Chapter 10

Sobolev Spaces
We now dene spaces H
1,p
(R
n
), known as Sobolev spaces. For u to belong
to H
1,p
(R
n
), we require that u L
p
(R
n
) and that u have weak derivatives
of rst order in L
p
(R
n
):
(10.1)
j
u = f
j
L
p
(R
n
),
where (10.1) means
(10.2)
_

x
j
udx =
_
f
j
dx, C

0
(R
n
).
If u C

0
(R
n
), we see that
j
u = u/x
j
, by integrating by parts, using
(7.67). We dene a norm on H
1,p
(R
n
) by
(10.3) u
H
1,p = u
L
p +

j

j
u
L
p.
We claim that H
1,p
(R
n
) is complete, hence a Banach space. Indeed, let (u

)
be a Cauchy sequence in H
1,p
(R
n
). Then (u

) is Cauchy in L
p
(R
n
); hence it
has an L
p
-norm limit u L
p
(R
n
). Also, for each j,
j
u

= f
j
is Cauchy in
L
p
(R
n
), so there is a limit f
j
L
p
(R
n
), and it is easily veried from (10.2)
that
j
u = f
j
.
We can consider convolutions and products of elements of H
1,p
(R
n
) with
elements of C

0
(R
n
) and readily obtain identities
(10.4)

j
( u) = (
j
) u = (
j
u),

j
(u) =

x
j
u + (
j
u),
129
130 10. Sobolev Spaces
and estimates
(10.5)
u
H
1,p
L
1 u
H
1,p,
u
H
1,p
L
u
H
1,p +

L
u
L
p,
for , C

0
(R
n
), u H
1,p
(R
n
). For example, the rst identity in (10.4)
is equivalent to

__

x
j
(x) (x y)u(y) dy dx
=
__
(x)

x
j
(x y) u(y) dy dx, C

0
(R
n
),
an identity that can be established by using Fubinis Theorem (to rst do
the x-integral) and integration by parts, via (7.67).
If p < , u H
1,p
(R
n
), and (
j
) is an approximate identity of the form
(7.64), with
j
C

0
(R
n
), then we can show that
(10.6)
j
u u in H
1,p
-norm,
using (10.4) and (7.65). Given > 0, we can take j such that

j
u u
H
1,p < .
Then we can pick C

0
(R
n
) such that (
j
u)
j
u
H
1,p < . Of
course,
j
u is smooth, so (
j
u) C

0
(R
n
). We have established
Proposition 10.1. For p [1, ), the space C

0
(R
n
) is dense in H
1,p
(R
n
).
Sobolev spaces are very useful in analysis, particularly in the study of
partial dierential equations. We will establish just a few results here, some
of which will be useful in Chapter 11. More material can be found in [EG],
[Fol], [T1], and [Yo].
The following result is known as a Sobolev Imbedding Theorem.
Proposition 10.2. If p > n or if p = n = 1, then
(10.7) H
1,p
(R
n
) C(R
n
) L

(R
n
).
For now we concentrate on the case p (n, ). Since C

0
(R
n
) is then
dense in H
1,p
(R
n
), it suces to establish the estimate
(10.8) u
L
Cu
H
1,p, for u C

0
(R
n
).
10. Sobolev Spaces 131
In turn, it suces to establish
(10.9) |u(0)| Cu
H
1,p, for u C

0
(R
n
).
To get this, it suces to show that, for a given C

0
(

B
1
) with (0) = 1,
(10.10) |u(0)| C(u)
L
p ,
where v = (
1
v, . . . ,
n
v) or, equivalently, that
(10.11) |u(0)| Cu
L
p, u C

0
(

B
1
).
In turn, this will follow from an estimate of the form
(10.12) |u(0) u()| Cu
L
p
(B
1
)
, u C

(R
n
),
given R
n
, || = 1. Thus we turn to a proof of (10.12).
Without loss of generality, we can take = e
n
= (0, . . . , 0, 1). We will
work with the set = {z R
n1
: |z|

3/2}. For z , let

z
be the
path from 0 to e
n
consisting of a line segment from 0 to (z, 1/2), followed
by a line segment from (z, 1/2) to e
n
, as illustrated in Figure 10.1. Then
(with A = Area )
(10.13) u(e
n
) u(0) =
_

_
_
z
du
_
dz
A
=
_
B
1
u(x) (x) dx,
where the last identity applies the change of variable formula. The behavior
of the Jacobian determinant of the map (t, z)
z
(t) yields
(10.14) |(x)| C|x|
(n1)
+ C|x e
n
|
(n1)
.
Thus
(10.15)
_
B
1/2
|(x)|
q
dx C
_
1/2
0
r
nq+q
r
n1
dr.
It follows that
(10.16) L
q
(B
1
), q <
n
n 1
.
Thus
(10.17) |u(e
n
) u(0)| u
L
p
(B
1
)

L
p
(B
1
)
Cu
L
p
(B
1
)
,
132 10. Sobolev Spaces
0
e
n

z
Figure 10.1
as long as p

< n/(n 1), which is the same as p > n.

