www.elsevier.com/locate/anihpb
Regularity of the diffusion coefficient matrix for the lattice gas with energy
Yukio Nagahata
Department of Mathematical Science, Graduate School of Engineering Science, Osaka University, Toyonaka, 560-8531, Japan Received 5 April 2003; received in revised form 4 November 2003; accepted 30 March 2004
Available online 11 September 2004
Abstract
In this paper we obtain the smoothness of the diffusion coefficient matrix for the lattice gas with energy. Furthermore we also obtain the smoothness of the central limit theorem variances for certain functions.
2004 Elsevier SAS. All rights reserved.
Résumé
Dans cet article, on montre la régularité de la matrice des coefficients de diffusion pour le gaz sur réseau avec énergie. On obtient également la régularité des variances associées à certaines fonctions par le théorème limite central.
2004 Elsevier SAS. All rights reserved.
1. Introduction
In our previous paper [6] we have introduced a lattice gas with energy and derived the fluctuation dissipation equation for it. In this paper we prove that the diffusion coefficient matrix appearing in the equation is smooth.
In the derivation of hydrodynamic limit, uniqueness of the Cauchy problem of the weak solution of limiting diffusion equation is needed. It seems unsolved in the existing literatures. But once smoothness and uniform ellip- ticity of the diffusion coefficient matrix is established and if there exists a Lipschitz continuous solution, then the uniqueness question is resolved.
The smoothness of the self-diffusion coefficient of the symmetric simple exclusion process is proved by Landim, Olla, and Varadhan [4], and the smoothness of the diffusion coefficient for a lattice gas reversible under the Bernoulli measures is proved by Bernardin [1].
It seems difficult to adapt to our model the method which introduced in [4] and developed in [1], since we do not have any suitable orthonormal basis (with respect to invariant measure) of functions on the configuration space.
E-mail address: nagahata@sigmath.es.osaka-u.ac.jp (Y. Nagahata).
0246-0203/$ – see front matter 2004 Elsevier SAS. All rights reserved.
doi:10.1016/j.anihpb.2004.03.006
In this paper, we choose a basis of the space of continuous functions on the configuration space which is not orthonormal (with respect to invariant measure). We also introduce a Markov process whose state space is a set of indexes of the basis and which may be regarded as a dual process. By using this process, the diffusion coefficient matrix is given by a finite linear combination of smooth functions whose coefficient is given by expectation of the total occupation time for a certain infinite set. We prove that the expectation of the occupation time for the infinite set converges ifd3, and diverges ifd=1,2. But if we examine the linear combination more carefully, then we find that it is a difference of expectations of the total occupations times of a certain infinite set for the processes starting at two different points. Such difference makes sense as in the same way that makes the potential function of one or two dimensional random walk well defined.
This paper is organized as follows: In Section 2 we state the model and results. In Section 3 we introduce a basis of continuous functions on the configuration space and compute the coefficient ofLf with respect to the basis.
We also introduce a Markov process which may be regarded as a dual process of our lattice gas, and solve the resolvent equation by using the process. In Section 4 we estimate the expectation of occupation time of certain set for the process in dimensionsd3. In Section 5, we estimate an expectation of occupation time for a finite state continuous time stochastic jump process. In Section 6, we estimate an expectation of occupation time for certain set for the process. In Section 7 we prove the main results.
2. Model and results
Our lattice gas is a Markov process on the state spaceX:= {0,1, . . . , k}Zd. Let η=(ηx)x∈Zd stands for a generic element ofX, so that for eachx,ηx∈ {0,1,2, . . . , k}. For any local functionf we define operatorsπ(x,y) andπx→yby
π(x,y)f (η):=1{ηx=0,ηy=0}
f (η(x,y))−f (η) , πx→yf (η):=1{ηx2,1ηyk−1}
f (ηx→y)−f (η) , whereη(x,y)andηx→yare the configurations defined by
(η(x,y))z:=
ηy, ifz=x, ηx, ifz=y, ηz, otherwise, (ηx→y)z:=
ηx−1, ifz=x, ηy+1, ifz=y, ηz, otherwise.
