• Nenhum resultado encontrado

Braz. J. Phys. vol.30 número4

N/A
N/A
Protected

Academic year: 2018

Share "Braz. J. Phys. vol.30 número4"

Copied!
6
0
0

Texto

(1)

Broad Histogram: Tests for a Simple and

EÆient Miroanonial Simulator

Paulo Murilo Castro de Oliveira

Instituto deFsia,UniversidadeFederalFluminense

av. Litor^anea s/n,BoaViagem,NiteroiRJ,Brazil24210-340

e-mail PMCOIF.UFF.BR

Reeivedon5August,2000

TheBroadHistogramMethod(BHM)allowsonetodeterminetheenergydegenerayg(E),i.e. the

energyspetrumofagivensystem,fromtheknowledgeofthemiroanonialaverages<N up

(E)>

and < N dn

(E) > of twomarosopi quantities N up

and N dn

dened within the method. The

fundamentalBHMequationrelatingg(E)tothequoted averagesis exatand ompletelygeneral

foranyoneivablesystem. Thus,theonlypossiblesoureofnumerialinauraiesresidesonthe

measurementoftheaveragesthemselves. Inthistext,weintrodueaMonteCarloreipetomeasure

miroanonialaverages. Inordertotestitsperformane,weappliedittotheIsingferromagneton

a3232squarelattie. Theexatvaluesofg(E)areknownuptothislattiesize,thusitisagood

standardtoompareournumerial resultswith. Measuring thedeviations relativeto theexatly

knownvalues,we veriedadeayproportionalto 1= p

ounts , by inreasingtheounter(ounts)

ofaveragedsamplesoveratleast6deades. Thatiswhywebelievethismiroanonialsimulator

presentsnobiasbesidesthenormalstatistialutuations.Forounts10 10

,wemeasuredrelative

deviationsnear10 5

for bothg(E)andthespeiheatpeak,obtainedthroughBHMrelation.

Monte Carlo methods are applied to statistial

physisinordertomeasure thethermalaverage

<Q>

T =

P

S Q

S

exp( E

S =T)

P

S

exp( E

S =T)

(1)

of some marosopi quantity Q (magnetisation,

den-sity, et). The temperature T is xed, and the

Boltz-mannonstantisset to unity. Bothsumsrun overall

possible mirostates S available for the system. The

energy(quantity Q)orrespondingto S isdenoted by

E

S (Q

S

). Withintraditionalomputersimulations,

in-steadoftakingallpossiblestates,onetakesonlyanite

setof them,i.e. aMarkovianhain ofstatesrandomly

tossedaordingtoprobabilitiesditatedbythe

Boltz-mann exponential fators exp( E

S

=T), the so-alled

importane sampling. Thus, one needsto xa

parti-ular value for the temperature T, before running the

omputer job. In order to determine the full

depen-dene of < Q >

T

upon T, one needs to run the job

againandagain,fordierentvaluesofT.

Analternativeistore-writethesameaverageas

<Q>

T =

P

E

<Q(E)>g(E)exp( E=T)

P

E

g(E)exp( E=T)

; (2)

with energy E, and both sums run over all possible

energies. Themiroanonialaverage

<Q(E)>= P

S[E℄ Q

S

g(E)

(3)

of thesamequantity Qorrespondsto axed energy,

i.e. the sumin (3) runs onlyoverthe statesS[E℄

be-longing to energy level E. This miroanonial

aver-age(3)issimplerthantheanonialounterpart(1)or

(2), presribingexatlythe sameweightto all

averag-ingstates,i.e. itisauniformaveragingproesswithin

eahenergylevelseparately.

