Aspects of such problems that we wish to consider are:
Existence: Under what conditions can we guarantee the existence of solutions to (#)?
Regularity: How smooth can we expect u to be, either in the interior (Ω) or up to the boundary (∂Ω)? Example: if −Δu=f with f∈Ck(Ω), is u∈Ck+2(Ω)?
Solution methods: When can we (rigorously) find explicit solutions to (#)?
Approximation methods: Can we 'discretise' the system (#) in some manner to approximate solutions and establish convergence of such a discretisation scheme?
Notation
Analysis
For x=(x1,…,xn)∈Rn and r>0, write B(x,r):={y∈Rn:∣y−x∣=(∑i=1n(xi−yi)2)1/2<r}.
Given a measurable set U⊂Rn, write μ(U) for its Lebesgue measure.
Write K for R or C. We will always deal with functions defined on (subsets of) Rn taking their values in K.
Ω will always be a nonempty open set in Rn!
Multiïndex notation
A multiïndex is an n-tuple α=(α1,…,αn)∈Zn, αi≥0. Addition is defined in the usual way. We set:
∣α∣:=∑i=1nαi
α!:=α1!⋅⋯⋅αn!
vα:=(v1)α1⋅⋯⋅(vn)αn for v=(v1,…,vn)∈Rn
We also introduce a partial ordering on multiïndices such that α≤β iff for all i∈{1,…,n}, αi≤βi.
Using multiïndices α with ∣α∣≤k, we may abbreviate derivatives of Ck functions u as
Dαu:=(∂x1)α1…(∂xn)αn∂∣α∣u=∂1α1…∂nαnu.
⇝Integration by parts: If φ vanishes close to ∂Ω, then
∫ΩDαf⋅φ=(−1)∣α∣∫Ωf⋅Dαφ.
⇝Taylor's theorem: Whenever Ω is open and connected, x,x0∈Ω and f∈Ck(Ω),
We will often interchange limits and integrals. To this end, we recall the following:
Fatou's Lemma: Suppose {fk}k=1∞ and k→∞liminffk are summable (i.e. ∫∣fk∣,∫k→∞liminf∣fk∣<∞). Then
∫k→∞liminffk≤k→∞liminf∫fk.
Monotone convergence theorem: Suppose {fk}k=1∞ is a sequence of measurable functions on Rn with
f1≤f2≤⋯≤fk≤fk+1≤….
Then ∫k→∞limfk=k→∞lim∫fk.
Dominated convergence theorem: Suppose {fk}k=1∞ is a sequence of integrable functions (i.e. measurable and ∫f− or ∫f+ finite) such that fk→f almost everywhere. If ∣fk∣≤g a.e. f for some summable g, then
k→∞lim∫fk=∫f.
Topological vector spaces
We will study and make heavy use of spaces of functions equipped with different notions of convergence.
Basic notions.
Let X be a (real or complex) vector space.
A seminorm is given by a nonnegative function ∣∣⋅∣∣:X→R s.t.
for all λ∈K and v∈X, ∣∣λv∣∣=∣λ∣⋅∣∣v∣∣ (positive homogeneïty); and
for all v,w∈X, ∣∣v+w∣∣≤∣∣v∣∣+∣∣w∣∣ (triangle inequality).
A norm is a semi-norm ∣∣⋅∣∣ such that ∣∣v∣∣=0⇒v=0.
Given a countable family {∣⋅∣i}i=0∞ of semi-norms such that ∣v∣i=0 for all i⇒v=0, we naturally obtain a metric on X:
ρ(v,w):=i=0∑∞2−i⋅1+∣v−w∣i∣v−w∣i.
The topology induced by this metric is independent of the ordering of the semi-norms. If X is equipped with such a family of semi-norms, call the pair (X,{∣⋅∣i}i=0∞) a countably normed space.
If X is equipped with only a single norm ∣∣⋅∣∣, call the pair (X,∣∣⋅∣∣) a normed space. Here we use the equivalent metric ρ(v,w)=∣∣v−w∣∣.
