with the
second term of (5.5). The remaining expression added to (5.4) is just the integral
term that we need to subtract in order to turn (M F (t))2 in a martingale. This
6. Estimates on the variance of additive functionals of Markov

processes
Consider a Markov process with generator L satisfying the assumptions of section

3 and having an invariant measure denoted by . Let S be the symmetric part of
the generator L : S = 2 1 (L + L ). For a function f in L2 ( ), denote by kf k1 its
H1 norm defined by
kf k21 = < f; ( S )f >
and notice that kf k21 = < f; Lf > . Recall from section 4 that we denote by
ps (x; y) the transition probability associated to the symmetric part of the generator.
With this notation,
1 X n o2
kf k21 =
2
(x)(x)ps (x; y) f (y) f (x) :
x;y2E
In particular, by Schwarz inequality, the H1 norm kf k21 of any function f in L2 ( )
is bounded above by 2kk1 kf k20 , provided kf k0 stands for the L2 ( ) norm of f .
Let H1 be the Hilbert space generated by L2 ( ) and the inner product <

f; g >1 =< f; ( S )g > : H1 = L2()N , if N stands for the kernel of the inner
product < ; >1 . (In the context of this chapter, N is the space generated by
constant functions because we assumed the skeleton discrete time Markov chain
to be indecomposable). Since kf k21 2kk1 kf k20 , L2 ( ) H1 .
Denote by kk 1 the dual norm of H1 with respect to L2 ( ) : for f in L2 ( ),
let n o
kf k2 1 = sup 2 < f; g > kgk21 : (6:1)
g2L2 ()
We claim that kf k2 1 (2kk1 ) 1 kf k20 for each function f in L2 ( ). Indeed,
since kg k21 2kk1 kg k20, the supremum on the right hand side of (6.1) is bounded
below by n o
sup 2 < f; g > 2kk1 kg k20 :
g2L2 ()
By Schwarz inequality this supremum is equal to (2kk1 ) 1 kf k20 .
Denote by H 1 the subset of L2 ( ) of all functions with finite kk 1 norm.
It follows from the definition of the H 1 norm that
> A1 kf k2 1 + Akgk21
2 < f; g (6:2)
for every f in H 1 , g in L2 ( ) and A > 0. Since L2 ( ) is dense in H1 ,

this inequality may be extended to functions g in H1 . Moreover, we claim that
SL2() = fSf; f 2 L2 ()g is contained in H 1 and
kSf k2 1 = kf k21
for all f in L2 ( ). Indeed, fix f in L2 ( ). A simple computation shows that
Variance of additive functionals of Markov processes 333
2 < g; Sf> < g; ( S )g > + < f; ( S )f > = kgk21 + kf k21

for every g in L2 ( ). In particular, kSf k 1 kf k1 . Taking g = f in the
supremum that defines kf k2 1 we deduce the reverse inequality.
We are now ready to state the main result of this section.
Proposition 6.1 For each function g in H 1 and t > 0,

h Z t 2 i
E p1 g(Xs) ds 20 kg k2 1 :
t 0
Proof. Since L is a generator, for every > 0, ( L) 1 is a bounded operator

in L2 ( ). In particular, since H 1 L2 ( ), there exists f in L2 ( ) such that
f Lf = g:
Taking on both sides of this equation the inner product with respect to f and
applying Schwarz inequality (6.2), we obtain that
< f ; f > kgk2 1 and kf k21 kgk2 1 : (6:3)
For > 0 consider the process M (t) defined by M (t) = f (Xt ) f (X0 )
Rt
0 Lf (Xs )ds. By Lemma 5.1, M (t) is a martingale if f is bounded. Approxi-
mating f , that belongs to L2 ( ), by bounded functions we deduce that M (t) is a
martingale in L2 ( ). With this notation we may rewrite the expectation appearing
in the statement of the lemma as
h Z t 2 i
t E M (t) f (Xt) + f (X0 ) + f (Xs )ds :
1
0
Since is an invariant measure for the Markov process, by Schwarz inequality,

this expression is bounded above by
4n h io
t f2 + ( t) g < f ; f > + E M (t) :
2 2
By Lemma 5.1 the quadratic variation of the martingale M (t) is equal to the
time integral of Lf (Xs )2 2f (Xs )Lf (Xs ). In particular, since the probability
measure is invariant, the expectation of M (t)2 is equal to 2tkf k21 because
< f; Lf > = kf k21 for every f in L2 (). Therefore, the expectation appearing
on the statement of the lemma is less than or equal to
4n o
t f2 + (n t) g < f ; f > + 2tkof k1
2 2
4f2 + ( t)2g(t ) 1 + 8 kgk2 1

in virtue of (6.3). To conclude the proof of the proposition, it remains to choose
= t 1.
Notice that we did not assume the process to be reversible.
Remark 6.2 Under the assumption of indecomposability, it is not difficult to

show that L2 ( ) = SL2 ( ) 1, where 1 is the subspace of L2 ( ) generated by
the constants and that H 1 = SL2 ( ). Thus, for each g in H 1 , there exists f in
L2 () such that Sf = g. This argument remains in force if S is replaced by L.
7. The Feynman–Kac formula
Consider a bounded function V : R+ E ! R satisfying assumption (5.1) and a

bounded function F0 : E ! R. Fix T > 0 and denote by F : [0; T ] E ! R the
solution of the differential equation
(
(@t u)(t; x) = (Lu)(t; x) + V (T t; x)u(t; x) ;
(7:1)
u(0; x) = F0 (x) :
Proposition 7.1 The solution F has the following stochastic representation :
h RT i
F (T; x) = Ex e V (s;Xs ) ds F0 (XT ) :
0
Proof. Consider the process fAt ; 0 t T g given by

nZ t o
At = F (T t; Xt ) exp V (s; Xs ) ds :
0
By Lemma 5.1, At can be rewritten as

n Z t o nZ t o
M0F (t) + ds (@s + L)F (T s; Xs ) exp V (s; Xs ) ds ;
0 0
Rt
where M0F (t) is the martingale F (T t; Xt ) 0( @s + L)F (T s; Xs )ds. Ito’s
formula now gives that
Z t R s n o
At ds e V (r;Xr ) dr F (T s; Xs )V (s; Xs )
0 + (@s + L)F (T s; Xs )
0
is a martingale. Since F is the solution of (7.1), the integral term vanishes showing
that At is martingale. In particular, Ex [AT ] = Ex [A0 ], which proves the lemma.

The proof of the previous lemma shows that
Rt Z t Rs n o
e V (r;Xr ) dr F (Xt )
0 ds e V (r;Xr ) dr (LF )(Xs) + V (s; Xs)F (Xs ) (7:2)
0
0
7. The Feynman–Kac formula 335
is a martingale for each bounded function F : E ! R.

For t 0, denote by Lt : E E ! R the operator defined by Lt (x; y ) =
L(x; y) + V (t; y)x;y , where x;y stands for the delta of Kronecker. Denote fur-
thermore, for 0 s t, by Ps;tV : E E ! R+ the operator given by
h Rt i
V (x; y) =
Ps;t Ee s V (r;Xr ) dr 1 fXt = yg Xs = x : (7:3)
A simple computation, relying on Markov property, permits to rewrite V (x; y)

Ps;t
as h Rt s i
V (x; y) = Ex e 0 V (s+r;Xr ) dr 1fXt s = y g :
Ps;t
The collection fPs;t
V ; 0 s tg can be extended to act on bounded functions :
h Rt s
V f )(x) =
(Ps;t Ex e V (s+r;Xr ) dr f (X )i :
0
t s
Property (7.2) applied to the bounded function F (z ) = 1fz = y g and a simple
computation shows that fPs;t
V ; 0 s tg is a semigroup associated to the operator
V P V = P V for all 0 s t u and
Lt : Ps;t t;u s;u
(
V )(x; y) = (P V Lt )(x; y) ;
(@t Ps;t s;t (7:4)
V (x; y)
Ps;s = x;y
for t s 0.
The arguments presented in section 3 permit to prove also the first Chapman–
Kolmogorov equations :
(
V )(x; y) =
(@s Ps;t V )(x; y) ;
(Ls Ps;t
Pt;tV (x; y) = x;y
for 0 s t.
Assume now that L is a reversible generator with respect to an invariant state
. Since V is bounded, Lt = L + V (t; ) is also a symmetric operator in L2( ).
Denote by t the largest eigenvalue of L + Vt :
n o
t = sup < Vt ; f 2 > + < Lf; f > :
kf k2 =1
By definition of the semigroup fPs;tV ; t s 0g,
h t R i
E e V (r;Xr ) dr = < P0V;t 1; 1 > :
0
On the other hand, by the first Chapman–Kolmogorov equation, for 0 s t,

d V V V 1; P V 1 >
ds < Ps;t 1; Ps;t 1 > = 2 < Ls Ps;t s;t
because Ls is a symmetric operator. By definition of s , this expression is bounded

below by
V 1; P V 1 >
2 s < Ps;t
s;t
Therefore, by Gronwall inequality and since Pt;t is the identity,
nZ t o
< P0V;t 1; P0V;t 1 > exp s ds :
0
Since, by Schwarz inequality,
< P0V;t 1; 1 > < P0V;t 1; P0V;t 1 >1=2 ;

we have proved the following lemma :
Lemma 7.2 Assume that the Markov process is reversible with respect to an
invariant probability measure . Let V : R+ E ! R be a bounded function. For
each t 0 denote by t the largest eigenvalue of the operator L + V (t; ). Then,
h nZ t oi nZ t o
E exp V (r; Xr ) dr exp s ds : (7:5)
0 0
In the case where is only an invariant measure, the previous argument shows
that (7.5) remains in force provided s is the largest eigenvalue of S = (1=2)fL +
L g, the symmetric part of L in L2 ( ).
The Feynman–Kac formula presented in Proposition 7.1 permits to obtain an
explicit formula for the Radon–Nikodym derivative of a time inhomogeneous
Markov process with respect to another, generalizing Proposition 2.6 to the inho-
mogeneous case. Fix a function F : R+ E ! R satisfying assumptions (5.1).
Denote by M F (t) the process defined by
n Z t o
M F (t) = exp F (t; Xt) F (0; X0) ds e F (s;Xs )(@s + L)eF (s;Xs ) :
0
We claim that M F (t) is a mean 1 positive martingale. Indeed, fix x0 in E and

define V , H : R+ E ! R by
V (t; x) = expf F (t; x)g(@s + L) expfF (t; x)g ;
(7:6)
H (t; x) = expfF (t; x) F (0; x0)g :
It follows from the proof of Proposition 7.1 that
Z t
M F (t) = H (t; Xt ) expf V (s; Xs )dsg
0
is a martingale. Since this martingale is equal to 1 Px 0 almost surely at time 0, it

is a mean 1 positive martingale.
7. The Feynman–Kac formula 337
Fix a time T > 0 and for each x0 in E , define on FT the probability measure
PxF by
Ex G M F (T )
ExF G (7:7)
0
0
= 0
for all bounded FT -measurable functions G. A simple computation shows that the
conditional expectation is

ExF G Fs
0
=
1
Ex
M F (s) 0
G M F (T ) Fs :
For 0 s t T , define the functions QF s;t : E E ! R+ by QFs;t (x; y) =
Px0 [Xt = yjXs = x]. Recall the definition of the function V introduced in (7.6)
F
and of the semigroup fPs;t V ; 0 s tg given in (7.3). It follows from the formula
for the conditional expectation that QF s;t (x; y) = Ps;t (x; y) expfF (t; y ) F (s; x)g.
V
We claim that fQF s;t ; 0 s t T g is a semigroup of transition probabilities.
Indeed, it is clear that QF s;t (x; y) 0 for all x, y inPE . It follows from the
F
explicit formula for the conditional expectation that y2E Qs;t (x; y) = 1 for
all 0 s t T and x in E . Finally, by definition of QF s;t and because
fPs;tV ; 0 s tg is a semigroup,
X
QFs;t (x; z )QFt;u(z; y) = QFs;u (x; y)
z2E
for all s t u and all x, y in E . This proves that fQF s;t ; 0 s t T g is a
semigroup of transition probabilities. We now compute the generator associated to
this semigroup. Since QF V (x; y) expfF (t; y ) F (s; x)g, by the second
s;t (x; y) = Ps;t
Chapman–Kolmogorov equations (7.4),
n o
@t QFs;t (x; y) V (x; y) expfF (t; y ) F (s; x)g
@t Ps;t
=
n o
= expfF (t; y ) F (s; x)g (Ps;t V Lt )(x; y) + P V (x; y)(@t F (t; y )) :
s;t
The explicit formula for Lt and some elementary computations lead to the formula
@t QFs;t (x; y) = (QFs;tLFt )(x; y) ;
where
LFt (x; y) =L(x; y) expfF (t; y) F (t; x)g x;y expfF (t; y)gL expfF (t; y)g :
In particular, for any bounded function H : E ! R,
X
(LF
t H )(x) = (x)p(x; y) expfF (t; y) F (t; x)g[H (y) H (x)] : (7:8)
y2E
We summarize in the next proposition the result just proved.
Proposition 7.3 Fix a function F : R+ E ! R satisfying assumptions (5.1). Fix

a time T > 0 and for each x0 in E , define on FT the probability measure PxF0
by (7.7). Under PxF0 , Xt is a time inhomogeneous Markov process with generator

fLFt ; 0 t T g given by (7.8) and starting from x0 .
8. Relative entropy
Let be a reference probability measure on E . For a probability measure denote

by H (j ) the relative entropy of with respect to defined by the variational
formula :
H (j) = sup < ; f > log < ; ef > :
f
In this formula the supremum is carried over all bounded functions f and < ; f >
stands for the integral of f with respect to . From now on, to keep notation and
terminology simple, we denote H (j ) by H () and refer to it as the entropy of
.
Notice that the addition of a constant to the function f does not change the
value of < ; f > log < ; ef >. We may therefore restrict the supremum
to bounded positive functions. The next result follows easily from the variational
formula for the entropy.
Proposition 8.1 The entropy is positive, convex and lower semicontinuous.
To show that the entropy is positive, just observe that < ; f > log <
; ef > vanishes for any constant function f .
We shall repeatedly use the entropy to estimate the expectation of a function
with respect to a probability measure in terms of integrals with respect to the
reference measure . Indeed, the entropy inequality gives that

< ; f > 1 log < ; ef > + H ()
for every positive constant . For indicator functions this inequality takes a simple
form.
Proposition 8.2 Let A be a subset of E .

log 2 + H ()
[A]
1

log 1 + A
[ ]
Proof. By the entropy inequality, for every > 0,

h i
[A] 1 log 1 + (e 1)[A] + 1 H () :
It is enough to choose in such a way that (e 1) [A] = 1, i.e., to take
8. Relative entropy 339

= log 1 +
1
[A] :
The next result presents an explicit formula for the entropy.
Theorem 8.3 The entropy H () is given by the formula
(x) ((xx)) (x)

X
H () =
(x)
log
x2E
(x) log ((xx))
X
=
x2E
if is absolutely continuous with respect to and is equal to 1 otherwise.
Proof. If the probability measure is not absolutely continuous with respect to
, since E is assumed to be countable, a simple choice of bounded functions f
shows that the entropy H () is infinite.
Assume now that is absolutely continuous with respect to and that the set
E is finite. The functional : RE ! R defined by
(f ) = < ; f > log < ; ef >
is concave and assumes its maximum where its gradient vanishes : for every
function f such that
f;h >
< ; h > = <<e ; ef >
for every h in RE . The invariance of by the addition of a constant permits to
choose among these functions one such that < ; ef > is equal to 1. In particular,
d
f = log d
and
sup (f ) = log d
d = < ; log
d >
d
f
To extend this result to the countable case, it is enough to remark that the
variational formula for the entropy can be rewritten as
H () = lim sup

k!1 f 2D(Ek )
(f ) ;
where (Ek )k1 stands for an increasing sequence of finite subsets of E whose
union is equal to E and D(Ek ) for the set of functions that are constant on the
complement of Ek . The arguments presented in the first part of the proof shows
that the supremum in last formula is equal to
c (Ekc )
(x) ((xx)) (x) (Ekc ) ((E
X
k)
x2Ek
log
(x) +
Ekc ) log
(EKc )
This expression is increasing in k because the function u log u is convex. More-
over, as k " 1, it converges, to
X (x)
x2E
(x) log
(x)
This explicit formula for the relative entropy involving the function u log u
explains the relation between the entropy and the expectation of functions of type
ef in the entropy inequality. Indeed, by Legendre duality, the convex functions
(u) = (1 + u) log(1 + u) u and (v) = ev 1 v form a pair of Young functions
(cf. Neveu (1972) for the terminology and an introduction to the Orlicz spaces
associated to pairs of Young functions). In this case we have that
uv (u) + (v )
for u, v 0. A simple computation shows that this inequality holds for u 1
and v 2 R. Changing variables, we obtain that
uv ev + u log u u (8:1)
for u 0 et v 2 R. Taking u as the density of with respect to and v as a
function f plus a constant, after an integration with respect to , we obtain the
inequality Z Z
f d eC0 ef d + H () 1 C0 :
Minimizing over the constant C0 , we obtain the entropy inequality.
9. Entropy and Markov processes
Consider a Markov chain on a countable space E with an invariant measure

denoted by . Let (Pt )t0 be the semigroup associated to the Markov chain.
The relative entropy with respect to the invariant measure plays an important
role in the investigation of the time evolution of the process. First of all, since
'(u) = u log u is strictly convex and vanish only at s = 0 and s = 1, the relative
entropy of Pt with respect to does not increase in time. This is the content of
the first proposition.
Proposition 9.1 For every probability measure , we have

H (Pt ) H () :
Moreover, H (Pt ) = H () < 1 implies that = if the chain is indecomposable.
9. Entropy and Markov processes 341
Proof. The proof relies on the explicit formula obtained in the previous section
and on the convexity of the function '(u) = u log u. Assume, without loss of
generality, that is absolutely continuous with respect to . A simple computation
shows that Pt shares the same property for all t 0. The explicit formula for
the entropy permits to write

X X
H (Pt ) = (x) ' (1x) (y) Pt(y; x)
x y

X X (y ) (y ) Pt (y; x)
(x) '
y (y ) (x)
=
x

X X
(x) ' (y) (y ) (y ) Pt (y; x)
x y (x)
= H () :
Since '(u) is strictly convex, there is equality only if (y )= (y ) is constant for
every y such that Pt (x; y ) > 0 for some (and therefore for all) x.
The entropy furnishes an upper bound for the distance between two probability
measures. Denote by k k the total variation distance :

k k := 2 sup [A] [A] :
AE
In this formula, the supremum is taken over all subsets of E . The reason for the
constant 2 is explained by the explicit formula for the total variation distance.
Indeed, a straightforward argument shows that the total variation distance can be
written as
X X
k k = j(x) (x)j = (x)jf (x) 1j
x2E x2E
provided f stands for the Radon–Nikodym derivative of with respect to . It
follows now from Schwarz inequality and the elementary inequality 3(a 1)2
(2a + 4)(a log a a + 1) for a 0 that
k k2 2H () :
On the other hand, if Pt stands for the adjoint of Pt in L2 ( ) and if f (resp.
ft ) stands for the density of (resp. Pt ) with respect to ,
ft (x) = (Pt f )(x) :
In particular, the density ft is solution of

(
f0 = f
(9:1)
@t ft = L ft
This observation permits to deduce a simple estimate for the time derivative
of the entropy of Pt .
Theorem 9.2 Let be a probability measure with finite entropy : H () < 1. For
every t, h 0, we have that
Z t+ h
H (Pt+h ) H (Pt ) = ds < fs ; L log fs >
t
Z t+h p p
ds 2 < fs ; L fs > :
t
Moreover,
p p X hp p i2
2< fs ; L fs > = (x) L(x; y) fs (y) fs (x) :
x;y2E
In these formulas we used the notation introduced in Proposition 4.3.
Proof. By the explicit formula for the entropy, the difference H (Pt+h ) H (Pt)
is equal to
X Z t+ h
(x) ds f1 + log fs (x)gLfs (x) :
x2E t
Recall that we denoted by the upper bound of the jump rate (). It is
easy to check that the absolute value of L fs is P
bounded above by the integrable
(with respect to ) positive function ffs (x) + y p (x; y )fs (y )g. In particular,
by Fubini lemma
X Z t+h Z t+ h X
(x) ds Lfs (x) = ds (x)L fs (x)
x2E t t x 2E
Z t+h
= ds < ; Lfs > :
t
Since L is the adjoint of L in L2 ( ), this expression vanishes.
On the other hand, log fs (x)L fs (x) can be rewritten as
X
L (x; y) log fs (x)fs (y) (x)fs (x) log fs (x) :
y2E
y 6= x
Since the entropy deceases in time, fs log fs is integrable for t s t + h. By
inequality (8.1), the positive part of log fs (x)fs (y ) is also integrable. In particular,
by Fubini’s lemma, we may interchange the time integral with the space integral
to obtain that
X Z t+h Z t+ h
( x) ds log fs (x)Lfs (x) = ds < log fs ; L fs > :
x2E t t
10. Dirichlet form 343
p pof
This concludes the proof of the first part of the theorem in view of thepdefinition
the adjoint L . To prove the inequality, recall that a[log b log a] 2 a[ b a]
for a, b 0. From this inequality it follows that the right hand side of the previous
identity is bounded above by
Z t+h p p
2 ds < fs ; L fs > :
t
It remains to show that
p p X hp p i2
2< fs ; L fs > = (x) L(x; y) fs (y) fs (x) : (9:2)
x;y2E
Since L is the adjoint of L in L2 (),
p p p p p p
2 < fs ; L fs > = < fs ; L fs > + < fs ; L fs > :
The right hand side is equal to
X p hp p i
(x) L(x; y) fs (x) fs (y) fs (x)
x;y2E
X p hp p i
+ (x) L(x; y) fs (x) fs (y) fs (x) :
x;y2E
Since L is the adjoint of L, (x)L(x; y ) = (y )L (y; x). A change of variables
in the second sum of the last expression permits to conclude.
10. Dirichlet form
At the end of the previous section we deduced that the time derivative of the
entropy of St is bounded above by a function of the density of Pt . It is therefore
natural to introduce, for every function f in L2 ( ), the Dirichlet form D(f ) of f
defined by
X
D(f ) := < f; Lf > = f (x) Lf (x) (x) :
x2E
This sum is well defined because the generator L is a bounded operator in L2 ( ).
The next result follows from the computation performed at the end of last section.
Proposition 10.1 The Dirichlet form of a function f in L2 ( ) is positive and equal

to
X
D(f ) = 12 (x) L(x; y) [f (y) f (x)]2 :
x;y2E
Notice that if the Dirichlet form of a function f vanishes, D(f ) = 0, and

if the process is indecomposable, then f is constant. Furthermore, if a function
F : R ! R is such that jF (a) F (b)j ja bj, then
D( F f ) D ( f ) : (10:1)
This inequality applies to the function F (x) = x ^ M where M is a fixed real and
to F (x) = jxj. Furthermore, the Dirichlet form is convex : let p be a probability
measure on N . By Schwarz inequality,
X X
D( pj fj ) pj D(fj ) : (10:2)
j j
In the case, where the probability measure is reversible, there exists a vari-
ational formula for the Dirichlet form D(f ).
Theorem 10.2 (Variational formula for the Dirichlet form) For every positive
function f of L2 ( ),
(x) fg((xx)) Lg(x) :

X 2
D(f ) := inf
g x2E
In this formula, the infimum is taken over all bounded positive functions bounded
below by a strictly positive constant.
Proof. We first show that the supremum is bounded above by the Dirichlet form.
Fix a function g that is bounded and is bounded below by strictly positive constant.
Set = fg=f g1ff > 0g so that vanishes where f vanishes. With this definition,
< fg ; Pt g > = < f ; Pt f > :

2
Since is reversible, this last sum is equal to

(x) f (x) f (y) Pt(x; y) ((xy))
X
x;y

1 X
(x) f (x) f (y ) Pt(x; y )
(y) + (x) :
=
2 x;y (x) (y)
Since for every positive real a, a + a 1 2, this expression is bounded below by
X
(x) f (x) f (y) Pt(x; y) = < f; Pt f > :
x;y
Subtracting < f 2 > to both terms and multiplying by t 1 , we deduce that
1
< f 2 ; (g P g) > 1 < f; (f P f ) > :
t g t t t
10. Dirichlet form 345
Letting t # 0, we obtain that

< fg ; Lg > D(f )
2
for every fixed function g because, by Lemma 3.1, t 1 (g Pt g ) converges bound-

edly pointwise to Lg and by (4.1) t 1 (f Pt f ) converges in L2 () to Lf .
Formally, to prove the converse inequality, it is enough to take g = f . To prove
it rigorously, however, we need to approximate f by bounded positive functions
bounded below by strictly positive constants.
For each positive integer M , let fM be the function defined by
fM (x) = M 1
+ f ( x) ^ M :
Notice that fM is bounded and bounded below by a strictly positive constant and
that LfM is bounded. Moreover, since is reversible, we have that
< ff ; LfM > = (x)L(x; y) ff (y()y) ff (x()x) [fM (y) fM (x)] :

2
1 X n 2 2 o
M 2
x;y2E M M
Since fM converges pointwisely to f as M " 1, by Fatou lemma and the explicit
formula for the Dirichlet form,
D(f ) lim inf <

f 2 ; Lf > ;
M !1 fM M
what concludes the proof of the theorem.
The next result is a simple consequence of this proposition.
Corollary 10.3 The functional

p
D (f ) = D( f )
defined for all densities with respect to is convex and lower semicontinuous.
p In principle the domain of the Dirichlet form corresponds to the domain of

L. The explicit form for the Dirichlet form given above in Proposition 10.1
permits, however, to extend the definition of the Dirichlet form to L2 ( ) if we
allow it to assume the value +1. From now on we will always implicitly assume
that D() is defined on L2 ( ).
11. A maximal inequality for reversible Markov processes
We conclude this chapter with a maximal inequality for reversible Markov pro-
cesses due to Kipnis and Varadhan (1986). We assume throughout this section that
Xt is a Markov process reversible with respect to some invariant state .
Theorem 11.1 Fix g : E ! R. For each T > 0 and A > 0, we have that
jg(Xt)j A Ae < g; g > + T D(g) :
h i p
P sup (11:1)
0 tT
Proof. Denote by G the subset G = fx; jg (x)j Ag and by the hitting time of
G : = infft 0; Xt 2 Gg. For each > 0, define the function ' : E ! R+ by
h i
' ( x) = '(; x) = Ex e :
Of course ' is identically equal to 1 on G. On the other hand, for x
not in
G, decomposing the chain according to the first site visited, the Markov property
gives that
Z t X
'(; x) = ds (x)e f(x)+gs p(x; y)'(; y) + e f(x)+gt '(; x) :
0 y2E
Dividing by t and letting t # 0, we get that
X
(x) p(x; y)['(; y) '(; x)] = '(; x) :
y2E
for x in Gc . We have thus proved that L' = ' on Gc .
By definition of the stopping time , the left hand side of (11.1) is equal to
P [ T ]. Since
Px [ T ] eT '(; x) ;
by Schwarz inequality we have that
h i X
P sup jg(Xt)j A eT (x) '(; x)
tT
0 x 2E
sX
eT (x) '(; x)2 :
x2E
The expression inside the square root is clearly bounded above by
X
(x) ' (x)2 +
D(' ) :
1
(11:2)
x2E
To conclude the proof of the theorem it remains to show that ' is the function
that minimizes the functional J (f ) defined by
11. A maximal inequality for reversible Markov processes 347
X
J ( f ) = (x) f 2(x) +
D(f )
1
(11:3)
x2E
among all functions that are equal to 1 on G. Indeed, since (jgj ^ A)A 1
is equal
to 1 on G, (11.2) is bounded above by
nX o
A 2
(x) (jg(x)j ^ A)2 +
D(jgj ^ A) :
1
x2E
By property (10.1) of the Dirichlet form, this expression is bounded above by
nX o
A 2
(x) g(x)2 +
D(g) :
1
x2E
It remains to choose = T 1 .
To conclude the proof of the theorem we need thus to show that ' is the
function that minimizes the functional J (f ) defined by (11.3) among all functions
that are equal to 1 on G.
A function h: E ! R that is equal to 1 on the set G and that minimizes the
functional J (f ) must satisfy the identity
X
h(x) = (x) p(x; y) [h(y) h(x)] = Lh(x)
y2E
for x 2 Gc . There is at most one function identically equal to 1 on G and satisfying
the previous relation on Gc . Indeed, assume h1 and h2 share this property. Then
h̄ = h1 h2 vanishes on G and Lh̄ = h̄ on Gc . Multiplying this identity by h̄
and integrating it with respect to we obtain that
< h̄; h̄ > = < h̄; Lh̄ >

because h̄ vanishes on G. Since the left hand side is positive and the right hand
side is negative in virtue of Proposition 10.1, both expressions vanish, so that
h1 = h2 .
Since we proved in the beginning that L' = ' on Gc , ' is the unique
function that minimizes the functional J (f ) among all functions that are equal to
1 on G. This concludes the proof of the theorem.
Appendix 2. The Equivalence of Ensembles, Large
Deviation Tools and Weak Solutions of Quasi–Linear
Differential Equations
In the first two sections of this chapter we prove a uniform local central limit
expansion for exponential families of independent and identically distributed ran-
dom variables. This expansion permits to prove the equivalence of ensembles :
the marginals of the canonical measures N;K converge to the marginals of the
grand canonical measure as N " 1 and K=N d ! , uniformly on compact
sets for the density .
In section 2, in the context of generalized exclusion processes, we prove a
second order expansion for the expectation of a cylinder function with respect to
the canonical measure. Such an expansion is needed in the proof of a sharp estimate
for the largest negative eigenvalue of the generator of a reversible generalized
exclusion process restricted to a finite cube, to be proved in the next appendix.
In section 3 we prove two general results on large deviations that are used
in the investigation of the large deviations of the empirical measure around its
hydrodynamic limit.
Finally, in sections 4 and 5, we fix the terminology of weak solutions of
quasi–linear parabolic and hyperbolic differential equations and state several re-
sults without proofs that are quoted throughout the book.
Note: Throughout this section, to keep notation simple, for a positive integer k ,
we denote by k? the number 2k + 1.
1. Local central limit theorem and equivalence of ensembles
Let p() be a probability distribution on N and assume that 0 < p(0) < 1. To fix
ideas one may consider the one site marginal of a zero range distribution : p(k ) =
Z (') 1f'k =g(k)!g for some ' > 0. Define the partition function Z : R+ ! R+
associated to this distribution :
X
Z (') = 'k p( k )
k0
and denote by ' the radius of convergence of this series. We shall assume that
Z () diverges at the boundary of its domain of definition :
350 Appendix 2. General tools
!' Z (') =
'lim 1: (1:1)
Define the exponential family of distributions fp'(); ' 2 [0; ')g by

p'(k) k
Z (') ' p(k) k in N .
1
= for
Denote by R(') the mean of the distribution p' :

X
R ( ') = kp'(k) = ' @' log Z (') :
k0
A simple computation shows that log Z (e ) is a strictly convex function in and
therefore that R() is strictly increasing. In particular, if [0; ) stands for the range
of R(), R is a bijection from [0; ' ) to [0; ).
Notice that is equal to 1 unless the probability distribution p() is finitely
supported. Indeed, assume there exists k 0 such that p(k ) = 0 for k k 0 . Denote by
the size of the support of p() : = maxfk; p(k) > 0g. It is easy to show that in
this case = . Otherwise, assumption (1.1) guarantees that = 1. The proof
presented in section 2.3 for zero range distributions applies without modifications
to the present general context. Hence, for each 0 in [0; ) there is a unique
distribution p'0 with expectation 0 and if we denote by () the inverse function
of R() : = R 1 ,
X
= k p()(k) for each in [0; ).
k 0
For each 0 ' < ' and each finite subset of Zd, denote by ¯' (resp.
¯' ) the product measure on N Zd (resp. N ) with one site marginal equal to p' .
d
Configurations of N Z or N are indifferently denoted by Greek letters , or
. For each positive integer K , let ;K be the measure ¯' conditioned on the
hyperplane ;K of all configurations of N with K particles and at most
particles per site :
n X o

;K = 2 N ; (x) = K and (y) for y 2 ;

x2 (1:2)
X
;K () = ¯' (x) = K :
x2
We shall often omit the superscript in ;K . Notice that the right hand side of
the second line does not depend on the parameter '. The probability measures
;K () are called the canonical measures and the product measures ¯' are called
the grand canonical measures. To keep notation simple, we abbreviate N ;K and
N ;K by N;K and N;K and we denote the linear size of N (equal to 2N + 1)
by N? . The purpose of this section is to show that the finite dimensional marginals
of the canonical measures N;K converge, as N " 1 and K=N?d ! , uniformly
1. Local central limit theorem and equivalence of ensembles 351
on compact sets of the parameter ', to the finite dimensional marginals of the
grand canonical measure with density :
Theorem 1.1 For every positive integer `, every bounded function f : N ` !R

and every B > 0,

lim sup
N !1 0K BN?
EN;K [f ] E¯ K=N?d [f ] = 0: (1:3)
d ( )
The proof of this result relies on a local central limit theorem for independent
random variables with distribution p' . To prove such an expansion uniformly over
compact sets in the parameter ', a set of three assumptions is required. To state
these hypotheses, for each positive integer k , denote respectively by k ('), !k (')
and k (') the central moment, the absolute central moment and the cumulant of
order k of the distribution p' :
h i h i
k (') = E¯' ((0) R('))k ; !k (') = E¯' (0) R(')k ;
X k 1
Y j mj :
k (') k! ( 1)m1++mk 1 (m1 + + mk
j =1 mj ! j !
= 1)!
In this last formula summation is carried over all nonnegative integer solutions of
m1 + + kmk = k and j stands for the j -th moment of (0) : j = E¯' [(0)j ].
We shall often write (')2 for 2 ('). Notice also that 3 (') = 3 (').
Denote by v' (t) the normalized characteristic function associated to the dis-
tribution p' :
v' (t) = E¯' exp it[(0)(')R(')] :
h n oi
For each 0 '0 < '1 ' and k in N consider the following hypotheses :
(CL1) There exists a finite constant A0 such that
sup !k (')= (')k A0

