introduction to statiscal physics_lecture_notes2009

8/4/2019 Introduction to Statiscal Physics_lecture_notes2009

http://slidepdf.com/reader/full/introduction-to-statiscal-physicslecturenotes2009 1/50

Introduction to statistical physics:Lecture notes

Eric BertinLaboratoire de Physique, ENS Lyon and CNRS

Lyon, France

March 16, 2010

Contents

1 Introduction 3

2 Equilibrium statistical physics 42.1 Microscopic dynamics of a physical system . . . . . . . . . . . 4

2.1.1 A particle attached to a spring . . . . . . . . . . . . . 42.1.2 Many-particle system . . . . . . . . . . . . . . . . . . 72.1.3 Case of discrete variables: spin models . . . . . . . . . 8

2.2 Statistical description of an isolated system at equilibrium . . 82.2.1 Statistical description: a toy model . . . . . . . . . . . 82.2.2 Fondamental postulate of equilibrium statistical physics 92.2.3 Computation of Ω( E ) and S (E ): some simple examples 102.2.4 Distribution of energy over subsystems and statistical

temperature . . . . . . . . . . . . . . . . . . . . . . . . 122.3 Equilibrium system in contact with its environment . . . . . 14

2.3.1 Exchanges of energy . . . . . . . . . . . . . . . . . . . 142.3.2 Canonical entropy . . . . . . . . . . . . . . . . . . . . 172.3.3 Exchanges of particles with a reservoir: the grand-

canonical ensemble . . . . . . . . . . . . . . . . . . . . 192.4 Phase transitions . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.4.1 Example of the Ising model . . . . . . . . . . . . . . . 202.4.2 Ising model in fully connected geometry . . . . . . . . 212.4.3 Ising model with nite connectivity . . . . . . . . . . . 232.4.4 Renormalization group approach: a brief introduction 25

2.5 Disordered systems and glass transition . . . . . . . . . . . . 29

1



1 Introduction

These lecture notes are a transcription of the introductory lecture on statis-tical physics given at the French Summer School on Complex Systems heldboth in Lyon and Paris during the summers 2008 and 2009. This lecture wasaimed at giving a basic knowledge of the concepts and methods of statisticalphysics to the participants, so that they could later on follow more advancedlectures on diverse topics in the eld of complex systems.

Generally speaking, the goal of statistical physics may be summarizedas follows:

• Study systems composed of a large number of interacting entities,

• Predict macroscopic behavior from microscopic laws.

These two items actually correspond to one of the generic goals of com-plex system sciences. However, the specicity of statistical physics can besummarized as follows:

• The entities considered are basically atoms or molecules (althoughmore complicated entities could also be considered),

• The entities are most often all the same, or only a few number of speciesare considered (in contrast with biological or sociological systems for

instance),

• The microscopic laws are known from fundamental physical theories.

For these reasons, systems studied in the framework of statistical physicsmay be considered as among the simplest examples of complex systems. Onefurther specicity of statistical physics with respect to other sciences aimingat describing the collective behavior of complex systems is that it allows fora rather well-developed mathematical treatment.

The present lecture is divided into two rather unequal parts. The rstone deals with equilibrium statistical physics, and we try to expose in a con-

cise way the main concepts, paying specic attention to those that could bemore generally relevant to complex system sciences. The second part, whichis shorter, presents a few examples of non-equilibrium models or situations,some of them being chosen outside physics. This part aims at illustratinghow a physicist’s approach could be of interest to describe different types of complex systems.

3



2 Equilibrium statistical physics

In this section, we shall discuss some of the fundamentals of equilibriumstatistical physics. For further reading on this topic, we refer the reader tostandard textbooks, like e.g. [1, 2, 3].

2.1 Microscopic dynamics of a physical system

2.1.1 A particle attached to a spring

As an elementary example, consider a particle constrained to move on aone-dimensional horizontal axis x, and attached to a spring, the latter beingpinned to a rigid wall. We consider the position x(t) of the particle at timet, as well as its velocity v(t). The force F exerted by the spring on theparticle is given by

F = −k(x −x0), (1)

where x0 corresponds to the position of repose of the particle, for which theforce vanishes. For convenience, we shall in the following choose the originof the x axis such that x0 = 0.

From the basic laws of classical mechanics, the motion of the particle isdescribed by the evolution equation:

mdvdt

= F (2)

where m is the mass of the particle. We have neglected all friction forces, sothat the force exerted by the spring is the only horizontal force (the gravityforce, as well as the reaction force exerted by the support, do not havehorizontal components in the absence of friction). In terms of x variable,the equation of motion (2) reads

md2xdt2 = −kx. (3)

The generic solution of this equation is

x(t) = A cos(ωt + φ), ω =

k

m. (4)

The constants A and φ are determined by the initial conditions.

Hamiltonian reformulation of the problem

Let us introduce the momentum p = mv, and the kinetic energy E c = 12 mv 2.

In terms of momentum, the kinetic energy reads E c = p2/ 2m. The potential

4



Given that ω2 = k/m , one nds

E (t) =12

kA2 sin2(ωt + φ) + cos 2(ωt + φ) =12

kA2 (12)

which is indeed a constant.

Phase-space representation

Physical space is described in the example by the coordinate x. The equa-tions of motion (7) allow the position and momentum of the particle to bedetermined at any time once the initial position and momentum are known.So it is interesting to introduce an abstract representation space contain-ing both position and momentum. In this example, it is a two-dimensionalspace, but it could be of higher dimension in more general situations. Thisrepresentation space is often called “phase space”. For the particle attachedto the spring, the trajectories in this phase space are ellipses. Rescallingthe coordinates in an appropriate way, one can transform the ellipse into acircle, and the energy can be identied with the square of the radius of thecircle. To illustrate this property, let us dene the new coordinates X andY in phase-space as

X = k2 x, Y =p

√2m. (13)

Then the energy E can be written as

E =p2

2m+

12

kx2 = X 2 + Y 2. (14)

As the energy is xed, the trajectory of the particle is a circle of radius √E in the ( X, Y )-plane.

Time reversibility

To illustrate the meaning of time reversibility, let us imagine that we lmthe motion of the particle with a camera, and that we project it backward.If the backward motion is also a possible motion, meaning that nothing isunphysical in the backward projected movie, then the equations of motionare time-reversible.

More formally, we consider the trajectory x(t), t = 0 , . . . , t 0, and denethe reversed time t = t0 − t. Starting from the equations of motion (7)expressed with t, x and p, time reversal is implemented by replacing t with

6



t0

−t , x with x and p with

− p , yielding

−dxdt

= −∂H ∂p

,dpdt

= −∂H ∂x

. (15)

Changing the overall sign in the rst equation, one recovers Eq. (7) forthe primed variables, meaning that the time-reversed trajectory is also aphysical trajectory.

Note that time-reversibility holds only as long as friction forces are ne-glected. The latter break time reversal invariance, and this explains why oureveryday-life experience seems to contradict time reversal invariance. Forinstance, when a glass falls down onto the oor and breaks into pieces, itis hard to believe that the reverse trajectory, in which pieces would cometogether and the glass would jump onto the table, is also a possible trajec-tory, as nobody has ever seen this phenomenon occur. In order to reconcilemacroscopic irreversibility and microscopic reversibility of trajectories, thepoint of view of statistical physics is to consider that the reverse trajectoryis possible, but has a very small probability to occur as only very few initialconditions could lead to this trajectory. So in practice, the correspondingtrajectory is never observed.

2.1.2 Many-particle system

In a more general situation, a physical system is composed of N particles

in a 3-dimensional space. The position of particle i is described by a vectorx i , and its velocity by v i , i = 1 , . . . , N . In the Hamiltonian formalism, itis often convenient to introduce generalized coordinates and momenta q jand p j which are scalar quantities, with j = 1 , . . . , 3N : (q1, q2, q3) are thecomponents of the vector x 1 describing the position of particle 1, ( q4, q5, q6)are the component of x 2, and so on. Similarly, ( p1, p2, p3) are the componentsof the momentum vector mv 1 of particle 1, ( p4, p5, p6) are the components of mv 2, etc. With these notations, the Hamiltonian of the N -particle systemis dened as

H (q1, . . . , q3N , p1, . . . , p 3N ) =3N

j =1

p2 j

2m+ U (q1, . . . , q3N ). (16)

The rst term in the Hamiltonian is the kinetic energy, and the last one isthe potential (or interaction) energy. The equations of motion read

dq j

dt=

∂H ∂p j

,dp j

dt= −

∂H ∂q j

, j = 1 , . . . , 3N. (17)

7



time. Now assume that there is an observable ε i on each site i. It could forinstance the height of the site, or any arbitrary observable that characterizesthe state of the particle when it is at site i.

A natural question would be to know what the average value

ε =1T

T

t=1εi(t ) (19)

is after a large observation time T . Two different answers to this questioncould be given:

•Simulate the dynamics of the model on a computer, and measure di-rectly ε .

• Use the concept of probability as a shortcut, and write

ε =L

i=1

P i εi (20)

where the probability P i to be on site i is dened as

P i =time spent on site i

total time T , (21)

namely the fraction of time spent on site i.

The probability P i can be calculated or measured by simulating the dynam-ics, but it can also be estimated directly: if the particle has turned a lot of times around the ring, the fraction of time spent on each site is the same,P i = 1 /L . Hence all positions of the particle are equiprobable, and theaverage value ε is obtained as a at average over all sites.

