Using a capacity approach and the theory of the measure’s perturbation of the Dirichlet forms, we give the probabilistic representation of the general Robin boundary value problems on an arbitrary domain Ω, involving smooth measures, which give rise to a new process obtained by killing the general reflecting Brownian motion at a random time. We obtain some properties of the semigroup directly from its probabilistic representation, some convergence theorems, and also a probabilistic interpretation of the phenomena occurring on the boundary.
1. Introduction
The classical Robin boundary conditions on a smooth domain Ω of ℝN (N≥0) is giving by
(1.1)∂u∂ν+βu=0on∂Ω,
where ν is the outward normal vector field on the boundary ∂Ω and β a positive bounded Borel measurable function defined on ∂Ω.
The probabilistic treatment of Robin boundary value problems has been considered by many authors [1–4]. The first two authors considered bounded C3-domains since the third considered bounded domains with Lipschitz boundary, and the study of [4] was concerned with C3-domains but with smooth measures instead of β. If one wants to generalize the probabilistic treatment to a general domain, a difficulty arise when we try to get a diffusion process representing Neumann’s boundary conditions.
In fact, the Robin boundary conditions (1.1) are nothing but a perturbation of ∂/∂ν, which represent Neumann’s boundary conditions, by the measure μ=β·σ, where σ is the surface measure. Consequently, the associated diffusion process is the reflecting Brownian motion killed by a certain additive functional, and the semigroup generated by the Laplacian with classical Robin boundary conditions is then giving by
(1.2)𝒫tμf(x)=Ex[f(Xt)e-∫0tβ(Xs)dLs],
where (Xt)t≥0 is a reflecting Brownian motion (RBM) and Lt is the boundary local time, which corresponds to σ by Revuz correspondence. It is clear that the smoothness of the domain Ω in classical Robin boundary value problem follows the smoothness of the domains where RBM is constructed (see [5–10] and references therein for more details about RBM).
In [6], the RBM is defined to be the Hunt process associated with the form (ℰ,ℱ) defined on L2(Ω) by
(1.3)ℰ(u,v)=∫Ω∇u∇vdx,∀u,v∈ℱ=H1(Ω),
where Ω is assumed to be bounded with Lipschitz boundary so that the Dirichlet form (ℰ,ℱ) is regular. If Ω is an arbitrary domain, then the Dirichlet form needs not to be regular, and to not to lose the generality we consider ℱ=H~1(Ω), the closure of H1(Ω)∩Cc(Ω¯) in H1(Ω). The domain H~1(Ω) is so defined to insure the Dirichlet form (ℰ,ℱ) to be regular.
Now, if we perturb the Neumann boundary conditions by Borel’s positive measure [11–13], we get the Dirichlet form (ℰμ,ℱμ) defined on L2(Ω) by
(1.4)ℰμ(u,v)=∫Ω∇u∇vdx+∫∂Ωu~v~dμ,∀u,v∈ℱμ=H~1(Ω)∩L2(∂Ω,dμ).
In the case of μ=β·σ (Ω bounded with Lipschitz boundary), (1.4) is the form associated with Laplacian with classical Robin boundary conditions and (1.2) gives the associated semigroup. In the case of an arbitrary domain Ω, we make use of the theory of the measure’s perturbation of the Dirichlet forms, see, for example, [14–23].
More specifically, we adapt the potential theory and associated stochastic analysis to our context, which is the subject of Section 2. In Section 3, we focus on the diffusion process (Xt)t≥0 associated with the regular Dirichlet form (ℰ,H~1(Ω)). We apply a decomposition theorem of additive functionals to write Xt in the form Xt=x+Bt+Nt, we prove that the additive functional Nt is supported by ∂Ω, and we investigate when it is of bounded variations.
In Section 3, we get the probabilistic representation of the semigroup associated with (1.4), and we prove that it is sandwiched between the semigroup generated by the Laplacian with the Dirichlet boundary conditions and that of the Neumann ones. In addition, we prove some convergence theorems, and we give a probabilistic interpretation of the phenomena occurring on the boundary.
2. Preliminaries and Notations
The aim of this section is to adapt the potential theory and the stochastic analysis for application to our problem. More precisely, it concerns the notion of relative capacity, smooth measures, and its corresponding additive functionals. This section relies heavily on the book of Fukushima [17], particularly, chapter 2 and 5, and the paper in [11]. Throughout [17], the form (ℰ,ℱ) is a regular Dirichlet form on L2(X,m), where X is a locally compact separable metric space, and m a positive Radon measure on X with supp[m]=X.
For our purposes, we take X=Ω¯, where Ω is an Euclidean domain of ℝN, and the measure m on the σ-algebra ℬ(X) is given by m(A)=λ(A∩Ω) for all A∈ℬ(X) with λ the Lebesgue measure; it follows that L2(Ω)=L2(X,ℬ(X),m), and we define a regular Dirichlet form (ℰ,ℱ) on L2(Ω) by
(2.1)ℰ(u,v)=∫Ω∇u∇vdx,ℱ=H1~(Ω),
where H~1(Ω)=H1(Ω)∩Cc(Ω¯)¯H1(Ω). The domain H~1(Ω) is so defined to insure the Dirichlet form (ℰ,ℱ) to be regular, instead of ℱ=H1(Ω) which make the form not regular in general, but if Ω is bounded open set with Lipschiz boundary, then H~1(Ω)=H1(Ω).
