Professional Documents
Culture Documents
(x
0
) = f(x
0
) ;
(iii) if F C
1
([a, b]), then
(ICI) F(x) F(a) = R
_
x
a
F
i=1
A
i
)
i=1
(A
i
) for any sequence (A
i
)
i
X (countable subadditivity) .
(ii) A set E X is called -measurable (with respect to an o.m. on X), if
(A) = (A E) + (A E) A X .
The class of -measurable sets will be denoted by /
.
(iii) A -algebra / on X is a class of subsets / T(X) satisfying the two
following properties:
(A1) if E /, then X E /;
(A2) for each sequence (E
i
)
i
/, then
i=1
E
i
/.
(iv) A measure on X is a set function : / [0, ], where / is a -algebra
on X, satisfying the following two properties:
(M1) () = 0 ;
(M2) (
i=1
E
i
) =
i=1
(E
i
) (countable additivity)
for each disjoint sequence (E
i
)
i
/. The structure composed by the triple
(X, /, ), or also the couple (X, /), is called measure space and the sets
contained in / are called measurable sets.
(v) Let (X, /, ) be a measure space. The measure is said to be nite, if
(X) < ; it is said to be -nite, if there exists a sequence (X
i
)
i
/ such
that X =
i=1
X
i
and (X
i
) < for each i.
(vi) A function f : X R := [, ] is called measurable with respect to an
o.m. (respectively with respect to a measure : / [0, ]) if f
1
(U) is
-measurable (respectively f
1
(U) /) for each open set U in R.
(vii) A simple function s : X R is one that assumes only a nite number of values.
More precisely, s is a simple function if and only if it can be represented as
s(x) =
k
i=1
a
i
A
i
(x) x X ,
with a
i
R, A
i
X (i = 1, . . . , k), X =
k
i=1
A
i
and A
i
A
j
= if i ,= j.
Now, we enrich the environment X, by adding a topology and we require compat-
ibility between topology and measure.
Denition 1.4. Let (X, ) denote a topological space and denote B(X) the -algebra
of Borel sets of X, i.e. the smallest -algebra of X which contains the open and closed
sets of X.
4
(i) An o.m. on X is called a Borel o.m. if the class of -measurable sets
/
B(X).
(ii) An o.m. on X is called a Borel regular o.m. if it is a Borel o.m. and for
each A X there exists B B(X) with B A and (A) = (B).
(iii) An o.m. on X is called a Radon o.m. if it is a Borel regualr o.m. and
(K) < for each compact set K X.
(iv) An o.m. on a metric space (X, d) is called a Caratheodory o.m. (or also a
metric o.m.) if
(A B) = (A) + (B) A, B X with d(A, B) > 0 ,
where d(A, B) := infd(a, b) : a A, b B.
(v) A measure : / [0, ] on X is called a Borel measure if / = B(X).
(vi ) A measure : / [0, ] on X is called a Radon measure if it is a Borel
measure and (K) < for each compact set K X.
The following basic properties of outer measures are well-known (see [De] or [GZ]).
Theorem 1.5. (i) Let be an o.m. on X. Then the class of -measurable sets
/
is a -algebra on X and : /
.
(ii) (Caratheodorys criterion) Let be a Carathedory o.m. on a metric space
(X, d). Then is a Borel o.m.
Example 1.6. (i) Let /
n
denote the Lebesgue o.m. on R
n
. Then /
n
is a Radon
measure on R
n
. The class /
n
/
L
n is called the class of Lebesgue measur-
able sets.
(ii) Let s [0, ) and 1
s
denote the s-dimensional Hausdor measure on R
n
.
Then 1
s
is a Borel regular o.m. on R
n
, but it is not a Radon o.m. unless
s n.
(iii) Let A be a non Borel set of R
n
(why does A exists?); let : T(R
n
) [0, +]
be the set function dened as
(E) :=
_
0 if E A
+ if E A ,=
.
It is easy to see, by the denition, that is a Caratheodory outer measure
on R
n
. However it is not Borel regular because it does not exist a Borel set
B A such that (B) = (A).
(iv) Let : T(R
n
) [0, +] be the set function dened as
(E) :=
_
0 if E =
1 if E ,=
.
It is easy to see, by the denition, that is an o.m. on R
n
and /
= , R
n
.
In particular, it is not a Borel o. m.
By Theorem 1.5 (i) we see that to every o.m. on X is associated the measure
space (X, /
, )
Question: Given a measure space (X, /, ) is there an associated o.m.
:
T(X) [0, ] such that
= on / ?
5
There is a simple procedure due to Caratheodory to generate from a measure
: / [0, ] an outer measure
. Moreover
is an o.m.;
(ii)
;
(iv) Let A be a -algebra with / A /
f
((a, b]) := f(b) f(a) .
Example: f(x) = x,
f
((a, b]) = b a.
Based on this notion of lenght, a measure analogous to the Lebesgue measure can
be generated. This establishes an important connection between measures on R and
monotone functions. To make this connection precise, it is necessary to use half-open
intervals rather than closed intervals, because of some technical reasons that will be
clear later.
Denition 1.8. The Lebesgue-Stieltjes (acronym LS) outer measure of an arbitrary
set E R is dened by
f
(E) := inf
_
i=1
f
((a
i
, b
i
]) : E
i=1
(a
i
, b
i
]
_
where the inmum is taken over all countable collections of half-open intervals (a
i
, b
i
]
such that E
i=1
(a
i
, b
i
].
Theorem 1.9. If f : R R is a nondecreasing function, then
f
is a Caratheodory
(or also metric) o.m.
Proof. We need only to show that
(1.1)
f
() = 0 ,
6
(1.2)
f
(A)
i=1
f
(A
i
) if A
i=1
A
i
(monotonicity and countable subadditivity) ,
(1.3)
f
(A B) =
f
(A) +
f
(B) if d(A, B) > 0 (metric property) .
where d(A, B) := inf[a b[ : a A, b B.
The property (1.1) follows immediatly from the denition of
f
.
For the proof of (1.2), without loss of generality we can assume that
f
(A
i
) < i = 1, 2, . . .
otherwise (1.2) is trivial. Fix > 0 (arbitrary). By denition of
f
, for given i =
1, 2, . . . , there exists a family of half-open intervals I
(i)
h
= (a
(i)
h
, b
(i)
h
] (h = 1, 2, . . . )
such that
i=1
I
(i)
h
A
i
and
h=1
f
(I
(i)
h
) <
f
(A
i
) +
2
i
.
Then, since A
i=1
A
i
i,h=1
I
(i)
h
,again, by denition of
f
,
f
(A)
i=1
h=1
f
(I
(i)
h
)
i=1
_
f
(A
i
) +
2
i
_
=
i=1
f
(A
i
) + > 0 ,
that establishes (1.2), since is arbitrary.
Eventually let us prove (1.3). By the countable subadditivity of
f
, we need only
to prove that
(1.4)
f
(A B)
f
(A) +
f
(B) if d(A, B) > 0 .
Without loss of generality we can assume that
f
(A B) < ,
otherwise (1.4) is trivial. Fix > 0 (arbitrary). By denition of
f
,there exists a
family of half-open intervals I
h
= (a
h
, b
h
] (h = 1, 2, . . . ) such that
(1.5)
i=1
I
h
A B.
(1.6)
h=1
f
(I
h
) =
i=1
(f(b
h
) f(a
h
)) <
f
(A B) +
Now notice that for each nite partition of [a, b], a = t
0
< t
1
< < t
m
= b,
(1.7)
f
((a, b]) = f(b) f(a) =
m
i=1
(f(t
i
) f(t
i1
)) =
m
i=1
f
((t
i1
, t
i
]) .
By (1.7), we can replace each interval (a
h
, b
h
] by a smaller one. Thus, it is not
restrictive to assume that
b
h
a
h
<
4
h
where = d(A, B) > 0. Let J
1
:= h N : I
h
A ,= and J
2
:= h N : I
h
B ,=
. Then J
1
and J
2
are disjoint and, by (1.5),
A
hJ
1
I
h
and B
hJ
2
I
h
.
7
Therefore, by denition of
f
,
f
(A) +
f
(B)
hJ
1
f
(I
h
) +
hJ
2
f
(I
h
)
h=1
f
(I
h
)
(1.6)
f
(A B) + ,
that establishes (1.4) since is arbitrary.
Remark 1.10. From Caratheodorys criterion, it follows that the family of
f
-
measurable sets /
f
B(R). Thus
f
is a Borel outer measure. Actually
f
is
a Radon outer measure, that is, it is a Borel regular outer measure and
f
(K) <
for each compact K R (see Exercise I.2).
In the case of Lebesgue outer measure, it holds that /
1
(I) = b a for each (open
or closed) interval I with endpoints a and b.
Question: Let f : R R be a nondecreasing function, does it holds that
f
((a, b]) =
f
((a, b]) = f(b) f(a) ?
Theorem 1.11. Let f : R R be a nondecreasing and right-continuous function,
i.e.
lim
yx
+
f(y) = f(x) x R.
Then
f
((a, b]) = f(b) f(a)
for all a, b R with a < b.
Proof. We skip the proof (see [GZ, Theorem 4.32]).
Exercise I.3 Let f(x) := x for x R. Then
f
= /
1
.
Remark 1.12. The Lebesgue outer measure on R is a special case of LS-outer mea-
sure. Moreover the conclusion of Theorem 1.11 fails if f is not right-continuous. For
instance, let f : R R be the function dened by f(x) :=
_
0 if x 0
1 if x > 0
. The
function f is nondecreasing and not right-continuous at x = 0. Now notice that, since
f
is a measure on B(R),
f
((0, 1]) = lim
h
f
__
1
h
, 1
__
.
On the other hand,
f
__
1
h
, 1
__
f
__
1
h
, 1
__
= f(1) f(1/h) = 0
and
f
((0, 1]) = f(1) f(0) = 1 .
Thus,
f
((0, 1]) = 0 < 1 =
f
((0, 1]) .
8
We have just seen that each nondecreasing f : R R gives rise to a Radon outer
measure
f
. The converse is readily seen to hold. Indeed, if is a nte Borel o.m.
on R, that is is a Borel o.m. and (R) < , let
f(x) := ((, x]) x R.
Then
Exercise I.4: f is nondecreasing and right-continuous. Moreover
((a, b]) = f(b) f(a) a, b R with a < b .
(Incidentally, this now shows why half-open intervals are used in the development.)
Theorem 1.13. Suppose is a nite Borel o.m. on R and let
f(x) := ((, x]) x R.
Then
(C)
f
(B) = (B) B B(R) .
Before the proof of Theorem 1.13, let us recall three important results on approxi-
mation of measures by open and closed sets. The rst result is also contained in [De]
or [GZ, Theorem 4.17]).
Theorem 1.14 (Approximation of outer measures by open and closed sets). Let
be a Borel (respectively a Borel regular) o.m. on a metric space (X, d) and let B X
be a Borel set (respectively a -measurable set).
(i) Suppose that (B) < , then for each > 0 there exists a closed set F B
such that
(B F) < .
(ii) Suppose
B
i=1
V
i
where each V
i
X is an open set with (V
i
) < . Then there is an open set
U B such that
(U B) < .
The rst important consequence of Theorem 1.14 is the following
Corollary 1.15. Let be a Borel (respectively a Borel regular) o.m. on a metric
space (X, d). Suppose there exists a sequence of open sets (V
i
)
i
X such that
() X =
i=1
V
i
with (V
i
) < i .
Then for each B B(X) (respectively B /
)
(i) (B) = inf(U) : U B, U open;
(ii) (B) = sup(C) : C B, C closed.
The second important consequence of Corollary 1.15 and the Caratheodory-Hahn
extension theorem (Theorem 1.7) is the following approximation result for Borel mea-
sures.
9
Corollary 1.16 (Approximation of Borel measures by open and closed sets). Con-
sider a measure space (X, B(X), ) where X is a metric space and is a Borel
measure. Suppose that the assumption () of Corollary 1.15 holds replacing with
. Then for each B B(X)
(i) (B) = inf(U) : U B, U open;
(ii) (B) = sup(C) : C B, C closed.
Proof of Corollary 1.15. (i) Assume (B) = . Then, by choosing U = X we get
that
= (B) (X) .
Thus U B and (U) = = (B).
If (B) < , for each > 0, from the approximation of o.m. by open and closed
sets there exists an open set U B such that (U B) < . Therefore
(B) (U) < (B) +
which completes the proof.
(ii) Firstly, assume that (B) < . The approximation theorem by open and
closed sets yields that > 0 a closed set C B such that (B C) < . In
particular,
(B) < (C) + .
Since
(B) sup (C) : C B, C closed ,
we obtain the desired conclusion.
Assume now (B) = . By () and Exercise I.5, there exists a sequence (A
i
) X
of disjoint Borel (or, respectively, -measurable, if B is -measurable) sets such that
(1.8) X =
i=1
A
i
and (A
i
) < i .
Dene B
i
:= B A
i
. Then
B =
i=1
B
i
and = (B) =
i=1
(B
i
) .
By (1.8), for each i, B
i
is -measurable and (B
i
) < . From the approximation
theorem by open and closed sets, there exists a closed set C
i
B
i
such that
(C
i
) (B
i
)
1
2
i
.
Now the sequence (C
i
)
i
is disjoint,
i=1
C
i
B and
lim
j
(
j
i=1
C
i
) = ((
i=1
C
i
) =
i=1
(C
i
)
i=1
_
(B
i
)
1
2
i
_
= .
Observe now that
j
i=1
C
i
is a closed set for each j, whence we get the desired assertion.
, beacuse the
assumptions of that result hold. Indeed, from Caratheodory- Hahn extension theorem
(see Theorem 1.7), /
B(X) and
= on B(X). Since
agrees with on
10
B(X), () of Corollary 1.15 is satised with =
i=1
V
i
with
(V
i
) < i .
Therefore
agrees with on
B(X), satises (i) and (ii), too.
Remark 1.17. When (X, d) is a separable, locally compact metric space and (re-
spectively ) is a Radon outer measure (respectively measure) on X, the assumption
() is satised. Thus the conclusions of Corollary 1.15 (respectively Corollary 1.16)
hold. Moreover the approximation from below by means of closed sets can be replaced
by compact sets.
Proof of Theorem 1.13. We need only to prove (C) for each open set U R. Then,
by Corollary 1.15 (i), the proof is accomplished.
If U R is open set, then, by Exercise I.6, U =
i=1
(a
i
, b
i
) with ((a
i
, b
i
))
i
suitable
sequence of disjoint open intervals. On the other hand, each open interval (a, b) can
be obtained as
(a, b) =
h=1
(a
h
, b
h
]
with (a
h
, b
h
] (a
k
, b
k
] = if h ,= k. For instance, choose
(a
h
, b
h
] :=
_
a +
h(b a)
h + 1
, a +
(h + 1)(b a)
h + 2
_
with h 1. Thus
f
((a, b)) =
f
(
h=1
(a
h
, b
h
]) =
h=1
f
((a
h
, b
h
]) =
h=1
(f(b
h
) f(a
h
)) =
h=1
((a
h
, b
h
]) = (
h=1
(a
h
, b
h
]) = ((a, b)) .
Therefore, by Exercise I.4, it follows that
f
(U) = (U).
Now we are in position to state the main problem in the setting of measure theory.
Denition 1.18. Let X be a set and let f : X R. The nonnegative and nonpositive parts
of f are, respectively, the functions f
+
, f
: X [0, ] dened by
f
+
(x) := max f(x), 0 , f
= (F
)
+
(F
,
it follows that
F(x) F(a) =
_
x
a
F
(t) dt =
_
x
a
(F
)
+
(t) dt
_
x
a
(F
(t) dt .
Thus, if we put F
1
(x) := F(a) +
_
x
a
(F
)
+
(t) dt, F
2
(x) :=
_
x
a
(F
(x) if x [a, b]
0 otherwise
. Notice also that
F
i) F C
0
(R)?
(1
ii) F
f
(E) :=
_
E
f d E /.
Which are the measures : / [0, ] verifying =
f
for some measurable
function f : X [0, ]?
The answer to this question will be the content of the Radon-Nikodym theorem.
1.2. Radon-Nikodym and Lebesgue decomposition theorems. Firstly, let us
introduce some denitions and preliminary results.
Denition 1.21. Let (X /) be a measure space and let , : / [0, ] be two
measures.
(i) The measure is said to be absolutely continuous with respect to the measure
, written << , if it holds that
(E) = 0 (E) = 0 .
(ii) The measures and are said to be mutually singular, written , if
there exists a measurable set E such that
(E) = (X E) = 0 .
The following result justies why the word continuity is used in this context.
Theorem 1.22. Let be a nite measure and a measure on a measure space
(X, /). Then the following are equivalent:
(i) << ;
(ii) lim
(A)0
(A) = 0, that is, for every > 0 = () > 0 such that (E) <
whenever (E) < .
Proof. see [De] or [GZ, Theorem 6.33].
Let us recall the following fundamental result in measure theory, wich is a particular
case of that proved in the course Analisi Matematica VI for signed measure (see
Theorem 1.33 below, see also [De] and [GZ, Section 6.6]).
Theorem 1.23 (Radon-Nikodym). Let and be two measures on (X, /). Suppose
that
(i) and are -nite, that is, there exists a sequence (X
i
)
i
/ such that
X =
i=1
X
i
and
(X
i
) < and (X
i
) < for each i .
(ii) << .
Then there exists a measurable function w : X [0, ] such that =
w
on / , that is,
(RN) (E) =
w
(E) :=
_
E
wd E /.
Moreover the function w in (RN) is -a.e. unique.
13
Denition 1.24. The function w in (RN) is called the Radon-Nikodym derivative of
with respect to and denoted by w =
d
d
.
Remark 1.25. Because is -nite, then w is also -integrable with respect to ,
that is,
(I) 0
_
X
i
wd < i ,
where (X
i
)
i
is the sequence in statement (i).
Remark 1.26. We can actually weaken the assumptions of the Radon-Nikodym
theorem. Indeed it is sucient to require that only is -nite in order that (RN)
holds (see, for instance,[Ro, Theorem 23, Chap. 11]). When is not -nite, the
Radon- Nikodym theorem fails (see Exercise I.8).
For Radon measures on locally compact, separable metric spaces, the Radon-
Nikodym theorem has the following simpler and stronger version.
Let us recall
Denition 1.27. Let (X, ) be a topological space.
(i) (X, ) is said to be separable if it has a countable dense subset.
(ii) (X, ) is said to be locally compact if for each x X there is an open set
O x such that the closure of O, denoted by O, is compact .
Theorem 1.28 (Radon-Nikodyms theorem for Radon measures). If X is supposed
to be a locally compact, separable metric space and and are Radon measures on X
with << , then (RN) holds and the Radon-Nikodym derivative w :=
d
d
is locally
integrable on X, i.e. w L
1
loc
(X, ), where
L
1
loc
(X, ) :=
_
f : X R : f is measurable,
_
K
[f[ d < for each compact K X
_
.
Before the proof of Theorem 1.28, let us recall the following well-known result of
topology (see [Ro, Proposition 7.6])
Lemma 1.29. Let (X, d) be a separable metric space and let D = x
i
: i N X
be dense. Then the family of open sets
| := B(x
i
, q) : i N, q Q (0, )
is a basis for the topology induced on X by the distance, where B(x, r) := y X : d(x, y) < r
if x X and r > 0.
Proof of Theorem 1.28. Firstly, let us prove there exists a sequence (V
i
)
i
X of open
sets such that
(1.11) X =
i=1
V
i
with V
i
compact set i .
Recall that, by denition, (X, d) is locally compact if and only if x X r
x
> 0
such that B(x, r
x
) is compact. Let D = x
i
: i N X be dense. From Lemma
1.29, the family | is a basis for the topology and let us enumerate |, that is assume
that | = B
i
: i N.Thus there exists a set I(x) N for which
B(x, r
x
) =
iI(x)
B
i
.
14
In particular, there exists a choice function : X N satisfying:
(1.12) x B
(x)
and B
(x)
B(x, r
x
) B(x, r
x
) .
Let J := (X) N and V
j
:= B
j
if j J.Then, by (1.12), (1.11) follows.
Because and are Radon measures, it follows that (V
i
) (V
i
) < and
(V
i
) (V
i
) < . Thus condition (i) of Theorem 1.23 holds with X
i
= V
i
. Then
there exists the Radon-Nikodym derivative w : X [0, ] that satises (I) with
X
i
= V
i
. Let K X =
i=1
V
i
be compact, then there exists i
0
N such that
(1.13) K
i
0
i=1
V
i
.
By (1.13), it follows that
0
_
K
wd
_
i
0
i=1
V
i
wd
i
0
i=1
_
V
i
wd <
and the proof is accomplished.
Historical remark: The rst version of Radon-Nikodyms theorem is due to H.
Lebesgue ([Le]) and to G. Vitali ([Vi]) when X = R in 1904. Radon extended the
result when X = R
n
([Ra]) in 1913. Eventually Nikodym ([Ni]) extended the result
to the abstract setting in 1930.
A consequence of the Radon-Nikodym theorem is the following
Lebesgue Decomposition Theorem 1.30. Let and be -nite measures on a
measure space (X, /). Then there is a decomposition of such that
=
ac
+
s
,
where
ac
and
s
are still measures on (X, /) with
ac
<< and
s
. The
decomposition is unique.
Proof. We employ the same device as in the proof of the Radon-Nikodym theorem.
More precisely let us dene the measure = + . Being << and << ,
consider the Radon-Nikodym derivatives
w
=
d
d
and w
=
d
d
.
Dene
A := x X : w
(x) = 0 .
Then A, B /, A B = and X = A B. Let
ac
(E) := (E A) =
_
EA
w
d E /,
s
(E) := (E B) E /,
and we will show they provide the desired decomposition.
First note that
(E) = (E (A B)) = (E A) + (E B) =
ac
(E) +
s
(E) .
15
Next we have that
(X A) = (B) =
_
B
w
d = 0 ,
s
(A) = (A B) = () = 0.
Thus
s
. To show that
ac
<< , consider E / with (E) = 0. Then
0 = (E) =
_
E
w
d =
_
EA
w
d.
Thus w
ac
+
s
,
with
ac
<< ,
ac
<< ,
s
,
s
and let us prove that
ac
=
ac
and
s
=
s
.
There exist S, S
/ such that
s
(X S) = (S) =
s
(X S
) = (S
) = 0 .
Thus, it also holds
(1.15)
s
(X (S S
)) = (S S
) =
s
(X (S S
)) = 0 ,
whence it follows that
(1.16)
s
(E) =
s
(E (S S
)) E / (since
s
(X (S S
)) = 0),
(1.17)
s
(E) =
s
(E (S S
)) E / (since
s
(X (S S
)) = 0),
(1.18)
ac
(E) =
ac
(E (X (S S
))) E / (since
ac
((S S
)) = 0),
(1.19)
ac
(E) =
ac
(E (X (S S
))) E / (since
ac
((S S
)) = 0),
On the other hand, from (1.14), we obtain that
ac
(E (X (S S
))) +
s
(E (X (S S
))) =
ac
(E (X (S S
))) +
s
(E (X (S S
))) .
From (1.16), (1.18) and (1.19), it follows that
ac
(E) =
ac
(E) E /.
Analogously we obtain
s
(E) =
s
(E) E /.
16
Exercise I.10 Let consider = /
1
, =
0
as measures on the - algebra /
1
of Lebesgue measurable sets in R,where
0
denotes the Dirac measure at 0, that is,
0
(E) :=
_
1 if 0 E
0 if 0 / E .
. Prove that the Lebesgue decomposition of with respect
to , =
ac
+
s
, is given by
ac
0 and
s
= .
Signed measures
Radon-Nikodyms theorem still holds for a more general class of measures. Namely
for set functions : / R which still verify basic properties of countable additivity.
Denition 1.31. (i) Let / be a -algebra on a set X. An extended real valued
set function : / R is a signed measure if it satises the following three
properties:
(SM1) assumes at most one of the values +, ;
(SM2) () = 0;
(SM3) For each sequence of disjoint sets (E
i
)
i
/, it holds that
(
i=1
E
i
) =
i=1
(E
i
)
where the series on the right either converges absolutely or diverges to
or +.
(ii) A signed measure : / R is said to be absolutely continuous with respect to
a measure : / [0, ] , written << , if (E) = 0 whenever (E) = 0.
(iii) A signed measure : / R is said to be -nite if there exists a sequence
(X
i
)
i
/ such that X =
i=1
X
i
and [(X
i
)[ < for each i.
Example: See Exercise I.11.
Remark 1.32. Observe that a measure is a signed measure. In some contexts we
will emphasize that a measure is not a signed measure by saying that it is a
positive measure.
Theorem 1.33 (Radon-Nikodyms theorem for signed measures). Let (X, /, ) be
a measure space with -nite, and : / R be a -nite signed measure such
that << . Then there exists a function w : X R measurable such that either
w
+
or w
is integrable and
(RN) (E) =
_
E
wd E /.
Proof. See [De] or [GZ, Theorem 6.38].
1.3. Lebesgue points and dierentiation theorem for Radon measures on
R
n
: Lebesgues dierentiation theorem for monotone functions. In this sec-
tion we are going to introduce the main results about the dierentiation of measures
which will be used later in order to solve the main problem stated at the beginning.
In this section the environment metric space will be X = R
n
and / = B(R
n
).
Notation: If A R
n
, x R
n
, r > 0, diam(A) := sup[x y[ : x, y A,
[A[ := /
n
(A),
_
A
f dx :=
_
A
f dx
[A[
, B(x, r) =: y R
n
: [y x[ < r , if B is a
17
(closed or open) ball in R
n
centered at x and radius r, by
B we will denote the ball
still centered at x and radius 5 r and it is called enlargement of the ball B.
Lebsegue points.
Let be a Radon measure on R
n
and assume that << := /
n
. Then, from the
Radon-Nikodym theorem for Radon measures, there exists w L
1
loc
(R
n
), w 0 such
that w =
d
d/
n
, that is,
(E) :=
_
E
wdx E B(R
n
) .
Remark 1.34. Assume w is continuous at a point x
0
R
n
, then
lim
r0
_
B(x
0
,r)
w(x) dx = w(x
0
)
(see Exercise I.12).
Question:
(D) lim
r0
(B(x, r))
[B(x, r)[
= lim
r0
_
B(x,r)
w(y) dy = w(x) (/
n
)a.e. x R
n
?
Theorem 1.35 (Lebesgue [Le2],1910). Let f L
1
loc
(R
n
). Then
lim
r0
_
B(x,r)
f(y) dy = f(x) a.e. x R
n
,
that is, by denition, there exists a negligible set N R
n
(i.e. [N[ = 0) such that
() lim
r0
_
B(x,r)
f(y) dy = f(x) x R
n
N .
Proof. We skip the proof (see [GZ, Theorem 7.8]).
Remark 1.36. If f L
1
loc
(R
n
), denote by
[f] :=
_
g : R
n
R : g measurable, g = f a.e. in R
n
_
.
Then the previous theorem actually states that there exists a negligible set N(f) such
that
lim
r0
_
B(x,r)
g(y) dy =
f(x) x R
n
N(f), g [f] ,
and
f(x) = f(x) x R
n
N(f) .
Thus the limit in () provides a way to dene the value of f at x that is independent
of the choice of representative in the equivalence class of f. Notice also that () can
be written as
lim
r0
_
B(x,r)
(f(y) f(x)) dy = 0 a.e. x R
n
.
It is rather surprising that Theorem 1.35 implies the following apparently stronger
result.
18
Theorem 1.37 (of Lebesgue points). : Let f L
1
loc
(R
n
), then a.e. x R
n
there
exists
(LP) lim
r0
_
B(x,r)
[f(y) f(x)[ dy = 0 .
Denition 1.38. A point x R
n
for which (LP) holds is called a Lebesgue point of
f.
Proof of Theorem 1.37. For each Q, apply the previous theorem to conclude that
there is a set N
R
n
with [N
[ = 0 such that
(1.20) lim
r0
_
B(x,r)
[f(y) [ dy = [f(x) [ x R
n
N
.
