We are now interested in haraterizing probability measures on the set of perfet mathings
M ( G )
oftheinnitegraphG
whihare,insomeappropriatesense,innitevolumeversionsofthe Boltzmann measureonM ( G n )
. Reallthat bydenition, theprobabilityof amathinghosen with respet to the Boltzmann measure onM ( G n )
is proportional to the produt of its edge weights. Thisdenitiondoesnot workwhenthegraph isinnite, andis replaedbythenotionof Gibbs measure, whih is a probability measure on
M ( G )
satisfying the DLR 1 onditions:if the perfet mathing inan annular region of
G
is xed, mathings inside and outside of the annulus are independent, and the probability of any interior mathing is proportional to theprodutof itsedge-weights.
3.3.1 Limit of Boltzmann measures
A natural way ofonstruting aGibbs measureis to take thelimit ofthe Boltzmann measures
on ylinder sets of
M ( G n )
, where a ylinder set onsists of all perfet mathings ontaining axedsubset of edgesof
G n.
Theorem 2 gives an expliit expression for the Boltzmann measure on ylinder sets when the
graph is planar and nite. In the ase of toroidal graphs, a similar but more ompliated
expressionholds: itisaombinationoffourstermssimilartothose ofEquation(2.2) ,involving
thematries
K n 00 , . . . , K n 11,andtheir inverses.
Usingtheblokdiagonalizationofthematries
K n στ oftheproofofTheorem9,oneanompute
the elements of the inverse expliitly and obtain Riemann sums. The onvergene of these
Riemann sums is againompliated bythe zerosof
P (z, w)
on thetorusT 2,but an be shown
toonvergeonasubsequeneofn
'stotherighthandsideofEquation(3.7) . UsingaTheoremof
Sheeld [She05 ℄ whih shows a prioriexistene of thelimit,one dedues onvergene for every
n
. Then, by Kolmogorov's extension theorem, there exists a unique probability measure on( M ( G ), σ( A ))
whihoinideswiththelimitoftheBoltzmannmeasuresonylinder sets,whereσ( A )
isthesmallestsigma-eldontainingylindersets. ThislimitingmeasureisofGibbstypeby onstrution. Wehave thus skethed theproof ofthefollowing theorem.
1
DLRstandsforDobrushin,LanfordandRuelle
Theorem12. [CKP01 ,KOS06 ℄Let
{ e 1 = w 1 b 1 , . . . , e k = w k b k }
beasubset ofedgesofG
. Then there existsa unique probability measureµ
on( M ( G ), σ( A ))
suh that:µ( e 1 , . . . , e k ) = Y k i=1
K( w i , b i )
!
det(K − 1 ( b i , w j ) 1 ≤ i,j ≤ k ),
(3.7)where
K
is a Kasteleynmatrix assoiated tothe graphG
,and assumingb
andw
are in a single fundamental domain:K − 1 ( b , w + (x, y)) = 1 (2πi) 2
Z
T 2
Q bw (z, w)
P (z, w) z x w y dw w
dz z ,
and
Q bw (z, w)
isthe( b , w )
element of the adjugatematrix (transpose of the ofatormatrix) ofK 1 (z, w)
. Itis a polynomial inz, w, z − 1 , w − 1.
3.3.2 Ergodi Gibbs measures
Inthe previous setion,wehave expliitlydetermined aGibbsmeasure on
M ( G )
. We now aimat haraterizing all of them. In order to this in a way whih is oherent with the model, we
introdue thefollowing notions.
A probability measure on
M ( G )
is translation-invariant, if the measure of a subset ofM ( G )
is invariant under the translation-isomorphism ation. An ergodi Gibbs measure (EGM) is a
Gibbs measure whih is translation invariant and ergodi, i.e. translation invariant sets have
measure
0
or1
.ForanergodiGibbsmeasure
µ
,dene theslope(s, t)
tobetheexpetedhorizontalandvertialheight hange in the
(1, 0)
and(0, 1)
diretions, that iss = E µ [h( v + (1, 0)) − h( v )]
, andt = E µ [h( v + (0, 1)) − h( v )]
.Let us denote by
µ n the Boltzmann measure on M ( G n )
. For a xed (s, t) ∈ R 2, let M s,t ( G n )
M s,t ( G n )
be thesetof mathings of
G n whih have height hange ( ⌊ sn ⌋ , ⌊ tn ⌋ )
. Assuming thatM s,t ( G n )
is non-emptyfor
n
suiently large, letµ n (s, t)
denote theonditional measure indued byµ n
on
M s,t ( G n )
. Then, a haraterization of all ergodi Gibbsmeasures onM ( G )
isgiven bythefollowing theoremofSheeld.
