On approximation of a Dirichlet problem for divergence form operator by Robin problems ∗

We show that under natural assumptions solutions of Dirichlet problems for uniformly elliptic divergence form operator can be approximated pointwise by solutions of some versions of Robin problems. The proof is based on stochastic representation of solutions and properties of reﬂected diﬀusions corresponding to divergence form operators.


Introduction
Let D be a bounded Lipschitz domain in R d , d ≥ 3, and L be the divergence form operator with measurable coefficients a ij : D → R such that for some Λ ≥ 1.For f : D → R, g : ∂D → R and n ≥ 1 we consider the following boundary-value problem where a = {a ij } 1≤i,j≤d and n(x) is the inward unit normal at x ∈ ∂D.Note that (1.2) is a particular version of Robin problem (also known as Fourier problem, Newton problem or the third boundary-value problem).It is known (see, e.g., [3, Appendix 1, Section 4.4]) that if f ∈ L 2 (D), g ∈ H 1 (D), then for each n ≥ 1 there exists a unique weak solution of (1.2) and u n → u in H 1 (D) as n → ∞, where u is the unique weak solution of the Dirichlet problem As noted in [3, p. 360], this approximation result is of practical interest, because many numerical schemes for solving boundary value problems (Dirichlet, Neumann, Neumann-Dirichlet) for L consist in computing the solution of a discrete version of (1.2) or its modification with large n.
If f ∈ L p (D) with p > d and g ∈ H 1 (D) ∩ C(∂D), then u n , u have continuous versions and one may ask whether u n → u for every x ∈ D. In this note, we give positive answer to this questions.Our proof is quite simple and is based on stochastic representation of solutions of (1.2), (1.3).But let us stress that in the proof of our convergence results we use deep results from [5,6] (see also [2] for the case L = (1/2)∆) saying that one can construct a process M on D (reflected diffusion) associated with L with a strong Feller resolvent.In fact, in these papers the strong Feller property is proved for less regular and possibly unbounded domains.

Preliminaries
In the paper, We assume that the matrix a satisfies (1.1) and consider the Dirichlet form (E, D(E)) on L 2 (D) defined by where H 1 (D) is the usual Sobolev space of order 1, and for λ > 0 set E λ (u, v) = E(u, v) + λ(u, v), where (•, •) is the usual inner product in L 2 (D; m).We denote by (T t ) t>0 the strongly continuous semigroup of Markovian symmetric operators on L 2 ( D) associated with E (see [4,Section 1.3]).
In the paper, we define quasi-notions (exceptional sets, quasi-continuity) with respect to (E, H 1 (D)).We will say that a property of points in D holds quasi everywhere (q.e. in abbreviation) if it holds outside some exceptional set.It is known (see [4, Lemm 2.1.4,Theorem 2.1.3])that each element of H 1 (D) admits a quasi-continuous m-version, which we denote by ũ, and ũ i q.e.unique for every u ∈ H 1 (D).
In [6, Theorems 2.1 and 2.2] (see also [5]) it is proved that there exists a conservative diffusion proces M = {(X, P x ), x ∈ D} on D associated with the Dirichlet form (2.1) in the sense that the transition density of M defined as has the property that where P t f (x) = D f (y)p t (x, dy) = E x f (X t ).Furthermore, (P t ) t>0 is strongly Feller in the sense that P t (B b ( D)) ⊂ C( D) and lim t↓0 P t f (x) = f (x) for x ∈ D, f ∈ C( D).In particular (see [4,Exercise 4.2.1]), the transition density satisfies the following absolute continuity condition: p t (x, •) ≪ m for any t > 0, x ∈ D.
We denote by (R α ) α>0 the resolvent associated with M (or with (P t ) t>0 ), that is For a Borel measure µ on D we also set whenever the integral makes sense.Let σ denote the surface measure on ∂D.By [6, Lemma5.1,Theorem 5.1], σ belongs to the space of smooth measures in the strict sense, and hence, by [4,Theorem 5.1.7],there is a unique positive continuous additive functional of M in the strict sense with Revuz measure σ.In what follows we denote it by A. Note that for any Indeed, by [4,Theorem 5.1.3]the above equality holds for m-a.e.x ∈ D, and hence for every x ∈ D, because p t satisfies the absolute continuity condition and for any nonnegative g ∈ B b ( D) both sides of the above equality are α-excessive functions.Also note that the support of A is contained in ∂D.Hence (for more details see the beginning of the proof of Lemma 4.1).It follows that in fact the right-hand side of (2.2) is well defined for g ∈ B b (∂D).
Remark 2.1.If, in addition, . ., X d ) has the following Skorohod representation: for i = 1, . . ., d and every x ∈ D where M i are martingale additive functionals in the strict sense with covariations and In case of the classical Dirichlet form defined by i.e. if a = 1 2 I, the process M is called a reflecting Brownian motion.By Lévy's characterization of Brownian motion, representation (2.3) reads where B = (B 1 , . . ., B d ) is a standard Brownian motion.For the proof of (2.4) see [4, Example 5.2.2] and for the general case (2.3) see [6,Theorem 2.3].In case a is a general function satisfying (1.2) some representation of X (Lyon's-Zheng-Skorohod decomposition) is given in [13] (for bounded C 2 domain D and x ∈ D). Let where ∂ is a point adjoined to D as an isolated point (cemetery state).We adopt the convention that every function We denote by M λ the canonical subprocess of M with respect to the multiplicative functional e −λt .For its detailed construction we refer to [4,Section A.2].Here let us only note that we may assume that M λ = (X λ , P x ) is defined on the same probability space on which M is defined and where Z is a nonnegative random variable independent of (X t ) t≥0 having exponential distribution with mean 1.