We denote for any α>0:ℰα(u,v)=ℰ(u,v)+α(u,v)m, for all u,v∈ℱ.
2.1. Relative Capacity
The relative capacity is introduced in a first time in [11] to study the Laplacian with general Robin boundary conditions on arbitrary domains. It is a special case of the capacity associated with a regular Dirichlet form as described in chapter 2 of [17]. It seems to be an efficient tool to analyse the phenomena occurring on the boundary ∂Ω of Ω.
The relative capacity which we denote by CapΩ¯ is defined on a subsets of Ω¯ by the following: for A⊂Ω¯ relatively open (i.e., open with respect to the topology of Ω¯) we set
(2.2)CapΩ¯(A):=inf{ℰ1(u,u):u∈H~1(Ω):u≥1a.e.onA}.
And for arbitrary A⊂Ω¯, we set
(2.3)CapΩ¯(A):=inf{CapΩ¯(B):BrelativelyopenA⊂B⊂Ω¯}.
A set N⊂Ω¯ is called a relatively polar if CapΩ¯(N)=0.
The relative capacity (just as a cap) has the properties of a capacity as described in [17]. In particular, CapΩ¯ is also an outer measure (but not a Borel measure) and a Choquet capacity.
A statement depending on x∈A⊂Ω¯ is said to hold relatively quasieverywhere (r.q.e.) on A, if there exist a relatively polar set N⊂A such that the statement is true for every x∈A∖N.
Now we may consider functions in H~1(Ω) as defined on Ω¯, and we call a function u:Ω¯→ℝ relatively quasicontinuous (r.q.c.) if for every ϵ>0 there exists a relatively open set G⊂Ω¯ such that CapΩ¯(G)<ϵ and u|Ω¯∖G is continuous.
It follows [13] that for each u∈H~1(Ω) there exists a relatively quasicontinuous function u~:Ω¯→ℝ such that u~(x)=u(x)m—a.e. This function is unique relatively quasieverywhere. We call u~ the relatively quasicontinuous representative of u.
For more details, we refer the reader to [11, 13], where the relative capacity is investigated, as well as its relation to the classical one. A description of the space H01(Ω) in term of relative capacity is also given, namely,
(2.4)H01(Ω)={u∈H~1(Ω):u~(x)=0r.q.e.on∂Ω}.
2.2. Smooth Measures
All families of measures on ∂Ω defined in this subsection were originally defined on X [17], and then in our settings on X=Ω¯, as a special case. We reproduce the same definitions, and most of their properties on ∂Ω, as we deal with measures concentrated on the boundary of Ω for our approach to the Robin boundary conditions involving measures. There is three families of measures, as we will see in the sequel: the familes S0, S00, and S. We put ∂Ω between brackets to recall our context, and we keep in mind that the same things are valid if we put Ω or Ω¯ instead of ∂Ω.
Let Ω⊂ℝN be open. A positive Radon measure μ on ∂Ω is said to be of finite energy integral if
(2.5)∫∂Ω|v(x)|μ(dx)≤Cℰ1(v,v),v∈ℱ∩Cc(Ω¯)
for some positive constant C. A positive Radon measure on ∂Ω is of finite energy integral if and only if there exists, for each α>0, a unique function Uαμ∈ℱ such that
(2.6)ℰα(Uαμ,v)=∫∂Ωv(x)μ(dx).
We call Uαμ an α-potential.
We denote by S0(∂Ω) the family of all positive Radon measures of finite energy integral.
Lemma 2.1.
Each measure in S0(∂Ω) charges no set of zero relative capacity.
Let us consider a subset S00(∂Ω) of S0 defined by
(2.7)S00(∂Ω)={μ∈S0(∂Ω):μ(∂Ω)<∞,∥U1μ∥∞<∞}.
Lemma 2.2.
For any μ∈S0(∂Ω), there exist an increasing sequence (Fn)n≥0 of compact sets of ∂Ω such that
(2.8)1Fn·μ∈S00(∂Ω),n=1,2,…,CapΩ¯(K∖Fn)→0,n→+∞foranycompactsetK⊂∂Ω.
We note that μ∈S0(∂Ω) vanishes on ∂Ω∖∪nFn for the sets Fn of Lemma 2.2, because of Lemma 2.1.
We now turn to a class of measures S(∂Ω) larger than S0(∂Ω). Let us call a (positive) Borel measure μ on ∂Ω smooth if it satisfies the following conditions:
μcharges no set of zero relative capacity;
there exist an increasing sequence (Fn)n≥0 of closed sets of ∂Ω such that
(2.9)μ(Fn)<∞,n=1,2,…,(2.10)limn→+∞CapΩ¯(K∖Fn)=0foranycompactK⊂∂Ω.
