Part I, Chapter 2

Weak derivatives and Sobolev spaces

We investigate in this chapter the notion of differentiation for Lebesgue inte-

grable functions. We introduce an extension of the classical concept of deriva-
tive and partial derivative which is called weak derivative. This notion will be
used throughout the book. It is particularly useful when one tries to differen-
tiate finite element functions that are continuous and piecewise polynomial.
In that case, one does not need to bother about the points where the classical
derivative is multivalued to define the weak derivative. We also introduce the
concept of Sobolev spaces. These spaces are useful to study the well-posedness
of partial differential equations and their approximation using finite elements.

2.1 Differentiation
We study here the concept of differentiation for Lebesgue integrable functions.

2.1.1 Lebesgue points

Theorem 2.1 (Lebesgue points). Let f ∈ L1 (D). Let B(x, h) be the ball
of radius h > 0 centered at x ∈ D. The following holds true for a.e. x ∈ D:
Z
1
lim |f (y) − f (x)| dy = 0. (2.1)
h↓0 |B(x, h)| B(x,h)

Points x ∈ D where (2.1) holds true are called Lebesgue points of f.

Proof. See, e.g., Rudin [170, Thm. 7.6]. ⊓
⊔
This result says that for a.e. x ∈ D, the averages of |f (·) − f (x)| are
small over small balls centered at x, i.e., f does not oscillate too much in
the neighborhood of x. Notice that if the function f is continuous at x, then
x is a Lebesgue point of f (recall that a continuous function is uniformly
continuous over compact sets).
14 Chapter 2. Weak derivatives and Sobolev spaces

Let H ⊂ R be is a countable set with 0 as unique accumulation point (the

sign of the members of H is unspecified). Let F : R → R. We say that F is
strongly differentiable at x if the sequence ( F (x+h)−F
h
(x)
)h∈H converges.

RTheorem 2.2 (Lebesgue’s differentiation). Let f ∈ L1 (R). Let F (x) :=

x
−∞ f (t) dt. Then F is strongly differentiable at every Lebesgue point x of f ,
and at these points we have F ′ (x) = f (x).

Proof. See Exercise 2.2. ⊓

⊔

In the above theorem, we have F ′ (x) = f (x) for a.e. x in R. Thus, it is

tempting to move away from the classical sense of differentiation and view
F ′ as a Rfunction in L1 (R). If we could make sense of F ′ in L1 (R), then
x
F (x) = −∞ F ′ (t) dt would be an extension of the fundamental theorem of
calculus in Lebesgue spaces. As an example of this possibility, let f := 1[0,∞)
be the Heaviside function (i.e., f (x) := 1 if x ≥ 0 and f (x) := 0 otherwise).
RNotice that f 6∈ L1 (R) but f ∈ L1loc (R) (see Definition 1.29), and F (x) :=
x
−∞
f (t) dt is well defined. Then F (x) = 0 if x < 0 and F (x) = x if x > 0
(notice that 0 is not a Lebesgue point of f ; see Exercise 2.1). We would like
to say that F ′ = f in L1loc (R). The objective of the rest of this section is to
make sense of the above argument.

2.1.2 Weak derivatives

Definition 2.3 (Weak derivative). Let D be an open set in Rd . Let u, v ∈
L1loc (D). Let i ∈ {1:d}. We say that v is the weak partial derivative of u in
the direction i if
Z Z
u∂i ϕ dx = − vϕ dx, ∀ϕ ∈ C0∞ (D), (2.2)
D D

and we write ∂i u := v. Let α ∈ Nd be a multi-index. We say that v is the

weak α-th partial derivative of u and we write ∂1α1 . . . ∂dαd u := v if
Z Z
u∂1α1 . . . ∂dαd ϕ dx = (−1)|α| vϕ dx, ∀ϕ ∈ C0∞ (D), (2.3)
D D

where |α| := α1 + . . . + αd . Finally, we write ∂ α u := ∂1α1 . . . ∂dαd u, and we set

∂ (0,...,0) u := u.

Lemma 2.4 (Uniqueness). Let u ∈ L1loc (D). If u has a weak α-th partial
derivative, then it is uniquely defined.

