This action might not be possible to undo. Are you sure you want to continue?
∗
Steven V Sam
ssam@math.mit.edu
June 30, 2010
Contents
1 Abstract Integration 1
2 Positive Borel Measures 5
3 L
p
Spaces 13
4 Elementary Hilbert Space Theory 17
Disclaimer: A majority of these solutions have not been proofread and are bound to
be incorrect. They are to be read at the reader’s risk.
1 Abstract Integration
1. Exercise. Does there exist an inﬁnite σalgebra which has only countably many members?
Solution. The answer is no. Let X be a measurable set with an inﬁnite σalgebra M. Since Mis
inﬁnite, there exists nonempty E ∈ Mproperly contained in X. Both E and E
c
are measurable
spaces by letting the measurable subsets of E (resp. E
c
) be the intersections of measurable
subsets of X with E (resp. E
c
). Since M is inﬁnite, at least one of these two σalgebras must
be inﬁnite.
Now we deﬁne a rooted binary tree inductively as follows. The root is our set X. Given a
vertex which is a measurable subset E of X, if it contains a proper measurable subset E
, pick
one such subset, and let its two successors be E
and E ` E
. The remarks above guarantee
that this tree is inﬁnite, and hence has inﬁnite depth. So pick an inﬁnite path consisting of
subsets E
0
E
1
E
2
. . . . Then the sets F
i
= E
i
` E
i+1
form an inﬁnite collection of disjoint
nonempty measurable subsets of X by construction. At the very least, Mneeds to contain every
union of such sets, and this is in bijection with the set of subsets of N, which is uncountable.
Thus, M must be uncountable.
2. Exercise. Prove an analogue of Theorem 1.8 for n functions.
Solution. We need to prove the following: if u
1
, . . . , u
n
are real measurable functions on
a measurable space X, and Φ is a continuous map of R
n
into a topological space Y , then
h(x) = Φ(u
1
(x), . . . , u
n
(x)) is a measurable function from X to Y .
∗
third edition, by Walter Rudin
1
1 ABSTRACT INTEGRATION 2
Deﬁne f : X → R
n
by x → (u
1
(x), . . . , u
n
(x)). By Theorem 1.7(b), to prove that h is measurable,
it is enough to prove that f is measurable. If R is any open rectangle in R
n
which is the Cartesian
product of n segments I
1
, . . . , I
n
, then f
−1
(R) = u
−1
1
(I
1
) ∩ ∩ u
−1
n
(I
n
), which is measurable
since u
1
, . . . , u
n
is measurable. Finally, every open set of R
n
is the countable union of such
rectangles, so we are done.
3. Exercise. Prove that if f is a real function on a measurable space X such that ¦x [ f(x) ≥ r¦
is measurable for every rational r, then f is measurable.
Solution. Let U ⊆ R
1
be an open set. First, U can be written as a union of countably
many open balls with rational radii that are centered at rational points. So to prove that
f
−1
(U) is measurable, it is enough to prove this when U is an open ball of this form, say
with radius r and center c. Since the set of measurable sets is closed under complements and
ﬁnite intersections, every set of the form ¦x [ r
1
> f(x) ≥ r
2
¦ is measurable for rational
r
1
, r
2
. Now note that ¦x [ c + r > f(x) > c − r¦ can be written as the countable union
¸
n≥1
¦x [ c +r > f(x) ≥ c −r + 1/n¦, so f
−1
(U) is measurable.
4. Exercise. Let ¦a
n
¦ and ¦b
n
¦ be sequences in [−∞, ∞], and prove the following assertions:
(a) limsup
n→∞
(−a
n
) = −liminf
n→∞
a
n
.
(b) limsup
n→∞
(a
n
+b
n
) ≤ limsup
n→∞
a
n
+ limsup
n→∞
b
n
provided none of the sums is of the form ∞−∞.
(c) If a
n
≤ b
n
for all n, then
liminf
n→∞
a
n
≤ liminf
n→∞
b
n
.
Show by an example that strict inequality can hold in (b).
Solution. The supremum A
k
of the set ¦−a
k
, −a
k+1
, . . . ¦ is the negative of the inﬁmum A
k
of
the set ¦a
k
, a
k+1
, . . . ¦. Hence inf
k
¦A
k
¦ = −sup
k
¦A
k
¦, which implies (a).
The relation
sup¦a
k
+b
k
, a
k+1
+b
k+1
, . . . ¦ ≤ sup¦a
k
, a
k+1
, . . . ¦ + sup¦b
k
, b
k+1
, . . . ¦
is clear, so this implies (b). To see that the inequality in (b) can be strict, consider a
1
= 1, a
i
= 0
for i > 1, and b
1
= −1, b
i
= 0 for i > 1. Then limsup(a
n
+b
n
) = 0, but limsup a
n
+limsup b
n
= 1.
Now suppose that a
n
≤ b
n
for all n. Then inf¦a
k
, a
k+1
, . . . ¦ ≤ inf¦b
k
, b
k+1
, . . . ¦ for all k, so (c)
follows.
5. Exercise.
(a) Suppose f : X → [−∞, ∞] and g : X → [−∞, ∞] are measurable. Prove that the sets
¦x [ f(x) < g(x)¦, ¦x [ f(x) = g(x)¦
are measurable.
(b) Prove the set of points at which a sequence of measurable realvalued functions converges
(to a ﬁnite limit) is measurable.
1 ABSTRACT INTEGRATION 3
Solution. Let Y
+
and Y
−
be the sets where g(x) = ∞ and −∞, respectively, and deﬁne Z
+
and Z
−
analogously for f. Then these subsets are measurable: for example, Y
+
is a countable
intersection of the sets ¦x ∈ X [ g(x) ≥ n¦ as n ranges over the positive integers. Let X
be the
complement of these sets, i.e., the subset where both f and g take ﬁnite values.
So we can deﬁne the function h = f −g on X
, and it is a measurable function. The ﬁrst set of
(a) is
h
−1
([−∞, 0)) ∪ (Y
+
` Z
+
) ∪ (Z
−
` Y
−
),
so is measurable. Also, the set where f and g agree is
(X
` h
−1
([−∞, 0) ∪ (0, ∞])) ∪ (Y
+
∩ Z
+
) ∪ (Y
−
∩ Z
−
),
which is also measurable.
As for (b), let f
n
be a sequence of measurable real functions, and let E be the set of x such that
f
n
(x) converges as n → ∞. Deﬁne f = limsup f
n
. Then f is measurable (Theorem 1.14), and
f agrees with limf
n
on E. For each n, the function f −f
n
is measurable (1.22), so the set E
n,r
which is deﬁned to be the preimage of f −f
n
of (−r, r) is measurable. Then E =
¸
∞
r=1
¸
∞
n=1
E
n,r
,
so is measurable.
6. Exercise. Let X be an uncountable set, let M be the collection of all sets E ⊂ X such that
either E or E
c
is at most countable, and deﬁne µ(E) = 0 in the ﬁrst case, µ(E) = 1 in the second.
Prove that M is a σalgebra in X and that µ is a measure on M. Describe the corresponding
measurable functions and their integrals.
Solution. Since X
c
= ∅ is at most countable, X ∈ M. Also, if E ∈ M, then either E or
E
c
is at most countable, so the same is true for E
c
since (E
c
)
c
= E, and so E
c
∈ M. Now
suppose E
n
∈ M for all n, and put E =
¸
n≥1
E
n
. Let I be the set of n for which E
n
is at most
countable, and let J be the set of n for which E
n
is uncountable, but E
c
n
is at most countable,
so that E =
¸
n∈I
E
n
∪
¸
n∈J
E
n
. If J = ∅, then E is a countable union of countable sets, and
hence is countable. Otherwise, E
c
=
¸
n∈I
E
c
n
∩
¸
n∈J
E
c
n
, so E
c
⊆
¸
n∈J
E
c
n
, which is countable
since J = ∅, so E ∈ M. Thus, M is a σalgebra.
Now write a measurable set A as a disjoint union of measurable sets A
n
. If A is at most
countable, then so is each A
n
, so µ(A) =
¸
µ(A
n
) = 0. In case A
c
is at most countable,
then A is uncountable, so at least one A
i
is uncountable. Suppose that A
i
and A
j
are both
uncountable for i = j. Then A
c
i
∪ A
c
j
is countable and equal to X since A
i
and A
j
are disjoint.
But this contradicts that X is uncountable, so exactly one A
i
is uncountable, which means that
µ(A) =
¸
µ(A
n
) = 1. Hence µ is a measure on M.
The measurable functions on Mconsist of those functions f : X → R
1
such that for each r ∈ R
1
,
f
−1
(r) is either at most countable, or f
−1
(R
1
` ¦r¦) is at most countable. If we let A ⊂ R
1
denote the set of points such that f
−1
(r) is not countable, then the integral of f is
¸
r∈A
r.
7. Exercise. Suppose f
n
: X → [0, ∞] is measurable for n = 1, 2, 3, . . . ; f
1
≥ f
2
≥ f
3
≥ ≥ 0,
f
n
(x) → f(x) as n → ∞, for every x ∈ X, and f
1
∈ L
1
(µ). Prove that then
lim
n→∞
X
f
n
dµ =
X
f dµ
and show that this conclusion does not follow if the condition “f
1
∈ L
1
(µ)” is omitted.
