HAL Id: jpa-00247059
https://hal.archives-ouvertes.fr/jpa-00247059
Submitted on 1 Jan 1995
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.
Constrained Annealing for Spin Glasses
Giovanni Paladin, Michele Pasquini, Maurizio Serva
To cite this version:
Giovanni Paladin, Michele Pasquini, Maurizio Serva. Constrained Annealing for Spin Glasses. Journal
de Physique I, EDP Sciences, 1995, 5 (3), pp.337-354. �10.1051/jp1:1995130�. �jpa-00247059�
Classification Physics Abstracts
05.50 02.50 75.10N
Constrained AnneaEng for Spin Glasses
Giovanni
Paladin(~),
MichelePasquini(1)
and MaurizioServa(2)
(~) Dipartimento di Fisica, Università dell'Aquila, 1-67100 Coppito, L'Aquila, Italy (~) Dipartimento di Matematica, Università dell'Aquùa, 1-67100 Coppito, L'Aquila, Italy
(Received
3 October 1994, accepted 29 November 1994)Abstract. The quenched free energy of spin glasses con be estimated by means of annealed averages where the frustration or other self-averaging variables of disorder are constrained to their average value. We discuss the case of d-dimensional Ising mortels with random nearest
neighbour coupling, and for +J spin glasses we introduce a new method to obtain constrained annealed averages without recurring to Lagrange multipliers- It requires to perform quenched
averages either
on small volumes in
an analytic way, or on finite size strips via products of random transfer matrices. We thus give a sequence of converging lower bounds for the quenched
free energy of 2d spm glasses-
l. Introduction
Disordered systems are made of random variables which relax to
equilibrium
on very diiferent time scales. For instance in the standard models ofspin glass,
the spins arrange themselvesto minimize the free energy while the random
couplings
are assumed to be frozen in their initialvalues,
since their evolution time is muchlonger.
As a consequence, one should consider diiferent kinds of averages:quenched
averages for the frozenvariables,
and annealed averages for the fast variables.However,
it,is much easier to compute annealed averages, which#ve
lower bounds of the
quenched
free energy, thephysically
relevantquantity.
These boundsare often very poor
estimates,
since in annealed averages the main contribution comes from a set of disorder realizations which bave zeroprobability
in triethermodynarniclimit.
In thoserealizations,
triequenched
random variables arearranged
in order to minimize trie free energy, at diiference with atypical sample
of trie system wherethey
are frozen in agiven
realization.Trie idea of constrained
annealing
is toperform
annealed averages where trie random vari- ables arepartly
frozenby
trie requirement ofsatisfying appropriate
constraints, related to somedisorder
self-avera#ng
variables. Trie standard way toimpose
a constraint isby
aLagrange multiplier. By
thatmethod,
trie constrained annealed average allows one to introduce a sort of Gibbsthermodynamic potential, depending
on trie temperature and on trieLagrange
mul-tipliers
whichplay
the role of trie chernicalpotential
inordinary
statistical mechanics [1]. In thethermodynamic
limit, the values of themultipliers
that maximize the Gibbspotential
arethose ones that select the realizations with a correct value of the disorder intensive
variables,
@ Les Editions de Physique 1995
and,
at trie sametime,
that minimize the diiference between trie mean free energydensity
and its annealedapproximations.
Constrained
annealing
for disordered systems bas beenproposed
andapplied
inparticular
contexts
by
many authors(see [2-6]).
Ageneral
formulation of trie method and itsphysical
meaning
can be found in [1].Trie purpose of this paper is to show that it is
possible
toperform
constrained annealed averages withoutLagrange multipliers,
m d-dimensionalspin glasses
with dichotomic randomcouplings
in terms of aquenched
average over a finite sizesub-system.
This is oftenpossible
either m an
analytic
or m a numeric way.Using
ourmethod,
we are able to obtain verygood
lower bounds of trie
quenched
free energy of two dimensional spinglasses.
Trie summary of trie paper is trie
following:
In section 2, we recall some basic facts about spm
glasses.
We alsoclassify
the set of trie constraints necessary to recover triequenched
free energyby
annealed averages. In section 3,we
study
trieIsing
model withcouplings
which con assume trie values J~j = +1 withequal probability (the
sc-called +J spinglass).
In section 4, we introduce our method to obtainconstrained annealed averages without
recurring
toLagrange multipliers
in trie +J spinglass.