2.2.2 Fondamental postulate of equilibrium statistical physics

We consider a physical system composed of N particles. The microscopicconguration of the system is described by ( x

i, p

i= mv

i), i = 1 , . . . , N , or

s i = ±1, i = 1 , . . . , N , for spin systems.The total energy E of the system, given by

E =N

i=1

p 2i

2m+ U (x 1, . . . , x N ) (22)

9



or

E = −J i,j

s i s j −hN

i=1

s i , (23)

is constant as a function of time. Accordingly, starting from an initial condi-tion with energy E , the system can only visit congurations with the sameenergy. In the absence of further information, it is legitimate to postulatethat all congurations with the same energy as the initial one have the sameprobability to be visited. This leads us to the fondamental postulate of equilibrium statistical physics :

All congurations with a given energy E have the same probability. Other congurations have zero probability.

The corresponding probability distribution is called the microcanonical dis-tribution or microcanonical ensemble for historical reasons (a probabilitydistribution can be thought of as describing an innite set of copies –anensemble– of a given system).

A quantity that plays an important role is the volume, or area, occupiedby all congurations with energy E . For systems with discrete congurations(spins), Ω( E ) is the number of congurations with energy E . The Boltzmannentropy is dened as

S (E ) = kB lnΩ(E ), (24)

where kB = 1 .3810− 23J/K is the Boltzmann constant. This constant hasbeen introduced both for historical and practical reasons, but from a theo-retical viewpoint, its specic value plays no role, so that we shall set its valueto kB = 1 in the following (this could be done for instance by working withspecic units of temperature and energy such that kB = 1 in these units).Let us now give a few simple examples of computation of the entropy.

2.2.3 Computation of Ω(E ) and S (E ): some simple examples

Paramagnetic spin model

We consider a model of independent spins, interacting only with an ex-

ternal eld. The corresponding energy is given by

E = −hN

i=1

s i , si = ±1 (25)

The phase space (or here simply conguration space) is given by the set of values s ii=1 ,...,N . The question is to know how many congurations there

10



are with energy E . It is easily seen that xing the energy E amounts toxing the magnetization M = N i=1 s i . Let us denote as N + the numberof spins with value +1 (’up’ spins). The magnetization is given by M =N + −(N −N + ) = 2 N + −N , so that xing M is in turn equivalent to xingN + . From basic combinatorial arguments, the number of congurations witha given number of ’up’ spins is given by

Ω =N !

N + !(N −N + )!. (26)

Using the relation

N + =12

N −E h

, (27)

one can express Ω as a function of E :

Ω(E ) =N !

12 (N −E/h ) ! 1

2 (N + E/h ) !(28)

The entropy S (E ) is given by

S (E ) = ln Ω( E )

= ln N !−ln12

N −E h

! −ln12

N +E h

! (29)

Using Stirling’s approximation, valid for large N

ln N ! ≈N ln N −N, (30)

one nds

S (E ) = N ln N −N + E/h

2ln

N + E/h2 −

N −E/h2

lnN −E/h

2. (31)

Perfect gas of independent particles

We consider a gas of independent particles conned into a cubic containerof volume V = L3. The generalized coordinates q j satisfy the constraints

0 ≤q j ≤L, j = 1 , . . . , L . (32)

The energy E comes only from the kinetic contribution:

E =3N

j =1

p2 j

2m(33)

11



The accessible volume in phase space is the product of the accessible vol-ume for each particle, times the area of the hypersphere of radius √2mE ,embedded in a 3N-dimensional space. The area of the hypersphere of radiusR in a D -dimensional space is

S D (R) =Dπ D/ 2

Γ D2 + 1

RD − 1, (34)

where Γ( x) = ∞

0 dt t x− 1e− t is the Euler Gamma function (a generalizationof the factorial to real values, satisfying Γ( n) = ( n −1)! for integer n ≥1).So the accessible volume Ω( E ) is given by

Ω(E, V ) = L3N 3Nπ3N/ 2

Γ 3N 2 + 1 √2mE 3N − 1

=3Nπ 3N/ 2

Γ 3N 2 + 1

√2m3N − 1

V N E 3 N − 1

2 (35)

The corresponding entropy reads, assuming N 1,

S (E, V ) = ln Ω( E ) = S 0 +3N 2

ln E + N ln V (36)

with

S 0

= ln3Nπ 3N/ 2

Γ 3N 2 + 1√2m

3N (37)

Note that in principle, some corrections need to be included to take intoaccount quantum effects, namely the fact that quantum particles are undis-tinguishable. This is important in order to recover the extensivity of theentropy, that is, the fact that the entropy is proportional to the number N of particles. In the present form, N ln N terms are present, making the en-tropy grow faster than the system size. This is also related to the so-calledGibbs paradox. However, we shall not describe these effects in more detailshere, and refer the reader to standard textbooks.

2.2.4 Distribution of energy over subsystems and statistical tem-perature

Let us consider an isolated system, with xed energy and number of parti-cles. We then imagine that the system is partitioned into two subsystems S 1and S 2, the two subsystems being separated by a wall which allows energyexchanges, but not exchanges of matter. The total energy of the system

12



E = E 1 + E 2 is xed, but the energies E 1 and E 2 uctuate due to thermalexchanges.

For a xed energy E , let us evaluate the number Ω( E 1|E ) of congura-tions of the system such that the energy of S 1 has a given value E 1. In theabsence of long-range forces in the system, the two subsystems can be con-sidered as statistically independent (apart from the total energy constraint),leading to

Ω(E 1|E ) = Ω 1(E 1)Ω2(E −E 1), (38)

where Ωk (E k ) is the number of congurations of S k .The most probable value E 1 of the energy E 1 maximizes by denition

Ω(E 1|E ), or equivalently ln Ω( E 1|E ):

∂ ∂E 1 E 1

lnΩ(E 1|E ) = 0 . (39)

Reporting Eq. (38) into Eq. (39), one nds

∂ ln Ω1

∂E 1 E 1=

∂ ln Ω2

∂E 2 E 2 = E − E 1. (40)

Thus it turns out that two quantities dened independently in each subsys-tem are equal at equilibrium. Namely, dening

β k =∂ ln Ωk

∂E k E k, k = 1 , 2, (41)

one has β 1 = β 2. This is the reason why the quantity β k is called thestatistical temperature of S k . In addition, it can be shown that for largesystems, the common value of β 1 and β 2 is also equal to

β =∂S ∂E

(42)

computed for the global isolated system.To idenfy the precise link between β and the standard thermodynamic

temperature, we notice that in thermodynamics, one has for a system thatexchanges no work with its environment:

dE = TdS, (43)

which indicates that β = 1 /T (we recall that we have set kB = 1). Thisis further conrmed on the example of the perfect gas, for which one ndsusing Eq. (36)

β ≡∂S ∂E

=3N 2

1E

, (44)

13



or equivalently

E = 3N 2

1β

. (45)

Besides, one has from the kinetic theory of gases

E =3N 2

T (46)

(which is nothing but equipartition of energy), leading again to the identi-cation β = 1 /T . Hence, in the microcanonical ensemble, one genericallydenes temperature T through the relation

1T =

∂S ∂E . (47)

We now further illustrate this relation on the example of the paramagneticcrystal that we already encountered earlier. From Eq. (31), one has

1T

=∂S ∂E

=1

2hln

N −E/hN + E/h

. (48)

This last equation can be inverted to express the energy E as a function of temperature, yielding

E = −Nh tanhhT

(49)

This relation has been obtained by noticing that x = tanh y is equivalent to

y =12

ln1 + x1 −x

(50)

In addition, from the relation E = −Mh , where M = N i=1 s i is the total

magnetization, one obtains as a byproduct

M = N tanhhT

. (51)

2.3 Equilibrium system in contact with its environment

2.3.1 Exchanges of energy

Realistic systems are most often not isolated, but they rather exchangeenergy with their environment. A natural idea is then to describe the system

S of interest as a macroscopic subsystem of a large isolated system S + R,where Ris the environment, or energy reservoir. The total energy E tot =

14



E + E R is xed. A conguration C tot of the total system can be written asC tot = ( C, C R ), where C is a conguration of S and C R is a congurationof R. The total system S + Ris isolated and at equilibrium, so that it canbe described within the macrocanonical framework:

P tot (C tot ) =1

Ωtot (E tot ), C tot = ( C, C R ). (52)

To obtain the probability of a conguration C of S , one needs to sumP tot (C tot ) over all congurations C R of Rcompatible with the total energyE tot , namely

P (C ) =C R :E R = E tot − E (C )

P tot

(C, C R

) =ΩR (E tot

−E (C ))

Ωtot (E tot ). (53)

We introduce the entropy of the reservoir S R (E R ) = ln Ω R (E R ). Under theassumption that E (C ) E tot , one has

S R (E tot −E (C )) ≈S R (E tot ) −E (C )∂S R∂E R E tot

. (54)

One also has∂S R∂E R E tot ≈

∂S R∂E R E R

=1T

(55)

where T is the temperature of the reservoir. Altogether, we have

P (C ) =ΩR (E tot )Ωtot (E tot )

e− E (C )/T . (56)

Introducing the partition function

Z =C

e− E (C )/T , (57)

one can eventually write the distribution P (C ) in the form

P (C ) =1Z e

− E (C )/T

, (58)

which is the standard form of the canonical distribution. At zero tempera-ture ( T = 0), only the congurations of minimal energy are occupied. Thusthe systems minimizes its energy in this case.