'0 '<'1
for all 2 k k + 1.
(CL2) For every > 0, there exists C () < 1 such that

sup sup v' (t) C ( ) :
'0 '<'1 jtj(')
(CL3) There exists > 0 so that

Z
v' (t) dt C < 1 :

sup
'0 '<'1 jtj(')
Remark 1.2 Assumptions (CL1)–(CL3) are verified on any compact subset of

(0; ' ) for any positive integer k .
To prove this claim fix an interval ['0 ; '1 ] contained in (0; ' ) and a positive
integer k . The proof of Lemma 2.3.5 can easily be adapted to show that ¯' is an
increasing family of probability measures. By assumption (1.1), E¯ '1 [expf0 (0)g]
is finite for some positive 0 . Therefore, the expectation E¯ ' [expf0 (0)g] is finite
for all '0 ' '1 and a fortiori !k (') is finite on ['0 ; '1 ] for each 2 k k .
Since (') is a smooth function strictly positive on (0; '), assumption (CL1) holds
on ['0 ; '1 ]. On the other hand, v' (t) is a continuous function of the pair ('; t).
Assumptions (CL2) and (CL3) follow therefore from the fact that (') is a smooth
strictly positive function on (0; ' ) and that jv' (t ('))j < 1 for each fixed (t; ')
in (0; ] (0; ' ) .
We may now state the local central limit theorem uniform over compact sets on
the parameter '. For each 0 ' < ' , let fXj ; j 1g be a family of independent
identically distributed random variables on a probability space (
; A; P' ) with
common distribution p' . For m 0, denote by Hm (x) the Hermite polynomial
of degree m :
m=
Hm (x) = ( 1)mex =2
2 dm e x =22
m!
X2]
[
( 1)k
xm 2k
dxm =
k=0 k!(m 2k)!2k
Here [m=2] stands for the integer part of m=2. Let q0 (x) denote the density of the
normalized Gaussian distribution and, for j 1, let
j km
qj (x) p1 e x2 =2 X Hj +2a (x) Y 1 m+2(') ;
m+2
m=1 km ! (m + 2)! (')
=
2
where the summation is carried out over all non negative integer solutions of
k1 + 2k2 + + jkj = j and k1 + k2 + + kj = a.
Theorem 1.3 Assume hypotheses (CL1)–(CL3) for some 0 '0 < '1 ' and
some k 2. There exist constants E0 = E0 ('0 ; '1 ; k ) and n0 = n0 ('0 ; '1 ; k )
such that
n 2
kX
n(kE01)=2
p hX i
sup sup n(' P'
)2 Xi = K n j=2 qj (x)
1
'0 ''1 K 0 i=1 j =0
for all n n0 . In this formula x = (K nR('))=(')pn.
The proof of this theorem is omitted since it follows closely the classical
arguments given for instance in Petrov (1975) (Theorem VII.12). There is only a
slight problem to control the integral I3 in the proof of this theorem but hypotheses
(CL2) and (CL3) are built to estimate this integral.
Theorem 1.1 under assumptions (CL1)–(CL3) is easily deduced from this re-
sult.
Corollary 1.4 Assume hypotheses (CL1)–(CL3) for '0 < '1 and k = 4. There
exists a constant C (E0 ; A0 ) depending only on A0 and E0 , the constants appearing
in assumption (CL1) and in the statement of Theorem 1.3, such that for every
positive integer ` and every cylinder function f : N ` ! R with finite second
moments with respect to ¯' , '0 ' '1 ,

EN;K [f ]

E f
¯(K=N?d ) [ ]
r
C (E0; A0 ) N`?
d h 2 i
E¯ K=N?d f E¯ K=N?d [f ]
? ( ) ( )
for all N 2` and all R('0 ) K=N?d R('1 ). In particular, for all bounded
cylinder function f ,

lim sup sup N d EN;K [f ]

E¯ K=N?d [f ] < 1 :
N !1 R('0 )K=N? R('1 )
d ( )
Proof. The difference jEN;K [f ] E¯ K=N?d [f ] j may be written as

( )
P
N x2N ` (x) = K M ( )
¯'
X `

f E¯' [f ]
¯' ( ) ( ) P 1 (1:4)
2N` ¯'N x2 (x) = K N
P
where M ( ) stands for x2` (x) and ' = (K=N?d ). Theorem 1.3 and ele-
mentary estimates give that the absolute value of the expression inside braces is
bounded above by

R(') d ¯ R(') 2
C (A0 ; E0 ) N`?
d ¯
?
1 +
(') + `? (')2
for all N large Penough. In this formula, ¯ stands for the density of particles in
` : ¯ = j` j 1 x2` (x) and C is a constant depending only on A0 and E0 ,
the constants appearing in hypothesis (CL1) and in Theorem 1.3. By Schwarz
inequality and assumption (CL1), expression (1.4) is therefore bounded above by
r
C 0 (A0 ; E0 ) N`? 1 + `? d=2 E¯' f
d n o h 2 i
E¯' [f ] :
?
To prove Theorem 1.1 it remains therefore to check assumptions (CL1)–(CL3)

for '0 = 0, '1 < ' and k = 4. We start with (CL1). By assumption (1.1),
!k (')=(')k is a smooth function on (0; ' ) and is therefore bounded on any
compact subset of (0; ' ). However, for ' close to the origin, a straightforward
expansion shows that
k (') p(1) ' O('2 ) k2:

=
p(0) + for all
Assumption (CL1) fails therefore close to the origin because k (')= (')k =
O('1 (k=2) ) and we are forced to consider separately the case of small densi-
ties. The analysis relies on the following alternative version of the local central
limit theorem.
Theorem 1.5 For all 0 < '0 < ' and k0 2, there exist finite constants
E1 = E1 ('0 ; k0 ) and E2 = E2 ('0 ; k0 ) such that
n kX
((')2nE)1(k
p hX i 0 2

sup n(')2 P' Xi = K 1
j= qj ( x )
=
j =0 n

K 0 i=1
2 0 1) 2
for all ' '0 such that (')2 n E2 . Here again x = (K nR('))=(')pn.
We omit the proof of this theorem because it follows closely the classical proof
(Theorem VII.12 of Petrov (1975)). There are only two modifications needed.
Firstly, the integral I1 must be redefined as the integral of jfn (t) uk;n (t)j over
the region jtj < (n (')2 ) for some < 1=6. Secondly, to estimate the integral
I3 we may proceed as follows. Denote by v~' (t) the characteristic function of
X1 R(') under P' : v~' (t) = E' [expfit(X1 R('))g]. It is easy to see that

v t
~' ( )2 1 P' [X1 = 0]P'[X1 = 1]fcos t 1g C'fcos t 1g
for some constant C that depends only on '0 and on the original distribution p()
because
P' [X1 = 0] = 1 + O(') P' [X1 = 1] p(1) ' O('2 )

and =
p(0) +
for ' close to the origin. These estimates permit to bound the integral I3 .
Recall that (')2 = p(1)p(0) 1' + O('2 ) and R(') = p(1)p(0) 1' + O('2 ).
Therefore, for all '0 < ' , there exists finite positive constants C1 = C1 (p; '0 )
and C2 = C2 (p; '0 ) such that
0 < C1
(')2 C < 1 (1:5)
R(') 2
for 0 ' '0 . In particular, taking in Theorem 1.5 ' = (K=n), we have that
n(')2 nR(') = K . The remainder in Theorem 1.5 is thus of the order of K 1 ,
the inverse of the total number of particles.
With the same arguments presented in the proof of Corollary 1.4, we obtain
the following estimate for the expectation of cylinder functions with respect to
canonical measures.
Corollary 1.6 Fix 0 < '0 < ' , a positive integer ` and a cylinder function
f : N` ! R with finite second moment with respect to ¯' for all ' '0 . There
exist finite constants E1 = E1 ('0 ) and E2 = E2 ('0 ) such that

EN;K [f ] E¯ K=N?d [f ]
( )

E1 N`? (1')2 E¯' f E¯' [f ]
d h i
? (1:6)
r
h 2 i
+
1
( ') E¯' f E¯' [f ]
for all N 2` and all K such that (K=N?d ) '0 , ((K=N?d ))2 N?d E2. On
the right hand side of the inequality ' = (K=N?d ).
Theorem 1.5 also provides a simple proof of the equivalence of ensembles.
Corollary 1.7 Fix 0 < '0 < ' , a positive integer ` and a cylinder function
f : N ` ! R with finite second moment :
h i
sup E¯' jf j2 < 1 :
''0
0
There exists a finite constant E1 = E1 ('0 ; f ) such that
E¯ K=N?d [f ] NE1d

EN;K [f ]

? ( )
for all large enough N and all 0 K=N? R('0 ).

d
Proof. Fix '0 > 0 and a cylinder functions f : N ` ! R with finite second
moment. Denote by E1 and E2 the constants introduced in Corollary 1.6. To
prove this lemma for N large enough and K so that ((K=N?d ))2 N?d E2 ,
we just need to show that the expression inside braces in (1.6) is bounded. This
expression is of course bounded for ' on any compact subset (0; '0 ] because f
has finite second moment and (')2 is strictly positive. We just need therefore
to investigate the behavior of this expression for ' close to the origin. Since by
assumption E¯ ' [jf j2 ] is finite, we have
f (dx) f (0) pp(1)

X h i
E¯' [f ] = f (0) +
(0)
' + Of ('2 ) ;
x2`
f (dx) f (0) pp(1)
h i X h i2
E¯' f E¯' [f ] 2 = (0)
' + Of ('2 )
x2`
and
h i
E¯' f E¯' [f ]

X
h
f f
i
X
f f p
(1)
= (dx )
x2`
(0) + (dx )
x2`
(0)
p(0) ' + Of ('2 ) :
In these formulas 0 stands for the configuration of N ` with no particles; for x in

` , dx stands for the configuration with no particles but one at site x and Of ('2 )
indicates a constant bounded in absolute value by C (f )'2 . The derivation of these
expansions is based on the following simple estimate :
j` j ' M ()
¯' () ¯' () Z (p'0 )

(1:7)
0
0 '0
that
P holds for every ' '0 and configuration of N ` provided M ( ) =
x2` ( x ).
Replacing f by (0) we obtain that (')2 = fp(1)=p(0g' + O('2 ). A straight-
forward computation shows then that the right hand side of (1.6) is bounded
above by C (E1 ; f )N? d . To conclude the proof of the corollary it remains to
consider the cases not covered by Corollary 1.6, i.e., densities K=N?d such that
0 ((K=N?d ))2 N?d E2 .
Fix such value of K . By inequality (1.5) it follows that K is less than or equal
to C1 ('0 ; p) 1 E2 . By estimate (1.7), the difference jEN;K [f ] E¯(K=N?d ) [f ] j
is bounded above by
h X i h X i

f
(0)
N;K ( x) = 0 ¯(K=N?d ) x
( ) = 0
x2` x2`
h X i
+ C (f; K )N;K (x) 1 + C (f; '0 ; `)E¯' jf j (K=N?d )
x2`
0
h X i h X i
2C (f; K ) N;K (x) 1 + ¯(K=N?d ) ( x) 1
x2` x2`
+ C (f; '0 ; K )N d
because R(') = p(1)p(0) 1' + O('2 ). In these formulas C (f; K ) stands for the
maximum value of the cylinder function f over all configurations of N `
with less than K particles : C (f; K ) = max;P
x2` (x)K
jf ( )j. By Cheby-
chev inequality this last expression is bounded above by C (f; '0 )N? d because
K C1 ('0 ; p) 1E2 . This concludes the proof of the corollary.
We have seen in this section that the equivalence of ensembles, just proved
for product measures or independent random variables, relies strongly on the local
central limit theorem. This later result has also been proved for Gibbs measures.
We refer to DelGrosso (1974), Dobrushin and Tirozzi (1977) and the references
therein.
2. On the local central limit theorem 357
2. On the local central limit theorem
We proved in the previous section a local central limit theorem, uniform over
the density, for exponential families of independent identically distributed random
variables. This uniform local central limit theorem provided some estimates on the
expectation of local functions with respect to canonical measures. Theorem 1.1
is a typical example of such an estimate. We prove in this section more refined
estimates that are needed in the proof of a spectral gap for conservative dynamics.
We start with a second order expansion of the expectation of cylinder functions
with respect to canonical measures. In order to deduce it we just need to recall
from Theorem 1.5 the second order expansion of the local central limit theorem
and the proof of Corollary 1.4.
Lemma 2.1 Fix '0 < ' , a positive integer ` and a cylinder function f : N ` !
R. There exist finite constants E1 = E1 ('0 ) and E2 = E2 ('0 ) such that

EN;K [f ] E¯ K=N?d [f ]
( )
`? d
3 (') < f ; ¯ > `d? < f ; f¯ g2 >

N? ( ') 4
¯'
(')2 ¯'
N 3d=E21 (')2 < f ; f >1¯='2

for all N 2` and all K such that (K=N?d ) '0 , ((K=N?d ))2 N?d E2 . In
this formula < f ; g >¯ ' represents the covariance of f and g with respect to ¯' ,
' stands for (K=N?d) and ¯ = `? dM ( ).
In the remaining of this section we consider generalized exclusion processes
introduced in section 2.4. Recall that in this model at most particles are allowed
per site and that the grand canonical measures f¯' ; ' 0g are given by
¯' f; (0) = j g 'j

1 + ' + + '
=
for 0 j .
This model presents a special feature, known as the particles–holes duality,
that simplifies some computations : for each 0 j ,
¯' f; (0) = j g = ¯1=' f ; (0) = j g :
This identity can be interpreted as follows. Under ¯' , the distribution of holes is
equal to the distribution of particles under ¯1=' . In particular, to prove any state-
ment concerning the the distribution of ' , it is enough to prove it for ' 1. This
property is used below in the proof of Lemmas 2.2 and 2.3 and in the investigation
of the spectral gap for generalized symmetric simple exclusion processes. This du-
ality between holes and particles provides also some simple relations among the
central moment of order k of different distributions that may be useful when

computing expansions around 0 or 1. It follows immediately from the previous
identity that
R(') = R (' 1
) and j (') = ( 1)j j (' 1
)
for all j 2.
Lemma 2.2 Fix a positive integer `. For all cylinder functions f : ` ! R,
there exists a constant E0 = E0 (f ) such that
E¯ K=N?d [f ] EN0 (df )

EN;K [f ] ( )
(2:1)
for all N ` and all 0 K N?d.

Proof. Since there is at most a finite number of particles per site, all cylinder
functions are bounded and therefore the assumptions of Corollary 1.7 are satisfied.
In particular, inequality (2.1) is satisfied for N large enough and 0 K
(2=3)N?d.
Consider now holes as particles and particles as holes, i.e., consider the prob-
ability measure q () on f0; : : : ; g defined by q (k ) = p( k ). Corollary 1.7
applied to q () states that inequality (2.1) is fulfilled for all N large enough and
all (=3)N?d K N?d . This proves inequality (2.1) for all N large enough.
Since f is a cylinder function (and therefore bounded), to extend inequality
(2.1) for small values of N , it is enough to modify, if necessary, the value of the
constant E0 (f ).
Lemma 2.3 Denote by h( ) the cylinder function [1+ (0)]1f (0) < g. For every
" > 0, there exists N 0 = N 0 (") and '0 = '0 (") such that

EN;K [h((0))]
+1 EN;K [h((0))]
N"d
h n oi

1
K +1 E N;K +1 (0) h( (0)) h((0) 1)
for all N N 0 and all K=N?d < R('0 ).

Proof. Fix " > 0. Here is the idea of the proof. In the case where N d (')2 is
large, this statement follows from the second order expansion for the marginal
probabilities of the canonical measures proved in Lemma 2.1. In the case where
N d(')2 (and therefore K ) is small we prove the result by inspection.
We start with the case where the total number of particles K is large. Take
'0 such that R('0 ) = 1, denote by E1 , E2 the two constants given by Lemma 2.1
and by E0 the constant introduced in Lemma 2.2
Let K 0 = K 0 (") be defined as
n o
K0 = max C1 1 E12 " 2 ; C1 1 E2 ; E0 ( + 1)2" 1
;
where C1 is the lower bound for (')2 =R(') obtained in (1.5) and set '~ =
(K=N?d). Notice that for K 0 K N?d we have ('~)2 N?d E2 because, by
inequality (1.5), ('~)2 N?d = K ('~)2 R('~) 1 KC1 and we chose K 0 C1 1 E2 .
We are thus entitled to estimate the expectation EN;K [h( (0))] with Lemma 2.1
for K 0 K N?d :
EN;K [h] = E¯' [h]
~
i
1 3 ('~) E hh; (0)i 1
E
h
h f g
('~)4 ¯' ('~)2 ¯'
2
N?d (2:2)
+ ~ ; (0) ~
E1 E [h ; h]1=2
¯
N? ('~)2 '
3d=2 ~
We claim that the expression inside braces in this formula vanishes as '~ # 0.
Indeed, performing a change of variables = + d0 , we obtain that
h i h i
E¯' h((0))f ()
~ =
1
'~E ¯ '~ (0)f (
d0 )
for every cylinder function f . Let = R('¯ ). Applying this identity to the cylinder
functions 1, (0) and ( (0) )2 we obtain that the expression inside braces
in (2.2) is equal to
3 E h(0)f(0) g
i 1
E
h
(0)f(0) g2
i
~ 4 '
' ¯ ~ 1
'
~ 2
¯ '
~ 1
1 n 4 o
+ E¯ ' [ (0)] 2 3 2 :
1
'~ ~ =
'
~ 4
Expanding the functions R('), (')2 and 3 (') around their value at the origin
we obtain that
R(') = ' + '2 + O('3 ) ; (')2 = ' + 2'2 + O('3 )
and 3 (') = ' + 4'2 + O('3 ) :
It is easy easy to prove from these estimates that ' 1 (') 4 f2 4 3 2 g
is of order O(') for ' close to the origin.
This proves that the expression inside braces in formula (2.2) is of order O(')
for ' close to the origin. A similar argument proves that E¯ ' [h; h]= (')2 is
bounded on compact intervals of R+ . Therefore, for all N 1, K 0 K N?d
s
R ('~)E12
EN;K [h] = E¯' [h] + N d O('~) ('~)2 K 1
~
because K=N?d = R('~). By inequality (1.5) and since K K 0 C1 1 E12 " 2 the
absolute value of the remainder is bounded by N d fO('~) + "g. From this estimate
it follows that
EN;K [h] EN;K [h] = h~

K + 1 h~ NKd

R";N;'~ (2:3)
+1
N?d ?
+
provided K 0 K N?d . In this formula h~() stands for the expectation of h

with respect to the grand canonical measure with density : h~() = E¯ () [h] and
R";N;'~ is a remainder, bounded in absolute value by N dfO('~) + "g.
We estimate now EN;K [ (0)fh( (0)) h( (0) 1)g]. This cylinder function
may be rewritten as (0) ( + 1)1f (0) = g. By Lemma 2.2,
E¯ K=N?d 1f(0) = g NE0d

h i h i
EN;K 1f(0) = g = ( )
for all N 1 and 0 K N?d . Notice that the expectation on the right hand
side is of order O('~ ). Therefore,
h i
E +1 (0)fh( (0))
h((0) 1)g
1
K +1 N;K
1 C ()O('~ )
NE0d((K++1)1)
2
=
N?d +
N?dR('~)
Since 2, R(') = ' + O('2 ) and K 0 K N?d , the last two terms are
bounded in absolute value by N d fO('~) + "g because by definition K K 0
( + 1)2 E0 " 1 .
In view of this estimate and (2.3), to conclude the proof of the lemma in the
case K 0 K N?d , it remains to show that
h~ KN+d 1 h~ NKd
1
? ? N ?d
is bounded by "=N d for '~ small enough. Denote by h~0 () the derivative of h~
with respect to the density . A straightforward computation shows that h~0 () =
1 + O(). In particular, by Taylor expansion, h~((K + 1)=N?d) h~(K=N?d ) 1=N?d is
bounded above by O('~)N? d . This concludes the proof of the lemma in the case
K 0 K N?d.
It remains to consider the case K K 0 . For a finite subset of Zd, denote by
Z (; K ) the total number of configurations of f0; : : : ; g with K particles. We
shall expand this expression in since the total number of particles K is bounded
by K 0 . The main contribution to Z (; K ) is the set of all configuration with no
site occupied by more than one particle. There are O(jjK ) of such configurations.
The second main contribution to Z (; K ) is the set of all configurations with one
site occupied by two particles and all other sites occupied by at most one particle.
There are O(jjK 1 ) of such configurations. We have therefore that

j j j j
Z(; K ) = K + jj K 2 + C (K )O(jjK 2)
1

j j K (K 1) C (K )
K !(jj K )! 1 + jj K + 1 jj2 :
!
=
Since the canonical measure N;K is the uniform measure on N;K , N;K f ;
(0) = ag is equal to the proportion of configurations with a particles at the origin :
N;K f; (0) = ag Z (N f0g; K a) : (2:4)

=
Z (N ; K )
This identity and the explicit formula for Z (; K ) shows that the probability
N;K f; (0) 2g is bounded by C (K )N? 2d. In particular, since h is bounded,
to prove the lemma in the case K K 0 , we have only to show that
h i
h(0) N;K +1f; (0) = 0g N;K f; (0) = 0g
h i
+ h(1) N;K +1 f ; (0) = 1g N;K f; (0) = 1g
K + 1 [h(0) h(1)]N;K +1f; (0) = 1g

1
+
is bounded by "N d for all N large enough. Since N;K f ; (0) = 0g is equal to
1 N;K f ; (0) 1g and since h(1) h(0) = 1, it is enough to prove that
K
K + 1 N;K +1f; (0) = 1g N;K f; (0) = 1g (2:5)
is bounded by "N d for all N large enough. From the explicit formula for
Z (; K ), we obtain that
N;K f; (0) = 1g = K 2(K 1)
n o
2d
L 1
N?d + 1 K C ( K ) N ? :
This identity shows that the difference (2.5) is equal to 2K=(N?d K )(N?d + 1
K ) C (K )N? 2d, what concludes the proof of the lemma.
Lemma 2.4 There exist a universal constant B2 = B2 () so that
N;K f; (0) = ag B
sup
N;K f; (0) = bg 2
for all a b. In this formula the supremum is taken over all N and K such that
that K=N?d =2.
Proof. Fix '0 > 0 so that R('0 ) = =2. Denote by E1 and E2 the con-
stants introduced in the statement of Theorem 1.5 and by C1 the lower bound
of inf0=2 ()2 = provided by inequality (1.5). Set K 0 = E2 C1 1 . Notice that
for K 0 K (=2)N?d,
((K=N?d))2N?d = ((K=N?d))2 KR((K=N?d)) 1 C1 K E2
by definition of K 0 . We are thus entitled to apply Theorem 1.5.
By definition of the canonical measure,
p nP o
N;K f; (0) = ag 2 (jN j 1)¯' x2N f0g (x) = K
a
N;K f; (0) = bg = 'a b p nP o ;
2 (jN j 1)¯' x2N f0g ( x ) = K b
where ' = (K=N?d ). Theorem 1.5 and straightforward computations show that
the last ratio is bounded by some constant C depending only on for all N large
enough and all K 0 K (=2)N?d.
It remains to examine the asymptotic behavior of N;K f ; (0) = ag=N;K f ;
(0) = bg for K K 0 . Recall the definition of Z (; K ) given in the proof of
Lemma 2.3. By formula (2.4), we have
N;K f; (0) = ag Z(N f0g; K a) C (K )O(jN jb a ) :
N;K f; (0) = bg =
Z(N f0g; K b) =
This concludes the proof of the lemma.
3. Remarks on Large Deviations
We prove in this section two general results on large deviations needed in Chapter
10. The first result is due to Varadhan (1966) and the second one to Donsker and
Varadhan (1975c).
Consider a sequence of random variables fXN ; N 1g taking values in some
metric space E . We say that the sequence fXN ; N 1g satisfies a large deviation
principle with rate function I and decay rate aN if for every closed set F of E
1
log P [XN 2 F] inf I (u)
N !1 aN
lim sup
u2F
and for every open set G of E
N !1 aN
lim inf
1
log P [XN 2 G] inf I (u)
u2G
for some positive function I : E ! R+ and some increasing sequence aN " 1.
Notice that we didn’t require the rate function to be lower semicontinuous or
the levels sets of I to be compact. The reason is that these assumptions are not
needed in the next theorem.
Theorem 3.1 (Laplace–Varadhan) Let fXN ; N 1g be a sequence of random

variables satisfying a large deviation principle with rate function I and decay rate
aN . Let F : E ! R be a bounded continuous function. Then,
h i
log E eaN F (XN ) = sup fF (u) I (u)g : (3:1)
1
N !1 aN
lim
u2E
Proof. We start with the lower bound which is easier. Fix " > 0. There exists u0
in E such that supu2E fF (u) I (u)g F (u0 ) I (u0 ) + ". Fix a neighborhood V
of u0 such that F (u) F (u0 ) " for all u in V . In particular,
3. Remarks on Large Deviations 363
h i h i
log E eaN F (XN ) a1 log E eaN F (XN )1fXN 2 Vg
1
aN N

F (u0) " +
a
1
log P XN 2 V :
N
By the lower bound of the large deviation principle and since u0 belongs to V,
the limit, as N " 1, of the previous expression is bounded below by
F ( u0 ) " inf I (u)

u2V
F (u0) I (u0) " :
Since supu2E fF (u) I (u)g F (u0 ) I (u0 ) + ", we proved that the left hand
side of (3.1) is bounded below by the right hand side.
We now turn to the upper bound. Let B = kF k1 . Fix a positive integer K and
divide the interval [ B; B ] in K contiguous closed intervals of length 2B=K .
Denote these intervals by Ij = [rj ; rj +1 ], 1 j K , and denote by Fj the pre-
image by F of Ij : Fj = fu 2 E ; F (u) 2 Ij g. The subsets Fj are closed because
F is continuous. On the other hand, by definition,
h i K
X h i
E eaN F (XN ) erj aN P XN 2 Fj :
+1
j =1
Since aN " 1,
n o
lim sup aN1 logfdN + bN g = max lim sup aN1 log dN ; lim sup aN1 log bN :
N !1 N !1 N !1
(3:2)
By the large deviations principle, the left hand side of (3.1) is bounded above by
n o
max
1j K
rj+1 + sup I (u)
u2Fj
n o
1max r + sup fF (u) I (u)g +
j K j +1
sup F (u) :
u2Fj u2Fj
Since on Fj F takes value in the interval [rj ; rj +1 ] and rj +1 rj = 2B=K , this
last expression is bounded above by supu2E fF (u) I (u)g + 2B=K . This shows
that the left hand side of (3.1) is bounded above by the right hand side.
Lemma 3.2 (Minimax lemma) Let K be a compact set of a polish space E , let
A be some parameter set and let fJ : E ! R; 2 Ag be a family of upper
semi–continuous functions. Then,
max inf sup J ( )

O1 ;:::;On 1jn 2A 2Oj
inf sup inf J () :
2K 2A
In the left hand side, the first infimum is carried over all finite open covers
fO1; : : : ; On g of K.
Proof. Denote the right hand side by R and fix > 0. Since for each in
K, inf2A J () R, there exists () such that J() () R + . Moreover,
since each function J is assumed to be upper semi–continuous, there exists a
neighborhood O of so that
sup J() ( ) R + 2 :
2O
The family fO ; 2 Kg constitutes an open cover of the compact set K and has
a finite subcover that we denote O1 ; : : : ; On . Since
sup J(j ) ( ) R + 2 for 1 j n;
2Oj
we have that
inf sup J ( ) R + 2 for 1 j n :
2A 2Oj
It is now easy to conclude the proof of the lemma.
This result must be understood as follows. Under very general conditions
since we do not impose any restriction on the index set A and require only the
upper semi–continuity of each function J , it allows the replacement of inf sup
by sup inf. This exchange constitutes one of the main technical difficulties in the
proof of the large deviations upper bound as can be seen in Chapter 10. The next
result illustrates how to obtain a large deviations upper bound for compact sets
from the minimax lemma and some upper bounds for open sets.
Lemma 3.3 Consider a polish space E and a sequence of probabilities PN on

E . Let fJ : E ! R; 2 Ag be a family of upper–semi continuous functions
indexed by some set A. Assume that we are able to prove upper bounds for every
open subset O of E :
log PN [O] inf sup J () : (3:3)

1
N !1 aN 2A 2O
lim sup
Then, for every compact set K,
log PN [K] sup inf J () :

1
N !1 aN
lim sup
2K 2A
Proof. Let O1 ; : : : ; On denote a finite open cover of K. By remark (3.2) and

assumption (3.3),
log PN [K] max inf sup J () :

1
a
lim sup
N !1 N 1j n 2A 2Oj
To conclude the proof it remains to minimize over all finite open covers of K and
recall the statement of the minimax lemma.
4. Weak solutions of nonlinear parabolic equations 365
4. Weak solutions of nonlinear parabolic equations
We briefly fix in this section the terminology of weak solutions of non linear
parabolic equations and we present the main existence, uniqueness and regularity
results of such equations. The reader is referred to Ladyženskaja et al. (1968) for
a complete treatment of the question. Hereafter is a smooth strictly increasing
function such that k0 k1 g < 1 and = fi;j ; 1 i; j dg is a symmetric
positive definite matrix : i;j = j;i and there exists > 0 such that v v jv j2
for all v in Rd .
We start considering the problem of the existence and uniqueness of classical
solutions.
Theorem 4.1 Fix " > 0. For each initial profile 0 : Td ! R of class C 2+" (Td),
there exists a unique classical solution of class C 1+";2+" (R+ Td ) of the Cauchy
problem 8 X
< @t =
> i;j @u2 i ;uj ()
1i;j d (4:1)
>
:
(0; ) = 0 () :
Moreover, a maximal principle holds :
inf
u2Td
0 (u) (t;u)2infR Td (t; u) sup (t; u) sup 0 (u) : (4:2)
+ (t;u)2R Td u2Td
+
We now turn to weak solutions
Definition 4.2 Fix a bounded initial profile 0 : Td ! R. A measurable function

: R+ Td ! R is a weak solution1;2of the Cauchy problem (4.1) if for every
function G: R+ Td ! R of class CK (R+ Td )
Z 1 Z n X o
dt d
du (t; u) @tG + ((t; u)) i;j @u2 i ;uj G
0 T i;jd
1
Z
+ du G(0; u)0(u) = 0:
Td
The following a priori estimate (cf. Oleinik and Kružkov (1961)) on bounded
weak solutions of quasi–linear parabolic equations due to Nash plays a central role
in the investigation of the existence and uniqueness of weak solutions. It states
that bounded weak solutions are uniformly Hölder continuous on each compact
subset of (0; 1) Td .
Theorem 4.3 Fix a bounded profile 0 and a bounded weak solution (t; u) of
(4.1). There exist constants a and A depending only on the dimension d, on and
on such that for every 0 < s t,
j(t; u) (s; v)j A kk1 jupsvj + tpss

n a a=2(1+a) o
:
It follows from this estimate and Theorem 4.1 that there exists a bounded weak
solution of the Cauchy problem (4.1) for bounded initial profiles 0 . In fact, it is
not difficult to prove the existence and uniqueness of weak solutions in the class
of measurable functions in L2 ([0; T ] Td ).
Theorem 4.4 Fix a bounded profile 0 . There exists a unique weak solution of the
quasi–linear parabolic equation (4.1) that belongs to L2 ([0; T ] Td ). Moreover,
the solution is uniformly Hölder continuous on each compact subset of (0; 1) Td
and satisfies the maximum principle (4.2).
Proof. Fix a bounded profile 0 and consider a sequence of smooth profiles

f"0 ; " > 0g bounded by k0k1 and converging weakly to 0 in the sense that
Z Z
lim
"!0
du H (u)"0 (u) = du H (u)0(u) (4:3)
for all continuous function H : Td ! R. For each " > 0, denote by " (t; u) the
unique classical solution of equation (4.1) with initial data "0 . By the maximum
principle (4.2) the sequence " is uniformly bounded. On the other hand, by
Theorem 4.3, on each compact subset of (0; 1) Td , the sequence f" ; " > 0g
is uniformly Hölder continuous. The sequence is therefore relatively compact (for
the uniform topology) on each compact set of (0; 1) Td and we may obtain a
subsequence "k that converges uniformly on each compact subset of (0; 1) Td
to a bounded function . It is very easy to show that is a weak solution of (4.1),
what proves the existence of a bounded weak solution. It satisfies, moreover, the
maximum principle (4.2) and is uniformly Hölder continuous on each compact
subset of (0; 1) Td .
To prove uniqueness in the class L2 ([0; T ] Td ), we need to introduce some
notation. For each z in Zd, denote by z : Td ! C the L2 (Td ) function defined by
(u) = expf(2i)z ug. Here z u stands for the inner product in Rd . It is well
known that f z ; z 2 Zdg forms an orthonormal basis of L2 (Td ). In particular,
any function f in L2 (Td ) can be written as
X
f = < z; f > z ;
z2Zd
where < ; > stands for the inner product in L2 (Td). Since f z ; z 2 Zdg is an
orthonormal basis, for f , g in L2 (Td ),
Z X
d
du f (u)g(u) = < z ; f >< z ; g > : (4:4)
T z2Zd
Moreover, an integration by parts shows that
4. Weak solutions of nonlinear parabolic equations 367
< z ; @uj f > = 2izj < z ; f >

for every function f in C 1 (Td ) and every z 2 Zd, 1 j d. Finally, for f , g in
L2 (Td), denote by f g the convolution of f and g :
Z
(f g)(u) = dv f (v)g(u v) : (4:5)
Td
It is easy to deduce from this definition that
< z ; f g > = < z ; f >< z ; g > : (4:6)
To keep notation simple, assume that is the identity matrix. Fix a positive
integer M , a > d and define the function FM : Td ! C by the series
F M ( u) =
X M (u) :
(1 + j z j2 )(M + jz ja ) z
z2Zd
FM is a well defined twice continuously differentiable real function because a > d.
Consider two weak solutions 1 , 2 of (4.1) such that
Z T Z
dt du jj (t; u)j2 < 1
0 Td
for j = 1, 2. Denote the difference 1 2 (resp. (1 ) (2 )) by ¯ (resp. ¯ )
and let RM : [0; T ] ! R be the function defined by
Z
RM (t) = du ¯t(u)(FM ¯t )(u) :
Td
RM is well defined because jt , j = 1, 2, belong to L2(Td) and FM is bounded.
By properties (4.5) and (4.6),
RM (t)
X M
a ) < z ; ¯t > :
2
=
z2Zd (1 + j z j M + j z j2 )(
Moreover, since 1 , 2 are in L2(Td),

X
R (t)
M !1 M
lim =
(1 +
1
jz j2) < z ; ¯t >2 :
z2Zd
Denote the right hand side by R(t). Since a > d and 1 , 2 are weak solutions,
RM is time differentiable and
d R )(t) 8 2
X M j z j2 < z ; ¯t >< z ; ¯ t > :
(
dt M =
z2Zd (1 + jz j2 )(M + jz ja )
We may rewrite the right hand side as

X
8 2 < z ; ¯t >< z ; ¯ t >
z2Zd
+ 8 2
X jz ja < ; ¯ >< ; ¯ >
z t z t
z2Zd M + jz j
a
+ 8 2
X M
a < z ; ¯t >< z ; t > :
¯
z2Zd (1 + jz j M + jz j )
2 )(
By Schwarz inequality the third expression is less than or equal to
8A 2
X M
a ) < z ; ¯t >
2
z2Zd (1 + j z j M + j z j 2 )(
2 2 X M
A z2Zd (1 + jz j2 )(M + jz ja) < z ; t >
+ ¯ 2
for every A > 0. In virtue of (4.4), the second term of this sum is bounded above
by
Z
2 2 X 2 2
A z2Zd < z ; ¯ t
> = A d du ¯ t (u)2 2
T
2 Z 2 X
2Ag d du ¯ t (u)¯t(u) = 2Ag < z ; ¯ t >< z ; ¯t >
T z2Zd
because ¯ t (u) = (1 (t; u)) (2 (t; u)), () is strictly increasing and 0 is
bounded in absolute value by g . Therefore, setting A = g , integrating in time
and applying identity (4.4), we obtain that RM (t) is bounded above by
Z t Z Z t
RM (0) + BM 6 2
ds du ¯(s; u)¯ (s; u) + 82 g ds RM (s) ;
0 Td 0
where
Z T X jz ja < ; ¯ >< ; ¯ > :
BM = 8 2 dt z t z t
z2Zd M + jz j
0
a
By Gronwall inequality,
Z t Z n o n o
RM (t) + 6 2
ds du ¯(s; u)¯ (s; u) RM (0) + BM exp 8 2 g t
0 T d
for every t T . Since 1 , 2 (and therefore ¯ ) belong to L2 ([0; T ] Td), by (4.4)
limM !1 BM = 0. Therefore, letting M " 1, from the definition of R(t),
Z t Z n o
R(t) + 6 2 ds d
du ¯(s; u)¯ (s; u) R(0) exp 82 g t ;
0 T
5. Entropy solutions of quasi–linear hyperbolic equations 369
what concludes the proof of the theorem.