Proof. Let v1 , v2 ∈ L1loc (D) be two weak α-th derivatives of u. We have

Z Z Z
v1 ϕ dx = (−1)|α| u∂ α ϕ dx = v2 ϕ dx, ∀ϕ ∈ C0∞ (D).
D D D
Part I. Elements of functional analysis 15

R
Hence, D (v1 − v2 )ϕ dx = 0. The vanishing integral theorem (Theorem 1.32)
implies that v1 = v2 a.e. in D. ⊓
⊔
If u ∈ C |α| (D), then the usual and the weak α-th partial derivatives are
identical. Moreover, it can be shown that if α, β ∈ Nd are multi-indices such
that αi ≥ βi for all i ∈ {1:d}, then if the α-th weak derivative of u exists in
L1loc (D), so does the β-th weak derivative. For instance, with d = 1 (writing
∂x instead of ∂1 ), if ∂xx u exists in L1loc (D), so does ∂x u; see Exercise 2.4.
Example 2.5 (1D). Let us revisit the heuristic argument at the end of
§2.1.1. Let D := (−1, 1). (i) Let us first consider a continuous function u ∈
C 0 (D; R), e.g., u(x) := 0 if x < 0 and u(x) := x otherwise. Then u has a
weak derivative. Indeed, let v ∈ L1 (D) be s.t. v(x) := 0 if x < 0 and v(x) := 1
otherwise. Let ϕ ∈ C0∞ (D). We have
Z 1 Z 1 Z 1 Z 1
u∂x ϕ dx = x∂x ϕ dx = − ϕ dx = − vϕ dx.
−1 0 0 −1

Hence, v is the weak derivative of u. (Notice that ṽ defined by ṽ(x) := 0 if

x < 0, ṽ(0) := 12 and ṽ(x) := 1 if x > 0 is also a weak derivative of u, but
v = ṽ a.e. in D, i.e., v and ṽ coincide in the Lebesgue sense.) (ii) Let us
now consider a function u ∈ L1 (D; R) that is piecewise smooth but exhibits
a jump at x = 0, e.g., u(x) := −1 if x < 0 and u(x) := x otherwise. Then u
does not have a weak derivative. Let us prove this statement by contradiction.
Assume that there is v ∈ L1loc (D) s.t. ∂x u = v. We have
Z 1 Z 1 Z 0 Z 1 Z 1
vϕ dx = − u∂x ϕ dx = ∂x ϕ dx − x∂x ϕ dx = ϕ(0) + ϕ dx,
−1 −1 −1 0 0

for all ϕ ∈ C0∞ (D). Let {ϕn }n∈N be a sequence of functions in C0∞ (D) s.t.
0 ≤ ϕn (x) ≤ 1 for all x ∈ D, ϕn (0) = 1, and ϕn → 0 a.e. in D. Lebesgue’s
R1
dominated convergence theorem implies that 1 = limn→∞ ( −1 vϕn dx −
R1
0
ϕn dx) = 0, which is a contradiction. ⊓
⊔
Lemma 2.6 (Passing to the limit). Let {vn }n∈N be a sequence in Lp (D),
p ∈ [1, ∞], with weak α-th partial derivatives {∂ α vn }n∈N in Lp (D). Assume
that vn → v in Lp (D) and ∂ α vn → gα in Lp (D). Then v has a weak α-th
partial derivative and ∂ α v = gα .
R R
Proof. The assumptions imply that limn→∞ D ∂ α vn ϕ dx = D gα ϕ dx and
Z Z Z
lim ∂ α vn ϕ dx = (−1)|α| lim vn ∂ α ϕ dx = (−1)|α| v∂ α ϕ dx,
n→∞ D n→∞ D D

for all ϕ ∈ C0∞ (D). The conclusion follows readily. ⊓

⊔
A function v ∈ L1loc (D)
is said to be locally Lipschitz in D if for all x ∈
D, there is a neighborhood Nx of x in D and a constant Lx such that
|v(z) − v(y)| ≤ Lx kz − ykℓ2 (Rd ) for all y, z ∈ Nx .
16 Chapter 2. Weak derivatives and Sobolev spaces

Theorem 2.7 (Rademacher). Let D be an open set in Rd . Let f be a

locally Lipschitz function in D. Then f is differentiable in the classical sense
a.e. in D. The function f is also weakly differentiable, and the classical and
weak derivatives of f coincide a.e. in D.

Proof. See [99, p. 280], [138, p. 44]. ⊓

⊔

2.2 Sobolev spaces

In this section, we introduce integer-order and fractional-order Sobolev
spaces. The scale of Sobolev spaces plays a central role in the finite element
error analysis to quantify the decay rate of the approximation error.

2.2.1 Integer-order spaces

Definition 2.8 (W m,p (D)). Let m ∈ N and p ∈ [1, ∞]. Let D be an open set
in Rd . We define the Sobolev space

W m,p (D) := {v ∈ L1loc (D) | ∂ α v ∈ Lp (D), ∀α ∈ Nd s.t. |α| ≤ m}, (2.4)

where the derivatives are weak partial derivatives. We write W m,p (D; Rq ),
q ≥ 1, for the space composed of Rq -valued functions whose components are
all in W m,p (D), and we write W m,p (D) whenever q = d.

