Anda di halaman 1dari 60

Contents

Contents 1
1 Metric Space Analysis. 3
1.1 Connectedness. . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Compactness in metric spaces. . . . . . . . . . . . . . . . . . 4
1.3 Functions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4 Riemann Integration. . . . . . . . . . . . . . . . . . . . . . . 34
1.5 Additional Topics . . . . . . . . . . . . . . . . . . . . . . . . 43
2 Multivariable Calculus 55
2.1 Dierentiation. . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.2 Implicit and Inverse Function. . . . . . . . . . . . . . . . . . 57
1
C
h
a
p
t
e
r
1
Metric Space Analysis.
1.1 Connectedness.
1.1 Connected metric space. Two subsets A, B of a metric space are
separated if both A B and A B are empty. A metric space is connected
if it cannot be written as union of two nonempty separated sets.
Connectedness can equivalently be dened in the following way: M is
a connected metric space if it contains a proper subset A (proper means
A M) which is both open and closed in M (we call such subsets
clopen). For, if A is proper and clopen in M, A
c
is also proper clopen, and
we write M = A A
c
, that is, as union of nonempty separated sets.
1.2 Theorem. Let M be a connected metric space and f : M N con-
tinuous and onto. Then N is connected.
Proof. Suppose not. Then N = A A
c
, where A is a proper clopen subset
of N. By continuity, the preimage of A, f
1
(A) is then nonempty (by the
fact that f is onto), open and closed in M. Hence M contains a proper
clopen subset, which is a contradiction.
Alternative proof: suppose N = A B with A, B nonempty and sepa-
rated, and set C = f
1
(A), D = f
1
(B). Then C D = M. Besides, since
A A, C f
1
(A), and the latter set is closed, thus, f(C) A. This
implies CD = , since f(D) = B A
c
. Similarly DC = , which leads
to a contradiction.
3
4 CHAPTER 1. METRIC SPACE ANALYSIS.
1.3 Exercise. Let S, T be connected subsets of a metric space M having
a point in common. Prove that S T is connected.
Proof. Suppose not, then there exist closed (in S T) nonempty separated
sets A, B such that AB = S T. Dene A
S
= AS, B
S
= B S. Since
S A B, we have S = A
S
B
S
with A
S
and B
S
separated. Then, at
least one of A
S
and B
S
is empty. Suppose B
S
= without loss of generality,
which leaves A
S
= A = S. We then have
T (S T)S = (A B)A = B = T S A B =
which yields a contradiction.
1.4 Exercise. Let R
22
be the metric space of 2 2 matrices endowed
with the norm
_
_
_
_
_
a b
c d
__
_
_
_
= (a
2
+ b
2
+ c
2
+ d
2
)
1/2
.
Let X R
22
be the subset of all invertible 22 matrices. Is X connected?
Solution. We may identify R
22
with R
4
and write X as the complement
of the set
Y = (a, b, c, d) R
4
: F(a, b, c, d) = 0, F(a, b, c, d) = ac bd.
We compute DF as DF(a, b, c, d) = (c, d, a, b) , which has rank 1 if
(a, b, c, d) is dierent from 0. Thus, the 0-level set of F, Y , is locally equiva-
lent to a 3-hyperplane at every point except the origin. Since a 3-hyperplane
disconnects R
4
, X is disconnected.
1.2 Compactness in metric spaces.
1.5 Sequential compactness. Let M, d be a metric space. K M is
said to be (sequentially) compact if every sequence (x
n
)
nN
of points of K
has a subsequence converging to some point of K.
1.6 Theorem. Every compact set is closed and bounded.
Proof. Let K M compact, and (x
n
)
nN
a sequence of points of K con-
verging to x M. Since (x
n
) has a convergent subsequence (x
n
k
) to some
x K, uniqueness of limit implies x = x K. Hence, K is closed.
1.2. COMPACTNESS IN METRIC SPACES. 5
Suppose that K is not bounded. Then, given p M there exists a
sequence (x
n
)
nN
K such that d(x
n
, p) n, which means that no subse-
quence of (x
n
) is bounded. But (x
n
) must possess a convergent subsequence
(x
n
k
), which need be bounded. Contradiction.
1.7 Theorem. The closed interval [a, b] R is compact.
Proof. Let (x
n
) be a sequence in [a, b] and dene the set
C = x [a, b] : x x
n
for all suciently large n.
The set C is nonempty (it contains a) and bounded above by b, hence,
there exists its l.u.b c. We now show the existence of a subsequence of
(x
n
) converging to c. Suppose not, then there exists > 0 such that
x
n
(c , c +) only nitely often. Thus, c + belongs to C, contrary to
c being an upper bound. Contradiction.
1.8 Theorem. The cartesian product of compact subsets K
1
M
1
, K
2

M
2
is compact in M
1
M
2
. In general, the cartesian product of a nite
number of compact sets is compact.
Proof. Let (a
n
, b
n
) be a sequence in K
1
K
2
. Then there exists a con-
vergent in K
1
subsequence (a
n
). Consider now the subsequence (a
n
, b
n
).
Then (b
n
) also has a convergent in K
2
subsequence (b
n
). The so-obtained
subsequence (a
n
, b
n
) converges in K
1
K
2
. To obtain the general case,
apply induction.
1.9 Heine-Borel Theorem. A R
n
is compact if and only if it is closed
and bounded.
Proof. Theorems 1.7 and 1.8 imply that n-cells (i.e. plurirectangles) are
compact in R
n
. If A is bounded in R
n
, it is contained in a k-cell, and closed
subsets of compact sets are compact.
1.10 Open cover. Let A M. A collection O = O

of open sets of M
is said to be an open cover for A if A

.
1.11 Covering compactness. K M is said to be covering compact if
every open cover of K possesses a nite subcover. That is, if O

are open
sets,
K
_

=
1
, . . .
n
: K O

1
. . . O

n
.
6 CHAPTER 1. METRIC SPACE ANALYSIS.
1.12 Theorem. Covering compactness implies sequential compactness.
Proof. Let K M be covering compact and (x
n
) a sequence of points of
K. Suppose (x
n
) does not possess any convergent subsequence. Hence, for
every a K there exists
a
> 0 such that d(x
n
, a) <
a
only for nitely
many n. The collection O = B

a
(a)
aK
is an open cover of K, thus, it
must possess a nite subcover B
i
= B

i
(a
i
), i = 1, . . . , n. Each B
i
contains
a nite collection of x
n
, hence x
n
possesses nitely many terms. A sequence
with nite range necessarily possesses a convergent subsequence, which is
then a contradiction.
1.13 Lebesgue number. Let O be an open cover of A M. O is said
to possess positive Lebesgue number if there exist > 0 such that for each
a A, there exists O O such that B

(a) O. Such a is called Lebesgue


number of O. If no > 0 possesses this property, O has Lebesgue number
0.
Notice that the choice of the open set O may depend on a, but is not
allowed to.
1.14 Example. Consider the set Q = Q(0, 1). We show that Q possesses
open covers with zero Lebesgue number. Suppose > 0 is a Lebesgue
number for the cover O = O

. This means that


_
qQ
B

(q)
_

.
Now, for every x (0, 1) there exists q Q with [x q[ < , that is
(0, 1)
_
qQ
B

(q),
which means that O covers (0, 1). We have just showed that every open
cover of Q with positive Lebesgue number also covers (0, 1). However, there
are open covers of Q which do not cover (0, 1): for example, given 0 < < 1,
O = O
m,n
, O
m,n
=
_
x R : [x m/n[ < 2
mn
_
,
for

OO
diam(O) = , while for every open cover of (0, 1) the sum of the
diameters of the open sets in the cover must be at least 1.
1.15 Lemma. If K M is sequentially compact, and O is an open cover
of K, O has positive Lebesgue number.
1.2. COMPACTNESS IN METRIC SPACES. 7
Proof. Let O be an open cover of K. Argue by contradiction: suppose
that, for each > 0, there exist a K such that B

(a) is not a subset of


any O O. Call (a
n
) the sequence of points of K obtained by choosing
= 1/n: (a
n
) must possess a subsequence (a
n
) converging to some b K.
Choose r small enough such that B
r
(b) is contained in some O O. Then,
for n

suciently large to have 1/n

< r/2 and d(a


n
, b) < r/2
d(a
n
, q) < 1/n

= d(b, q) d(a
n
, b) + d(a
n
, q) < r,
that is,
B
1/n
(a
n
) B
r
(b) O,
which gives a contradiction.
1.16 Theorem. Every sequentially compact subset K M is covering
compact.
Proof. Let O be an open cover of K. We prove the theorem by constructing
a nite subcover. By the preceeding lemma, O has positive Lebesgue num-
ber . Choose now a
1
K, and O
1
O such that B

(a
1
) O
1
. If K O
1
,
then O
1
is the sought nite subcover. Otherwise, choose a
2
KO
1
and
O
2
O such that B

(a
2
) O
2
. Either K O
1
O
2
, or else we can con-
tinue, producing a sequence (a
n
) K and a sequence O
n
in O such that
B

(a
n
) O
n
, and a
n+1
K
n
i=1
O
i
.
The sequence (a
n
) has a subsequence (a
n
) converging to some b K.
For n

greater than some integer N, d(a


n
, b) < , that is, b B

(a
n
) O
n
.
By our previous construction, a
n
/ O
n
for n

> N, which contradicts the


convergence. The reached contradiction implies that the process of choosing
points a
n
terminates after a nite number of steps, which yields the nite
subcover of K.
1.17 Example. Let M be an innite set with the discrete metric. The sets
m, with m M, are open sets, and the collection O = m : m M
is an open cover of M which does not reduce to a nite subcover (thus, M
is not compact). However, every open cover of M has positive Lebesgue
number. Indeed, each m need be contained in at least one member O of
the open cover, and, for every 0 < < 1, m O = B

(m) O, since
B

(m) = m.
1.18 Total boundedness. A subset A M of a metric space is totally
bounded if for every > 0 there exists a nite open covering of A of balls
B

(x
i
) : x
i
A
n
i=1
. It is immediate to see that any compact set K is
8 CHAPTER 1. METRIC SPACE ANALYSIS.
totally bounded, since, given > 0, B

(x) : x K is an open cover of K,


therefore, it admits a nite subcover.
1.19 Theorem. Let M be a complete metric space. Then, A M is
compact if and only if it is closed and totally bounded.
1.20 Theorem. Every compact metric space is separable (i.e. has a count-
able dense subset).
Proof. For every n N, let B
1/n
(x
n
i
) : i = 1, . . . , k
n
a nite open covering
of M with balls of radius 1/n. The set
P =
_
nN
x
n
i
: i = 1, . . . , k
n

is countable and dense in M.


Direct proofs of covering compactness facts.
This sections contains the basic theorems about compactness in metric
spaces, proved by means of the covering version of compactness. The proofs
below continue to hold in general topological spaces.
1.21 Lemma. Let I
n
be a nested (that is, I
n
I
n+1
) sequence of nonempty
intervals. Then
nN
I
n
is nonempty.
Proof. If I
n
= [a
n
, b
n
], let E = a
n
. The set E is nonempty and bounded
above by b
1
, hence x = l.u.bE exists. Then, a
n
x b
n
for each n, so
x I
n
for each n, hence x I
n
.
Proof of Theorem 1.7. Let O be an open cover of I
0
= [a, b], and suppose
it does not reduce to a nite subcover. Set c = (a + b)/2; then, I
0
=
[a, c] [c, b], and at least one of these two sets, say I
1
= [a, c
1
], cannot
be covered by a nite subcollection of open sets in O (otherwise also I
0
would be covered). Set c
2
= (a + c
1
)/2, and write I
1
= [a, c
2
] [c
2
, c
1
] and
iterate the procedure, obtaining a nested sequence of nonempty intervals I
n
,
neither of which can be cover by a nite subcollection of sets in O. Besides,
diam(I
n
) (b a)2
n
.
By the previous lemma, the intersection I
n
contains at least a point
x, which must belong to some open set O O. Being O open, B
r
(x) is
contained in O for r suciently small. However, choosing n large enough
so that (b a)2
n
< r, since x I
n
, we must have I
n
B
r
(x). This leads
to a contradiction.
1.2. COMPACTNESS IN METRIC SPACES. 9
1.22 Theorem. Let K

be a collection of compact sets such that the


intersection of every nite subcollection K

1
, . . . , K

n
is nonempty (K

is said to have the nite intersection property). Then

is nonempty.
Proof. Suppose that

= . Taking complements, and setting G

=
K
c

, we get that the collection of open sets G

covers K
1
. Being K
1
compact, there is a nite subcover G

1
, . . . , G

n
, that is
K
1
G

1
. . . G

n
K
1
K

1
. . . K

n
= ,
which is a contradiction.
1.23 Theorem. Let M be a metric space, and suppose that every collection
of closed subsets of M which has the nite intersection property has also
nonempty grand intersection. Then M is compact.
Proof. Let O = O

be an open cover of M. Then

O
c

= ,
which implies that the collection of closed sets O
c

does not have the nite


intersection property. Hence, for some
1
, . . . ,
n
,
O
c

1
. . . O
c

n
= O

1
. . . O

n
= M,
that is, O possesses a nite subcover.
1.24 Theorem. Let M, N be metric spaces and f : M N continuous.
Then f(A) is compact for each A M compact.
Proof. Let O = O

be an open cover of f(A). By continuity,


G

= f
1
(O

)
is open. Moreover, for each x A, y = f(x) f(A), thus y O

for some
, that is x G

. Thus, G

is an open cover of A, hence it must possess


a nite subcover G

1
, . . . , G

n
. To prove that O

1
, . . . , O

n
covers f(A), it
is enough to observe that y f(A) implies y = f(x) for some x belonging
to some G

i
, that is y f(G

i
), that is y O

i
. We have then found a
nite subcover.
1.25 Theorem. Let M be a compact metric space, and f : M N
continuous. Then f is uniformly continuous.
10 CHAPTER 1. METRIC SPACE ANALYSIS.
Proof. Denote with d the distance in M, with D the distance in N. Let
> 0. For each x M there exists
x
> 0 such that
d(x, y) < 2
x
= D(f(x), f(y)) <

