HAL Id: jpa-00247198
https://hal.archives-ouvertes.fr/jpa-00247198
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.
Glauber Dynamics in a Zero Magnetic Field and Eigenvalue Spacing Statistics
R. Mélin
To cite this version:
R. Mélin. Glauber Dynamics in a Zero Magnetic Field and Eigenvalue Spacing Statistics. Journal de
Physique I, EDP Sciences, 1996, 6 (4), pp.469-492. �10.1051/jp1:1996225�. �jpa-00247198�
Glauber Dynamics in
aZero Magnetic Field and Eigenvalue Spacing Statistics
R. Mélin
(*)
CRTBT-CNRS,
BP 166X, 38042 GrenobleCedex,
France(Received
21 J1Jne 1995, revised 13November1995, accepted
4January1996)
PACS.75.10.Hk Classical
Spin
ModelsPACS.75.10.Nr
Spin-glass
and other random modelsAbstract. We disc1Jss the
eigenvalue spacing
statistics of the Glauber matrix for various models of statistical mechamcsla
one dimensionalIsing
model, a two dimensionalIsing
model,a one dimensional model with a disordered
ground
state, and a SK model with and without aferromagnetic bios).
Thedynamics
of the one dimensionalIsing
modelare
integrable
and theeigenvalue spacing
statistics are non-umversal. In the other cases, theeigenvalue
statistics in thehigh
temperatureregime
are intermediate between Poisson and G-O-E(with P(0)
of the order ofo-à)-
In the intermediate temperature regime, the statistics are G-O-E-- In the Iow temperatureregime,
the statistics have apeak
at s = 0. In the Iow temperature regime, and for disordered systems, theeigenvalues
condense aroundintegers,
due to the fact that the local field on any spm nover vanishes. This property is still valid for theIsing
model on theCayley
troc, even if it is not disordered. WC also study the spacing between the twolargest eigenvalues
as a function of temperature. This quantity seems to be sensitive to the existence of a broken symmetry
phase.
Résumé. Nous discutons les
statistiques
d'écart entre valeurs propres de la matrice de Glau- ber pour différents modèles demécanique statistique (modèle d'lsing unidimensionnel,
modèled'lsing
bidimensionnel, modèle umdimensionnel avec un état fondamental désordonné, modèleSK avec ou sans biais
ferromagnétique).
Ladynamique
du modèled'lsing
umdimensionnel estintégrable,
et la statistique d'écart entre valeurs propres est non universelle. Dans les autres cas, la statistique de valeurs propres dans lerégime
haute temperature est intermédiaire entre la loi de Poisson et la loi G-O-E-(avec P(0)
de l'ordre deo-à)-
Dans lerégime
intermédiaire, lesstatistiques
sont G-O-E-- Dans lerégime
de bassestempératures,
lesstatistiques présentent
un pic pour s = 0. A basse
température
et pour dessystèmes
désordonnés, les valeurs propres condensent autour des entiers, àcause du fait que le
champ
localsur aucun spin ne s'annule
jamais.
Cettepropriété
est encore vraie pour le modèled'lsing
sur l'arbre deCayley
bienqu'il
nesoit pas désordonné. Nous étudions
également
l'écart entre les deuxplus grandes
valeurs propresen fonction de la
température.
Cettequantité
semble être sensible à l'existence d'unephase
avecbrisure de
symétrie.
(*)e-mail: mehn@crtbt.polycnrs-gre.fr
©
LesÉditions
de Physique 19961. Introduction
The ideas of level spacing statistics
emerged
for the first time in the context of nuclearphysics [1-3],
whereWigner proposed computing
statisticalquantities
from consideration of de- tern1inisticspectra. Later,
these ideas wereapplied
toquantunl systems
whose classicalanalogs
are chaotic
[4, Si.
The idea of level spacing statistics is to calculate the difference between two consecutivelevels,
and tostudy
theprobability
of occurrenceP(s)
of a levelspacing
s. The diiferentgeneric
behaviors ofP(s)
are dassifiedaccording
to random matrixtheory [4,
fil. Ageneric
case is theintegrable
spectrum. Each level is labeledby
a set of quantumnumbers,
the energy levels are decorrelated and the statistics are Poissonian:P(s)
= exp
(-s).
If the number of conservedquantities
is toosmall,
it is notpossible
to find a set ofquantum
numbers for eachlevel,
and the levels arecorrelated,
thatis,
there exîsts levelrepulsion.
Therepulsion
islinear and the level
spacing
statistics have the GaussianOrthogonal
Ensemble(G.O.E.) shape:
P(s)
=jse~l~~ (l)
If time reversal invariance is
broken,
and if thesystenl
ischaotic,
the levelspacing
statistics have a GaussianUnitary
Ensemble(G.Il.E.) shape:
Pis)
=~(s~e~l~~, (2)
7r
where the
repulsion
isquadratic.
The ideas ofquantum
chaos have beenapplied
to various fields of condensed marrerphysics,
such as disorderedsystems iii.
