HAL Id: jpa-00212443
https://hal.archives-ouvertes.fr/jpa-00212443
Submitted on 1 Jan 1990
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.
Damage spreading in Ising model with increased range
of interaction
S.S. Manna
To cite this version:
1261
LE
JOURNAL DE
PHYSIQUE
Short Communication
Damage spreading
in
Ising
model
with increased
range
of interaction
S.S. Manna
HLRZ,
c/oForschungszentrum (KFA)
Jülich,
D-5170 Jülich1,
F.R.G.(Reçu
le 30 mars1990,
accepté
le 4 avril1990)
Abstract. 2014 We have studied the
spreading
ofdamage
inIsing
models with increased range of interaction.Simulating
a square lattice of linear size 1000 we find that thespreading
temperature
and the Curietemperature
ofmagnetic
transition are the same for both nearest and next nearestneighbour
models whereas for third nearestneighbour
modelthey
may not be the same.1
Phys.
France 51(1990)
1261-1265Classification
Physics
Abstracts 05.50How the effect of a small
perturbation
grows,
particularly
near the criticalpoint
in asystem
showing
aphase
transition is theobject
of"Damage Spreading"
studies.They yield
informationfor the
dynamical
behaviour of thesystem
concerned. In this method one starts with twosystems
which are identical
except
for a small localisedregion (initial damage)
and then bothsystems
are allowed to evolve under identical
dynamics.
It is seen that the effect of initialperturbation
depends
on thetemperature, dynamical
process
of evolution and thegeometry
of thesystem.
Damage spreading
has been studied inIsing
systems
[1-4]
and also in othersystems
like cellularautomata
[5,6]
andspin
glasses
[7,8].
Consider a
square
lattice of size L withIsing spin
variables 1 at each site.Starting
froma
spin configuration
in which allspins
areup
thesystem
is allowed to evolve under aparticular
dynamical
process
and after along
timeequilibrium
is obtainedcorresponding
to a fixedassigned
temperature.
Areplica
of thisspin configuration
is constructed andonly
a small number ofspins
(for
sitedamage
it isonly
the centralspin,
for linedamage
it is a line ofspins
at the centreparallel
to oneedge)
of this lattice areflipped.
Now bothspin
systems
are allowed to evolveobeying
the samedynamics
with the same set of random numbers at the sametemperature.
Damage
isdefined as the fraction of sites in which
spins
of the two lattices are different. This is called the"Hamming
distance" in neural networklanguage.
With time evolution thedamage
mayspread
1262
to the whole
lattice,
it can befixed,
or it can vanish("healing")
in two dimensions.Stanley et
al.[1]
observed that there exists a criticaltemperature
Ts
which is the demarcation line for the finite and infinitedamage
but from their medium size numericalstudy they
were unable to sayvery
precisely
whether7s
is the same or different from the Curietemperature Te
offerromagnetic
transition. ForMetropolis dynamics damage
spreads
for T >Ts
and ceases to grow below thistemperature.
For heat bathdynamics
the behaviour isopposite
[7].
In this
paper
withlarge
scale numerical simulation weinvestigate
whetherTs
isequal
to ordifferent from the Curie
temperature Te
offerromagnetic
transition. For cubic latticesTs
Te
[2,3],
and thus one cansuspect
Ts/Tc
to be lower for alarger
number ofneighbours.
Thus it wassuggested
[11]
to increase the range of interaction for the two-dimensionalIsing
model.We have studied here
damage
spreading
in theIsing
model withincreasing
range of interac-tionusing Metropolis algorithm.
We find that thespreading
temperature
2§
isindistinguishable
fromTe
for both the nearest(nn)
and next nearestneighbour (nnn)
interactions ofIsing
model but for the third nearest(nnnn) neighbour
interactionthey
are found to beslightly
different(see
however the discussion
later).
We alsostudy
how averagespreading
time scales with lattice size. In a recentstudy
Poole and Jan[4]
checked that theexponent
characterising
thisscaling
in heat bathdynamics
isnothing
but thedynamical
exponent
of theIsing
model. Their simulation resultssupport
this idea for three dimensions but notvery
well in two dimensions.In
Metropolis dynamics spins
areupdated
with a Boltzmannweight
factorwhere DE is the
change
in energy of thespin
whenflipped.
At each site a random fraction iscalled and
compared
with theweight
factor.Spin
isflipped
when the fraction isgreater
than orequal
to theweight
factor;
otherwise it remains the same.We
study
the nearestneighbour Ising
model on asquare
lattice with the helicalboundary
conditionusing
asingle
array for the whole lattice. Forvectorisation,
we divide the whole latticeinto two sublattices as in a checker board
lattice,
of dimensions odd(L -1)
by
even(L).
