Part I, Chapter 4

Distributions and duality in Sobolev spaces

The dual space of a Sobolev space is not only composed of functions (defined
almost everywhere), but this space also contains more sophisticated objects
called distributions, which are defined by their action on smooth functions
with compact support. For instance the function x1 is not in L1 (0, 1), but the
R1
map ϕ 7→ 0 x1 ϕ(x) dx can be given a meaning for every smooth function
that vanishes at 0. Dual Sobolev spaces are useful to handle singularities
on the right-hand side of PDEs. They are also useful to give a meaning
to the tangential and the normal traces of Rd -valued fields that are not in
W s,p (D; Rd ) with sp > 1. The extension is done in this case by invoking
integration by parts formulas involving the curl or the divergence operators.

4.1 Distributions
The notion of distribution is a powerful tool that extends the concept of
integrable functions and weak derivatives. In particular we will see that every
distribution is differentiable in some reasonable sense.

Definition 4.1 (Distribution). Let D be an open set in Rd . A linear map

T : C0∞ (D) ∋ ϕ 7−→ hT, ϕi := T (ϕ) ∈ R or C, (4.1)

is called distribution in D if for every compact subset K of D, there exist an

integer p and a real number c (both can depend on K) s.t. for all ϕ ∈ C0∞ (D)
with supp(ϕ) ⊂ K, we have
|α|

|hT, ϕi| ≤ c max ℓD k∂ α ϕkL∞ (K) . (4.2)
|α|≤p

Example 4.2 (Locally integrable functions). Every function v in L1loc (D)

can be identified with the following distribution:
50 Chapter 4. Distributions and duality in Sobolev spaces
Z
Tv : C0∞ (D) ∋ ϕ 7−→ hTv , ϕi := vϕ dx.
D

This identification is possible owing to Theorem R 1.28, since

R two functions
v, w ∈ L1loc (D) are such that v = w a.e. in D iff D vϕ dx = D wϕ dx for all
ϕ ∈ C0∞ (D). We will abuse the notation by writing v instead of Tv . Notice
that the identification is also compatible with the Riesz–Fréchet theorem
(Theorem 1.37) in L2 (D), which allows one to identify L2 (D) with its dual
space by means of the L2 -inner product. ⊓
⊔

Example 4.3 (Dirac mass or measure). Let a be a point in D. The Dirac

mass (or Dirac measure) at a is the distribution defined by hδa , ϕi := ϕ(a)
for all ϕ ∈ C0∞ (D). There is no 1
R function f ∈ Lloc (D) such that δa = Tf .
Otherwise one would have 0 = D f ϕ dx for all ϕ ∈ C0∞ (D\{a}), and owing
to Theorem 1.28, this would imply that f = 0 a.e. in D\{a}, i.e., f = 0 a.e. in
D. Hence δa 6∈ T (L1loc (D)). This example shows that there are distributions
that cannot be identified with functions in L1loc (D). ⊓
⊔

Definition 4.4 (Distributional derivative). Let T be a distribution in D

and let i ∈ {1:d}. The distributional derivative ∂i T is the distribution in D
such that h∂i T, ϕi := −hT, ∂i ϕi for all ϕ ∈ C0∞ (D). More generally, for a
multi-index α ∈ Nd , the distributional derivative ∂ α T is the distribution in
D acting as h∂ α T, ϕi := (−1)|α| hT, ∂ α ϕi. We set conventionally ∂ 0 T := T ,
and ∇T := (∂1 T, . . . , ∂d T )T .

Example 4.5 (Weak derivative). The notion of distributional derivative

extends the notion of weak derivative. Let v ∈ L1loc (D) and assume that v
has a weak α-th partial derivative, say ∂ α v ∈ L1loc (D). Just like in Exam-
ple 4.2, we can identify
R v and ∂ α v with the distributions
R Tv and T∂ α v such
that hTv , ψi := D vψ dx and hT∂ α v , ϕi := (−1)|α| D v∂ α ϕ dx. This implies
that hT∂ α v , ϕi = (−1)|α| hTv , ∂ α ϕi, which according to Definition 4.4 shows
that ∂ α Tv = T∂ α v , i.e., the distributional derivative of Tv is equal to the
distribution associated with the weak derivative of v. ⊓
⊔