Thus, with N =
Q
N
19
Dierentiation of a Radon measure with respect to /
n
.
We are going to deal with the problem of the dierentiation in the case when is
a Radon measure on R
n
, even not absolutely continuous with respect to /
n
.
Recall that, if = /
n
, then the Lebesgue decomposition of with respect to
yields =
ac
+
s
with
ac
<< /
n
and
s
/
n
. Let us study the asymptotic
behaviour, as r 0, of
(B(x, r))
[B(x, r)[
=
ac
(B(x, r))
[B(x, r)[
+
s
(B(x, r))
[B(x, r)[
.
We proved (Corollary 1.39) that
D
L
n
ac
(x) =
d
ac
d/
n
(x) a.e. x R
n
Question: D
L
n
s
(x) a.e. x R
n
?
Lemma 1.40. Let be a Radon measure on R
n
that is (mutually) singular with
respect to /
n
, that is, /
n
. Then
D
L
n(x) = 0 a.e. x R
n
.
The proof of Lemma 1.40 relies on a fundamental covering type result do to G.
Vitali when n = 1, which will state without proof (see [GZ, Lemma 7.3]).
Lemma 1.41 (Vitali covering lemma). Let ( be a family of closed balls in R
n
such
that
R := sup diam(B) : B ( < .
Then there exists a countable subfamily T ( of pairwise disjoint elements such
that
BG
B
BF
B
Remark 1.42. Vitali covering lemma still holds for a collection of open balls whose
the radii are bounded. We emphasize that the point of the lemma is that the subcol-
lection consists of disjoint elements, a very important consideration since countable
additivity plays a central role in measure theory.
Proof of Lemma 1.40. Since /
n
, there exists a Borel set A R
n
such that
(R
n
A) = /
n
(A) = 0. For each positive integer k, let
E
k
:= (R
n
A)
_
x R
n
: limsup
r0
(B(x, r))
[B(x, r)[
>
1
k
_
.
It suces to show that
/
n
(E
k
) = 0 k
because it will follow that
lim
r0
(B(x, r))
[B(x, r)[
= 0 x (R
n
A)
k=1
E
k
.
Since (R
n
A) = 0, for xed k, referring to the approximation result of Borel
measures by open and closed sets (see Corollary 1.16), it follows that for every > 0
there is an open set U
R
n
such that
(1.21) U
E
k
and (U
) < .
20
For each x E
k
there exists a ball B
x
= B(x, r
x
) U
with
(1.22) r
x
< 1 and /
n
(B
x
) < k (B
x
) .
The collection of closed balls ( := B
x
: x E
k
provides a covering of E
k
. Now
employ Vitali covering lemma to obtain a countable, disjoint subfamily T ( such
that
E
k
BG
B
BF
B.
Then
/
n
(E
k
) /
n
_
BF
B
_
5
n
BF
/
n
(B)
(1.22)
5
n
k
BF
(B) = 5
n
k (
BF
B) 5
n
k (U
)
(1.21)
5
n
k .
Since is arbitrary, this shows that /
n
(E
k
) = 0.
From Corollary 1.39 and Lemma 1.40, the following result immediatly follows.
Theorem 1.43 (Derivative of a Radon measure with respect to /
n
). Suppose is
a Radon measure on R
n
. Let
ac
denote the absolutely continuous part of in the
Lebesgue decomposition with respect to /
n
. Then
D
L
n(x) =
d
ac
d/
n
(x) a.e. x R
n
.
Applications to the derivative of functions.
Theorem 1.44 (Lebesgues dierentiation theorem for monotone functions,1904).
Let a, b R with a < b. and let f : [a, b] R be a nondecreasing function. Then
(i) f
(x) a.e. x [a, b], that is , there exists a set N [a, b] such that /
1
(N) =
0 and
f
(x) := lim
yx
f(y) f(x)
y x
R x [a, b] N .
(ii) The function g : [a, b] R, dened by g(x) :=
_
f
(x) if x [a, b] N
0 if x N
is
integrable on [a, b] and
(WICI)
_
b
a
[f
(x)[ dx =
_
b
a
f
f
((x, y])
/
1
((x, y])
if x < y
f
((y, x])
/
1
((y, x])
if y < x
,
whence
f
(x) = c R lim
yx
+
f
((x, y])
/
1
((x, y])
= lim
yx
f
((y, x])
/
1
((y, x])
= c .
Obviously the half-open intervals
(x, y], (y, x]
are not balls in R centered at x. Indeed
B(x, r) = (x r, x + r) if r > 0 .
Therefore , the theorem which asserts the existence of the derivative of a Radon
measure with respect to /
1
(see Theorem 1.43) does not apply. However we are
going to show that a more general class of sets can be considered in the derivative of
measures, instead of balls.
Denition 1.46. Let x R
n
, a sequence of Borel sets (E
h
(x))
h
B(R
n
) is called a
(regular) dierentiation basis at x for /
n
provided there is
x
= (x) > 0 with the
following properties:
(i) there exists a sequence of balls (B(x, r
h
))
h
with r
h
0 such that
E
h
(x) B(x, r
h
) h;
(ii) [E
h
(x)[
x
[B(x, r
h
)[ h.
Theorem 1.47. Let be a Radon measure on R
n
, then there exists
lim
h
(E
h
(x))
[E
h
(x)[
=
d
ac
d/
n
(x) a.e. x R
n
,
whenever (E
h
(x))
h
is a dierentiation basis of /
n
at x.
Proof. Recall that
(1.23) =
ac
+
s
with
ac
<< /
n
and
s
/
n
,
(1.24) lim
r0
s
(B(x, r))
[B(x, r)[
= 0 a.e. x R
n
,
(1.25) lim
r0
(B(x, r))
[B(x, r)[
= lim
r0
ac
(B(x, r))
[B(x, r)[
=
d
ac
d/
n
(x) := w(x) a.e. x R
n
.
22
Let x R
n
be such that (1.24) holds and let (E
h
(x))
h
be a dierentiation basis at x.
Then
0
s
(E
h
(x))
[E
h
(x)[
1
s
(B(x
h
, r
h
))
[B(x
h
, r
h
)[
,
and, from (1.24), it follows that
(1.26) lim
h
s
(E
h
(x))
[E
h
(x)[
= 0 .
To conclude the proof we need only to prove that
(1.27) lim
h
ac
(E
h
(x))
[E
h
(x)[
= w(x)
when x R
n
satises (1.25). Indeed, (1.23), (1.26) and (1.27) conclude the proof.
From Lebesgue points theorem,
lim
h
_
B(x,r
h
)
[w(y) w(x)[ dy = 0 a.e. x R
n
.
Now observe that
_
E
h
(x)
[w(y) w(x)[ dy =
1
[E
h
(x)[
_
E
h
(x)
[w(y) w(x)[ dy
[B(x, r
h
)[
[E
h
(x)[
_
B(x,r
h
)
[w(y) w(x)[ dy
1
x
_
B(x,r
h
)
[w(y) w(x)[ dy 0 .
Thus, it follows that
ac
(E
h
(x))
[E
h
(x)[
w(x)
_
E
h
(x)
w(y) dy w(x)
_
E
h
(x)
(w(y) w(x)) dy
_
E
h
(x)
[w(y) w(x)[ dy 0 .
Then (1.27) holds.
Proof of Theorem 1.44. Without loss of generality, we can assume f : R R, by
extending it as f(x) = f(a) if x a and f(x) = f(b) if x b.
(i) Let us prove that f admits derivative a.e. on R. By Exercise I.13, there is a
right-continuous nondecreasing g : R R such that
g(x) = f(x) everywhere except possibly for a countable set ,
and
(1.28) if g
(x) = g
(x) .
Denote by
g
the LS o.m. induced by g. Recall that, being g right-continuous
(Theorem 1.11),
g
((a, b]) = g(b) g(a) a, b R with a b .
For each x R and each sequence (x
h
)
h
R x with lim
h
x
h
= x, denote
E
h
(x) :=
_
(x, x
h
] if x < x
h
(x
h
, x] if x
h
< x
.
23
It is easy to prove that (E
h
(x))
h
is a dierentiation basis at x for /
1
. The previous
result states that
lim
h
g(x
h
) g(x)
x
h
x
= lim
h
g
(E
h
(x))
[E
h
(x)[
=
d
g,ac
d/
1
(x) a.e. x R.
Since the sequence (x
h
)
h
is arbitrary, it follows that
(1.29) g
(x) =
d
g,ac
d/
1
(x) a.e. x R.
From (1.28) and (1.29), we obtain that
(1.30) f
(x) =
d
g,ac
d/
1
(x) a.e. x R,
then the desired conclusion.
(ii) Extend f : R R as before. From (1.30), it follows that f
is Lebesgue measur-
able because it agrees with the Borel function
d
g,ac
d/
1
a.e. on R. Note also that, since
f is nondecreasing,
f
(x) 0 a.e. x R.
Eventually let us prove that
_
b
a
f
(x) a.e. x R.
By Fatous lemma, we obtain
_
b
a
f
(x) dx =
_
b
a
liminf
h
g
h
(x) dx liminf
h
_
b
a
g
h
(x) dx =
liminf
h
_
h
__
b
a
f(x + 1/h) dx
_
b
a
f(x) dx
__
=
liminf
h
_
h
_
_
b+1/h
a+1/h
f(x) dx
_
b
a
f(x) dx
__
=
liminf
h
_
h
_
_
b+1/h
b
f(x) dx
_
a+1/h
a
f(x) dx
__
liminf
h
_
h
_
f(b + 1/h)
h
f(a)
h
__
= f(b) f(a) .
Another consequence of the above results is the following theorem concerning the
derivative of the indenite integral, which also yields an answer to the main problem
(1
i).
24
Theorem 1.48 (Derivative of the indenite integral). Suppose f : [a, b] R is a
Lebesgue integrable function, i.e. f is Lebsegue measurable and
_
b
a
[f[ dt < . For
each x [a, b] let
F(x) :=
_
x
a
f(t) dt .
Then
(i) F C
0
([a, b]);
(ii) F
_
x
h
x
f(t) dt
_
I
h
[f[ dt
where I
h
= [x, x
h
] if x x
h
and I
h
= [x
h
, x] if x > x
h
. Therefore, from Lebesgue
dominated convergence theorem, we obtain
lim
h
_
I
h
[f[ dt =
_
{x}
[f[ dt = 0 .
(ii) Firstly, notice that we ca assume that f 0 a.e. on [a, b]. Othewerwise we can
decompose it as f = f
+
f
25
1.4. Functions of bounded variation.
Lebesgues dierentiation theorem for monotone functions (see Theorem 1.44) ac-
tually extends to a more general class of functions, Namely, it holds for the functions
of bounded variation, a class of functions we will introduce in this section.
Denition 1.49. Let f : [a, b] R.
(i) The total variation of f on [a, b] is dened by
V
f
(a; b) := sup
_
m
i=1
[f(x
i
) f(x
i1
)[ : x
0
= a < x
1
< < x
m
= b
_
where the supremum is taken on over all nite sequences x
0
= a < x
1
< <
x
m
= b.
(ii) f is said to be of bounded variation (abbreviated BV ) on [a, b] if V
f
(a; b) < .
We will denote by BV ([a, b]) the set of functions with bounded variation on [a, b].
Notation: V
f
(a; a) = 0.
Examples:
(i) Let f : [a, b] R be a monotone function (nondecreasing or nonincreasing).
Then
V
f
(a; b) = [f(b) f(a)[ < .
(ii) Let f : [0, 1/] R, f(x) :=
_
_
_
x sin
_
1
x
_
if 0 < x
1
0 if x = 0 .
.
Then f / BV ([0, 1/]) (see Exercise I.16).
Remark 1.50. Example (i) implies that a function of bounded variation need not
be continuous. Example (ii) implies that a continuous function on [a, b] (also dier-
entiable on (a, b]) need not be with bounded variation.
Proposition 1.51. BV ([a, b]) is a vector space (on R).
Proof. See Exercise I. 17 (i) and (ii).
The following characterization of functions with bounded variations holds.
Theorem 1.52 (Jordans decomposition theorem). Let f : [a, b] R. Then the
following are equivalent:
(i) f BV ([a, b]);
(ii) there exist g, h : [a, b] R nondecreasing such that f = g h.
From the previous theorem and Lebesgues dierentiation theorem for monotone
functions (see Theorem 1.44), it immediatly follows the following result.
Corollary 1.53. Let f BV ([a, b]). Then f
Remark 1.54. Naively it might be expected that the set of points of a continuous
function where it is not dierentiable should be small in some sense. This might
be because it is dicult to draw or visualise a continuous function whose set of
nondierentiable points is something other than a nite set of points. This is not
true. Indeed there are several examples of continuous nowhere dierentiable function
f : R R (see Exercise I.18).
Historical remark: The rst example of a continuous nowhere dierentiable
function , along with the proof that it is nowhere dierentiable, was rst given by
Weierstrass in a paper presented to the K onigliche Akademie der Wissenschaften on
18 July 1872.
1.5. The fundamental theorem of calculus.
Come back to the main problem 2, that is, to characterize the class of functions
f : I R satisfying the identity.
(ICI) f(x) f(y) =
_
x
y
g(t) dt x, y I with y x ,
for some locally integrable function g : I R, where I R is a given interval. Some
considerations are now in order.
From Theorem 1.48, if (ICI) holds, then
f
(x) = 0 a.e. x I;
f is nondecreasing.
We are now going to select, by using the introduced tools of measure theory , the
properties in order that (ICI) holds.
Remark 1.55. Let f : [a, b] R satisfy (ICI) on I = [a, b]. Extend it to the whole
R in usual way, that is, f(x) = f(a) if x a and f(x) = f(b) if x b. Then it is
easy to verify that f veries (ICI) on I = R and f
(t) dt x, y R, x < y ;
(ii) f is continuous and
f
<< /
1
, where
f
denotes the Lebesgue-Stieltjes o.m.
on R induced by f.
Proof. Firstly, let us prove the implication (ii) (i). We know, because f is contin-
uous, that (see Theorem 1.11)
(1.31)
f
((x, y]) = f(y) f(x) x, y R, x < y .
Since << /
1
, from the Lebesgue decomposition theorem (Theorem 1.30) and
Lebesgues dierentiatiom theorem for monotone functions (Theorem 1.44), it follows
that
d
f,ac
d/
1
=
d
f
d/
1
= f
a.e. in R,
whence, by Radon- Nikodyms Theorem (Theorem 1.23),
(1.32)
f
(E) =
_
E
f
(t) dt E B(R) .
Choosing E = (x, y], from (1.31) and (1.32), (i) follows.
Let us prove the implication (i) (ii). The continuity of f immediatly follows from
derivative of the indenite integral result (Theorem 1.48). To conclude the proof it
is sucient to show that (1.32) holds. Denote by
f
the LS o.m. induced by f and
by /
1
f
: B(R) [0, ] the measure
/
1
f
(E) :=
_
E
f
(t) dt E B(R) .
Then (ICI) on I = R reads as follows
f
((y, x)) = /
1
f
((y, x)) x, y R with y < x ,
since f is continuous (see Exercise I.4 ii). From (ICI) on I = R and Exercise I.6, it
follows that
f
(U) = /
1
f
(U) for each open set U R.
By using the approximation of measures by open and closed sets, we obtain (1.32).
We are now in position to introduce the class of function satisfying (ICI).
Denition 1.58 (Vitali, 1904). Let I R be an interval. A function f : I R is
said to be absolutely continuous (briey, f AC(I)) if for every > 0 there exists
= () > 0 such that
m
i=1
[f(b
i
) f(a
i
)[ <
28
for any nite collection of nonoverlapping intervals (a
1
, b
1
), . . . , (a
m
, b
m
) in I with
m
i=1
[b
i
a
i
[ < .
Remark 1.59. (i) From the denition, it follows that an absolutely continuous
function is uniformly continuous. The converse is not true as proved by the
Cantor-Lebesgue-Vitali function (see Exercise I.15).
(ii) It is easy to see that AC(I) is a vector space.
Exercise I.19 Let f : [a, b] R be a Lipschitz function, that is, by denition,
L > 0 such that [f(x) f(y)[ L[x y[ for each x, y [a, b]. Prove that
f AC([a, b]).
Exercise I.9 Let and be two Radon measures on R. Suppose that (R) < +.
Then the following are equivalent:
(i) << ;
(ii) for each > 0 there exists = () > 0 such that
m
i=1
((a
i
, b
i
)) < ,
for each disjoint family of intervals (a
1
, b
1
), . . . , ( a
m
, b
m
), verifying
m
i=1
((a
i
, b
i
)) < .
A rst consequence of Theorem 1.57 and Exercise I.9 is the fundamental theorem
of calculus for monotone functions.
Theorem 1.60 (Fundamental theorem of calculus for monotone functions). Let f :
[a, b] R be a nondecreasing function. Then the following are equivalent:
(i) f satises (ICI) on I = [a, b], that is,
f(x) f(a) =
_
x
a
f
(t) dt x [a, b] ;
(ii) f AC([a, b]).
Proof. Let
f : R R denote the usual extension of f dened by
f(x) = f(a) if
x a and
f(x) = f(b) if x b. Then, it is easy to see that
(1.33) f AC([a.b])
f AC(R) .
Applying Exercise I.9 with =
f
and = /
1
, it follows that
(1.34)
f
<< /
1
f AC(R) .
Since
f is continuous, from the characterization of (ICI) by the absolute continuity
of measures (Theorem 1.57) , it follows that
(1.35)
f
<< /
1
f satises (ICI) on I = R.
29
Now notice that
(1.36)
f satises (ICI) on I = R f satises (ICI) on I = [a, b] .
From (1.33), (1.34), (1.35) and (1.36), the proof is accomplished.
Corollary 1.61. Let f : [a, b] R be a nondecreasing function. Suppose that
(i) f AC([a, b]);
(ii) f
i=1
[f(x
i
) f(x
i1
)[ : x
0
= a < x
1
< < x
m
= b
_
.
By assummption of the absolutely continuity, if = 1 there exists
1
= (1) > 0 such
that
(1.37)
m
i=1
[f(b
i
) f(a
i
)[ < 1
for any nite collection of nonoverlapping intervals (a
1
, b
1
), . . . , (a
m
, b
m
) in [a, b] with
m
i=1
[b
i
a
i
[ <
1
.
Divide the interval [a, b] in k subintervals with equal lenght and k > b a/
1
,
x
0
= a < x
1
< < x
k
= b .
From (1.37),then it follows that
(1.38) V
f
(x
i1
; x
i
) 1 i = 1, . . . , k .
By Exercise I.17 (iii) and (1.38), we obtain
V
f
(a; b) =
k
i=1
V
f
(x
i1
; x
i
) k < .
30
Proof of Theorem 1.62. From Lemma 1.63, it follows that, if f AC([a, b]), is well-
dened the variation function g : [a, b] R, g(x) := V
f
(a; x) if a x b. Moreover,
from Jordans decomposition theorem (Theorem 1.52), the function h : [a, b] R,
dened by h(x) := f(x) g(x) if a x b, is nondecreasing.
From Exercise I.20, it follows that, if f AC([a, b]) also g AC([a, b]). Since
AC([a, b]) is a v.s. h AC([a, b]), too. Thus the proof is complete.
Fundamental theorem of calculus 1.64. Let f : [a, b] R. Then f AC([a, b])
if and only if
(i) f is dierentiable a.e. in [a, b];
(ii) f
= g
a.e. in [a, b] .
Hence (i) follows. By (1.40), we obtain that
[f
[ [g
[ + [h
[ ,
whence
_
b
a
[f
[ dx
_
b
a
[g
[ dx +
_
b
a
[h
(t) dt
_
x
a
h
(t) dt =
_
x
a
(g
(t) h
(t)) dt =
_
x
a
f
(t) dt .
and (iii) also follows.
Let us now prove the suciency of conditions (i), (ii) and (iii). Notice that, if
(a
1
, b
1
), . . . , (a
m
, b
m
) is a collection of nonoverlapping intervals contained in [a.b], from
(iii) we have that
(1.41)
m
i=1
[f(b
i
) f(a
i
)[ =
m
i=1
_
b
i
a
i
f
(t) dt
i=1
_
b
i
a
i
[f
(t)[ dt .
Dene g : R R, g(x) :=
_
[f
(x)[ if x [a, b]
0 otherwise
, and let and be the measures
dened, respectively, by
(E) :=
_
E
g(t) dt, (E) := /
1
(E) if E /
1
.
31
Since is nite and << , by Exercise I.9 it follows that lim
(E)0
(E) = 0, that
is, for each > 0 there exists = () > 0 such that
m
i=1
((a
i
, b
i
)) =
m
i=1
_
b
i
a
i
g dt < ,
for each disjoint family of intervals (a
1
, b
1
), . . . , ( a
m
, b
m
), verifying
m
i=1
((a
i
, b
i
)) =
m
i=1
[b
i
a
i
[ < .
Therefore, from (1.41), it follows that f AC([a, b]).
2. Main spaces of functions and results on Banach and Hilbert spaces
([B, De, GZ, Ro, R1, R2]).
Motivation: In this chapter we are going to introduce and study some classical
spaces of functions of Analysis. The study concerns with topological and metric
features of concrete innite dimensional spaces as far as completeness, compactness
and separability. Meanwhile, we will recall some fundamental abstract results about
Banach and Hilbert spaces that you should already have studied in some previous
course, as Analisi Funzionale (see [De]).
2.1. The space of continuous functions C
0
().
Denition 2.1. (i) Let A R
n
be a set,
C
0
(A) := f : A R, f continuous at any x A .
(ii) Let K R
n
be a compact set and let f C
0
(K). We will denote by |f|
= |f|
,K
= sup
xK
[f(x)[ .
| |
) is an innite
dimensional Banach space (B.s.).
Proof. We restrict ourselves to the case when n = 1 and = (a, b) and leave as
exercise to prove that (C
0
(), | |
k N such that
|f
h
f
k
|
= sup
x
[f
h
(x) f
k
(x)[ < h, k >
k .
This implies that
(2.1) > 0
k N such that [f
h
(x) f
k
(x)[ < h, k >
k, x .
From (2.1) (f
h
(x))
h
R turns out to be a Cauchy sequence. Thus,
(2.2) f(x) := lim
h
f
h
(x) x .
32
From (2.2), passing to the limit in (2.1),as k , we obtain
> 0
k N such that [f
h
(x) f(x)[ h >
k, x ,
that is, by denition, f
h
f uniformly in . By Exercise II.2, f C
0
().
Question: Is (C
0
(), ||
= (f, f)
1/2
for all f C
0
()?
Exercise II.3: It doesnt!
Compactness in (C
0
(), | |
).
We are now going to characterize the compact subsets of (C
0
(), | |
). Firstly,
recall some important notions and results related to the compactness topic in a metric
space (see [GZ, Section 3.5].)
Denition 2.3. Let (X, d) be a metric space and denote by B(x, r) the (open) ball
in X centered at x X with radius r > 0.
(i) A point x
0
X is called limit point of a set A X provided A(B(x
0
, r) x
0
) ,=
for all r > 0.
(ii) A set A X is called bounded if there is R
0
> 0 such that d(x, y) R
0
for
all x, y A.
(iii) A set A X is called totally bounded if, for every > 0 , A can be covered
by a nite family of balls
B(x
1
, ), . . . , B(x
N
, ) ,
that is, A
N
i=1
B(x
i
, ).
(iv) A set A X is called sequentially compact if every sequence in A has a
subsequence that converges to a point in A.
(v) A set A X is said to have the Bolzano-Weierstrass (BW) property if every
innite subset of A has a limit point that belongs to A.
Remark 2.4. It is easy to see that a totally bounded set is bounded, the converse is
seen to be false (see Remark 2.8 below). In a topological space (X, ) compact sets
and sequentially compact sets satisfy (BW) property. No other implications hold in
general.
Historical remark: [P, Section 3.2.3] Compactness is certainly the most important
concept of general topology. In the context of metric spaces, it was introduced by
Frechet in 1906. However from our present point of view , its formulation in terms of
sequences did not express the real essence which is hidden in the covering properties.
This observation was emphasized by Alexandro/Urysohn in 1924.
In a metric space we have the following characterization of the compact sets (see
[GZ, Theorem 3.33]).
Theorem 2.5 (Characterization of the compact sets in metric spaces). If A is a
subset of a metric space (X, d), the following are equivalent:
(i) A is compact;
(ii) A is sequentially compact;
33
(iii) A is closed and totally bounded;
(iv) A has the BW property.
Corollary 2.6. Let A R
n
. Then
A is compact A is closed and bounded.
Theorem 2.7 (Riesz, 1918). Let (E, | |) be a normed vector space and denote
B
E
:= x E : |x| 1 . Then B
E
is compact if and only if dim
R
E < .
Proof. See [De], [B, Theorem 6.5] or [GZ, Exercise 8.9].
Remark 2.8. Theorem 2.7 yields that a bounded set A in a innite dimensional n.v.s
(E, | |) need not be totally bounded. For instance the set A = B
E
.
Denition 2.9. Let A R
n
. A family T C
0
(A) is said to be equicontinuous if
for every > 0 there exists = () > 0 such that for each f T, [f(x) f(y)[ <
whenever x, y A with [x y[ < .
We are now in position to give a characterization of the compact sets of (C
0
(K), |
|
) when K R
n
is compact.
Theorem 2.10 (Arzel` a -Ascoli). Let K R
n
be compact and let T C
0
(K). Then
T is compact in (C
0
(K), | |
) if and only if T is
(i) closed in (C
0
(K), | |
);
(ii) bounded in (C
0
(K), | |
);
(iii) equicontinuous.
Historical remark: The notion of equicontinuity was introduced at around the
same time by Ascoli (18831884) and Arzel` a (18821883). A weak form of the
theorem was proven by Ascoli (18831884), who established the sucient condition
for compactness, and by Arzel`a (1895), who established the necessary condition and
gave the rst clear presentation of the result. A further generalization of the theorem
was proven by Frechet (1906) for a space in which the notion of a limit makes sense,
such as a metric space or Hausdor space.
In many applications, it is not of great interest to know whether T itself is compact,
but whether a given sequence (f
h
)
h
T has a subsequence (f
h
k
)
k
that converges to an
element f C
0
(K), and not necessarily to f T. In other words, the compactness
of the closure of T is the critical question. It is easy to see that if T is equicontinuous
so is T. This leads to the following corollary.
Corollary 2.11. Let K R
n
be compact and let T C
0
(K). Suppose that T is
bounded and equicontinuous. Then T is compact in (C
0
(K), | |
).
In particular this corollary yields the following special result.
Corollary 2.12. Let f
h
: [a, b] R (h = 1, 2, . . . ) be a sequence of continuous
functions. Suppose that
(i) M > 0 such that [f(x)[ M x [a, b], h.
(ii) (f
h
)
h
is equicontinuous, that is, > 0 = () > 0 such that [f
h
(x)
f
h
(y)[ < x, y [a, b] with [x y[ < , h.
34
Then there are a subsequence (f
h
k
)
k
and a function f C
0
([a, b]) such that f
h
k
f
uniformly in [a, b].
Exercise II.5: Let M > 0 be a given constant and let T = f C
1
([a, b]) : [[f[[
C
1
M. Prove that
(i) T is a relatively compact set of (C
0
([a, b]), [[ [[
) ;
Outline of the proof: (i) Show that
(1) T is closed and bounded in (C
0
([a, b]), [[ [[
) (this is immediate!) ;
(2) T is equicontinuous .
To prove (2), it suces to note that,
[f(x) f(y)[ M [x y[ x, y [a.b], f T .
Proof of Theorem 2.10. Suciency: assume that (i), (ii) and (iii) hold and show that
T is compact. Referring to the characterization of compact sets (Theorem 2.5),
it suces to show that T is sequentially compact. Thus it suces to show that an
arbitary sequence (f
h
)
h
T has a convergent subsequence (f
h
k
)
k
to a function f T,
that is,
|f
h
k
f|
0 as k .
