Week 3: Approximation by Ck functions up to the boundary
We now take up the question of whether u∈Hk,p(Ω) may be approximated by a function v∈Ck(Ω), where we assume Ω is bounded. It turns out that this is possible provided Ω is sufficiently regular.
Definition 3.1. An open bounded set Ω⊂Rn is said to be a domain of class Ck if for every x∈∂Ω, there exists an open neighbourhood Nx∋x and a Ck-diffeomorphism Φx:Nx→B(0,1)⊂Rn with the following properties:
Φx(x)=0;
Φx(∂Ω∩Nx)=B(0,1)∩{xn=0}; and
Φx(Ω∩Nx)=B(0,1)∩{xn>0}.
In a certain sense, the approximation properties of Sobolev spaces on domains of class Ck may be deduced from those of Sobolev spaces on the upper half-ballB+(0,r):=B(0,r)∩{xn>0}. We establish this in a sequence of lemmas.
Lemma 3.2. Suppose {Ui}i=1N is a collection of open sets such that Ω⊂⋃i=1NUi and let Ωi:=Ui∩Ω. If u∈Hk,p(Ωi) for all i, then u∈Hk,p(Ω).
Proof. Let {ψi}i=1N be a partition of unity subordinate to the cover {Ui}. Now, u∈Hk,p(Ωi) iff there is a sequence {uji}j=1∞⊂Ck,p(Ω) with ujij→∞u in Wk,p(Ωi). Set
uj:=i=1∑suji⋅ψi.
Clearly uj∈Ck,p(Ω), and since we may write u=∑i=1su⋅ψi, by the triangle inequality
where we have used the Leibniz rule as in Theorem 2.11.
Remark 3.3. Lemma 3.2 may be viewed as the converse to the fact that u∈Hk,p(Ω)⇒u∈Hk,p(V) for any V⊂Ω.
Lemma 3.4. Let U⊂Rn be open and bounded such that ∂U=Γ1∪Γ2, Γ1∩Γ2=∅ and γ1=V×{0}, V⊂Rn−1 open. If u∈Hk,p(U), then for any A⊂U with A⊂U∪Γ1, there exists a sequence {um}⊂C∞(A) such that umm→∞u in Wk,p(A).
Proof. For u∈Lp(U) and ε>0, define the mapping Jε′u:A→K by
where as before we assume u≡0 outside U. Note that z↦ρ(εx+2εen−z) is supported in B(x+2εen,ε) so that it does not intersect Γ1 and is compactly contained in Ω for ε<31dist(A,Γ2).
Jε′u possesses the basic properties of Jεu of Theorem 1.1, specifically:
Jε′u∈C∞(A): Since the support of z↦ρ(εx+2εen−εz) is compact and this function is smooth, we may apply the dominated convergence theorem to interchange integral and limit to show smoothness.
∣∣Jε′u∣∣p,A≤∣∣u∣∣p,U: Exactly as before, we estimate using Hölder's inequality:
If ε<31dist(A,Γ2), u∈Wk,p(U) and ∣α∣≤k, we have DαJε′u=Jε′Dαu on A: As remarked earlier, for such ε we have that supp(z↦ρ(εx+2εen−z))⋐U so that by `integrating by parts' as in the case of Jεu,
Combining all of the above, we see that if u∈Wk,p(U), then {um:=J1/m′u} is a sequence with the desired property.
Corollary 3.5. If Ω is a domain of class Ck, then Hk,p(Ω)=Ck(Ω).
Proof. For each x∈∂Ω, let Φx be as in the definition of a domain of class Ck and set Ωδ:={x∈Ω:dist(x,∂Ω)>δ}. Clearly
Ω⊂x∈∂Ω⋃Φx−1(B(0,21))∪δ>0⋃Ωδ.
By compactness, we may find {xi}i=1N⊂∂Ω and a single δ>0 such that
Ω⊂Ωδ∪i=1⋃NΦxi−1(B(0,21)).