If (X,∣∣⋅∣∣) is a normed space and there exists a Hermitian form⟨⋅,⋅⟩:X→K such that for all v∈X, ∣∣v∣∣=⟨v,v⟩, call the pair (X,⟨⋅,⋅⟩) an inner product space. The Cauchy-Schwarz inequality holds in such spaces:
∣⟨v,w⟩∣≤∣∣v∣∣⋅∣∣w∣∣
A sequence {vn}⊂(X,{∣⋅∣i}i=0∞) is said to be
a Cauchy sequence if ρ(vn,vm)n,m→∞0 (⇔∀i∣vn−vm∣in,m→∞0);
a (strongly) convergent sequence if there exists a v∈X such that ρ(vn,v)n,m→∞0 (⇔∀i∣vn−v∣i→0). This v is unique and we write
′vnn→∞v in X.′
A convergent sequence is always a Cauchy sequence. If every Cauchy sequence in X converges, we call it a complete space. Complete countably normed, normed and inner product spaces are referred to as Fréchet, Banach and Hilbert spaces respectively.
A Fréchet, Banach or Hilbert space is separable if it contains a countable, dense subset, i.e. if there exists a countable subset {vn}n=1∞ such that for any ε>0 and v∈X, there exists an n such that ρ(vn,v)<ε.
Separable Hilbert spaces admit orthonormal bases, i.e. vectors {vn}n=1∞⊂X such that
⟨vn,vm⟩=δnm; and
for any v∈V, v=n=1∑∞⟨vn,v⟩vn.
The dual space of a Banach or Hilbert space X, which is itself a Banach space, is given by
X∗:={f:X→Klinear:∣∣v∣∣=1sup∣f(v)∣<∞}.
The condition ∣∣v∣∣=1sup∣f(v)∣<∞ is equivalent to the condition
vnn→∞vin X⇒f(vn)n→∞f(v).
A sequence {vn}n=1∞⊂X in a Banach space is said converge weakly to v∈X, written vnn→∞v if for allf∈X∗, f(vn)n→∞f(v). Weak limits are also unique by the Hahn-Banach Theorem.
Examples
Continuous and differentiable functions. For k∈N, we define the sets
C0(Ω)Ck(Ω):={u:Ω→Kcontinuous}:={u:Ω→K:∀∣α∣≤kDαuexists and is inC0(Ω)}
as well as
C0(Ω)Ck(Ω):={u:Ω→Kcontinuous}:={u:Ω→K:∀∣α∣≤kDαuexists, is in C0(Ω)and is continuously extensible toΩ}.
For i∈{0,…,k} and a family of {Kj⋐Ω}j=0∞ with ⋃j=0∞Kj=Ω, introduce the seminorm
∣u∣i,j:=∣α∣=isupKjsup∣Dαu∣.
Properties:
(Ck(Ω),{∣⋅∣i,j:i∈{0,…,k},j∈N∪{0}}) is a Fréchet space. The underlying topology is independent of the choice of compact sets. Convergence in this space is referred to as local uniform convergence.
If Ω is bounded, the norm∣∣u∣∣k,Ω:=∑i=0ksup∣α∣=isupΩ∣Dαu∣ gives Ck(Ω) the structure of a Banach space.
A very useful compactness criterion in C0(Ω) is the Arzéla-Ascôli theorem: If {un}n=1∞ is bounded, i.e. ∣∣u∣∣0,Ω≤C, and equicontinuous, i.e. for all ε>0 there is a δ>0 such that for all n∈N,
∣x−y∣<δ⇒∣un(x)−un(y)∣<ε,
then there is a subsequence {unk} of {un} and a u∈C0(Ω) such that unkk→∞u in C0(Ω) (i.e. unk converges to u uniformly).
Smooth functions: We similarly set C∞(Ω):=⋂k=0∞Ck(Ω) and C∞(Ω):=⋂k=0∞Ck(Ω). Both of these sets are Fréchet spaces when equipped with the family of semi-norms {∣⋅∣i,j}i,j∈N∪{0} and {∣⋅∣i,Ω}i∈N∪{0} respectively.