Recall that K is compact and separable. Let D := x
i
: i N denote a countable
dense set in K. The boundedness of T implies that there exists M
1
> 0 such that
|f g|
M
1
f, g T .
In particular, if we x an element f
0
T, then
|f
0
f
h
|
M
1
h N.
Moreover
| f
h
|
=|(f
h
f
0
) + f
0
|
|f
h
f
0
|
+ |f
0
|
M
1
+ |f
0
|
:= M
2
.
Thus, there is a suitable constant M
2
> 0 such that
[f
h
(x)[ M
2
x K, h.
We are now going to construct a convergent subsequence by means of Cantors diag-
onal process. More precisely,
Step 1: (f
h
(x
1
))
h
is a sequence of real number contained in [M
2
, M
2
]. It follows that
this sequence has a subsequence (f
(1)
h
(x
1
))
h
converging in R;
Step 2: Consider now the sequence (f
(1)
h
(x
2
))
h
[M
2
, M
2
]. Thus, it also has a
convergent subsequence (f
(2)
h
(x
2
))
h
. Note also that (f
(2)
h
(x
1
))
h
is still convergent since
it is a subsequence of (f
(1)
h
(x
1
))
h
.
Continuing this process we obtain at
Step k: a subsequence (f
(k)
h
)
h
of (f
(k1)
h
)
h
such that (f
(k)
h
(x
j
))
h
is convergent for each
xed j = 1, . . . , k.
Diagrammatically, we have the following situation:
35
Dene
g
k
:= f
(k)
k
: K R.
Observe that, for each i = 1, 2, . . . , the sequence (g
k
)
ki
is a subsequence of (f
i
k
)
ki
.
In particular this implies that (g
k
)
k
is a subsequence of (f
h
)
h
and, by construction,
(2.3) (g
k
(x))
k
is convergent in R, x D.
We now proceed to show that
(2.4) (g
k
)
k
is convergent in (C
0
(K), | |
) .
Use the assumption that T is equicontinuous, that is,
(2.5)
> 0 = () > 0 : x, y K and [x y[ < [f(x) f(y)[ < f T .
Fix an arbitrary > 0, then = () is also xed. Because K is totally bounded ,
for each > 0 there is a nite family of balls B(x
1
, ), . . . , B(x
N
, ) of R
n
such that
N = N(), x
i
K for each i = 1, . . . , N, and
K
N
i=1
B(x
i
, ) .
By the density of D in K, y
i
D B(x
i
, ) for each i = 1, . . . , N. In particular
K
N
i=1
B(y
i
, 2 ) .
Thus, choosing now = /2, we obtain there exists N = N() = N() = N() and
D
:= y
1
, . . . , y
N
D such that
(2.6) K
N
i=1
B(y
i
, ) .
Since, by (2.3), each of the sequences
(g
k
(y
1
))
k
, . . . , (g
k
(y
N
))
k
,
converges, there is an integer k = k() with
[g
k
(y
i
) g
r
(y
i
)[ < k, r > k, i = 1, . . . N .
By (2.6),
x K y
i
D
such that [x y
i
[ <
(2.5)
[g
k
(x) g
k
(y
i
)[ < , k N.
Therefore, we have
[g
k
(x) g
r
(x)[ [g
k
(x) g
k
(y
i
)[ + [g
k
(y
i
) g
r
(y
i
)[ + [g
r
(y
i
) g
r
(x)[
+ + = 3 x K ,
provided k, r k. This shows that for each > 0 there exists k = k() such that
|g
k
g
r
|
3 k, r > k .
That is (g
k
)
k
is a Cauchy sequence in (C
0
(K), ||
). Since (C
0
(K), ||
) is complete
(Thorem 2.2) and T is closed, it follows that f T such that lim
k
|g
k
f|
= 0.
Since (g
k
)
k
is a subsequence of original sequence (f
h
)
h
, we have shown that T is
sequentially compact, thus establishing the sueciency argument.
Necessity: show that, if T is compact in (C
0
(K), | |
), then,
from the characterization of the compact sets in metric spaces (Theorem 2.5), T is
36
closed and totally bounded and therefore bounded. Show that T is equicontinuous,
that is, we have to prove (2.5). By contradiction, assume
0
> 0 : > 0 f
T, x
, y
K with [x
[ < and
[f
(x
) f
(y
)[
0
.
Choose = 1/h, and denote by f
h
:= f
1/h
, x
h
:= x
1/h
and y
h
:= y
1/h
. Then we
have constructed three sequences (f
h
)
h
T, (x
h
)
h
, (y
h
)
h
K verifying
(2.7) [x
h
y
h
[ < 1/h, [f
h
(x
h
) f(y
h
)[
o
> 0 h.
Since T and K are compact, there exist three subsequences, that we will still denote
(f
h
)
h
T, (x
h
)
h
, (y
h
)
h
K such that
lim
h
x
h
= lim
h
y
h
= z K and f
h
f T uniformly in K .
By Exercise II.6, lim
h
f
h
(x
h
) = lim
h
f
h
(y
h
) = f(z). But, passing to the limit
in (2.7), we reach a contradiction. Thus, the proof is complete.
Remark 2.13. The Arzel`a -Ascoli theorem does not hold in C
0
(A) when A R
n
is not compact. For instance let C
0
b
(R) denote the space of bounded continuous
functions on R, i.e.,
C
0
b
(R) :=
_
f C
0
(R) : sup
R
[f[ <
_
.
Then it is easy to see that (C
0
b
(R), | |
) (see Corollary
2.11). Indeed, note that
f(x) := lim
h
f
h
(x) = 0 x R and |f
h
f|
= 1 h.
This implies that convergent subsequences of (f
h
)
h
in (C
0
b
(R), | |
).
Recall some relevant topological notions and results (see, for instance,[Ro, Chapters
7 and 8]).
Denition 2.14. Let (X, ) be a topological space. Then (X, ) is said to satisfy the
second axiom of countability if it has a countable basis for the topology .
Theorem 2.15. Let (X, d) be a metric space. Then
(i) (X, d) is separable if and only if it satises the second axiom of countability.
(ii) Every subspace of (X, d) is still separable, provided (X, d) is separable.
(iii) Let (Y, ) be an other metric space and let T : (X, d) (Y, ) be an homeor-
phism. Then (X, d) is separable if and only if so is (Y, )
37
Remark 2.16. We stress that this notion is very important and useful in Analysis
for approximation purposes. Think to the rational numbers and the proof of Arzel`a
-Ascolis theorem (Theorem 2.10).
Eventually, recall the following useful criterion to check out that a topological space
is not separable.
Proposition 2.17. Let (X, ) be a topological space. Assume there exists a family
U
i
: i I such that:
(i) U
i
is an open set for each i I;
(ii) U
i
U
j
= if i ,= j;
(iii) I is uncountable.
Then (X, ) is not separable
Proof. See [B, Lemma 4.2] or [Ro, Proposition 6, Section 7.2].
Exercise II.37(ii) Let
l
:= x R
N
: sup
nN
[x(n)[ < ) ,
equipped with the norm
|x|
l
:= sup
nN
[x(n)[ .
Show that (l
, [[ [[
l
) is a Banach space, but it is not separable.
Hint: Let
I = 2
N
:= x : N 0, 1 l
and
U
x
= B
l
_
x,
1
2
_
:=
_
y l
: |y x|
l
<
1
2
_
if x I .
Then consider the family U
x
: x I and apply Proposition 2.17.
Theorem 2.18. Let K R
n
be a compact set. Then (C
0
(K), | |
) is separable.
We are going to prove the previous theorem when n = 1 K = [a, b]. Before we need
to state one of the most important approximation results of Mathematical Analysis.
Theorem 2.19 (Weierstrass approximation theorem, 1885). Let f C
0
([a, b]). Then
there exists a sequence of polymomial functions p
h
: R R (h = 1, 2, . . . ) with real
coecients, that is p
h
R[x], such that
p
h
f uniformly in [a, b] .
Proof. We skip the proof (see [R1, Theorem 7.26]).
Remark 2.20. Because polynomials are the simplest functions, and computers can
directly evaluate polynomials, this theorem has both practical and theoretical rele-
vance, especially in polynomial interpolation.
38
Proof of Theorem 2.18. We only show the result when n = 1 and K = [a, b].
Let T denote the set of polynomial functions with rational coecients, that is, T :=
Q[x]. It is well-known that T is countable. Prove that T is dense in (C
0
([a, b]), ||
),
that is,
f C
0
([a, b]), > 0 q T such that |f q|
.
From the Weierstrass approximation theorem (Theorem 2.19), for given arbitrary
> 0 there exist p R[x], that is, p(x) =
m
x
m
+ +
1
x +
0
, with
i
R
i = 0, 1, . . . , m, such that
(2.8) |f p|
<
2
.
Dene q(x) :=
m
x
m
+ +
1
x +
0
with
i
Q such that
[
i
i
[ <
2
m
i=0
c
i
i = 0, 1, . . . , m,
where c := max[a[, [b[. Then
(2.9) [p(x) q(x)[
m
i=0
[
i
i
[ [x[
i
2
x [a, b] .
Thus, by (2.8) and (2.9), we obtain
|f q|
|f p|
+ |p q|
2
+
2
= .
||1
|D
f|
,
.
39
| |
C
1 is called C
1
norm.
Theorem 2.22. Let R
n
be a bounded open set. Then (C
1
(), | |
C
1) is an
innite dimensional Banach space (B.s.), but is not a Hilbert space.
Proof. We restrict ourselves to the case when n = 1 and = (a, b). Firstly, show
that it is complete and it is not a Hilbert space. Consider the linear map
(2.10) T : (C
1
(), | |
C
1) (C
0
() C
0
(), | |
C
0
()C
0
()
), T(f) := (f, f
)
where
|(f, g)|
C
0
()C
0
()
:= |f|
+ |g|
if (f, g) C
0
() C
0
() .
Note that T is an isometry, that is,
|T(f)|
C
0
()C
0
()
= |f|
C
1 f C
1
() .
In particular, if we dene M := T(C
1
()),then M is a vector space and the map
T : (C
1
(), | |
C
1) (M, | |
C
0
()C
0
()
) is an onto isometry.
Exercise II.7: Let (E, [[ [[
E
) and (F, [[ [[
F
) be B. s. Endow EF with the norm
[[(x, y)[[
EF
= [[x[[
E
+[[y[[
F
. Then (E F, [[(x, y)[[
EF
) is a B.s.
Thus, we need only to show that M is closed in (C
0
() C
0
(), | |
C
0
()C
0
()
),
in order to accomplish the proof.
Assume ((f
h
, f
h
))
h
M is a sequence that
(2.11) lim
h
|(f
h
f, f
h
g)|
C
0
()C
0
()
= 0
for some (f, g) C
0
() C
0
(), and prove that
(2.12) f
= g in [a, b] .
By (2.11), we obtain that f
h
f and f
h
g uniformly in [a, b] and therefore also
pointwise. Since, by the classical fundamental theorem of calculus,
f
h
(x) f
h
(a) =
_
x
a
f
h
(t) dt x [a, b], h,
we can pass to the limit, as h , in the previous identity and (2.12) follows.
Exercise II.4: Show that (C
1
(), [[ [[
C
1) is a B. s.,where [[u[[
C
1 :=
||1
[[D
u[[
and R
n
is a bounded open set.
C
1
() is trivially an innite dimensional vector space because it contains the set
of polynomials.
It is not a Hilbert space. Indeed we can argue as in Exercise II.3.
Compactness in (C
1
(), | |
C
1).
Theorem 2.23. Let T C
1
() and denote by T
i
:= D
i
f : f T i = 1, . . . , n.
Then T is compact in (C
1
(), | |
C
1) if and only if T and T
i
, for each = 1, . . . , n,
are
(i) bounded in (C
0
(), | |
C
0);
(ii) closed in (C
0
(), | |
C
0);
40
(iii) equicontinuous on .
Proof. We restrict ourselves to the case when n = 1 and = (a, b).
Necessity: Show that, if T is compact (C
1
(), | |
C
1), then (i), (ii) and (iii) hold. Let
T : (C
1
(), | |
C
1) (C
0
() C
0
(), | |
C
0
()C
0
()
) be the map dened in (2.10).
In the proof of Theorem 2.22, we showed that there exists T
1
: (M, ||
C
0
()C
0
()
)
(C
1
(), | |
C
1) continuous. Thus,
T is compact in (C
1
(), | |
C
1)
T(T) is compact in (C
0
() C
0
(), | |
C
0
()C
0
()
) .
Observe now that, if we denote by
i
: C
0
() C
0
() C
0
() (i = 1, 2) the
projections on the coordinate spaces, that is,
i
(f
1
, f
2
) = f
i
if (f
1
, f
2
) C
0
()
C
0
(), then
i
is continuous. Since T is compact (C
1
(), | |
C
1), T(T) is also
compact in (C
0
() C
0
(), | |
C
0
()C
0
()
). This implies that
1
(T(T)) = T and
2
(T(T)) = T
+ |f
1
_
.
Then T is not compact in (C
1
([a, b]), [|
C
1) beacuse of Rieszs theorem (Theorem 2.7)
(recall that C
1
([a, b]) is an innite dimensional space!). But T is relatively compact
in (C
0
([a, b]), | |
= 0 .
Separability of (C
1
(), | |
C
1).
Theorem 2.25. (C
1
(), | |
C
1) is separable.
Proof. Let T : (C
1
(), | |
C
1) (M, | |
C
0
()C
0
()
) be the map dened in (2.10).
Since it is a homeomorphism and the separability property is stable by homeomor-
phisms (see Theorem 2.15 (iii)), we need only to show that the subspace (M, |
|
C
0
()C
0
()
) of metric space (C
0
() C
0
(), | |
C
0
()C
0
()
) is separable . This
follows by the separability of (C
0
(), | |
([f[) [x y[ = L[x y[ x, y .
Remark 2.31. (i) The previous inclusion does not hold if is not convex.
Example: Let = (x, y) R
2
: y <
_
[x[, x
2
+ y
2
< 1 and f(x, y) :=
sign(x)
_
y
if y > 0
0 if y 0
, with 1 < < 2. Then f C
1
() Lip(). Indeed it
is easy to see that f C
1
(). Let us prove that f / Lip(). By contradiction,
42
assume f Lip(). Then there exists L > 0 such that, for each (x, y) with
x > 0, y > 0
[f(x, y) f(x, y)[ = 2 y
2 Lx y
_
x
L
_
1/
.
Since 1/2 < 1/, we can choose ( x, y) such that
x > y >
_
x
L
_
1/
, which
contradicts the previous inequality.
(ii) The inclusion is strict.
Example: Let = (1, 1) and f(x) = [x[. Then f Lip() C
1
().
Denition 2.32. Let be a bounded open set and let f Lip() . We will denote
by
|f|
Lip
= |f|
Lip,
:= |f|
,
+ Lip(f, ) .
| |
Lip
is called Lip norm.
Theorem 2.33. (Lip(), | |
Lip
) is an innite dimensional Banach space that is not
a Hilbert space, provided R
n
is a bounded open set.
Proof. It is easy to see that (Lip(), | |
Lip
) is a normed linear space noticing that
(2.13) Lip(f + g) Lip(f) + Lip(g) f, g Lip() .
Let us show that is complete. Let (f
h
)
h
be a given Cauchy sequence in (Lip(), ||
Lip
),
i.e. for every > 0 there exists h = h() N such that
[f
h
(x) f
k
(x)[ +
[f
h
(y) f
k
(y) f
h
(z) + f
k
(z)[
[y z[
|f
h
f
k
|
+ Lip(f
h
f
k
) = |f
h
f
k
|
Lip
(2.14)
k > h > h, x, y, z with y ,= z. By (2.13) and (2.14) , it follows that there exists
L > 0 such that
(2.15) Lip(f
h
) L h,
and, again by (2.14) ,
(f
h
)
h
is a Cauchy sequence in (C
0
(), | |
) .
Then, there exists f C
0
() such that f
h
f uniformly in . By (2.15), we obtain
that Lip(f) L, whence f Lip(). Taking the limit in (2.14), as k , we get
> 0 there exists h = h() N such that
[f
h
(x) f(x)[ +
[f
h
(y) f(y) f
h
(z) + f(z)[
[y z[
h > h, x, y, z with y ,= z .
This implies that lim
h
|f
h
f|
Lip
= 0.
Since the set of polynomial functions is still contained in Lip(), Lip() is innite
dimensional. Eventually, one can prove that it is not a Hilbert space, arguing like in
the previous cases, by using the parallelogram indentity.
As a consequence of Proposition 2.30 we also obtain the following result.
43
Corollary 2.34. The inclusion C
1
() Lip() is is a bilipschitz map, that is there
exists L > 0
1
L
|f|
C
1 |f|
Lip
L|f|
C
1 f C
1
() ,
and it is strict, provided R
n
is a bounded open convex set. In particular, C
1
()
is a closed subspace of (Lip(), | |
Lip
).
Proof. We prove that assertion only in the case n = 1 and = (a, b). By Proposition
2.30 and Remark 2.31 (ii), we need only to show that the inclusion is an isometry.
This follows proving that
Exercise II.10: if f C
1
([a, b]), then |f|
Lip
= |f|
C
1.
Compactness in Lip().
Theorem 2.35. Let R
n
be a bounded open set and let
T = B
Lip()
:= f Lip() : |f|
Lip
1 .
Then B
Lip()
is compact in (Lip(), | |
).
Proof. We need only to prove that T si compact in (C
0
(), | |
). We apply the
Arzel` a -Ascoli theorem (Theorem 2.10). Prove that:
(i) T is bounded in (C
0
(), | |
): it is trivial by denition.
(ii) T is closed in (C
0
(), | |
): that is, if (f
h
)
h
T with |f
h
f|
0, then
f T. Indeed,
f
h
T [f
h
(x)[ +
[f
h
(y) f
h
(z)[
[y z[
1 h, x, y, z with y ,= z .
Taking the limit, as h , in the right-hand side inequality, we obtain
[f(x)[ +
[f(y) f(z)[
[y z[
1 x, y, z with y ,= z ,
whence
f T .
(iii) T is equicontinuous in . Indeed, it is sucient to notice that, by denition,
[f(y) f(z)[ [y z[ y, z , f T .
This concludes the proof.
Remark 2.36. Note that B
C
1
()
:=
_
f C
1
() : |f|
C
1 1
_
is not compact in
(C
1
(), | |
): see Exercise II. 5 (ii). This is one of the good features enjoyed by
Lip() but not by C
1
().
Separability of (Lip(), | |
Lip
).
Theorem 2.37. Let R
n
be a bounded open set. Then (Lip(), | |
Lip
) is not
separable.
44
Proof. We need only to prove there exists an uncountable disjoint family U
: I
of open sets in (Lip(), | |
Lip
) (Proposition 2.17). We divide the proof in two steps.
1
th
step. Assume n = 1 and = (a, b) and let us prove the desired conclusion. Let
u
|
Lip
Lip(u
) 2 if ,= .
Indeed
Lip(u
) = sup
_
[u
(x) u
(x) u
(y) + u
(y)[
[x y[
: x, y (a, b), x ,= y
_
[u
() u
() u
() + u
()[
[ [
= 2
[ [
[ [
= 2 .
Therefore, the family
U
:= f Lip((a, b)) : |f u
|
Lip
< 1 I .
enjoys the desired properties.
2
nd
step. Assume is a bounded open set. Since is open, there exists an open
cube (a
1
, b
1
) (a
n
, b
n
) . Let f
: (a
1
, b
1
) Lip() be the family of
functions dened by
f
(x) := u
(x
1
) if x = (x
1
, x
2
, . . . , x
n
) , I := (a
1
, b
1
) ,
where u
is the one variable function dened in the rst step. By (2.16), we obtain
that, if ,=
Lip(f
, ) Lip(u
, (a
1
, b
1
)) 2 .
Therefore, the family
U
:= f Lip() : |f f
|
Lip
< 1 I .
enjoys the desired properties.
We have considered the class Lip() of Lipschitz continuous functions f : R
which by denition satisfy the estimate
(L) [f(x) f(y)[ C [x y[ x, y ,
for some constant C > 0. We have seen that (L) implies continuity , and more
importantly provides uniformly continuity. It turns out to be useful, for instance for
the study of PDEs, to consider also functions satisfying a variant of (L), namely
(H) [f(x) f(y)[ C [x y[
x, y ,
for some constants C, > 0.
Exercise II.11: Let R
n
be a connected open set and assume (H) holds for
some C > 0 and > 1. Then f const.
Thus the H older condition does not make sense for functions with exponents greater
than 1 on connected open sets.
Denition 2.38. Let A R
n
, a function f : A R is called Holder continuous
with exponent > 0 if it satises (H) for some constant C > 0.
45
Historical remark: ([P, Sections 6.7.1.2]) The notion of Lipschitz function oc-
curred for the rst time in the work of Lipschitz on trigonometric series in 1864 and
on ordinary dierential equations in 1876. As well H older introduced his notion in
1882 in a work of potential theory.
2.4. The space of p-integrable functions L
p
(). Let us recall the space of p-
integrable functions with respect to the n-dimensional Lebesgue measure.
Denition 2.39. Let A R
n
be a Lebesgue measurable set and p [1, ],
L
p
(A) :=
_
f : A R : f is Lebesgue measurable and |f|
L
p < +
_
where
|f|
L
p = |f|
L
p
(A)
:=
_
_
_
__
A
[f(x)[
p
dx
_
1/p
if 1 p <
inf M > 0 : [f(x)[ M a.e. x A if p =
.
The quantity |f|
L
p is called the L
p
norm of f on A.
Let us recall the following fundamental result (see [De] and [GZ, Theorem 6.24]).
Theorem 2.40 (Fisher-Riesz,1907). (L
p
(A), | |
L
p) is a B.s. if 1 p . More-
over L
2
(A) turns out to be a Hilbert space with respect to the scalar product
(f, g)
L
2 :=
_
A
f g dx f, g L
2
(A) .
As a consequence of the proof of Riesz- Fishers Theorem we have the following
useful result.
Theorem 2.41. Let R
n
be an open set, (f
h
)
h
L
p
() and f L
p
() with
1 p . Suppose that
(MC) lim
h
|f
h
f|
L
p
()
= 0 .
Then, there exist a subsequence (f
h
k
)
k
and a function g L
p
() such that
(i) f
h
k
(x) f(x) a.e. x ;
(ii) [f
h
k
(x)[ g(x) a.e. x , k.
Proof. See [De] or [GZ, Theorem 6.25].
Remark 2.42. It does not hold the implication (MC) f
h
(x) f(x) a.e. x
(see Exercise II.16).
Remark 2.43. Note that C
0
() L
p
() for each p [1, ], provided R
n
is
a bounded open set, otherwise the inclusion need not hold. Meanwhile it holds the
inclusion C
0
c
() L
p
() for each p [1, ] and any open set , where
C
0
c
() :=
_
f C
0
() : spt(f) is compact and contained in
_
and
spt(f) := closure x : f(x) ,= 0 .
Moreover recall that (C
0
(), | |
L
2) is a normed linear space, but is not a B.s. (see
Exercise II.12).
46
Historical remark: ([P, Sections 1.1.4,1.5.2,4.4.1]) Fisher and Riesz invented the
Hilbert space L
2
([a, b]) in 1907, by proving its completeness. Both authors observed
the signicance of Lebesgues integral as the basic ingredient. Subsequently, in 1909,
Riesz extended this denition to exponents 1 < p < and described how the
interval [a, b] can be replaced by any measurable set of R
n
. L
p
spaces are sometimes
called Lebesgue spaces, named after H. Lebesgue (Dunford & Schwartz 1958, III.3),
although according to Bourbaki (1987) they were rst introduced by Riesz.
Compactness in (L
p
(), | |
L
p).
In this section we are going to deal with some compactness results in L
p
spaces.
We will only state these results, without proofs which can be found in [B, Section
4.5].
Let f : R
n
R and v R
n
be given, then we dene by
v
f : R
n
R the
v-translated function of f dened by
(
v
f)(x) := f(x + v) .
Theorem 2.44 (M. Riesz- Frechet- Kolmogorov). Let T be a bounded subset in
(L
p
(R
n
), | |
L
p) with 1 p < . Suppose that lim
v0
|
v
f f|
L
p = 0 uniformly
for f T, that is
> 0 = () > 0 such that |
v
f f|
L
p < v R
n
with [v[ < ,
f T .
(EN
F
)
Then T[
:= f[
: f T is relatively compact in (L
p
(), | |
L
p), i.e. its closure
is compact in (L
p
(), | |
L
p), for each open set R
n
with nite Lebesgue measure.
From Theorem 2.44 it follows the following compactness criterion in (L
p
(), ||
L
p).
If f : R, let us denote by
f : R
n
R the function dened as
f(x) :=
_
f(x) if x
0 otherwise
.
Corollary 2.45. Let R
n
be an open set with nite measure, let T L
p
() and
let
T :=
f : f T. Assume that
(i) T is bounded in (L
p
(), | |
L
p) with 1 p < ;
(ii) lim
v0
|
v
f
f|
L
p = 0 uniformly for f T, that is,
T satises EN
F
.
Then T is relatively compact in (L
p
(), | |
L
p).
Proof. From Theorem 2.44,
T is relatively compact. Notice now that
T is relatively
sequentially compact in (L
p
(R
n
), | |
L
p) if and only if T is relatively sequentially
compact in (L
p
(), | |
L
p). Thus, the characterization of compact sets in metric
spaces (Theorem 2.5) completes the proof.
Eventually recall the following characterization of compactness in (L
p
(R
n
), | |
L
p).
Theorem 2.46. Let T L
p
(R
n
) with 1 p < . Then T is relatively compact in
(L
p
(R
n
), | |
L
p) if and only if
(i) T is bounded in (L
p
(R
n
), | |
L
p);
(ii) for each > 0 there exists r
))
< f T ;
47
(iii) lim
v0
|
v
f f|
L
p = 0 uniformly for f T.
Remark 2.47. (i) The assumption (EN
F
) is necessary in Theorem 2.44. Indeed,
consider the family T := f
h
: h N where f
h
: R R is dened as f
h
(x) :=
_
1
h
if 0 x
1
h
0 otherwise
and let := (0, 1). Then it is easy to see that |f
h
|
L
1
(R)
= 1
for each h N and T[
f
h
(x + v) dx =
_
v
0
f
h
(x) = 1 .
Thus, (EN
F
) does not hold for T.
(ii) If has not nite measure, then the conclusions of Theorem 2.44 need not hold.
Indeed , consider the family T := f
h
: h N where f
h
: R R is dened
as f
h
(x) := f(x + h) where f Lip(R) with spt(f) = [a, a], a > 0, and f not
identically vanishing. Then
(2.17) |f
h
|
L
1
(R)
= |f|
L
1
(R)
> 0 h.
Moreover T satises (EN
F
), because
[
v
f(x) f(x)[ = [f(x + v) f(x)[ L[v[
[a1,a+1]
(x) x R, v [1, 1] ,
and
|
v
f
h
f
h
|
L
1
(R)
= |
v
f f|
L
1
(R)
h
where L := Lip(f). Let := R and observe now that T = T[
is not relatively
compact in (L
1
(R), ||
L
1). Otherwise a contradiction arises by (2.17), since f
h
(x) 0
for each x R,
Historical remark: A rst compactness type-result was proved by Frechet in 1908
in the setting of l
2
. In 1931, Kolmogorov proved the rst result in this direction. It
characterizes compactness in L
p
(R
n
) for 1 < p < , in the case where all functions
are supported in a common bounded set. Condition (iii) of Theorem 2.46 is replaced
by the uniform convergence in L
p
norm of spherical means of each function in the
class to the function itself. (Clearly, our condition (ii) is automatic in this case.) Just
a year later, Tamarkin expanded this result to the case of unbounded supports by
adding condition (ii) of Theorem 2.46. In 1933, Tulajkov expanded the Kolmogorov-
Tamarkin result to the case p = 1. In the same year, and probably independently,
Riesz proved the result for 1 < p < , essentially in the form of our Theorem
2.46. Thus we feel somewhat justied in using the names Kolmogorov and Riesz in
referring to the theorem, though we are perhaps being a bit unfair to Tamarkin and
Tulajkov in doing so. In 1937, M.Frechet replaced conditions (i) and (ii) of Theorem
2.46 with a single condition (equisummability), and generalized the theorem to
arbitrary positive p.