15



The partition function Z is a useful tool in statistical physics. For in-stance, the average energy E can be easily computed from Z :

E =C

P (C ) E (C ) =C

E (C )1Z

e− E (C )/T

=1Z

C

E (C ) e− βE (C )

= −1Z

∂Z ∂β

= −∂ ln Z

∂β (59)

Instead of Z , one may also use the “free energy” F dened as

F = −T ln Z. (60)

Let us give a simple example of computation of Z , in the case of the para-magnetic spin model. The partition function is given by

Z =s i = ± 1

e− βE (s i ) , (61)

with E (s i) = −h N i=1 s i . Hence one has

Z =s i = ± 1

eβh P N i =1 s i

=s i = ± 1

N

i=1

eβhs i =N

i=1 s= ± 1eβhs (62)

so that one ndsZ = eβh + e− βh N

. (63)

Turning to the average energy, one has

E = −∂ ln Z

∂β = −N

∂ ∂β

ln eβh + e− βh , (64)

so that one obtains, recalling that β = 1 /T ,

E = −Nh tanhhT

. (65)

It is interesting to note that the above equation has exactly the same form asEq. (49), provided that one replaces E , which is xed in the microcanonical

16



ensemble, by its average value E in the canonical ensemble. This propertyis an example of a general property called the “equivalence of ensemble”: inthe limit of large systems, the relations between macroscopic quantities arethe same in the different statistical ensembles, regardless of which quantityis xed and which one is uctuating through exchanges with a reservoir.The interpretation of this important property is basically that uctuatingobservable actually have very small relative uctuations for large systemsizes. This property is also deeply related to the law of large numbers andto the central limit theorem. Indeed, the relative uctuations (quantied bythe variance) of a sum of independent and identically distributed randomvariables go to zero when the number of terms in the sum goes to innity.

Another example where the computation of Z is straightforward is theperfect gas. In this case, one has

Z = L

0dq1 . . .

3L

0dq3N

∞

−∞dp3N e− β P 3 N

j =1 p2j / 2m

= L3N 3N

j =1 ∞

−∞dp j e− βp 2

j / 2m

= V N ∞

−∞dp e− βp 2 / 2m

3N (66)

Given that

∞

−∞dp e− βp 2 / 2m = 2πmβ , (67)

one nds

Z = V N 2πmβ

3 N 2

. (68)

Computing the average energy leads to

E = −∂ ln Z

∂β =

3N 2β

=32

NT (69)

yielding another example of ensemble equivalence, as this result has the

same form as Eq. (45).

2.3.2 Canonical entropy

As we have seen above, the microcanonical entropy is dened as S (E ) =lnΩ(E ). This denition is clearly related to the equiprobability of accessi-ble microscopic congurations, since it is based on a counting of accessible

17



congurations. A natural question is then to know how to dene the entropyin more general situations. A generic denition of entropy has appeared ininformation theory, namely:

S = −C

P (C ) ln P (C ) (70)

where the sum is over all accessible congurations of the system. Thisentropy is called Boltzmann-Gibbs, von Neumann or Shannon entropy de-pending on the context. This denition of entropy is moreover consistentwith the microcanonical one: if P (C ) = 1 / Ω(E ) for congurations of energyE , and P (C ) = 0 otherwise, one nds:

S = −C :E (C )= E

1Ω(E )

ln 1Ω(E )

= lnΩ( E ) (71)

In this general framework, the canonical entropy reads

S can = −C

P can (C ) ln P can (C )

= −C

1Z

e− βE (C ) ln1Z

e− βE (C )

=

C

1Z

e− βE (C ) (ln Z + βE (C ))

= ln Z + β E . (72)

Recalling that the free energy F is dened as F = −T ln Z , one thus hasT S = −F + E , which is nothing but the well-known relation F = E −T S . Another standard thermodynamic relation may be found using E =

−∂ ln Z/∂β :

S = ln Z −β ∂ ln Z

∂β

= ln Z + T ∂ ln Z

∂T

= ∂ ∂T

(T ln Z )

so that one nds the standard thermodynamic relation

S can = −∂F ∂T

. (73)

18



2.3.3 Exchanges of particles with a reservoir: the grand-canonicalensemble

Similarly to what was done to obtain the canonical ensemble from the micro-canonical one by allowing energy exchanges with a reservoir, one can furtherallow exchanges of particles with a reservoir. The corresponding situationis called the grand-canonical ensemble.

We thus consider a macroscopic system S exchanging both energy andparticles with a reservoir R. The total system S + Ris isolated with totalenergy E tot and total number of particles N tot xed:

E R + E S = E tot , N R + N S = N tot (74)

Generalizing the calculations made in the canonical case, one has

P GC (C ) = K ΩR (E R , N R )= K ΩR (E tot −E (C ), N tot −N (C ))= K exp( S R (E tot −E (C ), N tot −N (C )))

As E (C ) E tot and N (C ) N tot , one can expand the entropy S R (E tot −E (C ), N tot −N (C )) to rst order:

S R (E tot −E (C ), N tot −N (C )) = S R (E tot , N tot ) (75)

−E (C )∂S R∂E R E tot ,N tot −N (C )

∂S R∂N R E tot ,N tot

As before, the derivative ∂S R /∂E R is identied with 1 /T . We also introducea new parameter, the chemical potential µ, dened as:

µ = −T ∂S R∂N R

. (76)

Similarly to the temperature which takes equal values when subsystemsexchanging energy have reached equilibrium, the chemical potential takesequal values in subsystems exchanging particles, when equilibrium is at-tained. Gathering all the above results and notations, one nds that

P GC (C ) =1

Z GC exp −

1T

E (C ) +µT

N (C ) (77)

which is the standard form of the so-called grand-canonical distribution. Thenormalization constant Z GC is called the grand-canonical partition function.

19



2.4 Phase transitions

2.4.1 Example of the Ising model

Phase transitions correspond to a sudden change of behavior of the systemwhen varying an external parameter across a transition point. This couldbe of interest in complex systems well-beyond physics, and is genericallyassociated with collective effects. To illustrate this last property, let usbriey come back to the paramagnetic model dened in Sect. 2.2.3, forwhich the mean magnetization per spin is given by

m ≡M N

= tanhhT

. (78)

The magnetization is non-zero only if there is a non-zero external eld whichtends to align the spins. A natural question is thus to know whether onecould obtain a non-zero magnetization by including interactions betweenspin tending to align spins between them (and not with respect to an externalsource). In this spirit, let us consider the standard (interaction) energy of the Ising model

E Ising = −J i,j

s is j , J > 0. (79)

This interaction energy is minimized when all spins are parallel. To computethe mean magnetization per spin, one would need to compute either thepartition function in presence of a external magnetic eld and take thederivative of the free energy with respect to the eld, or to compute directlythe mean magnetization from its denition. In any case, this is a verycomplicated task as soon as the space dimension D is larger than one, andthe exact calculation has been achieved only in dimensions one and two.The results can be summarized as follows:

• D = 1: m = 0 for all T > 0, so that there is no phase transition atnite temperature. Calculations are relatively easy.

• D = 2: there is a phase transition at a nite critical temperature T c,

namely m = 0 for T ≥ T c and m = 0 for T < T c. Calculations arehowever very technical.

• D ≥ 3: no analytical solution is known, but numerical simulationsshow that there is a phase transition at a nite temperature that de-pends on D .

20



2.4.2 Ising model in fully connected geometry

An interesting benchmark model, which can be shown analytically to exhibita phase transition, is the fully connected Ising model, whose energy is denedas

E fc = −J N

i<js i s j + E 0, (80)

where the sum is over all pairs of spins in the system. The 1 /N prefactoris included in order to keep the energy per spin nite in the large N limit.The term E 0 is added for later convinience, and is arbitrary at this stage (itdoes not modify the canonical distribution). Considering the magnetizationM = N

i=1 s i , one has, given that s2i = 1,

M 2 = 2i<j

s i s j + N (81)

from which one nds

E fc = −J

2N (M 2 −N ) = −

J 2N

M 2 +J 2

+ E 0. (82)

Choosing E 0 = −J/ 2, and introducing m = M/N , one nds

E fc = −J 2

Nm 2. (83)

A good way to detect the phase transition is to compute the distributionP (m) of the magnetization:

P (m) =1Z

C :m (C )= m

e− βE (C )

=1Z

Ω(m) e12 βJNm 2

=1Z

eS (m )+ 12 βJNm 2

(84)

where Ω(m) = eS (m )

is the number of congurations with magnetization m.Using the relation

Ω(m) =N !

N + !N − !(85)

withN + =

N 2

(1 + m), N − =N 2

(1 −m), (86)

21



one obtains for S (m) = lnΩ( m)

S (m) = −N 1 + m

2ln

1 + m2

+1 −m

2ln

1 −m2

(87)

Hence from Eq. (84) it turns out that P (m) can be written as

P (m) = e− Nf (m ) (88)

with f (m) given by

f (m) =1 + m

2ln

1 + m2

+1 −m

2ln

1 −m2 −

J 2T

m2. (89)

The function f (m) is called a large deviation function, or a Landau free en-ergy function. Hence the magnetization m 0 that maximizes the probabilitycorresponds to a minimum of f (m). Moreover, uctuations around m 0 areexponentially suppressed with N . For high temperature T , the term J/T is small, and the entropic contribution to f (m) should dominate, leading tom0 = 0. To understand what happens when temperature is progressivelylowered, it is interesting to expand f (m) for small values of m, up to orderm4, leading to:

f (m) = −ln2 +12

1 −J T

m2 +1

12m4 + O(m6). (90)

One can then distinguish two different cases:

• If T ≥T c ≡J , f (m) has only one minimum, for m = 0.

• If T < T c, f (m) has two symmetric minima ±m0. These minima areobtained as solutions of df /dm = 0:

df dm

= 1 −J T

m +13

m3 = − 1 −J T

m +13

m3 = 0 . (91)

The non-zero solutions are m =

±m 0 with

m0 = 3 J T −1 = √3

T c −T T

1/ 2. (92)

It can be checked easily that the solution m = 0 corresponds in this case toa local maximum of f (m), and thus to a local minimum of P (m).

22



Hence, there is a phase transition at T = T c

≡J , T c being called the crit-

ical temperature. The most probable magnetization m 0 is called the “orderparameter of the phase transition”, as the phase transition is characterizedby the onset of non-zero values of m 0. In addition, the order parametervaries as m0 (T c −T )β for T close to T c, with β = 1 / 2 here. The ex-ponent β is called a critical exponent, and the value β = 1 / 2 is called the“mean-eld value” of β , for reasons that will become clear in the next sec-tion. The notation β is standard for the critical exponent associated to theorder parameter, and should not be confused with the inverse temperatureβ = 1 /T .

