HAL Id: inria-00282107
https://hal.inria.fr/inria-00282107v2
Submitted on 28 May 2008
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of
sci-entific research documents, whether they are
pub-L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diffusion de documents
scientifiques de niveau recherche, publiés ou non,
models: mathematical analysis and computational
approaches
Xavier Blanc, Claude Le Bris, Frédéric Legoll, Carsten Patz
To cite this version:
Xavier Blanc, Claude Le Bris, Frédéric Legoll, Carsten Patz. Finite-temperature coarse-graining of
one-dimensional models: mathematical analysis and computational approaches. Journal of Nonlinear
Science, Springer Verlag, 2010, 20 (2), pp.241-275. �10.1007/s00332-009-9057-y�. �inria-00282107v2�
a p p o r t
d e r e c h e r c h e
Thème NUM
Finite-temperature coarse-graining of
one-dimensional models: mathematical analysis and
computational approaches
Xavier Blanc — Claude Le Bris — Frédéric Legoll — Carsten Patz
N° 6544
approa hes XavierBlan
∗†
, Claude LeBris‡†
, Frédéri Legoll§†
, Carsten Patz¶
ThèmeNUMSystèmesnumériques
ProjetMICMAC
Rapportdere her he n°6544May200840pages
Abstra t: Wepresentapossibleapproa hforthe omputationoffreeenergiesand
ensem-ble averages ofone-dimensional oarse-grainedmodels in materialss ien e. The approa h
isbaseduponathermodynami limitpro ess,andmakesuseofergodi theoremsandlarge
deviationstheory. Inadditiontoprovidingapossiblee ient omputationalstrategyfor
en-sembleaverages,theapproa hallowsforassessingthea ura yofapproximations ommonly
usedinpra ti e.
Key-words: ensemble averages, free energies, oarse-grainedmodels, materials s ien e,
thermodynami limit, omputationalstrategy
∗
Laboratoire J.-L.Lions, UniversitéPierre etMarie Curie,Boîte ourrier187,75252 ParisCedex 05,
Fran e. Conta t: blan ann.jussieu.fr
†
INRIA Ro quen ourt, MICMAC team-proje t, Domaine de Volu eau, B.P.105, 78153 LeChesnay
Cedex,Fran e.
‡
CERMICS,E oleNationaledesPontsetChaussées,6et8avenueBlaisePas al,CitéDes artes,77455 Marne-la-ValléeCedex2,Fran e. Conta t: lebris ermi s.enp .fr
§
LAMI, E oleNationale desPonts et Chaussées, 6 et 8 avenueBlaise Pas al, Cité Des artes, 77455 Marne-la-ValléeCedex2,Fran e. Conta t: legolllami.enp .fr
¶
Weierstrass-Institut fürAngewandte AnalysisundSto hastik, Mohrenstrasse 39, 10117 Berlin, Ger-many.Conta t: patzwias-berlin.de
mathématique et appro hes numériques
Résumé: Nousprésentonsune appro hepossiblepour le al ulde moyennesd'ensemble
et d'énergies libres de modèles réduits en s ien e des matériaux, dans un adre
mono-dimensionnel. L'appro he s'appuie sur une limite thermodynamique, et utilise des
théo-rèmesergodiqueset lathéorie desgrandes déviations. Nous obtenons ainsinon seulement
unestratégienumériquee a epourle al uldemoyennesthermodynamiques,maisaussi
unmoyendevériernumériquementlavaliditédeshypothèsesquisont ourammentfaites
danslalittératurepourdévelopperdetellesappro hes.
Mots- lés: moyennesd'ensemble,énergieslibres,modèlesréduits,s ien edesmatériaux,
1 Introdu tion
Computing anoni alaveragesis astandardtaskof omputationalmaterialss ien e.
Con-sider an atomisti , supposedly large, system onsisting of
N
parti les, at positionsu =
u
1
, . . . , u
N
∈ R
3N
. Providethissystemwithanenergy
E
µ
(u) = E
µ
u
1
, . . . , u
N
.
(1) Aprototypi alexampleofsu hanenergyisthepairintera tionenergyE
µ
(u) =
1
2
X
i6=j
W u
j
− u
i
.
(2)Thenitetemperaturethermodynami alpropertiesofthematerialareobtainedfrom
anon-i alensembleaverages,
hAi =
Z
Ω
N
A(u) exp(−βE
µ
(u)) du
Z
Ω
N
exp(−βE
µ
(u)) du
,
(3) whereΩ ⊂ R
3
is thema ros opi domainwhere thepositions
u
i
vary,
A
is theobservable of interest, andβ = 1/(k
B
T )
is the inverse temperature [15℄. The denominator of (3) is denoted byZ
and alled thepartition fun tion. Themajor omputationaldi ultyin (3) isof oursetheN
-foldintegrals,whereN
, thenumberofparti les, isextremelylarge. For integralsof thetype (3) to be quantitatively meaningful in pra ti e,N
does not need to approa htheAvogadronumber,but stillneedstobeextremelylarge(10
5
, say).Thethreedominant omputationalapproa hesfortheevaluationof(3)areMonteCarlo
methods, Markov hains methods, and Mole ulardynami s methods respe tively(see e.g.
[13℄forareviewandamathemati aland omputational omparison). Inthepresentarti le,
weusethelattertypeofmethods, andmorepre iselytheoverdampedLangevindynami s
(also alled biasedrandowwalk). The ensemble average(3) is al ulatedasthe long-time
average
hAi = lim
T→+∞
1
T
Z
T
0
A(u(t)) dt
(4)alongthetraje torygeneratedbythesto hasti dierentialequation
du = −∇
u
E
µ
(u) dt +
p
2/β dB
t
.
(5)It isoften the asethat theobservable
A
a tuallydoesnot depend onthepositionsu
i
ofalltheatoms,but onlyonsomeofthem. Think forinstan eofnanoindentation: weareespe iallyinterestedin thepositions oftheatomsbelowtheindenter, in thefor es applied
ontheseatoms,...Ourrstaim istodesignanumeri almethod thate iently omputes
TheQuasiContinuum Method (QCM)isa ommonlyusedexample ofapproa hesthat
allow for the al ulation of ensemble averages. In its original version, the method was
fo usedon thezerotemperaturesetting. It wasoriginallyintrodu edin [44, 45℄, andthen
further developed in [27, 34, 35, 41, 42, 43℄. It has been studied mathemati ally in e.g.
[1, 2,3,7, 8,9,18,20,21,22, 30, 31,37℄. See[10℄forare entreview. Anextensionofthe
originalideahasre entlybeendevelopedin[19℄and arriesthroughtothenite-temperature
ase, onsideredin thepresentarti le. Seealso[14,29℄forpriorstudiesdevelopingideasin
thesamevein.