This proves (10.12) and hence Proposition 10.2, for p (n, ). For
n = 1, (10.13) simplies to u(1) u(0) =
_
1
0
u

(x) dx, which immediately

gives the estimate (10.12) for p = n = 1.
We can rene Proposition 10.2 to the following.
Proposition 10.3. If p (n, ), then every u H
1,p
(R
n
) satises a
H older condition:
(10.18) H
1,p
(R
n
) C
s
(R
n
), s = 1
n
p
.
Proof. Applying (10.12) to v(x) = u(rx), we have, for || = 1,
(10.19) |u(r) u(0)|
p
Cr
p
_
B
1
|u(rx)|
p
dx = Cr
pn
_
Br
|u(x)|
p
dx.
This implies
(10.20) |u(x) u(y)| C

|x y|
1n/p
_
_
Br(x)
|u(z)|
p
dz
_
1/p
, r = |x y|,
which gives (10.18).
10. Sobolev Spaces 133
If u H
1,
(R
n
) and if C

0
(R
n
), then u H
1,p
(R
n
) for all p
[1, ), so Proposition 10.3 applies. We next show that in fact H
1,
(R
n
)
coincides with the space of Lipschitz functions:
(10.21) Lip(R
n
) = {u L

(R
n
) : |u(x) u(y)| K|x y|}.
Proposition 10.4. We have the identity
(10.22) H
1,
(R
n
) = Lip(R
n
).
Proof. First, suppose u Lip(R
n
). Thus
(10.23) h
1
_
u(x + he
j
) u(x)

is bounded in L

(R
n
).
Hence, by Proposition 9.4, there is a sequence h

0 and f
j
L

(R
n
)
such that
(10.24) h
1

_
u(x + h

e
j
) u(x)

f
j
weak

in L

(R
n
).
In particular, for all C

0
(R
n
),
(10.25) h
1

_
(x)
_
u(x + h

e
j
) u(x)

dx
_
(x)f
j
(x) dx.
But the left side of (10.25) is equal to
(10.26) h
1

_
_
(x h

e
j
) (x)

u(x) dx
_

x
j
u(x) dx.
This shows that
j
u = f
j
. Hence Lip(R
n
) H
1,
(R
n
).
Next, suppose u H
1,
(R
n
). Let
j
(x) = j
n
(jx) be an approximate
identity as in (10.6), with C

0
(R
n
). We do not get
j
u u in H
1,
-
norm, but we do have u
j
=
j
u bounded in H
1,
(R
n
); in fact, each u
j
is
C

, and we have
(10.27) u
j

L
K
1
, u
j

L
K
2
.
Also u
j
u locally uniformly. The second estimate in (10.27) implies
(10.28) |u
j
(x) u
j
(y)| K
2
|x y|,
since u
j
(x) u
j
(y) =
_
1
0
(x y) u(tx + (1 t)y) dt. Thus in the limit
j , we get also |u(x) u(y)| K
2
|x y|. This completes the proof.
We next show that, when p [1, n), H
1,p
(R
n
) is contained in L
q
(R
n
)
for some q > p. One technical tool which is useful for our estimates is the
following generalized H older inequality.
134 10. Sobolev Spaces
Lemma 10.5. If p
j
[1, ],

p
1
j
= 1, then
(10.29)
_
M
|u
1
u
m
| dx u
1

L
p
1 (M)
u
m

L
pm
(M)
.
The proof follows by induction from the case m = 2, which is the usual
H older inequality.
Proposition 10.6. For p [1, n),
(10.30) H
1,p
(R
n
) L
np/(np)
(R
n
).
In fact, there is an estimate
(10.31) u
L
np/(np) Cu
L
p
for u H
1,p
(R
n
), with C = C(p, n).
Proof. It suces to establish (10.31) for u C

0
(R
n
). Clearly
(10.32) |u(x)|
_

|
j
u| dy
j
,
where the integrand, written more fully, is |
j
u(x
1
, . . . , y
j
, . . . , x
n
)|. (Note
that the right side of (10.32) is independent of x
j
.) Hence
(10.33) |u(x)|
n/(n1)

j=1
_
_

|
j
u| dy
j
_
1/(n1)
.
We can integrate (10.33) successively over each variable x
j
, j = 1, . . . , n,
and apply the generalized H older inequality (10.29) with m = p
1
= =
p
m
= n 1 after each integration. We get
(10.34) u
L
n/(n1)
_
n

j=1
_
R
n
|
j
u| dx
_
1/n
Cu
L
1.
This establishes (10.31) in the case p = 1. We can apply this to v = |u|