Another field ofapplication
is
strongly
correlated electronsystems [8, 9],
where thehope
is to extract more information from finite sizesystenls.
In thepresent
paper we wish toanalyze
thedynamics
of classicalspin
systems
using eigenvalue spacing
statistics. We use here the term"eigenvalue"
rather than"level" since there are no energy
levels,
as in quantum mechanics. We consider the2~
x2~
Glauber
matrix,
with N the nunlber ofIsing
spms, anddiagonahze
it for snlall dusters. Theonly symmetries
are the latticesymmetries,
that we treatusing
grouptheory,
and theglobal 22 symmetry.
We can thusonly study
thedynamics
for a small number ofsites, typically
onthe order of10 sites. The
eigenstates
are notphysical,
except for the Boltzmann distribution whichcorresponds
to the uppereigenvalue
= 0. The othereigenvectors
are notprobability
distributions,
since the sum of their components is zero, so that theirinterpretation
is not obvious.We have studied two diiferent
quantities.
The firstquantity
is the distanceAN (fl)
between the twolargest eigenvalues (one
of thembeing
zero for allfil
for N sites. In the infinitetemperature limit, AN loi
= 1
(see
below for aproof
of thisfact).
Notice that in what followsone unit of time
corresponds
to asingle
spinflip step;whereas
a Monte CarloStep
wouldcorrespond
to Nsingle spin flip
steps. As a consequence, with the normalized lime units thespectrum
lies in the interval[-1,1],
and the normalized distance between the twolargest eigenvalues
isÀN(fl)
"
AN(fl)/N.
Inparticular, ÂN(0)
=
1IN
in the infinitetemperature
hmit. One wouldexpect
the functionAN (fl)
to decrease as the inversetemperature increases,
since the relaxation times areexpected
to belarger
for smallertemperatures.
In the presence ofa broken symmetry
(for
instance in the case of the two dimensionalIsing model),
oneexpects that,
in thethermodynamic limit, Acc(fl)
= 0 if
fl
>pc
since the brokensymmetry
statesbecomes
degenerate
with the Boltzmann distribution m the range oftemperatures p
>pc. By
contrast, in the absence of a brokensymmetry (for
instance for a one dimensionalIsing
modelor for a model with a disordered
ground state),
thequantity Acc (pi
should befinite, except
in the hmitp
= +cc.
Ho~ever,
our finite sizestudy
is far from thethermodynamic
hmit smce wecould
only diagonalize
the Glauber matrix for about ten sites. We were not able todistinguish
between the diiferent
conjectured
behaviors in thethermodynamic
limit.Nonetheless,
westudy
finite size eifects and show that
AN (fl)
decreases as the number of sites N increases. For some models(one
dimensionalIsing
model and the frustrated one dimensionalIsing model),
we find thatAN (fl)
is close to anexponential.
In the case of the two dimensionalIsing model, AN (fl)
is
dearly
nonexponential.
Another
quantity
of interest is theeigenvalue spacing
statistics of the fullspectrum.
We first consider the one dimensionalIsing
model. In this case, thedynamics
are shown to be inte-grable.
Thecorresponding eigenvalue spacing
statistics are found to benon-universal,
with apeak
at s = 0 which increases as thetemperature
decreases. In the case of the two dimensionalIsing
model with nearestneighbor coupling,
the statistics are intermediate between Poisson and G-O-E- for very smallfl,
withP(0)
ce o-à- Asfl increases,
the statistics evolve towards aG-O-E-
shape là
r-
1).
At low temperatures noeigenvalue repulsion
exists, and the statistics exhibit apeak
at s= 0. The
weight
of thepeak
increases withfl. Next,
we consider a frus- trated one dimensional model with an extensiveentropy
at low temperatures. The evolution of theeigenvalue spacing
statistics is similar to the case of the two dimensionalIsing
model.In the case of the
Sherrington-Kirkpatrick (SKI model,
we also have thefollowing
evolution of theeigenvalues
spacing statistics: norepulsion
at very lowtemperatures, repulsion
for inter-mediate
temperature là
r-
ii,
and norepulsion
at low temperatures. Animportant property
of disordered models is that theireigenvalues
condense aroundintegers
at lowtemperatures.
This is due to the fact
that, except
for disorder realizations with zeroprobability
measure,the local field on any site is never zero.
However,
such a behavior also exists for non randomsystems,
for instance for the nearestneighbor Ising
model on theCayley
tree.2. The Glauber Matrix
Glauber
dynamics [11]
are asingle
spinflip dynamics
with a continuous time. Ifp((a), t)
is theprobability
to find the spinsystem
in theconfiguration (a)
at time t, the masterequation
for thesingle spin flip dynamics
is~
p((OE),t)
"If ~((OE))~ p((OE),t)+( w~((OEl,
~OE~,,
CEN))P((OEl,
, ~OE~,..
,
UNI, t).
