HAL Id: jpa-00246397
https://hal.archives-ouvertes.fr/jpa-00246397
Submitted on 1 Jan 1991
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.
Information transfer in the lattice gas on random lattices
A. Kitaev
To cite this version:
A. Kitaev. Information transfer in the lattice gas on random lattices. Journal de Physique I, EDP
Sciences, 1991, 1 (8), pp.1123-1130. �10.1051/jp1:1991194�. �jpa-00246397�
Classification
Physics
Abstracts05.20
Information transfer in the lattice gas
onrandom lattices
A. Yu. Kitaev
L. D. Landau Institute for Theoretical
Physics, Academy
of Sciences of the U-S-S-R-, 117940,Kosygina
Str. 2, Moscow, U-S-S-R-(Received 5
April
1991,accepted
25 April 1991)Abstract. A lattice gas model on a random
graph
with vertexconnectivity
p is studied.Information~theoretical arguments are used to
verify
the TAPequations
and toidentify
the transition temperature. The transition from the lowdensity phase
to aglass phase
is continuous at small p and discontinuous atlarge
p.1. Introduction.
One of the earliest
approaches
to thespin glass problem
wasproposed by Thouless,
Anderson and Palmer[I].
The TAPapproach
hasplayed
animportant
role inunderstanding
the nature of thespin glass phase.
An unusual feature of this method is that the TAP free energy(as
afunction of m;
=
(s~) )
is validonly
for some restricted domain in the orderparameter
space while the rest of(m;)
space isunphysical.
For the TAP free energy to bevalid,
finite-size corrections to the mean-fieldtheory
should converge. For the SKmodel,
thisrequirement implies
thatpair spin-spin
correlations are small[1],
N~ '
£ ((s; m;)(s~ m~))~
l,,j
In this paper the TAP
approach
will beapplied
to a morecomplex
model with finiteconnectivity.
It will be shown that thepair
correlation test is insufficient forverifying
the TAP method. Global correlations between allspins together
areimportant
in some cases. Anaccount of these correlations is based on the concept of information transfer. A
key
role inour consideration will be
played by
so-called mutual entropy between aparticular spin
and allspins
atinfinity.
We conclude that if thisquantity
isequal
to zero then the TAP free energy is valid.The model to be studied is a finite
temperature
version of the well known combinatorialoptimization problem
« maximumindependent
set »[2].
It can also beinterpreted
as a latticegas model. Let us consider a
graph
rconsisting
of N vertices and some set ofedges
E.Suppose
that any vertex can be empty oroccupied (by
anatom)
which is describedby
N
spins
»f,,
,
f
~r(f;
=
0, ).
The Harniltonian ismerely
the number of atomsmultiplied by
some chemicalpotential,
3C=-P if;. (1)
l124 JOURNAL DE PHYSIQUE I lNf 8
Besides that, there is an infinite interatomic
repulsion~
that is thefollowing
hard restrictionsare
imposed
f, f~
= 0 for anyedge (I,
/ ~ E(2j
Note that these restrictions lead to frustration if the
graph
r contains oddloops. Finally,
wespecify
thegraph
r to be a randomgraph
with coordination numberp~p
~2 j Moreprecisely,
one chooses withequal probabilities
anygraph
with Nvertices,
each connected to pneighbours.
Calculation of anyquantity
for aparticular graph
should be followedby sample averaging.
The coordination number p is assumed to be constant while N tends toinfinity.
This model is rather similar to the Viana
Bray
model[3]
withantiferromagnetic
bondsonly.
However, the
spin-spin
interaction in our model is stronger,especially
forlarge
p. Another distinction is that the vertex coordination number p does notdepend
on the vertex orsample.
There are two
independent
parameters in the model: p anda =
~ Small values of
T
a
correspond
to a low lattice gasdensity
and a weak interaction(normal phase).
Atlarge
a the
system
condenses into aglass phase.