In section 5, we
apply
trie method toloops
of infinitelength using
trieLyapunov
exponent ofproduct
of random transfer matrices. We are thus able to get accurate estimates of trie freeenergy in two dimensions.
2. Constraints for d-Dimensional
Spin
GlassesSpin glasses
are triechallenge
of constrainedannealing.
In that case, trie relevant variables to be constrained should be related to frustration [7], as first remarkedby
Toulouse and Vannimenus in 1976 [5]. As far as weknow,
there areonly
Montecarlosimulations,
see e.g.[8,9],
which usedthat idea while no
analytic
calculations bave beenperformed
on thatground.
However, it ispossible
to getanalytic
results m dimension d > 2 with constraints connected to frustration.Let us consider d-dimensional
Ising
models with nearestneighbour couplings
J~j which areindependent identically
distributed(1.i.d.)
randomvariables,
in absence of externatmagnetic
field. Trie
partition
function iszN
=
~ fl expjpJ~jaiajj (2.1)
lai i J
where trie
product
runsover1, j
which are nearestneighbours. Using
standard relations of triehyperbohc functions, (2.1)
becomesZN "
~ fl cosh(flJ~j) il
+ a~ajtanin(flJ~j) (2.2)
lai '>J
This form is useful since shows that trie non-trivial part of trie partition function is
given by
fl,
~
il
+ a~ajtanin(flJ~j
)]. Atypical
term m thatproduct
is a~ajaj aitanin(flJ,j tanin(flJji)
where trie bonds
(1, j ), ( j,1)
can form either a closed or an open line on trie lattice. If trie bonds form an openline,
at least two distinct spin variables survive as factor of trieproduct
of the varioustanh(flJ,j ),
and the related termgives
a zero contribution to the partition function,after
jerformmg
the sum over the aconfigurations.
As a consequence, one bas to consideronly
trie terms
corresponding
to bonds which form a closed fine(where
trieproduct
of trie a's isequal
tounity
smcethey
appeartwice)
so that(2.2)
con be written as a sum over all triepossible loops
LÎ~~, where T > 4 indicates trielength
of trieloop
(T bas to beeven)
and s labelsdiiferent
loops
with the samelength:
ZN
"eÉ.,>
~" ~°~~~~"~2'~
l
+
~j fl tanh(flJ,j) (2.3)
lL[~>j j~,jjeL[~>
It follows that in a system with a finite number N of
spins,
the free energy of a disorder realization isgiven by
y~ = in z~ = B +
fl~ (t(4)
~(Rj ~~~)
flN fl
' 'where R is the maximum
possible length
of aloop
in the system(R
< constN),
andB = In 2 +
~j
Incosh(flJ~j (2.4a)
~
>,J
and
~~~~~~~' '~~~~~
Î
~~~
~
~Î~~
(2.4b)
jL[~>j This function
depends
on the new set of variablest(~) =
fl tanh(flJ~j) (2.4c)
j~ ~j~~(r>
given by products
which run on theloops
LÎ~~, 1-e-plaquettes
oflarger
andlarger perimeter
at
increasing
T. For instance, theloop
L(~) is theelementary plaquette
of four bonds. Let usstress that a
loop
L(~) can be built up as a set of two(or more)
distinct and separateloops
L(~'l
and L(~2) withTi + T2 " T. In the
following
we often label acoupling by only
one labelmstead of two ones, 1.e. Jki is denoted
by Jk,
since there is no risk ofambiguity
on a closed line.As the number of the nearest
neighbour couplmgs
in ahypercubic
d-dimensional lattice is dN, in thethermodynamic
limit thequenched
free energy readsf
e lim @ " d Incosh(flJ~j
+ F + In 2(2.5)
N-CC
fl
where
F = lim
j lui
1+
~j
tÎ~~(2.6)
~~~°
j~(r>j
The reduction of the
partition
function(2.1)
into(2.3)
is theprocedure
used in the Kac and Ward solution of the 2dIsing
model without disorder.Equation (2.5)
is fundamental since it shows that thequenched
free energydepends only
on the distributions of the variables
t(~),
apart the mean value Incosh(flJ,j).
The relevant variables to be fixed to their ~r~ean value m a constrained annealed average of the free energyare
related,
withincreasing difliculty 1)
In coshflJ~j
2) t(~l
=
tanh(flJi) tanh(flJ2) tanh(flJ3) tanh(flJ4)
3)
t~~l with T > 6(considering only
connectedloops).