~~
~=l ~=l
(3)
The
single
spinflip
transitionprobabilities
are defined as theprobabilities
that thespin
a~flips
from a~ to -a~ while the other
spins
remain fixed. Since the Boltzmann distribution is a fixedpoint
of thedynamics (3),
the transitionprobabilities
have the formw~jjajj
=
jjl
a~ tanh
jpJ ~ ajjj, j4j
JEV(~)
where
Vii)
is the set ofneighbors
of the site 1.Denoting
the2~
vector of thep((a), t)
asp(t), equation (3)
can be written asÎ~~~~ ~~~~~'
~~~where the matrix G is the Glauber matrix. Since the Boltzmann distribution is a
steady
state of thedynamics,
itscorresponding eigenvalue
is zeroregarless
of temperature. The matrix Gis not
symmetric.
It can however be related to asymmetric
matrix M. To do so, we notice that the Glauber matrix satisfies the detailedbalance,
that isGa,ppf~
=
Gp,apl~,
wherep(°)
is the Boltzmann distribution. As a consequence,
(0) ~ (0) (0)
~ (0)
jfi)
~~°
~~~
"~
~~fl
~~~
~~fl
~~~
~"
~~°
~~~
We call M the matrix defined
by
Map
=(p(°))
~~~G~p (p(°~)
~~~
ii)
Then,
M issymmetric.
If p is aright eigenvector
of the Glaubermatrix,
then£ Gappp
=
Àpp (8)
fl
is
equivalent
to[ Maô lP1°~)
~~~Pff =lPi°1)~~~
Pa, 191so that
(pΰ1)
~~~
p~ is an
eigenvector
of M. We condude that G isdiagonalizable,
and that all itseigenvalues
are real.The
spectrum
in the infinitetemperature
limit can be understood as follows. If we callllfil =
fl fllalllail
@ @ION), li°1
iai
then the
dynamics
read(ltfil
=)ltfil
+flalltfil, Îi Iii)
so that the
eigenvalues
of the Glauber matrix at infinitetemperature
are of the form=
(
+~i £
/t~,(12)
where /t~
= +l. The
spectrum
in the infinite temperature limit is thus made ofeigenvalues
atinteger
values between -N and0,
with adegeneracy
givenby
the binomial coefficients.Another
property
of G is that forbipartite lattices,
such as the square lattice or theCayley tree,
thespectrum
of G issymmetric:
ifbelongs
to the spectrum then -N is aneigenvalue
also. Theproof
is as follows. LetX(a)
be aneigenvector
ofM,
with aneigenvalue
À:~~~~~ i ~~
~~~~~~~~~~~~~~~~~~ ~
2coslÎflJh~~~°~"
'~°~">°N), (13)
where h~ is defined
by
h~ =
£
aj.
(14)
jEV(4 Let
Y(a)
be defined asY(a)
=(-1)~1°lX(à), (15)
where
u(a)
is the number of upspins
in theconfiguration (a). (à)
is deduced from(a) by flipping
thespins
of one of the two sublattices.Then,
N
(MY)(a)
=
-£-(1-a~tanh(flHh~))(-1)~1°lX(à) (16)
~=i~
~Î~
N ~~~~~~~ ~ ~~'~~2CoshÎpJh~j'~l~i~..~~~~,
,àml
=
(-1)~~°~ IN ~ (l
+ â~ tanin(flJh~ II X(à) (1il
~=l
~ coshÎflJh~)~~~~'"
'
~~' ~~~j
=
-iN
+Àii-ii~l°lXlàl
=
-iN
+ÀiYlal. i181
Given an
eigenvector
X for theeigenvalue
À, we have built aneigenvector
Y for theeigenvalue
-N À.
The diiference between
(Si
and theSchrôdinger equation
is that quantum mechanics pre-serves the scalar
product, leading
to Hermitian Hamiltonians. More over, the space ofphysical
states is a Hilbert space, and each state of the Hilbert state is
physical.
In the case of the Glaubermatrix,
no vector space is present in the sense that the sum of twoprobability
dis- tributions is not aprobability
distribution.However,
somequantities
are conservedby
thedynamics.
It is easy to show that theeigenvectors
of G for the non-zeroeigenvalues
have theproperty
that~ P(°)
" °.
(19)
loEl
This is a
simple
consequence of the fact that the Glauber matrix preserves thequantity
flPlal. 12°1
iai
3. One Dimensional
Ising
Model3.1. INTEGRABILITY OF THE DYNAMICS. In trie case of trie one dimensional
mortel,
trie Glauberdynamics
areintegrable.
To showthis,
we follow Glauber and write evolution equa- tions for the correlation functions. We callRÎ)~_
~
(t)
trien-point
correlationfunction,
withia ~
ip if a~ fl,
that is ' ' "j~~~~ (~) = ~
(~)) (~~)
~i ~~~
~
~~
~
Then
following Glauber,
we write trie evolutionequation
ofR))~
~
(t)
under trie formjRli,1
,~~lu =
-21a~~ iti..a~n
lu iW~~ilali
+ + W~nilaliii, (221
where the transitionprobabilities
aregiven by (4).