One site khas its nearest
neighbours
at(k+ 1),
( k - 1 ) ,
(k-L+1)
and at(k+L-1)
sites andtwo shadow layers
are used
parallel
to the odd direction for theboundary
conditions. Weupdate
first thespins
on onesublattice
using
thespins
on the other sublattice and then the second sublattice isupdated
with thespin configuration
of the first sublattice. In this way onegets
twonon-interacting
sublattices. For the next nearestneighbour Ising
model we need to devide the whole lattice into nine sublatticesto maintain the condition that the
updating
ofspins
on one sublattice is to be doneusing
thespins
of other sublattices. Inaddtion,
we made simulations with less sublattices. Forexample,
tostudy
the next nearest
neighbour Ising
model on the checker boardlattice,
onupdates
onespin using
the four
spins
on the same sublattice and four on the other sublattice. We studied how far thissublattice
mixing
affects the result. For that wecompared
themagnetisation
belowTe
with theresult obtained from
using
nine sublattices. The difference is small and it becomes smaller forbigger
lattices. Therefore we used checker board lattice forstudying large
next nearestneighbour
Ising
models. For third nearestneigbour Ising
model weupdate
the whole lattice in threeseparate
scans. In each scan we
update
every
third site. The condition for vectorisation of the inner mostloop
is that one cannotupdate
onespin
which is used toupdate
theprevious spin.
All these methods ofupdating
are vectorised and weget
aspeed
of 8.4spins
per
microsecond for nn and 7.0 for nnnIsing
models on oneCray-YMP
processor.
For all three
ranges
of interactions westudy
initial linedamage
and find the fractions ofsites
damaged
when thedamage
touches theboundary
with a siteby
sitecomparison
of thewhole lattice. For critical
temperatures
we use the valuesTc(nn)
=2/ln
1
+vf2-),
Tc(nnn) =
1/0.19019269(5) [9]
andTc(nnnn)
=1/0.1135
[10].
Westudy
damage
spreading
at six999 x 1000 for nn and nnn cases and 1000 x 1000 for nnnn case. For this lattice size we wait
10000 Monte Carlo time
steps
per
spin
fortemperatures
belowTe
and100000,
30000 and 15000steps
fortemperature
ratios1.005, 1.01
and 1.02respectively.
Wego
toequilibrium only
once and successivedamages
are considered one after the other. We average over 20 suchconfigurations
for eachtemperature.
We
plot
infigure
1 the averagedamage
as a function of thetemperature
ratioT/Tc.
We find that for nn and nnn interactions there is finite and zerodamage
at 1.005 and 0.995 ofT/Tc.
Therefore we concludeTs/Tc
= 1.000±0.005 much moreaccurately
than estimated earlier[1]
fornn
only.
However for nnnn interaction weget
15percent
damage
at 0.995 ofT /Tc.
It seems that thismay
be due to an inaccurate value ofTc(nnnn) = 1/0.1135. To
see if thissuspicion
is true westudied how the average
magnetisation
falls to zero as weapproach
the criticaltemperature
for L = 300(see
Fig.
2).
We see that for nn and nnn modelsthey
are still finite atT/Tc
= 1.005 but fornnnn it is zero.
Again
forT/Tc
= 0.995 themagnetisation
for both nn and nnn areconsiderably
higher
than nnnn model. From thisstudy
it seems that the actual value ofTc
in nnnn model is not at1/0.1135
butslightly
less and with an accurate value ofTc, Ts
may beequal
toTe.
Fig.
1. -Average damage
as function of thetemperature
ratioT/Tc :
(A) (nn), (p) (nnn)
and(o) (nnnn).
Next we
study
how the averagespreading
time Te scales with the lattice size as(r,) -
L2in nearest
neighbour
Ising
model.Recently
Poole and Janinvestigated
this[4]
with heat bathdynamics.
They
conclude that thisexponent z
is thedynamical
exponent
of theIsing
model.Their numerical calculations agree very well in three dimensions but it is
slightly
different in twodimensions from the
accepted
value of thedynamical
exponent.
Werepeat
their calculation with somewhatbigger
lattice size(up
to L =128)
withgreater
accuracy to see if this difference isreally
due to theinaccuracy
of finite simulation. Infigure
3 weplot
average
value of thespreading
time(r,)
with lattice size L in alog log
scale. We consideronly
four lattice sizes L =32, 64,
88 and 128. A leastsquare
fit of these fourpoints gives
the value of theslope z
= 2.27. Thus we1264
Fig.
2. -Average magnetisation
as function of thetemperature
ratioTITC :
(à) (nn), (p)
(nnn)
and(o)
(nnnn).
Fig.
3. -We do the similar
study
also forMetropolis dynamics.
In the sameplot (Fig. 3)
we see that the averagespreading
times in this case are much smallercompared
to the heat bathdynamics
and calculate a value for z = 1.15. This resultgives
anothersupport
of the conclusion that thedamage
spreading depends
on theparticular dynamical
process.
The whole
project
was doneusing
around 100 hours of CPU inCray-YMP
Acknowledgements.
’
1 like to thank J. Adler for
suggesting
this work. 1 also like to thank D. Stauffer for many usefulsuggestions
and comments.References
[1]
STANLEYH.E.,
STAUFFERD.,
KERTESZ J. and HERRMANNH.J.,
Phys.
Rev. Lett. 20(1987)
2326.[2]
STAUFFERD.,
Philos.Mag.
56(1987)
901;
COSTA