Fixint
all the constraintscorresponding
to thosequantities
in an annealed average isequiv-
alent to solve the system withquenched
disorder. This can be understoodnoting
that in(2.6)
we have a sum over ail the
possible
tÎ~~. Because of the law of thelarge
numbers the rescaledsum of the t-variables
corresponding
to the sametopological
kind ofloops self-averages
to itsquenched
mean value.Moreover,
in theappendix
we shall prove that this result holds even in a finite volume system for a +Jspin glass:
thequenched
free energy can be obtainedby
a minimization of the
logarithm
of the annealed average of aGrand-partition
function with respect to theLagrange multipliers
related to the variables tÎ~~ over ail thepossible loops
of the finite system. For ageneric
distribution of thecouplings,
an extension of this statement ispossible.
In one dimension there are no
loops,
so that the constraint related topoint 1)
is suflicient to obtain the exact solution. Ingeneral,
we expect that manyphysical
features of aspin glass
can be
reproduced already by considering
the first twopoints.
The case of random
couplings
J~j = +1 withequal probability
isparticularly simple
in this respect, since the variable In coshflJ,j
is a constant and thus can beignored. Moreover,
oneimmediately
sees that(2.4c)
becomest(~) =
tanh~(fl) fl
J~j(2.7)
j, jj~ ~(r>
implying
that thequenched
free energyonly depends
on the values of theproduct
ofJ,j
onclosed fines of bonds. The non trivial part of the free energy yN m
(2.4)
is thusFN
"ln(1
+ x4tanh~ là)
+ x6tanh~(fl)
+ .)(2.8)
N
where the coefficients
(x4,
x6;.)
arenon-mdependent
randomquantities.
As a consequence, thequenched
free energy isf
=fl (d
In coshfl
+ In 2 + limN-no éÎln(1
+ x4tanh~ (pi
+ x6tanh~ (fl)
+(2.9)
In the annealed model, the last non-trivial term of
(2.9)
vanishes since n= o for ail i's
(each
x, is a sum of
products
ofindependent
randomcouphngs).
In the limit ofhigh
temperaturelà
-o),
the diiference between annealed free energy andquenched
free energy thus is of ordertanh~ fl
~J
fl4.
Moreimportant,
thesimplest
relevant constraint in the +Jspin glass
is the frustrationJp
on theelementary
squareplaquette
of fourcouphngs Ji, J2, J3, J4,
that isJp
= -JiJ2J3J4(2.lo)
since t(~)
=
-Jp tanh~ fl.
3. +J
Spin
GlassesIn this section we consider a d-dimensional spin
glass
with 1.i.d. nearestneighbour couplings
which assume two
possible
valueJ,j
= +1 withequal probability.
From(2.9),
one sees that the annealed free energydensity
isIn one
dimension,
there are noloops
andfa
isequal
to thequenched
free energydensity f,
since In
coshflJ,j
is a constant, at diiference with the Gaussian case where that variable is random and has to be constrained in order to recoverf [loi.
In section2,
we have recalled standard results which show that it is necessary to constrain variables associated to theloop
variablet(~),
in order toimprove (3.1).
As first notedby
Toulouse and Vannimenus [Si, onehas to consider the intensive variable
a~r=
j j (
13.2) j~j,>j
which has zero mean
value,
Wi " o. Let us remark that it is aself-averaging quantity,
that islimN-co
ON" o for almost ail disorder realizations. The related constrained annealed average
con be
regarded
as theGrand-partition
functioniiN(à>
ÎL) ~ZN
e~'~""NIn the
thermodynamic
limit theLagrange multiplier
p that maximizes~~~'~~
ÀÎCO flÎÎ
~~ ~~~'~~
fixes the frustration on square
plaquettes
to its mean value. Moreover it issimple
to show thefoiiowing inequality:
f
~Ill
~ ~°~~girl, PI
In our case, we have to compute
QN(fl, Al
"
~exp(~j flJ,ja,aj ~j pJp) (3.3)
(a) i,j j
~l')j
Taking
into account that thecoupiings
are dichotomic withequal probabiiity,
one conperform
the gauge transformation J~j -
Jjja~aj
which leavesunchanged
the free energy of a disorder realization of the system as well asJp,
1-e-Jp
-Jp.
