HAL Id: jpa-00247181
https://hal.archives-ouvertes.fr/jpa-00247181
Submitted on 1 Jan 1996
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Finitely Generated Multifractals Can Display Phase Transitions
Thierrey Huillet, Bernard Jeannet
To cite this version:
Thierrey Huillet, Bernard Jeannet. Finitely Generated Multifractals Can Display Phase Transitions.
Journal de Physique I, EDP Sciences, 1996, 6 (2), pp.245-256. �10.1051/jp1:1996146�. �jpa-00247181�
Finitely Generated Multilfactals Can Display Phase Transitions Thierry Huillet (*)
and BernardJeannet
Laboratoire
d'Ingénierie
des Matériaux et des Hautes Pressions(**)
Institut Galilée,Université Paris 13, Avenue Jean-Baptiste Clément, 93430 Villetaneuse, France
(Received
11September
1995, received in final form 10 October 1995 andaccepted
25 October1995)
PACS.64.60.Ak
Renormahzation-group, fractal,
andpercolation
studies ofphase
transitionsPACS.05.50.+q
Latticetheory
andstatistics; Ising problems
PACS.75.10.-b Generaltheory
and mortels ofmagnetic ordering
Abstract. A new Mass of multifractal
objects ("skewed" multifractals)
is introduced, themultiplicative
generator of which has a finite number of branches of different real-valueddepths.
Both
microscopic
and macroscopic scales are represented by suchobjects,
each of these corre-sponding
to aspecific thermodynamical regime.
In the "diluted"regime,
thepartition
function Zt isexactly
renormalizable which means in thesequel,
as is the case in triegeneral
multifractaltheory,
that t~~log
Zt as a non trivial limit as t tends ta infinity. In trie "condensed" one thepartition
function converges. Details about the transition between these two regimes aregiven.
Résumé. Une nouvelle classe de "multifractales" est
introduite,
pourlaquelle
legénéra-
teur
présente
un nombre fini de branches delongueur
variable à valeurs réelles. Les échellesmacroscopiques
etmicroscopiques
sontreprésentables
par de telsobjets,
chacune d'elles cor-respondant
àun
régime
thermodynamiquespécifique.
Dans la phase "diluée", la fonction departition
Zt est exactementrenormalisable,
en ce sens(classique)
que la limitequand
t - oo de t~~log
Zt est non triviale. Dans laphase
"condensée" la fonction de partition converge. Lesdétails
thermodynamiques
concernant cette transition dephase
sont fournis.1. Introduction
Inhomogeneous
naturalobjects
cangenerally
be considered as assemblies offragments
or"atoms",
thesizes, weights
andspatial organization
of which cover anextremely
wide range of orders ofmagnitude,
rangmg from themicroscopic
scale to themacroscopic
one. A num-ber of common
examples
of suchcomposite objects,
such asbeaches, forests, galaxies
etc...,have
already
beenwidely
discussed in the hterature about fractals. Whenonly microscopic arbitrarily
smallfragments
areconsidered, then,
acomplete description
of theassembly
is notaccessible and a statistical
description
is thekey
point of the multifractaltheory.
In most cases however both
microscopic
andmacroscopic fragments
coexist in theobject.
Therefore the first
question
that arises is: what kind of model would becapable
ofdescribing (*)
Author for correspondence(e-mail: huillet@limhp.univ-paris13.fr)
(**)
UPR 1311 CNRS©
LesÉditions
dePhysique
1996such a situation? The second
question
is: how does one handle or understand the transition between these t~vophases,
for instance in a way similar to that used to understand the initial process ofcondensation,
1-e- theclustering
of molecules intodroplets
when the temperature of a gas is lowered. One of thegoals
of this paper is to ~vork out thisanalogy
and see its limitations.According
to anhypothesis commonly
used in natural sciences the structuresinvestigated
here will be assumed to present a character ofself-similanty.