Only the Boltzmann fator exp( E=T) appearing

in equation (2) depends on the temperatute,

arry-ingallthermodynami informationaboutthe

environ-ment whih ontinuously exhanges energy with the

system under study. Contrary to this, bothg(E)and

<Q(E) >are independent ofthe partiular waythis

energy exhange ours, independent of the

environ-ment: theyaremorefundamentalpropertiesofthe

sys-tem alone, dened onlybyits energyspetrum. They

arenotthermodynami quantities,and donotdepend

upon thermodynamionerns liketemperature,

equi-librium, et. In pratial terms, equation (2) allows

oneto determine<Q>

T

(2)

The Broad Histogram Method (BHM) [1℄ relates

g(E) with the miroanonial averages of two

maro-sopi quantities N up

and N dn

, measured at the

ur-rent state S. First, one needs to adopt some

proto-ol of allowed movements whih ould be performed

on S, leading to another possible state S 0

. One an

adopt any suh a protool, the only restrition being

its reversibility(if S ! S 0

is allowed, sois S 0

! S).

Considering the Ising model, for instane, the

proto-ol ould be hosen to be thewhole set of single-spin

ips (among many other alternative hoies). Given

suhaprotool,N up

(S)ountsthenumberof allowed

movementswhih ouldbeperformedonS,leadingto

a xed energy inrement E (whih must be hosen

a priori, although the method is also independent of

this hoie). Analogously, N dn

(S) ountsthe number

of allowed movements dereasing the energy of S by

the same xed amount E. The fundamental BHM

relation[1℄is

g(E)<N up

(E)>= g(E+E)<N dn

(E+E)> ;

(4)

wherethemiroanonialaveragesofN up

andN dn

are

denedbyequation(3). Byknowingtheseenergy

fun-tions,equation(4)allowsoneto determineg(E)along

alltheenergyaxis,instepsofE |aonstant,

unim-portant pre-fator is anelled out by performing the

average in equation (2). Relation (4) is shown to be

valid for any energy spetrum, under ompletely

gen-eral grounds [2℄. Also, the same reasoning ould be

applied foranotherbasiquantityq, insteadofthe

en-ergy, by onsidering the degeneraies g(q) instead of

g(E). Forsimpliity, we will restrit ourselvesto the

aseoftheenergy.

Thus, theBroadHistogramMethod onsistsin: i)

to hoosesomeprotoolofallowedmovements,aswell

asanenergyjumpE;ii)tomeasure,byanymeans,

themiroanonialaverage<Q(E)>asafuntionof

the energy, as well as < N up

(E) >and <N dn

(E) >

whihdetermineg(E)throughequation(4);iii)to

ob-tainthedesiredthermalaveragethroughequation(2).

Thereisnoapproximationatall,andthenal

numeri-alauraydependsexlusivelyuponthe

miroanon-ialmeasuringstrategyadoptedinstepii. Themethod

is reviewedin [3℄. Somerefereneswhereitisused are

[4-16℄.

Let's briey analyse hereafter some possible

om-puterstrategiesoneanadoptinorderto measurethe

miroanonialaverages<Q(E)>, <N up

(E)>and

< N dn

(E) >, as funtions of the energy. A

Marko-vianhainofstatesisobtainedbyperformingrandom

These movements are tossed among some previously

denedset of possibilities, anotherprotool whih has

nothingtodowiththeBHMprotoolofvirtual

move-ments. Both protools ouldeven behosentobethe

same,but notneessarily.

Therstdiret strategyistokeepalwaysthesame

xedvalueE: bystartingfromsomestate

orrespond-ingtothedesiredenergy,onesimplyrejetsanytossed

movementwhihhangestheurrentenergy. Afterone

hasalreadyalargeenoughnumberofvisitedstates

in-sidethispartiularenergylevel,thesameproessis

re-peatedforotherlevels. Dependingontheadopted

pro-tool of (real) movements, this strategy ould lead to

ergodiityproblems: duetoitshigh rejetionrate,one

risksto sampleonly abiasedsub-setof states

belong-ing to energy level E, violating the required uniform

visitation. Moreover,it is also an ineÆient strategy,

in what onerns the omputertime, againdue to its

highrejetionrate.