As a consequence,(3.3)
becomesiiN(à,
ÎL)"
~~ eXp(~ flÀj ~ pJp) (3.4)
1,j
j~l')j
This expression
(which
ismeaningful
for any dimension d >2)
cannot becomputed exactly including
all thepossible
squareplaquettes
LÎ~I of the d-dimensional lattice. It can be shown that theproblem
isequivalent
to solve a d-dimensional gauge model withoutdisorder,
whosesolution is not known for d > 2.
However, an exact solution can be achieved if we restrict the sum
(3.4) only
on a part(of
order
N)
of theplaquettes.
In other terms the constraint isimposed
over a fraction of theplaquettes,
that is we are able topartially
fix theplaquette
frustration to its mean value.As a consequence, the annealed system exhibits a "residual" frustration related to the non- constrained
plaquettes.
Let us consider the 2d model. It is convenient to limit ourselves to consider one half of the
plaquettes,
which must be chosen in such a way thatthey
do not share anycoupling,
as ithappens
for the black squares of a chessboard. With this restriction(3.4)
contains a sum over-1
_j .-.-.-.-.-.-.-. d=2
j=±1
~=--,
f'~ /~
i /
~ /
Î
~ /
/ /
/ f sq
Î / f~~
~
/ f cr
ac
0 0.5 1-5 2
Fig. 1. +J spin glass m 2d: the annealed free energy fa
(dashed
fine) and the constrained free ener- gies fil(fuit
fine) and fil(dot-dashed
fine) versus temperature T=
1là.
The constraintsare imposed, respectively, on
N/2
independent plaquettes and onN/8
independent crosses of rive plaquettes.products
ofN/2 independent
variables(corresponding
to trieN/2
blackplaquettes)
and trieGrand-partition
function can be rewritten as:QN
(à)
"2'~ (exp(fl(Ji
+ J2 +J3
+J4)
+pJi J2J3J4 ))
~~~(3.5)
where Ji
,
J2
,
J3
,
J4 are trie
couplings
of one of trieplaquettes.
Aftermaximizing
with respectto p trie Gibbs
potential
[1], one obtainsf])
= In 2 + In cosh2fl
+In(
3 +cosh4fl ) (3.6)
fl
2 4 4The functions
fa
andfil
of trie 2d spinglass
are shown inFigure
1. At diiference with trie annealed case trie entropy of trie constrained annealed system never becomesnegative.
We bave thus obtained a realqualitative
improvement.It is also
interesting
to consider trie heat capacity C which can beexplicitly computed.
Forfl
- co one hasC
~J
fl~e~~P (3.7)
The argument of the
exponential
has beenconjectured
inIll]
for thequenched
model and it is diiferent from the exact one-dimension result where the argument is-2fl.
Note that theannealed system also
gives
C~J
exp(-2fl).
Finally,
since we have fixedonly
one half of thesingle plaquette frustration,
it is reasonable to ask what is the residual frustration #a~ Since the blackplaquettes
have zerofrustration,
thisquantity
will be one half of the frustration of a whiteplaquette.
The fourcouphngs forming
a white
plaquette
areindependent
smcethey
are notcoupled by
theLagrange multiplier.
Therefore, the white
plaquette
frustration can be written as « /»~,
and theglobal
frustration #a~ reads:
#a~ = ~ « Jj
»~ (3.8)
2
/ / /
/
~
'( o.
(
/j
É j j]1
/
/
~/ 0
ig. . -
+J pin
lass
for the mortel. The
constraint is imposed on N/2
plaquettes.
where « . » is an verage on the Gibbs measure associated
(3-3).
One ~ ~~~ ~~ ~~~
~~
~ ~~ ~2~ôÎ~ ~'~~
where p* is
alue
which
aximizestotal umber of ouplings is 2N. Since ri(4) " 9(4,
Îl~(4)1
and
sincethe
derivativeUsing the xpression (3.6) for fil we
can
compute the residual
~~~
~ Î~~
~ ~~~ ~ Î Î~~~Î~~ ~ 4 ~ÎÎÎ~Î2fI)1~At
ture #a~ =3~ o.158.
Thisesult
indicates trie
non vanishing residual
frustration.
An alogous
xpression can be found from trie
annealed
free
energy
fa. In thiscase ail
plaquettes can bave negative rustration and
#a =
-
« Ji »~ =-th~(fl)
(3.12) heretrie is in
trie
Gibbs
measure
generated#a(T = o) = -1 since at zero temperature triespins are aligned and trie
is
Both #a and #a~ are reported in Figure 2.Trie
fact that intrie
constrained annealed
model trie residual rustration doesnot
anish indicates that triefree energy estimate could be improved by further
constraints
as we shall
discuss section.