It follows from the uniqueness of bounded weak solutions and from the first
part of the proof of this theorem that solutions of quasi–linear parabolic equations
depend continuously on the initial data.
Theorem 4.5 Fix a bounded profile 0 and a sequence of bounded profiles 0"
converging weakly to 0 in the sense (4.3). For each " > 0, denote by " (t; u) (resp.
(t; u)) the unique bounded weak solution of equation (4.1) with initial data 0"
(resp. 0 ). The sequence " converges uniformly on each compact set of (0; 1) Td
to .
5. Entropy solutions of quasi–linear hyperbolic equations
We review in this section some properties of weak and entropy solutions of the
conservation law @t + m r() = 0. Proofs, examples and further details can be
found in Lax (1957), Kružkov (1970) and Smoller (1983).
Definition 5.1 Fix a bounded initial profile 0 : Rd ! R and a vector m =

(m1 ; : : : ; md ) in Rd . A bounded function : R+ Rd ! R is a weak solution of
the Cauchy problem 8
> d
X
>
< @t + mj @uj () = 0
j =1 (5:1)
>
>
:
(0; ) = 0 ()
if for every function G: R+ Rd ! R of class CK1;1 (R+ Rd )
Z 1 Z n d
X o
dt d
du (t; u) @tG + ((t; u)) mi @ui G
0 R i=1
Z
+ du G(0; u)0(u) = 0:
Rd
It turns out that weak solutions are not determined uniquely by their initial
value. An additional criterion is therefore needed to select among the weak so-
lutions the physically relevant. The entropy condition presented below is due to
Kružkov (1970). In dimension 1, if is convex, this condition is equivalent to
require the weak solution to have only decreasing (resp. increasing) shocks in the
case m1 > 0 (resp. m1 < 0), what explains the terminology.
Definition 5.2 A bounded function : R+ Rd ! R is an entropy solution of the

1;1
Cauchy problem (5.1) if for every positive function G of class CK ((0; 1) Rd )
and for every a in R+ ,
Z 1 Z n d
X o
dt d
du j(t; u) aj @t G + j((t; u)) (a)j mi @ui G 0 (5:2)
0 R i=1
and Z
lim
t!0 K
du j(t; u) 0 (u)j = 0 (5:3)
for all compact sets K of Rd .
Kružkov (1970) proved the existence and uniqueness of entropy solutions :
Theorem 5.3 For every bounded profile 0 : Rd ! R, there exists a unique entropy
solution of equation (5.1).
P
The first condition imposes @t j(t; u) aj + 1id mi @ui j((t; u)) (a)j
to be negative in the weak sense on (0; 1) Rd for all a in R. It is only the
second condition that connects the solution to the initial data since in the first G
is taken with compact support on the open set (0; 1) Rd .
Kružkov also proved that the entropy solutions are monotone and stable in the
L1 norm :
Theorem 5.4 Consider two bounded profiles 10 , 20 : Rd ! R such that 10 20
and denote by i (t; u) the entropy solutions of (5.1) with initial data 0i , i = 1, 2.
For every t 0, 1 (t; u) 2 (t; u).
Theorem 5.5 Consider two bounded functions 1 , 2 defined on (0; 1) Rd that

satisfy inequality (5.2). Let n = maxi=1;2 ki k1 . For every positive real R and
every 0 < s t < 1,
Z Z
du j (t; u) (t; u)j
1 2
du j1 (s; u) 2 (s; u)j ;
jujR jujR+n (t s)
where, for each positive integer n, n = jmj supjajn j0 (a)j.
Since constants are entropy solutions of (5.1), it follows from Theorem 5.4 that
k(t; )k1 k0k1 for every t 0. On the other hand, Theorem 5.5 concerns
functions defined in the open set (0; 1) Rd and no assumption nor statement is
made on the behavior at time 0. Uniqueness of entropy solutions follows from this
stability result since by assumption (5.3) there is L1 local convergence at time 0 :
Corollary 5.6 Consider two bounded profiles 10 , 20 : Rd ! R. Let n = maxi=1;2
ki0k1 and denote by i (t; u) the entropy solutions of (5.1) with initial data i0,
i = 1, 2. For every positive real R and every 0 s t < 1,
Z Z
du j (t; u) (t; u)j
1 2
du j1 (s; u) 2 (s; u)j :
jujR jujR+n (t s)
5. Entropy solutions of quasi–linear hyperbolic equations 371
Moreover, the entropy solutions are continuous in L1loc :

Corollary 5.7 Consider the entropy solution (t; u) of equation (5.1) with initial
data 0 . For every positive real R and every 0 t < 1,
Z
lim
"!0 jujR
du j(t + "; u) (t; u)j = 0:
Proof. For t = 0 the result follows from assumption (5.3). Fix thus t > 0. For
each " > 0, the bounded function " (t; u) defined as (t + "; u) satisfies inequality
(5.2). Therefore, by Theorem 5.5, for every R > 0,
Z Z
du j(t + "; u) (t; u)j du j(2"; u) ("; u)j
jujR jujR+nt
that vanishes as " # 0 by assumption (5.3).
We conclude this section with a result on the continuous dependence on the
initial data of entropy solutions of one-dimensional equations due to P. Lax
(1957). Unless otherwise stated, up to the end of this section, we consider the
one-dimensional differential equation
(
@t + @u () = 0
(5:4)
(0; ) = 0 ()
and assume to be strictly concave or convex in the range of 0 . In this context
an explicit formula for the entropy solution was derived by E. Hopf (1950) for a
quadratic equation and by P. Lax (1957) for the general case.
To fix ideas denote by [a1 ; a2 ] the range of 0 and assume that is strictly
convex in [a1 ; a2 ]. If necessary we redefine in [a1 ; a2 ]c in order for to be
strictly convex on R. Denote by the Legendre transform of :
( u) = supfuv (v)g
v2R
and define R as the integral of 0 :
Z u
R ( u) = dv 0 (v) :
0
The following three results are taken from Lax (1957).
Lemma 5.8 Consider the expression
t u t v :

B (v ) = R (v ) +
For a fixed t > 0, with the exception of a countable set of values of u, B assumes
its minimum at a single point denoted by v0 (t; u).
We are now in a position to write an explicit formula for the entropy solution
of equation (5.4) :
Theorem 5.9 The function (t; u) defined by
(t; u) = (0 ) 1
u v0 (t; u)
t
is the entropy solution of (5.4).
This explicit formula permits to prove the continuous dependence on the initial
data of entropy solutions of one-dimensional equations :
Theorem 5.10 Let "0 a sequence of bounded functions converging weakly to a

bounded limit 0 as " # 0. Denote by " (resp. (t; u)) the entropy solution of
equation (5.4) with initial data "0 (resp. 0 ). The sequence " (t; u) converges to
(t; u) at all continuity points u of (t; ).
All analysis carried out above can be extended to weak or entropy solutions
defined on the torus Td since each bounded function on Td can be interpreted
as a periodic bounded function on Rd with period Td and equal to on Td . In
the periodic case, Corollaries 5.6 and 5.7 state that the L1 norm of the difference
of two entropy solutions decreases in time and that the entropy solution is L1
continuous in time :
Theorem 5.11 Consider two bounded profiles 10 , 20 : Td ! R and denote by
i (t; u) the entropy solutions of (5.1) with initial data i0 , i = 1, 2. For every
0 s t < 1,
Z Z
d
du j1(t; u) 2 (t; u)j du j1 (s; u) 2 (s; u)j :
T Td
Moreover, for every 0 t < 1,
Z
lim
"!0 Td
du j(t + "; u) (t; u)j = 0:
Appendix 3. Nongradient Tools : Spectral Gaps and
Closed Forms
We present in this chapter the main tools used in the proof of the hydrodynamic
behavior of reversible nongradient systems : estimates on the rate of convergence
to equilibrium of reversible Markov processes and closed and exact forms in the
context of interacting particle systems. The chapter is organized as follows. In
section 1 we prove a second order expansion for the largest eigenvalue of a small
perturbation of a reversible generator. In the hydrodynamic setting this expansion
reduces the proof of a two block estimates to the computation of some central
limit variances (cf. section 7.3). In sections 2 and 3 we prove that the spectral gap
of the generator of a symmetric generalized exclusion process on a cube of length
N shrinks as N 2 in any dimension and uniformly over the density. Finally, in
section 4 we investigate the closed and exact forms in the setting of interacting
particle systems.
In order to motivate the derivation of bounds on the largest negative eigenvalue
of the generator of a reversible process, we conclude this section with a brief
investigation of the rate of convergence to equilibrium in L2 of reversible Markov
processes. Consider an irreducible Markov process Xt on some countable state
space E . Denote by fPt ; t 0g the semigroup of the process and by L its
generator : X
(Lf )(x) = (x)p(x; y)[f (y) f (x)] :
y2E
Assume that the process is reversible with respect to an invariant probability
measure . Since L is a self adjoint operator on L2 ( ) all its eigenvalues are real
and nonpositive. Moreover, 0 is an eigenvalue associated at least to the constant
functions. We claim that 0 has multiplicity 1. To prove this statement, consider an
eigenfunction f in L2 ( ) associated to the eigenvalue 0 : Lf = 0. Multiply both
sides of this equation by f and integrate with respect to to get that the Dirichlet
form of f vanishes :
X
0 = D(f ) = < Lf; f > = (1=2) (x)(x)p(x; y)[f (y) f (x)]2 :
x;y2E
Since we assumed the process to be irreducible, this in turn implies that f is
constant.
Recall that the Dirichlet form is well defined in L2 ( ). Denote by 1 the lower
bound of the strictly positive part of the spectrum of the generator L :
374 Appendix 3. Nongradient Tools : Spectral Gaps and Closed Forms
1 < Lf; f > D(f )

f 2L ( ) < f; f > f 2L ) Var(; f )
:= inf = inf
2 2(
<f;1> =0
In this formula Var(; f ) stands for the variance of f with respect to and the first
infimum is taken over all L2 ( ) functions f which are orthogonal to the constants
(< f > =< f; 1 > = 0). We shall refer to 1 as the spectral gap of the generator
L. Notice that 1 is not necessarily an eigenvalue of L and that 1 may vanish
because L2 ( ) is an infinite dimensional space. The following result establishes
that 1 is closely related to the exponential rate of convergence to equilibrium in
L2 ( ).
Theorem 0.1 Denote by 0 the largest real such that for every function f in
L2 ( )
Pt f < f > 2 C (f )e t

for all t > 0 and some finite constant C (f ) depending only on f . 0 coincides with
the spectral gap : 0 = 1 .
R0
Proof. Denote by 1 dE the spectral decomposition of the self adjoint non-
positive operator L. For each function f in L2 ( ), denote by f the spectral
measure on R associated to f : f (d) = d < E f; f >. It follows from the def-
inition of 1 that f (( 1 ; 0]) = 0 for each function f orthogonal to the constants.
In particular,
Z 0

Pt f < f >
= et f <f> (d) e t kf < f > k2
1
2 1
for every f in L2 ( ). This shows that 0 1 .
We turn to the inverse inequality. Since for every L2( ) function f and for
every t 0,
Z 0
et f <f> (d)
= Pt f < f > 2 C (f )e t 0
1
for some finite constant C (f ), f <f> (( 0 ; 0]) = 0 for every f in L2 ( ).
Therefore, for each function f orthogonal to the constants, f (( 0 ; 0]) = 0. In
particular, 1 0 .
We have just shown that the spectral gap 1 is intimately connected to the rate
of convergence to equilibrium in L2 . It is therefore natural to try to prove lower
bounds for 1 . The next result provides such an estimate in a general context.
Proposition 0.2 Suppose f; g : E ! R+ are functions satisfying Lg + fg = 0. If

f (x) > 0 and g(x) is bounded below by some strictly positive constant, the
spectral gap is bounded below by .
1. On the spectrum of reversible Markov processes 375
Proof. By assumption, the function f (x) = (Lg )(x)=g (x) is bounded below by
. In particular, for every L2( ) function h with norm khk2 = 1,
Z
(Lg )(x) 2
g(x) h (x) (dx) :
If g was bounded above, by Theorem A1.10.2, the Dirichlet form of h would be
bounded below by for every h with L2 norm equal to 1. The proposition would
therefore follow from the definition of 1 . Hence, to conclude the proof it remains
to approximate g by bounded functions.
For each positive integer M , denote by gM the function defined by gM (x) =
g(x) ^ M . By assumption gM is bounded below by a strictly positive constant and
bounded above by M . In particular, by Theorem A1.10.2,
Z
(LgM )(x) 2
D(h) gM (x) h (x) (dx)
for every M . A straightforward analysis shows that (LgM )(x)=gM (x) is bounded
below by (Lg )(x)=g (x) if g (x) M and that (LgM )(x)=gM (x) is positive if
g(x) M . In particular, the right hand side of the last inequality is bounded
below by Z
(Lg )(x)
g(x) 1fg(x) M gh (x) (dx) :
2
Since by assumption (Lg )(x)=g (x) is a positive function bounded below by ,

by the monotone convergence theorem, as M " 1, this expression converges to
Z
(Lg )(x) 2
g(x) h (x) (dx) :
This concludes the proof of the proposition.
1. On the spectrum of reversible Markov processes
We consider in this section a reversible Markov process on a countable state space

E and keep the notation of the beginning of the chapter.
Theorem 1.1 Assume that the generator L has a spectral gap of magnitude 1
:
Var(; f ) D( f )
for every f in L2 ( ). Let V be a mean-zero bounded function such that <
( L) 1V; V > < 1. Denote by " the upper bound of the spectrum of L + "V :
n o n o
" = sup < f; (L + "V )f > = sup " < f; V f > D(f ) :
f ; kf k2 =1 f ; kf k2 =1
Then,
" "2 < ( L) 1V; V >

2kV k1 "
0
1
Proof. Fix " > 0. To keep notation simple, denote by L" the operator L + "V . Let
fG";n ; n 1g be a sequence of L2( ) functions that approaches the supremum :
kG";n k2 = 1 and
!1 < G";n ; L" G";n > = " :
nlim
Notice that < G";n ; (" L" )G";n > vanishes as n " 1 by definition of G";n .
We may assume without loss of generality that G";n has positive expectation :
< G";n > 0 :

The eigenfunctions associated to the largest eigenvalue of the generator L are
the constants. Since we are considering a small perturbation of the generator and
G";n is normalized to have positive mean, we expect G";n to be close to 1. The
idea of the proof is therefore to expand G";n around 1. By definition of G";n , we
have
" = < G";n ; (" L" )G";n > + < G";n ; L" G";n > :
Since L" = L + "V , we may rewrite the second term on the right hand side as
n o
" < V > + 2 < V [G";n 1] > + < V [G";n 1]2 > D(G";n ) :
Recall that V has mean 0. By Schwarz inequality the expression inside braces is
bounded above by
" A" < ( L) 1V; V > + A" < ( L)[G";n 1]; [G";n 1] >
n
o
+ < V [G";n 1]2 >
"A < ( L) 1V; V > + 2"kV k1 1 < G";n > + A D(G";n )

2
for every positive A.

To bound the second term of the last expression, notice that by Schwarz in-
equality and our choice of G";n ,
0 < G";n > < G2";n >1=2 = 1
so that
0 < G";n > 1 < G";n >2
1
= < G2";n > < G";n >2 D(G";n )
provided we have a spectral gap of magnitude 1 . Recollecting all previous
estimates we obtain that " is bounded above by
1. On the spectrum of reversible Markov processes 377
< G";n ; (" L" )G";n > + "A < ( L) 1V; V >
2
(1 A 2kV k1 " ) D(G";n ) :

We conclude the proof of the upper bound by choosing A to be equal to 1
2kV k1 " and letting n " 1 because < G";n ; (" L" )G";n > vanishes as
n " 1.
The lower bound follows from the fact that V has mean 0. We have just to
set f = 1 in the variational formula for " .
The following result is a simple consequence of the previous Theorem.
Corollary 1.2 Assume that the generator L has a strictly positive spectral gap of
magnitude 1 . Let V be a bounded function. For every sufficiently small ",
n o
sup " < V f 2 > < ( L)f; f >
f
" < V > "2 < ( L) 1V; V > ;
1 2kV k1 "
+
where the supremum is taken over all functions f in L2 ( ) such that kf k2 = 1.
We conclude this section presenting an alternative variational formula for the

largest eigenvalue of a perturbation of a generator which is reversible with respect
to some measure . This alternative version is constantly used in the proof of large
deviations principles for Markov processes.
Consider a continuous time Markov process Xt on a countable space E with
generator L reversible with respect to an invariant state . Let V : E ! R be
a bounded function and denote by V the largest eigenvalue of the symmetric
operator L + V in L2 ( ). We claim that
n p o
V = sup < V; f > D( f ) ; (1:1)
f
where
R the supremum is taken over all densities with respect to : f 0 and
fd = 1.
To prove this statement recall the variational formula for the largest eigenvalue
of a symmetric operator :
n o
V = sup < V; f 2 > D(f ) :
f
In this formula the supremum is carried over all functions f in L2 ( ) such that
kf k2 = 1. We have proved in Appendix 1 that D(V (f )) D(f ) for every function
V : R ! R such that jV (b) V (a)j jb aj. In particular, D(jf j) D(f ) and
the previous supremum is equal to
n o
sup < V; f 2 > D(f ) ;
f 0
p
where the supremum now is carried over all positive functions with L2 norm equal
to 1. To obtain (1.1) it remains to replace f by f .
2. Spectral gap for generalized exclusion processes
We investigate in this section the spectral gap of symmetric generalized exclusion

dynamics on finite d-dimensional cubes. For each positive integer N , denote by
N a cube of linear size N :

N = f1; : : : ; N gd
and by L
N the generator of the symmetric generalized exclusion process on
N :
X
(L
N f )( ) = (1=2) rx;y ()[f (x;y) f ()] :
x;y2
N
jx yj=1
In this formula rx;y ( ) is equal to 1 whenever a jump is possible from x to y:
rx;y () = 1f (x) > 0; (y ) < g : (2:1)
For each fixed total number of particles 0 K N d , the Markov process

with generator L
N and state space

N ;K is a finite state irreducible Markov
process. In particular, it has a unique ergodic invariant probability measure that we
denote by
N ;K or N;K . Since the transition probability is symmetric (p(y ) =
(1=2) if jy j = 1 and 0 otherwise), this measure is in fact reversible. Moreover,
a simple computation shows that N;K is the uniform probability measure on

N ;K , the space of all configurations of f0; : : :; g
N with K particles :
X

N ;K = f 2 f0; : : : ; g
N ; (x) = K g :
x2
N
Keep in mind that L2 (N;K ) is a finite dimensional space. Since L
N is a
self adjoint generator all its eigenvalues are real and nonpositive. Denote them
by 0 0 > 1 > > R . To keep notation simple, we omitted the
dependence of i on N and K . Since L
N is a generator, 0 is an eigenvalue
associated at least to the constant functions. We claim that 0 has multiplicity 1 :
let f be a eigenfunction in L2 (N;K ) associated to the eigenvalue 0 : L
N f = 0.
Multiply both sides of this equation by f and integrate with respect to N;K to
show that f is constant. In particular, for each N 1 and 0 K N d , L
N
has a positive spectral gap denoted by 1 = 1 (K; N ). We may express 1 by a
variational formula :
2. Spectral gap for generalized exclusion processes 379
< L
N f; f >N;K
1 (N; K ) = inf
< f; f >N;K ;
where the infimum is carried over all functions in L2 (N;K ) that are orthogonal
to the eigenspace associated to 0 , i.e., that are orthogonal to the constants :
< f; 1 >N;K = 0. Thus, if we denote by W (N; K ) the inverse of the spectral gap,
W (N; K ) = 1 (N; K ) 1,

Var (N;K ; f )
W (N; K ) = sup :
f 2L (N;K ) D(N;K ; f )
2
In this formula, for a finite subset of Zd and function f in L2 (;K ) Var (;K ; f )
and D(;K ; f ) denote respectively the variance and the Dirichlet form of f with
respect to ;K :
h 2 i
Var (;K ; f ) = E;K f E;K [f ]
X X Z n o2
D(;K ; f ) = (1=4) rx;y () f (x;y ) f () ;K (d) :
x2 y2
jy xj=1
We investigate in this section the asymptotic behavior, as N " 1, of the spec-
tral gap of the generator of a generalized symmetric exclusion process restricted
to a d-dimensional cube of linear size N . We prove that the spectral gap shrinks
as N 2 in all dimensions :
Theorem 2.1 There exists a universal constant C0 such that
Var (N;K ; f ) C0 N 2 D(N;K ; f )

for all N 1, 0 K N d and all functions f in L2(N;K ).
Fix a finite subset of Zd, a subset 1 of and a configuration of
f0; : : : ; g. Recall that M () = M () stands for the total number of particles
for the configuration : X
M ( ) = (x) :
x2
For a function f : ;K ! R and a configuration of f0; : : :; g1 with at most
K particles, denote by f the function on 1 ;K M ( ) whose value at a con-
figuration is equal to f (; ). Here (; ) is the configuration of ;K defined
by
(x) if x 2 1
(; )(x) =
(x) if x 62 1 :
This notation permits to express in a simple form the conditional expectation of
f : ;K ! R given a configuration of 1 . An elementary computation
shows that it is equal to the expectation of f with respect to the canonical measure
1 ;K M ( ) :
h i
E;K f () (x) = (x) for x 2 1 = E 1 ;K M () [f ] : (2:2)
Denote by W (N ) the maximum over all densities of the inverse of the spectral
gaps W (N; K ) :
W (N ) = 0Kmax
N d
W (N; K ) :
With this notation the statement of the theorem reduces to the existence of a
universal constant C0 such that
W ( N ) C0 N 2 :
for all positive integers N . This statement is proved by induction on N .
It follows from the discussion preceding Theorem 2.1 that for each fixed N
and 0 K N , L
N has a strictly positive spectral gap : W (N; K ) < 1. In
particular, since there are only a finite number of cases, for any N 0 , there exists a
finite constant d0 so that
Var (N;K ; f ) d0 N 2 D(N;K ; f ) (2:3)
for all N N 0 , all 0 K (N 0 )d and all f in L2 (N;K ). We may restate this
inequality saying that for any N0 there exists a finite constant d0 so that
W ( N ) d0 N 2 for all N N 0.
Proof of Theorem 2.1. To avoid unnecessary heavy notation and to detach the
main ideas of the proof, we first consider the problem in one dimension. We
indicate in the next section the ingredients needed to extend this proof to higher
dimensions.
For generalized exclusion processes with jump rate given by (2.1), there is a
duality between particles and holes, i.e., the holes evolve with the same dynamics
as particles do. We may therefore assume that the density of particles K=j
N j is
bounded above by =2. We shall do so without further comment.
The idea of the proof consists in using the induction hypothesis to set up a
recursive equation for W (N ). With this purpose in mind we write the identity
n h io
f ( ) EN;K [f ] = f ( ) EN;K f (N )
n h i o
+ EN;K f (N ) EN;K [f ] :
Hereafter EN;K stands for the expectation with respect to the measure N;K .
Through this decomposition and since the two terms on the right hand side of the
last identity are orthogonal in L2 (N;K ), we may express the variance of f as
h 2 i
Var (N;K ; f ) = EN;K f EN;K f (N )
h h i 2 i (2:4)
+ EN;K EN;K f (N ) EN;K [f ] :
The first term on the right hand side is easily analyzed through the induc-
tion assumption. Indeed, from identity (2.2) we have that EN;K [f (N )] =
EN 1;K (N )[f(N ) ]. To keep notation simple, we abbreviate N 1 by N1 and
K (N ) by K(N ) and denote configurations of N1;K(N ) by the Greek letter
. Taking conditional expectation with respect to (N ) and applying once more
identity (2.2), the first term on the right hand side of (2.4) becomes
h 2 i
EN;K f EN ;K N [f(N )] 1 ( )
h 2 i
= EN;K EN ;K N f(N ) EN ;K N [f(N ) ]
1 ( ) 1 ( ) :
h 2 i
Notice that EN1 ;K(N ) f(N ) EN1 ;K(N ) [f(N ) ] is the variance, with respect
to N1 ;K(N ) , of f(N ) , a function of N 1 variables. By the induction hypothesis,
this expression is bounded by W (N 1)D(N1;K(N ) ; f(N ) ). In particular, the first
term on the right hand side of (2.4) is bounded by
h i
W (N 1)EN;K D(N ;K N ; f(N ) )
1 ( ) =
W (N 1) X h
EN;K EN ;K N rx;y ( ) f(N )( x;y ) f(N ) ( )
h n o2 ii
:
1 ( )
4
x;y2
N1
jx yj=1
Since expectation with respect to N1 ;K(N ) corresponds to conditional expectation
with respect to (N ), the last expression is equal to
h n o2 i
X
(1=4)W (N 1) EN;K EN;K rx;y () f (x;y ) f () ( N)
x;y2
N1
jx yj=1
W (N 1)D(N;K ; f ) :
This last relation misses to be an equality only because in the Dirichlet form
D(N;K ; f ) there is a piece that measures the dependence of f on jumps over the
bond fN 1; N g that does not appear at the left hand side of the inequality.
Up to his point we showed by means of the induction hypothesis that the first
term on the right hand side of (2.4) is bounded by W (N 1)D(N;K ; f ) :
h 2 i
EN;K f EN;K f (N ) W (N 1)D(N;K ; f ) : (2:5)
We turn now to the second term on the right hand side of equation (2.4). Since

the expected value, with respect to N;K , of EN;K [f j (N )] is equal to EN;K f ,

this expression is the variance of EN;K [f (N )], a function of one variable. The
estimation of this variance constitutes the second step in our route to build a
recurrent formula for W (N ). We summarize it in the following proposition.
Proposition 2.2 There exists a finite constant C = C () depending only on such
that for each " > 0, there exists ` = `(") and N 0 = N 0 (") for which
h i
Var N;K ; EN;K f (N )
(2:6)
C () N + `W (N` + 1) D(N;K ; f ) "
n o
+
N Var (N;K ; f )
for all N N 0 (") and all 0 K (=2)N .
We conclude now the proof of Theorem 2.1 assuming Proposition 2.2. Take
" < 2 in Proposition 2.2. From identity (2.4), estimate (2.5) and Proposition 2.2,
for N large enough,
" Var ; f nW (N 1) + C () N oD( ; f )
1
N N;K N;K
for some constant C () that depends only on . There exists, therefore, a constant
C () such that
W (N ) 1 N"
1n o
W (N 1) + C ()N
for all N large enough. It is elementary to prove from this recursive inequality
and estimate (2.3) the existence of a constant C0 for which W (N ) C0 N 2 . for
all N 1.
We turn now to the proof of Proposition 2.2. The strategy may roughly be
described as follows. We shall first prove a spectral gap, uniform over the param-
eters K and N , for functions depending only on one site (hereafter called one site
functions). More precisely, we shall prove that there exists a universal constant
B1 = B1 () such that for every H : f0; : : : ; g ! R,

Var N;K ; H ((N )) B1 ()D N;K ; H ((N )) ;
where D is a slight modification of the Dirichlet form D. Applying this result to
the one site function EN;K [f j (N )], we shall reduce the proof of Proposition 2.2
to the proof that the Dirichlet form of EN;K [f j (N )] is bounded by the right hand
side of (2.6).
The statement of the uniform spectral gap for one site functions requires some
notation. Denote by N;K
1
the one site marginal of N;K :
N;K
1
(a) = N;K f; (N ) = ag for 0 a :
Consider the one site Dirichlet form D(N;K

1
; ) defined on L2(N;K
1
) by

X
D(N;K
1
; H) = [H (a 1) H (a)]2N;K
1
(a) :
a=1
We have the following one–coordinate Poincaré inequality :
Lemma 2.3 There exists a constant B1 = B1 (), so that for all functions
H : f0; : : :; g ! R,
h 2 i
EN;K H ((N )) EN;K [H ((N ))] B1 D(N;K
1
; H) : (2:7)
Proof. Since H ((N )) depends only on (N ), we may write the variance of H
as

X n o2
X n o2
N;K
1
(a) H (a) EN;K [H ] 1 N;K
1
(a) H (a) H (0)
a=0 a=0
By Schwarz inequality and a summation by parts, the right hand side of this last
formula is bounded above by

X n
o2 X N;K
1
( a)
N;K
1
(b) H (b 1) H (b)
a=b N;K (b)
1
b=1
Recall that we assumed the density K=N to be bounded above by =2. By Lemma
A2.2.4, uniformly over densities bounded by =2, the ratio N;K
1
(a)=N;K
1
(b) for
a b is bounded by some constant depending only on . The last expression is
therefore dominated by

X n o2
B1 () N;K
1
(b) H (b 1) H (b) :
b=1
This concludes the proof of the lemma.