For any local functionf we define the generator
Lf (η):=
x,y∈Zd:|x−y|=1
cex(ηx)π(x,y)f (η)+cge(ηx)πx→yf (η) ,
where| · |is the Euclidean norm of Zd andcex(r), cge(r)are positive functions ofr=0,1,2, . . . , k. Furthermore we suppose thatcex(0)=cge(0)=cge(1)=0 andcex(r) >0 for 1rkandcge(r) >0 for 2rk.
The process is thought to model a time evolution of the dynamics of a gas of particles having energy. The values ofηxare interpreted in such a way that ifηx=0, then the sitexis vacant, and ifηx=0, then there exists a particle having (discrete) energyηxat sitex. A particle at sitex moves to a nearest neighbor sitey at ratecex(ηx)ify is vacant. One unit of energy of the particle at sitexis transferred to the particle at a neighboring siteyat ratecge(ηx) if the energy of the particle at the siteyis less thank.
Consider the family of product measures on the product space{0,1,2, . . . , k}Zd with the marginal distribution
νp,α
{η: ηx=l} :=
1−p ifl=0,
pZ1
α ifl=1,
pZ1
α
αl−1
cge(2)cge(3)···cge(l) if 2lk
for allx∈Zd, where 0p1, 0α <∞andZα is the normalization constant. Put ρ:=ρ(p, α):=Eνp,α[η0],
thenρis a rational function of two variablesp, α, and for eachp,ρis a strictly increasing function ofα. Therefore there exists an inverse functionα˜= ˜α(p, ρ)say. We definePp,ρ by
Pp,ρ(·):=νp,α(p,ρ)˜ (·),
for 0p1 andpρ < kp. For any local functionf,Ep,ρ[f]is a smooth function ofpandρ. We can easily check thatLis symmetric with respect toPp,ρ. One can show that there exists a unique closed extension ofLin the space of continuous functionsC(X)with supremum norm (see [5]). We denote byet Lthe semigroup generated by the closed extension.
We define the shift operatorτxforx∈Zd, which acts on allA⊂Zd, and local functionsf as well as configu- rationsηas follows:
τxA:=x+A, τxf (η):=f (τxη), (τxη)z:=ηz−x.
LetD(p, ρ)˜ =(D˜i,j(p, ρ))i,j∈{1,2}be a 2×2 symmetric matrix defined via the variational formula a· ˜D(p, ρ)a
=
i,j
aiD˜i,j(p, ρ)aj
:=inf
u Ep,ρ
cex(η0)
π(0,e)(a1η0+a21{η0=0})+
x
π(0,e)τxu 2
+cge(η0)
π0→e(a1η0+a21{η0=0})+
x
π0→eτxu 2
, (1)
whereais any 2-dimensional vector,eis a unit vector and infuis taken over all local functions. Put χ (p, ρ):=
Ep,ρ[η02] −ρ2 (1−p)ρ (1−p)ρ p(1−p)
,
D(p, ρ)= ˜D(p, ρ)χ−1(p, ρ), (2)
for 0< p <1 andp < ρ < kp. (χ−1(p, ρ)denotes the inverse matrix ofχ (p, ρ).) Using these notations we can state our main result.
Theorem 2.1. Let(p, ρ)such that 0< p <1 andp < ρ < kp. Then the diffusion coefficient matrix for this model defined by (2) is a smooth function ofpandρ, and has the smooth extension up to the boundary.
The smoothness of the diffusion coefficient matrix is implied by that of the central limit theorem variance for certain functions.
We define the currentsweE, wPe by
wEe(η):=cex(η0)η01{η0=0,η1=0}(η)−cex(η1)η11{η1=0,η0=0}(η)
+cge(η0)1{η02,1η1k−1}(η)−cge(η1)1{η12,1η0k−1}(η), wPe(η)=cex(η0)1{η0=0,η1=0}(η)−cex(η1)1{η1=0,η0=0}(η).