4. Constraints without
Lagrange Multipliers
As far as we
know,
up to now constrainedannealing
bas beenperformed only using Lagrange multipliers.
This is a stronglimitation,
since in most cases it is too diflicult to deriveanalitically
an annealed average
using
more than one constraint. Triespin glass
with dichotomiccouplings
bas some
algebraic
features whichpermit
us to introduce an alternativemethod,
in order to obtain constrained annealed averagesavoiding
in this way trieproblem
of trie minimization of trieLagrange multipliers.
Let us consider a d-dimensional lattice of N points: trie total number of trie
couplings
isdN,
but trie system can befully
describedby
a lower number ofindependent
random variables. In fact trie non-trivial part of triequenched
free energy(2.6)
is a function of trie variables (tÎ~~),
see
(2.4c)
,
defined on trie closed
loops
of trie lattice.Indeed,
from atopological point
of view any closedloop
LÎ~~ can bethought
as trie union of someelementary
squareplaquettes,
say anumber
k,
so that trie related variabletl~l
can be written as:k
t(~) =
tanh~(fl) fl
J)~)(4.1)
1=1
where we bave introduced trie frustrations
Jp
=fl)~~
J, = +1 withequal probability
of trie kplaquettes (trie
index1 runs over trie four sites of eachsquare).
Let us stress that aloop
canbe either connected or
disconnected,
1-e- it is not necessary that trieplaquettes
areneighbours.
Notice that trie
plaquette
frustration waspreviously
defined with a minussign (sec (2.lo)),
butin this context we
neglect
it tosimplify
trie notation.Moreover,
we dot not assumeperiodic
boundary
conditions in this section.In other terms, we con consider trie
plaquette
frustrations as trie random variables of trie system, instead of triecoùphngs,
with no loss ofgenerality.
We shall see that trie number offrustrations necessary for a fuit
description
of trie system isalways
lower thandN,
trie total number ofcouplings.
For instance, in2d,
there are 2Ncouplings,
while trie number ofplaquette
frustrations is of order N when N - co.
However,
trie 2d case exhibits aspecial feature,
since ail triepossible plaquette
frustrations turn out to be a set ofindependent
random variables([et
us recall that we do not assumeperiodic boundary condition).
In triegeneral
case d > 3, this statement is nolonger
valid.Indeed,
[et us consider a cube in a d-dimensionai iattice: a moment of reflection shows that trieproduct
of trie reiated sixplaquette
frustrations bas to be 1, so thatonly
rive frustrations aremdependent
random variables. As a consequence, for d > 3 triechange
of random variables from triecouplings
to trieplaquette
frustrations invoivesoniy
a subset ofthem,
[et us sayN*,
chosen in such a way that
they
areindependent
random variables and thatthey
can build up everypossible
closedloop
on trie lattice.Let us illustrate this
change
of random variables for a 3d lattice of Npoints,
where we can introduce threeinteger
coordinates(x,y, z).
A squareplaquette
has three diiferent orienta- tions,belonging
to a x = constplane,
or y= const, or z = const. In the
thermodynamic
limitone has
Nl/~ planes
for each orientation, and eachplane
is a set ofN~/~ plaquettes,
for a total number of 3Nplaquettes
in the whole lattice.Consider a cube of six
plaquettes,
with one of them on the z= o
plane.
Aspreviously discussed,
one of the J's cannotbelong
to the subset ofplaquette
frustrations that describe trie whole system, [et us say trieÎrustration
of trieplaquette
on trie z = 1
plane.
Let usnow consider trie other cube
sharing
this"neglected" plaquette. By introducing
as random variables of the system the frustrations which are related to the fourplaquettes
that share acoupling
with theplaquette
on the z= 1
plane,
we have toneglect
the sixth frustration of theplaquette
on the z= 2
plane,
since theproduct
of ail theplaquette
frustrations of these twonearest
neighbour
cubes is fixed to(the
presence of theplaquette
on z= 1 in trie
product
isirrelevant,
since it appearstwice).