We shall concentrate on the normalphase
and itsstability.
When
dealing
withself-averaging quantities (free
energy~entropy~ etc.)
one canspeak
about a «typical sample
» and thus avoidexplicit sample averaging.
For aparticular graph
r one can define a distance between two vertices I and
j
as thelength
of the shortestpath
from I toj.
A veryimportant
property of atypical
sparse ~p= const., N
- co
graph
is that it islocally organized
as the Bethe lattice(Caley tree).
Almost allloops
consist of more thanlog~N edges. Therefore~
one can consider alarge enough (I
« radius«log~N)
tree- structuredvicinity
of some vertex and use the information about the entiregraph
toimpose
reasonableboundary
conditions on thispiece
of the system. Thisapproach
has beenalready applied
to the VianaBray
model[4].
2. TAP entropy end free energy.
The purpose of this section is to express the entropy and the free energy in terms of the thermal averages
(f,)
= m, for a
particular graph
r. It is convenient to use a moregeneral
HamiltonianKIT
=
-jja, f,.
The local fields a, are to beadjusted
toproduce
theprescribed
values of m,. The entropies of asingle spin (,
and a pair of spinsf,
and f~(f, f~
=
0)
areS,
=-m,Inm,-(I-m,)In(I-m,j
(3j
$~ =-m,lnm,-m~Inm~-jl-m,-m~)In(I-m,-m,).
For the Bethe lattice, the entropy and the free energy are given
by
the formulaeS
=jjS,- jj
AS,~(4j
, <,1>~E
(
=
La,
m, sisj
where AS,~ =
S,
+ S~ S,~. The TAPequations islim,
= a can be obtained from the free energy minimum condition.
To prove
equation (4),
let us start with the system of N disconnectedspins
and add thebonds step
by
step. One should alsoreadjust
a,through
this process to conservem;,
Any
newedge
added connects two disconnectedparts
of a currentgraph
since the resultantgraph
r has noloops.
Theentropy change
due toadding
oneedge (I, j
iss" s'
=
As;y
+£ (w>'(f;, fy) w'(f;, fj)) B(f,, f~)
<,,<j
where w' and w" are the
probability
distributionsoff;
andf~
before and after theedge
wasadded, B(f;, f j)
is the systementropy
for fixed valuesoff,
andfj. However,
the second termvanishes since
ml'=m), mj'=mj
andB(f;,f~)
can bedecomposed
into the sumBj (f~)
+B~(fj), Therefore,
anyedge
contributesAS~~ to the total
entropy
toyield equation (4).
One can try to extend this
proof
to the case of atypical
sparsegraph
r. Thekey point
in the above consideration is thepossibility
todecompose B(f,, f~ ).
Such adecomposition
isalways possible, provided
anyspin
correlates with at most oneoff;
andfj.
This condition also ensuresthat it is
possible
toreadjust
a; and aj so that to conserve all the averages m~. For the case of
a
typical
sparsegraph,
this conditionimplies
that there are nolong-range
correlations in the system.It is rather obvious that
long-range
correlations are absent for small a(normal phase).
Moreover,
the averages(f;)
= m; are
locally
determined in this case(in
theglass phase
m; may
depend
on theglobal
latticestructure).
The local environments of all vertices are the same, so all m; areequal
to aunique quantity
m~= m
~(
a).