A distinguished subset of both of these spaces is the space of compactly supported test functionsC0∞(Ω):
C0∞(Ω):={u∈C∞(Ω):suppu⋐Ω}
⟦Recall that the support of a function u:Ω→K is given by suppu={x∈Ω:u(x)=0}.⟧
This set may be given a natural topology with the following notion of convergence:
A sequence {un}n=1∞⊂C0∞(Ω) is said to converge in C0∞(Ω) to u∈C0∞(Ω) if there is some compact set K⊂Ω such (∀k)suppuk⊂K and uk∣Kk→∞u∣K in C∞(K).
Lebesgue spaces: For each p∈[1,∞] and measurable function u:Ω→K, set
∣∣u∣∣p:=⎩⎪⎨⎪⎧(∫Ω∣u∣p)1/p,ess supΩ∣u∣,p<∞p=∞
where ess supΩf:=inf{K∈R:f≤Ka.e.} for a real-valued function f.
We now define the Lebesgue spacesLp(Ω) (or Lp(Ω,K)) as
Lp(Ω):={fmeasurable:∣∣f∣∣p<∞}/∼,
where f∼g⇔f=g almost everywhere (a.e.). We will generally omit mention of equivalence classes and refer to the functions in question directly.
Properties:
Lp(Ω) is a Banach space for any p, separable for p<∞, and L2(Ω) is a Hilbert space with inner product
⟨f,g⟩:=∫Ωfˉ⋅g
C0(Ω) is dense in Lp(Ω), i.e. for all u∈Lp(Ω) and ε>0, there exists a v∈C0(Ω) such that ∣∣u−v∣∣p<ε. In fact, v may be taken to be compactly supported in Ω.
If unn→∞u in Lp(Ω), then there exists a subsequence {unk} of {un} such that unkn→∞u a.e.
For p∈[1,∞[, (Lp(Ω))∗=Lq(Ω), where q is s.t. p1+q1=1, i.e. if F∈(Lp(Ω))∗, then there exists an f∈Lq(Ω) s.t. for all g∈Lp(Ω),
F(g)=∫Ωf⋅g.
If {un}n=1∞⊂Lp(Ω) is a bounded sequence, i.e. ∣∣un∣∣p≤C, then there exists a subsequence {unk}k=1∞ of {un} and a u∈Lp(Ω) such that unkk→∞u.
The Hölder inequality holds: For p,q>1 with p1+q1=1, f∈Lp(Ω) and g∈Lq(Ω),
∫Ω∣f⋅g∣:≤∣∣f∣∣p⋅∣∣g∣∣q
For later purposes, we define the auxiliary spaces
Llocp(Ω):={u:Ω→Kmeasurable:∀K⋐Ωu∈Lp(K)}.
By Hölder's inequality, we have that Llocp(Ω)⊂Llocq(Ω) whenever q<p. We say that a sequence converges in Llocp(Ω) if its restriction to K converges in Llocp(K) for every K⋐Ω.
Hölder spaces: We call a function f:Ω→KHölder continuous with exponent γ∈]0,1] if there is a constant C≥0 s.t. for all x,y∈Ω,
∣f(x)−f(y)∣≤C∣x−y∣γ.
⟦Note that the case γ=1 corresponds to Lipschitz functions.⟧
We introduce the γth Hölder seminorm
[f]γ:=sup{∣x−y∣γ∣f(x)−f(y)∣:x,y∈Ω,x=y}
as well as the k+γth Hölder norm
∣∣f∣∣k+γ:=∣α∣≤k∑∣∣Dαf∣∣k,Ω+∣α∣=k∑[Dαu]γ.
The Hölder spaces are then given by
Ck+γ(Ω):={f∈Ck(Ω):∣∣f∣∣k+γ<∞}.
Exercise. (Ck+γ(Ω),∣∣⋅∣∣k+γ) is a Banach space.