The purpose of this paper is 1) to construct a
simple
mathematical model able to descnbecomposite objects
as those describedpreviously, ii)
to use the statisticalproperties
of this model in order toreproduce
the main features of theseobjects, namely:
coexistence of both diluted and condensedphases
andphase
transition.In order to
exploit
theanalogy
with acondensating
fluid mentionedpreviously,
thelanguage adopted
here will benaturally
that ofthermodynamics.
Apurely geometrical
model has been chosen. It is based upon aspecial partition
of the unit-interval1 ~~~(0
< x <1)
into a set ofsub-intervals,
thelengths
of which are either "observable" or not. This set is derived from aspecial
kind of "skewed" non-randomfinitely-generated multiplicative
cascade. In thiscascade,
thepartitiomng
of a finite number of the initial sub-intervals isstopped,
whereas it iscontinued for the other sub-intervals. Under these conditions the
limiting
distribution of the mtervals presents at the sanie time bothmicroscopic
andmacroscopic
scales. Some critical value of an externalparameter (separation point
orline) separates
these twophases.
We then sho~v that two differentthermodynamical
limits exist for these twophases.
In the dilutedphase, fragmentation
iscomplete, entropy diverges
but remainsrenormalizable,
orself-similar,
as it is m a standard multifi.actal
theory;
thisphase
is dominatedby
chaos in this sense. In the condensedphase,
the limit entropyexists;
we shallconversely
say that order dominates.The statistical
physics
of such hierarchicalsystems
is thendeveloped.
Particular attention ispaid
to thedivergence
mode of thethermodynamical
functions near the criticalpoint
or line.From this view
point
thepresent
model exhibitsspecial
structures which are worth exammmg in some detail.Similar
problems
were firstinvestigated
iniii by using
a reductible transfer matrix. As it will be seen later ouf conclusions are also in agreement with other recentinvestigations [2j.
The loss ofanalyticity
of thethermody.namical
functions has also been introduced in multifractal modelsdescnbing
diffusion-limitedaggregates [3,4j.
It is shownby
these Authors that the anomalous behaviour ofmultiphcatively generated
measures iscompatible
with theself-similanty
of thesemeasures which is a direct consequence of the fact that the multifractals considered there
are not
finitely generated.
This leads the Authors to derme left-sided multifractals. Such a situation is known to be asignature
of aphase transition,
which requires somegeneralization
of classical fractals
(see
for instance the multinomial measures mlà-?] ), keeping analytical
thethermodynamical
functions. Simflar treatments and conclusions cari be found in the field ofdynamical systems
[8].It is worth
mentiomng
however that theseapproaches
should not be confused with the present one, where it is shown thatfinitely generated
multifractals may. fail to be renormalizable.2. The Model
Now we come to a brief
description
of thephenomena
we aredeahng
withhere,
I.e. a hier- archicalmultiphcatively generated system
of intervals on: 1=
(0
< z <1).
At time zero, this initialinterval, generates offsprings
infinite
number(say AI).
There are twocategories
of"sons" in our
model, namely
the"productive"
ones that will repeat their ancestors'"growth
program",
and the "sterfle" ones that do trotspht
anymore.~ l
t xi x2 x3 x~ X> ~2 x3 x> x2
~ 0
t2
Fig.
l.First-generation
tree andfirst-generation
of intervals on 1.Each of the
productive first-generation
mtervals possesses its own real-valued lifetime before itgenerates
asecond-generation
in the same ratio of sizes.Thus,
thelength
of the intervals ismultiplicatively propagated
in the cascade. Asmultiplicative phenomena
are qmte diflicult tohandle,
one can m anequivalent
manner work with an additive structure, ratherconsidering
the
logarithm
of thelength
process.At a
given generation
the lifetimes are trot all the same. Trie different lifetimes of the firstgeneration
intervals are assumed to take their values into a finite set of I( real distinct values:(tk)Î=1
It is convenient torepresent
this firstgeneration
of intervalsby
agenerator tree,
as mFigure 1,
for which each branchsymbolizes
an mterval. Thelength
of each branch indicates the time at which the interval is cut into sub-intervals,The different
lengths
of the firstgeneration
intervals are assumed to take their values into a finite set of L real distinct values:(xi )Î=1
The "sterile"
fragments
remain forever.They
have infinite lifetimes. Whatonly
matters is tokeep
track of their distinctlength, taking
their differentpossible
values in a finite set ofÎ
real distinct values:
(11))~~.
