(J.-P. Rivet, J. P. Boon) Lattice Gas Hydrodynamic

Lattice gas hydrodynamics is the name used to describe the approach to fluid dynamics
starting from a simple micro-world constructed as an automaton universe. In this

universe, the microscopic dynamics is based not on a realistic description of interacting
particles, but merely on the laws of symmetry and of invariance of macroscopic physics.
We imagine point-like particles residing on a regular lattice, where they move from
node to node and undergo collisions when their trajectories meet at the same node. If the
collisions occur according to some simple logical rules, and if the lattice has the proper
symmetry, then the automaton shows global behavior very similar to that of real fluids.
This book carries two important messages. First, it shows how an automaton universe
with very simple microscopic dynamics - the lattice gas - can exhibit macroscopic
behavior in accordance with the phenomenological laws of classical physics. Second, it
demonstrates that lattice gases have spontaneous microscopic fluctuations which capture
the essentials of actual fluctuations in real fluids. The two aspects are closely related,
offering a microscopic approach to the description of fluid systems. Accordingly, the
book follows the philosophy of classical statistical mechanics, and introduces the reader
to the world of the discrete analogue of continuous molecular hydrodynamics.
This book will be of interest to graduate students and researchers in statistical
physics, computational physics, hydrodynamics, applied mathematics and engineering.
JEAN PIERRE BOON studied at the University of Brussels and at the University of
Chicago. He has been FNRS Research Fellow and Professor at the University of
Brussels since 1975. He is a former Visiting Professor at Bell Telephone Laboratories, at
the Kernforschungsanlage Julich, at the Massachusetts Institute of Technology and at
the University of Nice. J.P. Boon's main field of research is statistical physics. He is the
co-author of Molecular Hydrodynamics by J.P. Boon and S. Yip.
JEAN-PIERRE RIVET received his PhD in 1988 from the University of Nice for work on
theoretical and numerical aspects of lattice gas hydrodynamics. Since then, he has been
working on applications of the lattice gas method to computational fluid dynamics.
Since 1996 he has worked on relativistic stochastic processes and statistical physics, in
collaboration with researchers at the Laboratory of Radio-Astronomy of the Ecole
Normale Superieure (Paris). In addition to his research activity, he has given several
courses in Physics and Astronomy at the University of Paris VI, and at the University of
Nice. He is presently CNRS researcher at the Observatoire de la Cote d'Azur (Nice).
Cambridge Nonlinear Science Series 11
EDITORS
Professor Boris Chirikov

Budker Institute of Nuclear Physics, Novosibirsk
Professor Predrag Cvitanovic

Niels Bohr Institute, Copenhagen
Professor Frank Moss

University of Missouri, St Louis
Professor Harry Swinney

Center for Nonlinear Dynamics,
The University of Texas at Austin
Lattice Gas Hydrodynamics
TITLES IN PRINT IN THIS SERIES
G. M. Zaslavsky, R. Z. Sagdeev,
D. A. Usikov and A. A. Chernikov
1. Weak chaos and quasi-regular patterns
J. L. McCauley
2. Chaos, dynamics and fractals: an
algorithmic approach to deterministic
chaos
K. Nakamura
3. Quantum chaos: a new paradigm for
nonlinear dynamics
C. Beck and F. Schlogl
4. Thermodynamics of chaotic systems:
an introduction
P. Meakin
5. Fractals, scaling and growth far from
equilibrium
R. Badii and A. Politi
6. Complexity - hierarchical structures
and scaling in physics
H. Kantz and T. Schreiber
7. Nonlinear time series analysis
T. Bohr, M. H. Jensen, G. Paladin and A. Vulpiani
8. Dynamical systems approach to turbulence
P. Gaspard
9. Chaos, scattering and statistical mechanics
E. Scholl
10. Nonlinear Spatio-Temporal Dynamics and
Chaos in Semiconductors
J.-P. Rivet and J. P. Boon
11. Lattice Gas Hydrodynamics
Lattice Gas
Hydrodynamics
J.-P. Rivet
Observatoire de la Cote d'i4zur, France
and J. P. Boon
Universite de Bruxelles, Belgium
CAMBRIDGE
UNIVERSITY PRESS
CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, Sao Paulo
Cambridge University Press
The Edinburgh Building, Cambridge CB2 2RU, UK
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521419444
© Cambridge University Press 2001
This publication is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without
the written permission of Cambridge University Press.
First published 2001

This digitally printed first paperback version 2005
A catalogue recordfor this publication is available from the British Library
ISBN-13 978-0-521 -41944-4 hardback

ISBN-10 0-521-41944-1 hardback
ISBN-13 978-0-521 -01971 -2 paperback

ISBN-10 0-521-01971-0 paperback
Contents
Preface xiii
Chapter 1 Basic ideas 1
1.1 The physicist's point of view 1

1.2 The mathematician's point of view 2
1.2.1 Finite automata 2
1.2.2 Cellular automata 3
1.2.3 Lattice gases 3
1.3 Comments 5
1.3.1 The velocity vectors 5
1.3.2 Space and time 6
1.3.3 The exclusion principle 6
1.3.4 Bravais lattices 7
1.3.5 Local versus non-local collisions 8
1.3.6 Collision-propagation versus propagation-collision
1.3.7 Mathematical versus physical 9
Chapter 2 Microdynamics: general formalism 10
2.1 Basic concepts and notation 10

2.1.1 The lattice and the velocity vectors 10
2.1.2 The Boolean field 12
2.1.3 Observables 12
2.1 A Generalized observables 14
vn
viii Contents
2.2 The microdynamic equation 15

2.2.1 Formal expression 16
2.2.2 The propagation operator 16
2.2.3 The collision operator 18
2.2.4 Analytic expressions of the microdynamic equation 20
2.3 Microscopic properties of a lattice gas 20
2.3.1 Detailed and semi-detailed balance 20
2.3.2 Duality 21
2.3.3 Conservation laws 21
2.3.4 G-invariance 23
2.3.5 Crystallographic isotropy 26
2.3.6 Irreducibility 28
2.4 Special rules 29
2.4.1 Solid impermeable obstacles 30
2.4.2 Sources and sinks of observable quantities 33
2.5 Comments 33
Chapter 3 Microdynamics: various examples 34
3.1 The HPP model 34

3.1.1 The micro dynamical equation 36
3.1.2 Microscopic properties 38
3.2 The FHP-1 model 39
3.2.1 The micro dynamical equation 41

3.3.1 The microdynamical equation 44
3A.I The microdynamical equation 48
3.5 The 'colored' FHP model (CFHP) 50
3.6 The GBL model 53
3.7 Three-dimensional models 57
Contents ix
3.7.1 Models with multiple links 58

3.7.2 Models with biased collisions 59
3.7.3 The FCHC models 59
3.7.4 Collision rules 63
Chapter 4 Equilibrium statistical mechanics 68
4.1 The Liouville description 68

4.1.1 Macrostates 69
4.1.2 Ensemble-averages 69
4.1.3 The lattice Liouville equation 70
4.2 The Boltzmann description 71
4.2.1 The Boltzmann approximation 71
4.2.2 The lattice Boltzmann equation 72
4.3 The if-theorem 73
4.3.1 Some basics about communication and information 74
4.3.2 The H-theorem for lattice gases 77
4.4 Global equilibrium macrostates 80
4.4.1 The Liouville approach 81
4.4.2 The lattice Boltzmann approach 84
4.4.3 The variational approach 85
4.5 Natural parameterization of equilibria 86
4.5.1 Low-speed equilibria for single-species non-thermal
models 89
4.5.2 Nearly equally distributed equilibria for thermal models 93
4.6 Statistical thermodynamics 98
4.7 Static correlation functions 105
Chapter 5 Macrodynamics: Chapman-Enskog method 109
5.1 Local equilibria and the hydrodynamic limit 110

5.2 The multi-scale expansion for macrodynamics 111
5.2.1 The scale separation parameter 111
5.2.2 Perturbed local equilibrium 111
5.2.3 Macroscopic space and time scales 112
5.2.4 The averaged microdynamic equation 114
5.2.5 The expansion in powers of e 116
5.3 First order macrodynamics 118
x Contents
5.3.1 Solvability conditions for the first order problem 118

5.3.2 Solution of the first order problem 121
5.4 Second order macrodynamics 126
5.4.1 Solvability conditions for the second order problem 126
5.5 The macrodynamic equations 127
5.6 Transport coefficients within the Boltzmann
approximation 129
5.7 Non-thermal models 132
5.7.1 First order macrodynamics 132
5.7.2 Second order macrodynamics 133
5.7.3 The macrodynamic equation 134
5.7.4 The transport coefficients 135
5.8 Comments 135
Chapter 6 Linearized hydrodynamics 137
6.1 The linearized Boltzmann equation 138

6.2 Slow and fast variables 140
6.3 The hydrodynamic limit 144
6.3.1 The coupling function 145
6.3.2 The memory function 145
6.3.3 The random force term 147
6.3.4 The long-wavelength, long-time limit 148
6.4 The transport matrix 150
6.5 Comments 151
Chapter 7 Hydrodynamic fluctuations 153
7.1 The dynamic structure factor 154

7.2 Fluctuation correlations 155
13 The hydrodynamic modes 157
7.3.1 The spectral decomposition 157
7.3.2 The eigenvalues 160
7.4 The hydrodynamic spectrum 163
7.5 The eigenvalue spectrum 164
7.5.1 Hydrodynamic regime: k£f < 1 165
7.5.2 Generalized hydrodynamic regime: fe/f < 1 166
7.5.3 Kinetic regime: fc/f >, 1 167
7.6 Power spectrum 168
7.6.1 High density 168
Contents XI
7.6.2 Low density 171

7.6.3 Dispersion effects 172
7.7 Diffusion and correlations 174
7.7.1 The two-species lattice gas 175
7.7.2 The hydrodynamic limit 177
7.7.3 The power spectrum 178
Chapter 8 Macrodynamics: projectors approach 183
8.1 Preliminaries 184

8.2 Multiple scales analysis 188
8.3 The hydrodynamic equations 190
8.4 Linear response and Green-Kubo coefficients 193
8.5 Long-time tails 196
Chapter 9 Hydrodynamic regimes 199
9.1 The acoustic limit 200

9.2 The incompressible limit 203
9.3 Comments 205
9.3.1 Invariances 205
9.3.2 Four-dimensional models 205
9.3.3 Lattice gases to simulate real fluid dynamics 206
Chapter 10 Lattice gas simulations 207
10.1 Lattice gas algorithms on dedicated machines 208

10.2 Lattice gas algorithms on general purpose computers 210
10.2.1 Channel-wise vs. node-wise storage 210
10.2.2 Collision strategies 213
10.2.3 Obstacles 214
10.3 Essential features of a lattice gas simulation code 215
10.3.1 Initialization 215
10.3.2 Raw physical data extraction 216
10.3.3 Post-processing 217
10.4 Measurement of basic lattice gas properties 219
10.4.1 Measuring g(p) and v(p) 220
10.4.2 Measuring cs(p) and v'(p) 220
10.4.3 An example: the FCHC-3 model 221
10.5 Examples of lattice gas simulations 222
xii Contents
10.5.1 The Kelvin-Helmholtz instability 222

10.5.2 Particle aggregation 223
10.5.3 Two-dimensional flow past an obstacle 225
10.5.4 Three-dimensional flow past an obstacle 227
10.5.5 Two-dimensional flow of a two-phase fluid in a porous
medium 231
Chapter 11 Guide for further reading 233
11.1 The historical 'roots' 234

11.1.1 Discrete kinetic theory 234
11.1.2 The early days 235
11.1.3 Cellular automata 235
11.3 Theoretical analyses 236
11.3.1 General lattice gas theory 237
11.3.2 Statistical physics and thermodynamics 237
11.3.3 Violation of semi-detailed balance 239
11.3.4 Invariants and conservation laws 240
11.3.5 Obstacles and Knudsen layers 240
11A Models with particular features 241
11 A.I Fluid mixtures and colloids 241
11.4.2 Reaction-diffusion systems 241
11.4.3 Immiscible fluids and free interfaces 242
11A A Flow in porous media 243
11.4.5 Thermo-hydrodynamics 243
11.4.6 Elastic waves 243
11.4.7 Other models 244
11.5 Lattice Boltzmann method 244
11.6 Lattice Bhatnagar-Gross-Krook model 246
11.7 Numerical simulations and implementations 246
11.7.1 Implementation on dedicated hardware 246
11.7.2 Simulations on general purpose computers 247
11.8 Books and review articles 248
Appendix Mathematical details 250

References 275
Author index 281
Subject index 285
Preface
... Feynman told us to explain it like this: We have noticed in nature that the
behavior of a fluid depends very little on the nature of the individual particles in
that fluid. [...] We have therefore taken advantage of this fact to invent a type
of imaginary particle that is especially simple for us to simulate. This particle is
a perfect ball bearing that can move at a single speed in one of six directions.
The flow of these particles on a large enough scale is very similar to the flow of
natural fluids.
W.D. Hillis, Physics Today, February 1989
The story of lattice gas automata started around 1985 when pioneering studies
established theoretically and computationally the feasibility of simulating fluid
dynamics via a microscopic approach based on a new paradigm : a fictitious
oversimplified micro-world is constructed as an automaton universe based not
on a realistic description of interacting particles (as in molecular dynamics),
but merely on the laws of symmetry and of invariance of macroscopic physics.
Imagine point-like particles residing on a regular lattice where they move from
node to node and undergo collisions when their trajectories meet at the same
node. The remarkable fact is that, if the collisions occur according to some
simple logical rules and if the lattice has the proper symmetry, this automaton
shows global behavior very similar to that of real fluids. Then, to observe the
fluid-like behavior of the automaton, we look at the lattice from a distance and,
to measure its properties, we perform measurements over averaged quantities
to obtain for instance the fluid velocity field or to visualize its stream lines (see
Figure 1).
At the beginning, most of the effort was invested into lattice gas hydrody-
namics through theoretical developments recovering Navier-Stokes equations,
xin
Preface
simulations showing wakes and eddy formation, and hardware elaboration of

dedicated computers. By the late eighties, the field of lattice gas automata had
flourished into a broad range of theoretical studies and applications : statistical
mechanics theories had gone deeper into the basics of lattice gases and models
had been proposed for a variety of phenomena. The concomitant consequence
was the sophistication of the new models and of the methods developed for
their implementation. As a result, the field was also branching out into special-
ized sub-areas, such as statistical mechanics of lattice gases, three-dimensional
Figure 1 Beyond some threshold value of the Reynolds number, the flow
behind an obstacle produces a non-stationary wake, the Benard-von Karman
vortex street illustrated in (a) by a picture taken by Prandtl and Tietjens in
1934; (b) shows the result of a lattice gas simulation (performed on the
dedicated machine 'RAP-T, see Chapter 10) under conditions similar to those
of the laboratory experiment. In both (a) and (b), the Reynolds number is of
the order of 200. ((a) In An Introduction to Fluid Dynamics, G.K. Batchelor,
Cambridge University Press, 1967; (b) J.P. Rivet, 1987.)
Preface xv
models for relatively high Reynolds numbers and transition to turbulence, multi-
species and multi-phase models for low Reynolds number complex flows (with
or without surface tension), boundary layers, lattice Boltzmann approach to
turbulence, thermal lattice gases, reactive systems. Despite an apparent diver-
sity, lattice gas automata developed for generic or specific physical problems
all share a common basis, conceptually and operationally. The reason can be
found in the following considerations.
Often systems with a large number of degrees of freedom exhibit macroscopic
behavior where the details of the microscopic dynamics are relatively unimpor-
tant. This feature opens the way to a generic description: systems with different
microscopic characteristics can be described at the macroscopic scale with a
generic set of equations where the specific nature of the constituents of the
system enter only a restricted number of coefficients at the level of constitutive
relations. Since the macroscopic description does not provide any content to
these coefficients, their specificity can be (at least partly) eliminated when the
equations are properly recast into non-dimensional form. Then the global behav-
ior of the system depends on a limited number of universal control parameters
where the microscopic nature of the elementary constituents does not appear
explicitly. The macroscopic level of analysis therefore provides a phenomenolog-
ical description where (i) the complexity of identifying the microscopic degrees
of freedom is bypassed, and (ii) many different physical phenomena are recog-
nized to belong to a limited number of classes. Classical fluid mechanics offers
a striking example through Reynolds's dynamical similarity law (1883). Never-
theless, it is obvious that in this process of reasoning the connection between
the phenomenology and the underlying microscopic mechanisms has been lost.
Consequently, the macroscopic level of description which uses average quantities
prevents analysis of, for instance, how large-scale phenomena can be triggered
by local and/or transient deviations from averages, that is, fluctuations. This
is one of the objects of the statistical mechanical approach which establishes
the microscopic basis of the properties of many-body systems. But we have
observed that it is not necessary to have a complete knowledge of the details
of the microscopic interactions to understand how macroscopic phenomena
emerge. It will appear as logical conduct to develop a microscopic approach to
macroscopic phenomena by modeling the microscopic dynamics by means of
a simplified description provided the basic requirements of fundamental physics
are correctly incorporated, essentially - although further requirements may be
in order - the conservation laws and the symmetry properties.
Our description uses a class of simple discrete models, where space and time
are incremented by finite elementary units, as in the physical picture that we
presented where point particles move step by step on a regular lattice. A
logical question to ask is: why are we interested in such discrete models? The
fact is that despite their simplicity at small scale (at the microscopic level),
xvi Preface
these discrete systems can show extremely complex behavior at larger scales
(the mesoscopic and macroscopic levels), and this global behavior can be very
similar to what is observed in natural phenomena (see Figure 1). Now many
of these natural phenomena are very complicated (for instance the diversity
of forms that a fluid can exhibit in the turbulent regime), and often the best
one can do is to offer a phenomenological analysis of what is observed. Yet
the objective is to try to understand how things come about by providing
an analysis in terms of the elementary processes that are responsible for the
emergence of large-scale phenomena, a formidable task if one starts with a
realistic description of all the basic interactions that govern the underlying
microscopic dynamics.
The philosophy that constitutes the conceptual basis of the lattice gas au-
tomaton (LGA) approach is that a microscopically simple system which can
exhibit complex macroscopic behavior in accordance with observed phenomena
should contain, at the elementary level, the essentials which are responsible for
the emergence of complexity, and thereby can teach us something about the
basic mechanisms from which complexity builds up.
As a matter of fact, when LGAs were initially developed around 1985, the
goal was to produce simplified models that could provide new computational
accessibility to complicated problems in fluid dynamics such as 2-D and 3-D
turbulence. While earlier studies (in the seventies) had considered even simpler
lattice models for basic problems in statistical mechanics, subsequently a variety
of models were constructed for more involved physical applications: hydrody-
namic instabilities, flows in porous media, phase transitions, reactive systems,
etc. All these models share a common structure where point particles move
along the links of a regular lattice and interact on the nodes through collisions.
The key point is that these simple basic ingredients (along with some symmetry
requirement for the lattice structure) suffice for the emergence of complex global
behavior from elementary processes. Furthermore, the lattice gas automaton ex-
hibits two other important features: (i) it usually resides on a large lattice, and
so possesses a large number of degrees of freedom; (ii) its microscopic Boolean
nature (which we shall describe in detail in the next two chapters) combined with
the (generally) stochastic rules which govern its microscopic dynamics, result in
intrinsic fluctuations. Because of these spontaneous fluctuations and because of
its large number of degrees of freedom, the lattice gas can be considered as a
'reservoir of thermal excitations' in much the same way as an actual fluid (see
Figure 2).
The principal goal of this book is to carry two messages. The first is to
show how an automaton universe with simple microscopic dynamics can exhibit
macroscopic behavior in accordance with the phenomenological description of
classical physics. The second is to establish that the correlations of the lattice gas
intrinsic fluctuations capture the essentials of actual fluctuations in real fluids.
Preface
The two aspects are of course closely connected. Therefore our emphasis is on
the statistical mechanics of lattice gas automata as a microscopic approach to
the description of fluid systems, and in this sense we proceed very much along
the lines of the general philosophy of classical statistical mechanics.
-4.0 -3.0 -2.0 -1.0 0.0 1.0 2.0 3.0 4.0

co (GHz)
-2.0 -1.0 0.0 1.0 2.0

CO
Figure 2 The spectrum of density fluctuations in the hydrodynamic regime, (a)

The Rayleigh-Brillouin-Mandelstahm spectrum measured by light scattering
spectroscopy in liquid argon (Fleury and Boon, 1969); (b) the dynamic
structure factor S(k,co) for the GBL thermal lattice gas (see Section 3.6)
(Grosfils, Boon, Lallemand, 1992). In both (a) and (b), S(k,co) is in arbitrary
units, and in (b) co is in reciprocal simulation time units.
xviii Preface
After this simple presentation of the lattice gas picture and of the ideas
underlying the philosophy of the lattice gas automaton approach, we introduce
the basic notions of cellular automata and of lattice gas automata from the
physical point of view and from the mathematical point of view, and we
discuss the complementary aspects of the two viewpoints (Chapter 1). With
these basic elements, we construct the general formalism of the microscopic
dynamics of the lattice gas automaton leading to the microdynamic equations
and the corresponding microscopic properties (Chapter 2). The formalism is then
illustrated through the most used LGA models which are discussed in Chapter 3.
At this point we have all the material necessary to establish the basic statistical
mechanics of the lattice gas leading to the most important mathematical object
of our subject: the lattice Boltzmann equation, and its consequences, the H-
theorem and the equilibrium state (Chapter 4). A discussion of the equilibrium
properties of the lattice gas follows naturally providing interesting analogies
with the static correlation functions of classical statistical mechanics.
The macroscopic dynamics of the lattice gas is developed following two com-
plementary paths. In Chapter 5, we proceed according the classical Chapman-
Enskog method to derive the lattice gas hydrodynamic equations and the ex-
pressions for the transport coefficients. These results show how the connection
can be established between the lattice gas macrodynamics and the equations
of classical fluid dynamics; we also discuss the problem of Galilean invariance,
which calls for special attention. The hydrodynamic regimes which emerge from
the limits of the macrodynamic equations are explored in Chapter 9; in par-
ticular, in the incompressible fluid limit, we recover exactly the Navier-Stokes
equations of classical fluid dynamics.
When we consider small deviations from local equilibrium, we can use the
linearized Boltzmann equation to describe the dynamics of fluctuations, and
from the natural separation into fast and slow variables, we obtain the linearized
hydrodynamic limit (Chapter 6), which sets the stage for the investigation
of hydrodynamic fluctuations whose correlation functions are computed in
Chapter 7. In the hydrodynamic limit, we retrieve the results of the Landau-
Placzek theory, and we show that the lattice gas dynamic structure factor exhibits
the same structure as the power spectrum of density fluctuations in real fluids.
We then consider the lattice gas dynamics beyond the hydrodynamic regime -
that is at shorter wavelengths - where the LGA approach offers the interesting
possibility of exploring all the modes of the system up to the kinetic regime.
These results are also extended to the two-component system (the 'colored' lattice
gas) to incorporate mass diffusion. Larger deviations for systems arbitrarily far
from equilibrium are considered in Chapter 8 where, using a projector technique,
we derive the full non-linear hydrodynamic equations and the general form of
the Green-Kubo expression for the transport coefficients.
Numerical simulations are an important contribution to our subject: they
Preface xix
provide tests of validity of the LGA theory and, in cases such as those illustrated
in Figures 1 and 2, the results of lattice gas simulations can be compared with
laboratory experimental results; but most importantly they can also offer access
- theoretically or numerically - to regimes which cannot be easily reached by
other methods. Chapter 10 presents examples of lattice gas simulations which
are illustrated and discussed for various hydrodynamic flows, along with the
essentials of the procedure for computer implementation. Finally, we close with
an epilogue in the form of a guide for further reading (Chapter 11) indicating
topics and areas of research where lattice gas automata offer, technically or
conceptually, a particularly interesting and useful approach.1
The ideas expressed in this book and the material presented here reflect
our own choices and views. The materialization of these ideas has been influ-
enced in various ways by the interaction, the stimulation and the discussions
we had over the past fifteen years with our colleagues and friends. To all of
them we express our appreciation, with special thanks to David Dab, Matthieu
Ernst, Uriel Frisch, Patrick Grosfils, David Hanon, Michel Henon, Dominique
d'Humieres, Ray Kapral, Pierre Lallemand, Alain Noullez, Dan Rothman, Al-
berto Suarez, Sauro Succi, Olivier Tribel, Jorg Weimar, and Stephane Zaleski.
JPR acknowledges support by the Centre National de la Recherche Scientifique
(CNRS, France) and JPB acknowledges support by the Fonds National de la
Recherche Scientifique (FNRS, Belgium).
1
As much as possible, we avoided disrupting the course of the main text by deferring
mathematical details and technical developments to the appendix.
Chapter 1
Basic ideas
When the lattice gas was introduced in statistical physics around 1985 (see Frisch,
Hasslacher and Pomeau, 1986), it was originally constructed as a physical model
for hydrodynamics. In fact, the concept of the lattice gas is as much a physical
concept - and we shall indeed start with intuitive physical ideas - as it is a
mathematical concept, as a more formal definition can also be given. We first
present the point of view of the physicist (Section 1.1), then we describe the
lattice gas automaton from the mathematical viewpoint (Section 1.2), and in
Section 1.3 we discuss the two aspects.
l.i The physicist's point of view

A lattice gas can be viewed as a simple, fully discrete microscopic model
of a fluid, where fictitious particles reside on a finite region of a regular
Bravais lattice.1 These fictitious particles move at regular time intervals from
node to node, and can be scattered by local collisions according to a node-
independent rule that may be deterministic or non-deterministic. Thus, time,
space coordinates and velocities are discrete at the microscopic scale, that is, at
the scale of particles, lattice nodes and lattice links.
The stationary states in statistical equilibrium and thus the large-scale dynam-
ics of a lattice gas will crucially depend on its conservation properties, that is, on
the quantities preserved by the microscopic evolution rule of the system. Suitable
1
A Bravais lattice is a special case of regular lattice; see Section 1.3.4 for details.
2 1 Basic ideas
conservations are necessary, but not sufficient conditions in order that a lattice
gas exhibit physically realistic collective large-scale dynamics: an appropriate
crystallographic space group for the lattice2 is another necessary condition for
macroscopic dynamics with correct continuous invariances. These points will be
emphasized in later chapters.
Another important feature is the 'exclusion principle': two absolutely indis-
tinguishable particles cannot reside simultaneously on the same node. However,
two particles with different velocities can coexist at a node as well as parti-
cles with the same velocity as far as they can be distinguished by some other
property (mass, color, label, etc.). A consequence of the exclusion principle is
the following: provided there exists only a finite number b of distinguishable
kinds of particles, the occupation state of a node is completely determined by
b Boolean variables, each variable encoding the presence or the absence of a
particle of each kind on the node. This feature is of considerable importance
for the implementation of lattice gases on special-purpose machines as well as
on general purpose computers.
So the physicist would view a lattice gas as a universe of fictitious zero-
dimensional particles undergoing one-dimensional displacements on the links
and colliding on the nodes of a regular Bravais lattice embedded in a D-
dimensional space. But the Boolean encoding of the state of a lattice gas leads
naturally to a mathematically more formal point of view.
1.2 The mathematician's point of view
We shall first introduce the concepts of 'finite automata' and 'cellular automata'.
1.2.1 Finite automata

A 'finite automaton' is a mathematical model of a processor with a finite
number of possible internal states, which evolves and produces output data
according to a rule depending on an input symbol belonging to a finite alphabet
(for more details about finite automata and related notions, see Hopcroft and
Ullman, 1979). Notice that a processor with a finite number of states, whose
evolution depends on several input symbols belonging to several finite alphabets,
also qualifies as a finite automaton, provided the number of inputs is finite.
Indeed, a finite number of input symbols, where each one belongs to a finite
alphabet, can be viewed as a single symbol belonging to a larger, yet finite
2
The 'crystallographic space group' or 'global invariance group' of a Bravais lattice is the
group of all rigid (i.e. distance-preserving) transformations, including translations,
rotations, symmetries and their combinations, that leave the lattice globally unchanged
(see Ashcroft and Mermin, 1976).
1.2 The mathematician s point of view 3
alphabet. The above definition prescribes a deterministic evolution rule for the
internal state of the automaton. Nevertheless, a non-deterministic automaton
also fits the above definition if its non-deterministic evolution rule can be
considered as a deterministic rule depending on a finite number of additional
input symbols belonging to finite alphabets, produced by 'external' (pseudo-)
random generators.
Since the internal state of a finite automaton can change, the automaton
undergoes some kind of evolution, and there must be an underlying notion of
'before' and 'after': a 'past' and a 'future'. However, these primitive notions do
not really imply a temporal structure for the automaton, since the concept of
'time interval between events' is not included in the definition.
1.2.2 Cellular automata

A 'cellular automaton' (see Wolfram, 1986) is a set of synchronized identical
finite automata with identical evolution rules, which exchange data according
to a regular finite connection scheme. Each individual automaton exchanges
information (input and output symbols) with a finite sub-set of automata in
some prescribed neighborhood, according to a connection rule which is the
same for all individual automata. For example, if the automata are indexed by
a signed integer i, the connection rule could be 'automaton number i receives
input symbols from automata number i — 1, i and i -f 1'.
This definition does not imply any geometrical structure. Individual automata
must be labeled, for example by signed integers, in order to define a connection
rule, but the automata are not supposed to be attached to geometrical points of
a metric space; there is no notion of distance or angle contained in the definition.
However, the key notion of homogeneity is already present: homogeneity, in this
context, simply refers to the fact that all individual finite automata are identical
(same internal structure, same evolution rule, and same connection rule).
Another key notion in cellular automata is their synchronization, which im-
plicitly requires that the concept of 'past' and 'future' be universal, that is,
independent of the individual automaton: an event which is in the past (future)
for one automaton must be in the past (future) for all individual automata.
So, there must be a single clock for all individual automata, which justifies the
notion of simultaneity and thus of synchronization.
1.2.3 Lattice gases

From the mathematical point of view, a lattice gas is a particular class of
cellular automata which verify the following additional constraints:
4 1 Basic ideas
(i) The individual automata are tied geometrically to the nodes of a regular
Bravais lattice embedded in a Euclidean space of dimension D. This relation
between individual automata and lattice nodes yields a one-to-one mapping
between the set of all the individual automata of the cellular automaton
and a connected part of the Bravais lattice. Because of this one-to-one
mapping, the individual automata can also be called 'nodes'. The nodes
are labeled by their position vector r*; the star subscript indicates that the
vector takes only discrete values.
(ii) The elementary evolution process of the cellular automaton is repeated at
regularly spaced discrete times separated by a time increment At (the time
step), which is usually taken equal to unity by a proper choice of the time
scale.
(hi) The number of possible internal states of any individual automaton is
2b, where b is an integer. Any possible internal state of an individual
automaton can thus be represented by b Boolean variables, which will
be called 'channels'3 because they correspond to 'communication channels'
between nodes, as we shall see below. The b channels on any node are
labeled by an integer i ranging from 1 to b or from 0 to b — 1. Since all
the individual automata are taken identical, the number b of channels per
node and their labeling are node-independent.
(iv) The elementary evolution process of the cellular automaton that occurs at
each time step is a sequence of two distinct phases:
(a) The collision phase: during this phase, the internal state of each indi-
vidual automaton evolves according to a purely local, node-independent
rule which can be non-reversible and non-deterministic. More precisely,
the post-collision state of an individual automaton depends only on its
own pre-collision state and on afinitenumber of random input symbols
for non-deterministic models (see comments on local versus non-local
collisions in Section 1.3).
(b) The propagation phase: during this phase, the information present in the
automata's 'memory' (the post-collision state) is shifted from node to
node without loss or gain, according to a node-independent, reversible
and deterministic rule. In practice, each channel index i between 0
and b — 1 is associated to a unique lattice vector c,-, such that the
Boolean value of channel i on any node r* is transferred during the
propagation phase to channel i of node r* + c*. The vectors c,- are called
the 'velocity vectors', and must verify the following conditions: (i) the
set of velocity vectors remains globally invariant under the action of the
3
Channels are sometimes called 'cells', but the word 'cell' may be misleading because of
possible confusion with the crystallographic notion of (primitive) cell of a lattice.
1.3 Comments 5
crystallographic point group G of the lattice,4 (ii) they include at least

D generating vectors of the Bravais lattice, that is, D lattice vectors that
generate the whole lattice by linear combinations with signed integer
coefficients.
We shall consider the set of properties listed above as our definition of lattice
5
gases.
1.3 Comments
1.3.1 The velocity vectors
We first justify the name 'velocity vectors'. If the time scale is chosen in such
a way that the time step At is equal to unity, then the information (particle)
present in channel i at node r* goes to node r* + ct in one unit of time. The
CjS can thus be viewed as the microscopic velocity vectors of fictitious particles
occupying channels i.
All velocity vectors are not necessarily distinct: lattice gas models with mul-
tiple channels can be designed to model, for instance, gas mixtures or reaction-
diffusion systems where distinguishable particles representing different 'chemical
species' have the same kinematics (i.e. the same propagation rules). In addition,
some of the velocity vectors can be zero: this is the case for models with rest
particles which stay on a node during the propagation phase.
The velocity vectors, which must have the same local symmetries as the
lattice, are also constrained to include D generating vectors of the Bravais
lattice. This guarantees that any two nodes of the lattice can be connected by
a finite chain of velocity vectors, and thus, that the lattice cannot separate into
sub-lattices with independent evolutions. Note that for a given Bravais lattice,
there may exist several sets of D lattice vectors that generate the whole lattice,
and several sets of velocity vectors compatible with the symmetries of the lattice.
To illustrate this point, take for instance the three-dimensional body-centered
lattice ('bcc' lattice) with unit lattice constant (length of the spatial periodicity).
This Bravais lattice can be generated by the set £f\ containing the three lattice
an
vectors {\>—\,—\),{—\i\i—\)> d (i> i> \)- The smallest set of velocity vectors
that can be built around Sf\ contains eight vectors which can be constructed
by applying all isometries of the local invariance group of the bcc lattice, to
all three vectors of Sf\. These eight vectors with modulus *f have the form
(^15^25^35). Here, (^1,62,^3) are any of the eight possible combinations of'+'
4
The point group or local invariance group of a lattice is the subgroup of the space
group of the lattice which leaves individual nodes unchanged; see Ashcroft and Mermin
(1976).
5
This definition should not be considered as universal: with the same basic ideas, one can
relax, modify or add constraints to construct different definitions of lattice gases.
6 1 Basic ideas
and '—'. Now, the bcc lattice can also be generated by the set &\ containing the
three vectors (1,0,0), (0,1,0) and (5,5,5). The smallest set of velocity vectors
that can be built around ^2 contains fourteen vectors that include the eight
vectors with modulus ^ , plus six vectors with unit modulus: (±1,0,0), (0, ±1,0)
and (0,0, ±1). This example shows that with one single Bravais lattice, one can
construct several lattice gas models with different propagation rules and thus
different complexities. It motivates the introduction of the concept of 'minimal
model' for a given Bravais lattice. A minimal model built on a given Bravais
lattice is a model which has the smallest possible number b of channels per
node compatible with the local symmetries of the lattice. There may be several
minimal models associated to the same Bravais lattice, since they can differ
by their collision rules, but also by the 'physical' properties attached to the
channels.
1.3.2 Space and time

One of the most important features that distinguish lattice gas automata from
cellular automata is that lattice gases have a spatial structure and geometry.
Distances, angles, surfaces, volumes, and thus densities andfluxesare meaningful
quantities which can be defined. The same holds for time: it makes sense to speak
of time intervals for lattice gases. This feature renders possible the comparison
of collective motions in lattice gases with space- and time-dependent behaviors
of physical objects such as real fluids.
1.3.3 The exclusion principle

The constraint that two indistinguishable particles cannot reside at the same
node at the same time (see Section 1.1), has been imposed to enable a natural
Boolean description and encoding of the states of lattice gases. However, models
permitting the coexistence of up to n indistinguishable particles are also admis-
sible, provided this number n is of the form 2P — 1, with p integer. Indeed, such
models can be recast in a way that does not violate the exclusion principle, by
simply using the binary expression of the number of identical particles residing
on a node. Any number of indistinguishable particles less than or equal to
n = 2P — 1 can be viewed as the sum of zero or one single particle, plus zero
or one pair of particles, plus zero or one quadruplet of particles, and so on, up
to zero or one 2p~l-multiplet of particles. To get an intuitive picture, one can
think of a pair of particles as a single particle with double mass, a quadruplet as
a single particle with quadruple mass, etc. These 'super-particles' with multiple
mass are now distinguishable and their presence on a node can thus be fully
described by p Boolean quantities. This picture in terms of multiple particles is
1.3 Comments 7
correct provided the collision phase can 'break' particles with multiple mass in
such a way that it produces no spurious conservations.
Note that n distinguishable particles require n Boolean quantities, whereas n
indistinguishable particles only require p ~ log 2 n Boolean quantities, provided n
is of the form 2P — 1. The problem of distinguishable versus indistinguishable
particles occurs, for example, for some three-dimensional lattice gas models (see
Section 3.7).6
1.3.4 Bravais lattices

Because of the homogeneity principle inherent in the concept of cellular au-
tomata, the connection scheme between the nodes must be node-independent.
As a consequence, the set of velocity vectors of the lattice gas, and thus the
local neighborhood of a node, must be identical for all nodes. The only regular
lattices which verify this condition are, by definition, Bravais lattices. 7
To illustrate the difference between an ordinary regular lattice and a Bra-
vais lattice consider two regular lattices: the triangular lattice with hexagonal
symmetry, and the hexagonal lattice with triangular symmetry (also called the
'honeycomb lattice'). From Figure 1.1, it is clear that the triangular regular lattice
is a Bravais lattice, but the hexagonal one, while regular, is not a Bravais lattice.
Figure 1.1 (a) An example of a two-dimensional Bravais lattice: the triangular

lattice, (b) An example of a two-dimensional regular but not Bravais lattice:
the hexagonal 'honeycomb' lattice. For both lattices, the arrows represent the
neighborhood of a node. The hexagonal lattice has two different kinds of
neighborhoods, depending on the parity of the node number.
6
One of these three-dimensional models, the FCHC-8 model, admits up to three
indistinguishable particles with zero-velocity (rest particles). Only two Boolean quantities
are used for them, one encoding the presence of a particle of mass 1, the other encoding
the presence of a particle with mass 2 (that is, a pair of particles of mass 1).
7
Detailed definitions and discussions of Bravais lattices are given in books on solid state
physics; see for example Ashcroft and Mermin (1976).
8 1 Basic ideas
Indeed, for the latter, the set of nearest neighbors is different for 'odd' and 'even'
nodes. No node-independent connection scheme involving nearest neighbors can
be designed on the honeycomb lattice.
One can classify exhaustively and enumerate all possible Bravais lattices in
given dimensions, according to their symmetry groups: there exist only five
different Bravais lattices in two dimensions, and fourteen in three dimensions,
as first observed by Auguste Bravais in 1845 (see Cauchy, 1851). Notice that the
constraint of using a Bravais lattice is not too severely restrictive, since several
lattice gas models can be built on one single Bravais lattice by changing the
connection scheme and the collision rules. Moreover, discrete models within the
spirit of lattice gases, but involving non-regular or non-Bravais or quasi-lattices
can exist, but will not be considered here, as we keep with the definition of
lattice gases given in Section 1.2.3.
1.3.5 Local versus n o n - l o c a l collisions

According to the definition of lattice gases given in Section 1.2.3, the collision
phase should be local, that is, the post-collision state at any node r* depends
only on the pre-collision state at the same node r*. This constraint seems
to preclude all models with finite-range interactions which require non-local
collisions (see for example Appert and Zaleski, 1990). Such models can actually
fit the definition of Section 1.2.3. Indeed, as shown in the Appendix, Section A.I,
non-local interaction models can be 're-coded' using a larger number of channels
per node, in such a way that the evolution can be split into a more complex
propagation phase and a more complex but local collision phase. This holds
provided the neighborhood with which a node interacts during the non-local
collision phase is globally invariant under the point-group of the lattice. This
re-coding is neither very natural nor recommended for practical purposes, but
it shows that the definition of Section 1.2.3 is less restrictive than it appears.
1.3.6 Collision-propagation versus propagation-collision

In Section 1.2.3, the evolution sequence is presented as a collision phase followed
by a propagation phase. This order was chosen by anticipating its convenience
for writing the microdynamic equation (see Chapter 2). In fact, this choice is
arbitrary and not restrictive: models with the reverse order can be re-coded
to fit the chosen order. This re-coding is operationally similar to the re-coding
discussed above in Section 1.3.5.
In general, the order in the sequence is unimportant when long-time behaviors
and large-scale properties are considered. However, if one considers fluctuating
processes which are measured and analyzed via fluctuation correlations, the
order in the sequence is not irrelevant: the result of the correlation measurements
1.3 Comments 9
may depend on the phase where the iteration sequence is stopped (see Chapter 4).
Care must be taken in using proper averaging procedure.
1.3.7 Mathematical versus physical

The mathematical point of view, where lattice gases are considered as a special
class of cellular automata,8 is most useful when a formal definition is needed.
But this more abstract view may result in some loss in physical intuition.
For example, the crucial role of the conservation laws appears more clearly
by analogy with usual statistical mechanics when considering the lattice gas
as a gas of point particles. The picture which shows particles moving and
colliding, rather than information exchanges and local evolution, often yields
physical insight. We shall use both languages, but with some preference for the
physicist's language, as we want to put the emphasis on the physical properties
of lattice gases.
8
A different point of view, discrediting the concept of cellular automata for lattice gases,
is presented by Henon (Henon, 1989).
Chapter 2
Microdynamics: general formalism
In this chapter, we develop the 'microdynamic formalism', which describes the

instantaneous microscopic configuration of a lattice gas and its discrete-time
evolution. This exact description of the microscopic structure of a lattice gas is
the basis for all further theoretical developments, in particular for the prediction
of large-scale continuum-like behavior of LGAs.
We first introduce the basic tools and concepts for a general instantaneous
description of the microscopic configurations (Section 2.1). The time evolution
of the lattice gas is then given in terms of the 'microdynamic equations' (Sec-
tion 2.2). Thereafter, we define microscopic characteristics (e.g. various forms
of reversibility), which have a crucial incidence on the macroscopic behavior of
the gas, and therefore on its suitability to simulate real physical situations (Sec-
tion 2.3). The last section is devoted to special rules needed to handle boundary
problems (obstacles, particle injections, etc.).
This chapter deals with rather abstract concepts which will find their appli-
cation in Chapter 3, where lattice gas models are described at the microscopic
level.
2.1 Basic concepts and notation

2.1.1 The lattice and the velocity vectors
One of the most important features of lattice gases is the underlying Bravais
lattice structure which gives a geometrical support to the abstract notion of
10
a cellular automaton. Strictly, a Bravais lattice is by definition infinite. We

consider that the cellular automaton only occupies a connected subset if of the
D-dimensional underlying Bravais lattice. In practical applications, this subset
contains only a finite number Jf of lattice nodes. However, for theoretical
developments, if will be assumed to be infinite when necessary.
We label any given node of the Bravais lattice by its position vector r* with
respect to some arbitrary origin. The vectors r* belong to the D-dimensional
space R D in which the Bravais lattice is embedded. The subscript '*' will
be systematically used to distinguish discrete-valued quantities such as discrete
positions (r*) or discrete times (t*) from the corresponding continuous quantities.
The components of position vectors, and more generally of any D-dimensional
vector will be labeled by Greek indices ranging from 1 to D. For example, the
components of the position vector r* are denoted r*a, a = 1,..., D.
To each node of if corresponds a finite automaton (see Section 1.2); so the
cellular automaton contains Jf identical automata, and each of them has a
finite number of internal states coded with b binary digits. In more physical
terms, each node, i.e. each individual automaton, can be viewed as a set of
b 'channels' each of which can be 'occupied' by a fictitious particle (binary
value 1) or 'unoccupied' (binary value 0). The channels and, more generally, any
quantity related to channels will be labeled with a Roman index ranging from
1 to b. For example, the velocity vector corresponding to channel i is denoted
Ci, i = 1,..., b. For convenience, the channel index can also run from 0 to b — 1
(e.g. for models with one rest particle, see Section 3.3).
From the standard definition of a finite automaton, the internal state evolves
according to a rule that depends on some external inputs. In the case of a
lattice gas, the external input data come from a fixed finite number B (B < b)
of neighboring nodes, and, in the case of non-deterministic models, from a
certain number of random variable generators associated to each automaton.
The number B, which is called the 'connectivity' or 'connection number' of the
lattice gas model, is the maximum number of different non-zero vectors in the
set of b velocity vectors c* (see Section 1.2.3). These B distinct non-zero vectors
are called the 'connection vectors' of the model.
If the evolution rule involves exchanges of information with all nearest neigh-
bors, and only those, then the connectivity B of the lattice gas model is identical
to the coordination number of the Bravais lattice, as defined in solid state physics
(see e.g. Ashcroft and Mermin, 1976). If, in addition, all the velocity vectors are
non-zero, then the model is said to be 'homokinetic', all velocity vectors having
the same modulus.
Notice that the connectivity B of a lattice gas model may be equal to,
or may differ from, the number b of channels per node; it depends on
whether all c,s are identical and non-zero, or not (see the models described in
Chapter 3).
12 2 Microdynamics: general formalism
2.1.2 The Boolean field

Each channel has only two possible states: 0 or 1, which means physically that
each channel is either empty or occupied by at most one particle (exclusion
principle). For any given node, we denote by s = (st, i = 1,..., b) the b-bit word
that describes the instantaneous internal state of the individual automaton, or,
equivalently, the occupation of the b channels of the node. This fo-bit word is
called the 'Boolean state' of the node, and the set y of the 2b possible Boolean
states of one node defines what we call the 'single-node phase space' or the
'local phase space'.
Similarly, the instantaneous state of the whole lattice will be encoded by
a collection of Jf b-bit words, one word for each node. This collection of
Jf xb Boolean quantities S = (Sj(r*)5 i = 1,..., b, r* G if) is called the 'Boolean
configuration' of the lattice, and the set T of all possible Boolean configurations
defines the 'global phase space' or simply 'phase space'.
In order to simplify the algebraic treatment of Boolean configurations, we
use algebraic binary quantities whose values are ' 1 ' or '0', instead of logical
quantities with values 'TRUE' or 'FALSE'. The equation governing the time
evolution of the lattice will then be an algebraic relation with the same content
as the corresponding logical equation.
For any discrete time t*, we define the 'Boolean field' n as the function
mapping the set of all channels of all nodes onto the set {0,1}, associating with
each channel of each node, its occupation level (0 or 1):
n: JSPx{l fc} - {0,1}

(r*,i) •-> n,(r*),
where rii(r*) is the occupation of channel i at node r*. There is a semantic

distinction between the Boolean configuration Sj(r*) and the Boolean field
n,-(r*): the Boolean field is a generic notion, whereas a Boolean configuration is
a particular realization of the Boolean field.
At this stage, the Boolean field is a function of only two discrete-valued
variables: the channel index i and the node position r*. For the dynamic
description to be presented in Section 2.2, it is useful to include a third variable:
the discrete time £•, and so to work with the time-dependent Boolean field
2.1.3 Observables
In order to establish the connection between purely abstract automata and
concrete gas models, physical properties (mass, momentum, energy, etc.) must
be associated with each channel of each node of the lattice. We emphasize
that such physical quantities (observables) are not associated with the particles
themselves, but with the channels: particles loose their identity during collisions,
whereas channels do not. Nevertheless, for simplicity and intuition, we shall use
terms such as 'the mass of particle f or 'the momentum of particle f rather
than 'the mass associated with channel number i if occupied' or 'the momentum
associated with channel number i if occupied'. Mass, momentum and energy are
examples of observables with physical content; yet, the concept of observable, as
we define it below, is in principle not restricted to quantities with an equivalent
in the physical world.
It is systematically assumed that the amount of a given observable assigned
to a channel depends only on the channel index, and not on the position of the
node: if occupied, channel i of any node always carries the same amount of the
observable.
Under the above assumption, an observable is defined as a collection of
b values (qi9 i = l,...,fo), indicating the amount of the underlying observable
quantity which is present in the channel labeled by the index i9 when this channel
is occupied. The qfi can be scalars, vectors or more generally tensors of arbitrary
order.
The 'value of the observable measured at node r*' is the total amount of the
observable quantity present at node r*: X^=i 4ini(r*)- It is called the 'microscopic
density per node' of the observable or simply its 'microscopic density', and is
denoted by N*(r*) in the most general case:
1=1
where the superscript '*' indicates that it is a non-averaged microscopic quantity.

Now, when divided by t>o, the volume 1 of the primitive unit cell of the Bravais
lattice, this density becomes a 'microscopic density per unit volume', and when
divided by b, the number of channels per node, it becomes a 'microscopic density
per channel'. Similarly, the 'microscopic flux' of the observable q is defined by
Ensemble- and/or space- and/or time-averaging of the microscopic densities

N* and of the microscopic fluxes <I>* yields the 'macroscopic densities' Nq
and 'macroscopic fluxes' Q>q (without '*' superscript). These 'densities' and
'fluxes' (without further specification) are the physically relevant fields in the
macroscopic theory of lattice gases (see Chapters 4 and 5). In order to shed some
light on the above definitions, we now give examples of some useful observables,
frequently encountered in lattice gas theory.
1
Length, surface, volume or hyper-volume, depending on the dimensionality D.
2.1.3.1 Example 1
The collection g = ( l , l , l , . . . , l ) o f b quantities equal to 1 is a scalar observable
called the 'particle number observable', as its value measured at node r* is
J2 n,-(r*) which is clearly the total number of particles present at node r*.
2.1.3.2 Example 2
Denoting by m,- the individual mass assigned to any particle present in channel i,
the collection q = (mi,..., nib) is a scalar observable called the 'mass observable'.
Its value ^ m ;fy( r *)> measured at node r*, is the total mass present at node r*.
If the individual masses m,-, (i = 1,..., b) are all equal to a single value m (which
can be taken equal to 1 by a proper choice of the mass scale), then the mass
observable is equal to the particle number observable. This is the most frequent
case.
2.1.3.3 Example 3
Denoting by pia((x = 1,...,D) the components of the individual momentum as-
signed to any particle present in channel i, the collection qa = (pi a ,... ,p&a), (a =
1,...,D) is a vector observable called the 'momentum observable'. Its value
qa = 5^Pia»i(r*), 2 measured at node r*, is the total momentum present at node
r*. For many models, the individual momentum p, of channel i is chosen equal
to niiCi, which is physically consistent with the interpretation of the C;S as mi-
croscopic velocities. This choice also guarantees that the microscopic density of
momentum is just the microscopic flux of mass.
2.1.3.4 Example 4
Denoting by e\ the individual energy assigned to any particle present in channel i,
the collection q = (e\,..., e&) is a scalar observable called the 'energy observable'.
Its value J2eini(r*)> measured at node r*, is the total energy present at node r*.
If particles carry only kinetic energy, it is reasonable to take e\ equal to p?/(2mj).
2.1.4 Generalized observables

The observables, as defined above, are linear, local and homogeneous since their
value, measured at node r*, is a linear combination of the n^ at the same node,
with node-independent coefficients (the qts). Relaxing these constraints, we can
build the abstract notion of 'generalized observable'. A generalized observable q
is any physical quantity whose value measured at node r* is a function q(n9r*)
of the Boolean field n and of the space variable r*.
When the value g(n,r*) of a generalized observable q, measured at node r*,
2
Remember that the Greek subscript a refers to the components of the vectors, and the
Roman subscript i refers to the channel index.
does not explicitly3 depend on r*, then the generalized observable is said to be
'homogeneous'. When the value of the observable measured at node r* only
depends on thefy-sat this very node r*, then the generalized observable is said
to be 'local'.
The following simple examples should provide a more physical picture of
generalized observables.
2.1 A.I Example 1: space-averaged quantities

Let / b e a connected subset of <£. Consider the generalized observable whose
value, measured at node r* is:
«(n,r*)= X > ( r * + '*')• (2.4)

r*'eJ
This generalized observable is linear and homogeneous, but non-local, since its
value, measured at node r*, depends on the Boolean field at nodes r* + r*', with
x* e «/. The resulting value can be understood as the population of channel
number 1, space-averaged over the domain J surrounding node r*.
2.1.4.2 Example 2: potential energy

Suppose the fictitious particles are in interaction with a space-dependent force
field that derives from a scalar potential t/(r*). The potential energy measured
at node r* is a generalized observable whose value is:
This generalized observable is linear and local, but non-homogeneous, since its
value depends explicitly on r* through the scalar potential U(u).
2.2 The microdynamic equation
The cellular automaton rule governing the evolution4 of a lattice gas can be
viewed as a logical equation expressing the Boolean field at time U + 1 from
the Boolean field at time U. When written in algebraic form, the resulting
equation is called the 'microdynamic equation' of the lattice gas, as it governs
the microscopic evolution (microdynamics) of the lattice (Frisch, d'Humieres,
Hasslacher, Lallemand, Pomeau and Rivet, 1987). The microdynamic equation
is an exact equation which contains the complete dynamics of the lattice gas.
3
The value of q depends implicitly on the space variable r*, through the dependence in
the Boolean field.
4
The time scale is chosen such that the time increment of the cellular automaton is one.
2.2.1 Formal expression

The cellular automaton rule can be fully deterministic or not. For the deter-
ministic case, the Boolean field at time U + 1 is a well-defined function of the
Boolean field at time U. This function, mapping T (the cellular automaton phase
space) onto itself, is the 'evolution operator' S\
n{U + l) = £(n(U)). (2.6)
For the non-deterministic case, the transition between times U and U + l involves
random choices at each node. The microdynamic equation then reads:
n(U + l) = £(n(t*),Z(t*))9 (2.7)
where £(£*) is a collection of Jf random functions of the discrete time variable

U (one function for each node).
This formalism is quite general and involves few assumptions about the
discrete gas model. In the following, we give a more explicit expression for the
evolution operator, using the fundamental property that the evolution rule of a
lattice gas, as defined in Section 1.2, is a sequence of two phases: (i) a collision
phase during which the Boolean state of each node is modified locally (that
is, without exchanging any information between nodes), and (ii) a propagation
phase during which different nodes can exchange information in a deterministic
and reversible way. The evolution operator $ can thus be written as the product
9^ o ^ where the 'collision operator' ^ is local but may be non-deterministic
and non-invertible, and the 'propagation operator' or 'streaming operator' Sf is
non-local, but deterministic and invertible.
2.2.2 The propagation operator

According to the definition of the propagation phase given in Section 1.2, the
action of the propagation operator on the Boolean field is such that the state
of channel i at node r* — c, before propagation becomes, after propagation, the
state of channel i at node r*. The propagation operator carries the information
from channel i of node r* to channel i of node r* + c,. So the expression for the
propagation operator reads:
y(n).(r*) = ni(r*-ci). (2.8)
This definition of the propagation operator raises the problem of finite size
lattices. Indeed, if the subset <£ of the Bravais lattice under consideration is
finite, the node r* + c, may be outside j£f, even if the node r* is inside.
One solution to this problem is to introduce a periodic wrapping in such a
way that any particle driven out of S£ be re-injected elsewhere in if. More
precisely, we need a finite domain if' of the underlying Bravais lattice which
contains all nodes of <£ and which can perfectly tile the full Bravais lattice
by successive lattice-preserving translations. So there must exist a subgroup of
the lattice-preserving translations group such that the nodes 5 in $£' and their
images by the elements of this subgroup cover the full infinite Bravais lattice
without any gap or overlap (see Figure 2.1 (a)). It is then possible to wrap $£'
on itself and to 'connect' opposite boundaries in such a way that j§?' become a
topological torus (see Figure 2.1(b)). In terms of particle motions, this wrapping
leads to periodic boundary conditions6 at the boundaries of if'.
Tiling with U and its images
Periodic
boundary conditions
Figure 2.1 The construction of periodic boundary conditions is illustrated for

the two-dimensional triangular lattice, (a) Starting from a lattice domain j£?
with arbitrary shape, one constructs <£' which tiles perfectly the full Bravais
lattice. This tiling is realized by applying to <£' all the translations generated
by the two lattice-preserving translations with vectors A and B. (b) The sides
of $£' are reconnected in such a way that it becomes a topological torus.
5
Note that 5£ and <£' must be thought of as sets of lattice nodes and not as geometrical
surfaces or volumes.
6
In solid state physics, periodic boundary conditions are also called 'Born-von Karman
boundary conditions'.
For a given lattice gas model and a given domain j£f, there can exist several
different periodic boundary conditions. Indeed, the same domain <£ can lead to
several perfect tilings with different tile shapes, even if one restricts the choice
to tiles which contain a minimal number of nodes (see Figure 2.2(a)). Moreover,
the same tile shape can lead to several different tilings with different translation
subgroups (see Figure 2.2(b)).
We emphasize that this periodic wrapping is applied only to the propagation
operation; it does not imply that the macroscopic variables satisfy periodic con-
ditions in the usual sense. Indeed, special collision operators can be applied to
the boundary nodes in such a way that the boundaries of <£ behave macro-
scopically as solid walls, or as mass and/or momentum injectors for example.
These special collision operators can be applied as well to the boundaries of
$£ to produce particular macroscopic boundary conditions, and/or to localized
parts inside j£?, to simulate for instance fixed solid obstacles of almost any
arbitrary shape. Some of the most useful special collision rules are described
and discussed in Section 2.4.
2.2.3 The collision operator

To obtain a closed expression for the collision operator, we use the property
that the collision rule is identical for all nodes (although only statistically in
the non-deterministic case). So we can introduce a unique 'collision matrix'
A(s->s') for (5, sf) e y2, that contains complete information on the collision rule.
Each matrix element A[s^sf) of this 2b x 2b square matrix is the transition
probability from the Boolean state s to the Boolean state sf. Interpreted as
Figure 2.2 The same domain <£ can lead to several different tilings with (a)
different tile shapes and/or (b) different translation subgroups. The
two-dimensional triangular Bravais lattice is used to illustrate the point.
transition probabilities, the matrix elements are real numbers between zero and
one, and satisfy the normalization condition:
A(s^) = l9 Vsey. (2.9)
s'ey
Under the above assumption, one can write the collision operator in an explicit
form that involves the collision matrix indirectly:
«(n),(r*)= Y, # M ) ] I " / V ' (2.10)

(s,s>)ey2 ;=1
where Sj = 1 — Sj and H;- = 1 — rij. The rijS and n ; s are taken at node r*.
The £(s^>s') s are the elements of a 2b x 2 b square matrix of Bernoulli random
variables (values 0 or 1) whose averages are the A(s-+sf)s, and such that
J2Z{s-s') = l Vsey. (2.11)

s'ey
Equation (2.10) expresses, in a generic way, the post-collision Boolean field

^(n) as an explicit function of the pre-collision Boolean field n. It is an analytic
transcription of the cellular automaton collision rule. There are two important
points which justify the analytic structure of (2.10):
(i) For a given statistical realization of the Bernoulli variables £(s->s'), and for
a given Boolean state s, there exists one and only one Boolean state sf such
that £(s->s') = 1, all £(s->s")s for s" ^= s' being zero. This is a consequence
of (2.11) and of the fact that the £(s->s')s are Bernoulli random variables.
So, the t;(s^>sf) factor on the r.h.s. of (2.10) selects, among all possible pairs
of Boolean states (5, s'), those which are pre- and post-collision states of the
collision that takes place.
(ii) The product Yin/m/i i s a 'matching operator' which is one if n/(r*) = Sj
(for all j = 1,..., b), and zero otherwise. Operationally, it is equivalent to a
'Kronecker symbol' <5fcr(n(r*) — s). So, it selects, among all possible Boolean
states s, the one which is the actual value of the pre-collision Boolean field
n at node r*.
As an obvious consequence, the product
selects, among all possible pairs of Boolean states (5, s'\ the one for which s is
the pre-collision state of node r*, and s' is the actually realized post-collision
state. Equation (2.10) assigns to ^(n)(r*) the value of this selected post-collision
state s'.
The above formalism is applicable to both deterministic and non-deterministic
models. For deterministic models, the matrix elements A(s^>sf) are either 0 or 1,
and the £(s->s')s are deterministic and thus simply equal to the A(s^>s')s.
2.2.4 Analytic expressions of the microdynamic equation

Using the expressions (2.8) and (2.10) for the propagation and collision operators,
the complete microdynamic equation can be written as:
(s,s')ey2 ;=1
where the rijS on the r.h.s. are implicitly taken at node r* and at time U. An
alternative form, strictly equivalent to (2.12), reads:
b
m(u + c, 9t* + i) = m+ J2 ( i ~ M( ^) II n / V ' -s S S

< 2 - 13 )
(s,s')ey2 7=1
The equivalence between (2.12) and (2.13) follows immediately from the identity:
b
(s,s')ey2 7=1
which is a consequence of (2.11). The virtue of the alternative form (2.13) is that
it distinguishes 'effective collisions' with s' =/= s from 'passive collisions' which
leave the Boolean state unchanged.
The microdynamic equation plays for lattice gases the role of Hamilton's
equations for a system of interacting particles; it involves no approximation and
constitutes the starting point of all further theoretical developments.
2.3 Microscopic properties of a lattice gas
As for real gases, the macroscopic (large-scale) dynamics of lattice gases depends
crucially on their microscopic properties. We now define some of the most
important microscopic characteristics that a given lattice gas model may or may
not possess. The consequences of these microscopic characteristic properties on
the large-scale macroscopic behavior will be discussed in Chapters 4 and 5.
2.3.1 Detailed and semi-detailed balance

The collision process is said to be 'micro-reversible' if any collision has the same
probability as the reverse collision, that is, if and only if :
V(s, s') € y\ A(s-+tf) = A(s^s). (2.15)
This symmetry condition for the matrix ^4(s-»s') is usually called 'detailed bal-
ance'. A weaker form, called 'semi-detailed balance', reads:
sey sey
and may be viewed as a state-averaged form of the detailed balance (2.15).

The physical interpretation of semi-detailed balance is that for any state s' 9
the 'number of collisions'7 with output state sf is the same as the number of
collisions with input state s': when semi-detailed balance is satisfied, collisions
'create' any given state sr statistically as often as they 'destroy' it.
Using the normalization relation (2.9), semi-detailed balance can be written
as:
A(s^s') = l (2.17)
sey
which is the most commonly used expression for the semi-detailed balance
condition.
The semi-detailed balance will be crucial in deriving the macroscopic theory
of lattice gases. In particular, it affects deeply the nature of equilibrium states
and the validity of the Boltzmann approximation.
2.3.2 Duality
For any Boolean state s, the dual state s is obtained by exchanging particles and
'holes' (occupied and unoccupied channels) or Os and Is in s. If s is considered
as a collection of b logical variables, then the dual state is s, the logical bitwise
negation of s.
A lattice gas model is said to be 'self-dual' if the dynamics of holes is, at least
statistically, identical to the dynamics of particles. The propagation operator
moves information without any loss or gain, and is thus self-dual. Consequently,
a necessary and sufficient condition for a model to be self-dual is that the
collision matrix A(s^sf) satisfy the relation:
V(s, s') e y2, A(s->sf) = A(s->sf). (2.18)
2.3.3 Conservation laws

As we emphasized earlier, it is crucial to know the set of the configuration-
dependent quantities which are conserved by the evolution operator, for any
initial configuration of the lattice. Indeed, the nature and the number of these
conserved quantities have a direct influence on the existence and the nature
of statistical equilibrium states and on the large-scale collective behavior of
the lattice gas. Therefore, the conserved quantities are determinant elements for
the validity of the lattice gas as a model system for the simulation of physical
phenomena.
The accurate description of the conservation laws in a lattice gas model calls
7
weighted by their probability.
on the notions of 'collisional invariants' and of 'geometrical invariants', where

one distinguishes 'static' and 'moving' geometrical invariants (d'Humieres, Qian
and Lallemand, 1989, 1990). We now introduce these notions.
2.3.3.1 Collisional invariant

A collisional invariant, also called 'microscopic invariant' or 'local invariant' is
an observable quantity whose value, measured at a node in any state, remains
unchanged under the action of the sole collision operator. In other words, if
an observable is a collisional invariant, the transition probability between two
Boolean states corresponding to different values of the observable must be zero.
So, the observable q is a collisional invariant if, and only if, the collision matrix
A(s-*sf) satisfies the condition:
5, s') e y\
Of course, during the propagation phase, the value at individual nodes of an

observable which is a collisional invariant can change, since it is redistributed
over neighboring nodes. However, the sum over the whole lattice is conserved if
the standard collision operator is applied everywhere (no obstacles, no mass or
momentum creation or annihilation).
Now, considering the complete evolution operator, we express that an observ-
able q is a collisional invariant, in terms of the Boolean field:
b b
Y, QMT* + c-, u +1) = ]T qini(u, u). (2.20)
This equation, which must hold for all discrete times U and for all nodes in jSf,
expresses that the amount of the collisional invariant q present at time U at node
r* is conserved by the collision operator, and distributed over the b neighboring
nodes r*+c; by the propagation operator.8 Equation (2.20) displaying collisional
invariance will be explicitly used as a starting point for the macroscopic theory
of lattice gases.
2.3.3.2 Static geometrical invariant

A static geometrical invariant is a triplet (q,J>,k) where q is an observable, J
a sub-lattice of <£ and k an integer independent of the size of =£?, such that
for any initial configuration of the lattice, the value of the observable, summed
8
If a special treatment (reflection, injection, forcing, etc., see Section 2.4) is applied to the
boundaries or to localized parts of if, then Equation (2.20) holds only in the bulk of if.
over all nodes in J remains unchanged after k successive applications of the

full evolution operator. In terms of the Boolean field, this statement reads:
b b
For instance, the triplet formed by any collisional invariant, the whole lattice
J&f, and any integer k is a static geometrical invariant, since the propagation
operator only redistributes the total amount of the observable initially present in
JSf, without any loss or gain. Of course, this only holds when periodic conditions
are applied (with no boundaries elsewhere).
2.3.3.3 Moving geometrical invariant
In order to define a moving geometrical invariant, we need a triplet (^,«/,T),
where q is an observable, J> a sub-lattice of JSf, and T a lattice-preserving
translation which does not leave J globally invariant. It is then assumed that,
for any initial configuration of the lattice, the amount of observable present in
J be transferred completely to T(«/). Re-expressed in terms of the Boolean field,
this reads:
b b
2^(r*>U + 1
)=EE^(r*,**). (2.22)
2.3.3.4 Conserved quantities
A conserved quantity is simply a static geometrical invariant with k = 1. It is a
doublet formed by an observable q and a sub-lattice J> of if, such that the value
of the observable summed over all nodes in J remains unchanged under the
action of the full evolution operator, for any initial configuration of the lattice.
A collisional invariant, summed over the whole lattice is always a conserved
quantity (except if special conditions are applied, e.g. local injection or removal
of particles, biased velocity field).
2.3.4 G-invariance
The property of G-invariance is quite intuitive. It characterizes lattice gas mod-
els whose microscopic structure (propagation rules, collision rules, observables)
is compatible with the local symmetries of their underlying Bravais lattice.
G-invariant lattice gases are those whose physical structure has the same sym-
metries as the Bravais lattice.
2.3.4.1 Definition
A lattice gas model is said to be G-invariant whenever, for any isometry g in the
crystallographic point group (local invariance group) G of the Bravais lattice,
there exists a unique index permutation g acting on {1,2,...,b} such that:
the set of velocity vectors c I? i = 1 , . . . , b of the model satisfies:
g(Cf) = Cg(l> f = l,...,ft,

c
any relevant observable g,, i = 1 , . . . , b of the model satisfies:
g t e ) = 4g(/)> i = 1,...,&,
the collision matrix satisfies:
A(s^s'), V(s, s') e y\

where ^ is the permutation acting on y, such that
&(s)i = sg-i (l> V(s) G 7, 1 = 1,..., b.