An important remark is that the average value m of the magnetizationis still zero for T < T c, since the two values

±m0 of the magnetization

have the same probability. However, for a large system, the time neededto switch between states m 0 and −m0 becomes very large, and the time-averaged magnetization over a typical observation time window is non-zero,and equal either to m0 or to −m0.

2.4.3 Ising model with nite connectivity

We now come back to the nite-dimensional Ising model. As mentionedabove, the analytical solution is hard to obtain in dimension D = 2, andis not known in higher dimensions. However, useful approximations havedevelopped, the most famous one being called the mean-eld approximation.

The reason why the fully connected model can be easily solved analyti-cally is that its energy E is a function of the magnetization m only, as seenin Eq. (83). When the model is dened on a nite-dimensional lattice, thisproperty is no longer true, and the energy reads:

E = −J 2

N

i=1

s i j V (i)

s j . (93)

where V (i) is the set of neighboring sites of site i. The factor 1 / 2 comesfrom the fact that a given link of the lattice now appears twice in the sum.This last expression can be rewritten as

E = −DJ N

i=1

s i s j V (i) , (94)

s j V (i) being the local average magnetization of the set of neighbors V (i):

s j V (i) =1

2D j V (i)

s j . (95)

23



The parameter D is the space dimension, and the number of neighbors of agiven site i is 2D , given that we consider hypercubic lattices (square latticein D = 2, cubic lattice in D = 3,...).

As a rst approximation, one could replace the average magnetizationover the set of neighbors by the global magnetization per spin of the wholesystem, m = N − 1 N

i=1 s i :s j V (i) →m. (96)

This approximation, which is expected to be valid if D is large enough, leadsto the following expression of the energy

E

≈E mf =

−DJm

N

i=1

s i =

−DJNm 2. (97)

Then E mf depends only on the magnetization m, and it has a form similarto the energy E fc of the fully connected model. One can dene an effectivecoupling J mf = 2 DJ so that the form of the two energies become exactlythe same, namely

E mf = −12

J mf Nm 2. (98)

Now it is clear that the results of the fully connected model can be applied tothe present mean-eld approximation, yielding a phase transition at T mf

c =J mf = 2 DJ . For T > T mf

c , m = 0 while for T < T mf c , m (T mf

c −T )

1/ 2

. Qualitatively, the approximation is expected to be valid for largespace dimension D . It can be shown, using more involved arguments, thatfor D ≥4, the approximation is semi-quantitatively valid, in the sense thatthe value β = 1 / 2 of the critical exponent, obtained from the approximation,is correct. However, the value of the critical temperature T mf

c is not correctlypredicted by the mean-eld approximation, namely T c = T mf

c . For D < 4,the value of β differs from the mean-eld value 1 / 2, and the mean-eldapproximation breaks down. For D = 3, one has β ≈ 0.31, and for D = 2,β = 1 / 8. Finally, for D = 1, m = 0 except for T = 0, so that the exponentβ is not dened.

The discrepancy mentioned above between mean-eld predictions and

results in low-dimensional systems mainly comes from the presence of uc-tuations of the local magnetization j V (i) s j . Since on the other hand exactsolutions are very hard to obtain, there is need for a different approach, thatcould be generic enough and could be centered on the issue of correlation,which is at the heart of the difficulties encountered. This is precisely theaim of the renormalization group approach.

24



2.4.4 Renormalization group approach: a brief introduction

A standard observation on nite dimensional systems exhibiting a continu-ous phase transition is that the correlation length diverges when the tem-perature approaches the critical temperature T c. The correlation length isdened through the correlation function

C ij = (s i −m0)(s j −m0) = s i s j −m20. (99)

As soon as the length r = dij is large with respect to the lattice spacing a,the correlation function is generally isotropic, C ij = C (r ). In addition, thelarge distance behavior of C (r ) is often of the form

C (r )1

r α e− r/ξ , α > 0, (100)

which denes the correlation length ξ. The latter diverges for T →T c. Thisis the reason why direct calculations in the range T ≈T c are very difficult,due to the strong correlation between spins. A natural idea is to search foran approach that could reduce in some way the correlation.

This is basically the principle of the renormalization group (RG) ap-proach, in which one progressively integrates over the small scale degrees of freedom. The idea is that at the critical point, structures are present at allscales, from the lattice spacing to the system size. A RG transform mayintuitively be thought as defocusing the picture of the system, so that ne

details become blurred. This method is actually very general, and couldbe relevant in many elds of complex system sciences, given that issues likelarge scale correlations and scale invariance or fractals are often involved incomplex systems.

For deniteness, let us however consider again the Ising model. To im-plement the RG ideas in a practical way, one could make blocks of spins anddene an effective spin for each block, with effective interactions with theneighboring blocks. The effective interactions are dened in such a way thatthe large scale properties are the same as for the original (non-renormalized)model. This is done in practice by conserving the partition function, namelyZ = Z (the prime generically denotes renormalized quantities in the follow-

ing). One would then like to dene a renormalized interaction constant J such that

H = −J B 1 ,B 2

S B 1 S B 2 (101)

where B1 and B2 are generic labels for the blocks (the sites of the renormal-ized lattice). The problem is that very often, the RG transform generates

25



new effectives couplings, like next-nearest-neighbor couplings, that were ab-sent in the original model, and the number of couplings keeps increasingwith the number of iterations of the RG transform. However, in some sim-ple cases, the transformation can be performed exactly, without increasingthe number of coupling constants, as we shall see later on.

Yet, let us rst emphasize the practical interest of the RG transform.We already mentioned that one of the main difficulties comes from the pres-ence of long-range correlations close to the critical point. Through the RGtransform, the lattice spacing becomes a = 2 a (if one makes blocks of linearsize 2a). On the contrary, the correlation length remains unchanged, sincethe large scale properties remain unaffected by the RG transform. Hencethe correlation length expressed in unit of the lattice spacing, namely ξ/a ,decreases by a factor of 2 in the transformation, to become

ξa

=12

ξa

. (102)

Thus upon iterations of the RG transform, the effective hamiltonian becomessuch that ξ a , so that approximation schemes can be used. One thenneeds to follow the evolution of the coupling constant J under iterations.This is called the renormalization ow.

An explicit example can be given with the one-dimensional Ising chain,using a specic RG transform called decimation procedure. We start withthe energy (or Hamiltonian)

H = −J N

i=1

s i s i+1 + Nc (103)

where periodic boundary conditions are understood. The basic idea of thedecimation procedure is to perform, in the partition function, a partial sumover the spins of –say– odd indices in order to dene renormalized couplingconstants J and h . Then summing over the values of the spins with evenindices yields the partition function Z of the renormalized model, which isby denition of the renormalization procedure equal to the initial partitionfunction Z . To be more explicit, one can write Z as

Z =s 2 j s 2 j +1

exp −β

N

i=1H i,i +1 (s i , s i+1 ) (104)

and rewrite the above equality in the following form:

Z =s 2 j

exp −β N/ 2

j =1

H j,j +1 (s2 j , s2 j +2 ) (105)

26



where H (

s2 j

) = N/ 2

j =1 H j,j +1 (s2 j , s2 j +2 ) is the renormalized Hamiltonian,dened by

exp −βH (s2 j ) =s 2 j +1

exp −β N/ 2

j =1

H j,j +1 (s2 j , s2 j +2 ) . (106)

This last relation is satised if, for any given j = 1 , . . . , N / 2, and any givenvalues of s2 j and s2 j +2 ,

exp −βH j,j +1 (s2 j , s2 j +2 ) (107)

=

s 2 j +1 = ± 1

exp(−β [H 2 j, 2 j +1 (s2 j , s2 j +1 ) + H 2 j +1 ,2 j +2 (s2 j +1 , s2 j +2 )])

Assuming that H j,j +1 (s2 j , s2 j +2 ) takes the form

H j,j +1 (s2 j , s2 j +2 ) = −J s2 j s2 j +2 + c , (108)

one obtains

exp(βK s2 j s2 j +2 −βc) =s 2 j +1 = ± 1

exp(βJ s 2 j s2 j +1 ). (109)

We introduce the reduced variable

u = e− 4βJ , (110)

and Eq. (109) leads to the following recursion relation:

u =4u

(1 + u)2 . (111)

Let us denote as ξnd the non-dimensionalized correlation length

ξnd =ξa

. (112)

Then from Eq. (102) the recursion relation for ξnd reads

ξnd =12

ξnd , (113)

from which one deduces that the xed points of the renormalization proce-dure must satisfy either ξnd = ∞or ξnd = 0. The latter is called the trivialxed point, as it corresponds to the limit situation where no correlation

27



is present in the system. In constrast, the xed point ξnd =

∞is called

the critical xed point, as it corresponds to a situation of innite correla-tion, which is precisely the situation at the critical point. As ξnd decreasesthrough iteration of the RG transform, it is clear that the critical xedpoint is unstable through iteration of the RG transform, while the trivialxed point is stable.

Coming back to the iteration relation Eq. (111), let us rst look for thexed points of this equation, namely the solutions of

u =4u

(1 + u)2 . (114)

The value u = 0 is obviously a solution, and it is easy to check that u = 1is the other positive solution ( u = −3 is the third solution, but in viewof Eq. (110), we are seeking for positive solutions only). Then to identifywhich one of the two xed points is the critical point, we need to investigatethe stability of each xed point under iteration. The stability is studiedby introducing a small variation δu around a given xed point u1, namelyu = u1 ±δu, and writing the evolution equation for δu to leading order. Foru1 = 0, one nds, with u = δu,

δu =4δu

(1 + δu)2 ≈4δu, δu > 0, (115)

so that δu increases upon iteration: the xed point u1 = 0 is unstable, andthus corresponds to the critical xed point. Besides, the xed point u1 = 1is easily checked to be stable. Using u = 1 −δu, we have

1 −δu =4(1 −δu)(2 −δu)2 , (116)

leading after a straightforward calculation to

δu ≈14

δu2. (117)

Hence δu converges to 0 upon iteration, conrming the stability of the xedpoint u1 = 1. Coming back to the critical xed point, and recalling thedenition Eq. (110), one sees that u1 = 0 corresponds to an innite value of J/T . In the above framework, this case is interpreted as an innite couplinglimit, as the iteration was made on J . However, the xed point can bealso be interpreted as a zero-temperature xed point, keeping the couplingconstant J xed.