We notice that in trie one dimensional case, eachspin
has twoneighbors,
so thatw~((a))
can be written asW~llall
=(i )a~la~+i
+ a~-iIl
,
1231
with ~ = tanh
2flJ. Here,
we takeperiodic boundary conditions,
but the case of openboundary
conditions is similar.
Inserting (23)
into(22)
weget
~~ÎÎÎ
.,~n ~~~
~~ÎÎ~
>~n~~~ ~
~ Î~~'"~~ ~ 'P~'
~~~~e=+1n=1 p#n
The terms with
e = 1 collect the
right neighbors,
and e= -1
corresponds
to the leftneighbors.
The correlation function in
(24)
leads to ain 2)-point
correlator ifàfl,i~+~
=
ip
or to an-point
correlator if not. Theexpression (24)
can bebrought
under the formd (~) ~
l~
~~~~~
'~" ~~~~~~~~"'~"
~~~ ~(~ ~ ~ Ùp,~~+eRÎ)
~Î~_~,~~~~~ ~
j~~
Em OE=1 p=i
' P-1, p+1>. .,~n
n
~ i à
fl(n)
j~~
j~~~
~
M>~OE+E ii,- toe-i>tm+E>~n+i>. ->in'
~~~
If none of the sites
ii,
...,
in
areneighbors,
the termcontaining R("~~l
vanishes.However,
if at least two sites in the set ii,..., in are
neighbors,
we have to take into account a termcontaining R("~~l
in the evolution ofR(").
It is clear that(25)
isnothing
but arewritting
of(3)
in thecase where all the sites have
only
twoneighbors.
The number of distinct correlation functionsis
~j
~ ~=2~, (2G)
~=o
~"
which is
equal
to the number of spinconfigurations.
Thesystem (25)
isintegrable.
Glaubergives
theexplicit
solution forRÎ~I(t).
Theequation giving dR(~)(t)/dt
containsonly
linear combinations ofR(~).
In order to solve for the threepoints
correlationfunctions,
weinject
Glauber's solution into the evolutionequation
forR(~), diagonahze
the associated matrix andget
a first order diiferentialequation,
which isexplicitly integrable
andyields
the second order correlation functions. The entirehierarchy
can be solvedby
this method sincedR(~) /dt
does not containR(P)
with p > k.3.2. EIGENVALUE SPACING STATISTICS. In order to calculate the
eigenvalue spacing
statis-tics,
we need to take all thesymmetries
of the lattice into account.Here,
thesymmetries
are soobvious that we do not require a group
theory
treatment. We work with an openIsing
chain.This
graph
is invariant under the reflection and theidentity
operators. We denote the basis of our "Hilbert" space as(a)).
We use quotes since there is no vector space structure on theprobability
distributions.However,
todiagonalize
the Glaubermatrix,
we can use ananalogy
to quantum mechanics. If R is the reflection operator, we form the combinations
iialie
=
) niait
+
éRiialii
1271This
operation
leads to states with a well defined behavior under the reflection. Theresulting
state is either
symmetric le
=1)
orantisymmetric le
=-1).
Theantisymmetric
state may bezero if
((a))
is invariant under the reflection. The dimension of theantisymmetric
sector is(2~ 21~))
,
(28)
0 2 0
-l 3 -1
flfi
-2
3 5 3
-4 '_ -6
-5 -7 -5
-6 -8 -6
' fi
-7 '
8
.ol ù-1 1
inverse inverse emperature
(a) (b) (c)
Fig. i. Evolution of the
eigenvalues
of the Glauber matrix as a function of the inverse temperature for:a)
an 8 siteIsing
chain m the reflectionantiperiodic
sector and the 22antiperiodic
sector. Noavoided crossings are present, as
expected
for anintegrable
system;b)
the 3 x 4Ising
mortel on asquare
Iattice,
in the representation number 3 and theantisymmetric
22 sector. The evolution of 32eigenvalues
isplotted
in thisfigure; c)
trie frustratedone dimensional mortel with nearest
neighbor
interactions, for 8 sites in the antisymmetric reflection sector, and theantisymmetric
22 sector.and the dimension of the
symmetric
sector is~
ij
-(2
+2 2), (29)
where
[...]
denotes theinteger part. Finally,
we take into account theglobal 22 symmetry
of the Glauber matrix.(The
matrix elements of the Glauber matrix are invariant under thetransformation
(a~)
-(-a~)). Taking
into account all thesymmetries,
wediagonalize
the Glauber matrix in the reflectionsymmetric
and theantisymmetric
sectors, and in the22 symmetric
andantisymmetric
sectors. The evolution of theeigenvalues
of one sector as a function of the inverse temperature isplotted
inFigure
la. No avoided crossings are present, which is what isexpected
for anintegrable
system. The diiference between the twolargest eigenvalues
isplotted
inFigure
2 as a function of the inverse temperature.Eventhough
somedeviations are
visible,
the behavior ofAN là)
is close to anexponential decay.