Let {ψ0,ψ1,…,ψN} be a partition of unity subordinate to this cover and v∈Ck,p(Ω). Define
vm:=ψ0⋅v+i=1∑Nψi⋅J1/m′(v∘Φxi−1)∘Φxi,
where we take U:=B+(0,1) and A:=B+(0,21) in the definition of J1/m′(v∘Φxi−1). Clearly vm∈Ck(Ω), since J1/m′(v∘Φxi−1)∈C∞(B+(0,21)). Using the same tricks as before, we compute that
Now, noting that ∣detDΦ∣ is bounded from below by a positive constant, using the chain rule and liberally bounding from above, we may estimate this expression from above:
since we also have that v∘F−1∈Ck,p(A). Therefore, for any ε>0, we may choose m large enough so that ∣∣vm−v∣∣<ε.
Another useful characterisation of Hk,p(Ω) functions is in terms of compactly supported functions. We first establish the following lemmas.
Lemma 3.6. There exists a constant C>0 depending on k and p such that for any u∈Ck(B+(0,r)), we may find a u′∈Ck(B(0,r)) such that u′∣B+(0,r)≡u and ∣∣u′∣∣k,p≤C⋅∣∣u∣∣k,p.
where the {cj}j=1k+1 satisfy ∑j=1k+1cj⋅(−j1)m−1=1 for all m∈{1,…,k+1}. It follows that this relation uniquely determines the {cj} and ensure that u′ is Ck across {xn=0}. Moreover, it is clear that for each α with ∣α∣≤k, ∣∣Dαu′∣∣p≤C(k,p)∣∣Dαu∣∣p, which establishes the claim.
Lemma 3.7. Suppose Ω is a domain of class Ck. There exists a constant C>0 depending on k, p and Ω such that for each u∈Ck(Ω), we may find a u′∈C0k(Rn) with u′∣Ω≡u∣Ω and
∣∣u′∣∣k,p≤C⋅∣∣u∣∣k,p.
Proof. As before, we cover Ω⊂Ωδ∪⋃i=1NΦxi−1(B(0,21)) and let {ψ0,…,ψN} be a subordinate partition of unity. Note that for each i, u∘Φxi−1∈Ck(B+(0,21)). By the preceding lemmas, we may extend this function to a vi∈Ck(B(0,21)) such that ∣∣vi∣∣k,p≤C(k,p)⋅∣∣u∣∣k,p. Now let
u′=ψ0⋅u+i=1∑Nψi⋅vi∘Φxi.
It immediately follows that u′ possesses the desired properties.
Remark 3.8. Let Ω be a domain of class Ck and V⋐Rn an open set such that Ω⋐V and let ψ:Rn→[0,1] be a smooth function such that ψ≡1 on Ω and suppψ⋐V. Using the same techniques as before, we may establish the estimate
∣∣ψ⋅u′∣∣k,p≤C(ψ,k,p)⋅∣∣u′∣∣k,p
for u′ as in Lemma 3.7. In particular, replacing u′ with ψ⋅u′, we deduce that for any u∈Ck(Ω) and p>1, there exists a u′∈C0k(Ω) such that u′∣Ω≡u∣Ω, suppu′⋐V and ∣∣u′∣∣k,p≤C(k,p,V)⋅∣∣u∣∣k,p. A cursory inspection of the proofs of Lemmas 3.6 and 3.7 show that the mapping
Ck(Ω)∋u↦u′∈C0k(V)
is linear, and the condition on the norm of u′ amounts to its continuity. We call this an extension operator.
Corollary 3.9. For any u∈Hk,p(Ω) and ε>0, there exist u1∈C0k(Ω) and u2∈C0∞(Ω) such that ∣∣u−u1∣∣k,p<ε and ∣∣u−u2∣∣k,p<ε.
Proof. By Corollary 3.5, for any u∈Hk,p(Ω) and ε>0, we may find a v∈Ck(Ω) such that ∣∣u−v∣∣k,p,Ω<2ε, and by Lemma 3.7 this function is the restriction to Ω by a function v′∈C0k(Rn), establishing the equality. To establish the second, consider the sequence {wm:=J1/mv′}⊂C0∞(Rn), the inclusion following from Theorem 1.1. By Remark 2.6, we may choose N>0 large enough so that ∣∣wN−v′∣∣k,p,Rn<2ε, which immediately implies that ∣∣wN−v′∣∣k,p,Ω<2ε. By the triangle inequality, we have that ∣∣u−wN∣∣k,p,Ω<ε, establishing the second equality.