Approximation by smooth functions
Let ρ∈C0∞(Rn) be such that suppρ⊂B(0,1), ρ≥0 and ∫Rnρ=1; for example,
ρ(x):={c⋅e−1/(∣x∣2−1),0,∣x∣<1otherwise
for a suitable constant c>0.
Given u∈Lloc1(Ω), extend u to all of Rn s.t. u∣Rn\Ω≡0 and for each ε>0 define the mollifier of u to be the function
If suppu⋐Ω, then for ε<dist(suppu,∂Ω), Jεu∈C0∞(Ω).
For any u∈Lp(Ω), p∈[1,∞], ∣∣Jεu∣∣p≤∣∣u∣∣p.
If u∈C0(Ω), then Jεuε↘0u locally uniformly, i.e. for each K⋐Ω,
ε↘0limKsup∣Jεu−u∣=0.
If p∈[1,∞[, then ∣∣Jεu−u∣∣pε↘00.
If p∈[1,∞], then ε↘0limJεu=u almost everywhere.
Proof. Note that x↦ρ(εx−z) and all of its derivatives are bounded functions of z supported in B(x,ε), whence smoothness follows from the dominated convergence theorem. We tackle the other claims:
Let x∈Rn be such that dist(x,suppu)>ε and z∈B(0,1). The (reverse) triangle inequality implies that dist(x−εz,suppu)>dist(x,suppu)−ε>0 so that
Jεu(x)=∫B(0,1)ρ(z)u(x−εz)dz=0,
whence suppJεu⊂Bε(suppu). This establishes the claim for small ε.
We estimate, using Hölder's inequality using the conjugate pair (p−1p,p):
Let u∈Lp(Ω) and fix δ>0. By the density of C0(Ω) in Lp(Ω), we may find a v∈C0(Ω) with compact support in Ω such that ∣∣u−v∣∣<3δ. Therefore, by the triangle inequality,
By the preceding part, we may choose ε>0 sufficiently small so that ∣∣Jεv−v∣∣p<3δ.
Note that ∣Jεu(x)−u(x)∣=∣ε−n∫Ωρ(εx−z)⋅(u(z)−u(x))dz∣≤supρ⋅ε−n∫B(x,ε)∣u(z)−u(x)∣dzε↘00 by the Lebesgue differentiation theorem.
Using mollifiers, we may prove the following density result.
Lemma 1.2. C0∞(Ω) is dense in Lp(Ω), i.e. for any u∈Lp(Ω) and ε>0, there exists a v∈C0∞(Ω) such that ∣∣u−v∣∣p<ε.
Proof. First assume Ω is bounded. For any δ>0, let U={x∈Ω:dist(x,∂Ω)>δ} and set v:=u⋅χU, where χU is the characteristic function of U. Since
∣∣u−v∣∣p=(∫Ω\U∣u∣p)1/p
and μ(Ω\U)↘0 as δ↘0, we may choose δ small enough so that ∣∣u−v∣∣p<3ε, and suppv⊂U⋐Ω. On the other hand, by the preceding theorem, ∣∣Jzv−v∣∣pz↘00 and Jzv∈C0∞(Ω)z<δ. Therefore, for sufficiently small z, ∣∣Jzv−v∣∣p<3ε. The triangle inequality now implies ∣∣u−Jzv∣∣<32ε.
Now suppose Ω is unbounded. Since ∫Ω∣u∣p=limr→∞∫Ω∩B(0,r)∣u∣p, we may choose r>0 large enough so that ∣∣u−u⋅χΩ∩B(0,r)∣∣p<3ε. We may now treat u⋅χΩ∩B(0,r) according to the preceding case.
Sobolev spaces
We now turn our attention to Lp functions that admit derivatives in a weaker sense.
Weakly differentiable functions
Definition. Let u∈Lloc1(Ω). We say that Dαu exists in the weak sense if there exists a v∈Lloc1(Ω) such that for all φ∈C0∞(Ω),
∫Ωu⋅Dαφ=(−1)∣α∣∫Ωv⋅φ(*)
and write Dαu:=v. If for each ∣α∣≤k, Dαu exists in the weak sense, we say that u is k-times weakly differentiable.
