Sobolev spaces
Department of Mathematics, Aalto University 2021
1 SOBOLEV SPACES 1
1.1 Weak derivatives . . . 1
1.2 Sobolev spaces . . . 4
1.3 Properties of weak derivatives. . . 8
1.4 Completeness of Sobolev spaces . . . 9
1.5 Hilbert space structure . . . 11
1.6 Approximation by smooth functions . . . 12
1.7 Local approximation in Sobolev spaces. . . 16
1.8 Global approximation in Sobolev spaces . . . 17
1.9 Sobolev spaces with zero boundary values. . . 18
1.10 Chain rule. . . 21
1.11 Truncation. . . 23
1.12 Weak convergence methods for Sobolev spaces. . . 25
1.13 Difference quotients. . . 33
1.14 Absolute continuity on lines . . . 36
2 SOBOLEV INEQUALITIES 42 2.1 Gagliardo-Nirenberg-Sobolev inequality . . . 43
2.2 Sobolev-Poincaré inequalities . . . 49
2.3 Morrey’s inequality . . . 55
2.4 Lipschitz functions andW1,∞. . . 59
2.5 Summary of the Sobolev embeddings. . . 62
2.6 Direct methods in the calculus of variations . . . 62
3 MAXIMAL FUNCTION APPROACH TO SOBOLEV SPACES 70 3.1 Representation formulas and Riesz potentials . . . 71
3.2 Sobolev-Poincaré inequalities . . . 78
3.3 Sobolev inequalities on domains . . . 86
3.4 A maximal function characterization of Sobolev spaces. . . 89
3.5 Pointwise estimates . . . 92
3.6 Approximation by Lipschitz functions. . . 97
3.7 Maximal operator on Sobolev spaces . . . 102
4 POINTWISE BEHAVIOUR OF SOBOLEV FUNCTIONS 106
4.1 Sobolev capacity . . . .106
4.2 Capacity and measure . . . .109
4.3 Quasicontinuity . . . .116
4.4 Lebesgue points of Sobolev functions . . . 119
4.5 Sobolev spaces with zero boundary values. . . 124
Sobolev spaces 1
In this chapter we begin our study of Sobolev spaces. The Sobolev space is a vector space of functions that have weak derivatives. Motivation for studying these spaces is that solutions of partial differential equations, when they exist, belong naturally to Sobolev spaces.
1.1 Weak derivatives
Notation. LetΩ⊂Rnbe open,f:Ω→Randk=1, 2, . . .. Then we use the following notations:
C(Ω)={f :f continuous inΩ}
suppf={x∈Ω: f(x)6=0}= the support off
C0(Ω)={f∈C(Ω) : suppf is a compact subset ofΩ}
Ck(Ω)={f∈C(Ω) : f isktimes continuously diferentiable}
C0k(Ω)=Ck(Ω)∩C0(Ω) C∞=
\∞ k=1
Ck(Ω)= smooth functions C∞0(Ω)=C∞(Ω)∩C0(Ω)
= compactly supported smooth functions
= test functions
■
W A R N I N G : In general, suppf*Ω. Examples 1.1:
(1) Letu:B(0, 1)→R,u(x)=1− |x|. Then suppu=B(0, 1).
1
(2) Letf:R→Rbe
f(x)=
x2, xÊ0,
−x2, x<0.
Now f∈C1(R) \C2(R) although the graph looks smooth.
(3) Let us defineϕ:Rn→R, ϕ(x)=
e
1
|x|2−1, x∈B(0, 1), 0, x∈Rn\B(0, 1).
Nowϕ∈C∞0(Rn) and suppϕ=B(0, 1) (exercise).
Let us start with a motivation for definition of weak derivatives. LetΩ⊂Rn be open,u∈C1(Ω) andϕ∈C∞0(Ω). Integration by parts gives
ˆ
Ω
uÇϕ Çxj
dx= − ˆ
Ω
Çu Çxjϕdx.
There is no boundary term, sinceϕhas a compact support inΩand thus vanishes nearÇΩ.
Let thenu∈Ck(Ω), k=1, 2, . . ., and letα=(α1,α2, . . . ,αn)∈Nn (we use the convention that 0∈N) be a multi-index such that the order of multi-index|α| = α1+. . .+αnis at mostk. We denote
Dαu= Ç|α|u Çxα11. . .Çxαnn
= Çα1 Çx1α1. . . Çαn
Çxαnn
u.
TH E M O R A L: A coordinate of a multi-index indicates how many times a function is differentiated with respect to the corresponding variable. The order of a multi-index tells the total number of differentiations.
Successive integration by parts gives ˆ
ΩuDαϕdx=(−1)|α|
ˆ
ΩDαuϕdx.
Notice that the left-hand side makes sense even under the assumptionu∈L1loc(Ω).
Definition 1.2. Assume thatu∈L1loc(Ω) and letα∈Nnbe a multi-index. Then v∈L1loc(Ω) is theαth weak partial derivative ofu, writtenDαu=v, if
ˆ
Ωu Dαϕdx=(−1)|α|
ˆ
Ωvϕdx
for every test functionϕ∈C∞0 (Ω). We denoteD0u=D(0,...,0)=u. If|α| =1, then Du=(D1u,D2u. . . ,Dnu)
is the weak gradient ofu. Here Dju= Çu
Çxj=D(0,...,1,...,0)u, j=1, . . . ,n, (the jth component is 1).