Finite size eifectsare visible:
AN (fl)
decreases as the number of sites N increases.N=6 - N=7 -~-
N=8 a
N=9 x
N=10 -~-
o-i
O.oi
', u
'#
0 001
0 0 5 1.5 2 2.5
Inverse temperalure
Fig.
2. Evolution of the parameter AN(fl)
as a function of the inverse temperature fl, for differentsizes
(open Ising chain).
1.6
Poisson
14
12
f
08 '06
0.4
02
0
0 0.5 5 2 25 3 3 5 4
normalized separalion
Fig.
3.Eigenvalue
spacing statistics of the one dimensionalIsing
mortel with nearestneighbors couphngs.
The inverse temperature isfl
= i. The statistics are non umversal.
The
eigenvalue spacing
statistics are found to be non universal. Forinstance, Pis)
isplotted
in
Figure
3 forfl
= 1. The
height
of thepeak
at s = 0 decreases as thetemperature
decreases.Eventhough
thedynamics
areintegrable,
theeigenvalue spacing
statistics are not of trie Poisson formil).
Such a behavior for thespectrum
ofintegrable systems
hasalready
been found inthe context of
integrable quantum
fluids(see [9]).
4. Two dimensional
Ising
Model4.1. DYNAMICS. In this case, the
dynamics
are nolonger integrable.
The evolution of the correlation function is not a linearequation,
as it was in the case of theIsing
chain. Thisis
essentially
due to the factthat,
with fourneighbors,
one bas to introduce a cubic term inw~((a)) given by equation (23)
tanin
(flJ(ai
+ a2 + a3 +a4))
"
a(ai
+ a2 + a3 +a4)
+a'(ai
+ a2 + a3 +a4)~, (30)
with
o =
j (8tanh2flJ
tanh4flJ) (31)
a'
=
(tanh 4flJ
2 tanh2flJ). (32)
In the one dimensional case, we could
integrate
thedynamics
becausedR(~)/dt
wasonly
a
function of
R(~)
with k <n. In the two dimensional case
dR(~)/dt
is also a function ofR(~)
with k > n, so that thehierarchy
is nolonger integrable by
this method. It is not becauseone does not know how to solve trie
dynamics
that thesedynamics
are notintegrable.
Theanalysis
of thespectral
statistics of the Glauber matrix may be useful to determine whetheror not there exists some conserved
quantities
in the Glauberdynamics
of theIsing
model.4.2. USE OF GROUP THEORY. We use group
theory
to find thesymmetries
of the clusters for which we shalldiagonalize
the Glauber matrix. Notice that we are restricted to small sizes since the size of the "Hilbert" space isequal
to2~.
Inpractice,
and to have reasonableexecution
times,
we are restricted to N < 13. The first step is to determine thesymmetry
group of thelattice,
that is to enumerate all thepermutations
that leave the lattice invariant.To do so, we do not test all the NI
possible permutations
since triecomputation
time may behuge. Instead,
we use triefollowing procedure.
We first label trie lattice sites andgive
the list of bonds. We determine all thepossible images a(1)
of trie site1,
that is all the N -1 sites.Then,
for each of thepossible images
of trie site1,
we determine trieimages a(2)
of site 2 whichrespect
the latticesymmetry:
if there is a bond between 1 and2,
there must be a bond betweena(1)
anda(2).
If there is no bond between 1 and2,
there must be no bond betweena(1)
anda(2).
ht thispoint,
we bave a list ofpotential permutations beginning
withail)
anda(2).
Then,
we determine all thepossible images
of site 3 which leave the lattice invariant. We thusget
a tree structure,but, during
theconstruction,
some branches shallstop.
At the end of the process, that is whena(N)
has beendetermined,
weget
all triepermutations
which leave triegraph
invariant. The secondstep
is to determine the classes and the table of characters of the group. We use the program in[12],
whichautomatically
determines the classes and the table of characters. In a thirdstep,
we have to determine the size of the blocscorresponding
to the irreduciblerepresentations,
and how many times agiven
irreduciblerepresentation
appears.The dimension of the blocs
corresponding
to the irreduciblerepresentation (j)
isequal
toDim(J~
=
£ xlàlx~~~làl. 1331
geG
j
is therepresentation
of trie group element g in the "Hilbert" space, h is the cardinal of the groupG, x(§)
is the trace of§
andx(Jl(g)
is read from the table of characters at theintersection of the line
corresponding
to therepresentation (j)
and the column of the Mass ofj.
The fourth step is toimplement
a Gram-Schmidtprocedure
to determine the basis of one bloccorresponding
to therepresentation (j).
We first use theprojector
p(J)
=
fl ~(J)jjjj_ j~~j
gEG
A basis element of the Hilbert space is coded as a
binary
number of size N. Zerocorresponds
to a down
spin,
and 1 codes an upspin.
In order to label the basis vectors, we use the decimalrepresentation
of thebinary
number of size N. We denote thecorresponding
vector(~lk).