Theorem 13. [She05 ℄For eah
(s, t)
for whihM s,t ( G n )
is non-emptyforn
suientlylarge,µ n (s, t)
onverges asn → ∞
to an EGMµ(s, t)
of slope(s, t)
. Furthermoreµ n itselfonverges
to
µ(s 0 , t 0 )
where(s 0 , t 0 )
is the limit of the slopes ofµ n. Finally, if (s 0 , t 0 )
lies in the interior
of the set of
(s, t)
for whihM s,t ( G n )
is non-empty forn
suiently large, then every EGMof slope
(s, t)
is of the formµ(s, t)
for some(s, t)
as above; that isµ(s, t)
is the unique ergodiGibbs measure of slope
(s, t)
.Proof. TheexisteneisestablishedbytakinglimitsofBoltzmannmeasuresonlarger andlarger
tori while restriting height hange. The uniqueness is muh harder, and we won't disuss it
here.
3.3.3 Newton polygon and available slopes
Theorem13raisesthefollowingquestion: whatisthesetofpossibleslopesforGibbsmeasuresor
equivalently for limitsofonditional Boltzmannmeasures ? Theanswerisgiven bytheNewton
polygon
N (P )
dened as follows:N (P )
is the losed onvex hull inR 2 of the set of integer
exponents ofthe monomials of the harateristi polynomial
P (z, w)
,up to ontribution ofthe refereneowω M 0,thatis:
N (P ) =
onvexhull{ (i, j) ∈ Z 2 | z i+x 0 w j+y 0 is amonomial inP (z, w) } .
Proposition 14. [KOS06 ℄ The Newton polygon is the set of possible slopes of EGMs, that is
there existsan EGM
µ(s, t)
if and onlyif(s, t) ∈ N (P )
.Proof. Observing that hanging the referene ow merely translates the Newton polygon, we
assumethat
(x 0 , y 0 ) = (0, 0)
.Letus rst prove thatif
(s, t) ∈ N (P )
,thenthere isa Gibbsmeasure ofslope(s, t)
,orequiva-lently
M s,t ( G n )
is non-emptyforn
large enough. For onveniene, we will assumethat thesetof possible slopesis losed.
ByLemma8,theabsolutevalueoftheoeient
z i w j inP (z, w)
istheweightedsumof
math-ings of
G 1 with height hange (i, j)
, thus there is a mathing orresponding to eah extremal
pointofN (P )
,i.e. if(s, t)
isanextremalpointofN (P)
,thenM s,t ( G 1 ) 6 = ∅
. Itsuestoshow
thatif
M s 1 ,t 1 ( G n 1 )
andM s 2 ,t 2 ( G n 2 )
arenon-emptyfor somen 1 andn 2,thenM s 1+ s 2
M s 1+ s 2
2 , t 1+ 2 t 2 ( G m )
is also non-empty for some
m
. Indeed, by indution, this allows to prove existene of a Gibbsmeasure of slope
(s, t)
for a dense subset of the Newtonpolygon. The proofis ended by usingtheassumption thatthe set ofpossibleslopesis losed.
Withoutlossof generality,weanassumethat
n 1 = n 2,otherwisetake thelmoftheir periods.