2
.
The collection of open sets B

x
(x) : x M covers M, hence it must
reduce to a nite subcover B

i
(x
i
)
n
i=1
, with
i
=
x
i
. Set = min
i

i
> 0.
Choose x, y M such that d(x, y) < and x
i
such that d(x, x
i
) <
i
. Then
d(y, x
i
) < 2
i
, so that
D(f(x), f(y)) D(f(x), f(x
i
)) + D(f(x
i
), f(y)) <

2
+

2
= ,
which is uniform continuity.
1.26 Counterexample. It is not true that every function f between two
metric spaces M, N such that f
1
(K) compact in M for every compact K
of N is continuous. A counterexample is the real function on R given by
f(x) =
_

_
x [x[ 1
x
1
[x[ < 1, x ,= 0
0 x = 0;
computing the counterimage of every closed interval will do.
Equivalent denitions of compactness.
1.27 Theorem. Suppose that every continuous function f : M R is
bounded. Then M is compact.
Proof. Suppose not. Then M contains a sequence of distinct points (p
n
)
possessing no convergent subsequence. This easily implies that for each n,
there exists
n
> 0 such that
d(p
n
, p
m
) >
n
, m ,= n.
Call B
n
the open ball of radius
n
=
n
/3 centered at p
n
. Then, each
sequence (q
n
) such that q
n
B
n
for each n cannot have a convergent sub-
sequence. For, if it were so,
d(p
n
k
, p
n
h
) d(p
n
k
, q
n
k
) + d(q
n
k
, q
n
h
) + d(p
n
h
, q
n
h
)
n
k
,
1.2. COMPACTNESS IN METRIC SPACES. 11
for n
h
, n
k
suciently large to have d(q
n
h
, q
n
k
) <
n
k
/3. Now, dene the
function
f(p) =
_
_
_
n(r
n
d(p, p
n
))
r
n
p B
n
,
0 otherwise.
If a sequence (q
n
) converges to q B
k
, then, by the previous considerations,
all but nitely many q
n
lie outside B
k
, and f is uniformly continuous on
B
k
, thus f(q
n
) f(q).
If (q
n
) converges to q M

= M
n
B
n
, on which f 0, then, either all
but nitely many q
n
lie in M

, which easily gives f(q


n
) f(q), or all but
nitely many q
n
lie in B
k
M

and q B
k
, which implies f
(
q
n
) 0 = f(q).
Thus, f is continuous on M, but, since f(p
n
) = n, fails to be bounded.
1.28 Theorem. The following assertions hold.
(a) If every continuous bounded function f : M R achieves a maximum,
then M is compact.
(b) If every continuous function f : M R has compact image, then M is
compact.
(c) If every nested decreasing sequence of nonempty closed subsets of M has
nonempty intersection, then M is compact.
Proof. The proofs of (a) and (b) follow the same steps of the proof of
Theorem 1.27, with a dierent choice of f, namely
f(p) =
_
_
_
(n 1)(r
n
d(p, p
n
))
nr
n
p B
n
,
0 otherwise.
We now turn to the proof of (c). Suppose M is noncompact. Thus, there
exist a sequence (p
n
), with no convergent subsequence. For n 1, dene
the range sets
K
n
= p M : p = p
k
for some k n.
It is obvious that K
n
is nonempty for each n and that K
n
K
n+1
. More-
over, each K
n
is closed, for it contains all its limit points. To see this,
suppose p is a limit point for K
n
. Then, for each k N, there exists
q
k
K
n
such that d(q
k
, p) < 1/k. The sequence (q
k
) is a convergent subse-
quence of (p
n
), thus, it must have nite range (otherwise p
n
would possess
a convergent subsequence), that is q
k
= p for all k suciently large, hence
12 CHAPTER 1. METRIC SPACE ANALYSIS.
p K
n
. Using the fact that the intersection K
n
is nonempty, there exists
p M such that p K
n
for each n. This means that p
n
= p for innitely
many n, hence, (p
n
) has a subsequence converging to p. Contradiction.
1.29 Absolute properties. Let M be a metric space. We say that
h : M N is an embedding of M if it is an homeomorphism from M
onto h(M). A property of M that holds for each embedded copy of M
is said to be absolute. For example, as a consequence of the fact that
every compact set is closed and bounded and compactness is preserved by
continuous functions, every compact set is absolutely closed and absolutely
bounded. In this mood, the following two theorems provide other equivalent
denition for compactness.
1.30 Theorem. If M is absolutely bounded, then M is compact.
Proof. If M is noncompact, then there is an unbounded continuous function
f : M R (see Theorem 1.28a). Dene the map
p M F(p) = (p, f(p)) M R,
which is easily seen to be continuous and injective on M, and onto the set
N = F(M). The inverse function
(q, y) F
1
(q, y) = q,
is a projection, hence continuous. Thus, N is an embedding of M in MR.
Being N unbounded, M is not absolutely bounded.
1.31 Cone over a metric space. Let M be a metric space whose metric
d is bounded by 1. Given p
0
M, the cone over M is the metric space
C(M) = p
0
(0, 1] M,
with metric
((r, p), (s, q)) = [r s[ + minr, sd(p, q), ((r, p), p
0
) = r.
It is easy to verify that is a metric on M.
1.32 Theorem. If M is absolutely closed, then M is compact.
Proof. If M is noncompact, then there is a continuous function f : M R
whose range is (0, 1] and which does not attain its minimum (see Theorem
1.2. COMPACTNESS IN METRIC SPACES. 13
1.28b). This means that we can construct a sequence p
n
of points of M
which possesses no convergent subsequence and f(p
n
) 0. Dene the map
p M G(p) = (f(p), p) C(M),
which is easily seen to be continuous and injective on M, and onto the set
N = F(M). The inverse function
(y, q) G
1
(y, q) = q,
is a projection, hence continuous. Thus, N is an embedding of M in C; how-
ever, N is not closed. To see this, we observe that the sequence (f(p
n
), p
n
)
converges to the vertex of the cone p
0
, which is not a member of N.
1.33 Remark. If every function f : M R is uniformly continuous,
M need not be compact. For example, if M is an innite set with the
discrete metric, M is not compact, while every function on M is uniformly
continuous (for each > 0, any (0, 1) is sucient to guarantee [f(x)
f(y)[ < if d(x, y) < ). However, if we additionally require M to be
a subset of a compact metric space K, the condition is sucient, for it
entails the closedness of M. To see this, suppose M contains a sequence
(p
n
) converging to p and p / M. Then, the function f(q) = (d(p, q))
1
is
continuous on M but not uniformly continuous (evaluate what happens on
the sequence (p
n
)).
1.34 Counterexample. Even if the metric space M is bounded (that
is, d(x, y) 1 for all x, y M) there are uniformly continuous functions
from M to R which are unbounded. Let M = N with the discrete metric.
The function f(n) = n is evidently unbounded (in R). However, for every
> 0, we clearly have [f(x)f(y)[ < for d(x, y) < 1/2, thus f is uniformly
continuous.
Exercises.
1.35 Exercise. Let (X, d) be a complete metric space and F
n
a (decreas-
ingly) nested sequence of nonempty closed subsets of X, with the property
that each F
n
can be covered by a nite number of balls of radius 1/n. For
each n, choose x
n
F
n
. Show that (x
n
) has a convergent subsequence to
some x
n
F
n
.
Solution. Argue by contradiction. Suppose (x
n
) has no convergent subse-
quence. Being X complete, this is equivalent to (x
n
) possessing no Cauchy
14 CHAPTER 1. METRIC SPACE ANALYSIS.
subsequences, which means that there is > 0 such that d(x
n
, x
m
) > for
all but nitely many m, n. Fix n such that > 1/n. Then, F
n
can be cov-
ered by a nite number of balls of radius . Since (x
m
)
mn
is contained in
F
n
, there is at least one ball of radius which contains innitely many x
m
,
that is to say, d(x
m
, x
k
) < for innitely many m, k . Contradiction.
The above reasoning shows that (x
n
) has a subsequence (x
n
k
) converging
to some x X. Since F
n
are closed sets, and (x
n
k
)
n
k
n
is contained in F
n
,
x F
n
for each n, so that x
n
F
n
.
1.36 Exercise. Suppose K is a compact subset of a metric space M and
(x
n
) a sequence of points of K such that every convergent subsequence
converges to the same limit x. Show that x
n
x. Then, give an example
to show that, if K is not compact, the property need not hold.
Solution. Obviously, if (x
n
) converges, it must converge to x. Suppose (x
n
)
does not converge to x. This means that, for some > 0, d(x, x
n
) > for
innitely many n. Thus, the subsequence
y
k
: y
k
= x
n
k
and d(x
n
k
, x) >
contains innitely many terms. Since K is compact, y
k
must possess a
convergent subsequence (z

= y
k

); the sequence z

is a convergent subse-
quence of (x
n
), hence, it converges to x, which implies d(z

, x) < for all


suciently large. Contradiction.
For a counterexample, take a convergent sequence (x
n
) R and con-
struct the sequence (z
n
) R as
z
n
=
_
x
k
n = 2k,
n n = 2k + 1.
The only convergent subsequences of z
n
contain a nite number of terms
z
2k+1
, thus they keep converging to x, while z
n
obviously does not converge.
1.37 Exercise. Let f : M N be a continuous injective function be-
tween metric spaces, with M compact. Prove that the function
g : f(M) M, g(f(x)) = x,
is continuous.
Solution 1. We prove that, for every A open in M, g
1
(A) is an open set in
f(M), which is continuity of g. Being f injective, g
1
(A) = f(A), so that
1.2. COMPACTNESS IN METRIC SPACES. 15
the above claim is equivalent to have f(A) open in f(M) for each A open in
M. Thus, let A be an open set in M. Its complement A
c
is a closed subset
of M, hence compact, so that f(A
c
) is compact in f(M) (hence closed).
But, since f is a bijection from M to f(M).
f(A
c
) = f(x) : x A
c
= f(M) f(x) : x A = (f(A))
c
,
so that f(A) is open in f(M), its complement being a closed set.
Solution 2. Consider a sequence of points y
n
f(M) converging to y
f(M). We have y
n
= f(x
n
), y = f(x), for some x
n
M, x M. If x
n
k
is
a subsequence of x
n
converging to some x M, then, by continuity
f( x) = lim
k
f(x
n
k
) = lim
k
y
n
k
= y,
so that x = x is forced by injectivity of f. Thus, every convergent subse-
quence of x
n
converges to the same x M, and, being M compact, we may
use Exercise 1.36 to infer that the whole sequence x
n
= g(y
n
) converges to
x = g(y). Hence,
y
n
y f(M) = g(y
n
) g(y) M,
which proves continuity of g.
1.38 Exercise. If M is a compact metric space and A is dense in M,
show that for each > 0 there exist a nite subset a
i

n
i=1
of A such that
M =
j
B

(a
j
).
Solution. Fix > 0. By density, each x M belongs to B

(a) for some


a A, so that B

(a) : a A is an open cover of M. Using compactness,


extract a nite subcover.
1.39 Exercise. Suppose f : R
m
R possesses the two following proper-
ties:
(a) f(K) is compact for each K R
m
compact;
(b) for any nested decreasing sequence of compact sets (K
n
)
f
_