Example 4.6 (Step function). Let D := (−1, 1). Let w ∈ L1 (D) be

defined by w(x) := −1 if x < 0 and w(x) = 1 otherwise. Then for any ϕ ∈
R R0 R1
C0∞ (D) we have − D w∂x ϕ dx = −1 ∂x ϕ dx − 0 ∂x ϕ dx = 2ϕ(0) = hδ0 , ϕi.
This shows that the distributional derivative of w is the Dirac mass at 0; we
write ∂x w = δ0 . As established in Example 4.3, δ0 cannot be identified with
any function in L1loc (D); hence, w does not have a weak derivative but w has a
distributional derivative. Consider now the function v(x) := 1 − |x| in L1 (D).
By proceeding as in Example 2.5, one shows that v has a weak derivative
and ∂x v(x) = 1 if x < 0, and ∂x v(x) = −1 otherwise. As established in
Example 4.5 the distributional derivative of v and its weak derivative coincide.
Notice though that the distributional second derivative of v is ∂xx v = −2δ0
which is not a weak derivative. ⊓
⊔
Part I. Elements of functional analysis 51

Example 4.7 (Dirac measure on the unit sphere). (i) Let a ∈ Rd . Defi-
nition 4.4 implies that h∂ α δa , ϕi = (−1)|α| ∂ α ϕ(0). (ii) Let u : Rd → R be such
that u(x) := 1 if kxkℓ2 ≤ 1 and u(x) := 0 otherwise. Let B(0, 1) and S(0, 1)
d
be the unit ball and unit sphere in R R . We define the Dirac measure sup-
ported in S(0, 1) by hδS(0,1) , ϕi := S(0,1) ϕ ds. Let ei be one of the canonical
R R
unit vectors of Rd . Then h∂i u, ϕi = − B(0,1) ∂i ϕ dx = − B(0,1) ∇·(ϕei ) dx,
R
which proves that h∂i u, ϕi = − S(0,1) n·ei ϕ ds. Hence ∇u = −δS(0,1) n. ⊓ ⊔
Definition 4.8 (Distributional convergence). Let D be an open set in
Rd . We say that a sequence of distributions {Tn }n∈N converges in the sense
of distributions if one has limn→∞ hTn , ϕi = hT, ϕi for all ϕ ∈ C0∞ (D).
Example 4.9 (Oscillating functions). Let D := (0, 1) and fn (x) :=
sin(nx) for all n ≥ 1. This sequence does not converge in L1 (D), but
R1 R1
hTfn , ϕi = 0 sin(nx)ϕ dx = 0 n1 cos(nx)ϕ′ dx, so that limn→∞ hTfn , ϕi = 0
for all ϕ ∈ C0∞ (D), i.e., Tfn → 0 in the sense of distributions. Up to an
abuse of notation we say that fn converges to 0 in the sense of distribu-
tions. Likewise one can show that cos(nx) → 0 in the sense of distributions.
Let us now consider gn (x) := sin2 (nx) for all n ≥ 1. Using the identity
sin2 (nx) = 12 − 21 cos(2nx) and the above results, we conclude that gn → 12
in the sense of distributions. ⊓
⊔

4.2 Negative-order Sobolev spaces

Equipped with the notion of distributions, we can now define Sobolev spaces
of negative order by duality using W0s,p (D).
Definition 4.10 (W −s,p (D)). Let s > 0 and p ∈ (1, ∞). Let D be an open
′
′
set in Rd . We define the space W −s,p (D) := W0s,p (D) with 1p + p1′ = 1
(for p = 2, we write H −s (D) := W −s,2 (D)), equipped with the norm

|hT, wi|
kT kW −s,p(D) := sup . (4.3)
′
w∈W s,p (D)
kwkW s,p′ (D)
0

′
Identifying Lp (D) with the dual space of Lp (D) (see Theorem 1.37), we
infer that Lp (D) ֒→ W −s,p (D) (and both spaces coincide for s = 0 since
′ ′
W00,p (D) = Lp (D) by Theorem 1.34). Moreover any element T ∈ W −s,p (D)
is a distribution since, assuming s = m ∈ N, we have
 1
m + d p′
1
|α|

|hT, ϕi| ≤ kT kW −m,p(D) |D| p′ max ℓD k∂ α ϕkL∞ (K) , (4.4)
d |α|≤m

for all ϕ ∈ C0∞ (D) with supp(ϕ) ⊂ K. The argument can be adapted to the
case where s = m + σ, σ ∈ (0, 1).
52 Chapter 4. Distributions and duality in Sobolev spaces