Note that the second condition must hold for all relevant observables, whether
they are scalars, vectors, or higher order tensors. This definition of the G-
invariance gives a formal frame to the intuitive notion of compatibility between
the lattice symmetries and the 'physical' microscopic properties of the lattice
gas.
2.3.4.2 The classes of channels

When G-invariance is satisfied, or at least when its restricted version omitting
the third condition about collisions is satisfied, then it becomes possible to define
an equivalence relation within the indices i = l,...,b, leading to a classification
of the b channels. Indeed, consider that channel i and channel j are equivalent
whenever there exists an index permutation g associated to an isometry g in G
such that j = g(i). This relation is clearly an equivalence relation. This comes
directly from the fundamental properties of the group structure of G. This
equivalence relation naturally leads to a partition of the set of all channels into
one or more classes of equivalent channels. Conventionally, each class is labeled
by an integer index that will be denoted (when needed) by capital letters such
as / or J. The symbol C/(o denotes the index of the class to which channel i
belongs. In the case of the most simple lattice gas models, only one class exists,
and this notation is not necessary.
2.3.4.3 Examples of models with one or several classes

To illustrate the notion of class, we use the simple geometry of the two-
dimensional square lattice. The point group G related to this lattice is usually
denoted ' C ^ ' according to the standard Schoenflies convention (see Eliott and
Dawber, 1979). It contains the symmetries Sy, Sx-y, SX9 Sx+y with respect to
the directions (1,0), (1,1), (0,1), (-1,1) and the rotations RQ, Rn/i, Rn, #371/2
around an axis normal to the plane of the lattice. We present two possible
sets of velocity vectors compatible with the symmetries of this lattice. The first
set, defined in Figure 2.3(a), corresponds to the simple HPP model which will
be described in Chapter 3. It contains only four channels with four different

velocity vectors ci to C4. Each of these vectors has unit modulus and corresponds
to one of the four natural directions of the lattice. In this simple case, there
exists only one class, since any of these four vectors can be changed into any
other by an isometry of G. The second set of velocity vectors corresponds to an
extended square model as shown in Figure 2.3(b). It contains 24 channels with
24 different velocity vectors. These 24 channels can be collected into five classes
of equivalent channels, as listed in Table 2.3.4.3. It is clear that if a lattice gas
model is G-invariant, any associated scalar observable must have the same value
for all channels in the same class.
We must keep in mind that G-invariance is a microscopic property, that does
c15 CM
c21
c2
C22
\ L/ / c,
1\ clf / c, c
C}
c,
Cu
—^ —^ 9
\1
c4 c17
< ^
> c 2l
/ t
\ C2<
c23 c18 Cn cn
Figure 2.3 (a) A lattice gas model (the HPP model) with only one class of
channels, (b) A more complex model with five classes.
Table 2.1 The classes of channels for the extended

HPP model of Figure 2.3. The first column gives the
class index /. The second column describes the
channels in each class. The third and fourth columns
give the number of elements in each class and the
moduli of the corresponding vectors respectively.
Class index Channels No. of elements Modulus

1 1 to 4 4 1
2 5 to 8 4
3 9 to 12 4 2
4 13 to 20 8 V5
5 21 to 24 4 2V2
not imply directly the isotropy of macroscopic collective (large-scale) dynamics.

An example is the HPP model (see Section 3.1), which is G-invariant, but whose
large-scale dynamics is not isotropic.
2.3.5 Crystallographic isotropy

In lattice gases, the 'symmetries' (invariance properties not restricted to mathe-
matical symmetries) of large-scale dynamics depend crucially on the microscopic
symmetries of the underlying lattice. Although the microscopic structure of lat-
tice gases only displays discrete symmetries, the resulting macroscopic behavior
may possess continuous invariances. Obviously, a lattice cannot have the full
isotropy of the continuum since it has preferred directions. Still, the large-scale
dynamics of a lattice gas may 'not feel' these preferred directions and be fully
isotropic (at least to leading order), if suitable microscopic conditions are fulfilled
by the lattice geometry and the collision rules.
There is a simple method to find whether a given Bravais lattice with a
given set of velocity vectors is likely to be a natural candidate for a lattice
gas with isotropic large-scale properties. To describe the method, we need some
machinery that we shall first introduce (note beforehand that it provides neither
a necessary nor a sufficient condition for isotropy).
The method amounts to computing up to which order n the set of velocity
vectors c,- satisfies a geometrical microscopic property called the 'n th order
crystallographic isotropy'. Crystallographic isotropy (to be defined hereafter), is a
purely geometric, microscopic concept. It must not be confused with macroscopic
(full) isotropy, which is related to the absence of preferred directions in large-
scale dynamics.
That the crystallographic isotropy property be satisfied up to the desired or-
der is not sufficient to guarantee macroscopic isotropy for large-scale dynamics:
indeed, non-G-invariant collision rules can introduce some 'artificial' anisotropic
contributions to the large-scale equations of motion. Moreover, the requirement
of crystallographic isotropy may even be bypassed as a proper bias in the colli-
sion rule may compensate for the lack of crystallographic isotropy. However, for
simple G-invariant homokinetic models with collisions satisfying semi-detailed
balance, crystallographic isotropy up to desired order will be a necessary and
sufficient condition for macroscopic isotropy. This point will be discussed in
Chapter 5.
In order to describe the crystallographic isotropy property, we first define
some geometrical properties of lattices and tensors.
2.3.5.1 Isotropic tensors

A tensor of order n defined in the D-dimensional Euclidean space R D is said
to be 'isotropic' if, and only if, it remains invariant under the action of any
element of the orthogonal group Op in dimension D, that is, the group of all
distance-preserving linear transformations acting on R2* (see Tung, 1985). There
exists an equivalent and more intuitive criterion to characterize the isotropy of
a tensor: a tensor of order n is isotropic if, and only if, its contraction with
n arbitrary 'test' vectors of R^ can be expressed exclusively in terms of inner
products of these vectors. With this criterion, it is obvious that odd-order tensors
cannot be isotropic, unless they are null. It is also easy to convince oneself that
a second order tensor T aia2 is isotropic if, and only if, it is a multiple of the unit
tensor 5aiai (Kronecker symbol). In the same way, a fourth order isotropic tensor
T^WWA must be a linear combination of 5^3^, <5aia3(5a2a4, and <5aiaA2a3- More
generally, an nth rank isotropic tensor (with n even) must be a linear combination
of all products of the form <5aff(1)a<r(2)... <5a<y(fI_1)a<r(B), where a is any permutation on
(1,...,*).
If, in addition, the tensor is fully symmetric (i.e. invariant under any per-
mutation of its indices), then there exists a second isotropy criterion which is
equivalent to the previous one, but technically much simpler to use: A fully
symmetric nth order tensor T (n) is isotropic if, and only if, its n-fold contraction
with one single arbitrary test vector x of R D is proportional to the nth power
of the modulus ||x|| of the test vector, that is, if, and only if, there exists a real
number a such that:
VxeR", | ] . . . t l t U " ^ = * r (2-23)

ai=l an=l
2.3.5.2 Crystallographic tensor of order n

Consider a tensor T (n) of order n > 1 constructed from the c* vectors by summing
over the channel index i, the w-fold tensor product 9 of c,- with itself, weighted
by a set of coefficients 0(C7(o) which depend only on the class C/(o to which
channel i belongs (see Section 2.3.4):
Such a tensor will be called a 'crystallographic tensor' of the lattice gas.
2.3.5.3 Crystallographic isotropy of order n

A lattice gas model has the 'n th order crystallographic isotropy' property if all
crystallographic tensors with order less than or equal to n are isotropic. We
emphasize again that this property of crystallographic isotropy must not be
confused with ordinary isotropy: a lattice cannot be isotropic in the ordinary
9
The underlying metrics is supposed to be Euclidean. We thus make no distinction
between covariant and contravariant subscripts.
sense, but its local geometry (the c/s) may however yield crystallographic isotropy
up to some order.
Crystallographic tensors built on the c/ vectors of lattice gas models are
fully symmetric by construction. Hence, the second isotropy criterion provides
a simple algebraic method to check the crystallographic isotropy properties of
a given lattice gas model. Indeed, consider the nth order crystallographic tensor
j{n) = ^ = 1 c i a i ...Cian of a lattice gas model; if n is odd, the isotropy test is
simple: isotropy holds if, and only if, T^ is zero. If n is even, then the full
contraction of T(w) with an arbitrary vector x reduces to X^=i(c* ' XT which is a
polynomial function with even order of the coordinates (xi,..., xp) of the vector
x. The test for crystallographic isotropy is then to check whether this function
is proportional to (]Ca=i x«)2. Such algebraic manipulations, involving only the
use of multinomial expansions, can be carried out by hand at least for orders
lower than or equal to 6 or 8. (The use of a symbolic manipulator can simplify
the task.)
2.3.6 Irreducibility
Irreducibility is a symmetry-related property which is not as natural as the
notions defined previously. However, some of the calculations presented in
Chapters 4 and following, are dramatically simplified if this property is verified.
We thus wish to spend some time to define and visualize clearly this notion, and
to explain its consequences.10
2.3.6.1 Definition
A lattice gas model is said to be 'irreducible' if, and only if, for any channel index
i= 1,..., ft, the invariant subspace £; of the subgroup Gt of all lattice-preserving
isometries that also preserve c,-, is linearly generated by c,-. In other words, any
vector u preserved by all the isometries in G which also leave c,- unchanged,
must be a multiple of c,-.
We shall try to justify the name 'irreducibility' given to this property. Let us
consider the subgroup Gt and the hyper-plane 11/ of dimension D — 1, orthogonal
to c/. We construct the set G\ collecting the restrictions to 11/ of all elements in G/.
The irreducibility property simply states that G\ is an irreducible representation
of Gt on n*.11
2.3.6.2 Examples
To illustrate the notion of irreducibility for a lattice gas, we use once more
the simple HPP and extended HPP models presented in Section 2.3.4. Let us
10
On first reading, the reader can omit the details of the present section and just keep in
mind its conclusions and consequences.
11
For basics on irreducible representations of groups, the reader is referred to text books
on group theory, for example Eliott and Dawber (1979), or Tung (1985).
consider first the simple HPP model (see Figure 2.3(a) for notations). We pick
channel 1 corresponding to the velocity vector ci = (1,0). The subgroup G\ of G
leaving ci unchanged contains only the identity (Ro) and the mirror symmetry
Sy. The only vectors in R 2 that remain invariant under the action of G\ are
clearly proportional to ci. Since all four vectors ci to C4 are equivalent, what
is true for ci is true for all c* vectors. This completes the proof that the simple
HPP model possesses the property of irreducibility. Consider now the extended
HPP model defined in Figure 2.3(b). We pick channel 13 corresponding to the
velocity vector C13 = (2,1). The subgroup G13 of G leaving C13 unchanged only
contains the identity (Ro)- All vectors in ]RD are invariant under the action
of G13, not only those proportional to C13. This shows that the extended HPP
model does not possess the irreducibility property.
2.3.6.3 Consequences of the irreducibility property

As already stressed earlier, the irreducibility property can be used to greatly
simplify algebraic calculations on the statistical and macroscopic properties of
a lattice gas model. Indeed, suppose that, in some algebraic calculation, we
encounter a set of vectors T\ such that, for all g in G and for all i = l,...,b,
we have g(T/) = Tg(,) (for the definitions of g and g, see Section 2.3.4). If the
model considered is G-invariant and verifies the irreducibility condition, then
we can immediately state that there exists a collection of coefficients ((/) where
/ = C/(o, such that T,- = £(J)Cj, for all i = l,...,fc. This considerably simplifies
the algebra.
Moreover, the same kind of simplification also holds for second order tensors.
Consider a set of symmetric12 second order tensors T,- such that, for all g in
G and for all i = l , . . . , b , we have g(T,-) = Tg(;).13 If the lattice gas model is
G-invariant and verifies the irreducibility condition, then the tensors T, must
have the form T, = A(/)c; ® c* + fi(I)d9 where d is the Kronecker tensor, and
X(I) and n(I) are two collections of coefficients that depend only on the class
/ = Cl(i) of the channel.
The presentation of these technically very useful properties closes this section
on the microscopic properties of lattice gases.
2.4 Special rules
If the part of lattice i f actually occupied by the lattice gas were infinite in
all directions, then the cellular automaton rule described by the microdynamic
12
In this context, 'symmetry' means that Ti0Lp = T^ a , where Titxp are the Cartesian
components of the tensor Tj.
13
For simplicity, we denote by the same symbol g the isometry acting on vectors and on
second order tensors.
equation (2.12) could be applied everywhere. In practice however, that part

of the lattice is finite and the propagation phase requires periodic wrapping
such that no particle be lost through the boundaries of 5£. As explained in
Section 2.2.2, this is realized (i) by embedding if into a wider domain if', that
can perfectly tile the full Bravais lattice by successive translations, and (ii) by
reconnecting opposite sides in such a way that <£' becomes a topological torus.
In most cases, if can itself perfectly tile the Bravais lattice, and is thus identical
to ifr. From now on, we assume this to be the case, and we use the notation if
only.
If the collision operator described in Section 2.2.3 were applied everywhere in
^£, then the macroscopic quantities would verify periodic boundary conditions
in the usual sense. When different macroscopic boundary conditions are im-
posed, one must apply to the boundary nodes of 5£ special collision rules that
differ from those applied in the bulk. Note that special collision rules are not
necessarily restricted to boundary nodes: they can also be applied to localized
zones in the bulk, to simulate physical situations such as, the presence of solid
obstacles in a flow.
Several special collision rules are designed to simulate various macroscopic
boundary conditions. In the following subsections, we describe some of those
which are most frequently used in lattice gas simulations of fluid dynamics.
They can be used plainly or in a statistically mixed manner to obtain particular
effects. For example, a solid obstacle collision rule statistically mixed with the
normal collision rule can produce a semi-permeability effect, with a permeability
adjusted by tuning the respective probabilities for the two collision rules.
2.4.1 Solid impermeable obstacles

A fixed solid obstacle inside, or at the boundaries of, a fluid flow can be
implemented by applying a special collision operator to all the obstacle nodes.
This special collision operator, called the 'reflection operator', must act on the
Boolean state of obstacle nodes so as to send back any incoming particle to
regular bulk nodes. In other words, the channels connecting an obstacle node
to any other obstacle node must be empty after the application of the reflection
operator. This guarantees that the nodes inside the obstacle (those which cannot
be connected to any node of the bulk by one of the c,- vectors) will remain
empty if initially empty. Only the nodes of the surface layers of the obstacle
(those which can be connected to a bulk node by one of the c,-s) will not
remain empty. This raises the question of the position of the effective surface
of the obstacle, that is, the location where the normal mean velocity actually
vanishes. Cornubert, d'Humieres and Levermore (1991) have shown, for some
two-dimensional models, that the actual position of the effective surface may
depend on the orientation of the surface with respect to the lattice symmetry
axes, at least for specular reflections. If solid obstacle conditions are applied to
the boundaries of if, for example to simulate wall-bounded flows, then some
care must be taken to avoid leakage. The layer of boundary nodes where the
obstacle collision operator is applied must be sufficiently thick to guarantee that
during the propagation phase, no particle be sent directly outside the layer of
boundary nodes. This point is discussed in Chapter 10.
There exist several ways to realize particle reflection, leading to several dif-
ferent reflection operators, and to different macroscopic effects. In the next
paragraphs, we describe some of the simplest reflections: the 'bounce-back',
'specular', and 'diffusive' reflection operators.
2.4.1.1 Bounce-back reflection

The bounce-back reflection operator sends back all the particles impinging on
an obstacle node to where they come from (see Figure 2.4(a)). This operation
M7I/,
WW/,
Diffusive
Px t=> -PXr^
p
y P
Figure 2.4 The figure illustrates, in a two-dimensional case, the effect on the
total momentum of (a) the bounce-back reflection, (b) the specular reflection
and (c) the diffusive reflection. The solid arrows represent total momenta at
surface nodes. They do not represent the individual velocity vectors c,- as in
the figures of Chapter 3. Px and Py denote respectively the parallel and
orthogonal components of the total momentum P.
conserves the total mass and energy observables, but the momentum observable
is reversed.
The bounce-back reflection mode is by far the simplest one and does not
depend on the orientation of the surface. The easiest realization is to apply
the central symmetry to the state of obstacle nodes. A quite efficient numerical
strategy to implement this central symmetry is described in Chapter 10. At
the macroscopic level, the global effect of bounce-back reflections is that all
components of the mean velocity are canceled at the surface of the obstacle,
where 'no-slip' boundary conditions are thus realized.
2.4.1.2 Specular reflection

The specular reflection operator makes the particles bounce elastically on the
surface of the obstacle, like a light beam reflected on a perfect mirror (see
Figure 2.4(b)). This operation conserves the total mass and energy observables;
the component of the momentum parallel to the surface is also conserved, but
the normal component is reversed.
The specular reflection depends on the orientation of the surface and requires
that the direction normal to this surface be well defined, which is not obvious in
a discrete context where surfaces are not smooth. One solution is to define the
normal direction as the binary symmetry axis14 of the Bravais lattice which is the
closest to the sum of all the c, s connecting the obstacle node to other obstacle
nodes. The specular reflection is realized by applying successively the central
symmetry and the 180°rotation around the normal direction. This operation
reverses the normal component of the momentum but conserves the parallel
component.
At the macroscopic level, the effect of specular reflections on the surface of
an obstacle, is to cancel the normal component of the mean velocity, without
canceling its parallel component. We so obtain 'free-slip' boundary conditions
at (or near) the surface of the obstacle.
2.4.1.3 Diffusive reflection

The diffusive reflection is a statistical combination of bounce-back and specular
reflections, with tunable respective probabilities (see Figure 2.4(c)). This oper-
ation conserves the total mass and energy observables; the component of the
momentum normal to the surface is reversed, and the parallel component is
conserved or reversed according to a random choice.
14
A binary symmetry axis is such that a 180°rotation around the axis leaves the Bravais
lattice globally invariant.
2.5 Comments 33
2.4.2 Sources and sinks of observable quantities

It is often useful to simulate localized zones at the boundaries or in the bulk
of <£, where some observable quantity, which would be conserved otherwise, is
being injected or absorbed (sources or sinks). As an example, consider a thermal
lattice gas model designed to simulate fluid flow with thermal exchanges: local
heating or cooling can be simulated by locally injecting or absorbing kinetic
energy (otherwise conserved), while conserving mass and momentum. This kind
of effect can be achieved by applying suitable combinations of the usual collision
operator with some special collision operators. The latter have almost the same
conservation rules as the usual operator, except that each of these special
collision operators maximizes the increase or decrease of one of the usually
conserved observable quantities. A suitable tuning of the relative statistical
weights in the combination produces the desired rate of injection/absorption
for each of the usually conserved quantities. In the example of tunable local
heating (cooling), one combines the usual collision operator with a 'heating
operator' ('cooling operator'), with adjustable statistical weights. The heating
operator (cooling operator) must be designed so as to maximize (minimize)
kinetic energy while conserving mass and momentum. Another example is the
incorporation of body forces in a lattice gas model for fluid flow. The realization
proceeds along the same lines. Consider a body force in the direction x\, acting
over a more or less extended region of if. Two special operators are needed
that maximize respectively the increase and decrease of the component p\ of
the momentum, while conserving all other usually conserved quantities (mass,
momenta P2 to p#, energy, etc.). The sign and intensity of the body force is
adjusted by a proper tuning of the statistical weights of the normal and special
collision operators.
2.5 Comments
We have presented the framework and the tools for the description and classi-
fication of lattice gases at the microscopic level. These microdynamic concepts
and properties have a crucial importance for the theory to be developed in later
chapters. The material presented so far may be perceived as rather abstract.
Therefore, before engaging in the systematic study of the statistical properties of
lattice gases at equilibrium and out of equilibrium, we devote the next chapter
to the description of various examples of lattice gas models in order to illustrate
the microdynamic formalism.15
15
The reader familiar with the microdynamic formalism and with lattice gas models may
want to move directly to Chapter 4.
Chapter 3
Microdynamics: various examples
We now illustrate the abstract microdynamic notions of Chapter 2, with a

presentation of lattice gas models in terms of the microdynamic tools. The
models are chosen to illustrate the various microdynamical concepts; further
models will be considered briefly in Chapter 11.
We start with the simplest two-dimensional model based on the square lattice,
the earliest lattice gas model (1973) labeled HPP according to the initials of
the authors: Hardy, de Pazzis and Pomeau. Sections 3.2 to 3.4 are devoted
to models constructed on the triangular lattice and based on the FHP model
initially introduced by Frisch, Hasslacher and Pomeau (1986). A 'colored' version
of the FHP model, developed as a two-components lattice gas is presented in
Section 3.5. A slightly more complex model, also based on the triangular lattice,
but with thermal properties (Grosfils, Boon and Lallemand, 1992) is described
in Section 3.6. We then move to three-dimensional systems in Section 3.7, as we
introduce the basic (pseudo-four-dimensional) lattice gas model of d'Humieres,
Lallemand and Frisch (1986).
Except for the HPP model, all the models presented in this chapter, have
been designed to exhibit large-scale dynamics in accordance with the Newtonian
viscous behavior of isotropic fluids.
31 The HPP model
Historically, the first lattice gas model was introduced in the early seventies by
Hardy, de Pazzis and Pomeau (1973) with motivations focusing on fundamental
34
aspects of statistical physics (see also Hardy et al, 1972, 1976 and 1977). Their
model (HPP) is based on the two-dimensional square lattice with unit lattice
constant (see Figure 3.1).
The HPP model is a 4-bit model, where each node has b = 4 channels,
corresponding to the four directions of the square lattice (east, north, west,
south) labeled by integers from 1 to 4. The c,-s have unit modulus and their
Cartesian components are:
These vectors connect a node to its four nearest neighbors. So, the model's
connectivity B and the number b of channels per node are both equal to
the coordination number of the two-dimensional square lattice, namely four.
Furthermore, the four connection vectors are identical to the four velocity
vectors C*, which are all different and non-zero (see Section 2.1.1).
The masses m,- of particles are equal, and by a suitable choice of the mass
scale, their common value may be taken as one unit mass.1 The momentum p,
of each particle is equal to c,-, which is physically consistent with unit mass and
unit time step. The kinetic energy of each particle is taken equal to \, again
with physical consistency. The choice of unitary quantities for lattice constant,
time step, and particle mass completely defines a natural system of units for
microscopic mechanical quantities.
As prescribed by the general definition of lattice gases (see Section 1.2.3),
Figure 3.1 The HPP lattice with the corresponding c,s.

1
Most often, the mass will therefore not appear explicitly in the equations.
36 3 Microdynamics: various examples
the evolution rule of the HPP model is a two-phase sequence: a propagation

phase and a collision phase. During the propagation phase, a particle present
at node r* in channel i moves to node r* + c;, its nearest neighbor in the
direction i. During the collision phase, a pair of particles, arriving at the
same node from opposite directions, for example north-south, is turned into
a pair of particles heading in the two other directions, here east-west (see
Figure 3.2). The reverse collision transforming east-west into north-south also
has probability one, but is not shown on Figure 3.2 because it can be obtained
from the collision (north-south —• east-west) by the application of an isometry
(+n/2 rotation) belonging to the local invariance group of the lattice. All
other states which are not strictly 'north-south' (s = 0101) or 'east-west' (s =
1010) remain unchanged in the collision step. Notice that only two among
the 24 = 16 possible states, can undergo an 'effective collision', that is, with
a post-collisional state different from the pre-collisional state. The 'collisional
efficiency' (the percentage J2s=£s'A(s^>s')/Y^SS'A(s^sf) of efficient collisions) is
12.5 %. In other words, when the Boolean state of a given node r* is randomly
chosen among the 16 possible states of the single-node phase-space y (with
equal probabilities), effective collisions occur at node r* with a probability of
0.125.
3.1.1 The microdynamical equation

The number b of channels per node being quite small for the HPP model,
(namely four), the number of possible states per node is only 2b = 16. In
addition, the HPP model is deterministic and verifies detailed balance (2.15). So,
the collision matrix A(s^s') is a symmetrical 16 x 16 square matrix containing
only 0s and Is. This matrix is a bit cumbersome and can most easily be
constructed. The microdynamical equation (2.13) can be rewritten for the HPP
1
-I 1 l_ -4 _
l
1 1 1 1 1 1k 1
\ 1 J 1 1
1 ^1 1 T
1 1 1 1 ,. 1
-\ — 1— + -\ — i 1-
1
Figure 3.2 Collision rules for the HPP model, reduced by symmetry. The full
set of collisions can be reconstructed from the above reduced set by the
application of all the isometries of ^. The number T above the open arrow
indicates that the collisional change of configuration occurs with probability 1.
model by taking all the £(s^s')s equal to the ^l(s->s')s (since the model is
deterministic). It reads:
for i = 1,...,4:
- ^+1^+2^+3
+
Here, n; stands for (1 — nt) and the subscript i (channel index) is cyclic (modulo
4). All variables on the right hand side are taken implicitly at time U and node
r*.
Because of the simplicity of the HPP microdynamics, (3.2) can be derived
merely by intuition: the three terms on the right hand side simply express that
the collision phase can either:
v leave the occupation state of channel i unchanged, or

}-• modify its state from 1 to 0 if channels i and i + 2 are occupied and
channels i + 1 and i + 3 are unoccupied, or
• modify its state from 0 to 1 if channels i + 1 and i + 3 are occupied and
channels i and i + 2 are unoccupied.
We note that (3.2) is compatible with the exclusion principle, since it guaran-
tees that the nts at time U + 1 cannot be anything other than 0 or 1, provided
the n,s at time U are 0 or 1.
For simplicity, we have chosen to write (3.2) for a microscopic evolution law
where the collision phase is performed before the propagation phase. If the
reverse order is chosen, the microdynamical equation takes almost the same
form, except that each of the n,s and n^ on the right hand side must then be
taken at a different node (namely at nodes r* — c,-), instead of being all taken
at the same node r*. This minor difference is generally unimportant, since the
order of the two successive phases becomes macroscopically irrelevant when a
large number of steps are executed sequentially.2
The microdynamical equation can also be written in logical form; it suffices
to consider the n^s as logical variables with values T R U E ' or 'FALSE' (instead
of algebraic variables with values T or '0'), to obtain:
n;(r* + c,-, U + 1) = (iti and (n,-+i or n,-+2 or n I+3 ))

(3 3)
) \ '
or f ni+\ and (n,-+2 and ni+2 and n,-+3) J
Here, n,- now stands for the logical negation of 7?,. Although equations (3.2)
and (3.3) seem formally different, their content is the same. For many purposes,
however, the algebraic form is more convenient.
2
Note that care must be taken as to when the iteration sequence is stopped, when
performing correlation measurements on the lattice.
3.1.2 Microscopic properties

The properties listed below for the HPP model have been defined in a general
framework in Section 2.3. Proofs are given for these properties only when
non-trivial.
(i) The HPP model is deterministic, by construction.

(ii) It satisfies detailed balance and a fortiori semi-detailed balance (see Sec-
tion 2.3.1).
(iii) It is self-dual since the only permitted collisions (1010) —> (0101) and
(0101) —> (1010) are dual of each other, and have equal probabilities,
namely one (see Section 2.3.2).
(iv) The only linearly independent collisional invariants are the mass (particle
number) and the two components of the momentum:
o\
1 (3.4)
m Py 0
V-i/
In addition, the HPP model has several static and moving geometrical
invariants (see Section 2.3.3), such as the x- (or y-) component of the
momentum, summed over any line of lattice nodes with equal x- (or y-)
coordinates (d'Humieres, Qian and Lallemand, 1989, 1990).
(v) The model is clearly G-invariant (invariant under the crystallographic group
of the lattice, see Section 2.3.4), and all four c,- vectors clearly belong to the
same class, as defined in Section 2.3.4.
(vi) The HPP lattice gas model has third order crystallographic isotropy (see
Section 2.3.5). As a justification, Table 3.1 gives the crystallographic tensors
of order 1 to 4, contracted with a vector x with coordinates (xi,X2)-
Clearly, the contraction of the fourth order tensor is not proportional to
Table 3.1 Crystallographic isotropy properties of

the HPP model. The second column gives the
result of a full contraction of the nth order tensor
with an arbitrary test vector x. The third column
summarizes the conclusion about isotropy.
Order n ^ T^]Mnxai...*«„ Status

1 0 isotropic
2 2(xl + xl) (=2||x|| 2 ) isotropic
3 0 isotropic
4 2(xj+^2) (=^= a||x||4) anisotropic
3.2 TheFHP-1 model 39
||x||4; hence, by the criterion of Section 2.3.5, fourth order crystallographic

isotropy definitely does not hold for the HPP lattice.
(vii) The HPP model is irreducible (see Section 2.3.6). Indeed, the only isometry
in G that preserves ci, for example, is the mirror symmetry with respect
to the x-direction, whose invariant sub-space only contains vectors propor-
tional to ci. The same argument holds with C2, C3 and C4.
Since all C;S have the same modulus 1 and all the particles have the same
mass, the kinetic energy observable is simply proportional to the mass observable.
States with the same mass will also have the same kinetic energy, so, the energy
conservation and the mass conservation are equivalent. As a consequence, the
HPP model ignores thermal effects.
The degree of crystallographic isotropy of the HPP model (namely three)
is not sufficient to produce isotropic large-scale dynamics as described by the
Navier-Stokes equations for a Newtonian fluid.3 This requires at least fourth
order crystallographic isotropy (see Chapter 5). It is actually possible to construct
models with fourth order isotropy on the square lattice. However, such models
are not minimal, in the sense of the concept introduced in Section 1.3.1, and
require a larger set of c* vectors: in addition to the HPP c,s, which must now
be endowed with multiplicity four (4 channels for each of the c;s), four vectors
must be added* connecting each node to its next-nearest neighbors (1 channel
per vector). Such 'modified' HPP models require b = 20 channels per node, and
are not homokinetic. About ten years after the introduction of the HPP model,
the 'anisotropy disease' has been cured in a less expensive way by models based
on the triangular lattice, which are discussed in the next sections.
3.2 The FHP-1 model
The family of FHP models, based on an original idea of Frisch, Hasslacher,

and Pomeau, provides a class of two-dimensional lattice gas models with the
simplest structure compatible with the property of fourth order crystallographic
isotropy producing proper (isotropic) large-scale dynamics (Frisch, Hasslacher
and Pomeau, 1986). This family of models is based on the two-dimensional
triangular lattice with unit lattice constant (see Figure 3.3). Several versions of
the FHP model have been successively developed, with the same geometrical
lattice structure, but with different collision rules. In this section, we present the
earliest and simplest model of the family: the 'FHP-1' model.
The FHP-1 model is a 6-bit model. Each node has b = 6 channels, corre-
3
Notice that the square lattice gas is well suited to model scalar transport such as
diffusion, and has been used efficiently for a microscopic approach to reaction-diffusion
phenomena (see Boon, Dab, Kapral, and Lawniczak, 1996).
sponding to the six directions of the triangular lattice, labeled by integers from
1 to 6. The CjS have unit modulus and their Cartesian components are:
(3.5)
C4
=(~o 1
These vectors connect each node to its six nearest neighbors. Thus, the model's
connectivity B and the number b of channels per node are both equal to
the coordination number of the two-dimensional triangular lattice, namely six.
Furthermore, the six connection vectors are identical to the six velocity vectors
Cj, the latter being all different and non-zero (see Section 2.1.1).
The FHP-1 model shares with the HPP model the following properties: The
masses m* of all the particles are equal, and by a suitable choice of the mass
scale, the common value is taken as unity. The momentum p,- of each particle
is equal to c,-; this is physically consistent with unit mass and unit time steps.
The kinetic energy of each particle is taken equal to \, again for physical
consistency. The option that lattice constant, time step and particle mass be
equal to unity completely defines a natural system of units for microscopic
mechanical quantities.
The propagation phase in the FHP-1 model proceeds in exactly the same way
as for the HPP model. An essential difference appears in the collision phase:
while the collisions are deterministic for the HPP model, some stochasticity
enters the FHP microdynamics. Indeed, two particles coming from opposite
directions (for example 1 and 4) undergo a binary collision with an output state
Figure 3.3 The FHP lattice with the corresponding C;S.

rotated by +60° or —60° (as shown in Figure 3.4(a)), with probabilities p and
1 — p respectively. Another new feature is the introduction of deterministic triple
collisions: three particles coming from three directions forming 120° angles
between each other, will be deflected by 60° (as shown in Figure 3.4(b)). All
other states remain unchanged. Note that, among the 26 = 64 possible states,
the five states shown in Figure 3.4 are the only ones that can undergo effective
collisions. The collisional efficiency of FHP-1 is 7.81% (see Section 3.1 for
definition).

The microdynamical equation for the model FHP-1 exhibits the same structure
as the microdynamical equation for the HPP model. It contains a few more
terms because the number of possible collisions is larger. This equation can be
obtained systematically from the general form (2.13), but it is a rather tedious
operation, and the intuitive method described for the HPP model is preferable.
One obtains:
fori=l,...,6:
. + 1) = nt
(3.6)
Here, ni stands for (1 — n,) and the index i is cyclic (modulo 6). All variables
on the right hand side are taken implicitly at time U and node r*. £(+)(r*, £*)
--*—X--
\ / \ / \ /
Figure 3.4 Collision rules for the FHP-1 model, reduced by symmetry. The
full set of collisions can be reconstructed from the above reduced set by the
application of all the isometries of (§. The numbers above the open arrows are
the transition probabilities. The most commonly used value p = 1/2 has been
chosen here, since it is the only value compatible with the G-invariance.
(respectively £(_)(r*,£*)) denotes a collection of Bernoulli random variables

(with values 0 or 1) satisfying £(+)(r*,£*) + £(_)(r*,£*) = 1, at each time U and
at each node r*, and whose ensemble averages are all equal to p (respectively
1-p).

The properties listed below for the FHP-1 model have been defined in a general
framework in Section 2.3. Proofs are given for these properties only when
non-trivial.
(i) The FHP-1 model is non-deterministic, except if p = 1 or p = 0.

(ii) The collision rule satisfies semi-detailed balance (2.17) for any value of the
probability p between 0 and 1. Indeed, for any of the three Boolean states
(100100), (010010) and (001001), there exist two possible pre-collision states,
with respective probabilities p and 1 — p, and for all other Boolean states
there exists one and only one possible pre-collision state with probability 1.
The identity ^ s ^4(s-»s') = 1 is thus satisfied. Note that when p = 1/2, the
collision rule even satisfies detailed balance, since the relation A[s^s f) =
A(sf^s) then holds for any pair of Boolean states s and s'.
(iii) The model is clearly not self-dual, as the dual of two-particle collisions
are four-particle collisions which are ignored by the evolution rule of the
FHP-1 model.
(iv) The only linearly independent collisional invariants are the mass (that is,
the particle number) and the two components of the momentum:
/1 \ / 1 \ / 0 \
1 2 4
~2 ] 2T
m: !. ,, p x ':
Px , py : . . (3.7)
-1 0
\J
1 2 -4
\-4/
The geometrical invariants of FHP models are discussed by d'Humieres,
Qian and Lallemand (1989, 1990).
(v) The FHP-1 model is G-invariant only when p is equal to 1 — p, that is, if
p = 1/2. Indeed, consider the three states s<°> = (100100), s^ = (010010)
and s ( - ) = (001001). The mirror symmetry gx with respect to the x-axis
induces on the single-node phase-space y a, transformation (SX that leaves
s(0) unchanged and permutes s (+) and s^~\ Since &x changes the collision
s(0) _• s(+) j n t 0 th e collision s ^ —• s^ and reciprocally, the probabilities of
these two collisions must be equal in order that the model be G-invariant.
3.2 TheFHP-1 model 43
In addition, all six velocity vectors clearly belong to the same class, since
any of these six vectors can be changed into any other by a rotation of
some integer multiple of TC/3.
(vi) The FHP-1 model has the fifth order crystallographic isotropy (see Sec-
tion 2.3.5). As a justification, Table 3.2 gives the crystallographic tensors
of order 1 to 6, contracted with a test vector x of coordinates (xi,*2)-
Clearly, the contraction of the fourth order tensor is proportional to ||x|| 4 ,
and the fifth order tensor is zero. Hence, by the criterion of Section 2.3.5,
crystallographic isotropy holds up to the fifth order for the FHP lattice.
(vii) The FHP-1 model is irreducible (see Section 2.3.6). Indeed, the only isom-
etry in G that preserves ci, for example, is the mirror symmetry with
respect to the x-direction, whose invariant sub-space only contains vectors
proportional to ci. The same argument holds for C2 to C6.
As for the HPP model, the energy observable is proportional to the mass
observable and its conservation is a consequence of the mass conservation. So,
the FHP-1 model has large-scale behavior ignoring thermal effects.
The basic reason for incorporating triple collisions in the FHP-1 model follows
from the fact that without triple collisions the model would have an additional
collisional invariant, namely ( 1 , - 1 , 1 , - 1 , 1 , - 1 ) , which is physically irrelevant.
This spurious invariant would strongly affect the macroscopic dynamics of the
model and would make it unphysical.
The degree of crystallographic isotropy of FHP-1 is larger than four, which is
compatible with the possibility of having realistic large-scale dynamics governed
in certain limits by the Navier-Stokes equation, provided the model be G-
invariant, that is, for p = 1/2, which is the most commonly used value. Other
values produce chirality.
Table 3.2 Crystallographic isotropy properties of the FHP

models. The second column gives the result of a full
contraction of the nth order tensor with an arbitrary test
vector x. The third column summarizes the conclusion
about isotropy.
Order n £ r a ^ x 8 l ...x« n Status

1 0 isotropic
2 3(xf + x22) (= 3||x||2) isotropic
3 0 isotropic
4 \{x\ + 2x\x\ + x\) (= | ||x||4) isotropic
5 0 isotropic
6 ^(llxf + \5x\x\ + A5x\x\ + 9x\)
(=£ a||x||6) anisotropic
3.3 The FHP-2 model
The FHP-2 model is a variant of the FHP-1 model that includes the possibility
of one rest particle per node, in addition to the six moving particles of FHP-1.
Each node then has b = 7 channels, corresponding to particles moving along
the six directions of the triangular lattice and to the rest particle. The channels
corresponding to moving particles are labeled by integers from 1 to 6, and the
channel corresponding to the rest particle is labeled 0.
The six connection vectors are the same as for FHP-1. They do not include
the vector Co which is zero. As a consequence, the connectivity B, which is still
equal to the coordination number of the triangular lattice (that is, 6), is now
different from the number b = 1 of channels per node. The six vectors labeled
by i = 1 to 6 have equal unit moduli, and the vector Co corresponding to the
rest particle obviously has zero modulus.
The masses m* of all particles are equal to one and the momentum p, of each
particle is equal to c,. Thus, the momentum of the rest particle is zero. The
kinetic energy is taken equal to \ for moving particles, and to 0 for the rest
particle. As for the HPP and FHP-1 models, the above choices are physically
consistent with unit particle mass and unit time step.
The propagation phase for moving particles is the same as for FHP-1, and
does not affect the rest particles. The collision rules of the FHP-2 model are
similar to the collision rules of FHP-1 with four additional collisions coupling
moving and rest particles: (i) a moving particle arriving at a node occupied
only by a rest particle produces a pair of moving particles with angles +60°
and —60° measured from the direction of the incoming particle (Figure 3.5.e),
(ii) the reverse of the former (Figure 3.5.f), (iii) the collisions of FHP-1 with
a passive rest particle left unchanged by the collision (Figure 3.5.C and 3.5.d).
All possible collisions of FHP-2 that cannot be deduced from one another
by a lattice-preserving isometry are summarized in Figure 3.5. Only 22 among
the 27 = 128 possible states can undergo effective collisions. The collisional
efficiency of FHP-2 is 17.2% (see Section 3.1 for definition).

As for FHP-1, it is possible to derive the microdynamical equation of FHP-2
directly from the general form (2.13). The same result can be obtained more
rapidly by the intuitive method which consists of writing, for each n,-, the balance
for positive and negative contributions from each type of collision. The resulting
microdynamical equation for the model FHP-2 has the same structure as for
FHP-1, except that there is an additional equation for no, and a few more terms
on the right hand side. The result reads:
for i = 1, . . . , 6 :
n,-(r*+c,-,f* + l) = fit
+
+
/\ 7\
\ / \/
- -K - - > - -
\/ \ / \ \ / x / \
A** A /\
application of all the isometries of ^ . Black dots represent rest particles.
for z = 0:
n0
Here, the index i is cyclic (modulo 6), in such a way that index i = 7 is equivalent
to index 1 (and not 0); other notations are as for FHP-1 (see Section 3.2.1).
Note that in (3.8) five terms are independent of no, which reflects that the
binary and triple collisions of FHP-1 occur indifferently with or without a
passive rest particle. The strict application of (2.13) would have required two
pairs of random Bernoulli variables (6(H, 6^1) and (6^1, £p\), for binary head-
on collisions with and without passive rest particle respectively. In fact, since
these two pairs of random variables have the same statistics, they are taken
equal: this leads to the above form for the microdynamical equation of FHP-2.

framework in Section 2.3. As for HPP, proofs are given only when non-trivial.
(i) The model is non-deterministic, except if p = 1 or p = 0.

(ii) The collision rule satisfies detailed balance if p = 1/2, and semi-detailed
balance otherwise,
(iii) The model is not self-dual,
(iv) The only collisional invariants are the mass and the two components of
the momentum:
1\ / 0 \
1 1 0
l
1 2
1
m 1 , Px : 2 (3.10)
1 -1 0
1
1 2
1/ V5/
(v) The model is G-invariant only if p = 1/2.
The velocity vectors can belong to two different classes: The first class,
conventionally labeled by / = 0, contains only Co, the null velocity vector of
the rest particle; the second class, labeled by / = 1, contains the non-zero
velocity vectors ci to C6-
(vi) The model has fifth order crystallographic isotropy, for the same reason
as FHP-1, since the rest particles, which have zero velocity, do not modify
the crystallographic tensors of any order.
(vii) The model is irreducible. Indeed, the only isometry in G that preserves
ci, for example, is the mirror symmetry with respect to the x-direction,
whose invariant subspace only contains vectors proportional to ci. The
same argument holds for C2 to C6. For Co which belongs to Class 7 = 0
however, the reasoning is as simple, albeit different: all the isometries in G
leave Co unchanged. The only vector left unchanged by all the isometries
in the point group G of the triangular lattice is of course the null vector,
which is Co.
As for FHP-1, the degree of crystallographic isotropy of FHP-2 is compatible

with the possibility of large-scale behavior governed (within certain limits) by
the Navier-Stokes equation, provided that the model be G-invariant (that is, for
P = V2).
FHP-1 and FHP-2 have the same microscopic properties, and the proofs
are identical. The essential difference lies in the number of possible effective
collisions which is larger for FHP-2 (22 versus 5), and yields a higher collisional
efficiency (17.2% versus 7.81%). This difference in collisional efficiency has a
major consequence at the macroscopic level: The kinematic shear viscosity is
smaller for FHP-2 than for FHP-1. This decrease in shear viscosity allows
numerical simulations at higher Reynolds number (for equal numerical invest-
ment). Another difference between FHP-1 and FHP-2 is that the kinematic bulk
viscosity is zero for FHP-1 and non-zero for FHP-2.
The kinetic energy observable in FHP-2 is not proportional to the mass
observable because the velocity of rest particles is zero. The kinetic energy is
not a collisional invariant for FHP-2, since binary collisions between a rest
particle and a moving particle create some kinetic energy. Moreover, the only
collisions that conserve the total mass, momentum, and kinetic energy are those
which also conserve the number of rest particles, that is, those for which the rest
particles are only passive spectators. Recovering the kinetic energy conservation
by allowing only this type of collisions would cancel the benefit of rest particles:
they would remain static for ever, being affected neither by the propagation,
nor by the collision phase. To recover energy conservation without inhibiting
the dynamics of rest particles, we need to assign to the particles some kind
of 'internal degree of freedom' which is never excited (zero energy) for moving
particles, and always excited for rest particles. The energy associated with the
excitation of the internal degree of freedom of rest particles is chosen to be
equal to the kinetic energy of a moving particle (1/2 in natural lattice units). So,
the total energy observable (kinetic plus internal) is proportional to the mass
observable and is thus conserved. It is precisely because the FHP-2 model allows
for collisional transfer of internal energy to kinetic energy that bulk viscosity
exists in the FHP-2 gas. However, the energy conservation does not lead to
an additional relevant macroscopic equation, since mass and total energy are
simply proportional, and there can be no effective collisional transfer of kinetic

energy. Consequently there are no thermal effects in the FHP-2 lattice gas.
Triple collisions are not mandatory here, as they are for FHP-1 where they
remove the spurious collisional invariant. Indeed, the two-particle collisions
involving a rest particle are sufficient to eliminate the spurious invariance.4
3.4 The FHP-3 model
A further variant of the 7-bit FHP-2 model is FHP-3 where the collision rules
are designed to include as many collisions as possible, under the constraint
of having the same collisional invariants as with FHP-2. Figure 3.6 shows the
collision rules of FHP-3. As for Figures 3.2, 3.4 and 3.5, the lattice-preserving
isometries are used to reduce the number of collisions shown. In addition to the
isometries, the self-duality property, that holds for FHP-3 but not for FHP-1
and FHP-2, has also been used to further reduce the number of configurations
in the figure. A complete collision table could be reconstructed from Figure 3.6,
by applying duality combined with the isometries of G. When this reconstruction
is performed, it can be shown that 76 among the 27 = 128 possible states can
undergo effective collisions. The collisional efficiency of FHP-3 is 59.4 % (see
Section 3.1 for definition).

It is possible to write a microdynamical equation for FHP-3, but the operation
is tedious and useless. One would obtain an equation with the same form as for
FHP-2, but with many more terms to describe all possible collisions.

framework in Section 2.3. Proofs are omitted since they are either trivial or
identical for all FHP models.
(i) The model is non-deterministic, except if p = 1 or p = 0.
(ii) The collision rule satisfies detailed balance if p = 1/2 and semi-detailed
balance otherwise,
(iii) The model is self-dual, because the set of possible collisions is complete.
This is the major difference with FHP-1 and FHP-2, at least at the
microscopic level.
4
For a detailed discussion of the collisional and geometrical invariants of the FHP
models, see d'Humieres, Qian and Lallemand (1989) and (1990).
\/ \/
/ \ / \ / \ x / \ /
_\/_A/_ CrN -:4r-*--
A A
x / \ / \ / \ /
\ /\ /\ / 1 \ / \^\ /
~/ jif^^\ I ^ ~p TV \
/ \ / \ / \
V;
rX—V-
/\ / \
X / \ /
• * "
-iL—^~-
\ i/4
1/4 -V- ~^~1r~^
7\~~
x / \ /
^ -^
-xrV- ~ A ~ A""
\|2 -;\ / \ /
-?-*—
•£.*7\~- 7^ \ / \ /\ /
1/2 _^_J*__\^
application of all the isometries of ^, combined, when necessary, with the
duality operation (exchanges between particles and holes). Black dots
represent rest particles.
(iv) The only collisional invariants are the mass and the two components of
the momentum, as for FHP-2 and FHP-1.
(v) The model is G-invariant if, and only if, p = 1/2.
The class structure is the same as for FHP-2.
(vi) The model has fifth order crystallographic isotropy.
(vii) The model is irreducible, for the same reason as FHP-2.
As for FHP-1 and FHP-2, the degree of crystallographic isotropy of FHP-3 is

compatible with large-scale dynamical behavior governed (within certain limits)
by the Navier-Stokes equation, provided that the model be G-invariant (that is,
for p = 1/2).
The comment about the (non-)conservation of the kinetic energy observable
in the FHP-2 model holds for FHP-3.
FHP-3 has almost the same microscopic properties as FHP-1 and FHP-2,
except that FHP-3 is self-dual.
The number of possible effective collisions is higher: 76 for FHP-3 versus
22 for FHP-2. The larger momentum transfer resulting from the increase in
collision efficiency has the consequence that the kinematic shear viscosity has a
lower value for FHP-3 as compared to FHP-2.
3.5 The 'colored' FHP model (CFHP)
The models presented so far involve identical particles with identical masses, that
cannot be differentiated by any physical property but their velocity. Physically,
these models involve a single 'chemical species', and all the resulting dynamics,
at any scale, is purely mechanical. In order to study diffusion phenomena,
the FHP-3 model must be modified to include two different chemical species
(Bernardin and Sero-Guillaume, 1990; McNamara, 1990; Noullez, 1990; Hanon
and Boon, 1997). The version presented here is the simplest one: the two
species are chemically passive and mechanically identical (same mass); there
is no chemical reaction between species (the collisions conserve the number
of particles of each species), and the mechanical behavior of the particles is
not affected by the specificity of the species. Consequently, the macroscopic
motions of the mixture are identical to those of a single component fluid such
as FHP-3, and each species acts as a passively diffusing scalar with respect to
the mixture. To get a better picture, we can associate colors (say red and blue)
to the two species, and consider the CFHP ('colored' FHP) model as an FHP-3
model whose particles are painted red or blue, without modifying their motions
and collisions. Indeed, the FHP-3 collision rules apply to the particles arriving
at a node, independently of their color. Then, the color attribute is attached
3.5 The 'colored' FHP model (CFHP) 51
randomly to the outgoing particles, with the constraint to conserve the number
of red and blue particles.
There exist several variations on the theme of colored lattice gases, e.g.
incorporation of reactive processes (see Boon et a/., 1996) or surface tension
effects (see Rothmann and Zaleski, 1994). They will not be considered here (see
Chapter 11).
The encoding of the CFHP model raises a little problem. Indeed, each channel
can be a priori in three different states: empty, occupied by a red particle, or
occupied by a blue particle. This situation does not seem compatible with a
binary encoding. One way to recover the Boolean formalism is to consider that
each of the seven channels of the FHP-3 model is now a double channel: one
'sub-channel' for the red particles, and one 'sub-channel' for the blue ones.
Fourteen Boolean quantities per lattice node are now required, as illustrated in
Figure 3.7. The (sub-)channels 0 to 6 encode the presence of red particles, and
the (sub-)channels 7 to 13, the presence of blue particles. The microdynamical
equation of the CFHP model is designed to avoid the 'forbidden' situations
where both red and blue sub-channels are occupied simultaneously. Of course,
initially, the lattice configuration has to be preset so that each pair of sub-
channels contains at most one particle (either red or blue). This rule expresses
an exclusion principle per pair of sub-channels.
y y y y y y
/ \Y
V
A
A AyA& \7 A A A
Y
A V
A
Vc V
'4
;\/ \ /
/ \ A
y
A
V
4
A
\/ A A \ / \/ \
A A A A A A
Figure 3.7 The 14 channels and velocity vectors c, for the colored-FHP model
(CFHP). 'Red' particles reside on channels i = 0 to i = 6, and 'blue' particles
on channels i = 7 to i = 13. The velocity vectors of channels 0 (red) and 7
(blue) are null and are not labeled. They are represented by the two concentric
circles.

Rather than writing explicitly a voluminous microdynamical equation for CFHP,
we describe its evolution as the following sequence of operations:
(i) At each node, the number of red particles is computed,

(ii) At each node and for each of the seven pairs of subchannels, the 'OR'
operator is applied so as to obtain seven Boolean quantities,
(iii) The collision operator of the FHP-3 model is applied to these seven Boolean
quantities,
(iv) The color attribute is then randomly redistributed, while conserving the
number of red particles computed at the first step,
(v) The usual propagation operator is applied.
This evolution rule clearly guarantees that if the initial state does not have more
that one particle per pair of subchannels (red or blue, but not both), then so
will the final state. From now on, we will assume this is the case. Thus, the
single-node phase space y is restricted to the 3 7 'authorized' states, among the
2 14 a priori possible states of a model with b = 14.

The properties listed below for the CFHP model have been defined in a general
framework in Section 2.3. Proofs are omitted since they are either trivial or
identical for all FHP models.
(i) The model is non-deterministic.

(ii) The collision rule satisfies detailed balance if p = 1/2 and only semi-
detailed balance otherwise. Note that the summation appearing in the
definition (2.16) of the semi-detailed balance has to be extended to the 3 7
'authorized' states.
(iii) The model is not self-dual, because the dual of a state with no particle
would be a state with both red and blue particles in all channels, which is
forbidden.
(iv) The only collisional invariants are the total mass, the two components of
the momentum and the mass of red particles. Note that since the total
mass and the mass of red particles are conserved, then so is the mass of
blue particles.
(v) The model is G-invariant if and only if p = 1/2.
The 14 channels are displayed into four classes: the first class contains
the velocity vector Co (red rest particle); the second contains the velocity
vector C7 (blue rest particle); the third gathers the velocity vectors ci to C6
(red moving particles); and the fourth one gathers the velocity vectors eg
to C13 (blue moving particles).
(vi) The model has fifth order crystallographic isotropy, as FHP-3 and for the
same reasons,
(vii) The model is irreducible, for the same reason as FHP-3.
3.6 The GBL model
The two-dimensional HPP and FHP models do not contain thermal effects,
because there is no energy conservation independently of mass conservation
(the energy observable is proportional to the mass observable). So these models
are restricted to the description of non-thermal fluids.
We now introduce a lattice gas model with non-trivial thermodynamics:
the '19-bit model', constructed by Grosfils, Boon and Lallemand (1992); for
consistency in acronymic notation, we shall refer to this model as the 'GBL
model'. The main feature of the model is that it is a multi-speed lattice gas
residing on the two-dimensional triangular lattice. Figure 3.8 shows how the
velocity vectors connect any lattice node to (i) itself (one rest particle per node;
||Co|| = 0), (ii) its six nearest neighbors (||CJ|| = 1, i = 1,...,6), (iii) its six next-
nearest neighbors (||Cj|| = ^/3, i = 13,...,18), (iv) its six next-to-next-nearest
neighbors (||c;|| = 2, i = 7,...,12).
The number of channels b is equal to 19, whereas the connection number B
is 18, since there is one rest particle per node. Both numbers are independent
of the coordination number of the triangular lattice (that is, 6).
A7
Ci Y ^ ^7 y y v
Y Y
A
AA/ A/ v \/c'\/ N\ A
y V V v\ A
C / Y
A A V A/ A
/V\i
Y \ / \
c x
/ [7v\/ y
\ A A 11
Cl8
V« A A
Figure 3.8 The GBL lattice with the corresponding c,-s.

The CjS have moduli 0, 1, ^/3 and 2 and their Cartesian components are:
co =
0 '
^ 'l\ ^_(~\
c4 = I . , c5 = % ) , c6 =
C 7 = l A ) , C8=( R \ , C9 = ! / - ! - (3-H)
1
ClO = I I, Cn = I _ ), C12 =
n R
-V5
C13 = | t/i I , CM = I j | ], C1S =
C C
16 = I J I' 17 = I ^ I. C« =
\ 2 / \ 2 /
The masses m, of the particles are equal to unity; their momenta are equal
to the CjS, and their kinetic energy is equal to ^ ||c/1|2. It is clear that the kinetic
energy observable (0, 5,..., \,2,...,2, §,..., \) is not proportional to the mass
observable (1,...,1). In contrast to FHP-2 and FHP-3, rest particles carry no
internal energy.5 Collisions can couple the different velocity modulus levels, and
conserve the kinetic energy observable. In other words, there exist collisions
which conserve the mass, momentum and kinetic energy observables, without
conserving individually the population of each velocity modulus level. Figure 3.9
shows some of these 'population-mixing' collisions, which do not exist in the
FHP models. The GBL model also includes all FHP collisions at the different
velocity modulus levels. The complete set of collision rules involves too many
possible cases to be shown explicitly. However, the general rule can be defined
in a synthetic way as follows. We define a family of Boolean states as a subset
of the single-node phase space y that contains all Boolean states with the same
total mass, the same total momentum, and the same total kinetic energy. The
set of all these families defines a partition of the single-node phase space, since
any possible Boolean state belongs to one and only one family. The collision of
the GBL model can be defined by a collision matrix A(s^s') such that:
" A(s^sr) = 0 if s and sf do not belong to the same family.

v A(s^>sf) = (card(!F)) if both 5 and sr belong to the same family 3F
containing card(^) elements.
5
Similarly as the particles of the hard sphere gas of classical statistical mechanics.
An exhaustive count shows that 517 750 among the 219 = 524288 possible states
can undergo effective collisions. The collisional efficiency of GBL is 94.4%
(which is different from 517 750/524288 because ineffective collisions can occur
with probabilities different from 0 or 1). Note that an increase in collisional effi-
ciency (raising its value up to 98.9%) can be gained by imposing that ineffective
collisions are only admitted for Boolean states belonging to families with only
one element. The probability for an effective collision 5 —> s' is then A(s^s') =
(card(?F) — \)~ , if both s and sf belong to a family with more than one element.

It would be a formidable task to write down explicitly the microdynamic equa-
tions for the GBL model, under its extensive form (as for the FHP models).
x
V - \/_ v
^\ A A \ A 7
\/ \ / \ / \ / \/ \ /
A A
Figure 3.9 Examples of elementary 'population-mixing' collisions for the GBL

model. Black dots represent rest particles.
Furthermore, the operation would be useless, since all macroscopically relevant

quantities can be computed from expressions involving the collision matrix
A(s^s'), suitably summed. These sums can be computed by hand, or by a
symbolic manipulator, even if the number of collisions is too large to allow
the explicit writing of the microdynamical equations. The only useful micro-
dynamical equation for the GBL model is the general form given in (2.12)
or (2.13).

As for the FHP models, the properties listed below for the GBL model have
been defined in a general framework in Section 2.3. Proofs are given only when
non-trivial.
(i) The model is non-deterministic by construction.

(ii) The collision rules satisfy detailed and consequently semi-detailed balance,
since all transition probabilities between pairs of states belonging to the
same family are equal. Thus ^(s-^) = ^4(sr->s) is always true.
(iii) The model is self-dual. Indeed, the duality operation (exchanging holes
and particles) acting on Boolean states induces a duality relation acting
on families: the dual of family J^ of all states s with the same total mass
M, the same total momentum P, and the same total energy E, is the set
of all states s that are dual to some state s of # \ This set is also a family
denoted #", since it contains all states with total mass M = 19 — M, total
momentum P = —P and total energy E = 24 — E. The duality between
states yields a one-to-one relation between the elements of $F and those of
# \ Thus 8F and $F have the same number of elements, and the collision
probabilities A(s^sf) and A(s^sr) are always equal.
(iv) The only linearly independent coUisional invariants are the mass, the two
components of the momentum, and the kinetic energy:
/ 0 \ / 0 \ /0\
1 Cly l
2
1 Cly 2
m : . ,, px ::
Px *:. », p« '•:
Py . ,, ee :: . . (3.12)
3
1 2
\Cl8x/ "J v-y

(v) The model is G-invariant. The proof is quite similar to the proof of self-
duality. Any isometry g of G induces a relation between families with the
same number of elements. The G-invariance follows immediately.
The class structure within the c, vectors is more complex than for FHP
models. The class 7 = 0 only contains Co, the (null) velocity vector of the
rest particle. The class 1 = 1 contains the velocity vectors ci to C6, which
have unit modulus. The class 1=2 contains vectors Cj to C12, with modulus
equal to 2. The class 7 = 3 contains vectors C13 to cig, with modulus y/3.
(vi) The model exhibits fifth order crystallographic isotropy, since its set of
velocity vectors can be viewed as the superposition of a zero vector plus
three groups of six vectors, each group being obtained from the six vectors
of the FHP model by a similitude operation (rotation and rescaling). The
degree of crystallographic isotropy is thus the same as for FHP models.
(vii) The GBL is irreducible. Indeed, the only isometry in G that preserves ci
is the mirror symmetry with respect to the x-direction, whose invariant
subspace only contains vectors proportional to ci. The same argument
holds for vectors C2 to cig. For Co, the same argument as for FHP-2 holds.
The degree of crystallographic isotropy and the non-trivial energy con-

servation property renders the GBL model suitable for large-scale thermo-
hydrodynamic simulations of two-dimensional fluid dynamics.
3.7 Three-dimensional models
The models described so far are all constructed on two-dimensional lattices.