K
n
_
=

f(K
n
).
Prove that f is continuous.
Solution. Let x R
m
, > 0 and, for each n, call K
n
the closed ball of
radius 1/n centered at x. Clearly, the sets K
n
are nested and compact, and

n
K
n
= x. Thus,
f(x) =

n
f(K
n
).
16 CHAPTER 1. METRIC SPACE ANALYSIS.
Now, call B the open ball of radius centered at f(x). The sets W
n
=
B
c
f(K
n
) are closed subsets of the compact sets f(K
n
), hence compact,
and W
n
W
n+1
. Moreover,
n
W
n
is empty. This means that there exists
N such that W
n
is empty for all n N, that is
[y x[ < 1/N = [f(y) f(x)[ < ,
which shows f is continuous at x.
An alternative proof relying on Exercise 1.36 is the following. Let x
n
be
a sequence converging to x. Then, the sets
K
n
= y : y = x
k
for some k n x,
are nested compact sets. Besides,
n
K
n
= x, for if y K
n
for each n,
either y = x or y = x
k
for innitely many n. The latter claim implies (x
k
)
possesses a subsequence converging to y, which forces again y = x. Thus,
we conclude that

n
f(K
n
) = f(x),
and, reasoning as above, we get that every convergent subsequence of the
sequence (f(x
n
)) converges to f(x). Being f(K
1
) compact, we use Exercise
1.36 to conclude that the whole sequence (f(x
n
)) converges to f(x).
1.40 Exercise. Let X R
m
be compact and f : X R continuous.
Show that, given > 0, there is a constant M such that, for all x, y X,
[f(y) f(x)[ M[y x[ + .
Solution. Being X compact, f is uniformly continuous on X, that is, for
each > 0 there exists > 0 such that
x, y X, [y x[ < = [f(y) f(x)[ < .
Suppose that the thesis is false. Then, for some > 0, for each n n there
exist x
n
, y
n
X such that
[f(y
n
) f(x
n
)[ n[y
n
x
n
[ + .
Since f is also uniformly bounded,
[f(y
n
) f(x
n
)[ 2 max
xK
[f(x)[ = Q,
combining the two inequalities we see that [y
n
x
n
[ (Q )/n. This
means that, for every > 0, there exists n large enough such that
[y
n
x
n
[
Q
n
< and [f(y
n
) f(x
n
)[ ,
1.2. COMPACTNESS IN METRIC SPACES. 17
which contradicts uniform continuity.
1.41 Exercise. Let f : R
2
R satisfying the following two properties:
(a) for each x
0
R
2
, y f(x
0
, y) is continuous and for each y
0
R
2
,
x f(x, y
0
) is continuous;
(b) f maps compact sets into compact sets.
Show that f is continuous on R
2
.
Solution. We do not lose in generality in restricting to prove f continuous
at the origin. By translation, we can assume that f(0, 0) = 0. Argue by
contradiction, assuming that there exists > 0 so that [f(x
n
, y
n
)[ for
some sequence (x
n
, y
n
) converging to (0, 0). By continuity of x f(x, 0),
there exists N large enough such that n N implies [f(x
n
, 0)[ /2. By
continuity of y f(x
n
, y), there exists 0 < y

n
< y
n
such that, applying the
intermediate value theorem,
[f(x
n
, y

n
)[ =
n
n + 1
.
Now, the set B = x
n
, y
n
(0, 0) is compact, since (x
n
, y

n
) (0, 0),
so that f(B) need be compact. However, f(B) = n/(n + 1) 0, so
compactness fails, being a limit point of f(B) which does not belong to
. Contradiction.
1.42 Exercise. Let M be a compact metric space and f : M M an
isometry (that is, d(f(p), f(q)) = d(p, q) for every p, q M). Prove that
f(M) = M.
Solution. Choose q M and suppose q / f(M). Being f(M) compact
(hence closed), there exists < 0 such that B
2
(q) f(M) = . Let n be
the least number of open -balls necessary to cover M (by compactness,
there is such an n). If B is an -ball covering q, B cannot intersect f(M):
hence, n 1 -balls are enough to cover f(M). The inverse images of such
balls are n 1 open sets of diameter covering M, against the minimality
of n. Contradiction.
1.43 Exercise. Let f : R
n
R
k
be a continuous mapping. Show that, if
S R
n
is a bounded set, then f(S) is a bounded subset of R
k
.
Solution. Suppose not, so that there is a sequence of points x
n
S such
that [f(x
n
)[ > n for all n. The set S, the closure of S, is a compact set,
so that x
n
has a convergent subsequence x
n
k
to some x S. Continuity
then implies [f(x)[ > n
k
for every k, with n
k
. Being f(x) R
k
well
dened, this leads to a contradiction.
18 CHAPTER 1. METRIC SPACE ANALYSIS.
1.3 Functions.
1.44 Oscillation. Let f : M N a function between metric spaces.
Dene the oscillation of f at x M as
osc
x
(f) = inf
>0
diam[f(B

(x))] ,
where B

(x) is the open ball of radius centered at x. Notice that, if x is


a continuity point of f, osc
x
(f) is zero.
1.45 Proposition. Let f : M N a function between metric spaces.
The sets
D

= x M : osc
x
(f) , > 0,
are closed subsets of M, and it follows that the discontinuity set of f is an
F

-set.
Proof. For a given > 0, call (x
n
) a sequence of points of D

converging
to x M. Fix , > 0 and x
n
such that d(x
n
, x) < . We can choose > 0
such that B

(x
n
) is contained in B

(x) and diam(f(B

(x
n
))) > . It
follows that
diam(f(B

(x))) > .
Taking the inmum on both sides, being arbitrary, we conclude that
osc
x
(f) , that is, x D

. To get the remaining claim, it is enough to


observe that the set of discontinuities D may be written as
n
D
1/n
.
1.46 Baires Theorem. Let M be a complete metric space. If M is the
countable union of closed subsets C
n

nN
, then some C
n
has nonempty
interior.
1.47 Theorem. Let f be the pointwise limit of a sequence of continuous
functions f
n
: [a, b] R (f is said to be of Baire class 1). Then, f is
continuous on a dense set of points.
Proof. For a generic > 0, we prove that the closed set D

relative to f has
nonempty interior. After that, the thesis is achieved by Baires Theorem.
Suppose D

contains some interval (, ) [a, b]. Write R as the count-


able union of intervals (c

, d

) each of length strictly smaller than . Set


H

= f
1
((c

, d

)). Then

= [a, b], and, for each , D

has empty
interior. To see this, observe if I is an interval in D

, then f(I) has diam-


eter greater than or equal to , hence, no (c

, d

) can contain f(I), hence


1.3. FUNCTIONS. 19
H

cannot contain I. Now, dene the sets


F
mn
= f
1
n
_
[c

+ 1/m, d

1/m]
_
, E
mN
=

nN
F
mn
.
The sets F
mn
are preimages of closed sets through continuous functions,
hence closed; the sets E
mN
are intersections of closed sets, hence closed.
Besides,
H

=
_
m,NN
E
mN
.
To see this, choose x H

. Then f(x) (c

, d

), thus,
f(x) [c

+ 1/m, d

1/m]
for some m N. Since f(x) is the pointwise limit of f
n
(x), there exists
N N f
n
(x) [c

+ 1/m, d

1/m] for all n N. Hence, x E


mN
for
some m, N N, that is, x
m,N
E
mN
. The reverse inclusion uses the
same ideas and it is easier. In the end, we have obtained
[a, b] =
_
N
H

=
_
,m,NN
E
mN
,
which implies, taking intersections,
[, ] =
_
,m,NN
(E
mN
[, ]).
By Baires Theorem, some E
mN
has nonempty interior, that is, there is
an open interval I (, ) contained in some E
mN
. This leads to a
contradiction, since I D
k
, and E
mN
D
k
has empty interior.
1.48 Example. Suppose g : [a, b] R is dierentiable. Then g

is of
Baire class 1, which means it is continuous on a dense subset of [a, b]. To
see this, extend g to [a 1, b + 1] preserving dierentiability (for example,
one may set g(x) = g(a) + g

(a)(x a) for x < a...), and observe that


g

is the pointwise limit on [a, b] of the sequence of continuous functions


f
n
(x) = n(g(x + 1/n) g(x).
1.49 Zero sets. A R is called a zero set if for each > 0, there exist
countably many intervals (a
n
, b
n
) R such that
A
_
nN
(a
n
, b
n
) and

nN
(b
n
a
n
) < .
The following properties are of easy verication.
20 CHAPTER 1. METRIC SPACE ANALYSIS.
(a) Any subset of a zero set is a zero set.
(b) Any nite set is a zero set.
(c) Any countable union of zero sets is a zero set.
1.50 Exercise. (Sards Theorem, dimension 1) Let f : R R be a
continuously dierentiable function. Prove that the set of critical values
/
f
= y R : y = f(x) and f

(x) = 0
is a zero set.
Solution. There is no loss of generality in proving the result for a function f
dened on a closed and bounded interval [a, b] (to obtain the general case,
it suces to write R as a countable union of bounded intervals and use
property (c) of zero sets). Let C be the (closed, by the continuity of f

) set
of critical points of f. We start by noting that f

is uniformly continuous
on [a, b], hence, there exists a nondecreasing function : R
+
R
+
such
that
lim
s0
+
(s) = 0 and [f

(s) f

(t)[ ([s t[).


Let x C. For any z [a, b], by Lagranges Theorem we have
f(x) f(z) = f

()(x z), 0 < [x [ < [x z[,


so that, recalling that f

(x) = 0,
(1.1) [f(x) f(z)[ [f

()[[x z[ [x z[([x z[).


Now, x n N and call = (ba)/n, and partition [a, b] into n subintervals
I
j
of length . Obviously,
C A :=
_
j:I
j
C=
I
j
,
so that f(C) f(A). If I
j
C ,= , we have, using (1.1),
diam[f(I
j
)] < ().
Summing over j, we get
diam[f(C)] diam[f(A)] n() = (b a)().
As goes to zero when n , the proof is nished.
1.3. FUNCTIONS. 21
Uniform convergence.
1.51 The space ((M). Let M be a metric space. The space ((M) of
real-valued continuous functions on M equipped with the norm
|f| = sup
xM
[f(x)[
is a Banach space. A similar denition is given when M is a topological
space.
1.52 Theorem. Let M be a separable metric space. Then R and ((M)
have the same cardinality.
Proof. Clearly, ((M) contains every constant function, which implies that
the cardinality of ((M) is greater than or equal to the cardinality of R. We
now show that equality holds. Call D a countable dense subset of M. Start
by noting that if f, g ((M) and f = g on D, then f = g on M. For, if
x M, there exists a sequence (x
n
) of points of D converging to x and, by
continuity,
f(x) = lim
n
f(x
n
) = lim
n
g(x
n
) = g(x).
Thus, any function f ((M) is uniquely determined by its values on D.
Now, the cardinality of real-valued functions on D is equal to the cardinality
of R (the cartesian products of countably many sets of real cardinality has
real cardinality), which implies the desired result.
1.53 Equicontinuity. Let M, d, N, D be metric spaces. A family of
functions T = f

: M N

is pointwise equicontinuous if for every


p M and > 0 there exists
p
> 0 such that
d(p, q) <
p
= D(f(p), f(q)) < .
If
p
= for every p M, T is uniformly equicontinuous.
1.54 Remark. If M is a compact metric space, every pointwise equicon-
tinuous family of functions is uniformly equicontinuous.
1.55 Counterexample. In the vein of Remark 1.3 above, we present an
example of a sequence of uniformly continuous functions (on a noncompact
set), which are pointwise equicontinuous but not uniformly equicontinuous.
Let f
n
: [0, ) R dened as
f
n
(x) =
_
0 x n,
n(x n) x > n,
22 CHAPTER 1. METRIC SPACE ANALYSIS.
It is immediate to see that each f
n
is uniformly continuous. For pointwise
equicontinuity, we x x (0, +) and note that [f
n
(x) f
n
(y)[ n[x y[
for n < x, while f
n
(x) is constant in a neighborhood of x when x n.
Summarizing, for each > 0, for every n N,
[x y[ < min/[x[, 1 = [f
n
(x) f
n
(y)[ .
The inequality above is sharp and thus shows that (f
n
) is not uniformly
equicontinuous.
The main compactness result for the space (M, R is the following
theorem.
1.56 Arzel`aAscolis Theorem. Let M be a compact metric space. Ev-
ery pointwise bounded equicontinuous sequence of functions in (M, R pos-
sesses a uniformly convergent subsequence.
1.57 Remark. If M is also connected, the boundedness requirement may
be relaxed to requiring that there is x
0
M such that f
n
(x
0
) is a bounded
sequence of real numbers. Indeed, this, together with equicontinuity, guar-
antees uniform boundedness. To prove the assertion, call C = sup
n
[f
n
(x
0
)[
and x > 0 in order to have, for all n N,
d(x, y) < = [f
n
(x) f
n
(y)[ < 1.
Choose x M. By compactness, M can be covered by a nite number
of open balls B
i
, i = 1, . . . , N, of radius , centered at some x
i
. Observe
that for each center x
i
there exists at least one x
j
with j ,= i such that
d(x
i
, x
j
) < , for otherwise B
i
and
j=i
B
j
would be separated and M could
be written as union of nonempty separated sets. Thus, we may rearrange
the x
i
in order to have x
0
B
1
, d(x
i
, x
i+1
) < , x B
k
for some k N.
Using the equicontinuity property,
[f
n
(x)[ [f
n
(x
0
)[ +[f
n
(x
0
) f
n
(x
1
)[ +[f
n
(x
1
) f
n
(x
2
)[
+ . . . +[f
n
(x
k
) f
n
(x)[
< C + k + 1 < C + N + 1,
which is uniform boundedness.
The following result establishes the converse of Ascolis Theorem and gives
a topological interpretation to it.
1.58 HeineBorel Theorem for (M, R. Let M be a compact metric
space. T (M, R is compact if and only if it is closed, bounded and
equicontinuous.
1.3. FUNCTIONS. 23
1.59 Example. Let (f
n
) is a sequence of real functions of a real variable
such that, for each compact subset K R, (f
n