Solution. If we ﬁrst assume that f
1
(x) < ∞ for all x, then the conclusion is a consequence of
Lebesgue’s dominated convergence theorem (Theorem 1.34) using g(x) = f
1
(x) since f
1
(x) ≥
1 ABSTRACT INTEGRATION 4
f
n
(x) ≥ 0 implies that f
1
(x) ≥ [f
n
(x)[. Otherwise, let E = ¦x ∈ X [ f
1
(x) = ∞¦. If µ(E) > 0,
then
X
[f
1
[ dµ = ∞, which contradicts f
1
∈ L
1
(µ). So we conclude that µ(E) = 0, in which
case, we can ignore E when integrating over X, and we are back to the above discussion.
Now suppose that f
1
∈ L
1
(µ) no longer holds. Take X = R
1
, and µ(E) is the length of E. Then
deﬁne f
n
(x) = ∞ for x ∈ [0, 1/n], and 0 elsewhere, so that f
n
→ 0. Then
X
f
n
dµ = ∞ for all
n, but
X
0 dµ = 0.
8. Exercise. Put f
n
= χ
E
if n is odd, f
n
= 1 − χ
E
if n is even. What is the relevance of this
example to Fatou’s lemma?
Solution. This is an example where
X
(liminf
n→∞
f
n
) dµ < liminf
n→∞
X
f
n
dµ,
provided that µ(E) > 0 and µ(X ` E) > 0. To see this, ﬁrst note that liminf f
n
(x) = 0 for
all x ∈ X because f
n
(E) = 1 for n odd, f
n
(E) = 0 for n even, and f
n
(X ` E) = 0 for n
odd, f
n
(X ` E) = 1 for n even. So the integral on the lefthand side is 0. On the other hand,
X
f
n
dµ = µ(E) if n is odd, and
X
f
n
dµ = µ(X ` E) if n is even. Hence the righthand side is
min(µ(E), µ(X ` E)) > 0.
9. Exercise. Suppose µ is a positive measure on X, f : X → [0, ∞] is measurable,
X
f dµ = c,
where 0 < c < ∞, and α is a constant. Prove that
lim
n→∞
X
nlog(1 + (f/n)
α
) dµ =
∞ if 0 < α < 1,
c if α = 1,
0 if 1 < α < ∞.
Solution. First, we ignore the set where f(x) = ∞ since f is integrable and hence this set has
measure 0. When α ≥ 1, we claim that the integrand is bounded from above by αf, so we may
use Lebesgue’s dominated convergence theorem. To see this, deﬁne a function g : [0, ∞) → R
by g(x) = αx −nlog(1 +(x/n)
α
). We need to show that g(x) ≥ 0 for all x ≥ 0. First, g(0) = 0.
Then in general, we take the derivative of g to get
g
(x) = α −n
1
1 + (x/n)
α
αx
α−1
n
α
= α −
nαx
α−1
n
α
+x
α
.
So it is enough to show that nx
α−1
≤ n
α
+x
α
. But this is clearly true: if n ≤ x, then nx
α−1
≤ x
α
,
and if n ≥ x, then nx
α−1
≤ n
α
.
For α = 1, the integrand converges to log e
f
= f, so the integral is c. If α > 1, then rewrite the
integrand as
log(1 + (f/n)
α
)
n
α
n
α−1
.
As n → ∞, the numerator goes to f
α
and the denominator goes to ∞, so the limit is 0, and
hence the integral is also 0.
If 0 < α < 1, then the integrand approaches ∞ since n
α−1
→ 0 as n → ∞. Hence by Fatou’s
lemma, the limit of the integral is also inﬁnite.
2 POSITIVE BOREL MEASURES 5
10. Exercise. Suppose µ(X) < ∞, ¦f
n
¦ is a sequence of bounded complex measurable functions
on X, and f
n
→ f uniformly on X. Prove that
lim
n→∞
X
f
n
dµ =
X
f dµ,
and show that the hypothesis “µ(X) < ∞” cannot be omitted.
Solution. Since f
n
→ f uniformly, there exists N such that n ≥ N implies that [f
n
(x)−f(x)[ <
1 for all x ∈ X. Then since ¦f
1
, . . . , f
N−1
¦ is ﬁnite and consists of bounded sets, we can take C to
be the largest absolute value any of them obtains, and let C
be the maximum of the largest value
of [f(x)±1[ and C. Then C
≥ [f
n
(x)[ for all n, and C
∈ L
1
(µ) because
X
C
dµ = C
µ(X) < ∞.
So by Theorem 1.34,
lim
n→∞
X
f
n
dµ =
X
f dµ.
To see that µ(X) < ∞ is necessary, let X = R
1
with the usual measure. Deﬁne f
n
to be
the constant function 1/n. Then f
n
→ 0 uniformly, but the
X
f
n
dµ = ∞ for all n, while
X
0 dµ = 0, so the equality does not hold.
11. Exercise. Show that
A =
∞
¸
n=1
∞
¸
k=n
E
k
in Theorem 1.41, and hence prove the theorem without any reference to integration.
Solution. Denote the righthand side by B. Recall that A is the set of all x which lie in
inﬁnitely many E
k
. Pick x ∈ A. Then x ∈
¸
∞
k=n
E
k
for all k, so x ∈ B. If x / ∈ A, then x is
contained in ﬁnitely many E
k
, say ¦E
i
1
, . . . , E
ir
¦ with i
1
< < i
r
. So x / ∈
¸
∞
k=ir+1
E
k
, which
means x / ∈ B, and hence A = B.
Now set B
n
=
¸
∞
k=n
E
k
. Then µ(B
1
) < ∞ by assumption, and B
1
⊃ B
2
⊃ . By Theo
rem 1.19(e), µ(B
n
) → µ(B) as n → ∞. Since µ(B
n
) ≤
¸
∞
k=n
µ(E
k
), and the bounding sum
approaches 0 as n → ∞, we get that µ(B) = 0.
13. Exercise. Show that Proposition 1.24(c) is also true for c = ∞.
Solution. We wish to prove that if f ≥ 0, then
E
∞f dµ = ∞
E
f dµ.
Let F be the set where f is nonzero. Then F is measurable, being the preimage of an open set,
and we can integrate over F instead of E and get the same result since ∞ 0 is deﬁned to be 0.
Then the integral over F is ∞ on both sides of the above equation, so we are done.
2 Positive Borel Measures
1. Exercise. Let ¦f
n
¦ be a sequence of real nonnegative functions on R
1
, and consider the following
four statements:
(a) If f
1
and f
2
are upper semicontinuous, then f
1
+f
2
is upper semicontinuous.
(b) If f
1
and f
2
are lower semicontinuous, then f
1
+f
2
is lower semicontinuous.
2 POSITIVE BOREL MEASURES 6
(c) If each f
n
is upper semicontinuous, then
¸
∞
n=1
f
n
is upper semicontinuous.
(d) If each f
n
is lower semicontinuous, then
¸
∞
n=1
f
n
is lower semicontinuous.
Show that three of these are true and that one is false. What happens if the word “nonnegative”
is omitted? Is the truth of the statements aﬀected if R
1
is replaced by a general topological
space?
Solution. First suppose that f
1
and f
2
are upper semicontinuous. The set ¦x ∈ R
1
[ f
1
(x) +
f
2
(x) < α¦ is the union of the sets ¦x ∈ R
1
[ f
1
(x) < β¦ ∩ ¦x ∈ R
1
[ f
2
(x) < α − β¦ where
we range over all β ≤ α. Hence this set is open, so f
1
+ f
2
is upper semicontinuous. If both f
1
and f
2
are instead lower semicontinuous, then an analogous argument shows that f
1
+f
2
is also
lower semicontinuous. We have not used that the functions are nonnegative here, nor have we
used that f
1
and f
2
are deﬁned on R
1
.
Now suppose that we have a sequence ¦f
n
¦ of lower semicontinuous functions. For any given
α ∈ R, we claim that
¦x ∈ R
1
[
¸
i
f
i
(x) > α¦ =
¸
n≥1
¦x ∈ R
1
[
n
¸
i=1
f
i
(x) > α¦.
To see this, ﬁrst pick x such that
¸
i
f
i
(x) > α. Since the partial sums
¸
n
i=1
f
i
(x) converge
to
¸
i
f
i
(x), there must be some N such that
¸
N
i=1
f
i
(x) > α. This proves that the LHS is
contained in the RHS. Conversely, if
¸
n
i=1
f
i
(x) > α, then the same is true for
¸
m
i=1
f
i
(x)
whenever m ≥ n by the fact that the f
i
are nonnegative functions, so
¸
i
f
i
(x) > α.