Let
(ai,k)Î(~i
denote the number offirst-generation
sons oflength
xi among those that willsplit
at té,ànd (âi ~)~~~
L denote the number of sterile
first-generation
sons oflength fi.
L
There are thus: a_,k ~~~
~
ai k
Productive
mdividuals whose lifetime istk,k
= 1:.I( and
i=1 K
ai, ~~~
~ai,k
whoselength
isxi,1
=
1, ,L.
We shallspeak
ofdegeneracy
to describe thisk=1
K Î
fact. Let also A ~~~
~
a,k be the total number ofproductive
sons, and D ~~~~âi,co
thek=1 ~=i
total number of sterile sons. ive
get
of course, A + D= ilf.
"Length
conservation" is assumed,L £
which means:
~
ai xi +~
ài coi= 1.
Figure
1 is anexample
of such a generator forwhich,
i=i i=i
L =
3,
K=
2,
ai,1 " a2,1 " a3,1 #1, îli,co
# î12,co#
1,
ai,." a2, # a3,. #
1,
a1 "
2,
a 2 # 1,A = 3 and D
= 2.
No
special ordering
is chosen in thearrangement
oflengths
of intervals. It is howeveruseful,
fornotftional
convenienceonly,
to introduce someparticular
values of thesenumbers, namely:
~°~~
~~~max
~°~~~
Let
(1*, k~)
~~~Arg
max~°~~~
ive also define
t~ i,k=1 .L,K tk i k=i...L,K tk
0
~~ ~
~~~ +' % ~' ~~
t
~ ~ ~~ ~~
t ti
t
iii
12'((((( l Il Il 1Il1 t2
,
2
x2x3 x2
ii
x212,~2~( ~~ ' ' ,
Fig.
2. Generation andfragmentation
of trie unit~interval at any time t.an associated
value, ~°)/~
~~~°~~~'~
k i~,k~
In a similar
fashion,
if:logl*
~~~ max~(- log t ),
and 1* ~~~Arg
max~(- log A ),
we definei=i...L i=i...L
log
à~ the associatedlog-number
value of such sterile branches:logà*
~~~logài,co
i*
At any time
t, (see Fig. 2),
we thus have constructed asystem
of intervals embedded in the unit interval. We shall callX,
the limit set ofpoints
obtainedby considering
theendpoints
ofthese intervals when t - oo. Observe
that,
due to our constructionincluding
stenleindividuals,
there is an infinite number of
adjacent
sitesseparated by
a measurabledistance,
and of coursea
large
amount of"infinitely"
close pairs.We have
adopted
here the"population theory" language
because thepresentation
of our model appears to be easier this way. Wemight
have descnbed our model in ternis of "contactprocesses"
m a very similar fashion.So,
letX~
~~~ UB~(x)
stand for the"coarse-gramed"
~ex set obtained from X
by centering
balls,
of radius e on each
point
x of thisset,
andconsidenng
the reunion of these halls. When trie resolution is very
low,
say when e exceeds some value 0 < go <1, X~
is limited to asingle
connected comportent.By increasing resolution,
thatis to say
by positioning
e below go, this component issplit
into Msub-components
within which eachpoint
is e~indiscernible. Part of these Msub-components (A
ofthem)
will thensplit
intosub-sub-components
when e crossesrespectively
the value(eh )k=1..K, (not including degeneracy).
Part of these(D
ofthem)
will notsplit,
which means ci#
0,
=1...1,
inthis case.