Whenever it is possible to identify a length scale ℓD associated with D,

e.g., its diameter ℓD := diam(D) if D is bounded, we equip W m,p (D) with
the following norm and seminorm: If p ∈ [1, ∞), we set
X |α|p X
kvkpW m,p (D) := ℓD k∂ α vkpLp (D) , |v|pW m,p (D) := k∂ α vkpLp (D) ,
|α|≤m |α|=m

and if p = ∞, we set
|α|
kvkW m,∞ (D) := max ℓD k∂ α vkL∞ (D) , |v|W m,∞ (D) := max k∂ α vkL∞ (D) ,
|α|≤m |α|=m

where the sums and the maxima run over multi-indices α ∈ Nd . The advan-
tage of using the factor ℓD is that all the terms in the sums or maxima have
the same dimension (note that k·kW m,p (D) and |·|W m,p (D) have a different
scaling w.r.t. ℓD ). If there is no length scale available or if one works with
dimensionless space variables, one sets ℓD := 1 in the above definitions.

Proposition 2.9 (Banach space). W m,p (D) equipped with the k·kW m,p (D) -
norm is a Banach space. For p = 2, the space

H m (D) := W m,2 (D) (2.5)

Part I. Elements of functional analysis 17

is a real RHilbert space when equipped with the inner product (v, w)H m (D) :=
P α α m
|α|≤m D ∂ v ∂ w dx. Similarly, H (D; C) is a complex Hilbert space when
P R
equipped with the inner product (v, w)H m (D;C) := |α|≤m D ∂ α v ∂ α w dx.

Proof. We are going to do the proof for m = 1. See e.g., [3, Thm. 3.3], [99,
p. 249], or [189, Lem. 5.2] for the general case. Let {vn }n∈N be a Cauchy
sequence in W 1,p (D). Then {vn }n∈N is a Cauchy sequence in Lp (D) and the
sequences of weak partial derivatives {∂i vn }n∈N are also Cauchy sequences
in Lp (D). Hence, there is v ∈ Lp (D) and there are g1 , . . . , gd ∈ Lp (D) such
that vn → v in Lp (D) and ∂i vn → gi in Lp (D). We conclude by invoking
Lemma 2.6. ⊓
⊔
Example 2.10 (H 1 (D)). Taking m := 1 and p := 2 we have

H 1 (D) := {v ∈ L2 (D) | ∂i v ∈ L2 (D), ∀i ∈ {1:d}},

(notice that L2 (D) ⊂ L1loc (D)) and

  12 X
kvkH 1 (D) := kvk2L2 (D) + ℓ2D |v|2H 1 (D) , |v|2H 1 (D) := k∂i vk2L2 (D) .
i∈{1: d}

Let ∇v be the column vector in Rd whose components are the directional

weak derivatives ∂i v of v. Then a more compact notation is H 1 (D) := {v ∈
L2 (D) | ∇v ∈ L2 (D)} and |v|H 1 (D) := k∇vkL2 (D) . ⊓
⊔

Lemma 2.11 (Kernel of ∇). Let D be open and connected set in Rd . Let
v ∈ W 1,p (D), p ∈ [1, ∞]. Then ∇v = 0 a.e. on D iff v is constant.
Proof. We prove the result for D := (−1, 1) and we refer the reader to [138,
p. 24], [189, Lem. 6.4] for the general case.RLet u ∈ L1loc (D) be such that
R ∂x u =
0. Fix a function ρ ∈ C0∞ (D) such thatR D ρ dx = 1 and set cρ := D uρ dx.
Let now ϕ ∈ C0∞ (D) and set cϕ := D ϕ dx. Then the function ψ(x) :=
Rx
(ϕ(y) − cϕ ρ(y)) dy is by construction in C0∞ (D), and we have ∂x ψ(x) =
−1 R R
ϕ(x) − cϕ ρ(x). Since D u∂x ψ dx = − D (∂x u)ψ dx = 0 by assumption on
∂x u, we infer that
Z Z Z Z
uϕ dx = u(∂x ψ + cϕ ρ) dx = cϕ uρ dx = cρ ϕ dx,
D D D D

for all ϕ ∈ C0∞ (D). Theorem 1.32 shows that u = cρ . ⊓

⊔

Remark 2.12 (Lipschitz functions). Let D be an open set in Rd . The

space of Lipschitz functions C 0,1 (D) is closely related to the Sobolev space
W 1,∞ (D). Indeed, C 0,1 (D)∩L∞ (D) is continuously embedded into W 1,∞ (D).
Conversely, if v ∈ W 1,∞ (D), then |v(y) − v(z)| ≤ dD (y, z)k∇vkL∞ (D) for all
y, z ∈ D, where dD (y, z) denotes the geodesic distance of y to z in D, i.e.,
the shortest length of a smooth path connecting y to z in D (if D is convex,
18 Chapter 2. Weak derivatives and Sobolev spaces