Now by (a), each partial sum
¸
n
i=1
f
i
(x) is lower semicontinuous, so the set above is open, which
implies that
¸
f
n
is lower semicontinuous. Here we have we not used that these functions are
deﬁned on R
1
. But if we drop the nonnegativity requirement, we can construct a counterexample
as follows: let f
0
(x) = −1 if [x[ ≥ 1 and 0 otherwise, and for i > 0, let f
i
(x) be the function
which is −1 on [−
1
i
, −
1
i+1
] ∪ [
1
i+1
,
1
i
] and 0 elsewhere. Then f
i
(x) is lower semicontinuous for
all i, and the sum f(x) =
¸
i
f
i
(x) is 0 at x = 0 and negative elsewhere, so is not a lower
semicontinous function.
However, (c) is a false statement. Deﬁne f
1
(x) = 0 on (−1, 1) and f
1
(x) = 1 on the rest of R
1
.
For n > 1, deﬁne f
n
(x) = 1 on
1
n
,
1
n−1
∪
−
1
n−1
, −
1
n
and 0 elsewhere. Then each f
n
is upper
semicontinuous since the set of x such that f
n
(x) = 0 is open. However,
¸
f
n
is 0 at 0 and
greater than 0 elsewhere, so is not upper semicontinuous.
2. Exercise. Let f be an arbitrary complex function on R
1
, and deﬁne
ϕ(x, δ) = sup¦[f(s) −f(t)[ [ s, t ∈ (x −δ, x +δ)¦,
ϕ(x) = inf¦ϕ(x, δ) [ δ > 0¦.
Prove that ϕ is upper semicontinuous, that f is continuous at a point x if and only if ϕ(x) = 0,
and hence that the set of points of continuity of an arbitrary complex function is a G
δ
.
Formulate and prove an analogous statement for general topological spaces in place of R
1
.
Solution. We formulate the general statement and prove that. Let X be a topological space,
let f : X → C be an arbitrary function, and deﬁne
ϕ(x) = inf
Ux
sup¦[f(s) −f(t)[ [ s, t ∈ U¦,
2 POSITIVE BOREL MEASURES 7
where U ranges over open sets containing x. Then ϕ is upper semicontinuous and f is continuous
at x if and only if ϕ(x) = 0.
Pick a real number α, and consider the set E = ¦x ∈ X [ ϕ(x) < α¦. Pick x ∈ E and ε > 0 such
that ϕ(x)+ε < α. Then there exists an open set U ÷ x such that sup¦[f(s)−f(t)[ [ s, t ∈ U¦ < α.
In particular, this means that for every t ∈ U, ϕ(t) < α. So E is open, and hence ϕ is upper
semicontinuous.
Now suppose that f is continuous at x. Then for every ε > 0, there is a neighborhood U
ε
÷ x
such that f(U
ε
) ⊂ B
ε
(f(x)), where B
ε
(f(x)) denotes the ball of radius ε around f(x). In
particular, this means that ϕ(x) < ε, so we conclude that ϕ(x) = 0. Conversely, suppose that
ϕ(x) = 0. To show that f is continuous at x, it is enough to show that for every ε > 0, there is
an open set U
ε
÷ x such that f(U
ε
) ⊂ B
ε
(f(x)), but this is clear from the deﬁnition.
We conclude that the set of points for which f is continuous is a G
δ
since it is in the countable
intersection of open sets
¸
n≥0
E
n
where E
n
= ¦x ∈ X [ ϕ(x) < 1/n¦.
3. Exercise. Let X be a metric space, with metric ρ. For any nonempty E ⊂ X, deﬁne
ρ
E
(x) = inf¦ρ(x, y) [ y ∈ E¦.
Show that ρ
E
is uniformly continuous function on X. If A and B are disjoint nonempty closed
subsets of X, examine the relevance of the function
f(x) =
ρ
A
(x)
ρ
A
(x) +ρ
B
(x)
to Urysohn’s lemma.
Solution. Pick ε > 0, and put δ = ε/2. We claim that if ρ(x, y) < δ, then [ρ
E
(x) −ρ
E
(y)[ < ε
for all x, y ∈ X. We can ﬁnd z ∈ E such that ρ(y, z) < ρ
E
(y) +δ by deﬁnition of ρ
E
. Then
ρ(x, y) +ρ
E
(y) +δ > ρ(x, y), +ρ(y, z) ≥ ρ(x, z) ≥ ρ
E
(x),
so
ρ(x, y) +δ > ρ
E
(x) −ρ
E
(y).
By symmetry, we conclude that
ρ(x, y) +δ > [ρ
E
(x) −ρ
E
(y)[.
But the lefthand side is less than ε, so we have established that ρ
E
is a uniformly continuous
function on X.
Now let A and B be disjoint nonempty closed subsets of X, and consider f as deﬁned above.
Then f(x) = 1 for x ∈ B, f(x) = 0 for x ∈ A, and f(x) ≤ 1 on X`(A∪B). Then χ
B
≤ f ≤ χ
X\A
,
so that this is an analogous result to Urysohn’s lemma.
4. Exercise. Examine the proof of the Riesz theorem and prove the following two statements:
(a) If E
1
⊂ V
1
and E
2
⊂ V
2
, where V
1
and V
2
are disjoint open sets, then µ(E
1
∪ E
2
) =
µ(E
1
) +µ(E
2
), even if E
1
and E
2
are not in M.
(b) If E ∈ M
F
, then E = N ∪ K
1
∪ K
2
∪ , where ¦K
i
¦ is a disjoint countable collection of
compact sets and µ(N) = 0.
2 POSITIVE BOREL MEASURES 8
Solution. Recall that the deﬁnition is µ(E
i
) = inf¦µ(V ) [ E
i
⊂ V, V open¦. Note that Step I
of the proof of the Riesz theorem does not use the fact that E
i
∩K ∈ M
F
for every compact set
K. Since this is the only diﬀerence between sets in M and sets not in M, the proof follows just
as before, so µ(E
1
∪ E
2
) ≤ µ(E
1
) + µ(E
2
). Conversely, let U be a subset containing E
1
∪ E
2
.
Then µ(U) = µ(U ∩ V
1
) + µ(U ∩ V
2
) ≥ µ(E
1
) + µ(E
2
), where the ﬁrst equality follows since V
1
and V
2
are disjoint. Hence µ(E
1
∪ E
2
) ≥ µ(E
1
) + µ(E
2
) by deﬁnition, and we have established
(a).
Now pick E ∈ M
F
, and set E
0
= E. By Step V of the proof of the Riesz theorem, there is
a compact set K
1
and an open set V
1
such that K
1
⊂ E
0
⊂ V
1
and µ(V
1
` K
1
) < 1. Then
E
0
` K
1
∈ M
F
by Step VI, so set E
1
= E
0
` K
1
. Inductively, we can ﬁnd a compact set K
n
and
open set V
n
such that K
n
⊂ E
n−1
⊂ V
n
and µ(V
n
` K
n
) < 1/n, and deﬁne E
n
= E
n−1
` K
n
.
Then set N = E `
¸
n≥1
K
n
. Then µ(N) < 1/n for all n, so µ(N) = 0, and we have (b).
5. Exercise. Let E be Cantor’s familiar “middle thirds” set. Show that m(E) = 0, even though
E and R
1
have the same cardinality.
Solution. First deﬁne E
1
= [0, 1], and inductively deﬁne E
n
to be the the result of removing
the open middle third of each connected component of E
n−1
. Then m(E
n
) = (2/3)
n−1
. Letting
E =
¸
n≥0
E
n
, we then get m(E) = 0. However, E contains uncountably many points because
each decimal in base 3 with either no 1’s or exactly one 1 at the end is an element of E.
6. Exercise. Construct a totally disconnected compact set K ⊂ R
1
such that m(K) > 0.
If v is lower semicontinuous and v ≤ χ
K
, show that actually v ≤ 0. Hence χ
K
cannot be approx
imated from below by lower semicontinuous functions, in the sense of the Vitali–Carath´eodory
theorem.
Solution. Deﬁne K
0
= [0, 1], and inductively deﬁne K
n
to be K
n−1
with an open interval of
length 2
−2n
removed from the middle of each connected component, then take K =
¸
n≥0
K
n
.
Since K
n
has 2
n
connected components, we see that
m(K
n
) = 1 −
n
¸
i=1
1
2
2i
2
i−1
= 1 −
n
¸
i=1
1
2
i+1
= 1 −
1 −
1
2
n+2
1 −
1
2
−1 −
1
2
,
so m(K) = lim
n→∞
m(K
n
) = 1/2. Furthermore, K is bounded and closed, since it is the intersection
of closed sets, so K is compact. Finally, K is totally disconnected: if there were a connected
component of K consisting of more than a point, then K contains an interval (a, b) for a < b.
But for n suﬃciently large, 2
−2n
< b − a, so we have a contradiction. Hence K is also totally
disconnected.
Now let v be a lower semicontinuous function with v ≤ χ
K
. The set of x where v(x) > 0 lies
inside of K and is open, so must be empty, because K has no interior. So v ≤ 0.
7. Exercise. If 0 < ε < 1, construct an open set E ⊂ [0, 1] which is dense in [0, 1], such that
m(E) = ε.