Denoted byWthe linear space spanned by currentswEe, wPe for all positive unit vectorse.
Theorem 2.2. Supposef ∈W. Then∞
0
x∈ZdEp,ρ[feLtτxf]dtis a smooth function ofpandρ.
Remark. If the dimensionsdis greater than or equal to 3, the present method is applicable to generalized exclusion process which is introduced by Kipnis, Landim and Olla [2], namely Theorems 2.1 and 2.2 are valid for it in dimensionsd3.
3. Basis ofC(X)and dual process
In this section we will introduce a basis of the space of continuous functions on the configuration space and compute the coefficient ofLf with respect to the basis for a local functionf.
LetC(X)denote the space of continuous functions on X with supremum norm. It is convenient to define a subspaceC0(X)ofC(X)as follows: LetC0(X)be the set of functions such that there exist positive constantsc1 andc2with
sup
η,η:ηx=ηxfor|x|∞l
f (η)−f (η)c1e−c2l for alll0, where| · |∞is the supremum norm on Zd.
We defineAby A:=
A=(A1, . . . , Ak): AiZd,withAi∩Aj= ∅ifi=j ,
whereAZd meansA⊂Zdand|A|is finite. ForA, B∈A,B⊂AmeansBi ⊂Ai for alli,A\Bmeans that theith component isAi\Bi, andτxAmeans thatith component isτxAi (i=1,2, . . . , k). We define two types of cardinality by
#A:=
i
|Ai|,
#A˜ :=
i
i|Ai|,
and a family of functions{ΨA}A∈Aby ΨA(η):=
k i=1
x∈Ai
1{ηx=i}(η).
ForA∈Awe define the special configurationηAby (ηA)z:=
i ifz∈Aifor 1ik, 0 ifz /∈
iAi.
The cardinality #Aand#A˜ equals the number of particles and total energy for the special configurationηA, re- spectively. The functionΨAis the indicator function of the set of configurationsηfor which each site of
iAi is occupied by a particle and each particle onAihas the common energyi.
Lemma 3.1. The family of functions{ΨA}A∈Ais a basis of the linear spaceC(X). Furthermore, forf ∈C(X)if we definefˆ: A→R by
f (A)ˆ :=
B⊂A
(−1)#(A\B)f (ηB), then
f (η)=
A∈A
f (A)Ψˆ A(η). (3)
Proof. It is not difficult to see that{ΨA}A∈Aare linearly independent. Therefore we have only to prove (3).
FixA∈Aandx∈Zdsuch thatx /∈
iAi. For 1ikandB⊂A, defineBx,i∈Aby (Bx,i)j:=
Bi∪ {x} ifj =i, Bj ifj =i.
Then forf ∈C(X), f (Aˆ x,i)=
B⊂A
(−1)#(A\B)+1f (ηB)+
B⊂A
(−1)#(A\B)f (ηBx,i).
Letf be a local function. Then there existsΛ=Λ(f )Zdsuch thatf depends only on{ηx:x∈Λ}, and
f (ηA)=f (ηA∩Λ) (4)
is valid for allA∈A, whereA∩Λ∈Ais defined by(A∩Λ)i:=Ai∩Λ. These two equality shows that
f (A)ˆ =0 (5)
if
iAi∩Λc= ∅. Therefore right-hand side of (3) is a finite sum. By (4), we have only to check the truth of f (ηB)=
A∈A
f (A)Ψˆ A(ηB)
for allB∈Asuch that
Bi⊂Λ. By the binomial expansion
A⊂B
(−1)#(B\A)=(1−1)#B forB∈A, it holds that
A∈A
f (A)Ψˆ A(ηB)=
A⊂B
C⊂A
(−1)#(A\C)f (ηC)=f (ηB).