Let us briey detail the bottom line of oarse-grainingstrategies for the omputation
of anoni alaverages. For simpli ityof exposition,we lettheatomsvary in
Ω = R
3
. The
ideaistosubdivide theparti lesofthesystemintotwosubsets. Therstsubset onsistsof
theso- alledrepresentativeatoms (abbreviatedintheQCMterminologyasrepatoms,with
positionshen eforthdenotedby
u
r
). These ondsubsetisthatofatomsthatareeliminated in the oarse-grainedpro edure. Theirpositions aredenoted herebyu
c
. Weassume that theobservable onsideredonlydependsonthepositionsu
r
oftherepatoms,notonthoseof theotheratoms,u
c
. Morepre isely,onewritesu = (u
1
, . . . , u
N
) = (u
r
, u
c
),
u
r
∈ R
3N
r
,
u
c
∈ R
3N
c
,
N = N
r
+ N
c
.
Ouraimisto ompute(3)forsu hobservable,thatis,hAi = Z
−1
Z
R
3N
A(u
r
) exp(−βE
µ
(u)) du.
(6)Weobservethat,owingtoourassumptionon
A
,Z
R
3N
A(u
r
) exp(−βE
µ
(u)) du =
Z
R
3Nr
du
r
A(u
r
)
Z
R
3Nc
exp(−βE
µ
(u
r
, u
c
)) du
c
,
andlikewiseZ =
Z
R
3N
exp(−βE
µ
(u)) du =
Z
R
3Nr
du
r
Z
R
3Nc
exp(−βE
µ
(u
r
, u
c
)) du
c
.
Introdu ingthe oarse-grainedpotential(also alledfreeenergy)
E
CG
(u
r
) := −
1
β
ln
Z
R
3Nc
exp(−βE
µ
(u
r
, u
c
)) du
c
,
(7)theexpression(6)rewrites
hAi = Z
−1
r
Z
R
3
Nr
A(u
r
) exp(−βE
CG
(u
r
)) du
r
,
(8) whereof ourseZ
r
=
Z
R
3
Nr
exp(−βE
CG
(u
r
)) du
r
.
(9)Under appropriate onditionsensuringergodi ityof thesystem,the integral(8)is in turn omputedfrom
hAi = lim
T
→+∞
1
T
Z
T
0
A(u
r
(t)) dt
withdu
r
= −∇
u
r
E
CG
(u
r
) dt +
p
2/β dB
t
.
(10)Simulatingthedynami sin (10)isaless omputationallydemanding taskthansimulating
(5), owingto the redu eddimension
N
r
. Thissimpli ation omes at apri e: al ulating the oarse-grainedfreeenergy(7).Remark1 The present work on entrates on the omputation of ensemble averages
us-ing oarse-grainedmodels, and free energies. Pra ti e shows that the same oarse-graining
paradigmisusedtosimulate a tual oarse-graineddynami s atnite temperature. Wewill
notgo inthis dire tion asthe physi al relevan eof thelatterapproa h isun lear tous.
Inordertoapproximatethefreeenergy(7),state-of-the-artnitetemperaturemethods
perform a Taylor expansion of the position of the eliminated atoms
u
c
. In this Taylor expansion, a linear interpolation and a harmoni approximation of the positions of theatomsaresu essivelyperformed. Morepre isely,giventhepositions
u
r
oftherepatoms,the averagepositionsu
c
(u
r
)
oftheeliminatedatomsarerstdeterminedbylinearinterpolation betweentwo(ormore)adja entrepatoms. Thenitispostulatedthatu
c
= u
c
(u
r
) + ξ
c
wheretheperturbation
ξ
c
issmall. Theenergyisthen al ulatedfrom aTaylorexpansion trun atedatse ondorder:E
µ
(u
r
, u
c
) =
E
µ
(u
r
, u
c
(u
r
) + ξ
c
)
≈ E
µ
(u
r
, u
c
(u
r
)) +
∂E
µ
∂u
c
(u
r
, u
c
(u
r
)) · ξ
c
+
1
2
ξ
c
·
∂
2
E
µ
∂u
2
c
· ξ
c
=: e
E(u
r
, u
c
).
It follows (we skip the details of the argument and refer to the bibliography pointed outaboveforfurtherdetails) that
E
CG
(u
r
)
isapproximatedbyE
QCM
(u
r
) = −
1
β
ln
Z
R
3
Nc
exp(−β e
E(u
r
, u
c
)) du
c
,
(11)whi hisanalyti ally omputable. Withoutsu hsimplifyingassumptions,thea tual
ompu-tationof
E
CG
forpra ti alvaluesofN
r
andN
c
seemsundoable. Theapproa hhasproven e ient. Reportedly,itsatisfa torilytreatsthree-dimensionalproblems oflargesize.How-ever, from the mathemati al standpoint, it is an open question to evaluatethe impa t of
theabove oupleofapproximations(linearinterpolationofthe average positions, followed
byharmoni expansion). Thepurposeofthepresentarti leistopresentanapproa hthat,
Our approa h is based ona thermodynami limit. It wasrst outlined in [38℄ for the
spe ial aseofharmoni intera tions. Theapproa hisexa tinthelimitofaninnite
num-berof eliminatedatoms, andtherefore validwhenthis number
N
c
is largeas ompared to thenumberN
r
of representativeatoms thatare kept expli itin the oarse-grainedmodel. Thisregime, afterall, is theregime that allee tive oarse-grainingstrategiesshouldtar-get, although,in pra ti e,
N
c
isseeminglymu h smallerthanideally, andeven sometimes ofthe sameorderof magnitudeasN
r
. Inshort,the onsiderationof theasymptoti limitN
c
→ +∞
makestra tablea omputationwhi his nottra tableforniteN
c
(unless sim-pli ations, asthosementioned above,are performed). Wedo not laim for originality inourtheoreti al onsiderations on thethermodynami limit of the freeenergy of atomisti
systems. We providethem herefor onsisten y. However,ourspe i useofsu h
theoret-i al onsiderationsas a omputationalstrategyfor approximating oarse-grainedensemble
averagesin omputationalmaterials s ien eseems,to thebest ofourknowledge,new. We
werenotabletondany omparableendeavourintheexistingliteraturewehavea essto.
Let us on ludethis introdu tionby brieydes ribing ourapproa h. Assume for
sim-pli itythatthereisonlyonerepatom:
N
r
= 1
. Ourideato omputehAi
in(6)isto hange variables,thatisintrodu ey = (y
1
, . . . , y
N
) = Φ(u)
,andre ast(6) ashAi =
Z
R
3N
A
1
N
N
X
i=1
y
i
!
ν(y) dy
for some probability density
ν(y)
(see equation (17) below for an expli it example). We next re ognizehAi
as the expe tation valueE
"
A
1
N
N
X
i=1
Y
i
!#
,
whereY = (Y
1
, . . . , Y
N
)
arerandomvariablesdistributed a ordingtotheprobabilityν
. A LawofLargeNumbers providesthelimitofhAi
whenN → +∞
(whi h orrespondstoN
c
→ +∞
, sin eN
r
= 1
). Therateof onvergen emayalsobeevaluatedusingtheCentralLimitTheorem.The above approa h bypasses the al ulation of the free energy
E
CG
to ompute the ensembleaverage(8). ButitisalsointerestingtotryandevaluateE
CG
inthesameregime, in order to, again, both provide ane ientnumeri al approa h and assessthe validity ofommonlyused simplifyingapproa hes. First,itisto beremarkedthat
E
CG
s aleslinearly withthenumberN
c
ofeliminatedatoms. Therelevantquantityishen ethefreeenergyper parti leF
∞
(u
r
) :=
lim
N
c
→+∞
1
N
c
E
CG
(u
r
).