dD (y, z) = ky − zkℓ2 ); see [189, Lem. 7.8]. A set D ⊂ Rd is said to be

quasiconvex if there exists C ≥ 1 s.t. every pair of points x, y ∈ D can be
joined by a curve γ in D with length(γ) ≤ Ckx − ykℓ2 . If D is a quasiconvex
open set, then W 1,∞ (D) = C 0,1 (D) ∩ L∞ (D), and if D is also bounded, then
W 1,∞ (D) = C 0,1 (D); see Heinonen [113, Thm. 4.1]. ⊓
⊔
Remark 2.13 (Broken seminorms). Let D ⊂ Rd be an open set and let
{Di }i∈{1: I} be a partition
S of D, i.e., all the subsets Di are open, mutually
disjoint, and D \ i∈{1: I} Di has zero Lebesgue measure. Let v ∈ W 1,p (D)
P
and p ∈ [1, ∞). Then one can write |v|pW 1,p (D) = i∈{1: I} k(∇v)|Di kpLp (Di ) .
In this book, we are going to abuse the notation by writing |v|pW 1,p (D) =
P p
i∈{1: I} k∇vkLp (Di ) . This abuse is justified by observing that (∇v)|Di =
1,1
∇(v|Di ) for all v ∈ Wloc (D). We stress that it is important that the
weak derivative of v exists to make sense of the above identities. For in-
stance, letting H be the Heaviside function, we have k∇(H|(−1,0) )kpLp (−1,0) +
k∇(H|(0,1) )kpLp (0,1) = 0, but H 6∈ W 1,p (D); see Exercise 2.8. ⊓
⊔

2.2.2 Fractional-order spaces

Definition 2.14 (W s,p (D)). Let s ∈ (0, 1) and p ∈ [1, ∞]. Let D be an open
set in Rd . We define W s,p (D) := v ∈ Lp (D) | |v|W s,p (D) < ∞ , where
Z Z  p1
|v(x) − v(y)|p
|v|W s,p (D) := dx dy , p < ∞, (2.6)
D D kx − yksp+d
ℓ2

|v(x)−v(y)|
and |v|W s,∞ (D) := ess supx,y∈D kx−yks2 . Letting now s > 1, we define
ℓ

W s,p (D) := {v ∈ W m,p (D) | ∂ α v ∈ W σ,p (D), ∀α, |α| = m}, (2.7)

where m := ⌊s⌋ and σ := s − m. Finally, we denote H s (D) := W s,2 (D).

We write W s,p (D; Rq ), q ≥ 1, for the space composed of Rq -valued functions
whose components are all in W s,p (D), and we write W s,p (D) whenever q = d.
Definition 2.15 (Sobolev–Slobodeckij norm). Let s = m + σ with m :=
⌊s⌋ and σ := s − m ∈ (0, 1). For all p ∈ [1, ∞) and all v ∈ W s,p (D), we
set kvkpW s,p (D) := kvkpW m,p (D) + ℓsp p p
D |v|W s,p (D) with seminorm |v|W s,p (D) :=
P α p
|α|=m |∂ v|W σ,p (D) . We also set

kvkW s,∞ (D) := max(kvkW m,∞ (D) , ℓsD |v|W s,∞ (D) ),

with seminorm |v|W s,∞ (D) := max|α|=m |∂ α v|W σ,∞ (D) . Equipped with this
norm W s,p (D) is a Banach space (and a Hilbert space if p = 2).
Example 2.16 (Power functions). Let D := (0, 1) and consider the func-
tion v(x) := xα with α ∈ R. One can verify that v ∈ L2 (D) if α > − 12 ,
v ∈ H 1 (D) if α > 21 , and, more generally v ∈ H s (D) if α > s − 12 . ⊓
⊔
Part I. Elements of functional analysis 19

Example 2.17 (Hölder functions). If D is bounded and p ∈ [1, ∞), then

C 0,α (D) ֒→ W s,p (D) provided 0 ≤ s < α ≤ 1; see Exercise 2.9. ⊓
⊔

Example 2.18 (Step function). Let D := (−1, 1) and consider v(x) := 0

if x < 0 and v(x) := 1 if x ≥ 0. Then v ∈ W s,p (D) iff sp < 1 as shown by the
following computation (notice that sp > 0):
Z 0Z 1 Z 0  
1 1 1 1
|v|pW s,p = 2 sp+1
dx dy = 2 − − dx.
−1 0 |y − x| −1 sp (1 − x)sp |x|sp
R0 1
The integral −1 |x|sp
dx is convergent if and only if sp < 1. ⊓
⊔