Considertwomathingsof
M s 1 ,t 1 ( G n )
andM s 2 ,t 2 ( G n )
,respetively. Thesuperimpositionofthe two mathings being a set of disjoint alternating yles, one an hange from one mathingtotheotherbyrotatingalongtheyles. Iftheheight hanges
( ⌊ s 1 n ⌋ , ⌊ t 1 n ⌋ )
,( ⌊ s 2 n ⌋ , ⌊ t 2 n ⌋ )
ofthetwo mathings areunequal, some ofthese yles have non-zero homology in
H 1 ( T 2 , Z )
,sothatrotating along them will hange the height hange. On the toroidal graph
G 2n, onsider four opies ofthetwo mathings andshift halfofthenon-trivial yles;thisreates anewmathing
with height hange
( ⌊ (s 1 + s 2 )n ⌋ , ⌊ (t 1 + t 2 )n ⌋ ) = ⌊ s 1 +s 2 2 2n ⌋ , ⌊ t 1 +t 2 2 2n ⌋
, thus proving that
M s 1+ s 2
2 , t 1+ 2 t 2 ( G m ) isnon-empty form = 2n
.
Let us now suppose that there exists a Gibbs measure
µ(s, t)
of slope(s, t)
and prove that(s, t) ∈ N (P )
. Denoteby~ E 1 thesetofdiretededgesofthefundamentaldomain G 1 = ( V 1 , E 1 )
.
Realling thatthe divergene
div
is alinearfuntionof ows,thesetof non-negative, white-to-blakunit ows denes apolytopeofR ~ E 1:
{ ω ∈ R ~ E 1 : ∀ wb ∈ E 1 , ω( b , w ) = 0, 0 ≤ ω( w , b ) ≤ 1; ∀ w ∈ W 1 , div ω( w ) = 1, ∀ b ∈ B 1 , div ω( b ) = − 1 } .
The mapping
ψ
whih assigns to a owω
the total uxarossγ x and γ y isa linearmapping
fromthepolytopeto
R 2,implyingthattheimage ofthepolytopeunderψ
istheonvexhull of
theimagesof the extremalpointsof thepolytope.
Now, from Setion 3.1, we know that every dimer onguration of
G 1 denes a non-negative,
white-to-blakunitowtakingvaluesin{ 0, 1 }
oneverydiretededgeof~ E 1. Theonversebeing
also true, this implies that extremal points of thepolytope aregiven by dimer ongurations.
Sine therefereneow issuh that
(x 0 , y 0 ) = (0, 0)
,theimage ofa dimeronguration underψ
is its height hange. This means thatthe image of extremal pointsof the polytopeontainsthe extremal points of the Newton polygon
N (P )
; the image of the polytope underψ
is thusN (P )
.TheGibbs measure
µ(s, t)
of slope(s, t)
denesa non-negative, white-to-blak owω µ(s,t):
∀ e = wb ∈ E 1 , ω µ(s,t) ( w , b ) = µ(s, t)( e ), ω µ(s,t) ( b , w ) = 0.
Sine
µ(s, t)
isaprobabilitymeasure,theowω µ(s,t) hasdivergene1ateverywhitevertexand
-1ateveryblakvertex. Itthusbelongstothepolytopeanditsimageunder
ψ
belongstoN (P )
.Theproof isonluded byobserving thatthe image of
µ(s, t)
underψ
is theslope(s, t)
.Example 3.2. Figure3.7showstheNewtonpolygonofthedimermodelonthesquare-otagon
graph with weights 1 on the edges. Marked points represent monomials of the harateristi
polynomial
P (z, w) = 5 − z − 1 z − w − w 1.
1 1
Figure 3.7: Newton polygon of the dimer model on the square-otagon graph with uniform
weights.
3.3.4 Surfae tension
For every
(s, t) ∈ N (P )
,letZ s,t ( G n )
bethe partition funtion ofM s,t ( G n )
,thatis:Z s,t ( G n ) = X
M ∈M s,t ( G n )
ν(M ).
Then, by denition, thefree energy ofthemeasure
µ(s, t)
is:σ(s, t) = − lim
n →∞
1
n 2 log Z s,t ( G n ).