Solution. Note that in (Ex. 2.6), we could have replaced
1
2
with an arbitrary number in (0, 1)
(start with a smaller or larger set for K
1
). Then we just need to take the complement in [0, 1]
to get the desired example.
9. Exercise. Construct a sequence of continuous functions f
n
on [0, 1] such that 0 ≤ f
n
≤ 1, such
that
lim
n→∞
1
0
f
n
(x) dx = 0,
2 POSITIVE BOREL MEASURES 9
but such that the sequence ¦f
n
(x)¦ converges for no x ∈ [0, 1].
Solution. For a given n, deﬁne n functions g
n,i
for i = 0, . . . , n−1 by g
n,i
(x) = 1 on
i
n
,
i+1
n
and
deﬁne g
n,i
(x) = nx−(i −1) on
i−1
n
,
i
n
and g
n,i
(x) = −nx+i +2 on
i+1
n
,
i+2
n
, and 0 elsewhere.
Then think of g
n,i
as functions on [0, 1], note that they are continuous. Let ¦f
1
, f
2
, . . . , ¦ be the
sequence ¦g
1,0
, g
2,0
, g
2,1
, . . . , g
n,0
, . . . , g
n,n−1
, . . . ¦. Then
1
0
g
n,i
dx = 2/n if 0 < i < n − 1 and
otherwise the integral is equal to 3/2n. Hence
1
0
f
n
(x) dx → 0 as n → ∞. However, ¦f
n
(x)¦
does not converge for any x ∈ [0, 1] because there are inﬁnitely many values of n for which
f
n
(x) = 1 and inﬁnitely many values of n for which f
n
(x) = 0 for each x ∈ [0, 1].
11. Exercise. Let µ be a regular Borel measure on a compact Hausdorﬀ space X; assume µ(X) = 1.
Prove that there is a compact set K ⊂ X (the carrier or support of µ) such that µ(K) = 1 but
µ(H) < 1 for every proper compact subset H of K.
Solution. Let K be the intersection of all compact K
α
such that µ(K
α
) = 1. Each K
α
is
compact, and hence closed since X is Hausdorﬀ, so K is closed, and hence compact since X is
compact.
Let V be an open set which contains K. Then V
c
is closed and hence compact. Since the K
α
are compact, they are closed, so K
c
α
∩ V
c
forms an open cover of V
c
, and by compactness, we
can write V
c
= (K
c
1
∪ ∪ K
c
n
) ∩ V
c
. Since µ(K
c
1
) = 0, this shows that µ(V
c
) = 0, and hence
µ(V ) = 1. So all open sets containing K have measure 1, which implies µ(K) = 1 since µ is
regular. Finally, if H is a compact set properly contained in V , then µ(H) < 1. If not, then
µ(H) = 1, which contradicts the deﬁnition of K.
13. Exercise. Is it true that every compact subset of R
1
is the support of a continuous function?
If not, can you describe the class of all compact sets in R
1
which are supports of continuous
functions? Is your description valid in other topological spaces?
Solution. A point is a compact subset of R
1
, but cannot be the support of any continuous
function. Any nonzero continuous function f has some real number α in its image, and hence
its support is either all of R
1
. Otherwise, f contains 0 in its image, and so its support must
contain an open interval f
−1
(0, α). Restricting to each connected component of Supp f, we see
that each component needs to contain an open interval. Conversely, if K is a compact set with
nonempty interior, it is the support of some continuous function. To construct such a function,
we need only construct it on each connected component, so assume that K is connected. Let ϕ
be a homeomorphism of the interior of K to R
1
. Then we can deﬁne f : R
1
→ R
1
by x → e
−x
2
,
and we can extend the function f ◦ ϕ to K by deﬁning it to be 0 at the end points of K. It is
clear that the support of this function is K.
However, this description will not carry to general topological spaces, namely because compact
sets need not be closed. For an example, take the indiscrete topology on a set X = ¦a, b, c¦
with open sets ¦∅, ¦a¦, X¦. Then ¦a, b¦ is a compact set with nonempty interior, but it is not
closed.
14. Exercise. Let f be a realvalued Lebesgue measurable function on R
k
. Prove that there exist
Borel functions g and h such that g(x) = h(x) a.e. [m], and g(x) ≤ f(x) ≤ h(x) for every
x ∈ R
k
.
Solution. Pick I = (i
1
, . . . , i
k
) ∈ Z
k
, we will deﬁne F on the box B
I
= (i
1
, i
1
+1] (i
k
, i
k
+1].
Let E
n
be the set f
−1
((n, n + 1]) ∩ B
I
for all n ∈ Z. Then [f[ is bounded on E
n
, so there exist
compactly supported continuous functions F
n,r
(x) such that if we deﬁne F
n
= lim
r→∞
F
n,r
, then
2 POSITIVE BOREL MEASURES 10
F
n
= f a.e. on E
n
by the corollary to Theorem 2.24. Since continuous functions are Borel
measurable, F
n
is Borel measurable (Corollary to Theorem 1.14). Finally, deﬁne F on I to be
F
n
on the set E
n
. Then F is also Borel measurable. We repeat for every I ∈ Z
n
and deﬁne a
global function in this fashion. The set where F and f diﬀer is then a countable union of sets
of measure zero, so f = F a.e.
Let X = ¦x [ f(x) = F(x)¦. We can partition X into measurable sets X
n
= (f −F)
−1
((n, n +
1]) ∩ X. Now deﬁne a function ϕ: R
k
→ R
1
by ϕ(x) = n if x ∈ X
n
, and ϕ(x) = 0 if x / ∈ X.
Then ϕ is a measurable function. Now deﬁne g = F +ϕ and h = F +ϕ+χ
X
. Then we see that
g and h are Borel measurable, that g = h a.e., and that g(x) ≤ f(x) ≤ h(x) for all x ∈ R
k
.
15. Exercise. It is easy to guess the limits of
n
0
1 −
x
n
n
e
x/2
dx and
n
0
1 +
x
n
n
e
−2x
dx,
as n → ∞. Prove that your guesses are correct.
Solution. Deﬁne a function f
n
to be
1 −
x
n
n
e
x/2
on [0, n] and 0 for x > n. Then f
n
→ e
−x/2
as n → ∞, and furthermore, e
−x/2
≥ [f
n
(x)[ for all n, and e
−x/2
∈ L
1
(R
1
). So by Lebesgue’s
dominated convergence theorem,
lim
n→∞
∞
0
f
n
dx =
∞
0
e
−x/2
dx = 2,
and the lefthand side is the ﬁrst limit to compute. By similar considerations, the second integral
is
∞
0
e
−x
dx = 1.
16. Exercise. Why is m(Y ) = 0 in the proof of Theorem 2.20(e)?
Solution. In this case, Y lies in a proper linear subspace of R
k
. It is easy to see that the
measure of a proper linear subspace must be 0 because we can take arbtrarily thin open sets
that contain the subspace.
17. Exercise. Deﬁne the distance between two points (x
1
, y
1
) and (x
2
, y
2
) in the plane to be
[y
1
−y
2
[ if x
1
= x
2
, 1 +[y
1
−y
2
[ if x
1
= x
2
.
Show that this is indeed a metric, and that the resulting metric space X is locally compact.
If f ∈ C
c
(X), let x
1
, . . . , x
n
be those values of x for which f(x, y) = 0 for at least one y, and
deﬁne
Λf =
n
¸
j=1
∞
−∞
f(x
j
, y) dy.
Let µ be the measure associated with this Λ by Theorem 2.14. If E is the xaxis, show that
µ(E) = ∞ although µ(K) = 0 for every compact K ⊂ E.
Solution. Let ρ be the metric deﬁned. It is obvious that ρ(x, y) ≥ 0 for all x, y, and that
ρ(x, y) = 0 if and only if x = y. It is also obvious that ρ(x, y) = ρ(y, x). We just need to verify
the triangle inequality. Let α = (x
1
, y
1
), β = (x
2
, y
2
), and γ = (x
3
, y
3
). Then
ρ(α, γ) ≤ 1 +[y
1
−y
3
[ ≤ [y
1
−y
2
[ + 1 +[y
2
−y
3
[ ≤ ρ(α, β) +ρ(β, γ),
2 POSITIVE BOREL MEASURES 11
so ρ is a metric. Note that a set under ρ is open if and only if its intersection with each vertical
line is open when considered as a copy of R
1
. Every point in the plane has an open neighborhood
in its vertical line whose closure is compact when thought of as a set in R
1
. But each vertical
line is closed since it is the complement of the other vertical lines, so the closure of such a
neighborhood is the same as the closure when considering it as a subset of R
1
. So X is locally
compact.
Note that if f ∈ C
c
(X), then the values of x for which f(x, y) = 0 for at least one y must be
ﬁnite because any collection of vertical lines is open, and hence only a ﬁnite union of vertical
lines can be compact. The fact that Λ is a linear functional follows from the fact that one could
sum over all x ∈ R
1
and not change the value of Λ, and the fact that the integral is a linear
functional. Also, Λf < ∞ because f is compactly supported, and such a compact set in X is a
union of closed sets in ﬁnitely many vertical lines.