TH E M O R A L: Classical derivatives are defined as pointwise limits of differ- ence quotients, but the weak derivatives are defined as a functions satisfying the integration by parts formula. Observe, that changing the function on a set of measure zero does not affect its weak derivatives.
WA R N I N G We use the same notation for the weak and classical derivatives. It should be clear from the context which interpretation is used.
Remarks 1.3:
(1) Ifu∈Ck(Ω), then the classical partial derivatives up to orderkare also the corresponding weak derivatives ofu. In this sense, weak derivatives generalize classical derivatives.
(2) Ifu=0 almost everywhere in an open set, thenDαu=0 almost everywhere in the same set.
Lemma 1.4. A weakαth partial derivative ofu, if it exists, is uniquely defined up to a set of measure zero.
Proof. Assume thatv,ev∈L1loc(Ω) are both weakαth partial derivatives ofu, that
is, ˆ
ΩuDαϕdx=(−1)|α|
ˆ
Ωvϕdx=(−1)|α|
ˆ
Ωevϕdx for everyϕ∈C∞0 (Ω). This implies that
ˆ
Ω
(v−v)eϕdx=0 for every ϕ∈C0∞(Ω). (1.1) Claim:v=evalmost everywhere inΩ.
Reason. Let Ω0bΩ (i.e. Ω0 is open and Ω0 is a compact subset of Ω). The spaceC∞0(Ω0) is dense inLp(Ω0) (we shall return to this later). There exists a sequence of functions ϕi∈C0∞(Ω0) such that|ϕi| É2 inΩ0 andϕi→sgn(v−ev) almost everywhere inΩ0asi→ ∞. Here sgn is the signum function.
Identity (1.1) and the dominated convergence theorem, with the majorant
|(v−v)eϕi| É2(|v| + |ev|)∈L1(Ω0), give 0=lim
i→∞
ˆ
Ω0
(v−v)eϕidx= ˆ
Ω0
lim
i→∞(v−ev)ϕidx
= ˆ
Ω0(v−v) sgn(ve −v)dxe = ˆ
Ω0|v−ev|dx
This implies thatv=evalmost everywhere inΩ0for everyΩ0bΩ. Thus v=ve
almost everywhere inΩ. ■
From the proof we obtain a very useful corollary.
Corollary 1.5 (Fundamental lemma of the calculus of variations). Iff∈L1loc(Ω)
satisfies ˆ
Ωfϕdx=0
for everyϕ∈C∞0 (Ω), thenf=0 almost everywhere inΩ.
TH E M O R A L: This is an integral way to say that a function is zero almost everywhere.
Example 1.6. Letn=1 andΩ=(0, 2). Consider
u(x)=
x, 0<x<1, 1, 1Éx<2, and
v(x)=
1, 0<x<1, 0, 1Éx<2.
We claim thatu0=vin the weak sense. To see this, we show that ˆ 2
0
uϕ0dx= − ˆ 2
0
vϕdx for everyϕ∈C∞0 ((0, 2)).
Reason. An integration by parts and the fundamental theorem of calculus give ˆ 2
0
u(x)ϕ0(x)dx= ˆ 1
0
xϕ0(x)dx+ ˆ 2
1
ϕ0(x)dx
=xϕ(x)
¯
¯
¯
¯
1 0
| {z }
=ϕ(1)
− ˆ 1
0
ϕ(x)dx+ϕ(2)
| {z }
=0
−ϕ(1)
= − ˆ 1
0
ϕ(x)dx= − ˆ 2
0
vϕ(x)dx
for everyϕ∈C∞0 ((0, 2)). ■
1.2 Sobolev spaces
Definition 1.7. Assume that Ω is an open subset of Rn. The Sobolev space Wk,p(Ω) consists of functionsu∈Lp(Ω) such that for every multi-indexαwith
|α| Ék, the weak derivativeDαuexists andDαu∈Lp(Ω). Thus Wk,p(Ω)={u∈Lp(Ω) :Dαu∈Lp(Ω),|α| Ék}.
Ifu∈Wk,p(Ω), we define its norm kukWk,p(Ω)=
à X
|α|Ék
ˆ
Ω|Dαu|pdx
!1p
, 1Ép< ∞, and
kukWk,∞(Ω)= X
|α|Ék
ess sup
Ω |Dαu|.
Notice thatD0u=D(0,...,0)u=u. Assume thatΩ0is an open subset ofΩ. We say thatΩ0is compactly contained inΩ, denotedΩ0bΩ, ifΩ0is a compact subset of Ω. A functionu∈Wk,p
loc (Ω), ifu∈Wk,p(Ω0) for everyΩ0bΩ.
TH E M O R A L: Thus Sobolev spaceWk,p(Ω) consists of functions inLp(Ω) that have weak partial derivatives up to orderkand they belong toLp(Ω).
Remarks 1.8:
(1) As inLpspaces we identifyWk,pfunctions which are equal almost every- where.