Let us note that μ then satisfies
(2.11)μ(∂Ω∖∪nFn)=0.
An increasing sequence (Fn) of closed sets satisfying condition (2.10) will be called a generalized nest; if further each Fn is compact, we call it a generalized compact nest.
We denote by S(∂Ω) the family of all smooth measures. The class S(∂Ω) is quiet large and it contains all positive Radon measures on ∂Ω charging no set of zero relative capacity. There exist also, by [15, Theorem 1.1], a smooth measure μ on ∂Ω (hence singular with respect to m) “nowhere Radon” in the sense that μ(G)=∞ for all nonempty relatively open subset G of ∂Ω (see [15, Example 1.6]).
The following Theorem, says that, any measure in S(∂Ω) can be approximated by measures in S0(∂Ω) and in S00(∂Ω) as well.
Theorem 2.3.
The following conditions are equivalent for a positive Borel measure μ on ∂Ω.
μ∈S(∂Ω).
There exists a generalized nest (Fn) satisfying (2.11) and 1Fn·μ∈S0(∂Ω) for each n.
There exists a generalized compact nest (Fn) satisfying (2.11) and 1Fn·μ∈S00(∂Ω) for each n.
2.3. Additive Functionals
Now we turn our attention to the correspondence between smooth measures and additive functionals, known as Revuz correspondence. As the support of an additive functional is the quasisupport of its Revuz measure, we restrict our attention, as for smooth measures, to additive functionals supported by ∂Ω. Recall that as the Dirichlet form (ℰ,ℱ) is regular, then there exists a Hunt process M=(Ξ,Xt,ξ,Px) on Ω¯ which is m-symmetric and associated with it.
Definition 2.4.
A function A:[0,+∞[×Ξ→[-∞,+∞] is said to be an additive functional (AF) if
At is ℱt-measurable,
there exist a defining set Λ∈ℱ∞ and an exceptional set N⊂∂Ω with capΩ¯(N)=0 such that Px(Λ)=1, for all x∈∂Ω∖N, θtΛ⊂Λ, for all t>0; for all ω∈Λ, A0(ω)=0; |At(ω)|<∞ for t<ξ. A(ω) is right continuous and has left limit, and At+s(ω)=At(ω)+As(θtω)s, t≥0.
An additive functional is called positive continuous (PCAF) if, in addition, At(ω) is nonnegative and continuous for each ω∈Λ. The set of all PCAF’s on ∂Ω is denoted 𝒜c+(∂Ω).
Two additive functionals A1 and A2 are said to be equivalent if for each t>0, Px(At1=At2)=1r.q.ex∈Ω¯.
We say that A∈𝒜c+(∂Ω) and μ∈S(∂Ω) are in the Revuz correspondence, if they satisfy, for all γ-excessive function h, and f∈ℬ+(Ω¯), the relation
(2.12)limt↘01tEh·m[∫0tf(Xs)dAs]=∫∂Ωh(x)(f·μ)(dx).
The family of all equivalence classes of 𝒜c+(∂Ω) and the family S(∂Ω) are in one to one correspondence under the Revuz correspondence. In this case, μ∈S(∂Ω) is called the Revuz measure of A.
Example 2.5.
We suppose Ω to be bounded with Lipschitz boundary. We have [2]
(2.13)limt↘01tEh·m[∫0tf(Xs)dLs]=12∫∂Ωh(x)f(x)σ(dx),
where Lt is the boundary local time of the reflecting Brownian motion on Ω¯. It follows that (1/2)σ is the Revuz measure of Lt.
In the following we give some facts useful in the proofs of our main results. We set
(2.14)UAαf(x)=Ex[∫0∞e-αtf(Xt)dAt],RαAf(x)=Ex[∫0∞e-αte-Atf(Xt)dt],Rαf(x)=Ex[∫0∞e-αtf(Xt)dt].
Proposition 2.6.
Let μ∈S0(∂Ω) and A∈𝒜c+(∂Ω), the corresponding PCAF. For α>0, f∈ℬb+, UAα is a relatively quasicontinuous version of Uα(f·μ).
Proposition 2.7.
Let A∈𝒜c+(∂Ω), and f∈ℬb+, then RαA is relatively quasicontinuous and
(2.15)RαAf-Rαf+UAαRαAf=0.
In general, the support of an AF A is defined by
(2.16)supp[A]={x∈X∖N:Px(R=0)=1},
where R(ω)=inf{t>0:At(ω)≠0}.
Theorem 2.8.
The support of A∈𝒜c+(∂Ω) is the relative quasisupport of its Revuz measure.
In the following, we give a well-known theorem of decomposition of additive functionals of finite energy. We will apply it to get a decomposition of the diffusion process associated with (ℰ,ℱ).
Theorem 2.9.