(12) Thisenergyisrelatedtothe oarse-grained onstitutivelawofthematerialatnitetemper-ature(see[17,36℄forrelatedapproa hes). Evenif
F
∞
isagoodapproximationofE
CG
/N
c
forlargeN
c
,it anbeseenthatN
c
F
∞
isnotne essarilyagoodapproximationofE
CG
. It isnot leartoushowtousetheprobabilitymeasureZ
−1
N
c
exp(−βN
c
F
∞
)
to omputeinan
e ientmanneranapproximationoftheaverage
hAi
(seeRemark 6below).Wedevelopourapproa h in theone-dimensionalsetting, forsimple asesof pair
omputational strategy to approximate ensemble averages (see Se tion 2.1), and wenext
addressthe omputations of free energies (see Se tion 2.2). Numeri al onsiderationsare
olle tedin Se tion2.3.
We next turn to next-to-nearest neighbour intera tions, traditionally abbreviated as
NNN.For thismodel, wefo uson the omputation ofensembleaverages(see Se tion3.1).
AsexplainedinSe tion3.2,more ompli atedtypesofintera tionpotentialsand"essentially
one-dimensionalsystems"(in ludingpolymer hains)maybetreatedlikewise,althoughwe
donotpursuein thisdire tion.
Asimilarinterpretationofensembleaveragesastheonepresentedhere,usingaMarkov
hain formalism, shouldleadto an analogous strategyfor two-dimensional systems. Some
preliminarydevelopments,notin ludedinthepresentarti le,already onrmthis. However,
denite on lusionsareyettobeobtained,bothontheformalvalidityoftheapproa hand
onthebest possiblenumeri ale ien y a omplished. The fa tthat the two-dimensional
aseismu hmoredi ultthantheone-dimensional aseis orroboratedbytheliterature
onthissubje t: onlyverysimple ases,su hasspinsystems,orharmoni intera tions(with
zeroequilibriumlength)areknowntohaveexpli itsolutionsinthis ontext(seethereviews
[5, 39℄). Wetherefore prefer to postpone onsiderations onthe two-dimensional situation
untilafuturepubli ation[11℄.
2 The nearest neighbour (NN) ase
Asmentionedintheintrodu tion,ourapproa hisbasedontheasymptoti limit
N −→ +∞
. Wethereforerstres aletheproblemwiththeinteratomi distan eh
,su hthatN h = L = 1
. Theatomisti energy(2)in theres aled NN asewritesE
µ
u
1
, . . . , u
N
=
N
X
i=1
W
u
i
− u
i−1
h
.
(13) Wenowimposeu
0
= 0
to avoidtranslationinvarian e,and onsiderthat onlyatoms0and
N
are repatoms, while allthe other atomsi = 1, . . . , N − 1
are eliminated in the oarse-grainingpro edure(seeFigure1). Ourargument anbestraightforwardlyadaptedtotreatthe aseof
N
r
≫ 2
repatoms(seeFigure2). SeetheendofSe tion2.1forthis.L
0
N
Figure 1: We isolate a segment between two onse utive repatoms (in red). All atoms
0
N
Figure 2: The repatoms (in red) are expli itly treated, all other atoms (in blue) being
eliminatedintheCGpro edure.
Inthissimplesituation,theaverage(6)reads
hAi
N
= Z
−1
Z
R
N
A u
N
exp
−β
N
X
i=1
W
u
i
− u
i−1
h
!
du
1
. . . du
N
,
(14)where wehaveexpli itly mentionedthedependen e of
hAi
uponN
using asubs ript. We introdu ey
i
:=
u
i
− u
i−1
h
, i = 1, . . . , N,
(15) andnextremark thatu
N
= h
N
X
i=1
y
i
=
1
N
N
X
i=1
y
i
.
(16)Theaveragerewritesas
hAi
N
= Z
−1
Z
R
N
A
1
N
N
X
i=1
y
i
!
exp
−β
N
X
i=1
W (y
i
)
!
dy
1
. . . dy
N
,
(17) wherenowZ =
Z
R
N
exp
−β
N
X
i=1
W (y
i
)
!
dy
1
. . . dy
N
.Remark2 In(14),weletthevariables
u
i
varyonthewholereal line. Wedonot onstrain
them to obey
u
i−1
≤ u
i
, whi h en odes the fa t that nearest neighbours remain nearest
neighbours. The argumentprovided here and below arries through when this onstraintis
a ountedfor,basi allyrepla ing the intera tion potential
W
byW
c
(y) =
W (y)
wheny ≥ 0
+∞
otherwise.(18)
Likewise, we ould also imposethat all the
u
i
stayin agiven ma ros opi segment. If they
areorderedin reasingly,itsu estoimposethis onstrainton
u
0
and
u
N
. Thisisagaina
2.1 Limit of the average
Itisevidentontheexpression(17)that
hAi
N
= E
"
A
1
N
N
X
i=1
Y
i
!#
forindependentidenti allydistributed(i.i.d.) randomvariables
Y
i
,sharingthe ommonlawz
−1
exp(−βW (y))dy
,with
z =
Z
R
exp(−βW (y))dy
. Asimple omputationthus gives Theorem1 AssumethatA : R −→ R
is ontinuous, that for somep ≥ 1,
there existsa onstantC > 0
su hthat∀y ∈ R,
|A(y)| ≤ C(1 + |y|
p
),
(19)
andthat
Z
R
(1 + |y|
p
) exp (−βW (y)) dy < +∞.
(20) Then,
lim
N
→+∞
hAi
N
= A(y
∗
),
wherey
∗
:= z
−1
Z
R
y exp(−βW (y)) dy,
(21) withz =
Z
R
exp(−βW (y)) dy
. Inaddition, if
A
isC
2
andif (19)-(20)hold with
p = 2
,thenhAi
N
= hAi
approx,1
N
+ o
1
N
,
(22)with
hAi
approx,1
N
:= A(y
∗
) +
σ
2
2N
A
′′
(y
∗
),
whereσ
2
= z
−1
Z
R
(y − y
∗
)
2
exp(−βW (y)) dy.
(23) Theproofof(21)isadire tappli ationoftheLawofLargeNumbers,andthatof(22)isanappli ationoftheCentralLimitTheorem. Weskipthem. Thefollowing onsiderations,
formoreregularobservables
A
,indeed ontaintheingredientsforproving(21)-(22),simply bytrun atingtheexpansionatrstorder.Note that if
A
is more regularthan stated in Theorem 1,then it is of ourse possible topro eedfurther intheexpansionofhAi
N
in powersof1/N
. Indeed,assumeforinstan ethat
A
isC
6
,that
A
(6)
isgloballybounded andthat(19)-(20)holdwith
p = 6
. ThenA
1
N
N
X
i=1
Y
i
!
=
A
y
∗
+
1
N
N
X
i=1
D
i
!