Therefore the equality of (3) is valid for every local function.
For a continuous functionf we can approximate it uniformly by a local functionfnwhich is defined by fn(η):=f (ηΛn),
where
(ηΛn)x:=
ηx ifx∈Λn, 0 ifx /∈Λn,
andΛnis a cube centered at origin and of side 2n+1. Sincefnis a local function, we can write fn(η)=
A∈A
fˆn(A)ΨA(η).
It is not difficult to check that if
iAi⊂Λn thenfˆn(A)= ˆf (A). Therefore (3) holds for every continuous func- tion. 2
From now on, we regard the sequence{ ˆf (A)}Aas the coefficient of the continuous function off in the expan- sion (3).
We expand currentswEe, wPe in the expansion (3). Then we have wEe =
k i=1
cex(i)i
Ψ∅0,i−
k j=1
Ψ∅0,i;e,j
− k i=1
cex(i)i
Ψ∅e,i−
k j=1
Ψ∅0,j;e,i
+ k i=2
k−1
j=1
cge(i)Ψ∅0,i;e,j − k
i=2 k−1
j=1
cge(i)Ψ∅0,i;e,j, (6)
wPe = k i=1
cex(i)
Ψ∅0,i−
k j=1
Ψ∅0,i;e,j
− k
i=1
cex(i)
Ψ∅e,i−
k j=1
Ψ∅0,j;e,i
, where∅x,i,∅x,i;y,j∈Aforx, y∈Zd, x=yand 1i, jkare defined respectively by
(∅x,i)l:=
{x} ifl=i,
∅ otherwise,
(∅x,i;y,j)l:=
{x} ifl=i=j, {y} ifl=j=i, {x, y} ifl=i=j,
∅ otherwise.
(7)
Denote byθethe reflection operator with respect to 1/2ealong theedirection, namely forx∈Zd, (θex)e:=
xe ife=e,
−xe+1 ife=e. (8)
We may extendθetoAnaturally. Put
Be:= {∅0,i: 1ik} ∪ {∅0,i;e,j: 1i, jk} and
˜
wEe(A):=
cex(i)i ifA= ∅0,i for 1ik,
−cex(i)i+cge(i) ifA= ∅0,i;e,j for 2ik, 1jk−1,
−cex(1) ifA= ∅0,1;e,jfor 1jk,
−cex(i)i ifA= ∅0,i;e,kfor 2ik,
0 otherwise,
˜
wPe(A):=
cex(i) ifA= ∅0,ifor 1ik, cex(i) ifA= ∅0,i;e,jfor 1i, jk, 0 otherwise,
where∅x,iand∅x,i;y,jare defined by (7). In view of (6), it holds that wEe =
A∈Be
˜
wEe(A)(ΨA−ΨθeA), wPe =
A∈Bew˜Pe(A)(ΨA−ΨθeA).
(9)
We defineAx→yandA(x,y)by Ax→y:=
(A1, . . . , Ai−1∪ {x}, Ai\ {x}, . . . , Aj\ {y}, Aj−1∪ {y}, . . . , Ak), ifx∈Ai, y∈Aj such that 2ik,1j k−1,
A, otherwise, A(x,y):=
(A1, . . . , Ai\ {x} ∪ {y}, . . . , Ak), ifx∈Ai, y /∈
jAjsuch that 1ik, A, otherwise.
Givenx, y∈Zd, defineΥx,y⊂Xby Υx,y:= {η: ηx2,1ηyk−1}.
ThenΨA(ηx→y)=ΨAy→x(η)forη∈Υx,y. Therefore using summation by parts formula, we get πx→yf (η)=1{ηx2,1ηyk−1}(η)
A
f (A)ˆ
ΨAy→x(η)−ΨA(η)
=1{ηx2,1ηyk−1}(η)
A
f (Aˆ x→y)− ˆf (A) ΨA(η).