Remark 2.19 (Limits s↓0 and s↑1). The expression (2.6), which is usually
adopted in the literature to define |v|W s,p (D) , gives |v|W s,p (D) → ∞ as s↑1
even if v ∈ W 1,p (D). A remedy to this deficiency has been proposed in
Bourgain et al. [38], Maz’ya and Shaposhnikova [140]. It is shown in [38]
1
that by redefining |v|∗W s,p := (1−s) p |v|W s,p for all s ∈ (0, 1), and setting
|v|∗W 1,p (D) := |v|W 1,p (D) , there exists c, s.t. for all σ, s with 0 < σ < s ≤ 1 and
all v ∈ W σ,p (D), one has |v|∗W σ,p (D) ≤ c|v|∗W s,p (D) and lims→1 |v|∗W s,p (D) =
|v|W 1,p (D) (see Borthagaray and Ciarlet [34, Rmk. 2.3]). It has been proposed
1
[140, Thm. 3] to redefine |v|∗W s,p (D) := (s(1−s)) p |v|W s,p (D) to improve also
the behavior of the seminorm when s↓0. It is shown therein that if there is
W σ,p
σ > 0 s.t. v ∈ C0∞ (Rd ) , then lims↓0 s|v|pW s,p (Rd ) = 2p−1 |S d−1 |kvkpLp (Rd ) ,
where |S d−1 | is the measure of the unit sphere in Rd . ⊓
⊔

Remark 2.20 (Definition by interpolation). Fractional-order Sobolev

spaces can also be defined by means of the interpolation theory between
Banach spaces (see §A.5). Let p ∈ [1, ∞) and s ∈ (0, 1). Then we have

W s,p (D) = [Lp (D), W 1,p (D)]s,p ,

and more generally W m+s,p (D) = [W m,p (D), W m+1,p (D)]s,p for all m ∈ N,
with equivalent norms in all the cases; see Tartar [189, Lem. 36.1]. Using the
interpolation theory may not be convenient in finite element analysis if one
is interested in local approximation properties. Unless specified otherwise we
use the Sobolev–Slobodeckij norm in the book. ⊓
⊔

2.3 Key properties: density and embedding

This section reviews some key properties of Sobolev spaces: the density of
smooth functions and the (compact) embedding into Lebesgue spaces or into
spaces composed of Hölder continuous functions.
20 Chapter 2. Weak derivatives and Sobolev spaces

2.3.1 Density of smooth functions

Theorem 2.21 (Meyers–Serrin). Let D be an open set in Rd . Let s ≥ 0
and p ∈ [1, ∞). Then C ∞ (D) ∩ W s,p (D) is dense in W s,p (D).

Proof. See Meyers and Serrin [143] and Adams and Fournier [3, Thm. 3.17];
see also Evans [99, p. 251] for bounded D. ⊓
⊔

Remark 2.22 (p = ∞). Let m ∈ N. The closure of C ∞ (D) ∩ W m,∞ (D)

with respect to the Sobolev norm k·kW m,∞ (D) differs from W m,∞ (D) since
it is composed of functions whose derivatives up to order m are continuous
and bounded on D. ⊓
⊔

The density of smooth functions in Sobolev spaces allows one to derive

many useful results. We list here some of the most important ones.

Corollary 2.23 (Differentiation of a product). Let D be an open subset

of Rd . Then we have uv ∈ W 1,p (D) ∩ L∞ (D) and ∇(uv) = v∇u + u∇v for
all u, v ∈ W 1,p (D) ∩ L∞ (D) and all p ∈ [1, ∞].

Proof. See, e.g., [48, Prop. 9.4, p. 269]. ⊓

⊔

Corollary 2.24 (Differentiation of a composition). Let D ⊂ Rd be an

open set. Let G ∈ C 1 (R). Assume that G(0) = 0 and there is M < ∞
such that |G′ (t)| ≤ M for all t ∈ R. Then we have G(u) ∈ W 1,p (D) and
∇(G(u)) = G′ (u)∇u for all u ∈ W 1,p (D) and all p ∈ [1, ∞].

Proof. See, e.g., [48, Prop. 9.5, p. 270]. ⊓

⊔

Corollary 2.25 (Change of variable). Let D, D′ be two open subsets of

Rd . Assume that there exists a bijection T : D′ → D s.t. T ∈ C 1 (D′ ; D),
T −1 ∈ C 1 (D; D′ ), DT ∈ L∞ (D′ ; Rd×d ), and DT −1 ∈ L∞ (D; Rd×d ), where
DT and DT −1 are the Jacobian matrices of T and T −1 , respectively. Then
we have u ◦ T ∈PW 1,p (D′ ) for all u ∈ W 1,p (D) and all p ∈ [1, ∞], and
∂x′i (u ◦ T )(x′ ) = j∈{1: d} ∂xj u(T (x))∂x′i T (x′ ) for all i ∈ {1:d} and x′ ∈ D′ .