Thus they are appropriate for the investigation of phenomena which are relevant
in two dimensions. Obviously, many phenomena are not amenable to a two-
dimensional description (e.g. the onset of three-dimensional turbulent flow), and
require the introduction of lattice gas models with higher dimensionality. In two
dimensions, models based on the triangular lattice with hexagonal symmetry
satisfy the fourth order crystallographic isotropy, but we noticed that the HPP
(square) model does not, and thus produces anisotropic large-scale dynamics.
In three dimensions, the problem of isotropy is more complicated: none of the
fourteen Bravais lattices can sustain a lattice gas model that satisfies the fourth
order crystallographic isotropy, at least within the scope of minimal models,
that is, with the smallest number of channels per node compatible with the
lattice local symmetries (see Section 1.3.1). Indeed, minimal models based on the
three cubic Bravais lattices (simple, body-centered, and face-centered) do not
satisfy the crystallographic isotropy beyond the third order. Moreover, minimal
models based on the two tetragonal lattices (simple and body-centered), the four
orthorhombic lattices (simple, base-centered, body-centered, and face-centered),
the two monoclinic lattices (simple and body-centered), the triclinic lattice and
the trigonal lattice, do not even satisfy the second order crystallographic isotropy.
The three-dimensional hexagonal Bravais lattice can serve as a basis for minimal
models which satisfy the second order crystallographic isotropy only if its aspect
ratio is equal to \/3/2, but even in this case, the fourth order condition is not
satisfied.
So, strategies must be developed to overcome this difficulty, and three-
dimensional lattice gas models can indeed be constructed with fully isotropic
large-scale dynamics. These strategies are discussed in the next three sections.
3.7.1 Models with multiple links

A first solution to the anisotropy problem is to broaden the scope of the search
by including models that are not minimal, such as models where propagation
connects nodes to nearest and next-nearest neighbors with different 'weights' on
the different connecting links. These weights can be set by using multiple-link
models, that is, models with b > B where two or more distinguishable particles
can have the same velocity vector,6 so that some velocity vectors are counted
twice or more. The price to pay for this recovered isotropy is an increase in the
number of channels per node, leading to more complex collision rules. With this
idea, one can construct several models based on the three-dimensional cubic
lattices (simple, body-centered and face-centered), with propagation allowed to
nearest and next-nearest neighbors via multiple links with multiplicities tuned
to obtain fourth order crystallographic isotropy.
To clarify this type of solution, we describe in detail its application to
the body-centered cubic lattice with unit lattice constant. We build the set of
connection vectors with the vectors connecting one node to its nearest neighbors:
(±5, ±\, ±5) and to its next-nearest neighbors: (±1,0,0), (0,±l,0) and (0,0,±1).
The connection number B is thus equal to 14 (8 vectors with modulus ^ , and
6 with modulus 1). We now consider that the links to nearest neighbors and
to next-nearest neighbors have multiplicity £1 and £2 respectively (£1 and £2
are positive integers). In other words, each of the 8 connection vectors with
modulus Y corresponds to £1 different channels, and each of the 6 connection
vectors with modulus 1 corresponds to £2 different channels. The second and
fourth order crystallographic tensors, fully contracted with an arbitrary vector
x = (xi,X2,X3) yield respectively:
b
£ > • x)2 = (2d + 2£2)(x2 + x\ + x\\ (3.13)
1=1
6
This is not in conflict with the exclusion principle, since particles with equal velocities
are assumed to be distinguishable.
and
rx) 4 =(|+2£ 2 )(*? + 4 + 4) + ^

(3.14)
Clearly, the right hand side of (3.13) is proportional to ||x||2 for any choice
of Ci and £2, but the right hand side of (3.14) is proportional to ||x||4 only if
^Ci + 2^2 = §Ci, that is, for d = 2^2- The easiest way to satisfy this condition
is to choose £1 = 2 and £2 = 1, which leads to b = 8£i + 6^2 = 22 channels per
node.
The same idea can be applied to the simple cubic lattice, with propagation to
nearest and next-nearest neighbors with suitable weights. It yields a model with
B = 18 connections (6 to nearest neighbors and 12 to next-nearest ones) and
b = 24 channels per node (each of the six links to nearest neighbors is a double
channel).
3.7.2 Models with biased collisions

D. Levermore (personal communication, 1986) suggested that large scale isotropy
can be obtained without multi-channel connections (and thereby less pro-
hibitively since then b = B) by introducing a bias in the collision rules such
that the statistical weight (mean population) of the channels corresponding to
velocity vectors pointing to nearest neighbors be twice the statistical weight of
the channels corresponding to velocity vectors pointing to next-nearest neigh-
bors. Strictly, such non-homokinetic models do not satisfy the fourth order
crystallographic isotropy, but only a 'statistically weighted fourth order crys-
tallographic isotropy', which can be sufficient for fully isotropic large-scale
dynamics.
3.7.3 The FCHC models

The most efficient and most used three-dimensional lattice gas models with
correct isotropy follow from a detour into four-dimensional space. d'Humieres,
Lallemand and Frisch (1986) noticed that there exists at least one Bravais lattice
in four dimensions that has the proper isotropy properties. This Bravais lattice
known as the F4 lattice in field theory, and is frequently called the Tace-centered-
hyper-cubic' (FCHC) lattice. It can be defined as the set of all points of R 4 with
signed integer Cartesian coordinates with even sum. The name FCHC follows
from the analogy with the three-dimensional face-centered cubic lattice which is
defined as the subset of points of R 3 with signed integer Cartesian coordinates

with even sum.7
The FCHC lattice can be generated by the four vectors (1,1,0,0), (1, -1,0,0),
(0,0,1,1) and (0,0,1,-1). The point symmetry group of the FCHC lattice
contains 1152 isometries which can be generated by combinations of:
(i) The four symmetries Si, S2, S3, and S4 with respect to the four hyper-planes
x\ = 0, X2 = 0, X3 = 0, and X4 = 0 (coordinate reversals),
(ii) The six symmetries Pn, P13, Pu, P23, P24, and P34 with respect to the
hyper-planes x\—X2 = 0, x\ — X3 = 0, xi — X4 = 0, X2 — X3 = 0, xi — X4 = 0,
and X3 — X4 = 0 (coordinate exchanges),
(iii) The symmetries Zi and E2 with respect to the hyper-planes X1+X2—X3—X4 =
0, and xi — X2 — X3 — X4 = 0 respectively.
This set of symmetries generates the point group G of the FCHC lattice, but it
is not a minimal set since only five of them, for example (Si,Pi2,Pi3,Pi4,£i) are
sufficient to generate G. A complete discussion on the symmetry group of the
FCHC lattice is given by Henon (1987a,b).
The minimal set of velocity vectors that can be built on the four generating
vectors is:
±1\ (-
±1 0
C5,...,8 =
0
o/ Vo (3.15)
o\
±1
f°
1
0
±1 » Cl7,...,20 =
o C21,.,24= + 1
0 / \+l
Here, the numbering of the C;S is arbitrarily based on the lexicographic order
of the alphabet (+1,-1,0). Note that for operational convenience in numerical
simulations, another order may be chosen in such a way that c;+i2 = — c* for
i= 1,..., 12. The reason for this choice is related to technical details concerning
the practical implementation of bounce-back reflections on solid obstacles (see
Chapter 10).
A three-dimensional lattice gas model can be constructed from the four-
dimensional FCHC lattice by the following dimension-reducing procedure:
(i) We consider the portion of the four-dimensional lattice limited by the two
hyper-planes X4 = +1 and X4 = — 1. This (hyper-)slice of lattice corresponds
7
The acronym FCHC is slightly ambiguous; in the Appendix, Section A.2, we show that
the lattice should be named 'two-dimensional face-centered body-centered hyper-cubic'
lattice. For simplicity and convenience, we nevertheless use FCHC as commonly
accepted.
to one lattice period in the fourth direction, and tiles exactly the FCHC
lattice by successive translations of vector (0,0,0,2).
(ii) We impose periodic conditions in the fourth direction; in other words, we
cancel the fourth component of the c,-s and project all the nodes in this
portion onto the hyper-plane X4 = 0.
Figure 3.10 illustrates this procedure for a simple case: the construction of a
one-dimensional lattice from a two-dimensional square lattice.
Figure 3.11 shows how the three-dimensional (pseudo-four-dimensional) model
so constructed is merely the three-dimensional cubic lattice, with double chan-
nels connecting nearest neighbors and single channels connecting next-nearest
neighbors. There are still b = 24 channels per node in three dimensions, but six
of them are double channels (B = 18).
The three-dimensional model constructed by projection of a four-dimensional
lattice can also be viewed as a multiple-link model based on the simple cubic
lattice with propagation to nearest and next-nearest neighbors with multiplicity
2 and 1 respectively. While it is easier to describe this model as a multiple-link
Figure 3.10 This picture illustrates in a simple case the section-projection

procedure to construct a lattice gas model in D — 1 dimensions from a model
in D dimensions. We start with the HPP two-dimensional lattice gas model
(see Section 3.1) and construct a one-dimensional model. The same procedure
can be used to obtain a three-dimensional model from the FCHC
four-dimensional Bravais lattice.
model (which requires no four-dimensional gymnastics), it is more convenient for

theoretical analysis to work with a model originating from a four-dimensional
homokinetic model (all C/S having equal moduli).
Note that the section-projection procedure can be applied once more to
construct a two-dimensional model. The two-dimensional model so obtained
has 20 moving particles and 4 distinguishable rest particles corresponding to
the four-dimensional velocity vectors C21 to C24 whose projection on the two-
dimensional space (X3 = 0, X4 = 0) is zero. This two-dimensional model is
equivalent to the 'modified' HPP model described earlier (see Section 3.1) with
four additional rest particles.
Figure 3.11 Three-dimensional lattice gas model obtained by projection of the

four-dimensional FCHC lattice. Single arrows represent vectors with
zero-component in the fourth dimension. Double arrows represent pairs of
vectors such as (1,0,0,1) and (1,0,0,-1) which differ only by their component
(+1) in the fourth dimension and whose three-dimensional projections are
thus identical. Hidden vectors are not labeled. Black dots represent the nodes
of the three-dimensional lattice.
3.7.4 Collision rules

Designing and implementing collision rules is much more involved in three di-
mensions than for two-dimensional lattice gas models. The most obvious source
of difficulties is the value of b which is generally larger in three dimensions. The
number of possible states per node grows exponentially (2b) with the number of
channels. Even after reduction using the invariance properties of the lattice, the
number of non-trivial cases to be treated remains quite large. For example, the
FHP-3 model (see Section 3.4) with 7 channels per node has 2 7 = 128 possible
states. If the model is to be made G-invariant, one can restrict the analysis to 26
states which cannot be interchanged by isometries of the point group G. This
number reduces to 13 if, in addition, the model must be self-dual. It is then easy
to design, optimize, store and finally implement collision rules conserving the
correct observables, by simply considering each case separately. For the three-
dimensional FCHC model with 24 channels per node, the number of possible
states rises to 2 24 = 16777216. When the model is assumed to be G-invariant,
the number of cases to study is still as high as 18 736, which reduces to 10 805
if self-duality is imposed (Henon, 1992). It is clear that 'hand-manufactured'
collision rules are no longer possible. The design and optimization of collision
rules for three-dimensional models require new strategies to handle the problem
in a global and systematic way.
Even had the design problem been solved, the storage and implementation
problem would persist. Indeed, since the collision operator A(s^sf) is formally a
2b x 2b matrix, the linear size of the complete information to be stored for run-
time grows exponentially with b. The full collision matrix for a non-deterministic
model with 24 channels per node would represent 2 2 x 2 4 ~ 2.8 x 1014 real numbers
between 0.0 and 1.0, which by far exceeds the random access storage capacity
of the most powerful computer systems presently known, even if each entry
is stored on eight bits only. As for the design, new strategies are required to
drastically reduce the amount of data to be stored and accessed at run-time.
Solutions exist which overcome the problem of designing and implementing
collision rules for three-dimensional models. These solutions belong to three
classes: algorithmic strategies, look-up table strategies and compound strategies.
3.7.4.1 Algorithmic strategies

These strategies amount to storing (almost) nothing, and computing the post-
collision state for each node at each time step, according to a suitable algorithm.
For the FCHC model, a collision rule involving an algorithmic strategy was
introduced by M. Henon (1987a). It is called the 'isometric algorithm' because
it uses lattice-preserving isometries in a suitable way to perform collisions. This
strategy has been implemented and used for validation tests on the FCHC
lattice gas model (Rivet, 1987a), but because of its lack of numerical efficiency,
the isometric algorithm was not further exploited.
3.7.4.2 Look-up table strategies

For deterministic models, a unique post-collision state corresponds to each pre-
collision state. So the collision matrix contains only Os and Is, and each line
(or column) contains only one non-zero element. Therefore, instead of storing
2b x 2b real numbers between 0.0 and 1.0, it suffices to store only the unique
post-collision state assigned to each of the 2b pre-collision states. The storage
of this look-up table requires b x 2b bits, which is still a fairly large number for
most three-dimensional models. For instance, the look-up table for the simplest
FCHC model (with b = 24) requires about 50 megabytes storage capacity,
which represents a non-negligible part of the random access memory of most
supercomputers. These strategies involving a pre-computed look-up table are
very 'memory-consuming', but potentially rather fast (see Chapter 10). The
construction of the table can be quite time-consuming, but it is done once for
all.
Strategies based on pre-computed look-up tables can be adapted to 'weakly
non-deterministic' models by using more than one look-up table, with random
call to each table at each time step and possibly at each node. This solution
tends to become heavy and prohibitively bulky if the choice involves a large
number of tables.
A considerable drawback of deterministic look-up table strategies is that it
may be difficult to meet the important G-invariance property. A careful design
of the look-up table can however yield some 'approximate G-invariance' by
introducing some 'frozen randomness' in the design of the table.
3.7.4.3 Compound strategies (or reduced-table strategies)

Compound methods use features such as symmetries and/or duality to reduce
the number of entries in the look-up table. These strategies combine the advan-
tages and drawbacks of both look-up table and algorithmic strategies. They are
less 'space-consuming' than straight look-up table methods, since only a reduced
table has to be stored. On the other hand, they are more 'time-consuming' since
for each node some computation is required to recast the pre-collision state into
an entry of the reduced table.

As a consequence of the quite large number of non-trivial collisions in the
FCHC model, the microdynamical equation cannot be written, in practice, in an
extensive form as for HPP or FHP models. Of course, the concept of an exact
equation governing the FCHC microscopic dynamics still holds, but its explicit
form would involve more lines than the present book contains; the only useful
expression is the general form given by (2.12) or (2.13).

Several FCHC models can be designed with the same geometry and the same
connection vectors, but with different collision rules, and various numbers of rest
particles (i.e. with different numbers of channels per node). The simplest FCHC
models have 24 channels per node (no rest particles). More complex models
can have up to several tens of indistinguishable rest particles per node, in
addition to the 24 moving particles (see comment on indistinguishable particles
in Section 1.3.3). Only a few of these possible FCHC models have actually
been implemented and used for simulations. The interested reader will find in
Table 3.3 a summary of the microscopic properties of FCHC models. Most
important is that, independently of these specific properties, the whole FCHC
class exhibits the following properties:
(i) The class structure may depend on the number of rest particles. However,
the 24 velocity vectors of the moving particles always belong to the same
class. Indeed, let us take ci (see Equation (3.15) for the definition of the C;
vectors). It is clear that C2 to C4 belong to the same class as ci, since the
application of combinations of the symmetries Si and S2 to ci generates
these three vectors. The application of P23 to vectors ci to C4 generates
vectors C5 to eg, and so on. Thus, any vector ci to C24 can be changed into
any other vector c* by an isometry in G, which proves that all 24 vectors
belong to the same class.
(ii) The mass and the four components of the momentum are the only linearly
independent collisional invariants. The energy observable is proportional
to the mass observable and so is trivially conserved for models without
Table 3.3 Microscopic properties of the most important FCHC models.

The models with b = 26 admit up to three indistinguishable rest particles
coded with two Boolean quantities (see comment in Section 1.3.3).
Label b Strategy Deterministic S.D.B. Duality G-invar.

FCHC-1 24 Algo. No Yes Yes Exact
FCHC-2 24 Table Yes Yes Yes Approx.
FCHC-6 24 Table Yes No Yes Approx.
FCHC-7 26 Table Yes No Yes Approx.
FCHC-8 26 Table Yes No No Approx.
rest particles. Energy conservation in FCHC models with rest particles is

subject to the discussion presented for FHP-2 and FHP-3 (see Sections 3.3.2
and 3.4.2).
(iii) The models have up to fifth order crystallographic isotropy. Table 3.4 gives
the values of the crystallographic tensors of order 1 to 6, fully contracted
with an arbitrary vector.
(iv) The FCHC models are irreducible. Indeed, let us consider ci. The sub-group
of G which leaves ci unchanged contains, among others, the isometries S3,
S4 and Pn (see Section 3.7.3). The only vectors which are invariant under
S3, S4 and Pn have components X3 and X4 equal to zero, and components xi
and X2 equal one to each other. They are thus proportional to ci. The same
argument holds for vectors C2 to C24. For the velocity vectors of rest particles
(if any), the argument developed for the FHP-2 model is applicable.
The reader may anticipate the following question: in order that three-
dimensional lattice gas large-scale dynamics be governed by the incompressible
fluid Navier-Stokes equations, the collisional invariants must be the mass, the
energy, the three components of the momentum, and no other quantity ; but what
about the fourth component of the momentum which is also conserved in FCHC
models ? In fact, this additional invariant is not spurious, as it corresponds to an
additional macroscopic variable with its own large-scale dynamics, but which
does not perturb the large-scale dynamics of the three-dimensional relevant
macroscopic quantities. Indeed, consider a fully four-dimensional FCHC model
with large-scale collective dynamics governed by the four-dimensional Navier-
Stokes equations. Now, when cutting a (hyper-)slice in the four-dimensional
lattice and wrapping it onto itself in the fourth dimension (see Section 3.7.3),
the macroscopic variables no longer depend on the fourth component X4 of
Table 3.4 Crystallographic isotropy properties of the FCHC

models. The second column gives the result of a full contraction of
the nth order tensor with an arbitrary test vector x. The third
column summarizes the conclusion about isotropy.
Order rc £ Tj^x^ ...Xttn Status

isotropic
1 0
12
2
3 0
0> 2
(=12|W| ) isotropic
isotropic
4 nhr aX4
5 0 isotropic
6 12(5
l^ll 6) anisotropic
the position. As a result, the four-dimensional Navier-Stokes equations separate

into (i) the three-dimensional Navier-Stokes equations for the three-dimensional
variables, and (ii) an advection-diffusion equation for the fourth component of
the momentum density, which behaves as a passive (non-perturbing) diffusive
scalar advected by the three-dimensional flow. This 'charge-free' passive scalar
can be considered as a bonus offered by pseudo-four-dimensional models.
Chapter 4
Equilibrium statistical mechanics
In Chapter 2, we established the equations governing the microscopic dynamics

of the lattice gas. This microscopic dynamics provides the basics of the pro-
cedure for constructing automata simulating the behavior of fluid systems (see
Chapter 10). However, in order to use LGAs as a method for the analysis of
physical phenomena, we must go to a level of description which makes contact
with macroscopic physics. Starting from a microscopic formalism, we adopt the
statistical mechanical approach.
We begin with a Liouville (statistical) description to establish the lattice Liou-
ville equation which describes the automaton evolution in terms of configuration
probability in phase space. We then define ensemble-averaged quantities and,
in particular, the average population per channel whose space and time evo-
lution is governed by the lattice Boltzmann equation (LBE), when we neglect
pre-collision correlations as well as post-collision re-correlations (Boltzmann
ansatz). The LBE will be seen to play in LGAs the same crucial role as the
Boltzmann equation in continuous fluid theory; an H-theorem follows from
which the existence of a Gibbsian equilibrium distribution is established when
the semi-detailed balance is satisfied.
4.1 The Liouville description

In Chapter 2 we introduced the notion of Boolean field 7t;(r*, t*) to obtain
a complete microscopic description of the time-evolution of lattice gases. We
also introduced the phase space T which is the set of all possible Boolean
68
4.1 The Liouville description 69
configurations of the whole lattice. The Boolean configuration of the lattice is

a collection of Jf x b Boolean quantities (with values 0 or 1) which encodes
the full microscopic state of the Jf nodes of the lattice, with b channels per
node. The Boolean configuration of the lattice is called its 'microstate' because
it describes the state of the whole lattice at the microscopic level.
Now when we are interested in macroscopic behavior, such a detailed descrip-
tion in terms of microstates is generally unnecessary and often even useless. To
describe the macroscopic states we need averaged quantities. Following the same
line of reasoning as in classical statistical mechanics, it is logical to introduce a
statistical description of lattice gases in terms of its 'macrostates'.
4.1.1 Macrostates
Most common laboratory probes (thermometers, pressure sensors, velocimeters,
etc.) yield time- and space-averaged measures, simply because they have a finite
macroscopic size and a finite response time, that is, they operate on a length
scale and a time scale which are large compared to the characteristic microscopic
lengths and times. For theoretical developments however, it is logical and
convenient, following Gibbs, to consider ensemble-averages (Huang, 1963). This
procedure requires the ergodic hypothesis assuming that ensemble-averages
correctly describe macroscopic observables which are in practice time- and
space-averages.
Formally, we define a 'macrostate' of the lattice as a function acting on the
phase space F, such that it associates to each microstate S e F a real number
between 0 and 1:
*: F - [0,1]
V }
with the normalization property:
1. (4.2)
SGT
Because of this normalization property, the function &> defining a macrostate

can be viewed as a probability distribution function. The concept of macrostate
is the basic tool for the Liouville (statistical) description of lattice gases.
4. l .2 Ensemble - averages
Having defined a probability measure in the phase space, we can now define
ensemble-averaged quantities. Consider an arbitrary generalized observable (see
Section 2.1.4), that is, a physical quantity whose measure at node r* at time
U is an arbitrary function q(n,u) of the Boolean field n, and possibly of the
70 4 Equilibrium statistical mechanics
discrete space-variable. We define the 'ensemble-averaged' value (q) (r*) of the

generalized observable q measured at node r* as:
*)P(S). (4.3)
SeT
Consider now an ordinary observable1 q = (qt, i = l,...,fr) (see Section 2.1.3).

Its value measured at node r* is X^g*n*(r*)- The ensemble-averaged value reads:
b
SeT i=l
which reduces after summation inversion to:
b
, (4.4)
where (rc^r*)) is the (ensemble-)averaged population of channel i at node r*:
(mM) = ^2Si(u)^(S). (4.5)

SeT
We have now at our disposal the concept of ensemble-averaged quantities,

which are defined for an arbitrary probability measure (macrostate) & with no
restricting hypothesis, except for the normalization condition (4.2). The next
step in the statistical description of lattice gases is the derivation of the master
equation that governs the time-evolution of the macrostate ^ .
4.1.3 The lattice Liouville equation

When the lattice gas evolution rule is applied, the probability distribution
function is a priori time-dependent. The time-evolution of 3P is governed by
a relation that can be established from the following arguments. The collision
phase acts independently at each lattice node, and changes the state S(r*) of
node r* into the state S'(r*) according to the transition probability ^4(S(r*)->S'(r*))
(see Chapter 2). So, the conditional probability of obtaining the microstate S'
after the collision phase, knowing that the microstate was S before collision,
is Y[A(S(r*)-+Sf(r*)), where the product extends to all lattice nodes r* G JSf. We
now assume that before collision, the statistics of the lattice can be described
by the macrostate gP(S,U). So the probability to find the microstate S' after
collision is ^@>(S,U)Y[A(S{**)-*Sf{**)), where the sum is taken over all possible
microstates SeT. Since the propagation operator 5^ (notation introduced in
Chapter 2) is a fully deterministic operator, the probability of finding the lattice
in the microstate 5^(S") after the propagation phase is equal to the probability
1
An ordinary observable is a linear, local and homogeneous generalized observable (see
Section 2.1.4).
4.2 The Boltzmann description 71
of finding the lattice in the microstate Sf before propagation. We then obtain a

Chapman-Kolmogorov equation for the time-evolution of the macrostate:
9 {Sf (S'), U + 1) = Y^ & (s> **) I I ^(S(r*HS'(r*)), VS' G T. (4.6)

SeT ue^
Although formally different from the Liouville equation for continuous systems,
Equation (4.6) plays an equivalent role for lattice gases: it governs the time-
evolution of the probability measure in phase space T. We therefore call (4.6)
the lattice Liouville equation'.
We now address the problem of finding the equation governing the time-
evolution of the ensemble-averaged populations of each channel at each node.
4.2 The Boltzmann description

The probability distribution function 9 contains the full statistics of the whole
lattice, and the lattice Liouville equation which describes the evolution of 9 is
an exact equation. However, the ensemble-averaged populations of each channel
of each node are more directly useful than the full statistics.
For consistency with the usual notation in statistical mechanics, we denote by
/(Cj, r*, £•), or equivalently /,-(r*, U\ the ensemble-averaged populations (n,-) (r*,t*).
These ensemble-averaged populations play the same role as the single particle
distribution function /(p, r, t) in kinetic theory of continuous gases. Indeed, for
continuous gases, f(p,r,t)d3rd3p is the average number of particles present at
time t in a volume element d?r about r and within the volume element cPp
about p in momentum-space. For lattice gases, /(c;,r*,£*) is the average number
of particles present at time U at node r* with velocity c,- (no volume element
is needed for lattice gases since both position space and momentum space are
discrete).
4.2.1 The Boltzmann approximation

In classical statistical mechanics, one cannot derive an exact and closed equation
for the time-evolution of the single particle distribution function /(p, r, t) (see
Huang, 1963). The difficulty is bypassed by using the Boltzmann approximation
(the molecular chaos hypothesis) to obtain the closed Boltzmann equation. The
situation is very much the same in lattice gases. 2 We must assume that the
2
There is at least one exception: the diffusive lattice gas model constructed as a collection
of random walkers on the lattice (Boon, Dab, Kapral and Lawniczak, 1995) for which
an exact lattice Boltzmann equation is obtained without the factorization assumption.
However, one can argue that the assumption is contained in the model itself through the
randomization of the channel occupations at each time step, a procedure used to model
the random walk of the particles on the lattice.
propagation phase shuffles the particles in such a way that their correlations
vanish at each node, so that the pre-collision (post-propagation) state of each
node be totally factorized. Obviously, right after a collision, particles are in a
strongly correlated (post-collision) state precisely because of the effect of the
collisions, but we assume that during the subsequent propagation phase, these
correlations are completely damped out.
The Boltzmann approximation, for lattice gases as for real gases, is a strong
hypothesis whose validity is expected to hold at sufficiently low gas density.
For lattice gases however, the approximation is rather well verified even at
moderate densities, at least for models satisfying semi-detailed balance (like the
GBL model and the FCHC-3 model; see Chapter 7).
4.2.2 The lattice Boltzmann equation

By ensemble-averaging both sides of the microdynamic equation (2.12), we
obtain the evolution equation for /,-(r*,£*):
/,(r
s,s'ey
where the rij s on the r.h.s. are implicitly taken at time U and node r*. For clarity,
these arguments are not written explicitly, and will be omitted systematically
when no ambiguity results. We also use the more compact notation ft(r*,t*)
rather than /(c,-,r*, £*), which is formally closer to the classical notation /(p,r,t),
keeping in mind that /;(r*,£*) and /(c;,r*,f*) denote the same object: (w,-(r*,**)).
To arrive at (4.7), we have used the fact that the £(s-»s')s in (2.12) are random
Bernoulli variables, statistically independent from the Boolean field, and whose
averages are the ^4(s-»s')s.
Equation (4.7) is an exact equation but, as it stands, there is no way to express
explicitly the r.h.s. in terms of the fts. Therefore, we now use the Boltzmann
approximation (see Section 4.2.1), that is, we make the assumption that the
pre-collision configuration is totally uncorrelated. With this assumption, the
ensemble-average of the product on the r.h.s. of (4.7) can be factorized into the
product of the ensemble-averages, to obtain:
ft(u + c,, U + 1) = £ 44S-J) I ] ff'ff', (4-8)

s,s'ey j=l
where / ; - stands for 1 — / ; . This equation governs the time-evolution of the

averaged populations ft. It plays the same role as the Boltzmann equation in
4.3 The H-theorem 73
kinetic theory of continuous gases, and is therefore referred to as the lattice

Boltzmann equation'. 3
In the same way as there are alternative versions of the microdynamic equa-
tion, (2.12) and (2.13), the lattice Boltzmann equation can be rewritten as:
b
fi(T* + C;, U + 1) - /;(r*, U) =Y,^'i- S/MM) I I / / ' / A (49)

s,s'ey 7=1
The equivalence between (4.8) and (4.9) is an immediate consequence of the

identity:
fj', (4.10)
s,s'ey 7=1
which follows directly from the normalization condition (2.9).

The analogy with the classical Boltzmann equation is more visible from (4.9):
the l.h.s. is the first order finite difference expression of the convective derivative
(dt + v • V)/, and the r.h.s. is a collision term which plays a role similar to the
collision integral (see Huang, 1963). The main difference is that for ordinary
gases, only binary collisions are taken into account, so that the collision integral
is quadratic in the distribution function / , whereas for lattice gases collisions
involving up to b particles are admitted. Consequently, the collision term for
lattice gases is a polynomial of degree b in the /,-s.
Following Boltzmann's reasoning, we now prove an 'H-theorem' for lattice
gases satisfying semi-detailed balance. Namely, we show that one can define a
scalar function H of the average populations /,-(r*) that can only increase or
remain constant. The important consequence of the if-theorem is the existence
of equilibria (Section 4.4).
4.3 The if-theorem
Usually one presents the H-theorem - in the context of ordinary gases in

continuous space-time - by introducing the functional H of the single particle
probability distribution f,H = —k^ f flogfd 3pd3r, and one shows that this
quantity cannot decrease if the time-evolution of/ is governed by the Boltzmann
equation (Huang, 1963). The functional H is interpreted as the statistical entropy
of the system and the equilibrium states follow from the maximization of H
under constraints related to conservation laws.
Here, we shall steer a different - although equivalent - course, which makes a
3
The lattice Boltzmann equation can also be used - quite independently of its underlying
Boolean microscopic dynamics - as the starting point for the simulation of fluid
dynamics (Succi, Benzi and Higuera, 1991; Succi, 2000); see also Section 11.5 in
Chapter 11.
detour via the theory of communication as established in 1948 by C.E. Shannon

and subsequently formalized in a more rigorous presentation by A.I. Kinchin
(Kinchin, 1957). This presentation may seem rather abstract at first sight, but it
leads to a nice intuitive interpretation.4
We first briefly review the basic ideas of Shannon's theory of communication
and postpone the connection with lattice gases to Section 4.3.2.
4.3.1 Some basics about communication and information

The goal of Shannon was to provide a global mathematical description of
communication processes. A communication process includes the generation
of a message and its encoding, the (possibly noisy) transmission through a
physical channel, and its reception and decoding. No quantitative theory could
be built about communication without an accurate tool to measure the amount
of information contained in a message. We shall restrict our short description to
this crucial notion of information measurement, because it is the most relevant
feature for lattice gas statistics. The interested reader will find the complete
theory in Shannon and Weaver (1969), and in Kinchin (1957).
4.3.1.1 The deterministic case

We introduce the notions of 'quantity of information' and of 'entropy of infor-
mation', by considering first the simple case of a variable x whose value can be
either x(1) or x(2). Suppose for example that the value of this variable is known
to be x(1), then the quantity of information we have about the variable x is one
bit, by definition. In addition, since we fully know the value of x, no freedom
of choice is left concerning x. Therefore, we say that the freedom of choice,
which Shannon calls the 'entropy of information', is zero bits. The 'bit' 5 is the
standard unit of measure for the quantity of information. Conversely, if we do
not know anything about the value of x, then we say that we have zero bits of
information about x, and that the freedom of choice (entropy of information)
concerning x is one bit. The word entropy is used here by analogy with the
concept of entropy in thermodynamics, since in some sense it is a measure of
the 'disorder' of the system in both situations.
Consider now a discrete variable x which can take any of the n values
l
x^ \ i = l,...,rc. (Superscripts between parentheses label the possible values
taken by the variable x; subscripts will be used in further sections to label
different variables.) Suppose that an incomplete measure of x revealed that the
actual value of the variable x can be either x(1) or x^ or ... or x(m) (m < n). We
4
The material in this section is inspired by a work by M. Henon, published in an
appendix of the article by Frisch, d'Humieres, Hasslacher, Lallemand, Pomeau and
Rivet (1987). Henon takes the traditional route to present the H-theorem, and uses
information theory only as an interpretation.
5
The word 'bit', introduced by J.W. Tuckey, is the contraction of 'binary digit'.
4.3 The H-theorem 75
define the 'entropy of information' H associated with this incomplete measure

of x, as the logarithm to the base 2 of m: H = log 2 m. We also define the
'quantity of information' given by the incomplete measure as / = log 2 n — log2 m.
Note that if m = 1, the value of x is fully determined, and therefore the entropy
is zero and the quantity of information takes its maximum value log 2 n. On
the contrary, if m = n, the value of x is totally unknown; then the entropy
takes its maximum value log 2 n and the quantity of information is zero. We
emphasize the analogy with Gibbs' entropy: the entropy associated with a state
of a macroscopic system is proportional to the logarithm of the volume of phase
space actually accessible to the system. In the context of information theory, the
entropy is proportional to the logarithm of the number of accessible values of
the discrete variable x.
We can also give a more operational interpretation of the notion of freedom
of choice (entropy), by using the language of computer science: assume that the
number n of possible values of x is an integer power of 2, say 2b. Suppose that an
incomplete measure reveals that x can actually take 2C of the 2b a priori possible
values (c is an integer smaller than b). Then, the number of computer memory
bits necessary to describe the possible choices remaining after the measure is
log2 2C = c. Thus, the entropy of information H associated to this incomplete
measure of x is just the number of memory bits required to store the remaining
free choices. Here, the word 'bit' is used in the context of computer memory
rather than in the context of information theory, but the meaning is essentially
the same. The only difference is that, in computer science, memory sizes must be
an integer number of bits, whereas this constraint does not exist in information
theory, since the number of possible values of a variable need not necessarily
be an integer power of 2.
4.3.1.2 The non-deterministic case

Next comes the case of variables for which the only available information
is statistical. Consider as before a discrete variable x which can take any of
the n values x w , i = l,...,w, but now the only information we have at our
disposal is the set of probabilities p^l\ i = 1,..., n for each one of the n possible
values x^l\ i = l,...,n. Shannon shows that under some reasonable constraints,
the only possible measures of the freedom of choice and of the quantity of
information associated with this statistical knowledge about x are, respectively:
/ =klog2n-H,
where k is an arbitrary positive constant that will be chosen equal to one if H is
to be expressed in bits. We shall not reproduce Shannon's proof, but rather use
a simple argument to convince ourselves that this measure is 'reasonable' and
consistent with the intuitive and natural results of the deterministic case. Indeed,
suppose that an incomplete measure reveals that x can take indifferently any
of the values x^\..., x(m) (m < n), and no other value. This partial information
is equivalent to the statistical information that the probabilities p ( 1 ) ,...,p ( m ) are
equal to 1/m, and p(m+1\...,p^ are zero. With this set of probabilities, (4.11)
yields H = log 2 m, which is consistent with the result presented above for the
deterministic case.
4.3.1.3 Correlated random Bernoulli variables

With the context of lattice gases in mind, we now address the problem of the
quantity of information and entropy related to a collection X = (xt, i = 1,..., g)
ofq Boolean variables (with values 0 or 1) for which only statistical knowledge is
available. The composite variable X thus takes 2q possible values. We denote by
F the set of all these 2q values. Suppose that we know the statistics of X, namely
the probabilities 0>(Y) that X takes the value Y = (yu i = l,...,q) G T. Then,
the entropy and the quantity of information associated with this knowledge are,
respectively:
YeT
I =log22q + ^0>(Y)log20>(Y) (4.12)
Yer
= q-H.
Suppose now that the only available knowledge is not the full statistics of
X, but the individual probabilities pi, i = l,...,q that x\ take the value 1.
According to (4.11), the entropy and the quantity of information associated
with the knowledge of the individual variable xio are, respectively:
H
k = ~ I Pio lo g2 Pio + (! - Pio) loS2 (* ~ Pio) 1»
L J
(4.13)
since x;0 is a Boolean variable which can take two values: 1 or 0, with respective
probabilities pt and (1 —pi). The entropy and the quantity of information associ-
ated with the knowledge of all the individual probabilities are thus, respectively:
q (4.14)
I* =^2l+ [Pi loS2 Pi + Pi l°g2 P]
where pt stands for 1 — p,-. Since the knowledge of the individual statistics does
not account for possible correlations between the variables, it is logical to expect
that it contains less information than the knowledge of the full statistics of X.
The proof for this quite intuitive statement is given in the Appendix, Section A.3.
We now apply these concepts to lattice gases.
4.3.2 The H-theorem for lattice gases

Consider a lattice gas with b channels per node, residing on a finite portion <£
of a Bravais lattice, containing Jf connected nodes. At any discrete time £*, this
lattice gas can be seen as a collection of b x Jf Boolean variables. We focus
attention on a particular node r* for which we adopt a Liouville description:
the information available about the state of this node is the time-dependent
local probability distribution function p(r*,£*,s). At any discrete time U, we can
associate to this local statistical knowledge the following 'local entropy':
sey
Notice that we now use natural logarithms rather than logarithms to the base 2.
(The difference is a non-significant multiplicative constant that we shall ignore.)
To simplify the notation, we shall omit the arguments r*, U in p(r*, £*, s).
The effect of the collision phase on the local probability distribution function
p is a linear transformation involving the collisional transition probabilities
v4(s->s') (see Chapter 2), that is, the post-collision local probability distribution
function pf can be expressed as:
M). Vs'ey. (4.16)

sey
We now use a basic property of convex functions: for any convex function 0,
the following inequality holds:
¥ , 6 r ,
where the sums implicitly extend over all 5 in 7. If we assume that the semi-
detailed balance condition ^2sey A(s^s') = 1 is satisfied, then, after summation
over all s', (4.17) reads:
( )
s'ey \sey / s'ey sey
Using (4.16) and the normalization condition J2S'ey ^(s-Ô = 1> w e obtain:
s'ey sey
Equation (4.19), applied to the convex function (j)(x) = —xlnx, implies that:
(4.20)
s'ey sey
This result proves that under the assumption of semi-detailed balance the
collision phase increases the local entropy of the lattice gas, and therefore
decreases the quantity of statistical information we have about the state of the
lattice node.
Note that without the semi-detailed balance assumption the collision phase
can decrease the local entropy of the node. As a simple example, consider
a lattice gas model with a collision rule such that the post-collision state of
any collision is always the same state so, whatever the pre-collision state. This
unphysical model does not satisfy the semi-detailed balance condition. Suppose
that, initially, this lattice gas is in a macrostate such that all configurations are
equally probable, namely: p(s) = 2~b, for all s in y. Then, the local entropy
initially has its maximum value b x In 2. After the collision phase, the lattice is
in a completely determined configuration, where all the nodes are in state so-
Thus, the local entropy after the collision phase has decreased to its minimum
value: zero. We now make the Boltzmann approximation, i.e. we neglect the
correlations between particles entering a collision. Then, the local entropy at the
node before collision is equal to the sum of the entropies in each channel:
b
- £ [/,In/,, +/,ln/J = -5>(s)lnp(s). (4.21)
1=1 sey
Now, there is no reason that the post-collision state be uncorrelated; so, from
the results of Section 4.3.1.3, we have, after the collision:
b
- ^ [/>/;+/>/;] > -][>V)lnpV), (4.22)

i=l s'ey
Combination of (4.20), (4.21) and (4.22) leads to the local form of the if-theorem:
b b
- J2 [fi in/;+ft in/;] > - E if*ln^+ft ^ / J > <4-23)

i=l i=l
where the average populations are taken at node r*.

Local H-theorem: Under the Boltzmann assumption and with the semi-detailed
balance condition, the local quantity
b
h(u9 t*) = -J2 [/,-(r*, t*) ln/i(r*, U) + /,-(r*, **) ln/,-(r*, U)]
i=l
cannot decrease under the effect of the collision phase. This quantity is called the
local entropy.
We now introduce the 'global entropy' H which is the sum of the local
entropies h(u), extended over all lattice nodes r*. If the lattice is either infinite
or finite with periodic boundary conditions, the propagation phase only moves
information from node to node without loss or gain. It thus has no effect on
the global entropy. As a consequence, the local if-theorem implies that the
global entropy cannot decrease under the action of the full evolution operator
(collision and propagation).
Global if-theorem: Under the Boltzmann approximation and with the semi-detailed
balance condition, the global quantity
, **) +fi(r*9t*)]nfi(u9t*)]
x+eS£ i=\
cannot decrease under the action of the lattice gas evolution rule, at least if the
lattice is either infinite or finite with periodic boundary conditions. This quantity
is called the global entropy.
This theorem, proved under the Boltzmann approximation, implies that the
evolution of the statistics of a lattice gas is irreversible.
Let us illustrate the if-theorem with a simple example. We consider a lattice
gas residing on a finite lattice with periodic boundary conditions and prepared in
a homogeneous macrostate. In the homogeneous macrostate, the global entropy
H(u) is merely Jf times the local entropy /i(r*,£*) which no longer depends on
the position variable r*. So the statement that H cannot decrease when semi-
detailed balance is verified, at least within the Boltzmann approximation, also
applies to h. Suppose now that we prepare the lattice gas in a non-equilibrium
macrostate, for instance, by biasing in a systematic way one of the collision
rules; then, at some later time, we restore the usual rules (satisfying detailed
or semi-detailed balance), and in the course of time, we perform successive
measurements which give us an estimate of the entropy. We observe, as shown
in Figure 4.1, an increase in the numerical estimate of h(U) (see Tribel and Boon,
1997) towards an asymptotic plateau corresponding to the equilibrium value as
predicted by Boltzmann's theory. Now, in Figure 4.1, we also observe that on
top of the global increase of h(u) (in accordance with Boltzmann's theory) there
are fluctuations where h can decrease locally; the reason is that the Boltzmann
decorrelation hypothesis is only an approximation.
Along the same lines of reasoning, but without the Boltzmann approximation,
one can also prove a weaker form of the H-theorem:
Liouville if-theorem: With the semi-detailed balance condition, the quantity
cannot decrease under the action of the lattice gas evolution rule, at least if the
lattice is either infinite or finite with periodic boundary conditions.
Now that we have established the if-theorem, the next step in the statisti-
cal mechanical description of the lattice gas is to investigate the existence of
equilibrium situations, that is, of macrostates for which all statistically relevant
quantities are stationary.
4.4 Global equilibrium macrostates
Hardy, de Pazzis and Pomeau (1973) showed that the HPP lattice gas model
has very simple statistical equilibrium solutions that they call 'invariant states'.
These states have homogeneous (node-independent) properties, and they show
no equal time (static) correlations between channels of any pair, whether on
the same node or not. Such global (homogeneous) equilibrium macrostates are
analogous to the Maxwellian equilibrium states of real gases. We shall prove
that the equilibria, for a quite general class of lattice gases (including the
HPP model), are not characterized by the Maxwell-Boltzmann distribution of
1.960
1.955 -
1.950 -
1.945 -
1.940 -
1.935 -
1.930
20 40 60 80 100
U
Figure 4.1 The local entropy h{U) = - ^ = 1 [/*(£*) ln/i(r*) + /,(£*) In/,(£*)] as
a function of time, for an FHP-1 lattice gas model on a lattice with 256 x 256
nodes, u is in the natural microscopic time unit: the automaton time step.
(From Tribel and Boon, 1997.)
classical gases, but by the Fermi-Dirac distribution. As physical intuition would

indicate, the equilibrium macrostates (also called 'invariant measures') play a
crucial role in the theory of large scale collective motions.
In order to make the problem more easily tractable, we shall not attempt
to find and characterize all possible equilibrium macrostates. Rather, we shall
restrict the scope of our search to global (homogeneous) factorized macrostates,
that is, macrostates whose statistical properties are node-independent, and for
which no correlations exist between nodes and between channels. We shall adopt
three different, although equivalent approaches:
K7 The Liouville approach: the equilibrium macrostates are characterized by

their probability distribution functions ^(eq), which are required to be
stationary solutions of the lattice Liouville equation.
& The Boltzmann approach: the equilibrium macrostates are characterized
by their averaged populations ffq\ which are required to be stationary
solutions of the Boltzmann equation.
• The variational approach: the equilibrium macrostates are characterized
by their averaged populations ffq\ which are required to maximize the
global entropy of the lattice, under conservation constraints.
4.4.1 The Liouville approach

We look for stationary solutions of the lattice Liouville equation among the
restricted class of homogeneous and totally factorized distributions of the form:
where the /,-s should be taken for the moment as b real numbers between 0 and
1, and where overlined quantities ft and S,-(r*) denote the complements: 1 —/,,
and 1 — S/(r*), respectively. Note that the distribution (4.24) can be interpreted
as the probability distribution of a collection of b x Jf independent random
Bernoulli variables with mean values /,-, i = 1,..., b which depend on the channel
index i, but not on the node position r*. The expression (4.24) is obviously the
most general form for a fully factorized and spatially homogeneous equilibrium
measure.
Notice that the identity:
b
ft = J2 S^ I I I I fjS^f)fjSs^\ Vi = 1,..., b, Vr* e if, (4.25)
ser x*'ese ;=i
which follows straightforwardly from simple algebra, means that the probabili-
ties /; in (4.24) are the ensemble-averaged populations (n,-) of each channel.
A homogeneous factorized probability distribution such as (4.24) is completely

determined by b real numbers: the averaged populations /,-. In order that this
probability distribution be an equilibrium distribution, these b real numbers
must verify the following set of 2^bxjr^ relations (stationary lattice Liouville
equation):
9 {&> (Sf)) = J2 &$) I t 4S<"HSV)) > VS' G T. (4.26)

SeT r*e&
To proceed, we must assume that the lattice if is either infinite or finite with
periodic boundary conditions (see Chapter 2). Under this assumption, all the
nodes of the lattice are equivalent, and the free streaming operator only transfers
information from node to node without any loss or gain. Since the equilibrium
solutions are supposed to be spatially uniform (the /,-s do not depend on the
node position r*), it becomes clear that 9(3f(S')) = &(S'), for any microstate
Sf in F. If we now look for fully factorized solutions of the form given by (4.24),
we must solve the following set of equations:
n n / w = E n 4^0 n//'/A ê r> ^^

r*er i=l Serr*ej^ j=\
where S and S' are implicitly taken at node r*. As shown in the Appendix,
Section A.4, the above set of 2bxjr equations is strictly equivalent to the set of
2b equations:
i=l sey
f j=i
The proof of equivalence does not involve any restricting hypothesis on the
lattice gas model. Furthermore, the equivalence is rather intuitive and natural
since (4.28) can be viewed as a local (single-node) version of (4.27). The physical
content of (4.28) is that the single-node equilibrium probability distribution
Ilf=i fiSift is kft unchanged by the application of the collision operator. Indeed,
the l.h.s. of (4.28) is the probability of finding the node in state sf before collision,
and the r.h.s. can be understood as the probability of finding the node in state
s' after collision.
At this point, it is not obvious whether the set (4.28) has a solution, since
it imposes 2b conditions to only b variables (the /,-s). But we shall show that
(4.28) reduces to at most (and in fact less than) b independent relations. This
'little miracle' requires the assumption of the semi-detailed balance property
(Chapter 2). With this assumption, (4.28) is strictly equivalent to the following
set of b equations:
//' =0, Vi=l ft, (4.29)

s,s'ey 7=1
which is itself equivalent to
• - Si)A(s-+s f) = 0, Vs, s' e y. (4.30)

i=l
Proving the equivalence between (4.28), (4.29) and (4.30) is not a trivial matter;
the proof involves a trick introduced in the context of gas models with discrete
velocities (Gatignol, 1975), and is given in the Appendix, Section A.5.
Before proceeding any further, it is useful to discuss the physical content of
(4.29) and (4.30). Equation (4.29) expresses that for a global equilibrium, the
ensemble-averaged population of any channel is left unchanged by collisions.
This interpretation becomes clear when we use the identity (see proof in the
Appendix, Section A.6):
S
ft = £ ^ M ) II //'/A Vi = 1,..., i, (4.31)
s,s'ey j=l
to rewrite (4.29) under the strictly equivalent form:
J2 J,A{s-J) JJ fj'Jf/' =ft, Vi = 1,..., b. (4.32)

s,s'ey j=l
The meaning of (4.30) is that the observable (log f A ; J, i = 1,..., b) is a colli-

sional invariant, that is, an observable whose value measured at any node is left
unchanged by the collision phase (see Section 2.3.3). The important consequence
is that the conservation laws in a lattice gas model have a crucial incidence on
its ability to exhibit global statistical equilibria, and thus macroscopic collective
motions.
We must now consider the set of all coUisional invariants of the lattice gas (see
Section 2.3.3). According to the definition (2.19) of a coUisional invariant, it is
obvious that the set of all coUisional invariants of a lattice gas model is a vector
subspace of R b . Its dimension d lies a priori between 0 and b, but the limit cases
3=0 and d = b correspond to rather pathological (unphysical) models. The case
6=0 corresponds to lattice gas models with no other coUisional invariant than
the null observable (0,...,0), and so there is no factorized uniform statistical
equilibrium other than the trivial one: f}• = \, i = 1,..., b. The other limit, d = b,
corresponds to lattice gas models with no efficient collisions (the post-collision
state is always identical to the pre-collision state), for which any factorized
homogeneous distribution is an equilibrium.
Let us now consider a basis (q^K\ K = 1,...,<5) of the vector subspace of
coUisional invariants. Any given coUisional invariant is a unique linear combi-
nation of the form J2K=i ^ # • According to (4.30), (log f yfV ], i = 1,..., b) is

a collisional invariant. So, there exist S real numbers /l M such that:
s
Wq\K\ Vi = ! , . . . , £ > , (4.33)
which yields:
/*=/r =
(4.34)
\K=1 /
1
Here v>FD(x) = (1+e*)" is the Fermi-Dirac distribution. The fact that we obtain
a Fermi-Dirac distribution instead of the Maxwell-Boltzmann distribution of
continuous gases, is a consequence of the exclusion principle. Equation (4.34)
describes a family of homogeneous and fully factorized equilibrium solutions
of the lattice Liouville equation (4.6), depending on d arbitrary parameters A M,
where S is the number of linearly independent collisional invariants of the
model.
The parameterization by the l M s comes out naturally in passing from (4.30)
to (4.33). In Section 4.4.3, these arbitrary coefficients will be interpreted as
Lagrange multipliers in the global entropy maximization procedure.
Two important features must be emphasized:
We have proved, without recourse to the Boltzmann approximation, that

lattice gas models satisfying the semi-detailed balance condition possess
homogeneous, factorized, equilibrium macrostates with Fermi-Dirac
distributions. However, the proof does not preclude the possibility that
other kinds of equilibrium macrostates exist, neither does it give
information about their stability.
*; Provided the collision rules satisfy the semi-detailed balance condition,
the homogeneous factorized equilibrium macrostates do not depend on the
details of the collision rules, but only on the conserved quantities
(collisional invariants). Consequently, lattice gas models with different
collision rules sharing the same conservation properties and verifying the
semi-detailed balance have the same global equilibria.
4.4.2 The lattice Boltzmann approach

A lattice gas in a factorized uniform equilibrium state can be seen as a collec-
tion of b x Jf independent Bernoulli random variables. Since the probability
distribution of a Bernoulli variable is fully determined by a single number (its

averaged value), we can address the problem of equilibrium situations in terms
of the ensemble-averaged populations /,- as well as in terms of the probability
distribution SP. In other words, it is equivalent to searching for equilibrium prob-
ability distributions satisfying the stationary lattice Liouville equation, and to
searching for equilibrium averaged populations satisfying the stationary lattice
Boltzmann equation.
Remember that the lattice Liouville equation is an exact equation whereas the
lattice Boltzmann equation is not. However, when we consider only factorized
statistics, the lattice Boltzmann equation becomes exact.
From (4.8) it is clear that the time- and space-independent averaged popula-
tions fi9 i = 1,...,b must satisfy:
}*TfJ, Vi = 1.....6, (4.35)

s,s'ey j=\
or equivalently, with identity (4.31):
()f[f f = 0, V<=1 ft. (4.36)

s,s'ey ;=1
Since (4.36) is identical to (4.29), it follows that the lattice Boltzmann approach
is equivalent to the lattice Liouville approach if semi-detailed balance is satisfied;
consequently, the equilibrium solutions of the lattice Boltzmann equation are
Fermi-Dirac distributions.
4.4.3 The variational approach

As shown in Section 4.3.2, the global entropy H of the lattice gas cannot de-
crease under the action of the evolution rule, which result is valid only under the
Boltzmann approximation, and if the semi-detailed balance condition is satisfied.
We search for equilibrium distributions within the class of factorized homoge-
neous macrostates for which the expression for the global entropy reduces to
(see Section 4.3.2):
1=1
where /,• does not depend on the position variable r* because of the homogeneity
of the equilibrium state.
In order that the macrostate be an equilibrium solution, the global entropy
must be maximum, under constraints imposed by conservation. Indeed, consider
a basis (q^K\ k = 1,..., d) of the vector space of collisional invariants. The quan-
tities (q^) = ^2q\ fi are constant, because the q^ s are collisional invariants.
So we seek b average populations // which maximize H under the constraints

nave
that the averaged collisional invariants Yl^fi g i y e n values. We then
M
introduce d Lagrange multipliers A for these S constraints, and search for
average populations /,- that maximize the quantity:
i=l K=1 1=1
It is a simple matter to show that the solution is the Fermi-Dirac distribution

(4.34).
In conclusion, we have shown that the lattice Liouville approach, the lattice
Boltzmann approach and the variational approach yield strictly equivalent
results when semi-detailed balance is satisfied.
The expression (4.34) for the family of global factorized equilibrium distri-
butions gives the equilibrium mean populations /,- in terms of the Lagrange
multipliers A\K\ This is natural from a theoretical point of view, but practically
it would be more convenient to have these expressions in terms of physically
relevant and accessible quantities such as the macroscopic variables of the lattice
gas, that is, in terms of the macroscopic densities of the collisional invariants
(see Section 2.1.3). For example, consider a lattice gas model for standard fluid
dynamics, for which the collisional invariants are the mass, the momentum and
the energy. The physically relevant variables are the mass density, momentum
density and energy density, and it is useful (for example to set up the initial lat-
tice state in a simulation) to have the mean population of each channel in terms
of these mass, momentum and energy densities. The next section is dedicated to
this question.6
4.5 Natural parameterization of equilibria
We consider a lattice gas prepared in a homogeneous, fully factorized equilibrium

macrostate & such that:
(4.38)
where the equilibrium mean populations ffq) are given by the Fermi-Dirac
distribution (4.34).
We define the macroscopic variables p M as the macroscopic densities per node
of the collisional invariants q^K\ that is, the microscopic densities per node p* M ,
6
Section 4.5 contains material which is important for lattice gas practice, but it is rather
technical and may be omitted at first reading.
ensemble-averaged over the probability distribution ^ (eq) (see Section 2.1.3 for
the notion of density of an observable):
This definition of the macroscopic variables leads to the following set of S non-
algebraic relations between the macroscopic variables p M and the Lagrange
multipliers A M :
Vie = 1,...,*. (4.39)
If we are able to solve this set of equations, i.e. if we can write the A Ms as
functions of the p M s , then we can express the equilibrium distribution (4.34) in
terms of the macroscopic variables p M .
Unfortunately, the general solution to (4.39) cannot be cast in an explicit
form, and we have no exact expression for the equilibrium distribution in
terms of macroscopic variables. However, approximations can be obtained by
perturbation methods. Indeed, suppose that we have an exact solution ( J ^ K =
1,...,(5) for (4.39), corresponding to some particular values ( p ^ , K = 1,...,<5)
of the macroscopic variables. Then it is usually possible to find an approximate
solution (AM, K = 1,...,(5), corresponding to macroscopic variables (p M , K =
1,...,8) close to (pM, K = 1,...,<?).
We now sketch the asymptotic expansion leading to this approximate solution.
Let rj (<C 1) be an expansion parameter such that the perturbed macroscopic
variables read:
p M=pM+^M, K=1,...,S. (4.40)
(The subscript between parentheses refers to the order in the expansion scheme,
whereas the superscript between square brackets refers to the index of the
associated collisional invariant.)
We look for solutions of (4.39) with the following form:
l (4.41)
The expansion will be described in detail up to second order in rj, which is

sufficient for hydrodynamics. 7
For simplicity, we introduce compact notations for the Taylor expansion of
the Fermi-Dirac distribution xpFD(x), namely fou fu and fit for the values of
7
The method can be applied to any order, but with rapidly increasing algebraic
complexity.
the Fermi-Dirac distribution and of its first and second derivatives respectively,
computed for the unperturbed solution ( J ^ K = 1,...,<5):
\K=1 /
Using the explicit form 1/(1 + expx) for \pFD, it is easy to show that:
fit = -/o;(l - hi) and /2l- = /o,-(l - /oO(l - 2/ 0l )- (4.43)

The Taylor expansion of the equilibrium populations is straightforward:
ft = xpFD I V j A ^ g } I
\K=1 /
= hi
3
^W1 (4-44)
(
K=l
^ \K=1 )
i=l,...,b.
It then follows that (4.39), to first order in rj, leads to the following linear set
forOlM, K = 1 , . . . , < 5 ) :
3
£Mfc*UM = pM, (4.45)
K' = l
where M is a real S by (5 square matrix with elements:
<,MflJ«V u , K=1,...,S, K' = 1,...,5. (4.46)

1=1
This matrix is invertible (see proof in the Appendix, Section A.7); so the first
order linear set (4.45) has a unique solution (/ifj^, K = 1,..., 6), which is explicitly
given in terms of the macroscopic variables.
The second order of the expansion in powers of rj leads to a similar linear set
for the X^s, given in terms of the Afj^s:
EM M W = -\E^f»(E
This set also has a unique solution. The structure is similar for higher orders,
but the r.h.s. grows more and more cumbersome.
The next step is to re-introduce the expressions of the perturbations X^ and
2^ into (4.44) to obtain the desired approximate expression to the average
populations /,-.
This expansion is quite general, but rather formal. We now turn to examples so
that these notions become more practical. We first apply the general expansion
scheme to a class of simple lattice gases which contains most models with fluid-
like non-thermal macroscopic properties (Section 4.5.1). Then we deal with the
more complex case of thermal models (Section 4.5.2).
4.5.1 Low-speed equilibria for single-species non-thermal

models
We illustrate the formalism for a specific but wide class of lattice gases that dis-
play fluid-like purely kinetic (non-thermal) macroscopic behavior. These models,
which are of practical interest for fluid dynamics simulations, are defined by the
following set of properties:
(i) The model is G-invariant (see Section 2.3.4).

(ii) The model has crystallographic isotropy, at least up to second order (see
Section 2.3.5).
(iii) The model verifies the irreducibility condition (see Section 2.3.6).
(iv) The model verifies the semi-detailed balance condition (see Section 2.3.1).8
(v) All particles have unit mass. 9
(vi) The model has S = D + 1 linearly independent collisional invariants (see
Section 2.3.3), namely the mass and the D components of the momentum:
(4.48)
where the Greek index a refers to Cartesian coordinates and ranges from 1
to D.10
Property (v) - all particles have same mass - leads to considerably lighter
algebra. The formalism, however, extends very simply to the case where only
8
Property (iv) excludes models FCHC-6, -7, and -8 (see Table 3.3).
9
Property (v) excludes models FCHC-5, -7, and -8 (see Table 3.3), because they have rest
particles with mass 1 and 2.
10
Property (vi) excludes the thermal GBL model which has an additional independent
collisional invariant (the kinetic energy, see Section 3.6.2), and will be treated in
Section 4.5.2.
moving particles have the same mass, while the mass of rest particles can take
different values (for instance, for models FCHC-5, -7, and -8).
We now use property (ii) (crystallographic isotropy up to order 2) to define
a characteristic number, £2, which will be useful in the subsequent calculations.
Property (ii) implies that the crystallographic tensor J2ici« cip *s isotropic (see
Section 2.3.5 for a definition of the notion of isotropic tensors); so there exists
a number £2 such that:
Ctodp = 6<W- (4-49)

1=1
The characteristic number £2 depends on the geometrical structure of the lattice

gas automaton model, and not on the collision rules. Table 4.1 in Section 4.5.2
gives the numerical values of £2 for several models.
The macroscopic variables associated with the collisional invariants are /?, the
mass density per node and j , the momentum density per node:
i= l
h (4.50)
[a]
p = jSk/,- (a = 1,...,/)).
We choose to study nearly equally distributed equilibria, that is, global equi-
libria for which the average population of all channels is close to a common
value. The basic unperturbed equilibrium is a macrostate in which all channels
have equal probability to be occupied: /<>,• = d, Vi = l,...,fe. The parameter
d (0 < d < 1) can be interpreted as the 'particle density per channel'.
These equilibria are also called 'low-speed equilibria', because the mean ve-
locity of particles at any node is small, as soon as the populations of all the
channels on that node are almost equal. For instance, if there are approximately
as many upward moving particles as downward moving particles, the mean
velocity is obviously small.
The values of the macroscopic variables corresponding to the unperturbed
(fully equally distributed) equilibrium are: 11
(4.51)
11
Remember that the subscript between parentheses refers to the order in the expansion
scheme, whereas the superscript between square brackets refers to the index of the
associated collisional invariant.
For these values of the macroscopic variables, it is easy to verify that the
following Lagrange multipliers satisfy (4.39):
AM=0, a =1,...,£>.
We look for nearly equally distributed equilibrium states for which the mass
density is still p, but the momentum density is now j = 0 + rj]{1)9 where rj is an
expansion (small) parameter. The perturbations (p^\ K = 1,...,<5) are:
P,™ = 0,
(4.53)
From the unperturbed values (4.52) of the Lagrange multipliers, we can compute
the values of the expansion coefficients /o,- (using (4.42) and (4.48)), and fu and
f2i (using (4.43)):
hi = d,
fu = -d(l-d), (4.54)
As a consequence of the fact that all particles have equal mass, these quantities
are independent of the channel index i.
It is now straightforward to compute the 5 x 5 real matrix M defined in (4.46).
We find:
/b 0 . . 0\
0 i2 . . 0
(4.55)
\0 0
Since M^'^ is diagonal, the solution of the first order problem (4.45) is
simply:
AJJ = 0,
i / (4-56)
1 a
[] ^
To compute the second order perturbations Aj2^, we need to evaluate the

r.h.s. of the second order equation (4.47). This operation does not involve any
conceptual difficulty, but requires some care with the algebraic manipulations.
Section A. 8 in the Appendix sketches this operation for the more complicated
case of thermal models. For non-thermal models, the algebra is similar, but
simpler.
The resulting values for the second order corrections 1 ^ to the Lagrange
multipliers are:
l 2 d
; [0] = t
« W(\-Wb& (457)
The final expression for the perturbed mean populations /,- can be obtained
from (4.44), using the expressions (4.56) and (4.57) for the first and second order
perturbations 1^ and X^ to the Lagrange multipliers. The resulting expression
reads, in terms of the macroscopic variables p = bd and j : 1 2
U
~ b
Ql
(4.58)
We can rewrite (4.58) in terms of the 'fluid velocity' vector n = j / p , to obtain:
(4.59)
where the density-dependent factor G(p) is given by: 13

(b - 2p) b
(4.60)
L\p — p) gj
Notice the similarity between (4.59) and the low-speed expansion of the
Maxwell-Boltzmann distribution in classical statistical mechanics:
(4.61)
12
Remember that Einstein's convention is used for implicit summation over repeated
Greek indices.
13
This factor and the 'non-Galilean factor' g(p) in Chapter 5 differ by a multiplicative
constant.
which gives:
/(c) ~ const. 1 +
1 / m
cacp - ^&ap) + #(w3)] • (4.62)
2 VfeBT
4.5.2 Nearly equally distributed equilibria for thermal

models
We now apply the formalism to a class of single-species lattice gases with
thermal effects; this class is defined by the following properties:
(i) The model is G-invariant (see Section 2.3.4).

(ii) The model has crystallographic isotropy at least up to fourth order (see
Section 2.3.5).
(iii) The model verifies the irreducibility condition (see Section 2.3.6).
(iv) The model verifies the semi-detailed balance condition (see Section 2.3.1).
(v) All particles have unit mass,
(vi) The model has 3 = D + 2 linearly independent collisional invariants (see
Section 2.3.3), namely the mass, the D components of the momentum, and
the kinetic energy: 14
'H\
. . (4.63)
This set of properties is almost the same as in Section 4.5.1, except that fourth
order crystallographic isotropy is now required, and that energy conservation
is taken into account. The GBL model for thermo-hydrodynamics is a typical
example of this class of LGAs.
We use property (ii) (crystallographic isotropy up to order 4) to define two
additional characteristic numbers, £4 and ^ (£2 was defined in Section 4.5.1).
an
Indeed, property (ii) implies that the crystallographic tensors ^c^c^ d
c
J2i iotCif]CiyCid are isotropic (see Section 2.3.5 for a definition of isotropic tensors).
Moreover, two velocity vectors corresponding to two channels belonging to the
same class (see Section 2.3.4) are connected by an isometry of the crystallographic
point group G of the underlying lattice. So all velocity vectors in a given class
have the same modulus, and the quantity ^- depends only on the class index
/ of Cj. As a consequence, the fourth order tensor Yli(cV^)ci^c^ciyc^ ls a
crystallographic tensor (as defined in Section 2.3.5), and is therefore isotropic

14
The Greek index a refers to Cartesian coordinates and ranges from 1 to D.
according to property (ii). Since the three crystallographic tensors defined above
are isotropic, there exist three numbers, £2, £4 and ^ , such that:
(464)
!«-•'>
b 2
+ Say^ +
The three characteristic numbers £2, £4 and ^ depend on the geometrical

structure of the lattice gas automaton model, and not on the collision rules.
The apparently strange choice for the definition of £4 and & will become clear
below. Table 4.1 gives the values of £2, £4 and ^ for various LGA models.
It is technically more convenient to choose an orthogonal basis for the space
of collisional invariants, that is, a set of 5 collisional invariants g M such that
Table 4.1 The geometrical characteristic

coefficients {2, £4 and £6 for various lattice gas
models. (See Equation (4.64) for definitions,
and Chapter 3 for a description of the
models).
Model name D b 6
HPP 2 4 2 — —
FHP-1 2 6 3 0 0
3
FHP-2 and -3 2 7 3
165 795
GBL ^ i. 24
T9~ ~36l
FCHC-1 to -4 4 24 12 0 0
^• =1 q\ q\K = 0 when K =£ K!'. The simplest and most physical choice is:
\Cba/
(4.65)
= ic2
The collisional invariants q^ and g M are the mass and the momentum of
the particles, respectively; however, g [D+1] is not the kinetic energy, but a
combination of the kinetic energy and of the mass, and is also a collisional
invariant which we call the 'effective energy'. It can be verified that the definition
of the microscopic effective energies e\ guarantees the desired orthogonality
property.
The macroscopic variables associated with these collisional invariants are
respectively the 'mass density per node' p, the 'momentum density per node'
j , and the 'effective energy density per node' w, (a linear combination of the
kinetic energy and mass densities, which we shall call 'energy' for short):
b
ja = p W = ]T Ciafu (a = 1,..., D\ (4.66)
i=i
i=\ i=\ /=1
We study nearly equally distributed equilibria, and we proceed along the same
lines as in the previous section. Here the values of the macroscopic variables for
the unperturbed equilibrium are:
= 0, a = l,...,D, (4.67)
pV+1] =
MO)
and the Lagrange multipliers satisfying (4.39) are:
= 0, a = l,...,D, (4-68)
We now look for nearly equally distributed equilibrium states with mass density
p, momentum density j = 0 + rj\{l), and energy density w = 0 + rjw{l), where rj is
an expansion (small) parameter. The perturbations (pf^, K = 1,...,<5) are now:
= j m a , a=l,...,D, (4.69)
The values of the expansion coefficients /o,-, / n and fn are the same as given
in (4.54), and the d x S real matrix M defined in (4.46) reads:
(b 0 . . 0 0
0 . 0 0
&
:
= - d ( l - d) (4.70)
0 0 . . £2 0
Vo 0 . . 0
The choice of an orthogonal basis for the space of coUisional invariants has the
beneficial consequence to render M diagonal. In addition, all diagonal elements
are non-zero, since the coUisional invariants g M are linearly independent. The
particular choice for the definition of £4 in (4.64) is now seen to be justified as
we obtain simple expressions for the coefficients of M. The solution to the first
order problem (4.45) is simply:
« = 0.
d(l-d) U'
The computation of the second order perturbations /lf*] requires the evaluation
of the r.h.s. of the second order equation (4.47); the algebra is given in the
Appendix, Section A.8. The results read:
l 2d
AM = ~ (-L f + - U 2 ) ,
M = 2
rf)2(^W(1)J(1)^' a=1
'--" D ' <472)
!d / 2 :2 ,
With (4.71) and (4.72), the final expression for the perturbed mean populations
fi is obtained in terms of the macroscopic variables p = bd, j and w:
P
f -
1 1
+ -rJaCia + T-wet
S2 54
(4.73)
+ (9(j3) + fl?(/w) + (9(jw2) + d?(w3).
or alternatively, in terms of the velocity u = j/p:
1
« +- w e
PG(P) "a «/r(c

- (4.74)
2^2 W 2^4 \
P
where the factor G(p) is defined in Equation (4.60).