These arguments con beeasily repeated,
so that trie subset ofplaquette
frustrations thatfully
describes trie whole system can be found as follows: one bas to consider two distinct orientations and to take into account ail trieplaquettes
with theseorientations,
for a total number of 2N.Moreover,
trie third orientationgives only
aplane
ofN~/~ plaquettes (trie
z= o
plane
in ourexample),
so that in thethermodynamic
limit trie number ofindependent
random frustrations N* thatfully
describe trie systems grows as 2N.Trie
generalization
of this result is immediate: in triethermodynamic
limit trie d-dimensional lattice of N points(with
d >3)
is built upby
N distinctplaquettes,
since is trie number of diiferent orientations for aplane
m trielattice,
butonly
a number N*= 2N
of
plaquette
frustrations is necessary to describe trie system. In fact triechange
of random variables involvesonly
the frustrations of theplaquettes
with two well-defined orientations(2N plaquettes), together
with aplane
ofplaquettes
for each one of trie otherorientations,
butm trie
thermodynamic
limit trie main contribution to N* comes from trie former term.Using
this subset of
plaquettes,
whose related frustrations are allindependent
randomvariables,
itis
possible
to build up everypossible
closedloop
on trie d-dimensional lattice.In conclusion the described
change
of variablespermits
us to reduce trie number ofindepen-
dent random variables necessary to describe trie system, from dNcouplings
to N*frustrations,
with N*= N for d
= 2 or N*
= 2N for d > 3. In trie
following,
if notdiiferently specified,
trie term "frustration"always
indicates one of trie N* random variables of trie system.Recalling (4.1),
equation(2-6)
becomes:N*
F = lim
In[1
+~j ~j $"~ $'~~ tanh~~'" '"k)(fl)] (4.2)
~~~° ~
k=i i,.. i~
where one considers all trie
products
of k EIl, N*] elementary plaquettes (trie
indices ii~
12
~ ~
ik run from 1 toN*)
andT(ii,
,
ik)
is trielength
of trieloop
built upby
trie kplaquettes,
while trie overline represents now trie average over trie frustrations.In trie
following paragraph,
we consider trieplaquettes
of trie latticeonly
from atopolo#cal point
of view, 1.e., trie fact that a frustration of aplaquette
is a random variable of trie system(or not)
is irrelevant. At thispoint,
weperform
adecomposition
of trie set ofplaquettes
into subsets such that eachcoupling J,j belongs
to one andonly
one of these subsets. It follows that two distinct subsets can bave in commononly
isolated latticepoints.
For instance, three diiferentdecompositions
of trie bonds of a 2d lattice(trie
blackregions)
are illustrated inFigure
3. It is worth
stressing that,
after thedecomposition,
we get a collection ofsub-systems
whichdo not cover trie whole
original lattice,
e.g. inFigure
3aonly
one half of the 2d lattice iscovered. This kind of
decomposition
divides trieplaquettes
into two classes: trie "black" ones, which areorganized
in groupscorrespoiiding
to triesub-systems (e.g.
trie crosses inFig. 3b),
and trie "white" ones, which do not
belong
to anysub-system-
Coming
back to triefrustrations,
note that ingeneral
there are many "white" or "black"plaquettes
whose frustration is not a random variable of the system.At this
point
we can derive an upper estimate Fa~ of the function(4.2), by treating
ail the random variables of frustration associated to the "white"plaquettes
as annealedvariables,
1.e.by averaging
over these variablesonly
trie argument of trielogarithm
m(4.2),
instead of theai
hi
Cl
Fig. 3- Ising mortel with nearest neighbour interactions on a 2d lattice: decompositions of the set of the couplings in independent subsets
(the
blackareas): a)
squares,b)
crosses and c) elongated crosses.logarithm
itself:~~ lôj
Fa~ = lim
ln[1
+~j ~j 1"~ l'~~~'~~ tanh~~'" "~)(fl)] (4.3)
~~~° ~
k=i i, i~
where R~'°) and R~~) represent trie averages over trie frustrations related to,
respectively,
trie "white" and the "black" squares. In(4.3) only
the termsl~~~ l'~~
with all "black"frustrations do not vanish after
performmg
the "white" average:p~~
(b)
Fa~ = lim
jIn[1
+~j ~j l~~~ l~~~ tanh~~~'
~~)(fl)] (4.4)
~~~°
k=i ii. .>k
where now trie indices ii, .,ik run
only
over trie N~ "black" frustrations. It should be re- marked that afterperformmg
trie "white" average, triesurviving loops
do not connect diiferentsub-systems.