K
) is pointwise bounded and
equicontinuous. Then (f
n
) has a pointwise convergent subsequence on all
of R. To see this, use AscoliArzel` as Theorem to construct a subsequence
f
n

of (f
n
) which converges uniformly on every (compact) set I
k
= [k, k]
with k N. Then, each x R belongs to some I
k
, so that the sequence
f
n

converges uniformly on I
k
, hence, f
n

(x) converges. However, uniform


convergence on all of R is not guaranteed. For a counterexample, the same
sequence of Counterexample 1.3 works.
1.60 Exercise. Let f
n
: [0, 1] R be a sequence of continuous functions
f
n
converging uniformly to f. Suppose that, for each n, f
n
vanishes at some
point. Does this imply f vanishes at some point in [0, 1]?
Solution. Yes. For each n, let x
n
be a point at which f
n
(x
n
) = 0. Then the
sequence (x
n
) possesses a subsequence x
n
k
converging to some x
0
[0, 1].
For > 0 xed, choose N
1
large enough so that, if n N
1
,
[f
n
(x) f(x)[ < /2, x [0, 1].
Being f
n
a sequence of continuous function uniformly convergent to f, the
limit function f is also continuous on [0, 1], so that we may x N
2
large
enough to have [f(x) f(x
n
k
)[ < /2, for n
k
N
2
. Thus
[f(x
0
)[ [f
n
k
(x
n
k
)[ +[f
n
k
(x
n
k
) f(x
n
k
)[ +[f(x
n
k
) f(x
0
)[ < ,
for n
k
maxN
1
, N
2
. Since is arbitrary, we see that f(x
0
) = 0.
1.61 Weierstrass Approximation Theorem. The space of real-valued
polynomials is dense in (([a, b]). That is, for every function f (([a, b]),
there is a sequence of real valued polynomials converging uniformly to f.
1.62 Exercise. Show that
(1.2) lim
n
_
1
0
f(x) sin(nx) dx = 0,
for each f (([a, b]).
Solution. Observe that, if f = 1,
_
1
0
f(x) sin(nx) dx =
_
1
0
sin(nx) dx =
1 cos(n)
n
0,
as n . Moreover, if f is a continuous function and F(x) =
_
x
0
f(t) dt,
integrating by parts we have

_
1
0
F(x) sin(nx) dx

F(1) cos(n)
n

+
1
n

_
1
0
f(x) cos(nx) dx

M
n
0,
24 CHAPTER 1. METRIC SPACE ANALYSIS.
as n , where M = 2 max [f[. The two facts above, together with
linearity, imply that (1.2) holds for every polynomial, and that convergence
is uniform on equibounded families of polynomials. For a generic f
(([a, b]), select a sequence of polynomials p
k
uniformly convergent to f
(thus, equibounded). Then, for each n,
_
1
0
f(x) sin(nx) dx = lim
k
_
1
0
p
k
(x) sin(nx) dx,
so that, passing to the limit n and using the above result for p
k
, we
get (1.2) for f.
1.63 Exercise. Let g
j
, h : [0, 1] R (j N) be continuous functions
such that
_
1
0
[g
j
(x)[ dx 1000 for all j and
lim
j
_
1
0
x
n
g
j
(x) dx =
_
1
0
x
n
h(x) dx,
for each n N. Prove that
(1.3) lim
j
_
1
0
f(x)g
j
(x) dx =
_
1
0
f(x)h(x) dx
for every continuous function f on [0, 1].
Solution. By linearity, (1.3) holds for every real polynomial on [0, 1]. For
f ((R), use Weierstrass approximation theorem to nd a sequence of
polynomials p
n
converging to f uniformly on [0, 1]. Then, for every > 0,
there is N large enough so that sup [f p
k
[ < / max1000, sup [h[ if
k N, and, for every j

_
1
0
(f(x) p
k
(x))g
j
(x) dx

sup
x[0,1]
[f(x) p
k
(x)[
_
1
0
[g
j
(x)[ dx < .
Now, being p
k
convergent, it is uniformly bounded. Hence, also exploiting
(1.3) with n = 0, we nd J such that, for every k, j J implies

_
1
0
(g
j
(x) h(x))p
k
(x) dx

sup
x[0,1]
[p
k
(x)[

_
1
0
g
j
(x) h(x) dx

< ,
and nally, collecting the obtained relations,

_
1
0
(g
j
h)f

_
1
0
(f p
k
)g
j

_
1
0
(g
j
h)p
k

_
1
0
(p
k
f)h

< 3,
1.3. FUNCTIONS. 25
for j, k large enough, so that the proof is complete.
1.64 Exercise. Let f : R R be continuous and such that the sequence
of functions f
n
dened as f
n
(x) = f(nx) is equicontinuous. What can be
inferred about f?
Solution. Fix > 0 and x, y R. By uniform continuity, there is > 0
such that
[ [ < = [f
n
() f
n
()[ < .
Fix n large enough to have [x y[ < n. Then, setting x = n, y = n, we
have [ [ < , and
[f(x) f(y)[ = [f(n) f(n)[ = [f
n
() f
n
()[ < .
Being arbitrary, we conclude that f(x) = f(y). Therefore, f is a constant
function.
1.65 Counterexample. We show that, if f
n
: R R is a sequence of
equicontinuous functions, such that f
n
(x) 0 as n for each x R,
f
n
may fail to converge uniformly to 0 on all of R. Let f
n
: R R be
dened as
f
n
(x) =
_
0 [x[ n,
[x n[ x > n.
Since [f
n
(x) f
n
(y)[ [x y[ for all x, y R and n N, f
n
is a uniformly
Lipschitz-continuous (hence, equicontinuous) sequence of functions, point-
wise convergent to 0. Thus f
n
0 uniformly on each compact subset of R.
However,
sup
xR
[f
n
(x)[ = ,
so that f
n
does not converge uniformly on R.
Exercises.
1.66 Exercise. Let (f
n
) be a sequence of real continuous functions on
[a, b], converging uniformly on (a, b) to f : (a, b) R. Show that f
n
converges uniformly on all of [a, b].
Solution. We start by proving that the sequence f
k
(a) converges
1
to some
R. Fix > 0, and N such that sup
(a,b)
[f
k
f
m
[ < if k, m N. Then,
use continuity of f
k
and f
m
to x an x (which is allowed to depend on k, m)
1
ti amo tantissimo!!!!!!!!!!!
26 CHAPTER 1. METRIC SPACE ANALYSIS.
suciently close to a to have [f
k
(x) f
k
(a)[ < , and [f
m
(x) f
k
(a)[ < .
Thus, for k, m N,
[f
k
(a) f
m
(a)[ < [f
k
(a) f
k
(x)[ +[f
k
(x) f
m
(x)[ +[f
m
(x) f
m
(a)[ < 3,
whence f
k
(a) is a Cauchy sequence, hence it converges to some R. We
may proceed in a similar fashion to nd that f
k
(b) converges to some .
Calling = f(a), = f(b) and letting m in the above formula guar-
antees that, for k N, sup
[a,b]
[f
k
f[ < , which is uniform convergence
on [a, b] of f
k
to f. As a byproduct, we automatically gain the continuity
of f on [a, b].
1.67 Exercise. Let (f
n
) be a sequence of real dierentiable functions on
R such that, for each x R, f
n
(x) g(x) R, and sup
R
[f

n
[ 1 for all n.
Show that g is a continuous function on R.
Solution. The sequence (f
n
) is equicontinuous, (indeed, it is uniformly-
Lipschitz continuous) and pointwise convergent to g, so that, by Ascolis
Theorem, f
n
g uniformly on each closed interval [a, b] R. Thus g is
continuous on every closed interval [a, b], being the uniform limit of contin-
uous functions. This yields the continuity on all of R
1.68 Exercise. Let (f
n
) be a sequence of real continuous functions on
[a, b] such that, for each x [a, b]
f
n
(x) f
n+1
(x) n n, lim
n
f
n
(x) = 0.
Show that (f
n
) is equicontinuous.
Solution. We rst show that (f
n
) converges uniformly to f 0. Note
that the two hypotheses combined entail f
n
0 in [a, b] for every n. Given
> 0, consider the sets
B
n,
= x [a, b] : f
n
(x) , n N.
By continuity of f
n
, the sets B
n
are closed (hence, compact), and, since
x B
n+1
= f
n+1
(x) = f
n
(x) = x B
n
,
nested. Moreover,
n
B
n
is empty (by pointwise convergence to f). This
implies that B
n
is empty for some n, that is
[f
k
(x)[ < x [a, b], k n,
1.3. FUNCTIONS. 27
which is uniform convergence to f. To see that this implies equicontinuity,
x > 0 and N large enough to have [f
n
(x)f
N
(x)[ < /3 for all x [a, b],
n N. Since f
N
is uniformly continuous, there exists
N
> 0 such that
[x y[ <
N
= [f
N
(x) f
N
(y)[ < /3,
so, for all n N, and [x y[ <
N
,
[f
n
(x) f
n
(y)[ < [f
n
(x) f
N
(x)[ +[f
N
(x) f
N
(y)[ +[f
n
(x) f
N
(x)[ < .
Now, calling = min
1
, . . . ,
N
, where
i
> 0, i = 1, . . . , N 1, is chosen
in order to have [f
i
(x) f
i
(y)[ < if [x y[ <
i
, we have
[x y[ < = [f
n
(x) f
n
(y)[ < , n N,
which is equicontinuity.
1.69 Exercise. Let f
n
: [0, 1] [0, 1] be a sequence of nondecreasing
functions whose pointwise limit f is continuous. Show that f
n
f uni-
formly.
Solution. Fix > 0. Since f is uniformly continuous on [0, 1], there is > 0
such that [f(x) f(y)[ < /2 if [xy[ < . Let x
0
= 0, B
i
(x
i
) (i = 1, . . . , k)
be a nite covering of [0, 1] with open balls of radius , and x
k+1
= 1. Now,
choose n N large enough to have
[f
n
(x
i
) f(x
i
)[ <

2
i = 0, . . . , k + 1.
For x [0, 1], we have x [x
i1
, x
i
] for some i, with (without loss of
generality) 0 < x
i
x
i1
< so that, using monotonicity of f
n
,
f
n
(x) f(x) f
n
(x
i
) f(x) (f
n
(x
i
) f(x
i
)) + (f(x
i
) f(x)) < ,
and, similarly
f
n
(x) f(x) (f
n
(x
i1
) f(x
i1
)) + (f(x
i1
) f(x)) > .
The inequalities above directly establish uniform convergence.
1.70 Exercise. Let f
n
: [0, 1] [0, ) be a sequence of nondecreasing
functions whose pointwise limit f is continuous, nondecreasing and
lim
x
f(x) = 1.
Show that f
n
f uniformly.
28 CHAPTER 1. METRIC SPACE ANALYSIS.
Solution. Fix > 0, and M such that 1 f(x) < /3 for all x M.
Since f is uniformly continuous on [0, M], one can argue exactly as in the
previous Exercise 1.69 and show that
sup
x[0,M]
[f
n
(x) f(x)[ 0, n .
For x M, we have
[f
n
(x) f(x)[ [f
n
(x) 1[ + 1 f(x)
[f
n
(M) 1[ +

3
,
for all n large enough to have [f
n
(M) f(M)[ < /3. This is enough to
conclude.
1.71 Exercise. Call B the closed unit ball in R
n
and f
n
: B R a
sequence of continuous functions such that
f
n
f
n+1
on B, and lim
n
f
n
(x) = 0, x B.
Show that f
n
0 uniformly
Solution. We start by observing that f
n
(x) f
n+1
(x) and f
n
(x) 0 as
n imply f
n
(x) 0, for every x B and n N. Since f
n
is continuous
on the compact set B, for each n there exists x
n
B such that
f
n
(x
n
) = max
xB
f(x
n
) = M
n
.
We prove that M
n
0 as n , which is equivalent to uniform con-
vergence of f
n
to 0. Observe that M
n
is a nonincreasing sequence, since
f
n
f
n+1
, bounded below by 0. Thus M
n
converges to some M 0.
Now, x > 0 and use compactness of B to extract a subsequence x
n
k
of x
n
, converging to some x B. Since f
n
(x) tends to 0 as n goes to
, there is N such that 0 f
n
(x) for all n N. By continuity of
each f
n
, [f
n
(x
n
k
) f
n
(x)[ < for n
k
n large enough. Collecting these
observations, we have that
[f
n
k
(x
n
k
)[ [f
n
(x
n
k
)[ [f
n
(x)[ +[f
n
(x
n
k
) f
n
(x)[ < 2,
so that M
n
k
tends to 0, which in turn forces M = 0.
1.72 Exercise. Let f : (0, 1)
2
R possess the following properties:
for each x (0, 1), y f(x, y) is continuous;
1.3. FUNCTIONS. 29
the family T = x f(x, r) : r Q (0, 1) is equicontinuous.
Show that f is a continuous function on (0, 1)
2
.
Solution. Fix (x
0
, y
0
) (0, 1)
2
and > 0. Exploiting the equicontinuity
of T, choose > 0 small enough to have
[x x
0
[ < = [f(x, r) f(x
0
, r)[ <

2
, r Q (0, 1).
Since y f(x
0
, y) is continuous, there is
y
0
> 0 such that
[y y
0
[ <
y
0
= [f(x
0
, y) f(x
0
, y
0
)[ <

2
.
Hence, for each r Q (0, 1) with [r y
0
[ <
y
0
,
[f(x, r) f(x
0
, y
0
)[ [f(x, r) f(x
0
, r)[ +[f(x
0
, r) f(x
0
, y
0
)[ < .
Now, for y with [y y
0
[ <
y
, we have, by continuity
f(x, y) = lim
ry,rQ
f(x, r) = lim
ry,rQ,|ry
0
|<
y
0
f(x, r),
and, being the above estimate uniform in r, we can pass to the limit, ob-
taining
[f(x, y) f(x
0
, y
0
)[ < ,
if [x x
0
[ < and [y y
0
[ <
y
, which is continuity of f.
1.73 Exercise. Let
k
be nonnegative continuous functions on [1, 1]
with the following properties:

_
1
1

k
(t) dt = 1, for each k;
for each (0, 1),
k
0 uniformly on D

= [1, ] [, 1].
Show that
lim
k
_
1
1
f(t)
k
(t) dt = f(0),
for every continuous function f : [1, 1] R.
Solution. Let M = max [f[. Fix > 0 and > 0 such that [f(t)f(0)[ <
if [t[ < . Then, there is K such that
k K = max
xD

[
k
(x)[ <

4M
.
30 CHAPTER 1. METRIC SPACE ANALYSIS.
Now,

_
1
1
f(t)
k
(t) dt f(0)

_
1
1
(f(t) f(0))
k
(t) dt

_
D

[f(t) f(0)[
k
(t) dt +
_

[f(t) f(0)[
k
(t) dt
2M

4M
_
D

1 dt +
_

k
(t) dt + = 2
which proves the claim.
Contractions.
1.74 Contraction. A map f : M M, with M metric space, is called a
contraction if there is < 1 such that
d(f(x), f(y)) d(x, y).
for all x, y M. It is clear that a contraction is necessarily a (uniformly-
Lipschitz) continuous function.
1.75 BanachCaccioppoli Contraction Theorem. Let f be a contrac-
tion on a complete metric space M. Then f has a unique xed point p M
(that is, f(p) = p), and, for every q
0
M, the sequence (q
n+1
= f(q
n
))
converges to p.
Proof. Here, we give a proof which diers from the usual one, relying on
the following useful lemma.
1.76 Lemma. Let B
n
be a sequence of closed, bounded, nonempty, nested
(B
n
B
n+1
) subsets of a complete metric space M such that
lim
n
diam[B
n
] = 0.
Then,
n
B
n
consists of a single point.
Choose q
0
M and call R = d(q
0
, f(q
0
)). Then, x > R/(1 ). If
q B
0
= B

(q
0
), we have
d(q
0
, f(q)) R + d(f(q
0
), f(q)) R + d(q
0
, q) R + < ,
so that f(B
0
) B
0
. Now, inductively dene the sequence of closed and
bounded sets B
n+1
= f(B
n
). We have
B
n+1
= f
n+1
(B
0
) = f
n
(f(B
0
)) f
n
(B
0
) = B
n
,
1.3. FUNCTIONS. 31
so that (B
n
) is a nested sequence of sets. Moreover, if q, q

B
n
,
d(f(q), f(q)

) d(q, q

) = diam[B
n+1
] diam[B
n
],
which immediately implies diam[B
n
]
n
. Thus, the intersection
n
B
n
consists of a single point p and f
n
(q
0
) converges to p (immediate). Choosing
q
0
= p, we see that f
n
(p) converges both to p and f(p), so that p = f(p) as
desired.
1.77 Counterexample. If M is not a complete metric space, a contraction
need not have a xed point. As an example, consider M = Q [1, ) and
f(x) = 1 + 1/x. It is easy to verify that f is a contraction on M, however,
a xed point y for f satises
y = 1 +
1
y
,
which has no solutions in M (a solution to which f
n
(q) converges in R for
every q M is = (1 +

5)/2, the golden ratio number.


1.78 Theorem. Let M be a metric space and f : M M be a function
satisfying, for every p, q M, p ,= q
d(f(p), f(q)) < d(p, q)
(such a function is said to be a weak contraction). If M is compact, f
possesses a unique xed point.
Proof. We note that a weak contraction is necessarily continuous, so that
the function
p M g(p) = d(p, f(p)) [0, )
is a real continuous function on a compact set. Hence, g attains its minimum
g( p) 0 at some p M. Suppose p ,= f( p), then
g(f( p)) = d(f( p), f
2
( p)) < d( p, f( p)) = g( p),
against the minimality of g( p). Thus, f( p) = p, and p is a xed point for f
on M. Uniqueness is checked in the usual way.
1.79 Counterexample. If M is not a compact metric space, Theorem
1.78 fails. A weak contraction with no xed points is f(x) = x x
1
on
M = [1, ).
32 CHAPTER 1. METRIC SPACE ANALYSIS.
1.80 Counterexample. Even if M is a compact metric space, a weak
contraction is not necessarily a contraction. For example, take M = [0, 1]
and f(x) = sin x. For each x, y M, x < y, we have
(x, y) : [ sin x sin y[ = [ cos [[x y[ < [x y[,
so f is a weak contraction. However, for every < 1 there exists x (0, 1]
such that sin x > x, so f cannot be a contraction.
1.81 Exercise. Let (k
i
) be a sequence of positive reals, and dene
[k
i
] =
1
k
i
, [k
i
, k
i+1
] =
1
k
i
+ [k
i+1
]
, [k
i
, k
i+1
, k
i+2
] =
1
k
i
+ [k
i+1
, k
i+2
]
, . . .
[k
i
, . . . k
i+n
] =
1
k
i
+ [k
i+1
, . . . , k
i+n
]
.
Show that, if k
i
2 for all i, lim
n
[k
1
, . . . , k
n
] exists.
Solution. To get the result, we mimick the (usual) proof of the Banach
contraction theorem. For y R
+
, dene the map
x R
+
f
y
(x) =
x
xy + 1
.
With this choice, [k
i
, k
i
+ 1] = f
k
i
(k
i+1
), and inductively,
[k
1
, . . . , k
n
] = (f
k
1
f
k
2
f
k
n1
f
k
n
)(k
n+1
).
Let
i
= (f
k
ni
f
k
n
)(k
n+1
). We show that
i
is a strictly decreasing
sequence of positive numbers, so that lim
i

i
= lim
n
[k
1
, . . . , k
n
] exists.
Indeed, for y 2, we have
0 < f
y
(x)
x
1 + 2x
< x,
so that 0 <
i
= f
k
ni
(
i1
) <
i1
, for every i = 1, . . . , n, and the proof is
completed.
1.82 Exercise. Dene the functional : (([0, 1]) (([0, 1]) by
[(f)](t) = 1 +
_
t
0
s
2
e
f(s)
ds.
Let f
0
be a nonnegative continuous function on [0, 1], with |f
0
| = 1, and
f
n+1
= (f
n
) for n N.
1.3. FUNCTIONS. 33
(a) Prove that 1 f
n
(t) 4/3 for all t [0, 1] and n = 1, 2, . . ..
(b) Show that the sequence of functions (f
n
) converges uniformly to some
function f (([0, 1]).
Solution. To prove (a), observe that, if f
n
is a nonnegative continuous
function satisfying |f
n
| |f
0
|, we have
1 f
n+1
(t) 1 + max
s[0,1]
|e
f(s)
|
_
1
0
s
2
ds
4
3
We now turn to (b). The set
C = f (([0, 1]) : f(x) 0 x [0, 1], |f|
4
3

is a closed subset of (([0, 1]), hence a complete metric space, and the op-
erator maps C into itself. Observe that [e
x
e
y
[ [x y[, since
max
[0,1]
[(ex)

[ = 1. Hence, for any f, g C,


[(f)](t) [(g)](t)
_
0
t
s
2
[e
f(s)
e
g(s)
[ ds
|f g|
_
0
t
s
2
ds
=
|f g|
3
,
which immediately gives
|(f) (g)|
|f g|
3
,
that is, is contractive on C. Apply the contraction theorem to establish
the existence of lim
n
f
n
, which is the unique xed point f of , solution of
the integral equation (f) = f,
f(t) = 1 +
_
t
0
s
2
e
f(s)
ds.
With standard methods, we obtain f(t) = 1 log(t
3
/3 + 1).
34 CHAPTER 1. METRIC SPACE ANALYSIS.
1.4 Riemann Integration.
1.83 Partitions. Mesh functions. Let [a, b] R be a closed and
bounded interval. We call P, partition of [a, b], the nite set of points
a = x
0
< x
1
< . . . < x
n1
< x
n
= b.
The mesh of P is the quantity [P[ = max i = 1, . . . , n 1x
i
x
i1
. A
nondecreasing function : [a, b] R is said to be a mesh function. Given
a partition P, we can associate to it the nonnegative quantities

i
= (x
i
) (x
i1
), i = 1, . . . , n.
1.84 Riemann Sums. Let f : [a, b] R be a bounded function. Given a
partition P of [a, b] and a mesh function , we dene the lower and upper
sums of f with respect to (P, ) as
L(f, P, ) =
n

i=1
m
i

i
, U(f, P, ) =
n

i=1
M
i

i
,
where
m
i
= inf[x
i1
, x
i
]f, M
i
= sup[x
i1
, x
i
]f.
The (well-dened) quantities
L(f, ) = sup L(f, P, ) U(f, ) = inf U(f, P, ),
where the sup and inf are taken on all partitions of [a, b], are called respec-
tively the lower and upper -integral of f. If equality holds above, we say
that f is -integrable on [a, b], with integral
_
b
a
f d = L(f, ) = U(f, ).
We indicate the class of functions which are -integrable on [a, b] as 1

([a, b]).
1.85 Renement. If P and P

are two partitions of [a, b] such that


P P

, then P

is a renement of P. Given two partitions P


1
and P
2
,
their common renement is P

= P
1
P
2
.
1.86 Lemma. For every bounded f : [a, b] R, and every mesh function
, if P

is a renement of P, then
L(f, P

, ) L(f, P, ), U(f, P

, ) U(f, P, ).
1.4. RIEMANN INTEGRATION. 35
Proof. We only prove the rst claim, since the second follows in the exact
same fashion. Let P = x
j

n
j=0
be a generic partition and P

= P x
with x [x
j1
, x
j
]. Then
m

1
= inf
[x
j1
,x]
f, m

2
= inf
[x,x
j
]
f m
j
= inf
[x
j1
,x
j
]
f,
which easily implies L(f, P

, ) L(f, P, ). The general case follows by


induction.
1.87 Lemma. Let f : [a, b] R be a bounded function. Then f
1

([a, b]) if and only if for each > 0, there exists a partition P of [a, b]
such that
(1.4) U(P, f, ) L(P, f, ) < .
Proof. Suppose f is -Riemann integrable on [a, b]. Given > 0, there are
partitions P
1
, P
2
such that
L(f, P
1
, ) >
_
b
a
f d

2
, U(f, P
2
, ) <
_
b
a
f d +

2
.
Call P the common renement of P
1
, P
2
, then
U(f, P, ) L(P, f, ) U(f, P
2
, ) L(f, P
1
, ) < ,
as claimed.
Viceversa, let (1.4) hold. Since, for every partition P,
L(f, P, ) L(f, ) U(f, ) U(f, P, ),
for each > 0 we can nd P such that
U(f, ) L(f, ) U(f, P, ) L(f, P, ) < ,
which means U(f, ) = L(f, ), that is, f is -Riemann integrable on
[a, b].
1.88 Lemma. Let > 0 and P be a partition for which (1.4) holds. Then,
for every choice of points s
i
, t
i
[x
i1
, x
i
],
n

i=1
[(f(s
i
) f(t
i
))[
i
< .
36 CHAPTER 1. METRIC SPACE ANALYSIS.
Moreover, if f 1

([a, b]),

i=1
f(t
i
)
i

_
b
a
f d

< .
Proof. The rst part follows observing that, for each i, [f(s
i
) f(t
i
)[
M
i
m
i
. The second part follows from (1.4) and the obvious fact

i=1
f(t
i
)
i

_
b
a
f d

< U(f, P, ) L(f, P, ).


1.89 Riemann-Lebesgue Lemma.Let f : [a, b] R be a bounded func-
tion. Then, f 1([a, b]) if and only if the set of discontinuities of f is a
zero set.
Proof. Suppose f is Riemann integrable on [a, b]. We write the discontinuity
set of f as
D =
_
kN
D
k
, D
k
=
_
x [a, b] : osc
x
(f) >
1
k
_
.
If we show that each D
k
is a zero set, we are done, being D a countable
union of zero set, and thus a zero set itself.
Hence, x k N, > 0, and choose a partition P such that
U(f, P) L(f, P) <

k
.
We say that j J if [x
j1
, x
j
] D
k
is nonempty. If j J, then M
j
m
j
>
1/k, so that

k
> U(f, P) L(f, P)

jJ
(M
j
m
j
)
j
>
1
k

jJ

j
,
which in turn implies, being
D
k

_
jJ
[x
j1
, x
j
] with

jJ

j
< ,
and arbitrary, that D
k
is a zero set.
1.4. RIEMANN INTEGRATION. 37
Conversely, suppose that the set D of discontinuity points of f is a zero
set. Let M = sup
x[a,b]
[f[ < , and, for a given > 0, let (x
j1
, x
j
) :
j N be a countable collection of open intervals of total length lesser
than /4M(b a) covering D

. The set D is closed and bounded, hence


compact, so that there is a nite subcollection of open intervals I
1
, . . . , I
n

still covering D

with the endpoints not belonging to D

.
The set
S = [a, b]
n
_
j=1
I
j
is compact, and f is continuous on S. Hence, it is uniformly continuous so
that there is > 0 such that [f(x) f(y)[ < /2 for x, y S, [x y[ < .
Let P = J
1
, . . . J
m
be a partition of S of mesh and consider the common
renement P

of P and Q = I
1
, . . . I
n
considered as partitions of [a, b].
We then have
U(f, P

) L(f, P

) =

I
j
(M
j
m
j
)
I
j
+

J
j
(M
j
m
j
)
J
j
2M

I
j

I
j
+

J
j

2(b a)