This is the type of expression for the mean populations given in terms of
the macroscopic variables (mass, momentum and energy densities) that we
shall use to compute the fluxes of the same conserved quantities, in order to
obtain the macrodynamic equations of the lattice gas. This result is also of
practical interest for fluid dynamics simulations. Indeed (as in computational
fluid dynamics (CFD) methods) the system must be initialized with values of
the numerical variables which represent the desired physical initial situation, e.g.
describing a field of given mass density, velocity and temperature. To initialize
the Boolean field in a lattice gas simulation, we must decide the random
occupation level (0 or 1) of each channel at each node, with the constraint that
the resulting macroscopic fields match the desired initial physical situation. An
efficient way is to use the approximate equilibrium mean populations in (4.74)
as the probabilities for the random distribution of 0s and Is; this procedure
offers the advantage that the initial state of the Boolean field is close to a (local)
equilibrium state, so that spurious transient regimes are avoided (or at least
considerably shortened).
4.6 Statistical thermodynamics
Lattice gases with single (unitary) velocity modulus (such as the FHP and FCHC
models) are well suited for the investigation of a large class of hydrodynamic
phenomena. They are by nature non-thermal models: mass and momentum
are the conserved physical quantities, and energy conservation is simply a
consequence of mass conservation. So the notion of temperature is irrelevant
in lattice gases with single non-zero velocity modulus. For a thermal lattice
gas, obviously we must have a multi-speed model. While this requirement is
a necessary condition (as fulfilled e.g. by the GBL model), the definition of a
temperature in the lattice gas does not follow straightforwardly, and in fact it is
a rather subtle matter to establish a lattice gas thermodynamics.15
In classical statistical thermodynamics, temperature T is introduced via the
canonical distribution function where exp(—jgjf) (with J f the total energy)
contains the quantity j3 = (/CBT)" 1 (with /CB the Boltzmann constant). Can we
legitimately expect a similar relation in the thermal lattice gas? One of the
complications arises from the exclusion principle (as in the Fermi gas) which,
in the lattice gas, imposes an upper bound to the energy. Consider for instance
the equipartition:
p J d V p e / J dPe\ (4.75)
15
This problem was considered by Ernst (1991), Cercignani (1994), and Grosfils (1994).
whose analogue for the general class of models with a finite number of discrete
velocities (including the class of lattice gases) would read:
^ ^ / - ^ , (4.76)
where the sum is taken over all velocity channels. To yield a standard definition
of the temperature, the above expression should be satisfied for any value of /?,
which is clearly not the case when i is discrete and finite.
Let us consider a volume element in the lattice gas at thermodynamic equi-
librium; the size of the volume element - the region <£ of the lattice - is large
compared to vo, the elementary volume associated to a node. So we shall use the
grand canonical ensemble. We start from the probability density in phase space,
(4.38):
b
= n Y[ftq)Si{r*y;q)Si{t*\ (4.77)
where ffq) is the Fermi-Dirac distribution (4.34):
s \ r / $
(4.78)
and with the normalization: X^SGr^(eq)(^) = 1> which is easily verified with
(4.77) since the Sts take the value 0 or 1. We rewrite (4.77) as:
where
(4.80)
J l
' K=\
Since we consider the system at global equilibrium to be described with the

grand canonical ensemble, we set:
8
- J > M * j l c l =*-Pei9 (4.81)

K=l
where et is the energy of channel i, and a and j8 are the conjugate thermodynamic
variables to be determined.16 Then we can rewrite ^{eq)(S) as:
b
0
*"^ 5 '^, (4.82)
16
We assume that there are no spurious invariants, and the variable conjugate to the
momentum does not appear since the system is at global equilibrium.
where S is the grand canonical partition function. From the normalization of

), and with the definition: z = expa, we obtain :
b
We can commute the sum and the products provided we take exactly the S/S that
are compatible with the state S eT; since here Si = 0 or 1, H is the Fermi-Dirac
grand partition function given by:
b
r*ei? 1=1
It will be convenient to define:
Q = log HFD
r*,i
53z«-""), (4.85)
where we have taken into account that the Jf nodes in the volume element
V (i.e. in if) have identical energy levels et. From the knowledge of the grand
partition function, we can obtain the basic thermodynamic quantities of the
system. Let us first compute the average number of particles in V'P
\ doc Jv,p \dai/v,p
Since (N) = J^Yli (ni(r*))> the average density per channel is given by:
which is indeed the Fermi-Dirac distribution (per channel) for global

equilibrium.18
By analogy with classical statistical thermodynamics (see e.g. McQuarrie,
1976), we can now compute the thermodynamic variables. The average density
per node is obtained straightforwardly:
17
Note that it is not essential to indicate explicitly that the partial derivatives in the
thermodynamic expressions are to be taken at constant V since, in the lattice gas,
expansion or compression cannot be performed as in classical thermodynamics; here ££
is fixed and V is necessarily constant.
18
For simplicity, we shall omit the superscript in /| eq) .
where we have used the elementary volume vo, associated to a node.19 To

compute the average energy (E), the thermodynamic pressure p, and the entropy
S, it is convenient to use the following algebraic relations:
2-oL + flei
ft = e*-pei
We find:
{E) =r ^ m =-(§) =^$>/,, (4.89)

L
\ Op / V,<x \Op J V,OL *-?
and
p V = p~l log S = p-1 Q = -jr r 1 E log (1 - ft), (4.90)

i
which suggest that P can be interpreted as the reciprocal temperature.
Now starting from the expression for the global entropy (see Section 4.3.2):
S = -JT Y, Ui log /,• + (1 - ft) log (1 - /,-)] , (4.91)
and using (4.85), (4.86), and (4.89), it is easy to show that (4.91) can be written
as:
S= Q+P{E) -a(JV), (4.92)
which is consistent with the classical thermodynamic expression for the entropy
provided a = P\x\ the consistency criterion determines \i as the chemical potential
(and consequently z = Qxp(pfi) as the fugacity), and we have:
= 0+B-i\(?Q\
= Q+P(E) - Pfi(N). (4.93)
The basic relation in classical thermodynamics is the equation of state, p =

p(p,T), where T is defined through p = (/cpT)"1 (here /CB is an arbitrary
constant which can be set equal to one). However, it is not obvious how an
19
Often we can omit writing VQ explicitly as it can be taken equal to one; then Jf can be
interpreted as the volume of the lattice gas.
explicit relation can be obtained for p(p, T) in the lattice gas. Indeed, using
(4.87), let us rewrite Equation (4.90) as:
pv0 = r )_^log(l + ze-pe<),

i
where z (or alternatively a) should be expressed in terms of p and T, i.e.

z = z((N)/V9P); this amounts to inverting Equation (4.86), but no explicit
analytical solution can be obtained.
Nevertheless, there is an interesting limiting case that we can consider: the
case of low densities and high temperatures. Then the number of accessible states
is considerably larger than the number of particles and the mean occupation
of an energy channel e\ is very small, i.e. /,- —> 0. From Equation (4.87), we see
that this condition is realized if z —> 0, i.e.
= ze~Pet9 (4.95)
and the corresponding thermodynamic state is characterized by (see (4.88))

p —> 0 and positive T. Summation of (4.95) over i and using (4.88) allows us to
eliminate z by writing:
li
p
We have thus retrieved Boltzmann statistics in the low density limit. Notice
however that the status of the partition function Yli e~^6i *s different in lattice
gases where - despite the fact that there are no non-local interactions - particles
subject to the exclusion principle are not independent, and (4.85)) does not lead
to the classical result:
The reason is that in classical statistical mechanics one sums over indistinguish-
able particles (hence the factor 1/iV! in (4.97)), whereas in lattice gases one sums
over the lattice sites.
Another difference with classical statistical thermodynamics is the equation
of state of the ideal gas which we are now in a position to compute from the
above results. Using (4.90):
pV= P~XQ = y r j 8 ~ 1 ^ l o g ( l + e"-pet), (4.98)

i
and expanding the log on the r.h.s. as:

00 m
1 / Y \
(499)
E^(rT^ '
we obtain:
m= rlY,^YtfiT- (4-100)
m=l i
For the reason discussed above, the pressure of the 'ideal lattice gas' is larger
than the intuitively expected value which would read:
Indeed, from (4.100), we have:
P»o = r1 [Yji + lYjt+ ••• I > (4-102)

which we rewrite formally as:
1
P =f p ( l +B2p+ •••), (4.103)
with:
^2Tf- (4-104)
Equation (4.103) is analogous to the virial expansion of classical statistical
mechanics. However, the lattice gas is not a Newtonian system :20 the notion of
interaction potential is precluded in the classical mechanical sense, 21 and there
is no virial theorem; therefore (4.103) is not stricto sensu a virial expansion.
Nevertheless, we shall see in the next section that the coefficient B^ can be
expressed in terms of a correlation function and takes a form equivalent to the
expression of the classical second virial coefficient. In fact, at this point, we
can notice an interesting analogy. For lattice gases with single non-zero velocity
modulus, /j- e q ) = d, and (4.104) can also be written as pBi = \ d ; for the hard
disk gas, one has pBi = \ Nna2/V, where a is the hard disk diameter (see e.g.
McQuarrie, 1976). So in both cases, the result is one half the relative average
'volume' where there is one particle of the gas.
The kinetic pressure PK, defined as the density of momentum transfer:
(4105)
i
§^'
i
with D the space dimension, is different from the thermodynamic pressure p,
(4.94). To illustrate the point, let us define:
i
20
T h e lattice gas dynamics is n o t governed by Newton's equations of motion.
21
Models have been developed where non-local interactions are incorporated in the
d y n a m i c s ; one can then define a quantity which is identified as the potential of mean
force of the lattice gas (Tribel a n d Boon, 1995).
where the summation is taken over all velocity channels; for non-thermal models
(with |c| = 1), (4.106) reduces to c2s = 1/D, and (4.105) becomes:
pKv0 = bc\d = - d , or pK = pc2s. (4.107)
Now setting jS"1 = c2s in (4.90), we find for the thermodynamic pressure:
pv0 = -bc2s\og(l-d). (4.108)
In the low density limit (p —> 0), the thermodynamic pressure (p ~ /J" 1 p, see
(4.103)) is equal to the kinetic pressure (see (4.107)), and
(4109)
»-?5-S
is the compressibility coefficient.22 In this limit, and with P~{ = c2, one retrieves
the law of ideal gases p = p/J" 1 .
The difference between thermodynamic and kinetic pressures in thermal lattice
gases is illustrated for the GBL model (see Section 3.6) in Figure 4.2; the kinetic
pressure for the same model is shown as a function of 6 = exp(—jS/2) in
Figure 4.3.
Some authors (Chen, Zhao and Doolen, 1989, and Molvig et a/., 1989) have
used the expression PK = P ^ B ^ K , which defines a kinetic temperature TR. AS
suggested by Figure 4.3 the kinetic temperature is a complicated function of
p and of [I, and must be distinguished from the thermodynamic temperature
T = (/CBJS)"1. To discuss this point we consider the low density limit where
the mean occupation of the energy channels et is small and is given by (4.95);
the condition z -> 0 is realized for the thermodynamic state p -> 0 at P or
Figure 4.2 Kinetic

pressure (full lines) and
thermodynamic pressure
(dashed lines) as a
function of density (GBL
model); isotherms for
0 = exp(-j8/2) = 0.05
(upper curves), 0.8 (lower
curves).
15.0
22
The notation cs used here will be justified in Chapter 9 where we shall see that cs is the
speed of sound propagation, which is consistent with (4.109).
T fixed. Now z —> 0 also corresponds to the thermodynamic state where p is

fixed and T -> oo (/? —• 0); indeed, considering the expression z = p Ylie~êi
(obtained by summing (4.95) over i), we see that for fi —• 0 or T —> oo, i.e.
g-/?e,- = ^ ^ g condition is realized provided the sum in the denominator is
large, that is, when there is a large number of energy levels. From Figure 4.3,
we observe that p K / p = ^B^k goes to a constant value as 6 -> 1, that is,
when T —> oo, the kinetic temperature goes to a constant value: /CBTK —* Cg,
whereas when 6 -> 0, i.e. T - • 0, T K is a function of the density, TK(p).
That P K / P = ^B^K —• ^s when 0 -» 1 can be explained in physical terms: in
lattice gases, energy is transferred through collisions between particles residing
on channels with different energy levels; but when the temperature becomes
infinite the channel occupation probability no longer depends on the energy
level, so that the energy transfer is then governed by momentum transfer which
produces sound propagation. We can now appreciate why the definition of a
thermodynamic temperature in lattice gases (and more generally in discrete
systems) in accordance with classical statistical thermodynamics is a delicate
matter.
4.7 Static correlation functions
In statistical mechanics the static correlations of density fluctuations are de-

scribed by the function:
(8p(u)8p(u')) = £(,5n(r*,c)<5n(r/,c')> = ^(Sû) dn}(u')) ,(4.110)

c,c' ij
Figure 4.3 The kinetic

pressure as a function of
9 = exp(-£/2) for the
GBL model; lower to
upper curves are for
p = 1,7,13, and 18.
0.1
0.0 0.2
where (5rcz(r*) = fy(r*) — /*, or equivalently by the static structure factor S(k),
the space-Fourier transform of (4.110):
pS(k) =
r*,r*'
r*,r*' ij
Higher order correlation functions may also be of interest. Now, at equilibrium,

lattice gases exhibit no correlations between different states {r*,c}. So we are
interested in the quantities ((<5n(c))") = ((Sni)n). Therefore we consider the
single channel probability distribution, and use the cumulant expansion method,
because through the connection between the cumulants and the central moments,
we can readily obtain explicit expressions for the density fluctuation correlations.
The generating function for the cumulants %f\ i.e. the log of the characteristic
function of the random variable su reads:
= log(exp{xn,}> = 2 ^ ^ - , (4.112)
and the nth cumulant is given by:
with Qi = log(l + ea~Pei) = log(l + ea~Pe{c)). The relations between the first
cumulants and moments of interest here are (Abramowitz and Stegun, 1965):
*?] = (n,) = U
K? = (to-fif),
K?] = (to-fif),
*?] = (to - ft)4) - Hto - ft)2)2 • (4.H4)
Note that, for single speed models, (n,-) = /,- = d9 in which case the subscripts
of the KS are redundant. With (4.113) and (4.114), some straightforward algebra
yields the results: 23
({8m)3) = /,.(l-/0(l-2/,),
4
((3m) ) = /,(1 - ft)(l - 3/,- + 3ff). (4.115)
23
Higher order cumulants can be obtained from a general expression (see Appendix A. 11),
but in practice we shall not need explicit expressions beyond the fourth moment. Notice
also that (n,-) = ft and {{Snt)2) = ft(l — ft) are classical results for the mean and the
variance of the coin-tossing experiment: n,- = 0 or 1 (see e.g. Papoulis, 1965).
(See also (4.42) and (4.43).) It then follows from (4.111) and (4.114) that the
static structure factor is given by:
pS(k) = ^2e±r*e~lkr* Y2lcf)<W> (4.116)

r*,r*' i
or, with (4.115) :24
Ps(k) = j>< 2) = E ^ 1
- fi)- ( 4 - 117 )
i i
For LGAs with single velocity modulus (/,- = d), (4.117) reduces to:
S(k) = 1-d. (4.118)

We notice that the absence of structure in S'(k) (= const.) stems from the strictly
local nature of the interactions. 25
The 'viriaP coefficient obtained in the previous section (see (4.103) and (4.104))
can now be expressed in terms of the static structure factor; indeed we can
rewrite (4.117) as:
> (4.119)
c i i
and it follows from (4.104) that:
B2 = \p-l{\-S(k)), (4.120)
a result which is analogous to the classical expression obtained in continuous

theory.26
The static structure factor is also related to the thermodynamic quantities P
and x discussed in the previous section. From (4.98), we have:
f
where:
A log (!+««-/»")) =/-•• (4-122)
Combining this result with (4.113) to obtain:
24
Rigorously pS(k) = £ ] . icf\l — d(k)), which shows t h a t exactly a t |k| = 0 (i.e. at infinite
wavelength) 5 ( k ) = 0 (Forster, 1975). But for all practical purposes the factor (1 - <5(k))
can b e ignored.
25
L G A s with non-local interactions exhibit a k - d e p e n d e n t static structure factor reflecting
the effect of the interaction r a n g e (Tribel a n d Boon, 1995).
26
In classical statistical mechanics, Bi is expressed in terms of the radial distribution
function g(r), which is related to the static structure factor (see H u a n g , 1963, a n d Boon
and Yip, 1980): B2 = - \ /0°° dh(g(r) - 1) = \ p" 1 (1 - lim^ 0 S(k)).
we find:
?) (4124)
-
So the isothermal compressibility is given by:
and, using (4.117):

S(k) = pp-'x- (4-126)
This expression is an interesting result because it relates the mean squared density
fluctuations to the thermodynamic quantities of the system. In continuous
statistical mechanics, one has essentially the same relation (see e.g. Boon and
Yip, 1980), the major difference being that the l.h.s. of (4.126) must be taken
in the thermodynamic limit (k —> 0). In lattice gases, (4.126) is valid at all fcs,
which is a consequence of the lack of 'molecular structure' in LGAs.27
27
For lattice gases whose microscopic dynamics includes non-local interactions (Tribel and
Boon, 1995), the limit k - • 0 must also be taken in (4.126).
Chapter 5
Macrodynamics: Chapman-Enskog method
In Chapter 4, we gave a statistical mechanical analysis of lattice gases, and dis-

cussed the equilibrium properties; we described uniform uncorrelated statistical
equilibria, and we showed that they have a Fermi-Dirac probability distribution.
The existence of these uniform equilibrium solutions can be established without
recourse to the Boltzmann approximation; the only required conditions are the
semi-detailed balance and the existence of local invariants. These equilibrium
solutions - the analogue of global equilibria in usual statistical mechanics - are
uniform by construction: the macroscopic variables, i.e. the ensemble-averages
of the local invariants, are space- and time-independent.
Now, we address the problem of space- and time-varying macrostates in
the 'hydrodynamic limit', that is, macrostates for which macroscopic variables
vary over space and time scales much larger than the characteristic microscopic
scales (lattice spacing and time-step duration). This scale separation between
microscopic and macroscopic variables is a crucial physical ingredient in the
theory of lattice gas hydrodynamic equations, and it is at the core of the
forthcoming derivation, which rests upon a discrete version of the well-known
'Chapman-Enskog method' (see Chapman and Cowling, 1970).
Although the principles of the method described hereafter apply to a wider
class of models, we are led, at a certain point of the forthcoming algebraic
procedure, to particularize our study to the (still wide) class of 'single-species
thermal models' defined in Chapter 4, Section 4.5.2. The simpler case of non-
thermal models is also treated, but in a less detailed manner (see Section 5.7).1
1
We consider here single-species thermal or non-thermal lattice gases but the method can
be extended to systems such as, multi-component lattice gases (see e.g. Section 7.7).
109
110 5 Macrodynamics: Chapman-Enskog method
5.1 Local equilibria and the hydrodynamic limit
Consider a lattice gas with d collisional invariants q^K\ K = 1,...,S, and sat-
isfying the semi-detailed balance condition so as to be in accordance with
the requirements for the existence of factorized global equilibrium states (see
Chapter 4).
Its macrostate at time U = 0 is described by a probability distribution function
3P (see Section 4.1.1) with the following properties:
The macroscopic variables p M = ( g M ) , computed with the distribution

& are b given functions p^(r*) of the space variable r*.
The macroscopic fields p M (r*) vary slowly in space, which means that the
typical distance over which the macroscopic fields vary significantly is
much larger than the microscopic unit length, namely the mesh size.
The distribution & depends on the macroscopic fields p M (r*) exactly in

the same way as ^ (eq) depends on the macroscopic variables p^ in the
homogeneous case (see Section 4.5).
This kind of macrostate is clearly not an equilibrium state: it cannot be a

stationary solution of the Liouville equation since the macroscopic variables vary
with space. However, these macrostates are called local equilibria (in the same
sense as in continuous statistical mechanics) because the fluid can be considered
to be in equilibrium locally, that is, in a region which is small compared to
the smallest macroscopic scale (yet large with respect to the microscopic scale).
They are denoted by ^(loc). This terminology is justified by the fact that local
equilibria have the same analytical form as global (homogeneous) equilibria,
but with slowly varying parameters. So locally the fluid seems at equilibrium,
whereas globally it evolves slowly towards a true uniform (global) equilibrium
state.2
During this evolution, macroscopic phenomena can occur on different time
scales. When the lattice gas shows fluid-like behavior, there may be elastic effects,
such as sound propagation, with short (macroscopic) time scales, and viscous
effects such as damping that occur on longer time scales. Both phenomena are
of interest here. Since they have well-separated time scales, we can analyze them
with multi-scale asymptotic methods.
2
Of course, the situation is quite different in the presence of an external force (e.g.
pressure gradient, gravity) which maintains the system out of equilibrium.
5.2 The multi-scale expansion for macrodynamics
5.2.1 T h e scale separation parameter

The theory developed in this chapter is valid in the hydrodynamic limit, the limit
where the smallest macroscopic scale is large compared to the characteristic
microscopic length (here typically the distance between two nearest nodes). Be it
for lattice gases or for real fluids, the scale separation hypothesis is mandatory
for a hydrodynamical description of the collective motions of the microscopic
particles. So we introduce a scale separation parameter e, defined as the ratio of
the microscopic scale to the smallest macroscopic scale. The hydrodynamic limit
corresponds to e < 1. This parameter will serve as an expansion parameter in the
multi-scale expansion leading to the macrodynamic (hydrodynamic) description
of lattice gases. This expansion is inspired by the Chapman-Enskog expansion
of continuous gases.
5.2.2 Perturbed local equilibrium

In a weakly inhomogeneous situation, it is reasonable to consider that the
actual time-dependent solution ^ ( 5 , t) of the Liouville equation (4.6) deviates
very little from the local equilibrium ^ (loc) corresponding to the same space and
time dependent macroscopic fields p M(r*,£*). This deviation must vanish when
the macroscopic fields become homogeneous, that is, when the scale separation
parameter vanishes. So it is natural to introduce the following expansion for 0*
in powers of the scale separation e:
where ^ (0) is just ^(loc), and ^ (1) and ^ (2) are the first and second order deviations
from local equilibrium.
The average populations /,• can be expanded in the same way:
fi = fim + efim + e2fp + ..., (5.2)
where / / 0 ) , f^ and / / 2 ) are computed respectively with the local equilibrium

probability distribution ^ (0) = ^(loc), and its first and second order deviations
SeT
SeT
SeT
That the actual solution /,• and its lowest order approximation / / 0 ) must lead
to the same macroscopic fields implies that the first and second order corrections
must not contribute to the macroscopic variables. Thus, the corrections / / ^ and
/ / 2 ) must verify:
and (54)
In terms of the probability distributions, this gives the conditions:

b
Ser i=l
a n d
(5.5)
= 0, K = 1,...,5.
SeT i=l
5.2.3 Macroscopic space and time scales

Since, in the hydrodynamic limit, noticeable space variations of the macroscopic
variables take place typically over distances of order e" 1, it is natural to introduce
a macroscopic space variable ri defined by:
ri=er, (5.6)
where r is the continuous version of the discrete microscopic space variable r*.
Rescaling r into ri amounts to choosing e"1 as the new length unit, instead of
the lattice constant which is the natural length unit for microscopic phenomena.
When the macroscopic variables are inhomogeneous, with gradients of order
e"1, then, by physical intuition (based on our experience with real fluids), we
anticipate two generic classes of macroscopic phenomena, superimposed on the
underlying microscopic evolution: (i) non-linear and pressure effects such as
acoustic propagation, and (ii) linear diffusive effects like viscous damping.
For a real Newtonian fluid, with given sound speed and viscosity (see e.g.
Landau and Lifschitz, 1987), non-linear and pressure effects involve first order
space derivatives. So, for inhomogeneous macroscopic fields with length scales
of order e"1, the typical time scales over which non-linear and pressure effects
occur should be of order e~l. This is particularly clear for sound waves: a
macroscopic inhomogeneity with wavelength of order e~l will propagate with

a given speed such that, at some given point, noticeable variations of the
macroscopic variables (pressure, velocity) will occur over time scales of order
On the other hand, diffusive effects involve second order space derivatives.
Thus, the inhomogeneities with typical length scales of order e~{ undergo slow
damping over time scales of order e~2. This is the case for sound attenuation in
viscous fluids (viscous damping of sound waves).
From these physical arguments, we anticipate similar macroscopic dynamical
processes in lattice gases and we introduce two auxiliary time variables t\ and
t2:
ti=et and t2 = e2t. (5.7)
At this point t\ and t2 are not two different time variables, but simply two
expressions of the same time variable t with different scalings. To illustrate the
argument, imagine a lattice gas whose (microscopic) time step would be one
second. Then the second is the natural time unit for microscopic phenomena.
Suppose now that the lattice gas is in an inhomogeneous macrostate with a
space scale separation parameter e of 1/60. Then, t\ is just the time expressed
in minutes and t2 the time expressed in hours. For this lattice gas, the acoustic
phenomena would evolve on time scales of a few minutes, and thus are better
described in terms of a time variable expressed in minutes; diffusive effects
would only be noticeable after hours, and therefore should be analyzed with a
time variable expressed in hours.
Note that t\ and t2 are not independent variables: they are connected through
their relations with t. But it is usual in multi-scale methods (see Bender and
Orszag, 1978) to consider any function F(r, t) of space and time as a function
F{*ututi) of ri, t\ and t2, as if t\ and £2 were independent variables. 3 The time
and space derivatives of an arbitrary function F(r, t) can thus be expressed as
follows in terms of the derivatives of F(ri, t\, t2):
dF dF a n dJ
dF dF 2dF
-x-
dr
= e dfx
r- ^~
dt
= €~^-
dti
+ e dt
^--
2
All the tools necessary for the multi-scale expansion are now ready; it remains
to introduce the master equation to which the expansion will be applied.
3
Strictly (from the mathematical point of view) the two functions should be labeled
differently, but, as no ambiguity results, we keep the same notation.
5.2.4 The averaged microdynamic equation

A convenient way to derive the macrodynamic equations for the macroscopic
variables p ^ is to start from the microdynamic equation (2.13) rewritten as:
b (5.9)
Si-sdZis^Hn/in/L
(s,s')ey2 ;=1
Here, nj denotes the Boolean population after collision. Note that the Boolean
populations n\ and n, in the second equation are all taken at node r* and
time U. This formulation which separates the collision phase and the propagation
phase, is strictly equivalent to (2.13), but conceptually easier to handle for the
forthcoming averaging operations.
Ensemble-averaging the first equation in (5.9) with the probability distribution
8P of the corresponding macrostate gives:
/,(r*+c I -,f* + l) = /;(r*,t*), (5.10)
which merely expresses that the propagation (or free-streaming) phase just
moves information, without loss or gain, from one node to another node. The
second equation in (5.9), after ensemble-averaging over £P and over all possible
choices for non-deterministic collisions, yields:
ft(u9u) = /,(r*,t*) + Ai(P,u9t*)9 (5.11)

where the averaged collision term A,-(^,r*,r*) is:
A,-(^, r*, u ) = / J2 (si ~ Si)A(s-,sf) f[ n/m/J \ . (5.12)

\(s,s')ey2 7=1 /
Indeed, the random variables £(s^sf) involved in the collision phase (see Sec-
tion 2.2.3) are statistically independent from the Boolean field. So, the averaging
over all possible choices for non-deterministic collisions simply replaces the
random variables £(s->s') by their mean values A(s-*sf) (the collisional transition
probabilities). The brackets (• • •) denote ensemble-averaging over the probability
distribution ^ , that is, summing over all possible realizations S of the Boolean
field n, with each term weighted by its probability ^(S,U)\
b
Ai(<P, r*, u) = ^ 9{S) ] T (s[ - Si)A(s^sf) J J S,W Sj(up.
SeT (s,s')ey j=l
Since the product Y{SJ{T*) JSJ{T*) J is zero unless S(u) = s, the averaged collision
S S
term A,-(^,r*,£*) takes the simpler form:

A,.(^, r., U) = 5 3 9(S) 53(s; - Si{u))A(S(r^sf). (5.13)
SeT s'ey
Without any further hypothesis, the averaged collision term A,- depends a
priori on the full probability distribution ^ , and not just on the mean popula-
tions /j(r*,f*). Indeed, the average of the product II/=i n/ 7 '"/ 7 ' ^n (5-9) cannot
be factorized into the product of the averages, since the populations of different
channels at the same node may be correlated. Consequently, a closed form for
the averaged microdynamic equation cannot be obtained without the Boltzmann
approximation (see Chapter 4), which assumes that particles entering a collision
are uncorrelated; then the average of the product Il;=i njSjn/j c a n be approx-
imated by the product of the averages. With this ansatz, we obtain the lattice
Boltzmann equation as the averaged microdynamic equation (see Chapter 4).
Here we postpone the introduction of the Boltzmann approximation and keep
the full expression (5.13) for the averaged collision term.
Aj(^, r*,£*) is clearly linear in 3P\ it is thus a linear function of 2bmjr real
variables, the probabilities 0>(S, U) of each of the 2hmjr possible configurations
S of the full lattice gas at time U. We rewrite (5.13) as:
\ (5.14)
SeT
where L.s(r*) is a node-dependent b x 2b'-^ real matrix with coefficients:
" Si(r.))A(S(r^sf), Vi = 1,..., ft, VS G T. (5.15)

s'ey
To derive the macrodynamics of the lattice gas we thus start from the following
equation for the averaged microdynamics:
ft(u + ct ,t* + l ) - ft(u ,**) = £ L i S ( u ) &(S,u). (5.16)

SeT
Obviously the average collision term A, vanishes for any global equilibrium
)
. It is less obvious, but nonetheless true, that it also vanishes for any local
equilibrium ^(loc). Indeed, the collision process in a lattice gas, as defined in
Section 1.2.3, is purely local. Therefore, the value Aj(^,r*) of the averaged
collision term at node r* and time U depends only on the single node probability
distribution for node r* at time £*, and not on the full information contained
in 3P. On the other hand, the definition of a local equilibrium (see Section 5.1)
implies that for each node r* in j£f, there exists a global equilibrium ^ (eq) such
that the single node distribution extracted from ^ (eq) is identical to the single
node distribution extracted from ^(loc) at r*. Since all global equilibria make A,
vanish, it is clear that A,- also vanishes for local equilibria:
A;(^(loc), r*, U) = 0, Vi = 1,..., b, Vr* e &, W*. (5.17)
It is important to note that the conservation laws imply some constraints for
A,-. Indeed (as in Section 4.4.1) consider a basis (q^K\ K = 1,...,<5) of the vector
subspace of collisional invariants. Using (5.13) for Aj(^,r*), and Equation (2.19)
as a definition of a collisional invariant, we can straightforwardly establish the

following relation:
*,t*) = 0, Vic=l,...,a, V^, (5.18)
which must be verified by the averaged collision term A,-(^,r*,£*). In terms of

the matrix L.s introduced in (5.14) and (5.16), (5.18) implies:
= 09 VK=1,...,<S, VSeT, Vr* e &, (5.19)
which means that the rank of the b x 2bmjr matrix LiS is smaller than fo, and is
at best equal to b — S, 6 being the number of independent collisional invariants.
In other words, the subspace of collisional invariants is included in the kernel
of the adjoint of LiS. This property will play an important role in the derivation
of the macrodynamics of the lattice gas.
5.2.5 The expansion in powers of e

NOTE: We recall that Latin indices are used for channel labels, and that Greek
indices are used to label space coordinates. In order to simplify the algebraic ma-
nipulations, we follow Einstein's convention for implicit summation over repeated
Greek indices. For better readability, we use the compact notation dl(Xfor the par-
tial derivative with respect to the coordinate a of rescaled space variable ri as
defined in (5.6), and dh and dt2 for the partial derivatives with respect to the
rescaled time variables t\ and ti, respectively, as defined in (5.7).
When the scale separation parameter e is small (i.e. the averaged populations
vary little from one node to the next), we can validly approximate the averaged
population /j(r* + c^<, U + 1) in Equation (5.16), by its Taylor expansion around
(5.20)
+ (9(e3).
Here, /,- and its derivatives are taken implicitly at node r* and time U (unless
otherwise specified). To obtain this result, we have used Equation (5.8), which
connects the derivatives with respect to the lattice space and time variables
(r*, £*), and the derivatives with respect to the rescaled (macroscopic) space and
time variables (ri,ti,^).
We now introduce the fact that the average populations /, are close to their
local equilibrium values; so they can also be expanded in powers of e - see (5.2)
- with a zeroth order term corresponding to the local equilibrium. Combining
the two expansions, the l.h.s. of (5.16) becomes:
(5.21)
The r.h.s. of (5.16) can also be formally expanded in powers of e, using

expression (5.14) for the averaged collision term. The resulting expansion is
(argument r* omitted in the r.h.s. for simplicity):
Ser
) (5-22)
SeT
33
+ (9(e )).
Using (5.21) and (5.22) as the respective expansions of the l.h.s. and r.h.s. of the
averaged microdynamic equation (5.16), we obtain, to the first order in e:
SeT
and to second order:
+ dtjt^ + y^yctfict,^ (5.24)
We shall now derive the first and second order macrodynamical equations of
the lattice gas starting from Equations (5.23) and (5.24) which follow from
the averaged microdynamic equation (5.16) expanded in powers of the scale
separation parameter e.
5.3 First order macrodynamics

Formally, the first order equation (5.23) is a set of b linear equations for 2 b v r
unknown quantities, the probability perturbations 0>{1)(S) for each lattice con-
figuration S. Of course, these b equations are not sufficient to fully characterize
^ (1) (S), which must also meet the requirements of the Liouville equation (see
Chapter 4) governing the evolution of the probability distribution ^ . However,
(5.23) is a necessary condition for ^{l)(S), and it has to admit at least one
solution. Because of the conservation laws, the rank of the linear system (5.23)
is lower than b9 the number of equations. So the r.h.s. of (5.23) must sat-
isfy solvability conditions (Fredholm's criterion) which lead to the first order
macrodynamics. We now see clearly the direct one-to-one relation between the
conservation laws and the macrodynamics: the conservation of a microscopic
observable q imposes a solvability condition on the r.h.s. of (5.23), and so yields
a macrodynamic equation which is a continuity equation for the associated
macroscopic variable.
By construction, the continuity equations so obtained are valid to first order
in the scale parameter e. They do not take into account diffusive and viscous
effects that will come out of the second order equation (5.24).
5.3.1 Solvability conditions for the first order problem

We consider the first order equation (5.23) for ^ (1) ; applying Fredholm's criterion
of solvability (the l.h.s. must be orthogonal to all vectors in the kernel of the
adjoint of the linear operator L. s), we find that f^ must verify:
8ti £ # ] / , ( 0 ) + Slf J2<^Ci»SP = 0, K = 1,...,8. (5.25)

1=1 i=\
The sums over i are clearly the macroscopic (ensemble-averaged) density p M

and flux O M of the observable q^K\4 and Equation (5.25) takes the familiar
form:
dtlP[K]+dlfi<bf]=O9 K = 1,...,<5. (5.26)
Under this form, the solvability condition (5.26) for the first order problem (5.23)
appears clearly as a set of macroscopic continuity equations connecting densities
p M and fluxes <DM of the microscopically conserved observables (collisional
invariants) g M .
We now need to evaluate these fluxes and to express them as functions of
the macroscopic densities p M , so that the continuity equations (5.26) become a
closed set of evolution equations for the macroscopic variables, i.e. the densities
4
See Section 2.1.3 for the definition of macroscopic densities and fluxes, and Section 4.1.2
for the notion of ensemble-averaging.
p M of the coUisional invariants. To do so, we need information about the nature

of these coUisional invariants; and we also need to know the explicit form of
the mean populations ft. Thus, we now particularize the analysis to the class
of thermal lattice gases (see Section 4.5.2) for which the coUisional invariants
are mass, momentum and energy. (The case of non-thermal models is treated
in a less detailed manner in Section 5.7. Indeed, the analysis is fundamentally
similar, albeit technically different.)
In Section 4.5.2, we derived expression (4.73) for the mean equilibrium popu-
lations in terms of the mass density p, the momentum density j, and the effective
energy density w. This expression, which is valid for nearly equally distributed
states (low velocity, low effective energy) reads:
f.(0) _ (L . _ i / c . . _* .
0 C2 C4
b(b-2p)
T2 J*J
2p{b-p) S2 (5.27)
where et is the microscopic effective energy: e}•, = ^ — ^ (see Section 4.5.2).

In (5.27), the expansion parameter rj is not a scale separation parameter as e. It
governs the proximity to the truly equally distributed macrostate with the same
mass density. It is also the order of magnitude of both j and w (see Section 4.5.2).
Let us also recall the expressions defining the geometric coefficients £2, £4 and
U in (5.27):
X \SapSyS + SayS + d d h
Ac?
^ w T )
X (<><xB(>yd
aj + Say SRS +
A
These coefficients depend only on the geometric structure of the lattice gas and
not on the collision rules.5
We now use (5.27) to obtain the averaged mass, momentum and effective
energy fluxes O [0] , O M and O [D+1] , involved in the first order macrodynamic
equations (5.25). The details of the calculations are given in the Appendix,
Section A.9.
(i) The mass flux is just the momentum density; its components are:
b
Qf = ^ <W,-(0) = Jp = f>up, P = h...,D, (5.29)
where u = j / p denotes the macroscopic fluid velocity (see Section 4.5.2).

(ii) The momentum flux is a second order tensor whose components can also
be expressed in terms of the macroscopic variables:
b
*M — \ ^ f (0)
i=\
(530)
w)Sp
where
and
P(p,j 2 ,w) appears as a diagonal part of the momentum flux tensor: it is

interpreted as the hydrostatic pressure of the lattice gas.
(iii) The effective energy flux follows similarly; its components are:
i=\
^ 2£4 . (5.33)
+ J
where
5
See Section 4.5.2 and Table 4.1 for the values of these coefficients for some simple
models.
The density-dependent coefficient g(p) that appears in both (5.30) and (5.32)
is called the 'non-Galilean factor', or, more precisely, the 'non-Galilean factor
for momentum'. 6 This non-Galilean factor is very important both for theory and
for applications of lattice gases. Indeed, Equation (5.30) is not Galilean invariant
unless g(p) is exactly equal to 1, which a priori is not the case. However, as
we shall see in Chapter 9, Galilean invariance can be recovered, at least for
non-thermal models, in incompressible regimes where the density p is nearly
constant: g(p) is then a constant factor that can be eliminated by proper time
rescaling. The factor g\p) in Equation (5.33) is the 'non-Galilean factor for
energy' because it plays, for the energy, the same role as the non-Galilean factor
g(p) in the momentum equation.
Somewhat in the same way as transport coefficients are expressed in terms of
time-correlation functions, these non-Galilean factors may also be given a mi-
croscopic content: they can be expressed in terms of static correlation functions
(for details see Brito, Bussemaker and Ernst, 1992). For LGAs satisfying detailed
or semi-detailed balance - so possessing a universal equilibrium distribution -
the non-Galilean factors are given in terms of the correlation functions K^
and 7c|3) discussed in Section 4.7; so, in some way, they can be interpreted as a
thermodynamic property of the lattice gas.
We now return to the first order continuity equations (5.26) which, with the
above results, we can write in the standard Eulerian form:
09 (5.35)
Kk + dlP (g{p)kup) + diaP(j>, j 2 , w) = (9(rj3l (5.36)
dtl w + dlP (g'(p)wup) + dlfi (^Jfi) = ®tf\ (5.37)
The mass continuity equation is valid to any order in j , whereas the momentum
and effective energy continuity equations are approximations neglecting terms
(9(rj3). The reason is that the mass flux is also the momentum density, but
the momentum and effective energy fluxes are functions of the macroscopic
variables, which can be obtained only through an asymptotic expansion in
powers of j and w.
5.3.2 Solution of the first order problem

The first order macroscopic equations (5.35) to (5.37) appear as solvability
conditions for the first order averaged microdynamic equation (5.23). But we
have not solved this first order equation, and its solution 0>{l)(S) should be
necessary to compute the first order mean populations /, (1) involved in the
6
g(p) and the factor G(p) of Section 4.5.1 differ by a multiplicative constant.
second order problem (5.24). Nevertheless, it is possible to proceed further (to

the second order) without actually solving this difficult problem, at least if one
is interested in the general form of the second order macrodynamic equations,
rather than in the explicit expression (and the numerical value) of the coefficients.
Indeed, suitable symmetry arguments will provide sufficient information on the
first order mean populations f^ and on how they depend on the macroscopic
variables. In this manner, linear response coefficients will be introduced, whose
values can be either 'measured' by numerical simulations (see Chapter 10), or
evaluated with the Boltzmann approximation (see Section 5.6).
Let us start from the first order equation (5.23), and rewrite the r.h.s. in terms
of the gradients of the macroscopic variables, using the expression (5.27) for the
equilibrium distribution //°), and the solvability conditions (5.35) to (5.37). The
resulting first order equation is:
ser
+ (9(erj2).
It is relatively easy to convince oneself that the solution ^ (1) (S) of the first order
problem must be a linear combination of the gradients diyj$ + d^jy and d^w.
The first order corrections f^ to the mean populations, that result from ^ (1) (S),
must thus also be a linear combination of these gradients. The most general
form for this combination reads:
fP = M ^ ^ l + M + NiAyWy (5.39)
where N,- is a set of i-indexed D -dimensional vectors, and where M,- is a
set of i-indexed D -dimensional second order tensors that may depend on the
macroscopic variables (p,j 2 , w). Furthermore, as the second order tensors M,- are
contracted with the symmetric expression diyjs + d^jy, they can be assumed,
without loss of generality, to be symmetrical with respect to exchanges of their
first and second Greek indices.
So far, we have made no hypothesis on the value of these vectors and tensors,
and they can be regarded as a set of unknown quantities depending on the
macroscopic variables. However, we can be more specific if the lattice gas is
G-invariant and verifies the irreducibility property (see Sections 2.3.4 and 2.3.6).
Then, we can apply the simplification procedure described in Section 2.3.6.3:
the vectors N,- must be proportional to the velocity vectors c,-, with equal
proportionality coefficients for all channels belonging to the same class :7
N,- = N(j)c,, Vi = l,...,6, (5.40)

7
For the notion of class, see Section 2.3.4.2.
where / is the index of the class to which channel i belongs. Along the same
lines of reasoning, the second order tensors M, are linear combinations of c,- ® c,-
and (5, the Kronecker tensor, with coefficients which depend only on the class
of the corresponding channel:
Mf = M{i)Ci 0 a + M'{i)5, Vi = 1,..., b. (5.41)
This simplification is possible because the vectors and tensors M,- and N,- are
such that for all isometry g in G and for all i = 1,..., b, we have g(M,-) = Mg(,-)
and g(N,-) = Ng(,-) (for the definitions of g, G and g, see Section 2.3.4).
The resulting form for f^ reads:
d
fP = (Minces+M\i)dy3) «»+ dM
iy diyw.
As emphasized in Section 5.2.2, the first order corrections f^ to the mean

populations must not contribute to the mass, momentum and effective energy
densities. This severely constrains the possible values of the arbitrary set of
parameters M(/), M\i) and N(i). Indeed, the fact that f^ does not contribute
to the mass variable (J2ift^ = 0) implies that:
b
Y^ (M{i)ciycid + M'(i)Sy5) = 0, V 7, S9 (5.43)
that ft^ does not contribute to the momentum variable imposes that:
b
i5=0, Vy,5, (5.44)
and that f^ does not contribute to the effective energy requires that:
b
^2(M(i)ciyciS + M\i)5yS)ei = 0, V 7,8. (5.45)
The // (1) s contribute neither to the effective energy density nor to the mass
density, thus, they do not contribute to the kinetic energy density either, since
the latter is a linear combination of effective energy and mass densities. So, the
following relation holds between M and Mr:
= 0, Vy,<S. (5.46)
1=1
This relation, taken for 7 = 6 and summed over 7 gives the following
simpler form:
i=\
which will lead to some simplification in the forthcoming algebra.

To write the solvability conditions for the second order problem, we need
an expression for the contributions J2t ci<*cipfi^ a n d ^ etcwfi^ °f the first
order mean populations /; (1) , to the momentum and effective energy fluxes. The
contribution to the mass flux Y,i cwfi^ *s obviously zero, since the mass flux
is nothing else but the momentum density, to which the first order populations
/i ( 1 ) cannot contribute. We show hereafter how to obtain the functional form
of these contributions, without knowing the actual values of the coefficients
involved.
The first order contribution J ] - c^Cipf^ to the momentum flux can be com-
puted from (5.42), which gives the following expression:
(5.48)
The lattice gas model is assumed to verify the fourth order crystallographic
isotropy property (see Section 2.3.5). Thus, there exist two coefficients v(c) and
v^c\ depending on the macroscopic variables (p,j 2 ,w), such that:
ipCiyCis = —v (c)
+8as8py),
and (5-49)
b
= - v ( c ) 5ap.
The superscript '(<?)' is used to indicate that these coefficients come from the
first order perturbation f^ to the mean populations, which are solutions of the
first order problem, and thus depend on the collision rules. They are not purely
geometrical coefficients like £2, £4 and £5.
Because of (5.47), the two coefficients v(c) and v^ cannot be independent.

Indeed, it follows from their definition (5.49) that:
b
1 2 2
cC
and (5.50)
1=1
So, Equation (5.47) implies that:
= 0. (5.51)
With this result, one can express the first order contribution to the momentum
flux tensor (5.48) in terms of the coefficient v(c^:
1=1
( . (5.52)
\ /
which yields immediately:
SiaJp + dlPh_^Ma\ (553)
Thanks to (5.51) which connects v(c) and v^, the first order contribution to
the momentum flux tensor takes the simple form (5.53), involving only the
symmetric traceless part of the momentum density gradient tensor d j .
In a similar manner, the first order contribution J2i eictpfi^ t 0 the effective
energy flux is given by:
(5-54)
Observing that e\ depends only on the class to which channel i belongs, and
invoking again the crystallographic isotropy property, we see that there exists a
coefficient ^c\ which depend on the macroscopic variables (p,j2,w), such that:
(5.55)
The resulting expression for the first order contribution to the effective energy
flux vector reads:
b
^2eiCiPf^ ~ ~^ îpw- (5.56)

;=i
We emphasize again that the actual values of v(c) and £(c) are still unknown. At
this point, all we can say - without the Boltzmann approximation - is that these
coefficients do exist, and that the first order contributions to the momentum
and effective energy fluxes are of the form (5.53) and (5.56). In Section 5.6, we
shall use the Boltzmann approximation to compute these coefficients explicitly.
Their numerical value can also be 'measured' by computer simulations of wave
damping (see Chapter 10).
5.4 Second order macrodynamics
The second order problem (5.24) has exactly the same mathematical structure
as the first order problem (5.23), except that the r.h.s. is more complicated.
The analysis will also involve solvability conditions, that is, the r.h.s. must be
orthogonal to all vectors in the kernel of the adjoint of the linear operator L. s
(Fredholm's criterion).
5.4.1 Solvability conditions for the second order problem

We apply Fredholm's criterion to the second order problem (5.24): the r.h.s.
must be orthogonal to the local invariants of the model q^K\ K = 1,...,<5. We
obtain that f^ and ff^ must verify the following relations:
i=\
wy (5.57)
1=1 i=l
for all K = 1,...,<5, that is, for all collisional invariants. Since we treat the
case of single-species thermal lattice gases, the linear invariants are the mass,
the momentum and the effective energy. The algebraic manipulations leading
from (5.57) to equations (5.58), (5.59) and (5.61) hereafter are given in the
Appendix, Section A. 10. In these manipulations, we use the first order macrody-
namical equations, (5.35) to (5.37), and the fact that the first order perturbations
5.5 The macrodynamic equations 127
/i ( 1 ) do not contribute to the mass, momentum and effective energy densities

(see Equation (5.4)).
For the mass invariant q^ = (1,...,1), the solvability condition (5.57) leads
to the simple relation:
dt2p = 0, (5.58)
(1)
which shows that there is no mass diffusion. The first order perturbations /; do
not appear in the second order solvability condition (5.58). Indeed, by definition,
the /i (1) s contribute neither to the mass density, nor to the mass flux, since the
latter is also the momentum density (see Equation (5.4)).
The case of the momentum invariant is more difficult, because the first order
perturbations /; (1) now contribute to the second order solvability condition
through the momentum flux J2iciocctpfi^ ( s e e Equation (5.53)). The resulting
solvability condition is:
-v (Vay> + dlPja - ^d.yjyS^ j = (9{n\ (5.59)
where the coefficient v, which a priori depends on (p,j 2 , w), amounts to:
This coefficient v is a transport coefficient that will be interpreted in Chapter 9

as a 'shear viscosity coefficient'.
The effective energy invariant leads to the same kind of expression:
) (5.61)
where the coefficient £ is:
The coefficient £ is also a transport coefficient that will be interpreted in

Chapter 9 as a 'thermal conductivity coefficient'.
In both (5.60) and (5.62), the first term is the contribution of f^l\ which
depends on the collision rules, and the second term is a purely geometrical
contribution, which depends only on the propagation rule and on the lattice
geometry.
5.5 The macrodynamic equations
We can now merge the results of the first order problem, (5.35), (5.36) and (5.37),
and of the second order problem, (5.58), (5.59) and (5.61), to obtain the macro-
dynamical equations governing the time-evolution of the macroscopic variables
p, j and w. To do this, we multiply (5.35), (5.36) and (5.37) by e, and (5.58),

(5.59) and (5.61) by e2. After summation, we use (5.8) to recover the real time-
and space-derivatives dt and da from their rescaled versions dtl, dt2 and dl<x. We
obtain, for the mass invariant:
dtp + dp(pup) = 0, (5.63)
for the momentum invariant:
dtja + dp(gjaup) + daP = dp(v(djp + dpj« - idyjyd«p))

V J V y
(5.64)
+(9(e3rj) + (9(e2rj2)
and for the energy invariant:
dp(&jp) = dpUdpw)
V J V 7
(5.65)
(9{e2ri2)
The non-Galilean factors g and g', defined in (5.31) and (5.34) respectively,
depend on the mass density p. The pressure P is defined in (5.32), and depends
on all the macroscopic parameters (p,j2,w). The transport coefficients v and £
are defined in (5.60) and (5.62). Their values also depend a priori on all the
macroscopic parameters (p,j2, w). However, any dependence of v or £ on j 2 or w
would lead to terms of order ofe2rj2 or higher in the macrodynamic equations. To
be consistent with the order of approximation in the macrodynamic equations,
these terms will be neglected; from now on, we shall use the short-hand notation
v and £ for v(p,0,0) and £(p,0,0).
Equations (5.63), (5.64) and (5.65) give the general form of the macrodynamic
equations governing the long-time, long-wavelength behavior of the lattice gas.8
These macrodynamic equations show close resemblance with the equations
of continuous fluid dynamics (such as the Navier-Stokes equation). However,
there is a difference: the non-Galilean factors g(p) and g'(p), in (5.64) and (5.65)
respectively, do not appear in the usual equations of continuous fluid mechanics.
The lack of Galilean invariance in the fluid-like behavior of the lattice gas is
a direct consequence of the discrete nature of the microscopic velocity space,
which itself results from the lattice structure.9 In Chapter 9, we shall show
how Galilean invariance can be recovered in hydrodynamic regimes such as
incompressible flow and linear acoustic flow.
8
In Chapter 8, we give an alternative derivation of the lattice gas hydrodynamic
equations using the lattice Boltzmann approximation.
9
A factor also appears in front of the advection term in the hydrodynamic equations of
discrete kinetic models (Gatignol, 1975) where position space is continuous and velocity
space is discrete.
5.6 Transport coefficients within the Boltzmann approximation 129
5.6 Transport coefficients within the Boltzmann approx-

imation
To compute theoretically the transport coefficients v and £, we need a physical

approximation: the Boltzmann ansatz by which we assume that particles entering
a collision are statistically uncorrelated (of course they are correlated after the
collision). Physically this means that the 'memory' that particles may have of
a past collision is erased by the propagation phase, before these particles meet
again for a new collision. Clearly this approximation is acceptable at sufficiently
low density.
Under this strong physical hypothesis, the averaged collision operator
Ai(&,r*,t*) defined in (5.12) is replaced by the 'Boltzmann collision opera-
tor' A^B\ where the average of the products of TZ;S is factorized into the product
of the averages:
b
A^Cf,) = £ (s't - Si)A(s^) H f/Jfj*. (5.66)
(s,s')ey2 ;=1
Combining the Boltzmann collision operator and the propagation operator, we

obtain the 'lattice Boltzmann equation': 10
b
s
fi(r* + a, U + 1) - /;(r*, U) = J2 ( i - Si)A(s^) n / / > / / ' . (5.67)
(s,s')ey2 7=1
As for the averaged microdynamic equation in Section 5.2.5, we can again

introduce the expansion of /,- in powers of the scale separation parameter
e. Following the same procedure, we obtain the first order lattice Boltzmann
equation, which connects the as yet unknown first order mean populations /, (1)
to the local equilibrium mean populations /; (0) :
jj ^ w \ (5.68)
7=1
where Q,7 is the Boltzmann collision matrix, that is, the Boltzmann collision
operator linearized around the uniform, zero-speed, zero-effective energy equi-
librium populations /; (0) = d = p/b:
% = (5.69)
One may wonder why the linearization is performed around the zero-speed and
zero-effective energy uniform equilibrium /,• = d, rather than around /, = f^°\
as it logically should. The reason is that the difference between the Boltzmann
10
The Boltzmann equation was discussed in detail in Section 4.2.2; we should keep in
mind, it is an approximate equation: strictly, in (5.67) we should write ' ^ ' rather than '='.
collision operator linearized around ft = d and around ft = / / 0 ) is of order

rj. Taking this difference into account would result in e2t]2 corrections to the
macrodynamics that would not be consistent with the chosen approximation
scheme (see Equations (5.64) and (5.65)).
From the Boltzmann collision operator (5.66), it can be shown that the
Boltzmann collision matrix Q i; is given by: 11
s
i i ) A ( s ^ s l SJ dP l
\ ~ d)bP'> ( 5 - 7 °)
s,s'ey
where p is the number of particles in the state 5:

b
1=1
Expression (5.70) provides the basis of a procedure for the computation of the
values of the matrix coefficients Q,7, once the collision rules A(s^s') of the lattice
gas model are given.
The Boltzmann collision matrix is not regular, because of the existence of
conserved quantities. Yet, Equation (5.68) has a family of solutions provided
the first order solvability conditions, (5.35) to (5.37), are satisfied. The desired
solution // (1) will be the only element in this family that also satisfies the
condition to have zero contribution to the mass, momentum and effective
energy (see Equation (5.4) and the discussion preceding (5.4)).
Using the solvability conditions together with the expression (5.27) for the
local equilibria fp\ one can re-express the r.h.s. of (5.68) in terms of the
gradients of the macroscopic variables p, j and w as:
E diyj5 + dl5jy
C
O f (1) ^ (, n ^
j j = y 5
Ti ^ ~ D ' 2
(5.71)
(9(erj2).
The standard procedure for solving (5.71) is to introduce Q [~1], the 'pseudo-
inverse' of the Boltzmann collision matrix. More precisely, consider the restric-
tion Q of Boltzmann's collision matrix Q, to the subspace of R b , orthogonal to
its kernel. This restricted operator is regular and its inverse operator Q" 1 can
be defined. The 'pseudo-inverse' Q [11 of Q is defined as an operator which acts
as the null operator on the kernel of Q, and as Q" 1 , on the subspace of ]Rb,
orthogonal to the kernel.
11
The linearized Boltzmann equation is discussed in detail in Section 6.1.
5.6 Transport coefficients within the Boltzmann approximation 131
In terms of Q[~1], the first order corrections f^ are:
f .(i) - 1 V Q["1] (c-c-t-^S*} diyjs + SlSJy
£2
l
. STct-lU 2
^\ z ( 5 - 72 )
+ (9(erj2).
It is then straightforward to compute the contributions of the first order popu-

lations / j ( 1 ) to the momentum and effective energy fluxes:
(5 ?3)
2 . . -
and
(5 74
] „ ., ->
Considering Equations (5.52) and (5.56) as definitions for v(c) and £(c), the
collisional contributions to the transport coefficients v and C, we can use the
relations (5.73) and (5.74) to obtain the desired expressions for v(c) and £(c)*
1 b
1
and
4i
The resulting values for v(c) and C(c) depend on the density per channel d through
Q[~l]; these are approximate results, since we used the Boltzmann approximation
in the computation. They are important theoretical results, but in practice, it is
more convenient - and also more accurate - to 'measure' the transport coeffi-
cients using a wave relaxation technique by lattice gas automaton simulations
(see Chapter 10).
5.7 Non-thermal models
Non-thermal models (see Section 4.5.1) differ from thermal models by the
conservation laws: non-thermal models conserve mass and momentum, but do
not conserve a linearly independent kinetic energy observable. Examples are the
FHP-II and FHP-III models. Consequently, the macrodynamics of non-thermal
models is governed by only D + 1 macroscopic continuity equations, while
there are D + 2 equations for thermal models. However, it would be erroneous
to think that the macrodynamics of non-thermal models simply degenerates
from the macrodynamics of thermal models, by 'omitting' the energy continuity
equation and the contributions of the effective energy density w to the mass
and momentum continuity equations. Indeed, there are substantial differences.
The first difference already appears at the stage of the nearly equally distributed
equilibrium distributions, which, for non-thermal models, read:
=
~h ~^~ ~p~CifXâ
b b - 2p , &c \ . . (5 77)
(see Equation (4.58) in Section 4.5.1). This expression cannot be obtained from
the equivalent expression (4.73) for thermal models by setting w = 0. Indeed, the
non-linear jajp term for non-thermal models involves the second order tensor
(ciaCia — ^<5aj?)> whereas the same term for thermal models involves the tensor
(cfaCia — Tâjs)- These tensors are different in the most general case, for example
for models with rest particles. This slight difference in the equilibrium distri-
bution produces modifications in the first and second order macrodynamical
equations.
5.7.1 First order macrodynamics

For non-thermal models, the momentum flux tensor ®Jf] = J2i Wipf^ takes a
form that remains rather similar to the equivalent expression (5.30) for thermal
models, except for the diagonal pressure term which is slightly modified. The
resulting first order macrodynamical equations for the mass and momentum
density are respectively:
(5.78)
Kk + dlP (g(p)jaup) + <?iaP(p, j 2 ) = (9(rj3\ (5.79)

5.7 Non-thermal models 133
where the pressure term reads:
(5.80)
The coefficient A, which is equal to 1 for thermal models (see Equation (5.32)),
is given by:
The factor g(p) is not modified:
, , b-2p b /4<JU D2
g(p)= b_ £>(2) + 2)V7 2 " + T
As already emphasized, the non-thermal pressure (5.80) cannot be obtained

from its thermal counterpart (5.32) by making w vanish, since the value of A,
the coefficient of the j 2 term, is a priori modified, at least in the most general
case. However, there are lattice gas models, such as FHP-1, which do conserve
kinetic energy, but do not enter the category of thermal models because the
kinetic energy observable is proportional to the mass observable (thus they are
linearly dependent). This is the case for the models where all velocity vectors
C; have equal moduli (homokinetic models). For these models, the coefficient £4
vanishes, and so the coefficient A is equal to 1, as for thermal models. As far as
only macrodynamical aspects are concerned, these homokinetic models can be
viewed as special cases of non-thermal models (with A = 1); they can also be
viewed as degenerated cases of thermal models (with vanishing effective energy
density w).12
5.7.2 Second order macrodynamics

The lack of an independent kinetic energy invariant has more dramatic conse-
quences on the second order macrodynamics. In the thermal case, the second
order momentum continuity equation (5.59) involve only the symmetric traceless
part of the momentum gradient tensor:
2
For non-thermal models, there is an additional diagonal term, proportional

to the divergence of j . There are two reasons for this: first, the equilibrium
distribution is different (no energy term), second, the lack of a kinetic energy
invariant releases the constraint (5.46) and its direct consequence (5.51).
12
For details on homokinetic models, see Frisch, d'Humieres, Hasslacher, Lallemand,
Pomeau and Rivet, 1987, or Rivet, 1988.
The second order macroscopic equations for the mass and momentum densi-
ties read:
dt2p = 0, (5.82)
and
St2ja + Stf —V dl0Ljp + dipja — l^d^jySap

J I
(5.83)
where the coefficients v and v', which a priori depend on p , are:
and
* > ^ (5.85)
The coefficient V in the additional diyjy term will be interpreted in Chapter 9

as the 'bulk viscosity coefficient'; this coefficient is a specific feature of (non-
thermal) lattice gases where rest particles are assigned an 'internal energy'
(see the discussion in Section 3.3.2). The bulk viscosity coefficient vanishes
for kinetic thermal models (such as the GBL lattice gas). The terms v ^ and
v/(c) are the collisional contributions to the coefficients v and V. They can be
'measured' numerically by sound wave damping simulations (see Chapter 10),
and they can be evaluated theoretically within the Boltzmann approximation
(see Section 5.7.4).
5.7.3 The macrodynamic equation

In the same way as we did for the thermal lattice gas (see Section 5.5), we merge
the first and second order contributions to the macrodynamics to obtain the
'macrodynamic equations':
dtp + dp(pup) = 0, (5.86)
(djp + dpja - idyjydzp) + V'idyjyS
+(9(e3rj) + (9(e2rj2) + < W ) .

The non-Galilean factor g is defined in (5.31), the pressure P, defined in (5.80),
depends on the macroscopic quantities (p,j 2), and the transport coefficients v
5.8 Comments 135
and v' are given in (5.60) and (5.62), respectively. The values of v and V also
depend a priori on the macroscopic quantities (p,j 2 ). However, any dependence
of v or v' on j 2 would lead to terms at least of order (e2rj2) in the macrodynamic
equations. For consistency with the order of approximation in the macrodynamic
equations, these terms are neglected; for simplicity, we shall use the short-hand
notation v and v' for v(p,0) and v'(p,0).
5.7.4 The transport coefficients

We proceed essentially along the lines of Section 5.6. The linearized Boltzmann
equation for the first order perturbation f^ reduces to:
(5.88)
2
+ (9(erj ).
As discussed earlier, for non-thermal models the tensor ciyCi$ — ^d ys replaces

the tensor ciyc^ — j}S ys of thermal models.
The collisional contributions v(c) and v/(c) to the shear and bulk viscosity
coefficients can be computed once the pseudo-inverse Q^11 of the linearized
Boltzmann collision operator is known. These contributions are:
v'(0 = 4-Vc^llci (5.89)
and
b
1 / 1 __ r „ \
- Dv^j. (5.90)
For homokinetic models (see Section 5.7.1), the bulk viscosity coefficient van-
ishes, and the Boltzmann expression for the shear viscosity coefficient is identical
to the shear viscosity of thermal lattice gases.
5.8 Comments
In this chapter, we developed a procedure inspired from the Chapman-Enskog

method (see Chapman and Cowling, 1970) to obtain the macroscopic equations
of lattice gases, starting from the microscopic dynamics. In Chapter 8, we present
a different and complementary approach to the 'micro- to macro-' derivation of
lattice gas hydrodynamics. We close by summing up the characteristics of the
LGA as a fluid model:
m The model is G-invariant (see Section 2.3.4).

IS- The model has crystallographic isotropy, at least up to fourth order (see
Section 2.3.5).
3* The model verifies the irreducibility condition (see Section 2.3.6).
m The model verifies the semi-detailed balance condition (see Section 2.3.1).
& All particles have unit mass.
& The collisional invariants of the model are the mass, the momentum, and
the kinetic energy for thermal models, and the mass and momentum for
non-thermal models (see Section 2.3.3).
Two additional assumptions define the regimes covered by the analysis:
S* The scale separation hypothesis, which justifies the Chapman-Enskog

procedure, expresses the hydrodynamic limit.
$? The local equilibrium around which the Chapman-Enskog expansion is
performed is 'nearly equally distributed', i.e. at any node, the mean
populations of all channels are close to a common value, which means
physically that the fluid velocity and the effective energy density are small.
The resulting macrodynamical equations describe the hydrodynamics of the lat-

tice gas. These equations are quite similar - albeit not identical - to the equations
of classical fluid mechanics. They appear clearly as continuity equations, they
are translation-invariant in space and time, and they are isotropic; however,
they are not Galilean invariant because of the presence of the factors g(p) and
g'(p), a consequence of the underlying discreteness of the microscopic velocity
space. In Chapter 9, we shall see that in certain hydrodynamic regimes, which
from the physical viewpoint are most interesting, the lattice gas is governed by
Galilean invariant equations.
Chapter 6
Linearized hydrodynamics
As LGAs are constructed as model systems where point particles undergo

displacements in discrete time steps and where configurational transitions on
the lattice nodes represent collisional processes, one can view the lattice gas as a
discretized version of a hard sphere gas on a regular lattice where particles are
subject to an exclusion principle instead of an excluded volume.1 The advantage
with LGAs is that, starting from exact microdynamical equations, statistical
mechanical computations can be conducted rather straightforwardly in a logical
fashion with well controlled assumptions to bypass the many-body problem.
This is well exemplified by the development in Chapter 4 leading to the lattice
Boltzmann equation. For the moment we shall consider the lattice gas automaton
as a bona fide statistical mechanical model with extremely simplified dynamics.
Nevertheless we may argue that the lattice gas exhibits two important features:
(i) it possesses a large number of degrees of freedom;

(ii) its Boolean microscopic nature combined with stochastic microdynamics
results in intrinsic fluctuations.
Because of these spontaneous fluctuations and of its large number of degrees
of freedom, the lattice gas can be considered as a 'reservoir of thermal excitations'
in much the same way as a real fluid. Now the question must be raised - as
for the hard sphere model in usual statistical mechanics - as to the validity
of the lattice gas automaton to represent actual fluids. In Chapters 5 and 8
we consider full hydrodynamics and macroscopic phenomena. Here we address
1
See also the discussion on the second virial coefficient in Section 4.6.
137
138 6 Linearized hydrodynamics
the question whether the intrinsic fluctuations in LGA capture the essentials of
actual fluctuations in real fluids as described by linearized hydrodynamics.
6.1 The linearized Boltzmann equation
We are interested in the dynamics of fluctuations which create perturbations

that are sufficiently weak to justify first order perturbation analysis. We consider
small deviations (5/(r*,c;t*) from the equilibrium distribution ^
Sf(u9e;u) = /(r*,c;t*)-/<«*>, (6.1)
which are produced by internal (spontaneous) fluctuations in position - velocity

space (r*,c).2 Here (5/(r*,c;£*) is the non-equilibrium ensemble average of the
fluctuations in the number of particles with velocity c at node r* at time U:
<5/(r*,c;fc) = (n(r*,c;£*) - (n(c)> )NE, (6.2)
where (n(c)) = / ( e q ) is the Fermi distribution (see Chapter 4) at global equilib-

rium:
/<«*> = (n(*)) = [1 + exp(-a + M ) ] " 1 . (6.3)
The space- and time-evolution of <5/(r*,c;£*) is governed by the equation

which is obtained by substituting /(r*,c;£*) = / ( e q ) + (5/(r*,c;£*) in the lattice
Boltzmann equation (4.7) and applying first order perturbation analysis. To first
order we obtain the linearized lattice Boltzmann equation:
7=1
with the notation /,-(r*;t*) = /(r*,c,-;t*) and where we have used the property
that the action of the collision operator on f^ is zero. The b x b matrix Q i;
denotes the linearized collision operator (5.69):
ssf
ss'
_ ^(eq)\-s; _ /i _ f(eq)\-:
X jj ) v1 j ; ;
2
The equilibrium state /(eq) is a global equilibrium /(eq)(c). Fluctuations can also be
related to external (forced) perturbations when a constraint is imposed to the system;
then /(eq) must be taken as local equilibrium /(eq)(r*,c).
6.1 The linearized Boltzmann equation 139
> i
(/ ( e q ) )(l -
Ksi — s
i) siys —* $ )
n //-(eq)\sfc/-i
k \J k ' V
_
1—\ 5
ss' ^~fj )
where the last line on the r.h.s. vanishes because of the equilibrium condition
(4.36).
For non-thermal LGAs (models with single non-zero velocity modulus) the
equilibrium distribution function (6.3) is independent of the velocity: / (eq ) = d,
the average particle density per channel. Then the linearized collision operator
reduces to (see 5.70)
ss'
where p = J2k sk *s the number of particles in configuration s. When detailed

balance A(s —> sf) = A(sf —> s) is satisfied, (6.5) can be written in symmetrical
form:
1 ^ f / -i b- -l
2
ss'
For multi-speed models (like thermal LGAs) this symmetry property does not
hold. Nevertheless a symmetrical expression can be obtained (Ernst and Das,
1992) by considering the modified operator Slijxfp where KJ^ = / J q ( l — /J q )
is related to the static correlation function S(k) (see Section 4.7). Defining the
factorized equilibrium distribution function:
b
f (s) = I I [êq))S* t1 - f{kq)f-Sk)],

{eq)
(6-7)
k=\
(which is a- and /^-dependent) and using the expression obtained above for Q I; ,
we have:
ss'
( f (eq)
s/
- ) n 7Z7(
l
k \ h
where, with the explicit expression for the equilibrium distribution function (6.3),
Here p(s) = ^2k Sk and e(s) = ^2k e(ck) Sk are respectively the number of particles
in configuration s and the corresponding energy; these quantities are conserved
during collisions, i.e.
P(s) = J2Sk =
k k
e(s) = ^ P e(ck) Sk = J2e(ck)s'k. (6.11)

k k
So, we obtain an explicitly symmetrical expression for the modified linearized

Boltzmann operator:
ss'
which is valid for any model satisfying detailed balance.