Example 4.11 (Dirac measure). Some of the objects in W −s,p (D) are not
functions but distributions. For instance the Dirac mass at a point a ∈ D is
in W −s,p (D) if sp′ > d. ⊓
⊔

Theorem 4.12 (W −1,p (D)). Let p ∈ (1, ∞). Let D be an open, bounded set
′
in Rd . For all f ∈ W −1,p (D), there are functions {gi }i∈{0: d} , all in Lp (D),
such that kf kW −1,p (D) = maxi∈{0: d} kgi kLp′ (D) and
Z X Z
hf, vi = g0 v dx + gi ∂i v dx, ∀v ∈ W01,p (D). (4.5)
D i∈{1: d} D

More
P generally for all m ∈ N, one has v ∈ W −m,p (D) if and only if v =
α p′
|α|≤m ∂ gα where gα ∈ L (D).

Proof. See Brezis [46, Prop. 9.20] for the case (m = 1) and Adams and
Fournier [3, Thm. 3.9]. ⊓
⊔

Example 4.13 (Gradient). Let s ∈ (0, 1), p ∈ (1, ∞), and sp 6= 1. If D is a

Lipschitz domain in Rd , then the linear operator ∇ maps W s,p (D) boundedly
to W s−1,p (D), i.e., we have ∇ ∈ L(W s,p (D); W s−1,p (D)); see Grisvard [106,
Thm. 1.4.4.6]. ⊓
⊔

Remark 4.14 (Interpolation). Assuming that D is a Lipschitz domain,

an alternative definition of the negative-order spaces can be devised by means
of the interpolation theory between Banach spaces (see §A.5). Let p ∈ (1, ∞)
and s ∈ (0, 1). Recalling the space W −1,p (D) from Definition 4.10, let us set

W̌ −s,p (D) := [W −1,p (D), Lp (D)]1−s,p .

′
s,p
Theorem A.30 and the definition (3.6) of W00 (D) imply that
′ ′
s,p′
′
W̌ −s,p (D) = [Lp (D), W01,p (D)]′s,p′ = W00 (D) .

The arguments from Remark 3.23 imply that Ȟ −s (D) = H −s (D) if s 6= 12

s
since H00 (D) = H0s (D) in this case (see (3.7)). (One can also infer that
f s,p (D) = W s,p (D), as conjectured in
W̌ −s,p (D) = W −s,p (D) for sp 6= 1, if W 00
Remark 3.23.) ⊓
⊔

4.3 Normal and tangential traces

The goal of this section is to give a meaning to the normal or tangential
component of Rd -valued fields for which we only have integrability properties
on the divergence or the curl, respectively, but not on the whole gradient.
The underlying idea is quite general and consists of defining the traces in
Part I. Elements of functional analysis 53

a Sobolev space of negative-order at the boundary by extending a suitable

integration by parts formula valid for smooth functions. Recall that for any
field v = (vi )i∈{1: d} ∈ L1loc (D) := L1loc (D; Rd ), the divergence is defined by
X
∇·v := ∂i vi , (4.6)
i∈{1: d}

and for d =P3, the curl ∇×v is the column vector in R3 with components
(∇×v)i := j,k∈{1: 3} εijk ∂j vk for all i ∈ {1:3}, where εijk denotes the Levi-
Civita symbol (εijk := 0 if at least two indices take the same value, ε123 =
ε231 = ε312 := 1 (i.e., for even permutations), and ε132 = ε213 = ε321 := −1
(i.e., for odd permutations)). In component form we have

∇×v := (∂2 v3 − ∂3 v2 , ∂3 v1 − ∂1 v3 , ∂1 v2 − ∂2 v1 )T . (4.7)

Recall also that the following integration by parts formulas hold true for all
v, w ∈ C 1 (D) and all q ∈ C 1 (D):
Z Z Z
(v×n)·w ds = v·∇×w dx − (∇×v)·w dx, (4.8a)
∂D
Z ZD Z D
(v·n)q ds = v·∇q dx + (∇·v)q dx. (4.8b)
∂D D D

Let p ∈ (1, ∞) and let us consider the following Banach spaces:

Z c,p (D) := {v ∈ Lp (D) | ∇×v ∈ Lp (D)}, (4.9a)

d,p p p
Z (D) := {v ∈ L (D) | ∇·v ∈ L (D)}. (4.9b)