Weak and probabilistic solutions
For the convenience of the reader, below we recall variational formulation of problems (1.2), (1.3).For more details and comments we refer to [3, Appendix I].
The existence and uniqueness of weak solutions of (1.2), (1.3) is well known.For proofs by classical variational methods we refer for instance to [3, Appendix I].In Proposition 3.2 below we give proofs by using the probabilistic potential theory.The advantage of using these less classical methods lies in the fact that they provide probabilistic representations of quasi-continuous versions of weak solutions.We would like to stress that the proof of Proposition 3.2 is simply a compilation of known facts.We provide it for completeness and later use.
Then there exists a unique weak solution u n of (1.2) and ũn defined q.e. on D by Then there exists a unique weak solution u of (1.3) and ũ defined q.e. on D by Proof.(i) Let (E nσ , D(E nσ )) denote the form E perturbed by the measure nσ, that is By the classical trace theorem, D(E nσ ) = H 1 (D), so u n is a weak solution of (3.1) if and only if u n ∈ D(E nσ ) and Therefore we have to show that there is a unique u n ∈ H 1 (D) satisfying (3.4).Suppose that u 1 n , u 2 n ∈ H 1 (D) satisfy (3.4) and let u = u 1 n − u 2 n .Then from (3.4) with test function v = u we get E nσ λ (u, u) = 0, hence that E λ (u, u) = 0. Clearly, this implies that u = 0 m-a.e.To prove the existence and its representation, it suffices to note that ũn can be written in the form ũn = R nA λ f + nU λ n,A g, where and then use [4, (6.1.5),(6.1.12)].Furthermore, ũn is quasi-continuous because R nA λ f is quasi-continuous by [4, Lemma 5.1.5]and U λ n,A g is quasi-continuous by [4, Lemma 6.1.3].(ii) With our convention, ũ can be equivalently written in the form the second equality being a consequence of [4, Theorem 4.4.1],Furthermore, ũ Therefore, under these assumptions on f and g, the integrals on the right-hand side of (3.2) are well defined for every x ∈ D. Similarly, the right-hand side of (3.3) is well defined for every x ∈ D.
The above remarks and Proposition 3.2 justify the following definition of probabilistic solutions of (1.2), (1.3).An equivalent definition of a probabilistic solution of (1.2), resembling (3.1), will be given in Proposition 3.4 below.
For a deep study of connections between probabilistic solutions, weak solutions as well of other kind of solutions to the Dirichlet problem with possibly irregular domain we refer the reader to [8].Here let us only note that if D is bounded and Lipschitz (as in the present paper), then it satisfies Poincare's cone condition.Therefore modifying slightly the proof of [1, Proposition II.1.13](we use Aronson's estimates for the transition densities of M) one can show that each point x ∈ ∂D is regular for D c , i.e.
Using this, similarly to the proof of [1,Proposition II.1.11],one can show that H λ ∂D g ∈ C( D) if g ∈ C(∂D).For an analytical proof of this well known fact see, e.g., [12].Furthermore, it is known (see [14,Section 9] or [11] is the probabilistic solution if and only if it satisfies the equation Proof.Define u n , ũn as in Proposition 3.2 and set By this and (3.1), E λ (w n , v) = E λ (u n , v), v ∈ H 1 (D), which implies that w n = u n m-a.e., and hence w n = ũn q.e. on D. From this and (3.7) it follows that w n is a continuous solution of (3.6).It is the probabilistic solution of (1.2).To see this, we first note that (3.6), with v n replaced by w n , can be equivalently written as dA t exist and are finite for each x ∈ D, in much the same way as in [9, Remark 3.3(ii)] we show that there is a martingale additive functional M such that for each x ∈ D the pair (Y n , M ), where Integrating by parts we get Hence Letting T → ∞ gives for every x ∈ D. This shows that v n is continuous and satisfies (3.6), and moreover, any continuous solution of (3.8) coincides with v n .
Note that (3.6) is a very special case of equation with smooth measure data and (3.9) is the corresponding backward stochastic differential equation (BSDE).More general, semilinear equations of the form (3.6), (3.9) are considered in [10].Note also that one can prove the existence of a quasi-continuous v n satisfying (3.6) for q.e.x ∈ D by solving the corresponding BSDE, i.e. by probabilistic methods (we do not need know in advance that there is a weak solution u n ).For a general result of this kind see [10,Theorem 4.3].