In other terms, anyloop appearing
in trie average(4.4)
can bedecomposed
into a set of
"sub-loops",
each of them limited to asingle sub-system.
Thisimplies
that tanh~~~i> >~k~(fl) con befactorized,
1-e-,T(ii,
.,
ik)
is trie sum of trielengths
of trie vanous"sub-loops".
As a consequence, trie whole argument of trielogarithm
m(4.4)
is factorizedamong ail trie
sub-systems,
and itimmediately
follows:Fa~ =
~
In
il
+~j ~j l~~~ l~~~ tanh~~~"""~~)(fl)] (4.5)
'lJ
Îi
i~ ~~
where trie average is
performed
over trie np "black"plaquette
frustrations thatbelong
to asingle sub-system,
and where nj is trie number ofcouplings
in triesub-system,
so thaté
represents trie total number of
sub-systems,
rescaled with N in triethermodynamic limil.
Recalling (2.5),
from(4.5)
trie lower estimatefa~
of triequenched
free energyyields:
fl fa~
= d Incosh(fl)
+ ~ln[1
+~j ~ l~'~ l~~~ tanh~~'i'
""k)(fl)]
+ lu 2(4.6)
"J
~i
1, .>~
It is
preferable
to choose atopolo#cal decomposition
of trie system such that trie np "black"frustrations
fully
describe triesingle sub-system.
In this context Fa~ ispractically equivalent
to the
quenched
free energy of thesub-system,
a part a factorproportional
to In coshfl.
If npis not too
large,
thecomputation
of Fa~ can beperformed analitically
ornumerically.
Trie correct
multiplicative
factors to transform triequenched
free energy of asub-system
totrie
global
constrained free energy(4.6)
can also be obtainedby
asimple
argument. Indeed triepartition
function Z~z~ of asub-system
made of nsspins
and njcouplings
is trie sum of 2'~Sterms which are given
by
trieproduct
of njexponentials
while trieglobal
constrained Grand-partition
functionQN
of a system with Nspins
and dNcouplings
is trie sum of 2'~ terms which are givenby
trieproduct
of dNexponentials.
In order to compare twoquantities
oforder one, one bas to write trie
followmg equality:
2~~ilN )~/~~~~
= 2~"~
z»~ )~/"'
so that
1 ~ ns
(4
7)/
=--(1-
d")
ln2 + d-fsub
~~
fl
"J "JThis formula is
completely equivalent
to(4.6),
but bas triedavantage
that triequenched
free energyfsu~
of triesub-system explicitly
appears. Trieground
state energy can be estimatedby
trie saddlepoint
method whenfl
- co m triequenched
average of Zn~Ea~(T
=
o)
= ~ maxH((J)) (4.8)
nj j«j
where trie maximum is taken over ail trie a
configurations
for each disorder realization. Onecan also obtains trie residual entropy as
Sa~(T
=
o)
=
(1 ~d)
In 2 + Indeg((J)) (4.9)
"J
where
deg((J))
is trie number of disorderconfigurations
which bave energyH((J)) equal
toEa~(T
=o).
Let us
briefly
resume triediscussion,
m order toclarify
triemeaning
of trie result.Using
a
topological decomposition
of trie system intoindependent sub-systems,
we bave derived a lower estimate of triequenched
free energy thatdepends
from triequenched
free energy of triesingle sub-system.
As discussed in trieappendix,
triequenched
free energy of asub-system
can be obtained
by
a minimization of an annealed average of trieGrand-partition
functionwith
Lagrange multipliers
over ail itspossible loops. Using
ourprocedure,
we are able to get trie annealed averagefa~
of trieglobal
system where trie constraints areimposed
over ail triepossible
rescaled sums of frustration on ail trieloops
that appear in trie varioussub-systems.
To illustrate our
method,
weapply
it in trie case of Section 3, 1-e- triedecomposition
of a2d
spin glass
in "white" and "black" squareplaquettes,
like in achessboard,
seeFigure
3a.In this case ail trie N
plaquette
frustrations represent a set ofindependent
randomvariables,
i e. N* = N. One half of them is related to "white"
plaquettes,
so thatthey
are treated as annealedvariables,
while trie other half related to "black"plaquettes
asquenched
variables.We also bave nj = 4 and np = 1, so that
(4.6)
becomes:fl f])
= 2 Incosh(fl)
+ln[1 tanh~(fl)]
+ In 2(4.lo)
It is easy to check that after trivial
algebraic manipulations
one getsagain (3.6).