ForA∈A, ifΨA(η)=0 for someη∈Υx,yc , thenAx→y=A. Therefore πx→yf (η)=
A
f (Aˆ x→y)− ˆf (A) ΨA(η).
Similarly
π(x,y)f (η)=
A
f (Aˆ (x,y))− ˆf (A) ΨA(η).
If we definecˆge(x, A)andcˆex(x, A)by ˆ
cge(x, A):=
cge(i) ifx∈Ai for 1ik, 0 otherwise,
ˆ
cex(x, A):=
cex(i) ifx∈Aifor 1ik, 0 otherwise,
then we can rewriteLf in the form Lf (η)=
A∈A
Lˆf (A)Ψˆ A(η),
where
Lˆf (A)ˆ =
x,y∈Zd,|x−y|=1
cˆge(x, A)f (Aˆ x→y)− ˆf (A)
+ ˆcex(x, A)f (Aˆ (x,y))− ˆf (A)
. (10)
LetXsbe a Markov process onAwhose generator isLˆ defined by (10) andPAa distribution of the Markov process starting fromA. Then the process is equivalent to the original process which starts from the configurationηA. Therefore the Markov process generated byLˆ inherits the conserved quantities from the original process. Namely, if we defineAi,j⊂Aby
Ai,j:= {A∈A: #A=i,#A˜ =j}, (11) thenAi,j becomes the ergodic classes of the Markov process generated byL.ˆ
LetXs be a Markov process onAwhose generator isLˆ defined by (10) andPAthe distribution of the Markov process starting fromA. Then the process is equivalent to the original process which starts from the configura- tionηA.
Let the sequence{ ˆgλ(A)}Abe defined by ˆ
gλ(A):=EA
∞ 0
f (Xˆ s)e−λsds, (12)
for a local functionf. Put gλ(η):=
A
ˆ
gλ(A)ΨA(η).
Lemma 3.2. The functiongλis well-define and an element ofC0(X). Furthermore this function is the solution of the resolvent equation
λgλ−Lgλ=f. (13)
Proof. It is enough to prove thatgλis well-define and an element ofC0(X). By the definition ofL, #Xˆ s and#X˜ s
are conserved. By using (5),f (A)ˆ is zero if #Aor#A˜ is large enough, and similarly forgˆλ(A). We define the stopping timeσnby
σn:=
inf{s:
i(Xs)i⊂Λn} if{s:
i(Xs)i ⊂Λn} = ∅,
∞ if{s:
i(Xs)i ⊂Λn} = ∅,
whereΛn is a cube centered at origin and of side 2n+1. Pick up a largensuch that if
iAi ∩Λcn= ∅then f (A)ˆ =0. Therefore
ˆ
gλ(A)=EA
∞
σn
f (Xˆ s)e−λsds
1
λfˆMEA[e−λσn],
wherefˆM =maxAf (A). We defineˆ dn(A)the sum of the supremum distances fromΛn to elements of
iAi: formally if
iAi\Λn= {x1, . . . , xm}then dn(A):=
m i=1
|xi|∞−n .
Since the jump rate is bounded, it is not difficult to show that there exist constantsC1, C2>0, which may depend onλ, such that
EA[e−λσn]C1e−C2dn(A).
This shows thatgλ(η)is well-define and an element ofC0(X). 2
We give the reversible measure forL, which we will use in Section 7. We defineˆ m(A, p, ρ)by m(A, p, ρ):=Ep,ρ[ΨA],
forA∈A, 0< p <1 andp < ρ < kp. The discrete measure whose mass ofAism(A, p, ρ)is also denoted bym.
Lemma 3.3. For eachA∈A,mis a smooth function ofp, ρ, and for each pair ofp andρ,mis a reversible measure forL.ˆ
Remark. We decomposeAinto{Ai,j}the ergodic classes of the Markov process. On each ergodic class,m(·, p, ρ) for 0< p <1, p < ρ < kpare essentially the same. That is, for each ergodic class,m(·, p, ρ)andm(·, pρ)are absolutely continuous each other and the Radon–Nikodym derivative is a constant depending only onp, ρ, p, ρ and the ergodic class.