Separability of (L
p
(), | |
L
p).
Theorem 2.48. Let R
n
be an open set. Then (L
p
(), | |
L
p) is separable if
1 p < and it is not separable if p = .
48
We need two fundamental preliminary results in order to prove Theorem 2.48: the
former is a topological result (Urysohns Lemma) and the latter concerns with the
approximation in L
p
by continuous functions.
Denition 2.49. Let (X, ) be a topological space. Then
C
0
c
(X) := f : X R : f is continuous and spt(f) is compact in (X, d)
where spt(f) := closure x X : f(x) ,= 0 .
Urysohns Lemma 2.50 (1925). Let X be a locally compact metric space, let K X
and V X be, respectively, a compact set and an open set such that K V . Then
there exists a function C
0
c
(X) such that
0 1, 1 in K and spt() V .
Proof. We skip the proof (see [GZ, Lemma 9.7]).
Theorem 2.51 (Approximation in L
p
by continuous functions ). Let R
n
be an
open set. Then C
0
c
() is dense in (L
p
(), | |
L
p), provided that 1 p < .
The proof of Theorem 2.51 also relies in two fundamental preliminary results on
the approximability of measurable functions, we only recall here.
Theorem 2.52 (Approximation by simple functions). Let (X, /) be a measure space
and let f : X [0, +] be a measurable function. Then there exists a sequence of
measurable simple functions s
h
: X [0, +) (h = 1, 2, . . . ) satisfying the proper-
ties:
(i) 0 s
1
s
2
. . . s
h
. . . f;
(ii) lim
h
s
h
(x) = f(x) x X.
In particular, if f L
1
(X, ), that is
_
X
f d < , then s
h
f in L
1
(X, ), that
is,
|f s
h
|
L
1
(X,)
:=
_
X
[f s
h
[ d 0 .
Proof. See [De] and [GZ, Theorem 5.25].
Lusins Theorem 2.53 (1912, form on locally compact metric spaces). Let be a
Radon measure on a locally compact, separable metric space X . Let f : X R be
a measurable function such that there exists a Borel set A X with
(A) < , f(x) = 0 x X A and [f(x)[ < a.e. x X .
Then, for each > 0, there exists g C
0
c
(X) such that
(x X : f(x) ,= g(x)) < .
Moreover g can be chosen such that
sup
xX
[g(x)[ sup
xX
[f(x)[ .
Proof. See [De] or [R2, Theorem 2.23].
49
Proof of Theorem 2.51. We divide the proof in two steps.
1
st
step. Let us prove that, > 0, f L
p
() there exists a measurable simple
function s : R such that
(2.18) [ x : s(x) ,= 0 [ < (in particular s L
p
() p [1, ]) ;
(2.19) |f s|
L
p < .
Firstly, assume that f 0 in . By the approximation of a nonnegative measurable
function by means of simple function (Theorem 2.52), there exists a sequence of
measurable simple functions s
h
: [0, +] (h = 1, 2, . . . ) such that
(2.20) 0 s
1
s
2
. . . s
h
. . . f ;
(2.21) lim
h
s
h
(x) = f(x) x .
From (2.20), it follows that
(2.22) s
h
L
p
() and [ x : s
h
(x) ,= 0 [ < h,
and
(2.23) [s
h
f[ 2 f in , h.
By (2.21) and (2.23), we can apply Lebesgues dominated convergence theorem, thus,
if 1 p < , it follows that
(2.24) lim
h
|s
h
f|
L
p = 0 .
Given an arbitary > 0, from (2.24), there exists h = h() N such that |s
h
f|
L
p <
. If we dene s := s
h
, then (2.18) and (2.19) follow.
The general case f : R can be proved decomposing f = f
+
f
and applying
(2.18) and (2.19), separately, to f
+
and f
.
2
nd
step. Let us show that > 0, f L
p
() g C
0
c
() such that |f g|
L
p < .
Let s be a measurable simple function satisfying (2.18) and (2.19) with
2
and
denote A := x : s(x) ,= 0. It is not restrictive to assume that |s|
> 0,
otherwise s 0 C
0
c
() concludes the proof. Applying Lusins theorem to the
function s, we obtain the existence of a function g C
0
c
() such that
(2.25) [A
,
with
(2.26) [g(x)[ |s|
x .
Note that
(2.27) |f g|
L
p |f s|
L
p + |s g|
L
p <
2
+ |s g|
L
p .
Estimate now the term |s g|
L
p:
(2.28) |s g|
p
L
p =
_
[s g[
p
dx =
_
A
[s g[
p
dx
(2.26)
2
p
|s|
p
[A
[
(2.25)
<
p
2
p
.
50
Therefore, (2.27) and (2.28) complete the proof.
Proof of Theorem 2.48. We need only to prove that
(2.29) (C
0
c
(), | |
) is separable.
Indeed, let us show that, from (2.29), it follows that (L
p
(), | |
L
p) is also separable,
provided that 1 p < .
Firstly, assume that is bounded. Given T (C
0
c
(), | |
<
2 [[
1/p
.
Recall that
Exercise II.13. |f|
L
p [[
1/p
|f|
f L
<
2
.
Therefore estimates (2.31) and (2.32) imply (2.30).
Assume now is unbounded. By a well-known result of topology, there exists a
sequence (
h
)
h
of bounded open sets such that
h
h
h+1
and
=
h=1
h
.
Moreover notice that
(2.33) C
0
c
() =
h=1
C
0
c
(
h
) .
By (2.29), there exists a set T
h
(C
0
c
(
h
), | |
h=1
T
h
and show that (2.30) still holds. From Theorem 2.51, f L
p
(), > 0 g
C
0
c
() such that (2.31) holds. Since K := spt(g) is a compact set contained in ,
there exists
h =
h(g) =
h() N such that K
h
. This implies that g C
0
c
(
h
)
and we can conclude like in the previous step.
Let us now prove (2.29).
Recall that (C
0
(K), | |
h=1
T
h
(C
0
c
(), | |
T
h
:=
h
g : g T
h
(h N) and T :=
h=1
T
h
.
Let us prove that (2.35) holds. By construction, T is countable. Thus we only have
to prove that it is dense in (C
0
c
(), | |
).
Fix > 0, f C
0
c
() and let K := spt(f) compact. There exists h
0
N such
that
K
h
0
h
0
+1
.
This implies that
f C
0
c
(
h
0
) C
0
(
h
0
) C
0
(
h
0
+1
) .
By (2.34), there exists f
1
T
h
0
+1
such that
(2.36) |f f
1
|
,
:= sup
x
[f(x) f
1
(x)[ < .
Since f 0 in
h
0
+1
h
0
, from (2.36), it also follows that
(2.37) sup
x
h
0
+1
\
h
0
[f
1
(x)[ < .
Dene now f
2
:=
h
0
+1
f
1
T
h
0
+1
, by (2.36) and (2.36), we obtain
|f f
2
|
,
= sup
x
[f(x) f
2
(x)[ = sup
x
h
0
+1
[f(x)
h
0
+1
f
1
(x)[
max
_
sup
x
h
0
[f(x) f
1
(x)[, sup
x
h
0
+1
\
h
0
[f
1
(x)[
_
< .
Thus (2.35) follows.
Eventually let us prove that (L
(), | |
L
) is not separable. We look for an
uncountable, disjoint family U
i
: i I of open sets in L
() (Proposition 2.17).
Let a and let
a
:= B(a, r
a
) where r
a
> 0 with B(a, r
a
) , Dene
U
a
:=
_
f L
() : |f
a
|
L
<
1
2
_
a I :=
Notice that:
U
a
is open in (L
(), | |
L
) a : it is trivial !
52
U
a
U
b
= if a ,= b: indeed, if, by contradiction, f U
a
U
b
, this implies
that
|
b
|
L
|
a
f|
L
+|f
b
|
L
<
1
2
+
1
2
= 1 .
On the other hand
|
b
|
L
= 1 if a ,= b ,
and then a contradiction.
I = is uncountable.
()
is strict. Moreover, for each f C
0
()
() |f|
,
= |f|
L
()
Indeed, it is trivial that
|f|
L
()
:= infM > 0 : [f(x)[ M a.e.x sup
x
[f(x)[ := |f|
,
.
To prove the reverse inequality observe that, if N is negligible set with respect
to /
n
, then
N .
Thus, by the continuity of f, if there exists M > 0 such that
[f(x)[ < M a.e. x [f(x)[ M x .
In particular, from (), C
0
() turns out to be closed in (L
(), | |
L
()
).
The dual space of L
p
().
Let us recall the notion of dual space for a n.v.s.
Denition 2.55. Let (E, | |) be a n.v.s. The dual space E
, | |
E
) is a B.s.
Notation. If f E
E
in place of f(x) and we
will say that ,
E
E
is the scalar product in the duality E
, | |
E
) are topologically isomorphic.
When dim
R
E = , (E, | |
E
) and (E
, | |
E
) need not be topologically isomorphic
(see Exercise II.14 (iii)). However, we prove that some topological properties of
(E
, | |
E
), as, for instance, the separability still hold for (E, | |
E
).
Theorem 2.57. (E, | |
E
) is separable if (E
, | |
E
) is separable.
Before the proof of Theorem 2.57 we need a useful density criterion for subspaces
in n.v.s., that is a corollary of the Hahn-Banach theorem-second geometric form.
Proposition 2.58 (Density criterion for subspaces). Let (E, ||
E
) be a n.v.s.Assume
that M E is a subspace that is not dense in (E, | |
E
) and let x
0
E M. Then
there exists f E
such that
f, x
E
E
= 0 x M and f, x
0
E
= 1 .
Proof of Proposition 2.58. From the Hahn-Banach theorem-second geometric form
(see [De] or [B, Theorem 1.7], see also [GZ, Theorems 8.18 and 8.19]) there exists
g E
E
=
separates the sets M and x
0
in strict sense, that is,
(2.38) g, x
E
E
< < g, x
0
E
x M .
Since M is still a subspace, by (2.38), it follows that
g, x
E
E
< R,
whence
(2.39) g, x
E
E
= 0 x M .
Thus, if we dene the functional f E
by
f :=
1
g, x
0
E
g ,
we have the desired conclusion.
Proof of Theorem 2.57. Let
T := f
h
: h N (E
, | |
E
)
be a dense. For each h there is an element x
h
E with
|x
h
| = 1 and [f
h
(x
h
)[
1
2
|f
h
|
E
.
54
Let
D := span
Q
x
h
: h N and D := span
R
x
h
: h N
that is, respectively, the sets of all nite linear combinations of elements x
h
: h N
with rational and real coecients. Then
D is countable, D is a subspace of E and,
by construction,
D (D, | |) is dense .
In order to conclude the proof, we need only to show that
D (E, | |) is dense .
By contradiction, if D is not dense, let x
0
E D. Then, from Proposition 2.58,
there exists f E
such that
f, x
E
E
= 0 x D and f, x
0
E
= 1 .
Since T is dense, there is a subsequence (f
h
k
)
k
for which
lim
k
|f
h
k
f|
E
= 0 .
However, since |x
h
k
| = 1,
|f
h
k
f|
E
[f
h
k
(x
h
k
) f(x
h
k
)[ = [f
h
k
(x
h
k
)[
1
2
|f
h
k
|
E
k N.
Thus
|f
h
k
|
E
0 as k ,
which implies that
f 0 ,
contradicting the fact that f(x
0
) = 1. Thus
D = E .
, | |
E
) yields a better understanding of
the topological properties of (E, | |
E
), as we will see in Chapter III.
The characterization of the dual space (E
, | |
E
) is undoubtedly a very dicult
problem, given a n.v.s (E, | |
E
). In the case of L
p
spaces this problem was solved
by F. Riesz and it turns out to be a fundamental result in Functional Analysis.
Riesz Representation Theorem 2.59 (1909). Let 1 p < and denote
p
:=
_
p
p 1
if 1 < p <
if p = 1
(conjugate exponent of p).
Then the mapping T : L
p
() (L
p
())
, dened by
T(u), f
(L
p
())
L
p
()
:=
_
u f dx f L
p
() ,
is an isometric isomorphism and we will mean this feature by means of the identi-
cation
L
p
() (L
p
())
if 1 p < .
55
Proof. We divide the proof in three steps.
1
st
step. Let us prove that T is an isometry, that is,
(2.40) |T(u)|
(L
p
())
= |u|
L
p
()
u L
p
() .
By Holders inequality, it is immediate the inequality
(2.41) |T(u)|
(L
p
())
|u|
L
p
()
u L
p
() .
Show the opposite inequality. Firstly assume that 1 < p < , which implies that
1 < p
< . If |u|
L
p
()
= 0, then u = 0 a.e. in and the desired inequality is
clear. Assume 0 < |u|
L
p
()
< , then we can assume that 0 < [u(x)[ < a.e.
x . Dene
f
u
(x) := [u(x)[
p
2
u(x) a.e. x .
Observe that f
u
L
p
(), since
[f
u
(x)[
p
= [u(x)[
p
a.e. x |f
u
|
L
p
()
= |u|
1/p1
L
p
()
.
Therefore
(2.42) T(u), f
u
(L
p
())
L
p
()
=
_
u[u[
p
2
u dx = |u|
p
L
p
()
.
From (2.41) and (2.42), it follows that
|u|
p
L
p
()
= T(u), f
u
(L
p
())
L
p
()
|T(u)|
(L
p
())
|f
u
|
L
p
()
= |T(u)|
(L
p
())
|u|
1/p1
L
p
()
.
This implies
(2.43) |T(u)|
(L
p
())
|u|
L
p
()
u L
p
() .
Thus (2.41) and (2.43) yield (2.40), if 1 < p < . Finally, for the case p = 1, so
p
()
. Then the set
E
M
:= x : [u(x)[ > M /
n
and [E
M
[ > 0 .
Since the measure space (, /
n
, /
n
) is -nite, there is a set F /
n
such
that
0 < [E
M
F[ < .
Set
f
u
(x) :=
1
[E
M
F[
sign(u(x))
E
M
F
(x) a.e. x .
Then
|f
u
|
L
1
()
=
_
[f
u
(x)[ dx = 1
and
(2.43 bis)
T(u), f
u
(L
1
())
L
1
()
=
1
[E
M
F[
_
E
M
F
[u[ dx M M (0, |u|
L
()
) .
From (2.41) and (2.43 bis), it follows that
M T(u), f
u
(L
1
())
L
1
()
|T(u)|
(L
1
())
|f
u
|
L
1
()
= |T(u)|
(L
1
())
M (0, |u|
L
()
) ,
56
whence
(2.44 bis) |T(u)|
(L
1
())
|u|
L
()
u L
() .
From (2.41) and (2.44 bis), the identity (2.40) follows when p = 1.
2
nd
step. Suppose rst that [[ < and let us prove that T is onto, that is,
(L
p
())
u L
p
() such that
T(u) = T(u), f
(L
p
())
L
p
()
= (f) f L
p
() .
Because of [[ < ,
E
L
p
() whenever E / := /
n
, p [1, ). Dene
the set function : / R
(E) := (
E
) E /.
Show that
(2.44) is a -nite, signed measure;
(2.45) << /
n
on /.
Indeed [(E)[ < for each E /, thus is -nite. Suppose (E
h
)
h
/ is a
disjoint sequence, we have to prove that (
h=1
E
h
) =
h=1
(E
h
). Set E :=
h=1
E
h
.
Then for any positive integer m
(E)
m
h=1
(E
h
)
(
E
)
m
h=1
(
E
h
)
(
E
)
_
m
h=1
E
h
_
E
m
h=1
E
h
_
h=m+1
E
h
_
||
(L
p
())
|
h=m+1
E
h
|
L
p
()
= ||
(L
p
())
h=m+1
E
h
1
p
and
h=m+1
E
h
h=m+1
[E
h
[ 0 as m since [E[ < . Thus
(E) =
h=1
(E
h
)
and since the same result holds for any rearrangement of the sequence (E
h
)
h
, the
series converges asbolutely and (2.44) follows. Moreover, since
[(E)[ ||
(L
p
())
[E[
1
p
E /,
(2.45) follows,too.
Warning: Note that if p = in the previous estimate, then (2.45) need not hold!
By the Radon-Nikodym theorem for signed measures (Theorem 1.33), there exists
a /-measurable function u : R with u
+
or u
L
1
() such that
(2.46) (
E
) = (E) =
_
E
u dx E /.
It actually holds that u L
1
(). Indeed, let
E
+
u
:= x : u(x) 0 and E
u
:= x : u(x) 0 .
57
From (2.46), we obtain
0
_
dx =
_
E
u
u dx = (E
u
) < .
Thus u
L
1
(). From the linearity of and the integral it is clear that
(2.47) (s) =
_
u s dx
for each measurable simple function s : R. To conclude we need only to prove
that
(2.48) u L
p
() p [1, ) .
Indeed, for each f L
p
(), by the approximation by simple functions (Theorem 2.52),
there exists a sequence s
h
: R (h = 1, 2, . . . ) of measurable simple functions
such that
(2.49) s
h
f in L
p
() .
From (2.48), (2.49) and H olders inequality, it follows that
(2.50)
u(s
h
f) dx
|u|
L
p
()
|f s
h
|
L
p
()
0 .
The conditions (2.47), (2.50) and the continuity of yield
(f) = lim
h
(s
h
) = lim
h
_
u s
h
dx =
_
u f dx f L
p
() .
In particular, it holds that there exists u L
p
() such that
T(u) = .
which completes the proof.
Let us prove (2.48).
In case p = 1 suppose M > 0 and set
E
M
: x : u(x) > M .
Then
M [E
M
[
_
E
M
u dx = (
E
M
) ||
(L
p
())
[E
M
[ .
Thus [E
M
[ = 0 if M > ||
(L
p
())
, from which it follows that
0 u
+
(x) ||
(L
p
())
a.e. x |u
+
|
L
()
||
(L
p
())
.
Similarly
|u
|
L
()
||
(L
p
())
and hence u = u
+
u
() = L
1
().
In case 1 < p < , by the approximation by simple functions, let (s
h
) be a
sequence of measurable simple functions such that
(2.51) 0 s
1
s
2
. . . s
h
. . . [u[ in ,
and
(2.52) lim
h
s
h
(x) = [u(x)[ x .
58
Let us now prove the following crucial estimate
(2.53) |s
h
|
L
p
()
||
(L
p
())
h
Set
u
h
(x) := [s
h
(x)[
p
1
sign(u(x)) if x .
Then (u
h
) is still a sequence of simple functions and
|s
h
|
p
L
p
()
=
_
[s
h
[
p
dx
(2.51)
s
p
1
h
[u[ dx =
_
u
h
u dx
(2.47)
= (u
h
)
||
(L
p
())
|u
h
|
L
p
()
= ||
(L
p
())
__
[s
h
[
(p
1) p
dx
_1
p
= ||
(L
p
())
__
[s
h
[
p
dx
_1
p
= ||
(L
p
())
|s
h
|
p
p
L
p
()
.
If |s
h
|
L
p
()
= 0, then (2.53) is trivial. If |s
h
|
L
p
()
> 0, the previous inequality
implies (2.53) by dividing for |s
h
|
p
p
L
p
()
and noticing that p
(1
1
p
) = 1.
From (2.52), (2.53) and Fatous Lemma we have
|u|
p
L
p
()
=
_
[u[
p
dx liminf
h
_
[s
h
[
p
dx = liminf
h
|s
h
|
p
L
p
()
||
p
(L
p
())
< .
Therefore (2.48) also follows when 1 < p < .
3
rd
step. Suppose now that [[ = and let us prove that T is still onto.
Let (
h
)
h
be an increasing sequence of bounded open sets such that
=
h=1
h
.
Let us agree to identify L
p
(
h
) and L
p
(
h
)(h = 1, 2, . . . ) with the subspaces of
L
p
() and L
p
it follows that
(2.54) (L
p
(
h
))
and ||
(L
p
(
h
))
||
(L
p
())
h
From step 2, for each h there exists a unique u
h
L
p
(
h
) such that
(2.55) |u
h
|
L
p
()
= ||
(L
p
(
h
))
and
(2.56) (f) =
_
h
u
h
f dx f L
p
(
h
)
Notice that since L
p
(
h
) L
p
(
h+1
), by the uniqueness, u
h+1
= u
h
a.e. in
h
. Thus,
it is well-dened the function u : R
u(x) := u
h
(x) if x
h
.
By (2.54), (2.55) and the monotone convergence theorem
|u|
L
p
()
= lim
h
|u|
L
p
(
h
)
= lim
h
|u
h
|
L
p
()
||
(L
p
())
< ,
so u L
p
() . Moreover, if f L
p
(), then, by the dominated convergence theorem,
f
h
f in L
p
() ,
59
and hence, by the continuity of and (2.56),
(f) = lim
h
(f
h
) = lim
h
_
h
u f dx =
_
u f dx .
This completes the proof.
Remark 2.60. The Riesz representation theorem extends to a measure space (X, /, ).
More precisely, the identication
L
p
(X, ) (L
p
(X, ))
still holds if
1 < p < for a general measure (see [GZ, Theorem 6.43] or [F, Theorem
6.15] );
p = 1 provided is -nite (see [R2, Theorem 6.16]).
The identication may fail in the other cases: see [F, section 6.2]
Historical remark:[P, Section 2.2.7] The identity (L
p
(a, b))
= L
p
(a, b) with
1 < p < was proved by F. Riesz in 1909. The limit case p = 1 is due to Steinhuas
in 1919.
Support of a L
p
function.
It is well-known that, given a function f : R
n
R, its support is the set
(S) spt(f) := closure x R
n
: f(x) ,= 0 = x : f(x) ,= 0 .
This denition is not suitable for a function f L
p
(R
n
). Indeed we would like that
this notion satises the following property:
f
1
= f
2
a.e. in R
n
spt(f
1
) = spt(f
2
) , except for a negligible set.
But this is not the case. Indeed
Example: Let f
1
:=
Q
: R R and f
2
0. Then it is clear that
f
1
= f
2
a.e. in R
but
spt(f
1
) = Q = R and spt(f
2
) = .
Proposition 2.61 (Essential support of a function). Let f : R
n
R. Denote
/
f
:= R
n
: open set and f = 0 a.e. in
and let
A
f
:=
A
f
.
Then A
f
is an open set and
f = 0 a.e. in A
f
.
The closed set
(ES) spt
e
(f) := R
n
A
f
is called the essential support of f in R
n
.
Remark 2.62. (i) From denition (ES), it follows that, if f
1
= f
2
a.e. in R
n
,
then spt
e
(f
1
) = spt
e
(f
2
).
60
(ii) Denitions (S) and (ES) agree when the function is continuous. More precisely
Exercise II.23: If f : R
n
R is continuous, then
R
n
A
f
= x R
n
: f(x) ,= 0 .
Proof of Proposition 2.61. It is trivial that A
f
is open. Let us prove that
(2.57) f(x) = 0 a.e. x A
f
.
Since R
n
is a separable metric space, then it satises the second axiom of countability
(Theorem 2.15). Thus, there exists a countable family of open sets | = U
i
: i N
such that each open set of R
n
is countable union of elements in |. For each /
f
,
suppose that
=
iJ
U
i
for a suitable set of indices J
N and let J :=
A
f
J
(R
n
) ;
(Mo2) spt(
h
) B(0, 1/h) ;
(Mo3)
_
R
n
h
dx = 1 ;
(Mo4)
h
(x) 0 x R
n
.
Example of molliers: It is quite simple constructing a sequence of molliers,
starting from a given non vanishing function : R
n
R satisfying
C
c
(R
n
) , spt() B(0, 1), 0 .
For instance, let
(x) :=
_
_
_
exp
_
1
[x[
2
1
_
if [x[ < 1
0 if [x[ 1
.
61
Then it is easy to see that C
c
(R
n
). Moreover we yield a sequence of molliers
by dening
h
(x) := c h
n
(hx) x R
n
, h N
and
c :=
__
R
n
dx
_
1
.
Notation: If A, B R
n
, A B denotes the set
A B := a b : a A, b B
Exercise: Prove that
(i) if A is compact and B is closed, then is A + B is closed;
(ii) if A and B are compact so is A + B
Proposition 2.64 (Denition and rst molliers properties). Let f L
1
loc
(R
n
) and
let (
h
)
h
be a sequence of molliers. Dene, for given h N and x R
n
,
f
h
(x) = (
h
f)(x) :=
_
R
n
h
(x y) f(y) dy x R
n
.
Then
(i) the function f
h
: R
n
R is well dened;
(ii) f
h
(x) = (
h
f)(x) = (f
h
)(x) for all x R
n
and h N;
(iii) f
h
C
0
(R
n
) for each h.
The function f
h
is called h
th
- mollier of f.
Proof. Let us x h and, for sake of simplicity, denote
h
.
(i) Notice that, by (Mo2) and (Mo4) ,
spt() B(0, 1/h) .
Then
_
R
n
[f(y) (x y)[ dy =
_
B(x,1/h)
[f(y) (x y)[ dy
sup
R
n
_
B(x,1/h)
[f(y)[ dy < .
Thus, for xed x R
n
, the function
g
x
(y) := (x y) f(y) y R
n
is integrable on R
n
, so it is well-dened the integral
R
_
R
n
g
x
(y) dy =
_
R
n
(x y) f(y) dy = ( f)(x) x R
n
.
(ii) By the change of variables formula
(f )(x) =
_
R
n
f(x y) (y) dy
(z=xy)
=
_
R
n
f(z) (x z) dz = ( f)(x) .
(iii) Given x R
n
and x
r
x, let us prove that
(2.58) ( f)(x
r
) ( f)(x) .
62
Notice that
(2.59) ( f)(x
r
) ( f)(x) =
_
R
n
((x
r
y) (x y)) f(y) dy r N.
Since the sequence (x
r
)
r
is bounded in R
n
, there exists a compact set K R
n
such
that
B(x
r
, 1/h)) = x
r
B(0, 1/h) K, B(x, 1/h) K r N.
In particular
(2.60) (x
r
y) (x y) = 0 y / K, r N.
Because Lip(R
n
), by (2.60), there exists a constant L > 0 such that
[(x
r
y) (x y)[ L
K
(y) [x
r
x[ y R
n
, r N.
Thus we obtain the estimate
(2.61) [(x
r
y) (x y)[ [f(y)[ L
K
(y) [f(y)[ [x
r
x[ y R
n
, r N.
From (2.59), (2.61) and the dominated convergence theorem, (2.58) follows.
Remark 2.65. The symbol denotes the convolution product between two functions
dened on the whole R
n
. Notice also that the conclusions of Proposition 2.64 still
holds if f L
1
loc
(R
n
) and
h
C
0
(R
n
) satisfying (Mo2). Actually, it is possible
to dene the convolution product between two functions g L
p
(R
n
) with 1 p
and f L
1
(R
n
)
(g f)(x) :=
_
R
n
g(x y) f(y) dy
and it holds that (see [GZ, Theorem 6.51])
(g f) L
p
(R
n
) and |g f|
L
p
(R
n
)
|g|
L
p
(R
n
)
|f|
L
1
(R
n
)
.
Theorem 2.66 (Friedrichs-Sobolev, approximation by convolution in L
p
). Let f
L
1
loc
(R
n
) and (
h
)
h
be a sequence of molliers. Then
(i) f
h
C
(R
n
) for each h N.
(ii) |f
h
|
L
p
(R
n
)
|f|
L
p
(R
n
)
for each h N, f L
p
(R
n
), for every p [1, ].
(iii) spt(f
h
) spt
e
(f) + B(0, 1/h) for each h N.
(iv) If f L
p
(R
n
) with 1 p , then f
h
C
(R
n
) L
p
(R
n
) for each
h N, and f
h
f as h , in L
p
(R
n
), provided that 1 p < .