(2) There are several ways to define a norm onWk,p(Ω). The normk · kWk,p(Ω)
is equivalent, for example, with the norm X
|α|ÉkkDαukLp(Ω), 1ÉpÉ ∞. andk · kWk,∞(Ω)is also equivalent with
max|α|ÉkkDαukL∞(Ω). (3) Fork=1 we use the norm
kukW1,p(Ω)=³
kukLpp(Ω)+ kDukpLp(Ω)
´1p
= µˆ
Ω|u|pdx+ ˆ
Ω|Du|pdx
¶1p
, 1Ép< ∞, and
kukW1,∞(Ω)=ess sup
Ω |u| +ess sup
Ω |Du|. We may also consider equivalent norms
kukW1,p(Ω)= Ã
kukLpp(Ω)+
n
X
j=1
°°Dju°
°
p Lp(Ω)
!1p ,
kukW1,p(Ω)= kukLp(Ω)+
n
X
j=1
°°Dju°
°Lp(Ω), and
kukW1,p(Ω)= kukLp(Ω)+ kDukLp(Ω)
when 1Ép< ∞and kukW1,∞(Ω)=max©
kukL∞(Ω),kD1ukL∞(Ω), . . . ,kDnukL∞(Ω)
ª.
Example 1.9. Letu:B(0, 1)→[0,∞],u(x)= |x|−α,α>0. Clearlyu∈C∞(B(0, 1) \ {0}), butuis unbounded in any neighbourhood of the origin.
We start by showing thatuhas a weak derivative in the entire unit ball. When x6=0 , we have
Çu
Çxj(x)= −α|x|−α−1xj
|x|= −α xj
|x|α+2, j=1, . . . ,n.
Thus
Du(x)= −α x
|x|α+2. Gauss’ theorem gives
ˆ
B(0,1)\B(0,ε)
Dj(uϕ)dx= ˆ
Ç(B(0,1)\B(0,ε))
uϕνjdS,
whereν=(ν1, . . . ,νn) is the outward pointing unit (|ν| =1) normal of the boundary andϕ∈C∞0(B(0, 1)). Asϕ=0 onÇB(0, 1), this can be written as
ˆ
B(0,1)\B(0,ε)
Djuϕdx+ ˆ
B(0,1)\B(0,ε)
uDjϕdx= ˆ
ÇB(0,ε)
uϕνjdS.
By rearranging terms, we obtain ˆ
B(0,1)\B(0,ε)
uDjϕdx= − ˆ
B(0,1)\B(0,ε)
Djuϕdx+ ˆ
ÇB(0,ε)uϕνjdS. (1.2) Let us estimate the last term on the right-hand side. Sinceν(x)= −|xx|, we have νj(x)= −|x|xj, whenx∈ÇB(0,ε). Thus
¯
¯
¯
¯ ˆ
ÇB(0,ε)
uϕνjdS
¯
¯
¯
¯É kϕkL∞(B(0,1))
ˆ
ÇB(0,ε)
ε−αdS
= kϕkL∞(B(0,1))ωn−1εn−1−α→0 as ε→0,
ifn−1−α>0. Hereωn−1=Hn−1(ÇB(0, 1)) is the (n−1)-dimensional measure of the sphereÇB(0, 1).
Next we study integrability ofDju. We need this information in order to be able to use the dominated convergence theorem. A straightforward computation gives
ˆ
B(0,1)
¯¯Dju¯
¯dxÉ ˆ
B(0,1)|Du|dx=α ˆ
B(0,1)|x|−α−1dx
=α ˆ 1
0
ˆ
ÇB(0,r)|x|−α−1dS dr=αωn−1 ˆ 1
0
r−α−1+n−1dr
=αωn−1
ˆ 1
0
rn−α−2dr= αωn−1
n−α−1rn−α−1
¯
¯
¯
¯
1 0
< ∞, ifn−1−α>0.
The following argument shows thatDju is a weak derivative ofualso in a neighbourhood of the origin. By the dominated convergence theorem
ˆ
B(0,1)
uDjϕdx= ˆ
B(0,1)
limε→0
³
uDjϕχB(0,1)\B(0,ε)
´ dx
=lim
ε→0
ˆ
B(0,1)\B(0,ε)
uDjϕdx
= −lim
ε→0
ˆ
B(0,1)\B(0,ε)
Djuϕdx+lim
ε→0
ˆ
ÇB(0,ε)
uϕνjdS
= − ˆ
B(0,1)
limε→0DjuϕχB(0,1)\B(0,ε)dx
= − ˆ
B(0,1)
Djuϕdx.
Here we used the dominated convergence theorem twice: First to the function uDjϕχB(0,1)\B(0,ε),
which is dominated by|u|kDϕk∞∈L1(B(0, 1)), and then to the function DjuϕχB(0,1)\B(0,ε),
which is dominated by|Du|kϕk∞∈L1(B(0, 1)). We also used (1.2) and the fact that the last term there converges to zero asε→0.
Now we have proved thatuhas a weak derivative in the unit ball. We note that u∈Lp(B(0, 1)) if and only if−pα+n>0, or equivalently, α< np. On the other hand,|Du| ∈Lp(B(0, 1), if−p(α+1)+n>0, or equivalently,α<n−pp . Thus u∈W1,p(B(0, 1)) if and only ifα<n−pp.
Let (ri) be a countable and dense subset ofB(0, 1) and defineu:B(0, 1)→[0,∞], u(x)=
X∞ i=1
1
2i|x−ri|−α. Thenu∈W1,p(B(0, 1)) ifα<n−pp .