Remark. When Dαu exists, it is uniquely determined by (*) almost everywhere.
Notation. Whenever u is k-times weakly differentiable, we adopt the usual partial derivative notation, i.e. ∂iu:=Deiu and Dαu=∂1α1…∂nαnu. Note that there is no ambiguity in the order of differentiation!
Exercise. Show that weak differentiation is K-linear, i.e. given λ∈K and u,v∈Lloc1(Ω) such that both Dαu and Dαv exist in the weak sense, we have that Dα(λu)=λ⋅Dαu and Dα(u+v)=Dαu+Dαv.
Exercise. If u∈Lloc1(Ω) is such that Dαu∈Lloc1(Ω) exists weakly and if Dβ(Dαu)∈Lloc1(Ω) exists weakly, then Dα+βu exists and Dα+βu=Dβ(Dαu).
Exercise. If u∈Lloc1(Ω) is k-times weakly differentiable and v∈Ck(Ω), then for all β with ∣β∣∈[0,m],
Dβ(uv)=α≤β∑(βα)Dαu⋅Dβ−αu.
Examples
If u∈Ck(Ω), then (*) is just integration by parts formula ⇒u is k-times weakly differentiable.
Consider the function f:]0,2[→R defined by
f(x)={x,1,0<x≤11≤x<2.
We claim that f is weakly differentiable with weak derivative g:]0,2[→R given by
We now consider a function with a jump, viz. f:]0,2[→R such that
f(x)={x,2,0<x≤11≤x<2.
We ask whether f is weakly differentiable.
Suppose there exists a g∈Lloc1(]0,2[) such that (!) holds for all φ∈C0∞(]0,2[). A quick computation shows that the left-hand side of this equation is φ(1)−∫01φ, whereas the right-hand side evaluates to −∫01φ⋅g. Now, choosing φ∈C0∞(]0,1[), we find that g≡1 a.e. on ]0,1[, and choosing φ∈C0∞(]1,2[), we see that we must have g≡0. However, for general φ, we are left with the nonsensical statement
∀φ∈C0∞(]0,2[φ(1)=0!
Definition. For each k∈N∪{0} and p∈[1,∞], the Sobolev spaceWk,p(Ω) is defined by
Wk,p(Ω):={u∈Lp(Ω):∀∣α∣≤k,Dαuexists in the weak sense and is in Lp(Ω)}.
Lemma 1.3. (Wk,p(Ω),∣∣⋅∣∣k,p) is a Banach space and (Wk,2(Ω),⟨⋅,⋅⟩k) a Hilbert space.
Proof. We proceed in steps:
∣∣⋅∣∣k,p is a norm: Homogeneity and ∣∣u∣∣k,p=0⇒u=0 a.e. is clear. The triangle inequality in the case p=∞ is immediate; for p<∞ and u,v∈Wk,p(Ω), we use Minkowski's inequality and the triangle inequality for Lp functions:
By the completeness of Lp(Ω), we may, for each α with ∣α∣≤k, find a vα∈Lp(Ω) such that ∣∣Dαun−vα∣∣pn→∞0. Now, by definition of the weak derivative, we have that for each φ∈C0∞(Ω),
∫Ωun⋅Dαφ=(−1)∣α∣∫ΩDαun⋅φ.
Taking limits on both sides of this equation (why?), we obtain
∫Ωv0⋅Dαφ=(−1)∣α∣∫Ωvα⋅φ,
i.e. vα=Dαv0 for all such α so that ∣∣Dαun−Dαv0∣∣pn→∞0⇒∣∣un−v0∣∣k,pn→∞0.
Remark. If we wish to differentiate between Sobolev norms corresponding to Sobolev spaces over different domains, we shall write ∣∣⋅∣∣k,p,Ω for the norm on Wk,p(Ω); likewise, we shall sometimes write ∣∣⋅∣∣p,Ω for the norm on Lp(Ω).