For any u∈ℱ, the AF A[u]=u~(Xt)-u~(X0) can be expressed uniquely as
(2.17)u~(Xt)-u~(X0)=M[u]+N[u],
where Mt[u] is a martingale additive functional of finite energy and Nt[u] is a continuous additive functional of zero energy.
A set σ(u) is called the (0)-spectrum of u∈ℱ, if σ(u) is the complement of the largest open set G such that ℰ(u,v) vanishes for any v∈ℱ∩𝒞0(X) with supp[v]⊂G. The following Theorem means that supp[N[u]]⊂σ(u), for all u∈ℱ.
Theorem 2.10.
For any u∈ℱ, the CAF N[u] vanishes on the complement of the spectrum F=σ(u) of u in the following sense:
(2.18)Px(Nt[u]=0:∀t<σF)=1r.q.ex∈X.
3. General Reflecting Brownian Motion
Now we turn our attention to the process associated with the regular Dirichlet form (ℰ,ℱ) on L2(Ω) defined by
(3.1)ℰ(u,v)=∫Ω∇u∇vdx,ℱ=H1~(Ω).
Due to the theorem of Fukushima (1975), there is a Hunt process (Xt)t≥0 associated with it. In addition, (ℰ,ℱ) is local, thus the Hunt process is in fact a diffusion process (i.e., A strong Markov process with continuous sample paths). The diffusion process M=(Xt,Px) on Ω¯ is associated with the form ℰ in the sense that the transition semigroup ptf(x)=Ex[f(Xt)], x∈Ω¯ is a version of the L2-semigroup 𝒫tf generated by ℰ for any nonnegative L2-function f.
M is unique up to a set of zero relative capacity.
Definition 3.1.
We call the diffusion process on Ω¯ associated with (ℰ,ℱ) the general reflecting Brownian motion.
The process Xt is so named to recall the standard reflecting Brownian motion in the case of bounded smooth Ω, as the process associated with (ℰ,H1(Ω)). Indeed, when Ω is bounded with Lipschitz boundary, we have that H~1(Ω)=H1(Ω) and by [6] the reflecting Brownian motion Xt admits the following Skorohod representation:
(3.2)Xt=x+Wt+12∫0tν(Xs)dLs,
where W is a standard N-dimensional Brownian motion, L is the boundary local (continuous additive functional) associated with surface measure σ on ∂Ω, and ν is the inward unit normal vector field on the boundary.
For a general domain, the form (ℰ,H1(Ω)) needs not to be regular. Fukushima [9] constructed the reflecting Brownian motion on a special compactification of Ω, the so-called Kuramuchi compactification. In [6], it is shown that if Ω is a bounded Lipschitz domain, then the Kuramochi compactification of Ω is the same as the Euclidean compactification. Thus for such domains, the reflecting Brownian motion is a continuous process who does live on the set Ω¯.
Now, we apply a general decomposition theorem of additive functionals to our process M, in the same way as in [6]. According to Theorem 2.9, the continuous additive functional u~(Xt)-u~(X0) can be decomposed as follows:
(3.3)u~(Xt)-u~(X0)=Mt[u]+Nt[u],
where Mt[u] is a martingale additive functional of finite energy and Nt[u] is a continuous additive functional of zero energy.
Since (Xt)t≥0 has continuous sample paths, Mt[u] is a continuous martingale whose quadratic variation process is
(3.4)〈M[u],M[u]〉t=∫0t|∇u|2(Xs)ds.
Instead of u, we take coordinate function ϕi(x)=xi. We have
(3.5)Xt=X0+Mt+Nt.
We claim that Mt is a Brownian motion with respect to the filtration of Xt. To see that, we use Lévys criterion. This follows immediately from (3.2), which became in the case of coordinate function
(3.6)〈M[ϕi],M[ϕi]〉=δijt.
Now we turn our attention to the additive functional Nt. Two natural questions need to be answered. The first is, where is the support of Nt located and the second concern the boundedness of its total variation.
For the first question we claim the following.
Proposition 3.2.
The additive functional Nt is supported by ∂Ω.
Proof.
Following Theorem 2.10, we have that supp[Nt]⊂σ(ϕ), where σ(ϕ) is the (0)-spectrum of ϕ, which means the complement of the largest open set G such that ℰ(ϕi,v)=0 for all v∈ℱ∩Cc(Ω¯) with supp[v]⊂G.
Step 1. If Ω is smooth (Bounded with Lipschitz boundary, e.g.), then we have
(3.7)ℰ(ϕi,v)=-∫∂Ωv·nidσ.
Then, ℰ(ϕi,v)=0 for all v∈ℱ∩Cc(Ω¯) with supp[v]⊂Ω. We can then see that the largest G is Ω. Consequently σ(ϕ)=Ω¯∖Ω and then σ(ϕ)=∂Ω.