Thus,
eachsub-component splits
at different levels ofresolution,
if itsplits
at all.Increasing
the resolution ad infinitum(that
is to say,taking
the limit e -0+) subcomponents
shrink therefore into the continuum of points X.Thus, considering
thechange
of variables: e=
eoe~~, according
to which the sequence(ek)k=i:.K
con be obtained from(tk)k=i...K by:
eh#
eoe~~L,
these two models appear very much alike.Thus,
more and more details are discovered from X as resolution goes tomfinity (e
-0+
or t-
oo). Reversing
the arrow oftime,
detailsamalgamate
until weget
a finalblur,
just
asepidemics eventually
contaminates a givenpopulation.
3. Evidence of a Phase Transition in the
Thermodynamic
Limit3.1. STATISTICAL PHYSICS OF THE CASCADE. From the statistical
physics point
ofview,
all information on inter-distance distribution is enclosed in thepartition
function:Zt(fl)
~~~~
N,xt(1)~, fl
Elit,
t > 0 whereNt
stands for the total number of individuals available at timei=i
t,
both sterile andproductive
ones, and xi (1) the distanceseparating
the two consecutive sites1- 1 and 1, =
1,
...,
Ni,
at time t. We define the localpartition
functions of the sterile andL L
productive offsprings, a~(fl)
~~~~îli,colt, ak(fl)
~~~~
aikxf, respectively.
One can theni=i i=i
easily
check thatZt(fl)
meets the renewalequation:
K
Z~(fl)
=aco(fl)
+~ iak(fl)lt~>t
+ak(fl)Zi-i~(fl)li~çii,
t2 °, Zo(fl)
=
i, (1)
which
basically
describes the "self-similar" character of our model(see Fig. 2).
In thisequation,
"1"
represents
the indicator which is 1 if itsargument
istrue,
and 0 otherwise.K
We observe that
Zi(0)
=
Ni
is accurate with:Ni
" D +
~[a,k.11~>1
+ a_k.Ni-i~.li~<iÎ,
k=i
t >
0, giving
the number of intervalsNi,
in thecascade,
at time t.Hence, Zi(1)
=
1,
whichmeans that it is a
fragmentation
of the interval[0, ii,
at any time.Letting
now, :Î(fl, s)
~~~Î
~e~~~Zi(fl)dt,
onegets, taking
theLaplace
transform ofil),
and after
elementary computations:
K
acojfl)
+~
ah(fl) Il
e~~~L~~~'~~
~Î
' ~~~
s
~ ak(fl)e~~~k
k=1
provided
s >max(0, log~(fl))
~~~logz(fl), considering
the twosingularities
ofÎ(fl, s).
Those
singularities
are s= 0 and the
largest
solution s=
log~(fl)
to the"generating
L,K L,K
equation": ~j
a~
kxfe~"~
=~j
ai
kxf~(fl)~~~
= 1.Moreover, i(fl)
isanalytic
on IR. Iti,k=i i k=i
should be observed
that,
on thecontrary,
functionlog z(fl),
asdefined,
is not itselfanalytic.
Indeed, -logz(fl)
= 0 for
fl
>flc,
with 0 <flc
<1,
the unique cntical value solution toL
i(fl)
= 1; hence:
~j
ai
xf~
= 1.Moreover, -logz(fl)
isfinite, non-decreasing
and concavei=i
(see Fig. 3).
It should be noted that:log z(0)
=~°)~,
where T is solution of theequation:
~j a,kAÎ
K= 1.
(Note
that if allproductive
branches have the samelength
7, solution of thek=i
last
equation
is T=
7).
-é~yzl>)
à
r
fl~£
Fig.
3. Functionlogz(fl):
the "renormalizedpartition
function".The existence of a
phase
transition in Durmortel, crossing
the criticalpoint,
has thus beenproved.
Note that ifao~(p)
=
0, 1(p, s)
hasonly
one
singularity.
In this case there is nophase
transition.
It therefore follows from the
study
ofsingularities
that the"thermodynamic
limit" halas:lim
log Zt(fl)
"log z(p), (3)
~vhere function
log z(fl)
is the "renormalizedpartition
function".At
p
=
0,
formula(3)
reads:hm
Zt(0)1 Î~
lim(Nt)Î
=
Al Ù z(0)
> 1(4)
Thus,
the total number of intervals of the cascade grows like:Ai.