This results yields two important results. The former is the following.
Theorem 2.67 (Fundamental lemma of the calculus of variations). Let R
n
be
an open set and let f L
1
loc
(). Assume that
()
_
f dx = 0 C
c
() .
Then f = 0 a.e. in .
Proof. It is sucient to prove that
(2.62)
_
K
[f[ dx = 0 for each compact set K .
63
Indeed, by (2.62), it follows that
f = 0 a.e. in K, for each compact set K .
This implies the desired conclusion. Let us prove (2.62).
Given a compact set K , dene g : R
n
R by
g(x) :=
_
_
_
f(x)
[f(x)[
if x K and f(x) ,= 0
0 otherwise
.
Then g L
1
(R
n
) and spt
e
(g) K . Let
g
h
:= g
h
.
By Theorem 2.66 (iii), there exists
h =
h(K) N such that
spt(g
h
) spt
e
(g) + B(0, 1/h) K + B(0, 1/h)
for each h >
h. Thus, by Theorem 2.66 (i) and (ii),
(2.63) g
h
C
c
() if h >
h and [g
h
(x)[ |g|
L
(R
n
)
= 1 x R
n
, h N.
From () we obtain
_
f g
h
dx = 0 h >
h.
On the other hand, from Theorems 2.66 (iv) and 2.41, we can suppose, up to a
subsequence, g
h
g a.e. in R
n
. Thus,
0 =
_
f g
h
dx 0 =
_
f g dx =
_
K
[f[ dx .
functions in L
p
). Let R
n
be an open set.
Then C
c
() is dense in (L
p
(), | |
L
p), provided that 1 p < .
Proof of Theorem 2.68. Given f L
p
(), dene
f : R
n
R by
f(x) :=
_
f(x) if x
0 if x R
n
.
Note that
f L
p
(R
n
).
Let (
h
)
h
be an increasing sequence of bounded open set such that
=
h=1
h
h
h
h+1
h,
and dene
g
h
(x) :=
h
(x)
f(x), and f
h,r
(x) := (
r
g
h
)(x) if x R
n
, h, r N.
From Theorem 2.66 (iii), it follows that
(2.64) spt(f
h,r
) B(0, 1/r) +
h
h, r N.
Moreover, given h N, there exists r
h
= r(h) N such that
(2.65) r
h
h and B(0, 1/r
h
) +
h
.
64
Dene
f
h
(x) := (
r
h
g
h
)(x) x R
n
, h N,
and, for sake of simplicity, assume that r
h
= h.
Then, by Theorem 2.66 (i), (ii) (2.64), and (2.65), f
h
C
c
() and
|f
h
f|
L
p
()
= |f
h
f|
L
p
(R
n
)
|
h
g
h
h
f|
L
p
(R
n
)
+|
h
f
f|
L
p
(R
n
)
= |
h
(g
h
f)|
L
p
(R
n
)
+|
h
f
f|
L
p
(R
n
)
|g
h
f|
L
p
(R
n
)
+|
h
f
f|
L
p
(R
n
)
h.
(2.66)
From Theorem 2.66 (iv),
f
f in L
p
(R
n
) ,
and, by the dominated convergence theorem,
g
h
f in L
p
(R
n
) .
Then, by (2.66), we conclude the proof.
Before the proof of Theorem 2.66, let us recall Fubini- Tonellis integration theorem
(see [GZ, Corollary 6.47]) and we need a basic convergence result for molliers of
continuous functions.
Theorem 2.69 (G. Fubini, L. Tonelli,1907). Let F : R
2n
= R
n
R
n
[0, ] be a
measurable function (with respect to /
2n
). Then
(i) the function
R
n
y F(x, y)
is measurable (with respect to /
n
) for /
n
-a.e. x R
n
.
(ii) The function
R
n
x
_
R
n
F(x, y) dy
is measurable (with respect to /
n
).
(iii)
_
R
2n
F(x, y) dxdy =
_
R
n
dx
__
R
n
F(x, y) dy
_
=
_
R
n
dy
__
R
n
F(x, y) dx
_
in the sense that either three expressions are innite or all are nite and equal.
Lemma 2.70. Let f C
0
(R
n
). Then
h
f f uniformly on compact sets of R
n
.
Proof. Firstly, let us recall we already proved that
h
f C
0
(R
n
). Let K R
n
be
a xed compact set and let K
_
R
n
f(x y)
h
(y) dy f(x)
_
R
n
f(x y)
h
(y) dy
_
R
n
f(x)
h
(y) dy
_
R
n
(f(x y) f(x))
h
(y) dy
_
B(0,1/h)
[f(x y) f(x)[
h
(y) dy
_
B(0,1/h)
h
(y) dy
= .
This concludes the proof.
Proof of Theorem 2.66. (i) We will prove the result only when n = 1.
It is sucient to show that, if C
m
c
(R) then f C
m
(R) for each integer
m 1.
Let m = 1 and assume that, if C
1
c
(R), then
(2.68) f C
1
(R) and ( f)
(x) = (
f)(x) x R.
Then, let us prove, by induction on m, we have the desired conclusion. Indeed,
suppose C
m
c
(R) with m 2. Then both and
C
m1
c
(R). From induction
hypothesis, f C
m1
(R) and (2.68) holds as well. Since
C
m1
c
(R) even
f C
m1
(R), whence, by (2.68), ( f)
C
m1
(R), too. Thus, f C
m
(R)
and this completes the proof of statement (i).
Let us now prove (2.68). Let 0 < [t[ 1 and x x R, then
( f)(x + t) ( f)(x)
t
(
f)(x)
=
_
R
(x y + t) (x y) t
(x y)
t
f(y) dy
(2.69)
We are going to take the limit in (2.69), as t 0: the main problem is the convergence
of the term in right-hand side.
Now observe that, being
bounded on R,
(2.70)
[(z +t)(z) t
(z)[ =
_
z+t
z
(
(s)
(z)) ds
(s)
(x y) = 0 if y / x + B(0, 1) K
0
, t B(0, 1) ,
66
because x y + t / K
0
for each t B(0, 1). Denote K := x + B(0, 1) K
0
, from
(2.70), it follows that
[(x y + t) (x y) t
(x y)[ [f(y)[
[t[ ([t[)
K
(y) [f(y)[ y R, t [1, 1] .
(2.72)
By (2.69), (2.71), (2.72) and the dominated convergence theorem, (2.68) follows.
(ii) For sake of simplicity, denote
h
. Notice that, being f : R
n
R
continuous (Proposition 2.64 (iii)), it is also measurable.
First, assume 1 p < . Then
(2.73) |f |
p
L
p
(R
n
)
=
_
R
n
[(f )(x)[
p
dx =
_
R
n
dx
_
R
n
f(x y) (y) dy
p
Recall now the following
Exercise II.24: Let h : R
n
R and : [0, +) be Lebesgue measurable
functions and suppose that
_
R
n
dx = 1. Prove that for each p [1, +)
__
R
n
[h[ dx
_
p
_
R
n
[h[
p
dx.
By (2.73), Exercise II.24 and Fubini- Tonellis theorem, it follows that
|f |
p
L
p
(R
n
)
_
R
n
dx
__
R
n
[f(x y)[
p
(y) dy
_
=
_
R
n
(y) dy
__
R
n
[f(x y)[
p
dx
_
=
__
R
n
(y) dy
_ __
R
n
[f(x)[
p
dx
_
= |f|
p
L
p
(R
n
)
.
Now let p = . By the denition of convolution,
[(f )(x)[ = [( f)(x)[ =
_
R
n
(x y) f(y) dy
|f|
L
(R
n
)
_
R
n
(x y) dy
= |f|
L
(R
n
)
x R
n
.
Thus, the proof is accomplished.
(iii) Set
h
and x x R
n
. Then, since
(x y) f(y) = 0 for a.e. y /
_
x B(0, 1/h)
_
spt
e
(f) ,
( f)(x) :=
_
R
n
(x y) f(y) dy =
_
(
xB(0,1/h)
)
spt
e
(f)
(x y) f(y) dy .
Note that
_
x B(0, 1/h)
_
spt
e
(f) ,= x B(0, 1/h) + spt
e
(f) ,
Therefore,
( f)(x) = 0 whenever x / B(0, 1/h) + spt
e
(f) ,
whence, since f is continuous (Proposition 2.64 (iii)),
spt( f) B(0, 1/h) + spt
e
(f) .
67
(iv) It is immediate, from statements (i) and (ii), that
h
f C
(R
n
) L
p
(R
n
) for
each p [1, ]. Let us show that
(2.74) lim
h
|
h
f f|
L
p
(R
n
)
= 0 if 1 p < .
Since C
0
c
(R
n
) is dense in (L
p
(R
n
), | |
L
p), provided that 1 p < (Theorem
2.51), for each > 0 there exists f
1
C
0
c
(R
n
) such that
(2.75) |f f
1
|
L
p
(R
n
)
< .
From Lemma 2.70 and the previous statement (iii),
h
f
1
f
1
uniformly on compact sets of R
n
and
spt(
h
f
1
) B(0, 1/h) + spt(f
1
) B(0, 1) + spt(f
1
) h.
Thus, by the dominated convergence theorem,
(2.76) lim
h
|
h
f
1
f
1
|
L
p
(R
n
)
= 0 .
Now observe that
|
h
f f|
L
p
(R
n
)
|f f
1
|
L
p
(R
n
)
+| f
1
h
f
1
|
L
p
(R
n
)
+ |
h
f
1
h
f|
L
p
(R
n
)
h.
(2.77)
From previous statement (ii) and (2.75),
|
h
f
1
h
f|
L
p
(R
n
)
= |
h
(f
1
f)|
L
p
(R
n
)
|f f
1
|
L
p
(R
n
)
< h.
Thus, by (2.77), we obtain
|
h
f f|
L
p
(R
n
)
2 + | f
1
h
f
1
|
L
p
(R
n
)
h,
whence, by (2.76),
limsup
h
|
h
f f|
L
p
(R
n
)
2 > 0
which implies (2.74).
Historical remark: Molliers were introduced by K. Friedrichs in 1944, which
are, according to P. Lax, a watershed in the modern theory of PDEs. However, S.
Sobolev had used molliers in his epoch making 1938 paper [So] (the paper containing
the proof of the Sobolev embedding theorem), as Friedrichs himself acknowledged in
later papers.
2.5. Recalls on Banach- Steinhaus, open mapping an closed graph theo-
rems. In this section we recall three important results concerning continuous linear
operators between Banach spaces which are considered cornerstones of Functional
Analysis. The proofs of these results rely on the completeness of involved normed
vector spaces. This property is fundamental since the statements of these results are
no longer true if the spaces are just assumed to be normed spaces.
68
Denition 2.71. Let (E, | |
E
) and (F, | |
F
) be n.v.s and let T : E F be a
continuous linear operator. The norm of T is the nonnegative real number
|T| := sup |T(x)|
F
: x B
E
The set of all continuous operators from E to F will be denoted by /(E, F) (or,
sometimes, by B(E, F)).
Exercise II.26: (/(E, F), | |) is a n.v.s.
Banach-Steinhaus Theorem 2.72 (or Uniform Boundedness Principle, 1927). Let
(E, | |
E
) and (F, | |
F
) be, respectively, a Banach space and a normed vector space
and let T
i
iI
/(E, F). Suppose that for each x E there exists a constant
c(x) (0, +) such that
sup
iI
|T
i
(x)|
F
c(x) .
Then there exists L > 0 such that
|T
i
(x)|
F
L|x|
E
for each x E and i I .
Proof. See [De] or [GZ, Theorem 8.20].
Two important consequences of the Banach-Steinhaus theorem are the following.
Corollary 2.73. Let (E, | |
E
) and (F, | |
F
) be, respectively, a Banach space and
a normed vector space and let (T
h
)
h
/(E, F) be a sequence. Suppose there exists a
map T : E F such that
T(x) = lim
h
T
h
(x) x E ,
that is, lim
h
|T
h
(x) T(x)|
F
= 0 for each x E. Then
(i) sup
h
|T
h
| < ;
(ii) T /(E, F);
(iii) |T| liminf
h
|T
h
|.
Proof. See [De] or [GZ, Corollary 8.21].
Remark 2.74. The conclusions of corollary 2.73 may fail if (E, ||
E
) is not supposed
to be a B.s. For instance
Example: Let (R
, | |
l
2) be the n.v.s. dened in Exercise II.14. Let T
h
: R
such that
T(x) = lim
h
T
h
(x) x R
,
where y = T(x) is dened by
y(n) := nx(n) n = 1, 2, . . .
It is easy to see that T is linear but not bounded. Hence T is not continuous.
69
Corollary 2.75. Let (G, | |
G
) be a n.v.s. and let B G. Suppose that
(WB) f G
f(B) = f, b
G
G
: b B is bounded in R
Then B is bounded in (G, | |
G
).
Remark 2.76. A set B (G, | |
G
) verifying (WB) is said to be weakly bounded.
Obviously a bounded set B is also weakly bounded. Indeed let
R := sup|b|
G
: b B < .
Then, f G
[f, b
G
G
[ |f|
G
|b|
G
R|f|
G
= a ,
that is
f(B) [a, a] .
Corollary 2.75 also states the opposite implication. This implies that in a n.v.s
boundedness weak boundedness .
Note that this coincidence is straighforward in a nite dimensional n.v.s. In fact, if
(G, | |
G
) = (R
n
, [ [), the set e
1
, e
2
, . . . , e
n
denotes the standard basis of R
n
, then
B is bounded in R
n
i M > 0 such that
[(e
i
, x)
R
n[ = [x
i
[ M x = (x
1
, . . . , x
n
) B, i = 1, 2, . . . , n.
Proof of Corollary 2.75. Apply the Banach-Steinhaus theorem with (E, | |
E
) =
(G
, | |
G
), (F, | |
F
) = (R, [ [), I = B and T
b
: G
R dened by
T
b
(f) := f, b
G
G
f G
, b B.
By (WB),
sup
bB
[T
b
(f)[ < f G
.
From the Banach-Steinhaus theorem there exists L > 0 such that
[T
b
(f)[ = [f, b
G
G
[ L|f|
G
f G
, b B.
Thus, from Exercise II.35,
|b|
G
= sup
fG
[f, b
G
G
[
|f|
G
L b B.
|x|
2
|x|
1
c |x|
2
x E .
Proof. It suces to apply the bounded inverse theorem with | |
E
= | |
2
, F = E,
| |
F
= | |
1
and T = Id.
Denition 2.80. Let (E, | |
E
) and (F, | |
F
) be n.v.s. It is called product norm on
E F, written | |
EF
, one of the following norms:
(i) [[(x, y)[[
EF
= [[x[[
E
+[[y[[
F
;
(ii) [[(x, y)[[
EF
=
_
[[x[[
E
2
+[[y[[
F
2
;
(iii) [[(x, y)[[
EF
= max[[x[[
E
, [[y[[
F
;
if (x, y) E F.
Exercise II.7:(i) Show that | |
EF
isa norm on E F;
(ii) (E F, | |
EF
) is a B.s., provided that (E, | |
E
) and (F, | |
F
) are B.s.
Closed Graph Theorem 2.81 (Banach, 1932). Let (E, | |
E
) and (F, | |
F
) be
Banach spaces. If T : E F is linear and its graph G(T) := (x, T(x)) : x E is
closed in (E F, | |
EF
), then T /(E, F).
Proof. See [De] or [GZ, Theorem 8.25].
Remark 2.82. Obviously, the graph of any continuous function f : E F is closed
in (E F, | |
EF
). The opposite need not hold when f : E F is nonlinear.
Example: Let f : R R be the function f(x) :=
_
1
x
if x ,= 0
0 if x = 0
. Then it is
easy to see that G
f
:= (x, f(x)) : x R is closed, even if f is not continuous.
Application: well-posedness of Cauchy problem for linear ODEs.
71
Exercise II.31 Consider the following initial value problem (or Cauchy problem) for
the linear ordinary dierential equation (ode) of order n:
(P(f,w))
_
_
u
(n)
+
n1
i=0
a
i
(t)u
(i)
= f
u(t
0
) = w
1
u
(t
0
) = w
2
...................
u
(n1)
(t
0
) = w
n
,
where a
i
C
0
([a, b]) (i = 1, . . . , n 1), w = (w
1
, . . . , w
n
) R
n
and t
0
[a, b] are
given. From a well-known result, for each f C
0
([a, b]), w = (w
1
, . . . , w
n
) R
n
there
exists a unique solution u = T(f, w) C
n
([a, b]) of (P(f,w)).
Question: Does the solution u = T(f, w) C
n
([a, b]) depend on the initial data
(f, w) with continuity?
Dene the operator T : E := C
0
([a, b]) R
n
F := C
n
([a, b]), dened by
T(f, w) = u, where (f, w) are the given data problem and u denotes the (unique)
solution of (P(f,w)). Prove that:
(i) T is a linear operator: it is trivial !
(ii) T : (E, [[ [[
E
) (F, [[ [[
F
) is continuous, where [[ [[
E
and [[ [[
F
denote the
norms respectively dened by
|(f, w)|
E
:= |f|
+[w[
R
n if (f, w) C
0
([a, b]) R
n
,
|v|
F
:=
n
i=0
|v
(i)
|
if v C
n
([a, b]) .
(Hint: Prove that G(T) is closed in E F, then apply the closed graph theorem.)
Historical remark: ([D, Section V]) Between 1900 and 1910, there was a sudden
crystallization of all the ideas and methods which had been slowly accumulated during
the 19
th
century. This was essentially due to the publication of four fundamental
papers:
Fredholms 1900 paper on integral equations;
Lebesgues thesis of 1902 on integration;
Hilberts paper of 1906 on spectral theory;
Frechets thesis of 1906 on metric spaces.
Between 1904 and 1906, Hilbert published six papers on integral equations, subse-
quently to Fredholms results and later collected in a book published in 1924. The
fourth paper, published in 1906, is, following [D], a masterpiece and one of the best
paper he ever wrote. By the depth and novelty of its ideas, it is a turning point in
the history of Functional Analysis, and indeed deserves to be considered the rst pa-
per published in that discipline. More precisely, Hilbert here introduced the innite
dimensional Hilbert space l
2
(the notation is due to Banach, according to [P, Section
1.5.1]) as a domain of linear, bilinear and quadratic form. Moreover he studied the
topological properties of the unit closed ball B
l
2 as well as he introduced the concepts
of strong and weak convergence in l
2
as we will see in the next section. Hilbert did
not introduce the modern notion of Hilbert space as a structure: it was introduced
72
later by one of his student E. Schmidt in 1908 and, in 1906-1907, by two other young
mathematicians, E. Fisher and F. Riesz, who in the so-called Riesz-Fisher theorem
established an unsuspected link between Hilbert spaces and the theory of Lebesgue
integration. Both authors observed the signicance of Lebesgues integral as the basic
ingredient for the study of the space L
2
(I) over a compact interval of R ([P, Section
1.5.1]).
([D, Section VI.5]) In 1932 S. Banach published the book [Ba] containing a com-
prehensive account of all results known at that time in the theory of normed spaces.
In particular it contains the uniform bounded principle, the open mapping and closed
graph theorems, which have became three of the most powerful tools in all applica-
tions of Functional Analysis. Moreover a large part of this book is devoted to the
concept of weak convergence of which we will postpone the study to the next section.
Analysts all over world began to realize the power of the new methods and to ap-
ply them to a great varierty of problems. Banachs terminology and notations were
universally adopted, complete normed linear spaces became known as Banach spaces
(the suggestion is due to Frechet, see [P, Section 1.1.1]) and soon their theory was
considered compulsory part in most curricula of graduate students.
3. Complements on Banach and Hilbert spaces ([B], [GZ], [Ro]).
Motivation: The main goal of this chapter is the introduction, in (an innite di-
mensional) given n.v.s. (E, | |
E
), of a smaller topology on E than one induced from
the norm | |
E
in order to gain compactness of the of unit closed ball with respect to
this topology. We will call this topology weak topology . Then we will introduce on
a dual space (E
, | |
E
) a third topology, weaker than either the one induced by the
norm | |
E
and the weak topology, named weak* topology. Eventually we will give
an application of the weak topology in the calculus of variations.
3.1. Weak topology on a normed vector space and compactness.
Question: Why introducing a weak topology?
Dene by
s
the topology induced from the norm, that is,
Notation:
s
:= A E : A is open with respect to | |
In order to distinguish the weak topology from the topology induced by the norm,
we will refer to
s
as the strong topology.
Recall that, given a n.v.s. (E, | |) , by Riesz theorem (Theorem 2.7),
B
E
:= x E : |x| 1 is compact with respect to
s
dim
R
E < .
Thus, when dim
R
E = , we are looking for a weaker topology on E than
s
, which we
will denote by
w
, such that (B
E
,
w
) is a sequentially compact space. More precisely
Problem: Find out a Hausdor topology
w
s
on E such that, for any sequence
(x
h
)
h
B
E
there exist a subsequence (x
h
k
)
k
, extracted from (x
h
)
h
, and x B
E
with
x
h
k
w
x as k .
73
Idea (Hilbert, 1906): introducing the weakly convergent sequences by means of coor-
dinatewise convergence.
Consider
E = l
2
:= x : N R :
n=1
[x(n)[
2
<
equipped with the norm
|x|
l
2 :=
n=1
[x(n)[
2
.
Then, it is well-known that l
2
is a Hilbert space with respect to the scalar product
(x, y)
l
2 :=
n=1
x(n) y(n) if x, y l
2
(see Exercise II.14). Thus, by the Riesz representation theorem for Hilbert spaces
(see [De] or [GZ, Theorem 8.42]),
(l
2
)
l
2
,
that is, the mapping T : l
2
(l
2
)
dened by
T(x)(y) := (x, y)
l
2 y l
2
is a linear isometry of l
2
onto (l
2
)
.
We are going to dene the weak topology
w
on l
2
in such a way, given
(x
h
)
h
and x in B
l
2
(WC) x
h
w
x as h for given n N, x
h
(n) x(n) as h ,
that is, for each n N, the sequence of n
th
-component of (x
h
)
h
, (x
h
(n))
h
, converges
to the n
th
-component of x, x(n), in R. This idea goes back to Hilbert, who introduced
it in a fundamental paper about spectral theory in 1906.
Notice that if x
h
x (strongly) in l
2
, that is
lim
h
|x
h
x|
l
2 = 0 ,
then x
h
w
x as h in the above sense, beacuse of the inequality
[x
h
(n) x(n)[ |x
h
x|
l
2 h, n N.
Viceversa, if x
h
w
x as h , then the strong convergence need not hold.
Example: Let
e
h
: N N e
h
(n) :=
_
1 if h = n
0 if h ,= n
.
Then it is easy to see that e
h
w
0 as h , but , |e
h
|
l
2 = 1 for each h. Thus, it
cannot (strongly) converge to 0 in l
2
.
Question: How can (WC) be understood in B
E
, if (E, | |) is a B.s. that need not
have a (Hilbert) coordinate system? For instance, if E = L
p
() with p ,= 2 ?
74
Answer: It means that, given
(x
h
)
h
and x in B
E
x
h
w
x as h
f, x
h
E
f, x
E
E
as h , f E
.
(WC2)
Let us prove that (WC2) is equivalent to (WC) when E = l
2
.
Let
R
:=
_
x l
2
: N N with x(n) = 0 n N
_
l
2
(l
2
)
.
Then it is easy to see that
for given n N, x
h
(n) x(n) as h (f, x
h
)
l
2 (f, x
h
)
l
2 f R
.
On the other hand, since R
is dense in (l
2
, | |
l
2), it also follows that
(f, x
h
)
l
2 (f, x
h
)
l
2 f l
2
whence (WC2) follows when E = l
2
.
We are going to introduce a notion of weak convergence on B
E
, by using this
strategy adopted for E = l
2
.
Denition 3.1. The weak topology on a n.v.s (E, | |), denoted
w
or (E, E
), is
the smallest topology on E with respect to which each f E
is continuous.
That such a weak topology exists may be seen by observing that the intersection
of any family of topologies for E is a topology for E. In particlular, denote
T (E) :=
_
: topology on E, f
1
(U), f E
) :=
T (E)
is continuous.
(ii) (E, (E, E
) with
x
i
U
i
i = 1, 2.
(iii) Let (x
h
)
h
E. Then
x
h
(E,E
)
x as h
f, x
h
E
f, x
E
E
as h , f E
.
Proof. We skip the proof (see [B, Chap. 3] or [Ro, Chap. 10]).
Remark 3.3. The notation (E, E
are stressed
75
Remark 3.4. If dim
R
E < , then
(E, E
) =
s
.
Indeed, for instance, in the case E = R
n
, by denition of (E, E
)
s
.
If dim
R
E = , then the strong topology is strictly ner than the weak one, that is,
(E, E
)
s
.
For instance, it can be proved that the sets
S
E
:= x E : |x|
E
= 1 and U := x E : |x|
E
< 1
are, respectively, never closed and open in the topology (E, E
). This property
is more delicate (see [B, Remark 20, Chap. 3] or [P, Sect. 3.3.2.7]).
Recall that each closed set in the weak topology (E, E
) .
Proof. The implication () is trivial because of (E, E
)
s
.
() Let us show that E C is weakly open. Let x
0
E C, then, from the Hahn-
Banach theorem-second geometric form (see [De] or [B, Theorem I.7], see also [GZ,
Theorems 8.18 and 8.19]) there exists f E
E
=
separates the sets C and x
0
in strict sense, that is,
f, x
0
E
< < f, x
E
E
x C .
Dene
U := x E : , f, x
E
E
< ,
then, by dention of weak topology, U is open in (E, E
) and
x
0
U E C
and this concludes the proof.
76
Remark 3.7. Recall that in a topological space (X, )
closure sequential closure,
that is, if C X is closed in the topology and (x
h
)
h
C with x
h
x, as h ,
then x C. Instead of
compactness ,= sequential compactness .
We already know they agree in a metric space (Theorem 2.5). Nevetheless in the
topological space (E, (E, E
)
_
f
h
g dx
_
f g dx g L
p
() .
Proof. It follows at once by the Riesz reprentation theorem, that is,
L
p
() (L
p
())
,
and Theorem 3.27 (iii).
Comparison between the weak and strong convergence.
Notation: If (E, | |) is a n.v.s, and (x
h
)
h
E, we will write
x
h
x lim
h
|x
h
x| = 0 ,
x
h
x x
h
(E,E
)
x .
Theorem 3.9. Suppose that (E, | |) is a n.v.s and let (x
h
)
h
E and x E. Then
the following assertions hold:
(i) if x
h
x, then x
h
x.
(ii) If x
h
x, then (x
h
)
h
is bounded in (E, | |), that is, the sequence (|x
h
|)
h
is
bounded and
(*) |x| liminf
h
|x
h
| .
(iii) Let (f
h
)
h
E
. Assume that
f
h
f in E
and x
h
x in E .
Then
f
h
, x
h
E
f, x
h
E
.
77
Proof. (i) By Theorem 3.27 (iii),
x
h
x f, x
h
E
f, x
E
E
f E
.
Since f : E R is continuous, it turns out that
f, x
h
E
f, x
E
E
f E
.
(ii) As a consequence of the Uniform Boundedness Principle (Corollary 2.75), given
B E
B is (strongly) bounded in (E, |, |)
B is weakly bounded, that is, sup [f, x
E
E
[ : x B < f E
.
Let us show that, if
B := x
h
: h N ,
then B is weakly bounded.
Since (f, x
h
E
)
h
is a convergent sequence in R, we obtain the desired conclusion.
Show now (*). Note that, for given f B
E
,
f, x
h
E
|f|
E
|x
h
| |x
h
| h N.
Taking the limit, as h in the previous inequality,
(3.1) f, x
E
E
liminf
h
|x
h
| f B
E
.
Recall now that this fundamental identity
Exercise II.35: Let (E, | |) be a normed vector space. Prove that, for xed x E,
it holds that
|x| = supf, x
E
E
: f B
E
= max< f, x >
E
E
: f B
E
.