Now let E be the xaxis. Any open set V containing E must contain a segment of the form
((x, −δ(x)), (x, δ(x))) for each x ∈ R
1
where δ(x) > 0. To show that µ(E) = ∞, it is enough to
show that the open set U
δ
= ¦(x, y) [ −δ(x) < y < δ(x)¦ has inﬁnite measure for all arbitrary
functions δ : R
1
→ R
1
>0
. To see this, consider the sets X
1
= [1, ∞), and X
n
=
1
n
,
1
n−1
for
n = 2, 3, . . . . Then for all x ∈ R
1
, δ(x) lies in some set. Since we have countably many sets, and
R
1
is uncountable, there is some set X
i
such that δ
−1
(X
i
) is inﬁnite. In particular, let x
1
, x
2
, . . .
be distinct values inside of X
i
. Then for each n, we can ﬁnd a compactly supported continuous
function f that is 1 for points of the form (x, y) where x ∈ ¦x
1
, . . . , x
n
¦ and −δ(x)/2 < y <
δ(x)/2. Then Λf ≥
n
2i
. In particular, as n → ∞, this shows that µ(U
δ
) = ∞, so µ(E) = ∞.
However, every compact set K contained in E must be a ﬁnite set of points ¦(x
1
, 0), . . . , (x
r
, 0)¦,
so µ(K) = 0 necessarily because for all n, the set U
n
= ¦(x, y) [ x ∈ ¦x
1
, . . . , x
r
¦, −
1
n
< y <
1
n
¦
contains K and has measure µ(U
n
) = r/n.
20. Exercise. Find continuous functions f
n
: [0, 1] → [0, ∞) such that f
n
(x) → 0 for all x ∈ [0, 1]
as n → ∞,
1
0
f
n
(x) dx → 0, but sup
n
f
n
is not in L
1
.
Solution. For each n, deﬁne n
2
functions g
n,r
: R
1
→ [0, ∞) for r = 0, . . . , n
2
−1 by g
n,r
(x) = n
on
i
n
2
,
i+1
n
2
, g
n,r
(x) = n(n
2
x−i+1) on
i−1
n
2
,
i
n
2
, and g
n,r
(x) = n(i+2−n
2
x) on
i+1
n
2
,
i+2
n
2
. Then
1
0
g
n,r
(x) dx ≤
2
n
. Letting the sequence ¦f
1
, f
2
, . . . , ¦ be ¦g
1,0
, g
2,0
, . . . , g
2,3
, . . . , g
n,0
, . . . , g
n,n
2
−1
, . . . ¦,
we get
1
0
f
n
(x) dx → 0, and sup
n
f
n
= ∞, so is not L
1
.
21. Exercise. If X is compact and f : X → (−∞, ∞) is upper semicontinuous, prove that f attains
its maximum at some point of X.
Solution. The sets f
−1
((−∞, α)) are open for all α ∈ R
1
, so by compactness, there are ﬁnitely
many that cover X, and hence f is bounded. In particular, we only need to take one such set.
Now let α = sup¦f(x) [ x ∈ X¦. We claim that f(x) = α for some x ∈ X. If not, then we can
ﬁnd some sequence α
n
such that 0 < α−α
n
<
1
n
. In particular, the sets f
−1
((−∞, α−
1
n
)) cover
X, and there is no ﬁnite subcover, which is a contradiction. Hence f attains its maximum at
some point of X.
22. Exercise. Suppose that X is a metric space, with metric d, and that f : X → [0, ∞] is lower
semicontinuous, f(p) < ∞ for at least one p ∈ X. For n = 1, 2, 3, . . . ; x ∈ X, deﬁne
g
n
(x) = inf¦f(p) +nd(x, p) [ p ∈ X¦
and prove that
2 POSITIVE BOREL MEASURES 12
(i) [g
n
(x) −g
n
(y)[ ≤ nd(x, y),
(ii) 0 ≤ g
1
≤ g
2
≤ ≤ f,
(iii) g
n
(x) → f(x) as n → ∞, for all x ∈ X.
Thus f is the pointwise limit of an increasing sequence of continuous functions.
Solution. Pick x, y ∈ X and p ∈ X with f(p) < ∞. Without loss of generality, suppose that
g
n
(y) ≥ g
n
(x). We have
nd(x, y) ≥ nd(y, p) −nd(x, p) = nd(y, p) +f(p) −nd(x, p) −f(p) ≥ g
n
(y) −g
n
(x),
so this shows (i).
It is clear that g
1
≥ 0. Also, f(p)+nd(x, p) ≥ f(p)+(n−1)d(x, p) ≥ g
n−1
(x), so g
n−1
(x) ≤ g
n
(x)
for all x. Furthermore, if f(x) = ∞, then f ≥ g
n
for all n. Otherwise, g
n
(x) = f(x) by taking
p = x, so f ≥ g
n
for all n in this case, too. So (ii) is established.
Again, if f(x) = ∞, then d(x, p) > 0 for all p with f(p) < ∞, so g
n
(x) → ∞ as n → ∞.
Otherwise, g
n
(x) = f(x) for all n, so in both cases we have g
n
→ f as n → ∞.
24. Exercise. A step function is, by deﬁnition, a ﬁnite linear combination of characteristic functions
of bounded intervals in R
1
. Assume f ∈ L
1
(R
1
), and prove that there is a sequence ¦g
n
¦ of
step functions so that
lim
n→∞
∞
−∞
[f(x) −g
n
(x)[ dx = 0.
Solution. Let g
n
=
¸
n
2
i=−n
2 f(i)χ
[
i
n
,
i+1
n
]
. One can partition R
1
with sets E
r
for r ∈ Z where
r ≤ f(x) < r + 1 for x ∈ E
r
. For a given r, for n suﬃciently large, on each the subintervals of
length we can bound [f(x)−g
n
(x)[ on their intersection E
r
by an error directly proportional to r
and inversely proportional to n. We omit the precise details. Since
[f[ dx < ∞, the expression
S
n
=
−n
−∞
[f[ dx +
∞
n
[f[ dx → 0
as n → ∞. Hence for a given ε, we can choose n large enough so that S
n
< ε, and also so that
n
−n
[f −g[ dx < ε. This is enough to guarantee that
∞
−∞
[f −g[ dx < 2ε by noting that
S
n
≤
−n
−∞
[f −g[ dx +
∞
n
[f −g[ dx.
From this it is clear that the limit of the integral above is 0.
25. Exercise.
(i) Find the smallest constant c such that
log(1 +e
t
) < c +t (0 < t < ∞).
(ii) Does
lim
n→∞
1
n
1
0
log(1 +e
nf(x)
) dx
exist for every real f ∈ L
1
? If it exists, what is it?
3 L
p
SPACES 13
Solution. Since exp is an increasing function, log(1 + e
t
) < c + t becomes 1 + e
t
< e
c+t
, from
which we divide by e
t
to get e
−t
+ 1 < e
c
, and ﬁnally, we apply log (which is an increasing
function) to get log(e
−t
+ 1) < c. The lefthand side is decreasing with t, so the smallest c
satisfying this inequality is lim
t→0
log(e
−t
+ 1) = log 2.
Let X ⊂ [0, 1] be the set where f(x) ≥ 0. This implies
X
log(1 +e
nf(x)
) dx ≤
X
(log 2 +nf(x)) dx = log 2 +
X
nf(x) dx,
so as n → ∞, the integral becomes
X
f(x) dx since the lefthand side increases toward the
second term as n → ∞, and the integral approaches 0 on [0, 1] ` X as n → ∞.
3 L
p
Spaces
1. Exercise. Prove that the supremum of any collection of convex functions on (a, b) is convex on
(a, b) (if it is ﬁnite) and that pointwise limits of sequences of convex functions are convex. What
can you say about upper and lower limits of sequences of convex functions?
Solution. Let ¦f
α
¦ be a collection of convex functions on (a, b), let f = sup
α
f
α
, and assume
that f is ﬁnite. Pick λ ∈ [0, 1] and x, y ∈ (a, b). Then for all α, we have
(1 −λ)f(x) +λf(y) ≥ (1 −λ)f
α
(x) +λf
α
(y) ≥ f
α
((1 −λ)x +λy).
So by deﬁnition of supremum, we conclude that (1 − λ)f(x) + λf(y) ≥ f((1 − λ)x + λy), and
hence f is convex.
Now let ¦f
n
¦ be a sequence of convex functions that converges pointwise to f. Pick λ ∈ [0, 1]
and x, y ∈ (a, b). Since
(1 −λ)f
n
(x) +λf
n
(y) → (1 −λ)f(x) +λf(y)
and
f
n
((1 −λ)x +λy) → f((1 −λ)x +λy)
as n → ∞, and we have
(1 −λ)f
n
(x) +λf
n
(y) ≥ f
n
((1 −λ)x +λy)
for all n, we conclude that
(1 −λ)f
n
(x) +λf
n
(y) ≥ f
n
((1 −λ)x +λy),
so that f is convex.