Let now
Nt(e)
~ ~~ i=
1,.
,
Nt/ logzt(1)
> e denote tl~enumber,
among theseNt
in-tervals at time t;
sharing
the property within the bracket: ztIi)
<exp(-et).
It follows from thelarge
deviation theorem ongeneral
random processes [9jthat, provided
e >~~°~~~~~
~Ù
~~~lim
Nt(e)~/~
= exp
s(e) (5)
t-m
Formula
(5),
thus grues theasymptotic shape
of the distribution of intervallength.
In casee <
~
~°~~~~~
,
change
> into < in the formuladefining Nt(e).
dp
~~oIn tl~e last
formula,
sje)
~~~ infje p+log z(fl)
is the concaveLegendre
transform oflog z(fl),
peur
i e. "renormalized
entropy".
We shall calle(fl)
~~~-~~°~~~~~,
the "renormalized internaidfl energy". Thus,
formula(5)
is validprovided
e >e(0).
3.2. THE ItENORMALIzED ENTROPY SHAPE. What follows is a brief
description
of theshape
ofs(e),
whichessentially
diifers front the left-sided ones of[3,4j (see Fig. 4). Taking
5(e)
@
---~-T '
l '
.
'
~°8~
t* ~~~~~<~~~~~~~~~~~~
°
e(flc)
e(oje(-oeJ
eFig.
4. The renormalized entropyshape.
the
Legendre
transform oflogz(fl),
as definedby (3),
onegets: s(e)
isdefined, continuous, nonnegative
and concave when 0 < e <e(-ce)
< ce. On the "renormalized internaienergy"
interval: 0 < e <
e(-ce), entropy
isstrictly
concave, with a linearpart s(e)
=
fl~e
for aile E
[0,e(fl~)]. Here, e(-ce)
=
~°~~~, s(e(-ce))
=
~°~~~
(definitions
of these numbers aret* t*
given in Section
2)
ands'(e(-ce))
= ce. s attains its
maximum, logA,
ate(0).
Remarks
1)
One can show thatpc
has a naturalinterpretation,
in terms of boxcounting dimension, namely:
dim X ~~~limsup ~°~~j~~,
wl~ereN(e)
is the minimal number of intervals~_o+
log
j necessary to caver X.
2) Up
to now, we haveonly
beendealing
with the statisticalproperties
of setX, through
the
Laplace parameter fl.
At thisstep, indeed, fl
hasonly
been used as a distortionparameter, allowing
tostudy
thefamily
of distributions:(zt(1)P,1
=
1,...,Nt).
Forexample, negative
values offl
underhne small intervals in thisdistribution,
whereas atfl
=0,
it is netpossible
to discriminate the small ones from thelarge
ones at ail. The statisticalphysics language
can beadopted, provided
one gives a richermeaning
to this centralparameter fl.
TO do so, observe that Dur formulation amounts to consider tworepelling partiales
whosepotential
energy is E=
logz,
that is to say, thelogarithm
of the distance between them[2]). Suppose
that theonly
reachable energylevels,
at time t, areprecisely: Et Ii)
~~~log~t Ii),1
=
1,..., Nt.
Thispair
ofpartiales,
at"temperature"
T
=
(related
to its own vibrationalenergy),
may then be used to observe the diiferentfl
energy levels.
Indeed,
one of the axioms of statisticalphysics
tells us that theprobability
of
configuration
is:Zt(fl)~~ e~P~'fi)
=
Zt (fl)~~
zt(1)P.
The statisticalphysics
of this paircan thus be derived from the
study
of thepartition
functionZt(fl),
where the central parameterfl
is nowinterpreted
as the inverse of thetemperature.
This
analogy
allows us toskip
from the statisticallanguage
to the one of statisticalphysics. Any
allusion to statisticalphysics
in thesequel
should be understood in thisphysical
context,only.