,
=
A(y
∗
) + A
′
(y
∗
)
1
N
N
X
i=1
D
i
+
1
2
A
′′
(y
∗
)
1
N
N
X
i=1
D
i
!
2
+
1
6
A
(3)
(y
∗
)
1
N
N
X
i=1
D
i
!
3
+
1
24
A
(4)
(y
∗
)
1
N
N
X
i=1
D
i
!
4
+
1
5!
A
(5)
(y
∗
)
1
N
N
X
i=1
D
i
!
5
+
1
6!
A
(6)
(ξ)
1
N
N
X
i=1
D
i
!
6
,
whereD
i
= Y
i
− y
∗
and
ξ
liesbetweeny
∗
and
(1/N )
P
Y
i
. Wenowtakethe expe tation valueofthis equality. Letusintrodu ehAi
approx,2
N
:=
A(y
∗
) +
1
2
A
′′
(y
∗
)
1
N
E
D
2
1
+
1
6
A
(3)
(y
∗
)
1
N
2
E
D
3
1
+
1
24
A
(4)
(y
∗
)
1
N
3
E
D
4
1
+
N − 1
N
3
E
(D
2
1
)
2
+
1
5!
A
(5)
(y
∗
)
1
N
4
E
D
5
1
+
N − 1
N
4
E
(D
2
1
)E(D
3
1
)
.
(24) ThenhAi
N
− hAi
approx,2
N
≤
6!
1
kA
(6)
k
L
∞
E
1
N
N
X
i=1
D
i
!
6
.
(25) We now use the fa t that any i.i.d. variablesD
i
with mean value0
satisfy the following bounds:∀p ∈ N,
∃C
p
> 0,
E
"
1
N
N
X
i=1
D
i
!
p
#
≤
C
p
N
p
2
ifp
iseven;C
p
N
p+1
2
ifp
isodd. (26)This is proved by developping the power
p
of the sum, and then using the fa t that the variablesarei.i.dandhavemeanvaluezero. Wehen einferfrom(24),(25)and(26)thathAi
N
= A(y
∗
) +
σ
2
2N
A
′′
(y
∗
) +
1
N
2
a
3
6
A
(3)
(y
∗
) +
σ
4
24
A
(4)
(y
∗
)
+ O
1
N
3
,
(27) whereσ
isdened by(23)anda
3
= z
−1
Z
R
More generally, it is possible to expand
hAi
N
at any order in1/N
, provided thatA
is su ientlysmoothandexp(−βW )
su ientlysmallatinnity. Inviewofthebounds(26), we anseethatusingaTaylorexpansionoforder2p
aroundy
∗
for
A
givesanexpansionofhAi
N
oforderp
.The pra ti al onsequen eof Theorem 1is that, for omputational purposes, we may
taketheapproximation
hAi
N
≈ A
z
−1
Z
R
u
N
exp(−βW (u
N
)) du
N
.
(28)Aspointedoutabove,itispossibletoimprovethisapproximationifne essarybyexpanding
furtherin powersof
1/N
.We on ludethisse tionbyshowingthatour onsiderationofasingle"segment" arries
throughto the ase when there are 3repatoms, of respe tiveindex
0, M
1
andM
1
+ M
2
, withM
1
h = L
1
,M
2
h = L
2
,N h = L = 1
(seeFigure3). Theaverageto omputewriteshAi
N
= Z
−1
Z
R
N
A u
M
1
, u
M
1
+M
2
exp
−β
N
X
i=1
W
u
i
− u
i−1
h
!
du
1
. . . du
N
.
Intheregime
h → 0
,N, M
1
, M
2
→ +∞
withM
1
/N
andM
2
/N
xed,wehave,usingsimilar arguments,lim
N
→+∞
hAi
N
= A(L
1
y
∗
, L
2
y
∗
).
Thegeneralizationto
N
r
> 3
repatoms,intheappropriateasymptoti regime,easilyfollows.L1
L2
0
M1
M1+M2
N
Figure 3: When onsidering two onse utive segments -or more-, the argument may be
readilyadapted. Seethetext.
Remark3(The small temperature limit) It is interesting here to onsider the small
temperaturelimit ofthe above expansion, thatis, the limit
β → +∞.
In su ha ase, using the Lapla e method (see [6 ℄), it is possible to ompute the limit of the various terms thatappearinthe expansionof
hAi
N
(see(27)). Wegive asanexamplethe rstandthese ond terms:A(y
∗
) = A(a) + O
1
β
,
σ
2
2
A
′′
(y
∗
) =
1
β
A
′′
(a)
W
′′
(a)
+ O
1
β
2
,
where
a
isthepointwhereW
attainsitsminimum(inthisremark,weassumeforsimpli ity thatW
attains itsminimum atauniquepoint). Itispossibleto re overthese termsby ex-pandingtheenergyE
µ
aroundtheequilibrium onguration orresponding toy
i
= a.
Indeed, if we assume thatW (y) = W
′′
(a) (y − a)
2
/2
in (14), then a simple expli it omputation
gives
hAi
N
= A(a) +
1
2N βW
′′
(a)
A
′′
(a) + O
1
β
2
N
2
.
Hen e, expanding the rst termsof (23) in powers of
1/β
for largeβ
gives an expansion thatagrees withthatobtainedusingaharmoni approximation ofthe energy. Thisprovidesaquantitative evaluationof the latterapproa hin thisasymptoti regime.
2.2 Limit of the free energy
Wenowlookforamoredemanding result. For larity,letus returntothegeneral
oarse-grainedaverage(8),whi hof ourseequals(14)and(17)inoursimpleNN ase. Insteadof
sear hingforthelimitoftheaverage
hAi
forlargeN
c
,wenowlook forthelimitofthefree energyperparti le(see(7)and(12)).In the present se tion,
u
r
is in fa t equal tou
N
(the right end atom) sin e
u
0
= 0
,
although arepatom, isxed to avoid translationinvarian e. Thuswewish toidentify the
behaviourfor
N
largeofE
CG
u
N
= −
1
β
ln
Z
R
N −1
exp −βE
µ
u
1
, . . . , u
N
du
1
. . . u
N−1
.
(29)Note that
E
CG
is thefree energy orresponding to integratingoutN − 1
variables. From Thermodynami s,itisexpe tedthatE
CG
s aleslinearlywithN
. Thisis onrmedbythe onsiderationofanharmoni potentialW (x) =
k
2
(x − a)
2
,forwhi hE
CG
u
N
=
kN
2
u
N
− a
2
+ C(N, β, k),
whereC(N, β, k) =
1
β
N −
1
2
ln N −
N − 1
2β
ln
2π
βk
does not depend on
u
N
(see the detailsin [38℄). Therefore,weintrodu ethefreeenergyperparti leF
N
(x) :=
1
N
E
CG
(x),
sothathAi
N
= Z
r
−1
Z
R
A u
N
exp −βNF
N
u
N
du
N
.
(30)Thelimitbehaviourof
F
N
isprovidedbytheLargeDeviationsPrin iple. This laimis madepre iseinthefollowingtheorem.Theorem2 Assumethepotential
W
satises∀ξ ∈ R,
Z
R
exp (ξy − βW (y)) dy < +∞,
(31) andexp(−βW ) ∈ H
1
(R \ {0}).