< ,
by the properties above, so that (1.4) is fullled for any given > 0. We
conclude that f 1([a, b]).
1.90 Exercise. Let c
n
be an innite sequence of distinct points of [0, 1]
and dene the function
f(x) =
_
_
_
1
n
x = c
n
0 x / c
n
.
Show that f is Riemann-integrable on [0, 1].
Solution. Given > 0, x an integer N > 2/. Let P be a partition of mesh
= /2N. We have m
i
= inf
[x
i1
,x
i
]
f = 0 for each i, since the complement
of c
n
is a dense set, so that L(f, P) = 0. Then, we say that i J if there
is c
n
[x
i1
, x
i
] with n N, so that
M
i
= sup
[x
i1
,x
i
]
f
_
1 i J
1/N i / J.
38 CHAPTER 1. METRIC SPACE ANALYSIS.
By the pigeonhole principle, J has at most N elements. Hence,
U(f, P)

iJ
1 +

i / J

N
N +
1
N
< ,
so that we may apply Lemma 1.4.
1.91 Exercise. Let f : R R be continuous. Show that
f
n
(x) =
1
n
n1

i=0
f
_
x +
i
n
_
converges uniformly to a limit on each compact subset of R.
Solution. Let K be a compact subset of R and x K, > 0, and > 0
such that
y, z K, [y z[ < = [f(y) f(z)[ <
(this is possible thanks to the uniform continuity of f on K). Being f
continuous, it is integrable on every interval, and
F(x) =
_
x+1
x
f(y) dy = lim
k
S
k
(x), S
k
(x) =
1
k
k1

i=0
M
i
,
where, denoting with I
i
= [x+i/n, x+(i +1)/n], M
i
= max
I
i
f(y). We will
show that f
n
converges uniformly to F. Fix N suciently large to have
1/N < and 0 < S
N
(x) F(x) < for all x K. For n N, we have, by
uniform continuity,
0 M
i
f
_
x +
i
n
_
< = 0 < S
n
f
n
(x) < ,
so that we arrive at
[f
n
(x)F(x)[ < [S
n
(x)f
n
(x)[ +[S
n
(x)F(x)[ < +[S
N
(x)F(x)[ < 2,
for every x K, which is uniform convergence.
1.92 Exercise. Let f : [a, b] R such that
lim
yx
f(y) = L
x
, lim
yx
+
f(y) = R
x
exist for each x
0
[a, b]. Prove (directly) that f is Riemann integrable on
[a, b].
1.4. RIEMANN INTEGRATION. 39
Solution. Fix > 0 and call = b a. For each x
0
[a, b], there is
x
0
> 0
such that
0 < x y <
x
= [f(y) L
x
[ < /4, (1.5)
0 < y x <
x
= [f(y) R
x
[ < /4.
The collection B

x
(x) : x [a, b] is an open cover of [a, b]. Extract a nite
subcover B
i
= B

i
(x
i
), i = 1, . . . , n and call
K = max
i
max([f(x
i
)[, [L
x
i
[, [R
x
i
[),
i
= (x
i+1
x
i
), =

2n(4K + /)
.
Note that
i
<
i
(being B
i
an open cover), and that f is bounded, since
our construction implies f K for every x [a, b]. Consider the partition
T
a < z
0,2
< z
1,1
< x
1
< z
1,2
< z
2,1
< x
2
< . . . < x
n
< z
n,2
< b,
where z
i,1
= x
i
min,
i1
/2, z
i,2
= x
i
+ min,
i
/2. By (1.5),
M
i,1
= max
[x
i
,z
i,2
]
f(x) maxR
x
i
+ /2, f(x
i
) K + /4,
m
i,1
= min
[x
i
,z
i,2
]
f(x) minR
x
i
/2, f(x
i
) K /4
M
i,2
= max
[z
i,2
,z
i+1,1
]
f(x) R
x
i
+ /2
m
i,2
= min
[z
i,2
,z
i+1,1
]
f(x) R
x
i
/2
M
i,3
= max
[z
i+1,1
,x
i+1
]
f(x) maxL
x
i+1
+ /2, f(x
i
) K + /4,
m
i,3
= min
[z
i+1,1
,x
i+1
]
f(x) minL
x
i
/2, f(x
i
) K /4.
This allows us to write
U(f, T) L(f, T)
n

i=1
(M
i,2
m
i,2
)
i
+
n

i=1
(M
i,1
+ M
i,2
m
i,1
m
i,2
)


2
n

i=1

i
+ (4K + /)n
<

2
+

2
= ,
which shows the integrability of f.
A quick indirect proof is obtained exploiting that a function which has left
and right limit at every point is bounded and discontinuous at most at a
40 CHAPTER 1. METRIC SPACE ANALYSIS.
countable set of points (which is a zero set). RiemannLebesgues Lemma
then yields integrability of f.
1.93 Exercise. Let f : R R be continuous and such that
_
q
p
f(x) dx = 0, p, q S
c
,
where S R is countably innite. Prove that f is identically zero.
Solution. Suppose there is x R such that (without loss of generality)
f(x) = k > 0. Thus, for some > 0,
[x y[ < = f(y) >
k
2
.
Being S countable, it is union of closed subsets with empty interior, hence
it contains no interval, so that S
c
(x , x) and S
c
(x, x + ) contain
respectively two distinct points p and q; we would have
_
q
p
f(x) dx >
k
2
(q p) > 0,
contrary to our hypothesis.
1.94 Exercise. Let f : R R be continuous and such that
_

[f(x)[ dx < .
Prove that there is a sequence x
n
such that
x
n
f(x
n
) 0, x
n
f(x
n
) 0, n .
Solution. Argue by contradiction: suppose that there exists > 0 such
that
liminf
x
[xf(x)[ > .
This means that x[f(x)[ > , for every x > M, so that,
_

M
[f(x)[ dx
_

M
1
x
dx = +.
Contradiction. The other case is handled similarly.
1.4. RIEMANN INTEGRATION. 41
1.95 Exercise. Let f : [0, ) R be uniformly continuous and
lim
b
_
b
0
f(x) dx = L R.
Prove that lim
x
f(x) = 0.
Solution. Our hypothesis implies that for every > 0 there is M large
enough such that

_
y
x
f(u) du

< , M x < y.
Argue by contradiction, supposing that there is a sequence x
n
(as
n ) and > 0 such that inf
n
f(x
n
) . By uniform continuity, there
is > 0 independent of n such that
[x
n
u[ < = f(u) >

2
.
This implies
_
x
n
+
x
n
f(u) du >

2
,
which, if compared to the previous inequality, yields a contradiction.
1.96 Exercise. Let f
k
, g : (0, ) R (k N) be Riemann-integrable on
each interval [t, T] (0, ), f
k
f uniformly on every compact subset of
(0, ), [f
k
[ < g on (0, ) and
_

0
g(x) dx = K < .
Show that
_

0
f(x) dx = lim
k
_

0
f
k
(x) dx.
Solution. We will carry the proof using the additional assumption that
each f
k
is nonnegative on (0, ), and come back to the general case at the
end.
Observe that uniform convergence guarantees that f is Riemann-integrable
on every interval [t, T] (0, ) and
_
T
t
f(x) dx = lim
k
_
T
t
f
k
(x) dx.
42 CHAPTER 1. METRIC SPACE ANALYSIS.
We will carry the two limit processes t 0 and T in separate steps.
To begin with, we observe that
_

1
f
k
(x) dx = lim
T
_
T
1
f
k
(x) dx
exists nite for every k, since the function T
_
T
1
f
k
is monotone increasing
and bounded by K. The same occurs to T
_
T
1
f, since, for every T we
can choose N large enough to have, for all x [1, T], [f
k
(x) f(x)[ < 1/T,
for k N, hence
_
T
1
f(x) dx
_
T
1
[f
k
(x) f(x)[ dx +
_
T
1
f
k
(x) dx 1 + K,
so that
_

1
f exists. Finally, for > 0, x T large enough so that
_

T
g < /4
and N large enough to have
_
T
1
[f
k
f[ < /2 for k N. Now, for every
> 0
_
T+
T
[f
k
(x) f(x)[ dx 2
_
T+
T
g(x) dx <

2
,
so that
_
T+
1
[f
k
(x) f(x)[ dx < .
Passing to the limit for and being arbitrary we are done.
1.97 Exercise. Prove or provide a counterexample to the following state-
ments:
(a) if f : R R is a continuous function, then L R such that
lim
0
+
_
|x|1
f(x)
x
dx = L;
(b) (a) is true under the additional assumption that there are positive
constants C, such that [f(x) f(y)[ C[x y[

for all x, y R.
Solution. (a) is false. Consider, as a counterexample, the function
f(x) =
_

_
0 x < 0

1
log x
0 x 1/2
1
log 2
x 1/2,
1.5. ADDITIONAL TOPICS 43
which is clearly continuous on R. However, for 0 < < 1/2,
_
1

f(x) dx =
1
2 log 2

_
1/2

1
x log x
dx =
1
2 log 2
+ log log(1/) log log 2,
which tends to as 0
+
.
(b) is true. It is enough to notice that we can assume f(0) = 0 without
losing in generality, since
_
|x|1
f(x)
x
dx =
_
|x|1
f(x) f(0)
x
dx +
_
|x|1
f(0)
x
dx
and the last term is zero, by symmetry. At this point, we know that
_
|x|1
f(x)/x dx exists and using our additional hypothesis,

_
|x|1
f(x)
x

_
|x|1
[x[
1
dx
and the last integral converges as 0
+
.
1.5 Additional Topics
1.98 Exercise. Let f : M N be a function between metric spaces such
that
g(x) = lim
px
f(p)
exists for all x M. Show that the function g is continuous on M.
Solution. If p M denote with B
r
(p) the open ball of radius r centered at
p. Fix x
0
M and > 0. There is > 0 such that
p ,= x
0
, p B

(x
0
) = d(f(p), g(x
0
)) < /2.
Hence, if x B

(x
0
), there is
1
such that B

1
(x) B

(x
0
) and
p ,= x, p B

1
(x
0
) = d(f(p), g(x)) < /2.
Select p B

1
(x
0
) and observe that
d(g(x), g(x
0
)) d(g(x), f(p)) + d(f(p), g(x
0
)) < .
Thus, for all x B

(x
0
), d(g(x), g(x
0
)) < , which is continuity of g at x
0
.
44 CHAPTER 1. METRIC SPACE ANALYSIS.
1.99 Exercise. Let f : [a, b] R be a dierentiable function for which
f

(a) < f

(b). Show that for each (f

(a), f

(b)) there is a c (a, b) such


that f

(c) = .
Solution. Fix (f

(a), f

(b)). Dene g(x) = f(x) x. Then, g is


dierentiable in (a, b). Being g

(a) < 0, there is some t


1
> a with g(t
1
) <
g(a), and, being g

(b) > 0, there is some t


2
< b with g(t
2
) < g(b). Hence,
g attains its minimum at some c (a, b). By Fermats Theorem, g

(c) = 0,
that is, f

(c) = .
1.100 Exercise. Let f (
2
([0, 1]), f(0) = f(1) = 0 and f(x) > 0 for all
x (0, 1). Show that
_
1
0

(x)
f(x)

dx > 4.
Solution. Since f(0) = f(1) = 0 and f(x) > 0 for every x (0, 1), f
attains its positive maximum at some point (0, 1). Use the mean value
theorem to nd (0, ), (, 1) such that
f() = f

() = f

()( 1),
which gives
f

() f

()
f()
=
1
1

1

=
1
( 1)
,
and nally
_
1
0

(x)
f(x)

dx >

(x)
f()
dx

() f

()
f()

=
1
(1 )
4,
as desired.
1.101 Upper semicontinuity. Let M be a metric space and p M. A
function f : M R such that for each > 0 there is a > 0 such that
d(p, q) < = f(q) < f(p) +
is said to be upper semicontinuous at p.
1.102 Exercise. Let f : [0, 1] R be an upper semicontinuous function.
Show that f is bounded above on [0, 1] and that it attains its maximum.
Solution. For each x [0, 1], let
x
> 0 be such that f(y) < f(x) + 1 if
[x y[ <
x
. The collection B

x
(x) : x [0, 1] is an open cover of [0, 1].
1.5. ADDITIONAL TOPICS 45
Extract a nite subcover B
i
= B

i
(x
i
)
n
i=1
. Then, for each y [0, 1], y
belongs to some B
i
, thus f(y) < f(x
i
) + 1, which entails
f(y) max
i=1,...,n
f(x
i
) + 1, y [0, 1],
and the right hand side is nite. To see that f attains its maximum, call
M = sup
[0,1]
f, and, for n N, dene the sets
K
n
=
_
x [0, 1] : f(x) M
1
n
_
,
which are nested and nonempty for each n. To see that K
n
are closed sets
(hence compact, being subsets of [0, 1]), observe that upper semicontinuity
implies
(x

) K
n
, x

x = f(x) limsup

f(x

) M
1
n
= x K
n
.
Thus
n
K
n
contains a point x, and x K
n
for each n implies f( x) = M.
1.103 Exercise. Let A, B R
n
nonempty, disjoint subsets with A com-
pact and B closed. Show that there exist a A, b B such that
|a b| |x y|, x A, y B.
Solution. For x A, dene the function
f(x) = inf
yB
|x y|.
We show that f is upper semicontinuous on A. Fix x
0
A and > 0. There
is y B such that |x
0
y| < f(x
0
) + so that, if x A, |x x
0
| < ,
f(x) |x y| < |x
0
y| +|x
0
x| < f(x
0
) + 2,
which is upper semicontinuity at x
0
. Hence, f attains its positive maximum
R at some x
0
A. Now, consider the set
B

