arXiv:0905.4816v1 [cond-mat.dis-nn] 29 May 2009
K. Shwarz
1
, A. Karrenbauer
2,3
, G. Shehr
4
, and H. Rieger
1,4
1
TheoretishePhysik,UniversitätdesSaarlandes,66041SaarbrükenGermany
2
MaxPlankInstitute forInformatis, UniversitätdesSaarlandes,66041
SaarbrükenGermany
3
ÉolePolytehniqueFédéraledeLausanne,Lausanne,Switzerland
4
LaboratoiredePhysiqueThéorique,UniversitédeParis-Sud,91405OrsayFrane
Abstrat. Domain walls, optimal dropletsand disorderhaos at zerotemperature
arestudiednumeriallyforthesolid-on-solidmodelonarandomsubstrate. Itisshown
thattheensembleofrandomurvesrepresentedbythedomainwallsobeysShramm's
left passageformula with
κ = 4
whereas their fratal dimension isd s = 1.25
, andthereforeisnotdesribedbyStohasti-Loewner-Evolution(SLE).Optimaldroplets
with a lateral size between
L
and2L
have the same fratal dimension as domainwalls but an energy that saturates at a value of order
O (1)
forL → ∞
suh thatarbitrarilylargeexitationsexistwhihostonlyasmallamountofenergy. Finallyit
isdemonstratedthat thesensitivityofthegroundstateto smallhangesoforder
δ
inthedisorderissubtle: beyondaross-overlengthsale
L δ ∼ δ − 1
theorrelationsofthe perturbedgroundstatewiththeunperturbedgroundstate,resaledbytheroughness,aresuppressed andapproahzerologarithmially.
1. Introdution
Domainwallsindisorderedsystemsplayanimportantroleinunderstandingthestability
of the ordered phase, the energetis oflarge sale exitations, the asymptotidynamis
in and out of equilibrium as well as the sensitivity to hanges of external parameters.
They have been studied quite intensively in the reent years for Ising spin glasses
[1,2, 3,4,5, 6,7,8℄, XY spinglasses [9℄, random eldsystems [10,11,12, 13℄, random
ferromagnets[10℄,disorderedelasti manifolds[15, 16,17,18,19,20℄, andmanyothers.
Two domain wall properties are prominent: the rst onerns energy and an be
haraterized by the saling behavior of the domain wall energy with their lateral size,
whihgivesrise toa rst, sometimes universal exponent, the stiness exponent
θ
. Theseond onerns geometryand gives rise toanother, sometimes universal exponent, the
fratal dimension
d s
,or,inase the domainisnot fratal,a roughnessexponentζ
. Theinterplay between energetis and geometry of the domain walls (i.e. between stiness
exponentandfrataldimension)determineshowsensitivethesystemstateistohanges
ofeither externalparameterslikethe temperatureoraeld, orinternalparameters like
small disorder variations. This sensitivity is often extreme in glassy systems and goes
under the name of haos [1℄.
Domain walls of glassy systems in two spae dimensions represent fratal urves
in the plane and the question arises, whether they fall into the general lassiation
sheme for ensembles of random urves desribed by Stohasti Loewner Evolution
(SLE) [21, 22℄. Reently indiations were found that domain walls in 2d spin glasses
(at zero temperature) are indeed desribed by SLE [23, 24℄, at least for a Gaussian
distributionofthebonds,but apparentlynot forbinaryouplings[25℄. Alsothe domain
walls in the random-bond Potts model at the ritial point (i.e at nite temperature)
were found to be numerially onsistent with SLE [26℄. It appears natural to ask,
whether the domain walls in other two-dimensional disordered systems are potential
andidates for adesription by SLE.
In this paper we study domain walls and haos at zero temperature in the solid-
on-solid (SOS) model on a disordered substrate. This is a numerially onvenient
representation of a two-dimensional elasti medium, with salar displaement eld,
interating with quenhed periodi disorder. It has been studied to desribe various
physial situationsranging from vortex latties in superondutors to inommensurate
harge density wavesand rystalgrowth onadisorderedsubstrate [27,28,29℄. Here we
fous onthree questions: 1) are domainwalls inthis modeldesribed by SLE, 2)what
is the relation between size and energy of optimal exitations(droplets) inthis model,
3) doesdisorder haos exist in the ground state of this model? After a briefsummary
ofwhatisalreadyknown aboutthemodelandadesriptionofthenumerialmethodby
whihweompute the groundstateand thedomainwallsthese threeissues arestudied
in separate setions. The paper ends with adisussion of the resultsobtained.
1.1. Model
We onsider the solid-on-solid model on a disordered substrate dened by the
Hamiltonian
H = X
(ij)
(h i − h j ) 2 , h i = n i + d i ,
(1)with
i ≡ (x i , y i ) ∈ Z 2
. In Eq. (1) the height variablesn i
(i = 1, . . . , N
) take oninteger values
n i = 0, ± 1, ± 2, . . .
and the osetsd i
are independent quenhed random variables uniformly distributed between 0 and 1. The sum is over all nearest neighborpairs
(ij)
of aretangularlattie ofsizeL x × L y
(L x = L y = L
if not statedotherwise).The boundary onditions will be speied below in the ontext of domain walls. The
HamiltonianinEq. (1) desribes adisretemodelof atwo-dimensionalelasti medium
in a disordered environment. In the ontinuum limit, it is desribed by a sine-Gordon
model with random phase shifts (and in the absene of vorties), the so alled Cardy-
Ostlundmodel [30℄,
H CO = Z
d 2 r( ∇ u(r)) 2 − λ cos(2π[u(r) − d(r)]) ,
(2)with a ontinuous salar displaement eld
u(r) ∈ ( −∞ , + ∞ )
and quenhed randomvariables
d(r) ∈ [0, 1]
. Disretizing the integral and performing the innite strong ouplinglimitλ → ∞
one reovers (1).It is well known that this model (1, 2) displays a transition between a high
temperaturephase,
T > T g = 2/π
wherethedisorderisirrelevantandalowtemperature phase belowT g
, dominated by the disorder. The high-temperature phaseT > T g
isharaterized by a logarithmi thermalroughness
C(r) = h (h i − h i+r ) 2 i ∼ T log r ,
(3)where
h . . . i
denotes thethermalaverageand. . .