Now consider the sequence of functions on (0, 2) deﬁned by f
n
(x) = x if n is even and f
n
(x) =
2 −x if n is odd. Then the lower limit f is deﬁned by f(x) = x if x ∈ [0, 1] and f(x) = 2 −x if
x ∈ [1, 2], and this is not convex: pick x = 1/2, y = 3/2, λ = 1/2. Then f(x)/2 +f(y)/2 = 1/2,
but f(1) = 1. So we conclude that the lower limit of a sequence of convex functions need not
be convex.
However, the upper limit of convex functions will be convex. The proof is similar to the proof
for pointwise convergent functions, except that we use that for any ε > 0, there exist inﬁnitely
many values of n (rather than all suﬃciently large n) for which (1 −λ)f
n
(x) +λf
n
(y) is within
ε of (1 −λ)f(x) +λf(y), and similarly for f
n
((1 −λ)x +λy) and f((1 −λ)x +λy).
3 L
p
SPACES 14
2. Exercise. If ϕ is convex on (a, b) and if ψ is convex and nondecreasing on the range of ϕ, prove
that ψ ◦ ϕ is convex on (a, b). For ϕ > 0, show that the convexity of log ϕ implies the convexity
of ϕ, but not vice versa.
Solution. Pick λ ∈ [0, 1] and x, y ∈ (a, b). Then (1 −λ)ϕ(x) + λϕ(y) ≥ ϕ((1 −λ)x + λy), and
since ψ is convex and nondecreasing,
(1 −λ)ψ(ϕ(x)) +λψ(ϕ(x)) ≥ ψ((1 −λ)ϕ(x) +λϕ(y)) ≥ ψ(ϕ((1 −λ)x +λy)),
so ψ◦ϕ is convex on (a, b). For ϕ > 0, the convexity of log ϕ implies the convexity of exp ◦ log ϕ =
ϕ since exp is a nondecreasing and convex function. However, the converse is not true: the
identity function x is convex, but log x is not convex.
3. Exercise. Assume that ϕ is a continuous real function on (a, b) such that
ϕ
x +y
2
≤
1
2
ϕ(x) +
1
2
ϕ(y)
for all x, y ∈ (a, b). Prove that ϕ is convex.
Solution. Pick λ ∈ [0, 1] and x, y ∈ (a, b). Without loss of generality, assume that ϕ(y) ≥ ϕ(x).
By repeated iterations of the above inequality, if λ is a rational number whose denominator is a
power of 2, then we can conclude that ϕ((1 − λ)x + λy) ≤ (1 − λ)ϕ(x) + λy. The general case
follows by continuity of ϕ: we can arbitrarily approximate ϕ((1 −λ)x +λy) by ϕ((1 −r)x +ry)
where r is some rational number whose denominator is a power of 2, and we can choose r such
that (1 −r)ϕ(x) +rϕ(y) ≤ (1 −λ)ϕ(x) +λϕ(y).
10. Exercise. Suppose f
n
∈ L
p
(µ), for n = 1, 2, 3, . . . , and f
n
− f
p
→ 0 and f
n
→ g a.e., as
n → ∞. What relation exists between f and g?
Solution. Since f
n
− f
p
→ 0, we know that f
n
→ f a.e. Let E be the set where limf
n
= g
and let F be the set where limf
n
= f. Then f = g except possibly on E∪F, which has measure
0. Hence f = g a.e.
11. Exercise. Suppose µ(Ω) = 1, and suppose f and g are positive measurable functions on Ω such
that fg ≥ 1. Prove that
Ω
f dµ
Ω
g dµ ≥ 1.
Solution. First note that fg ≥ 1 implies that
√
fg ≥ 1. Using the Cauchy–Schwarz inequality
on
√
f and
√
g, we get
Ω
f dµ
Ω
g dµ ≥
Ω
fg dµ
2
≥
Ω
1 dµ
2
= µ(Ω)
2
= 1.
12. Exercise. Suppose µ(Ω) = 1 and h: Ω → [0, ∞] is measurable. If
A =
Ω
hdµ,
prove that
1 +A
2
≤
Ω
1 +h
2
dµ ≤ 1 +A.
3 L
p
SPACES 15
If µ is Lebesgue measure on [0, 1] and h is continuous, h = f
, the above inequalities have a
simple geometric interpretation. From this, conjecture (for general Ω) under what conditions on
h equality can hold in either of the above inequalities, and prove your conjecture.
Solution. The function ϕ(x) =
√
1 +x
2
is a convex function because its second derivative
√
x
2
+1
x
4
+2x
2
+1
is always positive. Hence the ﬁrst inequality follows from Jensen’s inequality. The
second inequality is equivalent to
Ω
(
√
1 +h
2
− 1) dµ ≤
Ω
hdµ since µ(Ω) = 1. This new
inequality follows from the fact that
√
1 +x
2
≤ x + 1 for all nonnegative x. To see this, square
both sides to get 1 +x
2
≤ x
2
+ 2x + 1.
In the case that Ω = [0, 1] and µ is the Lebesgue measure, and h = f
is continuous, then
1
0
1 + (f
)
2
dµ is the formula for the arc length of the graph of f. Then A = f(1) −f(0), and
the second inequality says that the longest path from (0, f(0)) to (1, f(1)) is following along the
line y = f(0) from x = 0 to x = 1, and then going up the line x = 1 until y = f(1). And
√
1 +A
2
is the length of the hypotenuse of the right triangle whose legs are the path just described, so
the ﬁrst inequality says that the straight path is the shortest path.
The intuition from this suggests that the second inequality is equality if and only if h = 0 a.e.,
and the ﬁrst inequality is equality if and only if h = A a.e. The ﬁrst claim is easy to establish,
we go back to the above discussion and note that
√
1 +x
2
< x + 1 if x > 0. If h is constant
a.e., then trivially the ﬁrst inequality holds. Conversely, if
√
1 +A
2
=
Ω
√
1 +h
2
dµ, then an
examination of the proof of Jensen’s inequality, namely equation (3), shows that ϕ(A) = ϕ(h(x))
a.e., so h = A a.e. since ϕ is injective on [0, ∞).
13. Exercise. Under what conditions on f and g does equality hold in the conclusions of Theorems
3.8 and 3.9? You may have to treat the cases p = 1 and p = ∞ separately.
Solution. The inequality in question for Theorem 3.8 is
fg
1
≤ f
p
g
q
for p and q conjugate exponents. If 1 < p < ∞, this is H¨older’s inequality, so assuming that
both quantities are ﬁnite, we know that equality holds if and only if there are constants α and
β, not both 0, such that αf
p
= βg
q
a.e. If p = ∞, then [f(x)g(x)[ = f
∞
[g(x)[ holds if and
only if for all x, either g(x) = 0 or f(x) = f
∞
. The case for p = 1 is analogous.
For Theorem 3.9, we are interested in the inequality
f +g
p
≤ f
p
+g
p
.
For 1 < p < ∞, this follows from Minkowski’s inequality. Examining the proof of Minkowski’s
inequality, this has equality if and only if equality is obtained for f and f + g in H¨older’s
inequality and is obtained for g and f + g. In the case p = ∞ or p = 1, this is the inequality
[f +g[ ≤ [f[ +[g[, and this obtains equality if and only if f and g are nonnegative functions.
15. Exercise. Suppose ¦a
n
¦ is a sequence of positive numbers. Prove that
∞
¸
N=1
1
N
N
¸
n=1
a
n
p
≤
p
p −1
p ∞
¸
n=1
a
p
n
if 1 < p < ∞.
3 L
p
SPACES 16
Solution. Set f =
¸
n≥1
a
n
χ
[n,n+1]
. Then f ∈ L
p
if and only if
¸
n≥1
a
p
n
< ∞. If f / ∈ L
p
, then
the above inequality trivially holds. Otherwise, we can use (Ex. 3.14(a)) to get F
p
≤
p
p−1
f
p
,
where
F(x) =
1
x
¸
x
¸
n=1
a
n
+ (x −x)a
n+1
¸
.
If we assume that a
n
≥ a
n+1
for all n, then this inequality implies the desired inequality by
noting that F(x) ≤ F(x).
In the general case, note that the righthand side stays the same if we rearrange the a
n
to be
nondecreasing since the a
n
are positive and hence the sum is absolutely convergent. Among
all permutations of the sequence ¦a
n
¦, the sum on the lefthand side is biggest when they are
nondecreasing because the earlier terms appear more often. Hence we deduce the general case
from the nondecreasing case.
16. Exercise. Prove Egoroﬀ’s theorem: If µ(X) < ∞, if ¦f
n
¦ is a sequence of complex measurable
functions which converges pointwise at every point of X, and if ε > 0, there is a measurable set
E ⊂ X, with µ(X ` E) < ε, such that ¦f
n
¦ converges uniformly on E.
Show that the theorem does not extend to σﬁnite spaces.
Show that the theorem does extend, with essentially the same proof, to the situation in which
the sequence ¦f
n
¦ is replaced by a family ¦f
t
¦, where t ranges over the positive reals; the
assumptions are now that, for all x ∈ X,
(i) lim
t→∞
f
t
(x) = f(x) and
(ii) t → f
t
(x) is continuous.