A convergence result
Recall that A is an additive functional (AF in abbreviation) of M in the strict sense with Revuz measure σ.We denote by F A the support of A, i.e.F A = {x ∈ D : P x (A t > 0 for all t > 0) = 1}.Lemma 4.1.P x (A t∧τ D = 0, t ≥ 0) = 1 and P x (A t+τ D > 0, t ≥ 0) = 1 for every x ∈ D.
Proof.In view of (3.5) the first part of the lemma is trivial for x ∈ ∂D.To show it for x ∈ D, we denote by F the quasi-support of σ.We may and will assume that F ⊂ ∂D (see [4, p. 190]).Since A is an AF in the strict sense, by [4,Lemma 5.1.11]we have [4,Theorem 5.1.5],F A = F , so P x (A t = (1 F • A) t , t > 0) = 1 for every x ∈ D. Since F ⊂ ∂D, it follows that for x ∈ D, A t = 0 P x -a.s. on [0, τ D ).Since A is continuous, in fact A t = 0 P x -a.s. on [0, τ D ] for x ∈ D, which proves the first part of the lemma.Let B be a standard Brownian motion appearing in (2.4).We have P y (τ D = 0) = 1 for y ∈ ∂D, where τD = inf{t > 0 : B t / ∈ D}.From this, (2.4) and the fact that the reflecting Brownian motion is a diffusion with sample paths in D it follows that the support of the additive functional appearing in (2.4), which we denote for the moment by Ā, equals ∂D.Let Cap L denote the capacity associated with E and Cap the capacity associated with D (see [4,   Proof.Recall that v n is defined by the right-hand side of (3.2).First assume that x ∈ D. By Lemma 4.1 and the dominated convergence theorem, for x ∈ D we have as n → ∞.We are going to show that for every as n → ∞.We know that (P t ) t>0 is a strongly Feller semigroup on C( D).Let ( L, D( L)) denote its generator.Since D( L) is dense in C( D), one can choose a sequence {g k } ⊂ D( L) such that sup x∈ D |g k −g| ≤ k −1 .By [7, Theorem 3.6.5],g k (X) is a semimartingale under P x for x ∈ D. In fact, is a martingale under P x for x ∈ D. Integrating by parts, for all k ≥ 1 and t ≥ 0 we obtain Since e −λt−nAt → 0 as t → ∞ and A τ D = 0 P x -a.s., we get is a bounded Lipschitz domain (for a definition see, e.g., [4, Exercise 5.2.2]),D = D ∪ ∂D.We denote by m or simply by dx the d−dimensional Lebesgue measure.B( D) is the set of Borel subsets of D, B b ( D) (resp.C( D)) is the set of bounded Borel (resp.continuous) functions on D. To shorten notation, we write L 2 (D) instead of L 2 (D; m) and L 2 (∂D) instead of L 2 (∂D; σ).