Reason.
kukW1,p(B(0,1))É X∞ i=1
1 2i
°
°|x−ri|−α°
°W1,p(B(0,1))
É X∞ i=1
1 2i
°
°|x|−α°
°W1,p(B(0,1))
=°
°|x|−α°
°W1,p(B(0,1))< ∞. ■
Note that ifα>0, thenuis unbounded in every open subset ofB(0, 1) and not differentiable in the classical sense in a dense subset.
TH E M O R A L: Functions inW1,p, 1Ép<n,nÊ2, may be unbounded in every open subset.
Example 1.10. Observe, thatu(x)= |x|−α,α>0, does not belong toW1,n(B(0, 1).
However, there are unbounded functions inW1,n,nÊ2. Letu:B(0, 1)→R,
u(x)=
log³
log³ 1+|x|1
´´
, x6=0, 0, x=0.
Then u∈W1,n(B(0, 1)) when nÊ2, but u∉L∞(B(0, 1)). This can be used to construct a function inW1,n(B(0, 1) that is unbounded in every open subset of B(0, 1) (exercise).
TH E M O R A L: Functions inW1,p, 1ÉpÉn,nÊ2, are not continuous. Later we shall see, that everyW1,p function with p>n coincides with a continuous function almost everywhere.
Example 1.11. The functionu:B(0, 1)→R,
u(x)=u(x1, . . . ,xn)=
1, xn>0, 0, xn<0, does not belong toW1,p(B(0, 1) for any 1ÉpÉ ∞(exercise).
1.3 Properties of weak derivatives
The following general properties of weak derivatives follow rather directly from the definition.
Lemma 1.12. Assume thatu,v∈Wk,p(Ω) and|α| Ék. Then (1) Dαu∈Wk−|α|,p(Ω),
(2) Dβ(Dαu)=Dα(Dβu) for all multi-indicesα,βwith|α| + |β| Ék, (3) for everyλ,µ∈R,λu+µv∈Wk,p(Ω) and
Dα(λu+µv)=λDαu+µDαv, (4) ifΩ0⊂Ωis open, thenu∈Wk,p(Ω0),
(5) (Leibniz’s formula) ifη∈C∞0 (Ω), thenηu∈Wk,p(Ω) and
Dα(ηu)=X
βÉα
Ãα β
!
DβηDα−βu,
where
Ãα β
!
= α!
β!(α−β)!, α!=α1!. . .αn! andβÉαmeans thatβjÉαjfor everyj=1, . . . ,n.
TH E M O R A L: Weak derivatives have the same properties as classical deriva- tives of smooth functions.
Proof. (1) Follows directly from the definition of weak derivatives. See also (2).
(2) Letϕ∈C∞0(Ω). ThenDβϕ∈C∞0(Ω). Therefore (−1)|β|
ˆ
ΩDβ(Dαu)ϕdx= ˆ
ΩDαuDβϕdx
=(−1)|α|
ˆ
Ω
uDα+βϕdx
=(−1)|α|(−1)|α+β|
ˆ
Ω
Dα+βuϕdx for all test functionsϕ∈C∞0 (Ω). Notice that
|α| + |α+β| =α1+. . .+αn+(α1+β1)+. . .+(αn+βn)
=2(α1+. . .+αn)+β1+. . .+βn
=2|α| + |β|.
As 2|α|is an even number, the estimate above, together with the uniqueness results Lemma1.4and Corollary1.5, implies thatDβ(Dαu)=Dα+βu.
(3) and (4) Clear.
(5) First we consider the case|α| =1. Letϕ∈C∞0 (Ω). By Leibniz’s rule for differentiable functions and the definition of weak derivative
ˆ
ΩηuDαϕdx= ˆ
Ω
(uDα(ηϕ)−u(Dαη)ϕ)dx
= − ˆ
Ω(ηDαu+uDαη)ϕdx
for allϕ∈C∞0(Ω). The case|α| >1 follows by induction (exercise). ä
1.4 Completeness of Sobolev spaces
One of the most useful properties of Sobolev spaces is that they are complete.
Thus Sobolev spaces are closed under limits of Cauchy sequences.
A sequence (ui) of functionsui∈Wk,p(Ω),i=1, 2, . . ., converges inWk,p(Ω) to a functionu∈Wk,p(Ω), if for everyε>0 there existsiεsuch that
kui−ukWk,p(Ω)<ε when iÊiε. Equivalently,
ilim→∞kui−ukWk,p(Ω)=0.
A sequence (ui) is a Cauchy sequence inWk,p(Ω), if for everyε>0 there exists iεsuch that
kui−ujkWk,p(Ω)<ε when i,jÊiε.
WA R N I N G: This is not the same condition as
kui+1−uikWk,p(Ω)<ε when iÊiε.
Indeed, the Cauchy sequence condition implies this, but the converse is not true (exercise).
Theorem 1.13 (Completeness). The Sobolev space Wk,p(Ω), 1ÉpÉ ∞, k= 1, 2, . . ., is a Banach space.
TH E M O R A L: The spacesCk(Ω),k=1, 2, . . . , are not complete with respect to the Sobolev norm, but Sobolev spaces are. This is important in existence arguments for PDEs.
Proof. Step 1: k · kWk,p(Ω)is a norm.
Reason. (1) kukWk,p(Ω)=0⇐⇒u=0 almost everywhere inΩ.
=⇒ kukWk,p(Ω)=0 implieskukLp(Ω)=0, which implies thatu=0 almost every- where inΩ.