, >
1, obtaining
(10.35)
_
_
|u|

_
_
L
n/(n1)
C
_
_
|u|
1
|u|
_
_
L
1
C
_
_
|u|
1
_
_
L
p

_
_
u
_
_
L
p
.
For p < n, pick = (n 1)p/(n p). Then (10.35) gives (10.31) and the
proposition is proved.
10. Sobolev Spaces 135
There are also Sobolev spaces H
k,p
(R
n
), for each k Z
+
. By denition
u H
k,p
(R
n
) provided
(10.36)

u = f

L
p
(R
n
), || k,
where

1
1

n
n
, || =
1
+ +
n
, and, as in (10.2), (10.36) means
(10.37) (1)
||
_

udx =
_
f

dx, C

0
(R
n
).
Given u H
k,p
(R
n
), we can apply Proposition 10.6 to estimate the
L
np/(np)
-norm of
k1
u in terms of
k
u
L
p, where we use the notation
(10.38)
k
u = {

u : || = k},
k
u
L
p =

||=k

u
L
p,
and proceed inductively, obtaining the following corollary.
Proposition 10.7. For kp < n,
(10.39) H
k,p
(R
n
) L
np/(nkp)
(R
n
).
The next result provides a generalization of Proposition 10.2.
Proposition 10.8. We have
(10.40) H
k,p
(R
n
) C(R
n
) L

(R
n
) for kp > n.
Proof. If p > n, we can apply Proposition 10.2. If p = n and k 2, since
it suces to obtain an L

bound for u H
k,p
(R
n
) with support in the unit
ball, just use u H
2,n
(R
n
) and proceed to the next step of the argument.
If p [1, n), it follows from Proposition 10.6 that
(10.41) H
k,p
(R
n
) H
k1,p
1
(R
n
), p
1
=
np
n p
.
Thus the hypothesis kp > n implies (k 1)p
1
> kp > n. Iterating this
argument, we obtain H
k,p
(R
n
) H
,q
(R
n
), for some 1 and q > n, and
again we can apply Proposition 10.2.
136 10. Sobolev Spaces
Exercises
1. Write down the details for the proof of the identities in (10.4).
2. Verify the estimates in (10.14).
Hint. Write the rst integral in (10.13) as 1/A times
_

_
1
0
v
+
(z) u(tz,
1
2
t) dt dz +
_

_
1
0
v

(z) u(tz, 1
1
2
t) dt dz,
where v

(z) = (z, 1/2). Then calculate an appropriate Jacobian de-

terminant to obtain the second integral in (10.13).
3. Suppose 1 < p < . If
y
f(x) = f(x y), show that f belongs to
H
1,p
(R
n
) if and only if
y
f is a Lipschitz function of y with values in
L
p
(R
n
), i.e.,
(10.42)
y
f
z
f
L
p C|y z|.
Hint. Consider the proof of Proposition 10.4.
What happens in the case p = 1?
4. Show that H
n,1
(R
n
) C(R
n
) L

(R
n
).
Hint. u(x) =
_
0

_
0

1

n
u(x + y) dy
1
dy
n
.
5. If p
j
[1, ] and u
j
L
p
j
, show that u
1
u
2
L
r
provided 1/r =
1/p
1
+ 1/p
2
and
(10.43) u
1
u
2

L
r u
1

L
p
1 u
2

L
p
2 .
Show that this implies (10.29).
6. Given u L
2
(R
n
), show that
(10.44) u H
k,2
(R
n
)
_
1 +||
_
k
u L
2
(R
n
).
7. Let f L
1
(R), and set g(x) =
_
x

f(y) dy. Continuity of g follows

from the Dominated Convergence Theorem. Show that
(10.45)
1
g = f.
10. Sobolev Spaces 137
Hint. Given C

0
(R), start with
(10.46)
_
d
dx
g(x) dx =
_ _
x

(x)f(y) dy dx,
and use Fubinis Theorem. Then use
_

(x) dx = (y).
Alternative. Write the left side of (10.46) as
lim
h0
1
h
_
_
(x + h) (x)

g(x) dx = lim
h0
_ _
x+h
x
f(y)(x) dy dx,
and use (4.64).
8. If u H
1,p
(R
n
) for some p [1, ) and
j
u = 0 on a connected open
set U R
n
, for 1 j n, show that u is (equal a.e. to a) constant on
U.
Hint. Approximate u by (10.6), i.e., by u

u, where

0
(R
n
)
has support in {|x| < 1/},
_

dx = 1. Show that
j
(

u) = 0 on
U

U, where U

U as .
More generally, if
j
u = f
j
C(U), 1 j n, show that u is equal
a.e. to a function in C
1
(U).
9. In case n = 1, deduce from Exercises 7 and 8 that, if u L
1
loc
(R),
(10.47)
1
u = f L
1
(R) =u(x) = c +
_
x

f(y) dy, a.e. x R,

for some constant c.
10. Let g H
2,1
(R), I = [a, b], and f = g

I
. Show that the estimate (9.75)
concerning the trapezoidal rule holds in this setting.