Theprocedure
consists ofscanning
all the states (~lk) and to determineko
such asP(J)(~lk)
= 0
if k <
ko
andP(J)(~lko) #
0. The stateP(J)(~lko)
is the first vector of the basis that we arelooking
for. Once we have found the first vector of thebasis,
we continue to scan all the states(~lk),
but weproject
them withPÎ~~
"Llilkolàlilko)à.
1351gEG
If
P)~~
(~lk) =0,
weforget
about (~lk) andproject (~lk+i).
IfP)~l
(~lk)~ 0,
wetry
toincorporate P)~~(~lk)
into the basisusing
a modified Gram-Schmidtprocedure [13].
IfP)~~(~lk)
is a linear combination of the basis vectors, then we discard it andproject (~lk+i).
If it isnot,
weincorporate it into the
basis,
afterhaving orthogonalized it,
and we make theprojection
test for(~lk+i).
At the end of theprocedure,
the dimension of the basis must beequal
to(33).
We note that it is notpossible
to store all the components of the orthonormal basis because of limitedstorage capacity.
In order to save memory, we storedonly
the non zero components.The fifth
step
is to take into account theglobal 22 symmetry.
The sixth and laststep
is todiagonalize
the Glauber matrixusing
the basis that has been determined at the fourth step.The size of the matrices to be
diagonalized
are smallenough,
so that we can use the Jacobi method.4.3. RESULTS. We work with a 3 x4 lattice with
periodic boundary
conditions. The number ofrepresentations
isequal
to15,
and the maximal bloc dimension is 335. The spectrum in agiven sector of
symmetry
of the 3 x 4 square lattice ispictured
inFigure
16 as a function of the inversetemperature.
In the limitfl
-0,
we recoverdegeneracies
forinteger eigenvalues (see
relation(12)).
As the inversetemperature increases,
thedegeneracies present
forfl
= 0
are
lifted,
but theeigenvalues
from two diiferentdegeneracies
are not free to cross, due toeigenvalue repulsion.
We studied the evolution of
AN(fl)
as a function offl
for the 3 x 4 lattice and for the 3 x 3 lattice. The result isplotted
inFigure
4. Finite size eifects arevisible; AN (fl)
decreases if Nincreases, and the evolution of
AN là)
isclearly
nonexponential.
In thethermodynamic limit,
Acc là)
= 0 if
fl
>Pc
andAcc là)
> 0 iffl
<pc,
which means that InAcc là)
- -cc iffl
-flj.
Dur finite size
study
is consistent with such a behavior.However,
it would be useful toanalyze larger samples,
which will be done in the near future.We now discuss the
shape
of theeigenvalue
spacing statisticsPis).
If the inverse temperature is verysmall,
thedegeneracies
of thefl
= 0 case are lifted and trie oncedegenerate eigenvalues
spread
outhnearly.
Triecorresponding
statistics areplotted
inFigure
5 forfl
= 0.01. In this case, we find
P(0)
ci o-à and trie statistics are close to trie Poisson law forlarge
s. Iffl
increases, one reaches the
eigenvalue repulsion regime (see Fig. lb).
Lineareigenvalue repulsion
3x3 - 3x4 -~-
o-i
o.oi
o coi
0 0001
~
',,,
'~+,~
l e45
",,
1e.06
0 O.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
inverse temperature
Fig.
4. Evolution of AN(P)
as a function of theinverse temperature fl for the 3 x 4 Iattice
(N
=12)
and the 3 x 3 Iattice
IN
=9).
Finite size effectsare
visible,
and the variations ofàN(fl)
are notexponential.
12
beta =001 beta=1 Po>sson GOE
08
j~ ~~
~
04
,~-f~~
0 2
0
0 0 5 5 2 2.5 3 3 5 4 4.5
normalized spacin9
Fig.
5.Eigenvalue spacing
statistics of the 3 x 4 square Iattice for fl= 0.01 and
fl
= 1. The
Poisson and G-O-E- distribution are
plotted
in dashed Iines. In the regime whereimtially degenerate eigenvalues spread hnearly (fl
=
0.01), P(0)
flf ù-à and the statistics are close to the Poisson Iaw forlarge
s. In theeigenvalue
repulsion regime(fl
=1),
the statistics exhibiteigenvalue
repulsion(P(o)
=
0),
but theshape
ofPis)
isdistinctly
different from the G-O-E-shape.
2
beta=10 Poisson 18
16
1.4
~ 0.6
0A
02
0
0
Fig.
6.Eigenvalue
spacing statistics of the 3 x 4 square Iattice forp
= 10. The Poisson distributionis
plotted
in dashed fines.15 visible m
Figure
5 smce, forfl
=
1, Pis)
r- s for small s.
However,
for s of orderunity, large
deviations to the G-O-E- law occur. For
large fl,
the statistics are not universal(see Fig.