theaverageoverthe quenhed disorder.The low-temperature or glassy phase is instead superrough, haraterized by an
asymptotiallystronger (log-square) inrease of
C(r)
:C(r) ∼ c(T ) · log 2 r + O (log(r)) ,
(4)whihmeans
ζ = 0
, asexpeted for arandom periodisystem. ClosetoT g
, aCoulombGasrenormalizationgroup(RG)analysis tolowest order gives
c(T ) ≃ (1 − T /T g ) 2 /2π 2
[29℄, in rather good agreement with numerial simulations [31℄. At
T = 0
, numerialsimulations give the estimate
c(T = 0) ≈ 0.5/(2π) 2 ≈ 0.012
[16, 32℄. While earlierstudies,based onnearly onformal eld theory [33℄,laimed anexat resultfor
A(T )
,prediting
A(T = 0) = 0
,in lear ontradition with numeris, amore reent approah basedonFRG,inorporatingnonanalytioperatorspreditsanon-zeroA(T = 0)
whihompares reasonably withnumeris[34℄.
For free or periodi boundary onditions, the Hamiltonians (1) and (2) have a
disrete symmetry, the energy is invariant under a global height (displaement) shift
n i → n i + ∆n
(u(r) → u(r) + ∆n
), where∆n
is an arbitrary integer. This symmetryFigure 1. Left: Ground state
n 0
of a200 × 200
system with the boundary sites(indiate in green) xed to
n i = 0
. The dierent height valuesn i
are grey-oded(dark=lowvalues,bright=highvalues). Middle: Groundstateongurationofthe
samesystemastotheleftwiththeupperhalfoftheboundarysites(indiatedinred)
xedto
n i = 1
and thelowerhalf (indiatedin green)ton i = 0
. Right: Diereneplot betweenthe Left and Middle plots: in the lowerwhite region the groundstate
ongurationis idential to theorrespondingsites in theleft gure, whereasin the
upper greyregion they dier by exatly
∆n = 1
from the orresponding site in the middle panel. The border between the whiteand the grey region is adomain wall,representingastepintheheightproleofthegroundstate.
will not be broken in the low temperature phase of the innite system and true long-
range order at
T < T g
is absent, i.e.h h i i = 0
. Conomitantly the model (1), with free or periodi boundary onditions, has innitely many ground states, whih dier by aglobal shift
∆n ∈ {± 1, ± 2, . . . }
.1.2. Domain walls
By an appropriate hoie of boundary onditions one an fore a domain wall into the
system, whih is most easily visualized at
T = 0
(.f. Fig. 1): onsider the squaregeometry and x the values of the boundary variables to
n i = 0
. This yields a uniqueground state onguration
n 0 i
. If one xes the boundary variables ton i = +1
, theorresponding ground state would be
n ′0 i = n 0 i + 1
. A domain wall induing boundaryondition isone, in whih the lowerhalf of the boundaryvalues are xed to
n i = 0
andthe upper half to
n 0 i = 1
. The ground state of this set-up is thenn ˜ 0 i = n 0 i
in some,mainlythe lowerregion of the system, and
n ˜ 0 i = n ′ 0 i
in the rest - both region separatedby a domainwall of non-trivialshape.
It turns out that these domain walls are fratal [16, 18℄, whih means that their
lengths
l dw
sales with linearsystem size asl path ∼ L d s ,
(5)with
d s > d − 1 = 1
. The numerial estimate ford s
isd s = 1.27 ± 0.02
[18℄. Suha fratal saling of zero-
T
domain walls is also found for spin glasses, in the 2d EAmodel with Gaussian ouplings it is
d s,SG = 1.27 ± 0.01
, and with binary ouplings itis
d s,SGB = 1.33 ± 0.01
. On the other hand zero-T
domain walls indisordered IsingorPotts ferromagnets are rough (i.e. are haraterized by algebrai orrelations) but not
fratal.
The energy for suh a domain wall, given by the dierene between the energy of
the groundstateof the systemwith the domainwallinduingboundaryonditions and
the one with homogeneous boundary onditions, inreases with
L
logarithmially∆E ∼ log L .
(6)This result, whih was obtained by numerial simulations [16℄, is onsistent with the
usual saling relation
∆E ∼ L θ
together with the exat resultθ = d − 2 + 2ζ = 0
(thanks tostatistialtilt symmetry[35℄). This logarithmi behavior isharateristi of
amarginalglassphase,desribed byalineofxedpointindexedbytemperature(whih
isheremarginalintheRGsense). Foromparisonthestinessexponentin2d(3d)spin
glassesis
θ SG2d = − 0.28 ± 0.01
andθ SG2d = 0.3 ± 0.1
(andthusharaterizedbyaT = 0
xed point),whereas for disordered Ising orPotts ferromagnets
∆E ∼ L d−1
.1.3. Method
The ground states of (1), i.e. the onguration
n 0 = (n 0 1 , . . . , n 0 N )
with the lowestvalue for the energy
H[n 0 ]
for a given disorder ongurationd = (d 1 , . . . , d N )
, anbe omputed veryeiently using aminimum-ost-ow-algorithm[36, 16, 37℄. For the
spei details in whih domain walls are indued in the ground state it is useful to
reapitulatethe mappingontoa minimum-ost-owproblem.
After introduing the height-dierenes
n ∗ ij = n i − n j
(integer) andd ∗ ij = d j − d i
(
∈ [ − 1, +1]
) along the linksk = (i, j)
on the dual lattieG ∗
one obtains a ost (orenergy) funtion that lives onthe dual lattie
H[n ∗ ] = X
k
(n ∗ k − d ∗ k ) 2 .