We close this section with a technical point. For later use we introduce the
thermal scalar product:3
(A\B) = ^^(cOBte)*,-. (6.13)
With this notation, the symmetry of Cl^rc^ implies:
(A\a\B) =
ij
a property that will prove to be quite useful.
6.2 Slow and fast variables
In the present chapter we are interested in describing the space and time
dynamics of fluctuating quantities, such as the number density, the current
density, and the energy density, in the lattice gas. The relevant objects for
evaluating the dynamics of these quantities are their autocorrelation functions,
which involve the value of the fluctuating quantity at some arbitrary initial
time, provided the system is in a stationary state. Therefore the computation
of the correlation functions can be formulated as an initial value problem, and
the projection operator technique is quite appropriate (Zwanzig, 1965; Mori,
1965a, 1965b). It provides a convenient separation of the dynamical variables
into slow variables and fast variables: the latter correspond to the kinetic modes
and they decay practically instantaneously in comparison with those variables
which vary slowly in space and time. These slow variables can be identified
when the constants of motion of the system are known; we shall see that they
correspond to the hydrodynamic modes.
3
The weight K\ is p dependent.
It is convenient to rewrite the deviation 5f{r*,Ci\U) from the equilibrium

distribution (see (6.1)) as:
(5/(r*,c,;t*) = fc^cOMr;*), (6.15)
and consequently the density fluctuations, the current density fluctuations, and
the energy density fluctuations respectively as:
5p(u;t*) = £)fcf^(r;t), (6.16)
j
fcjhint), (6.17)
8e(u;t*) = \
We shall describe the long-time behavior of the lattice gas in terms of </>;(r;t).
Long times are large compared to the microscopic time At over which collision
and propagation take place. So far we have omitted to write explicitly the time
interval At because it is just the automaton unitary time, but here it is important
to do so for the clarity of mathematical manipulations; then the discrete time
variable is U = nAt, where n is an integer. So the linearized Boltzmann equation
(6.4) now reads:
* + At) - <fc(r*;**)
j
where the linearized collision operator is redefined as:

Clij = K^ChjKj, (6.20)
with Kt = Kp\ci) = Kf\
In order to obtain the large-scale description that we are seeking, we eliminate
the fast variables by means of a projection operator ^ , which when operating
on a function g(c) projects the function onto the subspace iô spanned by the
eigenvectors of Q^ with zero eigenvalues:
4
^g(c)= ^2\aa)NAa«\g), (6.21)
a=l
where aa(c) = {l,c,e(c) — c2T) is a set of orthogonal vectors and Na is a

normalization factor defined by (aâp) = N'1 Sap. The constant cT is defined
such that:
<ao|fl 3 > = ( l \ ( \ c 2 - c 2 T ) ) = 0 ; (6.22)
its physical interpretation will be given subsequently. It is easy to check that the
operator ^ , as defined by (6.21), has the properties of a projector, i.e.
^2 = ^ , and 0>g = g, VgGô; (6.23)

the complementary operator SL = 1 — 9 projects onto the subspace orthogonal

to-T o .
It will be most convenient to perform the subsequent developments using
Laplace-Fourier transformations; therefore we define the double transform of
the discrete dynamical variable a(r*;£*) as:
00
a(k;s) = X) 5 1 a(r*;t*)e-lkue-snAtAt. (6.24)
Applying this definition to 0,-(r*;£*), we project $(k,c;s) onto "To to obtain the
Laplace-Fourier transformed hydrodynamic variables, i.e.
4
^0(k,c;s) = ^ K)iV a 2 a (k;s), (6.25)
a=l
with
Ax(k;s) = dp(k;s) = J2K?)kKc;s)9 (6.26)
i
A2(k;s) = 3/x(k;s) = ^ ; c f ) c , ^ ( k , c ; S ) , (6.27)

i
A3(k;s) = ^(k;s) = ^ K P c j > y # , c ; S ) , (6.28)

i
A4(k;s) = 3e(k;s) — c\dp(k;s)

; s ) (629)
4 - -
Now Laplace-Fourier transformation of the linearized Boltzmann equation
(6.19) yields:
Uk;s) - &(k;s) - ^ f l y ^(k;s) = e s A ( A t ^ ' ^ ^ ( k ) , (6-30)
j
where (/>/(k) = <£j(k;£* = 0) is the space-Fourier transform of </>,-(r; t) at initial

time.4
We introduce the following notation ^ ( k ; s ) = |c/>(k;s)), and we define the
operator S£ by:
JSf |0(k;s)> = ^ lk - c ' At ^ (5 y + Q y )^(k;5), (6.31)
4
The factor exp(s At) on the l.h.s. of (6.30) is a consequence of discreteness. Note that by
taking the limit
lim \- \esAtcl)(s)-<i)(s)-esAtAt<l)(O)\ = s<j>(s)

Af->0 At [ J
one retrieves the usual expression of the continuous transform
,00
/ e-stdt<t>{t) =
Jo
to rewrite Equation (6.30) as:
(esAt - <£) |0(k;s)> = esAtAt |0(k)>. (6.32)
This result exhibits an appropriate form for applying the projectors 9 and 2, to
separate the equation into a 'hydrodynamic equation':
(6.33)
and a 'transport equation':
esAt2\cl)(k;s)) - 5JSP5|0(k;s)) - 2 & & \(/>(k; s)) = esAtA*i2 |0(k)),

(6.34)
where we have used the property (3P + 1 ) \(j>(k;s)) = |$(k;s)).

Our goal is to solve the hydrodynamic equation, but we observe that the third
term on the l.h.s. of (6.33) requires the knowledge of the perturbation <H \<j>(k;s))
(which is orthogonal to the slow variables), that is we must first obtain the
solution to the transport equation (6.34). Therefore we consider the resolvent
operator:
R = [esAt- l£>]-\ (6.35)
with the property:
= \esAt- bSeV SL = \esAt- £S£2>\~ J , (6.36)
which follows from the operator identity: R = S + RiS'1 - R~{)S, with

S = [exp(sAt) - £&Q\-\ and {S~l - R~l) = 2^^. We can then write the
formal solution to Equation (6.34) as:
) (6.37)
and inject this result into Equation (6.33) to obtain:
(esAt - 0>&)0>\<j>(k;s)) - ^^R2^^\(t)(k;s))
= esAtAt(^\(j)(k)) + ^jgfjRJ|0(k))), (6.38)
where there remains a dependency on J|<^(k;t* = 0)), the initial value of a

quantity orthogonal to the slow variables.
We are now in a position to write, formally at least, the set of equations
for the hydrodynamic variables, by using explicitly (6.25) and taking the scalar
product of (6.38) with the vectors (a a |, (a = 1,2,3,4). The result is the relaxation
equation:
4
sAt
e Aa(k;s) - ^2 -S?«*4r(k;s)-
<7=1 a=\
= e^'AtA^kiO) + / a(k;s). (6.39)
where (i) the second term on the l.h.s. represents the coupling between the
hydrodynamic variables, and the matrix elements J£?ao- are given by:
JSP*, = (aa\g\aa)NG, (6.40)
or, with (6.31) and noting that £10>\4>) = 0:
^ = <flak" Ik>cAt K>N ff ; (6.41)

(ii) the third term on the l.h.s. expresses collisional effects, and the matrix
)NG (6.42)
is referred to as the memory function]

(iii) the term / a (k;s) on the r.h.s.
7a(k;s) = (aa\&£ResAtAt£\(l)(k;0)), (6.43)

contains the contributions from the initial state and also includes collision effects.
Note that to obtain (6.42) and (6.43), we have used the property R SL = £R£.5
The set of equations (6.39) describes the dynamics of the hydrodynamic
quantities J4 a (k;s); these equations are exact within the limits of validity of the
linearized Boltzmann equation.
6.3 The hydrodynamic limit
We now want to obtain an explicit analytical formulation of the dynamics of

the lattice gas automaton fluctuations in the hydrodynamic limit, that is for long
times and large spatial scales (t much larger than At and k = |k| much smaller
than the reciprocal lattice unit length). Therefore we consider the limit of small
At by expanding our results in s and k (to second order); for instance the factor
esAt e x p a n c [ s as:
esAt ~ I + sAt + i(sAf) 2 . (6.45)

5
This property follows from the definition J R 1 = R1 - 0>R J , and the observation
that the action of the second term vanishes; indeed, using the operator identity given
after Equation (6.36) and with the notation T = [esAt — i f ] " 1 , one has
= 0, (6.44)
because 0> commutes with i f and with T.
The expansion of the other quantities in (6.39) is not so trivial, and they will
now be considered separately.
6.3.1 The coupling function

For the matrix elements jS?aff, (6.41), we have:
~ (aa\(l-ik'cAt-±(k'cAt)2)\aa)N
At2ff
~ S^ - ikiAt<aa|J<)Na - fc/UJjl-ylJZ1)Nff. (6.46)
Here k\ (I = {x,y}) denotes the components of the wave vector k,6 and |J a ) the
current associate to |a a ), with:
(
0 ikx iky 0
0 ikx
0 iky
%% o
0 ikx(c2s-c2T) iky(c2s-c2T) 0
= Lm, (6.47)
where the constants c2 and c\ are given by:
4 - i^.
6.3.2 The memory function

The memory function matrix (6.42) is more complicated: its exact computation
would require the inversion of the resolvent matrix R, which is analytically not
possible; so we shall use a perturbation procedure.
We first consider the right vector:
J8J2? \aa) = 2L [e- l k ' c A t (l + O)] \aa) ; (6.50)

expansion to first order gives
2&\aff) ~ -ikiAt£\JlG). (6.51)

Similarly for the left vector:
<a,|J&?<2 = (aa\[e-lk-*At(l+&)]$, (6.52)
6
In (6.46), the terms O(k2) are related to the propagation step in the lattice dynamics;
such terms are not present in continuous theory; here these terms will be incorporated
in the transport matrix yielding an extra contribution to the transport coefficients.
we obtain:
(6.53)
We observe that the current density appears as:

4
a=l
this quantity is called the subtracted current, and is orthogonal to the hydrody-
namic eigenvectors.
We now expand the resolvent operator:
cAr)(l + Q)j] *. (6.55)
One can easily verify that J£X2|<£) = A|0) because J projects any vector
onto the subspace orthogonal to the subspace spanned by the vectors with zero
eigenvalue; so in practice we can substitute £l£ll by Cl. On the other hand,
we notice from Equation (6.34) that R~{ = (exp(s At) — 1 <£) operates on the
projected vector J \(f>), and since I21 \cj>) = 11(/>), we may therefore, in (6.55),
replace operationally J 2 by the identity operator. With these simplifications, we
have
R ~ [sAt-Q + JOk-cAOa + Q ) ^ ] . (6.56)

Combining (6.51) and (6.53) into (6.42), we have
Kw ~ -kkm (At)2 <J< |(1 + Q) R |J-) No, (6.57)
with JR given by (6.56).7 We now treat the resolvent by applying the operator
identity
l
A=1B-1B{A-B)^ (658)
to (6.56); with the notation
H(k) = 5(ik • cAt)(l + Q)&, (6.59)
we obtain
l l
R = —^ —^H(K)R, (6.60)
w
sAt-Q sAt-Q
7
Expression (6.57) has a structure similar to the memory function in continuous theory
(see e.g. Boon and Yip, 1980, Chapter 3).
which is iterated to yield
sAt-Q sAt-Q, sAt - Q

1 1
- ••• . (6.61)
We take (6.61) as a perturbative expansion in powers offc,so that the memory

function can be written as:
(6.62)
where Kj$ is proportional to kn, that is to lowest order in k:
Kaa ~ *:£> = -ktkm (At)2 <J< |(1 + Q) sAt{_a \Ja) Na. (6.63)
This memory function is well behaved in the limit s —> 0 and for all values of
fc, provided the non-zero eigenvalues of Cl be sufficiently different from zero at
k = 0. This 'correct' long-time behavior is a consequence of the combination
of the operator (s At — Q)"1 with the projector =2 which prevents singularities
(~ s"1) that would result from the action of Cl on the hydrodynamic variables
(which have zero eigenvalues at k = 0).
6.3.3 The random force term

The term / a (k;s) in (6.39) and given by (6.43), expresses the effect produced by
the fast variables from the initial state, arid is usually referred to as the 'random
force'.8 Since we are interested in the long-time behavior, we use (6.53) and the
resolvent expansion (6.56) in (6.43), to obtain:
Ja(k;s) ~ -ik (At)2 <J< |(1 + Cl) T ^ T5|0(k;O)>. (6.64)

[AJf(k)J
If the initial distribution function has the form (see (6.15)):
/(r*,c;0) = fM>(c)
+ Kf(c) (d(r*) + C2(r*) • c + C3(r*) (^c2 - c2T)) , (6.65)
where Cy depends only on the space variable, then J|</>(k;0)) = 0 (because =2

projects on the subspace orthogonal to the subspace spanned by {l,c, \c2 — c2T})\
this result holds Vt > 0, and / a (k;s) vanishes in the long time limit.
What about initial states with a distribution function different from (6.65)?
From (6.64), we observe that / a (k;s) is governed by the application of the
8
In the Langevin formulation of relaxation dynamics, the random force is often inserted
'by hand' in the evolution equation in order to account for the effect of the microscopic
fluctuations on the dynamics of the macroscopic variables; here we have an analytical -
although implicit - expression in terms of the kinetic variables.
resolvent operator on «2|0(k;O)); the result of the application depends on the

eigenvalue spectrum of the collision operator H, and we may expect that, for
small values of k, the operator [ H(k) — Q] exhibits an eigenvalue spectrum
equal to that of A plus a term proportional to k. Let Xn(k) < 0 be the non-zero
eigenvalues of the eigenvectors corresponding to the fast variables; then the
time behavior of Ja(k;£*) should be of the type:
b
Ia(k;t*) = ] T ^ ) < a w ( k ; 0 ) , (6.66)
n>5
where a n (k; 0), (n = 5,...,fo),corresponds to the fast modes. The negative eigen-
values (distinct from the zero eigenvalues, n = 1,...,4) are responsible for the
fast decay and eventual vanishing of J a (k;t*). 9 So we may safely assume that in
the hydrodynamic limit:
lim lim/ a (k;t*) = 0. (6.67)
££•00
-•00 / 0
6.3.4 The long-wavelength, long-time limit

We can now collect our results (6.46), (6.63), and (6.67), to write the relaxation
equation (6.39) in the long-wavelength, long-time limit, i.e. for fc|c|At <C 1 and
sAt < 1. Noting that:
2«, = Sm - ikiAtiatlJbN, + (9(k2), (6.68)

and that Kaa = (9(k2\ we obtain to first order, i.e. (9(s) and 0(k):
(1 + sAt)Aa(k;s)-(l + sAr)Ar^ a (k;O)- ] T Sa(7Aa(k;s)
+ ik,At ^2{aa\j'a)NaAa()t;s) = 0. (6.69)

a
This result is divided by At to yield, in the limit sAt <C 1:

sA«(k;s) + ikiîaMNaAîs) = A(k;0), (6.70)
a
which corresponds to the Euler equations.
To second order, (9{s2) and 0(k2), we have:
(1 + sAt + ^s2At2)Ax(k;s)
- ( 1 + sAt+ ^s2At2)AtAa(k;0)
J2 Lxa(k)AtAa(k;s)
9
The decay rate is model dependent; this point and the related question of the separation
of eigenvalues at low k will be examined in the next chapter in the light of specific LGA
model computations.
A/-2 \fl
— \J m) + (V\ \Jm)
2 sAt — Q
IffAff{k;s) = 0. (6.71)
Again we divide this equation by At, and take the limit sAt <C 1, to obtain:
52 a (k;s) + -s
+ fe/fcm
- Q
+ At(Jla\ - ^ |#> 1 N.A^kis) = A«(k;0), (6.72)

J
s — Q
where 10
£1 = SI/At (6.73)
We now use the first order result, Equation (6.70), to evaluate the second term
on the l.h.s. of (6.72), i.e.
s2A«(k;s) = sK(k;s)] = s [-ikt ]T (a^NpAp], (6.74)
where we have omitted the term on the r.h.s. of (6.70) because when inserted in
(6.72), it will give ^sAtA^k',0) which vanishes in the limit sAt < 1. Iteration of
(6.74) gives:
J ^ N c A ^ (6.75)
where
- <W). (6.76)
Substitution of these results into (6.72) yields:
*»*m 5 1 *„,*;«)]a,(k;s) = ^ a (k;0), (6.77)
with the memory function:
- J - ^ \J:) + A t ( J ^ + - ^ | J r ) l N*. (6.78)

10
The difference between fl and A is immaterial in practice (since At is the unitary time
interval). However, in the continuous Boltzmann equation the collision operator has the
dimension^ of a frequency, while in the discrete linearized Boltzmann equation (6.19), the
operator Q is dimensionless. So (6.73) re-establishes the correct dimensional content of
the collision operator.
Using (6.73) in the second term:
(6.79)
then taking the limit sAt —> 0 in (6.78), we obtain: 11
^ ^ | ) , ( 6 . 8 0 )
z J
s— Q
These results, (6.77) with (6.80), conclude the derivation of the linearized hydro-
dynamic equations.
6.4 The transport matrix
The results obtained by computation in Laplace space are naturally equivalent

to the results obtained by computation with the time variable expressions.
However, some physical insight can be gained from the memory function (6.78)
when expressed in real time. In the linear equations (6.77), the memory term
âa(k; s) Aa(k; s) is the Laplace transform of the convolution:
t=nAt
] T * a f f (k;T)4 T (k;fc-T)A T , (6.81)
T=0
with the kernel:
The memory function expresses the effect of the past history of the dynamics
related to the fast variables. But here we are interested in the dynamics of the
slow variables Aff(k; U) which relax over times {(9(k2)) very long compared to the
characteristic decay rate of Oa(7(k;T) ~ exp(Ai), where X denotes the non-zero
(negative) eigenvalues of the collision operator. Consequently, in (6.81), we may
consider, to good approximation, the following simplifications: (i) Aff(k;t*) does
not vary significantly over the summation time, and can therefore be taken out
of the sum; and (ii) Oa<T(k;r) decays sufficiently rapidly that it is practically zero
for times larger than T ~ (9(X~l), and the summation can be extended to oo. As
a result the convolution (6.81) becomes Aa(T Aff(k;t*) with:
T=0
=E H=0
11
The limit 5 —> 0 in Laplace space corresponds to the limit t —> oo in real time.
6.5 Comments 151
[0a\ <^h \%) - At^no (J'A\J?) ]KAZ, (6.82)

n=0
which defines the transport matrix Aaff. Note that the Laplace transform of the
bottom line in (6.82) gives exactly (6.80).
The transport matrix (6.82) can also be written formally as:
00
e St
K*(s) = N, £ ~ 01 \J(t)i) ^ . (6.83)
t=Q
This expression is the analogue of the Green-Kubo integral in continuous

theory; at zero frequency (s = 0) and in the continuous limit (At —• 0), one
retrieves the classical formulation of the transport coefficients:
Ko = AT, r dt(Jl\J(t)lff). (6.84)

Jo
There are two important distinctions between the continuous expression and
the lattice gas result: (i) in the latter, instead of a time integral, we have a
summation, which is a consequence of the persistence at the macroscopic scale
of the space and time discretization at the microscopic level; (ii) the transport
coefficients contain an additional (negative) term — \N a(j[\J^)At, (see (6.80) and
(6.82)), which is another consequence of the discrete nature of the dynamics. 12
The transport coefficients were discussed in Section 5.6; they will be further
discussed and evaluated explicitly in the subsequent chapters.
6.5 Comments
In this chapter, we derived the linearized equations describing the space-time

evolution of the fluctuating quantities in the lattice gas at the hydrodynamic
scale. Using the projection operator method, we obtained a logical separation
into slow and fast variables; the ensuing set of relaxation equations for the
hydrodynamic variables follows from the lattice Boltzmann equation. There is
no analytical procedure by which, starting from a microscopic formulation, one
could derive the hydrodynamic equations in an exact manner, but at least the
approximations used in the derivation can be justified physically on the basis
of scales separation. The resulting description of the dynamical behavior gives
a picture of the system as viewed through a space-time window whose width is
12
In the literature, the negative contribution to the transport coefficient has often been
called a propagative part invoking the representation of the lattice gas in terms of
particles undergoing, during the propagation step, free flight from node to node over the
time interval At.
set by the observer in the range of sufficiently large wavelengths. In the next
chapter we shall see that this picture, as provided by a simplified system such
as the lattice gas automaton, is in accordance with the behavior of spontaneous
fluctuations as observed in real fluids.
Chapter 7
Hydrodynamic fluctuations
The object of the present chapter is small deviations from local equilibrium
which are triggered by spontaneous fluctuations. In real fluids these fluctuations
which temporarily disturb the system from local equilibrium are such that a
fluid at global equilibrium can be viewed as a reservoir of excitations extending
over a broad range of wavelengths and frequencies from the hydrodynamic
scale down to the range of the intermolecular potential. Non-intrusive scat-
tering techniques are used to probe these fluctuations at the molecular level
(neutron scattering spectroscopy) and at the level of collective excitations (light
scattering spectroscopy) (Boon and Yip, 1980). The quantity measured by these
scattering methods is the power spectrum of density fluctuations, i.e. the dy-
namic structure factor S(k,co) which is the space- and time-Fourier transform
of the correlation function of the density fluctuations. The spectral function
iS(k, co) is important because it provides insight into the dynamical behavior
of spontaneous fluctuations (or forced fluctuations in non-equilibrium systems).
Whereas the fluctuations extend continuously from the molecular level to the
hydrodynamic scale, there are experimental and theoretical limitations to the
ranges where they can be probed and computed. Indeed, no theory provides
a fully explicit analytical description of space-time dynamics establishing the
bridge between kinetic theory and hydrodynamic theory. Scattering techniques
have limited ranges of wavelengths over which fluctuation correlations can be
probed. Numerical computational techniques can realize molecular dynamics
simulations (Boon and Yip, 1980) covering in principle the whole desired range,
but in practice there are computation time and memory requirement limitations
153
154 7 Hydrodynamic fluctuations
(Mareschal and Holian, 1992). So lattice gas automata provide an interesting

alternative from both computational and theoretical viewpoints.
7.1 The dynamic structure factor
The dynamic structure factor is defined as the double Fourier transform with
respect to space and time of the van Hove function G(r, £•), the correlation
function of density fluctuations (5p(r*,£*) around the equilibrium state (Boon
and Yip, 1980):
00
pS(k,co) = Yl Y, e-'°"*-'kl*(5p(T*,\u\)Sp(0,0))
r* t*=—oo
y1 J2 k,0)), (7.1)
t*=—oo
where
h(k) = J2 exp(-zk • r*)/z(r*) (7.2)
denotes the discrete spatial Fourier transform, and V = LD = Jf is the number

of nodes in the D -dimensional lattice $£. In the case of deterministic and invert-
ible dynamics, time reversal invariance of the microscopic equations of motion
guarantees that the van Hove function is even in time. In the case of determinis-
tic but non-invertible, or stochastic dynamics, S(k, co) is simply defined through
(7.1), as twice the half-sided temporal Fourier cosine transform, where the time
evolution is defined for positive time argument only. The density fluctuation can
be expressed in terms of the fluctuations of the channel occupation variables,
nt(i = 0 , . . . , 6 - l ) :
The average is taken over an equilibrium ensemble and the equilibrium distri-
bution has the form of the Fermi-Dirac distribution (see Chapter 4):
• (7.4)
Here we consider global equilibrium, where the system is macroscopically at rest

(u = y = 0).
The most important quantity in the non-equilibrium description of fluids and
LGAs is the kinetic propagator or Green's function:
r l7 (k, u)Kf = (SmiK u)dn*(K 0)> (7.5)
7.2 Fluctuation correlations 155
where drii(k, t*) is the spatial Fourier transform of <5ftj(r*, £*), and KJ^ is the equal
time correlation function of the drift (see Section 4.7). With (7.3), the dynamic
structure factor (7.1) is expressed in terms of the propagator (7.5) as:
(7.6)
and the static structure factor S(k) - the Fourier transform of the equal time
density correlation function, discussed in Section 4.7 - is obtained from (7.5)
taken at t = 0:
7.2 Fluctuation correlations
To study the fluctuations dynamics, the main quantity to be computed is

the kinetic propagator F ^ k , £•) which we will calculate in the mean field or
Boltzmann approximation. Starting from the lattice Boltzmann equation (see
Chapter 4):
fi(u + c, U + 1) = /;(r*, U) + AJB) (fa}), (7.8)
we want to study fluctuations which disturb the gas locally from its equilibrium
state. We proceed as in Chapter 6 by expanding the collision term AJ5) around
the stationary distribution / , (7.4):
A[ s ) (/ + Sn) = A[ B) (/) + ^ OijSnj + ^ 0(SnjSnk). (7.9)
Here / satisfies &\B\f) = 0 (because of mass conservation), the b x b matrix

n is the linearized Boltzmann collision operator, and the definition of drit is
contained in (7.3). Combination of (7.8) and (7.9) gives the linearized lattice
Boltzmann equation:
drii(u + ci? u + 1) = <5n,-(r*, U) + ^ Qy5w7-(r*, t*). (7.10)
This equation is equivalent to that obtained in Section 6.1. By space Fourier

transformation, we rewrite Equation (7.10) as:
<5n,(k,U + 1)
which expresses the fluctuation dnt in terms of its value at the previous time step.
Solving Equation (7.11) by iteration, we obtain Srii(k, t) from the knowledge of
its initial time value:
, (7.12)
ij
which is valid for U > 0. Here [exp(—ik • c)]; ; = <5;;-exp(—ik • c,-) is a diagonal
matrix. Substituting (7.12) in (7.5) and noting from (7.5) that r i ; (k,0) = 5iJ9 we
obtain the expression for the propagator:
Tl7(k, u)Kf = [e-lk'c (I - O) ]'; Kf. (7.13)

We now introduce the resolvent, i.e. the Laplace transform of (7.13):
t=0
r In,
[ e s + i k c _ j[ _ Q J .j* ' W-1^
where the second equality follows from X^JLo*" = (1 — x) —1 , |x| < 1. After
summation over (i9j)9 we rewrite (7.14) equivalently as:
ijl
then we make use of the relation J \ QjKj = 0 (which follows from mass
conservation) to obtain:
E iVM*? = E 42) + E V
The dynamic structure factor iSf(k,co) is now readily obtained from (7.6) by
setting s = ico in (7.16):
pS(k,co) = 2 ReF(k, co), (7.17)

+lkx 1 {
F(k,co) = ^ { [ ^ - l l - l l ] " + ^} K ? (7.18)
where Re denotes the real part.

This is the general expression for the dynamic structure factor within the
Boltzmann approximation. The inverse matrix in (7.18) is b-dimensional and
depends on to and k: it is a complicated mathematical object which in all
generality resists analytical inversion. Nevertheless (7.18) can be used for direct
numerical evaluation once the explicit collision matrix of the LGA model is
known, as will be illustrated later. There is one extreme case where the explicit
7.3 The hydrodynamic modes 157
computation of S(k,co) is straightforward, that is for very short wavelengths (or

when the particles are in free flight motion); then from (7.6) and (7.13) with
n = 0, we obtain
pS(k,co) = R ^ £
t=—oo ij
Y ^ ™ + pS(k); (7.19)
this is the spectrum for the ballistic regime. On the other hand, when one is
interested in the hydrodynamic regime, i.e. when the system is probed with a
wavelength large compared to the microscopic scale (the inter-particle distance
in real fluids; the lattice unit length in LGAs) and large compared to the mean
free path between collisions (this condition prevails for dilute fluids), then one
can compute S(k,co) analytically from (7.18) in the limit of weak spatial and
temporal variations (|k| —> 0, co -> 0).1 The hydrodynamic structure factor is
the subject of Section 7.4.
7.3 The hydrodynamic modes
7.3.1 The spectral decomposition

To compute S(k,co) analytically from (7.18) in the hydrodynamic regime, 2 -
the domain of small values of |k| and co - a perturbation method can be
used. We shall search the solutions of the linearized Boltzmann equation in
the hydrodynamic limit by an eigenvalue problem technique (Resibois and de
Leener, 1977) on the basis that the hydrodynamic modes correspond to the
eigenvectors whose eigenvalues approach zero when |k| —• 0. We shall find
that, up to second order in fe, these eigenvectors yield the searched propagative
and diffusive modes characterizing the dynamics of the fluid in the linearized
hydrodynamic regime.
In Section 7.2, to obtain (7.16), we have used one of the right eigenfunctions
of Qtj with zero eigenvalue, i.e.
ijKfaniCj) = 0, (7.20)
with fj given by (7.4), and where

an(ci) = {lchei=1^} (7.21)
1
Physically [ |k| —> 0, co —> OJ means [ |k|/f < 1, COT/ < 1 ] , where Sf is the mean free
path and Xf is the mean duration between collisions, that is we consider large
wavelength ( > Sf), long time ( > Xf) phenomena.
2
The regime considered here is for the fluid at global equilibrium where linearized
hydrodynamics holds.
are the collisional invariants.3 We will also need the left eigenvectors with zero
eigenvalue of the non-symmetric matrix O. They follow from the conservation
laws for the total number of particles, the total energy and the total momentum:
£,-an(c0A<*)({n7-}) = 0,or
^ a n ( C l ) Q l 7 = 0. (7.22)
i
We now consider the eigenvectors and eigenvalues of the matrix

[exp(—zk • c)(ll + £1)] which is non-symmetrical and non-Hermitian. The right
eigenvectors are defined through the relation:
, (7.23)
where \\p^) is the b-component vector |i/^)/ = K^xp^Ct) (i = 1,...,&), and z/i(k)
denotes the associated eigenvalue. The zero eigenvectors an(c) of O represent
the right eigenvectors xp^ in the long wavelength limit (k —• 0). Denoting the
eigenvectors by:
Mi = {\p)i9\e)u\g)i} = ic| 2 ) {l^c?,q}, (7.24)

the eigenvalue equations (7.20) and (7.22) can be written as:
(an\n = 0. (7.25)
The zero eigenvectors an(c) of £1 represent the right eigenvectors xp^ in the long
wavelength limit (k —> 0).
The left eigenvectors, ( ^ | , defined through the relation:
(^(k)\e-lk'e(l+a) = ^ k ) (^(k)|, (7.26)
differ from the right eigenvectors |y^(k)). Taking the transpose of (7.26) and
using the symmetry of the thermal scalar product introduced in Section 6.1 (see
Equations (6.13) and (6.14)):
= (B\a\A) = YÂiOijKfBj, (7.27)

ij
and comparison with (7.23) yields:
(/>,(k) = ^ k > , ( k ) M , ( k ) , (7.28)
where J^^k) is a normalization constant. Neither the right nor left eigenvec-
tors are orthogonal, but they form a complete bi-orthonormal set satisfying the
relations:
M V ^ = <V (7.29)
3
(7.20) follows directly from the observation that for a stationary distribution with
Lagrange multipliers bn (and also with bn + Sbn), A^B\f(bn)) = 0.
With the help of these eigenfunctions, we make the following spectral decom-
position of the Boltzmann propagator (7.13):
(k)|; t > o. (7.30)
With the notation introduced above, we rewrite the spectral function F(k, co)
as:
to» {[e!W-2"(k) - I]"1 + i }
(fl>), (7-31)
where, for later convenience, we have introduced the quantities:
At this point we can verify that the first swm rw/e is satisfied, that is the
integrated intensity of S(k,co) (at fixed k-value) yields the static structure factor
iS(k) (7.7). This is easily seen as follows: we evaluate the integral (by contour
integration):
2Re fK dcoS^co) = 2Re f dco {[el(0~z^ - I]" 1 + 1} = 2rc, (7.33)

J— n J—n
and we use the relations (7.7), (7.24) and (7.25) to obtain (p\p) = pS(k) and
hence:
27T S(k) " ^
Each mode contributes a spectral line in (7.34), and their total weight factors
add up to unity.4
The dominant contributions to F(k, co) come from small values of the de-
nominators in (7.32) which are obtained for the eigenvalues corresponding to
the hydrodynamic modes since in the hydrodynamic regime these eigenvalues
z^(k) are either (9(k) or (9(k2) and k(= |k|) is small. So the hydrodynamic modes
with z^(fc) —• 0 as k —> 0 are slow modes (that is they exhibit slow decay in
space and time). In the long wavelength limit, there are also kinetic modes,
with Rez /i (0) < 0, which give terms with fast exponential decay in (7.30) and
4
At finite |k|-values, 'weight factors' are not necessarily positive, because the matrices are
complex, but not Hermitian.
whose contribution to F(k,co) is therefore negligible with respect to the hy-

drodynamic modes. Consequently, when we are interested in the hydrodynamic
behavior, we can restrict the summation in (7.31) to the /is corresponding to the
hydrodynamic modes whose effect is dominant.
7.3.2 The eigenvalues

We start from the eigenvalue equation:
In the long wavelength limit the hydrodynamic modes ty^(k) and eigenvalues
z^(k) can be determined by a Taylor series expansion:
z,(k) = ikzM + {ik)2z<jP + --. (7.36)
Substitution of (7.36) into (7.35) gives the set of equations:
<°>> = 0
f ^ ^ f i (7.37)
with c\ = k • c, where k is the unit vector along the k-direction. The general
solution of the zeroth order equation is an arbitrary linear combination of the
collisional invariants (7.24):
(7.38)
Here c\ and c± = k± • c are respectively the longitudinal and transverse com-

ponents of the microscopic momentum g(c) = c.5 The microscopic entropy,
s(c) = e(c) — hp(c) = \c2 — Cg, and the microscopic pressure, p(c) = \c2, are
linear combinations of the collisional invariants p(c) = 1 and e(c) = \c 2 in
(7.24). The constant h:
h c = (P\P) _ Lî'-i^j_ (TS9)

(p\p)
is chosen such that the microscopic pressure and entropy are orthogonal, 6 i.e.
(s\p) — 0. We have also introduced the quantity c s which will appear to be the
'adiabatic' speed of sound.
5
This decomposition will appear quite logical in terms of the longitudinal and transverse
modes.
6
The choice of the couple of thermodynamic variables s and p is convenient here, but
density and temperature would also be appropriate.
The coefficients Bn in (7.38) and eigenvalues z^ are determined by solving

exactly the eigenvalue problem to order k in the subspace spanned by {s, p, cu c±}.
Multiplication of the second equation in (7.37) to the left with (am\ yields:
m\ci+zM\an)Bn=O9 (7.40)
with n,m = {s,p,/,_!_}. The matrix (am\ci + z^\an) reads:
s p I _L
s (s\s) 0 0 0
P 0 z^(p\p)
"n \r\r/ (pic?)
\F\Î / 0
" (1 A\\
0 (p\cj) zM(ci\ci) 0
± 0 0 0 zi
and is diagonal in the labels s and _L, yielding z^ = z^ = 0. In the subspace

{P. 0:
where we have used (7.39) and the relations (p\cf) = (p\p) and (c/|c/) =
The eigenvalues z^ to first order, the hydrodynamic modes xp^ to zeroth order,
and the currents j ^ = (c/ -f z^)\p^ are then given respectively by:
z ^) = 0; xpf = c±; h = *xy = W±; (7.43)
with TXX = \{c2x — c\) and a = + ; here s labels the entropy (or heat) mode, _L
the shear (or transverse momentum) mode, and a the two sound modes.
The solution to the second equation of (7.37) is:
(7.44)
The first term on the right hand side belongs to the orthogonal complement
of the null space of H. The coefficients B^ remain undetermined. By inserting
(7.44) into the third equation of (7.37) and multiplying it on the left by (t/^l
(with X = s,_L and <r), we obtain:
0 =
For A = fi9 the first term on the r.h.s. vanishes and we obtain the eigenvalues to
second order:
1
z(2) ^ v=-(T I - + -IT )(c \C r
Q 2
Here D T is the thermal diffusivity, v the kinematic viscosity, and F the sound
damping coefficient. To further simplify the expression for F, we use the relation:
(ip^xpW) = 2{p\p) = 2c2s(ci\ci) = 2c2s(p\p), (7.47)
which follows from (7.39), (7.42) and (7.43). For lattice gas models with sufficient
symmetry (like the triangular lattice with hexagonal symmetry, see Chapter 3),
a fourth rank tensor is isotropic, yielding the equalities:
v(c±\c±) = ~(Txy\- + -\Txy) = -(TXX\- + -\TXX).

+ (7.48)
+ ^ ^ ' 2'
Combination of the above results provides an expression for the sound damping
coefficient in terms of the viscosity and the thermal diffusivity,
Now by writing s = p — p(p\p)/(p\p) and introducing the isothermal speed of

sound CT through the relation:
(p\p) = (Sp/da)ij = (dp_ \ = 2

(P\P) (dp/da),, {dpi =-°
4T', (7.50)
we have:
(s\s) . (P\P)(P\P) = c2s(p\p) cl _

(P\P) (P\P)2 {P\P) c\~y- l /
-MJ
So we can rewrite the sound damping coefficient as:
i i (7.52)
with y > 1, as seen from (7.51).

It remains to evaluate the coefficients B^ which determine the first order
eigenvectors through (7.44). From (7.45) with / z ^ v , and after some rearrange-
ments, we find:
B^fof |Vf>(z« - z«) = - t a | ± + \\j,\ (7.53)

7.4 The hydrodynamic spectrum 163
from which we obtain the non-vanishing coefficients, i.e.
B,,_, = (v-r)/[2(7C 8 ]
BT,O = (y-l)DT/[2acs]. (7.54)

The coefficients B^dx = T, a, _L) remain undetermined, but they are unimportant
as they will not appear in the computation of the dynamic structure factor
S(k,co).
In summary, starting from the linearized lattice Boltzmann equation, we
solved the eigenvalue equation (7.35) to obtain explicit expressions for the
hydrodynamic modes: the shear mode, the heat mode, and the two sound modes.
The eigenvectors contain the transverse (_L) and longitudinal (/) components of
the momentum fluctuation g(c) = c, the entropy fluctuation s(c) = \c2 — c\, and
the pressure fluctuation p(c) = \c2. The eigenvalues contain the speed of sound
cs and the transport coefficients: kinematic viscosity v, thermal diffusivity Dj
and sound damping coefficient F, whose explicit expressions were obtained in
the Boltzmann approximation.
7.4 The hydrodynamic spectrum
In order to carry out a consistent perturbation expansion of F(k, co) in (7.31)

for smallfc,we observe that we are interested in co ~ (9(ka), with a up to 2. The
dominant non-vanishing contribution ((9(k)) yields the line-shifts, and the next
contribution {(9(k2)) yields the line-shape (width) of the spectrum. From (7.31),
using the relation (ex — I)""1 + \— x~ l + @{x), valid for small x, we observe that:
F(k, co) = ^ V ^ [ico - z,(k)} - 1 (1 + (9(k2)), (7.55)
where the asterisk on the summation sign indicates restriction to the hydrody-
namic modes. The coefficients JV\ in (7.32) are evaluated in Section 7.3.2 for
smallfc,and yield:
~ (p\p) fl^l\ + (9(k2). (7.56)

We have also introduced the ratio, Equation (7.51):
± (p\p)(p\p)
y
4 <PIP> 2 ' (
'
with cs and CT the adiabatic and isothermal speed of sound respectively.
Combination of (7.31), (7.55) and (7.56) yields the dynamic structure factor
in the Landau-Placzek approximation:
S(k,co) (l^±\ 2DTk2 ly> Tk2
S(k) ~ \ y ) w + (DTk2)2
2
y^(co±
^ csk)2 + {Tk2)
+ 1 [F + (y - 1)DT] *
y ccs *-f (at
For small fc-values and co ~ (9(k) this expression is consistent up to relative

(9(k2).
The power spectrum (7.58) has the standard form as obtained from the
linearized hydrodynamic equations in continuous theory (Boon and Yip, 1980).
The first line on the r.h.s. of (7.58) contains three symmetrical Lorentzians: the
first one is the Rayleigh peak with a width Djk2 and the other two Doppler
shifted (±csk) lines are the Brillouin peaks with a width Tk2. The last term in
(7.58) describes asymmetric contributions to the Brillouin lines. On the relevant
scale co ~ #(fc), the Brillouin shift of (9(k) is the dominant feature that determines
the structure of *S(k,co). The line-shape of the Brillouin line is determined by
the line-width and the asymmetric terms, which are of relative order (9(k)
with respect to the frequency shift. The power spectrum (7.58) consisting of
the Rayleigh line and the asymmetric Brillouin lines, derived here from the
Boltzmann equation for the lattice gas, are in complete agreement with the
results of the Landau-Placzek theory for continuous isotropic fluids (Boon and
Yip, 1980). All further corrections to the theory are at least of relative order
(9(k2) and would involve higher order transport coefficients, such as Burnett
coefficients. Finally we note that the sum rule for the integrated intensity in the
Landau-Placzek theory follows from the integration:
(7.59)
the same result was obtained for the lattice gas in Section 7.3.I.7
7.5 The eigenvalue spectrum
In order to analyze the dynamic structure factor S(k,co) over the full spectral
range of wavelengths and thereby to assess the limits of validity of the Landau-
Placzek theory, we must know both the eigenvalues, zAt(k) = Rez/z(k) + zlmz/x(k),
and the eigenfunctions of the propagator (7.13). At fixed wavelength k, each
eigenmode contributes, according to (7.31), a spectral line with a maximum
located approximately at Imz^(k) (^ 0 if the mode is propagating) and a width
7
Notice that the small-fc limit and the co-integration cannot be interchanged, because the
co-integral does not converge uniformly near co = 0.
determined by Rez^(k); the eigenfunctions enter the weight factors (p\y>)(4>\p)

in (7.31), and determine predominantly the amplitude of the spectral lines. In
the previous section, we obtained these quantities by a perturbative calculation
in the limit of small values of k. Beyond this limit, analytical calculation is no
longer possible, and one must have recourse to numerical computation to obtain
the eigenvalues of the matrix [e~lkc(ll —12)]. The collision operator 12 should
then be known explicitly, that is the lattice gas model must be specified: we
shall consider the GBL model described in Section 3.6 which, as will be seen,
exhibits all the general features observed in an actual mono-atomic fluid.
Figure 7.1 shows the real and imaginary parts of the eigenvalues. The total
number of eigenvalues equals the number b of allowed velocity states (b = 19
in the GBL model), and their values depend on the details of the collision rules
(see Section 3.6).
(i) There are fast kinetic modes with Rez^(k) =£ 0 at k = 0, which may be
purely diffusive or may become propagating at non-zero /c-values.
(ii) There are slow hydrodynamic modes with a damping Rez^(k) ~ (9(k2) as
fc-»0; these modes may be propagating with a (sound) speed Imz^(k)/fc =
c^(k), or they may be purely diffusive (if Imz^(k) = 0). At smallfe-valuesthe
four hydrodynamic modes can easily be identified as they are well separated
from the kinetic modes.
The eigenvalue spectra of the lattice gas automaton have a number of uni-
versal features in common with those found in the continuous theory of liquids
and gases; one interesting aspect of lattice gases is that there is a finite number
of eigenvalues which can all be computed explicitly (at least numerically). These
features are important in the analysis of the dynamic structure factor and they
are most appropriately interpreted by considering the ratio of the mean free
path, £f ~ 1/p, to the wavelength, A = 2n/k, of the probe used to analyze
the dynamics. In laboratory experiments, A is the characteristic wavelength de-
termined by the geometry of the scattering set-up used in neutron or photon
spectroscopy (Boon and Yip, 1980); in lattice gas automata, the same role is
played by the reciprocal wavenumber used in the analysis of the simulation
data. So here the appropriate quantity is the product fc//.
7.5.1 Hydrodynamic regime: k^f <C 1

In the hydrodynamic regime, that is k <, 0.6 (or A ^ 10 lattice units) in
Figure 7.1, there is a clear separation between the kinetic eigenvalues and the
hydrodynamic ones which correspond to the shear mode (/u =JL), the heat
mode (fi = s), and the two sound modes (JI = a = +). The values z^(k) of
the hydrodynamic modes are given by (7.46) and the macroscopic transport
coefficients are independent of the wavenumber k. In this regime the hypothesis
of the Landau-Placzek theory is satisfied: long wavelength fluctuations decay

essentially according to the linearized macroscopic hydrodynamic equations;
we shall see indeed, in Section 7.6, that the hydrodynamic power spectrum of
the lattice gas is accurately described by Equation (7.58). This is the domain of
validity of classical hydrodynamics which gives a local description of the response
of the system: energy and momentum fluxes are linear in the temperature and
velocity gradients respectively.
7.5.2 Generalized hydrodynamic regime: /c/f < 1

When we look at a smaller scale, that is if we probe the dynamics of the
fluid at a wavelength A which is not much larger than /f, we observe that the
classical hydrodynamic description with constant transport coefficients breaks
down: the non-local response of the system to spatial inhomogeneities renders
Figure 7.1 Eigenvalue

spectrum of the GBL
model: z^k) as function
of k. (a) Real eigenvalue
RezM(/c); the dotted lines
represent the
hydrodynamic relaxation
rates: the upper curve
corresponds to the
kinematic viscosity, and
the lower double curve to
the thermal diffusivity
and to the sound
damping (DT = 0.302,
T = 0.297); (b) Imaginary
eigenvalue Imz^/c); the
slope of the dotted line is
the hydrodynamic speed
of sound (cs = 1.26).
Density and energy per
node: p = 6.0 and e = 6.7
respectively; k is oriented
along a lattice axis and k
is given in 2n x reciprocal
lattice units; Rez^(/c) and
Imz^/c) are frequencies
expressed in units of 2n x
reciprocal automaton
time steps.
2.0
the transport coefficients k-dependent We see indeed, in Figure 7.1, that the
discrepancy between the macroscopic relaxation rates (dashed lines) and the
exact ones (full lines) becomes larger and larger as k increases, and that the
speed of sound cs(k) exhibits dispersion. The exact diffusivities, as we shall see
in Section 7.6, can be described in terms of generalized k-dependent transport
coefficients: D^k) = Rez^(k)//c 2 = (v(k),r(k),D T (k)}.
However, we observe that, in the range 0.6 <, k <, 1.25, there still is a clear
separation between hydrodynamic and kinetic real eigenvalues; so the contri-
butions of the fast modes may be ignored in the analysis of the hydrodynamic
spectrum. This defines the domain of generalized hydrodynamics where the gen-
eral structure of the hydrodynamic spectrum should be preserved, but where
important discrepancies will be observed between the Landau-Placzek theory
and the Boltzmann theory for S(k,co). Closer inspection of Figure 7.1 leads to
the following observation: for k parallel to a basis vector, say lx, the matrix
in (7.31) is invariant under the reflection cy <-• — c y. Consequently the matrix
in (7.30) can be decomposed in two subspaces with even and odd parity in cy.
Therefore S(k,co) in (7.31) only couples to the even subspace containing sound
fa = a = +) and heat (// = s) modes. It does not couple to the shear mode
(fi =_L) with odd Cy-parity. Therefore, when considering S(k,co), the domain of
generalized hydrodynamics extends to about k ~ 1.5, which corresponds to a
wavelength of four lattice units.
In Section 7.6, we shall test these concepts of generalized hydrodynamics by
substituting the k-dependent parameters c s(k), F(k) and Dj(k) (as obtained from
the numerical evaluation of the eigenvalues) into the Landau-Placzek expression
(7.58) and compare the result with the Boltzmann prediction (7.31) which does
not involve any long wavelength or slow mode approximation.
7.5.3 Kinetic regime: /c/f >, 1

For large fc-values (k > 1.5), that is for very short wavelengths (A < 4 lattice
units), we can no longer make a distinction between fast and slow modes, as
all decay rates have the same magnitude (see Figure 7.1). In the kinetic regime,
it is physically meaningless to introduce any parameterization to reconstruct a
hydrodynamic type spectrum, even with k-dependent coefficients: all modes of
even cy -parity couple to the dynamic structure factor and contribute a spectral
line (Doppler shifted if the mode is propagating).
To illustrate the importance of the mean free path, i{ ~ 1/p, as the relevant
length scale parameter for physical interpretation, let us mention that the
eigenvalue spectrum at low density (p ~ 1) as compared to the spectrum of
Figure 7.1 (p = 6.0) has a classical hydrodynamic regime (k <, 0.15 or A >, 42)
about a factor of 6 narrower, and a generalized hydrodynamic range about a
factor of 8 smaller. For k >, 0.4 one observes the so-called Knudsen regime
(Das, Bussemaker and Ernst, 1993), where the collision matrix in (7.31) is only
a small perturbation to the propagation term which constitutes the dominant
contribution to the spectral density (see Equation (7.19)).
7.6 Power spectrum
We shall now examine the spontaneous density fluctuations in the microscopic

simulations of the GBL lattice gas automaton. The lattice gas is at global
equilibrium (i.e. the system is initialized with zero total momentum) and is
subject to periodic boundary conditions. The GBL model has the interesting
feature that its equilibrium thermodynamic state can be specified by the mass
density p = J ^ / i , and the energy density e = J2ieiff> which are set by the
fugacity z = exp(a) and the reduced temperature 6 = exp(—jS/2) introduced in
Chapter 4. We shall consider two states, one of high density (p = 6.0, e = 6.7)
and one of low density (p = 1.1, e = 1.1).
The 'experimental' S(k,a>) is obtained by space- and time-Fourier transforma-
tion of the data of the local density fluctuations <5p(r*, U); in the analysis, S(k, co)
is treated as a spectral function, i.e. as a function of co taken at fixed values
of the wavenumber k = |k|, and spectra will be analyzed for different typical
values of k. We shall compare these 'experimental' spectra with the theoretical
predictions obtained in the previous sections for the dynamic structure factor:
the Boltzmann expression for 5(k,co) in (7.17) and (7.31), and, for small |k|,
the Landau-Placzek expression (7.58) with the thermodynamic quantities y and
cs computed from (7.57) and (7.39), and the Boltzmann transport coefficients
v and Dj, computed from (7.45). So the comparative analysis also provides a
test of validity for the evaluation of the transport coefficients as they can be
obtained from the measurement of the spectral line-widths.
In principle, measurements could be performed down to a value of the
wavenumber equal to fco = 2n/L, where L is the lattice dimension. But we
should note that the accuracy on the line-width measurement is set by the
frequency resolution Aco = 2n/To, where To is the total simulation time (usually
of the order of a few thousand time steps), and so it decreases with the width
of the spectral lines, i.e. when k decreases. Therefore in practice there is a lower
bound to thefc-rangewhere measurements are reliable; for instance in the results
discussed below, this lower limit is set at k = 4fco — 0.05, which corresponds to
a wavelength of 130 lattice units while the lattice has 512 x 512 nodes.
7.6.1 High density

Spectra S(k,co) measured at small wavenumbers are characteristic of the hydro-
dynamic behavior at long wavelengths: a typical example for the lattice gas at
high density is shown in Figure 7.2. According to the discussion in Section 7.5.1
(see also Figure 7.1), this spectrum is measured at a fc-value well inside the
hydrodynamic regime where the predictions of the Boltzmann theory (7.18) and
of the Landau-Placzek theory (7.58) coincide, and are in excellent agreement
with the 'experimental' spectrum as Figure 7.2 shows. The spectrum consists of
a doublet of spectral lines located at co = ±csk, corresponding to the acoustic
modes, and a peak centered at co = 0, which arises from fluctuations at constant
pressure and corresponds to the thermal diffusivity mode (for non-thermal LGA
models, this central peak is absent). Thus we observe that in the small fe-domain
the spectral density of lattice gas fluctuations is very well reproduced by the
Landau-Placzek theory.8 Most important is the observation that the dynamical
structure factor of the lattice gas exhibits the typical structure and line-shapes
of the Rayleigh-Brillouin spectrum measured in real fluids (Boon and Yip,
1980); it was precisely to explain how spontaneous hydrodynamic fluctuations
in continuous fluids give rise to this type of spectrum that the Landau-Placzek
-2.0
Figure 7.2 The LGA dynamic structure factor in the hydrodynamic regime.
S(k,co) as a function of the frequency co for the GBL lattice gas at high
density and low k. Comparison between the simulation results (solid line) and
the theoretical predictions: Landau-Placzek theory, Equation (7.58), (dashed
curve), and Boltzmann theory, Equation (7.18), (dotted curve).
Thermodynamic state: p = 6.0; e = 6.7. Automaton universe: L x L — (512)2.
k = 20/co ^ 0.245 reciprocal lattice units, co is expressed in units of 2n/T0,
where To is the total number of time steps, and S(k,a>)/S(k) is expressed in
reciprocal co units.
The slight discrepancy in the height of the doublet peaks (see Figure 7.2) is probably
due to the weaker resolution at small /c-values of the peaks where fluctuations in the
simulation data are most pronounced.
theory was initially developed. The comparative inspection of Figure 7.2 and
Figure 7.3 is instructive as the latter shows the power spectrum measured by
light scattering spectroscopy in a real fluid at k ~ 2 x 105 cm" 1 , i.e. a wavelength
of a few thousands angstroms.
When the value of the wavenumber increases, the Landau-Placzek theory
progressively fails to reproduce the measured spectra, because, as we discussed
in Section 7.5.2, linearized hydrodynamics with constant transport coefficients
becomes invalid at short wavelengths. For the GBL lattice gas, at k = 5Ofco — 0.6
one enters the regime of generalized hydrodynamics where the Boltzmann theory
gives very good account of the simulation data.
When k increases further, the deviations from classical hydrodynamics become
larger, as exemplified in Figure 7.4, which shows a comparison between the
measured and computed dynamical structure factors at k = 136fco — 1-67 for
the GBL model. The figure shows that the Boltzmann prediction remains valid
down to quite short wavelengths (here A ~ 4 lattice units). Here we are at
the border between generalized hydrodynamics and the kinetic regime (see
Figure 7.1): the Landau-Placzek theory with constant transport coefficients is
expected to fail completely, but when combined with the fe-dependent sound
speed and fc-dependent transport coefficients, as computed from the eigenvalue
spectrum, considerable improvement is obtained, as illustrated in Figure 7.4.
Although it is clear that hydrodynamics becomes invalid in this high k domain,
it is remarkable that the general structure of the hydrodynamic power spectrum
persists down to such short wavelengths. Similar observation can be made from
the spectra measured by neutron scattering spectroscopy in actual fluids, e.g.
liquid neon at k ~ 107 cm" 1 , i.e. a wavelength shorter than 100 angstroms (Bell
et al, 1975).
!- 5 |
1 1 1
' 1I ' I ' I ' I ' iI ' I1 '' I ' | Figure 7.3 The power
spectrum of liquid argon
in its normal liquid range
measured by light
I Q \-- II -I
- scattering spectroscopy
(Fleury and Boon, 1969).
3' [ ^ 22.883
- GHz
883 GHz 1
JjJ 22.883
- 883 GHz
GHz ^ o) is given in GHz and
S{k,co) in arbitrary units.
'}
0.5 -
\H
0.0
-4.0 -3.0 -2.0 -1.0 0.0 1.0 2.0 3.0 4.0
co (GHz)
7.6.2 Low density

The various types of behavior discussed above are also found from the analysis
of the spectra of the lattice gas at low density (p ~ 1), with the restriction
that the domain where the hydrodynamic approximation is valid and even the
generalized hydrodynamic domain are narrower in fc, as can be inferred from
the discussion in Section 7.5. In Figure 7.5, the density fluctuation spectrum
is taken at k = 12fco — 0.15 which is really the border between classical and
generalized hydrodynamics for the low density lattice gas; here the Landau-
Placzek and Boltzmann predictions are in good agreement with the spectrum
obtained from the simulation data. On the other hand, when k = 52fco — 0.64,
the dilute gas is inside the kinetic regime, close to the Knudsen regime (Das,
Bussemaker, and Ernst, 1993); the simulation data become very noisy and it is
difficult to separate spurious oscillations from the actual kinetic spectral lines,
Figure 7.4 The LGA

dynamic structure factor
in the generalized
hydrodynamic regime.
Same conditions and
symbols as for Figure 7.2;
here k = 20/c0 ~ 0.245
reciprocal lattice units.
Figure 7.5 The dynamic

structure factor of the
GBL lattice gas at low
density at the border of
the classical and
generalized hydrodynamic
regimes. Thermodynamic
state: p = 1.1; e = 1.1.
/c = 12/c0 — 0.15 reciprocal
lattice units. Other
specifications as for
Figure 7.2.
-0.4
see Figure 7.6. In this almost collisionless regime, correlation effects through
the periodic boundary conditions might be quite substantial. Nevertheless the
Boltzmann theory yields predictions which follow more or less the general trends
of the spectral features. In fact, from the eigenvalue spectrum and a numerical
analysis of the eigenfunctions, one can see that 5(k,co) is strongly coupled to
six modes with Rez^(k) ~ 2, four of which are propagating, which explains
qualitatively why the 'Brillouin lines' in Figure 7.6 have a complex structure
resulting from the superposition of several modes.
7.6.3 Dispersion effects

The results presented in the previous sections offer the possibility to examine the
question, often discussed in statistical mechanics, of dispersion effects in simple
fluids. The interesting point is that in lattice gas automata these effects can be
analyzed quantitatively. Dispersion appears when a thermodynamic quantity,
like the speed of sound, and transport coefficients, e.g. the thermal conductivity,
deviate from their constant value predicted by classical theory, and become
fc-dependent functions, because at short wavelengths the response of the fluid
to spontaneous fluctuations or forced perturbations becomes non-local (see
Section 7.5.2).9
According to classical hydrodynamics the position of the (Brillouin) sound
peaks - the Doppler shift cos(k) in the power spectrum S(k,a>) - varies linearly
with k since in the hydrodynamic regime Im z±(k) = ±cs/c, where cs is a constant.
In Figure 7.7, the LGA measured Doppler shift cos(k) of the Brillouin lines
4.U 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
Figure 7.6 The dynamic
structure factor of the
GBL lattice gas at low
'.'/// \\\\;;; --
3.0 ;' :A J-i It \ - density in the kinetic
regime. Same conditions
and specifications as for
Figure 7.5 except
^ 2.0 -
k = 52/co — 0.64 reciprocal
\
lattice units.
1.0
-3.0
; J. . .k J -2.0 -1.0 0.0 1.0 2.0 3.0
0)
9
At very high frequencies, the transport functions can also become co-dependent as the
system exhibits viscoelastic behavior when the response becomes non-local in time.
(open circles) is compared with Imz+(/c) = ±cs(k)k (solid line) computed from
the lattice Boltzmann equation; the classical behavior is shown by the dotted
line, the thermodynamic speed of sound cs being calculated from (7.39). The
measured dispersion in the sound speed is seen to be remarkably well accounted
for (up to k ~ 1.5) by the Boltzmann theory for cs(k) confirming the validity of
generalized hydrodynamics.
From the expression for the classical hydrodynamic spectrum (7.58), we can
compute the ratio of the integrated intensity of the central Rayleigh peak to
those of the Brillouin peaks, and obtain the Landau-Placzek ratio, which is
a thermodynamic quantity equal to (y — 1). For the lattice gas, the measured
value yexp — 1-34 is in very good agreement with the theoretical value 1.32,
calculated from (7.51): y = c\/c\. In the regime of generalized hydrodynamics,
the Landau-Placzek ratio then becomes fc-dependent, but it is very hard to mea-
sure quantitatively the dispersion effects on y, and thereby make a comparison
with theoretical predictions, because of the difficulty of evaluating the integrated
intensities of the spectral peaks in a quantitative manner at high /c-values (as it
is in real fluids).
In the long wavelength limit, the value of the transport coefficients can be
obtained from the line-widths of the spectral lines of S(k,co) since the Landau-
Placzek theory predicts three Lorentzians of width DTk2 (for the central Rayleigh
peak) and Tk2 (for the two shifted Brillouin peaks). In Figure 7.8, the simulation
data for Rayleigh width (black dots) and for the Brillouin width (open circles)
are shown as functions of k2. The dashed line represents the classical values
Djk2 and F/c2 and corresponds in fact to two coinciding lines because here
(by mere chance) D T ^ F whose values are given by the Boltzmann transport
coefficients in the limit k —• 0, derived in (7.46). The solid curves D T(k)k2 and
T(k)k2 computed from the Boltzmann eigenvalues (see Section 7.5) indicate that
Figure 7.7 Sound

dispersion in the lattice
gas. Doppler shift cos(k)
of the Brillouin lines
measured from the
dynamic structure factor
o f the G B L lattice
§a s at
high
density (°pen
circles), classical
hydrodynamic value
cs x k (dotted line), and
Boltzmann theory
prediction (solid curve).
Same units as in
Figure 7.5.
the dispersion observed in the generalized hydrodynamic domain (k <, 1.5) is

well accounted for by the lattice Boltzmann theory.10
In conclusion, the spectral analysis of the spontaneous fluctuations in the
lattice gas automaton shows, theoretically as well as 'experimentally', that one
finds in the LGA dynamic structure factor the essential features of the scattering
function measured in real fluids.
7.7 Diffusion and correlations
In a fluid, spontaneous fluctuations also induce particle diffusion, but since

particles diffuse in their own medium, diffusion cannot be observed unless one
can, in one way or another, distinguish particles amongst themselves. This can
be realized in a two-species fluid with say red and blue particles, which ideally
do not differ from each other except by their color.11 In technical terms, this
means that there is an additional quantity which is necessary to characterize
Figure 7.8 Transport coefficients dispersion in the lattice gas. Line-widths

measured from spectra of the GBL lattice gas at high density as a function of
k2: the open circles refer to Brillouin lines (~ T{k)) and the black dots to the
Rayleigh line (~ DT(k)). Comparison with the Boltzmann theory predictions
(solid lines). The dashed straight line corresponds to the classical value
|Rez(fc - • 0)| = 0.3/c 2 (DT ~ 0.302 and T ~ 0.297). Same units as in Figure 7.5.
10
At least for the thermal diffusivity; for the sound attenuation the Boltzmann prediction
deviates from the observed dispersion when k i 0.6, but the measurements of peak
position and line-width are very sensitive to 'experimental' noise in the spectra at large
/c-values.
11
In experimental observations, one uses for instance isotopes where a tracer can be
conveniently identified.
the fluid, the concentration or species density, and which will couple to the other
thermodynamic and mechanical variables.
The FHP model has been generalized to study diffusive phenomena in binary
fluids using 'macroscopic' experiments (Bernardin and Sero-Guillaume, 1990;
McNamara, 1990; Noullez, 1990). Typically the measurement would be per-
formed in a lattice gas composed of red and blue particles (where the color
is a passive property used to distinguish species which otherwise do not differ
from one another) starting from an initial condition with a step function color
density profile. The observer would then measure the evolution of the density
profile, and evaluate the diffusion coefficient by fitting the 'experimental' profile,
obtained after some time, to the solution of the diffusion equation subject to
the appropriate boundary conditions (McNamara, 1990; Noullez, 1990).
Here we shall consider a more microscopic measure along the lines of the
spectral analysis developed in this chapter. But the mixture of two real fluids
exhibits a power spectrum where the central peak is not a simple Lorentzian,
even in the long wavelength limit (Boon and Yip, 1980): it has a spectral structure
where it is difficult to separate the contributions from entropy fluctuations and
from concentration fluctuations which, in general, are not decoupled. 12 In this
respect, the lattice gas offers an interesting approach to probe diffusion in a
simple fluid mixture, because we can consider a non-thermal LGA and define an
observable - the weighted difference of the species densities - whose fluctuation
correlations yield the diffusive mode independently of the other modes, so that
the corresponding power spectrum provides a measure of diffusion dynamics
solely (Hanon and Boon, 1997).
7.7.1 The two-species lattice gas

The two-species lattice gas is the CFHP model described in Section 3.5: the
particles are tagged as either 'red' or 'blue', and the proportion of red versus blue
particles can be tuned to any preset value of concentration. Color is conserved
by the dynamics, and, during the collision phase, is redistributed randomly
amongst the channels, independently of the mass redistribution.
The technical analysis of the 'colored' lattice gas follows essentially the lines
of the development presented in the previous sections for the single component
lattice gas. The major distinction is the introduction of a 'colored' density, as
we will be interested in the dynamics of the mass density fluctuations of one of
the species to probe diffusive transport in the gas mixture. So we consider (i)
the red mass density pred(r*, £•), which is the number of red particles at node r*
at time t*, and (ii) the corresponding fluctuations (5pred(r*,£*), defined in terms
of the red channel occupations n,-(i e { 1 , . . . , ^ } , with b the total number of
12
Unless one of the two components is in trace amounts, in which case the two modes can
be identified as they produce two independent central Lorentzians.
channels per node; b = 14 in the CFHP model):

b/2 b/2
i(u9U) = ^[ Wl -(r*,t*)-<Hi(r*,t*)>], (7.60)
i=l
where (ftj(r*, £•)) is given by:

rf
/ \_ f f %r for i = 1, ... , ife (red channels),
U J l }
' ' | d (1 - xr) for i = \b + 1, ... , b (blue channels), '
with d the average density per channel, and Xr the concentration of red particles.
Note that:
b
1 1
fo M(1 Zr) =
^ + ^rCd + ^1UC = p (7 62)
'
defines the respective average densities per node.