28



This one-dimensional example is of course only a very simple case, whichcan be solved through other more direct methods. However, it is a good il-lustration of the way the concept of RG can be implemented in practice. Intwo- or three-dimensional models, exact treatments like the above one aremost often not available. Yet, many approaches based on different approx-imation schemes have been developped. A typical situation in dimensionD > 1 is that there is a nite value K c of the ratio K = J/T which cor-responds to a critical xed point, and both values K = 0 and K = ∞correspond to trivial xed points, where no correlation is present. Quiteimportantly, linearizing the iteration equation in the vicinity of the criticalxed point allows the determination of the critical exponent β , as well asother critical exponents. In the Ising chain studied above, this is not pos-sible because the critical temperature is zero, so that there is no extendedtemperature region where the magnetization is non-zero. But this approachcan be used in higher dimension.

2.5 Disordered systems and glass transition

2.5.1 Disorder in complex systems: from social sciences to spin-glasses

Disordered systems are systems where each particle or agent has specicproperties, which are qualitatively the same for all particles or agents, butdiffer quantitatively from one to the other. In theoretical models, thesequantitative properties are most often drawn from a given probability dis-tribution for each particle or agent, and stay constant in time. Disorderedsystems should be very relevant in complex systems studies, like social sci-ence for instance, as each human being has its specic skills or tastes.

In a physical context, the concept of disordered system could seem lessnatural, since all particles within a given class are identical. In this case,disorder rather comes from the possibly random position of the particles. Astandard example is that of magnetic impurities (that carry a magnetic mo-ment, or spin) diluted in a non-magnetic material. The interaction betweenmagnetic atoms (which have to be described in the framework of quantummechanics) is mediated by the non-magnetic atoms, and acquires an oscil-latory behavior, depending on the distance r ij between the two spins:

H = −i,j

J (r ij )s i s j . (118)

The interaction constant J (r ij ) is a given function of the distance r ij , whichoscillates around 0, thus taking both positive and negative values. The

29



amplitude of the oscillations decays as a power-law of the distance. As thedistances between atoms are random, the interactions between atoms havea random sign, which is the basic property of spin-glasses.

2.5.2 Theoretical spin-glass models

In order to propose a simplied model for spin-glass materials, it has beenproposed to replace the position disorder by an interaction disorder, themagnetic atoms being now situated on a regular lattice. To this purpose,we consider an Ising-like model in D -dimensions, where the spins are placedat each node of a lattice. Spins on neighboring sites ( i, j ) interact through acoupling constant J ij , drawn from a distribution P (J ). These couplings are

called quenched disorder, as the J ij are kept constant in time. In D = 1, themodel is equivalent to the standard Ising model, up to a redenition of thespins. In D > 1, analytical solutions are not known, and results have thusbeen obtained through numerical simulations. A fully connected version,called the Sherrington-Kirkpatrick model, has been proposed and solved,but the techniques involved are already rather difficult, even at the levelof the fully connected model. The main qualitative picture emerging fromthese models is that below a given energy level, the phase space decomposesinto a lot of valleys, or metastable states, from which it takes a long time toescape.

2.5.3 The simplest disordered model: the Random Energy Model

A very simple disordered model, which already captures a lot of the phe-nomenology of realistic disordered systems, has been proposed by Derrida in1980, and has been called the Random Energy Model (REM) [6]. It can bethought of as a spin model, although spins do not play an essential role inthe description. Considering of system of N spins, the corresponding num-ber of conguration is 2 N . We shall label as α a microscopic conguration,α = s ii=1 ,...,N . To each conguration α is attached an energy E α , takenat random from a distribution P (E ):

P (E ) =1

√NπJ 2 exp −E 2

NJ 2 (119)

All the energies E α are independent random variables. We denote as n(E )dE the number of congurations with energy in the interval [ E, E + dE ], so thatn(E ) is the density of congurations with energy E . The density n(E ) is arandom quantity, but its uctuations are small if n(E ) is large, namely

30



n(E )

≈ n(E ) . By denition, P (E ) = n(E ) / 2N , so that n(E ) =

2N P (E ), leading to

n(E ) = exp N ln 2 −E 2

NJ 2

= exp N ln 2 −ε2

J 2(120)

where the energy density ε = E/N has been introduced. One sees that if ln 2−ε2/J 2 > 0, corresponding to |ε| < ε 0 = J √ln2, n(E ) is exponentiallylarge with N , so that there are a large number of congurations at energydensity ε, and the assumption n(E )

≈ n(E ) is justied. In contrast, if

ln 2−ε2/J 2 < 0, which corresponds to |ε| > ε 0, n(E ) is exponentially smallwith N . This means that in most samples, there are no congurations atenergy density |ε| > ε 0. The non-zero, but small value of n(E ) comes fromthe contribution to the average value of very rare samples, which includesome congurations with exceptionally low (or high) energy.

We can now evaluate the partition function of the REM, dened as

Z =2N

α =1e− E α /T . (121)

As all the energies E α are random variables, the partition function Z is alsoa random variable, which uctuates from one realization of the disorder toanother. Yet, we can evaluate the typical value of Z as follows:

Z ≈Z typ = ε0

− ε0

dε n(ε) e− Nε/T , (122)

with the notation ˜ n(ε) = Nn (Nε ). In the above equation, we have replacedn(ε) by n(ε) for |ε| < ε 0, and by 0 for |ε| > ε 0. We can then write

Z typ = ε0

− ε0

dε eNg (ε) (123)

withg(ε) = ln 2 −

ε2

J 2 −εT

. (124)

In the large N limit, we can evaluate Z typ through a saddle point calculation,namely

Z typ ≈eNg max (ε0 ) (125)

31



where gmax (ε0) is the maximum value of g(ε) over the interval [

−ε0, ε0]. Let

us rst consider the maximum ε of g(ε) over the entire real line. Takingthe derivative of g(ε), one has

g (ε) = −2εJ 2 −

1T

. (126)

From g (ε) = 0, we nd

ε = −J 2

2T 2. (127)

As g(ε) is a parabola, it is increasing for ε < ε and decreasing for ε > ε .If ε > −ε0, then gmax (ε0) = g(ε ), so that

Z typ ≈eNg (ε )

. (128)The condition ε > −ε0 translates into T > T g, where the so-called glasstransition temperature T g is dened as

T g =J

2√ln 2. (129)

For ε < −ε0, or equivalently T < T g, g(ε) is a decreasing function of ε overthe entire interval [ −ε0, ε0], so that gmax (ε0) = g(−ε0), and

Z typ ≈eNg (− ε0 ) . (130)

From these estimates of Z typ , one can compute the free energy F = −T ln Z typ ,and the entropy S = −∂F/∂T . For T > T g , one nds

F = −N T ln2 +J 2

4T , (131)

leading for the entropy to

S = N ln 2 −J 2

4T 2. (132)

For T < T g, we have

F = −T Ng(−ε0) = −T N ln 2 −ln2 + J T √ln 2 = −NJ √ln 2. (133)

The free energy does not depend on temperature in this range, so that thecorresponding entropy vanishes:

S = 0 , T < T g . (134)

32



It can also be checked that the entropy given in Eq. (132) for T > T gvanishes continuously for T →T g . Hence the temperature T g corresponds toa glass transition temperature, where the entropy goes to zero when loweringtemperature, and remains zero below T g . Actually, to make the statementsharper, only the entropy density S/N goes to zero for T < T g, in the inniteN limit. Computing subleading corrections to the entropy, one nds that theentropy S is independent of N , but non-zero, for T < T g . The entropy is thenintensive in this temperature range, meaning that only a nite number of congurations, among the 2 N ones a priori available, are effectively occupied:the system is quenched in the lowest energy congurations.

3 Non-equilibrium statistical physicsThe aim of the present lecture notes is of course not to review such a hugetopic as non-equilibrium statistical physics, but rather to briey highlightsome specic issues. In the present form, these notes essentially reectthe content of the oral lecture, and they still need to be completed. Notealso that the choice of the topics presented is obviously biased by my ownresearch activity. However, and even though many other choices could bemade, I think these topics are somehow representative of the contributionsof physicists to the modelization and understanding of complex systems.

3.1 Markovian stochastic processes and master equationLet us start with some basic considerations on stochastic processes. For moreadvance reading on this topic, we refer the reader to [5] for instance. Roughlyspeaking, a stochastic process is a dynamical process whose evolution isdetermined randomly to some extend, and also depends on the presentlyoccupied state and possibly on the history of the system. A Markovianstochastic process is a stochastic process for which the evolution dependsonly on the present conguration and not on past congurations. The evo-lution of Markovian stochastic processes is characterized by transition ratesW (C →C ), that are such that W (C →C )dt is the probability to go fromconguration C to conguration C in a time interval [ t, t + dt]. The rateW (C →C ) depends only on C and C , and neither on previously occupiedcongurations, nor on time t.

The probability P (C, t ) to be in conguration C at time t evolves ac-

33



cording to the so-called master equation:

dP dt

(C, t ) = −C = C

W (C →C )P (C, t ) +C = C

W (C →C )P (C , t ). (135)

The rst term in the r.h.s. corresponds to a loss term (sum of all the possi-bilities to exit conguration C ), while the second term is a gain term (sumof all the possibilities to arrive at conguration C starting from any otherconguration).