Then the limit behaviour of
F
N
is given by the following Legendretransform:lim
N→+∞
F
N
(x) +
1
β
ln
z
N
= F
∞
(x)
(32) withF
∞
(x) :=
1
β
sup
ξ
ξx − ln
z
−1
Z
R
exp(ξy − βW (y)) dy
(33) andz =
Z
R
exp(−βW (y)) dy
.Remark4 The assumption
exp(−βW ) ∈ H
1
(R \ {0})
allows for
W
to be pie ewise on-tinuous, with dis ontinuity at the origin. This in parti ular allows to deal with the type ofpotentialsmentionedin Remark2.
Proof: Letusrstrewritethefreeenergy
F
N
(x)
asfollows:F
N
(x)
= −
1
βN
ln
"Z
R
N −1
exp
−β
N
X
−1
i=1
W
u
i
− u
i−1
h
−βW
x − u
N
−1
h
!
du
1
. . . du
N
−1
#
= −
N − 1
βN
ln h −
βN
1
ln
"Z
R
N −1
exp
−βW
N x −
N
X
−1
i=1
y
i
!
−β
N
X
−1
i=1
W (y
i
)
!
dy
1
. . . dy
N
−1
#
= −
1
β
ln h −
1
β
ln z −
1
βN
ln µ
N
(x),
where
µ
N
is the law of the random variable(1/N )
P
N
i=1
Y
i
, andY
i
is a sequen e of i.i.d. randomvariableswithlawµ = z
−1
exp(−βW (y))
. A tually,wehave
µ
N
(x) = N µ
∗N
(N x) ,
whereµ
∗N
denotes the
(N − 1)
-fold onvolutionprodu tofµ
(µ
∗2
= µ ∗ µ
).Thesequen eof measures
µ
N
satisesalargedeviations property(seeforinstan e [23,1
ln µ
.
lowerbound,whi hisasimple onsequen eoftheresultsof[46℄. Theupperboundismore
involved: weneedto reprodu ethe orrespondingproofof[46℄,andusearenedversionof
theCentralLimitTheorem[32℄.
Weintrodu ethefun tion
G
N
(x) = −
1
βN
ln µ
N
(x),
(34) whi h satises,inviewoftheabove omputation,F
N
(x) = −
1
β
ln
z
N
+ G
N
(x).
(35) Firststep: lower bound. Wewriteµ
N
+1
(x) = (N + 1)
Z
R
µ(N (x − t) + x) µ
N
(t) dt.
(36) Letus deneJ
N
(t) = −
1
N
ln µ (N (x − t) + x) .
Thisfun tion learlysatisesthefollowing onvergen e:
lim inf
u→t,N →+∞
J
N
(u) = J
∞
(t) :=
(
+∞
ift 6= x,
0
ift = x.
Hen e,wemayapplyTheorem2.3of[46℄,whi himpliesthatlim inf
N
→+∞
−
1
N
ln
Z
R
exp (−NJ
N
(t)) µ
N
(t)dt
≥ inf
t∈R
(J
∞
(t) + βF
∞
(t)) = βF
∞
(x).
(37)Sin etheleft-handsideof(37)is equalto
β(N + 1)
N
G
N+1
(x) +
ln(N + 1)
N
,
weinferlim inf
N
→+∞
G
N
(x) ≥ F
∞
(x).
(38)Se ondstep: upper bound. Wenowaim at bounding
G
N
from above. Were all that the fun tionwemaximizein (33)is on ave,sothereexistsauniqueξ
x
∈ R
su hthatF
∞
(x) =
1
β
ξ
x
x − ln
z
−1
Z
R
exp (ξ
x
y − βW (y)) dy
.
TheEuler-Lagrangeequationofthemaximizationproblemalsoimplies
x =
Z
R
y exp (ξ
x
y − βW (y)) dy
Z
exp (ξ
x
y − βW (y)) dy
.
(39)Weintrodu ethenotations
˜
µ(t) =
Z
exp(ξ
x
t − βW (t))
R
exp(ξ
x
t − βW (t)) dt
andM (ξ) = z
−1
Z
R
exp(ξt − βW (t)) dt,
and omputeµ
N
(x)
= N
Z
R
N −1
µ
N x −
N−1
X
i=1
y
i
!
µ(y
1
) . . . µ(y
N−1
) dy
1
. . . dy
N−1
= N M (ξ
x
)
N
−1
Z
R
N −1
µ
N x −
N
X
−1
i=1
y
i
!
exp
−ξ
x
N
X
−1
i=1
y
i
!
ט
µ(y
1
) . . . ˜
µ(y
N−1
) dy
1
. . . dy
N
−1
≥ N M(ξ
x
)
N
−1
Z
|N x−P y
i
|≤δ
µ
N x −
N
X
−1
i=1
y
i
!
× exp
−ξ
x
N
X
−1
i=1
y
i
!
˜
µ(y
1
) . . . ˜
µ(y
N
−1
) dy
1
. . . dy
N
−1
≥ N M(ξ
x
)
N
−1
inf
[−δ,δ]
µ
exp(−ξ
x
N x − |ξ
x
|δ)
×
Z
|N x−P y
i
|≤δ
˜
µ(y
1
) . . . ˜
µ(y
N
−1
) dy
1
. . . dy
N−1
.
Hen e,G
N
(x)
≤ −
1
βN
ln N −
N − 1
βN
ln(M (ξ
x
)) +
ξ
x
x
β
+ |ξ
x
|
δ
βN
−
1
βN
ln
inf
[−δ,δ]
µ
−
βN
1
ln P
1
N
N
X
−1
i=1
Y
i
− x
≤
δ
N
!
,
(40)where therandomvariables
Y
i
are i.i.d. of lawµ.
˜
Theequation (39)impliesthatE
(Y
i
) =
x.
A ording to the hypotheses onW
, we haveµ ∈ H
˜
1
(R \ {0}) ,
hen e we may apply
Theorem5.1of[32℄. Itimpliesthatthelaw
θ
N
ofthevariableN
X
i=1
Y
i
− Nx
!
/
√
N
onverges inH
1
(R)
tosomenormallaw. Inparti ular,wehave onvergen ein
L
∞
,
hen eP
1
N
N−1
X
i=1
Y
i
− x
≤
δ
N
!
=
Z
x+δ
√
N −1
x−δ
√
N −1
θ
N
−1
(t)dt ≥
2γδ
√
N − 1
,
for
N
largeenough,whereγ > 0
doesnotdepend onN
. Insertingthisinequalityinto(40), wendG
N
(x)
≤ −
1
βN
ln N −
N − 1
βN
ln(M (ξ
x
)) +
ξ
x
x
β
+ |ξ
x
|
δ
βN
−
1
βN
ln
inf
[−δ,δ]
µ
−
1
βN
ln
2γδ
√
N − 1
.
(41) Hen e,lim sup
N
→+∞
G
N
(x) ≤ −
1
β
ln(M (ξ
x
)) +
ξ
x
x
β
,
whi h implies,a ordingtothedenitionof
M
andξ
x
,thatlim sup
N
→+∞
G
N
(x) ≤ F
∞
(x).