⇐= u=0 almost everywhere inΩimplies ˆ
Ω
Dαuϕdx=(−1)|α|
ˆ
Ω
uDαϕdx=0
for allϕ∈C∞0(Ω). This together with Corollary1.5implies thatDαu=0 almost everywhere inΩfor allα,|α| Ék.
(2) kλukWk,p(Ω)= |λ|kukWk,p(Ω),λ∈R. Clear.
(3) The triangle inequality for 1Ép< ∞follows from the elementary inequal- ity (a+b)αÉaα+bα,a,bÊ0, 0<αÉ1, and Minkowski’s inequality, since
ku+vkWk,p(Ω)= Ã
X
|α|Ék
kDαu+DαvkLpp(Ω)
!1p
É Ã
X
|α|Ék
¡kDαukLp(Ω)+ kDαvkLp(Ω)
¢p
!1p
É Ã
X
|α|Ék
kDαukLpp(Ω)
!1p +
à X
|α|Ék
kDαvkLpp(Ω)
!1p
= kukWk,p(Ω)+ kvkWk,p(Ω). ■ Step 2:Let (ui) be a Cauchy sequence inWk,p(Ω). As
kDαui−DαujkLp(Ω)É kui−ujkWk,p(Ω), |α| Ék,
it follows that (Dαui) is a Cauchy sequence inLp(Ω),|α| Ék. The completeness ofLp(Ω) implies that there existsuα∈Lp(Ω) such thatDαui→uαinLp(Ω). In particular,ui→u(0,...,0)=uinLp(Ω).
Step 3:We show thatDαu=uα,|α| Ék. We would like to argue ˆ
ΩuDαϕdx=lim
i→∞
ˆ
ΩuiDαϕdx
=lim
i→∞(−1)|α|
ˆ
ΩDαuiϕdx
=(−1)|α|
ˆ
Ω
uαϕdx
for every ϕ∈C∞0 (Ω). On the second line we used the definition of the weak derivative. Next we show how to conclude the first and last equalities above.
1<p< ∞ Letϕ∈C∞0 (Ω). By Hölder’s inequality we have
¯
¯
¯
¯ ˆ
ΩuiDαϕdx− ˆ
ΩuDαϕdx
¯
¯
¯
¯=
¯
¯
¯
¯ ˆ
Ω(ui−u)Dαϕdx
¯
¯
¯
¯
É kui−ukLp(Ω)kDαϕkLp0(Ω)→0
and consequently we obtain the first inequality above. The last inequality follows in the same way, since
¯
¯
¯
¯ ˆ
Ω
Dαuiϕdx− ˆ
Ω
uαϕdx
¯
¯
¯
¯É kDαui−uαkLp(Ω)kϕkLp0(Ω)→0.
p=1,p= ∞ A similar argument as above (exercise).
This means that the weak derivativesDαuexist andDαu=uα,|α| Ék. As we also know thatDαui→uα=Dαu,|α| Ék, we conclude thatkui−ukWk,p(Ω)→0.
Thusui→uinWk,p(Ω). ä
Remark 1.14. Wk,p(Ω), 1Ép< ∞is separable. In the case k=1 consider the mappingu7→(u,Du) fromW1,p(Ω) toLp(Ω)×Lp(Ω)nand recall that a subset of a separable space is separable. However,W1,∞(Ω) is not separable (exercise).
1.5 Hilbert space structure
The spaceWk,2(Ω) is a Hilbert space with the inner product
〈u,v〉Wk,2(Ω)= X
|α|Ék〈Dαu,Dαv〉L2(Ω), where
〈Dαu,Dαv〉L2(Ω)= ˆ
Ω
DαuDαv dx.
Observe that
kukWk,2(Ω)= 〈u,u〉
1 2 Wk,2(Ω).
1.6 Approximation by smooth functions
This section deals with the question whether every function in a Sobolev space can be approximated by a smooth function.
Defineφ∈C0∞(Rn) by
φ(x)=
c e
1
|x|2−1, |x| <1, 0, |x| Ê1, wherec>0 is chosen so that
ˆ
Rnφ(x)dx=1.
Forε>0, set
φε(x)= 1 εnφ³x
ε
´ .
The functionφis called the standard mollifier or Friedrich’s mollifier. Observe thatφεÊ0, suppφε=B(0,ε) and
ˆ
Rnφε(x)dx= 1 εn
ˆ
Rnφ³x ε
´ dx= 1
εn ˆ
Rnφ(y)εnd y= ˆ
Rnφ(x)dx=1 for allε>0. Here we used the change of variabley=xε,dx=εnd y.
Notation. IfΩ⊂Rnis open withÇΩ6= ;, we write
Ωε={x∈Ω: dist(x,ÇΩ)>ε}, ε>0.
If f∈L1loc(Ω), we obtain its standard convolution mollification fε:Ωε→[−∞,∞], fε(x)=(f∗φε)(x)=
ˆ
Ωf(y)φε(x−y)d y.
■
TH E M O R A L: Since the convolution is a weighted integral average off over the ballB(x,ε) for everyx, instead ofΩit is well defined only inΩε. IfΩ=Rn, we do not have this problem.
Remarks 1.15:
(1) For everyx∈Ωε, we have fε(x)=
ˆ
Ωf(y)φε(x−y)d y= ˆ
B(x,ε)
f(y)φε(x−y)d y.