Proof of Lemma 3.3. We have only to prove thatmis a reversible measure forL, or equivalently thatˆ ˆ
cge(x, A)m(A, p, ρ)= ˆcge(y, Ax→y)m(Ax→y, p, ρ), forA=Ax→y, ˆ
cex(x, A)m(A, p, ρ)= ˆcex(y, A(x,y))m(A(x,y), p, ρ), forA=A(x,y),
for eachp, ρ. But this relation immediately follows from the reversibility of the original process. 2
4. Estimate of the expectation of occupation time in dimensionsd3
In this section we will estimate an expectation of occupation time for a certain set in dimensionsd3.
Firstly we prove a general result on the expectation of occupation time for the Markov process.
Lemma 4.1. LetB⊂A, and suppose that there exist two subsetsB1,B2ofAand functionsf1andf2onAsuch that
B⊂B1⊂B2, (14)
f10, (15)
f1(A)1 ifA∈B1, (16)
Lfˆ 1(A)0 ifA∈B1c, (17)
sup
A∈B2c
f1(A) <1, (18)
f20, (19)
supf2<∞, (20)
ˆ
Lf2(A)−1 ifA∈B2. (21)
Then we have sup
A
∞ 0
EA
1B(Xs)
ds <∞.
Proof. We consider the martingale defined by
Mi,s:=fi(Xs)−fi(X0)− s 0
ˆ
Lfi(Xu)du,
fori=1,2. SinceMi,0=0, applying the Doob’s optional sampling time theorem
f1(A)=EA
f1(XσB1∧N)−
σB1∧N 0
f1(Xu)du
for allN >0, whereσB1 is a first hitting time toB1. Substituting the conditions (15)–(17) the right-hand side is greater than
EA[I{σB1<N}]. Therefore by condition (18)
sup
A∈B2cPA[σB1<∞] sup
A∈Bc2f1(A) <1.
Similarly, by using the conditions (19)–(21), we get sup
A∈B2
EA[σBc
2] sup
A∈B2
f2(A) <∞, whereσBc
2 is the first hitting time to B2c. On using the Markov property, the supremum of the expectation of occupation time forBis finite. 2
From now to the end of the section, we suppose that the dimensionsd is greater than or equal to 3. In order to apply Lemma 4.1, we prepare some functions. We define functionsh1,l, h2,l from Zd into R+, which are a small perturbation of the potential function on Zd, by
h1,l(x):=
1 if|x|l, √ 1
1+(|x|−l)2
d−2
if|x|l,
h2,l(x):=
√ 1+l 1+ |x|2
d−2
,
forl0, where| · |is the Euclidean norm of Zd. Fora=(a1, a2, a3, a4)such thata1, a2, a3>0 anda40, we define a functiongafrom R+into R+satisfying the following conditions
• gais a continuously differentiable function on R+.
• On the interval[0, a1],gais a quadratic function with−a2for the coefficient of the second order term.
• There existsba1such that:
– On the interval[a1, b],gais a quadratic function witha3for the coefficient of the second order term.
– On the interval[b,∞),gais zero.
• The differential coefficient from the right atx=0 is−a4. Formally
ga(x):=
−a2
x+2aa422
+(2a1a24+a4)21
a3 +a12
, if 0xa1, a3
x−2a1a2+a4
2a3 +a12
, ifa1xa1+2a12aa23+a4,
0, ifxa1+2a12aa23+a4.
Let us defineAM by
AM:= {A∈A: #A=M}.
Assume thatΛis a Euclidean ball with radiusl, and defineB:=BM,lby B:=
A∈AM:
Ai∩Λ= ∅or there existx, y∈
i
Aisuch that 1|x−y|∞l
. (22)
Here| · |∞is the supremum norm of Zd. ForA∈AM, we define{xi: 1iM} := {xi(A): 1iM} ⊂Zdby {xi: 1iM} =
i
Ai,
andxi,j(A)by thejth component ofxi(A).