(7)The ongurations
n ∗ k = (n ∗ 1 , . . . , n ∗ M )
, whereM
is the number of links (or bonds) ofthe original lattie, onstitute a ow on the graph
G ∗
. Suppose the original model(1) has free boundary onditions. Then the sum of the height dierenes along any
direted yle in the original lattie vanishes. Therefore the divergene of
n ∗
vanishesatall sites
i
:( ∇ · n ∗ ) i = 0 ,
(8)whih means that the ow
n ∗
onG ∗
, in order to give rise to a height eldn
on theoriginal lattie
G
has to be divergene-less, i.e. without soures orsinks. The problem of determining the groundstaten
of (1) isthus equivalentto nd the own ∗
with theminimum ost (7) under the mass-balane onstraint (8) - i.e. a minimum ost ow
problem,for whih there exist very powerfulalgorithms[36,16, 37℄.
Enforing one domain wall, or step of height one, into the ground state of (1)
by appropriate boundary ondition is then equivalent to modify the onstraint (8) at
exatly two sites, the start and end point of the domain wall (see Fig. 2a-d). As an
d) a) b)
c)
Figure 2. Dierentgeometries andonstraintsondomain wallsonsideredhere: a)
boundaryonditionsinduingastep/domainwallasinFig.1.b)Boundaryonditions
induingastep/domain wall runningdiagonallyfrom oneorner oftheretangular
lattieto theopposite one. ) Boundary onditionsforairular domaininduing a
boundary alongtheequator withtwodierent orientationsof theunderlying lattie.
d)Boundaryonditionsforahalf irledomainandadomainwallwithonexedend
attheoriginandafreeendontheouterhalfirle.
example onsider the ase in whih one wants the domain wall to start at the point
(x, y) = (1, L/2)
and end at(x, y) = (L, L/2)
of asquare lattie. Then one hooses theboundary onditions for
n i
as follows (.f. Fig. 2 a): one xes the values forn i
at thelower half of the boundary (i.e. at
i = (x, 1)
forx = 1, . . . , L
andi = (1, y )
and(L, y)
for
y = 1, . . . , L/2
) ton i = 0
, and the values forn i
at the upper half of the boundary(i.e. at
i = (x, L)
forx = 1, . . . , L
andati = (1, y)
and(L, y )
fory = L/2 + 1, . . . , L
)ton i = 1
. Translatingtheseboundaryonditionsfortheheightvariablesn
intoonstraintsfor the ow variables
n ∗
one immediately sees that at the point(x, y) = (1, L/2)
and(x, y) = (L, L/2)
,wherethestepintheheightprolestartsandterminates,respetively, the onstraint (8)is modiedinto( ∇ · n ∗ ) (1,L/2) = +1 , ( ∇ · n ∗ ) (L,L/2) = − 1 .
(9)In other words: the indued step sends a unit of ow from the starting point of the
domainwall,whihisthe asoureof unit strength, arossthe sampletothe end point,
the sink, along an optimal (minimum ost/energy) path. In what follows we identify
domainwallsimmediatelywith the optimal path forthe extra ow unit dened by the
modied mass balaneonstraints (9).
With the help of this onept one an then also onsider situations in whih the
startingpointofthedomainwallisxed buttheendingisonlyforedtobeonaspei
region of the boundary, opposing the starting point (see Fig. 2d). Suppose one wants
he domain wall to start at
i s = (1, L/2)
, and terminate somewhere on the opposingboundary
i t = (L, y)
withy ∈ { 1, . . . , L }
. Then one introdues an extra node into thedual graph
G ∗
, denoted as the target node, onnets it with bonds of zero ost to allsites onthe terminalboundary,andassignsasinkstrength
− 1
toit. The sourenodeisthe one losest to
i s
in the dual graph and has sourestrength+1
. The minimum ostow of this arrangement is then the desired groundstate onguration with a domain
wall startingat
i s
and ending somewhereon the opposite boundary.2. Shramm-Loewner evolution (SLE)
Sine the domain walls as dened above represent fratal urves embedded in a two-
dimensional spae the question arises whether they fall into the lassiation sheme
of Shramm-Loewnerevolution (SLE)likeloop-erasedrandomwalks, perolation hulls,
and domainwalls atphase transitions in2din the salinglimit[22,21℄. The neessary
(and suient) ondition for a set of random urves onneting two points on the
boundary
D
of a domainto be desribed by SLE are 1) the measure for these randomurves has to fulll a Markov property, 2) the measure has to be invariant under
onformalmappingsofthe domain. Reentlyitwassuggestedthat alsodomainwallsin
2dspin glasses an bedesribed by SLE [23, 24℄althoughneither onformalinvariane
isfullledforeahindividualdisorderrealizationnortheMarkovpropertyafterdisorder
averaging. However, onformal invariane might hold for the disorder averaged model
and the Markov property mightbe fullled almost always ina statistialsense.
A single parameter
κ
parameterizes all SLEs and it is related to the fratal dimension of the urve viad s = 1 + κ/8 .
(10)The parameter
κ
is the diusion oeient of the Brownian motion that underlies theSLEand generates viaa randomsequene ofsimple onformalmaps thefratalurves.
Inadditiontothefrataldimensionitdeterminesvariousothergeometriandstatistial
propertiesoftheSLEurves. Oneofthemisforinstanethe probabilitythataurvein
the upperhalfplane
H
generatedbySLEwillpass totheleftofagiven pointz = x + iy
is given by Shramm'sleft passageformula [38℄
P κ (z) = 1
2 + Γ κ 4
√ π Γ 8−κ 2κ 2 F 1
1 2 , 4
κ ; 3 2 ; −
x y
2 ! x
y ,
(11)where
2 F 1
is the hyper-geometri funtion2 F 1 (a 1 , a 2 ; b; z) =
∞
P
k=0
Γ(k+a 1 ) Γ(a 1 )
Γ(k+a 2 ) Γ(a 2 )
Γ(b) Γ(k+b)
z k k!