Thus, from Exercise II.35, taking the supremum in (3.1) over all f B
E
, (*)
follows.
(iii) It is Exercise III.1.
(Notice that
[f
h
, x
h
E
f, x
E
E
[
=[ (f
h
, x
h
E
f, x
h
E
) + (f, x
h
E
f, x
E
E
) [
[f
h
f, x
h
E
[ + [f, x
h
E
f, x
E
E
[
|f
h
f|
E
sup
h
|x
h
| + [f, x
h
E
f, x
E
E
[ .
Taking the limit in the previuos inequality, as h , by statement (ii), we obtain
the desired conclusion.)
Reexive spaces.
Let (E, | |) be a n.v.s., we will denote by E
, that is,
E
= (E
.
We will show that E is isometrically isomorphic to a subspace of E
.
Introduce the evaluation function between E and E
dened by
78
(3.2) J
E
(x)(f) := f, x
E
E
if f E
.
Since
(3.3) [J
E
(x)(f)[ |f|
E
|x| f E
,
then J
E
(x) is continuous, that is, J
E
(x) E
, in fact
|J
E
(x)|
E
|x|
E
.
It is readily veried that the mapping
J
E
: E E
) with |J
E
| 1 .
Theorem 3.10. The mapping J
E
: (E, | |) (E
, | |
E
) is an isometric isomor-
phism of E onto J
E
(E).
Proof. In view of Exercise II.35 we have
|J
E
(x)|
E
= sup< f, x >
E
E
: f B
E
= |x| x E .
The mapping J
E
is called the natural imbedding of E in E
.
Denition 3.11. A n.v.s. (E, | |) is said to be reexive if
J
E
(E) = E
,
in which case E is isometrically isomorphic to E
.
Remark 3.12. Since (E
, | |
E
) is a B.s. (see Exercise II.18) we see that every
reexive n.v.s. is in fact a B.s.
Proposition 3.13. A H.s. (H, (, ))
H
is reexive.
Proof. Recall, from the Riesz representation theorem for H.s.(see [De] or [GZ, Theo-
rem 8.42]),
H
H ,
that is,the mapping T : H H
dened by
(3.4) T(u), v
H
H
:= (u, v)
H
if v H ,
is a linear isometry of H onto H
.
It is easy to see that (, )
H
is a scalar product on H
and (H
, (, )
H
) turns out to
be a H.s and
|f|
H
=
_
(f, f)
H
f H
.
Let us now show that the natural imbedding J
H
: H H
is onto.
79
Referring again to the Riesz representation theorem for H.s., for each H
there
exists a unique f H
such that
(g) = (f, g)
H
= (T
1
(f), T
1
(g)) g H
.
Again, there exists a unique x H such that T(x) = f. Thus,
(g) = (f, g)
H
= (T
1
(f), T
1
(g))
H
= (x, T
1
(g))
H
(3.4)
= g, x
H
H
= J
H
(x)(g) g H
,
that is
J
H
(x) = .
Let us now list some elementary properties of reexive spaces we state without
proof (see [B, Chap. III] or [GZ, Theorem 8.33])
Theorem 3.14 (Stability properties of the reexive spaces). Let (E, | |
E
) be a
reexive B.s. Then
(i) If M E is a closed vector subspace then (M, , | |
E
) is also reexive.
(ii) (E
, | |
E
) is reexive i (E, | |
E
) is reexive.
We are now in order to state the weak compactness result for reexive spaces.
Theorem 3.15 (Weak sequential compactness in reexive spaces). If (E, | |) is
a reexive space, then (B
E
, (E, E
, ||
E
) is separable and, by Theorem 2.57, (E
, ||
E
)
is separable, too.
Let
T = f
i
: i N (E
, | |
E
) dense ,
and let (x
h
)
h
B
E
be a given sequence. Consider now, roughly speaking, the se-
quence of i
th
-components of the vectors x
h
, that is, the sequence
(f
i
, x
h
E
)
h
R i xed .
Then (f
i
, x
h
E
)
h
is bounded, since
(3.5) [f
i
, x
h
E
[ |f
i
|
E
|x
h
|
E
|f
i
|
E
h.
By the Cantor diagonal process, we can construct a subsequence (x
h
k
)
k
of (x
h
)
h
such
that
(3.6) lim
k
f
i
, x
h
k
E
= l
i
R i N,
and, by (3.5),
(3.7) [l
i
[ |f
i
|
E
i N.
Let us now show that there exists a functional : E
R such that
(3.8) lim
k
f, x
h
k
E
= (f) R f E
,
80
(3.9) [(f)[ |f|
E
f E
,
(3.10) is linear .
Notice that, from (3.9), (3.10), it follows that E
E
= (f) = lim
k
f, x
h
k
E
f E
,
whence x
h
k
x weakly in E and the proof is complete.
Let us prove (3.8). It is sucient to prove that, for given f E
, (f, x
h
k
E
)
k
is
a Cauchy sequence, that is,
(3.11) > 0 r = r() N : [f, x
h
r
x
h
s
E
[ < s, r > r .
Fix > 0, then by the density of T in E
, there exists
i =
E
[ =[f f
i
, x
h
r
x
h
s
E
+ f
i
, x
h
r
x
h
s
E
[
[f f
i
, x
h
r
x
h
s
E
[ + [f
i
, x
h
r
x
h
s
E
[
|f f
i
|
E
|x
h
s
x
h
r
|
E
+, [f
i
, x
h
r
x
h
s
E
[
2
+ [f
i
, x
h
r
x
h
s
E
[ .
By (3.6), (f
i
, x
h
k
E
)
k
R is a Cauchy sequence, then (3.11) follows.
Let us prove (3.9). Since
[f, x
h
k
E
[ |f|
E
|x
h
k
|
E
|f|
E
,
by (3.8),
[(f)[ = lim
k
[f, x
h
k
E
[ |f|
E
f E
.
Condition (3.10) immediately follows.
Step 2. Now suppose that E is a reexive B.s., not necessarily separable.
Let
M := span
R
x
h
: h N
s
,
that is, the closure, in the strong toplogy
s
of E, of the subspace of E spanned by
x
h
: h N. Then (M, | |
E
) is a separable B.s. (see Exercise II.38) and, by the
stability properties for reexive spaces (Theorem 3.14 (i)),it is also reexive. Thus,
from step 1, there are an element x B
M
= B
E
M and a subsequence (x
h
k
)
k
of
(x
h
)
h
such that
x
h
k
x in (M, M
), as k ,
that is
g, x
M
M
= lim
k
g, x
h
k
M
g M
.
81
For any f E
, let f
M
: M R denote the restriction to M. It is easy to see that
f
M
M
. Thus
f, x
E
E
= f
M
, x
M
M
= lim
k
f
M
, x
h
k
M
= lim
k
f, x
h
k
E
f E
,
whence x
h
k
x weakly in E.
Let C E be a closed, convex and bounded set and let (x
h
)
h
C. Then, because
C is bounded, there is a positive constant r
0
> 0 such that
|x
h
| r
0
h N.
Denote
y
h
:=
1
r
0
x
h
h N,
then (y
h
)
h
B
E
. Thus, there exist a subsequence (y
h
k
)
k
and an element y
B
E
such that
y
h
k
y
weakly in E, as k .
Of course, it also holds that
x
h
k
x
weakly in E, as k .
with x
= r
0
y
). Thus, C is sequentially
closed in (E, E
), whence x
), [[ [[
, (f
h
)
h
E, f E. Prove that:
(i) if f
h
f in E, then
(a) M > 0 such that [f
h
(x)[ M x [0, 1], h N;
(b) f
h
(x) f(x), x [0, 1];
(ii) f
h
(x) := x
h
, x [0, 1] does not weakly converge in E;
(iii) (C
0
([0, 1]), [[ [[
) is not reexive.
Hint: (i) a) It is a general property. In fact (Theorem 3.9 (ii)), if f
h
f in E, then,
[f
h
(x)[ sup
h
|f
h
|
x
(f) := f(x) if f C
0
([0, 1]) (Dirac delta at x) .
Then,
x
(C
0
([0, 1]))
and, if f
h
f in E,
f
h
(x) =
x
(f
h
)
x
(f) = f(x) x [0, 1] .
(ii) By contradiction, if there exists f C
0
([0, 1]) such that f
h
f in E, then, by
previous statement (ii),
(3.12) f(x) = lim
h
x
h
=
_
0 if x [0, 1)
1 if x = 1
/ C
0
([0, 1]) .
(iii) By contradiction, if it is reexive, the previous sequence (f
h
)
h
B
E
admits
a weakly convergent subsequence (f
h
k
)
k
to a function f C
0
([0, 1]). Arguing as
82
before we obtain that f turns out to be the function in (3.12) and then we reach a
contradiction.
2. (C
1
(
), [[ [[
C
1) is not reexive: see Exercise III.3.
3. (Lip(), [[ [[
Lip
) is not reexive: see Exercise III.4.
4.
Theorem 3.16. Let R
n
be an open set. Then (L
p
(), [[ [[
L
p) is reexive if
1 < p < ; it is not reexive if p = 1, .
Proof. Firstly, prove that L
p
() is reexive if 1 < p < . Let
J
L
p : L
p
() (L
p
())
(L
p
)
= J
L
p(u)(f) := f(u) = f, u
(L
p
)
L
p f (L
p
())
.
Let us show that
(3.13) J
L
p(L
p
()) = (L
p
())
if 1 < p < .
Recall that, by the Riesz representation theorem for L
p
spaces (Theorem 2.59),
L
p
() (L
p
())
if 1 p < with p
:=
_
p
p 1
if 1 < p <
if p = 1
.
This means that the mapping
T
p
: L
p
() (L
p
())
,
dened by
T
p
(u), v
(L
p
())
L
p
()
:=
_
u v dx v L
p
() ,
is a linear isometry from L
p
() onto (L
p
())
. In particular, since (p
= p, it holds
that
T
p
(u), v
(L
p
())
L
p
()
=
_
u v dx
= T
p
(v), u
(L
p
())
L
p
()
u L
p
(), v L
p
() .
(3.14)
Given (L
p
())
() R, dened by
(u) := , T
p
(u)
(L
p
)
(L
p
)
if u L
p
() .
It is easy to see that
(L
p
())
(L
p
())
= T
p
(v), u
(L
p
())
L
p
()
u L
p
() .
From (3.14) and (3.15), we have
(3.16) , T
p
(u)
(L
p
())
(L
p
())
= T
p
(u), v
(L
p
())
L
p
()
u L
p
() .
Beacause of
T
p
(L
p
()) = (L
p
())
,
83
let f = T
p
(u) with u L
p
(L
p
())
= f, v
(L
p
())
L
p
()
= J
L
p(v)(f) f (L
p
())
J
L
p(v) = .
Thus (3.13) follows.
Let us now show that L
p
() for p = 1, are not reexive. Indeed, it can be proved
that (see Exercise III.6)
B
L
1
()
is not weakly sequentially compact
if n = 1 and = (0, 1), then L
1
() is not reexive. On the other hand, by the
properties of reexive spaces (Theorem 3.14 (ii)), E is reexive i E
is reexive.
Thus, since
L
() = (L
1
())
,
L
f
h
k
g dx
_
f g dx as k , g L
p
() .
3.3. Weak topology and convexity: an application to the Calculus of Vari-
ations.
Problem: Extend Weierstrass theorem on the existence of the minimum for a
function : A E (, +], where (E, | |) is an innite dimensional n.v.s.
More precisely, which assumptions on A and do they assure that
min
A
?
Denition 3.18. Let E be a v.s. and let
: A E (, +] .
(i) is called a convex function if
(i
1
) A is convex ;
(i
2
) (t x + (1 t) y) t (x) + (1 t) (y) x, y A, t [0, 1] .
(ii) is called lower semicontinuous (lsc) with respect to a toplogy on E if,
for each t R the t-sublevel of is closed in A with repsect to the relative
topology, that is,the set
x A : (x) t = A F
with F E closed set in the topology .
Remark 3.19. Each continuous function : (E, ) R is also lsc
84
Lemma 3.20. Let (E, | |) be a n.v.s. Assume that A E is a closed set and
: A (, +] is lsc (with respect to the strong topology
s
). Assume also that
is convex. Then is still lsc with respect to the weak topology (E, E
).
Proof. Taking into account the semicontinuity denition and Lemma 3.6, it is su-
cient to prove that, for each t R, the set
x A : (x) t is convex .
This feature immediately follows from the convexity of .
We are in order to state a generalization of Weierstrass theorem to reexive spaces.
Generalized Weierstrass Theorem in reexive spaces 3.21. Let (E, | |) be a
reexive normed space and let : A E (, +]. Suppose that
(i) A is closed and is convex ;
(ii) A is bounded or A is unbounded but there exists
lim
xA,x+
(x) = + (coercivity condition) ;
(iii) is lsc (with respect to
s
) .
Then there exists
min
A
.
Let us introduce two interesting consequences of the generalized Weierstrass theo-
rem in reexive spaces.
Corollary 3.22. In a reexive n.v.s. (E, | |), each f E
E
.
Thus, there exists x
0
B
E
such that
f(x
0
) = (x
0
) (x) = f(x) x B
E
,
whence, there exists x
0
B
E
such that
f(x
0
) f(x) x B
E
.
Remark 3.23. The conclusion of Corollary 3.22 need not hold when E is not reexive
(see Exercise III.7).
Corollary 3.24 (Projection on a convex set in reexive spaces). Let (E, | |) be a
reexive n.v.s. and let A E be closed and convex. Then for each x
0
E there
exists
min
xA
|x x
0
| .
85
Remark 3.25. Corollary 3.24 geometrically means that the distance function from
A attains its minimum, that is
d(x
0
, A) := inf
xA
|x x
0
| = | x x
0
|
for some x A and the point x can be understood like a projecton of x on A. When E
is a H.s. this projection is unique and gives rise to the classical orthogonal projection
in H.s. (see [De] or [GZ, Theorem 8.40]). This conclusion need not hold when E is
not reexive (see Exercise III.12).
Proof of Corollary 3.24. By means of the translation
T : E E T(x) = x x
0
,
it is not restrictive to assume that
x
0
= 0 and (x) := |x| .
Then, it is easy to see that assumptions (i), (ii) and (iii) of the generalized Weierstrass
theorem hold, and this concludes the proof.
Proof of Theorem 3.21. If +, the proof is accomplished. Then, assume that
a A such that (a) < +
and denote
A := x A : (x) (a) .
Let us now show that
(3.17) min
A
,
Notice that, if (3.17) holds, then we trivially obtain
min
A
= min
) ,
and
(3.19) : (
A, (E, E
)) R is lsc .
Indeed, since is (strongly) lsc and convex and A is closed,
A is convex and (strongly)
closed. Thus, it is also weakly closed (Lemma 3.6). On the other hand, from assump-
tion (ii), it turns out to be that
Exercise:
A is bounded.
Therefore (3.18) follows from the weak sequential compactness in reexive spaces
(Theorem 3.30). In order to prove (3.19), apply Lemma 3.20.
Show now that, from (3.18) and (3.19), (3.17) follows, by means of the minimizing
sequence argument due to Weierstrass. Let (x
h
)
A a minimizing sequence for the
function :
A R, that is, a sequence satisfying
(3.20) lim
h
(x
h
) = inf
A
= m [, +) .
86
By (3.18), there exist a subsequence (x
h
k
)
k
and an element x
0
A such that
(3.21) x
h
k
x
0
in (E, E
), as k , .
Prove that
(3.22) (x
0
) m = lim
h
(x
h
) ,
whence (3.17) will follow. By contradiction, assume
(x
0
) > m = lim
h
(x
h
) ,
and let t R such that
(x
0
) > t > lim
h
(x
h
) .
Without loss of generality, by the limit denition, we can assume that
(x
h
)
h
C :=
_
x
A : (x) t
_
.
Since C is weakly closed, by (3.20), it follows that x
0
C and then we reach a
contradiction, beacuse
t R and (x
0
) t < (x
0
) .
Therefore (3.22) follows.
3.4. Weak* topology on a dual space and compactness. So far we have two
topologies on E
:
the usual (strong) topology
s
associated to the norm | |
E
;
the weal topology (E
, E
), obtained by performing on E
, the so-called
weak* topology. More precisely, for every x E, let J
E
(x) :=
x
: E
R denote
the linear continuous functional dened by
x
(f) := f, x
E
E
f E
where J
E
: E E
.
Denition 3.26. The weak* topology on a dual space (E
, | |
E
), denoted (E
, E),
is the smallest topology on E
. In particular, denote
T
(E
) :=
_
: topology on E
,
1
(U), J
E
(E) and U R open set
_
.
and
(E
, E) :=
T
(E
, E) is the smallest
topology on E
, (E
, E)) is a topological Hausdor space, that is, for each pair of distinct
points f
1
and f
2
in E
, E) with
f
i
U
i
i = 1, 2.
(iii) Let (f
h
)
h
E
. Then
f
h
(E
,E)
f as h
f
h
, x
E
E
f, x
E
E
as h , x E .
(iv) If E is reexive, then (E
, E
) = (E
, E).
Proof. We skip the proof (see [B, Chap. 3] or [Ro, Chap. 10]).
Remark 3.28. It can be proved (see, for instance, [B, Chap. 3, Remark 11]) that,
if E is not reexive, i.e., by denition, the natural imbedding J
E
: E E
is not
surjective, then
(E
, E) (E
, E
)
s
.
Moreover
(E
: |f|
E
1
is compact in the weak* topology (E
, E).
Proof. We skip the proof (see [B, Chap. 3]).
The weak* sequential compactness of B
E
does not hold for a general v.n.s ( see
Exercise III. 5 (iii)). However it holds that
Theorem 3.30 (Weak* sequential compactness in separable spaces). If (E, | |) is
a separable space, then (B
E
, (E
()
is weakly* sequentially
compact, that is, for each sequence (u
h
)
h
B
L
()
there exist a subsequence (u
h
k
)
k
and u B
L
()
such that
_
u
h
k
v dx
_
u v dx v L
1
() .
Historical remark: ([D, Chap. VIII, Sect. 1]) In his thesis Frechet (1906) had
already noticed that convergence in a metric space could not always correspond to
some classical types of convergence for functions. There was thus an obvious need
for a generalization of the concept of metric spaces, but none proved adequate for
Functional Analysis until Hausdor, in 1914, created General Topology as we un-
derstand it now, based on the concept of neighborhood. Ever since Hilbert, weak
88
convergence of sequences had became a central theme, rst in Hilbert spaces, then
with F. Riesz and E. Helly in some types of normed spaces. Eventually it was at the
center of the celebrated Banachs book [Ba] and the results he obtained concerning
that notions can be considered as his deepest results. Theorem 3.29, in this form,
was rst proved by N. Bourbaki in 1938, and independtly little later by L. Alaoglu
in 1940.
4. An introduction to the Sobolev space and an application to
Poissons equation ([B], [E])
Motivation: The study of a functional Hilbert space, called Sobolev space, which
turns out often to be the proper setting in which to apply ideas of functional analysis
to glean information concerning PDEs. In particular, we are going to introduce
this space in order to get existence and uniqueness for the so-called boundary-value
problem of Poissons equation.
4.1. Reference example : general electrostatic problem (Gauss, 1839). Let
R
3
a bounded open set of the space with regular boundary = (for instance,
the unit open ball of R
3
).
Denote by
u = u(x
1
, x
2
, x
3
) : R
and
= (x
1
, x
2
, x
3
) : R
respectively, the electrostatic potential inside and the electric charge density in .
The denition of electrostatic potential, combined with the dierential form of
Gauss law , provides a relationship between the potential u and the charge density
u(x) := div(u)(x) =
2
u
x
2
1
(x) +
2
u
x
2
2
(x) +
2
u
x
2
3
(x)
=
(x)
0
x = (x
1
, x
2
, x
3
) ,
where
0
is a positive constant which denotes the vacuum permittivity. This relation-
ship is a form of the so-called Poissons equation. When 0 it is called Laplaces
equation and a function u, satisfying Laplaces equation, is called harmonic in .
Now assume that the electric potential u on the boundary is prescribed by a
function
g : R
and assume that also the charge density is given.
Then the general electrostatic problem amounts to nd out a function u : R
satisfying
(EP)
_
u =
0
in
u = g on
,
Problem (EP) is called boundary-value or Dirichlet problem for Poissons equation.
89
Well-posedeness of (EP) and its classical solutions.
The informal notion of well-posedeness for the problem (EP) captures many desir-
able features of what it means to solve a PDE.
We say that problem (EP) is well-posed if
(WP1) the problem has (at least) a solution;
(WP2) this solution is unique;
(WP3) the solution depends continuously on the data given in the problem.
Condition (WP3) is particularly important for problem arising from physical ap-
plications: we would prefer that our (unique) solution changes only a little when the
conditions specifying the problem change a little.
Now clearly it would be desirable to solve the problem in such a way that (WP1)-
(WP3) hold. But
Question: what do we mean by a solution of problem (EP)?
Assume that the boundary datum g C
0
() and the density charge C
0
(),
then we call classical solution of (EP) a function
u C
2
() C
0
())
satisfying
_
_
_
u(x) =
(x)
0
x
u(x) = g(x) x
.
We call classical such a solution because it is certainly the most obvious notion of
solution.
So by solving a PDE in the classical sense we mean, if possible, to write down a
formula for classical solution satisfying (WP1)-(WP3) above, or at least to show such
a solution exists, and to deduce various of its properties.
Problem (EP) can be generalized to any dimension. Indeed, let R
n
be a
bounded open set, then introduce the Laplace operator
: C
2
() C
0
(), u(x) :=
n
i=1
2
u
x
2
i
(x) x .
Thus, the boundary value problem for Poissons equation turns out to be
(P)
_
u = f in
u = g on
,
for given f : R, g : R and = .
Example: Let n = 1, = (a, b). Let g : = a, b R and f C
0
([a, b]).
Then, it is easy to see that the unique classical solution u of (P) is given by
u(x) = c
1
x + F(x) + c
2
x [a, b] ,
where
F(x) :=
_
x
a
(t x) f(t) dt x [a, b] ,
90
and c
i
(i = 1, 2) are the constants
c
1
:=
g(b) g(a) F(b)
b a
, c
2
:= g(a) c
1
a .
Question: How to attack problem (P) when n 2 ?
Sometimes, it is possible to give explicit or more-or-less explicit formulas for the
solution of (P). This is in general a dicult matter, and can be done only when
has simple geometry (see, for instance, [E, Section 2.2]).
4.2. Energy functional and classical Dirichlets principle.
We are going to use now a strategy for solving (P) through the Functional Analysis.
The main idea goes back to Gauss and consists in demostrating that a solution of the
boundary-value problem (P) can be characterized as the minimizer of an appropriate
functional. For this, let R
n
be a bounded open set, f C
0
() and g C
0
(),
we dene Dirichlets energy (or integral) functional
I : C
1
() R, I(v) :=
1
2
_
[v[
2
dx
_
f v dx v C
1
() ,
and let us introduce the class of admissible functions
/ :=
_
v C
2
() : v = g on
_
Denition 4.1. Let R
n
be a bounded open set. We say that = is of class
C
m
if for each x
0
there exist r > 0 and a C
m
function : R
n1
R such that-
upon relabeling and reorienting the coordinates axes if necessary- we have
B(x
0
, r) = x = (x
1
, . . . , x
n
) B(x
0
, r) : x
n
> (x
1
, . . . , x
n1
)
Theorem 4.2 (Dirichlets principle). Let R
n
be a bounded open set with =
of class C
1
and let f C
0
() and g C
0
(). Assume u / solves the boundary-
value problem (P). Then
(DP) I(u) = min
vA
I(v) .
Viceversa, if u / satises (DP), then u solves the boundary-value problem (P).
Remark 4.3. In other words, if u /, the PDE u = f is equivalent to the
statement that u minimizes the energy functional I.
The proof of Theorem 4.2 relies in several steps, which we collect in dierent pre-
liminary results because they are worthy and important in the general framework of
the PDEs.
Lemma 4.4. The functional I : / R is strictly convex, that is,
I(t w + (1 t) u) < t I(w) + (1 t) I(u) t (0, 1) w, u / with w ,= u .
In particular, if u / satises (DP), it is unique.
Proof of Lemma 4.4. Assume that / , = , otherwise the proof is accomplished. First,
notice that / is convex and, since the functional
C
2
() v
_
f v dx
91
is linear, we need only to prove that the functional
/ v
_
[v[
2
dx
is strictly convex, that is,
(4.1)
_
[t w + (1 t) u[
2
dx < t
_
[w[
2
dx + (1 t)
_
[u[
2
dx
t (0, 1), w, u / with w ,= u.
Because the function R
n
[[
2
is strictly convex, we infer that
(4.2) [t
1
+ (1 t)
2
[
2
t [
1
[
2
+ (1 t)[
2
[
2
for each
1
,
2
R
n
, t (0, 1) and, if for some
1
,
2
R
n
,
t (0, 1)
(4.3) [
1
+ (1
t)
2
[
2
=
t [
1
[
2
+ (1
t), [
2
[
2
1
=
2
.
From (4.2), by choosing
1
= w,
2
= u and integrating over , it follows that
(4.4)
_
[t w + (1 t) u[
2
dx t
_
[w[
2
dx + (1 t)
_
[u[
2
dx
t [0, 1], w, u /.
Let us now show that, if for some
t (0, 1) and u, w /, it holds that
(4.5)
_
t w + (1
t) u[
2
dx =
t
_
[w[
2
dx + (1
t)
_
[u[
2
dx ,
then
(4.6) w = v in .
Because of (4.4), (4.1) will follow and the proof is accomplished.
By (4.2) and (4.5), it follows that
[
t w + (1
t) u[
2
=
t[w[
2
+ (1
t) [u[
2
a.e. in ,
and, by (4.3), since w and u are continuous,
(4.7) w = u in .
Since w = u = g on , by (4.7), (4.6) follows.
If w / is a dierent minimizer of I, that is, w ,= u and I(w) = min
A
I, then by
the strict convexity of I, t w + (1 t) u / and
I(t w + (1 t) u) < t I(w) + (1 t) I(u) = min
A
I t (0, 1)
and we reach a contradiction
(0) = 0.
Then
i(0) = min
R
i .
Proof of Lemma 4.5. It is an elementary result of standard calculus.
92
Theorem 4.6 (Gauss-Green). Assume that R
n
is a bounded open set with
= of class C
1
and let w C
1
(; R
n
). Then
(GG)
_
div(w) dx =
_
(w, )
R
n ds ,
where ds and : S
n1
respectively denote the n 1-dimensional area element
and the outward normal to and
div(w)(x) :=
n
i=1
w
i
x
i
(x) x .
Proof. See [E, Appendix C].
Remark 4.7. Notice that it can be proved that (GG) still holds, even though =
is not regular, but provided that w C
1
c
(; R
n
) and in this case
_
div(w) dx = 0 .
.
Lemma 4.8 ( Euler-Lagrange equation for Dirichlets functional). If u / satises
(DP), then
(EL)
_
(u, v)
R
n dx =
_
f v dx v C
c
() .
Remark 4.9. (EL) is said to be the weak form of the Euler-Lagrange equation for
Dirichlets functional.
Proof of Lemma 4.8. Fix v C
c
() and let
i(t) := I(u + t v) t R.
It is easy to see, by an explicit calculation, that
i(t) =
_
_
1
2
[u[
2
f u
_
dx + t
_
((u, v) f v) dx +
t
2
2
_
[v[
2
dx .
Since u+t v / for each t R, the scalar function f : R R must have a minumum
at t = 0. Thus
0 = i
(0) =
_
((u, v) f v) dx
and the proof is complete.
Proof of Theorem 4.2. 1. Let u / be such that
u(x) = div(u)(x) = f(x) x .
Multiply each term of the previous identity by u w for an (arbitrary) w / and
integrate on . This implies that
(4.8) 0 =
_
(u f) (u w) dx =
_
(div(u) f) (u w) dx w /.