Step 2. If Ω is arbitrary, then we take an increasing sequence of subset of Ω such that ⋃n=0∞Ωn=Ω. Define the family of Dirichlet forms (ℰΩn,ℱΩn) to be the parts of the form (ℰ,ℱ) on each Ωn as defined in Section 4.4 of [17]. By Theorem 4.4.5 in the same section, we have that ℱΩn⊂ℱ and ℰΩn=ℰ on ℱΩn×ℱΩn. We have that Ωn is the largest open set such that ℰΩn(ϕi,v)=0 for all v∈ℱΩn∩Cc(Ωn¯). By limit, we get the result.
The interest of the question of boundedness of total variation of Nt appears when one needs to study the semimartingale property and the Skorohod equation of the process Xt of type 3.2. Let |N| be the total variation of Nt, that is,
(3.8)|N|t=sup∑i=1n-1|Nti-Nti-1|,
where the supremum is taken over all finite partition 0=t0<t1<⋯<tn=t, and |·| denote the Euclidian distance. If |N| is bounded, then we have the following expression:
(3.9)Nt=∫0tνsd|N|s,
where ν is a process such that |ν|s=1 for |N|-almost all s.
According to § 5.4. in [17], we have the following result.
Theorem 3.3.
Assume that Ω is bounded and that the following inequality is satisfied:
(3.10)|∫Ω∂v∂xidx|≤C∥v∥∞,∀v∈H~1(Ω)∩Cb(Ω¯),
for some constant C. Then, Nt is of bounded variation.
A bounded set verifying (2.10) is called strong Caccioppoli set. This notion is introduced in [8], and is a purely measured theoretic notion. An example of this type of sets is bounded sets with Lipschitz boundary.
Theorem 3.4.
If Ω is a Caccioppoli set, then there exist a finite signed smooth measure ν such that
(3.11)∫Ω∂v∂xidx=-∫∂Ωvdμ,∀v∈H~1(Ω)∩Cb(Ω¯).
And ν=ν1-ν2 is associated with the CAF -Nt=-At1+At2 with the Revuz correspondence. Consequently ν charges no set of zero relative capacity.
To get a Skorohod type representation, we set
(3.12)ν=∑i=1N|μi|,ϕi=dμidνi=1,…,N.
We define the measure σ on ∂Ω by
(3.13)σ(dx)=2(∑i=1N|ϕi(x)|2)1/2ν(dx)
and the vector of length 1 at x∈∂Ω by
(3.14)ni(x)={ϕi(x)(∑i=1N|ϕi(x)|2)1/2if∑i=1N|ϕi(x)|2>0;0if∑i=1N|ϕi(x)|2=0.
Thus, μi(dx)=(1/2)ni(x)σ(dx), i=1,…,N.
Then
(3.15)Nt=∫0tn(Xs)dLs,
where L is the PCAF associated with (1/2)σ.
Theorem 3.5.
If Ω is a Caccioppoli set, then for r.q.e x∈Ω¯, one has
(3.16)Xt=x+Bt+∫0tn(Xs)dLs,
where B is an N-dimensional Brownian motion, and L is a PCAF associated by the Revuz correspondence to the measure (1/2)σ.
Remark 3.6.
The above theorem can be found in [9, 24]. In particular, Fukushima proves an equivalence between the property of Caccioppoli sets and the Skorohod representation.
4. Probabilistic Solution to General Robin Boundary Value Problem
This section is concerned with the probabilistic representation to the semigroup generated by the Laplacian with general Robin boundary conditions, which is, actually, obtained by perturbing the Neumann boundary conditions by a measure. We start with the regular Dirichlet form defined by (3.1), which we call always as the Dirichlet form associated with the Laplacian with Neumann boundary conditions.
Let μ be a positive Radon measure on ∂Ω charging no set of zeo relative capacity. Consider the perturbed Dirichlet form (ℰμ,ℱμ) on L2(Ω) defined by
(4.1)ℱμ=ℱ∩L2(∂Ω,μ),ℰμ(u,v)=ℰ(u,v)+∫∂Ωuvdμu,v∈ℱμ.
We will see in the following theorem that the transition function
(4.2)𝒫tμf(x)=Ex[f(Xt)e-Atμ]
is associated with (ℰμ,ℱμ), where Atμ is a positive additive functional whose Revuz measure is μ; note that the support of the AF is the same as the relative quasisupport of its Revuz measure.
Proposition 4.1.
𝒫tμ is a strongly continuous semigroup on L2(Ω).
Proof.
The proof of the above proposition can be found in [14].
Theorem 4.2.
Let μ be a positive Radon measure on ∂Ω charging no set of zero relative capacity and (Atμ)t≥0 be its associated PCAF of (Xt)t≥0. Then 𝒫tμ is the strongly continuous semigroup associated with the Dirichlet form (ℰμ,ℱμ) on L2(Ω).
Proof.
To prove that 𝒫tμ is associated with the Dirichlet form (ℰμ,ℱμ) on L2(Ω), it suffices to prove the assertion
(4.3)RαAf∈ℱμ,ℰαμ(RαA,u)=(f,u),f∈L2(Ω,m),u∈ℱμ.