Solution. Deﬁne
S(n, k) =
¸
i,j>n
¦x ∈ X [ [f
i
(x) −f
j
(x)[ < 1/k¦.
For each k, note that S(n, k) ⊂ S(n + 1, k), and that
¸
n≥1
S(n, k) = X by our assumptions
that the f
n
converge pointwise. Hence µ(S(n, k)) → µ(X) as n → ∞. So for ε > 0, we can
ﬁnd n
k
for each k such that µ(S(n
k
, k)) > µ(X) − ε/2. Then take E =
¸
k≥1
S(n
k
, k). Note
that µ(E) ≥ µ(X) − ε/2, so µ(X ` E) < ε. Also, by deﬁnition, ¦f
n
¦ will converge uniformly
on E because for every ε > 0, there is a k such that 1/k ≤ ε, and then every x ∈ E satisﬁes
[f
i
(x) −f
j
(x)[ < 1/k for all i, j > n
k
.
If we drop the condition that µ(X) < ∞and replace it with X is σﬁnite, the conclusion does not
necessarily hold. For an example, take X = R
1
with the Lebesgue measure. Then the functions
f
n
(x) = x/n converge pointwise to 0, but cannot converge uniformly on any unbounded set.
For the extension to functions ¦f
t
¦ as t ranges over positive real numbers, we can make the same
deﬁnitions, and now we just used that t → f
t
(x) is continuous for all x to ﬁnd the n
k
used in
the above proof.
21. Exercise. Call a metric space Y a completion of a metric space X if X is dense in Y and Y
is complete. In Sec. 3.15 reference was made to “the” completion of a metric space. State and
prove a uniqueness theorem which justiﬁes this terminology.
Solution. We claim that if (Y, d) and (Y
, d
) are both completions of a metric space X, then
Y and Y
are isomorphic metric spaces. More precisely, there exists bijective maps f : Y → Y
4 ELEMENTARY HILBERT SPACE THEORY 17
and g : Y
→ Y which preserve the metric, i.e., d(x, y) = d
(f(x), f(y)) for all x, y ∈ Y , and such
that f and g are inverses of one another.
So suppose that X can be embedded in both Y and Y
, and identify these two images. We
deﬁne f as follows. If y ∈ X, then f(y) = y. Otherwise, choose a sequence ¦x
n
¦ where x
n
∈ X
that converges to y. Then deﬁne f(y) to be the limit of this sequence in Y
. This makes sense
since X is dense in Y and since Y
is complete. That it is welldeﬁned is a consequence of the
uniqueness of limits. We can deﬁne g : Y
→ Y in an analogous manner. From our deﬁnition it
is obvious that f and g are inverses of one another. Also, note that both f and g are continuous
since they preserve limits.
We just need to check that f and g preserve the metrics. By symmetry, it is enough to do so for
f. Pick x, y ∈ Y . If x, y ∈ X, then d(x, y) = d
(f(x), f(y)) by our identiﬁcation of the image of
X in both Y and Y
. If x ∈ X and y / ∈ X, then let ¦x
n
¦ be a sequence converging to y. We then
have d(x, x
n
) = d
(f(x), f(x
n
)) for all n, so by continuity of the function z → d
(f(x), f(z)),
and taking n → ∞, we see that d(x, y) = d
(f(x), f(y)). The case x / ∈ X and y ∈ X is handled
by the symmetric property of metrics. Finally, suppose that x / ∈ X and y / ∈ X. We can show
that the metric is preserved in this case by taking a double limit instead of a single limit in the
previous argument. Hence we have shown that Y and Y
are isomorphic.
22. Exercise. Suppose X is a metric space in which every Cauchy sequence has a convergent
subsequence. Does it follow that X is complete?
Solution. Let ¦x
n
¦ be a Cauchy sequence in X with a convergent subsequence ¦x
n
i
¦ with limit
x. We claim that x is the limit of ¦x
n
¦. Pick ε > 0. Then there exists N such that n, m ≥ N
implies that d(x
n
, x
m
) < ε/2, and there exists I such that i ≥ I implies that d(x, x
n
i
) < ε. Then
for n suﬃciently large, there exists i such that d(x, x
n
) ≤ d(x, x
n
i
) + d(x
n
i
, x
n
) < ε, so we are
done. Hence X is complete.
4 Elementary Hilbert Space Theory
Notation. In this section, H denotes a Hilbert space, and T denotes the unit circle.
1. Exercise. If M is a closed subspace of H, prove that M = (M
⊥
)
⊥
. Is there a similar true
statement for subspaces M which are not necessarily closed?
Solution. The inclusion M ⊆ (M
⊥
)
⊥
is obvious. Conversely, pick x ∈ (M
⊥
)
⊥
. We can write
x = y+z where y ∈ M and z ∈ M
⊥
(Theorem 4.11(a)). Since 0 = (x, z) = (y, z)+(z, z) = (z, z),
we conclude that z = 0, so x ∈ M. Hence M = (M
⊥
)
⊥
.
Now suppose M is not necessarily closed. Then one can conclude that M
⊥
= ((M
⊥
)
⊥
)
⊥
because
M
⊥
is a closed subspace of H. Also, one can say that M = (M
⊥
)
⊥
. Indeed, M ⊆ (M
⊥
)
⊥
, and
(M
⊥
)
⊥
is a closed set, so M ⊆ (M
⊥
)
⊥
. On the other hand, if x ∈ M
⊥
, then x ∈ M
⊥
because
y → (y, x) is a continuous map.
2. Exercise. Let ¦x
n
[ n = 1, 2, 3, . . . ¦ be a linearly independent set of vectors in H. Show that the
following construction yields an orthonormal set ¦u
n
¦ such that ¦x
1
, . . . , x
N
¦ and ¦u
1
, . . . , u
N
¦
have the same span for all N.
Put u
1
= x
1
/x
1
. Having u
1
, . . . , u
n−1
deﬁne
v
n
= x
n
−
n−1
¸
i=1
(x
n
, u
i
)u
i
, u
n
= v
n
/v
n
.
4 ELEMENTARY HILBERT SPACE THEORY 18
Note that this leads to a proof of the existence of a maximal orthonormal set in separable Hilbert
spaces which makes no appeal to the Hausdorﬀ maximality principle.
Solution. It is obvious from the deﬁnition, that u
n
 = 1 for all n. Now we need to show
orthogonality, and we do showing that (u
n
, u
m
) = 0 for m < n by induction on n. If n = 1,
there is nothing to show, so pick n > 1. Then we will show that (u
n
, u
m
) = 0 for all m < n
by induction on m. Since they u
n
and v
n
diﬀer only by scalars, it is enough to show that
(v
n
, u
m
) = 0. We get
(v
n
, u
m
) =
x
n
−
n−1
¸
i=1
(x
n
, u
i
)u
i
, u
m
= (x
n
, u
m
) −(x
n
, u
m
) = 0
since (u
i
, u
m
) = 0 for i < n.
It is clear that ¦x
1
, . . . , x
N
¦ lies inside of the span of ¦u
1
, . . . , u
N
¦ for each N by deﬁnition of
the u
i
. Since u
1
= x
1
/x
1
, we can inductively build up the set ¦u
1
, . . . , u
N
¦ from ¦x
1
, . . . , x
N
¦
also by the deﬁnition of the u
i
.
If H is separable, then there is a countable dense subset for H, from which we can extract an
at most countable basis for it. Doing the above transformation, we may assume that this basis
is orthonormal. Then by Theorem 4.18, we know that it is a maximal orthonormal basis for
H.
4. Exercise. Show that H is separable if and only if H contains a maximal orthonormal system
which is at most countable.
Solution. If H is separable, then H contains a maximal orthonormal system which is at most
countable by (Ex. 4.2). Conversely, suppose that H contains a maximal orthonormal system
which is at most countable. Then the subspace spanned by this basis is dense by Theorem 4.18,
and must be countable, since it consists of ﬁnite linear combinations of an at most countable
set, so H is separable.
5. Exercise. If M = ¦x [ Lx = 0¦, where L is a continuous linear functional on H, prove that
M
⊥
is a vector space of dimension 1 (unless M = H).
Solution. By Theorem 4.12, there is a unique y ∈ H such that Lx = (x, y) for all x ∈ H. We
claim that M
⊥
= N, where N is the subspace spanned by y. Indeed, the N is closed because
any sequence in this subspace consists of multiples of y, and hence can only converge to some
multiple of y. Then N = (N
⊥
)
⊥
by (Ex. 4.1), and N
⊥
= M. In the case that M = H, y = 0,
so dimN = 1, and we are done.
7. Exercise. Suppose ¦a
n
¦ is a sequence of positive numbers such that
¸
a
n
b
n
< ∞ whenever
b
n
≥ 0 and
¸
b
2
n
< ∞. Prove that
¸
a
2
n
< ∞.