Theopposedalternativeistoaeptalsomovements

leadingtoenergyjumps,storingseparatedaveragesfor

eah energy level, in parallel. Obviously, this option

is muh moreeÆient in what onerns theomputer

time. However, one annot simply aept any tossed

movement: energyinrements wouldourmoreoften

than derements, beause g(E) is normally a fast

in-reasing funtion of the energy. As a result, at the

end, only states orresponding to the region near the

maximum of g(E) would be sampled. Thus, some

movement-rejetionpresriptionmustbeadopted,and

weget bakto theuniformityviolation problem. One

an adopt some already well established

movement-rejetion presriptionbased on detailed balane

argu-ments. Forinstane, anonial,xed-temperature

dy-namis ould be adopted [4,6℄, sampling statesinside

thenarrow energy windoworrespondingto the xed

valueofT. Inordertogetresultsonabroaderenergy

window, one an simply superimpose the histograms

obtainedfordierentomputer runs orresponding to

dierentvaluesofT.

Another possibility is to adopt one of the many

multianonialdynamis[7,9,11,13,14℄. Withinthis

ap-proah, onetunestheE-dependent rejetionrate

dur-ing the omputer run, in order to obtain the same

visit probability for all energy levels, i.e. a at

dis-tribution along the energy axis, at the end. In this

ase,thevisitation probabilitytoeahpartiularstate

wouldbeproportionalto1=g(E). Theso-alled

multi-anonialmethods[17-19℄arebasedjustonthisfeature:

by reording the aeptane probability for

energy-inreasing movements E ! E 0

, aumulated during

therun andwhihmustbeequaltog(E)=g(E 0

(3)

tantglobalfatorwhihanelsinequation(2). BHM

is ompletely distint from multianonial approahes

in many features. In partiular, the infomations

ex-trated from eah E-state S belonging to the

averag-ing Markovian hain are the marosopi values of

N up

(S) and N dn

(S), not themere one-more-visit

up-grade V(E) ! V(E)+1. That is why BHM gives

moreaurateresultsthanmultianonialapproahes,

even taking into aount the same Markovian hain

of averagingstates, asshown in [15℄ where the

multi-anonialdynamis[18℄isadoptedinordertomeasure

< N up

(E) > and < N dn

(E) >. At the end, g(E) is

determined twie, byfollowingthemultianonial

tra-ditionalwayor,alternatively,bytheBHMequation(4),

bothusing data takenfrom thesameset of averaging

states. AlearaurayadvantageforBHMisreported

[15℄. Moreover,duetothemarosopiharaterofthe

BHM quantities N up

and N dn

this advantagestill

in-reasesforlargerandlargersystems.

Another ruial dierene relative to

multianoni-al approahes is that BHM requires only the

unifor-mity of visits among the states inside eah energy

level, separately, in order to get the orret

miro-anonial averages. BHM doesnot need any detailed

balanebetweenvisitstodierentlevelsEandE 0

,nor

the multianonial at distribution along the energy

axis. Any dynami strategy whih is good for

multi-anonialmethodswillbealsogood forBHM(besides

theaurayadvantagequotedin thelast paragraph),

but the reverseis nottrue. Thus, within BHM,other

not-so-restriteddynami strategiesouldbeused.

Protingfromthisfeature,wedeidedtotestavery

simpledynamistrategyinspiredbyreferene[20℄

(al-thoughwithinadierentmethod,thedynamirule

in-trodued in [20℄ is essentially the same as presented

hereafter). The idea is to avoid movement rejetions

at all,within theenergylevelurrentlybeingsampled

foraveragingpurposes. Rejetionswillberestritedto

otherenergies,whosestatesareneverinludedintothe

averagingstatistis. Let's onsiderthat themaximum

energyjump allowedby theadoptedprotoolof (real)

movementsorrespondston levelsaboveorbelowthe

urrent energy E. Then, let'stakean energy window

of 2n+1adjaentlevels, startingfrom somestate

in-sideit. Anymovementwhihkeepsthesystemstill

in-sidethiswindowwillbeaepted. Thisisthedynami

strategyweproposehere. Averages aretakenonlyfor

the entral level E inside the hosen energy window:

the systemisallowedto visit theother nlevelsabove

it, as well asthe n levels below it, nevertheless

with-outmeasuringanythingduring these side visits. Note

that no tossed movement will be rejeted, if the

ur-ompletely rejetion-free,avoidinganysystematibias

duetoartiialrejetionsrules. Thesamestrategyan

beeasilyappliedalsoforontinuousenergyspetra,by

takingaveragesonlywithinanarrow,rejetion-free

en-ergywindowenteredinsideanotherbroader,free-visit

window: anymovementtooutsidethislatterwouldbe

rejeted.