Applying this lemma to the one site function E f N )] we obtain an
N;K [ (
estimate for the variance of EN;K [f (N )] :
h i
Var N;K ; E N;K ( f N)

X n o2 (2:8)
B1() N;K
1
(b) EN ;K b+1 [fb
1 1] EN ;K b [fb ] :
1
b=1
In this formula, for each 0 b , fb : N ;K b ! R stands for the function
defined by fb ( ) = f (; b).
1
We have to estimate the difference EN1 ;Kb +1 [fb 1 ] EN1 ;Kb [fb ]. Since the
measure N;Kb +1 is concentrated on configurations with Kb + 1 particles, we have
that X h i
EN1 ;Kb+1 [fb 1 ] = K 1+ 1 EN1;Kb +1 fb 1 ( ) (x) : (2:9)
b x2
N1
Performing a change of variables 0 = dx , where dx denotes the configuration
with no particles but one at site x, we obtain that the last expression is equal to
RN ;Kb X E h i
N ;K f b 1 ( + dx ) [1 + (x)] 1f (x) < g ; (2:10)
1
Kb + 1 x2
N b 1
where RN1 ;Kb is the factor N1 ;Kb +1 ( )=N1 ;Kb ( ) coming from the change of
measures. It is given by
RN;K Z (
N ; K ) ;
=
Z (
N ; K + 1)
provided Z (N; K ) stands for the total number of configurations of
N ;K :
X N!
Z (
N ; K ) =
Ji 0;0i J0 ! J !
J1 ++J =K
J0 ++J =N
Denote the cylinder function [1 + (x)] 1f (x) < g by h( (x)). Replacing fb 1
by 1 in formulas (2.9) and (2.10), we get that the expected value, with respect to
N1 ;Kb , of fN1 RN1;Kb =(Kb + 1)gh( (1)) is equal to 1. To keep notation simple
we shall abbreviate (NRN;K =K + 1) by SN;K and denote by gN1 ;Kb ( (x)) the
cylinder function SN1 ;Kb h( (x)). We may now write
EN ;Kb +1 [fb
1 1] EN ;Kb [fb]
1
n o
X
=
1
N1 x2
N EN ;Kb fb 1 ( + dx ) fb ( ) gN ;Kb ( (x))
(2:11)
1 1
1
n o
X
N1 x2
N EN ;Kb fb ( ) gN ;Kb ( (x)) :
1
+ 1 1 1
1
Notice that the second term is the covariance of fb and gN1 ;Kb ( (x)) because
gN1;Kb ( (x)) has mean one. We shall estimate each term of (2.11) separately. We
start with the first one which is simpler.
Lemma 2.4 There exists a constant C () depending only on such that

X

X
n o 2
N;K (b) N
1
EN ;Kb fb 1 ( + dx ) fb ( ) gN ;Kb ( (x))
1
1 1
b=1 1
x2
N 1
C () N D(N;K ; f ) :
(2:12)
Proof. By Schwarz inequality and since gN1 ;Kb has mean 1 with respect to the
measure N1 ;Kb , the left hand side of (2.12) is bounded above by

X X
n o2
N;K
1
(b)
1
N EN ;Kb fb 1 ( + dx ) fb ( ) gN ;Kb ( (x)) :
1 1
b=1 1
x2
N1
We shall prove at the end of this lemma that SN;K is uniformly bounded over all
densities less than (2=3) :
sup SN;K C () (2:13)

N 1
0K (2=3)N
for some constant C depending only on . From this estimate it follows that the
function gN1 ;Kb ( (x)) is bounded above by C ()1f (x) < g because there are
at most particles per site. In particular, the left hand side of (2.12) is bounded
above by
C () X X n o2
1
b EN ;K f b dx fb 1f (x) < g
N1 x2
N b=1 N;K ( ) b 11( + ) ( )
1

C () X E E
hn
f
o2

N;K N ;K N (N ) 1 ( + dx ) f (N ) ( )
=
N1 x2
N 1 ( )
1
i
1f(N ) > 0; (x) < g :
In these formulas C () is a constant depending only on that may change from
line to line. Recall that f(N ) ( ) = f (; (N )) and that N;K –conditional expecta-
tion with respect to (N ) corresponds to expectation with respect to N1 ;K (N ) .
Therefore, the last sum is equal to
C () X E hr ()nf (N;x) f ()o2 i : (2:14)
N1 x2
N N;K N;x
1
It remains to estimate this expression by the Dirichlet form. The difficulty here
is to evaluate the effect of a long range jump from N to x with a Dirichlet form
that measures only modifications due to nearest neighbor jumps. We already faced
this problem when proving the two block estimates for zero range processes in
section 5.5. The indicator function rN;x ( ) adds here a minor difficulty. In sake
of completeness we shall prove at the end of this section that for each fixed sites
x and y,
h n o2 i
EN;K rx;y () f (x;y ) f ()
x_X
y 1 h n o2 i
C ()jx yj EN;K rz;z+1 () f (z;z+1) f ()
z=x^y
for some finite constant C () depending only on . It follows from this estimate
that (2.14) is bounded above by
C () X (N x) NX1 E hr ()nf (y;y+1) f ()o2 i

N1 x2
N N;K y;y+1
1
y =x
X h n o2 i
C ()N EN;K ry;y+1 () f (y;y+1) f ()
y2
N 1
= C () N D(N;K ; f ) :
To conclude the proof of the lemma it remains to show that SN;K is bounded.
Recall that this expression is equal to (EN;K [h( (1))]) 1, because the expectation
of gN;K ( (1)) with respect to N;K is equal to 1. For each fixed N and K this
expression is of course bounded because h(0) = 1. To prove the statement we have
therefore to investigate the asymptotic behavior as N " 1. By the equivalence of
ensembles (Lemma A2.2.2), there exists a finite constant B () depending only on
such that

EN;K [h( (1))] EK=N [h( (1)] B ()=N
for all 0 K=N 2=3. In this formula and below E indicates expectation with
respect to the grand canonical invariant measure with density , that we denoted
by . Changing variables we get that E [h( (0))] = =(). A straightforward
expansion around the origin shows that this function is bounded below by a strictly
positive constant on any compact subset of [0; 2). In particular, EN;K [h( (0))] is
bounded below by a positive constant for N large enough and 0 K < (2=3)N .

We turn now to the second line in decomposition (2.11). Since gN1 ;Kb ( (x))
has mean 1 with respect to N1 ;Kb the second line reduces to the covariance

X
EN ;Kb fb ( ) ; N
1
1
gN ;Kb ( (x)) :
1
1
x2
N 1
Proposition 2.5 For each fixed " > 0, there exists ` = `(") and N 0 = N 0 (") such
that,

X

X
2
N;K (b) EN ;Kb fb ( ) ; N
1
1
1
gN ;Kb ( (x))
1
b=1 1
x2
N 1 (2:15)
C ()`W N
(` + 1)
D (N;K ; f ) +
" Var( ; f )
N N;K
for all N N 0 (") and all 0 K (=2)N .
Before proving this statement, notice that Proposition 2.2 follows from in-
equality (2.8), formula (2.11), Lemma 2.4 and Proposition 2.5. We turn now to
the proof of Proposition 2.5. We first single out the case of small densities.
Lemma 2.6 For every " > 0, there exists a positive integer N 0 and a density 0
such that
N"
1 X
Var N;K ; g N1 ;Kb ( (x))
N1 x2
N
1
for all N N 0 and all K=N 0 .

Proof. The variance may be written as
h i
N1 EN ;Kb gN ;Kb ( (1)) ; gN ;Kb ( (1))
1
1 1 1
1 h i
+ 1
N E N ;K b g N ;Kb ( (1)) ; gN ;Kb ( (2)) :
1 1 1
1
Since gN1 ;Kb ( (1)) has mean 1, a change of variables 0 = + d1 permits to rewrite
the variance of gN1 ;Kb ( (1)) as
N1 E h i h i
N ;Kb +1 gN ;Kb ( (1) 1) (1) EN ;Kb gN ;Kb ( (1)) :
Kb + 1 1 1 1 1
By similar reasons the covariance of gN1 ;Kb ( (1)) and gN1 ;Kb ( (2)) is equal to
N1 E h
g
i h
EN ;Kb gN ;Kb ( (1)) :
i
Kb + 1 N ;Kb +1 N ;Kb
1 ( (1)) (2) 1 1 1
In order to obtain a simpler expression for this covariance observe that the first
expectation may be rewritten as
N NX1 h i
(Kb + 1)(N
1
2)
EN ;Kb +1 gN ;Kb ( (1)) (x)
1 1
x=2
. Since N;K
because N;K is the uniform measure over all configurations of N;K
PN 1
is concentrated on configurations with K particles, x=2 (x) = Kb + 1 (1).
Therefore, last expression is equal to
N 1
E
h
g
i
(N
N ;K +1 N1 ;Kb
2) 1 b
( (1))
N 1 E h i
N 1 ;Kb +1 gN1 ;Kb ( (1)) (1) :
(Kb + 1)(N 2)
P
Up to this point we showed that the variance of x2
N1 gN1 ;Kb ( (x)) with
respect to N1 ;Kb is equal to

SN ;Kb EN ;Kb +1 [h( (1))]
1 1 EN ;Kb [h( (1))]
1
h n oi
1
E
Kb + 1 N1 ;Kb +1 (1) h( (1)) h( (1) 1) ;
where h( (1)) is the cylinder function [1 + (1)]1f (1) < g. To conclude the
proof of the lemma it remains to recall that SN;K is bounded and apply Lemma
A2.2.3.
Lemma 2.7 For each fixed "> 0 and 0 > 0, there exists ` = `(0 ; ") and
N = N (0; ") such that

X
X 2
N;K
1
(b) EN ;Kb fb ( ) ;
1
1
N gN ;Kb ( (x)) 1
b=1 1
x2
N 1 (2:16)
C ()`W
N
(` + 1)
D(N;K ; f ) + N" Var(N;K ; f ) :
for all N N (0; ") and all 0 N K (=2)N .
p
Proof. Fix a positive integer ` N and divide the set f1; : : : ; N1 g in non
overlapping intervals fBa ; 1 a pg of length ` or ` + 1 :
p
[

N 1 Ba and Ba1 \ Ba2 = for a1 == a2 :
=
a=1
Denote by Ma = Ma ( ) P the total number of particles in the interval Ba for the
configuration : Ma = x2Ba (x), by jBa j the cardinality of Ba , by à the
density of particles in Ba : à = jBa j 1 Ma and by g~a (à ) the expected value of
the cylinder function gN1 ;Kb with respect to the measure Ba ;Ma :
h i
g~a (a` ) = EBa ;Ma gN1;Kb ( (x)) for any x 2 Ba :
Since the covariance is bilinear, the left hand side of (2.16) is bounded by

X
p 2
1 X X
2 N;K (b) N
1
jBa jEN ;Kb fb ( ); jB j gN ;Kb ( (x)) 1
g~a (a` )
b=1 1
a=1 a x2Ba
1 1

X
p 2
1 X
+ 2 N;K (b) EN ;Kb fb ( ) ; N
1
1 jBa j g~a (a` ) :
b=1 1
a=1
(2:17)
We consider these two expressions separately.
To estimate the first line, apply Schwarz inequality and take conditional ex-
pectation with respect to Ma . Since to take conditional expectation with respect to
the total number of particles Ma corresponds to integrate with respect to Ba ;Ma ,
the first expression in (2.17) is equal to

X p
1 X
2 N;K (b) N
1
jBa j
b=1 1
a=1
h i 2
X
EN ;Kb EBa ;Ma fb ( ) ; jB j gN ;Kb ( (x)) 1
1
a x2Ba 1
because, by definition, g~a (à ) is the expected value of gN1 ;Kb () with respect to
Ba ;Ma . In this formula fb should be understood as a function of f (x); x 2 Ba g
with all remaining coordinates frozen.
By (2.13), the function gN1 ;Kb () is bounded by a constant depending only on
. Therefore, by Schwarz inequality and by the induction hypothesis, for each a
we have that
h i2
X
EBa ;Ma fb ( ) ; jB j gN ;Kb ( (x))
1
C ()Var(Ba ;Ma ; fb )
a x2Ba 1
C ()W (jBaj)D(Ba ;Ma ; fb ) :

Moreover,
p
X h i
EN ;Kb D(Ba ;Ma ; fb ) D(N ;Kb ; fb ) :
1 1
a=1
This inequality misses to be an equality because the Dirichlet form on the left hand
side does not take into account modifications due to jumps from one cube Ba to
another cube Bb for a = = b. Therefore, the first expression in (2.17) is bounded
above by

X
C ()N 1(` + 1)W (` + 1) N;K
1
(b)D(N ;Kb ; fb )
1
b=1
C ()`W
N
(` + 1)
D(N;K ; f ) :
because all cubes Ba have length at most (` + 1).
We turn now to the second expression of formula (2.17). Denote by h~() the
expectation, with respect to , of the cylinder function h( ) = [1 + (0)]1f (0) <
g, by g~() the expectation of gN1;Kb with respect to and by g~0 () the derivative
of the smooth function g~ calculated at :
h i
h~() = E [1 + (0)]1f (0) < g ; g~() = SN ;Kb h~()
1
g~0(0 ) d
d g =

and = ~( )
0
P
Consider now the second term in formula (2.17). Since a jBa jà is equal to
the total number of particles on
N1 , we may add g~(Kb =N1 ) g~0 (Kb =N1 )[à
Kb =N1 ] to the second term of the covariance without modifying the value of the
covariance. The second expression in (2.17) is therefore equal to

X p
X
2
2 N;K
1
(b) EN ;Kb fb ( ) ;
1
N
1 jBa jFa (a` )
b=1 1
a=1

X p 2
1 X
2 N;K (b)Var N ;Kb ; fb EN ;Kb N
1
1 1 jBa jFa (a` ) ;
b=1 1
a=1
(2:18)
where
Fa (a` ) = g~a (a` ) g~(Kb=N1 ) g~0 (Kb =N1)[a` Kb =N1] : (2:19)
To conclude the proof of the lemma it remains to show that the second variance
is bounded by "N 1 for some ` = `(0 ; "), all N N (0 ; ") and all 0 N K
(=2)N . This variance is equal to
X jBa j2 E h
Fa (a` )2
i X jBa jjBb j E h i
Fa (a` )Fb (b` ) :
N12 N1 ;Kb +
N12 N1 ;Kb
a a==b
Consider a total number of particles K so that the density Kb =N1 is bounded
below by 0 and above by =2. By Remark A2.1.2, Corollary A2.1.4 and since
the grand canonical measures are product, for N large enough,
h i
E ( 0 ) ` E h i1=2
EN1 ;Kb [Fa (a` )2 ] EKb =N1 Fa (a` )2 + N
`
Kb =N Fa (a )
1
4
and, for a == b,
h i h i
EN ;Kb [Fa (a` )Fb (b`)] EKb =N Fa (a` ) EKb =N Fb (b` )
1 1 1
E (0 )`
h i h i1=2
` `
N EKb =N Fa (a ) EKb =N Fb (b ) :
2 2
+ 1 1
By definition of g~a () and of g~(), the expectation of Fa (à ) with respect to the
grand canonical measure Kb =N is equal to 0. In particular, applying the elemen-
tary inequality 2ab a2 + b2 , we obtain that the second variance in (2.18) is
1
bounded by
E (0 )` X jBa j E h
F `
i
N a N Kb =N a a
2
( ) 1
i1=2 (2:20)
E (0 )`2 X jBa j h
`
N 2 a N EKb =N Fa (a ) :
4
+ 1
We shall estimate these two terms separately. Notice, however, that we reduced the
original problem involving canonical measures to a simpler estimate concerning
only product measures.
Recall that g~a (a ) (resp. g~()) denotes the expectation of SN1 ;Kb h with respect
to Ba ;Ma (resp. ). In particular, by (2.13), g~a and g~ are bounded by a con-
stant depending only on . Moreover, we get an explicit formula for h~ changing
variables : h~() = =(). This formula permits to compute the derivative of
g~() :
g~0 () = SN1;Kb h~0 () and h~0 () = (()(())())2
2
Hence, h~0 () is a smooth function on (0; ). A straightforward expansion permits
to describe its behavior at the boundary :
lim h~0 () = 1 ; ~0
!0 ! h () =
lim :
In particular, h~0 () may be extended as a continuous function on [0; ]. This es-
timate together with the previous bounds on g~a (a ) and g~() show that Fa is
bounded by a constant depending only on . In particular, the second expression
in (2.20) is bounded above by E (0 ; )`2 =N 2 .
On the other hand, by Lemma A2.2.2,
g~(a` ) SN ;Kb C (`)

g a` )
~a ( 1
for some constant C () depending only on . In particular, by (2.13) and Schwarz
inequality, the first expression in (2.20) is bounded above by
E (0 ; ) E (0 )` X jBa j E h
G
i
(a` )2 ;
`N +
N a N Kb =N 1 a
where Ga (à ) = g~(à ) g~(Kb =N1 ) g~0 (Kb =N1 )(à Kb =N1 ). By Taylor formula,
the absolute value of Ga is bounded above by SN1 ;Kb kh~00 k1 (à Kb =N1 )2 . The
first expression of (2.20) is thus bounded above by
0 ; )` km k
SN ;Kb kh~00 k21 E (N`
n o
1 2 4 1 + k2k21 :
In this formula k k1 stands for the sup norm of functions defined on (0; ).
To conclude the proof it remains to show that kh~00 k1 , km4 k1 and k 2 k1 are
bounded. As approaches 0 (resp. ), the one site marginal of converges to
the Dirac measure concentrated on the configuration with 0 (resp. ) particles.
Therefore, m4 (()) and (())2 must vanish as converges to 0 or . This
behavior can also be checked expanding m4 and 2 around the origin and around
. This proves that m4 and 2 are bounded functions because they are smooth on
(0; ).
On the other hand, straightforward computations show that
h~00 () (())2m3 (()) 2(())4

+
=
()(())6
From this formula, one can show that lim!0 h~00 () = 0 and lim!0 h~00 () =
2( + 1). Therefore h~00 can be extend as a continuous function on [0; ] and the
lemma is proved
We conclude this section with an estimate on the Dirichlet form of generalized
symmetric simple exclusion processes with long jumps that was used in the proof
of Lemma 2.4. Since this estimate reappears in the sequel for processes evolving
in higher dimensions, we prove it in this more general setup.
Denote by
N;K the cube f1; : : : ; N gd and by N;K the uniform measure on

N;K . For each pair of sites x = (x1 ; : : : ; xd ) and yP= (y1; : : : yd), let n = n(x; y)
stand for the distance from x to y : n = jjx y jj = 1id jxi yi j and denote
by (x; y ) a path from x to y , i.e., a sequence x = x0 ; : : : ; xn = y of sites such
that jjxi+1 xi jj = 1 for 0 i n 1. Note that we used the same symbol
xi to represent a sequence of sites in Zd and the i-th coordinate of a site x. In
the following, to avoid confusions and whenever necessary, we will clarify the
meaning of xi .
Lemma 2.8 For each fixed pair of sites x, y in

N;K and each path (x; y) =
(x = x0 ; : : : ; xn = y ),
h n o2 i
EN;K rx;y ( ) f ( x;y ) f ( )
nX1 h n o2 i
42jjx yjj EN;K rxk ;xk ( ) f ( xk ;xk
+1
+1
) f ( ) :
k=0
Proof. The proof of this lemma relies on two observations. On the one hand, the
Dirichlet form associated to exchange of occupation variables (rather than jumps
of particles) is bounded above by the Dirichlet form associated to exclusion jumps.
This is clear because to exchange the occupation variables of two distinct sites
z0 and z1 , one may just perform few jumps from one site to the other. On the
other hand, to displace one particle from site x to site y , we may simply exchange
the occupation variables of sites x = x0 and x1 , than exchange the occupation
variables of sites x1 and x2 and repeat this procedure up to the point where
particles originally at x sit at site xn 1 . Then, one may let one particle jump from
site xn 1 to site xn = y and repeat back the exchange procedure to retrieve the
original configuration with one particle less at site x and one additional particle
at site y .
A rigorous proof requires some notation. For each pair of neighbor sites z0 ,
z1 and each configuration , denote by Tz0;z1 the configuration obtained from
changing the occupation variables (z0 ) and (z1 ) :
8
z < ( ) if z == z0 ; z1 ;
(Tz ;z )(z ) = (z1 ) if z = z0 and
0 1
:
(z0 ) if z = z1 :
We first claim that
Z h i2
f (Tz ;z ) f () N;K (d) 42 < Lz ;z f; f >N;K
0 1 0 1 (2:21)
for every f in L2 (N;K ) and every bond fz0 ; z1 g.

To prove this statement, for fixed 0 a < b , consider the set of con-
figurations such that (z0 ) = a, (z1 ) = b. For 0 k b a, let k = k ( ) =
+ kdz kdz and notice that 0 = , b a = TP z0 ;z1 and k+1 = (k )z1 ;z0 . We may
thus rewrite the difference f (Tz0;z1 ) f ( ) as 0kb a 1 [f ((k )z1 ;z0 ) f (k )].
0 1
In particular, by Schwarz inequality,

Z h i2
1f (z0 ) = a; (z1 ) = bg f (Tz ;z ) f () N;K (d)
0 1
bX
a 1Z h i2
(b a) 1f (z0 ) = a; (z1 ) = bg f ((k )z ;z ) f (k ) N;K (d) :
1 0
k=0
Performing a change of variables = k we obtain that the last expression is equal
to
bX
a 1Z h i2
(b a) 1f (z0 ) = a + k; (z1 ) = b kg f ( z ;z ) f ( ) N;K (d )
1 0
k=0
because N;K is the uniform measure. Summing over all 0 a < b gives
that Z h i2
1f (z0 ) < (z1 )g f (Tz0 ;z1 ) f ( ) N;K (d )
Z h i2
2 rz ;z () f (z ;z ) f () N;K (d) :
1 0
1 0
To conclude the proof of the claim it remains to repeat the same argument in
the case (z0 ) > (z1 ) and recall the explicit expression for the Dirichlet form
< Lz0;z1 f; f >N;K .
Recall now the definition of the path (x; y ) = (x0 ; : : : ; xn ) from x to y .
For a fixed configuration and for 0 k n 1, denote by k = k ( ) the
configuration Txk 1 ;xk Tx0 ;x1 , 0 = . Let n = n ( ) be the configuration
obtained from n 1 letting one particle jump from xn 1 to y : n = (n 1 )xn 1;y .
For 1 j n 1, let n+j = Txn 1 j ;xn j Txn 1;xn 2 n . Notice that 0 =
= x;y
P that 2n 1 . We may therefore rewrite the difference f ( ) f ( ) as
and x;y
0k2n 2 [ f ( k+1 ) f ( k )]. By Schwarz inequality,
n o2 n
2X2 n o2
f ( x;y ) f ( ) (2n 1) f (k+1 ) f (k ) :
k=0
Notice that k+1 = Txk ;xk+1 k for 0 k < n 1 and k+1 = Tx2n k 2 ;x2n k 1 k
for n k < 2n 1. Therefore, if we define zk as xk for 0 k n 1 and as
x2n k 2 for n k < 2n 1, we have that
k+1 = Tzk ;zk k
+1
for 0 k < 2n 1, k == n 1:
With this notation and since the jump rate rx;y is bounded by 1,
X h n o2 i
EN;K rx;y ( ) f (k+1 ) f (k )
0k2n 2
k==n 1
X hn o2 i
EN;K f (Tzk ;zk k ) f (k ) +1
:
k2n
0 2
k==n 1
Performing the change of variables = k , applying inequality (2.21) and recalling

the definition of zk , we obtain that the last expression is bounded above by
X h n o2 i
4 2 EN;K rxk ;xk ( ) f ( xk ;xk
+1
+1
) f ( ) :
0 kn 2
k==n 1
On the other hand, n = (n 1 )xn 1 ;y and rx;y ( ) = rxn 1 ;y (n 1 ) because
n 1 (xn 1 ) = (x). Therefore, performing the change of variables = n 1 ,
we get that
h n o2 i
EN;K rx;y ( ) f (n 1) f (n )
h n o2 i
= EN;K rxn ;y (n 1 1) f ((n 1 )xn ;y ) f (n
1
1)
h n o2 i
= EN;K rxn ;y () f (xn ;y ) f ()
1
1
This estimate together with the previous one concludes the proof of the lemma.

In dimension one the previous lemma states that for each fixed x > 0,
h n o2 i
EN;K r0;x ( ) f ( 0;x) f ( )
x 1
X h n o2 i
4 2 x EN;K ry;y+1 ( ) f ( y;y+1) f ( ) :
y=0
3. Spectral gap in dimension d2

We indicate in this section the main modifications required to extend the proof of
the spectral gap in dimension 1 to higher dimensions. To fix ideas we investigate
the 2-dimensional case.
To set up a recursive equation for the inverse of the spectral gap W (N ) we
need to introduce some notation. For each fixed N consider the sequence of sites
fxN;k ; 1 k 2N 1g defined by

(k; N ) for 1 k N
xN;k
k) for N + 1 k 2N
=
(N; 2N 1
3. Spectral gap in dimension d 2 395
so that
N =
N 1 [ fxN;k ; 1 k 2N 1g. To keep notation simple, we
shall denote xN;k simply by xk when no confusion arises.
For each 1 k 2N 1, denote by Fk = FN;k the -algebra generated
by f (xj ); 1 j k g, by k the configuration ( (x1 ); : : : ; (xk )), by
~k the set
N fx1 ; : : : ; xk g and by fk the conditional expectation EN;K [f jFk ]. Recall

the definition of the function fk introduced just before (2.2). Equation (2.2) states
that fk is equal to E
~ ;K M ( ) [fk ].
k k
Since Fk is an increasing sequence of -algebras, with the convention that
F0 is the trivial -algebra and that f0 = EN;K [f ], ffk ; 0 k 2N 1g is a
martingale. We may thus express the variance of f as
h i 2X2N h i
Var (N;K ; f ) = EN;K (f f2N 1)
2
+ EN;K (fk+1 fk )2 : (3:1)
k=0
The very same arguments presented in the previous section to derive inequality
(2.5) and the induction assumption permit to estimate the first term on the right
hand side of this identity :
h i
EN;K (f f2N 1)
2
W (N 1) D(N;K ; f ) :
To estimate the second term on the right hand side of (3.1), recall that ffk ; 0
k 2N 1g is a martingale. Therefore, for each 0 k 2N 2, we may rewrite
the expectation EN;K [(fk+1 fk )2 ] as
h 2 i
EN;K EN;K fk+1 EN;K [fk+1 j Fk ]
Fk
h 2 i
= EN;K E
k ;K ~ M (k ) fk+1;k ((xk+1 )) E
k ;K ~ M (k ) [fk+1;k ] :
In this formula fk+1;k ( (xk+1 )) is the function fk+1 evaluated at the configuration
(k ; (xk+1 )). We write it in this way to indicate that (xk +1) only is integrated with
respect to
~k ;K M (k ) . In particular, the term E
~ ;K M ( ) [(fk+1;k ( (xk+1 ))
k k
E
~k ;K M (k ) [fk+1 ])2] is the variance of a function depending only on one site.
Applying Lemma 2.3 we obtain that this variance is bounded above by

X n o2
B1 ()
1~k ;K M (k ) (b) fk+1;k (b 1) fk+1;k (b)
b=1

X n o2
= B1 ()
1~k ;K M (k ) (b) fk+1 (k ; b 1) fk+1 (k ; b) ;
b=1
where
1
~k ;K M (k ) stands for the one site marginal of
~k ;K M (k ) .
Since fk+1 is the conditional expectation of f given Fk+1 , by equation (2.2),
fk+1 (k+1 ) = E
k ~
+1 ;K M (k+1 ) [fk+1 ] :
In particular,
fk+1;k (b 1) fk+1;k (b)

= E
;K M b [fk ;b 1] E
k ;K M (k ) b [fk ;b ] :
k k~
+1 ( ) +1 ~
+1
Adapting the arguments presented between formula (2.9) and (2.11) to the 2-
dimensional setting, we show that this difference is equal to
n o
X
j
~k+1 j x2
~k+1 E
~k+1 ;K M (k ) b fk ;b 1( + dx ) fk ;b ( ) gk;K;k ;b ( (x))
1
n o
X
1
E f ;b g k;K; ;b x :
+
j
~k+1 j x2
~k+1
~k+1 ;K M (k ) b k ( ) k ( ( )) 1
In this formula gk;K;k ;b stands for the mean-one cylinder function defined by
gk;K;k ;b ( (x)) = Sj
~k+1 j;K M (k ) b [1 + (x)]1f (x) < g. Notice that the sec-
ond term is the covariance of fk ;b and gk;K;k ;b ( (x)) because the function
gk;K;k ;b ( (x)) has mean one.
The following result permits to estimate the first term in the above decompo-
sition. This lemma is the main difference between the proof in dimension one and
the proof in higher dimension.
Lemma 3.1 There exists a constant C () depending only on such that
N
2X2 X X

1~k ;K M (k )(b) ~1
k=0 b=1 j
k+1 j x2
~k +1
n o 2
E
k~ ;K M k b fk ;b 1 ( + dx ) fk ;b ( ) gk;K;k ;b ( (x))
+1 ( )
C () N D(N;K ; f ) :
Proof. With the same arguments presented before formula (2.14) in the proof
of Lemma 2.4, we obtain that the expression on the left hand side of the above
inequality is bounded by
N
2X1 X h n o2 i
C () xk ;x
j
~k+1 j x2
~k+1 N;K xk ;x f ( )
E r f ( )
1
( )
k=1
2N 1
C
N2( ) X X h n
EN;K rxk ;x () f (xk ;x ) f ()
o2 i
k=1 x2
N
for some constant C () depending only on and that may change from line to
line.
3. Spectral gap in dimension d 2 397
It remains to estimate this expression by the Dirichlet form of f . We shall

consider the case 1 k N . By symmetry the arguments apply to N < k
2N 1.
Recall the terminology and the notation introduced just before Lemma 2.8.
To each site x in
N and each 1 k N , consider a path (xk ; x) from xk
to x moving first along the ordinate and than along the abscissa. More precisely,
let (xk ; x) = (x = z0 ; z1 ; : : : ; zn = x) for n = jjxk xjj. Among all possible
paths from xk to x we choose the one with the following property : there exists
0 n0 < n so that zj +1 zj = e2 for 0 j n0 1 and yj +1 yj = e1 for
n0 j n 1.
By Lemma 2.8, the summation over 1 k N in last expression is bounded
above by
N X jjxkX
C () X xjj 1 h n o2 i
E N;K r zj ;z j ( ) f ( zj ;zj f ( )
N k=1 x2
N j=0 )
+1
+1
because all paths have length at most 2N .