Trie
ground
state energy and trie zero temperature entropy can bedirectly computed
from(4.lo),
andrespectively
are:E])(T
=
o)
= -1.5S])(T
=
o)
= o(4.Il)
The method of constrained annealed averages without
multipliers
can beapplied
to non-elementary sub-systems
in an easy numerical way, since one bas not to introduce a set ofcorresponding Lagrange multipliers.
In order to improve(4-la)
and(4.Il),
we bave considereda
partition
of trie 2d lattice intoindependent
crosses of np = 5 squareplaquettes,
as shownm
Figure
3b. We thus obtain a constrained annealed average where ail trie relevant variableson trie non connected
loops
inside trie crosses are frozen. Since nj = 16, from(4.6)
trie free energy iseasily evaluated,
and is shown inFigure
1. Trieground
state energy and trie residual entropy arerespectively
E([(T
=
0)
= -1.484375S([(T
=o)
= 0.00882...(4.12)
The next step is to considered trie
elongated
cross of trie type shown inFigure
3c(np
= 8 and nj =24).
In this case we getE[[~(T
=
o)
= -1.477865S~~(T
=
o)
= 0.0130...(4.13)
At
increasing
trie size of trie subsets of triedecomposition,
trie convergence to triequenched ground
state energy Eo is rather slow as trie numerical result ofIll] #ves
Eo " -1.404 + 0.002 and So" o.075 +
o.oo4). However,
trie mainqualitative
feature(positive
residualentropy)
isreproduced by
our approximations.We bave also
applied
ourtechnique
to trie +J spinglass
in three dimensions. From atopological point
of view, a 3d lattice with Npoints
can bethought
as trie union of distinct cubes distributed in such a way that two of them baveonly
one latticepoint
in common(trie
3d
equivalent
of trie 2dchessboard).
In triethermodynamic
limit their number is N. Let us stress that in this case trie sixplaquettes
of each cube represent trie "black"re#on)for
a total of
~N,
while thereare
~N
"white"plaquettes
between cubes.2 2
It is easy to realize that there exists a
diiferent,
butequivalent, change
of random variables with respect to trie onepreviously described,
which involves np = 5 "black"plaquette
frustra- tions for each cube(trie neglected plaquette
must bealways
triesame), together
with otherplaquette
frustrations in trie "white" areas, for a total of 2Nindependent
random variables.After
performing
trie "white" annealed average, trieproblem
is reduced to trie calculation of/ / / / / /
~ -2.
t /
~ /
/ / /
/
-f~
/ ~ cube
~
0 0.5 1-5 2 2.5 3
T
Fig. 4- +J spin glass in 3d: the annealed free energy fa
(dashed fine)
and constrained free energyf(i~~ (fuit
fine versus T=
1là.
the
quenched
free energy of asingle
cube. In this case we obtain a constrained annealed av- erage of the free energy of theglobal
systemf(f~~,
where the constraints areimposed
over theproducts
of theJ,j's
on all the closedloops
of bonds in the~N independent
cubes. The4
resulting f(f~~
has a verylong analytic expression
and isplotted
inFigure
4, where it is also drawn for comparison the annealed free energy.Unfortunately,
the residual entropy is stillnegative,
thusindicating
that in three dimensions more constraints are necessary to get a fairapproximation
of thequenched
system at low temperature- However, trie free energy should bea
non-decreasing
function of trie temperature, so that trieground
state energy can be estimatedas trie supremum of an annealed
approximation although
trie free energy basnegative
deriva- tive at low temperature [5]. Our constraints on all trieloops
inside trie alternated cubes allows to estimateEo(d
=
3)
> -1.917 whichimproves
trie lower boundsgiven by
trie supremum of trie annealed free energy, Eo > -1.956.5. Constrained Annealed
Averages
on InfiniteLoops
The method introduced in trie previous section could allow us to obtain very accurate estimates
(lower bounds)
of triequenched
free energy. Indeed trie main limitation we bave met is that trie number ofplaquettes
np of triesub-system
should be not toolarge
m order toperform
an
analytic
calculation of itsquenched
free energy. For instance m 2d we bavestopped
trie estimates at trieelongated
cross with np = 8plaquettes.
However, we can consider infiniteloops by estimating
trie free energy ofsub-systems
of infinite sizeonly
in onedirection,
via trieLyapunov
exponent of trieproduct
of random transfer matrices [12].To