The 'red mass' dynamic structure factor Sred(k, co), defined as the space- and
time-Fourier transform of the van Hove correlation function:
Gred(r*,£*) = ((5p r e d (r.,^)^ r e d (0,0)) (7.63)
is given by:
oo
p red S red (k,co) = Y, J2 e~i(ur~'k'rGred(r*, |M) (7-64)

t,&se t=-oo
= ±f2e-"°t(5(ri(k,\t.\)5p™ i'(k0)), (7.65)

t=—oo
where <Spred(k, U) is the spatial Fourier transform of (5pred(r*,£*), and V = Jf

is the total number of nodes.13 As for the single component lattice gas, the
dynamic structure factor is expressed in terms of the kinetic propagator defined
by:
r y (k, t)K; = (dni(k,U)dnj(k,O)), (7-66)
where (5n,-(k, £*) is the spatial Fourier transform of 5n,(r*, t*) and Kj = /j e q '(l —
/j e q ) ). Using (7.66), we rewrite (7.64) as:
oo b/2 b/2
r y (M,)> 9 , (7.67)
and the static structure factor (the Fourier transform of the equal-time van Hove
13
Remember that V is also interpreted as the volume of the lattice i f universe, and here
the lattice is finite and has periodic boundary conditions.
function; see Section 4.7) as:

b/2 b/2
r i7 (k,0)K ; - (7.68)
b/2 b/2 b/2
or
S red (k) = 1-dxr- (7.70)
The kinetic propagator is then evaluated in the Boltzmann approximation

following the same lines as in Section 7.2, and we obtain:
p red S red (k,co) = 2ReF r e d (k,o;), (7.71)

b/2 b/2
Fred(k,(o) = Y^Yll ico+ik-c $—n + K

J' (7 72)
'
This expression for the dynamic structure factor is similar to the result given
in (7.17) and (7.18), and is exact within the Boltzmann approximation, but
again the explicit analytical inversion of the b x b matrix in (7.72) cannot
be performed in all generality. However, perturbation methods can be used
to compute analytically S red(k,co) in the hydrodynamic limit: |k| — • 0 and
co ~ $(|k|), $(|k| 2 ); beyond the long-wavelength, long-time domain, one has
recourse to numerical evaluation of (7.72) to compute the Boltzmann power
spectrum. Since the developments proceed essentially as for the single component
lattice gas, we shall not present the detailed analysis 14 and will focus on the
main results.
7.7.2 The hydrodynamic limit

In the large wavelength limit, one can identify four hydrodynamic modes in
the two-species (non-thermal) lattice gas: the shear mode and the acoustic
modes which are independent of color related properties, and one mode which
describes color diffusion only. To first order in k, the non-zero imaginary part of
the eigenvalues yields the frequency shift of the spectral peaks corresponding to
the propagating (acoustic) modes, and to second order in /c, the real parts of the
eigenvalues yield the dissipation coefficients which determine the line-widths of
the spectral components. These properties are reflected in the dynamic structure
14
The major technical difference is that, instead of the thermal scalar product (7.27), here
one has the colored scalar product (A\B) = Y2i=i ^( c») Ki ^(c0> where the weight K\
depends on density and concentration. We refer the reader to the original literature
(Hanon and Boon, 1997) for explicit computations.
factor of the 'colored' lattice gas which reads (Hanon and Boon, 1997) :15
2 KT + Kb co2 + (Dk2)2
b K2 ^ / r/c 2 r/c cs/c ± i
2 * r + Kb ^ V (« ± c*k) 2 + (r/c 2 ) 2 cs (co + csk)2
(7.73)
Here ?cr (red) and Kb (blue) are given by:
Kr = dXr(l -dXr), Kb = d(l- Xr )[l - d(l - &)] , (7.74)
cs is the speed of sound (here cs = y 3/7), F is the sound damping coefficient
(see (7.52), but without the second term since the colored LGA is non-thermal),
and D is the diffusion coefficient. We recognize in (7.73), the typical structure
of the hydrodynamic structure factor (compare with (7.58)): at fixed value of
k, the spectrum consists of a Brillouin-doublet centered around +/ccs, and of
a central peak characterizing color diffusion. It is straightforward to verify
that for Xr = 1, (7.73) reduces to the usual Landau-Placzek expression of the
hydrodynamic spectrum for the non-thermal single-component fluid.
Now if we consider the fluctuations of the observable defined by:
pdlff = /c b p r e d -/c r P b l u e , (7.75)

its corresponding spectral density can be computed straightforwardly along the
lines of the evaluation of the power spectrum S red(k,co), and we obtain the
interesting result that the spectrum is then a single Lorentzian characterizing
color diffusion alone:
, a,) = -2 j ^ ^ (o2 + {Dk2)2 • (7-76)
Notice that purely diffusive behavior of color is related to an observable which

is neither the concentration of one of the components nor the straight difference
between the concentrations of the two species.
7.7.3 The power spectrum

The results obtained in the hydrodynamic limit are in accordance with the
Landau-Placzek theory for continuous fluids (Boon and Yip, 1980). Now the
hydrodynamic theory breaks down at short wavelengths, but the Boltzmann
theory should remain valid down to k values where the purely kinetic domain
starts. In order to characterize the wavelength domain of the various regimes,
we consider the quantitiesfc*ff,where ^f(~ 1/p) is the mean free path, and
15
As for the single component lattice gas, the shear mode (or transverse momentum
mode) decouples from the density fluctuations modes (longitudinal modes), and so does
not show up in the dynamic structure factor.
d = p/pmax, the reduced density (which is also the average density per channel).
Accordingly the hydrodynamic regime is defined by fc/f <C 1, the generalized
hydrodynamic regime (Boltzmann regime) by k/f < 1, and the kinetic regime
by faff > 1.
We shall discuss the power spectra obtained from automaton simulation
data and will compare the results to the analytical Landau-Placzek expressions
and to the predictions of the lattice Boltzmann theory. For the latter, one
uses the eigenvalue spectrum of the propagator T which can be evaluated
numerically over the complete fc-domain, so extending the computation of
the power spectrum to the region of k values where analytical evaluation
can no longer be performed. Figure 7.9 shows a typical eigenvalue spectrum
as computed numerically, where the 14 modes of the CFHP model can be
distinguished.
We observe that in the range k < 0.4, corresponding to wavelengths X > 15/o
(with A) the lattice unit length), the four slow modes are well separated from
Figure 7.9 Eigenvalue

spectrum of the CFHP
model propagator: real
parts (a) and imaginary
parts (b). Full lines are
the result of Boltzmann
computation; the
hydrodynamic limit is
shown by dotted lines.
The reduced density and
concentration are
d = 0.15 and Xr = 30%
respectively. The
wavenumber k = |k| is
given in reciprocal lattice
units x 2n.
2.0
the kinetic modes, and their behavior is correctly given by the hydrodynamic
expressions. We therefore expect that in this domain, the Landau-Placzek theory
should provide a correct description of the dynamic structure factor. This
is indeed confirmed by the comparison between the results obtained from
simulation data and theoretical predictions as shown in Figure 7.10, where
we also notice that the hydrodynamic spectrum is indistinguishable from the
Boltzmann spectrum.
From the hydrodynamic spectrum one can directly measure the diffusion
coefficient D(d,xr); the expression (7.76) of Sdiff(k,co) is a single Lorentzian
with a half-width Aco = Dk2, so that plotting Aco as a function of fe2, as in
Figure 7.1 l(a), a linear least-squares fit yields a slope whose value gives a direct
measure of the diffusion coefficient. This measure is in agreement with the
predictions of the lattice Boltzmann theory up to reduced densities d ~ 0.25
as Figure 7.11(b) illustrates. Deviations for larger values of the density are not
Figure 7.10 Power

spectra of 'red' density
fluctuations (a) and of
Pdiff fluctuations (b), at
low density and small k.
Comparison of
experimental data (full
lines) with theoretical
predictions: the
Boltzmann results and the
Landau-Placzek spectra
coincide (dashed lines).
Density d — 0.15;
concentration Xr = 30%;
0.10 wavenumber |k| = 0.098
reciprocal lattice units; co
is given in reciprocal time
units (2n/T9 where T is
total number of
time-steps); the spectral
functions are given in
reciprocal co units.
-0.10 -0.05 0.00 0.05 0.10

surprising since the Boltzmann theory does not take into account correlations
which are known to become important at moderate and high densities.
As k increases from 0.4 to 1.4, there is still a distinct scale separation between
slow and fast modes (see Figure 7.9(a)), but the eigenvalues of the slow modes
depart significantly from the hydrodynamic prediction, indicating the breakdown
of the local response hypothesis: the transport coefficients become /c-dependent.
As a result, the hydrodynamic theory no longer describes the power spectrum
correctly,16 but the complete Boltzmann spectrum is in rather good agreement
with the simulation results, as Figure 7.12(a) shows. At higher densities there
is a discrepancy between the Boltzmann spectral density and the experimental
power spectrum (see Figure 7.12(b)) which reflects the failure of the Boltzmann
Figure 7.11 Microscopic

measure of the diffusion
coefficient: (a) Spectral
line-width as a function
of/c 2 : A(o = D(d,Xr)k\ for
d = 0.3 and Xr = 0.3
(squares) and %x = 0.5
(black dots); least-squares
fit (full line), (b)
Measured values of
D(d,Xr = 0.3) (black dots)
and theoretical prediction
from Boltzmann theory
(full line).
0.5 0.6
3.0
2.5
2 0
1.5
1.0
0.5
0.0
0.0 0.2 0.4 0.6 0.8 1.0
d
16
Note that even at rather short wavelength (A ~ lOô) the experimental data can still be
reasonably fit with a Landau-Placzek spectral function if the transport coefficients are
parameterized, indicating that a hydrodynamic type description holds qualitatively down
to quite short wavelengths, provided the transport coefficients are renormalized.
theory to evaluate correctly the transport coefficients in the high density region
where correlations are important.
With this discussion we close our investigation of the hydrodynamic fluctu-
ations in the lattice gas (with one and two species). We shall return briefly to
the two-component case in Chapter 10 when we discuss two-fluid simulations.
But we now turn to the macroscopic dynamics of the LGA; after showing in
this chapter that the lattice gas exhibits spontaneous fluctuations which are
compatible with those observed in real fluids, we go back to the important
question: are the equations describing the macroscopic behavior of the lattice
gas compatible with the Navier-Stokes equations of classical fluid dynamics?
This question was addressed in Chapter 5 and will now be revisited with a
different complementary approach.
Figure 7.12 (a) Power

spectrum at low density
d = 0.15, and high/c:
|k| = 0.49, i.e. X « 12/ 0 ,
for Xv = 30%. (b) Power
spectrum at high density
d = 0.5 and low k:
|k| = 0.098, i.e. A « 60/ 0 ,
for Xr = 30%.
Experimental data (full
line), Boltzmann spectrum
(dashed line),
Landau-Placzek theory
(dotted line).
0.05
Chapter 8
Macrodynamics: projectors approach
It is one of the main objectives of statistical mechanics to provide a microscopic

content to the phenomenologically established macroscopic properties and be-
havior of systems with many degrees of freedom. Although it is not necessary
to have a complete knowledge of the details of the microscopic interactions to
describe macroscopic phenomena in fluid systems, these phenomena emerge as a
consequence of the basic dynamical processes. However, to establish rigorously
the connection between the phenomenology and the underlying microscopic
processes amounts to solving the many-body problem. Even for systems with
oversimplified microscopic dynamics, such as lattice gas automata, this is an
impossible task: approximations are unavoidable.
In this chapter we derive the equations governing the macroscopic dynamics
of LGAs satisfying the semi-detailed balance condition; we shall start from
the microscopic dynamics of the automaton, and use the lattice Boltzmann
approximation (Suarez and Boon, 1997a,b).The main objective is to obtain the
non-linear hydrodynamic equations, where the Euler and dissipative contribu-
tions are expressed in terms of the microscopic evolution rules of the automaton,
and whose validity is not restricted to regions close to equilibrium,1 so that they
1
Most of the derivations of the LGA hydrodynamic equations given in the literature
(Frisch et al, 1986, 1987; Ernst, 1990a; Rothman and Zaleski, 1994) include
non-linearities in the Euler part of the equations, but the dissipative term is calculated
by linearizing around global equilibrium, thus limiting the validity of the results to a
regime in which linear response and simple fluctuation-dissipation relations hold. In the
development given by Wolfram (1986), the first few non-linear corrections are included,
but this derivation remains at a phenomenological level, as the non-linear coefficients
are assumed to be correctly obtained by direct Taylor expansion.
183
184 8 Macrodynamics: projectors approach
can be used to analyze phenomena taking place in systems arbitrarily far from
equilibrium, for instance in thermal LGAs under large temperature gradient.
In order to derive the hydrodynamic equations, we make use of the Boltzmann
hypothesis (see Section 4.4.2) that particles entering a collision are uncorrelated.
The average outcome of the collision is assumed to be a non-equilibrium
distribution which is close to the local equilibrium one. A local equilibrium
distribution has the same analytical form as an equilibrium one, but with space-
and time-dependent thermodynamic potentials (temperature, chemical potential,
etc.). The non-equilibrium configuration can then be split into a local equilibrium
component, which depends only on the local values of the set of densities of
conserved quantities of the system, and a non-local equilibrium contribution, a
'projection' onto the space orthogonal to the subspace spanned by the constants
of motion. This procedure is intimately related to the projection operator
methods developed to derive the hydrodynamic equations from the microscopic
dynamics in continuous fluids (Zwanzig, 1960; Mori, 1960; Procaccia, Ronis
and Oppenheim, 1979).2 We shall effectively use a simple projector that neglects
the coupling of the linear hydrodynamic variables (mass, momentum, energy
densities) to bilinear and higher order modes. With the treatment used here -
and perhaps at the cost of a somewhat formal analysis - we shall obtain a
general formulation of the lattice gas hydrodynamics valid for non-equilibrium
states arbitrarily far from global homogeneous equilibrium. We shall also make
contact with the linearized equations and show that the dissipative coefficients
are given in terms of time-correlation functions in the form of the Green-
Kubo expression of the transport coefficients. The present analysis is within the
logical continuation of Chapters 6 and 7; in Chapter 5 we give an alternative,
less formal, derivation of the lattice gas hydrodynamic equations using the
Chapman-Enskog method. Notice that, whatever the method - be it for discrete
systems or continuous fluids - there is always a common, unavoidable major
step in the analysis: the multiple scales hypothesis, by which one anticipates that
there is a natural space- and time-separation between hydrodynamic phenomena
and microscopic processes.
8.1 Preliminaries
Before proceeding to the derivation of the hydrodynamic equations, it is useful

to give some definitions and summarize some material from earlier chapters.3
2
For lattice gas automata, projection operator techniques have been employed to derive
the linearized hydrodynamic equations (see Chapter 6) which describe the time evolution
of small perturbations around a global equilibrium state (Zanetti, 1989; Ernst and
Dufty, 1990; Ernst, 1990).
3
When, for the sake of clarity in the course of the development, results are given without
derivation, the reader is explicitly referred to the appendix for algebraic details.
The observables in the automaton are defined as averages over the stochastic
dynamics and over an ensemble of initial conditions. The averaging is denoted by
angular brackets, (.. .) N E , where the subindex stresses the fact that the ensemble
is a non-equilibrium one. The collection of local quantities preserved under
collision are labeled by the Greek index cp running from cp = 1 to cp = iVhydro,
where A/hydro denotes the number of hydrodynamic variables. This set should
contain the quantities conserved in the dynamics of the physical system modeled
by the automaton. The index cp = 1 corresponds to the number of particles,
which is conserved by the collisions.
Before the collision step, the configuration of a node r* e <£ is given by :4
{n(T,t)} E= {m(T9t)}ll9 (8.1)
and after collision by:
{n'(r,t)} = K(M)}ti. (8-2)

The average of these quantities over the stochastic process and the non-
equilibrium ensemble is given by:
fi(i,t) = (^(r,0) N E , (8.3)
and:
ffat) = (n;(r,0)NE, (8.4)
respectively. The hydrodynamic variables, that is the densities of conserved

quantities, are collected in a vector:
b b
A(i,t) = {^(r,0}Jr = E "5k') 4 = E "*(r>')4r (8.5)

1=1 i=\
The last equality follows from the fact that the value of the hydrodynamic
variables is preserved by the collision step. The hydrodynamic variables include
some or all of the following: particle number, momentum, and, in thermal
automata, kinetic energy per node. For example, in non-thermal automata whose
collision rules preserve the number of particles and the momentum per node,
the hydrodynamic variables are the number density, p(r, t), and the momentum
density, pu(r, t), which can be collected in a two-dimensional vector:
ni(r t)Ai = ni{r t]

4
' t
For simplicity, we shall omit the subscript * in the notation of the discrete variables r*
' t
and U.
with the vector of collisional invariants:5
-' = ( c j • (8J)
The average of the hydrodynamic variables is:
The microscopic prescription of the collision step reads:
where £{s}{s'} is equal to 1 when, starting from a configuration {s}, the collision
produces configuration {s'}, and is 0 otherwise. For stochastic automata, ^ is a
random Boolean variable. In this case the average post-collisional configuration,
starting from a given pre-collisional configuration {ft(r, £)} is:
JiA(s^)5({n(T9t)},{s}), (8.10)
where A(s^sf) is the matrix whose entries are the probabilities of having a
configuration {s'} as the outcome of a collision starting from a configuration
{s}. The quantity S ({n}, {5}) is equal to one if the configurations {s} and
{ft} are equal, and zero if they are different. A representation of this Kronecker
delta in configuration space is (see Section 2.2.4):
nskk(l-nkf-^. (8.11)
k=l
With the semi-detailed balance condition (see Section 2.3.1):
an arbitrary distribution of the Fermi-Dirac form for the single particle distri-
bution (see Section 4.4):
ft = T - \. Al , (8-13)
1 + e x p { 0 * A t\
without correlations between fluctuations in different channels, is invariant under
the collision step. The quantities (j) are the thermodynamic fields conjugate to
5
We ignore effects associated with the presence of spurious global invariants (see Brito,
Ernst and Kirkpatrick, 1991).
the conserved quantities. The scalar product denoted by * is a contraction of

the indices labeling the conserved quantities:
iVhydro
<p=l
Furthermore, the H-theorem guarantees the existence of a stable global equi-

librium state characterized by a probability distribution which is a product of
Fermi-Dirac distributions for every channel on every node of the lattice (see
Chapter 4).
We define a local equilibrium ensemble, for which the average occupation of
channel i is:
•exp{0(r •,t)*A,} '

such that:
(4(r,0)LE = 7^ fjLE(l
1
A
-'
= <A(r,t))m. (8.16)
Equation (8.16) defines the thermodynamic potentials implicitly in terms of the
local values of the densities of conserved quantities: (j)(r,t) = <>
/ ((A(r, t))). For
conserved quantities, since the local equilibrium ensemble has been defined in
such a way that (A(r, t))LE = {A(r9 f))NE, we may omit the label indicating the
ensemble over which the average is taken.
The lattice Boltzmann equation (see Section 4.2.2):
t)Yk (l-A(r,0) ( 1 - S f c ) , (8.17)

{s}{s>} k=l
constitutes the starting point for the derivation of the macroscopic equations that
describe the evolution of the hydrodynamic fields. Multiplying Equation (8.17)
by At, summing over the index i and using the displacement operator defined
by:
/,(r + c , t) = exp {c; • ^ } /,(r, t), (8.18)
where the scalar product denoted by • is a contraction of the spatial indices, we
obtain:
b
ex
U(r, r + 1)> = Yl P {~c< • Vf} ffa ^ ^f (8.19)
1=1
Making use of the equality:

b
(A(T, t)) = ^2 //(r» 0 4j» (8.20)
we can rewrite (8.19) as:

b
(A(r,t+1)) - (A(r,t)} = ^ (exp{-c r V r } - l) f[{r,t) At. (8.21)
Equation (8.21) has a suitable form for the subsequent development.
8.2 Multiple scales analysis
The microscopic evolution equations (8.21) are finite difference equations con-
taining the full dynamics of the automaton able to describe its behavior at all
scales. The microscopic scales are given by the automaton lattice spacing and, for
time, by the lapse between two consecutive propagation steps. There is another
characteristic microscopic time TLE, of the order of a few collisions, which corre-
sponds to the relaxation to local equilibrium. In order to derive hydrodynamic-
type equations from the evolution equations (8.21), we introduce the parameter
e, which represents the ratio between characteristic microscopic and hydrody-
namic length scales. This ratio is assumed to be small, a fact which allows us to
carry out a multiple scale analysis (see also Chapter 5). Thus, we anticipate the
scales of physical interest in the solution of (8.21), and introduce the following
variables: a space variable ri = er, and two time variables, t\ = e t, which is
of order e° in theregime where Euler equations are valid, and ti = e2t, which
is of order e° in thedissipative regime. We further assume that deviations from
local equilibrium at each node are small. This corresponds to the Chapman-
Enskog picture of a non-equilibrium state (Chapman and Cowling, 1970), where
the deviations from local equilibrium are on the order of the gradients of the
conserved quantities, i.e. of order e. The substitutions (introduced in Chapter 5):
dt - edtl + e2dt2, (8.22)
in Equation (8.21) lead to the hierarchy of equations labeled by the different

powers of e:
e1 : dh(A(r,t)) + Vri • (J(r, f))LE = 0, (8.23)
= 2 ^ 1 :(M(M))LE)
-c f • Vr } - 1) At SfKr, t\ (8.24)
i=l
8.2 Multiple scales analysis 189
with:
6f't(r,t) = f[{x,t) - f?(t,t), (8.25)

and the definition of the currents associated to the conserved fields:
It = CiAi, (8.26)
b b
<J(M))LE = E AO"*') = E aAifFfrtl (8.27)

i=i i=i
and:
b b
(c J(M))LE = E c' i ffW = E ^ A / " ^ O - (8-28)

i=l i=l
Equations (8.23) are the Euler equations describing the streaming part of the
dynamics, and Equations (8.24) contain the dissipative contributions.
We note that we can rewrite the second term on the l.h.s. of Equation (8.24)
by making use of the first order equation (8.23):
= v ri \ \ • <I(r, t))LE) * (A\A)-J (r, t) * (A\J)hE (r, t)] , (8.29)
with the definitions:
J2 ^(r,t) B^ Bf\ (8.30)

1=1
where (see Section 4.7):
K>,0 = /J*(r, 0 ( 1 - / ^ , 0 ) , (8.31)
and:
[(A\A)LE(r,t)}-1 * (A\A)LE(r,t) = (A\A)LE(r,t) * [(4|4> L E (M)]"' = i-

(8.32)
The scalar product defined in (8.30) is in fact a cumulant average (see Ap-
pendix A.11):
, t) _ ( B ( D ( r , t ) ) L E ( B (2) (r> t ) ) L E . (8.33)

The last equality in (8.29) is a consequence of the relation:
a/?, .u (I(r, 0)LE = (4I4)" 1 (r, t) * (A\J)LE (r, *), (8.34)
which is proved in Appendix A. 11.
8.3 The hydrodynamic equations

Our starting point is the lattice Boltzmann equation (8.17), linearized around
local equilibrium. The deviations of the average occupation of channel i at node
r* e S£ at time t, from its local equilibrium value are:
Sfiirj) = /;(r,t)-/ L E (r,r), (8.35)

Sffat) = /;(r,f)-O-,t), (8.36)
before and after collision, respectively. We carry out an expansion of Equa-
tion (8.17) to obtain:
^ jSf^(r, 0 (5/7(r, 0 + <%2), (8.37)
7
with:
J ] (/r(r,0) S t (W fc LE (M)) ( *'• (8.38)

/c=l
We have assumed that 5fu and consequently Sf'i9 is small (of order e); this
assumption will be discussed below.
The matrix ifLE(r, t) has the form (see Appendix A.12):
JÊ(r, 0 = *f (r, 04- * (4I4)" 1 (r, t) * ^ + JÊ(r, t). (8.39)

The matrix ^ LE (r, t) has the same set of eigenvectors as ^fLE(r, r). The eigenvalues
of i?LE(r, r) corresponding to the set of collisional invariants are zero. The
remaining eigenvalues (i.e. the kinetic eigenvalues) coincide with those of J2?LE(r, t)
which have an absolute value strictly smaller than one. We also notice that the
first term of the r.h.s. of (8.39) has the form of a projection operator onto the
set of constants of motion (see Appendix A.13).
Given that:
b
^ AjSfj(f9t) = 0, (8.40)
7=1
8.3 The hydrodynamic equations 191
we can replace ^LE(r,t) by J?LE(M) in Equation (8.37):

Sffat) = J2 &y(f,t) 5fj(T,t), (8.41)
j
and perform the following manipulations:

b
Sffct) = ^^ E ( r '0[/;(r,0-/ 7 L E (r,0]

7=1
b
= £ &%(*> t) [f'ft - c;, * - 1) - /}E(r, 0]

7=1
(r, 0 [5/;(r - c,-, f - 1) + ff(f - c j 9 1 - \ ) - /} E (r, t
[
^ r , r) [(5/;(r - c;-, t-l)
7=1
(8.42)
where e~dt is the time-displacement operator. We rewrite Equation (8.42) as:
b b
E
Sffat) = Yl îf(rj) f) (r,t) + ^ ^ ( r , t ) 3 / ; ( r , t - l ) , (8.43)
7=1 7=1
where the matrix operators ^~LE and y L E are given by the expressions:
<n/(M) = ^ / ( r , 0 exp{-c r V r }, (8.44)
5^(r,t) = ^ ( r , t ) (exp{-c r V r }exp(-^) - l ) . (8.45)
Equation (8.43) is iterated to yield:
Sf'i(T,t) =
E E ( l l ^ LE ( r ^-^)l ^5f(r,t-T)/P(r,t-T)
y,/ T=0 I T'=0 J fy
b ( t-\ \
+ E i H ^LE(r>x- A W'°)' ( 846)

7=1 I ^=0 Jy
where ^f| is an anti-chronologically time-ordered product, with the convention:
( <r L E (M-T)l = dtj. (8.47)
The requirement that 5/-(r, t) be of order e, and the vanishing of the 'random
force' term depending on the initial conditions, <5/j(r, 0), are fulfilled by choosing
an initial local-equilibrium ensemble in which the values of the thermodynamic
variables vary significantly only over a hydrodynamic length scale. In the case
that this were not so, there should be an initial transient regime where the
hydrodynamic equations are not valid. The duration of this regime is on the
order of TLE, the time necessary to relax to local equilibrium.
We notice that:
T>TLE, (8.48)
which is a consequence of the fact that all the eigenvalues of J^ E (r, t) are strictly
smaller than one in absolute value. So, given that the kernel in Equation (8.46)
is non-zero only on a microscopic time scale TLE ~ (9(e°), and that during this
time the local equilibrium distribution does not change significantly (since it
depends only on slow variables, whose variation is on a time scale (9{e~1)),
Equation (8.46) can be approximated by:
t b
dffat) = - Y^ J2 {[^LEMT}.. (c; • v + dt) /}E(r'O

1J
T=1 y=l
+ (9(e2TLE). (8.49)
Within the same level of approximation we can further replace the upper limit
in the summation over T by infinity. Hence, the last term in (8.24) becomes:
^(exp{-c,-V r }-l) 5 f fat) A,

^
4 £ {[ dh) ff(t,t)
1
[
oo
£ c-4- £{[
ij TT=1
=1
(dji - Kf(r, t) Aj * (4I4L1 (r, t) * 4/) c, • \ ffE(r, t)

(8.50)
where we have used Equation (8.23) and the relation:
3
ffE(r,t) = Knr,t)A^ (A\A)-X {r,t), (8.51)
d(A(r,t))
proved in Appendix A.12 (see Equation (A.58) with n = 1). Notice that the
quantity on the r.h.s. of (8.50):
- = (<$,7 - Kf(r, t) Aj * (A\A)-El (r, t) * ^ / ) (8.52)
is a projection operator onto the set orthogonal to the constants of motion, as

discussed in Appendix A. 13.
Recombining Equations (8.23) and (8.24), using (8.22) and with the results
(8.29) and (8.50), we rewrite the equations of motion to order e2 in terms of the
variables r, t, which are now taken as continuous variables:
dt(A(r,t)) + V r -(J(r,0))LEL
(8.53)
The prime in the summation indicates that the first term of the sum (T = 0) is
multiplied by 1/2. This factor appears as a consequence of the discreteness of
time in the automaton. We have also replaced the matrix J? LE(r, t) by J^LE(r, £),
since the quantity upon which it acts contains no projections onto the conserved
quantities.
Equation (8.53) is the most general form6 of the hydrodynamic equations in
the lattice Boltzmann approximation; the expression of the dissipative term (on
the r.h.s. of Equation (8.53)) is valid even far from global equilibrium. So these
equations are the fundamental equations describing the macroscopic dynamics
of a lattice gas automaton in the hydrodynamic regime; in particular, they can
be used to explore non-equilibrium phenomena such as far from equilibrium
steady states and hydrodynamic instabilities in lattice gas automata.
8.4 Linear response and Green-Kubo coefficients
Now we shall show that linearization of Equation (8.53) yields the usual linear
response results, including the Green-Kubo formulae for the transport coeffi-
cients. We first recover the equations of linear hydrodynamics (see Chapters 6
and 7) by carrying out an expansion of (8.53) around the global equilibrium
state characterized by the values of the hydrodynamic variables:
{A){eq) = J2 At /}"•>. (8.54)

i=l
6
Notice one restriction: the validity of Equation (8.53) is limited by the fact that we have
neglected the coupling of the evolution of the single-particle distribution to ^-particle
distributions.
The deviation from equilibrium is given by:
(SA(r9t)) = (A(rj)) - (4) (eq) . (8.55)
Making use of Equation (A.58) (see Section A.ll), we expand the single particle
distribution ffE(r, t) around equilibrium:
/,LE(r,t) = /}«•> + KÂ, * {A\A)^ q) * (SA(t,t)) + • • • , (8.56)
and we use this expression to linearize the hydrodynamic equations (8.53). In

the Euler term we have:
(J(r,0) LE « (l\A){Qq)* (m)-1^* (3A(r,t)), (8.57)
and in the dissipative term:
]T) I { [i?LE(r, t)]'}.. (Sji - Kf(r, t)Aj * ( 4 I . C 1 (r, t) * 4,)
xcrVr/PM)
x c r V r (Kj
^ ) ^ 4(r, 0)
where the subtracted current is defined as (see also Section 6.3.2):
lj = lj - Aj * (A\A)^ * U|J) ( e q ) = (1J)J . (8.58)
In the second equality of Equation (8.58), we have used the definition of the
equilibrium projectors (Zanetti, 1989; Ernst, 1990a; Ernst and Dufty, 1990)
(^ + 1 = 1) acting on an arbitrary vector {Bj}bj=1:
(8.59)
j j j t o > { d = Bj. (8.60)
The projected quantities have the following properties (see Appendix A. 13):
(8.61)
(A\2B)(eq) = ( 4 | B ) ( e q ) = 0. (8.62)
In particular, for the subtracted current:
A | j) = (A | 1 J) (eq) = 0, (8.63)
which means that J has no projections onto the constants of motion.

Using the definition:
b
7=1
we can write:
£ i {M\- êq) *J = (q)
where the replacement of J, by J, is possible given that J ; has no projection

onto the set of constants of motion.
Using these results in (8.53), we obtain the linearized equations:
dt (5A(r, t)) + \f • [<J|4)(eq) * (A\A)^ * (3A(r,

_ {SA(T, t))] , (8.66)
where the matrix of linear transport coefficients is given by the expression:
A(eq) = E'(*iHJ* ul4>s>-

|_T=0
(8 67)
-
If the decay of the correlation function that appears on the r.h.s. in (8.67)
were slow on the microscopic time scale, one would be able to use the Euler-
Maclaurin sum formula (Bender and Orszag, 1978) to replace the summation
over time by a time integral. However, for lattice gas automata the decay of
J( T )) takes place in times of the order of a few automaton time steps,
/ (eq)
and the discrete sum must be retained.
Equation (8.67) shows that A(eq) is proportional to the sum over time (an
integral in the continuous time limit) of the equilibrium time correlation function
(see details in Appendix A. 14):
T=0 ^ 4J
= £' T=0
= £'
T=0 rr'
(8.68)
where V = Jf is the number of nodes of the lattice, and the superscript

(B) indicates that the time-correlation function is evaluated in the Boltzmann
approximation. The transport coefficients are then well-defined thanks to the

fact that:
{SniiTrfSnfaO))^ (8.69)
decays to zero sufficiently fast for the sum in (8.68) to be finite (Schmitz and
Dufty, 1990). The result is that the dissipative coefficients exhibit a Green-Kubo
form (Rivet, 1987b; Ernst, 1990b) analogous to the continuous theory expression
(Kubo, 1958; Zwanzig, 1965).
This concludes the derivation of the non-linear hydrodynamic equations
for lattice gas automata satisfying the semi-detailed balance condition, and of
their subsequent linearization yielding correlation function expressions for the
dissipative coefficients.
8.5 Long-time tails
Before closing this chapter, an important comment is in order. There is a

restriction in our derivation of the hydrodynamic equations in the sense that
by using the Boltzmann approximation we have neglected all mode-coupling
contributions (Ernst, 1990b; Brito and Ernst, 1992). The consequence of mode-
coupling effects is that the correlation function:
(dmir^SrijÔ)) (8.70)
generally exhibits long-time behavior usually in the form of algebraic decay:

long-time tails ~ t~D^2, where D is the space dimension.7 This implies that (i) in
dimensions lower than or equal to 2 the sum in (8.68) diverges, and the hydro-
dynamic equations are valid only for regimes in which mode-coupling effects are
negligible, and (ii) in dimensions 3 and higher, the form of the hydrodynamic
equations remains valid, but the transport coefficients are renormalized.
In mode-coupling theory, one starts with the idea that the long-time behavior
can be explained on the basis of hydrodynamic arguments. Consider the case
of the velocity of a tagged particle; the mode which describes the decay of
its velocity correlations, the shear mode, and the mode which describes particle
displacements, the diffusion mode, are coupled. The assumption is that eventually
the particle velocity will be equal to the fluid velocity, so that the velocity of
the particle is expressed in terms of the particle probability density and of the
fluid velocity fluctuations. The former obeys the diffusion equation and the
latter the linearized Navier-Stokes equation. So the basic assumption combines
7
The first observation of long-time tails was reported by Alder and Wainwright (1970)
who measured the time decay of the velocity autocorrelation function for hard disks and
hard spheres, by molecular dynamics simulations.
8.5 Long-time tails 197
the solutions of the two equations, and the result for the normalized velocity
autocorrelation function \p(t) reads (Ernst, 1991):
xp(t) ~ ^ = ^ (1 - d) ^ [4n(y + Ds)t]-D/2, (8.71)
where bd/vo is the number density per elementary unit volume of the lattice (e.g.
v
o = \/3/2 m * n e triangular lattice), and Ds denotes the self-diffusion coefficient.
This result is, in fact, the same as in continuous fluid theory (Ernst, Hauge
and van Leeuwen, 1971), apart from the factor (1 — d) which comes from the
exclusion principle.
10°
101 r
io- 2 -N
101-3 =-
101-5 -
101-6 =-
101-7 -
10*
10° 101 102 103 104
t
Figure 8.1 Normalized velocity autocorrelation function xp(t) of a tagged

particle computed with the moment-propagation method in the FHP-3 lattice
gas: the long-time behavior ~ t~~D/1 (here D = 2) is observed over more than
two decades; reduced density d = 0.8; error ~ 10~8. (van der Hoef and
Frenkel, 1990).
In order to compute the velocity autocorrelation function of a particle in the

lattice gas where all particles are indistinguishable, 'the' particle must be labeled
differently so that one can follow its dynamics. A very efficient procedure was
developed by Frenkel et al. (Frenkel and Ernst, 1989; van der Hoef and Frenkel,
1990, 1991): the 'moment-propagation' method, which exploits the fact that in a
collision a particle loses its identity. So the collision rules for the tagged particle
are stochastic, and the stochastic rules can be chosen such that any outgoing
particle is equally likely to be the tagged particle. Consequently there are many
different possible paths for the particle, each with a certain probability that can
be evaluated as the product of the scattering probabilities. Then the velocity
autocorrelation function can be computed by averaging over all possible paths
of one tagged particle and in turn over all possible particles. This procedure
improves considerably the accuracy of the computation which, for the detection
of long-time tails, requires high precision.
Simulations were performed for 2-D and 3-D lattice gases and the decay
of the computed velocity autocorrelation function ~ r~D/2 was found to be in
excellent agreement with the mode-coupling prediction. An example is shown
in Figure 8.1 for the FHP-3 model. Extended mode-coupling theory has even
improved these results by obtaining almost perfect agreement with the simulation
results for the amplitude factor (Naitoh et al, 1991). Now it is true that the
amplitude of the long-time tails is in general quite weak, and for most practical
purposes their effect can be neglected. Nevertheless, the evidence of the algebraic
decay ~ t~D^2 of the velocity autocorrelation function as shown in lattice gas
simulations constitutes by far the most accurate verification to date of mode-
coupling predictions for hydrodynamic long-time tails: this stays as one of the
beautiful realizations of the lattice gas approach to statistical mechanics, and
shows that LGAs are well suited to serve as a testing ground for concepts in
kinetic theory.
Chapter 9
Hydrodynamic regimes
In Chapters 5 and 8, we derived the macrodynamic equations for single species

thermal and non-thermal lattice gases. These macrodynamic equations govern
the dynamics of the lattice gas in the hydrodynamic limit, i.e. they describe the
long-time evolution of large-scale fluctuations of the macroscopic variables.1
'Long-time' and 'large-scale' should be understood as long and large compared
to the microscopic natural time and space scales, the automaton time step and
the lattice mesh, respectively.
Now we will show, at least for non-thermal models, that there exist regimes
for which the macrodynamical equations of the lattice gas become identical
to the standard hydrodynamic equations describing real fluids. We restrict our
analysis to non-thermal models for two reasons. First, because the concepts are
more easily and efficiently put to work (the case of thermal models is much
more complicated); and second, because most of lattice gas simulations of
hydrodynamic phenomena - in particular those reported in Chapter 10 - have
been performed with non-thermal models.
We start from the macrodynamical equations for non-thermal models (see
Section 5.7.3):
= 0, (9.1)
The macroscopic variables are the ensemble-averages of the conserved quantities (see
Section 4.5).
199
200 9 Hydrodynamic regimes
and
* + dp [gjaupj + daP =
^ +
where g(p) is the non-Galilean factor:
and P is the hydrostatic pressure:
pip j 2) = — p — — ^—j 2
(9.4)
with
A = ~"22 ""**". (9.5)
The geometrical coefficients ^2 and £4 are defined in Section 4.5.2, and their
values are given for various models in Table 4.1.
The expansion parameter e in (9.2) is the space-scale separation parameter;
the hydrodynamic limit corresponds to e < 1, the condition under which the
macrodynamical equations were derived. The expansion parameter rj represents
the order of magnitude of the macroscopic momentum flux j , that is, the (small)
'deviation' from zero-velocity basic equilibrium (see Section 4.5).
9.1 The acoustic limit
The hydrodynamic equations for continuous fluids, such as the Navier-Stokes

equations, are known to exhibit acoustic solutions, that is, solutions for which
the fluid velocity and the density fluctuations are small enough so that the non-
linear terms become negligible. The problem was considered in Chapter 6; we
reiterate the reasoning here in a slightly different presentation for non-thermal
fluids.
Consider the density fluctuation dp defined by:
where po is the constant and uniform mean density, and the velocity fluctuation
9.1 The acoustic limit 201
u(t, r) (with zero mean velocity). When p(t, r) and u(f, r) are small, we can ignore
the non-linear terms (in dp and u) in Equations (9.1) and (9.2) to obtain:
dp
6 3 D-
+ -yda— = vdpd pua + (V + —
where v and v' stand implicitly for v(po) and v'(po) respectively. Using the
standard 'nabla' notation, the preceding set of equations can be rewritten as:
(9.7)
_
^ V ( ^ ) = vAu + (V + ^-jr^v) V(V • u),
b po \ D /
where we clearly recognize the standard equations of viscously damped acoustics.
Here a comment is in order about Galilean invariance. The acoustic equa-
tions (9.6) or (9.7) are Galilean invariant while the full lattice gas macrodynam-
ical equations are not (because of the non-Galilean factor g(p)). In the acoustic
limit, Galilean invariance is recovered because the non-linear term dp(g(p)upj a)
is absent in (9.6) and (9.7).
Proceeding one step further in simplification, we consider the inviscid flow,
i.e. we consider short (yet macroscopic) time scales where viscous damping is
negligible. The dynamics is then governed by the standard Eulerian acoustic
equations:
Po
. , (9-8)
This set of equations has sound wave solutions 2 of the following form:
u(t,r) = u[ exp (ikcst — zk • r j k,

(9.9)
Sp(t,r) = k u(r,r)
Po cs
where u[ is an arbitrary (small) velocity amplitude, k = k/fc is the longitudinal
unit vector (parallel to the wave vector k), and cs is the sound speed of the
lattice gas which depends only on geometric parameters:
(9.10)
2
These sound waves are called 'longitudinal waves' because the fluid velocity u is
everywhere parallel to the wave vector.
As in classical hydrodynamics, we define the 'Mach number' Ma as the ratio of

the fluid velocity to the sound speed:
Ma=-. (9.11)
From (9.9), we observe that the relative density fluctuation Sp/po is of the same
order of magnitude as the Mach number.
On longer time scales, the viscous terms become relevant, and Equations (9.7)
exhibit solutions describing sound wave propagation with an amplitude decaying
exponentially slowly in time:
u(t,r) = u[ exp (iky/cs2 — a2k2 t — ik- rje~ afc2t k,

(9.12)
Sp(t,r) = k u(t,r) ^
Po cs
Here the damping coefficient a is a combination of the kinematic viscosity and
bulk viscosity coefficients:
D 2 '
and the phase shift <f> is given by:
ock
sin© = —,
Cs
cos 4> > 0.

In Chapter 10, we explain carefully how post-processed numerical simulation
data of damped sound waves provide an efficient way to 'measure' the damping
coefficient a, and the sound speed cs in a lattice gas.
In addition to longitudinal sound waves, there exists a second family of
solutions: 'shear waves' or 'transverse waves'. They are characterized by a
uniform density field, and a velocity field orthogonal to the wave vector:
u(£,r) = u[ exp(—ik • r) e~vk2t kj_,
Po
These are not propagating waves (there is no phase rotation); their amplitude
decays exponentially over long time scales. The shear wave solution is not
of direct interest for lattice gas simulations of fluid dynamics, but appears to
be very useful to obtain an independent measure of the kinematic viscosity
coefficient v. From the damping coefficient a of sound waves and the damping
coefficient v of shear waves, we can then compute the shear and bulk viscosities
separately (see Chapter 10).
9.2 The incompressible limit 203
9.2 The incompressible limit
For continuous 'real' fluids, a hydrodynamic regime is said to be 'incompressible'

when the density of any material element of the fluid stays constant (= p 0 ) during
its motion. In other words, the derivative Dp/Dt of the density of a fluid element
following the motion of the fluid must vanish:
dtp + u • V(p) = 0.
The continuity equation and the momentum equation for real fluids then be-
come:
V-u = 0,
p (9-14)
dtu + u V(u) = - V ( —) + vAu.
Po
The first equation is the 'incompressibility condition', and the second is the
'incompressible Navier-Stokes' equation (see for example Batchelor (1967),
page 174).
For 'real' fluids, the incompressible Navier-Stokes equation can be seen as
a low Mach number approximation of the fully compressible Navier-Stokes
equations. 3 The main feature of incompressible regimes is that the density can
be taken as approximatively equal to a constant po, except in the pressure term
which must fluctuate so that the two equations in (9.14) be compatible. This
amounts to saying that pressure fluctuations are slaved to the fluid velocity field
so that the divergence of u remains zero.
The same kind of limit can be taken for the lattice gas; the macrodynamical
equations (9.1) and (9.2) then reduce to:
dpup = 0,
Stua + g(po)dp(uaup) H d aP =
v J
Po
or, equivalently:
V • u = 0,
p (9.15)
dtu + g(po)u • V(u) = - V ( — ) + v(p o)Au.
Po
These equations are not identical to the continuous fluid incompressible Navier-
Stokes equations (9.15) because of the non-Galilean factor g(po) which is con-
stant but (in general) not equal to one. We now show that there exists a simple
rescaling by which we recover the standard incompressible Navier-Stokes equa-
tions.
3
There are some tricky points in the Mach number expansion (see e.g. Majda, 1984).
Let us denote by UQ and by ô the typical fluid velocity and the typical space
scale of the problem (expressed in lattice units) respectively. For example, con-
sider the flow around a fixed solid obstacle in a wind tunnel-like configuration:
the typical velocity wo would be the up-stream mean velocity in the simulated
wind tunnel, and the typical length £§ would be some transverse size of the
obstacle. With a length scale and a velocity scale, we can build the typical time
scale:
to =
g(po)uo
This time scale is equivalent to the 'circulation time' commonly used in con-
tinuous fluid dynamics. The presence of the non-Galilean factor g(po) in this
definition is necessary to render the rescaled equations of motion Galilean-
invariant.
With space, time and velocity scales, we can define the following dimensionless
variables:
T = f, R=-^, U=-, n =P | 2. (9.16)

ô ô w0 g\Po)PoUQ
In terms of these rescaled variables and fields, the incompressible macrodynam-

ical equations (9.15) are rewritten as:
V • U = 0,
(9.17)
dTU + U • V(U) = - V ( n ) + —AU,
Re
where the operators V and A are taken with respect to the rescaled space variable
R. The parameter:
U
Re = ^ ^ . (9.18)
v(po)
will be called the 'Reynolds number', because it plays, in lattice gases, the same
role as the Reynolds number for real fluids.
The equations (9.17) are identical to the dimensionless form of the 'real
world' incompressible Navier-Stokes equations (9.14). They involve only one
dimensionless control parameter: the Reynolds number. The rescaling (9.16) is
sufficient to transform the incompressible macrodynamical equations of non-
thermal lattice gases into the classical incompressible Navier-Stokes equations
for real Newtonian fluids, which are Galilean-invariant.
9.3 Comments 205
9.3 Comments
9.3.1 Invar iances

At the microscopic level, a lattice gas is neither fully rotation-invariant, nor fully
translation-invariant, neither is it Galilean-invariant. However, these invariances,
which are verified for real fluids, are recovered for certain hydrodynamic regimes.
The rotation invariance is recovered at the level of the macrodynamical equa-
tions, provided the lattice gas satisfies fourth order crystallographic isotropy (see
Chapters 2 and 5). The translation invariance is also recovered at the level of the
macrodynamical equations, at least in the hydrodynamic limit (weak gradients,
or, equivalently, large-scale distances). The Galilean invariance is only recovered
for particular hydrodynamic regimes: the acoustic regime and the incompress-
ible regime. In the incompressible regime, the Galilean invariance is recovered
in a less trivial way than for acoustic regimes where the non-Galilean factor
disappears simply because the non-linear term is absent; in the incompressible
regime, its effect is eliminated by a proper 'time stretching' (rescaling).
9.3.2 Four-dimensional models

We should pay special attention to four-dimensional (4-D) models such as
FCHC models. These models can be used, without modification, to simulate
three-dimensional incompressible fluid dynamics, with passive scalar advection-
diffusion. Indeed, let us call wa, a = 1,...,3 the first three components of the
4-D fluid velocity vector u, and w4 its fourth component. Now, consider a piece
of 4-D lattice that is extremely thin in the fourth direction (say, one lattice
unit length wide), with periodic conditions in this fourth direction, so that the
macroscopic fields have vanishing dependence on the fourth space-coordinate.
Then, the 4-D incompressible macrodynamical equations (9.15) degenerate into
3-D incompressible macrodynamical equations for p and wa, a = 1,..., 3, plus
an advection-diffusion equation for the fourth component w4:
dtp + V • u = 0,
dtu + g(po)n • V(u) = - V f — ) + v(p o)AU, (9.19)

\PoJ
dtu4 + g(po)u • V(w4) = v(po)Aw4,
where the divergence, gradient and Laplacian are three-dimensional operators.
The rescaling (9.16), applied to (9.19) yields the 'real-world' three-dimensional
incompressible Navier-Stokes equations with a passively advected diffusing
scalar quantity, which can be viewed, for example, as a passive tracer (like a
dye) concentration with no effect on the fluid density.
9.3.3 Lattice gases to simulate real fluid dynamics

We have shown that, within certain limits, lattice gases can exhibit large-
scale (hydrodynamic) behavior identical to the dynamics of real continuous
Newtonian fluids. Therefore, lattice gas automata can be used, and have been
used as an alternative to standard computational fluid dynamics techniques to
simulate fluid motions. Chapter 10 is devoted to various aspects of lattice gas
simulations.
Chapter 10
Lattice gas simulations
One of our main objectives has been to show that single-species non-thermal
lattice gases can exhibit large-scale collective behavior governed by the same
continuous, isotropic and Galilean-invariant equations as real Newtonian fluids.
This is true despite the intrinsically Boolean, spatially discrete, anisotropic and
non-Galilean invariant structure of lattice gases. Moreover, in the past 10 years,
further lattice gas models have been designed to incorporate more complicated
physical features such as reactive processes, magneto-hydrodynamic phenomena
or surface tension (see Section 11.4 in Chapter 11).
On one hand, there has been considerable effort in basic research to under-
stand the subtleties of the statistical mechanics of lattice gases and on the other
hand intense work has been accomplished to take advantage of the similarities
between lattice gases and real fluids in order to simulate fluid motions with
simple and easily implemented lattice gas algorithms. Indeed, because of their
fully Boolean cellular automaton structure, lattice gases are excellent candidates
for efficient implementations on both dedicated and general purpose computers
with serial, vectorial, parallel or even massively parallel architecture. In addition,
various physical effects can be added at low cost. For example, the presence
in a flow of a rigid fixed obstacle is extremely easy to take into account: it
just requires replacing the standard collision rule by a bounce-back rule (see
Section 2.4.1) on all nodes covered by the obstacle. Modifying the shape or
the position of the obstacle is almost immediate, and no mesh modification is
necessary.
Besides technicalities, the lattice gas methods for fluid dynamics offer an-
other fundamental advantage: the simulation of flows where bifurcations are
207
208 10 Lattice gas simulations
suspected. A bifurcation occurs when a given regime loses its stability in favor
of another regime, as soon as a dimensionless control parameter exceeds a
critical value. The new regime originates from small perturbations which grow
instead of being damped. This growth saturates because of non-linear effects,
and leads to the new regime. These small perturbations operate some kind
of 'natural selection' by eliminating unstable regimes. In natural fluid systems,
thermal noise plays this role. In standard (non-microscopic) computational fluid
dynamics methods, the unstable regime can survive artificially, and an externally
imposed disturbance is sometimes necessary to trigger or at least to fasten the
onset of the new regime. Since the lattice gas has fundamentally a microscopic
structure and possesses spontaneous microscopic fluctuations (see Chapter 7),
the macroscopic variables are naturally noisy: they are averages of naturally
fluctuating microscopic quantities. Thus no external perturbation is needed to
trigger bifurcations. Examples of lattice gas simulations displaying bifurcations
are described in Section 10.5.
Since this chapter is dedicated to the application of the lattice gas method
to simulate fluid dynamics, we start with a rapid overview of lattice gas im-
plementation strategies and of frequently encountered programming difficulties.
We then give some notions on how to perform physically relevant numerical
simulations offluidflows.Lattice gas numerical experiments are then described
to illustrate these concepts.
10.1 Lattice gas algorithms on dedicated machines
The possibility of building dedicated machines, that is, electronic devices spe-
cially designed to run lattice gas algorithms was considered as early as 1984.
Several machines were constructed and successfully used, like the RAP family
of D. d'Humieres and A. Cloucqueur (ENS, Paris), and the Cellular Automaton
Machines (CAM) of T. Toffoli and N. Margolus (MIT).
We briefly describe one of the earliest of these dedicated machines: the
RAP-1 prototype (October 1985).1 The acronym 'RAP' stands for 'Reseau
d'Automates Programmables' (Programmable Automata Array). The RAP-1
machine is a desk-top, relatively versatile, dedicated hardware, which can handle
two-dimensional lattice gas models, with up to 16 channels per node. It is not,
strictly speaking, a computer, since it contains no microprocessor. It is essentially
built with:
• 16 arrays of video random access memory (VRAM) chips to store and

1
For articles on dedicated machines, see Section 11.7.1 of the 'Guide for further reading'
(Chapter 11).
10.1 Lattice gas algorithms on dedicated machines 209
manipulate the binary digits encoding the instantaneous occupation level

(0 or 1) of the 16 channels on the 512 x 256 nodes of the lattice.
• two arrays of static random access memory (SRAM) chips to store the
look-up table needed for the collision phase, and the look-up table for
the color-encoding of the occupation levels to be displayed.
• Shift registers to perform the propagation phase.
• Logical functions on programmable array logics (PAL) chips, for various
control and data manipulation tasks.
• An interface board for communication with an external personal
computer.
A video board for display on an external monitor of the instantaneous
state of the lattice nodes, 50 times per second.
A power supply.
The RAP-1 machine can update the Boolean configuration of 512 x 256 nodes
with 16 channels each, at a rate of 50 times per second, so that the instantaneous
state of the whole lattice can be displayed at the European standard video
frequency. Figure 10.1 shows the RAP-1 facility, with its external personal
computer and its video display monitor.
The automaton behavior of the RAP-1 machine can be programmed from
the external personal computer, through a set of three tables:
Figure 10.1 The RAP-1 facility. On the left is the external personal computer
required to program and run the RAP machine. On the right is the RAP-1
machine itself with its display monitor.
A propagation table that drives the shift registers machinery so as to

produce the correct information streaming during the propagation phase.
; A collision table, accessed as a look-up table, to perform the desired local
information reorganization node by node during the collision phase.
: A color coding table, which tells the video board how to display, by color
blending, the Boolean state of each node.
The simplest one-dimensional and two-dimensional lattice gas models (HPP,
FHPs, Conway's game of life, colored FHP, etc.) fit easily within the machine
capacity. Note that the look-up table strategy for the collision phase does not
imply that the machine can handle only deterministic models. Indeed, one or
more of the 16 channels per node can be set at random, and used to select
among two or more sub-tables within the same collision table. This trick is used
to implement FHP models for which binary collisions involve a random choice
between two output states (see Chapter 3, Section 3.2).
Several fluid dynamics experiments were performed on the RAP-1 machine,
to provide early demonstrations of the capabilities of the lattice gas method for
fluid dynamics simulations. Figure l(b) in the preface is a good example which
shows theflowbehind afixedsolid flat plate in a wind tunnel-type geometry; the
streamlines are obtained after post-processing the raw data produced by RAP-1
running the FHP-3 lattice gas model. Further results obtained with RAP-1 are
given in Section 10.5.
10.2 Lattice gas algorithms on general purpose computers

Lattice gas algorithms have been implemented, since 1985, on a large variety
of computers, ranging from personal computers to powerful massively paral-
lel supercomputers. Implementations were done on machines running various
operating systems, with programs written in various programming languages.
Obviously it would be out of place to describe all existing implementations; we
simply want to shed some light on the main programming problems and strate-
gies in general, independently of the language used, of the operating system,
and, as far as possible, of the particular hardware architecture.
10.2.1 Channel-wise vs. node-wise storage

'How to store the Boolean field in the memory of the machine?' is the first
problem that the lattice gas code designer has to face. The random access
memory of a general purpose computer is organized in words of typically 32
or 64 bits, and the basic operations recognized by most processors manipulate
words rather than individual bits. Consider a lattice gas model with b channels
per node, residing on a piece of lattice with Jf nodes. The storage problem is the
following: how to dispatch the b x Jf bits necessary to encode the instantaneous
Boolean configuration into the available memory words, so as to minimize
memory wasting and to optimize the update speed of the automaton. Two
strategies are commonly used; each strategy has advantages and drawbacks,
that we discuss hereafter. To be simple, suppose we want to implement a lattice
gas model such as HPP (see Section 3.1) which requires 4 channels per node,
on a (non-realistic) computer with 4 bits per memory word. The essentials of
the discussion hold for any number of channels per node, and for computers
with any number of bits per word. We denote by Jf the total number of lattice
nodes that the lattice gas simulation code must manage.
The first strategy is the 'channel-wise storage': each 4-bit word stores the same
channel of 4 different nodes. Thus, 4 arrays of memory words are necessary to
store the whole Boolean field, one for each channel, and each array must contain
Jf 14 words. Figure 10.2(a) illustrates this storage mode. If the number of node
Jf is not an integer multiple of the number of bits per memory word, then
the number of words per array must be Jf 14 + 1, if Jf '/4 denotes the integer
division. The last word of each array is then only partially used. This strategy
has minor memory wasting.
The second strategy is the 'node-wise storage', which, in some way, is 'or-
thogonal' to the the channel-wise storage strategy. Each 4-bit word stores the
4 different channels of the same node. For a lattice with Jf nodes, Jf words
"3 "3 %
1 1 1 1
5 6 S
Nodel 11,(1) n 2 d) n 3 d) n 4 d) Nodel 11,(1) n 2 (D n 3 d) n 4 d)
Node 2 n,(2) n 2 (2) n3(2) n4(2) Node 2 n,(2) n2(2) n3(2) n4(2)
Node 3 11,(3) n 2 (3) n3(3) n4(3) Node 3 11,(3) n2(3) n3(3) n4(3)
Node 4 n,(4) n2(4) n3(4) n4(4) Node 4 n,(4) n2(4) n3(4) n4(4)
Figure 10.2 The two main storage modes of a Boolean lattice gas in the memory words
of a computer. The number of bits per memory words is taken to be 4 for simplicity.
The solid rectangular boxes symbolize the 4-bit memory words, (a) Bit-wise storage;
(b) node-wise storage.
are necessary to store the whole Boolean field. This storage mode is illustrated
in Figure 10.2(b). If the number of bits per memory word is smaller than the
number of channels per node, then two or more words may be necessary to store
the Boolean state of one single node. If, on the contrary, the number of bits per
memory word is larger than b, then the Boolean state of two or more nodes can
be stored in one single memory word. There may be some waste of memory if
the number of bits per memory word is not close to an integer multiple of b.
However, the wasted bits can be used to store auxiliary information (obstacle,
body force, etc.).
Neither of these two storage modes can be preferred a priori; the right choice
depends on several factors. The channel-wise storage wastes very little memory.
The propagation phase is conceptually simple, since it reduces to systematic
data shifts within each of the b arrays of words. If the collision rule is simple
enough to be coded with a non-prohibitive number of logical operations, then
the collision phase is also very efficient, since the logical operations (AND, OR,
NOT ...) act on all the bits of their data words at the same time, so that a
single 'turn off the loop' performs the collision on 64 nodes on a computer with
64-bit words for example. However, when the collision is too complex to be
efficiently coded with combined logical operations,2 one must resort to a look-
up table strategy which operates as follows: the b bits of a single node must be
picked from b different memory words, and gathered into a single auxiliary word
encoding the pre-collision state of the node under consideration. This word is
then sent as an address to the collision look-up table, which delivers the post-
collision state. Then, the converse operation must be performed: the b bits of
the output word must be dispatched into the b arrays of words. These gathering
and dispatching operations can be time-consuming, and the counterpart of an
efficient propagation phase is the time spent in the collision phase.
The node-wise storage can be less memory-efficient, since the number of bits
per words may not be close to an integer multiple of the number b of channels
per node. The propagation phase is less straightforward than in the channel-wise
storage. Indeed, the post-propagation state of one node must be built from b
bits extracted from b neighboring nodes, and collected in a single word. The
result of this operation must be stored in a buffer copy of the whole lattice
configuration. Indeed, if the post-propagation state of the node were stored back
in its address in the original array storing the pre-propagation state, then the
post-collision state of the neighboring nodes would be computed with already
propagated data. This is a consequence of the non-locality of the propagation
phase. The collision phase however is much simpler, at least for look-up table
strategies, since each word (or fraction of word) of the buffer copy can directly
serve as an address for the look-up table. The results furnished by the look-up
2
This is the case for most models, except the simplest HPP and FHP models.
table are then stored back in the original array, so that they are ready for the
next propagation phase.
To sum up the discussion, the channel-wise storage is in general preferable for
models with simple collisions and/or if the available amount of memory is the
bottle-neck of the computer to be used; the channel-wise storage offers better
efficiency for models with complicated collision schemes, and/or on a computer
where the real bottle-neck is the CPU-time.
10.2.2 Collision strategies

Several strategies have been designed to handle the difficult problem of the
collisions. In the previous section, we mentioned the implementations with
logical operators and with a look-up table. Here we give some further details
for these strategies, and describe other algorithms that have been used in some
cases.
10.2.2.1 Collision algorithms with logical operators

When the number of channels per node is relatively small, and if the collision
rule is sufficiently simple, it may be efficient to design a combination of logical
operators that transforms the pre-collision state into the post-collision state.
With this combination, one can handle random Boolean variables for non-
deterministic models. The solution is equivalent to a transcription with logical
operators of the polynomial collision term of the microdynamic equation (see
for example Equation (3.6) for the FHP-1 model). This transcription simply
amounts to changing the multiplications and the additions by AND and OR
operations, respectively. (The OR substitution is safely possible because in the
microdynamic equations only one term in each sum can be non-zero.) Except
for simple models such as HPP, FHP-1 and FHP-2, this method is prohibitively
slow.
10.2.2.2 Collision algorithms with look-up tables

This solution is memory-consuming but rather fast, especially on machines where
indirect addressing is efficient. The principle is simple: take the pre-collision state
as an address in a table, and the information stored at this address gives the
post-collision state. The table must have 2h entries for a model with b channels
per node. This rapidly becomes cumbersome for models such as FCHC which
require up to 24 or even 26 channels per node. On a 64-bit machine, a look-up
table for FCHC-8 requires 512 megabytes of random access memory, just to
hold the table. Of course, a single table makes the model deterministic. This
is often incompatible with the condition of G-invariance required for correct
macrodynamics. Three solutions exist: (i) use several tables with a random
swap between each table, (ii) keep a single table designed to incorporate some
kind of 'frozen randomness' to reduce the resulting lack of G-invariance to

an acceptable level, (iii) use a multi-entry table (see next section). The first
solution is often prohibitively memory-consuming, the second solution works
relatively well, especially for models such as FCHC for which the huge number
of possible states introduces some natural shuffling. The third solution works
well for models such as GBL, for which only a few states lead to random choices
with only a few possibilities. The next section is devoted to this kind of collision
algorithms with multiple entries.
10.2.2.3 Collision algorithms with multi-entry look-up tables

Very frequently, only a small fraction of the 2b possible Boolean states leads
to several post-collision states, but the number of these possible post-collision
states may be too large for a randomization strategy based on a random choice
between several complete look-up tables. The multi-entry look-up table method
solves this problem as follows: a first table stores, in a compact format, a
starting address and an ending address for each of the 2b possible Boolean
states. These starting and ending addresses, obtained for a given (pre-collision)
state, define a zone in a second table, where all possible post-collision states are
stored. It is then easy to select randomly an address between the starting and
ending addresses, and to pick the post-collision state stored at this address. This
mechanism can be improved to allow for non-equally probable post-collision
states.
10.2.2.4 Collision algorithms using symmetries

It is possible but time consuming to use the lattice symmetries to reduce the
size of the collision tables. The principle is simple: the possible Boolean states
are sorted into families, gathering states which can be transformed one into the
other by a lattice isometry. For each family, one single 'family leader' is selected,
and its possible output states are stored in a table.
10.2.3 Obstacles
When the node-wise storage mode is selected, there exists a simple and efficient
way to implement bounce-back boundary conditions (see Section 2.4.1). Suppose
that the b microscopic velocity vectors Cj are labeled such that Ci+b/2 = — c,-. If
the Boolean states are stored node-wise in computer words, so that the most
significant bit of each word stores the occupation level of channel 1, and so
on in sequence up to channel b, then the bounce-back collision is simply a
permutation of the b/2 most significant bits with the next b/2 following bits.
This requires only two masking operations, two shifts, and an 'OR' operation.
In addition, it is useless to apply this bounce-back rule in the bulk of an
obstacle, provided all obstacle nodes are initialized with zero particle density. It is
sufficient to apply the bounce-back rule to the 'surface' of the obstacle only, that
is, to the obstacle nodes that can be occupied because of particle propagation.
Indeed, the bounce-back rule sends back (to where it came from) any particle
arriving at the surface, so preventing any fluid particle 'contaminating' the
bulk of the obstacle. The thickness of the 'surface' layer depends on the set
of microscopic velocity vectors: if some of these vectors connect a node to its
second or third nearest neighbors, then the layer must be two or three times
thicker.
10.3 Essential features of a lattice gas simulation code
In the previous section, we addressed the main problems and choices involved in
designing the hard core of a lattice gas simulation code: the automaton updating
machinery. However, such a code cannot be reduced, even conceptually, to its
hard core. Many auxiliary software tools are needed to properly operate the
code and to extract qualitative physically relevant data. In addition to pure
programmer skill, these auxiliary tools sometimes require deep physical insight.
These tools can be roughly classified into three main categories: initialization
tools, raw physical data extraction, and post-processing.
10.3.1 Initialization
Like any computational fluid dynamics method, the lattice gas method requires
one to decide what should be the initial macroscopic fields (uniform, harmonic,
etc.). Furthermore, for lattice gas methods, there is an additional task before
starting the time-evolution: the initial microscopic Booleanfieldmust be set so as
to correspond to the desired initial macroscopic fields. This involves a random
initialization of the Jt;(r*) with 0s or Is, with the minimal constraint that the
resulting averaged macroscopic fields be actually the desired initial macroscopic
fields. Often there is an additional constraint: the random initialization of the
tt;(r*) must be realized with probabilities that correspond to (local) equilibrium
distributions, such as those discussed in Chapter 4. This eliminates (or short-
ens) the transient regime during which local equilibrium would have to settle.
Therefore the average populations fi(f*) corresponding to the desired initial
macroscopic fields must be computed according to the relations obtained in
Section 4.5. Then, at each node and for each channel, a pseudo-random number
generator must be used to assign to n^r*) the value 0 or 1, with a probability
/i(r*). A convenient way to do this is to use a random number generator deliv-
ering values cp with a uniform distribution between 0.0 and 1.0 (most built-in
random number generators do so); the integer part of cp +/,(r*), is then either
0 or 1, with respective probabilities /j(r*) and 1 — /*(r*).
If the initial macroscopic fields are uniform, the probabilities /j(r*) do not
depend on r*. Then they can be pre-computed once, for all the nodes (instead
of once for each node), leading to considerable speed-up.
Note also that if CPU time is a crucial issue - especially for non-homogeneous
initial fields that require computation of the probabilities /,-(r*) for each node
- a simplified procedure can be used to obtain the //(r*), usually with no
dramatic physical consequences: one can use approximation formulae for the
probabilities /;(r*), such as those obtained in Sections 4.5.1 and 4.5.2, but taking
into account the first order correction terms only. This simplifies the computation
considerably.
10.3.2 Raw physical data extraction

The lattice gas method for fluid dynamics simulations uses a fictitious mi-
croscopic word. Like in the 'real world', macroscopic relevant quantities are
recovered through some averaging procedure, and the resulting averaged quan-
tities are noisy (like in the 'real world'!), with a 'signal to noise ratio' that
increases roughly as the square root of the number of samplds used for aver-
aging. But this residual noise is much larger in the lattice gas context than for
real fluids because, in general, the number of samples over which the averages
can be realistically computed is much smaller (by several orders of magnitude)
than the number of particles in the smallest real macroscopic volume. Three
kinds of averaging can be performed, separately or in combination, in lattice
gases: ensemble-averaging, space-averaging, and time-averaging. The choice of
the right method is non-trivial and may involve subtle physical considerations.
10.3.2.1 Ensemble - averaging

Ensemble-averaging requires the simulation to be run with many different re-
alizations of the Boolean fields, initialized independently with the same macro-
scopic fields, that is, with the same probability distribution. The macroscopic
fields are then computed at any time, by summation over all these micro-
scopically different replicas. This strategy is particularly well suited to parallel
computing, since all the replicas evolve independently, without any need for
information exchanges. This noise-reduction technique has however a serious
drawback for fluid flows when symmetry-breaking is expected: different repli-
cas may break the symmetry differently, and the averaging over these replicas
may spuriously destroy the broken symmetry. An example is discussed in Sec-
tion 10.5.3.
10.3.2.2 Space- and time-averaging

By space- and/or time-averaging one computes the mean value of the macro-
scopic fields by averaging over small regions ('cells') of space or space-time. The
larger the cells, the smaller the residual noise, and the poorer the effective space-
and/or time-resolution. So there should be a compromise between the loss in
resolution and the loss in accuracy due to the residual noise. This averaging
procedure can also be viewed as a 'poor man's' filtering process, where the
microscopic data are convoluted with a square function that is equal to one in a
small space or space-time domain, and zero everywhere else. This interpretation
gives a natural generalization of the space- and/or time-averaging: the square
function can be replaced by any smooth function whose Fourier transform de-
cays smoothly at high space- and/or time-frequencies. However, the compromise
between noise reduction and effective resolution still remains: setting the cut-off
frequency too low will efficiently reduce the noise, but may reject physically
interesting small structures; conversely, setting the cut-off frequency too high
will produce prohibitively noisy fields.
10.3.3 Post-processing
Like standard floating point computational fluid dynamics methods, lattice gas
simulations often require serious processing for drawing physically significant
results from raw data, i.e. the averaged macroscopic fields. This may involve
physical aspects such as computing macroscopic fields different from those
delivered naturally by the lattice gas simulation. For example, one may wish to
compute the pressure field and the vorticity field from the density and velocity
fields. This phase may also involve graphical aspects, such as choosing the
right way to display the results to highlight a particular phenomenon, which is
of crucial importance, especially for three-dimensional simulations. Since these
problems are common to all computational fluid dynamics methods, we shall
not embark on a detailed discussion here. We shall focus attention on post-
processing problems involving specific aspects of the lattice gas method, such as
drag and lift coefficients measurements, and Strouhal number evaluations.
10.3.3.1 Drag and lift coefficients measurements

A viscous fluid moving around a fixed solid obstacle always exerts a force on
the object. This force can have longitudinal and transverse components with
respect to the mean flow direction. The longitudinal component, expressed in a
suitable dimensionless form, is the 'drag coefficient':
CO..)
where p is the mass density, u is the mean flow velocity, and S is the area of
transverse section of the obstacle, that is, the area of projection of the obstacle
onto a plane orthogonal to the mean flow velocity.
The transverse component of the force, expressed in the same way, yields the
'lift coefficient':
CL = y % - . (10.2)
These quantities are of obvious interest, and can be rather accurately extracted
from a lattice gas simulation (see Rivet, 1993). First, one has to compute the
momentum variation Ap during one time step, due to the collisions on the
obstacle. This must be done at each time step during the simulation, and
the values must be saved to be subsequently post-processed. Then one has to
compute the area of transverse section of the obstacle, in natural lattice units,
and the average mass density per unit volume, in natural lattice units. The volume
of the unit lattice cell must be taken into account if it differs from 1 (in natural
lattice units). The drag and lift coefficients are obtained from:
f^' (10.3)
g(p)\pu2S
and
CL = , f/\. (10.4)
g{p)\pu2S
The factor g(p) appears to express the momentum variation per natural lattice
time unit (time step) as a force, that is, a momentum variation per natural
physical time unit (see Section 9.2). It was found that the values of the drag
coefficients computed for spheres and cylinders with the lattice gas method fall
within a small percentage of experimental values.
10.3.3.2 Strouhal number evaluation

When the fluid velocity around an obstacle increases so that the Reynolds
number exceeds a critical value, the initially laminar stationary flow can pro-
duce an oscillating wake, such as the Benard-von Karman vortex street (see
Section 10.5.3). The frequency fw of the oscillation (in dimensionless form) is
the 'Strouhal number':
St = ^ , (10.5)
where L is the typical size of the obstacle, and u is the mean flow velocity.
The Strouhal number - and how it varies as a function of the Reynolds
number - may be a precious indicator of what happens in the wake (Konig,
Eisenlohr and Eckelmann, 1990). This is the case for example in the wake of
a cylinder, where the three-dimensional structure of the wake may be complex.
The Strouhal number can be computed by 'lattice gas hot-wire anemometry',
i.e. by implementing the numerical analogue of a 'hot-wire velocity probe', a
laboratory device that delivers the local fluid velocity as a function of time (see
Bonetti and Boon, 1989). In the lattice gas simulation, the 'numerical' hot-wire
probe is an averaging cell, over which the averaged velocity is computed at
each time step during the simulation, and stored for further post-processing.
The resulting (discrete) signal is very noisy, but a low frequency mode clearly
dominates. The next step is to get the best possible estimate of the frequency of
this dominant mode, which is done with standard signal processing algorithms
such as modulation and filtering. The frequency so obtained yields the Strouhal
number through the formula:
^ (10.6)
where L and u are expressed in natural lattice units. The lattice gas-specific non-
Galilean factor g(p) appears as a scaling factor between the physical time unit
and the natural lattice time unit (see Section 9.2). Strouhal numbers obtained
from LGA simulations of the wake behind a cylinder are also found to be in
agreement (within a small percentage) with experimental values.
10.4 Measurement of basic lattice gas properties
In order to perform reliable quantitative lattice gas simulations, it is essential

to have correct evaluations of the non-Galilean factor g(p), of the sound speed
cs, and of the kinematic and bulk viscosity coefficients (v and v'). The theoreti-
cal viscosity coefficients are obtained within the Boltzmann approximation (see
Section 5.7.4), and thus may yield inaccurate values. Therefore it is important to
obtain an 'experimental measure' of these coefficients by numerical simulations.
For the non-Galilean factor and the sound speed, the macrodynamic theory (see
Chapter 5) provides more accurate values (here the Boltzmann approximation is
not necessary) at least for models satisfying the semi-detailed balance condition.
For models without semi-detailed balance, we must use the Boltzmann approx-
imation from the beginning, even to evaluate g(p) and cs(p) (cs may depend
on the density when semi-detailed balance is not verified). So it is important
to use reliable methods to 'measure' numerically the values of these coefficients
for several densities. This calibration operation is a prerequisite that should be
carried once for all for any lattice gas model that is going to be used for hy-
drodynamic simulations. A convenient and efficient method for this calibration
is to 'measure' the lattice gas quantities from sound or shear wave simulations
(see Section 9.1 on acoustic regimes).
10.4.1 Measuring g(p) and v(p)

Consider a lattice gas simulation running on a lattice domain with periodic
boundary conditions along each direction. The initial state of the lattice gas is
prepared with the following values of macroscopic variables:
uy(0, r) = u\y cos (k x).