Fix a bond b = (b1 ; b2 ). From the way we construct the paths, there are at
most N 2 different paths using the bond b. In particular, changing the order of
summation, we rewrite last formula as
C () X E hr ()nf (b ;b ) f ()o2 i X
N b=(b ;b ) N;K b ;b 1 2
1 2
1 2 k;x; fb ;b g (xk ;x) 1 2

X h n o2 i
C ( ) N EN;K rb ;b () f (b ;b ) f ()
1 2
1 2
b=(b1 ;b2 )
C ()N D(N;K ; f ) :
=
In the firstformula, summation over b is carried over all oriented bonds b in
N and the second summation is carried over all sites x and xk whose path
(xk ; x) contains b1 and b2 . Since by symmetry the same argument applies to
N < k 2N 1, the lemma is thus proved
The proof of Proposition 2.5 is easily adapted to higher dimensional processes
because the geometry is almost irrelevant. In the present context this proposition
states that for each fixed " > 0, there exists ` = `(") and N 0 = N 0 (") such that,

X

X

;K M (k )(b) ~1
1
b=1
~k
j
k+1j x2
~k +1
n o2
E
k ~
+1
fk ;b ( ) gk;K;k ;b ( (x)) 1
;K M (k ) b
C ()`W N2
(` + 1)
D(N;K ; f ) + N"2 Var(N;K ; f )
for all N N 0 (") and all 0 K (=2)N 2.
Summing over 1 k 2N 1, we get from this result and Lemma 3.1 that
for every " > 0, the second term on the right hand side of (3.1) is bounded above
by
C () N + `W (` + 1) D(N;K ; f ) + " Var(N;K ; f )
N N
for some ` = `(") and all N N 0 ("), 0 K (=2)N 2. To conclude the proof
of the spectral gap in dimension 2 it remains to argues as we did in the previous
section just after the statement of Proposition 2.2.
4. Closed and exact forms
We investigate in this section the closed and exact forms on , the space of all
configuration with at most particles per site. To justify some definitions and to
clarify the ideas, we briefly overview in the beginning of the section the closed
and exact forms on Zd. P
Consider Zd endowed with the distance jj(x1 ; : : : ; xd )jj = 1id jxi j and
recall from section 4 that a path (x; y ) = (x = x0 ; : : : ; xn = y ) from x to y is a
sequence of sites fzi ; 0 i ng such that z0 = x, zn = y and jjzi+1 zi jj = 1
for 0 i n 1. The integer n is called the length of the path and x, y the end
points. A closed path is a path whose end points are equal and a n-step path is a
path of length n. By analogy with the continuous case we define a closed form as
Definition 4.1 A collection u = fux = (u1x ; : : : ; udx ); x 2 Zdg is a closed form

on Zd if
uix + ujx+ei = ujx + uix+ej (4:1)
for every 1 i; j d and x 2 Zd.
uix should be interpreted as a i-th partial discrete derivative at x. It represents

therefore the price to jump from x to x + ei . With this interpretation condition (4.1)
means that the price to jump from x to x + ei and then from x + ei to x + ei + ej is
the same as the one to jump from x to x + ej and then from x + ej to x + ei + ej .
In other words, condition (4.1) imposes that the price of a 2-step path depends on
the path only through its end points.
Since uix stands for the price to jump from x to x + ei , uix is the price to
jump from x + ei to x. Moreover, condition (4.1) considers only “increasing" paths
(paths from x to x + ei + ej for some x in Zd and 1 i; j d). Nevertheless,
it follows from (4.1) that the price of any 2-step path depends on the path only
through its end points. Fix, for instance, x in Zd, 1 i = 6 j d and consider the
paths 1 = (x; x + ei ; x + ei ej ) and 2 = (x; x ej ; x + ei ej ). By definition,
the price of 1 is uix ujx+ei ej and the price of 2 is ujx ej + uix ej . These
prices are the same if ujx ej + uix = uix ej + ujx+ei ej . This equality follows from
(4.1) setting y = x ej .
4. Closed and exact forms 399
Like in the continuous case each function F : Zd ! R gives rise to a closed

form. These closed forms play a particular role and deserve a special terminology.
Definition 4.2 An exact form is a closed form for which there exists F : Zd ! R
such that
uix = F (x + ei ) F (x) (4:2)
for x in Zd and 1 i d. Such form is denoted by uF = (uF;1 ; : : : ; uF;d ).
Notice that condition (4.1) follows from identity (4.2), for uF;i F;j
x + ux+ei = F (x +
ei + ej ) F (x) = uF;j F;i
x + ux+ej . Notice, furthermore, that the exact form associated
to F + C coincides with the one associated to F for any function F and constant
C.
At this point we need to elucidate whether all closed forms are exact forms. To
prove that a closed form is an exact form we need to exhibit a function F : Zd ! R
for which u = uF . In this case, by (4.2), uix would be equal to the i-th partial
discrete derivative of F at x. The strategy to build such integral F of the closed
form u is thus clear. We fix a site, say the origin, and assign an arbitrary value to
F at this site since we have seen that the closed forms associated to F and F + C
are equal. Then, for a site x we find a path from 0 to x and define F (x) as the path
integral of u along this path. More precisely, let a 2 R and set F (0) = a. Fix a site
x and select a path (0; x) =P(0 = x0 ; x1 ; : : : ; xn = x) from the origin to x. Write
F (x) = a + F (x) F (0) = a + 0jn 1 F (xj+1 ) F (xj ). Since, by definition of a
path, sites xj , xj +1 are neighbors and since u should be the discrete derivative of F ,
the difference F (xj +1 ) F (xj ) can be expressed with the form u. In the case where
xk+1 = xk + ei for some 1 i d, F (xk+1 ) F (xk ) = uixk =< uxk ; xk+1 xk >,
provided < ; > stand for the inner product in Rd . In the case where xk+1 = xk ei
for some 1 i d, F (xk+1 ) F (xk ) = uixk+1 =< uxk+1 ; xk+1 xk >. Therefore,
if for two sites x, y , we denote by x ^ y the largest site smaller than x and y for the
natural partial order of Zd ((x1 ; : : : ; xd ) (y1 ; : : : ; yd ) if xi yi for 1 i d),
F (xk+1 ) F (xk ) =< uxk ^xk+1 ; xk+1 xk > and
nX1
F ( x) = a + < uxk ^xk ; xk+1 xk >
+1
(4:3)
k=0
The second term on the right hand side of this expression is the path integral of
the closed form u along the path (0; x) = (0 = x0 ; x1 ; : : : ; xn = x) that we shall
denote by I (0;x) (u). Therefore, if the closed form u is the exact form uF associated
to F , F should satisfy the above relation. In principle nothing prevent I (0;x) (u)
to depend on the particular path (0; x) chosen. If we can prove, however, that
the path integral I (0;x) (u) does not depend on the particular path chosen, formula
(4.3) defines a function F : Zd ! R and a straightforward computation shows that
F (x + ei ) F (x) = uix for all x 2 Zd and 1 i d, i.e., that u = uF .
The previous argument shows that if the state space is countable and simply
connected, the statement that all closed forms are exact forms follows from the
statement that the path integral I (0;x) (u) does not depend on the particular path
chosen but only on the end points 0, x.
Of course, we may extend the definition of a path integral for any pair of sites
x, y in Zd : for a closed form u and a path (x; y) = (x = x0 ; : : : ; xn = y) from
x to y, denote by I (x;y) (u) the path integral of u along (x; y) :
nX1
I (x;y) (u) = < uxk ^xk ; xk+1 xk >
+1
k=0
We claim that on Zd the integral I (x;y) (u) depend on the path (x; y ) only through
its end points x and y :
Lemma 4.3 Consider a closed form u on Zd and two paths 1( x; y), 2( x; y) from
x to y. Then,
I 1 (x;y)(u) = I 2 (x;y)(u) :
Proof. A straightforward induction argument, left to the reader, shows that it is
enough to prove the statement for the elementary paths 1 = (x; x ei ; x ei ej ),
2 = (x; x ej ; x ei ej ) for 1 i; j d. In this case the identity I 1 (u) =
I 2 (u) follows immediately from property (4.1) of the closed form u. We shall
prove in Lemma 4.9 below this statement in greater generality.
Corollary 4.4 On Zd all closed forms are exact forms.
Proof. Fix a closed form u. We have just proved that we may define unam-
biguously the path integral of u. In particular, the function F : Zd ! R given by
F (x) = I (0;x) (u) is well defined. Moreover, by definition of the path integral,
F (x + ei ) F (x) =< ux ; ei >= uix for every x in Zd and 1 i d. Therefore u
is the exact form uF .
We conclude the examination of closed and exact forms on a countable simply
connected space with an example of space that admits closed forms that are not
exact. For a positive integer N , consider TN the one-dimensional torus with N
points and the closed form u identically equal to 1. u is clearly not an exact form.
In fact the path integral now depends on the path chosen since I(0;1;:::x) (u) = x
and I(0; 1;:::x N ) (u) = x N for every 0 < x N .
To investigate closed and exact forms in the context of infinite particle systems,
we summarize the concepts and the main ideas introduced up to this point through
another perspective.
We started with a topological space (Zd; jj jj ) endowed with a discrete metric.
This discrete metric permitted to define paths between two sites. We then intro-
duced the concept of a closed form. Condition (4.1) can be interpreted as requiring
the path integral of the closed form u along any 2-step path (x; y ) to depend only
on the end points x, y . In Lemma 4.3 we extended this property to a finite length
path (x; y ). This result permitted to integrate u unambiguously and to prove that
all closed forms are exact forms.
Consider now the state space . For two configurations and , denote by
D(; ) the minimum number of nearest neighbor jumps in order to obtain from
. For example, is at distance 1 from a configuration if is obtained from
by a jump of a particle to a nearest neighbor site : = x;xei for some site x in
Zd and some 1 i d.
A path (; ) = ( = 0 ; : : : ; n = ) from a configuration to is a sequence
of configurations k such that every two consecutive configurations are at distance
1:
0 = ; n = and D(k ; k+1 ) = 1 for 0 k n 1:
To avoid confusion, we should point out that we consider always endowed
with the product topology and not the discrete topology generated by the distance
D. In particular, when referring to continuous functions, we mean continuous
functions with respect to the product topology.
To keep notation simple, for two sites x, y , denote by Hx0 (resp. Hx ) the
set of configurations with at least one (resp. at most 1) particles at site x :
Hx0 = f 2 ; (x) > 0g (resp. Hx = f 2 ; (x) < g) and by Hx;y the
set Hx0 \ Hy . Let x;y : Hx;y ! Hy;x be the operator that moves a particle from
x to y : 8
< (z ) z == x ; y ;
x;y
( )(z ) = (x) 1 z = x ;
:
(y ) + 1 z = y :
We may now introduce the closed forms. Consider a family u = f(u1x ; : : : ;
udx ); x 2 Zdg of continuous functions uix : Hx;x+ei ! R and interpret uix () as the
price to move a particle from site x to site x + ei when the configuration is . In
particular, the price to move a particle from x to x ei when the configuration is
is equal to uix ei (x;x ei ). We have seen in the first part of this section that
a closed form gives the same price for any 2-step path with equal end points. In
the present context of particle systems with the distance adopted above, there are
two types of 2-step paths. We may either move a particle two times or move two
particles one time each.
Fix a site x, 1 i; j d and a configuration in Hx;x+ei +ej . There are four
possible different 2-step paths from to dx + dx+ei +ej . The first one is obtained
letting a particle jump from x to x + ei and then from x + ei to x + ei + ej . Formally
this becomes 1 = (; x;x+ei ; x+ei ;x+ei +ej x;x+ei ). This path is possible only if
belongs to H1 = Hx+ei and its price, denoted by I 1 (u), is uix ()+ujx+ei (x;x+ei ).
The second path is obtained letting a particle jump from x to x + ej and then from
x + ej to x + ei + ej : 2 = (; x;x+ej ; x+ej ;x+ei +ej x;x+ej ). This path is defined
on H2 = Hx+ej and its price is I 2 (u) = ujx ( ) + uix+ej ( x;x+ej ). We may also let
first a particle jump from x + ei to x + ei + ej and then let a particle jump from x to
x + ei . We obtain in this way 3 = (; jx+ei ;x+ei +ej ; x;x+ei x+ei ;x+ei +ej ) defined
on H3 = Hx0 +ei with price I 3 (u) = ux+ei ( ) + uix ( x+ei ;x+ei +ej ). Finally, we may
let first a particle jump from x + ej to x + ej + ei and then let a particle jump from
x to x + ej : 4 = (; x+ej ;x+ei +ej ; x;x+ej x+ej ;x+ei +ej ). This path is defined
on the set H4 = Hx0 +ej and its price is I 4 (u) = uix+ej ( ) + ujx ( x+ej ;x+ei +ej ).
In the spirit of the beginning of this section, a closed form u has to assign the
same price for all different paths constructed above : for all fixed 1 i; j d,
I (u)
k
= I (u)
l
for all 1 k; l 4 and
2 Hk \ Hl \ Hx;x+e +e : (4:4) i j
Furthermore, for any 1 i; j d, any two sites x, y such that x + ei 6= y

and y + ej 6= x, and any configuration in Hx;x+e \ Hy;y+e , there are two
i j
ways to move a particle from site x to x + ei and from site y to site y + ej . We

imposed x + ei (resp. x) to be different from y (resp. y + ej ) because these cases
belong to the first type of 2-step paths where a particle moves twice. We may
first move the particle at x and then the particle sitting at y or in the other way
around. In the first case the path is ~1 (x; y ) = (; x;x+ei ; y;y+ej x;x+ei ). This
path is possible only if x;y dx + x+ei ;y+ej dx+ei belongs to Hx;x+ei \ Hy;y+ej .
This additional restriction must be imposed because, in the case where x = y for
instance, two particles leave site x. This path is thus possible only if has at least
two particle at x or, equivalently, if dx belongs to Hx0 . The price of this path,
denoted by I ~1 (x;y) (u), is equal to uix ( ) + ujy ( x;x+ei ). In the second case, the
path is ~2 (x; y ) = (; y;y+ej ; x;x+ei y;y+ej ). This path is possible under the
same restrictions and its price is I ~2 (x;y) (u) = ujy ( ) + uix ( y;y+ej ). Once more, for
u to be a closed form it must assign the same price for these two paths :
uix () + ujy (x;x+e ) = ujy () + uix (y;y+e )
i j
(4:5)
for every 1 i; j d, every sites x, y such that x + ei 6= y and x 6= y + ej , and every

configuration such that x;y dx + x+e ;y+e dx+e belong to Hx;x+e \Hy;y+e .
i j i i j
Notice that in conditions (4.4) and (4.5) we considered only increasing paths.
We leave to the reader to check that it follows from (4.4) and (4.5) that the price
of any 2-step path depends on the path only through its end points. We give just
an example to illustrate. Fix x, 1 i = 6 j d and assume that is a configuration
such that (x) > 0, (x + ei ) < , (x ej ) < , (x + ei ej ) < . We
want to show that the price for moving a particle from x to x ej and then
from x ej to x ej + ei is the same as the price for moving a particle from
x to x + ei and then from x + ei to x + ei ej when the configuration is .
The price of the first path is ujx ej ( x;x ej ) + uix ej ( x;x ej ) and the second
is uix ( ) ujx+ei ej ( x;x+ei ej ). These two expressions are equal if and only
if uix ( ) + ujx ej ( x;x ej ) = uix ej ( x;x ej ) + ujx+ei ej ( x;x+ei ej ). Setting
= x;x ej and y = x ej , we see that (y) > 0, (y + ej ) < , (y + ei ) < ,
(y + ei + ej ) < . Moreover, the last equality holds if and only if
ujy ( ) + uiy+e (y;y+e )
j
j
= uiy ( ) + ujy+e (y;y+e ) :
i
i
This last identity follows from (4.5).

We are now ready to define closed forms in :
Definition 4.5 A collection u = f(u1x ; : : : ; udx ); x 2 Zdg of continuous functions

uix : Hx;x+ei ! R is a closed form if it satisfies equations (4.4) and (4.5).
Here again uix ( ) has to be interpreted as a discrete derivative at for jumps
from x to x + ei , i.e., as the price to move a particle from x to x + ei when the
configuration is . For this reason we defined uix only for configurations with at
least one particle at x and less than particles at x + ei , otherwise the jump would
not be allowed.
Like in the previous setting, to each continuous function W : ! R is
associated a closed form :
Definition 4.6 A closed form u is said to be an exact form if there exists a

continuous function W : ! R such that
uix () = W (x;x+e ) W ()

i
(4:6)
for every x 2 Zd, 1 i d and configuration in Hx;x+e . This closed form is

denoted by uW = (uW;1 ; : : : ; uW;d ).
i
Conditions (4.4) and (4.5) follow from relation (4.6). On the one hand, for
every site x, 1 i; j d, 1 k 4 and configuration in Hk \ Hx;x+ei +ej ,
I k (u) = W ( dx + dx+ei +ej ) W (). This proves (4.4). On the other hand,
for every site x, y , 1 i; j d and configuration such that x;y dx +
x+ei ;y+ej dx+ei belongs to Hx;x+ei \ Hy;y+ej , I 1 (x;y)(u) = I 2 (x;y)(u) = W (
dx dy + dx+ei + dy+ej ) W ().
We now present two examples of closed forms that will play a central role in
the sequel.
Example 4.7 Fix 1 i d and denote by ai the closed form defined by
(ai )jx ( ) =
ai;j
x ( ) = i;j
for 1 j d, x in Zd and configurations in Hx;x+e . ai is a closed form since
j
it adds 1 whenever a particle jumps from some site x to x + eiP . This closed form
corresponds to the formal function Wi ( ) defined by Wi ( ) = x < x; ei > (x)
since Wi ( x;x+e ) Wi ( ) = i;j . This last observation indicates that ai is not
j
an exact form. To prove this statement, assume by contradiction that ai is the

closed form associated to a continuous function W . Denote by 1 the configuration
with all sites occupied by 1 particle and by 10 the configuration with no particles
at the origin and all other sites occupied by one particle. For a positive integer
k, set xk = kei and k as the configuration such that k (x) = 1 for x == 0,
xk ; k (0) = 0 and k (xk ) = 2. Since W is continuous and k converges to 10 ,
W (10 ) = limk W (k ). On the other hand, since k is obtained from 1 moving
k times a particle in the i-th direction, W (k ) = k + W (1). This leads to the
contradiction W (10 ) = 1.
Example 4.8 Let h be P

a cylinder function. Recall from Chapter 7 that we denote
by h the formal sum x x h. Define vh as
(vh )ix ( ) = vh;i

x ( ) = h ( x;x+e )
i
h ( )
P
for x in Zd, 1 i d and configurations in Hx;x+ei . Though x x h is a
formal sum, the difference h ( + dx+ei ) h ( + dx ) is well defined. We leave to
the reader to check that vh is a closed form that is not exact, unless h is constant.
We have just presented two examples of closed forms on that are not exact.
On finite cubes, however, all closed forms are exact forms. To prove this claim,
consider a finite cube in Zd and recall that designates the configuration
space f0; : : : ; g . We have seen in the beginning of this section that for countable
simply connected state spaces, the proof that all closed forms are exact forms
reduces to the proof that the path integrals of the closed form depend on the path
chosen only through its end points.
In order to define a path integral in our particle system setting, consider two
configurations and at distance 1. Denote by b+ (; ) the site at which the
configuration has one particle more than the configuration and by b (; ) the
site at which has one particle less than . With this notation,
= b ;b +
where b+ = b+ (; ) and b = b (; ). Denote by = (; ) the configuration

defined by
if b+ b ;
(; ) =
otherwise :
In this formula we adopted the order introduced in the first part of this section. In
one dimension, is equal to if is obtained from letting one particle jump
to the right and is equal to if in order to obtain from we need to move a
particle to the left.
Consider the exact form associated to a continuous function W : ! R. Fix
two configurations , at finite distance and a path (; ) P = ( = 0 ; : : : ; n = ).
We have that W ( ) W ( ) is equal to W (n ) W (0 ) = 0kn 1 [W (k+1 )
W (k )]. Fix 0 k n 1 and suppose that k+1 = x;x+ej k for some site
x and some 1 j d. In this case, with the notation just introduced, k =
(k ; k+1 ) = k , bk; = b (k ; k+1 ) = x, bk;+ = b+ (k ; k+1 ) = x + ej . On the
other hand, by (4.6), W (k+1 ) W (k ) = uW;j x (k ). We may rewrite this expression
as < uW ;
bk; ^bk;+ k k;+ k;
( ) b b > . In this formula < ; > stands for the inner
product on Rd and uW W;1 W;d
bk; ^bk;+ for the vector (ubk; ^bk;+ ; : : : ; ubk; ^bk;+ ). In contrast,
in the case where k+1 = x;x ej k for some site x and some 1 j d, k =
(k ; k+1 ) = k+1 , bk; = b (k ; k+1 ) = x, bk;+ = b+ (k ; k+1 ) = x ej . Since,
by (4.6), W (k ) W (k+1 ) = uW;j x ej (k+1 ), we have that W (k+1 ) W (k ) =<
uW ;
bk; ^bk;+ k k;+ k;
( ) b b > . In conclusion, we have that
nX1
W ( ) W ( ) = < uW
b ^b (k ); bk;+ bk; >
k; k;+
k=0
By extension, if u is a closed form on and (; ) = ( = 0 ; : : : ; n = ) is
a path from to , we denote by I (;) (u) the path integral of u along (; )
defined by
nX1
I (;) (u) = < ub ^b (k ); bk;+ bk; >
k; k;+
(4:7)
k=0
Lemma 4.9 On the path integral of a closed form depends on the path chosen
only through its end points.
Proof. We shall prove this lemma in dimension 2. Fix a closed form u. To prove
the lemma, we have to show that the path integral of u along any closed path
vanishes. Consider a configuration and a closed path (; ) = (0 ; : : : ; n ).
For each 0 k n 1, let xk = b (k ; k+1 ), yk = b+ (k ; k+1 ) so that k+1 =
xk ;yk k . There exists ` large enough so that xk , yk belong to the cube ` for
every 0 k n 1. In order to enumerate all sites of ` , define J : ` !
f0; : : : ; (2` + 1)2 1g by

(2` + 1)(x2 + `) + (x1 + `)
x2 + ` is even ;
if
J (x1 ; x2 ) =
(2` + 1)(x + `) + ( x + `) if
2
x2 + ` is odd
1
and define implicitly the sequence fzk ; 0 k (2` + 1)2 1g by zJ (x) = x so

that ` = fz0 ; : : : ; z(2`+1) 1 g. 2
The proof is divided in two steps. We first reduce the problem to a one-
dimensional problem by constructing a new closed path ~(; ) = (0 ; : : : ; m )
that uses only bonds (zj ; zj 1 ) and whose path integral I ~(;) is equal to the
original path integral I (;) :
I ~(;) = I (;) and k+1 = kz ;z (4:8)
j j 1
for some 0 j (2` + 1)2 1 and all 0 k m 1.

In order to construct such new path, fix 0 k n 1 and recall that
k+1 = x ;y k . It is enough to show that there exists a path ~k (k ; k+1 ) =
k k
(k = 0k ; : : : ; m
k = k+1 ) fulfilling the second requirement of (4.8) and such that
I ( ; ) = I ~ ( ; ) , where k (k ; k+1 ) stands for the one step path leading
k k k+1
from k to k+1 . By definition of the sequence fzj ; 0 j (2` + 1)2 1g, if

k k k+1
yk = xk e1 , there exists j such that xk = zj and yk = zj1 . In this case nothing

has to be done and we may take ~k (k ; k+1 ) = k (k ; k+1 ).
Assume now that yk = xk + e2 . The case where yk = xk e2 is treated in
an analogous way. Let xk = (x1k ; x2k ) and assume without loss of generality that
x2k + ` is an odd number. In this case the path (k ; ( `;x );( `;x +1) k ) fulfills the 2
k
2
k
second property in (4.8) because J ( `; x2k + 1) = 1 + J ( `; x2k ). There is therefore

nothing to prove if x1k = ` since we may set ~k (k ; k+1 ) = k (k ; k+1 ).
Fix now ` < a ` and assume that x1k = a so that k+1 = (a;xk );(a;xk +1) k .
2 2
We shall construct a new 3-step path ~k;a (k ; k+1 ) that uses the bonds (xk ; xk
e1 ), (xk e1 ; yk e1) and (yk ; yk e1) and whose path integral coincides with
I k (k ;k+1 ) (u). Notice that this new path does not use the bond (xk ; yk ) but the
bond (xk e1 ; yk e1 ) instead.
There are four possible cases that must be treated separately. According to the
values of k (xk e1 ) and k (yk e1 ), k belongs to at least one of the following
sets : Hx0 k e1 \Hy0 k e1 , Hx0 k e1 \Hyk e1 , Hxk e1 \Hy0 k e1 and Hxk e1 \Hyk e1 .
We consider the first one and leave the other cases to the reader. Assume that k
belongs to Hx0 k e1 \ Hy0 k e1 . Consider the path ~k;a (k ; k+1 ) given by
(k ; yk e1 ;yk k ; xk e1 ;yk e1 yk e1 ;yk k ;

xk ;xk e1 xk e1 ;yk e1 yk e1 ;yk k = k+1 ) :
We claim that the path integral of the closed form u along this path is equal to
I k (k ;k+1 ) (u). To prove this claim, notice that the path integral along ~k;a (k ; k+1 )
is equal to
u1y e (k ) + u2x e (y e ;y k ) u1x e (k+1 )

k 1 k 1
k 1 k
k 1
because k+1 = x ;x e x e ;y e y e ;y k . By (4.4), this expression is

k k 1 k 1 k 1 k 1 k
equal to u2x (k ) = I ( ; ) (u). Therefore both path integrals are the same.
On the other hand, the new path k;a (k ; k+1 ) does not use the bond (xk ; yk )
k k k k+1
but the bond (xk e1 ; yk e1 ) instead. Repeating this procedure x1k + ` times
we obtain a path ~k (k ; k+1 ) fulfilling the second requirement of (4.8) and such
that I ~ ( ; ) (u) = I ( ; ) (u). Juxtaposing these paths we prove (4.8) and
k k k+1 k k k+1
conclude the first step.

Consider now a closed path (; ) = (0 ; : : : ; n ) such that for all 0 k
n 1, k+1 = kzj ;zj1 for some j . We want to prove that the path integral along
this path vanishes. Notice that this is a one-dimensional problem since only jumps
from sites zj to zj 1 are allowed.
The strategy consists in constructing a new path with length n 2 and same
path integral. The idea to construct such a new path is simple. Suppose that a
particle jumps from some site x to y and immediately after from y to x. A new
closed path of length n 2 can be constructed suppressing these two jumps and it
is easy to show that both path integrals are equal. If there are no such consecutive
jumps, a new closed path can still be constructed. We just need to change the
order of the jumps, preserving the path integral by property (4.5), up to obtain
two consecutive jumps as described before.
Recall that we are now considering a path ( = 0 ; : : : ; n = ) such that for
each 0 k n 1 k+1 = xk ;yk k where xk = zj and yk = zj 1 for some j .
Assume without loss of generality that J (y0 ) = J (x0 ) + 1. Since (; ) can be
interpreted as a closed, one-dimensional path on f0; : : : ; gZ and a particle jumped
from zJ (x0 ) to zJ (x0 )+1 , later a particle must jump from zJ (x0 )+1 to zJ (x0 ) . Let k0
be the first time when this happen : k0 = minfk 0; xk = zJ (x0 )+1 ; yk = zJ (x0 ) g
and let k1 be the last time before k0 that a particle jumps from zJ (x ) to zJ (x )+1 :
k1 = maxfk k0 ; xk = zJ (x ) ; yk = zJ (x )+1 g.
0 0
We shall assume without loss of generality that between k1 and k0 , there are
0 0
no jumps from some site x1 to some site y1 and then a jump from site y1 to x1 .
Otherwise we can repeat the same arguments to x1 , y1 in place of x0 , y0 .
Consider the path (; ) between its k1 and k0 -step. and denote it by
k ;k +1 (k ; k +1 ). Keep in mind that k +1 = y ;x k by definition of k0 . Denote
0 0
by ~1 (k ; k +1 ) the path (k ; : : : ; k 1 ; k ; k +1 ), where k = y ;x k 1 . The

1 0 1 0 0 0
0 0
1 0 1 0 0 0 0 0
difference between this path and the original one is that in the penultimate step
instead of moving a particle from xk0 1 to yk0 1 and then a particle from y0 to
x0 , we inverted the order of these jumps. We leave to the reader to check that the
assumption made in last paragraph guarantees that we may change the order of the
jumps and that property (4.5) ensures that the path integrals along both paths are
the same. Repeating this procedure k0 k1 times we obtain at the end a new closed
path ~k0 k1 (; ) = (0 ; : : : ; n ) whose path integral is equal to the path integral
of (; ) and such that 0 = , 1 = x0 ;y0 and 2 = y0 ;x0 1 = . Consider
now the (n 2)-step path obtained from ~k0 k1 suppressing the first two jumps :

1 (; ) = (; 3 ; : : : ; n ). By definition of the path integral, the price to move a
particle from x to y for the configuration is equal to the price to move a particle
from y to x for the configuration x;y . In particular, the path integral of u along

1 (; ) is equal to the path integral along (; ). Repeating this argument n=2
times we obtain that the path integral along (; ) vanishes.
From this result and the proof of Corollary 4.4, it follows that on finite cubes,
all closed forms are exact forms :
Corollary 4.10 Let be a finite cube in Zd. All closed forms on are exact
forms.
Proof. Let u be a closed form on . Fix 0 K jj and a configuration

. Set an arbitrary value F (K ) for F at K . Since
K in the hyperplane ;K
there is a
is simply connected, for every configuration in the hyperplane ;K
path (k )0kn from K to . Define F ( ) by
F ( ) F (K ) + I ( ;)(u) :
= K

It follows from the previous lemma that F is a well defined function on ;M ( ) .
F
K
Moreover, a simple computation shows that u( ) = u ( ) for all configurations

of ;M ( ) . This shows that u is an exact form.
K

Notice that in the proof of the previous Corollary, we have a degree of freedom
in the definition of F on each hyperplane ;K because the value of F (K ) is
arbitrary. We may choose, for instance, F (K ) so that F has mean zero with
respect to all canonical measures ;K .
Since all closed forms are exact forms on simply connected, finite subsets of
Zd, on the cube ` = f `; : : :; `gd , the closed forms fai ; 1 i dg and vh
introduced in Examples 4.7 and 4.8 are exact forms. It is easy to check that the
closed
P form ai is the exact form associated to the continuous function Wi ( ) =
x2` < x; ei > (x).
To introduce a special class of closed forms, we needed to extend its definition
to include collections of L2 ( ) functions.
Definition 4.11 A collection u = f(u1x ; : : : ; udx ); x 2 Zdg of L2 ( ) functions uix :

Hx;x+ei ! R is a closed form if it satisfies equations (4.4) and (4.5) in L2( ).
We may now introduce the main object of this section.
Definition 4.12 A class of L2 ( ) functions g = fgi : H0;ei ! R; 1 i dg is

a germ of a closed form if the collection f(g1x; : : : ; gdx ); x 2 Zdg defined through
translations of g by the formula
gix() = (x gi )( ) (4:9)
is a closed form. Notice that gix is defined on Hx;x+e i
as required. The space of
germs of closed forms is denoted by G.
For a cylinder function h and a bond (x; y ), denote by (rx;y h): Hx;y ! R
the cylinder function defined by (rx;y h)( ) = h( x;y ) h( ). We extend the
definition of rx;y h to setting (rx;y h)( ) = 0 if 62 Hx;y : (rx;y h)( ) =
1f (x) > 0; (y ) < g[h( x;y ) h( )]. We sometimes abbreviate r0;ei h by
ri h and we denote by rh the vector (r1h; : : : ; rd h).
Examples 4.7 and 4.8 provide two types of germs of closed forms. For 1 i
d, consider the class of cylinder function Ai = fAi;1 ; : : : ; Ai;dg with Ai;j = i;j ,
1 j d. The collection f(Ai; x ;i: : : ; Ax ); x 2 Z g obtained from A through
1 i;d d i
formula (4.9) is the closed form a of Example 4.7.
For a cylinder function h, the collection f(rx;x+e1 h ; : : : ; rx;x+ed h ); x 2
Zdg obtained from formula (4.9) through the cylinder function (r1 h ; : : : ; rd h )
is the closed form vh of Example 4.8. These germs of closed forms associated to
translations of cylinder functions are called germs of exact forms :
Definition 4.13 The germ g of a closed form is said to be the germ of an exact
form if there exists a cylinder function h such that g = r h . The space of germs
of exact forms is denoted by E:
E = fr h; h is a cylinder function g:
The goal of the remaining of this section is to prove that in L2 ( ) the germs
of closed forms are generated by the germs of exact forms and by the germs
fAi ; 1 i dg introduced above. Let G and E denote, respectively, the
closure in L2 ( ) of germs of closed forms and of germs of exact forms :
G = G in L2( ) and E = E in L2( ) :
Theorem 4.14 In L2 ( ) the space of germs of closed forms is a direct sum of the
linear space generated by the germs fAi ; 1 i dg and the closure of germs of
exact forms:
G = fAi ; 1 i dg E :
Proof. G , E are closed linear subspaces of L2 ( ). We have already seen that

for any 1 i d and any cylinder function h, the germs Ai , r h belong to G.
In particular, fAi ; 1 i dg + E is a subset of G . To conclude the proof of
the theorem it remain to show that the sum is direct and that
G fAi; 1 i dg + E :
We first prove the inclusion. The set E = fr h ; h cylinder functiong is a
linear subspace of L2 ( ) and so is fAi ; 1 i dg + E. It is well known that
in a Hilbert space the strong closure of a linear subspace coincides with its weak
closure (cf. Theorem V.1.11 of Yosida (1997)). Therefore, we have just to prove
that for every germ g of a closed form there is a sequence of cylinder functions
( nP ) and constants C1 ; : : : ; Cd such that r n converges weakly in L2 ( ) to
g + i Ci Ai .
The strategy of the proof of the Theorem is easy to understand. Consider a
germ g of a closed form f(g1x ; : : : ; gdx ); x 2 Zdg and recall that gix = x gi . Assume,
to fix ideas, that g1 ; : : : gd are cylinder functions. We proved in Corollary 4.10 that
all closed forms on are exact forms if is a cube of finite length. We are thus
tempted to project the closed form f(g1x ; : : : ; gdx ); x 2 Zdg on to recover a
cylinder function n whose discrete partial derivatives are gix . More precisely, fix
a positive integer n and denote by Fn the -algebra generated by f (x); x 2 n g.
For 1 i d and a site x in n;i = fy 2 n ; y + ei 2 n g, define gi;n x as the
conditional expectation of gix given Fn :
h i
gi;n
x = E gix Fn :
It is easy to check that fgi;n
x ; x 2 n;i g is a closed form on n =
n
. By
Corollary 4.10, there exists a function en Fn -measurable such that
rx;x+e en = gi;n for all sites in i;n and

x
i
h
x i
X
E enj x
( ) = K= 0 for all 0 K jn j :
x2
n
This last requirement picks a function among the one-parameter family of integrals
.
constructed in Corollary 4.10 on each hyperplane n;K
Since g is a cylinder function, gx = x g is Fn -measurable for x not too
i i i
close from the boundary of n . In particular, there exists n0 = n0 (g) such that
gi;n i
x = x g for all x in n n0 and
x r0;e x en
i
= rx;x+ei
en = gi;n
x = x gi
for x in n n . The first equality follows from the identity z ry;y+e = ry+z;y+z+e
z for z , y in Zd and 1 j d, to be used repeatedly hereafter. Averaging over
0 j j
x in n n , this identity shows that

0
X
gi = r0;e i
[2(n
1
n0 ) + 1]d x en :
x 2 n n0
The difference between this last average and r0;e i [2( n n0 )+1] d en = [2( n n0 ) +
1] d r0;e i en is equal to
X
~ni = r0;e [2(n n0 ) + 1] d
i
x en
x 2 n+1 n n0
because en is Fn -measurable and thus r0;ei x en vanishes for x 62 n+1 . In

particular, the inclusion G fA1 ; : : : ; Ad g + E will be proved as soon as we
show that the boundary term ~ni is a weakly relatively compact sequence whose
limit points belong to the linear space generated by Ai . This is the strategy of the
proof.
The proof is divided in several steps to detach the main ideas. We first project
the closed form f(g1x ; : : : ; gdx ); x 2 Zdg on finite cubes and apply Corollary 4.10
to obtain the cylinder function en . In steps 2 and 3 we estimate the L2 ( ) norm
of en and show that a modification Rni of the boundary term ~ni is a weakly
relatively compact sequence. In step 4 we prove that the non boundary term in
(2n) dr0;ei e converges strongly in L2 ( ) to the germ gi . Finally, in steps 5
and 6 we show that all limit points of the sequence Rni belong to the linear space
n
generated by Ai .
Step 1. Projection to finite boxes. Fix a germ (g1 ; : : : ; gd ) of a closed form
and recall the definition of Fn , gi;n x and en 2 Fn given in the beginning of
the proof. For 1 i d, consider the expression (2n) dr0;ei e . Since en is
Fn -measurable, this sum is
n
X
1
(2n)d
r0;e i en = (2 )d n r0;e
1
i
x en :
nx n+1
jx jn; j==i
i
j
P
It is not difficult to see that (2n) dr0;ei x2n ; n<xi n x en converges in
L2 ( ) to gi (the proof given in Step 4 for a slight modification of en applies
also to en ). The problem comes from the boundary terms
X X
ni =
(2n)d
1
r0;e i
y en +
1
(2n)d
r0;e i
y en : (4:10)
y2 ; y = n
n i y2e + ; y =n+1
i n i
The proof of the existence of a subsequence ni k that converges weakly to a

multiple of Ai is derived in two steps. We first show that the sequence fni ; n 1g
is weakly relatively compact and then prove that all limit points of this set are in
the linear subspace generated by Ai .
By Alaoglu’s theorem, the unit ball for the strong topology is weakly rela-
tively compact (cf. Theorem V.2.1 of Yosida (1997)). Therefore, to prove that the
sequence fni ; n 1g is weakly relatively compact we just need a bound on the
L2 ( ) norm of ni . To prove such a bound we first estimate the L2 ( ) norm of
en .
Step 2. Bound on the L2 ( ) norm of en . Since by construction, for each 0
K jnj, en has mean 0 with respect to n ;K , by the spectral gap proved in
sections 2 and 3, there exists a universal constant C1 such that
D E
( en )2

C1 n2D ; en ; n
where D(n ; ) is the Dirichlet form with respect to n . This is the unique
point in the proof of hydrodynamic limit of nongradient systems where a sharp
estimate on the spectral gap for the generator restricted to finite boxes is needed.
An estimate of the Dirichlet form en is easy to derive. By definition of en , for
x in n;i , rx;x+ei en = gi;nx . By Schwarz inequality and by translation invariance
of , < (gi;n
x ) 2
> is bounded above by < ( gi )2 > =< (gi )2 > . Therefore,
x
d X D
X 2 E
D( ; en )
n
= (1=2) rx;x+e i
en

i=1 x2 n;i
X
Cnd < (gi )2 > :
1 id
Henceforth we use the shorthand kgk2 for P1id (gi )2. The two previous esti-
mates show that
D E D E D E
D( ; en ) Cnd kgk2 ;
n
( en )2

Cnd+2 kgk2
for some universal constant C .
Step 3. Bound on the L2 norm of the boundary terms. We start with an estimate
of the L2 ( ) norm of ni in terms of the Dirichlet form of en . To understand
the difficulty, fix a site x in ei + n such that xi = n + 1. x en is therefore a
function of f (x); xi 1g and r0;e x en ( ) = (x en )( + de ) (x en )( ). We
i i
have thus to estimate the modification caused on en by the creation of a particle

with a Dirichlet form that measures only displacement of particles. Next lemma
shows that such an estimation is possible provided we smooth out en by taking
conditional expectations.
Lemma 4.15 Consider a F3n -measurable function f . Denote by h the conditional

expectation of f with respect to Fn : h( ) = E [f j (x); x 2 n ]. For each
1 i d, denote by @i;+ n the positive boundary of n in the i-th direction :
@i;+ n = fy 2 n ; yi = ng. There exist finite constants C1 (), C2 () such that
X hn o2 i
E h( dy ) h() 1f (y ) > 0g
y2@ +
i; n
C1() n 1 E [f 2] + C2 () n D( ; f ) : 3n
A similar statement holds if the positive boundary @i;+ n , dy and 1f (y ) >
0g are replaced by the negative boundary @i; n = fy 2 n ; yi = ng , + dy
and 1f (y ) < g.
Proof. Set i = 1 and fix a site y in @1;+ n . To keep notation simple, we shall
denote the configurations of 3n (resp. n , 3n n ) by the symbol (resp. ,
). Recall from section 4 that for each function f on 3n and each configuration
of n, f : 3n n ! R stands for the function defined by f ( ) = f (; ) for
all configurations of 3n n .
Since h is defined as the conditional expectation of f , we have that h( ) =
E [f j ] = E [f ( )]. In this last expectation it should be understood that inte-
gration is made with respect to the variable , while is kept frozen. In particular,
if we denote dy by y , the difference h( dy ) h( ) can be rewritten as
E [f y ( )] E [f ( )].
Denote by y;n the hypercube n 1 translated by y + (n + 1)e1 : y;n =
y + (n + 1)e1 + n 1 . Notice that y;n 3n for all y in @1;+ n . We may rewrite
E [f y ( )] as
X h i X h i
1
jy;nj z2 E f ( ) (z )
y +
1
jy;nj z2 E f ( )[ (z )] :
y
y;n y;n
In the first term, for each z in y;n, we perform a change of variables ~= dz
to rewrite it as
() X E hf ( + d )f1 + (z )g1f (z ) < gi
jy;n j z2 z y
y;n
X hn o i
=
1
jy;n j z2 E f ( + dz ) f ( ) g( (z ))
y
y;n
X h i
+
1
jy;nj z2 E f ( )g( (z )) :
y;n
In this last formula g ( (z )) stands for (()=)[1 + (z )]1f (z ) < g. Subtracting
h( ) = E [f ] we obtain that the difference h( dy ) h( ) is equal to
X hn o i
1
jy;n j z2 E f ( + dz ) f ( ) g( (z ))
y
y;n
h X i
+ E f ( ) j1 j fg( (z )) 1g
y;n z2 y;n
h X i
+ E f ( ) [1 ( (z )=)] :
1
j j
y
y;n z2y;n
The expected value of g ( (z )) and of (z )= with respect to is equal to 1.