.Sinetheprobabilitydependsjustontheratiobetween
Re(z)
andIm(z)
,itissometimesuseful toreplaethis ratioby afuntion ofanangle. Wedeided touse:
tan (φ) = x/y
.So
φ ∈ ] − π 2 ; π 2 [
isthe angleattheoriginbetween theimaginaryaxisandz
. Thisformulaholds for the domain of the SLE being the upper halfplane with the start point of the
urvebeingidentialtotheoriginofthe oordinatesystemandtheendpointatinnity.
A standard hek whether an ensemble of random urves is a potential andidate
for SLE therefore is to simultaneously determine their fratal dimension
d s
and the10 100 1000 10000
10 100 1000
L path
L
∝L 1.25 L•L geometry L•2L geometry
Figure 3. Averagelength ofdomain wallsspanning thesystemfrom oneend tothe
oppositeoneasafuntionofthesystemsize
L
inalog-logplot. Thestraightlinesareleastsquaretsto(5)andyieldtheestimateforthefrataldimension
d s = 1.25 ± 0.01
.left passage probabilities, and to test whether the latter t (11) with
κ = 8(d s − 1)
[23, 24,25℄.
In Fig. 3 we show our data for the average length of a domain wall in the
onguration depited in Fig. 2 b, i.e. starting at the point
i s = (1, 1)
and endingat
i t = (L, L)
in aL × L
geometry and ati t = (L, 2L)
in aL × 2L
geometry. Leastsquare ts to the saling law (5)yield the estimate
d s = 1.25 ± 0.01
whih agrees withthe value found in[18℄.
This value for the fratal dimension would imply
κ = 2.00 ± 0.08
if the domainwallsare desribed bySLE. Next wedetermined fordierentpoints
(x, y )
of thelattiethe frequeny that adomain wall passes tothe left of it, yielding a probability
p(x, y )
,whih we ompared with Shramm's left passage formula
P κ (x, y)
for xedκ
. For theirledomainand the hoieof thestart and endpointsof the domainwallasshownin
Fig. 2, the formula (11) ismodied:
Let
E
be the unit irle in the omplex plane. The Cayley funtiong : E → H , z 7→ i 1+z 1−z
maps the unitirle onformallyintothe upperhalfplaneH
. Furthermoreg( − 1) = 0, g(1) = ∞
, thus urves inE
starting atz = − 1
and ending atz = 1
areonformallymapped onurvesin
H
. Forthe latter, if the urvesare desribed by SLE,Shramm'sformula(11) holds, sothat for the former, the modiedformula
P κ,g (z) = 1
2 + Γ κ 4
√ π Γ 8−κ 2κ 2 F 1
1 2 , 4
κ ; 3 2 ; −
Re g (z) Im g(z)
2 !
Re g(z)
Im g(z) .
(12)holds. Hene we lled the unit irle with ner and ner grids
G
approximating better and better the ontinuum limitfor the urves inE
that we want to hek for SLE. Forthis we dene the funtion
f G (κ) = X
(x,y)∈G
[p(x, y ) − P κ,g (x, y )] 2
(13)that measures the umulative squared deviation of the probabilities
p(x, y)
from the0 20 40 60 80 100 120 140
3 3.5 4 4.5 5
f G ( κ )
κ
0 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.01
-1 -0.8 -0.6 -0.4
-0.2 0 0.2 0.4
0.6 0.8 1 -1 -0.8
-0.6 -0.4
-0.2 0
0.2 0.4
0.6 0.8
1
0.002 0 0.004 0.006 0.008 0.01
Figure 4. Left: The umulative squared deviation
f G (κ)
of the omputed leftpassage probabilities
p(x, y)
from the valuesP κ,g (x, y)
given by (12) as a funtionof
κ
. The underlying lattie geometry is the irle as skethed in Fig. 2., theorrespondingonformalmap
g(z)
entering(12)isgiveninthetext. Theminimumisat
κ = 4.00 ± 0.01
withasquareddierenepergridpointof about2 · 10 − 5
. Right:Absolutedierenebetweenthealulatedleftpassageprobability
p(x, y )
andtheSLEexpetation
P κ,g (x, y)
(12)forκ = 4
asafuntion ofthe2dlattieoordinates(x, y)
.Forthe whole unit irle thedeviation is almost everywheresmaller than1%. Note
thatthedomainwallisxedat
( − 1, 0)
and(1, 0)
,wherethelargestdeviationsour.0 50 100 150 200 250 300 350 400 450 500
3 3.5 4 4.5 5
f G ( κ )
κ
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.02
0 50 100 150 200 250 300 350 0 50 100
150 200 250 300
350 0
0.005 0.01 0.015 0.02
Figure5. Cumulativedeviation
f G (κ)
asfuntionofκ
andabsolutedierenebetweenp(x, y)
andP κ,g (x, y)
forκ = 4
asin Fig.4 but forthe squaregeometry asdepitedin Fig.2b. Notethat the domainwallis xedat
(0, 0)
and(L, L)
, where thelargestdeviationsour.
SLE-value for given
κ
. The result is shown in Fig. 4. The minimal deviation of thedata from the expeted SLE result is at
κ = 4.00 ± 0.01
whih learly diers from thevalue
κ = 2.00 ± 0.08
that one would expet fromthe fratal dimension if the domainwalls would be desribed by SLE.