(2) By a change of variablesz=x−ywe have ˆ
Ωf(y)φε(x−y)d y= ˆ
Ωf(x−z)φε(z)d z
(3) For everyx∈Ωε, we have
|fε(x)| É
¯
¯
¯
¯ ˆ
B(x,ε)
f(y)φε(x−y)d y
¯
¯
¯
¯É kφεk∞
ˆ
B(x,ε)|f(y)|d y< ∞. (4) Iff∈C0(Ω), thenfε∈C0(Ωε), whenever
0<ε<ε0=1
2dist(suppf,ÇΩ).
Reason. If x∈Ωεs.t. dist(x, suppf)>ε0(in particular, for everyx∈Ωε\ Ωε0) thenB(x,ε)∩suppf= ;, which implies thatfε(x)=0. ■ Lemma 1.16 (Properties of mollifiers).
(1) fε∈C∞(Ωε).
(2) fε→f almost everywhere asε→0.
(3) Iff∈C(Ω), thenfε→f uniformly in everyΩ0bΩ.
(4) Iff∈Lploc(Ω), 1Ép< ∞, then fε→f inLp(Ω0) for everyΩ0bΩ.
W A R N I N G : Assertion (4) does not hold forp= ∞, since there are functions inL∞(Ω) that are not continuous.
Proof. (1) Letx∈Ωε,j=1, . . . ,n,ej=(0, . . . , 1, . . . , 0) (the jth component is 1). Let h0>0 such thatB(x,h0)⊂Ωεand leth∈R,|h| <h0. Then
fε(x+hej)−fε(x)
h = 1
εn ˆ
B(x+hej,ε)∪B(x,ε)
1 h
· φ
µx+hej−y ε
¶
−φ³x−y ε
´¸
f(y)d y.
LetΩ0=B(x,h0+ε). ThenΩ0bΩandB(x+hej,ε)∪B(x,ε)⊂Ω0. Claim:
1 h
· φ
µx+hej−y ε
¶
−φ³x−y ε
´¸
→1 ε
Çφ Çxj
³x−y ε
´
for everyy∈Ω0ash→0.
Reason. Letψ(x)=φ¡x−y
ε
¢. Then Çψ Çxj
(x)=1 ε
Çφ Çxj
³x−y ε
´
, j=1, . . . ,n and
ψ(x+hej)−ψ(x)= ˆ h
0
Ç
Çt(ψ(x+tej))dt= ˆ h
0
Dψ(x+tej)·ejdt.
■
Thus
|ψ(x+hej)−ψ(x)| É ˆ |h|
0 |Dψ(x+tej)·ej|dt É1
ε ˆ |h|
0
¯
¯
¯
¯Dφ
µx+tej−y ε
¶¯
¯
¯
¯ dt É|h|
ε kDφkL∞(Rn).
This estimate shows that we can use the Lebesgue dominated convergence theorem (on the third row) to obtain
Çfε Çxj
(x)=lim
h→0
fε(x+hej)−fε(x) h
=lim
h→0
1 εn
ˆ
Ω0
1 h
· φ
µx+hej−y ε
¶
−φ³x−y ε
´¸ f(y)d y
= 1 εn
ˆ
Ω0
1 ε
Çφ Çxj
³x−y ε
´ f(y)d y
= ˆ
Ω0
Çφε
Çxj
(x−y)f(y)d y
= µÇφε
Çxj ∗f
¶ (x).
A similar argument shows thatDαfεexists and Dαfε=Dαφε∗f inΩε
for every multi-indexα. (2) Recall that´
B(x,ε)φε(x−y)d y=1. Therefore we have
|fε(x)−f(x)| =
¯
¯
¯
¯ ˆ
B(x,ε)φε(x−y)f(y)d y−f(x) ˆ
B(x,ε)φε(x−y)d y
¯
¯
¯
¯
=
¯
¯
¯
¯ ˆ
B(x,ε)φε(x−y)(f(y)−f(x))d y
¯
¯
¯
¯ É 1
εn ˆ
B(x,ε)φ³x−y ε
´
|f(y)−f(x)|d y ÉΩnkφkL∞(Rn)
1
|B(x,ε)| ˆ
B(x,ε)|f(y)−f(x)|d y→0
for almost every x∈Ω asε→0. Here Ωn= |B(0, 1)|and the last convergence follows from the Lebesgue’s differentiation theorem.
(3) LetΩ0bΩ00bΩ, 0<ε<dist(Ω0,ÇΩ00), andx∈Ω0. BecauseΩ00is compact andf∈C(Ω),f is uniformly continuous inΩ00, that is, for everyε0>0 there exists δ>0 such that
|f(x)−f(y)| <ε0for everyx,y∈Ω00with|x−y| <δ.
By combining this with an estimate from the proof of (ii), we conclude that
|fε(x)−f(x)| ÉΩnkφkL∞(Rn)
1
|B(x,ε)| ˆ
B(x,ε)|f(y)−f(x)|d y<ΩnkφkL∞(Rn)ε0 for everyx∈Ω0ifε<δ.
(4) LetΩ0bΩ00bΩ.
Claim: ˆ
Ω0|fε|pdxÉ ˆ
Ω00|f|pdx whenever 0<ε<dist(Ω0,ÇΩ00) and 0<ε<dist(Ω00,ÇΩ).