Therefore, the second and third terms are covariances of f and f y with averages
of cylinder functions. In particular, by Schwarz inequality, the square of the second
and third term in the above decomposition are bounded by C ()jy;n j 1 fE [f2 ]+
E [f2y ]g for some finite constant C () depending only on . On the other hand,
by Schwarz inequality, the square of the first term is bounded by
X hn o2 i
C () j1 j E f ( + dz ) f ( ) 1f (z ) < g
y
y;n z2 y;n
for some finite constant C () because g ( ) is bounded. Hereafter the constant

C () may change from line to line.
Recall that E [f ( )] = E [f j ]. From the previous bounds and Schwarz
inequality, we obtain that the expected value, with respect to , of [h( dy )
h( )]21f (y) > 0g is bounded by
n o
C ()jy;nj 1
E [f 2] + E [f 2 ( dy )1f(y) > 0g]
X h n o2 i
C ( ) j 1 j
+ E ry;z () f (y;z ) f () :
y;n z2 y;n
Performing a change of variables 0 = dy in the second expectation and

summing over y in the positive boundary of n , we obtain that the first two terms
are bounded by C ()n 1 E [f 2 ]. On the other hand, by Lemma 2.8 and like in
the proof of Lemma 3.1, the third expression is bounded by C ()nD( ; f ) for 3n
some constant C () depending only on .

We have now all elements to prove that a slight modification of the sequence
ni defined in (4.10) is bounded in L2 ( ). In view of Lemma 4.15, for each
positive integer n, let h i
n = E e3n jFn :
By Schwarz inequality and the translation invariance of , for each 1 i d,
2
nCd+1
X X h 2 i
E 1
(2n)d
r0;e i
y n E ry;y+e n i
y2@ i; n y 2@ +
i; n
because y r0;ei = ry;y+ei y for all y in Zd. By Step 2, < ( e3n )2 > is bounded
above by Cnd+2 < kgk2 > and the Dirichlet form D(n ; e3n ) is bounded by
Cnd < kgk2 > . Therefore, Lemma 4.15 applied to the function n shows that
the right hand side ofPthe last formula is bounded by C () < kgk2 > . The same
argument applies to y2ei +n ; yi =n+1 y n .
In view of this estimate, for 1 i d, define Rn = (Rn1 ; : : : ; Rnd ) by
X
Rni =
1
(2n)d
r0;e i n
=
1
(2n)d
r0;e i
x n :
x
Rni can be decomposed as
X X
1
(2n)d
r0;e i
y n +
(2n)d
1
r0;e i
y n
y2 y 2
n<y n y= n
n n
i i
X
+
1
(2n)d
r0;e i
y n :
y2e +
y =n+1
i n
i
We shall refer to the first sum as the non boundary term, to the second one, denoted
by ni; , as the negative boundary term and to the third sum, denoted by ni;+ , as
the positive boundary term. We have just proved that the sequences ni;+ , ni; are
bounded in L2 ( ) and are thus weakly relatively compact. In the next step we
show that the non boundary term converges to gi in L2 ( ) and in the Steps 5 and
6 we prove that all limit points of the sequence ni; + ni;+ belong to the subspace
generated by Ai .
Step 4. Convergence of the non boundary term to gi . For a fixed y in n such
that n < yi n we have

r0;e y n = y r y; y+e n = y r y; y+e E e3n j Fn :
i i i
Since n < yi n we may move the gradient r y; y+e inside the expectation.
By definition, r y; y+e e3n = gi;3yn = E [gi y j F3n ]. Therefore,
i
i h
r0;e y
i
E gi y j F3n j Fn
n = y E = y gi;ny
and the first term in the decomposition of Rni is

X
1
(2n)d
y gi;ny :
y2
n<y n
n
i
By the martingale convergence theorem g0i;n ! g0i = gi in L2 ( ), as n " 1.

Therefore, for every " > 0, there exists n0 2 N such that < (g0i;n gi )2 > "
for every n n0 . In particular, since is translation invariant, for every n n0
and every y in n n0 ,
D 2 E D 2 E
gi;n
y giy
= E y gi j Fn y gi
D 2 E
E y gi j Fy+ n0
y gi
D 2 E
= y E gi j Fn 0 y gi
":
Thus, for n n0 , we have that
D X 2 E
1
(2n)d
y gi;ny

gi
y2 n;n<y n i
C (d)n0
(gi )2 + 1 X D
gi;n giy
2 E
n (2n)d y
y 2 n n0
C (dn)n0 gi 2

+ ":
This estimate shows that
X X
1
(2n)d
r0;e i
y n = (2n) d y gi;ny
y2 y2 n; n<y n
n<y n
n i
i
converges to gi in L2 ( ).
We turn now to the boundary terms. By Step 3 the sequences (n1; ; : : : ; nd; )
and (n1;+ ; : : : ; nd;+ ) are weakly relatively compact. To prove that all limit points
belong to the linear space generated by fAi ; 1 i dg, consider a weakly con-
vergent subsequence. To keep notation simple assume that the sequences converge
weakly and denote by b+ (resp. b ) the weak limit of the positive (resp. negative)
boundary term.
We start proving that for each 1 i d, b+i and bi depend on only through
(0) and (ei ).
Step 5. For 1 i d, bi ( ) = bi ( (0); (ei)). We shall prove this statement
for bi ; the same arguments apply to bi+ . Notice that the negative boundary term
ni; is measurable with respect to the -algebra generated by f(ei ); (x); x 2
Zd; xi 0g. Of course the weak limit bi inherits this property.
We claim that rz;z+ej bi = 0 for all bonds fz; z + ej g in the half space
Zi; = fx 2 Zd; xi 0g that do not intersect the origin, i.e., such that z ,
d
z + ej == 0. We prove at the end of this step that this property implies that bi
depends on only through (0), (ei ). Notice that we do not require here j to be
different from i.
Fix such a bond and assume that n is large enough so that z , z + ej belong to
n . Since rz;z+ej is continuous with respect to the weak topology, it is enough
to show that the sequence rz;z+ej ni; converges to 0 in L2 ( ). Since fz; z + ej g
does not intersect f0; ei g, rz;z+ej and r0;ei commute. In particular, for each y in
n ,
h i
rz;z+e r0;e y
j
r0;e rz;z+e y E e3n j Fn :
i n = i j
Therefore, by Schwarz inequality, < (rz;z+e ni; )2 > , which is equal to j
D X 2 E
(2n) d r0;e rz;z+e y n
y2
i j

y= n
n
i
Cnd(+1d)
X D E
(r0;ei rz;z+ej y n )2
y2
y= n
n
i
for some finite constant C (d) that hereafter may change from line to line.
To estimate the right hand side, we have to consider three different cases.
Either the bond z , z + ej belong to y + n or z , z + ej belong to (y + n )c or
fz; z + ej g links y + n to its complement. In the first case, since z , z + ej belong
to y + n , z y , z y + ej belong to n . In particular, rz y;z+ej y and the
conditional expectation with respect to Fn commute. Thus,
h i
rz;z+e y
j n = rz;z+e y E j
e3n j Fn
h i h i
rz j;3n
= y E y;z+e y e3n j Fn = y E gz y j Fn
j
because rz;z+ej y = y rz y;z y+ej . The L2 norm of the last expression is

1=2
bounded above by < kgk2 > . This proves that
C ( d) X D
E
d
n y2 ; y = n
+1
y;z;z+e 2y+
n i
j n
vanishes as n " 1.
On the other hand, if z , z + ej belong to (y + n )c , rz;z+ej y E [ e3n j Fn ]
vanishes because y E [ e3n j Fn ] is measurable with respect to f (x); x 2 y + n g.
Therefore,
C (d) X D
E
= 0:
n d +1
y2 ; y = n
y;z;z+e 2(y+
n i
c
j n)
It remains to consider the case where fz; z + ej g links y + n to (y + n )c .

Assume that z 2 y + n and z + ej 2 (y + n )c , the other possibility is handled in
a similar way. In this case, yj = zj n and we are reduced to estimate
C (d) X D(r r E
:
nd+1 y2 ; y = n 0;e y z y;z y + e n ) 2
i j
y =z n
n i
j j
Notice that j 6= i because fz; z +ej g and f0; ei g are disjoints. Since < (r0;ei f )2 >
4 < f 2 > and is translation invariant, setting x = z y, we have that the
last expression is bounded above by
C (d) X D(r E
)2 :
n d +1 x;x+e n j

x =xn;2zx+=z +n
j i
n
i
It is not hard to adapt the proof of Lemma 4.15 to show that this expression
vanishes as n " 1.
Up to this point we proved that rz;z+ej bi vanishes for all bonds fz; z + ej g in
the half space Zdi; that do not intersect the origin. We claim that this property in
addition to the fact that bi is measurable with respect to f (ei ); (x); x 2 Zdi; g
implies that bi is a function of (0) and (ei ) only : bi = E [bi j (0); (ei )].
To prove this statement, consider a sequence bi;(n) of finite approximations of
bi : bi;(n) = E [bi j Fn ]. It is easy to check that sequence bi;(n) inherits from bi
both properties : bi;(n) is measurable with respect to f (ei ); (x); x 2 n \ Zdi; g
and rz;z+ej bi;(n) = 0 for all bonds fz; z + ej g in n \ Zdi; that do not intersect
the origin. In particular,
h X i
bi;(n) = E bi;(n) j (0); (ei); ( x) ; (4:11)
x 2 i;
n
where i;n = n \ Zi; .

d
Fix now a cylinder function g . Denote by g the conditional expectation of
g with respect to the -algebra generated by f(ei); (x); x 2 Zdi; g. Since bi
is measurable with respect to this -algebra, < g; bi > = < g ; bi > . More-
over, since g is a cylinder function, for n sufficiently large, < g ; bi > = <
g ; bi;(n) > . Therefore, by (4.11),
D h X i E
< g ; bi > = E bi;(n) j (0); (ei); (x) ; g
x 2 i;
n
D h X iE
= bi;(n) ; E g j (0); (ei); (x)
x2 i;
n
D h X iE
= bi ; E g j (0); (ei); ( x)
x2 i;
n
h P i
Since g is a cylinder function, as n " 1, E g j (0); (ei); x2 (x) i;
n
con-
verges in L2 ( ) to E [g j (0); (ei )]. Therefore,
D E D E
< g ; bi > = E [g j (0); (ei)] ; bi = g ; E [bi j (0); (ei)]
This concludes the proof of Step 5.
It remains to prove that b +b+ belongs to the linear space spanned by fAi ; 1
i dg. This identity is a simple consequence of the following two relations.
Step 6. Conclusion. For 1 i d, we have that
r2e ;e bi = 1f(ei) = 1g1f(2ei) > 0gbi
(4:12)
i i
r0; e bi = 1f( ei) < g1f(0) = 1gbi :

i
We leave to the reader to check that these relations implies that bi = C Ai .
We shall prove these relations for bi , all arguments apply to bi+ . Since r2ei ;ei
is continuous with respect to the weak topology and bi is the weak limit of the
negative boundary term, we have that
X
r2e ;e bii i
= lim (2n) d
n!1
r2e ;e r0;e y n :
i i i
(4:13)
y2
y= n
i
n
Since yi = n and n is Fn -measurable, we have that

r2e ;e r0;e y n
i i i
n h io
= r2e ;e 1f (0) > 0g1f (ei) < g (y n )( d0 )
i i
(y n )( )
h i
= 1f (0) > 0g (y n )( d0 ) (y n )( ) r2e ;e 1f (ei ) < g : i i
A trivial computation gives that r2e ;e 1f(ei) < g = 1f (2ei) > 0g1f (ei) =
1g. In particular,
i i
r2e ;e r0;e y
i i i n = 1f (2ei) > 0g1f (ei ) = 1gr0;ei y n :
Replacing in equation (4.13) rP 2ei ;ei r0;ei y n by the right hand side of last iden-
tity and recalling that (2n) d y2n ; yi = n r0;ei y n weakly converges to bi ,
we obtain the first relation in (4.12).
We turn now to the second identity in (4.12). We need to compute r0; ei r0;ei
y n for y in n with yi = n. The difference, with respect to the proof of the
first identity, is that y n depends on the occupation variables (0), ( ei ). For
this reason, computations are slightly more troublesome. It is easy to show that
r0; ei r0;ei y n is equal to
1f ( ei ) < g1f (0) = 1gr0;ei y n
+ 1f (ei ) < ; (0) > 1g(r0; ei y n )( d0 )
1f (ei ) < ; (0) > 1g(r0; ei y n )( ) :
P
On the one hand, since (2n) d y2n ; yi = n r0;ei y n weakly converges to bi ,
we have that
X
lim (2n) d 1f ( ei) < g1f (0) = 1gr0;e y n
n!1 y2
i
y= ni
n
= 1f ( ei ) < g1f (0) = 1gbi :

On the other hand, by definition of n , (r0; ei y n )( ) is equal to

h i
y r y; y e n () = y E r y; y e e3n j Fn
i i
= y gi;ny e ( y; y e ) :
i
i
Since the closed forms fgi;n

x g are uniformly bounded in L ( ), applying Schwarz
2
inequality, it is easy to show that

X n o
(2n) d 1f (0) > 1g (r0; e y n )( d0 )
i
(r0; ei y n )( )
y2
y= n
n
i
vanishes as n " 1. This concludes the proof of step 6 and shows that G is included
in E + fAi ; 1 i dg.
To conclude the proof of Theorem 4.14 we have to show that the intersection
of E and fAi ; 1 i dg in L2 ( ) is trivial.PAssume that there exists a cylinder
function h and a vector C in Rd such that 1id Ci Ai = r h . In particular,
Ci Ai = r0;ei h . taking inner product in L2 ( ) with respect to (ei ) (0), we
obtain that Ci = 0 because < (ei ) (0) ; r0;ei > = < r0;ei > == 0 and
< (ei ) (0) ; r0;ei h > = 0. This proves that the sum is direct and concludes
the proof of the theorem.
The estimate of the spectral gap presented in Proposition 0.2 is taken from Landim,
Sethuraman and Varadhan (1996). The second order expansion for the largest
eigenvalue of a small perturbation of a reversible generator stated in Theorem 1.1
is due to Varadhan (1994a). The proof that the spectral gap for the generator of the
generalized symmetric simple exclusion process restricted to a cube of length ` is
of order ` 2 is based on the martingale approach introduced by Lu and Yau (1993).
The proof presented here is taken from Landim, Sethuraman and Varadhan (1996).
The characterization of the closed and exact forms in the context of interacting
particle systems is due in dimension 1 to Quastel (1992). It is extended to higher
dimensions by Funaki, Uchiyama and Yau (1995).
Spectral gap of conservative dynamics. Lu and Yau (1993) proved that the
generator of the symmetric exclusion process with speed change, the so called
Kawasaki dynamics, restricted to a cube of length ` as a gap of order ` 2 in
any dimension and for any boundary condition provided some mixing conditions
are satisfied. Landim, Sethuraman and Varadhan (1996) applied the martingale
approach introduced by Lu and Yau to prove a spectral gap for zero range dynamics
for which the jump rate g () satisfies the Lipschitz condition supk1 jg (k + 1)
g(k)j < 1 and for which there exists > 0 and k0 in N such that g(k +j ) g(k)
for j k0 . Posta (1997) obtains a sharp estimate for the spectral gap of a two-
dimensional Kawasaki dynamics of unbounded discrete spins on finite cubes.
Nash inequalities, L2 decay to equilibrium. We have seen that the spectral
gap of the generator is closely related to the rate of convergence to equilibrium in
L2 . In fact, it is easy to see that for reversible systems the existence of a gap is
equivalent to an exponential rate of convergence to equilibrium in L2 . Consider a
Markov process with reversible invariant measure . Denote by the gap of the
spectrum :
= inf < ( L)f; f > ;
f < f; f >
where the infimum is carried out over all L2 ( ) functions orthogonal to the con-
stants and where < ; > stands for the inner product of L2 ( ). Then,
= t> 1 n kSt f k22 o ;

0 t f < f; f >
inf inf log
where the infimum is carried out over all L ( ) functions orthogonal to the con-
2
stants. Here fSt ; t 0g stands for the semigroup of the process and k k2 for the
L2 ( ) norm. We refer to Liggett (1989) for a proof of this statement.
As we have seen in this chapter, there is no spectral gap for conservative
dynamics in infinite volume. In this case a polynomial decay to equilibrium is
expected :

2
St f < f > 2 ta
V (f )
for some exponent a and some function V on L2 ( ). Here < f > stands for the
expected value of f with respect to . As already noticed by Deuschel (1989), the
exponent a might depend on the class of functions considered, which introduces
an additional difficulty in the analysis.
A simple computation on the symmetric simple exclusion process using du-
ality shows that for conservative, infinite volume interacting particle systems the
expected algebraic decay for cylinder functions is t d=2 .
There is a general method to prove polynomial decay to equilibrium in L2
based on Nash inequalities. This type of inequality consists in estimating the
variance of a L2 function by the Dirichlet form and a norm V defined in L2 :
Var(; f ) C D(f )1=pV (f )1=q :
for some universal constant C . In this formula D is the Dirichlet form (D(f ) =<
( L)f; f > ), 1 < p; q < 1 are conjugates p 1 + q 1 = 1 and V is a function
in L2 ( ) satisfying 0 V (f ) 1, V (rf + s) = r2 V (f ) for all r, s in R. It is
not difficult to prove (cf. Liggett (1991)) that if the semigroup is a contraction for
V , i.e., if V (St f ) V (f ) for all f in L2 ( ) and t > 0, then there is polynomial
convergence to equilibrium in L2 ( ) with exponent a = q 1 :
St f < f > 22 C Vtq(f1)

for all f in L2 , t > 0. The converse of this statement is also true because we
assumed the process to be reversible.
Nash inequalities have been proved by Bertini and Zegarlinski (1996a,b) for
symmetric simple exclusion processes and for symmetric exclusion processes with
speed change, the so called Kawasaki dynamics, with V given by the square of a
triple norm :
V (f )
X
@f 2 ;
1
x2Z @ (x)
=
d
where @f=@(x) measures the dependence of the function f on the site x :

@f=@(x) = f (x ) f () and x is the configuration with the occupation
variable at x flipped : ( x )(y ) = x;y (1 (y ))+(1 x;y ) (y ). Janvresse, Landim,
Quastel and Yau (1997) proved Nash inequality for zero range dynamics with the
triple norm V (f ) given by
V (f ) @f 2 ;
X

x2Z @ (x)
= 2
d
where @f=@ (x) is now f ( + dx ) f ( ).

The proof that the semigroup is a contraction for the triple norms defined above
has only been achieved for the symmetric simple exclusion process and seems to
be a very difficult problem for conservative particle systems. Janvresse, Landim,
Quastel and Yau (1997) presented a direct method to prove the algebraic decay
to equilibrium for zero range processes. Cancrini and Galves (1995) obtained
an upper bound for the rate of convergence to equilibrium for d-dimensional
symmetric exclusion processes starting either form a periodic configuration or
a stationary mixing distribution. They proved the existence of a finite constant
C (d; N ) such that
pt ;
d

x ; x N ] N C (d; N ) log
P [ ( ) = 1 1
t
where is either a periodic configuration or a stationary mixing distribution with
density . Galves and Guiol (1997) extended this result to the one-dimensional
nearest neighbor symmetric zero range process with jump rate given by g (k ) =
1fk 1g. Deuschel (1994) develops a general theory to prove polynomial L2
decay to equilibrium for attractive critical processes constructed by means of tran-
sient random walks on Zd.
The question of polynomial decay to equilibrium of infinite volume conserva-
tive dynamics is far to be well understood and constitutes one of the main open
problems in the field.
Non conservative dynamics. L2 decay to equilibrium has also been considered
for non conservative dynamics. Aizenman and Holley (1987) proved exponential
L2 convergence for infinite volume Glauber dynamics satisfying the Dobrushin–
Shlosman uniqueness condition (cf. Dobrushin and Shlosman (1987)). Thomas
(1989) proved that at low enough temperature the generator of reversible stochas-
tic Ising models restricted to a cube of length ` with free boundary condition is less
than C ( ) expf C0 `d 1 g for some universal constant C0 . Here is the inverse
of the temperature and C ( ) is a function of only. Liggett (1989) proved expo-
nential L2 convergence to equilibrium for supercritical reversible nearest particle
systems and proved in Liggett (1991) an algebraic L2 decay to equilibrium in the
critical case. Neuhauser (1990) proves the exponential decay to equilibrium of a
stochastic Ising model superposed with a stirring process multiplied by a small
constant.
Logarithmic Sobolev inequalities and convergence to equilibrium in L1 .
Closely related to the spectral gap inequalities are the logarithmic Sobolev in-
f with respect to a
equalities. It consists in proving that the entropy of a density p
canonical measure N;K is bounded by the Dirichlet form of f :
Z
p p
f log f dN;K C0 N 2 < f; ( LN ) f >N;K
for some universal constant C0 , all 0 K N d and all N 1. We refer to
Deuschel and Stroock (1989), Davies (1989) Davies et al. (1992) for the relation-
ship between spectral gaps, logarithmic Sobolev inequalities and the hypercon-
tractivity of the semigoup.
Logarithmic Sobolev inequalities have been introduced by Gross (1976). He
showed the connection between this type of inequality and the hypercontract-
ivity of the semigroup. By the same time Holley and Stroock (1976a,b) started
to investigate exponential L2 and L1 convergence to equilibrium of stochastic
Ising model. Holley (1985a,b) proved exponential L1 convergence of finite range,
translation invariant, attractive Glauber dynamics. Holley and Stroock (1987b)
extended this result to non attractive, one-dimensional dynamics.
In the context of interacting particle systems on a lattice, logarithmic Sobolev
inequalities were first derived by Holley and Stroock (1987a) for continuous spin
Ising models. Deuschel and Stroock (1990) investigated the relations between spec-
tral gap, logarithmic Sobolev inequalities and hypercontractivity for translation
invariant, finite range, continuous spins Glauber dynamics. Zegarlinski (1990a,b),
(1992) presented a general method to prove logarithmic Sobolev inequalities for
Glauber dynamics based on the Gibbs structure and on Dobrushin uniqueness con-
dition (cf. Dobrushin (1968), (1970)). He applied this approach to one-dimensional
Gibbs measures with bounded, finite range interactions and to continuous or finite
spin systems satisfying the Dobrushin uniqueness condition. Stroock and Zegarlin-
ski (1992a,b,c) proved the equivalence of a uniform logarithmic Sobolev inequal-
ity and the Dobrushin–Shlosman mixing conditions (cf. Dobrushin and Shlosman
(1987)) for finite range lattice gases with compact and continuous or finite spin
spaces. From this result they deduced exponential L1 convergence to equilibrium
for stochastic Ising models satisfying the Dobrushin–Shlosman mixing conditions.
Martinelli, Olivieri and Scopolla (1990) proved exponential convergence to
equilibrium of ferromagnetic stochastic Ising model in two dimensions at low
temperature and with a small external field, uniformly on the volume and on
the boundary conditions. Mountford (1995) proves exponential convergence to
equilibrium of attractive reversible subcritical nearest particle systems
Lu and Yau (1993) proved a logarithmic Sobolev inequality for Glauber dy-
namics by the martingale method, which is the only one that has been successfully
adapted to the conservative case by Yau (1997) for generalized symmetric exclu-
sion processes and Yau (1996) for Kawasaki dynamics. Martinelli and Olivieri
(1994a) analyze the connections between several mixing conditions and prove
uniform exponential convergence for finite range attractive stochastic Ising mod-
els. Martinelli and Olivieri (1994b) prove a logarithmic Sobolev inequality for
Glauber dynamics satisfying strong mixing conditions and deduce the hypercon-
tractivity of the semigroup and exponential L1 convergence to equilibrium in
infinite volume.
Logarithmic Sobolev inequalities has not yet been proved for conservative
dynamics with unrestricted number of particles per site. A version of L1 conver-
gence to equilibrium for conservative dynamics is a major problem in the field.
References
Aizenman, M., Holley R. (1987): Rapid convergence to equilibrium of stochastic Ising

models in the Dobrushin–Shlosman regime. In H. Kesten, editor, Percolation Theory and
Ergodic Theory of Infinite Particle Systems, volume 8 of IMA Volumes in Mathematics,
pages 1–11, Springer-Verlag, Berlin.
Aldous, D.J. (1978): Stopping times and tightness. Ann. Probab. 6, 335–340
Aldous, D.J. (1982): Markov chains with almost exponential hitting times. Stoch. Proc.
Appl. 13, 305–310
Aldous, D.J. (1989): Probability Approximations via the Poisson Clumping Heuristics. Vol-
ume 77 of Applied Mathematical Sciences. Springer-Verlag, New York
Aldous, D.J., Brown, M. (1992): Inequalities for rare events in time reversible Markov
chains I. In M. Shaked and Y. L. Tong, editors, Stochastic Inequalities, volume 22 of
IMS Lecture Notes, pages 1–16. IMS
Aldous, D.J., Brown, M. (1993): Inequalities for rare events in time reversible Markov
chains II. Stoch. Proc. Appl. 44, 15–25
Aldous, D.J., Diaconis, P. (1995): Hammersley’s interacting particle process and longest
increasing subsequences. Probab. Th. Rel. Fields 103, 199-213
Alexander, F.J., Cheng, Z., Janowsky, S.A., Lebowitz, J.L. (1992): Shock fluctuations in
the two–dimensional asymmetric simple exclusion process. J. Stat. Phys. 68, 761–785
Andjel, E.D. (1982): Invariant measures for the zero-range process. Ann. Probab. 10, 525–
547
Andjel, E.D. (1986): Convergence to a non extremal equilibrium measure in the exclusion
process. Probab. Th. Rel. Fields 73, 127–134
Andjel, E.D., Bramson, M., Liggett, T.M. (1988): Shocks in the asymmetric exclusion
Andjel, E.D., Kipnis, C. (1984): Derivation of the hydrodynamical equation for the zero-
range interaction process. Ann. Probab. 12, 325–334
Andjel, E.D., Vares, M.E. (1987): Hydrodynamic equations for attractive particle systems
Z
on . J. Stat. Phys. 47, 265–288
Arratia, R. (1983): The motion of a tagged particle in the simple symmetric exclusion
Z
system on . Ann. Probab. 11, 362–373
Asselah, A., Brito, R., Lebowitz, J.L. (1997): Self diffusion in simple models : systems
with long range jumps. J. Stat. Phys. 87 1131–1144
Asselah, A., Dai Pra, P. (1997): Sharp estimates for the occurrence of rare events for
symmetric simple exclusion. Stoch. Proc. Appl. 71, 259–273
Bahadoran, C. (1996a): Hydrodynamical limit for non-homogeneous zero range processes,
preprint
Bahadoran, C. (1996b): Hydrodynamical limit for spatially heterogeneous simple exclusion
processes, preprint
Bahadoran, C. (1997): Hydrodynamics of asymmetric misanthrope processes with general
initial configurations, preprint
426 References
Bellman, R., Harris, T.E. (1951): Recurrence times for the Ehrenfest model. Pacific J. Math.
1 179–193
Benassi, A., Fouque, J.P. (1987): Hydrodynamical limit for the asymmetric exclusion pro-
cess. Ann. Probab. 15, 546–560
Benassi, A., Fouque, J.P. (1988): Hydrodynamical limit for the asymmetric zero-range
process. Ann. Inst. H. Poincaré, Probabilités 24, 189–200
Benassi, A., Fouque, J.P. (1991): Fluctuation field for the asymmetric simple exclusion
process. In U. Hornung, P. Kotelenez and G. Papanicolaou, editors, Random Partial
Differential Equations, volume 102 of International Series of Numerical Mathematics,
pages 33–43, Birhäuser, Boston
Benassi, A., Fouque, J.P., Saada, E., Vares, M.E. (1991): Asymmetric attractive particle
Z
systems on : hydrodynamical limit for monotone initial profiles. J. Stat. Phys. 63,
719–735
Benjamini, I., Ferrari, P.A., Landim, C. (1996): Asymmetric processes with random rate.
Stoch. Proc. Appl. 61, 181–204
Benois, O. (1996): Large deviations for the occupation times of independent particle sys-
tems. Ann. Appl. Probab. 6 269–296
Benois, O., Kipnis, C., Landim, C. (1995): Large deviations for mean zero asymmetric
zero-range processes in infinite volume. Stoch. Proc. Appl. 55, 65–89
Benois, O., Koukkous, A., Landim, C. (1997): Diffusive behavior of asymmetric zero range
processes. J. Stat. Phys. 87, 577–591
Bertini, L., Landim, C., Olla, S. (1997): Derivation of Cahn–Hilliard equations from
Ginzburg–Landau models. J. Stat. Phys. 88, 365–381
Bertini, L., Presutti, E., Rüdiger, B., Saada, E. (1994): Dynamical fluctuations at the critical
point : convergence to a non linear stochastic PDE. Theory Probab. Appl. 38, 586–629
Bertini, L., Zegarlinski, B. (1996a): Coercive inequalities for Kawasaki dynamics : the
product case. University of Texas Mathematical Physics archive preprint 96-561
Bertini, L., Zegarlinski, B. (1996b): Coercive inequalities for Gibss measures. University
of Texas Mathematical Physics archive preprint 96-562.
Billingsley, P. (1968): Convergence of Probability Measures, John Wiley & Sons, New
York
Boldrighini, C., De Masi, A., Pellegrinotti, A. (1991): Nonequilibrium fluctuations in par-
ticle system modeling reaction–diffusion equations. Stoch. Proc. Appl. 42, 1–30
Boldrighini, C., De Masi, A., Pellegrinotti, A., Presutti, E. (1987): Collective phenomena
in interacting particle systems. Stoch. Proc. Appl. 25, 137–152
Boldrighini, C., Dobrushin, R.L., Suhov, Yu.M. (1983): One–dimensional hard rod carica-
ture of hydrodynamics. J. Stat. Phys. 31, 577–616
Bonaventura, L. (1995): Interface dynamics in an interacting spin system. Nonlinear Anal.
25, 799–819
Bramson, M. (1988): Front propagation in certain one dimensional exclusion models. J.
Stat. Phys. 51, 863–870
Bramson, M., Calderoni, P., De Masi, A., Ferrari, P.A., Lebowitz, J.L., Schonmann, R.H.
(1989): Microscopic selection principle for a diffusion–reaction equation. J. Stat. Phys.
45, 905–920
Bramson, M., Cox, J.T., Griffeath, D. (1988): Occupation time large deviations of the voter
model. Probab. Th. Rel. Fields 73, 613-625
Bramson, M., Lebowitz, J.L. (1991): Spatial structure in diffusion–limited two–particle
reactions. J. Stat. Phys. 65, 941–951
Breiman, L. (1968): Probability. Addison–Wesley, New York
Brezis, H., Crandall, M.G. (1979): Uniqueness of solutions of the initial value problem for
u u
t ( ) = 0. J. Math. Pure Appl. 58, 153–163
Brox, T., Rost, H. (1984): Equilibrium fluctuations of stochastic particle systems: the role
of conserved quantities. Ann. Probab. 12, 742–759
References 427
Buttà, P. (1993): On the validity of an Einstein relation in models of interface dynamics.