Let us define constantsa1, a2andc±by a1:=
M2+l, 2 c−, 1
c+M2,0
, (23)
a2:=
M2+l, 1 c−, 1
2c+M2,4M2c+ c−
, (24)
and
c−:=min
r cex(r), (25)
c+:=max
r cex(r). (26)
Lemma 4.2. Define setsB1,B2and functionsf1, f2by B1:=B,
B2:=
A∈AM: there existsx∈Asuch that|x|∞M2+l
or there existsx, y∈Asuch that 1|x−y|∞M2+l , f1(A):=
M i=1
h1,l(xi)+
1i<jM
h2,l(xi−xj),
f2(A):=
M i=1
d j=1
ga1
|xi,j|
+
1i<mM
d j=1
ga2
|xi,j−xm,j| ,
wherea1, a2andc±are defined by (23)–(26). Then the conditions of Lemma 4.1 hold, namely (14)–(21) hold.
Corollary 4.3. Suppose thatd3. Then forBdefined by (22), sup
A
∞ 0
EA
1B(Xs)
ds <∞.
Proof of Lemma 4.2. We have only to check the truth of the conditions (17) and (21). We defineB0by B0=B0,M:=
A∈AM: xi(A)−xj(A)>1 fori=j .
Then on the setB0,Lˆ behave as a discrete Laplacian on Zd for eachxi(A). Formally, if we denote byd the discrete Laplacian on Zd, then
Lfˆ 1(A)= M
i=1
ˆ
cex(xi(A), A)(dh1,l)(xi)+
1i<mM
cˆex
xi(A), A + ˆcex
xm(A), A
(dh2,l)(xi−xm), forA∈B0. Sinceh1,l, h2,lare small perturbations of thed-dimensional potential function, each term of the right- hand side is non-positive. SinceBc1⊂B0, the condition (17) is valid.
We also have Lfˆ 2(A)=
M i=1
d j=1
ˆ cex
xi(A), A
(1ga1)xi,j(A)
+
1i<mM
d j=1
cˆex
xi(A), A + ˆcex
xm(A), A
1ga1xi,j(A)−xm,j(A), forA∈B0. From the choice of the second component ofa1, a2, if|xi,j(A)|M2+l, then
ˆ cex
xi(A), A
(1ga1)xi,j(A)−2 and if 2|xi,j(A)−xm,j(A)|M2+l, then
cˆex
xi(A), A + ˆcex
xm(A), A
(1ga1)xi,j(A)−xm,j(A)−2.
From the choice of the third component ofa1, a2, each terms of the last equality is less than or equal to 1/M2. On the setB2∩B0, there existsisuch that|xi,j(A)|M2+lfor allj, or there existsi, msuch that 2|xi,j(A)− xm,j(A)|M2+lfor allj. Therefore on the setB2∩B0, the condition (21) is valid.
On the setB2\B0, without loss of generality we assume that there existn2 and 1jd such thatxl= x1+(l−1)ejfor 1ln, whereejis a unit vector for thejth coordinate, and ifn+1iM, then|xi−xm|2 for all 1mM,i=m. Furthermore, we can assume that the choice of thexi(A(x,y))is similar toxi(A), namely we assume that
xi(A(y,z))=
xi(A) ify=xi(A), z ify=xi(A), for all 1iM,y∈
iAi andz /∈
iAi. ThenLgˆ a1(|xi,j(A)|)andLgˆ a2(|xi,j(A)−xm,j(A)|)make sense for 1i < mM. By elementary computation if 1i < mn, then
Lgˆ a2xi,j(A)−xm,j(A)0, and
Lgˆ a2x1,j(A)−x2,j(A)−4c+M2 c− , Lgˆ a2xn−1,j(A)−xn,j(A)−4c+M2
c− .