Since ||RαAf||L2(Ω)≤||Rαf||L2(Ω)≤(1/α)||f||L2(Ω), we need to prove (4.7) only for bounded f∈L2(Ω). We first prove that (4.7) is valid when μ∈S00(∂Ω). According to Proposition 2.7 we have
(4.4)RαAf-Rαf+UAαRαAf=0,α>0,f∈ℬ+(Ω¯).
If μ∈S00(∂Ω), and if f is bounded function in L2(Ω), then ||Rαf||<∞, and UAαRαAf is a relative quasicontinuous version of the α-potential Uα(RαAf·μ)∈ℱ by Proposition 2.6. Since ||Uα(RαAf·μ)||∞≤||RαAf||∞||Uαμ||∞<∞ and μ(∂Ω)<∞, we have that
(4.5)RαAf=Rαf-UAαRαAf∈ℱμ
and that
(4.6)ℰα(RαAf,u)=ℰα(Rαf,u)-ℰα(UAαRαAf,u)=(f,u)-(RαAf,u)μ,u∈ℱμ,then (4.7) follows.
For general positive measure μ charging no set of zero relative capacity, we can take by virtue of Theorem 2.3 and Lemma 2.2 an increasing sequence (Fn) of generalized nest of ∂Ω, and μn=1Fn·μ∈S00(∂Ω). Since μ charges no set of zero relative capacity,μn(B) increases to μ(B) for any B∈ℬ(∂Ω).
Let An=1Fn·A. Then An is a PCAF of Xt with Revuz measure μn. Since μn∈S00(∂Ω) we have for f∈L2(Ω):
(4.7)RαAnf∈ℱμn,ℰαμn(RαAn,u)=(f,u),f∈L2(Ω,m),u∈ℱμn.
Clearly |RαAnf|≤Rα|f|<∞ r.q.e, and hence limn→+∞RαAnf(x)=RαAf(x) for r.q.e x∈Ω¯. For n<m, we get from (4.8)
(4.8)ℰαμn(RαAnf-RαAmf,RαAnf-RαAmf)≤(f,RαAnf-RαAmf),
which converges to zero as n,m→+∞. Therefore, (RαAnf)n is ℰ1-convergent in ℱ and the limit function RαAf is in ℱ~. On the other hand, we also get from (4.8)
(4.9)∥RαAnf∥L2(∂Ω,μ)≤(f,RαAnf)L2(Ω)≤(1α)∥f∥L2(Ω).
And by Fatou’s lemma: ||RαAf||L2(Ω)≤(1/α)||f||L2(Ω), getting RαAf∈ℱμ. Finally, observe the estimate
(4.10)|(RαAnf,u)μn-(RαAf,u)μ|≤∥RαAnf-RαAf∥L2(∂Ω,μn)∥u∥L2(∂Ω,μ)+|(Rαf,u)μ-μn|
holding for u∈L2(∂Ω,μ). The second term of the right-hand side tends to zero as n→+∞. The first term also tends to zero because we have from (4.8) ||RαAnf-RαAmf||L2(∂Ω,μn)≤(f,RαAnf-RαAmf), and it suffices to let first m→+∞ and then n→+∞. By letting n→+∞ in (4.7), we arrive to desired equation (4.3).
The proof of Theorem 4.2 is similar to [17, Theorem 6.1.1] which was formulated in the first time by Albeverio and Ma [14] for general smooth measures in the context of general (X,m). In the case of X=Ω¯, and working just with measures on S0(∂Ω), the proof still the same and works also for any smooth measure concentrated on ∂Ω. Consequently, the theorem is still verified for smooth measures “nowhere Radon,” that is, measures locally infinite on ∂Ω.
Example 4.3.
We give some particular examples of 𝒫tμ.
If μ=0, then
(4.11)𝒫t0f(x)=Ex[f(Xt)],
the semigroup generated by Laplacian with, Neumann boundary conditions.
If μ is locally infinite (nowhere Radon) on ∂Ω, then
(4.12)𝒫t∞f(x)=Ex[f(Bt)1{t<τ}],
the semigroup generated by the Laplacian with Dirichlet boundary conditions (see [13, Proposition 3.2.1]).
Let Ω be a bounded and enough smooth to insure the existence of the surface measure σ, and μ=β·σ, with β is a measurable bounded function on ∂Ω, then Atμ=∫0tβ(Xs)dLs, where Lt is a boundary local time. Consequently
(4.13)𝒫tμf(x)=Ex[f(Xt)exp(-∫0tβ(Xs)dLs)]
is the semigroup generated by the Laplacian with (classical) Robin boundary conditions given by (1.1).
The setting of the problem from the stochastic point of view and the stochastic representation of the solution of the problem studied are important on themselves and are new. In fact before there was always additional hypothesis on the domain or on the class of measures. Even if our approach is inspired by the works [14, 15] and Chapter 6 of [17], the link is not obvious and give as rise to a new approach to the Robin boundary conditions. As a consequence, the proof of many propositions and properties become obvious and direct. The advantage of the stochastic approach is, then, to give explicitly the representation of the semigroup and an easy access of it.