Next we varied the geometry and the domain wall onstraints and studied the
ase depited in Fig. 2b, i.e. a quadrati domain
D
with orners at0
,p
,p + ip
,ip
(p
real and positive). The funtiong : D → H : z 7→ − p(z; S)
withS =
{ 2n 1 p + i · 2n 2 p | n 1 , n 2 ∈ Z }
denes a onformal map fromD
intoH
, wherep(z; S) =
-0.02 -0.01 0 0.01 0.02
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 p( φ )-P 4 ( φ )
φ/π
R=50 R=75 R=100 R=125
Figure 6. Deviation of the left passageprobability
p(φ)
from the SLEexpetationP κ (φ)
(11) forκ = 4
in thehalf irle geometry depitedin Fig. 2 d (p = 510, q = 255, R = 500
),i.e.onefreeendofthedomainwall.1
z 2 + P
ω∈S\{0}
h 1
(z−ω) 2 − ω 1 2 i
istheWeierstraÿ
p
-funtion. Furthermoreg(p+ip) = 0, g(0) =
∞
,thusurvesinD
startingatz = 0
andendingatz = p + ip
areonformallymappedon urves in
H
starting at the origin and extending to innity. Usingg
in (12) wedetermined
κ
in the same way it was done in the ase of the unit irle. The result isshown in Fig. 5and yields
κ = 4.00 ± 0.01
.Wealsoompared
p(x, y)
diretlywithP κ=4 (x, y)
forurvesintheupperhalfplaneH
starting at the origin. For this we onsidered the geometry depited in Fig. 2d, inwhihtheurvesstartingattheoriginanendeverywhereonthehalfirle. Weheked
the predition (11) for dierent radii
R
and anglesΦ = arctan
x y
∈
− π 2 ; π 2
. The
resultisshowninFig. 6,wherethedeviationsfrom
P κ (φ)
(11)forκ = 4
are everywhereless than 2% for the size
L = 500
and radii shown in Fig. 6. We alsoobserve that thedeviationsystematiallydereaseswith thesystem size
L
forxed ratioR/L
indiatingvanishing deviations
P (Φ) − P κ=4 (Φ)
inthe limitL → ∞
.For allthe geometries that we studied we found that the data an be niely tted
with
κ = 4
,butthefrataldimensionsofthe domainwallsisunhangedbythe dierentset-ups:
d s = 1.25 ± 0.01
. The onlusion is that domain wallsin the SOS model onadisordered substrate are not desribed by SLE. The questionarises whihondition for
SLE is atually violated. The fat that for all geometries the left passage probability
we studied is well desribed by a ommon expression, Shramm's formula ontaining
theonformalmapofthehalfplanetothe speigeometryunderonsideration,would
atuallyhintatonformalinvariane. Butobviouslythisdoesnotexludethepossibility
thatabreakingofonformalinvarianemanifestsitselfinotherquantities,orevenother
geometries. We did not attempt to hek the domain Markov property, as was tried
in [24℄.
3. Exitations
In this setion we study large sale exitationsthat ost a minimum amountof energy,
alsodenoted as droplets[2℄, and the size dependene of their energy. Aording to the
usual argumentsindropletsaling theory [2℄this exitationenergy is expeted tosale
inthesame way asa domainwallof lateralsize
L
,i.e.like∆E ∼ L θ
withθ
the stinessexponent,whih forthe system we onsider is
θ = 0
. The domain wallenergy sales as∆E DW ∼ log L
.f.(6),and ifitisorretthat allexitationsofsaleL
displaythesamebehavior the question arises: howan thermalfutuations destabilize the ground state
suhthatthe presentmodelisindeedharaterizedby alineofxed points,indexed by
temperature(andthusdierentfroma
T = 0
xed pointlikeinspinglasseswithd > 2
?A
log L
saling of exitations still implies that larger and larger exitations, neessary to our to deorrelate the system's state from the ground state, need more and moreenergy and are thereforemore and more unlikely,i.e. ourring with a probability that
deays as
exp( − ∆E/T ) ∼ L −c/T
atT > 0
.The solution to this problem lies in the freedom that domain walls of exitations
of sale
L
have to optimize their energy (.f. for a similar disussion in the ontext ofoptimized disloationsin [18℄). Findingthe optimalexitation of agiven sale requires
ahighereortthansearhing adomainwallwithgiven start andendpoints(aswe have
seen above, there are even eient ways to optimize the positions of these endpoints).
Thereisasimplewaytoindueexitationsofanarbitrarysale,asrstproposedinthe
ontextof spin glasses[4℄and used again in[8℄: we ouldompute the groundstate for
xed boundaryonditionsasdepitedintheleftpanelofFig.1,thisyields
n 0
. Thenwehoose aentralsite(say
i = (x, y) = (L/2, L/2)
),xitton i = n 0 i + 1
, xthe boundaryas before and ompute the groundstate again,giving
˜ n 0
. This willdierfromn 0
onlyby aompat luster
C
that ontainsthe sitei
and whihhasn ˜ 0 i = n 0 i + 1
. This isthenin fatan optimal ordroplet exitation, but its size
V = # { i ∈ C }
an vary from 1toL 2
. Droplets of axedsize annotbe generated inthis way.The SOSmodelonadisordered substrate atuallyallowsfor aneientsearhfor
optimalexitations of agiven sale, as wedesribein the following:
A droplet exitations in the SOS model is a simply onneted luster
C
of siteswhoseheightvaluesareallinreased(oralldereased)byoneasomparedtotheground
state:
n i = n 0 i + 1 ∀ i ∈ C
, see Fig. 7. Pitorially it is, in the height representation, an extra-mountain (valley)of height+1
(− 1
). Its boundaryis adireted yle inthe duallattie,and rememberingthe mappingto theminimum-ost-owproblem(7-8),adding
a yle to a feasible solution maintains the divergene-free onstraint (8). Hene, the
transitionfromone state tothe other isdeterminedby suh ayle. Thus rst one has
to ompute the ground state of a given disorder realization. Then in order to nd a
dropletexitationof this ground state thathas a given lateral size one an for instane
fore this extra-yle to run within an annulus of inner radius
L/4
and outer radius3L/4
(i.e it's average diameter isL/2
). This an be ahieved by simply removing allsites / bonds outside this annulus and then omputing the optimal yle within this
t
a) b)
c)
Figure7. a)Skethofadropletexitationrepresentingasimplyonnetedluster
C
ofsites
i
whose height valuesare inreasedby one asompared tothe groundstate:n i = n 0 i + 1 ∀ i ∈ C
. b) Direted yleinthe duallattieorrespondingto theluster in a) representingthe step in height prole indued by raisingthe lusterC
by oneheight unit. ) Sketh of an
s − t
ut of thenodes of theoriginal lattiegraphsuhthat those sites onneted to the external node
s
are foredto be in the setS
andthoseonnetedto
t
inthesetT
(seetext).modied graph, the ost for whih depends on the ground state onguration and the
substrate heights. One assigns osts to eah direted edge that orresponds to the
energy ost for inreasing the height dierene between its left and right side by one
unit [36, 16, 37℄. Note that in pratie one may use the redued osts emerging from
the groundstate. Thatis,if and onlyif afeasible owhas minimumenergy,then there
are non-negative redued osts suh that the osts of eah yle remainsunhanged. If
the suessive shortest path algorithmis used to solve the minimum-ostow-problem
when omputing the groundstate
n 0
, then no extra work is neessary to ompute thedesired non-negative redued osts. Hene, Dijkstra's shortest path algorithm an be
used to nd the shortest direted yle around the annulus, i.e. the one separating the
inner and the outer ring ofthe annulus. To this end, the annulus is ut from the outer
to the inner ring, i.e.the orresponding edges are removed from the graph, to prevent
theshortestpathfromshort-utting. Foreahoftheseremoved edges,the shortestpath
Figure8.Examplesofoptimalexitationsofsale
L
,whoseboundary(domainwall)is fored to lie in the interior of the area indiated in green (an annulus in square
geometry). Fromthe toprowto thebottom itis
L = 32
,64
,128
and256
. Pituresaresaled tohavethesamesize.