3)
It should beemphasized
that the presence of aphase
transition is a direct consequenceof our structure
presenting macroscopic scales,
due to the presence of sterile individuals.Thus,
fl~separates
twophases:
the "cola" oneIi.e.
whenfl
>pc)
for which macroscopic orderprevails,
and the "hot" one for which disorder is the rule and which is identified with the exact "renormalization" of thepartition
function.3.3. BEHAVIOUR OF THE THERMODYNAMIC FUNCTIONS NEAR THE CRITICAL POINT
Going deeper
into theanalogy
withthermodynamics,
mentionedabove,
it ispossible
to intrc- duce thefollowing quantities: fl
can thus beinterpreted
as the inverse of some"temperature".
Occurrence of a
phase
transition appears in the limit: t ~ oe.At
given (finite)
t, the statisticalproperties
of this mortel are enclosed into the "freeenergy"
function=
Ft (4)
~~~j iog zt (4)
Entropy St(E),
theLegendre
transform oflog Zt(fl),
in the variablefl,
satisfies:~-pE,(i) ~pE,(i)
St(E)[~
~~~ d<o~z jp) =
flEt(fl)
+logZt(fl)
"
~ log
,
where
Et(fl)
is the' d~
~~~
~t(~) Zt(~)
average of the
"energy
levels" :log
zt(1),
as a function offl.
Looking
at t as the resolution power, its dualvariable,
in the sense ofLegendre, P,
will represent"thermodynamical pressure" [loi.
This allows todefipe
thethermodynamical
"freeenthalpy" G(fl, P),
theLegendre
transform ofwhich,
in the variableP,
isnothing
butFt(fl).
Pressure as a function of
fl
and t is:Pt(fl)
~~~~~~~~
ôt As t ~ oe, hm
St (E)[E,(p)
~~~s(e)[
~~~ diogzjp) =
fie(fl) +log z(fl),
where the "renormal-t-m t e "- do
ized internai
energy" e(fl)
~~~ lim~Et(fl)
"
-~~°~~~~~
is the hmit
(t
~oe)
average of thet-m t
à~
"normalized energy levels":
logzt(1).
One can show that pressure:P(fl)
~~~ limPt(fl)
t t-m =
~°~~~~~
(> 0). Moreover,
the"specific
heat per unit"~~~~
=-p~~~~ °(~~~~~
can be
fl d(p- dp
defined.
We shall now
briefly
discuss the behaviour of these functions asp
~pl.
From the definition oflog z(p)
andP(fl),
it follows that"pressure"
vanishes forfl
>pc,
isregular
anddecreasing
as
fl
<pc.P(flc)
= 0. Thus
"pressure"
is anon-increasing,
continuous function offl.
On the
contrary,
"internaienergy" e(fl)
is anon-increasing
function offl, vanishing
whenfl
>pc
but with ajump
atpc,
whoseamplitude
caneasily
becomputed.
Occurrence of ajump
at
p
=
pc
is sharedby
the"specific
heat per unit".3.4. INCLUDING THE GENERATION NUMBER. Additional information on the hierarchical
structure under
study
is enclosed in thebi-partition (grand canonical)
function:Et là, p)
~~~P~(t)
N<(P)~j e~~P ~j zt(1)°,
À,p
EIR,
t >0,
whereNt(p)
stands for the number of individuals at tp=i i=1
In the Cascaàe WÎtÎI
eXactly
p ancestors(file generatlon nUmber).
p+(t)
~~~Integer part
indicates the maximal number of such ancestors at t.min
(tk)
=1...K
P+(n)
Indeed,
bathmarginais: Bt(0, fl)
=Zt(fl),
andBt(À, 0)
=~j e~~PNt(p)
are of
interest,
in p=ithis case. This is one of the main interest of the "skewed" character of our multifractal
family.