Inwhat follows,weonsideraLLsquare lattie

Ising ferromagnet, with L = 32 for whih the exat

funtion g(E)isknown[21℄. Theenergyoftheurrent

stateS isountedasthetotalnumberofitsunsatised

bonds, i.e. the total numberof neighbouringpairs of

spinspointingoneupandtheotherdown. Theenergy

spetrum orresponds to all even numbers between 0

and2;048,i.e. E=0,2,4,6,8:::2,048,whihanalso

berepresented byenergy densitiese =E=(2L 2

). The

degeneraies are g(E) = 2, 0, 2,048, 4,096, 1,057,792

:::2,respetively. Thisspetrumissymmetriin

rela-tionto itsenterat E

max

=1;024(ore=0:5),where

g(E

max

)6:310 306

. Weneedonlyitsrsthalf,

orre-spondingtopositivetemperatures. Theritialenergy

(atthethermodynamilimit)orrespondstoe

0:147

(E

300,forL=32).

We will adopt single-spin ips as the protool of

movements(bothfortherealmovementstossedduring

theomputerrun,aswellasforthevirtualoneswe

on-siderinordertoountN up

andN dn

). Startingfromthe

urrentstateSwithenergyE,1;024movementswould

beavailable,bytossingone randomspintoip. They

anbelassiedaordingtothepossibleenergyjumps,

E!EE,where E ouldbe0;2or 4. Thus, our

energy window will have5adjaent levels, theentral

one, E, where the averages will be measured, whih

isrejetion-free, plustwoneighbouringlevelsaboveit,

andtwoothersbelowit. Duringtherandomwalk

per-formed inside this window, we measure the values of

N up

and N dn

for the urrentstate, everytime its

en-ergyisE,andaumulatethemintotwoE-histograms

H up

(E)and H dn

(E). AlsoavisitounterV(E)is

up-gradedtoV(E)+1. Iftheenergyoftheurrentstateis

notE,nothingismeasuredoraumulated. Attheend,

wetaketheaverages<N up

(E)>=H up

(E)=V(E)and

< N dn

(E) >= H dn

(E)=V(E). Note that this is the

onlyroleplayedbythenalV(E): noomparisonwith

theneighbouringvaluesV(EE)isneeded,no

fur-therinformation is extrated from thesevalues. They

must onlybe largeenoughin order to provide agood

statistis. ForeahnewE, anew5-levelsenergy

win-dowenteredonitissampled,andthewholeproessis

repeated.

(4)

E)=g(E)℄for15adjaentlevelsE=300,302:::328,

at the ritial region. The deviations from the exat

valueswereaveraged(rootmeansquare)overthese 15

levels,andareshownasafuntionofthenumberof

av-eragingstatessampledinsideeahenergylevel(ounts),

in Fig. 1. Thesquares orresponds to E =4, while

the diamonds representE =2. Thedashedstraight

line (1= p

ounts) indiates that no systemati errors

besidesthenormalstatistialutuationsour,giving

redit to oursimpledynami rule. Aording tothese

results, to improvethe numerialaurayis asimple

matteroftakingmoreandmoreaveragingstatesinside

eahenergylevel,uptotheomputertimeavailable.

Figure1.Deviationsbetweenmeasuredg(E)andtheexat

values,asafuntionofthenumber(ounts)ofsampled

av-eragingstates. Thedashedlineorrespondsto1= p

ounts .

Datafor3232Isingferromagnet.

Inorder toperformthermalaverages,onedoesnot

need the same auray along the whole energy axis.