= y B : x A with |x y| 2R.
Then B

is clearly a bounded set. It is also closed, since, if y


n
B

, and
y
n
y, then, by the closedness of B, y B. Moreover, there is a sequence
46 CHAPTER 1. METRIC SPACE ANALYSIS.
x
n
of points of A such that |y
n
x
n
| 2R. The sequence x
n
converges,
up to a subsequence, to some x A, so that
|y x| |y y
n
| +|y
n
x
n
| +|x
n
x| < 2R + ,
for big enough n. Being arbitrary, |y x| 2R, and y B

. We
conclude that B

is compact and that, for every x A,


f(x) = inf
yB
|x y| = min inf
yB

|x y|, inf
yB\B

|x y| = inf
yB

|x y|.
The function
(x, y) A B

F(x, y) = |x y|,
is continuous on the compact set AB

, so that it attains its minimum at


some (a, b) A B

. Finally, if x A, y B,
|x y| inf
y

B
|x y

| = inf
y

|x y| min
(x

,y

)AB

|x

| = |a b|,
as desired.
1.104 Exercise. Let f : [0, 1] R be nondecreasing. Is it true that f is
the pointwise limit of a sequence of continuous functions f
n
: [0, 1] R?
Prove or supply a counterexample?
Solution. True. Let d
n
be an enumeration of the discontinuities of f and,
for k = 0, . . . , 2
n
, x
k,n
= k/2
n
. Dene f
n
as the piecewise linear function
through the (at most 2
n
+ n) points
(x
k,n
, f(x
k,n
) (d

, f(d

) : = 1, . . . n.
It is clear that f
n
(d

) = f(d

) for all n , so that f is the pointwise limit


of f
n
at each discontinuity point of f. It is also easy to see that, if x is a
continuity point for f, f
n
(x) converges to f(x), so that we are done.
Solution. Fix x
0
[0, 1] and > 0. Let y
0
[0, 1] be such that g(x
0
) =
f(x
0
, y
0
). Being f continuous on a compact set, it is uniformly continuous,
hence there exists > 0 such that
[(x, y) (x

, y

)[ < = [f(x, y) f(x

, y

)[ < .
This property easily fetches that
[x x
0
[ < = g(x) f(x, y
0
) > f(x
0
, y
0
) .
1.5. ADDITIONAL TOPICS 47
Now, we claim that
[x x
0
[ < = g(x) < f(x
0
, y
0
) + ,
which is what remains to be proved. Suppose it is not so. Calling z [0, 1]
the point for which g(x) = f(x, z), we use the uniform continuity property
to get
f(x
0
, z) > f(x, z) > f(x
0
, y
0
),
which is a contradiction.
1.105 Exercise. Let M, N be metric spaces. A map f : M N is said
to be open if for each open set A M, the image set f(A) is open in N.
Prove or supply counterexamples:
(1) f open = f continuous;
(2) f homeomorphism = f open;
(3) f open, continuous bijection = f homeomorphism;
(4) f continuous surjection = f open;
(5) f : R R continuous, open surjection = f homeomorphism;
(6) replace R with S
1
(unit circle in R
2
) in (5).
Solution. (1) No. A counterexample is f(x) = 1/x for x ,= 0, f(0) = 0, on
[0, ).
(2) Yes. For, denoting g = f
1
, g is continuous, hence g
1
(A) = B is open
for each A open in M, which means that f(A) = B is open.
(3) Yes. For, denoting g = f
1
if A M open then f(A) = B is open.
Since f is a bijection, g(B) = A, that is, g
1
(A) = B is open for each A
open in M. Hence g is continuous, and f is an homeomorphism.
(4) No. A counterexample is f(x) = x x
3
on R.
(5) Yes. We show that f is injective, hence a bijection. Then one may
apply point (3). Suppose f is not injective, hence a, b R, a ,= b, for
which f(a) = f(b) = c R. Being f continuous, f assumes minimum
and maximum value on [a, b]. If both were equal to c, a contradiction is
reached, since this means f((a, b)) = c, contrary to the openness of f. We
can assume with no loss of generality that max
[a,b]
f = f(d) > c, for some
d (a, b). This also contradicts the openness of f, since the image of any
interval (a

, b

) with a < a

< c < b

< b is not open.


(6) No. A counterexample is the map (1, ) (1, 2).
48 CHAPTER 1. METRIC SPACE ANALYSIS.
1.106 Exercise. Let f : M N be a function from one metric space
to another, fullling the following condition: for every convergent sequence
(p
n
) M, (f(p
n
)) converges in N. Prove that f is continuous.
Solution. Let (p
n
) M, p
n
p. Consider the sequence
q
k
=
_
p
n
k = 2n 1,
p k = 2n,
which obviously converges to p. Thus, f(q
k
) converges to some q N,
and every subsequence of f(q
k
) must converge to the same q N. Since
f(q
2n
) = f(p), for all n, q = f(p), and f(q
2n1
) = f(p
n
) f(p). This shows
that f is continuous.
1.107 Exercise. Let f : R R be uniformly continuous. Prove that
there are constants A, B such that [f(x)[ A + B[x[ for all x R.
Solution. Fix x R (suppose x > 0, without losing generality), and > 0
such that
y, z R, [y z[ < = [f(y) f(z)[ < 1;
observe that does not depend on x. Set x
0
= 0, and construct a nite
sequence of points x
1
, . . . , x
n
such that x
i
x
i1
< and x
n
= x. Then, n
is the lower integer part of [x[/, and
[f(x) f(0)[
n

i=1
[f(x
n(i1)
) f(x
ni
)[ < n
[x[

,
that is
[f(x)[ [f(0)[ +
[x[

,
as desired.
1.108 Exercise. Let D be the half-plane (x, y) R
2
: x > 0 and
f : D R a (
1
function such that

f
x
(x, y)

x
,

f
y
(x, y)


1
.
Show that f is uniformly continuous on D.
1.5. ADDITIONAL TOPICS 49
Solution. Indeed, f is 1/2-H older continuous. For, if (x, y), (x
0
, y
0
) D,
we have
[f(x, y) f(x
0
, y
0
)[ [f(x, y) f(x, y
0
)[ +[f(x, y
0
) f(x
0
, y
0
)[

_
y
y
0

f
y
(x, t)

dt +
_
x
x
0

f
y
(t, y
0
)

dt
[y y
0
[ + 2([

x
0
[)
[y y
0
[ + 2
_
[x x
0
[,
which is the desired estimate.
1.109 Exercise. Let f : R R be a dierentiable function such that
lim
x
f(x)
x
= 1.
Show that there exists a sequence of points (x
n
) such that x
n
and
f

(x
n
) 1.
Solution. By contradiction: suppose there is > 0 such that [f

(x)1[ >
for all suciently large x, say x greater than some M > 0. Then, for
x > M, either f

(x) > 1 or f

(x) < 1 for all x, since f

satises the
intermediate value property. We can assume to be in the rst case without
loss of generality. By the mean value theorem, for each x > M there is

x
(M, x) such that
f(x) = f(M) + f

(
x
)(x M) > f(M) + (1 )(x M).
Divide by x and pass to the limit on both sides, so to get
lim
x
f(x)
x
lim
x
f(M) + (1 )(x M)
x
= 1 > 1.
Contradiction.
1.110 Exercise. Let f : [a, b] R be monotone increasing.
(1) Show that lim
tx
f(t) = f(x

) and lim
tx
+ f(t) = f(x
+
) exist for each
x (a, b), and that f(x

) f(x) f(x
+
).
(2) Show that the set of discontinuity points of f is at most countable.
(3) Is the same true if [a, b] is replaced with R?
Solution. Consider the set V
x
= f(t) : t [a, x), which is nonempty
and bounded above by f(x). Let f(x

) be the l.u.b of V
x
. Fixed > 0,
50 CHAPTER 1. METRIC SPACE ANALYSIS.
then there is

t [a, x) such that f(x

) < f(

t) f(x

). Now, setting
= x

t, monotonicity entails
f(x

) < f(

t) f(t) f(x

), tin[x , x),
which in turn proves the rst part of (1). A similar proof holds for the
second claim, while the third part follows from monotonicity and persistence
of sign.
We now turn to (2). Calling f(a

) = f(a), f(b
+
) = f(b), for x [a, b],
we dene f(x) = f(x
+
) f(x

). Preliminarily observe that, for any given


k points a x
1
. . . x
k
b, setting x
k+1
= b,
(1.6)
k

j=1
f(x
j
)
k

j=1
[f(x
j+1
) f(x
j
)] f(b) f(a).
If x is a discontinuity point of f, the monotonicity assures f(x) > 0.
Denote with D the set of such points. Then, D =
nN
D
n
, where
D
n
=
_
x [a, b] : f(x) >
1
n
_
.
Inequality (1.6) implies that D
n
contains at most N points, where N is the
greatest integer lesser than n(f(b) f(a)). Hence, D is a countable union
of nite sets, hence countable itself.
To deal with (3), call D the discontinuity points of f : R R. Then f
restricted on [n, n+1], n Z has a countable set of discontinuities D
n
, and
D
_
nZ
D
n
Z,
which is the countable union of countable sets, hence countable.
1.111 Exercise. Let U R
m
be an open set and h : U R
m
be a
uniformly continuous homeomorphism. Prove that U = R
m
.
Solution. U is open and nonempty. We prove that U is closed, and these
three facts imply U = R
m
. Let (x
n
) be a sequence of points of U converging
to x U and call y
n
= h(x
n
). Being uniformly continuous, h admits
a unique continuous extension to U, so that y = h(x) is well dened as
lim
n
h(x
n
). Since h is onto from U to R
m
, there is x U such that y = h( x).
We now show that x = x. Suppose not, then
y
n
y and h
1
(y
n
) x ,= x = h
1
(y),
1.5. ADDITIONAL TOPICS 51
which leads to a contradiction, for h
1
is continuous. Thus x U, and U
is closed.
1.112 Exercise. Suppose F : R
m
R
m
is continuous and satises
[F(x) F(y)[ [x y[
for every x, y R
m
, with > 0. Prove that F is an homeomorphism.
Solution. We divide the proof in three steps.
F is injective: let x, y R
m
such that F(x) = F(y) = u R
m
. Then
0 = [F(x) F(y)[ [x y[ , so that x = y.
F has a continuous inverse G : U = F(R
m
) R
m
. Let u, v U and
x, y R
m
such that F(x) = u, F(y) = v. Then
[G(u) G(v)[ = [x y[
1

[F(x) F(y)[ =
1

[u v[,
so that G is (uniformly Lipschitz) continuous. This, in particular, implies
U homeomorphic to R
m
, thus U is a open subset of R
m
.
F is surjective: we show that U is also closed in R
m
, which implies
U = R
m
. To see this, use that G is a uniformly continuous homeomorphism
from U to R
m
, and proceed as in Exercise 1.111.
1.113 Exercise. Let f : R
n
R
n
be a function of class (
1
such that
[Df(x)[ ,= 0 for all x R
n
. Show that, if the set x R
n
: [f(x)[ M is
bounded for every M > 0, f is onto.
Solution. We show that f(R
n
) is nonempty (obvious), open, and closed in
R
n
, so that it coincides with R
n
.
To see that f(R
n
) is open, choose y f(R
n
), (that is, y = f(x) for
some x R
n
). Since [Df(x)[ , = 0, by the inverse function theorem, f is
a dieomorphism near x, in particular, there are neighborhoods B

(y) and
B

(x) such that f(B

(x)) = B

(y), so that B

(y) is contained in f(R


n
),
which establishes openness of f(R
n
).
To see that f(R
n
) is also closed, consider a sequence of points y
n

f(R
n
), (that is y
n
= f(x
n
), x
n
R
n
), y
n
y R
n
. Since y
n
is a
convergent sequence, it is contained in some closed ball C = B
R
(0); the
set f
1
(C), which contains the sequence (x
n
) is bounded (by hypothesis)
and also closed, being the counterimage of a closed set through a continu-
ous mapping. This implies that x
n
has a convergent subsequence to some
x R
n
, and by continuity,
y = lim
k
y
n
k
= lim
k
f(x
n
k
) = f
_
lim
k
x
n
k
_
= f(x),
52 CHAPTER 1. METRIC SPACE ANALYSIS.
hence y f(R
n
), which establishes closedness of f(R
n
).
1.114 Exercise. Let f : (a, b) R. Show that f is convex if and only if
(1.7)
f(y
1
) f(x
1
)
y
1
x
1

f(y
2
) f(x
2
)
y
2
x
2
,
for arbitrarily given a x
i
< y
i
b, i = 1, 2 i = 1, 2, satisfying x
1
x
2
,
y
1
y
2
. Use this fact to prove that a convex function on (a, b) is uniformly
Lipschitz continuous on every closed interval contained in (a, b).
Solution. Suppose (1.7) holds. Given x, y [a, b], and t [0, 1] suppose
without loss of generality x < y. Then, setting z = x + t(y x), since
y z = (1 t)(y x), (1.7) gives
(1 t)(f(y) f(x)) f(y) f(z) = f(z) f(x) + t(f(y) f(x)),
which is convexity. Conversely, suppose f convex, and choose x
i
, y
i
as in
(1.7). Suppose x
2
y
1
(the case x
2
< y
1
is treated similarly). Then
x
2
= x
1
+ t(y
1
x
1
), with t = (x
2
x
1
)/(y
1
x
1
). Convexity of f then
implies
f(x
2
)
1 t
f(x
1
) +
tf(y
1
)
1 t
,
hence, writing y
1
x
2
= (1 t)(y
1
x
1
),
f(y
1
) f(x
1
)
y
1
x
1

f(y
1
) f(x
2
)
y
1
x
2
.
A similar proof yields
f(y
1
) f(x
2
)
y
1
x
2

f(y
2
) f(x
2
)
y
2
x
2
,
so that, by comparing the last two inequalities, (1.7) follows.
To prove the second claim, let [c, d] (a, b) and x x, y [c, d], sup-
posing x < y without losing generality. Choose c