For p = 2, we write

H(curl; D) := Z c,2 (D), H(div; D) := Z d,2 (D). (4.10)

1 1 ′
Let h·, ·i∂D denote the duality pairing between W − p ,p (∂D) and W p ,p (∂D).
′ 1 ′
The trace operator γ g : W 1,p (D) −→ W p ,p (∂D) being surjective (see The-
1
orem 3.10), we define the linear map γ c : Z c,p (D) → W − p ,p (∂D) by
Z Z
hγ c (v), li∂D := v·∇×w(l) dx − (∇×v)·w(l) dx, (4.11)
D D

1 ′ ′
for all v ∈ Z c,p (D) and all l ∈ W p ,p
(∂D), where w(l) ∈ W 1,p (D) is such
that γ (w(l)) = l. Note that (4.8a) shows that γ c (v) = v|∂D ×n when v is
g

smooth. A direct verification shows that the map γ c is bounded. Moreover

the definition (4.11) is independent of the choice of w(l); see Exercise 4.5.
1
We also define the linear map γ d : Z d,p (D) → W − p ,p (∂D) by
54 Chapter 4. Distributions and duality in Sobolev spaces
Z Z
d
hγ (v), li∂D := v·∇q(l) dx + (∇·v)q(l) dx, (4.12)
D D

1 ′ ′
for all v ∈ Z d,p (D) and all l ∈ W p ,p (∂D), where q(l) ∈ W 1,p (D) is such that
1
γ g (q(l)) = l, and h·, ·i∂D now denotes the duality pairing between W − p ,p (∂D)
1 ′
and W p ,p (∂D). Reasoning as above, one can verify that: γ d (v) = v|∂D ·n
when v is smooth; the map γ d is bounded; the definition (4.12) is independent
of the choice of q(l).

Theorem 4.15 (Normal/tangential component). Let p ∈ (1, ∞). Let

1
D be a Lipschitz domain in Rd . Let γ c : Z c,p (D) → W − p ,p (∂D) and γ d :
1
Z d,p (D) → W − p ,p (∂D) be defined in (4.11) and (4.12), respectively. The
following holds true:
(i) γ c (v) = v|∂D ×n and γ d (v) = v|∂D ·n whenever v is smooth.
(ii) γ d is surjective.
Z c,p (D) Z d,p (D)
(iii) Density: setting Z0c,p (D) := C0∞ (D) , Z0d,p (D) := C0∞ (D) ,
we have
Z0c,p (D) = ker(γ c ), Z0d,p (D) = ker(γ d ). (4.13)

Proof. Item (i) is a simple consequence of the definition of γ c and γ d . See [181,
Lem. 20.2] for item (ii) when p = 2. See Ern and Guermond [92, Thm. 4.7]
for item (iii) (see also Exercise 23.9). ⊓
⊔

Example 4.16 (Normal derivative). In the context of elliptic PDEs, one

often deals with functions v ∈ H 1 (D) such that ∇·(∇v) ∈ L2 (D). Since ∇v ∈
H(div; D), Theorem 4.15 with p = 2 shows that one can give a meaning to
1
the normal derivative of v at the boundary as γ d (∇v) ∈ H − 2 (∂D). Assuming
s 3
more smoothness on v, e.g., v ∈ H (D), s > 2 , and some smoothness of the
boundary ∂D, one can instead invoke Theorem 3.16 to infer that γ ∂n (v) ∈
3
H s− 2 (∂D) ֒→ L2 (∂D), i.e., the normal derivative is integrable. However this
smoothness assumption is often too strong for elliptic PDEs, and one has to
use γ d (∇v) to define the normal derivative. ⊓
⊔

Example 4.17 (Whitney’s paradox). Let us show by a counterexample

(see [190, p. 100]) that the normal component of a vector field with integrable
divergence over D may not be integrable over ∂D. The two-dimensional field
v(x1 , x2 ) := ( x−x x1
2 +x2 , x2 +x2 )
2 T
in D := (0, 1)2 satisfies kv(x)kℓ2 = kxkℓ−1
2 ,
1 2 1 2
p
v ∈ L (D) for all p ∈ [1, 2), and ∇·v = 0. However v·n is not integrable, i.e.,
v·n 6∈ L1 (∂D). ⊓
⊔