(42)Estimates(38)and(42)imply
lim
N
→+∞
G
N
(x) = F
∞
(x)
. Inviewof(35),thisimplies(32).
♦
Remark5(The small temperature limit) As in Remark 3, it is possible to omputethe expansion of
F
∞
(x)
asβ → +∞
. Using the Lapla e method, and assuming thatW
is onvex, onendsthatF
∞
(x) = W (x) +
1
2β
ln W
′′
(x) + O
1
β
2
.
Let us now onsider another strategy to nd an approximation of
F
N
. In the spirit of the QuasiContinuum Method, we expandE
µ
(u
1
, . . . , u
N
)
around the equilibrium onguration
u
i
= iu
N
/N
,for agiven
u
N
. More pre isely, weset
u
i
= u
i
+ ξ
i
,assume thatξ
i
issmall, and expand the energy at se ond order with respe t toξ
i
, as explained in the Introdu tion (see (11)). We next insert this approximated energyE
e
in (29). Due to the harmoni approximation, the resulting oarse-grained energy, that we denoteE
QCM
, is analyti ally omputableandwritesE
QCM
(x) = N W (x) +
N − 1
2β
ln W
′′
(x) +
N − 1
2β
ln
β
2π
+
1
2β
ln N.
(43) Hen e,F
QCM
(x) :=
lim
N→+∞
1
N
E
QCM
(x) = W (x) +
1
2β
ln W
′′
(x) +
1
2β
ln
β
2π
.
(44) Thus,uptoa onstant,F
QCM
(x)
orrespondstotherst-orderapproximation (inpowersof1/β
)ofF
∞
(x)
.Slightly improving theproof of Theorem 2above, it is also possibleto provethe
on-vergen eofthederivativeofthefreeenergy,aquantitywhi hisindeedpra ti allyrelevant
Corollary1 Assumethat thehypotheses ofTheorem2aresatised. Then,wehave
F
N
(x) +
1
β
ln
z
N
−→ F
∞
(x)
inL
p
loc
,
∀p ∈ [1, +∞).
(45) Inparti ular, this impliesthatF
′
N
onvergestoF
′
∞
inW
−1,p
loc
.
Proof: A ordingto Theorem 2,wealreadyknowthepointwise onvergen eof
G
N
(x) =
F
N
(x) + β
−1
ln(z/N )
. WethereforeonlyneedtoprovethatG
N
isboundedinL
∞
loc
toprove our laim.Lowerbound: Wegoba kto(36),andpointoutthat
µ ≤ 1/z
. Hen e,µ
N+1
(x) ≤
N + 1
z
Z
R
µ
N
=
N + 1
z
,
whi h implies,using (34),that
G
N
+1
(x) ≥ −
1
β(N + 1)
ln
N + 1
z
,
whi h isbounded frombelowindependentlyof
N
.Upper bound: We return to (41), and noti e that a ording to the denition of
ξ
x
, the fun tionx 7→ ξ
x
is ontinuous. Inaddition,the onstantγ
in (41)isa ontinuousfun tion ofξ
x
. Therefore,(41)providesanupperbound onG
N
.Asa on lusion,
G
N
isboundedinL
∞
loc
,whi hallowsto on lude.♦
Remark6 Considering the above theoreti al results, it ould be temptingto approa h theaverage (30),that is,
hAi
N
= Z
r
−1
Z
R
A u
N
exp −βNF
N
u
N
du
N
,
byZ
∞
−1
Z
R
A u
N
exp −βNF
∞
u
N
du
N
.
(46)Notethat
F
N
hasbeenrepla edbyF
∞
intheexponentialfa tor. Thisstrategyisnote ient sin e this approximation does not provide the expansion (22)-(23) ofhAi
N
in powers of1/N
. Indeed, itispossible touse the Lapla e method to ompute the expansion of (46)asN → +∞
. ItreadsA(y
∗
) +
1
2N
σ
2
A
′′
(y
∗
) +
d
3
σ
2
A
′
(y
∗
)
+ o
1
N
,
where
σ
isdenedby (23) andd
3
= z
−1
Z
R
(y − y
∗
)
3
exp(−βW (y)) dy
. This expansion o-in ideswith (22)-(23) onlyfor the rst term,that isA(y
∗
)
. The se ondonediers, unless
Toimprove the approximation (46), onemay usethe pre isedlarge deviationsprin iple
(see[16,Th.3.7.4℄ or [4 ℄). In su ha ase,onerepla es (46)by
Z
∞
−1
Z
R
A u
N
q
F
′′
∞
(u
N
) exp −βNF
∞
u
N
du
N
.
(47)Thisquantityiswell-denedsin e
F
∞
isa onvexfun tion. Thenitisseenthattheexpansion of (47) in powers of1/N
agrees with (22)-(23) up tothe se ond term. Note however that using (47)leads toamu hmore expensive omputationthan using (23).The above onvergen e of thefreeenergy
F
N
isuseful e.g. forthe omputation ofthe freeenergyofa hainofatomswithapres ribedlength. Indeed,insu ha ase,weimposeu
N
= ℓ,
where
ℓ
is xed, and aimat omputing thefreeenergyF
N
asafun tion ofℓ
, in thelimitN → +∞.
WehaveF
N
(ℓ) = −
1
βN
ln
"Z
R
N −1
exp
−β
N
X
i=1
W
u
i
− u
i−1
h
!
du
1
. . . du
N−1
#
,
whereu
N
= ℓ
. Thelimitof
F
N
isprovidedbyTheorem 2.Another interest of the approa h is to provide an approximationof
F
′
N
(ℓ)
, aquantity relatedtothe onstitutivelawofthematerialunder onsideration,atthenitetemperature1/β
. Indeed,notethatF
N
′
(ℓ) =
hA
N
i
N−1
hB
N
i
N−1
,
(48)where
h·i
N
−1
is the average with respe t to the Gibbs measure asso iated to the energyN−1
X
i=1
W
u
i
− u
i−1
h
,andtheobservables
A
N
andB
N
aredenedbyB
N
u
1
, . . . , u
N
−1
= exp −βW N ℓ − u
N−1
,
A
N
u
1
, . . . , u
N
−1
= W
′
N ℓ − u
N−1
exp −βW N ℓ − u
N
−1
.
Hen eF
′
N
(ℓ)
anbeinterpretedastheaveragefor ebetweenatomsN − 1
andN
,whenthe positon ofatomN
is pres ribedatu
N
= ℓ
. Corollary1providesthe onvergen eof
F
′
N
(ℓ)
toF
′
∞
(ℓ)
in aweaknorm.Remark7 Note that, in (48), both observables
A
N
andB
N
depend onN
. Hen e, the resultsofSe tion2.1(obtainedusingthe Law ofLargeNumbersandnotinvolvingthe Large2.3 Numeri al tests
Forournumeri altests,we hoosethepairintera tionpotential
W (x) =
1
2
(x − 1)
4
+
1
2
x
2
(49)shownonFigure 4. Notethat
W (x)
growsfastenoughto+∞
when|x| → +∞
, su h that assumptions (20) and (31) are satised. Note also that wehave made no assumption onthe onvexityof
W
in Theorems1and2. We onsiderherea onvexpotential. Attheend ofthisse tion, wewill onsideranon- onvexexample(see(50)),andshowthatweobtainsimilar on lusions.