The TAP entropy andequations
become
~=
-m~lnm.+~p-I)(I -m~)In(I-m~) (6)
~P/2)(1-2m*)in(1-2m~)
a = Inm~+
~p-I)In(I-m~)-pln (1-2m~). (7)
We conclude this section
by
a self-consistent calculation of the correlation functionG;j
=((f, m;)(f~ m~))
under theassumption
that the TAP free energy(5)
isvalid, am;
~iY
"$
~~'J
jJ
R;y
=~
= ~
~~ (8)
m~ m; m~
l m
~
~J
~i~. '- ~
~
~ l
-~
m ~
~'J
where ~__
~
(l
if(I, j
~ E'J 0 otherwise
A
simple
calculation based on the Bethe latticeapproximation yields
m~ ~>J
~'J '~*~~ '~*~
l m~ ~~~
where d;~ is the distance between the vertices I and
j. Equation (9)
is anapproximate
one.Actually, G;j
includes the contributions from all thepaths
from I toj (not only
from the shortestone). However, equation (9)
becomesprecise
in N- co
limit, provided
the suml126 JOURNAL DE PHYSIQUE I lNf 8
N~~jjG,)
converges atlarge
distancesd,,.
This requirement leads toinequality
,,
~p
)(ni«,<(I m~))~
~ l or
m~~m~ where m~= (101
,p-I
+IIt is
tempting
toidentify
thecorresponding
value a~ fromequation (7j
with aphase
transitionpoint.
The trouble is that the entropy obtainedby
a substitution of m~ intoequation (6)
isnegative
forlarge
p ~p~ 29 ).
Therefore~ pair spin-spin
correlations reveal nosingularity
at the true critical value a
= a~ if
a tends to a~ from below.
3. Information
propagation
on a tree.In this section more
complex
correlations will be studied. Let us choose anyspin (denoted by 0)
and consider alarge (of
radius r »I)
tree-structuredpiece
of the lattice(Fig. I).
We are tomeasure correlations between
fo
and allspins beyond
the selected part of the lattice. It isobvious that if
to
does not correlate with the system of spins s at theboundary
of the selectedpiece
then it does not correlate with more distantspins
as well. This fact can beexpressed
interms of so-called mutual entropy which is
commonly
used in information theory[5].
)
F>g I. -A tree-structured p>ece of the fait>ce.
Let .~j and
~~ be discrete variables and
>I(xj~ xi)
be theirprobability
measure. Theprobability
measures for the random variables xi and ~j consideredseparately
will be denotedby >vj(.~j)
and >t,~(.~~). The mutual entropy(mutual information) Ij~
is defined as follows~12"~t +~2~~12
where
Sj~
=jj
ii~(.<j,
.<~ ln it-(.;j, .;~)
II,, ,~
Sj
=
jj
w-j(xi
In w>ii-ii Sj
=
jj
>I.z(.;~ In w~~(,<~)< »
Ij~
is the average amount of information about xi which can be obtainedby
a measurement of x~(or
viceversa).
The mostimportant properties
of thisquantity
are as follows1.
Ii~
m o.2.
Ij~
=
0 if and
only
if xj and x~ arestatistically independent.
3.
Suppose
xj and J~ interact with each otherthrough
some intermediate variable J~, that isw(xj,
x~,x;)
=
u(xj, x;)
v(x;, x~).
ThenIj~
wIj,, Ij~
w1;~.
Property
3being applied
to our model means that the mutualentropy Io~
betweento
and s can be used as an upper bound for alllong-range
correlations. For a TAP free energyverification,
it is sufficient to show thatIo~
- 0 as r
- co.
We start from the
assumption
that thelong-range
correlations are absent. This allows todescribe the interaction of the selected
piece
with the restpart
of the latticeby
an effectiveboundary
field a~~~d. Then it ispossible
tocompute Io~
and check theoriginal assumption.
A calculation of any
quantity
for the Bethe lattice involves some recursion from the leavesto the root. We say that a vertex
j
is situated above a vertex I if thepath
from 0 toj
passesthrough
I. Each vertex(except
the leaves and theroot)
has p Ineighbours
above and Ineighbour
below. The root vertex 0 has pneighbours
above. For any vertex I one can consider the subtreeT; consisting
of I and all the vertices above I(Fig. I),
LetZ; (f;)
be thepartition
function ofT;
calculated for a fixed valueoff,.