By the Gauss-Green theorem (Theorem 4.6), integrating by parts in (4.8), we obtain
93
(4.9) 0 =
_
(0) =
_
((u, v)
R
n f v) dx =
_
(div(u) v f v) dx
=
_
(u f) v dx v C
c
() .
Thus, from the fundamental lemma of the calculus of variations (Theorem 2.67),
u f = 0 a.e. in .
Beceause of u f C
0
(), it follows that
u f = 0 in .
Therefore u satises the boundary-value problem (P).
An immediate consequence of Theorem 4.2 and Lemma 4.4 is the uniqueness for
the boundary-value problem (P).
Theorem 4.10 (Uniqueness of boundary-value problem (P)). Let R
n
be a
bounded open set with = of class C
1
and let f C
0
() and g C
0
().
There exists at most one solution u C
2
() of boundary-value problem (P).
Question: How can we prove the existence of a minimizer u / for the minimiza-
tion problem (DP)?
Historical remark: [MS, Section 12.3] Since it can be proved that the Dirichlets
energy is bounded from below on the admissible class of competitors / (see (4.20)
below), the existence of an inmum is guaranteed, that is,
inf
vA
I(v) R.
94
That this inmum is attained was taken for granted by Riemann (who coined the
term Dirichlets principle) and others. Until Weierstrass in 1869 gave an example of
a functional that does not attain its minimum. Arzel` a in 1896 and later Hilbert in
1897 justied Riemanns use of Dirichlets principle.
On the other hand rstly Prym in 1871 and then, independently, Hadamard in
1906 noticed that the correct admisslble class of functions where looking for classical
solutions of (P) is
/
:=
_
v C
2
() C
0
() : u = g on
_
which is, a priori, larger than /. They actually proved that there are a (not regular)
bounded open set R
2
and a classical soution u /
||2
|D
f|
,
.
Then (C
2
(), | |
C
2) is a B.s.
(ii) (C
2
(), | |
C
2) is not a reexive B.s.
(Hint: (ii) Let n = 1, = (0, 1) and f
h
(x) =
x
h+1
(h + 1) h
. Prove that (f
h
)
h
is a
bounded sequence in (C
2
(), | |
C
2) which does not admit any weakly convergent
sequence (compare with Exercise III.2).
4.3. Dirichlets principle in the Sobolev space and weak solutions. .
Idea: we look for a suitable reexive space of functions (actually a H.s.) cointaining
/, where setting our minimization problem. This space will turn out to be the Sobolev
space.
Focus now our attention on the homogeneous boundary-value problem (P), that is
the problem
(P
0
)
_
u = f in
u = 0 on
,
with boundary datum g 0.
Note that in this case the class of admissible functions above introduced
/ :=
_
v C
2
() : v = 0 on
_
is a v.s.
95
Consider the Euler-Lagrange equation (in weak form) (EL) associated to this prob-
lem, that is, if u /,
(EL)
_
(u, v)
R
n dx =
_
f v dx v C
c
() .
Dene on / the following bilinear form
(u, v)
A
:=
_
(u, v)
R
n dx u, v /.
Exercise IV.2: Let R
n
be a bounded open set. Prove that :
(i) (, )
A
is a scalar product on /;
(ii) (/, | |
A
) is a n.v.s, but not a B.s. with |u|
A
:=
_
(u, u)
A
.
Let us now introduce the following imbedding of / into L
2
() (L
2
())
n
,
(4.13) T : / L
2
() (L
2
())
n
, T(u) = (u, u) .
It is immediate that T is a linear and 1-to-1 mapping.
Notation: From now on R
n
will denote a bounded open set with =
of class C
1
and we will endow, respectively, (L
2
())
n
and L
2
() (L
2
())
n
by the
following product norms:
|w|
(L
2
())
n :=
_
n
i=1
|w
i
|
2
L
2
()
if w = (w
1
, . . . , w
n
) (L
2
())
n
,
|(u, w)|
L
2
()(L
2
())
n :=
_
|u|
2
L
2
()
+|w|
2
(L
2
())
n
if (u, w) L
2
() (L
2
())
n
.
By means of those norms, (L
2
())
n
and L
2
() (L
2
())
n
turn out to be H.s. with
respect to the following scalar products:
(w, z)
(L
2
())
n :=
n
i=1
(w
i
, z
i
)
L
2
()
if w = (w
1
, . . . , w
n
), z = (z
1
, . . . , z
n
) (L
2
())
n
,
((u, w), (v, z))
L
2
()(L
2
())
n := (u, v)
L
2
()
+(w, z)
(L
2
())
n if (u, z), (v, z) L
2
()(L
2
())
n
.
Eventually let
L
2
()
: L
2
() (L
2
())
n
L
2
() denote the rst component pro-
jection, that is,
L
2
()
(u, v) := u if (u, w) L
2
() (L
2
())
n
.
Remark 4.11. Note that by this notation
(u, v)
A
= (u, v)
(L
2
())
n u, v /.
Denition 4.12. Let us denote by
H
1
0
() :=
L
2
()
_
T(/)
L
2
()(L
2
())
n
_
H
1
0
() is called Sobolev space.
Theorem 4.13 (Characterization of the Sobolev space). The following are equivalent:
96
(i) u H
1
0
();
(ii) u L
2
() and there exist a sequence (u
h
)
h
/ and w (L
2
())
n
such that
(*) u
h
u in L
2
() and u
h
w in (L
2
())
n
.
Moreover the vector w = (w
1
, . . . , w
n
) (L
2
())
n
in the statement (ii) satises the
following integration by parts identity
(IP)
_
u
x
i
dx =
_
w
i
dx C
c
(), i = 1, . . . , n.
In particular the vector w is uniquely dened a.e. in .
Denition 4.14. A function u L
1
loc
() satisfying (IP) for a vector w (L
1
loc
())
n
is said to admit weak gradient
Du := (D
1
u, . . . , D
n
u) = (w
1
, . . . , w
n
) in ,
and D
i
u is called the i
th
-weak derivative of u in .
Remark 4.15. In other words Theorem 4.13 provides that a function u H
1
0
()
admits weak gradient Du in .
Proof of Theorem 4.13. The equivalence between the statements (i) and (ii) immedi-
ately follows from the denition of H
1
0
() and the characterization, in a metric space,
of a closed set by means of sequences.
Let us prove the identity (IP). Let u H
1
0
(). By the previous equivalence,
there exist a sequence (u
h
)
h
/ and w (L
2
())
n
such that (*) holds. By the
Gauss-Green theorem, it follows that
_
u
h
x
i
dx =
_
u
h
x
i
dx C
c
(), h, i = 1, . . . , n.
By (*), taking the limit as h in the previous identity, (IP) follows.
Assume there is an other vector w = ( w
1
, . . . , w
n
) (L
2
())
n
which satises (IP).
Then it follows that
_
w
i
dx =
_
w
i
dx C
c
(), i = 1, . . . , n.
From the fundamental lemma of the calculus of variations, we obtain that
w
i
= w
i
a.e. in , i = 1, . . . , n,
and the proof is accomplished.
Norm in H
1
0
().
We are going to introduce now a norm in H
1
0
(). Firstly let us introduce the
following important inequality.
Theorem 4.16 (Poincare inequality). There exists a positive constant C = C() >
0 such that
|u|
L
2
()
C |Du|
(L
2
())
n u H
1
0
() .
Proof. We skip the proof (see [E, Theorem 3, Section 5.6] or [B, Corollary 9.19]).
97
Corollary 4.17. Denote by
(u, v)
H
1
0
:= (Du, Dv)
(L
2
())
n if u, v H
1
0
() ,
where Dz denotes the weak gradient of z H
1
0
().
Then (H
1
0
(), (, )
H
1
0
) is a H.s. and we will denote
|u|
H
1
0
()
:=
_
(u, u)
H
1
0
=
_
(Du, Du)
(L
2
())
n if u H
1
0
() .
In particular the inclusion mapping
incl : H
1
0
() L
2
(), incl(u) := u ,
is continuous.
Proof. From the Poincare inequality, it immediately follows that the bilinear map
(, )
H
1
0
: H
1
0
() H
1
0
() R
is a scalar product. In particular (H
1
0
(), | |
H
1
0
()
) is a n.v.s. Thus, we need only to
prove that
(4.14) (H
1
0
(), | |
H
1
0
()
) is a B.s.
Let
T : (H
1
0
(), | |
H
1
0
()
) (L
2
() (L
2
())
n
, | |
L
2
()(L
2
())
n)
T(u) := (u, Du) .
It is easy to see that T extends the imbedding dened in (4.13), it is still linear and
1-to-1. Moreover
(4.15) |u|
H
1
0
()
|T(u)|
L
2
()(L
2
())
n
C
2
+ 1 |u|
H
1
0
()
u H
1
0
() ,
where C is the Poincare constant in Theorem 4.16 and
(4.16) T(H
1
0
()) = T(/)
L
2
()(L
2
())
n
.
Indeed, (4.15) and (4.16) follow, respectively, from the Poincare inequality and the
dention of H
1
0
().
Since
(T(/)
L
2
()(L
2
())
n
, | |
L
2
()(L
2
())
n) is a B.s.
and, by (4.15) and (4.16),
(H
1
0
(), ||
H
1
0
()
) and (T(/)
L
2
()(L
2
())
n
, ||
L
2
()(L
2
())
n) are topologically isomorphic ,
(4.14) follows.
Theorem 4.18 (Dirichlets principle in the Sobolev space). Let f L
2
() and
consider the Dirichlet energy functional
I : H
1
0
() R, I(v) :=
1
2
_
[Dv[
2
dx
_
f v dx v H
1
0
() .
Then there exists a unique u H
1
0
() such that
(DP) I(u) = min
H
1
0
()
I .
98
Moreover u is characterized by the following property: u H
1
0
() is the unique
solution of the Euler-Lagrange equation, in weak form,
(EL)
_
(Du, Dv)
R
n dx =
_
f v dx v H
1
0
() .
Proof. We are going to apply the generalized Weierstrass theorem with
A = E = H
1
0
() and = I .
Therefore we need only to prove that
(4.17) I : H
1
0
() R is convex;
(4.18) lim
v
H
1
0
+
I(v) = +;
and
(4.19) I is lsc (with respect to the strong topology of H
1
0
()) .
The property (4.17) can be proved like in the case of the functional I : / R
(Lemma 4.4) .
The property (4.19) follows from Corollary 4.17 noticing that I : H
1
0
() R is
actually continuous.
Let us prove (4.18). Indeed, by the Holder and Poincare inequalities, we obtain
I(v) =
1
2
_
[Dv[
2
dx
_
f v dx
1
2
|v|
2
H
1
0
|f|
L
2 |v|
L
2
1
2
|v|
2
H
1
0
C
|f|
L
2 |v|
H
1
0
v H
1
0
() .
(4.20)
Taking the limit in (4.20), as |v|
H
1
0
+, (4.18) follows.
If u H
1
0
() satises (DP), arguing as in Lemma 4.8, one can prove that u satises
(EL). Viceversa, if u H
1
0
() satises (EL), and i : R R denote
i(t) := I(u + t (u w)) t R,
for a given w H
1
0
(), one can prove that
i
(0) =
_
f (u w) dx = 0 .
Applying Lemma 4.5, we infer that
i(t) = I(u + t (u w)) i(0) = I(u) t R,
whence, for t = 1,
I(w) I(u) w H
1
0
() .
Let us now prove the uniqueness of the function u H
1
0
() satisfying the equation
(EL).
Let
: H
1
0
() R, (v) :=
_
f v dx if v H
1
0
() .
99
Then, by Corollary 4.17, (H
1
0
())
() and w (L
2
())
n
such
that
u
h
u in L
2
() and u
h
w in (L
2
())
n
.
Moreover w = Du a.e. in .
Proof. We skip the proof (see [E, Theorem 2, Section 5.3.2] or [B, Corollary 9.8].
Moreover the following simple characterization holds for 1-dimensional Sobolev
spaces, which points out a deep link with the classical space of absolutely continuous
functions.
Theorem 4.23 (Characterization of the 1-dimensional Sobolev spaces). Let n = 1,
= (a, b) with a, b R, a < b.
100
(i)
H
1
((a, b)) =
_
u AC([a, b]) : u
L
2
(a, b)
_
that is, by denition, given u L
2
(a, b), then u H
1
((a, b)) i there exists
u AC([a, b]) such that u = u a.e. on (a, b), u
L
2
(a, b).
Moreover, if Du and u
a.e. in .
(ii)
H
1
0
((a, b)) =
_
u AC([a, b]) : u
L
2
(a, b), u(a) = u(b) = 0
_
that is, by denition, given u L
2
(a, b), then u H
1
0
((a, b)) i there exists
u AC([a, b]) such that u = u a.e. on (a, b), u
L
2
(a, b) and u(a) = u(b) =
0.
Proof. See Exercise IV.4.
Theorem 4.24 (Regularity of weak solutions). Let R
n
be a bounded open set.
Assume that
(i) = is of class C
m+2
with m >
n
2
;
(ii) f H
m
();
(iii) u H
1
0
() is a weak solution of problem (P
0
), that is,
_
(Du, Dv)
R
n dx =
_
f v dx v H
1
0
() .
Then u C
2
() and u(x) = 0 for each x . In particular, u is a classical solution
of problem (P
0
).
Proof. We skip the proof (see [E, Section 6.3] or [B, Theorem 9.25]).
Historical remark:[P, Section 6.7.2] In the 1-dimensional case, the Sobolev spaces
were invented already by Banach in 1922. He introduced H
1
([a, b]) as the space of
functions u AC([a.b]) such that u
L
2
((a, b)) (see Exercise IV.4). The denition
of the Sobolev space H
m
() over a domain R
n
is due to Sobolev in 1938 ([So]).
He was able to overcome the obstacle of dening a right notion of partial derivatives
in higher dimesional setting, by introducing the notion of weak derivative.
101
Exercises
(*)= demanding exercise , (**)= very demanding exercise
I. Derivation of functions and measures.
I.1 (Dirichlet function) Let f : R R be f(x) :=
Q
(x). Prove that
(i) f is Lebesgue measurable;
(ii) f is not Riemann integrable on any interval [a, b], for each < a < b < +;
(iii) f is discontinuous at every point x R;
(iv) Let us consider X = [0, 1] as the topological space endowed with the classical Eu-
clidean topology and consider f : (X, /
1
[0, 1]) R, where by /
1
[0, 1] we denote
the -algebra of Lebesgue measurable sets of R contained in [0, 1]. Suppose Q[0, 1] =
q
i
: i = 1, 2, . . . . For given > 0, let F := [0, 1]
i=1
(q
i
2
(i+1)
, q
i
+ 2
(i+1)
).
Prove that f is measurable, F is closed, f 0 on F and L
1
([0, 1] F) < .
Remark: Notice that f is continuous on F with respect to the relative topology on
F induced by the topology of [0, 1], even though f is discontinuous at every point of
F with respect to the topology of [0, 1].
I.2 Let f : R R be a nondecreasing function and let
f
denote the Lebesgue-
Stieltjes outer measure on R induced by f. Prove that
(i) E R a Borel set B R such that E B and
f
(E) =
f
(B);
(ii) for each compact set K R it holds that
f
(K) < +.
Remark: From exercise I.2 it follows that
f
is a Radon outer measure on R.
(Hint: (i) If
f
(E) = + choose B = R. If
f
(E) < + , for each k N
there exists a family I
(k)
h
: h N, I
(k)
h
= (a
(k)
h
, b
(k)
h
] such that E
h=1
I
(k)
h
and
f
(E)
h=1
f
(I
(k)
h
)
f
(E) +
1
k
. By choosing B :=
k=1
(
h=1
I
(k)
h
) we obtain
the desired assertion.
(ii) There exist a, b R con a < b such that K (a, b]. Then
f
(K)
f
((a, b]) <
+.)
I.3 Let f(x) := x if x R. Then
f
(E) = /
1
(E) for any E R.
(Hint: By denition, it is immediate that /
1
(E)
f
(E). To prove the opposite
inequality, observe that if
i=1
[a
i
, b
i
] E, then
i=1
(a
i
2
i
, , b
i
] E and
f
(E)
i=1
(b
i
a
i
) + for each > 0.)
I.4 Let be a Borel measure ( or also a Borel outer measure) nite on R (that is
(R) < +). Dene f(x) := ((, x]) if x R. Prove that
(i) f : R R `e is nondecreasing and right-continuous;
(ii) f is continuous at a point x
0
i (x
0
) = 0;
(iii) ((a.b]) = f(b) f(a) for all a, b R with a < b.
(Hint: (i) To prove that f is right-continuous at x observe that if (x
h
)
h
R is a
decreasing sequence, converging to x, then
lim
h
f(x
h
) = lim
h
((, x
h
]) = (
h=1
(, x
h
]) = f(x) .)
I.5 Let be an outer measure on a set X. Let (E
i
)
i
be a sequence of subsets
of X. Then there exists a sequence of disjoint sets (A
i
)
i
such that A
i
E
i
and
i=1
E
i
=
i=1
A
i
. In case the sets E
i
, for each i N, are -measurable, so are A
i
.
102
(Hint:For each j N, dene S
j
=
j
i=1
E
j
. Note that
i=1
E
i
= S
1
(
k=1
(S
k+1
S
k
)) .
Now take A
1
:= S
1
and A
i+1
:= S
i+1
S
i
for all i 1. )
I.6 Let U R be an open set. Then U is a countable union of disjoint open intervals.
(Hint: For each x U, let I
x
denote the biggest open interval containing x and
contained in U. Prove that U =
xU
I
x
and I
x
I
y
= or I
x
= I
y
. Also prove that
the the family I
x
: x U is countable.)
I.7 Let f : [a, b] R be a nondecreasing function. Prove that f has at most a
countable set of discontinuity points.
I.8 Let X = (0, 1), = /
1
, = # and / = /
1
(0, 1). Prove that:
(i) is not -nite;
(ii) << ;
(iii) there is no measurable function w : (0, 1) [0, ] such that (E) =
_
E
wd
for each E /.
(Hint: (iii) First prove that, for given x (0, 1),
_
{x}
wd = w(x) for any measurable
function w : (0, 1) [0, ] . Then conclude by contradiction.)
I.9 Let and be two Radon measures on R. Suppose that (R) < +. Then the
following are equivalent:
(i) << ;
(ii) for each > 0 there exists = () > 0 such that
m
i=1
((a
i
, b
i
)) < ,
for each disjoint family of intervals (a
1
, b
1
), . . . , ( a
m
, b
m
), verifying
m
i=1
((a
i
, b
i
)) < .
(Hint: Recall that << if and only if
(AC)
> 0 = () > 0 such that A B(R) with (A) < (A) < .
From (AC) it follows the implication (i) (ii). To prove the reverse implication, use
again (AC). Recall that, for each A B(R) with (A) < , there exists an open set
U A tale che (U) < . Then apply exercise I.6.)
I.10 Let consider = /
1
, =
0
as measures on the - algebra /
1
of Lebesgue
measurable sets in R, where
0
denotes the Dirac measure at 0, that is,
0
(E) :=
_
1 if 0 E
0 if 0 / E .
Prove that the Lebesgue decomposition of with respect to , =
ac
+
s
, is given by
ac
0 and
s
= .
I.11 Let (X, /, ) be a measure space and let f : X R be a measurable function.
Suppose at least one of f
+
:= f0 or f
(x) and g
(x) = f
(x).
(Hint: Dene g(x) := lim
yx
+ f(y) R if x R.
(i) Prove that g is nondecreasing, right-continuous and keep in mind exercise I.7.
(ii) Prove that: the continuity points of f and g agree, at each point x of continuity
f(x) = g(x) and the following inequality holds
g
_
x +
_
1
h
[h[
_
h
_
g(x) f(x + h) f(x) g(x + h) g(x)
for each 0 < < 1, h ,= 0.)
I.14* (Cantor set, see also [GZ, section 4.4]) Let C R denote the set, called Cantor
(ternary) set, constructed in stages as follows.
At the rst step let I
1,1
:= (
1
3
,
2
3
). Thus I
1,1
is the open middle third of the interval
[0, 1]. Also denote C
1
:= [0, 1] (
1
3
,
2
3
) = [0,
1
3
] [
2
3
, 1]. The second step involves
performing the rst step on each of the two remaining intervals of C
1
. That is, we
produce two open intervals I
2,1
and I
2,2
, each being the open middle third of one
of the two intervals comprising C
1
. Again denote C
2
:= C
1
(I
2,1
I
2,2
). At the
i
th
step we produce 2
i1
open intervals, I
i,1
, . . . , I
i,2
i1, each of lenght (
1
3
)
i
. Denote
C
i
:= C
i1
(I
i,1
. . . I
i,2
i1) and dene C :=
i=1
C
i
. Then
(i) C is a compact set and /
1
(C) = 0;
(ii) C is nowhere dense (that is, C does not contain any (non empty) open set of R);
(iii) C has the cardinality of the continuum (that is, card(C) = c = card(R)).
(Hint: (iii) Prove that the numbers x C agree with x [0, 1] whose ternary
expansion admits the digits either 0 or 2. Here we adopt the convenction to represent
each numver x [0, 1] by means of a ternary expansion which has the smallest number
of digits equal to 1. Therefore C has the cardinality of the set 2
N
.)
I.15* (Cantor-Lebesgue-Vitali function, see also [GZ, section 5.6]) Let C
i
(i N)
denote the sets dened in the exercise I.12. Let J
i,1
, . . . J
i,2
i
1
denote the open intervals
such that [0, 1] C
i
=
2
i
1
h=1
J
i,h
and order them in the obvious way from left to right.
104
Given i N, let f
i
: [0, 1] [0, 1] be the continuous function dened as follows:
f
i
(0) = 0, f
i
(1) = 1, f
i
(x) =
h
2
i
if x J
i,h
, and f
i
linearly on C
i
. Prove that the
sequence of functions (f
i
)
i
converges uniformly to a function f : [0, 1] [0, 1] (called
Cantor-Lebesgue-Vitali function) verifying the following properties:
(i) f is continuous, nondecreasing and f(0) = 0, f(1) = 1.
(ii) f
(x) = 0 for each x [0, 1] C where C denotes the Cantor set dened in the
exercise I.14. Moreover f
(x) dx does not hold, even though the derivative of f vanishes a.e. in [0, 1]! Also
notice that, from the assertion (iii), f carries a set of measure 0 onto the interval
[0, 1].
(Hint: (i): Prove that the sequence (f
i
)
i
converges uniformly by means of the esti-
mate
[f
i
(x) f
i+1
(x)[
1
2
i
x [0, 1], i N.
(ii) The assertion follows noticing that f(x) =
h
2
i
if x J
i,h
.
(iii) The continuity of f implies that f([0, 1]) = [0, 1] and f(C) is compact. On the
other hand, by construction, f([0, 1] C) is denumerable, thus f(C) = [0, 1].
(iv) The assertion follows from the previous assertion (ii), the Lebesgue decomposition
of the measure
f
(with respect to /
1
) and the dierentiation theorem applied to
f
.)
I.16 Let f : [0,
1
] R be the function
f(x) :=
_
_
_
x sin
_
1
x
_
if 0 < x
1
0 if x = 0 .
Prove that f has unbounded variation.
I.17 Let f : [a, b] R and let V
f
(a; b) denote the total variation of f on [a, b]. Prove
that
(i) if g : [a, b] R, then V
f+g
(a; b) V
f
(a; b) + V
g
(a; b).
(ii) V
c f
(a; b) = [c[ V
f
(a; b) for each c R.
(iii) V
f
(a; b) = V
f
(a; c) + V
f
(c; b) for each c [a, b].
(iv) if f is of bounded variation, then the function g : [a, b] R, g(x) := V
f
(a; x) if
x [a, b] is nondecreasing.
I.18 (Example of a continuous function nowhere dierentiable) Let g : R R be the
function dened by g(x) = [x[ if x [1, 1], extented with period 2 outside [1, 1]
(that is, g(x + 2) = g(x) for each x R).
(i) Prove that the series
n=0
_
3
4
_
n
g(4
n
x)
converges for each x R.
105
(ii) Let f : R R be the sum of the previous series, that is
f(x) :=
n=0
_
3
4
_
n
g(4
n
x) x R, .
Prove that f is continuous.
(iii)* Prove that f is nowhere dierentiable on R.
(iv) Given a, b R with a < b, is f : [a, b] R of bounded variation?
(Hint: (i), (ii) and (iii): see [R1], Theorem 7.18.)
I.19 Let f : [a, b] R be a Lipschitz function, that is, by denition, L > 0 such
that [f(x) f(y)[ L[x y[ for each x, y [a, b]. Prove that f AC([a, b]).
Remark: From the previous exercise it follows that a Lipschitz function f : [a, b]
R is dierentiable a.e. in [a, b] and it satises f(x) f(a) =
_
x
a
f
(x) dx .
(Hint: (i) Use the denition of absolutely continuous function. (ii) From the assertion
(i), prove that (f g)
(x) = f
+ |f + g|
2
= 2
_
|f|
2
+|g|
2
_
f, g C
0
([a, b]) .
For instance, consider [a, b] = [1, 1], f(x) = 1 x
2
and g(x) = 1 f(x).)
II.4 Show that (C
1
(
), [[ [[
C
1) is a Banach space, where [[u[[
C
1 :=
||1
[[D
u[[
.
and R
n
is a bounded open set.
II.5 Let M > 0 be a given constant and let T = f C
1
([a, b]) : [[f[[
C
1 M.
Prove that
(i) T is a relatively compact set of (C
0
([a, b]), [[ [[
) ;
(ii) T is not a compact set either of (C
1
([a, b], [[ [[
C
1) and (C
1
([a, b], [[ [[
C
0).
(Hint: (ii) It is sucient to prove that T is not a compact set of (C
1
([a, b], [[ [[
C
0).
Consider [a, b] = [1.1] and T := f
h
: h = 1, 2, . . . where f
h
(x) := [x[
1+
1
h
.)
II.6 Let f
i
: A R
n
R (i = 1, 2, . . . ) be a sequence of continuous functions and
let (x
h
)
h
A. Suppose that
(i) there exists f : A R such that f
i
f uniformly on A;
(ii) there exists x A such that x
h
x.
Then there exists lim
i
f
i
(x
i
) = f(x).
Does the assertion still hold if in the assumption (i), in place of the uniform con-
vergence, we assume the pointwise one?
II.7 Let (E, [[ [[
E
) and (F, [[ [[
F
) be normed vector spaces. Endow E F with one
of the following norms:
[[(x, y)[[
EF
= [[x[[
E
+[[y[[
F
[[(x, y)[[
EF
=
_
[[x[[
E
2
+[[y[[
F
2
[[(x, y)[[
EF
= max[[x[[
E
, [[y[[
F
.
(i) Show that [[(x, y)[[
EF
is actually a norm.
(ii) Show that, if (E, [[[[
E
) and (F, [[[[
F
) are Banach spaces, then (EF, [[(x, y)[[
EF
)
is still a Banach space.
(iii) Show that, if (E, [[ [[
E
) and (F, [[ [[
F
) are separable, then (E F, [[(x, y)[[
EF
)
is still separable.
II.8 Let (f
i
)
i
C
1
([a, b]). Suppose that
(i) (f
i
)
i
e (f
i
)
i
are bounded sequences in (C
0
([a, b]), [[ [[
);
(ii) (f
i
)
i
e (f
i
)
i
are equicontinuous sequences on [a, b].
Then (f
i
)
i
is relatively compact sequence in (C
1
([a, b]), [[ [[
C
1), that is, there exists
a subsequence of (f
i
)
i
converging to a function f C
1
([a, b]) with respect to the norm
[[ [[
C
1.
(Hint: Apply the Arzel` a -Ascoli theorem to both sequences (f
h
)
h
and (f
h
)
h
.).
II.9 Let f : A R
n
R be a Lipschitz function. Prove that
(i) f is uniformly continuous on A;
(ii) there exists a unique Lipschitz function
f :
A R (with same Lipschitz constant
of f) such that
f = f on A.
107
II.10 (i) Let f C
1
([a, b]). Prove that f Lip((a, b)) and [[f[[
C
1 = [[f[[
Lip
.