J. Stat. Phys. 72 1401–1406
Buttà, P. (1994): Motion by mean curvature by scaling a nonlocal equation : convergence
at all times in the two–dimensional case. Lett. Math. Phys. 31, 41–55
Calderoni, P., Pellegrinotti, A., Presutti, E., Vares, M.E. (1989): Transient bimodality in
interacting particle systems. J. Stat. Phys. 55, 523–577
Cancrini, N., Galves, A. (1995): Approach to equilibrium in the symmetric simple exclusion
process. Markov Proc. Rel. Fields 2, 175–184
Caprino, S., De Masi, A., Presutti, E., Pulvirenti, M. (1989): A stochastic particle system
modeling the Carleman equation. J. Stat. Phys. 55, 625–638
Caprino, S., De Masi, A., Presutti, E., Pulvirenti, M. (1990): A stochastic particle system
modeling the Carleman equation : Addendum. J. Stat. Phys. 59, 535–537
Caprino, S., De Masi, A., Presutti, E., Pulvirenti, M. (1991): A derivation of the Broadwell
equation. Commun. Math. Phys. 135, 443–465
Caprino, S., Pulvirenti, M. (1995): A cluster expansion approach to a one–dimensional
Boltzmann equation : a validity result. Commun. Math. Phys. 166, 603–631
Caprino, S., Pulvirenti, M. (1991): The Boltzmann–Grad limit for a one–dimensional Boltz-
mann equation in a stationary state. Commun. Math. Phys. 177, 63–81
Carlson, J.M., Grannan, E.R., Swindle, G.H. (1993): A limit theorem for tagged particles
in a class of self–organizing particle systems. Stoch. Proc. Appl. 47, 1–16
Carlson, J.M., Grannan, E.R., Swindle, G.H., Tour, J. (1993) Singular diffusion limits of a
class of reversible self–organizing particle systems. Ann. Probab. 21, 1372–1393
Carmona, R.A., Xu, L. (1997): Diffusive hydrodynamic limit for systems of interacting
diffusions with finite range random interaction. Commun. Math. Phys. 188, 565–584
Cassandro, M., Galves, A., Olivieri, E. Vares, M. E. (1984): Metastable behavior of stochas-
tic dynamics: a pathwise approach. J. Stat. Phys. 35, 603–628
Chang, C.C. (1994): Equilibrium fluctuations of gradient reversible particle systems. Probab.
Th. Rel. Fields 100, 269–283
Chang, C.C. (1995): Equilibrium fluctuations of nongradient reversible particle systems. In
T. Funaki and W. Woyczinky, editor, Proc. On Stochastic Method for Nonlinear P.D.E.,
volume 77 of IMA vol. in Mathematics, pages 41–51, Springer-Verlag, New York
Chang, C.C., Yau, H.T. (1992): Fluctuations of one dimensional Ginzburg–Landau models
in nonequilibrium. Commun. Math. Phys. 145, 209–234
Chayes, L., Swindle, G. (1996): Hydrodynamic limits for one dimensional particle systems
with moving boundaries. Ann. Probab. 24, 559–598
Chow, Y.S., Teicher, H. (1988): Probability Theory, Springer-Verlag, New York
Chung, K.L. (1974): A Course in Probability Theory, 2nd edition, Academic Press, New
York.
Cocozza, C. (1985): Processus des misanthropes. Z. Wahrsch. Verw. Gebiete 70, 509–523
Cogburn, R. (1985): On the distribution of first passage and return times for small sets.
Ann. Probab. 13, 1219–1223
Comets, F. (1987): Nucleation for a long-range magnetic model. Ann. Inst. H. Poincaré,
Probabilités 23, 135–178
Comets, F., Eisele, Th. (1988): Asymptotic dynamics, noncritical and critical fluctuations
for a geometric long-range interacting model. Commun. Math. Phys. 118, 531–567
Covert, P., Rezakhanlou, F. (1996): Hydrodynamic limit for particle systems with non–
constant speed parameter. J. Stat. Phys. 88, 383–426
Cox, J.T., Griffeath, D. (1984): Large deviation for Poisson systems of independent random
walks. Z. Wahrsch. Verw. Gebiete 66, 543–558
Davies, E.B. (1989): Heat kernels and spectral theory, Cambridge University Press, Cam-
bridge
428 References
Davies, E.B., Gross, L., Simon, B. (1992): Hypercontractivity : a bibliographical review.

In S. Albeverio, J.E. Fenstand, H. Holden and T. Lindstrom, editors, In memoriam
of Raphael Hoegh-Krohn. Ideas and Methods of Mathematic and Physics, Cambridge
University Press, Cambridge
De Masi, A., Ferrari, P. A. (1984): A remark on the hydrodynamics of the zero range
processes. J. Stat. Phys. 36, 81–87
De Masi, A., Ferrari, P. A. (1985): Self-diffusion in one-dimensional lattice gases in the
presence of an external field, J. Stat. Phys. 38, 603–613
De Masi, A., Ferrari, P. A., Goldstein, S., Wick, W. D. (1985): Invariance principle for
reversible Markov processes with application to diffusion in the percolation regime.
Contemp. Math. 41, 71–85
De Masi, A., Ferrari, P. A., Goldstein, S., Wick, W. D. (1989): An invariance principle for
reversible Markov processes. Applications to random motions in random environments.
J. Stat. Phys. 55, 787–855
De Masi, A., Ferrari, P. A., Ianiro, N., Presutti, E. (1982): Small deviations from local
equilibrium for a process which exhibits hydrodynamical behavior II. J. Stat. Phys. 29,
81–93
De Masi, A., Ferrari, P. A., Lebowitz, J.L. (1986): Reaction-diffusion equations for inter-
acting particle systems, J. Stat. Phys. 44, 589–644
De Masi, A., Ferrari, P. A., Vares, M.E. (1989): A microscopic model of interface related
to the Burger’s equation. J. Stat. Phys. 55, 601–609
De Masi, A., Ianiro, N., Pellegrinotti, A., Presutti, E. (1984): A survey of the hydrodynam-
ical behavior of many-particle systems. In J. L. Lebowitz and E. W. Montroll, editors,
Nonequilibrium phenomena II : From stochastics to hydrodynamics Volume 11 of Studies
in Statistical Mechanics, pages 123–294, North-Holland, Amsterdam
De Masi, A., Ianiro, N., Presutti, E. (1982): Small deviations from local equilibrium for a
process which exhibits hydrodynamical behavior. I. J. Stat. Phys. 29, 57–79
De Masi, A., Kipnis, C., Presutti, E., Saada, E. (1989): Microscopic structure at the shock
in the asymmetric simple exclusion. Stochastics 27, 151–165
De Masi, A., Orlandi, E., Presutti, E., Triolo, L. (1994): Glauber evolution with Kac poten-
tials : I. Mesoscopic and macroscopic limits, interface dynamics. Nonlinearity 7, 633–696
De Masi, A., Orlandi, E., Presutti, E., Triolo, L. (1996a): Glauber evolution with Kac
potentials : II. Fluctuations. Nonlinearity 9, 27–51
De Masi, A., Orlandi, E., Presutti, E., Triolo, L. (1996b): Glauber evolution with Kac
potentials : III. Spinodal decompositions. Nonlinearity 9, 53–114
De Masi, A., Pellegrinotti, A., Presutti, E., Vares, E. (1994): Spatial patterns when phases
separate in an interacting particle system. Ann. Probab. 22, 334–371
De Masi, A., Presutti, E. (1991): Mathematical methods for hydrodynamic limits, volume
1501 of Lecture Notes in Mathematics, Springer-Verlag, New York
De Masi, A., Presutti, E., Scacciatelli, E. (1989): The weakly asymmetric simple exclusion
process. Ann. Inst. H. Poincaré, Probabilités 25, 1–38
De Masi, A., Presutti, E., Spohn, H., Wick, D. (1986): Asymptotic equivalence of fluctuation
fields for reversible exclusion processes with speed change. Ann. Probab. 14, 409–423
De Masi, A., Presutti, E., Vares, M.E. (1986): Escape from the unstable equilibrium in a
random process with infinitely many interacting particles. J. Stat. Phys. 44, 645–696
DelGrosso, G. (1974): On the local central limit theorem for Gibbs processes. Commun.
Math. Phys. 37, 141–160
Derrida, B., Domany, E., Mukamel, D. (1992): An exact solution of a one–dimensional
asymmetric exclusion model with open boundaries. J. Stat. Phys. 69, 667–687
Derrida, B., Evans, M.T., Mallick, K. (1995): Exact diffusion constant for a one–dimensional
asymmetric exclusion model with open boundaries. J. Stat. Phys. 79, 833–874
Derrida, B., Janowsky, S.A., Lebowitz, J.L., Speer, E.R. (1993): Exact solutions of the
totally asymmetric simple exclusion process : shock profiles. J. Stat. Phys. 73, 813–842
References 429
Deuschel, J.D. (1989): Invariance principle and empirical mean large deviations of the
critical Ornstein–Uhlenbeck process. Ann. Probab. 17, 74–90
L
Deuschel, J.D. (1994): Algebraic 2 decay of attractive critical processes on the lattice.
Ann. Probab. 22, 264–283
Deuschel, J.D., Stroock, D.W. (1989): Large Deviations. Academic Press, New York
Deuschel, J.D., Stroock, D.W. (1990): Hypercontractivity and spectral gap of symmetric
diffusions with applications to the stochastic Ising model. J. Funct. Anal. 92, 30–48
DiPerna, R.J. (1985): Measure-valued solutions to conservation laws. Arch. Rational Mech.
Anal. 88, 223–270
Dittrich, P. (1987): Limit theorems for branching diffusions in hydrodynamical rescaling.
Math. Nachr. 131, 59–72
Dittrich, P. (1988a): A stochastic model of a chemical reaction with diffusion. Probab. Th.
Rel. Fields 79, 115–128
Dittrich, P. (1988b): A stochastic particle system : fluctuations around a non-linear reaction-
diffusion equation. Stoch. Proc. Appl. 30, 149–164
Dittrich, P. (1990): Travelling waves and long-time behaviour of the weakly asymmetric
exclusion process. Probab. Th. Rel. Fields 86, 443–455
Dittrich, P. (1992): Long-time behaviour of the weakly asymmetric exclusion process and
the Burgers equation without viscosity. Math. Nachr. 155, 279–287
Dittrich, P., Gartner, J. (1991): A central limit theorem for the weakly asymmetric simple
exclusion process. Math. Nachr. 151, 75–93
Dobrushin, R.L. (1968): The problem of uniqueness of a Gibbs random field and problem
of phase transition. Funct. Anal. Appl. 2, 302–312
Dobrushin, R.L. (1970): Prescribing a system of random variables by conditional distribu-
tions. Theory Probab. Appl. 15, 453–486
Dobrushin, R.L. (1989): Caricatures of hydrodynamics. In B. Simon and A. Truman and
I. M. Davies, editors, IXth International Congress on Mathematical Physics, pages 117–
132, Adam Hilger, Bristol
Dobrushin, R.L., Pellegrinotti, A., Suhov, Yu.M. (1990): One dimensional harmonic lattice
caricature of hydrodynamics : a higher correction. J. Stat. Phys. 61, 387–402
Dobrushin, R.L., Pellegrinotti, A., Suhov, Yu.M., Triolo, L. (1986): One dimensional har-
monic lattice caricature of hydrodynamics. J. Stat. Phys. 43, 571–607
Dobrushin, R.L., Pellegrinotti, A., Suhov, Yu.M., Triolo, L. (1988): One dimensional har-
monic lattice caricature of hydrodynamics : second approximation. J. Stat. Phys. 52,
423–439
Dobrushin, R.L., Shlosman, S.B. (1987): Completely analytical interactions constructive
descriptions. J. Stat. Phys. 56, 983–1014
Dobrushin, R.L., Siegmund-Schultze, R. (1982): The hydrodynamic limit for systems of
particles with independent evolution. Math. Nachr. 105, 199–224
Dobrushin, R.L., Tirozzi, B. (1977): The central limit theorem and the problem of equiva-
lence of ensembles. Commun. Math. Phys. 54, 173–192
Donsker, M.D., Varadhan, S.R.S. (1975a): Asymptotic evaluation of certain Markov process
expectations for large time I. Comm. Pure Appl. Math. XXVIII, 1–47
Donsker, M.D., Varadhan, S.R.S. (1975b): Asymptotic evaluation of certain Markov process
expectations for large time II. Comm. Pure Appl. Math. XXVIII, 279–301
Donsker, M.D., Varadhan, S.R.S. (1975c): Asymptotic evaluation of certain Wiener integrals
for large time. In A. M. Arthurs, editor, Functional Integration and Its Applications,
Proceedings of the International Conference held at the Cumberland Lodge, Windsor
Great Park, London 1974, pages 15–33, Clarendon Press, Oxford
Donsker, M.D., Varadhan, S.R.S. (1976): Asymptotic evaluation of certain Markov process
expectations for large time III. Comm. Pure Appl. Math. XXIX, 389–461
Donsker, M.D., Varadhan, S.R.S. (1989): Large deviations from a hydrodynamic scaling
limit. Comm. Pure Appl. Math. XLII, 243–270
430 References
Doob, J.L. (1953): Stochastic Processes, John Wiley & Sons, New York
Durrett, R. (1991): Probability : Theory and Examples, Wadsworth, Belmont
Durrett, R., Neuhauser, C. (1994): Particle systems and reaction–diffusion equations. Ann.
Probab. 22, 289–333
Ekhaus, M., Seppäläinen, T. (1996): Stochastic dynamics macroscopically governed by the
porous medium equation for isothermal flow. Ann. Acad. Sci. Fenn. Math. 21, 309–352
Esposito, R., Marra, R. (1994): On the derivation of the incompressible Navier–Stokes
equation for Hamiltonian particle systems. J. Stat. Phys. 74, 981–1004
Esposito, R., Marra, R., Yau, H.T. (1994): Diffusive limit of asymmetric simple exclusion.
Rev. Math. Phys. 6, 1233–1267
Esposito, R., Marra, R., Yau, H.T. (1996): Navier–Stokes equations for stochastic lattice
gases. Comm. Math. Phys. 182, 395–456
Ethier, S.N., Kurtz, T. G. (1986): Markov Processes, Characterization and Convergence.
John Wiley & Sons, New York
Evans, L.C., Rezakhanlou, F. (1997): A stochastic model for growing sandpiles and its
continuum limit, preprint
Evans, L.C., Soner, H.M., Souganidis, P.E. (1992): Phase transition and generalized motion
by mean curvature. Comm. Pure Appl. Math. XLV, 1097–1123
Evans, L.C., Spruck, J. (1991): Motion of level sets by mean curvature. I. J. Diff. Geom.
33, 635–681
Evans, M.R., Foster, D.P., Godrèche, C., Mukamel, D. (1995): Asymmetric exclusion model
with two species : spontaneous symmetry breaking. J. Stat. Phys. 80, 69–102
Eyink, G.L. (1990): Dissipation and large thermodynamic fluctuations. J. Stat. Phys. 61,
533–572
Eyink, G.L., Lebowitz, J.L., Spohn, H. (1990): Hydrodynamics of stationary non-equilibrium
states for some stochastic lattice gas models. Commun. Math. Phys. 132, 253–283
Eyink, G.L., Lebowitz, J.L., Spohn, H. (1991): Lattice gas models in contact with stochastic
reservoirs: local equilibrium and relaxation to the steady state. Commun. Math. Phys.
140, 119–131
Eyink, G.L., Lebowitz, J.L., Spohn, H. (1996): Hydrodynamics and fluctuations outside of
local equilibrium : driven diffusive systems. J. Stat. Phys. 83, 385–472
Farmer, J., Goldstein, S., Speer, E.R. (1984): Invariant states of a thermally conducting
barrier. J. Stat. Phys. 34, 263–277
Feller, W. (1966): An Introduction to Probability Theory and its Applications, John Wiley
& Sons, New York
Feng, S., Iscoe, I., Seppäläinen, T. (1997): A microscopic mechanism for the porous medium
equation. Stoch. Proc. Appl. 66, 147–182
Ferrari, P.A. (1986): The simple exclusion process as seen from a tagged particle. Ann.
Probab. 14, 1277–1290
Ferrari, P.A. (1992): Shock fluctuations in asymmetric simple exclusion. Probab. Th. Rel.
Fields 91, 81–101
Ferrari, P.A. (1996): Limit theorems for tagged particles. Markov Proc. Rel. Fields 2, 17–40
Ferrari, P.A., Fontes, L.R.G. (1994a): Current fluctuations for the asymmetric simple ex-
clusion process. Ann. Probab. 22, 820–832
Ferrari, P.A., Fontes, L.R.G. (1994b): Shock fluctuations in the asymmetric simple exclusion
Ferrari, P.A., Fontes, L.R.G. (1996): Poissonian approximation for the tagged particle in
asymmetric simple exclusion. J. Appl. Prob. 33, 411–419
Ferrari, P.A., Fontes, L.R.G., Kohayakawa, Y. (1994): Invariant measures for a two species
asymmetric process. J. Stat. Phys. 76, 1153–1177
Ferrari, P.A., Galves, A., Landim, C. (1994): Exponential waiting times for a big gap in a
one-dimensional zero range process. Ann. Probab. 22, 284–288
References 431
Ferrari, P.A., Galves, A., Liggett, T.M. (1995): Exponential waiting time for filling a large
interval in the symmetric simple exclusion process. Ann. Inst. H. Poincaré, Probabilités
31, 155–175
Ferrari, P.A., Goldstein, S. (1988): Microscopic stationary states for stochastic systems with
particle flux. Probab. Th. Rel. Fields 78, 455–471
Ferrari, P.A., Goldstein, S., Lebowitz, J.L. (1985): Diffusion, mobility and the Einstein
relation. In, J. Fritz, A. Jaffe, D. Száz, editors, Statistical Physics and Dynamical Systems,
pages 405–441, Birkhäuser, Boston
Ferrari, P.A., Kipnis, C. (1995): Second class particles in the rarefaction fan. Ann. Inst. H.
Poincaré, Probabilités 31, 143–154
Ferrari, P.A., Kipnis, C., Saada, E. (1991): Microscopic structure of traveling waves in the
asymmetric simple exclusion process. Ann. Probab. 19, 226–244
Ferrari, P.A., Presutti, E., Vares, M.E. (1987): Local equilibrium for a one dimensional zero
range proces. Stoch. Proc. Appl. 26, 31–45
Ferrari, P.A., Presutti, E., Vares, M.E. (1988): Nonequilibrium fluctuations for a zero range
process, Ann. Inst. H. Poincaré, Probabilités 24, 237–268
Filippov, A.F. (1960): Differential equations with discontinuous right–hand side. Mat. Sb.
(N.S.) 51 (93), 99–128
Foster, D.P., Godrèche, C. (1994): Finite–size effects for phase segregation in a two–
dimensional asymmetric exclusion model with two species. J. Stat. Phys. 76, 1129–1151
Fouque, J.P., Saada, E. (1994): Totally asymmetric attractive particle systems on Z : hy-
drodynamical limit for general initial profiles. Stoch. Proc. Appl. 51, 9–23
Fritz, J. (1982): Local stability and hydrodynamical limit of Spitzer’s one–dimensional
lattice model. Commun. Math. Phys. 86, 363–373
Fritz, J. (1985): On the asymptotic behavior of Spitzer’s model for evolution of one–
dimensional point systems. J. Stat. Phys. 38, 615–645
Fritz, J. (1987a): On the hydrodynamic limit of a one dimensional Ginzburg–Landau lattice
model : the a priori bounds. J. Stat. Phys. 47, 551–572
Fritz, J. (1987b): On the hydrodynamic limit of a scalar Ginzburg–Landau Lattice model :
the resolvent approach. In, G. Papanicolaou, editor, Hydrodynamic behaviour and in-
teracting particle systems, volume 9 of IMA volumes in in Mathematics, pages 75–97.
Springer-Verlag, New York
Fritz, J. (1989a): On the hydrodynamic limit of a Ginzburg–Landau model. The law of
large numbers in arbitrary dimensions. Probab. Th. Rel. Fields 81, 291–318
Fritz, J. (1989b): Hydrodynamics in a symmetric random medium. Commun. Math. Phys.
125, 13–25
Fritz, J. (1990): On the diffusive nature of entropy flow in infinite systems : remarks to a
paper by Guo, Papanicolaou and Varadhan. Commun. Math. Phys. 133, 331–352
Fritz, J., Funaki, T., Lebowitz, J.L. (1994): Stationary states of random Hamiltonian systems.
Probab. Th. Rel. Fields 99, 211–236
Fritz, J., Liverani, C., Olla, S. (1997): Reversibility in infinite Hamiltonian systems with
conservative noise. Commun. Math. Phys. 189, 481–496
Fritz, J., Maes, C. (1988): Derivation of a hydrodynamic equation for Ginzburg–Landau
models in an external field. J. Stat. Phys. 53, 1179–1206
Fritz, J., Rüdiger, B. (1995): Time dependent critical fluctuations of a one dimensional local
mean field model. Probab. Th. Rel. Fields 103, 381–407
Funaki, T. (1989a): The hydrodynamical limit for a scalar Ginzburg–Landau model on . In R
M. Métivier and S. Watanabe, editors, Stochastic Analysis, Lecture Notes in Mathematics
1322, pages 28–36, Springer-Verlag, Berlin
Funaki, T. (1989b): Derivation of the hydrodynamical equation for a one–dimensional
Ginzburg–Landau model. Probab. Th. Rel. Fields 82, 39–93
Funaki, T., Handa, K, Uchiyama, K. (1991): Hydrodynamic limit of one dimensional ex-
clusion processes with speed change. Ann. Probab. 19, 245–265
432 References
Funaki, T., Spohn, H. (1997): Motion by mean curvature from the Ginzburg–Landau r
interface model. Commun. Math. Phys. 185, 1–36
Funaki, T., Uchiyama, K., Yau, H.T. (1995): Hydrodynamic limit for lattice gas reversible
under Bernoulli measures. In T. Funaki and W. Woyczinky, editors, Proceedings On
Stochastic Method for Nonlinear P.D.E.. IMA volumes in Mathematics 77, pages 1–40,
Springer-Verlag, New York
Gabrielli, D., Jona–Lasinio, G., Landim, C. (1996): Onsager reciprocity relations without
microscopic reversibility. Phys. Rev. Lett. 77, 1202–1205
Gabrielli, D., Jona–Lasinio, G., Landim, C. (1998): Onsager symmetry from microscopic
TP invariance, preprint
Gabrielli, D., Jona–Lasinio, G., Landim, C., Vares, M.E. (1997): Microscopic reversibility
and thermodynamic fluctuations. In C. Cercignani, G. Jona–Lasinio, G. Parisi and L. A.
Radicati di Brozolo, editors, Boltzmann’s Legacy 150 Years After His Birth. volume 131
of Atti dei Convegni Licei, pages 79–88, Accademia Nazionale dei Lincei, Roma
Galves, A., Guiol, H. (1997): Relaxation time of the one dimensional zero range process
with constant rate. Markov Proc. Rel. Fields 3, 323–332
Galves, A., Kipnis, C., Marchioro, C., Presutti, E. (1981): Nonequilibrium measures which
exhibit a temperature gradient: study of a model. Commun. Math. Phys. 81, 121–147
Galves, A., Martinelli, F., Olivieri, E. (1989): Large density fluctuations for the one dimen-
sional supercritical contact process. J. Stat. Phys. 55, 639–648
Gärtner, J. (1988): Convergence towards Burger’s equation and propagation of chaos for
weakly asymmetric exclusion processes. Stoch. Proc. Appl. 27, 233–260
Gärtner, J., Presutti, E. (1990): Shock fluctuations in a particle system. Ann. Inst. H.
Poincaré, Physique Théorique 53, 1–14
Giacomin, G. (1991): Van der Waals limit and phase separation in a particle model with
Kawasaki dynamics. J. Stat. Phys. 65, 217–234
Giacomin, G. (1994): Phase separation and random domain patterns in a stochastic particle
model. Stoch. Proc. Appl. 51, 25–62
Giacomin, G. (1995): Onset and structure of interfaces in a Kawasaki + Glauber interacting
particle system. Probab. Th. Rel. Fields 103, 1–24
Giacomin, G., Lebowitz, J.L. (1997): Phase segregation dynamics in particle systems with
long range interactions I. Macroscopic limits. J. Stat. Phys. 87, 37–61
Giacomin, G., Lebowitz, J.L. (1998): Phase segregation dynamics in particle systems with
long range interactions II. Interface motion. preprint.
Giacomin, G., Olla, S., Spohn, H. (1998): Equilibrium fluctutation for r interface model,
preprint
Gielis, G., Koukkous, A., Landim, C. (1997): Equilibrium fluctuations for zero range pro-
cesses in random environment. To appear in Stoch. Proc. Appl.
Gikhman, I.I., Skorohod, A.V. (1969): Introduction to the Theory of Random Processes. W.
B. Saunders Company, Philadelphia
Goldstein, S. (1995): Antisymmetric functionals of reversible Markov processes. Ann. Inst.
H. Poincaré, Probabilités 31, 177–190
Goldstein, S., Kipnis, C., Ianiro, N. (1985): Stationary states for a mechanical system with
stochastic boundary conditions, J. Stat. Phys. 41, 915–939
Goldstein, S., Lebowitz, J.L., Presutti, E. (1981): Mechanical systems with stochastic bound-
ary conditions. In J. Fritz, J.L. Lebowitz and D. Szász, editors, Random Fields, volume
27 of Colloquia Mathematica Societatis Janos Bolyai, pages 421–461, North-Holland,
Amsterdam
Goldstein, S., Lebowitz, J.L., Ravishankar, K. (1982): Ergodic properties of a system in
contact with a heat bath : a one-dimensional model. Commun. Math. Phys. 85, 419–427
Gravner, J., Quastel, J. (1998): Internal DLA and the Stefan problem. Preprint
Greven, A. (1984): The hydrodynamical behavior of the coupled branching process, Ann.
Probab. 12, 760–767
References 433
Grigorescu, I. (1997a): Self–diffusion for Brownian motion with local interaction. Preprint.
Grigorescu, I. (1997b): Uniqueness of the tagged particle process in a system with local
interactions. Preprint.
Gross, L. (1976): Logarithmic Sobolev inequalities. Am. J. Math. 97, 1061–1083
Guo, M.Z., Papanicolaou, G.C., Varadhan, S.R.S. (1988): Nonlinear diffusion limit for a
system with nearest neighbor interactions. Commun. Math. Phys. 118, 31–59
Hammersley, J.M. (1972): A few seedlings of research. Proceedings of the 6th Berkeley
symposium in Mathematics, Statistics and Probability, pages 345–394, University of Cal-
ifornia Press.
Harris, T.E. (1952): First passage and recurrence distributions. Trans. Amer. Math. Soc. 73,
471–486
Harris, T.E. (1965): Diffusions with collisions between particles. J. Appl. Prob. 2, 323–338
Harris, T.E. (1967): Random measures and motions of point processes. Z. Wahrsch. Verw.
Gebiete 9, 36–58
Helland, I. (1982): Central limit theorems for martingales with discrete or continuous time.
Scand. J. Stat. 9, 79–94
Holley, R. (1985a): Rapid convergence to equilibrium in one-dimensional stochastic Ising
models. Ann. Probab. 13, 72–89
Holley, R. (1985b): Possible rates of convergence in finite range, attractive spin systems.
In Particle Systems, Random Media and Large Deviations, Contemporary Mathematics
41, pages 215–234
L
Holley, R., Stroock, D.W. (1976a): 2 -Theory for the stochastic Ising model. Z. Wahrsch.
Verw. Gebiete 35, 87–101
Holley, R., Stroock, D.W. (1976b): Applications of the stochastic Ising model to the Gibbs
states. Commun. Math. Phys. 48, 249–265
Holley, R., Stroock, D.W. (1978): Generalized Ornstein-Uhlenbeck processes and infinite
particle branching Brownian motions. RIMSA Kyoto Publ. 14, 741–788
Holley, R., Stroock, D.W. (1979a): Central limit phenomena of various interacting systems.
Ann. Math. 110, 333–393
Holley, R., Stroock, D.W. (1979b): Rescaling short–range interacting stochastic processes
in higher dimensions. In J. Fritz, J.L. Lebowitz and D. Szász, editors, Random Fields,
volume 27 of Colloquia Mathematica Societatis Janos Bolyai, pages 535–550, North-
Holland, Amsterdam
Holley, R., Stroock, D.W. (1987a): Logarithmic Sobolev inequalities and stochastic Ising
models. J. Stat. Phys. 46, 1159–1194
L
Holley, R., Stroock, D.W. (1987b): Uniform and 2 convergence in one dimensional
stochastic Ising models. Commun. Math. Phys. 123, 85–93
u uu
Hopf, E. (1950): The partial differential equation t + x = u xx . Comm. Pure Appl.
Math. III, 201–230
Janowski, S.A., Lebowitz, J.L. (1994): Exact results for the asymmetric simple exclusion
proces with a blockage. J. Stat. Phys. 77, 35–51
Janvresse, E. (1997): First order correction for the hydrodynamic limit of symmetric simple
exclusion processes with speed change in dimension d 3, preprint
Janvresse, E., Landim, C., Quastel, J., Yau, H.T. (1997): Relaxation to equilibrium of
conservative dynamics I : zero range dynamics. Preprint
Jensen, L., Yau, H.T. (1997): Hydrodynamical Scaling Limits of Simple Exclusion Models.
To appear in Park City-IAS Summer School Lecture Series.
Jockusch, W., Propp, J., Shor, P. (1995): Random domino tillings and the artic circle
theorem, preprint
Jona–Lasinio, G. (1991): Stochastic reaction diffusion equations and interacting particle
systems. Ann. Inst. H. Poincaré, Physique Théorique 55, 751–758
434 References
Jona–Lasinio, G. (1992): Structure of hydrodynamic fluctutations in interacting particle