Proposition 4.4.
𝒫tμ is sub-Markovian, that is, 𝒫tμ≥0 for all t≥0, and
(4.14)∥𝒫tμf∥∞≤∥f∥∞(t≥0).
Proof.
It is clear that if f∈L2(Ω)+, then 𝒫tμf≥0 for all t≥0. In addition we have |𝒫tμf(x)|≤Ex[|f|(Xt)], and then ||𝒫tμf||∞≤||f||∞(t≥0).
Remark 4.5.
The analytic proof needs the first and the second BeurlingDeny criterion [11, Proposition 3.10] while our proof is obvious and direct.
Let Δμ be the self-adjoint operator on L2(Ω) generator of the semigroup 𝒫tμ, we write
(4.15)𝒫tμf(x)=e-tΔμf(x).
Following [13], we know that Δμ is a realization of the Laplacian. Then we call Δμ the Laplacian with General Robin boundary conditions.
Theorem 4.6.
Let μ∈S(∂Ω), then the semigroup 𝒫tμ is sandwiched between the semigroup of Neumann Laplacian, and the semigroup of Dirichlet Laplacian. That is
(4.16)0≤e-tΔD≤𝒫tμ≤e-tΔN
for all t≥0, in the sense of positive operators.
Proof.
Let f∈L2(Ω)+. Since Atμ≥0 we get easily the following: 𝒫tμf(x)≤Ex[f(Xt)] for any x∈Ω¯. On the other hand, we have 𝒫tμf(x)≥Ex[f(Xt)e-Atμ1{t<σ∂Ω}], where σ∂Ω is the first hitting time of ∂Ω. Since the relative quasisupport of Atμ and Nt are in ∂Ω, then in {t<σ∂Ω}, Nt and Atμ vanishes. Consequently, Xt=Bt in {t<σ∂Ω} and 𝒫tμf(x)≥Ex[f(Bt)1{t<σ∂Ω}]. The theorem follows.
Remark 4.7.
The fact that the semigroup 𝒫tμ is sandwiched between the Neumann semigroup and the Dirichlet one as proved in [13, Theorem 3.4.1] is not obvious and needs a result characterizing the domination of positive semigroups due to Ouhabaz, while our proof is simple and direct.
Proposition 4.8.
Let μ,ν∈S(∂Ω) such that ν≤μ (i.e., ν(A)≤μ(A), for all A∈ℬ(∂Ω)), then
(4.17)0≤e-tΔD≤𝒫tμ≤𝒫tν≤e-tΔN
for all t≥0, in the sense of positive operators.
Proof.
It follows from the remark that if ν≤μ, then Atν≤Atμ, which means that (Atμ)μ is increasing, and then (𝒫tμ)μ is decreasing.
There exist a canonical Hunt process XtA possessing the transition function 𝒫tμ which is directly constructed from Xt by killing the paths with rate -dLt, where Lt=e-At.
To construct the process associated with 𝒫tμ, we follow A.2 of [17], so we need a nonnegative random variable Z(ω) on (Ξ,ℳ,Px) which is of an exponential distribution with mean 1, independent of (Xt)t≥0 under Px for every x∈Ω¯ satisfying Z(θs(ω))=(Z(ω)-s)∨0. Introducing now a Random time ξA defined by
(4.18)ξA=inf{t≥0:At≥Z}.
We define the process (XtA)t≥0 by
(4.19)XtA={Xtift<ξA;Δift≥ξA,
where Δ is a one-point compactification.
And, the admissible filtration of the process (XtA)t≥0 is defined by
(4.20)ℱtA={Λ∈ℱ∞:Λ∩{At<Z}=Λt∩{At<Z},∃Λt∈ℱt}.
Since {At<Z}∩{At=∞}=∅, we may and will assume that Λt⊃{At=∞}.
Now, we can write
(4.21)Ex[f(XtA)]=Ex[f(Xt):t<ξA]=Ex[f(Xt):At<Z]=Ex[f(Xt)e-At]=𝒫tμf(x).
The Hunt process (XtA)f≥0 is called the canonical subprocess of (Xt)t≥0 relative to the multiplicative functional Lt. In fact, (XtA)t≥0 is a Diffusion process as (ℰμ,ℱμ) is local.
In the literature, the Diffusion process XtA is called a partially reflected Brownian motion [25], in the sense that, the paths of Xt are reflected on the boundary since they will be killed (absorbed) at the random time ξA with rate -dLt.
Theorem 4.9.
Let μ,μn∈S(∂Ω) such that μn is monotone and converges setwise to μ, that is, μn(B) converges to μ(B) for any B∈ℬ(∂Ω), then Δμn converges to Δμ in strongly resolvent sense.
Proof.