fromitshead toitstailisomputed,where onlythethe remainingedgesare used. The
obtained shortest paths are ompleted with the orresponding direted edges to form
ylesaroundthe annulus. Thisproedurehastoberepeatedforall(orarepresentative
number of)positions ofthe annulus withinthe originallattie(inpratie one xes the
annulusandshiftsthedisorderongurations, wrappingitaroundatoroidalgeometry).
The proedure of nding the optimal yle in a given annulus an be simplied
by observing that the dropletexitation inthe heightrepresentation orresponds to an
s
-t
-ut of the underlying graph [37℄‡
in suh a way that one fores all nodes of theinner irle of the annulus to belong to
S
and all nodes belonging to the outer ring ofthe annulus to
T
, see Fig. 7 . The minimums
-t
-ut with respet to the non-negative redued edge osts of the groundstaten 0
is then exatly the boundary of the optimalexitation (or the optimal yle) one is searhing for a given annulus arrangement.
‡
Ans
-t
-utisapartition ofthenodesofagraphG
intotwodisjointsetsS
andT = G / T
,suh thats ∈ S
andt ∈ T
.0 100 200 300 400 500 600 L
3 3.5 4 4.5 5
∆Ε
10 100 1000
L 3
3.5 4 4.5 5
∆Ε
Figure9. Disorderaveragedenergyoftheoptimalexitationsofsale
L
asafuntionof
L
-leftinlinearsale,rightinalog-logsale.Aording to the famous Min-Cut-Max-Flow theorem one an ompute the minimum
s
-t
ut in polynomial time by solving the assoiated maximum-ow problem [16, 37℄.Wehave implementedthis proedure andshowfor illustrationanumberofexamples in
Fig. 8.
Fig. 9 shows our result for the disorderaveraged energy of the optimal exitations
ofsale
L
thatwe obtainwith theproeduredesribed above. Thisrepresentsanupperbound for the optimal exitations of sale
L
sine the annulus arrangement does not inludeallpossibleexitationsofsaleL
. Asoneansee thisboundsaturatesataniteenergy of order
O (1)
in the limitL → ∞
. Consequently arbitrarily large exitations exist that ost onlya smallamount of energy, whihrenders the ground state unstableatan non-vanishing temperature.
Fig. 10 shows the distributionof optimalexitation energies for dierentvalues of
L
. As an be seen the distribution is nearly Gaussian with a nite width, i.e. it does not display long, e.g. algebraially deaying, tails, whih implies that the average isrepresentative foralmost alldisorderongurations. It an alsobeseen that the whole
probability distribution
P L (∆E)
beomes independent of the length saleL
for largeL
. This is an important observation sine droplet saling theory [2℄ one would expetP L (∆E) ∼ l −θ p(∆E/L ˜ θ )
. Forθ = 0
,asitistheasehere,itisnotapriorilearwhetherthis implies
P L (∆E) ∼ (ln L) −1 p(∆E/ ˜ ln L)
,i.e.asalingwith the averagedomainwallenergy
ln L
,orP L (∆E) ∼ p(∆E) ˜
,i.e.dropletsizeindependene. Fig.10shows thatthe latter is orret.This has important impliations for the saling of the average droplet energy
within a system of lateral size
L
as determined in [20℄. There the average energy ofdroplets of size
l < L
,(∆E) min L
, i.e. without a lower bound, was estimated to behavelike
(∆E) min L ∼ ln L
. In order to make ontat with our result one should note thatthis energy is the minimum among droplets of the kind we determined here, with size
l ∈ [L/2, L]
,l ∈ [L/4, L/2]
,l ∈ [L/8, L/4]
,. . .
, i.e a minmum of approximatelyln L
randomnumbers. Onlyiftheprobabilitydistributionoftheseenergiesisa)independent
and b)identiallydistributed, theirminimum goeslikethe inverse of their number, i.e.
0 1 2 3 4 5 6 7 8 9 10
∆Ε 0
0.0005 0.001 0.0015 0.002 0.0025 0.003
p( ∆Ε)
L = 32 L = 128 L = 512
Figure10. Distributionoftheoptimalexitationenergiesofsale
L
forthree valuesof
L
. ForL ≥ 128
thisdistribution beomesindependentofL
.(∆E) min L = min { (∆E) L/2 1 , (∆E) L/2 2 , . . . , (∆E) L/2 k } ∼ 1/k
withk ∼ ln L
. Aording toour result depited in Fig. 10 the assumption b, whih is impliit in the reasoning of
[20℄, is indeedfullled.