Reason. Takex∈Ω0. Hölder’s inequality implies
|fε(x)| =
¯
¯
¯
¯ ˆ
B(x,ε)φε(x−y)f(y)d y
¯
¯
¯
¯ É
ˆ
B(x,ε)φε(x−y)1−1pφε(x−y)1p|f(y)|d y É
µˆ
B(x,ε)φε(x−y)d y
¶p01 µˆ
B(x,ε)φε(x−y)|f(y)|pd y
¶1p
By raising the previous estimate to powerpand by integrating overΩ0, we obtain ˆ
Ω0|fε(x)|pdxÉ ˆ
Ω0
ˆ
B(x,ε)φε(x−y)|f(y)|pd y dx
= ˆ
Ω00
ˆ
Ω0φε(x−y)|f(y)|pdx d y
= ˆ
Ω00|f(y)|p ˆ
Ω0φε(x−y)dx d y
= ˆ
Ω00|f(y)|pd y.
Here we used Fubini’s theorem and once more the fact that the integral ofφεis
one. ■
SinceC(Ω00) is dense inLp(Ω00). Therefore for everyε0>0 there existsg∈C(Ω00) such that
µˆ
Ω00|f−g|pdx
¶1p Éε0
3. By (3), we havegε→guniformly inΩ0asε→0. Thus
µˆ
Ω00|gε−g|pdx
¶1p Ésup
Ω0 |gε−g|¯
¯Ω0¯¯1p<ε0 3,
whenε>0 is small enough. Now we use Minkowski’s inequality and the previous claim to conclude that
µˆ
Ω0|fε−f|pdx
¶1p É
µˆ
Ω0|fε−gε|pdx
¶1p
+ µˆ
Ω0|gε−g|pdx
¶1p +
µˆ
Ω0|g−f|pdx
¶1p
É2 µˆ
Ω00|g−f|pdx
¶1p +
µˆ
Ω0|gε−g|pdx
¶1p
É2ε0 3+ε0
3 =ε0.
Thusfε→f inLp(Ω0) asε→0. ä
1.7 Local approximation in Sobolev spaces
Next we show that the convolution approximation converges locally in Sobolev spaces.
Theorem 1.17. Letu∈Wk,p(Ω), 1Ép< ∞. then (1) Dαuε=Dαu∗φεinΩεand
(2) uε→uinWk,p(Ω0) for everyΩ0bΩ.
TH E M O R A L: Smooth functions are dense in local Sobolev spaces. Thus every Sobolev function can be locally approximated with a smooth function in the Sobolev norm.
Proof. (1) Fixx∈Ωε. Then
Dαuε(x)=Dα(u∗φε)(x)=(u∗Dαφε)(x)
= ˆ
ΩDαxφε(x−y)u(y)d y
=(−1)|α|
ˆ
ΩDαy(φε(x−y))u(y)d y.
Here we first used the proof of Lemma1.16(1) and then the fact that Ç
Çxj
³φ³x−y ε
´´
= − Ç Çxj
³φ³y−x ε
´´
= − Ç Çyj
³φ³x−y ε
´´
. For everyx∈Ωε, the functionϕ(y)=φε(x−y) belongs toC∞0(Ω). Therefore
ˆ
ΩDαy(φε(x−y))u(y)d y=(−1)|α|
ˆ
ΩDαu(y)φε(x−y)d y.
By combining the above facts, we see that Dαuε(x)=(−1)|α|+|α|
ˆ
Ω
Dαu(y)φε(x−y)d y=(Dαu∗φε)(x).
Notice that (−1)|α|+|α|=1.
(2) LetΩ0bΩ, and chooseε>0 s.t. Ω0⊂Ωε. By (i) we know thatDαuε= Dαu∗φεinΩ0,|α| Ék. By Lemma1.16, we haveDαuε→DαuinLp(Ω0) asε→0,
|α| Ék. Consequently
kuε−ukWk,p(Ω0)= Ã
X
|α|Ék
kDαuε−DαukLpp(Ω0)
!1p
→0.
ä
1.8 Global approximation in Sobolev spaces
The next result shows that the convolution approximation converges also globally in Sobolev spaces.
Theorem 1.18 (Meyers-Serrin). If u∈Wk,p(Ω), 1É p< ∞, then there exist functionsui∈C∞(Ω)∩Wk,p(Ω) such thatui→uinWk,p(Ω).
TH E M O R A L: Smooth functions are dense in Sobolev spaces. Thus every Sobolev function can be approximated with a smooth function in the Sobolev norm.
In particular, this holds true for the function with a dense infinity set in Example 1.9.
Proof. LetΩ0= ;and Ωi=
½
x∈Ω: dist(x,ÇΩ)>1 i
¾
∩B(0,i), i=1, 2, . . . . Then
Ω= [∞ i=1
Ωi and Ω1bΩ2b. . .bΩ.
Claim:There existηi∈C∞0 (Ωi+2\Ωi−1),i=1, 2, . . ., such that 0ÉηiÉ1 and X∞
i=1
ηi(x)=1 for every x∈Ω. This is a partition of unity subordinate to the covering {Ωi}.
Reason. By using the distance function and convolution approximation we can constructeηi∈C∞0 (Ωi+2\Ωi−1) such that 0ÉeηiÉ1 andeηi=1 inΩi+1\Ωi(exercise).