Proof. See, e.g., [48, Prop. 9.6, p. 270]. ⊓

⊔

2.3.2 Embedding
We use the notation V ֒→ W to mean that the embedding of V into W is
continuous, i.e., there is c such that kvkW ≤ ckvkV for all v ∈ V (see §A.2).
The main idea of the results in this section is that functions in the Sobolev
space W s,p (D) with differentiability index s > 0 do have an integrability
index larger than p (i.e., they belong to some Lebesgue space Lq (D) with
q > p), and if s is sufficiently large, for all u ∈ W s,p (D) (recall that u is
actually a class of functions that coincide almost everywhere in D), there is
a representative of u that is continuous (or even Hölder continuous). How
Part I. Elements of functional analysis 21

large s must be for these properties to hold true depends on the space di-
mension. The case d = 1 is particularly simple since W 1,1 (R) ֒→ C 0 (R) and
W 1,1 (D) ֒→ C 0 (D) for every bounded interval D; see [189, Lem. 8.5] (see
also Exercise 5.7). In the rest of this section, we assume that d ≥ 2. We first
consider the case where D := Rd .
Theorem 2.26 (Embedding of W 1,p (Rd )). Let d ≥ 2 and let p ∈ [1, ∞].
The following holds true:
(i) (Gagliardo–Nirenberg–Sobolev): If p ∈ [1, d), then

pd
W 1,p (Rd ) ֒→ Lq (Rd ), ∀q ∈ [p, p∗ ], p∗ := . (2.8)
d−p
∗
In particular, kukLp∗ (Rd ) ≤ p1∗ k∇ukLp(Rd ) with 1∗ := d−1
d
for all u ∈
∗
W (R ). Hence, W (R ) ֒→ L (R ), and the embedding into Lq (Rd )
1,p d 1,p d p d

for all q ∈ [p, p∗ ) follows from Corollary 1.43.

(ii) If p = d, then

W 1,d (Rd ) ֒→ Lq (Rd ), ∀q ∈ [d, ∞). (2.9)

(iii) (Morrey): If p ∈ (d, ∞], then

d
W 1,p (Rd ) ֒→ L∞ (Rd ) ∩ C 0,α (Rd ), α := 1 − . (2.10)
p

Proof. See [48, Thm. 9.9, Cor. 9.11, Thm. 9.12], [99, p. 263–266], [180, §I.7.4,
§I.8.2], [189, Chap. 8-9]. ⊓
⊔
Remark 2.27 (Continuous function). The embedding (2.10) means that
there is c, only depending on p and d, such that

|u(x) − u(y)| ≤ ckx − ykα

ℓ2 (Rd ) k∇ukLp (Rd ) , for a.e. x, y ∈ Rd , (2.11)

for all u ∈ W 1,p (Rd ). In other words, there is a continuous function v ∈

C 0,α (Rd ) such that u = v almost everywhere. It is then possible to replace u
by its continuous representative v. We will systematically do this replacement
in this book when a continuous embedding in a space of continuous functions
is invoked. ⊓
⊔
The above results extend to Sobolev spaces of arbitrary order.
Theorem 2.28 (Embedding of W s,p (Rd )). Let d ≥ 2, s > 0, and p ∈
[1, ∞]. The following holds true:

 q d pd
L (R )
 ∀q ∈ [p, d−sp ], if sp < d,
W s,p (Rd ) ֒→ Lq (Rd ) ∀q ∈ [p, ∞), if sp = d, (2.12)


L∞ (Rd ) ∩ C 0,α (Rd ) α := 1 − d
sp , if sp > d.
22 Chapter 2. Weak derivatives and Sobolev spaces

Moreover, W d,1 (Rd ) ֒→ L∞ (Rd ) ∩ C 0 (Rd ) (case s = d and p = 1).