W (x)
x
2.5 2 1.5 1 0.5 0 -0.5 6 5 4 3 2 1 0Figure4: Thepotential
W
hosenforthetests.Werst onsiderthe omputationofensembleaverages,andweagainrestri tourselves
to the aseoftworepatoms
u
0
= 0
and
u
N
. This isjust for simpli ityand forthesakeof
demonstratingthefeasibilityandtheinterestofourapproa h. The aseof
N
r
repatomsmay betreatedlikewise. It isof oursemore omputationallydemanding,althoughaordable.We hooseanobservable
A(x)
,andwe omparethefollowingfourquantities:(i) theexa taverage
hAi
N
dened by(14). Following(4)-(5), thisquantityis omputed asthelong-timeaverageofA(u
N
(t))
alongthefullsystemdynami s
du = −∇
u
E
µ
(u) dt +
p
2/β dB
t
inR
N
.
This equationis numeri allyintegratedwith theforwardEuler s heme, withasmall
timestep. Inpra ti e,wehavesimulatedmanyindependentrealizationsofthisSDE,
in orderto omputeerrorbarsfor
hAi
N
.(ii) a QuasiContinuumtype approximation of
hAi
N
, based on the 'interpolation + har-moni expansion'pro edure outlinedabove. That is, weintrodu eE
QCM
dened by(43),andweapproximate
hAi
N
byhAi
QCM
N
:=
Z
R
A(x) exp [−βE
QCM
(x)] dx
Z
R
exp [−βE
QCM
(x)] dx
.
(iii) a Law of Large Numbers (LLN) type approximation of
hAi
N
, whi h onsists in ap-proximatinghAi
N
byA(y
∗
)
, followingTheorem1.
(iv) arenedapproximation,whi h onsistsinapproximating
hAi
N
byhAi
approx,1
N
dened by(23),followingTheorem1.Notethat onlyone-dimensional integralsare neededforapproximations(ii), (iii) and(iv).
They anbe omputedwithahigh a ura y.
We plotonFigure 5these four quantities, forin reasing valuesof
N
(the temperature isxedat1/β = 1
),fortheobservableA(x) = exp(x)
. OnFigure 6,we omparethesame quantities,nowasfun tionsofthetemperature,forN = 100
andforN = 10
. Weherework withA(x) = x
2
,forwhi h
hAi
N
= hAi
approx,1
N
. QCM renedLLN LLN exa tN
100 80 60 40 20 2.3 2.2 2.1 2 1.9 renedLLN LLN exa tN
100 80 60 40 20 1.95 1.94 1.93 1.92 1.91 1.9Figure 5: Convergen e, as
N
in reases, ofhA(u
N
)i
N
(exa t), ofhA(u
N
)i
approx,1
N
(rened LLN)andofhA(u
N
)i
QCM
N
(QCM)and omparisontoA(y
∗
)
(LLN) (temperature
1/β = 1
, observableA(x) = exp(x)
; wehaveperformed omputationsforN = 10
,25,50and100;on therightgraph,weshowerrorbarsforhA(u
N
)i
N
).Asexpe ted,thethermodynami limitstrategies(iii)and(iv)betteragreewiththefull
atom al ulation, whateverthe temperature, provided the number of eliminated atoms is
large (note that the strategy (iv) is very a urate even for the small value
N = 10
, at thetemperature1/β = 1
). Approximation(ii) is learlyinee tive forhigh temperatures. Onthe otherhand,for asu ientlysmall temperatureand asu ientlysmall numberofQCM LLN exa t
1/β
1 0.8 0.6 0.4 0.2 0.7 0.6 0.5 0.4 0.3 0.2 QCM LLN exa t1/β
1 0.8 0.6 0.4 0.2 0.7 0.6 0.5 0.4 0.3 0.2Figure 6: We plot
hA(u
N
)i
N
= hA(u
N
)i
approx,1
N
(exa t),hA(u
N
)i
QCM
N
(QCM) andA(y
∗
)
(LLN) as fun tions of the temperature1/β
: on the left,N = 100
; on the right,N = 10
(observableA(x) = x
2
).eliminatedatoms,thisapproximationis losetothefullatomresult. However,evenforthe
small values
N = 10
and1/β = 0.2
, our asymptoti resulthA(u
N
)i
approx,1
N
= 1.6299
(forA(x) = exp(x)
) is loser to the exa tresulthA(u
N
)i
N
= 1.6303 ± 0.0008
than the QCM resulthA(u
N
)i
QCM
N
= 1.6469
.Remark8 As in Remark 2, we emphasize that the omputations reported on here do not
a ountfor the onstraintsonthepositionsofatoms. Analogous omputations,thata ount
for onstraints, may be performed. They provide similar on lusions, as an be seen on
Figure7,whi h isvery similar toFigure 6.
Wenow onsider the omputation of freeenergies,morepre isely,of thederivativesof
freeenergies. Thefull atom value
F
′
N
(x)
is omputedasaratioof ensembleaverages(see (48)). We omparethisquantitywith(i) itslarge
N
limitF
′
∞
(x)
,whereF
∞
isdenedby(33),ontheonehand, (ii) and, on the other hand, its QuasiContinuum type approximationF
′
QCM
(x)
, whereF
QCM
isdenedby(44). ItreadsF
QCM
′
(x) = W
′
(x) +
1
2β
W
′′′
(x)
W
′′
(x)
.
Webriey detailhowwe omputeF
′
∞
(x)
. Letξ
x
bethe uniquerealnumberat whi hthe supremumin(33)isattained. WehaveF
′
∞
(x) =
ξ
x
QCM LLN exa t
1/β
1 0.8 0.6 0.4 0.2 0.7 0.6 0.5 0.4 0.3 0.2Figure 7: We plot
hA(u
N
)i
N
= hA(u
N
)i
approx,1
N
(exa t),hA(u
N
)i
QCM
N
(QCM) andA(y
∗
)
(LLN)asfun tionsofthetemperature1/β
(N = 100
,observableA(x) = x
2
). Thepotential
energyisoftype(18): itis equalto
W (x)
denedby(49)ifx > 0
,and+∞
otherwise. TheEuler-Lagrangeequationsolvedbyξ
x
is(39),thatwere astasz
−1
Z
R
(x − y) exp(ξ
x
y) exp(−βW (y)) dy = 0.