6 forfl
=10),
with apeak
at s= 0. The
weight
of thispeak
increases asfl
increases.5. Frustrated One Dimensional Model
5.1. THE MODEL. We consider the one dimensional
antiferromagnetic Ising
model withantiferromagnetic next-nearest-neighbor
interactions. This model can be seen as a succession oftriangles,
aspictured
inFigure
7 and can be solved ~ia a transfer matrixformalism,
withtrie sites
gathered
as shown inFigure
7. Trie transfer matrix bas trie formT=( ~), (36)
~~~~
4pJ 2pJ 2pJ
À
"
~2pJ
~ i , ~ ~~ÎJ (3i)
In
(36),
trie states are ordered in trie formÎ, Î), Î, i), i, i), i, ii.
Because of trie form(36)
of trie transfer
matrix,
if(~l,
çJ) is aneigenvector
ofT,
then (~l + çJ, ~l + çJ) and (~l çJ, ~l çJ)are
eigenvectors
of T for trie sameeigenvalue,
so that trieeigenvalues
of T are trieeigenvalues
of A + B and A
B,
and trie initial 4 x 4problem
is reduced to two 2 x 2problems,
due to trie time reversal invariance. Triepartition
function issimply ZN
"
TrT~/~
fora N site chain
IN
is assumed to be
even).
The zerotemperature entropy
is found to beextensive,
of the formS(0) IN
= In2/2,
whereas in triecorresponding ferromagnetic problem,
theentropy
is finite atzero
temperatures.
This one dimensionalantiferromagnetic
model bas thus trie sameproperties
as the
triangular antiferromagnet [14], namely
the number of zero temperatureground
states isproportional
to expaN,
with o a constant.Fig.
7. The one dimen5ionaIIsing
model withnext-nearest-neighbor coupling,
and its representa~lion as a succession of
triangles.
The dashed Iines represent how the sites aregathered
in the transfer matrix formalism.5.2. RESULTS. We work with an open chain version, so that the
only symmetries
aretrie inversion and trie
global 22 syInmetry.
We bavealready explained
how to treat thesesymmetries (Sec. 3.2).
The evolution of
AN (fl)
as a function of the inversetemperature fl
isplotted
inFigure
8 for diiferent values of N. We observe finite size eifects:AN(fl)
decreases with N. Trie variationsN=6 -
N=7 -~~
N=8 u
N=9 x
N=10 -~-
o-i
o.oi
0001
0 0.5 5 2
inverselemperature
Fig.
8. Evolution of AN(PI
as a function ofp
for N= 6,7, 8, 9,10 in the case of the disordered one dimensional model of
Figure
7. If N >N', AN(P)
<AN'(fl)1
AN(P)
decreasesexponentially
with theinverse temperature
p.
beta=001 beta=1 Poisson G-O-E 08
06
04
02
0
0 05 1.5 2 2.5 3 3 5 4 45
normaiized spacin9
Fig.
9.Eigenvalue
spacing statistics of the frustrated one dimensional model with next nearestneighbor interactions,
for an inverse temperaturefl
= 0.01 and
fl
= 1.of
AN (fl)
as a function offl
are linear in asemi-log plot,
so thatAN (fl)
decreasesexponentially
with
fl.
The evolution of the
eigenvalues
of the Glauber matrix as a function of the inversetemper-
ature is
plotted
inFigure
lc. Levelrepulsion15
visible for 0.2 <fl
< 1.Trie
eigenvalue spacing
statistics in thehigh temperature regime
areplotted
inFigure
9là
=0.01).
In this range oftemperatures,
trie statistics are intermediate between Poisson and G-O-E-Pis)
bas amaximum,
but is occurs for smallerseparations
than in the G-O-E-case. If the inverse
temperature
increases, the maximum occurs forlarger separations,
andthe
eigenvalue
spacing statistics are close to the G-O-E- law forfl
= 1
(see Fig. 9).
Forlarger temperature là
= 3 in
Fig. loi, P(0)
is not zero(for
instance,P(0)
t 0A forfl
=3)
andPis)
is close to the Poisson law for
large spacings.
6. SK Model
6.1. THE MODEL. This model was
proposed
in 1975 as an"exactly
solvable"spin glass
model
[15].
Forgeneral
reviews on theproblems
ofspin glasses,
we refer the reader to references[16-18].
The SK model is definedby
thedisordered,
infinite range interaction HamiltonianH
=
£ J~a~aj, (38)
(~Jl
with
quenched
random interactionsJ~j
with a Gaussian distributionP(J~j)
=
~~)
exp(-~~~~
~
~~), (39)
27rJ
~
2J
beta=3 Poisson G OE
0.8
0.6
j~
~ ,,~
0.4
'
0 2 "'~,,
0
0 0.5 1.5 2 2.5 3 3.5 4
normaiized separation
Fig.
10.Eigenvalue
spacing statistics of the frustrated one dimensional model with next nearestneighbor
interactions, foran inverse temperature fl
= 3.