We shall also use the notationsZ;
=
Z~(0)
+Z;(I),
y~=
Z~(I)/&~ (in particular,
yo=
(to)
=
mo).
The recurrent relations for all thisquantities
are verysimple,
Z;(0)
=
fl (Zj(0)
+Zj(I )) Z;(I)
= e~
fl Z~(0) (12)
j -I
yi=
(I
+e~~fl
j
~Yj
Here
j
runs over all the nearestneighbours
above the vertex I. Thefix-point
fory is
m~
y~ =
(13)
The
boundary
field a~n~d isself-adjusted
toyield precisely
this value of y. Note that the fix-point
is unstable if m~~
lip-
A small deviation of a~~~~d leads toantiferromagnetic ordering
in the lattice domain under consideration.However,
thisinstability
issuppressed
in the entire lattice due to the presence of oddloops. Therefore,
one maymerely ignore
thisinstability.
The next step is to freeze the
boundary spins
at random and tostudy
the distribution of Y;,f, ~y;)
=
z y,
~j~(j '~ w~(s,) (14)
~, s i
Here s~ stands for the collection of the
boundary spins
which fall intoT;, Z,~(f;, s;)
is thepartition
function ofT;
for fixedf;
and s;,Z~(s,)
=
Z;~(0,s~)
+Z;~(l,s;), w~(s;)
is theprobability
of agiven boundary spin configuration.
We use the thermalboundary spin
distributionw~(s;)
=
Z~(s;)/Z;, w;~(f;, s;)
=
Z;~(f;, s;) IT,
which makes a diffierence from theapproach [4].
Theprobabilities wo~(fo, s)
arejust
thequantities
to be used for a mutualentropy calculation,
so one obtains1128 JOURNAL DE PHYSIQUE t4" 8
"
Hi. Hi
In,
" m in + I i?i in
f~(m
dm(15)
m I m
The functions
f, ii-,
have thefollowing properties (for
I #0)
~f(J') dJ'
= '~
J'flJ')
dv= ~
ii
6j
For I
=
0, one should
replace
y~ with m~.Recurrent relations for
Z,~(f,, s,)
andy~(s,)
=
Z,~(l,
s,)/Z~ is,
differ fromequations (12) only by
the presence of the additional function argumentss~ and s, =
is,
~,s~~,
).
From these relations andequations (14), (16)
one can obtain the relation forf,(v, ),
1>~
~
j '
f, (J'j )
= j '~ 3 JJ, i + e
fl fl f~ u
jj
d_>,~ i 7
j
j
J'i
j
where
j
runs over p- Ineighbours
of the vertex I.(For 1=0,
one shouldreplace
y~ and p with m~ and
p).
Assuming
allf,
inequation (17j
to be the samefunction,
one can write this equation in thep~3 '~rn~
,~0 11 42
ioc
l
nn~=O
414~
2
nn~=04145
3
nn~=Q
4154
nn,=Q
d16' so
I
~
50
~
~
~
~d 40
~
i
20
, ,
/
~'~
~
~ 'oo
00 02 04 05 08 10
rTl
Fig. 2. Local magnetization distribution for p 3.
p=37 [m~ =0.1429)
o.o
m~=Q.127
2
m~=0.1275
3
m~=0.128
6.0
~
/~~/
_p
4:0
2.0
3
O-O
0.0 0.4 0.6 O-B 0
m
Fig.
3. Local magnetization distribution for p =37.
symbolic
formf,=Afj,
where A is a nonlinearintegral operator.
Theoperator
A should beapplied repeatedly
to the initial functionfinit(y)
"
(i y*)
3(y)
+ y* 3(y 1) (18)
If this process converges to the trivial
fix-point f~y)
= 3(y y~)
then the mutualentropy Io~
tends to zero.Numerical
computations
have been done for different p and a. The restrictions(16)
wereimposed
to avoid an undesirableinstability. Any
deviation from these restrictions wasbeing
corrected
through
thecomputation
process(after
eachiteration).