Proof. See [110, Thm. 1.4.4.1], [88, Thm. 4.47] for p ∈ (1, ∞). For s = d and
p = 1, see, e.g., Ponce and Van Schaftingen [160] and Campos Pinto [55,
Prop. 3.4] (if d = 2). ⊓
⊔
Our aim is now to generalize Theorem 2.28 to the space W s,p (D), where
D is an open set in Rd . A rather generic way to proceed is to use the concept
of extension.
Definition 2.29 ((s, p)-extension). Let s > 0 and p ∈ [1, ∞]. Let D be an
open set in Rd . The set D is said to have the (s, p)-extension property if there
is a bounded linear operator E : W s,p (D) → W s,p (Rd ) such that E(u)|D = u
for all u ∈ W s,p (D).
Theorem 2.28 can be restated by replacing Rd with any open set D in Rd
that has the (s, p)-extension property. A rather general class of sets that we
consider in this book is that of Lipschitz sets in Rd . A precise definition is
given in the next chapter. At this stage, it suffices to know that the boundary
of a Lipschitz set can be viewed as being composed of a finite collection of
epigraphs of Lipschitz functions.
Theorem 2.30 (Extension from Lipschitz sets). Let s > 0 and p ∈
[1, ∞]. Let D be an open, bounded subset of Rd . If D is a Lipschitz set, then
it has the (s, p)-extension property.
Proof. See Calderón [54], Stein [181, p. 181] (for s ∈ N), [110, Thm. 1.4.3.1]
and [88, Prop. 4.43] (for p ∈ (1, ∞)), [141, Thm. A.1&A.4] (for s ∈ [0, 1],
p ∈ [1, ∞] and s > 0, p = 2) [189, Lem. 12.4] (for s = 1). ⊓
⊔
Theorem 2.31 (Embedding of W s,p (D)). Let d ≥ 2, s > 0, and p ∈
[1, ∞]. Let D be an open, bounded subset of Rd . If D is a Lipschitz set, then
we have

 q pd
L (D)
 ∀q ∈ [p, d−sp ], if sp < d,
s,p q
W (D) ֒→ L (D) ∀q ∈ [p, ∞), if sp = d, (2.13)


L∞ (D) ∩ C 0,α (D) α := 1 − d , if sp > d.
sp

Moreover, W d,1 (D) ֒→ L∞ (D) ∩ C 0 (D) (case s = d and p = 1).

Remark 2.32 (Bounded set). Note that W s,p (D) ֒→ Lq (D) for sp ≤ d
and all q ∈ [1, p] since D is bounded. The boundedness of D also implies that
d
W s,p (D) ֒→ C 0,α (D), with sp > d and α := 1 − sp , and W d,1 (D) ֒→ C 0 (D),
i.e., there is (Hölder-)continuity up to the boundary. ⊓
⊔
Example 2.33 (Embedding into continuous functions). In dimension
one, functions in H 1 (D) are bounded and continuous, whereas this may not
be the case in dimension d ≥ 2 (see Exercise 2.10). In dimension d ∈ {2, 3},
Theorem 2.31 says that functions in H 2 (D) are bounded and continuous. ⊓ ⊔
Part I. Elements of functional analysis 23

Example 2.34 (Boundary smoothness). Let α > 1, p ∈ [1, 2), and D :=

β 1+α
{(x1 , x2 ) ∈ R2 | x1 ∈ (0, 1), x2 ∈ (0, xα
1 )}. Let u(x1 , x2 ) := x1 with 1− p <
β < 0 (this is possible since p < 2 < 1+α). Then u ∈ W 1,p (D) and u ∈ Lq (D)
for all q such that 1 ≤ q < pα where p1α := p1 − 1+α 1
. Let us set p1∗ := p1 − 12 ,
β−1
ǫ := 1+α + p1 > 0, and p1β := p1α − ǫ. Notice that pα < p∗ and also pα < pβ
since ǫ > 0. Since one can choose β s.t. ǫ is arbitrarily close to zero, we pick
β so that pβ < p∗ . Hence, pβ ∈ (pα , p∗ ). But u 6∈ Lq (D) for all q ∈ (pβ , p∗ ),
which would contradict Theorem 2.31 if the Lipschitz property had been
omitted. Hence, D cannot be a Lipschitz set in R2 (α > 1 means that D
has a cusp at the origin). This counterexample shows that some smoothness
assumption on D is needed for Theorem 2.31 to hold true. ⊓
⊔
We conclude this section with important compactness results. Recall
from §A.4 that the embedding V ֒→ W between two Banach spaces is com-
pact iff from every bounded sequence in V, one can extract a converging
subsequence in W.
Theorem 2.35 (Rellich–Kondrachov). Let s > 0 and p ∈ [1, ∞]. Let D
be an open, bounded subset of Rd . If D is a Lipschitz set, then the following
embeddings are compact:
pd
(i) If sp ≤ d, W s,p (D) ֒→ Lq (D) for all q ∈ [1, d−sp ).
s,p 0
(ii) If sp > d, W (D) ֒→ C (D).
′
(iii) W s,p (D) ֒→ W s ,p (D) for all s > s′ .
Proof. See [3, Thm. 6.3], [48, Thm. 9.16], [99, p. 272], [138, p. 35], [110,
Thm. 1.4.3.2]. ⊓
⊔