systems. In F. Guerra, M. I. Loffredo and C. Marchioro, editors, Probabilistic Methods
in Mathematical Physics, pages 262–263, World Scientific, Singapore
Jona–Lasinio, G., Landim, C., Vares, M.E. (1993): Large deviations for a reaction-diffusion
model. Probab. Th. Rel. Fields 97, 339–361
Katsoulakis, M.A., Souganidis, P.E. (1994): Interacting particle systems and generalized
evolution fronts. Arch. Rational Mech. Anal. 127, 133–159
Katsoulakis, M.A., Souganidis, P.E. (1995): Generalized motion by mean curvature as a
macroscopic limit of stochastic Ising models with long range interaction and Glauber
dynamics. Commun. Math. Phys. 169, 61–97
Kemeny, J.G., Snell, J.L., Knapp, A.W. (1966): Denumerable Markov Chains, Van Nostrand
Company, Princeton
Kipnis, C. (1985): Recent results on the movement of a tagged particle in simple exclusion.
Kipnis, C. (1986): Central limit theorems for infinite series of queues and applications to
simple exclusion. Ann. Probab. 14, 397–408
Kipnis, C., Landim, C., Olla, S. (1994): Hydrodynamical limit for a nongradient system:
the generalized symmetric simple exclusion process. Comm. Pure Appl. Math. XLVII,
1475–1545
Kipnis, C., Landim, C., Olla, S. (1995): Macroscopic properties of a stationary non-
equilibrium distribution for a non-gradient interacting particle system. Ann. Inst. H.
Poincaré, Probabilités 31, 191–221
Kipnis, C., Léonard, C. (1995): Grandes Déviations pour un système hydrodynamique
asymétrique de particules indépendantes. Ann. Inst. H. Poincaré, Probabilités 31, 223–
248
Kipnis, C., Marchioro, C., Presutti, E. (1982): Heat flow in an exactly solvable model J.
Stat. Phys. 27, 65–74
Kipnis, C., Olla, S., Varadhan, S.R.S. (1989): Hydrodynamics and large deviations for
simple exclusion processes, Comm. Pure Appl. Math. XLII, 115–137
Kipnis, C., Varadhan, S.R.S. (1986): Central limit theorem for additive functionals of re-
versible Markov processes and applications to simple exclusion, Commun. Math. Phys.
106, 1–19
Komoriya, K. (1997): Hydrodynamic limit for asymetric mean zero exclusion processes
with speed change. preprint
Koukkous, A. (1997): Comportement hydrodynamique de differents processus de zero
range. Thèse de doctorat de l’Université de Rouen.
Korolyuk, D.V., Sil’vestrov, D.S. (1984): Entry times into asymptotically receding domains
for ergodic Markov chains. Theory Probab. Appl. 28, 432–442
Kružkov, S.N. (1970): First order quasilinear equations in several independent variables.
Math. USSR Sbornik 10, 217–243
Ladyženskaja, O.A., Solonnikov, U.A., Ural’ceva, N.N. (1968): Linear and Quasi–Linear
Equations of Parabolic Type. Volume 23 of Translations of mathematical monographs,
A. M. S., Providence, Rhode Island
Z
Landim, C. (1991a): Hydrodynamical equations for attractive particle systems on d . Ann.
Probab. 19, 1537–1558
Landim, C. (1991b): Hydrodynamical limit for asymmetric attractive particle systems on
Z d . Ann. Inst. H. Poincaré, Probabilités 27, 559–581
Landim, C. (1991c): An overview on large deviations of interacting particle systems. Ann.
Inst. H. Poincaré, Physique Théorique 55, 615–635
Landim, C. (1992): Occupation time large deviations for the symmetric simple exclusion
process. Ann. Probab. 20, 206–231
Z
Landim, C. (1993): Conservation of local equilibrium for attractive particle systems on d .
Ann. Probab. 21, 1782–1808
References 435
Landim, C. (1996): Hydrodynamical limit for space inhomogenuous one dimensional totally
asymmetric zero range processes. Ann. Probab. 24, 599–638
Landim, C., Mourragui, M. (1997): Hydrodynamic limit of mean zero asymmetric zero
range processes in infinite volume. Ann. Inst. H. Poincaré, Probabilités 33, 65–82
Landim, C., Olla, S., Volchan, S.B. (1997): Driven tracer particle and Einstein relation in
one dimensional symmetric simple exclusion process. Resenhas IME–USP 3, 173–209
Landim, C., Olla, S., Volchan, S.B. (1998): Driven tracer particle in one dimensional sym-
metric simple exclusion. To appear in Commun. Math. Phys.
Landim, C., Olla, S., Yau, H.T. (1996): Some properties of the diffusion coefficient for
asymmetric simple exclusion processes. Ann. Probab. 24, 1779–1807
Landim, C., Olla, S., Yau, H.T. (1997): First order correction for the hydrodynamic limit of
asymmetric simple exclusion processes in dimension d 3, Comm. Pure Appl. Math.
L, 149–203
Landim, C., Sethuraman, S., Varadhan, S.R.S. (1996): Spectral gap for zero range dynamics.
Ann. Probab. 24, 1871–1902
Landim, C., Vares, M.E. (1994): Equilibrium fluctuations for exclusion processes with speed
change. Stoch. Proc. Appl. 52, 107–118
Landim, C., Vares M.E. (1996): Exponential estimate for reaction diffusion models. Probab.
Th. Rel. Fields 106, 151–186
Landim, C., Yau, H.T. (1995): Large deviations of interacting particle systems in infinite
volume. Comm. Pure Appl. Math. XLVIII, 339–379
Landim, C., Yau, H.T. (1997): Fluctuation–dissipation equation of asymmetric simple ex-
clusion processes. Probab. Th. Rel. Fields 108, 321–356
Lax, P.D. (1957): Hyperbolic systems of conservation laws II. Comm. Pure Appl. Math.
X, 537–566
Lebowitz, J.L., Neuhauser, C., Ravishankar, K. (1996): Dynamics of a spin–exchange
model. Stoch. Proc. Appl. 64, 187–208
Lebowitz, J.L., Orlandi, E., Presutti, E. (1991): A particle model for spinodal decomposition.
J. Stat. Phys. 63, 933–974
Lebowitz, J.L., Penrose, O. (1966): Rigorous treatment of the Van der Waals-Maxwell
theory of liquid vapour transition. J. Math. Phys. 7, 98–113
Lebowitz, J.L., Presutti, E., Spohn, H. (1988): Microscopic models of hydrodynamic be-
havior. J. Stat. Phys. 51, 841–862
Lebowitz, J.L., Rost, H. (1994): The Einstein relation for the displacement of a test particle
in a random environment. Stoch. Proc. Appl. 54, 183–196
Lebowitz, J.L., Schonmann, R.H. (1987): On the asymptotics of occurence times of rare
events for stochastic spin systems. J. Stat. Phys. 48, 727–751
Lebowitz, J.L., Spohn, H. (1982a): Microscopic basis for Fick’s law of self–diffusion. J.
Stat. Phys. 28, 539–556
Lebowitz, J.L., Spohn, H. (1982b): Steady-state diffusion at low density. J. Stat. Phys. 29,
39–55
Lebowitz, J.L., Spohn, H. (1983): On the time evolution of macroscopic systems. Comm.
Pure Appl. Math. XXXVI, 595–613
Lebowitz, J.L., Spohn, H. (1997): Comment on “Onsager reciprocity relations without
micrsocpic reversibility". Phys. Rev. Lett. 78, 394–395
Liggett, T.M. (1973): An infinite particle system with zero range interactions. Ann. Probab.
1, 240-253
Liggett, T.M. (1975): Ergodic theorems for the asymmetric simple exclusion process. Trans.
Amer. Math. Soc. 213, 237–260
Liggett, T.M. (1985): Interacting Particle Systems, Springer-Verlag, New York
L
Liggett, T.M. (1989): Exponential 2 convergence of attractive reversible nearest particle
systems. Ann. Probab. 17, 403–432
436 References
L
Liggett, T.M. (1991): 2 rates of convergence of attractive reversible particle systems : the
critical case. Ann. Probab. 19, 935–959
Liverani, C., Olla, S. (1996): Ergodicity in infinite Hamiltonian systems with conservative
noise. Probab. Th. Rel. Fields 106, 401–445
Lu, S.L. (1994): Equilibrium fluctuations of a one-dimensional nongradient Ginzburg–
Landau model. Ann. Probab. 22, 1252–1272
Lu, S.L. (1995), Hydrodynamic scaling limits with deterministic initial configurations. Ann.
Probab. 23, 1831–1852
Lu, S.L., Yau, H.T. (1993): Spectral gap and logarithmic Sobolev inequality for Kawasaki
and Glauber dynamics. Commun. Math. Phys. 156, 399–433
Maes, C. (1990): Kinetic limit of a conservative lattice gas dynamics showing long range
correlations. J. Stat. Phys. 61, 667–681
Malyshev, V.A., Manita, A.D., Petrova, E.N., Scacciatelli, E. (1995): Hydrodynamics of
the weakly perturbed voter model. Markov Proc. Rel. Fields 1, 3–56
Marchand, J.P., Martin, Ph. A. (1986): Exclusion process and droplet shape. J. Stat. Phys.
44, 491–504, J. Stat. Phys. 50, 469–471 (1988)
Martin-Löf, A. (1976): Limit theorems for the motion of a Poisson system of independent
Markovian particles at high density. Z. Wahrsch. Verw. Gebiete 34, 205–223
Martinelli, F., Olivieri, E. (1994a): Approach to equilibrium of Glauber dynamics in the
one phase region I : the attractive case. Commun. Math. Phys. 161, 447–486
Martinelli, F., Olivieri, E. (1994b): Approach to equilibrium of Glauber dynamics in the
one phase region II : the general case. Commun. Math. Phys. 161, 487–514
Martinelli, F., Olivieri, E., Scoppola, E. (1990): Metastability and exponential approach to
equilibrium for low-temperature stochastic Ising models. J. Stat. Phys. 61 1005–1119
Morrey, C.B. (1955): On the derivation of the equation of hydrodynamics from statistical
mechanics. Comm. Pure Appl. Math. VIII, 279–326
Mountford, T.S. (1995): Exponential convergence for attractive reversible subcritical nearest
particle systems. Stoch. Proc. Appl. 59, 235–249
Mourragui, M. (1996): Comportement hydrodynamique et entropie relative des processus
de sauts, de naissances et de morts. Ann. Inst. H. Poincaré, Probabilités 32, 361–385
Naddaf, A., Spencer, T. (1997): On homogenization and scaling limit of some grandient
perturbations of a massless free field. Commun. Math. Phys. 183, 55–84
Nappo, G., Orlandi, E. (1988): Limit laws for a coagulation model of interacting random
particles. Ann. Inst. H. Poincaré, Probabilités 24, 319–344
Nappo, G., Orlandi, E., Rost, H. (1989): A reaction–diffusion model for moderately inter-
acting particles. J. Stat. Phys. 55, 579–600
Neuhauser, C. (1990): One dimensional stochastic Ising models with small migration. Ann.
Probab. 18, 1539–1546
Neveu, J. (1972): Martingales à Temps Discret, Dunod, Paris
Noble, C. (1992): Equilibrium behavior of the sexual reproduction process with rapid dif-
fusion. Ann. Probab. 20, 724–745
Oelschläger, K. (1984): A martingale approach to the law of large numbers for weakly
interacting stochastic processses. Ann. Probab. 12, 458–479
Oelschläger, K. (1985): A law of large numbers for moderately interacting diffusion pro-
cesses. Z. Wahrsch. Verw. Gebiete 69, 279–322
Oelschläger, K. (1987): A fluctuation theorem for moderately interacting diffusion pro-
cesses. Probab. Th. Rel. Fields 74, 591–616
Oelschläger, K. (1989): On the derivation of reaction-diffusion equations as limit dynamics
of systems of moderately interacting stochastic processes. Probab. Th. Rel. Fields 82,
565–586
Oleinik, O.A., Kružkov, S.N. (1961): Quasi–linear second–order parabolic equations with
many independent variables. Russian Math. Surveys 16-5, 105–146
References 437
Olivieri, E., Vares, M.E. (1998): Large deviations and Metastability. To be published by
Cambridge Universtiy Press, Cambridge
Olla, S., Varadhan, S.R.S. (1991): Scaling limit for interacting Ornstein-Uhlenbeck pro-
cesses. Commun. Math. Phys. 135, 355–378
Olla, S., Varadhan, S.R.S., Yau, H.T. (1991): Hydrodynamic limit for a Hamiltonian system
with weak noise. Commun. Math. Phys. 155, 523–560
Onsager, L. (1931): Reciprocal relations in irreversible processes I, II. Phys. Rev. 37, 405–
426; 38, 2265–2279
Onsager, L., Machlup, S. (1953): Fluctuation and irreversible processes I, II. Phys. Rev.
91, 1505–1512, 1512–1515
Penrose, O., Lebowitz, J.L. (1987): Towards a rigorous molecular theory of metastability.
In E. W. Montroll and J. L. Lebowitz, editors, Fluctutation Phenomena, second edition.
North-Holland Physics Publishing, Amsterdam
Petrov, V.V. (1975): Sums of Independent Random Variables, Springer-Verlag, New York
Port, S.C., Stone, C.J. (1973): Infinite particle systems. Trans. Amer. Math. Soc. 178, 307–
340
Posta, G. (1995): Spectral gap for an unrestricted Kawasaki type dynamics. European Series
App. Ind. Math. Prob. Stat. 1, 145–181
Presutti, E. (1987): Collective phenomena in stochastic particle systems. In S. Albeverio,
Ph. Blanchard and L. Streit, editors, Stochastic Processes – Mathematics and Physics II.
Volume 1250 of Lecture Notes in Mathematics, pages 195–232, Springer-Verlag, Berlin
Presutti, E. (1997): Hydrodynamics in particle systems. In C. Cercignani, G. Jona–Lasinio,
G. Parisi and L. A. Radicati di Brozolo, editors, Boltzmann’s Legacy 150 Years After His
Birth. Volume 131 of Atti dei convegni Licei, pages 189–207, Accademia Nazionale dei
Lincei, Roma
Presutti, E., Spohn, H. (1983): Hydrodynamics of the voter model. Ann. Probab. 11, 867–
875
Quastel, J. (1992): Diffusion of color in the simple exclusion process. Comm. Pure Appl.
Math. XLV, 623–679
Quastel, J. (1995a): Large deviations from a hydrodynamical scaling limit for a nongradient
system. Ann. Probab. 23, 724–742
Quastel, J. (1995b): Diffusion in disorder media. In T. Funaki and W. Woyczinky, editors,
Proceedings On Stochastic Method for Nonlinear P.D.E., IMA volumes in Mathematics
77, pages 65–79. Springer-Verlag, New York
Quastel, J., Rezakhanlou, F., Varadhan, S.R.S. (1997): Large deviations for the symmetric
simple exclusion process in dimension d 3. Preprint
Quastel, J., Yau, H.T. (1997): Lattice gases, large deviations and the incompressible Navier–
Stokes equation, preprint
Ravishankar, K. (1992a): Fluctuations from the hydrodynamical limit for the symmetric
Z
simple exclusion in d . Stoch. Proc. Appl. 42, 31–37
Ravishankar, K. (1992b): Interface fluctuations in the two–dimensional weakly asymmetric
simple exclusion process. Stoch. Proc. Appl. 43, 223–247
Reed, M., Simon, B. (1975): Methods of Modern Mathematical Physics. Academic Press,
New York
Revuz, D., Yor, M. (1991): Continuous Martingales and Brownian Motion, Springer-Verlag,
Berlin
Rezakhanlou, F. (1990): Hydrodynamic limit for a system with finite range interactions.
Commun. Math. Phys. 129, 445–480
Z
Rezakhanlou, F. (1991): Hydrodynamic limit for attractive particle systems on d . Com-
mun. Math. Phys. 140, 417–448
Rezakhanlou, F. (1994a): Evolution of tagged particles in non-reversible particle systems.
438 References
Rezakhanlou, F. (1994b): Propagation of chaos for symmetric simple exclusion. Comm.

Pure Appl. Math. XLVII, 943–957
Rezakhanlou, F. (1995): Microscopic structure of shocks in one conservation laws. Ann.
Inst. H. Poincaré, Analyse non Linéaire 12, 119–153
Rezakhanlou, F. (1996a): Kinetic limits for a class of interacting particle systems. Probab.
Th. Rel. Fields 104, 97–146
Rezakhanlou, F. (1996b): Propagation of chaos for particle systems associated with discret
Boltzmann equation. Stoch. Proc. Appl. 64, 55–72
Rezakhanlou, F. (1997): Equilibrium fluctuations for the discrete Boltzmann equation. To
appear in Duke Math. J.
Rezakhanlou, F., Tarver III, J.E. (1997): Boltzmann–Grad limit for a particle system in
continuum. Ann. Inst. H. Poincaré, Probabilités 33, 753–796
Rost, H. (1981): Nonequilibrium behavior of many particle systems : density profile and
local equilibria. Z. Wahrsch. Verw. Gebiete 58, 41–53
Rost, H. (1983): Hydrodynamik gekoppelter diffusionen : fluktuationen im gleichgewicht.
In P. Blanchard and L. Streit, editors, Dynamics and Processes. Volume 1031 of Lecture
Notes in Mathematics, pages 97–107. Springer-Verlag, Berlin
Rost, H. (1984): Diffusion de sphères dures dans la droite réelle : comportement macro-
scopique et équilibre local. In J. P. Azéma and M. Yor, editors, Séminaire de Probabilités
XVIII. Volume 1059 of Lecture Notes in Mathematics, 127–143. Springer-Verlag, Berlin
Rost, H. (1985): A central limit theorem for a system of interacting particles. In L. Arnold
and P. Kotelenez, editors, Stochastic space–time models and limit theorems, 243–248, D.
Reidel Publishing Company
Rost, H., Vares, M.E. (1985): Hydrodynamics of a one dimensional nearest neighbor model.
Saada, E. (1987a): A limit theorem for the position of a tagged particle in a simple exclusion
process. Ann. Probab. 15, 375–381
Saada, E. (1987b): Processus de zero-range avec particule marquée. Ann. Inst. H. Poincaré,
Probabilités 26, 5–18
Schütz, G. (1993): Generalized Bethe ansatz solution of a one–dimensional asymmetric
exclusion process on a ring with blockage. J. Stat. Phys. 71, 471–505
Schütz, G., Domany, E. (1993): Phase transition in an exactly soluble one–dimensional
exclusion model. J. Stat. Phys. 72, 277–296
Sellami, S. (1998): Equilibrium density fluctuations field of a one–dimensional non gradient
reversible model: the generalized exclusion process, preprint
Seppäläinen, T. (1996a): A microscopic model for the Burgers equation and longest in-
creasing subsequences. Electronic J. Probab. 1, 1–51
Seppäläinen, T. (1996b): Hydrodynamic scaling, convex duality and asymptotic shapes of
growth models, preprint
Seppäläinen, T. (1997a): A scaling limit for queues in series. Ann. Appl. Probab. 7, 855–872
Seppäläinen, T. (1997b): Increasing sequences of independent points on the planar lattice.
Ann. Appl. Probab. 7, 886–898
Shiga, T. (1988): Tagged particle motion in a clustered random walk system. Stoch. Proc.
Appl. 30, 225–252
Siri, P. (1996): Inhomogeneous zero range process. Ph.D. thesis, Politecnico di Torino
Smoller, J. (1983): Shock Waves and Reaction-Diffusion Equations, Springer-Verlag, New
York
Speer, E. (1994): The two species totally asymmetric simple exclusion process. In M.
Fannes, C. Maes and A. Verbeure, editors, On Three Levels: Micro-, Meso- and Macro-
Approaches in Physics. Volume 324 of Nato ASI series B, pages 91–102.
Spitzer, F. (1970): Interaction of Markov processes. Adv. Math. 5, 246–290.
References 439
Spohn, H. (1985): Equilibrium fluctuations for some stochastic particle systems. In J. Fritz
and A. Jaffe and D. Szász, editors, Statistical Physics and Dynamical Systems, pages
67–81. Birkhäuser, Boston
Spohn, H. (1986): Equilibrium fluctuations for interacting Brownian particles. Commun.
Math. Phys. 103, 1–33
Spohn, H. (1987): Interacting Brownian particles : a study of Dyson’s model. In G. C.
Papanicolaou, editor, Hydrodynamic Behavior and Interacting Particle Systems, IMA
volumes in Mathematics 9, pages 151–179. Springer-Verlag, New York
Spohn, H. (1990): Tracer diffusion in lattice gases. J. Stat. Phys. 59, 1227–1239
Spohn, H. (1991): Large Scale Dynamics of Interacting Particles, Springer-Verlag, Berlin
Spohn, H. (1993): Interface motion in models with stochastic dynamics. J. Stat. Phys. 71,
1081–1132
Spohn, H., Yau, H.T. (1995): Bulk diffusivity of lattice gases close to criticality. J. Stat.
Phys. 79, 231–241
Stroock, D.W., Varadhan, S.R.S. (1979): Multidimensional Diffusion Processes. Springer-
Verlag, New York
Stroock, D.W., Zegarlinski, B. (1992a): The equivalence of the logarithmic Sobolev in-
equality and the Dobrushin–Shlosman mixing condition. Commun. Math. Phys. 144,
303–323
Stroock, D.W., Zegarlinski, B. (1992b) The logarithmic Sobolev inequality for continuous
spin systems on a lattice. J. Funct. Anal. 104, 299-326
Stroock, D.W., Zegarlinski, B. (1992c): The logarithmic Sobolev inequality for discrete
spin systems on a lattice. Commun. Math. Phys. 149, 175–193
Suzuki, Y., Uchiyama, K. (1993): Hydrodynamic limit for a spin system on a multidimen-
sional lattice. Probab. Th. Rel. Fields 95, 47–74
Thomas, L.E. (1989): Bound on the mass gap for finite volume stochastic Ising models at
low temperature. Commun. Math. Phys. 126, 1–11
Uchiyama, K. (1994): Scaling limits of interacting diffusions with arbitrary initial distribu-
tions. Probab. Th. Rel. Fields 99, 97–110
Varadhan, S.R.S. (1966): Asymptotic probabilities and differential equations.
Comm. Pure Appl. Math. XIX, 261–286
Varadhan, S.R.S. (1991): Scaling limits for interacting diffusions. Commun. Math. Phys.
135, 313–353
Varadhan, S.R.S. (1994a): Nonlinear diffusion limit for a system with nearest neighbor in-
teractions II. In K. D. Elworthy and N. Ikeda, editors, Asymptotic Problems in Probability
Theory : Stochastic Models and Diffusion on Fractals. Volume 283 of Pitman Research
Notes in Mathematics, pages 75–128. John Wiley & Sons, New York
Varadhan, S.R.S. (1994b): Regularity of self diffusion coefficient. In M. Freidlin, editor, The
Dynkin Festschrift, Markov Processes and their Applications, pages 387–397, Birkhäuser,
Boston
Varadhan, S.R.S. (1995): Self diffusion of a tagged particle in equilibrium for asymmetric
mean zero random walk with simple exclusion. Ann. Inst. H. Poincaré, Probabilités 31,
273–285
Varadhan, S.R.S., Yau, H.T. (1997): Diffusive limit of lattice gases with mixing conditions,
preprint
Vares, M.E. (1991): On long time behavior of a class of reaction–diffusion models. Ann.
Inst. H. Poincaré, Physique Théorique 55, 601–613
Venkatraman, R. (1994): Hydrodynamic limit of the asymmetric simple exclusion pro-
S
cess with deterministic initial data and the Hammersley process on (1). Ph.D. Thesis,
Courant Institute, New York University.
Wick, W.D. (1985): A dynamical phase transition in an infinite particle system. J. Stat.
Phys. 38, 1015–1025
440 References
Wick, W.D. (1989): Hydrodynamic limit of nongradient interacting particle processes. J.

Stat. Phys. 89, 873–892
Xu, L. (1993): Diffusion limit for the lattice gas with short range interactions. Ph.D. thesis,
New York University
Yau, H.T. (1991): Relative entropy and hydrodynamics of Ginzburg–Landau models. Lett.
Math. Phys. 22, 63–80
Yau, H.T. (1994): Metastability of Ginzburg–Landau model with a conservation law. J.
Stat. Phys. 74, 705–742
Yau, H.T. (1996): Logarithmic Sobolev inequality for lattice gases with mixing conditions.
Yau, H.T. (1997): Logarithmic Sobolev inequality for generalized simple exclusion pro-
cesses. Probab. Th. Rel. Fields 109, 507–538
Yosida, K. (1995): Functional Analysis. Reprint from the sixth edition, Springer-Verlag,
Berlin
Zegarlinski, B. (1990a): On logarithmic Sobolev inequalities for infinite lattice systems.
Lett. Math. Phys. 20, 173–182
Zegarlinski, B. (1990b): Logarithmic Sobolev inequalities for infinite one dimensional lat-
tice systems. Commun. Math. Phys. 133, 147–162
Zegarlinski, B. (1992): Dobrushin uniqueness theorem and logarithmic Sobolev inequalities.
J. Funct. Anal. 105, 77–111
Zhu, M. (1990): Equlibrium fluctuations for one dimensional Ginzburg–Landau lattice
model. Nagoya Math. J. 117, 63–92.
Subject Index
Central limit theorem Fick’s law of transport 111

– For additive functionals of Markov Filippov solutions 225
processes 63
– For second class particles 252 Generator 322
– For the empirical measure 229 – Adjoint 328
– For tracer particles 226 – Symmetric part 329
– Local central limit theorem 352, 354 Green–Kubo formula 144
Closed form
L
– Closed form in 2 ( ) 408 Hydrodynamic equation
– Closed form in Zd 398 – Boltzmann equations 133
– Continous closed form 403 – Broadwell equation 133
– Germ of a closed form 408 – Cahn–Hilliard equations 131
Conservation of local equilibrium 16, 222 – Carleman equation 133
Convergence to non-extremal measures – Euler equations 130
254 – Motion by mean curvature 137
Current 61, 141 – Navier–Stokes equations 185
– Porous medium equation 110
Decay to equilibrium 420
– Reaction–diffusion equations 131
Degenerate diffusions 134
– Stefan problems 132
Detailed balance condition 326
Hyperbolic equation
Dirichlet form 343
– Entropy solution 193
Dynamical phase transition 254
– Measure valued solution 196
Einstein relations 65 – – Dirac solution 197
Empirical measure 43 – – Entropy solution 197
Entropy 338
– Entropy inequality 338 Initial state
– Entropy production 342 – Local equilibrium 13
– Explicit formula 339 – Probability measure associated to a
– Specific entropy 113 profile 43
Equivalence of ensembles 351, 355 – Product measure with slowly varying
– Second order expansion 357 parameter 12, 42
Escape from unstable equilibrium points – Weak local equilibrium 43
283 Interacting diffusions 138
Exact form Interacting particle systems
– Continuous exact form 403 – Asymmetric processes 27
– Exact form in Zd 399 – Glauber dynamics 136, 281
– Germ of an exact form 408 – Gradient systems 61
Exit points from a basin of attraction 282 – Hammersley process 223
– Mean-zero asymmetric processes 27
Feynman–Kac formula 334 – Processes in random environment 223
442 Subject Index
– Spatially inhomogeneous processes Onsager–Machlup time–reversal relation

222 280
– Symmetric processes 27 Oriented bond 143
– Weakly asymmetric processes 223
Interface motion 134 Partition function 28
Invariant measure 326 Propagation of chaos 65, 227
Invariant states
– Canonical measure 28 Quasi–potential 280
– Grand canonical measure 28
Radon–Nikodym derivative
Kac Potential 136 – Homogeneous case 320
– Inhomogeneous case 337
Laplace–Varadhan theorem 362 Rare events 284
Largest eigenvalue Replacement lemma 77
– Of small perturbation of reversible – One block estimate 82
generators 375 – Two blocks estimate 83
Reversible measure 327
– Variational formula 377
Lebowitz–Penrose limit 137
Second class particle 251
Lipschitz cylinder function 33
Self diffusion 62
Logarithmic Sobolev inequalities 422 Stationary measure of asymmetric systems
255
Markov chain 312 Stochastic reservoirs 110
Markov processes 319 Sublinear jump rate 82
– Chapman–Kolmogorov equation 322
– Forward equation 341 Tightness
Metastability 281 – Aldous criterion 51
Microscopic structure of the shock 251 – Exponential tightness 271
Minimax lemma 363 Tracer particle 62, 224
Modified modulus of continuity 50 Trotter–Kato formula 323
Nash inequalities 420 Weak convergence 22

Navier–Stokes corrections 185 Weakly interacting diffusions 139
Onsager reciprocity relations 112 Young measure 198

Livro Claudio

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Livro Claudio

Uploaded by

Copyright:

Available Formats

Claude Kipnis and Claudio Landim

Scaling Limit of Interacting

The idea to write up a book on the hydrodynamic behavior of interacting particle

formulations of local equilibrium. These two chapters can be skipped by someone

Campos dos Goytacazes, April 1998. Claudio Landim

2. Some interacting particle systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3. Weak formulations of local equilibrium . . . . . . . . . . . . . . . . . . . . . . . 43

5. An example of reversible gradient system : symmetric zero range

6. The relative entropy method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

7. Hydrodynamic limit of reversible nongradient systems . . . . . . . . . . 147

8. Hydrodynamic limit of asymmetric attractive processes . . . . . . . . . 197

9. Conservation of local equilibrium for attractive systems . . . . . . . . . 237

10. Large deviations from the hydrodynamic limit . . . . . . . . . . . . . . . . . 263

11. Equilibrium Fluctuations of Reversible Dynamics . . . . . . . . . . . . . . 293

1. Markov chains on a countable space . . . . . . . . . . . . . . . . . . . . . . . . . . 319

1.1 Discrete time Markov chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320

2. The equivalence of ensembles, large deviation tools and weak

3. Nongradient tools : spectral gap and closed forms . . . . . . . . . . . . . . 383

M+ (Td R+ ) 200 Tz ;z  392

" 69, 77 0 214

@uNi 142, 300

To avoid degeneracies, we assume that c(1;  ) > 0 if  (0) > 0 so that a

N() f; (x) = kg = (x=N ) f; (0) = kg :

In this chapter we investigate in detail the case of indistinguishable particles mov-

Let pt (x; y ) represent the probability of being at time t on site y for a

Recall that a Poisson distribution of parameter  0 is the probability measure

Proposition 1.1 If particles are initially distributed according to a Poisson mea-

We announced that the passage from microscopic to macroscopic would be done

Definition 2.1 (Product measure with slowly varying parameter associated to

(x  )(y ) = (y + x); y 2 TdN :

Definition 2.2 (Local equilibrium) A sequence of probability measures (N )N 1

that it is mean–zero asymmetric if p() is not symmetric but m = 0 and that it is

([uN ]) = 0 (u mt) = : (t; u)

if r denotes the gradient of  : r = (@u ; : : : ; @ud ).

Proposition 3.1 A system of particles evolving as independent asymmetric ran-

When the expectation m vanishes, the solution of this differential equation is

Proposition 3.2 A system of particles evolving as independent mean–zero asym-

Let fS N (t); t  0g be the semigroup on M1 associated to the Markov pro-

Definition 3.3 (Conservation of local equilibrium) The local equilibrium (N )N 1

In Section 2 we chose a class of invariant measures to describe the equilib-

that, as N " 1, converges to

It is in this sense, known to the physicists as the “equivalence of ensembles",

5. Comments and References

The mathematical formulation of the derivation of macroscopic evolution equa-

and @t  = Bi;j @u2 i ;uj  C +

Stationary measures. Relying on the hydrodynamic behavior of a superposition of

We introduce in this chapter the interacting particle systems we consider through-

1. Some remarks on the topology of N Z and

functions can not be uniformly approximated by cylinder functions, we present an

the weak topology defined as follows :

To prove the weak convergence of probability measures, we need only to

Proof. Consider a sequence of probability measures k such that Ek [ ] con-

while the absolute value of the second is bounded above by

for all functions f in Cb;m .

second) marginal is equal to 1 (resp. 2 ) and ¯ is concentrated on configura-

and then impose the measure to be periodic :

Remark 1.7 All analysis developed above extends in a straightforward man-

2. Simple exclusion processes (hard core interaction)

Definition 2.1 Let p be a finite range, translation invariant, irreducible transition

p(x; y) = p(0; y x) =: p(y x)

The interpretation is clear. Between 0 and dt each particle tries, independently

density . In particular, under  , the variables f (x); x 2 TdN g are independent

Proposition 2.2 The Bernoulli measures f ; 0   1g are invariant for simple

M+ (Td R+ ) 200 Tz ;z 392

" 69, 77 0 214

To avoid degeneracies, we assume that c(1; ) > 0 if (0) > 0 so that a

N() f; (x) = kg = (x=N ) f; (0) = kg :

Recall that a Poisson distribution of parameter 0 is the probability measure

(x )(y ) = (y + x); y 2 TdN :

Definition 2.2 (Local equilibrium) A sequence of probability measures (N )N 1

([uN ]) = 0 (u mt) = : (t; u)

if r denotes the gradient of : r = (@u ; : : : ; @ud ).

Let fS N (t); t 0g be the semigroup on M1 associated to the Markov pro-

Definition 3.3 (Conservation of local equilibrium) The local equilibrium (N )N 1

and @t = Bi;j @u2 i ;uj C +

Proof. Consider a sequence of probability measures k such that Ek [ ] con-

second) marginal is equal to 1 (resp. 2 ) and ¯ is concentrated on configura-

density . In particular, under , the variables f (x); x 2 TdN g are independent

Proposition 2.2 The Bernoulli measures f ; 0 1g are invariant for simple

For each 0 K N d , denote by N;K

(the family fN;K ; 0 K N d g in the context of simple exclusion processes,

for each k 0 and x in TdN .

() = ¯() () : (3:6)

for every 0. Moreover, a simple computation shows that the function () is

@' log Z (') C0 ' 1 :

g(a) f; (x) = ag = ' f; (x) = a 1g :

Lemma 3.5 Suppose that 1 2 . Then . 1 2

whose second marginal is equal to 2 and which is concentrated on configurations

gE ; (0) (0) :

Since the measure 1 ;2 is concentrated on configurations 1 2 we can remove

With the notation just introduced, we have the identity g~ = .

Lemma 3.8 f ; 0g is a continuous family : ~() = E [ ] is a continuous

Proof. Fix 0 in R+ and a sequence fk ; k 1g converging to 0 . By Lemma

Lemma 3.9 For every 0, the set of probability measures A = f; g

¯' f; (x) = kg 'k

generator L̄N starting from ¯ . Expectation with respect to P̄¯ is denoted by Ē ¯ .

Since (t ; t ) is right continuous (cf. section 2 of Appendix 1) and F0 =

because . This proves the claim.

p(x; y) = p(0; y x) =: p(y x) 0

Theorem 6.1 There exists a semigroup fS (t); t 0g defined on Ll such that

The existence of a probability P on D(R+ ; N Z ), for every in El , corre-

0 = g(0) < g(k) g(k + 1)

Theorem 6.3 A translation invariant probability measure bounded above by

N() f; (x) = kg

the expected value of under :

Definition 0.2 . (Weak local equilibrium) A sequence (N )N 1 of probability

In this formula, for a d-dimensional vector u, u represents the Dirac measure

we shall also denote by the measure ( N ) 1 on M+ induced by and N

random measures N (du) converges in probability to the deterministic measure

Proposition 0.4 Let (N )N 1 be a local equilibrium of profile 0 almost surely

at all continuity points u of 0 . By the law of large numbers, when ` tends to

Proposition 0.6 Let (N )N 1 be a weak local equilibrium of profile 0 . Assume

lim sup sup jk ; (v) 0(u)j = 0:

Then the sequence (N )N 1 is a strong local equilibrium of profile 0 .

lim sup [uN ] EN [ ] ~(0(u)) = E u[ ] :

The same arguments replacing the sequence N N