It is interesting to introduce a time-dependent entropy dened as

S (t) = −C

P (C, t ) ln P (C, t ). (136)

This denition closely follows the denition (70). Under the assumptionW (C →C ) = W (C →C ), which is a specic form of detailed balancerelation (a property related to time-reversibility), one can show that S (t) isincreasing with time. Let us start by computing the time-derivative of theentropy:

dS dt

= −C

dP dt

(C, t ) ln P (C, t ) −C

dP dt

(C, t ). (137)

The last term cancels due to the normalization condition C P (C, t ) = 1.Using the master equation, one has:

dS dt = − C ln P (C, t )

C = C −W (C →C )P (C, t ) + W (C →C )P (C , t )

= −C,C (C = C )

ln P (C, t ) −W (C →C )P (C, t )

+ W (C →C )P (C , t ) (138)

Exchanging the notations C and C in the last equation, we also have

dS dt

= −C,C (C = C )

ln P (C , t ) −W (C →C )P (C , t ) + W (C →C )P (C, t ) .

(139)Summing Eqs. (138) and (139), and using the reversibility property W (C →C ) = W (C →C ), we obtain

dS dt

=12

C,C (C = C )

ln P (C , t ) −ln P (C, t ) P (C , t ) −P (C, t ) W (C →C ).

(140)

34



As ln P (C , t )

−ln P (C, t ) and P (C , t )

−P (C, t ) have the same sign, one

concludes that dS dt ≥0. (141)

This is one possible statement, in the context of stochastic processes, of the second principle of thermodynamics. Moreover, in the stationary state,dS/dt = 0, and one necessarily has, for all pairs ( C, C ), either P st (C ) =P st (C ), or W (C →C ) = 0. One then recovers, in a stochastic framework,the fundamental postulate of equilibrium statistical mechanics, stating thatmutually accessible congurations have the same probability.

More generally, for Markovian stochastic processes described by the mas-ter equation (135), it is always possible to dene a functional S (P (C, t ))

that increases with time, without need for detailed balance or microre-versibility properties [5]. The general denition of S (P (C, t )) is

S (t) = −C

P (C, t ) lnP (C, t )P st (C )

, (142)

where P st (C ) is the stationary distribution. However, it turns out that thestationary distribution needs to be known in order to dene S . Hence, thefunctional S does not allow to determine the stationary solution.

3.2 Aging dynamics and trap models

Although many systems converge to a stationary state on times shorterthan or comparable to the observation time, it turns out that some systemsdo not reach a steady state and keep evolving on time scales that can bevery large compared to standard observation times. This is the case forinstance of glasses, which keep aging for years or more [7]. The same typeof mechanism is at play in laser cooling experiments [7]. It is also likely thataging mechanisms play a signicative role in many types of complex systems,like biological systems for instance. Even though the aging mechanismsmay be much more complex in these situations, it is certainly of interest toinvestigate one of the simplest known aging mechanisms.

Let us consider a model system in which to each conguration C isassociated in given lifetime τ . This lifetime τ is the mean time spent in con-guration C before moving to another conguration. As we consider onlytemporal aspects of the dynamics, and not other types of observables (en-ergy, magnetization,...), we simply label the congurations by their lifetimeτ . We then choose a simple form for the transition rate W (τ →τ ), namely:

W (τ →τ ) =1τ

ψ(τ ). (143)

35



The function ψ(τ ) is the (a priori) probability distribution of the congura-tions τ , meaning that the selected new conguration is chosen completelyat random. From the normalization condition

∞0 dτ ψ(τ ) = 1, we have

∞

0dτ W (τ →τ ) =

1τ

, (144)

so that the characteristic escape time from a conguration with lifetime τ is precisely τ . For simplicity, we also assume that all lifetimes τ are greaterthan a value τ 0, that we set to τ 0 = 1 in the following. The master equationthen reads:

∂P ∂t (τ, t ) = −P (τ, t )

∞

1 dτ W (τ →τ ) + ∞

1 dτ W (τ →τ )P (τ , t )

= −1τ

P (τ, t ) + ψ(τ ) ∞

1

dτ τ

P (τ , t ). (145)

At equilibrium, the probability to be in a conguration with lifetime τ isproportional to τ and to the a priori distribution ψ(τ ) of congurations:

P eq(τ ) =1τ

τψ(τ ), (146)

where τ is dened as

τ = ∞

1 dτ τ ψ(τ ). (147)An interesting situation appears when the distribution ψ(τ ) takes a power-law form, namely

ψ(τ ) =α

τ 1+ α , τ > 1, α > 0. (148)

This is realized for instance in the case of a particle trapped into potentialwells of random depth E , with an exponential distribution

ρ(E ) =1

E 0e− E/E 0 . (149)

The lifetime τ is given by the standard Arrhenius law

τ = τ 0 eE/T , (150)

where τ 0 = 1 is a microscopic time scale. Using the relation ψ(τ )|dτ | =ρ(E )|dE |, one precisely nds the form (148) for ψ(τ ), with α = T/E 0.

36



In the case α > 1, τ is nite, but if α

≤1 then τ is innite, so that

the equilibrium distribution does not exist (it is not normalizable). As aresult, no stationary state is reached, and the system keeps drifting towardcongurations with larger and larger lifetimes τ .

It is then of interest to determine the time-dependent probability distri-bution P (τ, t ) in the long-time regime. We postulate the following scalingform

P (τ, t ) =1t

φτ t

. (151)

From the normalization condition of P (τ, t ), one has

∞

1dτ P (τ, t ) =

1

t ∞

1dτ φ

τ

t= 1 , (152)

from which one gets, with the change of variable u = τ /t ,

∞

1/tdu φ(u) = 1 . (153)

As φ(u) does not depend explicitely on time t, the above condition cannotbe satised for all t. But we are looking for an asymptotic large- t solution,so that we impose that Eq. (153) is satised in the innite t limit, namely

∞

0du φ(u) = 1 . (154)

As a result, the scaling form (151) is an approximate solution that becomesexact when t → ∞. From Eq. (151), one obtains for the time derivative of P (τ, t ):

∂P ∂t

= −1t2 φ

τ t −

τ t3 φ

τ t

, (155)

where φ is the derivative of φ. Multiplying Eq. (145) by t2, one obtains,with the notations u = τ/t and v = τ /t ,

−φ(u) −uφ (u) = −1u

φ(u) + ψ(ut ) t

∞

1/t

dvv

φ(v). (156)

Using the specic form (148) of ψ(τ ), we nd

1 −1u

φ(u) + uφ (u) +α

(ut )1+ α t ∞

1/t

dvv

φ(v) = 0 . (157)

37



For the above equation to be well-dened in the innite t limit in which it issupposed to be valid, the explicit t-dependence has to cancel out. One thusneeds to have

∞

1/t

dvv

φ(v) tα , t → ∞, (158)

which requires that φ(v) has the following asymptotic form at small v:

φ(v) ≈φ0

vα , v →0. (159)

Here, φ0 is an arbitrary constant, to be determined later on from the nor-malization condition of φ(u). The master equation is then nally written asthe following differential equation:

1 −1u

φ(u) + uφ (u) +φ0

u1+ α = 0 . (160)

This equation is a linear differential equation, and its solution can be foundusing standard techniques. The solution of this equation satisfying the nor-malization condition (154) reads [8]

φ(u) =sin παΓ(α)

1u

e− 1/u 1/u

0dv vα − 1ev . (161)

It is rather easy to show that φ(u) u− α for u

→0 as expected, and that

φ(u) u− 1− α for u → ∞, leading for P (τ, t ) to

P (τ, t ) τψ(τ ), τ t, (162)P (τ, t ) ψ(τ ), τ t. (163)

These asymptotic behaviors can be interpreted rather easily: congurationswith lifetimes τ t have been visited a large number of times, so that theyare quasi-equilibrated; in constrast, congurations with lifetimes τ t havebeen visited at most once, and the precise value of τ is not yet felt by thedynamics ( τ appears as essentially innite).

3.3 Collective motion of active particlesActive “particles” are particles that are able to sustain a continuous motionthanks to some external energy input. This concept is used to model forinstance the motion of animals or bacteria. A simplied physicist’s model of active particle is a point-like particle with a velocity vector of constant mod-ulus, but arbitrary direction. When two (or more) active particles interact,

38



they tend to align their velocities. Such very simple physicists’ models havebeen studied through numerical simulations by Vicsek and coworkers, andlater by Gregoire and Chate [9, 10, 11]. They have found a transition fromdisordered motion when the density of active particles is low, to orderedcollective motion when the density is high. This transition exhibits someproperties similar to that of phase transitions observed in physical systems.It is also possible to develop analytical approaches, either by postulatingphenomenological equations of motion at the macroscopic scale (hydrody-namic equations) [12], or by using a Boltzmann equation approach to derivethese equations of motion [13]. We present here a brief summary of theresults obtained from this later approach.

3.3.1 Denition of the model

We consider self-propelled point-like particles moving on a continuous plane,with a velocity vector v of xed magnitude v0 (to be chosen as the velocityunit) in a reference frame. The velocity of the particles is simply dened bythe angle θ between v and an arbitrary reference direction. Particles movein straight line, following their velocity vector, until they experience eithera self-diffusion event (a random “kick”, Fig. 1(a)), or a binary collision thattends to align the velocities of the two particles –see Fig. 1(b). Self-diffusionevents are dened as follows: the velocity angle θ of any particle is changedwith a probability λ per unit time to a value θ = θ+ η, where η is a Gaussian

noise with distribution p0(η) and variance σ20 . Binary collisions, that are theonly interactions between particles, occur when the distance between twoparticles becomes less than d0 (in the following, we set d0 = 1

2 ). The velocityangles θ1 and θ2 of the two particles are then changed into θ1 = θ + η1 andθ2 = θ+ η2, as shown on Fig. 1(b). In the last expression, θ = arg( eiθ 1 + eiθ 2 )is the average angle, and η1 and η2 are independent Gaussian noises with thesame distribution p(η) and variance σ2. Note that these binary collisions aredifferent from the collisions in usual gases, as in this latter case, collisionsare ruled by energy and momentum conservation laws.