The last value4c+c−M2 is given by the choice of the fourth component ofa2. We also compute that Lˆ
1in
ga1xi,j(A)= ˆcex
x1(A), A
ga1x1,j(A)−1−ga1x1,j(A) + ˆcex
xn(A), A
ga1xn,j(A)+1−ga1xn,j(A).
Since we assume thatxi,j(A)=xi−1,j(A)+1 for 2in, the right-hand side of the last equality is equal to ˆ
cex
x1(A), A
1in
(1ga1)xi,j(A) +
ˆ cex
xn(A), A
− ˆcex
x1(A), A
ga1xn,j(A)+1−ga1xn,j(A).
From the choice of the third component ofa1, the first line is less than or equal tom/M2. Since the maximal value ofga1(x)−ga1(x+1)is given by the maximal value of differential coefficient ofga1, from the choice ofa1the second line is less than or equal to 4c+(M2+l)/c−. From the assumption of thexi forin+1, the other terms forLfˆ 2is less than or equal to 1/M2. Thus we have checked the truth of the condition (21). 2
5. Occupation time for some stochastic processes
In this section, we consider finite state continuous time stochastic jump processes, which may not be Markovian.
LetSbe a finite set, which is a state space of our stochastic process. For eachi∈S, there is given a probability measureµi onS×R+. Let(Yi,n, σi,n), i∈S, n=0,1,2, . . .be a system of independent random variables taking values inS×R+such that for eachithe joint distribution of(Yi,n, σi,n)is given byµi for everyn.
Then our process is described as follows. Suppose that the process starts ati0∈S. Then one takes the random variable(Yi0,0, σi0,0) and let the process stay at i0 up to the time σi0,0 and jump to Yi0,0 at time σi0,0. The procedure is repeated over by letting it start ati1=Yi0,0 and taking up(Yi1,1, σi1,1)place of(Yi0,0, σi0,0). We denote the distribution of the process which starts fromi∈SbyPi. We suppose several conditions for the joint distributionµi. Let us define
Fi,j(λ):=Eµi
1{Yi=j}e−λσi . (A.1) It holds that
Fi,i(λ)=0 for alli∈S.
(A.2) There exist constantspi,j, fi,j,and a functiong(λ)such that Fi,j(λ)=pi,j−fi,jg(λ)+o
g(λ) , wherepi,j 0 and for alli∈S
j∈S
pi,j=1;
fi,j is non-negative; andg(λ)is a positive and increasing function from R+to R+which vanishes atλ=0. Here o(·)is Landau’s symbol asλtends to 0. In this paperg(λ)will be√
λor−1/logλaccordingly asd=1 or 2.
(A.3) The stochastic matrixpi,j in (A.2) is irreducible.
(A.4) There exists a pair(i, j )such thatfi,j is strictly positive.
(A.5) There existsi0such thatfi0,j=0 for allj. Furthermore, there exists a positive constantfi0 such that
j
Fi0,j(λ)=1−fi0λ+o(λ).
Lemma 5.1. Suppose that (A.1)–(A.5) hold. Then for alli∈S, the Laplace transform of occupation time fori0is asymptotically equivalent to 1/g(λ)(asλtends to 0), namely there exists a constantC1which depends onisuch that
Ei ∞ 0
1{i0}(Xs)e−λsds= C1 g(λ)+o
1 g(λ)
.
Furthermore for alli, j ∈S, there exists a limit of the difference of Laplace transform of occupation time fori0, namely there exists a constantC2which depends oniandj such that
lim
λ→0
Ei
∞ 0
1{i0}(Xs)e−λsds−Ej ∞ 0
1{i0}(Xs)e−λsds
=C2.
Proof. LetQi,i0(λ)be the Laplace transform of the occupation time fori0with starting pointi∈S. Letσ1be the first jump time. Then on using the conditional independence of the process after the timeσ1and the jump timeσ1 givenXσ1,