The funtion [g(E)exp( E=T)℄ 2

is displayed by the

dotted linein Fig. 2,whereT

=2:293930istheexat

loationofthe spei heat peak. This urvedisplays

the (squared) relative ontribution of eah energy to

thepartitionfuntion. Asthenumberofsampled

aver-aging statesinside eahenergylevelisproportionalto

the squarednumerialauray, thedottedline in

g-ure2showstheidealproleofvisitsoneneedsinorder

tohaveequallyaurateontributionsfromeahenergy

level. Itisasharppeakedurve,aordingtowhihthe

omputationaleortanbeonentratedonlyinsidea

narrowenergywindow. Protingfrom thisfeature, we

line,ingure2. Thepossibilityofdesigningthisprole

ofvisits,samplingdierentnumbersofaveragingstates

for dierentenergies, aordingto the relative

ontri-butionofeahenergyregion,isafurtherbigadvantage

of BHM overmultianonial methods. Almost all the

omputationaleortisonentratednear thepeak.

Figure2. Proleofvisitsalongtheenergyaxis,whihould

beshapedaordingtotheimportaneofeahenergy

on-tributiontothepartitionfuntion(dottedline).

Fig. 3showsadetailed omparisonof our

simula-tional results with the exat spei heat urve, near

its peak. Being a derivative, whih orresponds to a

mathematialproedureinwhihnumerialaurayis

stronglyompromised,thisquantityisagoodstandard

for worst-ase omparisons, moreover near its peak.

Nevertheless, the relative deviations we obtained are

ompatiblewiththenumber(ounts=810 9

)of

av-eragedstatesperenergylevel,at thepeak. Moreover,

betteryet auraieswere obtainedalong allthe

tem-peraturerangefrom0to1,alwaysbyusingthesame

simulationaldata,i.e. thesameaveragedvaluesforthe

BHMquantitiesN up

andN dn

(5)

sin-Figure3. Detailofthespeiheatpeak,aworst-ase

omparison.

Inshort, we havetested asimple dynamis whih

is veryeÆientin measuringmiroanonialaverages.

It is essentially the same dynamis as introdued in

[20℄, for other purposes. Here, the aim is to measure

themiroanonialaveragesof somepartiular

maro-sopi quantities dened within the broad histogram

method [1-3℄. One one knows these averaged

quan-tities asfuntions of the energy, the method provides

theenergydegenerayfuntiong(E)throughanexat

relation. During the same omputer run, the

miro-anonialaverage<Q(E)>ofthequantityQof

inter-est is alsomeasured. Then, one oneknowsg(E)and

<Q(E)>,theanonialthermalaverage<Q>

T an

bedeterminedbyequation(2),forany,ontinuously

varyingtemperatureT,withoutresortingagainto

further omputer simulations. Aording to our

tests,theurrentdynamiruledoesnotintrodueany

systematiaveragingbias,besidesthenormalstatistial

utuationswhihdeayproportionallyto1= p

ounts.

Thus, by applying this dynami rule to BHM, to

im-provemoreand morethenumerialaurayisa

sim-ple matter of inreasing the omputer time. Among

the further advantages of the miroanonial

dynam-is tested here, we an quote: i) its implementation

simpliity, withoutdetailed balane andother

ompli-ations; ii) no movement-rejetions at all, within the

averagingenergylevel;iii)thepossibilitytoshapethe

desired auray; iv) no randomness at all is used in

order to deideto perform ornot theurrentlytossed

movement[22℄;v)short andnon-periodiwaitingtime

betweenonseutiveaveragingstates[24℄.

I am indebted to Kim Doohul who warned me

about referene [20℄, reading whih I have had the

ideatoapplyitsdynamirulestothebroadhistogram

method. D. Stauer was kind enough to perform a

ritial reading of the manusript. This work is

par-tially supported by Brazilianagenies CAPES,CNPq

andFAPERJ.

Referenes

[1℄ P.M.C. deOliveira,T.J.P.PennaandH.J. Herrmann,

Braz. J. Phys. 26, 677 (1996) (also in Cond-Mat

9610041).

[2℄ P.M.C.deOliveira,Eur.Phys.J.B6,111(1998)(also

inCond-Mat9807354).

[3℄ P.M.C.deOliveira,Braz.J.Phys.30,195(2000)(also

inCond-Mat0003300).

[4℄ P.M.C. deOliveira,T.J.P.PennaandH.J. Herrmann,

Eur. Phys. J. B1, 205 (1998); P.M.C. de Oliveira,

inComputerSimulationStudiesinCondensed Matter

PhysisXI,169,eds.D.P.LandauandH.-B.Shuttler,

Springer,Heidelberg/Berlin(1998).