3.3.2 Description of the dynamics through a Boltzmann equation

A useful mathematical tool to describe statistically the dynamics of the sys-tem is the one-particle phase-space distribution f (r , θ , t ), namely the prob-ability to nd a particle at position r and with a velocity angle θ, at timet. The evolution of this one-particle phase-space distribution is ruled by the

39



η

η

ηθ

θ θθ ’ 1

2’

θ 2

11θ

2’

θ

Binary interactionsSelf−diffusion

Figure 1: Schematic representation of the dynamics of the model.

Boltzmann equation, which reads

∂f ∂t

(r , θ , t ) + e (θ) · f (r , θ , t ) = I dif [f ] + I col[f ]. (164)

The functionals I dif [f ] and I col[f ] respectively account for the self-diffusionand collision phenomena. The vector e (θ) is the unit vector in the directionθ. I dif [f ] is given by

I dif [f ] = −λf (r , θ , t ) + λ π

− πdθ

∞

−∞dη p0(η) (165)

×∞

m = −∞δ(θ + η −θ + 2 mπ )f (r , θ , t ).

The evaluation of the collision term I col[f ] is more subtle. We know thattwo particles collide if their distance becomes less than the interaction ranged0. In the frame of particle 1, particle 2 has a velocity v 2 = e (θ2) −e (θ1).Hence, particles that collide with particle 1 between t and t + dt are thosethat lie, at time t, in a rectangle of length |v 2|dt and of width 2 d0. Thisleads to

I col[f ] = −f (r , θ , t ) π

− πdθ |e (θ ) −e (θ)|f (r , θ , t ) (166)

+ π

− πdθ1

π

− πdθ2

∞

−∞dη p(η) |e (θ2) −e (θ1)|

× f (r , θ1, t )f (r , θ2, t )

∞

m = −∞δ(θ + η −θ + 2 mπ ),

with θ = arg( eiθ 1 + eiθ 2 ). It can be checked easily that the uniform distri-bution f (r , θ , t ) = ρ/ 2π is a solution of Eq. (164) for any density, and forarbitrary noise distributions p0(η) and p(η).

40



3.3.3 Hydrodynamic equations

In order to deal with more convenient physical quantities, we introduce thehydrodynamic density and velocity elds ρ(r , t ) and u (r , t )

ρ(r , t ) = π

− πdθ f (r , θ , t ) (167)

u (r , t ) =1

ρ(r , t ) π

− πdθ f (r , θ , t ) e (θ). (168)

Integrating the Boltzmann equation (164) over θ yields the continuity equa-tion for ρ(r , t )

∂ρ∂t + · (ρu ) = 0 . (169)

The operator is the vector differential operator of components ( ∂/∂x ,∂/∂y ). The derivation of a hydrodynamic equation for the velocity eld isless straightforward, and involves an approximation scheme. The reader isreferred to Ref. [13] for more details on the derivation. Introducing for con-venience the momentum eld w (r , t ) = ρ(r , t )u (r , t ), we nd the followinghydrodynamic equations:

∂ w∂t

+ γ (w · )w = −12

(ρ −κw 2) (170)

+ ( µ

−ξw 2)w + ν 2w

−κ(

·w )w .

It is interesting to give a physical interpretation of the different terms ap-pearing in this hydrodynamic equation. The rst term in the r.h.s. of Eq. (170) may be thought of as a pressure gradient, introducing an effectivepressure p = 1

2 (ρ −κw 2). The second term describes the local relaxation of w , whereas the third term is simply the usual viscous term. Finally, the lastterm corresponds to a feedback on the ow from the compressibility effects.

The different coefficients appearing in Eq. (170) can be computed ex-plicitely as a function of the microscopic parameters of the model. They aregiven by

ν = 14

λ 1 −e− 2σ 20 + 4π

ρ 1415

+ 23

e− 2σ 2− 1

, (171)

γ =8ν π

1615

+ 2 e− 2σ 2

−e− σ 2 / 2 , (172)

κ =8ν π

415

+ 2 e− 2σ 2+ e− σ 2 / 2 , (173)

41



0 2 4 6 8 10ρ

0

0.2

0.4

0.6

0.8

1

σ

u > 0

u = 0

Figure 2: Phase diagram of the model in the noise-density plane. Atransition line separates the domains with zero hydrodynamic velocity,u = |u | = 0 , from the domain of collective motion, u > 0. Parameters:λ = 1 , d0 = 0 .5, v0 = 1 .

µ =4π

ρ e− σ 2 / 2 −23 −λ 1 −e− σ 2

0 / 2 , (174)

ξ =64ν π2 e− σ 2 / 2 −

25

13

+ e− 2σ 2. (175)

Note that ν , γ and κ are always positive; µ can change sign, and ξ > 0whenever µ > 0.

3.3.4 Phase diagram

We turn to the study of the spontaneous onset of collective motion in thepresent model. We thus look for possible instabilities of the spatially ho-mogeneous system, that is the appearance of a uniform nonzero eld w .Considering a time-dependent, but spatially homogeneous ow, we get

∂ w∂t

= ( µ −ξw 2)w . (176)

Obviously, w = 0 is a solution for arbitrary values of the coefficients. How-ever, this solution becomes unstable for µ > 0, when a nonzero solutionw 0 = µ/ξ e appears ( e is a unit vector pointing in a arbitrary direction).Turning to the expression (174) of µ, it turns out that µ = 0 corresponds toa threshold value ρt given by

ρt =πλ (1 −e− σ 2

0 / 2)4(e− σ 2 / 2 − 2

3 ). (177)

42



The transition line dened by ρt in the plane ( ρ/λ, σ ) is plotted on Fig. 2,for σ0 = σ and for a xed value σ0 = 1. If σ0 = σ, the instability occursat any density, provided the noise is low enough. On the contrary, at xedσ0, the instability disappears below a nite density ρ0

t = 3 πλ (1 −e− σ 20 / 2)/ 4.

Both transition lines saturate at a value σt = (2ln 32 )1/ 2 ≈0.90.

In conclusion, the transition to collective motion can be obtained fromthe hydrodynamic equations. We note however that further instabilitiesleading to more complicated patterns are also observed [10, 13], and analyt-ical approaches to these instabilities are still an open issue.

3.4 Dynamics of residential moves

Another example of the dynamics of complex systems is the Schelling modelwhich modelizes the dynamics of residential moves in a city [14, 15]. The cityis modeled as a checkerboard, divided into cells. Two types of agents (sayred and green) live in the city. They reside in the cell of the checkerboard,with at most one agent in each cell. Agents characterize their degree of sat-isfaction regarding their environment by a utility, which is a given function(the same for all agents) of the number of agents of the same type in theirneighborhood. The neighborhood can be dened in different ways. Onepossibility would be to consider the set of nearest neighbors. However, moststudies rather use the Morse neighborhood, that is the 3 ×3 (or sometimes5 ×5) square surrounding the current cell.

Before moving, an agent chooses at random an empty cell, and evaluatesthe utility unew associated to this new location. The agent compares thisquantity to the utility uold of his present location, by computing the utilitydifference ∆ u = unew −uold . The move is then accepted with probability1/ (1 + e− ∆ u/T ). Here, T is a parameter analogous to the temperature inphysical systems, that characterizes the inuence of other factors, like thepresence of facilities, shops, or friends, that are not explicitely taken intoaccount in the model, but could bias the decision of moving or not. At lowT , and for a large class of utility functions such that agents have a (possiblyslight) preference for being with agents of the same type, a segregation phe-nomenon is observed when simulating the model numerically: two types of

domains form, namely domains with a majority of red agents and domainswith a majority of green agents. Quite surprisingly, this segregation phe-nomenon seems quite robust, and is also observed in the case where agentshave a marked preference for mixed neighborhood.

The Schelling model in its standard form is very hard to solve analyti-cally, and solutions are not presently known. The reason for these difficulties

43



is mainly that the neighborhoods of two neighboring nodes on the grid over-lap, generating complicated correlations in the system. In order to nd ananalytical solution, a standard strategy is to dene a variant of the modelon a specic geometry that avoids these correlations. This strategy wasfor instance successful in the Ising model: assuming that all spins interacttogether, the phase transition could be obtained analytically in a simpleway.

A straightforward application of this idea to the Schelling model a prioriseems to lead to a deadlock. If an agent evaluates his utility by consideringthe whole city as his neighborhood, this utility will not change when movingwithin the city. A more interesting strategy is then to divide the city into alarge number of blocks, so that agents evaluate their utility within blocks,and move from blocks to blocks. In this way, correlations between blocksmay be suppressed.

3.4.1 Denition of a simplied model

In order to implement this strategy, we consider the following model, witha single type of agent to further simplify the derivation (the case of twodifferent types of agents could be dealt with in the same way). The segrega-tion phenomenon then corresponds to the formation of domains of differentdensities. The city is divided into a large number Q of blocks, each blockcontaining H cells (a cell may be thought of as representing a at). We

assume that each cell can contain at most one agent, so that the number n qof agents in a given block q (q = 1 , . . . , Q ) satises n q ≤H . A microscopicconguration C of the city corresponds to the knowledge of the state (emptyor occupied) of each cell. For each block q, we also introduce the density of agents ρq = nq/H . Each agent has the same utility function u(ρq), whichdescribes the degree of satisfaction concerning the density of the block it isliving in. The collective utility is dened as the total utility of all the agentsin the city: U (C ) = q nqu(ρq).