Aftersufficiently
manyiterations the process
always converged
to somefix-point f~~y).
Then a modifiedoperator Ao
wasapplied
to thisfix-point
to obtainfo(m ).
The results for p = 3 and p =37 are shown in
figures 2,
3(different
curves are indexed withm*(a )
rather thana).
The transition from the trivialfix-point
to a nontrivial one have different nature for small andlarge
p. For smallp, the critical value aA is
equal
to a~. Thepeak gradually
become wider as a(or
m~)
increases so the transition seems to be continuous in this case. On thecontrary,
forlarge
p a
sharp
transition occurs at aA < a~. The critical value p~separating
« small » and «large
»p is found to be
greater
than 25 and less than 30.(A
moreprecise
identification of p~requires
a modification of the numericalalgorithm
to reducerounding errors).
l130 JOURNAL DE PHYSIQUE I M 8
4. Discussion.
It is shown that information-theoretical arguments can be used to
verify
the TAPapproach
and to
identify
thetemperature
of thenormal-to-glass phase
transition. For small p the transition is similar to the continuous transition in the SK model. Forlarge
p~ the transition issharp
and seems to be similar to that of the Pottsglasses [6, 7].
It can beinterpreted
asergodicity breaking.
Just above a~ theconfigurational
space is subdivided into manyweakly
connected parts which are refer to as system states. For this reason any vertex I is
characterized
by
a distribution of m, =(f,)
rather than aunique
value of m,. One canidentify (at
least if a-
a()
this distribution with the functionfo(m)
calculated above.However~ the Potts
spin
systemundergoes
twophase
transitions as a temperatures decreases[7].
The first transition at T=T~
is apurely dynamical
one. Ananalytical expression
for the free energy athigh
temperature remains valid up to the second transitiontemperature
T~
~
T~.
As to ourmodel,
a finite mutual information between anyspin
andinfinity
does notnecessarily-
mean that the TAPapproach
is notapplicable.
Theonly physical
consequence one can
definitely
conclude in this case is that there arelong-range, long-time dynamical
correlations. The transition at a~ islikely
adynamical
one while the free energy has asingularity
at a~ ~ a ~. In the intermediatephase (a~
~ a ~ a~)
the TAP entropy(6)
isprobably
valid but it must bedistinguished
from the average entropy of a system state. Thedifference S~~~j =
Sr~p £
is called theconfigurational
entropy[7].
The number of states which contribute to the total free energy isexp(S~~~i).
Thequantity £
can be obtainedby averaging
the TAP entropy(4)
over m,. The result looks as follows~
=
(l p/2) lo
+~p/2j (1 m,) S,
~
where
lo
= (- m In m I m In m
fo(m
j dmj19)
S~ = j- v In i'- i'j In
jl i')j j,(>')
di~.The
configurational
entropy perspin
ispositive
at a~. It becomes zero at some a~ ~ a~. So the number of states isexponential (e~~)
in the intermediatephase
and finite orpolynomial
in theordinary glass phase (a
~a~).
Acknowledgments.
I wish to thank M.
Feigelman
and L. Levitov forhelpful
discussions on this work.References
ill
THOULESS D. J., ANDERSON P. W. and PALMER R. G., Philos. Mag. 35 (1977) 593[2] GAREY M. R. and JOHNSON D. S., Computers and
Intractability
(W. H. Freeman andCompany,
San Francisco, 1979).
[3] ViANA L. and BRAY A. J., J. Phys. C18 (1985) 3037.
[4] CARLSON J. M., CHAYES J. T., CHAYES L., SETHNA J. P. and THOULESS P. J., Europhjs. Lelt. 5
(1988)
355.[5] For example, see KoLmoGoRov A. N., Information Theory and
Theory
of Algorithms (Nauka, 1987) (in Russian).[6] GRoss D. J., KANTER I. and SOMPOLINSKY H.,