One can
easily
check thatBt(À, fl)
meets the extended renewalequation:
~~~~ ~~ ~~~°~i~~
+e> ia~i~)it~>~
+
a~i~~~~-~~ j~, ~~i~~<~j,
> ~Bo(À,fl)
= 1
(6)
This follows from the
straightforward
recurrence formuladefining Nt(p)1
K
Nt(p)
= D
ip~i
+~ la
~
ip=i it~>t
+ a,~Nt-t~ (p i) it~~tj
,
t
2 o,
p=
i,.
,
p+(t).
k=1
Letting
now,:É(À, fl, s)
~~~Î~ e~~~Bt(À, fl)dt,
onegets, taking
theLaplace
transform of(6):
e~~ lK
are
(~)
+~
ak
(~)(Î
e~~~~j
É(À>4> s)
=
~)l
(71S Î e-~
~
ak(~)e~~~~
k=1
provided
s >max(0, log~y(À, fl))
~Î~logf(À, fl).
Here s= ~y(À,
fl)
is(for
each(À, fl))
thelargest
K L,K
solution to the
"generating equation": e~~ ~jak(fl)e~~~L
=e~~ ~j ai,kzfe~~~L
= thatis,
k=i 1k=i
L,K
satisfying: ~j
ai
k~f~y(À, fl)~~L
=e~.
l,k=i
Function
log f(À, fl)
thus faits to beanalytic,
which is thesignature
of aphase
transition.This function is
finite, non-decreasing
and concave. Note thatf(0,fl)
=
z(fl).
It satisfieslogf(0,0)
=
log Al,
withAl ~ÎÎ f(0, 0)
>1,
and T defined as in Section 3.1.Moreover,
logf(À, fl)
=
0,
when >T(fl),
withT(fl)
thestrictly
convex critical hne ofequation:
L,K
T(fl)
=log ~j
ai
kzf, (8)
k=1
for which:
T(0)
=
logA, T(flc)
= 0. One has thusproved
the existence of aphase transition, crossing
the critical fine thistime,
theequation
of which is=
T(fl),
in the(fl,
À)plane.
It aise follows that the
following thermodynamic
limit halas: hm-~
logBt(À,fl)
=
t-m t
logf(À, fl),
and that the behaviour of ailthermodynamic
variables can of course be derived from this moregeneral
situation.The behaviour of the
thermodynamic functions,
asfl
~pi
caneasily
be obtained in asimilar way.
&àÇ*
ôc
/
r' à
-logâ*1'
'Fig.
5. Functionlog Z(fl).
4.
Thernlodynamics
of trie Condensed Phase4.1. CONVERGENCE OF THE PARTITION FUNCTION. Iii case
fl
>pc,
it should now be clearthat the limit of the
partition
functionZ(p)
~~~ hmZt(fl)
itself exists and is:t-m
~ al,mÎÎ
Z(fl)
" lilll S
Z(S, fl)
#~)
~~~"
~~~
~
(9)
s-o+
~j ak(flj ~j
ai,xl
k=i 1=1
from the initial value theorem of
Laplace. Figure
5gives
theshape
of the functionlogZ(fl).
Thus,
in the "cola"phase (fl
>fl~),
orderprevails
and thepartition
function has itselfa
thermodynamic
limit. Thisphase
thus deserves its own statisticalphysics.
Forexample, defining
theLegendre
transform:S(E)
= inf
(Efl
+log Z(fl)),
theentropy,
it should be clear P>P~that the whole
"machinery"
of statisticalphysics
can be defined. Inparticular,
the "internaienergy" E(fl) ~ÎÎ -@
the "freeenergy" F(fl)
~~~-) logZ(fl)
and the"specific
heat"C(fl) ~ÎÎ -fl~~
can be of interest, in the condensed
phase.
4.2. THE SHAPE OF THE ENTROPY. We first locus on the
shape
ofentropy,
as a function of the internai energy(see Fig. 6).
For obvions reasons, this function is concave,non-negative,
defined for E >-logl*,
where Î" is thelargest
value among the(fli)i=i
z.Moreover,
S(-logl~)
=
logé'
~vith à* the number of individuals attached to thisvllul. S(E(1))
=
E(1)
> 0 whereE(1)
is the bounded derivative atfl
= 1 oflogZ(fl).