Remark 4.18 (2D). In dimension two (d = 2), the tangential component

1
is defined using the linear map γ c : Z c,p (D) → W − p ,p (∂D) as follows:
Z Z
c ⊥
hγ (v), li∂D := v·∇ w(l) dx + (∇×v)w(l) dx,
D D
Part I. Elements of functional analysis 55

1 ′ ′
for all v ∈ Z c,p (D) and all l ∈ W p ,p (∂D), where w(l) ∈ W 1,p (D) is such
that γ g (w(l)) = l. Here ∇⊥ v := (−∂2 v, ∂1 v)T and ∇×v := ∂1 v2 − ∂2 v1 . Note
that ∇⊥ v = R π2 (∇v) and ∇×v = −∇·(R π2 (v)) where R π2 is the rotation of
angle π2 in R2 (i.e., the matrix of R π2 relative to the canonical basis of R2 is
0 −1

c
1 0 ). Whenever v is smooth, we have γ (v) = v|∂D ·t where t := R 2 (n)
π

is a unit tangent vector to ∂D. ⊓

⊔

Exercises
Exercise 4.1 (Distributions). Let D be an open set in Rd . Let v be a
distribution in D. (i) Let ψ ∈ C ∞ (D). Show that the map C0∞ (D) ∋ ϕ 7→
hv, ψϕi defines a distribution in D (this distribution is usually denoted ψv).
(ii) Let α, β ∈ Nd . Prove that ∂ α (∂ β v) = ∂ β (∂ α v) in the distribution sense.

Exercise 4.2 (Dirac measure on a manifold). Let D be a smooth

bounded and open set in Rd . Let u ∈ C 2 (D; R) and assume that u|∂D = 0. Let
ũ be the extension by zero of u over Rd . Compute ∇·(∇ũ) = ∂11 u + . . . ∂dd u
in the distribution sense.

Exercise 4.3 (P.V. x1 ). Let D := (−1,R1). Prove that the linear map T :
C0∞ (D) → R defined by hT, ϕi := limǫ→0 |x|>|ǫ| x1 ϕ(x) dx is a distribution.

Exercise 4.4 (Integration byR parts). Prove

R the two identities in (4.8) by
using the divergence formula D ∇·φ dx = ∂D (φ·n) ds for all φ ∈ C 1 (D).

Exercise 4.5 (Definition (4.11)). Verify that the right-hand side of (4.11)
is independent of the choice of w(l). (Hint : consider two functions w1 , w2 ∈
′
W 1,p (D) s.t. γ g (w1 ) = γ g (w2 ) = l and use the density of C0∞ (D) in
′
W01,p (D).)
56 Chapter 4. Distributions and duality in Sobolev spaces

Solution to exercises
Exercise 4.1 (Distributions). (i) Let K be a compact subset of the open
set D. Since v is a distribution, there exist c ∈ R and p ∈ N (both depending
|α|

on K) such that |hv, ϕi| ≤ c max|α|≤p ℓD k∂ α ϕkL∞ (K) for all ϕ ∈ C0∞ (D)
(i.e., ϕ is a smooth function with compact support in K). Let ψ ∈ C ∞ (D)
and ϕ ∈ C0∞ (D). Since ψϕ ∈ C ∞ (D) and supp(ψϕ) ⊂ K, we have ψϕ ∈
C0∞ (D). The product rule implies that there exists c′ depending on p such
|β|

that k∂ α (ψϕ)kL∞ (K) ≤ c′ max|β|≤|α| ℓD k∂ β ψkL∞ (K) k∂ α ϕkL∞ (K) for all α
s.t. |α| ≤ p. For all ϕ ∈ C0∞ (D) we infer that
|α|

|hv, ψϕi| ≤ c max ℓD k∂ α (ψϕ)kL∞ (K)
|α|≤p
|β|

≤ c c′ max max ℓD k∂ β ψkL∞ (K) k∂ α ϕkL∞ (K)
|α|≤p |β|≤|α|
′ |β|

≤ c c max ℓD k∂ β ψkL∞ (K) max k∂ α ϕkL∞ (K) ,
|β|≤p |α|≤p

thereby proving that the linear map C0∞ (D) ∋ ϕ 7→ hv, ψϕi ∈ R is a distri-
bution.
(ii) Let α, β ∈ Nd . Let ϕ ∈ C0∞ (D). Using Clairaut’s theorem for ϕ, we infer
that

h∂ β (∂ α v), ϕi = (−1)|β| h∂ α v, (∂ β ϕ)i = (−1)|α|+|β| hv, ∂ α (∂ β ϕ)i

= (−1)|α|+|β| hv, ∂ β (∂ α ϕ)i = (−1)|α| h∂ β v, (∂ α ϕ)i
= h∂ α (∂ β v), ϕi.