Exercises
Exercise 2.1 (Lebesgue point). Let a ∈ R. Let f : R → R be defined by
f (x) := 0 if x < 0, f (0) := a, and f (x) := 1 if x > 0. Show that 0 is not a
Lebesgue point of f for all a.
Exercise 2.2 (Lebesgue differentiation). The goal is to prove Theo-
rem 2.2. (i) Let h ∈ H (the sign of h is unspecified). Show that R(x, h) :=
F (x+h)−F (x) R x+h
h − f (x) = h1 x (f (t) − f (x)) dt. (ii) Conclude.
Exercise 2.3 (Lebesgue measure and weak derivative). Let D :=
(0, 1). Let C∞ be the Cantor set (see Example 1.5). Let f : D → R be
defined by f (x) := x if x 6∈ C∞ , and f (x) := 2−5x if x ∈ C∞ . (i) Is f measur-
able? (Hint : use Corollary 1.11.) (ii) Compute supx∈D f (x), ess supx∈D f (x),
inf x∈D f (x), ess inf x∈D f (x), and kf kL∞ (D) . (iii) Show that f is weakly dif-
Rx
ferentiable and compute ∂x f (x). (iv) Compute f (x) − 0 ∂t f (t) dt for all
c 0 c
x ∈ D. (iv) Identify R xthe function f ∈ C (D) that satisfies f = f a.e. on D?
c
Compute f (x) − 0 ∂t f (t) dt for all x ∈ D.
24 Chapter 2. Weak derivatives and Sobolev spaces

Exercise 2.4 (Weak derivative). Let D := (−1, 1). Prove that if u ∈

L1loc (D) has a second-order weak derivative,
Rx it also has a first-order weak
derivative. (Hint : consider ψ(x) := −1 (ϕ(t) − cϕ ρ(t)) dt for all ϕ ∈ C0∞ (D),
R R
with cϕ := D ϕ dx, ρ ∈ C0∞ (D), and D ρ dx = 1.)

Exercise 2.5 (Clairaut’s theorem). Let v ∈ L1loc (D). Let α, β ∈ Nd and

assume that the weak derivatives ∂ α v, ∂ β v exist and that the weak derivative
∂ α (∂ β v) exists. Prove that ∂ β (∂ α v) exists and ∂ α (∂ β v) = ∂ β (∂ α v).

Exercise 2.6 (Weak and classical derivatives). Let k ∈ N, k ≥ 1, and

let v ∈ C k (D). Prove that, up to the order k, the weak derivatives and the
classical derivatives of v coincide.

Exercise 2.7 (H 1 (D)). (i) Let D := (−1, 1) and u : D → R s.t. u(x) :=

3
|x| 2 − 1. Determine whether u is a member of H 1 (D; R). (ii) Let u1 ∈
C 1 ((−1, 0]; R) and u2 ∈ C 1 ([0, 1); R) and assume that u1 (0) = u2 (0). Let
u be such that u|(−1,0) := u1 and u|(0,1) := u2 . Determine whether u is a
member of H 1 (D; R). Explain why u 6∈ H 1 (D; R) if u1 (0) 6= u2 (0).

Exercise 2.8 (Broken seminorm). Let D be an open set in Rd . Let

{D1 , . . . , Dn } be a partition of D as in Remark 2.13. (i) Show that (∇v)|Di =
1,1
∇(v|Di ) for all i ∈ {1:n} and P all v ∈ Wploc (D). (ii) Letp p ∈ [1, ∞) and
1,p
v ∈ W (D). Show that i∈{1: n} |v|Di |W 1,p (Di )P= |v|W 1,p (D) . (iii) Let
s ∈ (0, 1), p ∈ [1, ∞), and v ∈ W s,p (D). Prove that i∈{1: n} |v|Di |pW s,p (Di ) ≤
|v|pW s,p (D) .

Exercise 2.9 (W s,p ). Let D be a bounded open set in Rd . Let α ∈ (0, 1].
Show that C 0,α (D; R) ֒→ W s,p (D; R) for all p ∈ [1, ∞) if s ∈ [0, α).

Exercise 2.10 (Unbounded function in H 1 (D)). Let D := B(0, 12 ) ⊂ R2

1
be the ball centered at 0 and of
radius 2 . (i) Show that the (unbounded)
function u(x) := ln −ln(kxkℓ2 ) has weak partial derivatives. (Hint : work
on D\B(0, ǫ) with ǫ ∈ (0, 12 ), and use Lebesgue’s dominated convergence
theorem.) (ii) Show that u is in H 1 (D).

Exercise 2.11 (Equivalent norm). Let m ∈ N, m ≥ 2, and let p ∈

1
[1, ∞). Prove that the norm kvk := (kvkpLp + ℓmp p
D |v|W m,p (D) )
p is equiva-

m,p
lent to the canonical norm in W (D). (Hint : use the Peetre–Tartar lemma
(Lemma A.20) and invoke the compact embeddings from Theorem 2.35.)