A dynamical rule allows the agents to move from one block to another.At each time step, one picks up at random an agent and a vacant cell, withintwo different blocks. The agent moves in that empty cell with probability:

W (C →C ) =1

1 + e− ∆ u/T , (178)

where C and C are the congurations before and after the move respectively,and ∆ u is the variation of utility associated to the proposed move. Theparameter T has the same interpretation as in the standard Schelling model.

44



It is interesting at this stage to emphasize the difference between thepresent model and standard physical approaches. It could seem at rstsight that the utility is simply the equivalent, up to a sign reversal, of theenergy in physics. In the present model however, an economics perspectiveis adopted, so that the agents are considered as purely selsh. They takedecision only according to their own utility change ∆ u, and do not considerthe potential impact of their decision on the other agents. In contrast, inphysical models, the probability for a particle to move depends on the energyvariation of the whole system, and the effect on the other particles is thustaken into account from the outset. This has important consequences, as weshall see below.

3.4.2 Equilibrium congurations of the model

We wish to nd the stationary probability distribution P (C ) of the micro-scopic congurations C . This is not an easy task in general. Yet, if wewere able to show that a detailed balance relation holds in this model, wewould straightforwardly get the solution. Let us assume that the cost ∆ ucan be written as ∆ u = F (C ) −F (C ), where F is a function on congura-tion space. From Eq. (178), we nd that the dynamics satises a detailedbalance relation:

W (C →C )P (C ) = W (C →C )P (C ), (179)

with a distribution P (C ) given byP (C ) =

1Z

e− F (C )/T . (180)

It can be shown that a function F satisfying this condition is given by

F (C ) = −q

n q

m =0u(m/H ). (181)

To characterize the “segregation” phenomenon, the full statistical informa-tion on the occupation number of each cell is not necessary. Instead, anaggregated description in terms of densities of the blocks turns out to bemore useful. Such a coarse-grained description is obtained by aggregating allcongurations with the same number of agents in each block. As there areH !/ (n!(H −n)!) way of ordering n agents in H cells, we obtain the followingcoarse-grained probability distribution:

P (nq) =1Z

exp −H T q

f (nq) , (182)

45



where we have introduced the function f :

f (n) =T H

lnn!(H −n)!

H ! −1H

n q

m =0u

mH

. (183)

The above expression suggests to consider the limit of large H in order toget a continuous formulation for f . Keeping constant the density of eachblock ρq = nq/H (ρq hence becoming a continuous variable), and expandingthe factorials using Stirling’s formula, one obtains for H → ∞

1H

lnnq!(H −nq)!

H ! →ρq ln ρq + (1 −ρq)ln(1 −ρq). (184)

Similarly, the last term in the expression of f converges to an integral:

1H

n q

m =0u

mH →

ρq

0u(ρ )dρ . (185)

In terms of density ρq, the stationary distributions Π(nq) turns into aprobability density Ψ( ρq) given by (with Q

q=1 ρq = Qρ0 held xed):

Ψ(ρq) = K exp −H T

Q

q=1f (ρq) (186)

where K is a normalization constant, and where the function f (ρ) is denedas

f (ρ) = T ρ ln ρ + T (1 −ρ)ln(1 −ρ) − ρ

0u(ρ )dρ . (187)

The function Φ( ρq) = q f (ρq) may be called a potential, or a largedeviation function. It is also the analogue of the free energy functions usedin physics. The congurations ρq that minimize the potential Φ( ρq)under the constraint of xed Q

q=1 ρq are the most probable to come up.In the limit H → ∞, these congurations are the only ones that appear

in the stationary state, as the probability of other congurations vanishesexponentially with H .

3.4.3 Condition for phase separation

Focusing on the large H case, the problem gets back to nding the set ρqwhich minimize the potential Φ( ρq) with the constraint q ρq xed. We

46



Figure 3: Phase separation: the system of density ρ0 splits into two phasesof densities ρ1 and ρ2 to lower its potential.

are interested to know whether the stationary state is statistically homo-geneous or inhomogeneous. Following standard physics textbooks methods[16], the homogeneous state at density ρ0 is unstable against a phase sepa-ration if there exists two densities ρ1 and ρ2 such that

γf (ρ1) + (1 −γ )f (ρ2) < f (ρ0). (188)

The parameter γ (0 < γ < 1) corresponds to the fraction of blocks thatwould have a density ρ1 in the diphasic state. This condition simply meansthat the value of the potential Φ is lower for the diphasic state than for thehomogeneous state, so that the diphasic state has a much larger probabilityto occur. Geometrically, the inequality (188) corresponds to requiring thatf (ρ) is a non-convex function of ρ. The values of ρ1 and ρ2 are obtained byminimizing γf (ρ1) + ( 1 −γ )f (ρ2) over all possible values of ρ1 and ρ2, withγ determined by the mass conservation γρ 1 + (1 −γ )ρ2 = ρ0 (see Fig. 3)

We now try to translate the convexity condition (188) into a conditionon the utility function u(ρ). Phase separation occurs if there is a range of density for which f (ρ) is concave, namely f (ρ) < 0. We thus compute thesecond derivative of f , yielding

f (ρ) =T

ρ(1 −ρ) −u (ρ). (189)

For a given utility function, this condition can be checked explicitely. Wenote that in the limit T →0, f (ρ) = −u (ρ), so that the homogeneousstate is stable if u(ρ) is a monotonously decreasing function of ρ.

47



3.4.4 Peaked utility function

The specic form of the utility function is an input of the model, and itcan be postulated on a phenomenological basis, or rely on a theory of theinteractions among agents. In order to analyze an explicit example of anon-linear utility function, we consider the asymmetrically peaked utilityfunction dened as:

u(ρ) = 2 ρ if ρ ≤12

u(ρ) = 2(1 −ρ) if ρ >12

.

The expression of f (ρ) can be easily deduced from u(ρ), and is illustrated onFig. 4 for different values of T . To study the stability of the homogeneousphase, we look at the sign of f (ρ). One has for ρ < 1/ 2

f (ρ) = T/ρ + T/ (1 −ρ) −2, (190)

and for ρ > 1/ 2:f (ρ) = T/ρ + T/ (1 −ρ) + 2 . (191)

It is easy to check that f (ρ) is minimum for ρ →12

− , the correspondingvalue being

limρ→ 12

− f (ρ) = 4 T −2 (192)

Thus, for T > 1/ 2, the fonction f (ρ) is convex on the whole interval 0 <ρ < 1 as f (ρ) > 0 on this interval, and the homogeneous phase is stable.On the contrary, for T < 1/ 2, there exists an interval ρ1 < ρ < ρ 2 wheref (ρ) is concave (f (ρ) < 0), so that in the stationary state, the system issplit into two phases of respective densities ρ1 and ρ2.

The surprising phenomenon here is that a phase separation occurs evenin the case ρ = 1 / 2, although all agents have a signicant preference fora half-lled neighborhood. This can be understood intuitively as follows.At a small, but non-zero temperature T , small uctuations of density in theblocks are possible. Let us assume that we start from the homogeneous stateof density ρ = 1 / 2, with some small uctuations of this density around themean value 1 / 2. If a block has a density smaller that 1 / 2, then this blockbecomes less attracting for the agents living in it. So some agents will startto move to the most attractive blocks which have exactly the density 1 / 2.In doing so, the initial block becomes less and less attractive, thus making

48



0 0.2 0.4 0.6 0.8 1ρ

-1

-0.5

0

f ( ρ )

T = 1T=0.2T = 0

Figure 4: Plot of the function f (ρ) for different values of temperature, T = 1,T = 0 .2 and T →0, illustrating that f (ρ) becomes non-convex for T < 0.5,leading to a phase separation.

more and more agent leave it. This avalanche process, which is related tothe selsh behavior of the agents, qualitatively explains the instability of the homogeneous state with density 1 / 2.

Acknowledgements

It is a pleasure to thank Guillaume Beslon and Sebastian Grauwin for in-teresting discussions and for their help with Netlogo simulations during thelectures.

References

[1] D. A. McQuarrie, “Statistical Mechanics”, (Harper and Row, 1976).

[2] T. L. Hill, “An Introduction to Statistical Thermodynamics”, (Dover,1986).

[3] D. Chandler, “Introduction to Modern Statistical Mechanics”, (Oxford,

1987).[4] H. Goldstein, “Classical Mechanics”, 2nd Edition (Addison-Wesley,

1980).

[5] N. G. Van Kampen, “Stochastic Processes in Physics and Chemistry”,(North Holland, 1992).

49



[6] B. Derrida, Physical Review Letters 45 , 79 (1980); Physical Review B24 , 2613 (1981).

[7] E. Bertin and F. Bardou, American Journal of Physics 76 , 630 (2008).

[8] C. Monthus and J.-P. Bouchaud, Journal of Physics A 29 , 3847 (1996).

[9] T. Vicsek et. al. , Physical Review Letters 75 , 1226 (1995).

[10] G. Gregoire and H. Chate, Physical Review Letters 92 , 025702, (2004).

[11] J. Toner, Y. Tu and S. Ramaswamy, Annals of Physics 318 , 170 (2005).

[12] J. Toner and Y. Tu, Physical Review Letters 75 , 4326 (1995).

[13] E. Bertin, M. Droz and G. Gregoire, Physical Review E 74 , 022101(2006).

[14] T.C. Schelling, Journal of Mathematical Sociology 1 , 143 (1971).

[15] S. Grauwin, E. Bertin, R. Lemoy, P. Jensen, Proceedings of the NationalAcademy of Science 106 , 20622 (2009).

[16] H. Callen, Thermodynamics and an introduction to thermostatistics (J.Wiley and sons, New York, 1985).

50

introduction to statiscal physics_lecture_notes2009

Documents