Letusintrodu e
G(y, ξ) = (x−y) exp(ξy)
. Wehen elookforξ
x
su hthatE
µ
[G(y, ξ
x
)] = 0
, where the random variabley
is distributed a ordingto the probability measureµ(y) =
z
−1
exp(−βW (y))
. TheRobbins-Monroealgorithm [28℄ anbeused to ompute
ξ
x
, hen eF
′
∞
(x)
.Werststudythe onvergen eof
F
′
N
(x)
toF
′
∞
(x)
asN
in reases,foraxed hainlengthx = 1.4
andaxedtemperature1/β = 1
. ResultsareshownonFigure8. Weindeedobserve thatF
′
N
(x) → F
∞
′
(x)
whenN → +∞
.Wenow omparethetwoapproximations(i)and(ii)of
F
′
N
(x)
,forN = 100
and1/β = 1
. Results are shown on Figure 9. We observethatF
′
∞
(x)
is a very good approximationofF
′
N
(x)
. Asexpe ted,thetemperatureistoohighfortheharmoni approximationtoprovide ana urateapproximationofF
′
N
(x)
. On Figure 10, weplotF
′
∞
(x)
forseveral temperatures,aswellasits zerotemperature limit,whi hisW
′
(x)
(seeRemark5).
Up to here,wehaveused the onvexpotential (49). Forthe sakeof ompleteness, we
nowbriey onsiderthe aseofanon- onvexpotential
W
. We hoosethetoy-modelW (x) = (x
2
− 1)
2
,
(50) whi h orresponds to a double-well potential. On Figure 11, we plotF
′
∞
(x)
for several temperatures, for this double-well potential. Although we have not yet ompared theseF
′
N
(x = 1.4)
F
′
∞
(x = 1.4)
N
50 40 30 20 10 2.15 2.14 2.13 2.12 2.11 2.1 Figure8: Convergen eofF
′
N
(x)
(shownwitherrorbars)toF
′
∞
(x)
asN
in reases (temper-ature1/β = 1
,xed hainlengthx = 1.4
).F
′
∞
(x)
F
′
N
(x)
F
′
QCM
(x)
x
1.6 1.4 1.2 1 0.8 0.6 0.4 3 2.5 2 1.5 1 0.5 0 -0.5 -1 -1.5 Figure9: WeplotF
′
N
(x)
,F
′
∞
(x)
andF
′
F
′
∞
(x)
,1/β = 1
F
′
∞
(x)
,1/β = 0.5
W
′
(x)
x
1.4 1.2 1 0.8 0.6 0.4 2.5 2 1.5 1 0.5 0 -0.5 -1 Figure10:F
′
∞
(x)
fordierenttemperatures.oarse-grained omputationswiththefullatom omputations,thenumeri alresultsreported
hereare onsistent withthe small temperature limit
lim
T
→0
F
′
∞
(x) = (W
∗
)
′
(x)
, whereW
∗
is
the onvexenvelopof
W
.F
′
∞
(x)
,1/β = 0.5
F
′
∞
(x)
,1/β = 1
(W
W
∗
)
′
(x)
′
(x)
x
1.5 1 0.5 0 -0.5 -1 -1.5 6 4 2 0 -2 -4 -6 -8 Figure11:F
′
∞
(x)
fordierenttemperatures,inthe aseofthedouble-wellpotential(50).3 The NNN ase and some extensions
In this Se tion, we rst onsider the ase of a NNN intera ting system. The analysis is
NNNN ase(stillforone-dimensionalsystems)andse ondthe aseoflinearpolymer hains,
whereatomssamplethephysi alspa e
R
3
.
3.1 The next-to-nearest neighbour (NNN) ase
Wenow onsiderthenext-to-nearestneighbour ase. Itturnsoutthat,forthe omputation
ofensembleaveragesaswell asfor otherquestions,this aseissigni antlymoreintri ate
thantheNN ase. Ourstrategy,basedontheLawofLargeNumbers,willbesimilartothat
usedfortheNN ase,buttheobje tmanipulatedarenotindependentrandomvariablesany
longer. Markov hainsaretherightnotionformalizingthesituationmathemati ally.
Webeginbyintrodu ingtheres aledatomisti energy,similarly to(13):
E
µ
u
1
, . . . , u
N
=
N
X
i=1
W
1
u
i
− u
i−1
h
+
N
X
−1
i=1
W
2
u
i+1
− u
i−1
h
.
(51)As above, weintrodu e the hange ofvariables (15), repla ing
(u
i
− u
i−1
)/h
bythe
inter-atomi distan es
y
i
. Re allfrom (16)thatu
N
=
1
N
N
X
i=1
y
i
. TheensembleaveragehAi
N
of anobservablethat dependsonlyontheright-endatomthereforewriteshAi
N
=
Z
−1
Z
R
N
A u
N
exp −βE
µ
u
1
, . . . , u
N
du
1
. . . du
N
=
Z
−1
Z
R
N
A
1
N
N
X
i=1
y
i
!
e
−β P
i
W
1
(y
i
)
e
−β P
i
W
2
(y
i
+y
i+1
)
dy
1
. . . dy
N
.
(52)Thekeyingredientisnowtoseetheaboveexpression,as
N
goestoinnity,asanasymptoti s foradis rete-timeMarkov hain. Theasymptoti sofMarkov hainsbeingamathemati alproblem mu h more involved than that of i.i.d. sequen es, we restri t ourselves to the
omputationof theaverage ofanobservable. Theasymptoti behaviourofthefreeenergy
may bestudied, applying aLargeDeviations Prin iple forMarkov hains(see forinstan e
[26,Th.IV.3℄). Wewillnotpursueinthis dire tion.
Se tion 3.1.1deals withthe aseoftworepatoms (namely
u
0
= 0
and
u
N
), while
Se -tion3.1.2 indi ate the hanges in order to dealwith morethan tworepatoms. Numeri al
resultswill bereportedin Se tion3.1.3.
3.1.1 Limit ofthe average,the aseof tworepatoms
Inorderto ompute
lim
N
→+∞
hAi
N
,weintrodu ethenotation
Equation(52)rewrites
hAi
N
= Z
−1
Z
R
N
A
1
N
N
X
i=1
y
i
!
e
−βW
1
(y
1
)
f (y
1
, y
2
) . . . f (y
N
−1
, y
N
) dy
1
. . . dy
N
.
(53)Ourmethod onsistsin onsideringthesequen eofvariables
(y
1
, . . . , y
N
)
in(53)asa real-izationofaMarkov hainwithkernelf (·, ·)
. However,theslightte hni aldi ultyatthis stageisthat thekernelf
isnotnormalized,sin eingeneralZ
R
f (y
1
, y
2
) dy
2
=
Z
R
exp(−βW
2
(y
1
+ y
2
)) exp(−βW
1
(y
2
)) dy
2
6= 1.
Astandardtri kofProbabilitytheoryallowsto ir umventthisdi ulty. Introdu e¯
f (x, y) := exp
−βW
2
(x + y) −
β
2
W
1
(x) −
β
2
W
1
(y)
.
Notethat
f
¯
isasymmetri fun tion (whereasf
isnot), hen etheoperatorP φ(y) =
Z
R
¯
f (y, z)φ(z) dz
(54) isself-adjointonL
2
(R)
. Considerthenψ
1
:=
argmaxZ
R
2
ψ(y) ψ(z) ¯
f (y, z) dy dz;
Z
R
ψ
2
(y) dy = 1
,
(55) andsetλ =
Z
R
2
ψ
1
(y) ψ
1
(z) ¯
f (y, z) dy dz.
(56) Using standardtoolsof spe tral theory ofself-adjoint ompa t operators, it is possible toprovethat theeigenvalue