The funtion
σ : N (P ) → R
is knownas thesurfae tension. Sheeld [She05 ℄ provesthatit is stritly onvex.Asa onsequene of this denition and of Theorem 13, one dedues thatthe measure
µ(s 0 , t 0 )
of Theorem13 isthe onewhih hasminimal freeenergy. Moreover, sine thesurfae tensionis
stritly onvex, the surfaetension minimizing slope isunique and equal to
(s 0 , t 0 )
.3.3.5 Construting Gibbs measures
Theorem12ofSetion3.3.1provesanexpliitexpressionfortheGibbsmeasure
µ(s 0 , t 0 )
ofslope(s 0 , t 0 )
. Our goalnow isto obtainanexpliit expressionfor theGibbsmeasuresµ(s, t)
withallpossibleslopes
(s, t)
.ReallthatbyTheorem 13,theGibbsmeasure
µ(s, t)
isthelimit oftheonditionalBoltzmannmeasures
µ n (s, t)
onM s,t ( G n )
. The problem is that onditional measures are hard objetsto work with in order to obtain expliit expressions. But we know how to handle the full
Boltzmann measure, whih onverges to the Gibbs measure of slope
(s 0 , t 0 )
. So the idea toavoidhandling onditionalmeasures isto modifytheweightfuntionon theedgesof
G ninsuh
a waythatmathings withanother slope than
(s 0 , t 0 )
get favored. Hene, we arelooking for aweight funtion whih satisesthe following: thenew weight of a mathingis equal to theold
weight multipliedbyaquantity whih dependsonly on its height hange. This anbe done by
introduing magneti eld oordinates asfollows.
Reallthat
γ x,n,γ y,n areorientedhorizontal and vertial yles inthe dualgraph G ∗
G ∗
n
obtainedby taking
n
times the basisvetore x of the underlying lattie Z 2, n
times the basisvetor e y
n
times the basisvetore y
respetively,embeddedonthetorus. Then,on
G ∗
n
therearen
horizontalopiesoftheyleγ x,n,
and
n
vertial opies of the yleγ y,n. Let (B x , B y )
be two real numbers known as magneti
eld oordinates. Multiply all edges rossing the
n
opies of the horizontal yleγ x,n by e ± B x,
depending on whether the white vertex is on the left or on the right. In a similar way, edges
rossing the
n
opies of thevertial yleγ y,n are multipliedbye ± B y. This denes amagneti
altered weight funtion,denoted by
ν (B x ,B y ) satisfying our requirement. Indeed, let M 0 be the
periodireferenemathingof
G n,anddenotebyx n 0,y 0 nthetotaluxofω M 0 throughγ x,n,γ y,n.
y 0 nthetotaluxofω M 0 throughγ x,n,γ y,n.
γ x,n,γ y,n.
Then, arguing ina way similarto the proofof Lemma 8,one an express themagneti altered
weight funtion
ν (B x ,B y ) astheweight funtionν
,multiplied bya quantitywhih onlydepends
on theheight hange:
∀ M ∈ M ( G n ), ν (B x ,B y ) (M ) = ν(M)e nB x (h M x +x n 0 ) e nB y (h M y +y n 0 ) .
(3.8)Let
P (B x ,B y ) (z, w)
betheharateristipolynomialofthegraphG 1orrespondingtothemagneti
altered weight funtion. The key fat is that P (B x ,B y ) (z, w)
an easily be expressed using the
harateristipolynomial
P (z, w)
ofthegraphG 1: expressingP (B x ,B y ) (z, w)
usingLemma8and
replaing
ν (B x ,B y ) (M )
bythe right handside of (3.8)inthease wheren = 1
,yields:P (B x ,B y ) (z, w) = P(e B x z, e B y w).
Let
Z (B x ,B y ) ( G n )
bethe partitionfuntionandµ (B n x ,B y )betheBoltzmannmeasureofthegraph
G n withthemagneti alteredweight funtion. Denote byµ (B x ,B y ) theGibbsmeasureobtained
asweak limit of the Boltzmann measures
µ (B n x ,B y ). Then, asa diret orollaryof Theorems 11
and 12,we have:
So,itisquiteremarkablethattheresultsobtainedfortheweightfuntion
ν
alsoyieldtheresultsfor themagneti alteredweight funtion
ν (B x ,B y ).
Note that we have not yet related the magneti eld oordinates to the slope of the Gibbs
measure