We prove the theorem for μn increasing, the proof of the decreasing case is similar. Let An (resp. A) be the additive functional associated to μn (resp. μ) by the Revuz correspondence. Similarly to the second part of the proof of Theorem 4.2, we have limRαAnf(x)=RαAf(x) for r.q.e x∈Ω¯. Consequently limn→+∞||RαAnf-RαAf||L2(Ω)=0. For n<m, we have ℱμm⊂ℱμn, and then (4.22)ℰαμn(RαAnf-RαAmf,RαAnf-RαAmf)≤(f,RαAnf-RαAmf),
which converges to zero as n,m→+∞. Therefore, (RαAnf)n is ℰ1-convergent in ℱ and the limit function RαAf is in ℱ~. The result follows.
Corollary 4.10.
Let μ∈S(∂Ω) finite and let k∈ℕ*. We defined for u,v∈ℱμ:
(4.23)ℰμk(u,v)=∫Ω∇u∇vdx+1k∫∂Ωu~v~dμ,
then Δμk→ΔN in the strong resolvent sense.
Intuitively speaking, when the measure μ is infinity (locally infinite on the boundary), the semigroup 𝒫tμ is the Dirichlet semigroup as said in the example 2 in section 4, which means that the boundary became “completely absorbing,” and any other additive functional in the boundary cannot influence this phenomena, which explain why Nt does not appear yet in the decomposition of Xt, which means that the reflecting phenomena disappear, and so any path of Xt is immediately killed when it arrives to the boundary.
When μ is null on the boundary, then the semigroup 𝒫tμ is the Neumann one, and then the boundary became completely reflecting, but for a general measure μ the paths are reflected many times before they are absorbed at a random time.
Acknowledgments
This work was the result of a project that got the DAAD fellowship for a research stay in Germany. The author would like also to thank his thesis advisor Prof. Dr Omar El-Mennaoui for his guidance.
SatoK.UenoT.Multi-dimensional diffusion and the Markov process on the boundary196545296050198547PapanicolaouV. G.The probabilistic solution of the third boundary value problem for second order elliptic equations1990871277710.1007/BF012177461076956ZBL0688.60063RamasubramanianS.Reflecting Brownian motion in a Lipschitz domain and a conditional gauge theorem20016321781931897449ZBL1006.60079RenmingS. Probabilistic approach to the third boundary value problem19899185100BassR. F.BurdzyK.ChenZ.-Q.Uniqueness for reflecting Brownian motion in lip domains200541219723510.1016/j.anihpb.2004.06.0012124641ZBL1067.60036BassR. F.HsuP.The semimartingale structure of reflecting Brownian motion199010841007101010.2307/20479601007487ZBL0694.60075BassR. F.HsuP.Some potential theory for reflecting Brownian motion in Hölder and Lipschitz domains1991192486508110627210.1214/aop/1176990437ZBL0732.60090ChenZ. Q.FitzsimmonsP.J.WilliamsR. J.Quasimartingales and strong Caccioppoli set19932281315FukushimaM.A construction of reflecting barrier Brownian motions for bounded domains196741832150231444ZBL0317.60033HsuE. P.1984Stanford UniversityArendtW.WarmaM.The Laplacian with Robin boundary conditions on arbitrary domains200319434136310.1023/A:10241816088631988110ZBL1028.31004ArendtW.WarmaM.Dirichlet and Neumann boundary conditions: what is in between?20033111913510.1007/s0002803000051977430ZBL1044.31005WarmaM.2002University of UlmAlbeverioS.MaZ. M.Perturbation of Dirichlet forms—lower semiboundedness, closability, and form cores199199233235610.1016/0022-1236(91)90044-61121617ZBL0743.60071AlbeverioS.MaZ. M.Additive functionals, nowhere Radon and Kato class smooth measures associated with Dirichlet forms19922922472651173989ZBL0786.60104BlanchardPh.MaZ. M.New results on the Schrödinger semigroups with potentials given by signed smooth measures19901444Berlin, GermanySpringer21324310.1007/BFb00836181078851ZBL0717.60075FukushimaM.ŌshimaY.TakedaM.199419Berlin, GermanyWalter de Gruyterx+39210.1515/97831108897411303354KatoT.1995Berlin, GermanySpringerxxii+619Classics in Mathematics1335452SimonB.Schrödinger semigroups19827344752610.1090/S0273-0979-1982-15041-8670130ZBL0524.35002StollmannP.Smooth perturbations of regular Dirichlet forms1992116374775210.2307/21594421107277ZBL0768.31009StollmannP.VoigtJ.Perturbation of Dirichlet forms by measures19965210913810.1007/BF003967751378151ZBL0861.31004VoigtJ.Absorption semigroups1988201117131972185ZBL0702.47024VoigtJ.Absorption semigroups, their generators, and Schrödinger semigroups198667216720510.1016/0022-1236(86)90036-4845197ZBL0628.47027FukushimaM.Dirichlet forms, Caccioppoli sets and the Skorohod equation199723Boston, Mass, USABirkhäuser59661636827ZBL0885.60064GrebenkovD. S.Partially reflected Brownian motion: a stochastic approach to transport phenomena2006New York, NY, USANova Science Publishers1351692553673