(ii) C
1
([a, b]) is a closed set of (Lip((a, b)), [[ [[
Lip
).
II.11 Let R
n
be a connected open set and let f : R
n
R. Suppose there
exist two constants L > 0 and > 1 such that
[f(x) f(y)[ L[x y[
x, y .
Then f is constant on , that is, there exists c R such that f(x) = c for each x .
II.12 Given f C
0
([a, b], let [[f[[
L
2 :=
_
_
b
a
f(x)
2
dx. Prove that:
(i) (C
0
([a, b], [[ [[
L
2) is a normed vector space;
(ii) (C
0
([a, b], [[ [[
L
2) is not a Banach space.
(Hint:(ii) For instance, let [a, b] = [1, 1] and let f
h
: [1, 1] R be (h = 1, 2, . . . )
be dened as f
h
(x) :=
_
_
0 if 1 x
1
h
1 + hx if
1
h
x 0
1 if 0 x 1
. Show that (f
h
)
h
is a Cauchy se-
quence in (C
0
([1, 1], [[[[
L
2). On the other hand, if f(x) :=
_
0 if 1 x < 0
1 if 0 x 1
. ,,
then f
h
f in L
2
([1, 1]) and f / C
0
([1, 1].)
II.13 Let 1 p q . Suppose that R
n
is a bounded open set. Show that:
[[f[[
L
p [[
(
1
p
1
q
)
[[f[[
L
q
(Hint: Use Holders inequality.).
II.14 Let R
n=1
[x(n)[
2
< and
[[x[[
l
2 :=
_
n=1
[x(n)[
2
.
Prove that
(i) (R
, [[ [[
l
2) is a normed vector space, but not a Banach space.
(ii) (l
2
, [[ [[
l
2) is a Hilbert space;
(iii) (R
, [[ [[
l
2)
l
2
(Hint: (iii) Prove that for each f (R
, [[ [[
l
2)
there exists y l
2
such that
f(x) =
n=1
y(n) x(n) for every x l
2
.)
II.15 Let (E, [[ [[) be a normed vector space and let f : E R be a linear functional.
Show that f is continuous if and only if f is bounded, that is,
[[f[[
E
:= sup
xE\{0}
[f(x)[
[[x[[
< +.
II.16 Find a sequence (f
n
)
n
L
p
([0, 1]), with 1 p < , satisfying:
(i) f
n
0 in L
p
([0, 1]);
(ii) x [0, 1], (f
n
(x))
n
R does not converge to 0.
108
II.17 Let (E, [[ [[) be a normed vector space and let f : E R be a linear functional.
Show that, if dimE < , then f is continuous.
II.18 Let (E, [[ [[) be a normed vector space and let E
:= f : E R : f linear
and continuous. Prove that:
(i) E
is a vector space;
(ii) if [[ [[
E
denotes the canonical norm, then (E
, [[ [[
E
) is a Banach space.
II.19 Let (E, [[ [[) be a normed vector space. Prove that, if n = dimE < , then
E ed E
, [[ [[
(EF)
) (E
, [[ [[
E
F
)
f (f(, 0), f(0, )) .
Prove that T is a linear and topological ismorphism.
II.21 Let (E, [[ [[
E
) be a normed vector space and let H denote the hyperplane
H = f = . Show that
H is a closed set f E
.
II.22 Let (E, [[ [[)be a normed vector space and F E a subspace.
Show that:
(i) if dim
R
F < +, then F is closed;
(ii) all the subspaces of R
n
are closed in R
n
;
(iii) F is also a subspace.
II.23 Let f : R
n
R be a continuous function and denote by supp
e
(f) the essential
support of f, that is supp
e
(f) := R
n
A
f
, where
A
f
:=
_
A
f
and /
f
:= R
n
: open , f = 0 a.e. on .
Prove that
supp
e
(f) = supp(f) ,
that is,
R
n
A
f
= closure x R
n
: f(x) ,= 0
II.24 Let R
n
be an open set, let h : R and : [0, +) be Lebesgue
measurable functions and suppose that
_
[h[ dx
_
p
[h[
p
dx
109
II.25 Let u L
p
(R
n
) L
q
(R
n
), with 1 p < q . Prove that u L
r
(R
n
) for
every r [p, q] and
|u|
L
r |u|
L
p |u|
1
L
q
provided that
1
r
=
p
+
1
q
.
II.26 Let (E, [[ [[
E
) and (F, [[ [[
F
) be normed vector spaces. Denote
/(E, F) := T : E F such that T is llinear and continuous
Show that (/(E, F), [[ [[
L(E,F)
) is a normed vector space, endowed with the norm
[[T[[
L(E,F)
:= sup[[T(x)[[
F
: x B
E
.
II.27 Let (E, [[ [[
E
) and (F, [[ [[
F
) be, respectively, a Banach and a normed vector
space. Let (T
n
)
n
/(E, F). Suppose there exists
Tx := lim
n
T
n
x x E
Show that:
(i) sup
nN
[[T
n
[[
L(E,F)
< +;
(ii) T /(E, F);
(iii) [[T[[
L(E,F)
liminf
n
[[T
n
[[
L(E,F)
.
(Hint: Use the Banach- Steinhaus theorem.)
II.28** Given (E, [[ [[) normed vector space with dim
R
E = , prove there exists a
nonlinear operator T : E E continuous, bijective but with inverse T
1
: E E
discontinuous.
(Hint (proposed by M. Degiovanni): Denote with S = u E : |u| = 1 the
unit sphere of E. Then S is not compact.
Prove, because of the Tietze extension theorem, that there exists an unbounded
function f : S [1, +).
Dene T : E E by T(u) =
u
f(u/u)
if u ,= 0 and T(0) = 0. Prove that T is
continuous and bijective.
Prove there exists a non convergent sequence (u
h
)
h
S with lim
h
f(u
h
) = +
such that v
h
:= lim
h
T(u
h
) = 0. Hence the sequence (u
h
= T
1
(v
h
))
h
does not
converge, even though the sequence lim
h
v
h
= 0 . )
II.29 Let (E, [[ [[
E
) and (F, [[ [[
F
) be Banach spaces. If T : E F is a linear
operator, denote by G(T) its graph, that is G(T) := (x, Tx) : x E.
Show that G(T) is closed if and only if (E, [[ [[
1
) is a Banach space, where[[x[[
1
:=
[[x[[
E
+[[Tx[[
F
.
(Hint: Use the closed graph theorem.)
II.30 Let E = C
0
([a, b]) R
n
, [[v[[
E
:= [[f[[
+[[w[[
R
n if v = (f, w) E. Show that
(E, [[ [[
E
) is a Banach space.
110
II.31 Consider the following initial value problem (or Cauchy problem) for the linear
ordinary dierential equation (ode):
u
(n)
+
n1
i=0
a
i
(t)u
(i)
= f
with initial values: u(t
0
) = w
1
, . . . , u
(n1)
(t
0
) = w
n
, where a
i
C
0
([a, b]) (i =
1, . . . , n 1), w
i
R (i = 1, . . . , n) and t
0
[a, b] are given. Dene the operator
T : E := C
0
([a, b]) R
n
F := C
n
([a, b]), dened by T(f, w) = u, where (f, w) are
the given data problem and u denotes the (unique) problem solution. Prove that:
(i) T is a linear operator.
(ii) T : (E, [[ [[
E
) (F, [[ [[
F
) is continuous, where [[ [[
E
denotes the norm dened
in exercise III.30 and [[ [[
F
:= [[ [[
C
n.
Remark: From exercise II.31 (ii), it follows the continuity of the solutions with
respect to the initial data: the so-called well-posedness of the Cauchy problem for a
linear ode of order n.
(Hint: Use the closed graph theorem.)
II.32 Let E = C
1
([a, b]), F = C
0
([a, b]), [[ [[
E
= [[ [[
F
= [[ [[
u Tu
where
< Tu, v >
H
H
:= (u, v)
H
.
Show that T is a linear isometry.
II.34 Let
B :=
_
1
2
_
_
cos hx
: h = 1, 2,
_
_
sin hx
: h = 1, 2,
_
.
Show that B is a Hilbert basis of L
2
(, ), that is:
(i) B is an orthonormal system of H;
(ii)* span B is dense in L
2
(, ).
(Hint: (ii): see [R2] section 4.24.)
II.35 Let (E, [[ [[) be a normed vector space. Prove that, for xed x E, it holds
that
[[x[[ = sup< f, x >
E
E
: f B
E
= max< f, x >
E
E
: f B
E
.
111
II.36 Let (X, [[ [[
X
) and (Y, [[ [[
Y
) be Banach spaces, S /(X, Y ). let S
: Y
(f) is dened by
g(x) := f(S(x)) x X ,
if f Y
. Prove that:
(i) S
/(Y
, X
);
(ii) if S is dense, that is, S(X) is dense in Y , then S
: Y
is injective.
II.37 Let l
1
:= x R
N
:
n=1
[x(n)[ < ) := [[x[[
l
1and
l
:= x R
N
: sup [x(n)[ < ), equipped with the norm [[x[[
l
:= sup
nN
[x(n)[.
Show that:
(i) (l
1
)
, that is, f (l
1
)
!y l
n=1
y(n)x(n) x
l
1
.
(ii) (l
1
, [[ [[
l
1) is a separable Banach space, but (l
, [[ [[
l
) is not a separable Banach
space.
II.38 Let (E, [[ [[
E
) be a Banach space and let (x
n
)
n
B
E
.
Denote M
0
:= span
R
x
h
: h N. Show that M
0
is separable.
II.39 Let (X, /, ) be the measure space where X = N, / = T(X), = # and
# denotes the counting measure on N.
Show that:
(i) L
p
(X, ) = l
p
:= f : N R :
n=1
[f(n)[
p
< if 1 p < ;
(ii) L
(X, ) = l
:= f : N R : sup
n
[f(n)[ < if p = .
III. Complements on Banach and Hilbert spaces.
III.1 Given (f
h
)
h
E
, f E
, (x
h
)
h
E and x E, show that, if f
h
f e
x
h
x,, it follows that
< f
h
, x
h
>< f, x > .
III.2 Let E = C
0
([0, 1]), [[ [[
E
= [[ [[
, (f
h
)
h
E, f E. Prove that:
(i) if f
h
f in E, then
(a) M > 0 such that [f
h
(x)[ Mx [0, 1], h N;
(b) f
h
(x) f(x), x [0, 1].
(ii) f
h
(x) := x
h
, x [0, 1] does not weakly converge in E;
(iii) (C
0
([0, 1]), [[ [[
) is not reexive.
III.3 Let E = C
1
([0, 1]), [[ [[
E
= [[ [[
C
1, (f
h
)
h
E, f E. Prove that:
(i) if f
h
f in E, then f
h
(x) f(x) e f
h
(x) f
:= x R
N
: sup [x(n)[ < ) equipped with the norm [[x[[
l
:=
sup
nN
[x(n)[.
(i) Show that (l
, [[ [[
l
) is a Banach space, but it is not separable.
(ii) Consider (
n
)
n
(l
dened by
<
n
, x >:= x(n) .
Prove that [[
n
[[
(l
)
= 1 n N.
(iii) Show that it does not exist any subsequence of (
n
)
n
weakly converging in (l
.
(Hint: (i) By contradiction: suppose D = d
i
: i N l
dense. Dene z l
in
the following manner: x(h) = (1)
k
if h = h
k
and x(h) = 0 otherwise. Then,
<
h
k
, x >= (1)
k
does not converge.)
Remark: From exercise III.5 (iii) it follows that B
E
is not sequentially compact
with respect to (E
, E
).
III.6 Let f
h
: [0, 1] R (h = 1, 2, . . . ) be the functions dened by f
h
(x) = h if
0 x 1/h and f
h
(x) = 0 otherwise. Prove that it does not exist a subsequence
(f
h
k
)
k
and a f L
1
((0, 1)) such that f
h
k
f in L
1
((0, 1)).
(Hint: By contradiction: there exists (f
h
k
)
k
and a f L
1
((0, 1)) such that f
h
k
f
in L
1
((0, 1)). Then, there exists a sequence (g
i
)
i
L
;
(ii) [[f[[
E
= 2;
(iii)
max
B
E
f, that is, there is no u B
E
such that f(u) = 2.
(Remark: Exercise III.7 (iii) implies that E is not reexive: why?
Furthermore, compare exercise III.7 with exercise II.35)
113
III.8 Let (E, [[ [[)
E
and (F, [[ [[)
F
be reexive Banach spaces. Show that (E
F, [[ [[
EF
) (equipped with any one of the norms introduced in exercise II.7) is still
a reexive Banach space.
(Hint: Use exercise II. 20.)
III.9 Let (E, [[ [[
E
) and (F, [[ [[
F
) be normed vector spaces and let T : (E, [[ [[
E
)
(F, [[ [[
F
) denote a linear continuous operator. Prove that
x
h
x in E T(x
h
) T(x) in F .
(Hint: see [B] Theorem 3.9.)
III.10 Let (H, (, )) be a Hilbert space and let (x
h
)
h
H, x H. Prove that:
(i)
x
h
x (x
h
, y) (x, y) y H ;
(ii)
x
h
x x
h
x and [[x
h
[[ [[x[[ .
III.11 Let H = L
2
((, )), (, ) = (, )
L
2 and let f
h
(x) := sin(hx) if x (, ).
Prove that:
(i)
f
h
0 ;
(ii) (f
h
)
h
does not (strongly) converge to 0 in L
2
((, )).
(Hint: (i) use exercise II.34; (ii) use exercise III.10 (ii).)
III.12 Let E = L
1
((0, 1)), [[ [[ = [[ [[
L
1 and let
C :=
_
u L
1
((0, 1)) : u(x) 0 a.e. x (0, 1) ,
_
1
0
x u(x)dx 1
_
.
Show that:
(i) C is a nonempty, closed and convex set of E;
(ii) d(0, C) := inf [[u[[ : u C = 1;
(iii)* there is no u C such that [[u[[ = d(0, C) = 1.
III.13 Let E = R
n
and [ [ denote the Euclidean norm. Prove that:
(i) a : E E R is a bilinear map if and only if there exists a unique real matrix
n n A = [a
ij
]
ij
such that
a(u, v) = (Au, v) u, v R
n
,
where (, ) denotes the canonical scalar product on R
n
.
(ii) If a : E E R is a bilinear map, it is also continuous.
(iii) Let A be a real n n matrix such that (Au, u) > 0 u R
n
0. Then A is
invertible.
(iv) The inverse implication of the statement (iii) does not hold.
(Hint: (i) It follows from well known results of linear algebra. (ii) It follws by
recalling the following property : if dimE < and f : E R is linear, then f is
114
also continuous. (iii) Prove that the linear operator T : R
n
R
n
T(u) := Au is both
injective and surjective. (iv) Find a counterexample through a 2 2 matrix.)
III.14 Let (H, (, )) be a Hilbert space and let A : H H denote a linear continuous
operator such that there exists a constant > 0 for which
(Au, u) [[u[[
2
u H .
Then A is invertible and [[A
1
[[
L(H,H)
1
.
(Hint: Prove that R(A) := Au : u H, the range of A in H, is closed and dense
in H, whence R(A) = H.)
IV. An introduction to the Sobolev space and an application to Poissons
equation.
Exercise IV.1 (i) If f C
2
(), denote
|f|
C
2 :=
||2
|D
f|
,
.
Then (C
2
(), | |
C
2) is a B.s.
(ii) (C
2
(), | |
C
2) is not a reexive B.s.
(Hint: (ii) Let n = 1, = (0, 1) and f
h
(x) =
x
h+1
(h + 1) h
. Prove that (f
h
)
h
is a
bounded sequence in (C
2
(), | |
C
2) which does not admit any weakly convergent
sequence (compare with Exercise III.2).
Exercise IV.2 Let R
n
be a bounded open set and let
/ :=
_
v C
2
() : v = 0 on
_
, (u, v)
A
:=
_
(u, v)
R
n dx u, v /.
Prove that :
(i) (, )
A
is a scalar product on /;
(ii) (/, | |
A
) is a n.v.s, but not a B.s. with |u|
A
:=
_
(u, u)
A
.
Exercise IV.3 (Heavised function) Prove that the function H : R R, H(x) :=
[0,+)
(x) does not admit weak derivative in L
1
loc
(R).
(Hint: Notice that, for each C
1
c
(R)
_
R
dx = (0) =
_
R
d
0
.
Prove there is no function v L
1
loc
(R) such that
_
R
d
0
=
_
R
v dx C
1
c
(R) .)
Exercise IV.4* (Characterization of the one-dimensional Sobolev spaces) Let a,
b R with a < b. Prove that
(i)
H
1
((a, b)) =
_
u AC([a, b]) : u
L
2
(a, b)
_
115
that is, by denition, given u L
2
(a, b), then u H
1
((a, b)) i there exists u
AC([a, b]) such that u = u a.e. on (a, b), u
L
2
(a, b).
Moreover, if Du and u
a.e. in .
(ii)
H
1
0
((a, b)) =
_
u AC([a, b]) : u
L
2
(a, b), u(a) = u(b) = 0
_
that is, by denition, given u L
2
(a, b), then u H
1
0
((a, b)) i there exists u
AC([a, b]) such that u = u a.e. on (a, b), u
L
2
(a, b) and u(a) = u(b) = 0.
(Hint : (i) The inclusion H
1
((a, b)) u AC([a, b]) : u
L
2
(a, b) follows by
Meyers- Serrins theorem and the fundamental theorem of calculus. Indeed, if
(u
h
)
h
C
(), u
h
u and u
h
v in L
2
(a, b) ,
then, up to a subsequence, it follows that
(1) u
h
(x) u(x) a.e. x (a, b) .
Fix x
0
(a, b) such that (1) holds and consider the identity
u
h
(x) = u
h
(x
0
) +
_
x
x
0
u
h
(t) dt x (a, b) .
By passing to the limit in the previous identity we get that u(x) = u(x) a.e. x (a, b),
where
u(x) = u(x
0
) +
_
x
x
0
v(t) dt x [a, b] .
Thus, from the fundamental theorem of calculus, u AC([a, b]), u
= v = Du a.e. in
(a, b). The reverse inclusion can be proved by arguing as follows. Let u AC([a, b]),
with u
L
2
(a, b). Notice that, by the fundamental theorem of calculus,
u(x) = u(a) +
_
x
a
u
(t) dt x [a, b] .
Let (
h
)
h
be a sequence of molliers in R and let u
= u
in (a, b) and u
0 otherwise. Let v
h
, u
h
: R R be the functions
dened respectively by
v
h
(x) := (
h
u
)(x) and u
h
(x) := u(a) +
_
x
a
v
h
(t) dt if x R.
Prove that
(u
h
)
h
C
(), u
h
u and u
h
u
in L
2
(a, b) .
(ii) The inclusion H
1
0
((a, b)) u AC([a, b]) : u
L
2
(a, b), u(a) = u(b) = 0 still
follows by the fundamental theorem of calculus. Indeed, if
(u
h
)
h
/ :=
_
u C
2
([a, b]) : u(a) = u(b) = 0
_
, u
h
u and u
h
v in L
2
(a, b) ,
then it follows that
u(x) =
_
x
a
v(t) dt x [a, b] and u(b) =
_
b
a
v(t) dt = 0 .
Thus, from the fundamental theorem of calculus, u AC([a, b], Du = u
= v a.e. in
(a, b). The reverse inclusion can be proved by arguing as before with slight changes.
116
More precisely, let u AC([a, b]), with u
L
2
(a, b) and u(a) = u(b) = 0. Notice
that, by the fundamental theorem of calculus,
u(x) =
_
x
a
u
(t) dt = 0 .
Let (
h
)
h
be a sequence of molliers in R and let u
)(x) and u
h
(x) :=
_
x
a
v
h
(t) dt if x R.
Prove that
(u
h
)
h
/, u
h
u and u
h
u
in L
2
(a, b) .)
Exercise IV.5 Let n = 1, = (a, b) with < a < b < +, f L
2
() and
R be given. Consider the functional I : H
1
0
() R dened by
I(v) :=
1
2
_
(v
2
+ v
2
) dx
_
f v dx v H
1
0
()
and the boundary value problem
(P
0
)
_
u
+ u = f in
u = 0 on
,
Prove that:
(i)(Dirichlets principle) for each 0, there exists a unique u H
1
0
() solution of
the problem
(DP) I(u) = min
H
1
0
()
I ,
and u satises the (weak) Euler-Lagrange equation
(EL)
_
(u
+ uv) dx =
_
f v dx v H
1
0
() .
Viceversa, if u H
1
0
() satises (EL), then u is a solution of (DP). A function
u H
1
0
() satisfying (EL) is said to be a weak solution of problem (P
0
).
(ii) (Regularity of the weak solutions) If f C
0
() and u H
1
0
() is a weak solution
of (P
0
), then u is a classical solution of (P
0
), that is, u C
2
() and u satises (P
0
)
in a pointwise sense.
(iii) If u C
2
() is a classical solution of (P
0
), then u is a weak solution of (P
0
).
(iv) If = (0, ), = 1 and f 0, then the functions u
1
(x) 0 and u
2
(x) sin x
are weak solutions of the problem (P
0
).
(v) If = (0, ), = 1 and f 1, then there are no weak solutions of the problem
(P
0
).
Remark: From Exercise IV.5 (iv) and (v) it follows that neither uniqueness and
existence of weak solutions for problem (P
0
) can be taken for granted whenever < 0.
(Hint : (i) About the existence, apply the generalized Weierstrass theorem to the
functional I : H
1
0
() R, like in the case of the Dirichlet energy functional. In order
that u satises (EL), consider the function i : R R, dened by i(t) := I(u + t v)
117
for a given v H
1
0
(). Since u is a minimizer for I, t = 0 is a minimum point for i,
then
0 = i
(0) =
_
(u
+ uv) dx
_
f v dx .
Now assume u satises (EL) and consider the function i : R R, dened by i(t) :=
I(u + t (u w)) for a given w H
1
0
(). Notice that, being I convex (to be proved),
i is also convex. Because
i
(0) =
_
(u
(u w)
+ u(u w)) dx
_
f (u w) dx = 0 ,
at t = 0 there is a minumum point for i. Thus u satises (DP).
(ii) First, let us prove that u C
2
([a, b]). Recall that, by the characterization of
1-dimensional Sobolev spaces (Exercise IV.4), Du = u
= Dv C
0
([a, b]), whence u
= Du = v C
1
([a, b]).
Therefore u C
2
([a, b]). Let us now prove that u satises (P
0
) in a pointwise sense.
Integrating by parts in (EL) we get
_
(u
+ u) v dx =
_
(u
+ uv) dx =
_
f v dx v H
1
0
() .
By the fundamental lemma of the calculus of variations,
u
+ u = f a.e. in ,
and, by the continuity,
(1) u
+ u = f in .
(iii) Multiply (1) by a function v H
1
0
() and perform an integration by parts.
(iv) By (ii) and (iii), it is sucient to prove that u
1
and u
2
are classical solutions.
(v) By (ii) and (iii), it is sucient to prove there are no classical solutions.)
References
[De] S. Delladio, Diario del corso di Analisi Funzionale, A.A. 2011/12, available at
http://www.science.unitn.it/ delladio/
[Ba] S. Banach, Theorie des operations lineaires, Warszawa,1932.
[B] H. Brezis, Functional Analysis, Sobolev Spaces and Partial Dierential Equations, Springer,
2011. An Italian version is also available: H. Brezis, Analisi Funzionale- Teoria e applicazioni,
Liguori, Napoli, 1986. .
[D] J. Dieudonne, History of Functional Analysis, Mathematical Studies 49, Notas de Matematica
(77), North-Holland,1981.
[E] L.C. Evans, Partial Dierential Equations, Graduate Text in Math. 19, AMS, Providence,
Rhode Island,1998.
[F] G.B. Folland , Real analysis. Modern techniques and their applications, Second Edition, Pure
and Applied Mathematics (New York). A Wiley-Interscience Publication. John Wiley & Sons,
Inc., New York, 1999.
[GZ] R. F. Gariepy, W. P. Ziemer, Modern Real Analysis, PWS Publishing Company, Boston, 1994.
[Le] H. Lebesgue, Lecons sur lintegration et la recherche de fonctions primitives, Deuxieme Edition,
Gauthier-Villars, Paris, 1926.
118
[Le2] H. Lebesgue, Sur lintegration des fonctions discontinues, Ann. Sci.
Ecole Norm. Sup. (3) 27
(1910), 361450.
[MS] Mazya V., Shaposhnikova T. Jaques Hadamard. A Universal Mathematician. History of Math-
ematics 14. AMS and LMS, 1998
[Ni] O. Nikodym, Sur une genesalization des integrales de M. J. Radon, Fund. Math. 15 (1930),
131179.
[P] A. Pietsch, History of Banach Spaces and Linear Operators, Birkhauser, Boston, 2007.
[Ra] J. Radon, Theorie und Anwendungen der Theorie der absolut additiven Mengenfunktionen,
Sitzungsber. Kaiserl. (
Osterreich) Akad. Wiss. Math.- Nat. Kl. , Abtteilung IIa 122, (1913),
12951438.
[Ro] H. L. Royden , Real Analysis, Macmillan Publishing Company, New York, 1988.
[R1] W. Rudin, Principles of Mathematical Analysis, Third Edition, McGraw-Hill International
Editions, Singapore, 1989.
[R2] W. Rudin, Analisi Reale e Complessa, Bollati Boringhieri, Torino,1991. An English version
is also available: W. Rudin, Real and Complex Analysis, Third Edition, McGraw-Hill Book
Co.,New York,1987.
[So] S.L. Sobolev, On a theorem of functional analysis (English) , Amer. Math. Soc. Transl. (2) 34
(1963), 3968.
[Vi] G. Vitali, Sulle funzioni integrabili, Atti Accad. Sci. Torino Cl. Sci. Fis. Mat. 40 (1904-05),
10211034.
119
INFORMATION ABOUT SOME QUOTED MATHEMATICIANS
Biographical and scientic information more detailed may nd at the website
http:///www-history.mcs.st-andrews.ac.uk/
ALAOGLU Leonidas (1914, Red Deer, Alberta, Canada- 1981): Alaoglu was a
Canadian-American mathematician, most famous for his widely-cited result called
Alaoglus theorem on the weak-star compactness of the closed unit ball in the dual
of a normed space, also known as the Banach-Bourbaki-Alaoglu theorem.
ASCOLI Giulio ( 1843, Trieste, Italy - 1896, Milan, Italy): He made contributions
to the theory of functions of a real variable and to Fourier series.
ARZELA Cesare (1847, Santo Stefano di Magra, La Spezia, Italy - 1912, Santo
Stefano di Magra): He conducted a deep research in the eld of functional theory.
BAIRE Rene-Louis (1874, Paris, France- 1932, Chambery, France): Baire worked
on the theory of functions and the concept of a limit. He is best known for the Baire
category theorem, a result he proved in his 1899 thesis.
BANACH Stefan (1892, Krakow, Austria-Hungary (now Poland) - 1945 in Lvov,
(now Ukraine)): Banach founded modern functional analysis and made major con-
tributions to the theory of topological vector spaces. In addition, he contributed to
measure theory, integration, and orthogonal series.
BORSUK Karol (1905, Warsaw,Poland - 1982, Warsaw,Poland) He was mainly
interested in topology. His topological and geometric conjectures and themes stimu-
lated research for more than half a century.
BOREL Emil F.E.J. (1871, Saint Arique, Aveyron, Midi-Pyrenees, France -
1956, Paris): Borel created the rst eective theory of the measure of sets of points,
beginning of the modern theory of functions of a real variable.
BOURBAKI Nicolas: Nicolas Bourbaki is the pseudonym of a group of (mainly)
French mathematicians who published an authoritative account of contemporary
mathematics.
CANTOR George F.L.P. (1845, St Petersburg, Russia - 1918, Halle, Germany):
Cantor founded the set theory and introduced the concept of innite numbers with
his discovery of cardinal numbers. He also advanced the study of trigonometric series.
CARATH
ECH