In this
section,
westudy
thesymmetric
case à= 0. The case à > is the
object
of Section 7.We refer the reader to the reviews
previously quoted
for the solution of this model. For T <J,
the model is a spinglass,
and aparamagnet
for T > J. In theglass phase,
the model has alarge
number of
thermodynamic phases,
with nosymmetry connecting ground
states. The diiferentground
states areseparated by large barriers, proportional
to the system size. Another feature ofglassiness
is the presence ofageing,
associated with slow relaxation processes.6.2. EIGENVALUE SPACING STATISTICS. In the case of the SK
model,
we do not need tolook for the lattice
symmetries since, except
for somespecial
cases of zeroprobability,
the random infinite range interactions break all the latticesymmetries.
Theonly symmetry
to be taken into account is theglobal 22 symmetry.
The evolution of the
eigenvalues
as a function of the inverse temperature isplotted
inFig-
ure 1la in a
given symmetry
sector. We observethat,
in the limitfl
- +cc, theeigenvalues
condense around
integers.
Thisproperty
is due to the factthat,
in thishmit,
thesymmetric representation M,
ofG,
isdiagonal
since the local field never vanishes(except
for some disor- derconfigurations
of zeroprobability measure),
and that thediagonal
coefficients areintegers.
This cluster
property
m the zerotemperature
limit is an eifect ofdisorder,
and is not related to the existence ofglassiness. However,
as we shall see in Section8,
we find a similar behavior for aglassy ferromagnetic system
with no disorderlThe evolution of
AN là)
as a function offl
isplotted
inFigure
12. The variations ofAN là)
asa function of
fl
areapproximately
linear in asemi-log plot,
so thatAN là) decays exponentially
as a function of
fl.
The
eigenvalue
spacing statisticsPis)
in thehigh
temperatureregime
are intermediate be- tween the Poisson law and the G-O-E- law(see Fig.
13 forfl
=
0.01).
In this regime,P(0)
t ù-à andPis)
is close to the Poisson law forlarge separations.
If the inversetemperature fl
in-o o
-i -1
2
=. -2
-3 -3
-4 -4
-5 -5
-6 6
-7 7
0.01 Ô.l 10 100 0.01 Ô.l 10 100
inverse temperature inverse temperature
(a)
(b)Fig.
ii. Evolution of theeigenvalues
as a function of the inverse temperaturefl, a)
in the 22antisymmetric
sector of a 7 site SKmodel,
and for a given disorderconfiguration; b)
m the 22 antisymmetric sector of the 7 site SK model with aferromagnetic
bios, and for agiven
disorderconfiguration.
Theferromagnetic
bios is à= I.à.
creases, the maximum of
Pis)
occurs forlarger separations (see Fig.
13 forfl
=
0.1)
andfinally
@
reaches the G-O-E- law(see Fig.
14 forfl
=
1).
In the very lowtemperature limit,
the statistics are close to the Poissonlaw,
with apeak
for s= 0
(see Fig.
14 forfl
=
10).
Theheight
of thispeak
increases asfl
increases.6.3. ZERO TEMPERATURE DENSITY OF ZERO EIGENVALUES. At zero temperature, the
eigenvalues
of the Glauber matrix of the SK model areintegers,
and the Glauber matrix isdiagonal
in the naturalbasis, excepted
for some disorderconfigurations
of zeroprobability
measure. As a matter of
fact,
someeigenvalues
are zero. The statescorresponding
to the zerotemperature
zeroeigenvalue
are the metastable states of the SKmodel,
and theirdegeneracy
is
expected
to increaseexponentially
with thesystem
size, asexpected
from the number of solutions to the TAPequations [16].
It is thusinteresting
to calculate thedensity
of zeroeigenvalues
in the zero temperature limit as a function of the systemsize,
even if ourapproach
is restricted
only
to smallsystem
sizes.First,
weanalytically
treat the case of the 3 site SKmodel,
as a warm up exercise.',,
delta 0 -
&,
delta = -~-
delta l.5 u
'~
x delta=2 x, ..
delta 3 -~
X u
',
~, X a
O,I ', ~
' X
~,
~
~
', ~ ",~
o ",,
à X
~,,
"~
'
~ "',~
', ~ 13 '~"~
n ù ",,
', a
~",
'+
0 01
'n ',
'&
', ',
'~
O.COI
0 0 2 0.4 0.6 0.8 1.2 4 6 1.8 2
inverse temperature
Fig.
12. Evolution ofàN(P)
a5 a function of p for N= 9 in the case of the SK
model,
for different values of theferromagnetic
bios à.àNfias
beenaveraged
over 200 realizations of the disorder. Weplotted àN(P)
for à= 0,1,1.à, 2,3.
àN(P)
decreases if theferromagnetic
bios à increases, which isconsistent with the fact that the transition temperature mcreases as a function of à.
beta O.Ol beta=0.1
Poisson G O.E
08
0.6 [[
;lÎ
ÎÎq~ Ii
ÉÎ 1,11Î(j
0 4
Ill
0.2
0
0 0.5 5 2 2 5 3 3 5 4 4 5
normalized spacin9
Fig.
13.Eigenvalue
spacing statistics of the SK model for an inverse temperaturep
= 0.01 and