Hence ∂ β (∂ α v) = ∂ α (∂ β v), which is Clairaut’s theorem for distributions.

Exercise 4.2 (Dirac measure on a manifold). We denote ∆ := ∇·(∇)

By definition we have
Z Z
h∆ũ, ϕi = hũ, ∂11 ϕ + . . . ∂dd ϕi = ũ∇·(∇ϕ) dx = u∇·(∇ϕ) dx
Rd D
Z Z Z
=− ∇u·∇ϕ dx = ∇·(∇u)ϕ dx − (n·∇u)ϕ ds
D D ∂D
Z Z
= f dx −
∆uϕ (n·∇u)ϕ ds.
Rd ∂D

Hence we have proved that ∆ũ = ∆u f − δ∂D ∇u·n where δ∂D is the Dirac
measure whose support is ∂D. Note that the notation δ∂D (n·∇u) makes
sense owing to the result of Exercise 4.1(i).
1
Exercise 4.3 (P.V. x ). Let ǫ > 0. We have
Part I. Elements of functional analysis 57

Z Z −ǫ Z 1
1 1 1
ϕ(x) dx = ϕ(x) dx +
ϕ(x) dx
|x|>|ǫ| x −1 x ǫ x
Z −ǫ Z 1
=− ln(|x|)ϕ′ (x) dx + ϕ(−ǫ) ln(ǫ) − ln(x)ϕ′ (x) dx − ϕ(ǫ) ln(ǫ)
−1 ǫ
Z 1
= 1|(−1,−ǫ)∪(ǫ,1) ln(|x|)ϕ′ (x) dx + (ϕ(−ǫ) − ϕ(ǫ)) ln(ǫ),
−1

where 1E is the indicator function of the set E. Clearly |(ϕ(−ǫ)−ϕ(ǫ)) ln(ǫ)| ≤

kϕ′ kL∞ ǫ ln(ǫ). Moreover the sequence 1|(−1,−ǫ)∪(ǫ,1) ln(|x|)ϕ′ (x) converges
a.e. in D to ln(|x|)ϕ′ (x) and we also have 1|(−1,−ǫ)∪(ǫ,1) ln(|x|)ϕ′ (x) ≤
ln(|x|)ϕ′ (x) ∈ L1 (D). Lebesgue’s dominated convergence implies that
Z Z 1
1
hT, ϕi := lim ϕ(x) dx = − ln(|x|)ϕ′ (x) dx,
ǫ→0 |x|>|ǫ| x −1

i.e., the limit process with respect to ǫ is well defined. Moreover we have
Z 1
|hT, ϕi| = x(ln(|x|) − 1)ϕ′′ (x) dx ≤ kϕ′′ (x)kL∞ (D) ,
−1

thereby proving that T is indeed a distribution. Note that we have actually

proved that T = ∂x ( x1 ), which makes sense after all.

Exercise 4.4 (Integration by parts). The identity (4.8a) follows from the
divergence formula for φ by using φ = v×w (since ∇·(v×w) = (∇×v)·w −
v·∇×w) and φ·n = (v×w)×n = −(v×n)·w), whereas the identity (4.8b)
follows from the divergence formula for φ by using φ = vq (since ∇·(vq) =
v·∇q + (∇·v)q and φ·n = (v·n)q).

Exercise 4.5 (Definition (4.11)). Let v ∈ Z c,p (D). Following the hint let
′ ′
w1 , w2 ∈ W 1,p (D) be s.t. γ g (w1 ) = γ g (w2 ) = l. Then w1 − w2 ∈ W01,p (D).
Invoking a density argument, let (φn )n∈N be a sequence in C0∞ (D) converging
′
to w1 − w2 in W01,p (D). Then we have
Z Z
0= v·∇×φn dx − φn ·∇×v dx.
D D

Passing to the limit n → ∞ yields

Z Z
0= v·∇×(w1 − w2 ) dx − (w1 − w2 )·∇×v dx.
D D

c g c g
Hence hγ (v), γ (w1 )i∂D = hγ (v), γ (w2 )i∂D , which establishes the claim.