Finally we note that also determined the fratal dimension of the boundary of
the optimalexitations and found that it is idential with the frataldimension of the
domainwallswith xed start and endpoints:
d s = 1.25 ± 0.01
.4. Disorder Chaos
In this setion, we study the sensibility of the ground state to a small hange of the
quenheddisorderonguration. Tothispurposewegenerateaongurationofrandom
osets
d 1 i
and ompute the assoiated ground stateh 1 i
. Then we slightly perturb thisongurationofrandomosets
d 2 i = d i + δǫ i
withδ ≪ 1
andwhereǫ i
'sareindependent and identially distributed Gaussian variables of unit variane and we ompute theassoiated ground state
h 2 i
. The question we ask is : how dierent are these twoongurations of the systems
h 1 i
andh 2 i
?Suh questions rst arose inthe ontext of spin glasses [1℄, where it was proposed
that disorder indued glass phases may exhibit stati haos, i.e. extreme sensitivity
to suh small modiations of external parameters (like disorder onsidered here or
temperature). Suh small perturbations are argued to deorrelate the system beyond
the so alled overlap length
L δ
whih diverges for smallδ
asL δ ∼ δ −1/α
, withα
thehaos exponent. As an example let us rst onsider the ase of an Ising spin-glass
with a ontinuous Gaussian distribution of random exhange interations, of width
J
. The system has two ground states related by a global spin reversal. Within thephenomenologialdroplettheory[1,14,41℄,alow-lyingenergyexitationsofthesystem
involves anoverturned droplet of linearsize
L
and osts anenergyJL θ
. If wenow adda small random bond perturbation, say Gaussian of width
δJ
, the exess energy ofa droplet is modied. For suh spin-glass system, this energy omes only from the
bonds whih are at the surfae of the droplet. Their ontribution is thus the sum of
L d s
independent random variables of widthδJ
, whered s
is the fratal dimension ofthe droplet : it is thus of order
± δℓ d s /2
. Therefore the ground state is unstable tothe perturbation on length sales
L
suh thatδJL d s /2 > JL θ
, i.e.L > L δ
whereL δ ∼ δ −1/α SG
withα SG = d s /2 − θ
. One thussees that, for spin glasses, disorder haosis losely relatedto the (geometrial)properties of the domain walls.
The situation is rather dierent for elasti systems in a random potential as
onsidered here. Here we onsider the SOS model on a disordered substrate dened
in Eq. (1) as
H = X
(ij)
(h i − h j ) 2 , h i = n i + d i ,
(14)with
i ≡ (x i , y i ) ∈ Z 2
. In Eq. (14) the height variablesn i
(i = 1, . . . , N
) take oninteger values
n i = 0, ± 1, ± 2, . . .
and the osetsd i
are independent quenhed random variables uniformly distributed between 0 and 1. This system (with free of periodiboundary onditions) has innitely many ground states whih dier by a global shift
∆n ∈ {± 1, ± 2, ... }
. Again, within the dropletargument [14, 42, 43℄ a low-lying energyexitationof the system involves adroplet of size
L
where the height eld isshifted byunity, say
∆n = 1
and it osts an energyL θ
. But now if we had a small perturbationd 2 i = d 1 i + δǫ i
,theexess energyofthis dropletomesfromthebulkof thedroplet,whihexperienes a random fore eld. This ontribution is thus the sum of
L d
randomvariables of width
δ
and therefore in this ase the ground state is unstable to theperturbation on length sales
L > L δ
whereL δ ∼ δ −1/α
withα = d/2 − θ
. Hered = 2
andθ = 0
and thusoneexpetsL δ ∝ δ −1
. Atvarianewithspin-glasses disussed above, one thus sees that for disordered elasti systems, disorder haos is not diretlyrelated tothe properties of domainwalls.
For the present model(14), disorderhaos wasdemonstrated analytially atnite
T
near the glass transitionT g
using a Coulomb Gas Renormalization Group [44℄. AtT = 0
,someindiationsofdisorderhaoswere alsofoundnumeriallyinRef. [16℄whereglobal orrelations between
h 1 i
andh 2 i
where studied throughχ(δ) = P
i (h 1 i − h 2 i ) 2
. Inthis paper, followingRef. [43℄, weharaterizethe loalorrelationsbetween these two
ongurations by the orrelation funtion
C ij (r)
withr ≡ (x, y)
(withi, j = 1, 2
)C ij (r) = (h i k − h i k+r )(h j k − h j k+r ) ,
(15)where
k + r ≡ (x k + x, y k + y)
and for a rotationnaly invariantsystem onsidered here one hasC ij (r) ≡ C ij (r)
withr = | r |
. In the following we will talk about intralayerorrelationsfor
C ii (r)
and interlayerorrelationsforC i6=j (r)
. Equivalently,onean also study suh orrelations (15) inFourier spae and deneS ij (q)
withq ≡ (q x , q y )
as:S ij (q) = ˆ h i q ˆ h j −q , ˆ h j q = 1 L 2
X
k
h j k e iq·k
(16)where
q · k = q x x k + q y y k
. Forarotationnalyinvariantsystem onsidered here, one hasS ij (q) ≡ S ij (q)
whereq = | q |
.Disorder haos at
T = 0
in generi disordered elasti systems in dimensiond
wasreently studied analytially using the Funtional Renormalization Group (FRG) [43℄.