[5℄ P.M.C.deOliveira,Int.J.Mod.Phys.C9,497(1998).

[6℄ J.-S.Wang, T.K.Tay andR.H.Swendsen,Phys.Rev.

Lett.82,476(1999).

[7℄ J.-S.Wang,Eur.Phys.J.B8,287(1999)(alsoin

Cond-Mat9810017).

[8℄ J.D. Mu~noz and H.J. Herrmann, Int. J. Mod. Phys.

C10, 95 (1999); Comput. Phys. Comm. 121-122, 13

(1999).

[9℄ R.H.Swendsen,B.Diggs,J.-S.Wang,S.-T.Li,C.

Gen-oveseandJ.B. Kadane,Int.J.Mod. Phys.C10,1563

(1999)(alsoinCond-Mat9908461).

[10℄ P.M.C. de Oliveira, Comput. Phys. Comm.121-122,

16 (1999), presented at the APS/EPS Conferene on

ComputationalPhysis,Granada,Spain(1998).

[11℄ J.-S. Wang, Comput. Phys. Comm. 121-122, 22

(1999).

[12℄ A.R. de Lima, P.M.C. de Oliveira and T.J.P. Penna,

SolidStateComm.114,447(2000)(alsoinCond-Mat

9912152).

[13℄ J.-S.WangandL.W.Lee,Comput.Phys.Comm.127,

131(2000)(alsoinCond-Mat9903224).

[14℄ J.-S.Wang,Prog.Theor.Phys.Supp.138,454(2000)

(alsoinCond-Mat9909177).

[15℄ A.R.deLima,P.M.C.deOliveiraandT.J.P.Penna,J.

Stat.Phys.99,691(2000)(alsoinCond-Mat0002176).

(6)

[17℄ B.A. Berg and T. Neuhaus, Phys. Lett. B267, 249

(1991); A.P. Lyubartsev, A.A. Martsinovski, S.V.

ShevkunovandP.N.Vorontsov-Velyaminov,J.Chem.

Phys.96,1776(1992);E.Marinari andG.Parisi,

Eu-rophys. Lett.19, 451(1992); B.A. Berg, Int.J. Mod.

Phys.C4,249(1993).

[18℄ J.Lee,Phys.Rev.Lett.71,211(1993).

[19℄ B. Hesselbo and R.B. Stinhombe, Phys. Rev. Lett.

74,2151(1995).

[20℄ K.-C.Lee,J.Phys.A23,2087(1990).

[21℄ P.D.Beale,Phys.Rev.Lett.76,78(1996).

[22℄ Withinsingle-spinips,forinstane,randomnumbers

are used only to toss whih spin would be ipped, a

oarse grained deision onerning onlyinteger

num-bers. One doesnot need to ompare them with

pre-iselydenedaeptaneratios,adeliateomparison

betweenrealnumberswhihouldintrodueundesired

biasesasin

[23℄ A.M.Ferrenberg, D.P.LandauandY.J. Wong, Phys.

Rev.Lett.69,3382(1992).

[24℄ Inthepresentaseofa3232lattie,anewstatewas

sampledafterevery40spinips,inaverage,instead

ofL 2

Imagem

Figure 2. Prole of visits along the energy axis, whih ould
Figure 3. Detail of the spei heat peak, a worst-ase

Referências

Documentos relacionados

ritial behavior is haraterized at high magneti onentrations..

body transverse Ising model with random bonds and elds. The ouplings and the magneti elds are

transition in suh a model, and study the orresponding ritial temperature as the spin value ande. the

the master equation on the phase spae, and the evolution of loal and global spin autoorrelation.. funtions, in terms of independent relaxation modes, whih are eigenvetors of the

We propose an improvement of a Monte Carlo method designed to treat the Ising model in a

universality lass at the ritial point marking

We alulate the spatial and temperature dependenes of various spin orre-.. lation funtions, as well as the wave-vetor dependene of the spin suseptibility for

relations has been applied to obtain the time evolution of a non-ommuting spin