Then we define
ηi(x)= ηei(x) P∞
j=1eηj(x), i=1, 2, . . . .
Observe that the sum is only over four indices in a neighbourhood of a given
point. ■
Now by Lemma1.12(5),ηiu∈Wk,p(Ω) and supp(ηiu)⊂Ωi+2\Ωi−1. Letε>0. Chooseεi>0 so small that
supp(φεi∗(ηiu))⊂Ωi+2\Ωi−1
(see Remark1.15(4)) and
kφεi∗(ηiu)−ηiukWk,p(Ω)< ε
2i, i=1, 2, . . . .
By Theorem1.17(2), this is possible. Define v=
X∞ i=1
φεi∗(ηiu).
This function belongs toC∞(Ω), since in a neighbourhood of any pointx∈Ω, there are at most finitely many nonzero terms in the sum. Moreover,
kv−ukWk,p(Ω)=
°
°
°
°
° X∞ i=1
φεi∗(ηiu)− X∞ i=1
ηiu
°
°
°
°
°Wk,p(Ω) É
X∞ i=1
°
°φεi∗(ηiu)−ηiu°
°Wk,p(Ω)
É X∞ i=1
ε
2i=ε. ä
Remarks 1.19:
(1) The Meyers-Serrin theorem1.18gives the following characterization for the Sobolev spacesWk,p(Ω), 1Ép< ∞: u∈Wk,p(Ω) if and only if there exist functions ui∈C∞(Ω)∩Wk,p(Ω), i=1, 2, . . . , such that ui→u in Wk,p(Ω) asi→ ∞. In other words,Wk,p(Ω) is the completion ofC∞(Ω) in the Sobolev norm.
Reason. =⇒ Theorem1.18.
⇐= Theorem1.13. ■
(2) The Meyers-Serrin theorem1.18is false forp= ∞. Indeed, ifui∈C∞(Ω)∩ W1,∞(Ω) such that ui→u inW1,∞(Ω), thenu∈C1(Ω) (exercise). Thus special care is required when we consider approximations inW1,∞(Ω).
(3) LetΩ0bΩ. The proof of Theorem1.17and Theorem1.18shows that for everyε>0 there existsv∈C∞0(Ω) such thatkv−ukW1,p(Ω0)<ε.
(4) The proof of Theorem1.18shows that not onlyC∞(Ω) but alsoC0∞(Ω) is dense inLp(Ω), 1Ép< ∞.
1.9 Sobolev spaces with zero boundary values
In this section we study definitions and properties of first order Sobolev spaces with zero boundary values in an open subset ofRn. A similar theory can be developed for higher order Sobolev spaces as well. Recall that, by Theorem1.18, the Sobolev spaceW1,p(Ω) can be characterized as the completion ofC∞(Ω) with respect to the Sobolev norm when 1Ép< ∞.
Definition 1.20. Let 1Ép< ∞. The Sobolev space with zero boundary values W01,p(Ω) is the completion of C∞0 (Ω) with respect to the Sobolev norm. Thus u∈W01,p(Ω) if and only if there exist functionsui∈C∞0 (Ω),i=1, 2, . . . , such that ui→u inW1,p(Ω) as i→ ∞. The spaceW01,p(Ω) is endowed with the norm of W1,p(Ω).
TH E M O R A L: The only difference compared toW1,p(Ω) is that functions in W01,p(Ω) can be approximated byC∞0(Ω) functions instead ofC∞(Ω) functions, that is,
W1,p(Ω)=C∞(Ω) and W01,p(Ω)=C∞0 (Ω),
where the completions are taken with respect to the Sobolev norm. A function inW01,p(Ω) has zero boundary values in Sobolev’s sense. We may say thatu,v∈ W1,p(Ω) have the same boundary values in Sobolev’s sense, ifu−v∈W01,p(Ω). This is useful, for example, in Dirichlet problems for PDEs.
WA R N I N G: Roughly speaking a function inW1,p(Ω) belongs toW01,p(Ω), if it vanishes on the boundary. This is a delicate issue, since the function does not have to be zero pointwise on the boundary. We shall return to this question later.
Remark 1.21. W01,p(Ω) is a closed subspace ofW1,p(Ω) and thus complete (exer- cise).
Remarks 1.22:
(1) ClearlyC∞0(Ω)⊂W01,p(Ω)⊂W1,p(Ω)⊂Lp(Ω).
(2) Ifu∈W01,p(Ω), then the zero extensionue:Rn→[−∞,∞],
u(x)e =
u(x), x∈Ω, 0, x∈Rn\Ω, belongs toW1,p(Rn) (exercise).
Lemma 1.23. If u∈W1,p(Ω) and suppu is a compact subset of Ω, then u ∈ W01,p(Ω).
Proof. Letη∈C∞0(Ω) be a cutoff function such thatη=1 on the support ofu.
Claim:Ifui∈C∞(Ω),i=1, 2, . . . , such thatui→uinW1,p(Ω), thenηui∈C∞0 (Ω) converges toηu=uinW1,p(Ω).
Reason. We observe that kηui−ηukW1,p(Ω)=
³
kηui−ηukLpp(Ω)+ kD(ηui−ηu)kLpp(Ω)
´1p
É kηui−ηukLp(Ω)+ kD(ηui−ηu)kLp(Ω),