(a, c), d

(d, b), then


c

< x < d, c < y < d

and (1.7) entails


f(c) f(c

)
c c


f(y) f(x)
y x

f(d

) f(d)
d

d
,
which is (uniform) Lipschitz continuity.
1.115 Corollary. Exercise 1.114 immediately yields that
f

(x) = lim
h0

f(x + h) f(x)
h
1.5. ADDITIONAL TOPICS 53
exist for every x (a, b), and that f

(x) f

+
(x).
1.116 Exercise. Let x
1
< x
0
and consider the iterative method
x
n+1
= x
n
f(x
n
)
f(x
n
) f(x
0
)
x
n
x
0
.
Prove that, if f is convex and strictly increasing on [x
1
, x
0
], x
n
converges
to a limit x which is the unique solution of f(x) = 0 in [x
1
, x
0
].
Solution. Suppose x
n
converges to some x [x
0
, x
1
], and x ,= x
0
. Then,
passing to the limit and exploiting the continuity of f (which is a conse-
quence of convexity),
x = x f( x)
f( x) f(x
0
)
x x
0
,
which implies f( x) = 0.
We now show that x
n
converges to some x ,= x
0
. To prove this claim,
we show inductively that f(x
n
) < 0 for each n, and x
n
is an increasing
sequence bounded above by x
0
. This is true for n = 1. Suppose it is true
for a generic n, then
0 <
n
=
f(x
n
)
f(x
n
) f(x
0
)
< 1,
so that x
n+1
=
n
x
0
+ (1
n
)x
n
is a convex combination of x
0
and x
n
,
whence x
n
< x
n+1
< x
0
. Moreover, using convexity of f
f(x
n+1
)
n
f(x
0
) + (1
n
)f(x
n
) =
(f(x
n
))
2
f(x
n
) f(x
0
)
< 0,
so that the inductive step is complete. This shows that x
n
converges to x
and f( x) 0, whence x ,= x
0
.
1.117 Exercise. Let U be an open, bounded, connected subset of R
n
and
f : U R be of class ((U) (
2
(U). Suppose f 0 on U and that, for
all p U,

2
f
x
2
i
(p) < 0, for some i 1, . . . , n.
Prove that f 0 on U.
Solution. We show that, under the above assumptions,
min
U
f = min
U
f,
54 CHAPTER 1. METRIC SPACE ANALYSIS.
so that the sought result follows immediately. Suppose that p U is a
point where f attains its minimum. Being U open, p is a local minimum
and thus a critical point of f. Hence, f(p) = 0, and, for every v R
n
of
unit norm
0 f(p + tv) f(p) =
t
2
2
(Hf(p)v) v + o(t
2
), (Hf)
ij
=

2
f
x
i
x
j
.
Taking v = e
k
, dividing by t
2
and letting t 0 yields,

2
f
x
2
k
(p) 0, for each k = 1, . . . , n,
which leads us to a contradiction.
C
h
a
p
t
e
r
2
Multivariable Calculus
2.1 Dierentiation.
In the following discussion, U is an open subset of R
n
.
2.1 Theorem. Let f : U R
m
be of class (
1
. If the segment
[0, 1] t x(t) = p + t(q p)
is contained in U, then
[f(q) f(p)[ M[q p[, M = sup
t[0,1]
[Df(x(t))[.
Proof. Let u be a unit vector in R
m
. Dene the function g(t) = u f(x(t)),
t [0, 1]. By the mean value theorem for real-valued functions,
u (f(q) f(p)) = g(1) g(0) = g

() = u Df(x())(q p) M[q p[,


for some (0, 1), so that, by the dual characterization of the norm in
euclidean spaces, [f(q) f(p)[ M[q p[.
2.2 Theorem. Let f : U R
m
be of class (
1
. If the segment
[0, 1] t x(t) = p + t(q p)
55
56 CHAPTER 2. MULTIVARIABLE CALCULUS
is contained in U, then
f(q) f(p) = T
pq
(q p), T
pq
=
_
1
0
Df(x(t)) dt.
Conversely, if there is a continuous family of linear maps T
pq
L(R
n
, R
m
)
for which the above formula holds, f is of class (
1
and Df(p) = T
pp
.
Proof. Fix an index i in 1, . . . , m. Dene g(t) = f
i
(x(t)), t [0, 1]. By the
fundamental theorem of calculus,
f
i
(q) f
i
(p) = g(1) g(0) =
_
1
0
g

(t) dt
=
_
1
0
f
i
(x(t)) (q p) dt
=
n

j=1
__
1
0
f
i
(x(t))
x
j
dt
_
(q
j
p
j
),
which is the i-th component of T
pq
(q p).
To check the converse, assume that our formula holds for a continuous
family of linear maps T
pq
. Call q = p + v and use Taylors theorem to get
f(p + v) f(p) T
pp
v = (T
pq
T
pp
)v,
and observe that, thanks to the continuity of q T
pq
, (T
pq
T
pp
)v = o([v[).
Hence T
pp
= Df(p).
2.3 Corollary. If U is connected, and f : U R
m
is dierentiable with
Df(x) = 0 for all x U, then f is constant.
Proof. Let B = B

(x) U be an open ball. If y in B, then the segment


tx + (1 t)y is contained in B, so that by the mean value theorem,
[f(y) f(x)[ max(|Df(tx + (1 t)y)|)[y x[ = 0.
Thus f is constant on each open ball contained in U. Fix x, y U and
call r = [x y[. The set W = U B
2r
(x), which contains x and y, is
connected and with compact closure, so that there exist a sequence of open
balls B
i
= B
r
i
(x
i
), i = 1, . . . n contained in W such that x B
1
, y B
n
and B
i
B
i+1
is nonempty. This, together with the previous step, yields
the result.
2.2. IMPLICIT AND INVERSE FUNCTION. 57
2.2 Implicit and Inverse Function.
2.4 Theorem. Let U be an open subset of R
n
R
m
and f : U R
m
be
a function of class (
r
, r 1. Let (x
0
, y
0
) U with x
0
R
n
, y
0
R
m
and
z
0
= f(x
0
, y
0
). If the matrix
B
ij
=
f
i
y
j
(x
0
, y
0
), i, j = 1, . . . , m
is nonsingular, then there exists , > 0 and a unique function
g : B

(x
0
) B

(y
0
),
such that g(x
0
) = y
0
and
f(x, g(x)) = z
0
, x B

(x
0
).
Moreover, g is of class (
r
.
Proof. By translation, there is no loss in generality in assuming that (x
0
, y
0
)
is the origin in R
n
R
m
and z
0
= 0. Let A be the matrix
A
ij
=
f
i
x
j
(x
0
, y
0
), i = 1, . . . , m, j = 1, . . . n.
Then, by Taylor expansion at (0, 0),
f(x, y) = Ax + By + R(x, y), R(x, y) = o([(x, y)[).
Exploiting the nonsingularity of B, we solve for y the equation f(x, y) = 0,
obtaining
(2.1) y = B
1
(Ax + R(x, y)).
Either R(x, y) does not depend on y, in which case we reached an explicit
formula for y = g(x). Otherwise, we mean to show that, for x suciently
close to the origin, the function
K
x
(y) : y B
1
(Ax + R(x, y))
has a xed point g(x), which is automatically a solution of (2.1). Observe
that, since R is sublinear and of class C
1
, DR(0, 0) = 0, so that there is
r > 0 small enough to have
[(x, y)[ < r =
_
_
_
_
R
y
(x, y)
_
_
_
_
<
1
2|B
1
|
.
58 CHAPTER 2. MULTIVARIABLE CALCULUS
Apply the mean value theorem to get, for [(x, y
1
)[, [(x, y
2
)[ < r,
[K
x
(y
1
) K
x
(y
2
)[ |B
1
|[R(x, y
1
) R(x, y
2
)[
= |B
1
|
_
_
_
_
R
y
(x, y

)
_
_
_
_
[y
1
y
2
[
1
2
.
Moreover, since K
x
is continuous at y = 0, we have [K
x
(0)[ < 1/2 for
[x[ < < r. In the end, K
x
contracts the closure of B
r
(0
R
m) into itself, for
all x B

(0
R
n). By the Banach contraction theorem, K
x
has a unique xed
point y = g(x), so that
f(x, g(x)) = 0 x B

(0
R
n).
We now show that g is a C
1
function. We start by proving that g is Lipschitz
at zero, since
[g(x)[ = [K
x
(g(x))[ LipK
x
[g(x)[ +[K
x
(0)[

1
2
[g(x)[ +[B
1
(Ax + R(x, 0))[
1
2
[g(x)[ + 2|B

1A|[x[,
and hence [g(x)[ 4L[x[, with L = |B

1A|. By the chain rule, we need


to have Dg(0) = B
1
A if g is dierentiable at the origin. To show this
holds, write up the Taylor estimate
[g(x) g(0) + B
1
Ax[ = [B
1
R(x, g(x))[
|B
1
|(x, g(x))([x[ +[g(x)[)
|B
1
|(x, g(x))(1 + 4L)[x[,
where (x, y) goes to zero as (x, y) does, so that the remainder is sublinear
with respect to x.
Analogously, one can prove the same results for every point on the zero
locus near the origin. Observing that Dg(x) = B
1
x
A
x
for [x[ < , and
that x B
1
x
, A
x
are continuous functions, Dg is also continuous, so that
g is of class (
1
. Apply induction to get that g is (
r
.
2.5 Theorem. Let U be open in R
n
and f : U R
n
be a function of
class (
r
, r 1, and p U. If Dg(p) is a linear isomorphism, then f
is a (
r
dieomorphism from a neighborhood B

(p) onto some neighborhood


B

(f(p)).
Proof. Relies on the Implicit Function Theorem. For x U, y R
n
, set
F(x, y) = f(x) y. F is a (
r
function from U R
n
into R
n
, F(p, f(p)) = 0
2.2. IMPLICIT AND INVERSE FUNCTION. 59
and D
x
F(p) = Df(p) is nonsingular. Apply the Implicit function Theorem
to nd neighborhoods V = B

(f(p)), W = B

1
(p) and a function g : V
W such that
F(g(y), y) = f(g(y)) y = 0, y V,
which means that f g = id
V
. We now show that f bijects
W
1
= x W : f(x) V
onto V and g = f
1
. Obviously, W
1
is open, being the counterimage with
respect to f, which is a continuous function, of the open set V . Besides,
for each y V , g(y) W and f(g(y)) = y V , so that g(y) W
1
and
f(g(y)) = y. Now, for x W
1
, f(x) V and there is a unique point in
W such that F(g(f(x)), f(x)) = 0. Observe that x is such a point, since
F(x, f(x)) = 0, so that g(f(x)) = x.
Exercises.
2.6 Exercise. Let E R
n
be an open convex set, f : E R.
(1) Show that f convex on E implies that f is Lipschitz continuous on every
convex set compactly contained in E.
(2) Suppose f dierentiable on E. Show that f is convex if and only if
f(y) f(x) +f(x) (y x), x, y E.
(3) Suppose f twice dierentiable on E. Show that f is convex if and only
if H
f
is positive denite on E.
Solution. (1) We rst prove the assertion for a closed cube Q E. Call
Q
0
= x
1
, . . . , x
2
n the set of vertices of Q: then the partial left-right
derivatives of f,

i
f, which exist everywhere, are bounded on Q
0
, which is
a nite set, by some constant C. Let Q
1
be the set of points on the edges of
Q. Then f restricted to an edge is a convex function of one variable, hence

i
f are bounded by the same constant C. Let Q
2
be the set of points on
the faces of Q. Then, a point x Q
2
is a convex combination of at most
two points x
1
, x
2
of Q
1
and f restricted to the segment x
1
, x
2
is a convex
function of one variable... Iteratively proceeding this way, we show that

i
f are uniformly bounded on Q. Now, given x, x

Q, with x x

parallel to the i-th coordinate axis, we conclude that


C
f(x) f(x

)
x x

C,
60 CHAPTER 2. MULTIVARIABLE CALCULUS
since x, x

may be written as convex combination of points on the boundary


Q. To conclude that f is Lipschitz, it is enough to observe that any two
points may be joined by means of a polygonal path composed of at most n
segments parallel to the coordinate axes.
Given a generic compact convex K E, it is enough to observe that
we may cover it by a nite number of cubes.
(2) Suppose rst that f is convex and choose x, y E. For t [0, 1],
consider the (dierentiable) function
(t) = f(x + t(y x)) f(x) t(f(y) f(x)).
Then (0) = 0, and, by convexity, (t) 0 for every t [0, 1]. This, in
particular, entails

(0) 0, that is
f(x) (y x) f(y) f(x),
as desired. To prove the converse statement, x x, y E, t (0, 1) and
call z = x + t(y x). The hypothesis then reads
f(y) f(z) +f(z) (y z), f(x) f(z) +f(z) (x z),
so that
tf(y) + (1 t)f(x) f(z) +f(z) (ty + (1 t)x z) = f(z).

Anda mungkin juga menyukai