The size of the domain in the x direction must be an integer multiple of
wavelength 2n/k9 and both u\x and u\y should be sufficiently small so that the
non-linear terms in the LGA hydrodynamic equations become unimportant (see
Section 9.1). These conditions correspond to a uniform flow in the x direction,
superimposed to a sinusoidal shear flow in the y direction, with a wave vector
in the x direction. The same arguments as in Section 9.1 lead to the conclusion
that the shear wave propagates in the x direction with a phase velocity g{p)u\ x
and decays exponentially as exp(—v(p)k 2t).
Obtaining g(p) and v(p) is then easy: at regularly spaced time intervals, the
macroscopic velocity field is computed, then a spatial Fourier transform is ap-
plied. The phase and the logarithm of the amplitude of the mode corresponding
to k are extracted, and their time-variations are adjusted to linear functions (by a
standard least squares fit method): the slopes yield respectively the non-Galilean
factor g(p), and the kinematic viscosity v(p). The root mean square difference
between the simulation values and the best fitting straight line can serve as an
estimate of the measurement accuracy.
10.4.2 Measuring cs(p) and v\p)

The same method applied to sound (compression) waves provides a measurement
of the sound damping coefficient. As discussed in Section 9.1, the amplitude of
sound waves decays as exp(—Ffc 2t), where
The damping coefficient F is obtained from a least squares fit of the logarithm of
the amplitude as a function of time. Knowing v(p) from shear wave experiments,
and F from sound wave experiments, the value of the bulk viscosity v'(p) follows
straightforwardly.
The sound velocity cannot be accurately extracted by fitting the phase of the
Fourier mode to a linear function. Indeed, the phase speed in the viscous case is
not cs but \/cs2 — T2k2 (see Equation (9.12)). One convenient way to obtain cs
is to compute the Fourier transform of both velocity and density fluctuations,

and to select the mode corresponding to k. The ratio of their moduli gives the
sound speed cs.
10.4.3 An example: the FCHC-3 model

As an illustration of the calibration procedure, we consider the FCHC-3 lattice
gas. Remember that the FCHC-3 model is a four-dimensional FCHC model
with no rest particle (b = 24), with semi-detailed balance and self-duality (see
Section 3.7.3 on the FCHC-3 model). The collision rule involves a look-up table
Table 10.1 Non-Galilean factor g(p) and sound speed cs(p) for the
FCHC-3 lattice gas model, d is the density per channel, p = d x b is the
density per node. gth is the theoretical value of g(p) (see Equation (9.3)),
gexp is the value obtained by shear wave numerical simulations, gerr is the
root mean square estimated accuracy of the simulation. The same
subscripts are used for the sound speed cs(p) (see Equation (9.10) for the
theoretical value).
d P gth gexp gerr Csth Csexp Cserr
0.32 7.680 0.3529 0.3531 0.0040 0.7071 0.7097 0.0040

0.34 8.160 0.3232 0.3230 0.0040 0.7071 0.7079 0.0040
0.36 8.640 0.2916 0.2906 0.0040 0.7071 0.7068 0.0040
0.38 9.120 0.2580 0.2571 0.0040 0.7071 0.7079 0.0040
0.40 9.600 0.2222 0.2215 0.0040 0.7071 0.7073 0.0040
0.42 10.080 0.1839 0.1827 0.0040 0.7071 0.7065 0.0040
0.44 10.560 0.1428 0.1416 0.0040 0.7071 0.7046 0.0040
0.46 11.040 0.0987 0.0980 0.0040 0.7071 0.7053 0.0040
0.48 11.520 0.0512 0.0509 0.0040 0.7071 0.7055 0.0040
0.50 12.000 0.0000 0.0005 0.0040 0.7071 0.7067 0.0040
Table 10.2 Kinematic and bulk viscosity coefficients v(p) and v'(p) for
the FCHC-3 lattice gas model. The convention for the meaning of the
subscripts is the same as in Table 10.1.
d p V th V exp V err v'th V'exp V'err
0.32 7.680 0.0351 0.0392 0.0007 0.0000 0.0014 0.0015

0.34 8.160 0.0321 0.0359 0.0007 0.0000 0.0001 0.0011
0.36 8.640 0.0296 0.0329 0.0006 0.0000 0.0021 0.0013
0.38 9.120 0.0274 0.0301 0.0006 0.0000 0.0022 0.0013
0.40 9.600 0.0256 0.0280 0.0006 0.0000 0.0010 0.0011
0.42 10.080 0.0242 0.0262 0.0006 0.0000 0.0014 0.0015
0.44 10.560 0.0230 0.0245 0.0006 0.0000 0.0006 0.0009
0.46 11.040 0.0222 0.0234 0.0005 0.0000 0.0007 0.0008
0.48 11.520 0.0217 0.0226 0.0005 0.0000 0.0010 0.0008
0.50 12.000 0.0215 0.0225 0.0005 0.0000 0.0003 0.0005
strategy with one single table. So the FCHC-3 lattice gas is only approximately
G-invariant. Simulations are performed over a domain of 256 x 64 x 64 lattice
nodes (see Dubrulle et al, 1990). Tables 10.1 and 10.2 give the theoretical values,
the measured values and the estimated accuracies, for g(p) and cs(p), and for
the coefficients v(p) and v\p) of the FCHC-3 lattice gas model.
10.5 Examples of lattice gas simulations

We now turn to more concrete applications as we present numerical experiments
performed by lattice gas methods for some classical fluid dynamical phenomena. 3
10.5.1 The Kelvin-Helmholtz instability

Consider two fluid layers initially separated by a flat interface, and moving with
respect to each other with parallel velocities. When the velocity difference is suffi-
ciently large, the flat interface becomes unstable, and a sinusoidal disturbance of
the interface grows and eventually saturates because of the non-linearities. This
instability produces an array of identical vortices rolling 'between' the two fluid
layers, a phenomenon known as the 'Kelvin-Helmholtz instability' (see Drazin
and Reid, 1981). This type of flow was simulated in two dimensions on the
RAP-1 machine using the FHP-3 lattice gas model, and variants incorporating
two different species with diffusion, or reaction-diffusion. The advantage is that
Figure 10.3 Kelvin-Helmholtz instability simulated on RAP-1, with a diffusive

two-species version of the FHP-3 model. Snapshot after 1200 automaton steps.
For further illustrations of lattice gas simulations, see Section 11.7 in Chapter 11.
the second species may serve as a tracer for better visualization of the vortices.
Figure 10.3 shows the simulation of the Kelvin-Helmholtz instability after 1200
time steps. The upper layer flows to the left, and is more concentrated in one of
the two species, whereas the lower layer flows to the right and is more concen-
trated in the other species. The gray levels label the concentration differences.
The two layers are initially separated by a flat interface, and after 1000 time
steps the Kelvin-Helmholtz instability is clearly visible. The patterns in Fig-
ure 10.3 can be compared (qualitatively) to those in Figure 10.4 showing (real)
clouds in two atmospheric layers undergoing a Kelvin-Helmholtz instability.
Figure 10.5 illustrates a simulation with a slightly different collision rule where
the two species are subject to a reactive process according to a majority rule:
if there are more particles of species A on a node, then the particles of the less
abundant species B are transformed into particles of species A. Consequently,
the gradients are sharper, and the dispersive effect of particle diffusion is re-
duced; so one can run the simulation for longer times (4000 time steps) and
avoid fuzzy contours.
10.5.2 Particle aggregation

With a simple modification of the FHP-3 rules, one can study qualitatively the
fractal aggregation of solid particles. A randomly chosen rest particle plays the
Figure 10.4 Photograph of real clouds between two atmospheric layers with
differential motion: the Kelvin-Helmholtz instability occurring between these
layers is made visible by the clouds. (Drazin and Reid, 1981.)
Figure 10.5 Snapshot (after 4000 automaton steps) of the Kelvin-Helmholtz

instability simulated on RAP-1, with a two-species reaction-diffusion version of
the FHP-3 model. The reactive process is used to reduce the dispersive effect
of diffusion, in order to obtain sharper contours. Note that this lattice gas
model differs from the model used in the simulation illustrated in Figure 10.3.
Figure 10.6 Fractal aggregation of particles starting with a 'seed' located at

the center of the figure, simulated with a FHP-3 lattice gas model with 'sticky'
rest particles.
role of the initial seed: it is 'sticky' in the sense that any moving particle arriving
at one of the neighboring sites (next to the seed) stops and stays, and becomes
sticky too. In the course of time, an aggregate of fixed sticky particles forms and
grows with an apparently fractal structure. Figure 10.6 showing such a fractal
aggregate was obtained after just a few seconds of simulation on RAP-1.
10.5.3 Two-dimensional flow past an obstacle

The FCHC lattice gas models described in Chapter 3 are basically four-
dimensional models which were initially designed to simulate three-dimensional
incompressible flow with, in addition, the possibility of tracking a passive scalar
in the flow (see Section 9.3.2). With one more dimensional reduction (as de-
scribed in Section 9.3.2), four-dimensional models can also be used - in a
degenerate way - to simulate two-dimensional hydrodynamics with two passive
scalars.
Having recourse to four-dimensional models to simulate three-dimensional hy-
drodynamics is quite logical since there exists no simple really three-dimensional
lattice gas model with the required symmetry giving correct isotropy in the hy-
drodynamic equations. But it may seem somewhat odd to use four-dimensional
models for two-dimensional hydrodynamics, since simpler really two-dimensional
hydrodynamically correct lattice gases exist. In fact there is a good reason for
this apparently paradoxical idea. Each model has an intrinsic kinematic viscos-
ity coefficient that depends on the details of the collision rules. This coefficient
cannot be made indefinitely small, at least for models with semi-detailed bal-
ance (see Section 2.3.1). This limitation imposes an upper bound to the range
of achievable Reynolds numbers, with a given amount of computer memory.
Because FCHC models have much lower viscosity than usual two-dimensional
lattice gases, using FCHC for two-dimensional hydrodynamics allows us to
reach higher Reynolds numbers, and thereby to broaden the range of achiev-
able hydrodynamic regimes, with a more efficient use of computer resources.
The price to pay is the complexity of the model.
Figures 10.7 and 10.8 illustrate the capability of FCHC models to simulate
two-dimensional hydrodynamics. The FCHC-8 lattice gas was implemented on
a four-processor CRAY-2 computer. The simulation domain is a rectangle of
4096 x 1024 lattice nodes. The upper and lower boundaries are solid walls, with
bounce-back rules. A wind-tunnel type flow is driven by momentum injection at
the inlet (left boundary) and momentum removal at the outlet (right boundary)
so as to maintain a roughly constant overall left to right fluid velocity with
modulus ~ 0.3 in lattice units. To the left (at 100 nodes of the inlet) and at mid-
height, there is a rectangular solid obstacle spanning over 32 x 100 nodes. The
resulting Mach number and Reynolds number are Ma = 0.47, and Re = 564,
respectively. Although this value of the Mach number may seem relatively high
to be compatible with incompressibility, the density fluctuations were checked

to stay below 3.5% of the mean value d = 0.38.
The simulation domain is divided into 32 x 32 square cells, over which the fluid
velocity and density are space-averaged (with no time-averaging). No ensemble-
averaging is performed because the fluid flow is likely to break spontaneously
the 'top-bottom' symmetry. Statistically, the independent realizations over which
the ensemble-average would be computed may break this symmetry differently,
and consequently, ensemble-averaging would 'erase' the symmetry-breaking.
Figure 10.7 shows the macroscopic velocity field 40000 time steps after the
impulsive start of the flow. This corresponds to 69.2 circulation times (see
Chapter 9). The velocity is coded with one arrow per averaging cell, and the
mean flow is subtracted in order to emphasize the wake structure. A long
regular Benard-von Karman vortex street is clearly visible in the wake of
the obstacle (for basics about Benard-von Karman vortex streets, see Lamb,
1945). The vorticity field is not a direct product of the lattice gas simulation.
It must be computed from the velocity field during the post-processing phase.
Figure 10.8 is a view of the vorticity field corresponding to the velocity field
shown in Figure 10.7. The gray scale is such that positive values of the vorticity
(counter-clockwise vortices) appear in white, and negative values (clockwise
Figure 10.7 Two-dimensional incompressible flow around a rectangular obstacle in

wind-tunnel geometry, simulated with the FCHC-8 lattice gas model on a CRAY-2
computer. The fluid velocity field with mean flow subtracted is represented 40 000 time
steps (69.2 circulation times) after the impulsive start of the flow.
Figure 10.8 Gray-level coding of the vorticity field corresponding to the velocity field in
Fig. 10.7.
vortices) appear in black. The square 'pixels' clearly visible in Figure 10.8 are
the averaging cells. Purposely, no smoothing was performed in this picture,
because it would have hidden the coarse discretization of the physical space,
imposed by the necessity of space-averages. As explained in Section 10.3.2, the
size of the averaging cells results from a compromise between the accepted noise
level, and the desired physical resolution: larger cells would have led to less noisy
values of the velocity field, at the expense of a poorer space-resolution; smaller
cells would have led to a better resolution, with an increased noise level. With
both space- and time-averaging, one would have a gain in 'signal-to-noise' ratio
for a given space-resolution, but at the cost of a loss in time-resolution.
10.5.4 Three-dimensional flow past an obstacle

The initial vocation of the FCHC family is obviously the simulation of three-
dimensional hydrodynamics. Several models of this family have been successfully
used to recover subtle, typically three-dimensional fluid dynamics phenomena,
such as the loss of translation-invariance in the wake behind a cylinder. The
details of the vortex patterns in the wake, and their time-evolution were obtained
from the lattice gas simulation, and several features observed in real fluids have
been recovered.
Experimental studies of incompressible flows around circular cylinders reveal
that the two-dimensional wake behind the cylinder undergoes a Hopf bifurcation
when the Reynolds number exceeds a critical value close to 47. This bifurcation
characterizes a regime transition from a stationary wake with a symmetrical
recirculation zone, to the oscillating Benard-von Karman vortex street illustrated
in Section 10.5.3. The three-dimensional structure of the wake raises complex
questions. In the limit of an infinitely long cylinder, is the wake translation-
invariant in the direction of the axis of the cylinder? Is there a range of Reynolds
numbers where the translation-invariant wake remains stable (a Benard-von
Karman array of straight vortices parallel to the cylinder)? In that case, what
is the critical Reynolds number beyond which the translation-invariant wake
becomes unstable, and how does the transition to a more complex regime
occur? There are several reasons why laboratory experiments are difficult to
carry out and their results hard to exploit. For example, in practice, cylinders
are necessarily finite and their ends must be firmly held to avoid vibrations
and spurious coupling between fluid oscillations and obstacle vibrations. The
question is then: how and how much do the boundary conditions at the ends
of the cylinder influence the wake patterns and the relative stability of each
regime? For reviews of this interesting problem, see Williamson (1989) for the
experimental aspects, and Huerre and Monkewitz (1990) for the theoretical
aspects.
We now describe a lattice gas simulation of this flow performed with the
FCHC-8 model. Figure 10.9 illustrates the geometry of the simulation. The
solid arrow labeled 'I/' shows the direction of the incoming flow. The boundary
conditions are periodic in the transverse direction X and in the span-wise
direction Y. In the stream-wise direction Z, the inlet and outlet conditions are
as described in the two-dimensional simulation of Section 10.5.3. The upstream
velocity is 0.19 in lattice units, and the corresponding Mach number and
Reynolds number are Ma = 0.29 and Re = 73, respectively. The Mach number
has a low enough value to maintain the density fluctuations within 1.5% of the
mean value d = 0.42 (for more details, see Rivet, 1993).
Figure 10.10(a) is a snapshot taken 16000 time steps (74.4 circulation times)
after the impulsive start of the flow. It shows a two-dimensional cross-section
through the fluid velocity field, in a plane perpendicular to the cylinder axis. A
Benard-von Karman vortex street is clearly visible in the wake of the cylinder,
as in the two-dimensional case. The vortices are evidenced in Figure 10.10(b)
with a gray-level encoding of the span-wise (Y) component of the vorticity, in
the same plane. The macroscopic velocity field is obtained by space-averaging
over 8 x 8 x 8 cubic cells, and the vorticity field is computed a posteriori from the
macroscopic velocity field, during the post-processing phase; the square pixels
are the cross-sections of the cubic space-averaging cells. No smoothing procedure
has been applied to the picture for the reason discussed in Section 10.5.3.
Any cross-section by a plane perpendicular to the cylinder would show the
same kind of structure: oscillating vortex streets with alternately positive and
negative vortices detaching from the obstacle. But are all these oscillating vortex
streets phase-synchronized? In other words, are the vortex rolls straight and
parallel to the cylinder, as we would naively expect? To answer this question,
we make cuts by planes parallel both to the cylinder and to the mean velocity
direction. Figures 10.1 l(a), 10.11(b) and 10.1 l(c) are three snapshots taken
respectively 10000, 16000 and 24000 time steps (46.4, 74.4 and 111.7 circulation
times) after the impulsive start of the flow. The pictures give a finite-distance
perspective view of the simulation domain, seen from a direction roughly at
Figure 10.9 Geometry of

the simulation box for
three-dimensional
incompressible flow
around a cylinder (all
dimensions in lattice
nodes).
55° from the X axis. They show a two-dimensional section in a plane tangent
to the cylinder and parallel to the stream-wise (Z) direction. In this plane, the
velocity field is represented so as to make the span-wise structure of the vortex
rolls clearly visible; the transverse (X) component of the macroscopic velocity
Figure 10.10 Three-dimensional incompressible flow around a cylinder in wind-tunnel

geometry, simulated with the FCHC-8 lattice gas model on a CRAY-2 computer. The
velocity field (a), and the span-wise component of the vorticity (b) are shown in a
cut-plane transverse to the cylinder, 16000 time steps (74.4 circulation times) after the
impulsive start of the flow.
field is represented by a gray-scale coding. The alternating dark and bright

stripes are the signatures of the vortex rolls. Without symmetry breaking, they
would be straight and parallel to the cylinder, which is obviously not the case:
they first appear wavy but roughly parallel to the cylinder (Figure 10.11 (a));
then a dislocation occurs (Figure 10.11(b)) in the vortex pattern, leading to an
oblique shedding mode (Figure 10.1 l(c)), with vortices tilted with respect to
the cylinder axis. The tilting angle cannot be arbitrary: it is 'quantified' by the
periodic boundary conditions imposed in the span-wise direction. This kind of
Figure 10.11
Three-dimensional
incompressible flow
around a cylinder in
wind-tunnel geometry,
simulated with the
FCHC-8 lattice gas model
on a CRAY-2 computer.
The X component of the
velocity is displayed by
gray levels in a plane
tangent to the cylinder
and parallel to the
stream-wise (Z) direction.
(a) 10000 time steps (46.4
circulation times) after
the impulsive start; a
sequence of wavy vortex
rolls is visible in the wake.
(b) 16000 time steps (74.4
the impulsive start; a
dislocation appears in the
vortex pattern, (c) 24000
time steps (111.7
the impulsive start; the
dislocation has
disappeared and the
vortex rolls are now tilted.
V
oblique vortex shedding is frequently reported in the literature on experimental

and theoretical work on cylinder wakes.
10.5.5 Two-dimensional flow of a two-phase fluid in a

porous medium
Fluid flows with moving interfaces, especially in complex geometries (such as
a porous medium), are particularly challenging problems in computational and
theoretical fluid dynamics (see for instance Homsy, 1987). They are also clearly
important for basic and applied research in geology and soil mechanics.
The problem of two-dimensional flow of a two-phase fluid in a porous medium
has been efficiently and successfully addressed by lattice gas methods (see for
example Lutsko, Boon and Somers, 1992). The model is a variant of the Rothman
and Keller model (see Rothman and Keller, 1988) implemented on a network
of 400 Transputers. The Rothman and Keller model is an extension of the two-
species FHP model, modified to include surface tension. Incorporating surface
tension effects a priori requires non-local interactions between particles (see
Rothman and Zaleski, 1997). In the implementation used by Lutsko, Boon and
Somers, the surface tension effect is obtained by transporting the information
about the locality of the fluid phases, along the channels during the propagation
phase, so that the collision phase remains local.
Here the simulation runs on a lattice domain of 2000 x 800 nodes, where the
porous medium effect is simulated byfixedscattering nodes distributed randomly
over the simulation domain (here 1% of the lattice nodes are occupied by fixed
scatterers). The viscosity ratio between the two fluids is ~ 10, and the surface
tension coefficient is 0.28. The noise reduction of the data uses space-averaging
over 10 x 10 cells, and time-averaging over 2500 time steps. Figure 10.12 shows
the destabilization of the initially flat interface when the low viscosity (black)
fluid is pushed through the medium filled with the highly viscous (grey) fluid:
viscous fingering occurs spontaneously, like in the Saffman-Taylor instability.
In natural situations and in laboratory experiments, it is difficult to access
the porous medium in order to observe the destabilization of the interface and
to analyze its structure whose details are therefore not well known. In the
numerical experiments described above, Lutsko, Boon and Somers exploited the
lattice gas simulation data to show that the destabilized interface between the
two fluids exhibits a multi-fractal structure (Lutsko, Boon and Somers, 1992), a
result which should be of interest for further experimental investigations.
Figure 10.12
Two-dimensional lattice
gas simulation of a
two-phase fluid with
surface tension in a
porous medium. The
interface is initially flat;
an instability develops
and produces the typical
fingering patterns of the
less viscous fluid (in
black) penetrating the
most viscous fluid (in
gray). The four snapshots
are taken after 30000 (a),
50000 (b), 100000 (c),
and 112 500 (d) time
steps.
Chapter 11
Guide for further reading
Since 1985, lattice gas automata have become a widely and actively explored
field. Academic groups and industrial laboratories around the world have in-
vested considerable effort in their research activities to drive the subject in
various new and promising directions. As a result, the literature on lattice gases
and related topics has grown so rapidly and has become so voluminous that
an exhaustive list and a detailed review of all relevant lattice gas publications
would practically make a book by itself.1
Here we select research areas where lattice gases have played an important
role, and, for each of them, we quote articles considered as representative,
historically or presently. Our review is by no means complete and our choices
are certainly selective; unavoidably we haven't done justice to those whose work
escaped our selection. Nevertheless we have attempted to cover a range of works
expanding over various aspects of the subject in the available literature. Our
goal will be reached if the reader finds this chapter a helpful tool for exploring
the subject beyond the scope of this book.
1
Systematic explorations of the literature were performed by G.D. Doolen and by
D. d'Humieres from 1986 to 1995; alphabetical lists of references (often with abstracts)
can be found in special issues of Physica D, 47, pp. 299-337, 1991 (for the period
1986-1990), J. Stat. Phys., 68, pp. 611-669, 1992 (for the period 1990-1992), and Fields
Inst. Comm., 6, pp. 275-346, 1996 (for the period 1992-1995).
233
234 11 Guide for further reading
n.i The historical 'roots'

nil Discrete kinetic theory
In the early sixties, the problem of shock waves in dilute gases was a subject of
increasing research effort, for fundamental as well as industrial reasons. Because
a shock wave - for example in supersonic flow - may have a typical thickness
of the same order as the mean free path of the particles in the dilute gas,
the problem cannot be treated with continuous fluid mechanics (as described
by the Navier-Stokes equations) which is valid in the 'hydrodynamic limit',
that is, when the smallest excited scale (macroscopic scale) is much larger than
the mean free path (microscopic scale). A logical approach then is gas kinetic
theory, and in particular the Boltzmann equation. Now solving an integro-
differential equation such as the Boltzmann equation is far from a trivial matter.
To circumvent the problem, 'discrete kinetic models' were proposed; in these
models, the dilute gas is modeled as a set of particles with continuous position
space, but with discrete velocity space: the particles are assumed to have only a
finite number of velocities. The resulting discrete Boltzmann equations are no
longer integro-differential, but partial differential equations instead. This type of
discrete kinetic theory can be seen as the 'ancestor' of the lattice gas approach.
In fact the idea goes even further back in time: in the mid-nineteenth century,
J.C. Maxwell proposed a model gas based on a similar idea. Discrete kinetic
theories produced a vast literature (mostly since 1960); the following list gives
some references.2
: Maxwell, J.C, Scientific Papers II, Cambridge University Press,

Cambridge (UK), 1890, p. 26.
-: Broad well, J.E., 'Shock structure in a simple discrete velocity gas', Phys.
Fluids, 7(8), p. 1243, 1964.
-; Broadwell, J.E., 'Study of rarefied shear flow by the discrete velocity
method', J. Fluid Mech., 19(3), p. 401, 1964.
-: Gatignol, R., 'Etude de la structure d'une onde de choc pour un gaz a
repartition discrete de vitesses, C. R. Acad. Sci. Paris, 261, p. 2841, 1965.
-: Gatignol, R., 'Theorie Cinetique des Gaz a Repartition Discrete de
Vitesse', in Lecture Notes in Physics, 36, Springer-Verlag, Berlin, 1975.
; Longo, E., Monaco, R. and Platkowski, T, 'Sound propagation and shock
wave structure by the semi-discrete Boltzmann equation', J. Mech. Theo.
AppL, 1, pp. 233-243, 1988.
-; Cabannes, H., 'On the initial-value problem in discrete kinetic theory', Fur.
J. Mech. B, 10, pp. 207-224, 1991.
2
The references are organized by chronological order in each sub-section.
11.1 The historical'roots' 235
11.1.2 The early days

Discrete kinetic theory describes model gases with discrete velocity space, but
where time and position space are continuous. The idea came about, in the early
seventies, to further simplify these models by restricting particle displacements
to the geometry of a regular lattice where the particles move from node to node
at regular time intervals. The concept of the lattice gas was born. The earliest
model, residing on a two-dimensional square lattice, was proposed as a simple
model for statistical mechanics. But it was not until ten years later that the
idea arose to use lattice gases for numerical simulations of fluid dynamics. A
quite possible reason for this delay is that computer science and technology
were not sufficiently developed for fluid dynamical applications of lattice gases
in the early seventies. By 1985 the concept of massively parallel computers had
spread in the scientific community, which technically paved the way for lattice
gas simulations.
The following short list presents some basic references on the first lattice
gas model (HPP), and the first 'isotropically correct' model (FHP) for fluid
simulation.3
-: Hardy, J. and Pomeau, Y., 'Thermodynamics and hydrodynamics for a

modeled fluid', J. Math. Phys., 13, pp. 1042-1051, 1972.
: Hardy, J., Pomeau, Y. and de Pazzis, O., 'Time evolution of a
two-dimensional model system: Invariant states and time correlation
functions', J. Math. Phys., 14, pp. 1746-1759, 1973.
-: Hardy, J. and Pomeau, Y, 'Microscopic model for viscous flow in two
dimensions', Phys. Rev. A, 16, pp. 720-726, 1977.
Frisch, U., Hasslacher, B. and Pomeau, Y, 'Lattice gas automata for the
Navier-Stokes equation, Phys. Rev. Lett, 56, pp. 1505-1508, 1986.
11.1.3 Cellular automata

Lattice gas automata can always be cast in such a way that they constitute a
sub-class of cellular automata (see the discussion in the Preface). The reader
interested in the general class of cellular automata, as well as in some aspects
of basic computer science, should refer to:
; von Neumann, J., Theory of Self-Reproducing Automata, University of

Illinois Press, 1966.
• Hopcroft, J.E. and Ullman, J.D., Introduction to Automata Theory,
Languages and Computation, Addison-Wesley, New-York, 1979.
3
These models are discussed in detail in Chapter 3.
-: Wolfram, S., Theory and Applications of Cellular Automata, World

Scientific, Singapore, 1986.
11.2 Three-dimensional models
The promising results of the lattice gas method for fluid dynamics simulations
in two dimensions stimulated the search for three-dimensional models. Unfor-
tunately, the 'miracle' that renders the macroscopic behavior of FHP models
isotropic does not happen so naturally in three dimensions. Several solutions
were proposed, including the 'FCHC models' that reside on a four-dimensional
lattice (see Section 3.7.3). But these models have at least 24 channels per node,
and the design of collision rules soon becomes a subtle optimization puzzle.
Indeed, the collision rules must satisfy the correct physical properties (conserva-
tion laws, G-invariance, possibly semi-detailed balance, etc.), and in addition -
for high Reynolds number flow - they should yield the smallest possible value
for the kinematic viscosity. Moreover, they must be numerically efficient, and
shouldn't of course exceed the computer's technical limitations. These questions
are discussed in detail in the following papers where the interested reader will
also find elaborations on collision rules strategy.
& d'Humieres, D., Lallemand, P. and Frisch, U., 'Lattice gas models for 3-D
hydrodynamics', Europhys. Lett, 2, pp. 291-297, 1986.
& Henon, M., Isometric collision rules for the 4-D FCHC lattice gas',
Complex Systems, 1, pp. 475-494, 1987.
¥> Henon, M., 'Optimization of collision rules in the FCHC Lattice Gas and
addition of rest particles', in the proceedings of the workshop (Discrete
Kinetic Theory, Lattice Gas Dynamics and Foundations of Hydrodynamics',
Torino (Italy), 20-24 September 1988, ed. R. Monaco, World Scientific,
Singapore, 1989.
& Somers, J.A. and Rem, P.C., 'The construction of efficient collision tables
for fluid flow computations with cellular automata', in Cellular Automata
and Modeling of Complex Physical Systems, pp. 161-177,
ed. P. Manneville, Springer-Verlag, Berlin, 1989.
n.3 Theoretical analyses
Initially, rather than tools for computer simulations, lattice gas models were
considered as 'test-benches' for theoretical ideas and methods. So it is not
surprising that the literature on lattice gas theory is so rich, since many aspects
of the subject have been explored by various approaches. In what follows we
offer the reader our selection of references which we have organized in five
categories.
n.3.1 General lattice gas theory

This set of articles and theses is merely a sample of the basic literature on lattice
gas theory. As there is a close connection between theory and computational
issues, most of the items listed below address both aspects. Note that the articles
defining lattice gas models, listed in Section 11.1.2, also belong to the present
list, since they contain a theoretical analysis of the lattice gas models originally
introduced in these papers.
: d'Humieres, D. and Lallemand, P., 'Lattice gas automata for fluid

mechanics', Physica, 140A, pp. 337-347, 1986.
,-: Wolfram, S., 'Cellular automaton fluids 1: basic theory', J. Stat. Phys., 45,
pp. 471-525, 1986.
-; Henon, M., 'Viscosity of a lattice gas', Complex Systems, 1, pp. 762-789,
1987.
: Frisch, U., d'Humieres, D., Hasslacher, B., Lallemand, P., Pomeau, Y. and
Rivet, J.P., 'Lattice gas hydrodynamics in two and three dimensions',
-: Rivet, J.P., 'Hydrodynamique par la methode des gaz sur reseaux', Ph.D.
Thesis, Universite de Nice (France), 1988.
-; Noullez, A., 'Automates de gaz sur reseaux: aspects theoriques et
simulations', Ph.D. Thesis, Universite Libre de Bruxelles (Belgium), 1990.
11.3.2 Statistical physics and thermodynamics

As emphasized in the Preface, lattice gases are as much models for hydrodynam-
ics (describing collective fluid motions) as they are model systems for statistical
mechanics, and both aspects are still actively explored today. The standard tools
of statistical mechanics (mode coupling theory, fluctuation-dissipation methods,
projection operator techniques, etc.) are quite naturally applicable to lattice gases
(as several chapters show in this book). Fundamental as well as complementary
information on the statistical mechanical aspects of LGAs can be found in:
; Hardy, J., de Pazzis, O. and Pomeau. Y., 'Molecular dynamics of a

classical lattice gas: Transport properties and time correlation functions',
Phys. Rev. A, 13, pp. 1949-1961, 1976.
-: Rivet, J.P., 'Green-Kubo formalism for lattice gas hydrodynamics and
Monte-Carlo evaluation of shear viscosities', Complex Systems, 1(4),
pp. 838-851, 1987.
: Dufty, J.W. and Ernst, M.H., 'Hydrodynamics modes and Green-Kubo

relations for lattice gas cellular automata', J. Phys. Chem., 93, p. 7015,
1989.
; Ernst, M.H., van Velzen, G.A. and Binder, P.M., 'Breakdown of the
Boltzmann equation in cellular-automata lattice gases', Phys. Rev. A, 39,
pp. 4327-4329, 1989.
• Ernst, M.H., 'Statistical mechanics of cellular automata fluids', in Liquids,
Freezing and the Glass Transition, eds. J.P. Hansen, D. Levesque and
J. Zinn-Justin, North Holland, Amsterdam, 1991.
: Ernst, M.H. and Dufty, J.W., 'Green-Kubo relations for lattice gas cellular
automata', J. Stat. Phys., 58, p. 57, 1990.
: van der Hoef, M.A. and Frenkel, D., 'Long-time tails of the velocity
autocorrelation function in 2-D and 3-D lattice gas cellular automata: a
test of mode-coupling theory', Phys. Rev. A, 41, p. 4277, 1990.
-: Ernst, M.H., 'Mode coupling theory and tails in CA-fluids\ Physica D, 47,
p. 198, 1991.
; Naitoh, T., Ernst, M.H., van der Hoef, M.A. and Frenkel, D., 'Extended
mode coupling and simulations in cellular automata fluids', Phys Rev. A, 44,
p. 2484, 1991.
• Bussemaker, H.J. and Ernst, M.H., 'Biased lattice gases with correlated
equilibrium states', J. Stat. Phys., 68, pp. 431^55, 1992.
• Brito, R., Bussemaker, H.J. and Ernst, M.H., A fluctuation formula for the
non-Galilean factor in lattice gas automata', J. Phys. A, 25,
pp. L949-L954, 1992.
• Grosfils, P., Boon, J.R, Brito, R. and Ernst, M.H., 'Statistical
hydrodynamics of lattice gas automata', Phys. Rev. E, 48(4),
pp. 2655-2668, 1993.
• Boghosian, B.M. and Taylor, W., Correlations and Renormalization in
Lattice Gases, M.I.T. internal report no. MIT-CTP-2265, April 1994.
• Grosfils, P., 'Hydrodynamique statistique des gaz sur reseaux, Ph.D.
Thesis, Universite Libre de Bruxelles (Belgium), 1994.
• Suarez, A., Boon, J.P. and Grosfils, P., 'Long-range correlations in
nonequilibrium systems: Lattice gas automaton approach', Phys. Rev. E,
54(2), pp. 1208-1224, 1996.
• Suarez, A. and Boon, J.R, 'Non-linear hydrodynamics of lattice gases with
semi-detailed balance', Int. J. Mod. Phys. C, 8(4), pp. 653-674, 1997.
• Suarez, A. and Boon, J.R, 'Nonlinear lattice gas hydrodynamics', J. Stat.
Phys., 87(5/6), pp. 1123-1130, 1997.
: Hanon, D. and Boon, J.P., 'Diffusion and correlations in lattice-gas

automata', Phys. Rev. E, 56(6), pp. 6331-6339, 1997.
There have been several attempts to extend lattice gas automata to the
quantum domain. Models have been proposed for the Dirac equation and for
the Schrodinger equation. Also motivated by the observation that a natural
architecture for nanoscale quantum computation is that of a quantum cellular
automaton, authors have developed theoretical analyses and algorithmic schemes
for quantum lattice gas automata. Here are some references on a subject which,
although still in its infancy, looks promising.
-; Succi, S. 'Numerical solution of the Schrodinger equation with discrete

kinetic theory, Phys. Rev. E, 53(2), pp. 1969-1975, 1996.
-: Meyer, D. 'From quantum cellular automata to quantum lattice gases',
J. Stat. Phys., 85, pp. 551-574, 1996.
: Boghosian, B.M., and Taylor, W. 'A quantum lattice gas model for the
many-particle Schrodinger equation', Phys. Rev. E, 56, pp. 705-716, 1997.
-: Meyer, D. 'Quantum lattice gases and their invariants', Int. J. Mod. Phys.
C, 8(4), pp. 717-725, 1997.
: Boghosian, B.M., and Taylor, W. 'Simulating quantum mechanics on a
quantum computer' Physica D, 120, pp. 30-42, 1998.
-: Meyer, D. 'Quantum mechanics of lattice gas automata', J. Phys. A, 31,
pp. 2321-2340, 1998.
11.3.3 Violation of semi-detailed balance

AH theoretical developments in Chapters 4 to 8 are based on the semi-detailed
balance assumption (see Section 2.3.1): the collision rules are 'democratic' with
respect to all possible microscopic states, so that if initially all microscopic states
are equally probable, they remain so after collision. Under this assumption, the
value of the kinematic viscosity coefficient (evaluated within the Boltzmann
approximation) has a lower bound that may severely limit the achievable upper
value of the Reynolds number (for a given computational effort). This lower
bound disappears if the semi-detailed balance constraint is relaxed, but then
there is a price to pay on the theoretical side: the Boltzmann hypothesis must
be imposed ab initio. The reader interested in this aspect is referred to:
• Dubrulle, B., 'Method of computation of the Reynolds number for the two
models of lattice gas involving violation of semi-detailed balance', Complex
Systems, 2, pp. 577-609, 1988.
* Dubrulle, B., Frisch, U., Henon, M. and Rivet, J.R, 'Low-viscosity lattice
gases', J. Stat. Phys., 59, pp. 1187-1226, 1990.
: Ernst, M.H. and Bussemaker, H.J., 'Algebraic spatial correlations in lattice

gas automata violating detailed balance', J. Stat. Phys., 81(1/2),
pp. 515-536, 1995.
• Bussemaker, H.J., 'Pattern formation and correlations in lattice gas
automata, Ph.D. Thesis, Universiteit Utrecht (The Netherlands), 1995.
: Tribel, O. and Boon, J.P., 'Entropy and correlations in lattice gas automata
without detailed balance', Int. J. Mod. Phys. C, 8(4), pp. 641-652, 1997.
11.3.4 Invariants and conservation laws

The following articles treat the problem of identifying and classifying in a
systematic way the various kinds of invariants in lattice gases (see Section 2.3.3).
• Pomeau, Y., Invariants in cellular automata', J. Phys. A, 17,

pp. L415-L418, 1984.
• d'Humieres, D., Qian, Y.H. and Lallemand, P., 'Finding the linear
invariants of lattice gases', in the proceedings of the workshop
Computational Physics and Cellular Automata, Ouro Preto (Brazil), 8-11
August 1989.
• Bernardin, D., 'Global invariants and equilibrium states in lattice gases',
J. Stat. Phys., 68(3/4), pp. 457-496, 1992.
11.3.5 Obstacles and Knudsen layers

There is a wide variety of problems in hydrodynamics where flows are studied in
the presence of obstacles and boundaries with simple or complicated shapes. It
is an advantage of the lattice gas approach to offer convenient implementation
of solid bodies with non-trivial geometries, by using 'special collision rules' on
the boundary nodes defining the shape of the solid body (see Section 2.4.1).
However, a careful analysis based on the Boltzmann equation shows that the
location where the macroscopic mean velocity actually vanishes is not exactly on
the surface of the obstacle, but rather a fraction of a lattice mesh size outside.
Furthermore, the value of this shift may depend on the orientation of the surface
of the obstacle. This is a delicate matter that must be taken into account for
accurate simulations, in particular for obstacles with moderate size (compared
to the full size of the automaton). This rather technical point is discussed in:
• Lavallee, P., Boon, J.P. and Noullez, A., 'Boundaries in lattice gas flows',
Physica D, 47, pp. 233-240, 1991.
• Cornubert, R., d'Humieres, D. and Levermore, D., A Knudsen layer
theory for lattice gases', Physica D, 47, pp. 241-259, 1991.
11.4 Models with particular features 241
Ginzbourg, I. and d'Humieres, D., 'Local second-order boundary method

for lattice Boltzmann models', J. Stat. Phys., 84(5/6), pp. 927-971, 1996.
Ginzbourg, I. and Adler, P.M., 'Boundary flow conditions analysis for the
three-dimensional lattice Boltzmann model', J. Phys. II, 4, pp. 191-214,
1994.
11.4 Models with particular features

The success of the lattice gas method for the study of macroscopic motions
of athermal fluids stimulated efforts to incorporate further physical features to
describe systems such as multi-component fluids, reactive processes, reaction-
diffusion systems, surface tension and free interfaces, magnetic fields, as well
as acoustic phonons and elastic seismic waves. Examples can be found in the
following categories:
n.4.1 Fluid mixtures and colloids

-; Burges, C. and Zaleski, S., 'Buoyant mixtures of cellular automaton gases',
; Chen, H. and Matthaeus, W.H., 'Cellular automaton formulation of passive
scalar dynamics', Phys. Fluids., 30, pp. 1235-1237, 1987.
Ladd, A.J.C., Colvin, M.E. and Frenkel, D., 'Application of lattice-gas
cellular automata to the Brownian motion of solids in suspension', Phys.
Rev. Lett., 60, p. 975, 1988.
Bernardin, D. and Sero-Guillaume, O.E., 'Lattice gases mixtures models
for mass diffusion, Eur. J. Mech. B, 9, p. 21, 1990.
van der Hoef, M.A. and Frenkel, D., 'Tagged particle diffusion in 3-D
lattice gas cellular automata', Physica D, 47, p. 191, 1991.
• van der Hoef, M.A., 'Simulation study of diffusion in lattice gas fluids and
colloids', Ph.D. Thesis, FOM-Instituut voor Atoom- en Molecuulfysica,
Amsterdam (The Netherlands), December 1992.
11.4.2 Reaction-diffusion systems

; Clavin, P., Lallemand, P., Pomeau, Y. and Searby, G., 'Simulation of free
boundaries inflow systems by lattice gas models', J. Fluid Mech., 188,
pp. 437^64, 1988.
M Dab, D. and Boon, J.P., 'Cellular automata approach to reaction-diffusion
systems', pp. 257-273, in Cellular Automata and Modeling of Complex
Physical Systems, eds. P. Manneville, N. Boccara, G.Y. Vichniac and

R. Bidaux, Springer-Verlag, Berlin, 1989.
Sero-Guillaume, O.E. and Bernardin, D., 'A lattice gas model for heat
transfer and chemical reaction', Eur. J. Meek B, 9(2), pp. 177-196, 1990.
; Dab, D., Lawniczak, A., Boon, J.P. and Kapral, R., 'Cellular automaton
model for reactive systems', Phys. Rev. Lett, 64(20), pp. 2462-2465, 1990.
; Boon, J.P., Dab, D., Kapral, R. and Lawniczak, A., 'Lattice gas automata
for reactive systems', Phys. Rep., 273, pp. 556-647, 1996.
11.4.3 Immiscible fluids and free interfaces

Clavin, P., d'Humieres, D., Lallemand, P. and Pomeau, Y., 'Automates
cellulaires pour les problemes a frontieres libres en hydrodynamique a 2 et
3 dimensions', C. R. Acad. Sci. Paris II, 303, pp. 1169-1174, 1986.
• d'Humieres, D., Lallemand, P. and Searby, G., 'Dynamics of
two-dimensional bubbles by the lattice gas method', Complex Systems, 1,
pp. 333-350, 1987.
; Clavin, P., Lallemand, P., Pomeau, Y. and Searby, G., 'Simulation of free
boundaries inflow systems by lattice gas models', J. Fluid Mech., 188,
pp. 437-464, 1988.
• Rothman, D.H. and Keller, J.M., Immiscible cellular automaton fluids',
J. Stat. Phys., 52, pp. 1119-1127, 1988.
: Rothman, D.H. and Zaleski, S., 'Spinodal decomposition in a lattice-gas
automaton', J. Phys. Paris, 50, pp. 2161-2174, 1989.
; Somers, J.A. and Rem, P C , 'Analysis of surface tension in two-phase
lattice gases', Physica D, 47, pp. 39-46, 1991.
Gunstensen, A.K. and Rothman, D.H., A Galilean-invariant immiscible
lattice gas', Physica D, 47(1/2), pp. 53-63, 1991.
Appert, C, Rothman, D.H. and Zaleski, S. A liquid-gas model on a
lattice', Physica D, 47(1/2), pp. 85-96, 1991.
; Rothman, D.H. and Zaleski, S., 'Lattice-gas models of phase separation:
interfaces, phase transitions, and multiphase flow', Rev. Mod. Phys., 66(4),
pp. 1417-1479, 1994.
. Adler, C , d'Humieres, D. and Rothman, D.H., 'Surface tension and
interface fluctuations in immiscible lattice gases', J. Phys. I, 4, pp. 29-46,
1994.
Olson, J.F. and Rothman, D.H., 'Three-dimensional immiscible lattice gas:
application to sheared phase separation', J. Stat. Phys., 81(1/2),
pp. 199-222, 1995.
11.4 Models with particular features 243
: Flekkoy, E.G. and Rothman, D.H., 'Fluctuating fluid interfaces', Phys.

Rev. Lett, 75(2), pp. 260-263, 1995.
11.4.4 Flow in porous media

-: Rothman, D.H., 'Cellular automaton fluids: a model for flow in porous
media, Geophysics, 53, pp. 509-518, 1988.
-; Rothman, D.H., 'Macroscopic laws for immiscible two-phase flow in porous
media: results from numerical experiments', J. Geophys. Res., 95, pp.
8663-8674, 1990.
-: Kohring, G. A., 'Calculation of the permeability of porous media using
hydrodynamic cellular automata', J. Stat. Phys., 63(1/2), pp. 411^18,
1991.
; Chen, S., Diemer, K., Doolen, G., Eggert, K., Fu, G, Gutman, S. and
Travis, B.J., 'Lattice gas automata for flow through porous media', Physica
D, 47(1/2), pp. 72-84, 1991.
• Lutsko, J.L., Boon, J.R and Somers, J.A., 'Lattice gas automata
simulations of viscous fingering in porous media,', Lecture Notes in Physics,
398, pp. 124-135, Springer-Verlag (Berlin), 1992.
11.4.5 Thermo-hydrodynamics
-: Chopard, B. and Droz, M., 'Cellular automata model for heat conduction
in fluid', Phys. Rev. Lett. A, 126, pp. 476-480, 1988.
; Luo, L.S., Chen, H., Chen, S., Doolen, G.D. and Lee, Y.C., 'Generalized
hydrodynamic transport in lattice-gas automata', Phys. Rev. A, 43,
pp. 7097-7100, 1991.
: Chen, S., Chen, H., Doolen, G.D., Gutman, S. and Lee, Y.C., A lattice
gas model for thermodynamics', J. Stat. Phys., 62, p. 1121, 1991.
: Ernst, M.H. and Das, S.R, 'Thermal cellular automata', J. Stat. Phys., 66,
pp. 465-483, 1991.
: Grosfils, R, Boon, J.R and Lallemand, R, 'Spontaneous fluctuation
correlations in thermal lattice gas automata', Phys. Rev. Lett., 68,
pp. 1077-1080, 1992.
: Das, S.R and Ernst, M.H., 'Thermal transport properties in a square
lattice gas', Physica A, 187, pp. '191-209, 1992.
11.4.6 Elastic waves

; Rothman, D.H., 'Modeling seismic P-waves with cellular automata',
Geophys. Res. Lett., 14(1), pp. 17-20, 1987.
-; Mora, P., 'The lattice Boltzmann phononic lattice solid', J. Stat. Phys.,
68(3/4), pp. 591-610, 1992.
-: Maillot, B., 'Modeles semi-microscopiques d'ondes elastiques, Ph.D. Thesis,
Institut de Physique du Globe, Paris (France), June 1994.
11.4.7 Other models

-; Hatori, T. and Montgomery, D.C., 'Transport coefficients for
magneto-hydrodynamics cellular automata, Complex Systems, 1,
pp. 735-752, 1987.
-; Montgomery, D.C. and Doolen, G.D., 'Two cellular automata for plasma
computations', Complex Systems, 1, pp. 831-838, 1987.
-: Chen, H. and Matthaeus, W.H., 'New cellular automaton model for
magnetohydrodynamics, Phys. Rev. Lett, 58, pp. 1845-1848, 1987.
-: Markus, M. and Hess, B., Isotropic cellular automaton for modeling
excitable media', Nature, 347, pp. 56-58, 1990.
: Chen, S., Chen, H., Doolen, G.D., Lee, Y.C. and Brand, H., 'Lattice gas
models for non-ideal gas fluids', Physica D, 47, p. 97, 1991.
-: Eloranta, K. 'Cellular automata for contour dynamics', Physica D, 89(1/2),
pp. 184-203, 1995.
11.5 Lattice Boltzmann method

We have stressed the fact that, because of the Boolean nature of the lattice
gas variables combined with the (usually) stochastic collision rules, LGAs pos-
sess intrinsic fluctuations, and we have discussed in detail the insight that can
be gained from the analysis of these fluctuations and their correlations (see
Chapter 7). Now when we are interested in macroscopic observables, we must
perform time- and/or space- and/or ensemble-averaging of microscopic vari-
ables. The macroscopic quantities so obtained are 'noisy' because averaging over
a finite sample, while reducing the microscopic fluctuations, does not eliminate
them completely. So we face a trade-off between the size of the averaging sam-
ples, the desired space- and time-resolution for the macroscopic variables, and
the available size of computer memory. In the early nineties, the idea emerged to
implement directly the lattice Boltzmann equation (see Section 4.2.2) rather than
the microdynamic equations of the lattice gas (see Section 2.2). Indeed, while
the microdynamic equations are for Boolean variables that need averaging, the
lattice Boltzmann equation governs the time-evolution of the mean populations
ft, which are 'floating-point' noiseless variables, and can be simulated directly
to obtain average quantities. So the lattice Boltzmann method can be seen as
11.5 Lattice Boltzmann method 245
a finite-difference scheme for fluid dynamics numerical simulations. The price

to pay is some loss of robustness: a lattice Boltzmann simulation can 'blow up'
when the simulation parameters are chosen too close to their limits (which is
not the case for lattice gas simulations).
In most practical cases, it is not the full lattice Boltzmann equation which is
used, but rather its linearized version. The linearization is performed around an
equilibrium solution, so that the collision operator (in the r.h.s. of Equation (4.9))
reduces to a linear operator (the 'linearized collision operator'), which acts on
the first order perturbations to the mean populations.
Basic references to the lattice Boltzmann method as well as several applications
are found in:
-: Succi, S., Benzi, R. and Higuera, F. 'The lattice Boltzmann equation: A

new tool for computational fluid dynamics', Physica D, 47(1/2),
pp. 219-230, 1991.
: Gunstensen, A.K., Rothman, D.H., Zaleski, S. and Zanetti, G., 'Lattice
Boltzmann model of immiscible fluids.', Phys. Rev. A, 43(8), pp. 4320-4327,
1991.
-: Benzi, R., Succi, S. and Vergassola, M., 'The lattice Boltzmann equation:
theory and applications', Phys. Rep., 222(3), pp. 145-197, 1992.
; Alexander, F.J., Chen, H., Chen, S. and Doolen, G.D., A lattice Boltzmann
model for compressible fluids', Phys. Rev. A, 46(4), pp. 1967-1970, 1992.
: Chen, S., Wang, Z., Shan, X. and Doolen, G.D., 'Lattice Boltzmann
computational fluid dynamics in three-dimension', J. Stat. Phys., 68(3/4),
pp. 379-400, 1992.
-; Mora, R, 'The lattice Boltzmann phononic lattice solid', J. Stat. Phys.,
68(3/4), pp. 591-610, 1992.
-; Somers, J.A., 'Direct simulation of fluid flow with cellular automata and the
lattice Boltzmann equation', Appl. Sci. Res., 51, pp. 127-133, 1993.
; Martinez, D.O., Matthaeus, W.H., Chen, S. and Montgomery, D.C.,
'Comparison of spectral method and lattice Boltzmann simulations of
two-dimensional hydrodynamics', Phys. Fluids, 6(3), pp. 1285-1298, 1994.
-: Wagner, L., 'Dependence of drag on a Galilean invariance-breaking
parameter in lattice Boltzmann flow simulations', Phys. Rev. E, 49(3),
pp. 2115-2118, 1994.
Oxaal, U., Flekkoy, E.G. and Feder, J., 'Irreversible dispersion at a
stagnation point: Experiments and lattice Boltzmann simulations', Phys
Rev Lett., 72(22), pp. 3514-3517, 1994.
: Ginzbourg, I. and d'Humieres, D., 'Local second-order boundary method
for lattice Boltzmann models', J. Stat. Phys., 84(5/6), pp. 927-971, 1996.
11.6 Lattice Bhatnagar-Gross-Krook model
The linearized Boltzmann collision operator (see Section 11.5) can be simplified
by reducing its matrix to a diagonal form with one single non-zero eigenvalue.
This method proposed initially in 1954 by Bhatnagar, Gross and Krook amounts
to replacing the full spectrum of eigenvalues by one single relaxation time. The
method applies straightforwardly to the lattice Boltzmann scheme and is quite
useful in particular for tuning, within certain limits, the kinematic viscosity to
lower values than can be achieved by the LGA method. The procedure known
as the lattice BGK method (LBGK) is discussed from the point of the theory
as well as for various applications in:
: Bhatnagar, P., Gross, E.P. and Krook, M.K., A model for collision
processes in gases', Phys. Rev., 94, pp. 511-525, 1954.
: Qian, Y.H., d'Humieres, D. and Lallemand, P., 'Lattice BGK models for
Navier-Stokes equation', Europhys. Lett., 17, pp. 479-484, 1992.
-: Qian, Y.H. and Orszag, S.A., 'Non-linear correction to Navier-Stokes
equation derived from lattice BGK models', Europhys. Lett., 21,
pp. 255-259, 1992.
: Succi, S., d'Humieres, D., Qian, Y.H. and Orszag, S.A., 'On the small-scale
dynamical behavior of lattice BGK and lattice Boltzmann schemes', J. Sci.
Comp., 8, pp. 219-230, 1993.
u Qian, Y.H., 'Simulating thermohydrodynamics with lattice BGK models',
J. Sci. Comp., 8, pp. 231-242, 1993.
• Flekkoy, E.G., 'Lattice Bhatnagar-Gross-Krook models for miscible
fluids', Phys. Rev. E, 47(6), pp. 4247-4257, 1993.
n.7 Numerical simulations and implementations
Because of the binary and discrete nature of cellular automata, and because
of the global updating of the LGA, lattice gases are very good candidates
for easy and efficient implementation on general purpose massively parallel
(super)computers, as well as on dedicated computers. The next two sub-sections
give references to articles on dedicated hardware implementations, and general
purpose computer implementations and simulations.
n.7.1 Implementation on dedicated hardware

• Margolus, N., Toffoli, T. and Vichniac, G., 'Cellular-Automata
Supercomputers for Fluid Dynamics Modeling', Phys. Rev. Lett., 56,
pp. 1694-1696, 1986.
11.7 Numerical simulations and implementations 247
: Toffoli, T. and Margolus, N., Cellular Automata Machines: a New

Environment for Modeling, M.I.T. Press, Cambridge (MA), 1986.
,: Cloucqueur, A. and d'Humieres, D., 'RAP-1, a cellular automaton machine
for fluid dynamics', Complex Systems, 1, pp. 584-597, 1987.
: Cloucqueur, A. and d'Humieres, D., 'RAP, a family of cellular automaton
machines for fluid dynamics', Helv. Phys. Acta, 62, pp. 525-541, 1989.
; Margolus, N. and Toffoli, T., 'Cellular automata machines', pp. 219-249, in
Lattice Gas Methods for Partial Differential Equations, Addison-Wesley,
1990.
; Adler, C , Boghosian, B.M., Flekkoy, E., Margolus, N. and Rothman,
D.H., 'Simulating three-dimensional hydrodynamics on a cellular-automata
machine', J. Stat Phys., 81(1/2), pp. 105-128, 1995.
11.7.2 Simulations on general purpose computers

: d'Humieres, D., Lallemand, R and Shimomura, T., Cellular Automata, a
New Tool for Hydrodynamics, Los Alamos National Laboratory, Report
LA-UR-85-4051, 1985.
; d'Humieres, D., Pomeau, Y. and Lallemand, R, 'Simulation d'allees de
Von Karman 2-D a Vaide d'un gaz sur reseau', C. R. Acad. Sci. Paris II,
301, pp. 1391-1394, 1985.
; d'Humieres, D. and Lallemand, R, 'Ecoulement d'un gaz sur un reseau
dans un canal bidimensionnel: developpement du profile de Poiseuille', C. R.
Acad. Sci. Paris II, 302, pp. 983-988, 1986.
d'Humieres, D. and Lallemand, R, 'Numerical simulations of
hydrodynamics with lattice gas automata in two dimensions', Complex
Systems, 1(4), pp. 598-631, 1987.
• d'Humieres, D., Lallemand, R and Searby, G., 'Numerical experiments on
lattice gases mixtures and Galilean invariance, Complex Systems, 1(4),
pp. 632-647, 1987.
-: Rivet, J.R, 'Simulation d'ecoulements tri-dimensionels par la methode des
gaz sur reseaux: premiers resultats, C. R. Acad. Sci. Paris II, 305,
pp. 751-756, 1987.
; Rivet, J.R, Henon, M., Frisch, U. and d'Humieres, D., 'Simulating fully
three-dimensional external flow by lattice gas methods', Europhys. Lett., 1,
pp. 231-236, 1988.
: Kadanoff, L.R, McNamara, G.R. and Zanetti, G., 'From automata to fluid
flow: comparisons of simulations and theory', Phys. Rev. A, 40,
pp. 4527^541, 1989.
-: Somers, J.A. and Rem, P.C., 'Flow computation with lattice gases', Appl.
Sci. Res., 48, pp. 391-435, 1991.
v Henon, M. and Scholl H., 'Lattice-gas simulation of a nontransverse
large-scale instability for a modified Kolmogorov flow', Phys. Rev. A, 43,
pp. 5365-5366, 1991.
-; Somers, J.A. and Rem, P.C., 'Obtaining numerical results from the 3D
FCHC-lattice gas', Lecture Notes in Physics, 398, pp. 59-78,
Springer-Verlag, Berlin, 1992.
-; Henon, M., 'Implementation of the FCHC lattice gas model on the
Connection Machine', J. Stat. Phys., 68(3/4), pp. 353-378, 1992.
-: Rivet, J.P., 'Spontaneous symmetry breaking in the 3-D wake of a long
cylinder simulated by the lattice gas method and drag coefficient
measurements', Appl. Sci. Res., 51, pp. 123-126, 1993.
11.8 Books and review articles
We close with a list of books and proceedings volumes published since 1986;
we also include general review articles on lattice gases and related topics.
; Wolfram, S. (editor), Collection of contributions to the workshop on
Large Non-Linear Systems, Santa Fe, (NM, USA), 27-29 October 1986;
Published as a special issue of Complex Systems, 1(4), 1987.
-: Frisch, U., 'Une nouvelle strategie pour Vhydrodynamique: les reseaux
d'automates, J. Astr. Fr., 32, pp. 17-20, 1988.
-: Boon, J.P., 'Physique, complexity et automates', Bull. Soc. Fr. Phys.,
October 1989.
; Monaco, R. (editor), Proceedings of the workshop on 'Discrete Kinetic
Theory, Lattice Gas Dynamics and Foundations of Hydrodynamics, Torino
(Italy), 20-24 September 1988, World-Scientific, Singapore, 1989.
-; Doolen, G.D. (editor), Lattice Gas Methods for Partial Differential
Equations, Addison-Wesley, 1990.
-; Boon, J.R, 'Lattice gas automata: a new approach to the simulation of
complex flows', in Microscopic Simulations of Complex Flows, ed.
M. Mareschal, Plenum Press, New York, 1990.
• Doolen, G.D. (editor), Lattice Gas Methods for Partial Differential
Equations: Theory, Applications and Hardware, special issue of Physica D,
47(1/2), 1991. (Proceedings of the NATO Advanced Research Workshop
on Lattice Gas Methods for Partial Differential Equations: Theory,
Applications and Hardware, Los Alamos National Laboratory (NM,
USA), 6-8 September 1989).
11.8 Books and review articles 249
Chen, S., Doolen, G.D. and Matthaeus, W.H., 'Lattice gas automata for
simple and complex fluids', J. Stat. Phys., 64(5/6), pp. 1133-1162, 1991.
Ernst, M.H., 'Statistical mechanics of cellular automata fluids', in the
proceedings of the NATO ASI on Liquids, Freezing and the Glass
Transition, Les Houches (France), July 3-28, 1989, eds. J.P. Hansen,
D. Levesque and J. Zinn-Justin, North Holland, Amsterdam, 1991.
Boon, J.P. (editor), special issue of J. Stat. Phys., 68(3/4), 1992.
(Proceedings of the NATO Advanced Research Workshop on Lattice Gas
Automata Theory, Implementation, and Simulation, Observatory of Nice
(France), 25-28 June 1991).
Boon, J.P., Frisch, U. and d'Humieres, D., 'L'hydrodynamique modelisee
sur reseau, La Recherche, 253, pp. 390-399, April 1993.
Rothman, D.H. and Zaleski, S., Lattice-Gas Cellular Automata,
Cambridge University Press, Cambridge (UK), 1997.
Chopard, B. and Droz, M., Cellular Automata Modeling of Physical
Systems, Cambridge University Press, Cambridge (UK), 1998.
Wolf-Gladrow, D., An Introduction to Lattice-Gas Cellular Automata and
Lattice Boltzmann Models, Springer-Verlag, Berlin, 1999.
Succi, S., The Lattice Boltzmann Equation for Fluid Dynamics and Beyond,
Oxford University Press, (to be published).
Appendix
Mathematical details
A.I A non-local collision model can be re-coded as a local collision

model (see Section 1.3.5)
Proof
Consider a non-local collision model with b channels per node. The collision
phase is such that the post-collision state n^-(r*) of any node labeled by its
position vector r* is a function J^ of the pre-collision state n,-(r* — e^) of B nodes
(the node itself plus B — 1 other nodes), with position vectors r* — e&:
rift*) = &j(ni(T*-ek)(i=it...tb;k=it...tB)>), for j = l,...,ft. (A.I)
In the most general case, the non-local collision process at a given node can
also involve the state of the node itself. So, one of the e^ s (say ei for example)
must be the null vector. In addition, the number B of 'collision vectors' e& is
finite, and the e^ s are globally invariant under the action of the point-group of
the lattice.
The propagation phase involves the set of b propagation vectors c* which
are also globally invariant under the point-group of the lattice. The final (post-
propagation) state nj(r*)(i=iv.^) of node r* is given by:
n\(x*) = n{(r* - c,), for i = 1,..., b. (A.2)
We define a new model with the same lattice structure, but with b = b x B
250
A.2 251
channels per node. As before, each node is labeled by its position vector r*, but
now each channel is labeled by a composite label ? = (i, k) with i = 1,..., b and
fc = l , . . . , B .
The collision phase of the new model is such that the post-collision state
n'fj^r*) of any node r* is given by:
)(/=i,...,6;fc=i,...,B))? for j = l,...,fo; l = l,...,B.