Definition 3 . 3 .
Let f ∈ L p (D) with p > d and g ∈ B b (∂D).The function v n : D → R defined by the right-hand side of (3.2) is called the probabilistic solution of (1.2).The function v : D → R defined by the right-hand side of (3.3) is called the probabilistic solution of (1.3).
Section 2.1] for the definitions).Assumption (1.1) implies that 2λ −1 Cap ≤ Cap L ≤ 2λCap.Therefore F is a quasisupport of σ considered as a smooth measure with respect to Cap L if and only if it is a quasi-support of σ considered as a smooth measure with respect to Cap.By what has already been proved and [4, Theorem 5.1.5],F = F Ā = ∂D, so by [4, Theorem 5.1.5]again, F A = ∂D.From this and the definition of F A we get the second part of the lemma.

Theorem 4 . 2 .
Assume that f ∈ L p (D) with p > d and g ∈ C(∂D).Then v n (x) → v(x) for every x ∈ D.

e 0 ee 0 e
−λs−nAs g k (X s ) dA s = E x e −λτ D g k (X τ D ) − λE x ∞ τ D e −λs−nAs g k (X s ) ds + E x ∞ τ D e −λs−nAs ( Lg k )(X s ) ds.Since g k , Lg k ∈ C( D), applying Lemma 4.1 and the dominated convergence theorem shows that the second and third term on the right-hand side of the above equality converge to zero as n → ∞.This proves thatnE x ∞ −λs−nAs g k (X s ) dA s → E x e −λτ D g k (X τ D ). −λs−nAs dA s ≤ nE x e −λτ D ∞ −nAs dA s ≤ e −λτ D (1 − e −A∞ ), so nE x ∞ τ D e −λs−nAs |g k − g|(X s ) dA s ≤ k −1 E x e −λτ D .(4.4)Clearly, we also haveE x e −λτ D |g k − g|(X τ D ) ≤ k −1 .(4.5)From (4.3)-(4.5)we get (4.2), which together with (4.1) shows the desired convergence for x ∈ D. Since P x (τ D = 0) = 1 for x ∈ ∂D, the above arguments also show thatv n (x) → E x g(X 0 ) = g(x) = v(x) for x ∈ ∂D, which completes the proof.Remark 4.3.(i) Let f ∈ L 2 (D), g ∈ C(∂D)and ũn , ũ be defined as in Proposition 3.2.Then ũn → u q.e. because the proof of Theorem 4.2 shows that then (4.1) holds for q.e.x ∈ D and (4.2) holds for every x ∈ D. In particular, if f ∈ L 2 (D) and g ∈ H 1 (D) ∩ C(∂D), then {u n } converges q.e. to the weak solution u of (1.3).If f ∈ L 2 (D), g ∈ H 1 (D), then the convergence holds in H 1 (D) and hence a.e.For an analytical proof of this fact we refer the reader to [3, Appendix 1, Section 4.4].(ii) In [3] some results on the rate of convergence of {u n } to u in the norm of H 1 (D) are given.Estimating the rate of pointwise convergence of {v n } to v presents a more delicate open problem L p (D) with p > d, then R λ |f | ∈ C( D) by [6, Theorem 2.1], and if g .7)By the remarks following the proof of Proposition 3.2, w n (x) is well defined and finite for each x ∈ D.Moreover, there is C > 0 such that |ũ n | ≤ C q.e.Since σ is smooth, |ũ n | ≤ C σ-a.e. on ∂D.From this and [4, Theorem 2.1] it follows that in fact w n ∈ C( D).For every v ∈ H 1 (D) we have