Theslope
ofS(E)
is infinite at E=
logfl*, (fl
~+oe),
one at E =E(1).
The behaviour of
S(E)
near E= ce is:
S(E)
mfl~E+log
E+ cste,traducing
the fact thatE(fl) diverges
aifl
E-m=
fl~. Thus,
thisentropy hnearly diverges,
as a function of E. TÎ~is resultis in accordance with the linear
part
of the renormalizedentropy s(e),
defined in Section 3.2.Divergence
of E is due to thedivergence
of the average value of thelogzt Ii),
wl~enfl
> fl~, In tl~is caseEt(fl)
"
~~°~~~~~
"
-t~~°~~~~~
=
te(fl) provided e(fl)
<e(fl~). TÎ~us,
Eàfl
t-mà~
SIC
,
, ,
Iogi* '_
,
E@J £
Fig.
6. The entropyshape.
diverges
likete,
where t ce ,vith e <e(p~)
ands(e)
= lim ~~~~~
=
p~e,
underhypothesis
t-m t
~ <
~(flc).
4.3. ASYMPTOTIC BEHAVIOUR NEAR THE CRITICAL REGION. We
briefly
stress here howit can be obtained as
fl
~fl/,
Observe from(9)
that the denominator ofZ(fl)
is aise:L L
~j ai.(z/~ zf), referring
to theequation defining
the critical value. Thus:~j ai.(zf~
i=i i=1
zf)
Lm
-(fl pc) ~j
ai,z/~ log
xi, so that:Z(fl)
mI(.(fl fl~)~~,
with K as apositive
4=4t~~~ 4=4t
known constant. In a similar ~vay, one con show that the "internai
energy" behavej
like:E(fl)
mK'+ (fl fl~)~~,
that the"specific
heat" behaves hke:c(fl)
m 1 ~~,
and
o=ot o=Pf 4
that the "free
energy"
grows like:F(fl)
m(K"(fl fl~)),
whereK',
K" arepositive P=Pt flc
known constants.
Finally,
the entropyS(fl) ~# S(E)j
=flE(fl)
+logZ(fl),
as a functionE(P)
of temperature grows like:
S(fl)
mKi(fl flc)~~ K2 log(fl pc ),
whereKi
andK2
areP=àf positive
known constants.5.
Concluding
RemarksIn this
article,
weemphasize
a basicproperty
of nonrandommultifractals,
under whichphase
transition occurs at finite
temperature, namely
the coexistence of micro andmacroscopic
fea-tures. It is indeed the presence of
macroscopic
characteristics within them which induces lossof
analyticity
of thethermodynamic
functions. We illustrate these ideas on a continuous-time(resolution) family
of skewedmultifractals,
which in itself deserves interest. Part of the "ther-modynamics"
of such structures is studied in somedetail, including
behaviour near trie criticalpoint (or hne).
References
iii
Huillet T. and JeannetB.,
J.Phys.
A 27(1994)
l191.[2j Mendès-France M. and Tenenbaum
G., preprint
to appear in Math.Phys. (1994).
[3] Mandelbrot
B-B-, Physica
A 168(1990)
95.[4] Mandelbrot
B-B-,
Evertsz C.J.G. andHayakawa Y., Phys.
Reu. A 8(1990)
4528.[5]
Halsey T.C.,
JensenM.H-,
KadanoifL.P.,
Procaccia I. and ShraimanB-I-, Phys.
Reu. A 33(1986)
l141.[6] Huillet T. and
Klopotowski A.,
J.App. Probabiiity
31(1994)
333.[7] Mandelbrot B-B- and Evertsz
C.J-G.,
in Chaos andfractals,
H-O-Peitgen,
H.Jurgens
and D.Saupe
Eds.(Berlin Springer, 1992)
921.[8]
Wang X.J., Phys.
Reu. A 40(1989)
6647.[9] Ellis