(A3)
The new collision process is now purely local, since the right hand side of (A. 3)
involves only the state n^k) of node r*.
The propagation phase of the new model involves the propagation vectors
c(l/c) = a + e/c, with i and k ranging from 1 to b and from 1 to B, respectively.
The final (post-propagation) state h^(u)(i=it^b;k=u...9B) of node r* is given as
usual by:
n'm(u) = n'm(u -cm), for i = 1,...,b; k = l,...,B. (A.4)
It is clear that if the initial state %/c)(r*, U = 0) of the new model is set
such that %/c)(r*, U = 0) = 7i;(r* — tÛ = 0), then, at any further discrete time
£•, the state n^k)(u,t*) will verify W(i,i)(r*,£*) = ni(r*,t*). In other words, the
exact dynamics of the initial model with non-local collisions is 'contained' in
the dynamics of the new model with purely local collisions. This shows that a
model with non-local collisions can be re-coded as a model with more channels
per node, but with local collisions. •
A.2 The so-called 'face-centered hyper-cubic' four-dimensional lat-

tice should be named '2D- face -centered body-centered hyper-
cubic' lattice (see section 3.7)
Proof
The four-dimensional FCHC lattice is the set of points of R 4 whose coordinates
are signed integers with even sum (see Section 3.7). The basic simple hyper-
cubic lattice on which the FCHC lattice is constructed is the set of all points
with coordinates (2/?i,2w2,2ft3,2ft4) where (^1,^2,^3,^4) are signed integers. Con-
sider the elementary hyper-cube limited by 16 lattice points with Cartesian
252 Appendix
coordinates:
0
0
\2j
/2\ f2\ /2\ ( \ r\
( \ ( \
2 2
P 2 2 2
P 2
2
\2/
2
\2/
'\2P ' 0 ' 2 ' 2 ' 0
\ 2 / \o/ \ 2 / \2/
2
\0/
The center of this hyper-cube is the point (1,1,1,1), which clearly belongs to
the FCHC lattice since the sum of its coordinates is even. So, the FCHC lattice
is 'body-centered'.
This hyper-cube has 8 (hyper-)faces which are obtained by setting successively
each of the 4 coordinates equal to either 0 or 2. These hyper-faces are three-
dimensional cubes. For instance, the hyper-face which verifies X4 = 0 is limited
by the 8 following lattice points:
2
0\ /2\ /o\
0 0 2
^°\0 /22\ /2\
0
/2^/o2 (A.6)
2
0 ' 0 ' 0
' p 'w 2 •2 •> 2
5
0/ \0 \o/ \o
The center of this hyper-face is the point (1,1,1,0) which does not belong to
the FCHC lattice. Thanks to the invariances of the lattice, it is easy to convince
oneself that none of the 8 hyper-faces are centered. So, the FCHC lattice is not
'face-centered', at least if 'face' is understood as three-dimensional 'hyper-face'.
The hyper-cube has 24 2D-faces which are obtained by setting successively
each pair of coordinates equal to either (0,0), (2,0), (0,2) or (2,2). For instance,
the 2D-face which verifies (^3,^4) = (0,0) is limited by the 4 following points:
/2\
0 0
(A.7)
w 0 0
\oj
The center of this 2D-face is the point (1,1,0,0) which clearly belongs to the
FCHC lattice. Using again the invariances of the FCHC lattice, it is easy to
show that all 24 2D-faces are centered. Thus the FCHC lattice is also '2D-face-
centered'. •
A.3 253
A.3 The full statistics of a collection of Boolean variables contains

more information than their individual statistics; when the
variables are uncorrelated, the knowledge of either statistics
yields the same amount of information (see Section 4.3.1)
The following proof is inspired by Fano (1961).
Proof
We compute the difference S between the quantity of information contained
in the knowledge of all the individual statistics and the knowledge of the full
statistics (see Section 4.3.1 for notation):
8 = r-i
q
1=1
with p. = (1 — p^ We now use the fact that for a Bernoulli variable xt, the
probability p\ that x\ be equal to 1 is also the averaged value of xf, namely:
and ft =
YeT YeT
We then rewrite d as:
YeT i=l YeT
Now considering the inequality In a < a — 1, which holds for any strictly positive
value of a, the difference 5 verifies the inequality:
YeT 1=1 YeT
The normalization condition for & and p imposes that

S r e r Ilf=i PiyiPi^ be equal to one, from which it follows that:
(5<0.
It is clear that d = 0 when 0>(Y) = n?=iPi J'PA t h a t is w h e n t h e

' individual
Boolean variables x, are statistically independent. •
254 Appendix
A.4 The following global equilibrium conditions:
r * e ^ i=l SeT r*e£> 7=1
and
7=1 sey
f
7=1
are strictly equivalent (see Section 4.4.1)
Proof
We first factorize the r.h.s. of (A.9) as follows: we single out a particular node
ro in i^, and denote by if (o) the set of all nodes of if except i*o. The r.h.s. of
(A.9) can then be rewritten as:
E E ^s'^mu'f/1 n Mtîiftf?' ,
sey SeT, 7=1 r*ej?(0) 7=1
S()
iS and 5' taken at node r*
which is also equal to:
sey 7=1 \ Ser, r*e^(0) 7=1

S(ro)=s N v
S and Sr taken at node r*
We now denote by F(o) the phase space of the sub-lattice if(o). The sum over
all configurations <S in T such that S(r0) = s can be replaced by a sum over S in
F(o). The above expression then becomes:
sey
S and S' taken at node r*

The expression in parentheses has exactly the same form as the r.h.s. of (A.9),
except that it concerns if (o) rather than 5£. The procedure can be repeated with
a new particular node, and so on, leading to the following form for the r.h.s. of
(A.9):
n z^^n^//7-
sey 7=1
A.5 255
Thus Equation (A.9) is equivalent to:
E, a ^sv.i)n; : ,//'//• =1_ vs,sr

•
Now consider an arbitrary Boolean state sf of the single-node phase space y.

Equation (A. 11), which holds for all Boolean configurations S', also holds for
the Boolean configuration S' for which S\u) = sr, for all r* in if. Thus (A.ll)
implies:
(A,2)
where Jf is the number of nodes in ££. The expression in parentheses is real

and non-negative; so it must be equal to one to satisfy (A. 12). This implies
(A. 10) immediately.
The reverse proof is clear: Equation (A. 10), which is true for all s' in y
guarantees that (A.ll) holds for all Sf in F, from which it follows that (A.9) is
true. •
A.5 The following equilibrium conditions:
7=1 sey 7=1
b
Y^ Wi - Si)A(s^s') J ] fffj> =0, Vi = 1,..., b, (A. 14)
s,s'ey ;=1
Z)log (T^J) ti
are strictly equivalent when the semi-detailed balance condi-
tion is satisfied (see Section 4.4.1)
Proof
• We prove that (A.13) ^> (A.14):
We use the normalization condition (2.9) to write (A.13) under the following
equivalent form:
sey 7=1 sey ;=1

256 Appendix
We multiply both sides by sj and sum over s' to deduce:
£4A(^)i[f/'7fJ = Es'As^flf/q?, vi = i,...,b.

s,s'ey 7=1 s,s'ey 7=1
(A.17)
On the l.h.s., both s and sr are dummy variables, which we can permute to
obtain:
s,s'ey 7=1 s/ey 7=1

(A.18)
Equation (A. 14) follows immediately.

Note that we have reduced a set of 2b equations to a set of b equations.
Yet we made no restrictive assumption about the lattice gas model, and the
semi-detailed balance condition (see Section 2.3.1) has not been used.
• We prove that (A.14) => (A.15):
We divide both sides of (A.14) by n l L i /fc?1 anc ^ introduce the notation ft = UIU
to obtain:
b
2 (s; - Si)A(s^) n ~fj = 0 , V i = l fc. (A.19)
s,s'ey 7=1
We now use a trick introduced by Gatignol in the context of gas models with
discrete velocities (Gatignol, 1975): we multiply both sides by log// and sum
over I Equation (A.19) becomes:
b b
X>g/,- 5>; - Si)A(s^) n/;j' =0, (A.20)

i=l s,s'ey 7=1
which also reads:
s,s'ey \LLj=lJj ) 7=1
We now introduce the semi-detailed balance condition (see Section 2.3.1):
sey sey
an
and we multiply both sides by n^=i fj d sum over
s,s'ey 7=1 s,s'ey j
1
We exclude the pathological case /& = 1.
A.5 257
Interchanging the dummy variables 5 and sf on the r.h.s. yields:
s,s'ey \;=1
By adding this identically null expression to the l.h.s. of (A.21) we obtain:
\ w; -wl+wi\=0.
l
(A.23)
where all products are taken from j = 1 to b. The expression between braces is
always negative or zero. Indeed, denoting by x the argument of the logarithm,
this expression has the form log x — x + 1 which is zero for x = 1 and whose
first derivative (= 1/x — 1) is positive for x < 1, null for x = 1 and negative
for x > 1. Thus, logx — x + 1 must be negative for any value of x ^ 1. As
a consequence, (A.23) implies that the expression between braces must be zero
whenever A(s^>sf) is non-zero. In other words, Il/=i Tj a n d Ily=i Tj m u s t be
equal for any states s and sr with a non-zero collisional transition probability
A(s^sf), that is:
(
5 6 \
log J J /; ; - log n /? = 0, Vs, 5' G 7, (A.24)

i=i ;=i /
or equivalently:
b
Y A(s-+s') (5; - Si) log I = 0, Vs, s ' G y . (A.25)
With the definition of/,-, this result yields (A. 15).

• We prove that (A.15) => (A.13):
Equation (A.15) implies that for all s and sf in y, the products Ylf*1 and
are equal when A(s->s') is non-zero. Therefore, we deduce that:
4^) n fW=A(S-S') n f
;=i 7=1
Summing both sides over all s in y and using the semi-detailed balance condition
(2.17) gives (A.13). •
258 Appendix
A.6 The following relation holds:
yjf*J = fu v* = 1,..., b (A.27)

s,s'ey 7=1
(see Section 4.4.1)
Proof
We use the normalization condition (2.9) to write the l.h.s. of (A.27) under the
following form:
sey 7=1
The sum over s can be restricted to the Boolean states 5 with s* = 1. The l.h.s.
of (A.27) then reads:
sey, 7=i,
The sum is obviously equal to one, since it is taken over all possible Boolean
states of the b — 1 remaining channels (channel i excluded). •
A.7 The square matrix M whose elements are:

b
^qf^fiu K = 1,..., d, Kr = 1,..., 5, (A.28)
is invertible (see Section 4.5)
Proof
We denote by E the canonical real vector space of dimension b, and by /
the S -dimensional vector subspace of collisional invariants. The set of vectors
(q[K\ K = 1,...,<5) is a basis of 7. The vectors in E will be denoted by Roman
letters rather than bold faced symbols, which we use for vectors in the physical
space IRD.
We introduce the symmetric quadratic form acting on E x £, which associates
to any pair of vectors (x,y) € E x E, the real number:
x
' y =
A.8 259
The fn s are given in terms of the unperturbed average populations /o* by

fu = — /o;(l —foi) (see Section 4.5). Thus, the coefficients (—fa) of the quadratic
form are all positive and at least one of them is non-zero, if we exclude the
pathological situations where the average populations are all zero or all one
(completely empty or completely full lattice). This symmetric quadratic form
is therefore positive definite and defines an inner product on E. The matrix
elements of M can be expressed in terms of this inner product as:
We now consider in / , a basis (p M , K = 1,..., 3) of orthogonal unitary vectors

(orthogonal and unitary with respect to the inner product defined above). In
other words, these vectors are such that:
P P — °KKr> K
> K
— 1,...,0.
We introduce the real d x d square matrix R which transforms the q^ s into the
p M s. Its matrix elements R^ are such that:
Consider the matrix —R M RT, where RT is the transpose of R. The matrix

elements of —RMR T are the inner products p M • p^K'\ which are equal to 1
when K = Kf and 0 otherwise (the matrix — R M R T is just the opposite of the
unit matrix). Since the q^ s and the p M s are two bases of /, the matrix R has
a non-zero determinant. Therefore M also has a non-zero determinant. •
A.8 Computation of the r.h.s. of (4.47) for the class of models de-
fined in Section 4.5.2
Let us call S[K] the r.h.s. of (4.47):
\f4t^r'])\ (A-29)
i=i V=i /
where the first order corrections k^ to the Lagrange multipliers are, according
to (4.71):
(A30)
260 Appendix
We recall the definitions of the characteristic coefficients {2, £4 and & (see
Equation (4.64)) which depend on the geometric structure of the lattice gas:
1=1
b
+ SgcySps + SafiSpy), ^ ' '
;
)
We also recall that the collisional invariants q^ of the class of models under
consideration are (see Equation (4.65)):
(A.32)
V e . = iicc 2 __ Mi
2H 2b
The expansion coefficient fii is d(l — d)(l — 2d), according to (4.54). Using the
definitions (A.31) and (A.32), one can prove the following useful identities:
**
t=
e, = ^6, 2^ ciaCijiei = [— + — J
1=1 1=1
which we will need to compute the r.h.s. of (4.47).

A.8 261
Starting from (A.29), and taking into account (A.30), we can rewrite S^ as:
The expression in parentheses can be written in terms of a sum over component

indices, running from 1 to D. These summations are implicit: with Einstein's
convention, jpjyCipCiy stands for ^^jpjyCtpCiy. Thus we have:
P,y
S[K] 1 2 J
1
(A.34)
We first examine the case K = 0. The summation over i can be performed using
the explicit expression (A.31) for the crystallographic tensor J2C^CW- Because
crystallographic tensors of odd order are null, the second term in (A.34) does
not contribute, and the expression for S^ reads:
(A 35)
-
For the case K = a, a = 1,...,D, the algebra is quite similar. The first and
third terms in (A.34) do not contribute since for q^ = c^ they involve odd
order crystallographic tensors. The second term is computed with the second
identity in (A.33). This gives:
The case K = D + 1 is also straightforward. The first and third identities

in (A.33) are used, which yields:
The second order perturbations X^ to the Lagrange multipliers must satisfy

the relation (4.47) which reads:
Jf = S[K\ (A.38)
K' = l
262 Appendix
where the r.h.s. terms S M are given by (A.35), (A.36) and (A.37). Since the
matrix M^K'^ is diagonal (see Section 4.5.2), the second order perturbations /l|*']
follow immediately:
[0] = l-2d /J_2 J_ 2\

(2)
2d (l-d) \bfrm
2 2
bi4 m
r
a = 1
>-'D> (A39)
A.9 Computation of the momentum and effective energy fluxes in-

volved in the first order macrodynamic equations (see Sec-
tion 5.3.1)
By definition (see Section 2.1.3), momentum and effective energy fluxes are:
b
1=1
(A.40)
where the /i (0) s are the equilibrium mean populations derived in Section 4.5.2
for the nearly equally distributed states of single-species thermal models:
f.(0) _ P
~ b
1 1
-rJyCiy + T-wei
Q2 C4
b(b-2p)
2p(b — p) ^2 ^ (A 41)
2
2/ 2
We recall that the expansion parameter rj represents the order of magnitude of

j and w, that is, the proximity of the considered equilibrium state to the 'ideal'
A.9 263
equally distributed equilibrium state with zero momentum density and effective
energy density.
The key-ingredients to compute the fluxes are the following relations:
1=1
)x
^ =l
\J><xpdy8 + <
( A - 42 )
1=1
• E<
These relations are direct consequences of the definition (4.64) of the geometrical
parameters £2, £4 and & (see Section 4.5.2). To these relations, we must also
add the fact that the sum over i of any product containing an odd number of
components of c, is zero (see Section 2.3.5).
With these relations, the computation of the momentum flux tensor is rel-
atively straightforward: all terms in (A.41) with odd powers of j vanish, and
using the relations (A.42) yields the desired expression (5.30):
®Jga] = g(p)j*up + PQ?J2,w)(5ais +

where
b-2p b /4£4 D2\
and
For the energy flux, all terms in (A.41) with even powers of j vanish, and one
obtains (5.33):
with
264 Appendix
A. 10 Computation of the second order solvability conditions for the

Chapman-Enskog expansion (see Section 5.4.1)
We recall the second order solvability conditions (5.57) for the Chapman-Enskog
expansion:
r]fim + Ufa J2JStsdyf^ (A-43)

1=1 1=1
for all K = 1,..., (5, that is, for all collisional invariants.
Before analyzing separately the cases of the mass, momentum, and effective
energy invariants corresponding respectively to K = 0, K = 1,..., D and K = D+l,
one can simplify (A.43) without any loss of generality.
First, the term J2i Q^ft^ vanishes for all K, since it represents the contribution
of the first order corrections /, (1) to the collisional invariants' densities, and this
contribution is zero by the definition of f^ (see Equation (5.4) and the preceding
discussion).
Second, one can use the first order solvability condition (5.25):
to simplify the last two terms in (A.43). The resulting simplified second order
solvability condition reads:
o = dt2p^ + di?J2^î
t '"' b ( A - 44 )
for all K = 1,...,(5, that is, for all collisional invariants. In (A.44), p^ is the
density Yli^fi^ °ftri e collisional invariant q^K\ and O M is the corresponding
We now address successively the cases of mass, momentum and effective

energy invariants. The mass invariant leads to the simplest algebra. Indeed, the
first order solvability condition (5.35) for the mass invariant makes the last
A. 10 265
two terms in (A.44) vanish. In addition, the contribution ]T^ c^f^ of /, (1) to
the mass flux (i.e. momentum density) vanishes. This leads to the second order
solvability condition (5.58) for the mass invariant:
Kv = o.
To handle the case of the momentum invariant, we need the expression (5.30)
of the zeroth order momentum flux J2t cia.CiPfP\ that we truncate at the second
order in r\:
{0)
( ) h «, P = 1,..., D.
We also need the expression (5.53) for the first order momentum flux ^ . c^c^f^:
Finally, we must compute J2i CiaCipCiyfi^ up to the first order in rj. This requires
the second identity in (A.42), that is, the very definition of £4. The resulting
expression is:
b
. C.RC. f . ( o > _
j 1 + -j1) (JM + UK + uhi) +

After incorporating these ingredients in (A.44), the second order solvability
condition for the momentum invariant becomes:
(c)
0 = dt2ja - dlP ( v
We then use the first order solvability conditions (5.35) and (5.37) for the mass
266 Appendix
and effective energy respectively, to re-express the time-derivatives of p and w

in terms of space-derivatives. This yields:
0 = dt2ja - dlP i v(c) (dl0Ljp + dtfj* - ^diyjy3a^j j
or in compact form:
0 = 3 t J a -5 1 ^(v ( c ) (3 i a ^ + 5 1 ^ a - - 3 i y
+
2) (it ^ ) (5lJ/? + dlfij* ~ l
+ (9{n2).
This gives the second order solvability condition (5.59) for the momentum
invariant:
where the coefficient v, which a priori depends on (p,j 2 ,w), is given by:
v v
2D(D + 2)
The case of the effective energy invariant is somewhat simpler. We need
expression (5.33) of the zeroth order effective energy flux ^ . etCipf^0\ that we
truncate at the second order in rj:
and we use expression (5.56) for the first order effective energy flux:
b
A. 10 267
Finally we must compute J2ieicWciyft^ t0

fifSt o r d e r in rj, which requires the
fourth and fifth identities in (A.42); the result reads:
b
i=l
When all these results are incorporated in (A.44), the second order solvability
condition for the effective energy invariant becomes:
/ \
0 = dhw-c
Using the first order solvability condition (5.36) for the momentum, we re-express
the time-derivatives of jp in terms of space-derivatives, and we obtain:
or in compact form:
0 = dt2w-dJ^d^
This leads to the second order solvability condition (5.61) for the effective energy
invariant:
268 Appendix
with:
This completes the computation of the second order solvability conditions.
A.11 Cumulants (Section 8.2)
Consider the definition of the commutative rc-product:
(BW I B ( 2 ) | ... | B ( " ' ) L E ( r , 0 EE J2 K n) r f

' ( ' ) Bi1] B ? ] ••• Bin)' ( A -
1=1
with:
/c|n)(r, t) = (-1)" | - log [1 + exp {-#•, t) * 4,}], (A.46)
where <^(r,t) = (/>((A(r,t))) is given implicitly by relation (8.16), and 0o is

the thermodynamic potential conjugate to the number of particles per node
(proportional to the chemical potential).
We remark that K\n\r, t) is thercthcumulant of the local-equilibrium distribu-
tion, and therefore therc-productdefined by (A.45) is a cumulant average:
...B^))Jf9t). (A.47)
The first few cumulants (see also Chapter 4) are given explicitly by the expres-
sions (Abramowitz and Stegun, 1965):
4\v,t) = f\*{T,t),
= /^(r, t) (1 - /,LE(r, t)) (1 - 2 * /-(r, t)) ,
ic{4)(r,t) = ( K E ( r , 0 ) 4 ) - 3((KE(r,0)2)2
= /^(r, t) (1 - /fE(r, t)) (l - 6/,LE(r, t) + 6 (/fE(r, r)
(A.48)
Higher order cumulants can be obtained by evaluating the determinant

A. 11 269
(Risken, 1984):
/, LE (M) 1 0 0
ff*(r,t) ft*(r,t) 1 0
ft*(t,t) /,LE(r,t) /,LE(M)
(A.49)
J i Vr5l) Ji lr?l) /,LE(r,0 " /H'.O

n
where we have used the fact that n,- is a Boolean variable, and therefore:
<(fU(r, *))">„ = <n,-(M))LE = /,LE(r,f). (A.50)
Alternatively we can use the recursion formula:
m \ )
m=l \ /
which is obtained from the observation that the generating function for the
cumulants:
oo n
G(x) = log(exp{xn,(r,t)})LE = X > ! " W ) ^ y , (A.52)
satisfies the differential equation:
G"(x) = G\x)(l-G'(x)). (A.53)
It is easy to verify that the cumulants are connected by the relation:
Ai (A 54)
- -
Consider this relation for n = 1:
3 ^ / H r , t) = - K?(r, t) [JJ^ t(r, t)) * A, (A.55)

If we multiply (A.55) by At on the right on both sides of the equation and then
sum over the channels, we obtain the equality:
which implies:
(r, t). (A.57)

270 Appendix
Thus, Equation (A.54) becomes:
K n) r f K
d{A(Ttt)) ' (') = ;"+1V>f) ^ * UI4>^ (*,t). (A.58)
Multiplying (A.58) by Bf]Bf] ... B,!n) and summing over i yields:
= UI4)" 1 (r, t)*(A\ B™ | B (2)

| ... | B(n))LE (r, t), (A.59)
and, as a particular example of this general relation:
(I(r> 0)LE = (r>

d{A(tt))
(
-'-)L~E1 °* (4I->LE (f' ° ' (A 60)
-
which proves Equation (8.33).
A.12 Matrices (Section 8.3)
Here we discuss the properties of the non-symmetric matrix <£LE(r, t) , defined

in Section 8.3 by:
where:
I I (/"(M)r (l-/r(r,0) ( 1 ^ } . (A.62)

are
First, we prove that the conserved quantities {dL}f=i left eigenvectors of
J2?LE(r, t), with eigenvalue one:
^ A^f^t) = Ap (A.63)
The proof makes use of the fact that the average collision matrix connects only
those configurations with the same value for the conserved quantities. Hence:
-E
A.12 271
MM
J
J *»> {s}{s>}
= Aj • (A.64)
The right eigenvectors of j£?LE(r, t) with eigenvalue one are:
^2 ^j(^ 0 KT(*> t)£j = ^ E ( r , t)Af (A.65)
The proof of this relation is similar to the previous one:

b
7=1
=E ^ 7=1
4- E
{s}{^}
(1-/^,0) E sM(s^s')F{LsE}(r,0
{S}{5'}
E
= ^ (r,t)4-. (A.66)
The other right eigenvectors of J?LE{r,t), which constitute the set of kinetic
eigenvectors, are:
= l (n) Kf(r, 0 C^n); n = 1,..., (fe - iVhydro), (A.67)
with eigenvalues | 2 ^ | < 1, for n = l,...,(fc — iVhydro), a property that follows

from the fact that the equilibrium state is stable. The kinetic eigenvectors are
orthogonal to the constants of motion, in the sense that:
(A\C^)LE (r, t) = 0 ; n = 1,..., (6 - iVhydro), (A.68)
where the scalar product has been defined in section A.ll. This property can be
272 Appendix
derived from the following set of equalities for n = 1,..., (b — Nhydro) •

b / b \
E —*
TC
v v /
I
7
n - (n\
7
V
/
> I V
v I /
>
v— l
T E •
v
V 1
l
T E X
J
^ ^,f»ll
J
Uj b 7=1 \i=l /
Y2*T(*>UjCf = (4|C(W))LE(M), (A.69)

7=1
^=1
) L E (r,0. (A.70)
From (A.69) and (A.70), we obtain:
(A(n) - 1) (41 C (W) ) LE (r, t) = 0; n = 1,..., (6 - N hydro ), (A.71)
which implies (A.68), since X^ =£ 1.

Using the properties given by Equations (A.63), (A.65) and (A.70), we can
decompose J^LE(r, t) in the form given in (8.39):
<?%(!, t) = K-(r, t)At * (4I4)- 1 (r, t) * Aj + i^ E (r, t); (A.72)
\J(r,t) has the same set of eigenvectors as <£)J(x,t\ with eigenvalues:
^ E ( r , t) Kf(r, t) Aj = 0, (A.73)
7=1
7=1
where |/l(n) | < 1 for n = 1,..., (b - Nhydl0).
A.13 Projectors (Section 8.3)
We define the projection operator ^ , the projector onto the set of constants of
motion, by its action on an arbitrary vector {£/(r, t)ffE(r, t)}f=1:
{PB)j(T9t)f?(T9t)
b
— > IA^(Y t\ A * / AI A\ (T t\ * >d RI(Y t\ v f^(r t\
/ ^ / v*5 ^/ _ / ' \±i|£i/LE V' / _/-*-'/V 5 / J I v J /
= KjB(r, 0 4 - * (414)ITE1 (r, 0 * (4 B(r, t))LE. (A.75)

A. 14 273
In particular, we notice that:

b
(A(&B)(r,t))LE = ^2Aj(^B)j(T9t)ff(T9t) = (AB(r,t))LE. (A.76)
7=1
The complementary projector i> = 1 — & projects onto the set orthogonal to the
constants of motion, i.e.:
1
(r, t) * 4/) *z(r, t)f\*(r91)
= BJ(T, 0/} E(r, t) - Kfir, t) Aj * {A\A)~^ (r, t) * {A B(r, 0) L E , (A.77)

such that:
b
(4(^)(r,0) LE = ^ ^ ( ^ ( M l / ^ M ) = 0, (A.78)
;=i
that is, the quantity (IB) (r, t) has no projections onto the set of constants of
motion. The quantity that appears on the r.h.s. of Equation (8.50) is precisely
of this form, with B/(r, t)ff(r, t) = cr Vri /^(r, t).
A.14 Time correlation functions (Section 8.4)
Within the Boltzmann approximation the evolution of the equilibrium two-point

correlation function (see also Section 6.1) is given by:
b
= ] T exp {-C/ • Vr} ^ (dnk(r, % - \)bntf 9 0)>gJ), (A.79)
with:
Teq)
JJ M«or /'i_^q)) (1"Sl
K
j k=l
(A.80)
Equation (A.79) is iterated to yield:
{-Cl- • Vr}
(A.81)
274 Appendix
with:
= êxp{crVr}. (A.82)
The equal time correlation function is:
(5ni(T,0)8nj(r'90)){« d = tf* h &M, (A.83)

where nfp* denotes the second cumulant KA2) (see Section 4.7). Using Equations
(A.81) and (A.83), we have:
rr'
rr'
Therefore:
00
T=0 rr'
00
T=0 ij
(A.85)
References
Abramowitz, M. & Stegun, LA. 1965. Boon, J.P., editor 1992. Lattice Gas
Handbook of Mathematical Functions. Automata Theory, Implementation and
Dover, New York. Simulation (special issue of J. Stat. Phys.
Alder, BJ. & Wainwright, T.E. 1970. Decay WW))- Plenum Press, New York
of the velocity autocorrelation function, (Proceedings of a NATO Advanced
Phys. Rev. A 1(1), 18-21. Research Workshop, Nice (France),
Appert, C. & Zaleski, S. 1990. Lattice gas
Boon
with a liquid-gas transition, Phys. Rev. > J P > D a b > D > Kapral, R.
Lett. 64(1), 1-4. & Lawniczak, A. 1996. Lattice gas
A i_ rv XTAir o A* • x t n 1 n-»r automata for reactive systems* Phys. Rep.
Ashcroft, N.W. & Mermin, N.D. 1976. - . 55147
Solid State Physics. Holt-Saunders,
Philadelphia. Boon, J.P. & Yip, S. 1980. Molecular
Batchelor, G.K. 1967. An Introduction to Hydrodynamics. McGraw-Hill, New York
Fluid Dynamics. Cambridge University (Reprinted by Dover, New York, 1991).
Press, Cambridge. Brito, R., Bussemaker, H.J. & Ernst, M.H.
1991 A
Bell, H., Mceller-Wenghoffer, H , fluctuation formula for the
Kollmar, A., Stockmeyer. R., Springer, T. non-Galilean factor in lattice gas
& Stiller, H. 1975. Neutron Brillouin automata, J. Phys. A Math. Gen. 25,
scattering in fluid neon, Phys. Rev. A L949-L954.
11(1), 316-327. Brito, R. & Ernst, M.H. 1992. Ring kinetic
theor
Bender, CM. & Orszag, S.A. 1978. y f o r tagged-particle problems in
lattice
Advanced Mathematical Methods for gases> phys- Rev- A 46(2)> 875-887.
Scientists and Engineers. McGraw-Hill, Brito, R., Ernst, M.H. & Kirkpatrick, T.R.
Philadelphia. 1991. Staggered diffusivities in lattice
Bernardin, D. & Sero-Guillaume, O.E. 1990. gas cellular automata, J. Stat. Phys.
Lattice gases mixtures models for mass 62(1/2), 283-295.
diffusion, Eur. J. Mech. B 9, 21. Bussemaker, H., Ernst, M.H. & Dufty, J.W.
Bonetti, M. & Boon, J.P. 1989. Chaotic 1995. Generalized Boltzmann equation
dynamics in open flows : the excited jet, for lattice gas automata, J. Stat. Phys.
Phys. Rev. A 40(6), 3322-3345. 78(5/6), 1521-1554.
275
276 References
Cauchy, A. 1851. Rapport sur un memoire Ernst, M.H. 1990b. Linear response theory
presente par A. Bravais, et intitule for cellular automata fluids,
'Etudes sur la cristallographie', C. R. in Fundamental Problems in Statistical
Acad. Sci. Paris 32(8), 284-289. Mechanics VII, ed. H. van Beijeren.
North Holland
Cercignani, C. 1994. On the ^ ^ Ams terdam.
thermodynamics of a discrete velocity Ernst, M.H. 1991. Temperature and heat
gas, Transp. Theor. Stat. Phys. 23(1-3), conductivity in cellular automata fluids,
1-8. in Discrete Models of Fluid Dynamics,
186
Chapman, S. & Cowling, T.G. 1970. The ~ 1 9 7 ' e d A S ' A l v e s - W o r l d Scientific >
Mathematical Theory of Non-uniform Singapore.
Gases. Cambridge University Press, Ernst, M.H. & Das, S.P. 1992. Thermal
Cambridge. cellular automata fluids, J. Stat. Phys.
Chen, S., Zhao, K.H. & Doolen, G.D. 1989. 66(1/2), 465-483.
A lattice gas model with temperature, Ernst, M.H. & Dufty, J.W. 1990.
Physica D 37(1/3), 42-59. Hydrodynamics and time correlation
fu nctioM f c e l l u l a r
Clavin, P., d'Humieres, D., Lallemand, P. , °' automata, J. Stat.
& Pomeau, Y. 1986. Automates ^ WW), 57-86.
cellulaires pour les problemes a frontieres Ernst, M.H., Hauge, E.H. & van
libres en hydrodynamique a 2 et 3 Leeuwen, J.MJ. 1971. Asymptotic time
dimensions, C. R. Acad. Sci. Paris, behavior of correlation functions, Phys.
serie II 303, 1169-1174. Rev. A 4(5), 2055-2065.
Cornubert, R., d'Humieres, D. Fano, R.M. 1961. Transmission of
& Levermore, D. 1991. A Knudsen layer Information. Wiley, New York (p. 43).
theory for lattice gases, Physica D 47, Fleury, P.A. & Boon, J.P. 1969. Brillouin
241-259. scattering in simple liquids : argon and
Dab, D., Lawniczak, A., Boon, J.P. neon, Phys. Rev. 186, 244-254.
& Kapral, R. 1990. Cellular automaton Forster, D. 1975. Hydrodynamic
model for reactive systems, Phys. Rev. Fluctuations, Broken Symmetry, and
Lett. 64(20), 2462-2465. Correlation Functions. Benjamin,
Das, S.P., Bussemaker, HJ. & Ernst, M.H. Reading, Massachusetts.
1993. Generalized hydrodynamics and Frenkel, D. & Ernst, M.H. 1989.
dispersion relations in lattice gases, Phys. Simulation of diffusion in a
Rev. A 48(1), 245-255. two-dimensional lattice-gas cellular
Doolen, G., editor 1990. Lattice Gas automaton: a test of mode-coupling
Methods for Partial Differential theory, Phys. Rev. Lett. 63, 2165-2168.
Equations. Addison-Wesley, New York Frisch, U., Hasslacher, B. & Pomeau, Y.
(Reprints collection). 1986. Lattice gas automata for the
Drazin, P.G. & Reid, W.H. 1981. Navier-Stokes equation, Phys. Rev. Lett.
56
Hydrodynamic Stability. Cambridge ' 1505-1508.
University Press, Cambridge. Frisch, U., d'Humieres, D., Hasslacher, B.,
Dubrulle, B., Frisch, U., Henon, M. Lallemand, P., Pomeau, Y. & Rivet, J.P.
1987
& Rivet, J.P. 1990. Low viscosity lattice - L a t t i c e § a s hydrodynamics in two
and three
gases J Stat Phys 59 1187-1226 dimensions, Complex Systems
' 1, 649-707.
Eliott, J.P. & Dawber, P.G. 1979. Symmetry
in Physics. MacMillan Press Ltd., Gatignol. R. 1975. Theorie Cinetique des
London âz a Repartition Discrete de vitesses.
Springer-Verlag, Berlin.
Ernst, M.H. 1990a. Statistical mechanics of
Grosfils
cellular automata fluids, in Liquids, ' R 1 9 9 4 ' Hydrodynamique
Freezing and the Glass Transition, 1-94, statistique des gaz sur reseau, Ph.D.,
eds. D. Levesque, J.P. Hansen & Universite Libre de Bruxelles.
J. Zinn-Justin. Elsevier Science, Grosfils, P., Boon, J.P. & Lallemand, P.
Amsterdam. 1992. Spontaneous fluctuation
References 277
correlations in thermal lattice gas van der Hoef, M.A. & Frenkel, D. 1991.
automata, Phys. Rev. Lett. 68, 1077-1080. Evidence for faster-than-r 4 decay of the
velocit
Grosfils, P., Boon, J.P., Brito, R. & Ernst, y autocorrelation function in a 2D
fluid P /
M.H. 1993. Statistical hydrodynamics of ' ^ Rev< Lett 6 6 ' 1591-1594.
lattice gas automata, Phys. Rev. E 48, Homsy, G.M. 1987. Viscous fingering in
2655-2668. porous media, Annu. Rev. Fluid Mech. 19,
Hanon, D. & Boon, J.P. 1997. Diffusion 271-311.
and correlations in a lattice gas Hopcroft, J.E. & Ullman, J.D. 1979.
automata, Phys. Rev. E 56(6), 6331-6339. Introduction to Automata Theory,
Hardy, J. & Pomeau, Y. 1972. Languages and Computation.
Thermodynamics and hydrodynamics for Addison-Wesley, New York.
a modeled fluid, J. Math. Phys. 13, Huang? K 1963 statistical Mechanics.
1042-1051. w i l e y ? N e w York.
Hardy, J., Pomeau, Y & de Pazzis, O. 1973. Huerre> p & Monkewitz? pA 1990 Local
Time evolution of a two-dimensional a n d global instabilities in spatially
model system: Invariant states and time developing flows, Annu. Rev. Fluid Mech.
correlation functions, J. Math. Phys. 14, 22 473-537
1746-1759.
XT , _ . _ . _ n _ __ 4 r _ d'Humieres, D., Lallemand, P. & Frisch, U.
Hardy, J., de Pazzis, O. & Pomeau, Y 1976. i n o , T ' A. . . ,, ~
^/' . t . ^ t .' . 1986. Lattice gas models for 3-D
Molecular dynamics
J of a classical lattice , , , ^ , r -
_ • j • hydrodynamics, Europhys. Lett. 2,
gas: Transport properties and time 291-297
correlation functions, Phys. Rev. A 13,
1949-1961. d'Humieres, D., Qian, YH. & Lallemand, P.
u , T p „ , , 1 f t _ Ayr. . 1989. Invariants in lattice gas models,
Hardy, J. & Pomeau, Y 1977. Microscopic . n 3. . , TTr .6,
J 1f • a • 4. in Proceedings of the Workshop on
model tor viscous now in two ,
,. . D, D A iiC n^r. n~c Discrete Kinetic Theory,y Lattice Gas
dimensions, Phys. Rev. A 16, 720-726. , .
Dynamics and Foundations oj
Henon, M. 1987a. Isometric collision rules Hydrodynamics, Torino (Italy), September
for the 4-D FCHC lattice gas, Complex 20-24, 1988, 102-113, ed. R. Monaco.
Systems 1, 475-494. World Scientific, Singapore.
Henon, M. 1987b. Viscosity of a lattice gas, d'Humieres, D., Lallemand, P. & Qian, YH.
Complex Systems 1, 763-789. 1990 Finding the linear invariants of
Henon, M. 1989. On the relation between lattice gases, in Computational Physics
and
lattice gases and cellular automata, Cellular Automata, eds. A. Pires,
in Proceedings of the Workshop on D.P. Landau & H. Herrman. World
Discrete Kinetic Theory, Lattice Gas Scientific, Singapore.
Dynamics and Foundations of Kinchin, A.I. 1957. Mathematical
Hydrodynamics, Torino (Italy), Foundations of Information Theory.
September 20-24, 1988, 160-161, ed. Dover? N e w York
R. Monaco. World Scientific, Singapore.
Konig, M., Eisenlohr, H. and Eckelmann, H.
Henon, M. 1992. Implementation of the 1 9 9 Q J h e flne s t r u c t u r e i n ^
FCHC lattice gas model on the Strouhal-Reynolds number relationship
connection machine, in Proceedings of rf ^ , a m i n a r w a k e rf a d r c u l a r
the NATO Advanced Research Workshop ^ ^ phys FMds A 1(% 1607_1614
on Lattice Gas Automata Theory,
Implementation, and Simulation, Nice Kubo. R. 1958. Some aspects of the
(France), June 25-28, 1991, J. Stat. Phys. statistical mechanical theory of
68(3/4), 353-377, ed. J.P. Boon. irreversible processes, in Lectures in
J TrrA/rAPX7 11T-, m™ Theoretical Physics, vol. 1, 120-140, eds.
van der Hoef, M.A. & Frenkel, D. 1990. __,„
W E _ . . o\ „ _ ' _Tri
. ., r 1 1 • - Bnttin & L.G. Dunham. Wiley,
Lone-time
& tails ol the velocity ^T ^T t
, . c . . / , New York,
autocorrelation function in two- and
three-dimensional lattice-gas cellular Lamb, H. 1945. Hydrodynamics, 6 th edition,
automata, Phys. Rev. A 41, 4211-42S5. Dover, New York.
278 References
Landau, L.D. & Lifschitz, E.M. 1987. Fluid states. IV. Far-from-equilibrium
Mechanics, 2nd edition, vol. 6, in 'A stationary states and the regression of
Course of Theoretical Physics'. Pergamon fluctuations, Phys. Rev. A 20(6),
Press, Oxford. 2533-2546.
Lutsko, J.F., Boon, J.P. & Somers, J.A. Resibois, P. & de Leener, M. 1977.
1992. Lattice gas automata simulations Classical Kinetic Theory of Fluids. John
of viscous fingering in a porous medium, Wiley, New York.
in Numerical Methods for the Simulation ~. , TT irkO/) ^ r ,, „, ,
J
. Risken, H. 1984. The Fokker-Planck
of Multi-Phase and Complex Flow, F .. c . ,, ,
;,, n f J ™ , * , , , ^ Equation. Springer- Verlag, DBerlin.
r
124-135, ed. T.M.M. Verheggen.
Springer-Verlag, Berlin. Rivet, J.P. 1987. Simulation d'ecoulements
,,., A ino. _ .L1 _. ., _. tri-dimensionnels par la methode des gaz
Majda, A. 1984. Compressible Fluid Flow , . , . _ „&
,' /•„ • • sur reseaux: premiers resultats, C. .R.
ana Systems of Conservation Laws in , . . , . TT _A
,o T^ • 11 o • ^r i Acad. Sci. Pans, sene II 305, 151-156.
Several Space Variables. Springer-Verlag,
Berlin. Rivet, J.P. 1987. Green-Kubo formalism
AT uiAyrour r>T J-* 1 nni for lattice gas hydrodynamics and
Mareschal, M. & Hohan, B.L., editors 1992. ^r _ ° ;, /
,,. . c . , .. r ~ , Monte-Carlo evaluation of shear
Microscopic Simulations of Complex .
TJ J J • r>i r»i viscosities, Complex Systems 1, 839-851.
Hydrodynamic Phenomena. Plenum, > r * •>
New-York. Rivet, J.P. 1988. Hydrodynamique par la
m 6 t h o d e des gaz sur r6seaux P h D
Molvig, K, Donis, P., Mffler.IL, ' '
Myczkowski, J. & Vichniac, G. 1989. Umversite de Nice.
Multi-species lattice-gas automata for Rivet, J.P. 1993. Spontaneous
realistic fluid dynamics, In Cellular symmetry-breaking in the 3-D wake of a
Automata and Modeling of Complex long cylinder simulated by the lattice gas
Systems, 206-231, eds. P. Manneville, R. method and drag coefficient
Bideaux, N. Boccara & G. Vichniac. measurements, Appl. Sci. Res. 51,
Springer-Verlag, Berlin. 123-126.
Mori, H. 1965a. Transport, collective Rothman, D.H. & Keller, J.M. 1988.
motion, and Brownian motion, Prog. Immiscible cellular automaton fluids,
Theor. Phys. 33, 423-^55. J. Stat. Phys. 52, 1119-1127.
Mori, H. 1965b. A continued fraction Rothman, D.H. & Zaleski, S. 1994.
representation of the time-correlation Lattice-gas models of phase separation:
function, Prog. Theor. Phys. 34, 399-416. interfaces, phase transitions, and
Naitoh, T, Ernst, M.H., van der Hoef, M.A. multi-phase flow, Rev. Mod. Phys. 66,
& Frenkel, D. 1991. Extended mode 1417-1479.
coupling and simulations in Rothman, D.H. & Zalseki, S. 1977. Lattice
cellular-automata fluids, Phys. Rev. A 44, Gas Cellular Automata, Cambridge
2484^2494. University Press, Cambridge.
McNamara, G.R. 1990. Diffusion in a Schmitz, R. & Dufty, J.W. 1990. Positivity
lattice gas automaton, Europhys. Lett. 12, o f transport coefficients for cellular
329
- automata, Phys. Rev. A 41(8), 4294-4297.
McQuarrie, D.A. 1976. Statistical Shannon, C.E. & Weaver, W. 1969. The
Mechanics. Harper and Row, New York. Mathematical Theory of Communication,
Noullez, A. 1990. Automates de Gaz sur fourth edition. The University of Illinois
Reseau: Aspects Theoriques et Press, Urbana.
Simulations, Ph.D., Universite Libre de Suarez? A & Boon> j p 1997a Non_linear
Bruxelles. lattice gas hydrodynamics, J. Stat. Phys.

Papoulis. A. 1965. Probability, Random 87(5/6), 1123-1130.
Variables, and Stochastic Processes. Suarez> A & Boon? j p m i b Non.linear
McGraw-Hill, New York. hydrodynamics of lattice gas automata

Procaccia, I., Ronis, D. & Oppenheim, I. with semi-detailed balance, Int. J. Mod.
1979. Statistical mechanics of stationary Phys. C 8(4), 653-674.
References 279
Succi, S. 2000. The Lattice Boltzmann Reynolds numbers, J. Fluid Mech. 206,
Equation for Fluid Dynamics and Beyond. 579-627.
Oxford University Press, Oxford (U.K.) Wolfram, S. 1986. Theory and Applications
(to be published). of CeiMar Automata. World Scientific,
Succi, S. Benzi, R. & Higuera, F. 1991. The Singapore.
lattice Boltzmann equation: A new tool Wolfram, S. 1986. Cellular automaton
for computational fluid mechanics, fluids x. b a s i c t h j S m Ph 45
Physica D 47 (1/2), 219-230. 471-525.
Tribel, O. & Boon, J.P. 1995. Lattice gas Zanetti, G. 1989. Hydrodynamics of lattice
automata with interaction potential, automata, Phys. Rev. A 40(3),
J. Stat. Phys. 81, 361-377. 1539-1548
Tribel, O. & Boon, J.P. 1997. Entropy and Zwanzig, R. 1960. Statistical mechanics of
correlations in lattice gas automata irreversibility, in Lectures on Theoretical
without detailed balance, Int. J. Mod. physics> w / J > 1 0 6 - 1 4 1 , eds. WE. Brittin,
Phys. C 8(4), 641-652. B w D o w n s & j . Downs. Wiley, New
Tung, W.K. 1985. Group Theory in Physics. York.
World Scientific, Singapore. Zwanzig, R. 1965. Time-correlation
Williamson, C.H.K. 1989. Oblique and functions and transport coefficients in
parallel modes of vortex shedding in the statistical mechanics, Ann. Rev. Phys.
wake of a circular cylinder at low Chem. 16, 67-102.
Author index
Abramowitz, M , 106, 268 Cabannes, H., 234

Adler, C, 242, 247 Cauchy, A., 8
Adler, P.M., 241 Cercignani, C, 98
Alder, B.J., 196 Chapman, S., 109, 188
Alexander, R, 245 Chen, H., 104, 241, 243-245
Appert, C, 8, 242 Chen, S., 104, 243-245, 249
Ashcroft, N.W., 2, 5, 7, 11 Chopard, B., 243, 249
Clavin, P., 241, 242
Batchelor, G.K., xiv, 203 Cloucqueur, A., 208, 247
Bell, H., 170 Clovin, M.E., 241
Benard, H., 218, 226 Cornubert, R., 30, 240
Benard, H., xiv Cowling, T.G., 109, 188
Bender, CM., 113, 195
Benzi, R., 245 ^ D ^
Bernardin, D., 175, 24(^242 ^ s p ^ m m > ^
Bhatnagar, P., 246 Dawber, P.G., 24, 28
Binder, P.M., 238 Diemer, K., 243
Boghosian, B.M., 238, 239, 247 Donis, P., 104
Boon, J.R, xvii, 34, 53, 103, 107, 108, 146, 153, _ 1 _ _ 1n/| . . . _ _ _ 0 . . .
Doolen, G.D., 104, 243-245, 248, 249
154, 164, 165, 169, 170, 175, 177, 178, 183, _ ' ' . .'
238-240,242,243,248,249 nT T'f ô
' Dubrulle, B., 239
™ n .' ; o Dufty, J.W., 184, 194, 196, 238
Bravais, A., 8
Brito, R., 186, 196, 238
Broadwell, J.E., 234 Eggert, K, 243
Burges, C, 241 Eliott, J.P., 24, 28
Bussemaker, H.J., 168, 171, 238, 240 Eloranta, K., 244
281
282 Author index
Ernst, M.H., 98, 139, 168, 171, 183, 184, 186, Kubo, R., 196
194, 196-198, 238, 240, 243, 249
Ladd, A.J.C., 241
Feder, J., 245 Lallemand, P., xvii, 15, 22, 34, 38, 42, 48, 53, 59,
Feynman, R., xiii 74, 133, 236, 237, 240-243, 246, 247
Flekkoy, E.G., 243, 245-247 Landau, L.D., 112
Fleury, P.A., xvii, 170 Lavallee, P., 240
Forster, D., 107 Lawniczak, A., 242
Frenkel, D., 198, 238, 241 Lee, Y.C., 243, 244
Frisch, U., 15, 34, 39, 59, 74, 133, 183, 235-237, Leener (de), M., 157
239, 247-249 Leeuwen (van), J.M.J., 197
Fu, C , 243 Levermore, D., 30, 59, 240
Lifschitz, E.M., 112
Gatignol, R., 83, 234, 256 Longo, E., 234
Ginzbourg, I., 241, 245 Luo, L.S., 243
Godreche, C , 249 Lutsko, J.L., 243
Groos, E.R, 246
Grosfils, P., xvii, 34, 53, 98, 238, 243 Maillot, B., 244
Gunstensen, A.K., 242, 245 Majda, A., 203
Gutman, S., 243 Mareschal, M, 154
Margolus, N., 208, 246, 247
Hanon, D., 175, 177, 239 Markus, M., 244
Hardy, J., 34, 35, 80, 235, 237 Martinez, D.O., 245
Hasslacher, B., 15, 34, 39, 74, 133, 235, 237 Matthaeus, W.H., 241, 244, 245, 249
Hatori, T., 244 Maxwell, J.C., 234
Hauge, E.H., 197 McNamara, G.R., 175, 247
Helmholtz (von), H., 222 McQuarrie, D.A., 103
Henon, M., 9, 60, 63, 74, 236, 237, 239, 247, 248 Mermin, N.D., 2, 5, 7, 11
Hess, B., 244 Meyer, D., 239
Higuera, R, 245 Miller, R., 104
Hillis, W.D., xiii Molvig, K, 104
Hoef (van der), M.A., 198, 238, 241 Monaco, R., 234, 248
Holian, B.L., 154 Monkewitz, P.A., 227
Homsy, G.M., 231 Montgomery, D.C., 244, 245
Hopcroft, J.E., 2, 235 Mora, P., 244, 245
Huang, K., 69, 107 Mori, H., 140, 184
Huerre, P., 227 Myczkowski, J., 104
Humieres (d'), D., 15, 22, 30, 34, 38, 42, 48, 59,
74, 133, 208, 236, 237, 240-242, 245-247, Naitoh, T., 198, 238
249 Neuman (von), J., 235
Noullez, A , 175, 237, 240
Kadanoff, L.R, 247
Kapral, R., 242 Olson, J.F., 242
Karman (von), T, 218, 226 Orszag, S.A, 113, 195, 246
Karman (von), T., xiv Oxaal, U., 245
Keller, J.M., 242
Kelvin, (Lord), 222 Papoulis, A., 106
Kinchin, A.I., 74 Pazzis (de), O., 34, 35, 80, 235, 237
Kirkpatrick, T.R., 186 Platkowski, T, 234
Kohring, G.A, 243 Pomeau, Y., 15, 34, 35, 39, 74, 80, 133, 235, 237,
Krook, M.K., 246 240-242, 247
Author index 283
Prandtl, L., xiv Tietjens, O., xiv

Procaccia, L, 184 Toffoli, T., 208, 246, 247
Tribel, O., 103, 107, 108, 240
Qian, Y.H., 22, 38, 42, 48, 240, 246 Tung, W.K., 28
Rem, P.C., 236, 242, 247, 248 Ullman, J.D., 2, 235

Resibois, P., 157
Risken, H., 269 Velzen (van), G.A., 238
Rivet, J.P., xiv, 15, 64, 74, 133, 196, 237, 239, Vergassola, M., 245
247, 248 Vichniac, G., 104, 246
Rothman, D.H., 183, 242, 243, 245, 247, 249
Wagner, L., 245
Schmitz, R., 196 Wainwright, T.E., 196
Scholl, H., 248 Wang, Z., 245
Searby, G., 241, 242, 247 Weaver, W, 74
Sero-Guillaume, O.E., 175, 241, 242 Williamson, C.H.K., 227
Shan, X., 245 Wolf-Gladrow, D., 249
Shannon, C.E., 74, 75 Wolfram, S., 3, 183, 236, 237, 248
Shimomura, T., 247
Somers, J.A., 236, 242, 243, 245, 247, 248 Yip, S., 107, 108, 146, 153, 154, 164, 165, 169,
Stegun, LA., 106, 268 175, 178
Suarez, A., 183, 238
Sued, S., 239, 245, 246, 249 Z a l e s k i ? S .,
8, 183, 241, 242, 245, 249
Zanetti, G., 184, 194, 245, 247
Taylor, W, 238, 239 Zwanzig, R., 140, 184, 196
Subject index
Page numbers in bold type indicate main sections.
acoustic limit, 200 description, 71

aggregation, 223 equation, lattice (LBE), 68, 72
alphabet, 2 Boolean
autocorrelation function, 140, 198, 238 configuration, 12, 22
automaton field, 12, 98, 114, 215
cellular, 3, 235 state, 12
deterministic, 3 variable, 2, 76
finite, 2, 11 bounce-back reflection, 31
individual, 3, 11 boundary conditions
non-deterministic, 3 Born-von Karman, 17
average bounce-back, 31, 214
ensemble-, 68, 69, 70, 216 diffusive, 32
population, 68, 70 macroscopic, 18, 30
space-, 69, 216 no-slip, 32
time-, 69, 216 periodic, 17, 30, 82
specular, 32
balance Bravais lattice, 1, 4, 7, 10
bulk
detailed, 20 viscosity, 134, 202
semi-detailed, 20, 82
violation of, 239 cell, 4
Bernoulli random variable, 19, 42 Cellular Automata Machine (CAM), 247
Bhatnagar-Gross-Krook models (lattice-), 246 cellular automaton, 3, 235
bifurcation, 208 channel, 4, 11
bit, 74, 75, 210 index, 4, 11
Boltzmann multiple, 5
approximation, 68, 71 channel-wise storage, 211
collision matrix, 129 Chapman-Enskog method, xviii, 109, 111, 135
285
286 Subject index
Chapman-Kolmogorov equation, 71 self-, 21

circulation time, 204, 226 state, 21
CM (Connection Machine), 248 duality, 21
collision
deterministic, 1 effective collision, 36
effective, 20, 36 effective energy, 95, 119
matrix, 18 density per node, 95
non-deterministic, 1, 4 eigenvalue spectrum, 164
non-reversible, 4 elastic waves, 243
operator, 16, 138 energy, 12
passive, 20 effective, 95
phase, 4, 16 kinetic, 95
rule, 1, 18, 236 entropy
collisional efficiency, 36 global, 79
colloid, 241 local, 77, 78
colored lattice gas, 50, 175 of information, 74
communication (theory of), 74 statistical, 73
compressibility coefficient, 104 equation of state, 101
configuration, 12, 22 equilibrium
connection global, 80
number, 11 local, 110
rule, 3 low speed, 90
scheme, 3 nearly equally distributed, 93
vector, 11 ergodic hypothesis, 69
Connection Machine (CM), 248 Euler equations, 148, 189
connectivity, 11 evolution
conserved quantities, 21, 23, 84, 130, 185 operator, 16
contour dynamics, 244 rule, 1, 3
coordination number, 11 exclusion principle, 2, 6
correlation function, xviii, 105, 237
crystallographic F4 lattice, 59
isotropy, 26, 27 FCHC lattice, 59
point group, 5, 23 FCHC models, 34, 57, 205, 236
space group, 2 Fermi-Dirac distribution, 81, 84, 99, 154
tensor, 27, 38, 43, 58 FHP models, 34, 235
cumulants, 268 CFHP, 50
FHP-1, 39
density, 13 FHP-2, 44
energy, 86 FHP-3, 48
mass, 86 finite automaton, 2, 11
momentum, 86 fluctuations, xvi, 137, 153
of an observable, 13 flux, 13
detailed balance, 20 of an observable, 13
diffusion, 174 Fredholm's criterion, 118, 126
coefficient, 180 free-slip boundary conditions, 32
diffusive reflection, 32
discrete kinetic theory, 234 G-invariance, 23
disorder, 74 Galilean invariance, xviii, 128, 201, 205
distinguishable particles, 7 GBL model, 34, 53
drag coefficient, 217 generating vector, 5
dual global equilibrium, 80
Subject index 287
grand canonical ensemble, 99 Lagrange multipliers, 84, 86

Green-Kubo formalism, 151, 184, 193, 196, 237 Landau-Placzek approximation, xviii, 164, 178
group lattice
global invariance, 2 Bhatnagar-Gross-Krook models, 246
local invariance, 5 body-centered cubic (bcc), 5
point, 5, 23 Boltzmann equation, xviii, 72
space, 2 linearized, 138
Boltzmann models, 244
tf-theorem, 68, 73, 77 Bravais, 1, 4, 7, 10
hard sphere gas, 137 constant, 5
homogeneity, 3 F4, 59
homokinetic model, 11, 133 FCHC, 59
honeycomb lattice, 7 gas, 3
HPP model, 34, 235 hexagonal, 7
hydrodynamic honeycomb, 7
equations, 109, 190 Liouville equation, 68, 70, 71, 82
fluctuations, xvi, 137, 153 quasi-, 8
limit, 109, 110, 111, 199 regular, 7
spectrum, 163, 180 simple cubic, 61
hydrodynamics, linearized, 138 square, 35
triangular, 7
immiscible fluids, 242 LB models, 244
incompressibility condition, 203 LBGK models, 246
incompressible limit, 203 lift coefficient, 218
indistinguishable particles, 7 light scattering spectroscopy, 153
individual automaton, 3, 11 link, 2
information Liouville
entropy of, 74 H-theorem, 79
measurement, 74 description, 68, 68, 69
quantity of, 74 equation, 68, 70, 71, 82
interfaces, 231 local equilibrium, 110
invariant long-time tails, 196
collisional, 22, 83 longitudinal wave, 201
geometrical
moving, 23, 240 Mach number, 202
static, 22, 240 macroscopic boundary conditions, 30
local, 22 macroscopic density, 13
microscopic, 22 macroscopic flux, 13
invariant measure, 81 macroscopic variables, 86, 87
invariant states, 80 macrostate, 69
irreducibility, 28, 29, 122 equilibrium
isometric algorithm, 63 global, 80
isotropy . local, 110
crystallographic, 26, 89, 90, 93, 124, 125, 136, homogeneous factorized, 81
205 magnetohydrodynamics, 244
full, 26 mass, 12
density per node, 95
Kelvin-Helmholtz instability, 222 Maxwell-Boltzmann distribution, 80
kinetic propagator, 154 Maxwellian equilibrium state, 80
kinetic temperature, 104 memory function, 144, 145
Knudsen layer, 240 micro-reversible, 20
288 Subject index
microdynamic global, 12
equation, 10, 15 local, 12
formalism, 10 single-node, 12
microscopic density phonons, 243
per channel, 13 plasmas, 244
per node, 13 point group, 5, 23
per unit volume, 13 porous medium, 231, 243
microstate, 69 position vector, 4, 11
minimal model, 6 post-collision state, 4
mixture (of fluids), 241 power spectrum, 178
mode-coupling theory, 196, 238 pre-collision state, 4, 72
molecular chaos hypothesis, 71 pressure, 120, 200
molecular dynamics, xiii, 153 kinetic, 103
moment propagation method, 198 thermodynamic, 103
momentum, 12 primitive cell, 13
density per node, 95 probability distribution function, 69
multi-scale, 110, 188 projection operator, 140, 184
multinomial expansion, 28 projectors, 272
multiple channel, 5 propagation
operator, 16
Navier-Stokes equations, xiii, 128, 203 phase, 4, 16
neutron scattering spectroscopy, 153
Newtonian fluid, 34, 112, 204, 207 quasi-lattice, 8
node, 2, 4, 11
node-wise storage, 211 Reseau d'Automates Programmables (RAP),
non-Galilean factor, 92, 121, 200 208, 247
for energy, 121 random force, 147
for momentum, 121 Rayleigh-Brillouin spectrum, 169
non-thermal models, 89, 199 reaction-diffusion, 241
normalization, 19, 69 reflection
bounce-back, 31
observable, 12 diffusive, 32
energy, 14 operator, 30
generalized, 14, 69 specular, 32
homogeneous, 15 regular lattice, 7
local, 15 rest particles, 5
mass, 14 Reynolds number, 204
momentum, 14 Reynolds's similarity law, xv
ordinary, 70
particle number, 14 Saffman-Taylor instability, 231
obstacle, 30, 240 scale
operator macroscopic, 110, 112
collision, 16 microscopic, 109, 110
linearized, 138 multi-, 110, 188
evolution, 16 separation, 109
projection, 140, 184 space- and time-, 109
propagation, 16 seismic waves, 243
self-dual, 21
passive scalar, 67, 225 semi-detailed balance, 20, 82
phase separation, 242 shear viscosity, 127, 202
phase space, 12 shear wave, 202
Subject index 289
signal to noise ratio, 216 shear, 127, 202

sound damping coefficient, 162
sound wave, 201 words (computer), 210
specular reflection, 32
speed of sound, 162, 163, 201
spurious invariant, 43
square lattice, 35
state, 12
dual, 21
internal, 2, 4
macro-, 69
micro-, 69
post-collision, 4
pre-collision, 4, 72
storage
channel-wise, 211
node-wise, 211
Strouhal number, 218
structure factor
dynamic, xvii, xviii, 153, 156, 164, 169, 180
static, 106, 107, 155, 159, 176
subtracted current, 146, 194
surface tension, 242
synchronization, 3
tagged particles, 196

tensor
crystallographic, 27
isotropic, 26
thermal conductivity, 127
thermal diffusivity, 162
thermal models, 93, 199, 243
thermal scalar product, 140
thermodynamics, 98
time step, 4, 35, 80, 199
transition probability, 18
transport coefficients, 127, 128, 135, 163
transport matrix, 150, 151
transputer, 231, 247
transverse wave, 202
turbulence, xvi
van Hove function, 154

vector
connection, 11
generating, 5
position, 4, 11
velocity, 4
virial coefficient, 103, 107
viscosity, 162, 237
bulk, 134, 202

(J.-P. Rivet, J. P. Boon) Lattice Gas Hydrodynamic

Uploaded by

Copyright:

Available Formats

(J.-P. Rivet, J. P. Boon) Lattice Gas Hydrodynamic

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(J.-P. Rivet, J. P. Boon) Lattice Gas Hydrodynamic

Uploaded by

Copyright:

Available Formats

Lattice gas hydrodynamics is the name used to describe the approach to fluid dynamics

starting from a simple micro-world constructed as an automaton universe. In this

Professor Boris Chirikov

Professor Predrag Cvitanovic

Professor Frank Moss

Professor Harry Swinney

Cambridge University Press

The Edinburgh Building, Cambridge CB2 2RU, UK

This publication is in copyright. Subject to statutory exception

First published 2001

A catalogue recordfor this publication is available from the British Library

ISBN-13 978-0-521 -41944-4 hardback

ISBN-13 978-0-521 -01971 -2 paperback

Chapter 1 Basic ideas 1

1.1 The physicist's point of view 1

Chapter 2 Microdynamics: general formalism 10

2.1 Basic concepts and notation 10

2.2 The microdynamic equation 15

Chapter 3 Microdynamics: various examples 34

3.1 The HPP model 34

3.3 The FHP-2 model 44

3.7.1 Models with multiple links 58

Chapter 4 Equilibrium statistical mechanics 68

4.1 The Liouville description 68

Chapter 5 Macrodynamics: Chapman-Enskog method 109

5.1 Local equilibria and the hydrodynamic limit 110

5.3.1 Solvability conditions for the first order problem 118

Chapter 6 Linearized hydrodynamics 137

6.1 The linearized Boltzmann equation 138

Chapter 7 Hydrodynamic fluctuations 153

7.1 The dynamic structure factor 154

7.6.2 Low density 171

Chapter 8 Macrodynamics: projectors approach 183

8.1 Preliminaries 184

Chapter 9 Hydrodynamic regimes 199

9.1 The acoustic limit 200

Chapter 10 Lattice gas simulations 207

10.1 Lattice gas algorithms on dedicated machines 208

10.5.1 The Kelvin-Helmholtz instability 222

Chapter 11 Guide for further reading 233

11.1 The historical 'roots' 234

Appendix Mathematical details 250

simulations showing wakes and eddy formation, and hardware elaboration of

-4.0 -3.0 -2.0 -1.0 0.0 1.0 2.0 3.0 4.0

-2.0 -1.0 0.0 1.0 2.0

Figure 2 The spectrum of density fluctuations in the hydrodynamic regime, (a)

l.i The physicist's point of view

1.2 The mathematician's point of view

1.2.1 Finite automata

1.2.2 Cellular automata

1.2.3 Lattice gases

crystallographic point group G of the lattice,4 (ii) they include at least

1.3.2 Space and time

1.3.3 The exclusion principle

1.3.4 Bravais lattices

Figure 1.1 (a) An example of a two-dimensional Bravais lattice: the triangular

1.3.5 Local versus n o n - l o c a l collisions

1.3.6 Collision-propagation versus propagation-collision

1.3.7 Mathematical versus physical

where rii(r) is the occupation of channel i at node r. There is a semantic

«(n,r)= X > ( r + '*')• (2.4)

n(U + l) = £(n(t),Z(t))9 (2.7)

y(n).(r) = ni(r-ci). (2.8)