At one loop order in a dimensional expansion in
d = 4 − ǫ
it was found that for shortrangedisorder(likerandombond problems)and randomperiodisystems indimension
d > 2
(inluding one omponentBragg-Glass), one has [43℄C 12 (r) = r 2ζ Φ(δr α ) with Φ(x) ∼
( c st , x ≪ 1 , x −µ , x ≫ 1 ,
(17)
with
c st
aonstant,ζ
theroughnessexponentandwhereµ
isthedeorrelationexponent.Translated intoFourier spae,this yields
S 12 (q) = L (d+2ζ) δ ϕ(qL δ ) with ϕ(x) ∼
( x −d−2ζ+µ , x ≪ 1 , x −d−2ζ , x ≫ 1 ,
(18)
with
L δ ∼ δ −1/α
and where the behavior for largex
is then suh that the dependeneon
L δ
anels inthis limit,as itshould.The two-dimensional disordered SOS model we are onsidering here (14)
orresponds preisely to the marginal ase
d = 2
(withζ = 0
and thusθ = 0
) where,as disussed in Ref. [43, 45℄, the analysis yielding the result in Eq. (17) eases to be
valid. Indeed in that ase non loal terms, irrelevant in
d > 2
are generated underoarse-grainingandthese additionaltermshavetobehandled withare at
T = 0
. Onethusonsiders theHamiltonianassoiated tothe twoopiesofthe system parametrized
by the salarelds
u i ≡ u i (r)
:H 2 copies = 1
T X
i=1,2
Z d 2 r
1
2 ( ∇ r u i ) 2 + V i (u i , r) − µ i (r) · ∇ r u i
,
(19)where
µ i ≡ (µ i x , µ i y )
are two-omponent random tilt elds, whih are generated upon renormalization. Whilethey areirrelevantind > 2
they beomerelevantind = 2
wherethey playaruial role. Thesetwoopies
u 1 , u 2
arethusindependent(19)butthey feel two mutually orrelated random potentialsV i (u, r)V j (u ′ , r ′ ) = R ij (u − u ′ )δ 2 (r − r ′ ) ,
(20)µ i ρ (r)µ j ρ ′ (r ′ ) = σ ij δ ρρ ′ δ 2 (r − r ′ ) ,
(21)where
i = 1, 2
is the index of theopy andρ = x, y
is aspatial index. This leads totherepliated Hamiltonian,
Z n = exp ( − H rep )
:H rep = 1 2T
Z d 2 r
2
X
i=1 n
X
a=1
1
2 ( ∇ r u i a ) 2 − 1 2T 2
2
X
i,j=1 n
X
a,b=1
Z
d 2 r[R ij (u i a − u j b )
(22)− 1
2 ∇ r u i ab ∇ r u j ab G ij (u i a − u j b )] ,
where weused the notation
u i ab = u i a − u i b
. In the bare model, one hasG ij (u) = σ ij
.Close to the transition
T . T g
, this model (22) was studied using Wilson RGanalysis by varying the short sale momentum uto
Λ ℓ = Λe −ℓ
, withℓ
the log-sale.It was shown, using Coulomb Gas tehnique at lowest order, that
G ii ∝ σℓ
. Thisleads to the orrelation funtion in Fourier spae
S ii (q) ∝ σℓ/q 2
, whih yields, settingℓ = log(1/q)
, tothelog 2 (r)
behaviorofthe intralayerorrelations(thisresult forC ii (r)
an be derived ina more ontrolled way using the Exat RenormalizationGroup [46℄).
Conerning haos properties, it was shown in Ref. [44℄ that
G 12 (0)
grows linearly withℓ
for smallℓ
beforeit saturates to aonstant for largeℓ
,G 12 (0) ∼ σ > ˆ 0
, whih yieldsS 12 (q) ∝ σ/q ˆ 2
. These resultslose toT g
an besummarized asS 12 (q) ∼
σ log 1/q
q 2 , q ≫ L −1 δ ˆ
σ
q 2 , q ≪ L −1 δ
(23)
whileinreal spae,the behaviorof theinterlayerorrelationfuntion
C 12 (r)
forT . T g
is thus
C 12 (r) ∼
( σ log 2 (r) , r ≪ L δ
ˆ
σ log(r) , r ≫ L δ .
(24)
At
T = 0
, it was reently shown [34℄, using FRG to one loop inluding the termG 11 (u)
thattheintralayerorrelationfuntionalsobehaveslikeC(r) ∝ log 2 (r)
,inrathergood agreement with numeris [16, 32℄. One thus also expets that
C 12 (r) ∝ log 2 (r)
for
r ≪ L δ
[16, 32, 34℄. Forr ≫ L δ
, a behavior ofC 12 (r) ∼ σ ˆ log r
as in Eq. (23)was disussed inRef. [43℄. To determine analytially whether
σ > ˆ 0
atT = 0
requiresa detailed and diult analysis of the oupled FRG equations for
R ij (u), G ij (u)
(22),whihgoesbeyondthe previousstudiesdoneinthat diretioninRef. [34,43,45℄. Here,
we willanswer this question using numerial simulations.
The purpose of our study is atually to answer the two main questions : (i) what
are the residualorrelations beyond
L δ
and in partiular isσ ˆ
alsonite atT = 0
? (ii)what is the saling form of this orrelation funtion
C 12 (r)
, i.e. the analogous of Eq.(17) from whih one an extrat the overlap length
L δ
and hek the value of haosexponent
α = 1
as expeted fromdroplet saling?Here will use the minimum-ost-ow algorithm desribed above to ompute the
two ground states
h 1 i
andh 2 i
with free boundary onditions. Instead of the orrelationfuntion
C 12 (r)
we ompute numerially the Fourier transformS 12 (q)
of the overlapbetween the ongurations (16). Our simulations have been performed on a square
lattie of linear size
L = 256
and we have hosenq = (q, 0)
withq = 2πn/L
withn = 0, 1, 2, · · · , L − 1
. The disorder average has been performed over10 6
independentsamples. In Fig. 11 (left) we show our numerial data for
S 12 (q)
. Its behavior loseto
T g
(23) suggests strongly to plotq 2 S 12 (q)
as a funtion ofq
. Given that we areworking on a disrete lattie of nite size, it is more onvenient to work with the
variable
y = sin (q/2) = [(1 − cos (q))/2] 1/2
insteadofq
(of ourse for smallq
it makesno dierene). In Fig. 11(left) we atually show a plot of
S 12 (q)[sin (q/2)] 2 /A(δ)
as afuntion of
sin (q/2)
on a log-linear plot for dierent values ofδ = 0, 0.1, 0.2, 0.3
. Onthis plot the amplitude
A(δ)
is hosen suh that the urves for dierent values ofδ
dooinide for
sin (q/2) ∼ 1
, withA(0) = 1
. Forδ = 0
, whereS 12 (q) = S 11 (q)
, this plotis almost a straight line, whih suggests indeed that for small