Solution. Suppose that
¸
a
2
n
= ∞. Then one can ﬁnd an inﬁnite number of disjoint sets
E
1
, E
2
, . . . such that S
k
=
¸
n∈E
k
a
2
n
> 1. Now put c
k
=
1
k
√
S
k
, and deﬁne b
n
= c
k
a
n
when
n ∈ E
k
. Then
¸
n≥1
a
n
b
n
=
¸
k≥1
c
k
¸
n∈E
k
a
2
n
=
¸
k≥1
S
k
k
√
S
k
≥
¸
k≥1
1
k
= ∞,
but
¸
n≥1
b
2
n
=
¸
k≥1
¸
n∈E
k
c
2
k
a
2
n
=
¸
k≥1
1
k
2
=
π
2
6
,
which contradicts our hypothesis on ¦a
n
¦, and hence
¸
a
2
n
< ∞.
4 ELEMENTARY HILBERT SPACE THEORY 19
8. Exercise. If H
1
and H
2
are two Hilbert spaces, prove that one of them is isomorphic to a
subspace of the other.
Solution. Let A
1
and A
2
be the cardinalities of maximal orthonormal bases β
1
and β
2
of both
H
1
and H
2
. Then either A
1
≤ A
2
or A
2
≤ A
1
. Without loss of generality, suppose that A
1
≤ A
2
.
Then we can ﬁnd a subset of β
2
with cardinality A
1
, and let H be the closure of the subspace
generated by this subset. Then H is a Hilbert space that has a maximal orthonormal basis of
cardinality A
1
, so we can ﬁnd an isomorphism from H
1
to H.
10. Exercise. Let n
1
< n
2
< n
3
< be positive integers, and let E be the set of all x ∈ [0, 2π] at
which ¦sin n
k
x¦ converges. Prove that m(E) = 0.
Solution. Let f(x) = lim
k→∞
sin n
k
x on E. From the relation 2 sin
2
α = 1 − cos 2α, we see that
2f(x)
2
= 1 − lim
k→∞
cos 2n
k
x. The integral of the righthand side is 0 by (Ex. 4.9), so 2f(x)
2
= 1
a.e. on E, and hence f(x) = ±
1
√
2
a.e. on E. Let E
1
be the set where f(x) =
1
√
2
and let E
2
be
the set where f(x) = −
1
√
2
. Using (Ex. 4.9) again, we have
0 =
E
1
f(x) =
m(E
1
)
√
2
,
so we conclude that m(E
1
) = 0. Similarly, m(E
2
) = 0, and so m(E) = m(E
1
) +m(E
2
) = 0.
11. Exercise. Find a nonempty closed set E in L
2
(T) that contains no element of smallest norm.
Solution. Let u
n
(t) = e
int
, and deﬁne x
n
=
n+1
n
u
n
for all positive integers n. Then the set
X = ¦x
1
, x
2
, . . . ¦ contains no element of smallest norm. For n = m,
x
n
−x
m
 = x
n
 +x
m
 =
n + 1
n
+
m+ 1
m
≥ 2
because the x
n
form an orthogonal set. If x is a limit point of the sequence ¦x
n
i
¦, then for
ε > 0, and for suﬃciently large i, x − x
n
i
 < ε. But x
n
j
− x
n
i
 = x − x
n
i
− (x − x
n
j
) ≤
x − x
n
i
 + x − x
n
j
 < 2ε, which is a contradiction for ε < 1/2 if i = j. Hence ¦x
n
i
¦ is a
constant sequence, which means that X is closed.
16. Exercise. If x
0
∈ H and M is a closed linear subspace of H, prove that
min¦x −x
0
 [ x ∈ M¦ = max¦[(x
0
, y)[ [ y ∈ M
⊥
, y = 1¦.
Solution. By Theorem 4.11, we can write x
0
= Px
0
+ Qx
0
where Px
0
∈ M and Qx
0
∈ M
⊥
,
and furthermore, that Px
0
− x
0
 = min¦x − x
0
 [ x ∈ M¦. We claim that y = Qx
0
/Qx
0

maximizes the quantity [(x
0
, y)[ where y ∈ M
⊥
and y = 1. Pick y
∈ M
⊥
such that y
 = 1.
Then
[(x
0
, y
)[ = [(Qx
0
, y
)[ = Qx
0
[(y, y
)[ ≤ Qx
0
,
where the last inequality is the Schwarz inequality. But [(x
0
, y)[ = Qx
0
, so we have proved
our claim. Finally, Qx
0
 = Px
0
−x
0
, so we have established the desired equality.
17. Exercise. Show that there is a continuous onetoone mapping γ of [0, 1] into H such that
γ(b) − γ(a) is orthogonal to γ(d) − γ(c) whenever 0 ≤ a ≤ b ≤ c ≤ d ≤ 1. (γ may be called a
“curve with orthogonal increments.”)
[ I think this is to be interpreted as there exists some Hilbert space H such that this is true,
because this is impossible for H = R
1
. ]
4 ELEMENTARY HILBERT SPACE THEORY 20
Solution. Take H = L
2
([0, 1]), and deﬁne γ(a) = χ
[0,a]
. This is clearly continuous and injective,
and furthermore, γ(b) − γ(a) = χ
(a,b]
whenever a < b. So in the case that a < b ≤ c < d, we
have (χ
(a,b]
, χ
(c,d]
) =
1
0
χ
(a,b]
χ
(c,d]
dt = 0 since the integrand is 0. If a = b, then γ(b) −γ(a) = 0,
so the orthogonality relation is obvious (similarly if c = d).
18. Exercise. Deﬁne u
s
(t) = e
ist
for all s ∈ R
1
, t ∈ R
1
. Let X be the complex vector space
consisting of all ﬁnite linear combinations of these functions u
s
. If f ∈ X and g ∈ X, show that
(f, g) = lim
A→∞
1
2A
A
−A
f(t)g(t) dt
exists. Show that this inner product makes X into a unitary space whose completion is a
nonseparable Hilbert space H. Show also that ¦u
s
[ s ∈ R
1
¦ is a maximal orthonormal set in
H.
Solution. To show that (f, g) is welldeﬁned, it is enough to do so when f = u
r
and g = u
s
for
r, s ∈ R
1
. In this case, for r = s,
(u
r
, u
s
) = lim
A→∞
1
2A
A
−A
cos((s −r)t) +i sin((s −r)t) dt = lim
A→∞
2i sin((s −r)A)
2A(s −r)
= 0.
For r = s, we get
(u
r
, u
r
) = lim
A→∞
1
2A
A
−A
dt = 1,
so this shows that (f, g) is welldeﬁned, as well as showing that ¦u
s
[ s ∈ R
1
¦ is an orthonormal
set.
Now we verify that this is an inner product structure on X. By deﬁnition, it is clear that
(f, g) = g, f, and by linearity of integrals, (f
1
+ f
2
, g) = (f
1
, g) + (f
2
, g), and (αf, g) = α(f, g)
for α ∈ C. Since f(x)f(x) = [f(x)[
2
, it follows that (f, f) ≥ 0, and that (f, f) = 0 implies that
f = 0 a.e. on R
1
, but since f is a ﬁnite linear combination of exponential functions, this implies
f = 0. Hence X is a unitary space.
Now let H be the completion of X. Since X is dense in H, it follows that ¦u
s
[ s ∈ R
1
¦
is a maximal orthonormal set in H by Theorem 4.18. If H were to have a countable dense
subset, then one could ﬁnd a countable basis for H, which contradicts the fact that we have an
uncountable orthonormal set. Hence H is a nonseparable Hilbert space.
19. Exercise. Fix a positive integer N, put ω = e
2πi/N
, prove the orthogonality relations
1
N
N
¸
n=1
ω
nk
=
1 if k = 0
0 if 1 ≤ k ≤ N −1
and use them to derive the identities
(x, y) =
1
N
N
¸
n=1
x +ω
n
y
2
ω
n
that hold in every inner product space if N > 3. Show also that
(x, y) =
1
2π
π
−π
x +e
iθ
y
2
e
iθ
dθ.
4 ELEMENTARY HILBERT SPACE THEORY 21
Solution. The identity
1
N
¸
N
n=1
ω
0
= 1 is obvious. If k is relatively prime to N, then the sum
¸
N
n=1
ω
nk
is the sum of all primitive Nth roots of unity, which is 0 for N > 1, and otherwise, if
k is not relatively prime, then the sum is k times the sum over all primitive
N
k
th roots of unity,
which is also 0.
To show the second identity, expand the righthand side:
1
N
N
¸
n=1
x +ω
n
y
2
ω
n
=
1
N
N
¸
n=1
(x +ω
n
y, x +ω
n
y)ω
n
=
1
N
N
¸
n=1
(x, x)ω
n
+
N
¸
n=1
(x, y)ω
−n
+
N
¸
n=1
(y, x)ω
2n
+
N
¸
n=1
(y, y)ω
3n
= (x, y)
using the ﬁrst identity.
The last identity is obtained by taking the limit as N → ∞ of the sum on the righthand side
of the second identity. The result is the Riemann integral
1
2π
π
−π
x + e
iθ
y
2
e
iθ
dθ. Of course,
since the lefthand side of the second identity is independent of N, we get the desired result.
Acknowledgements.
I thank Jorun Bomert, Caio Guimar˜ aes Souza, and Brian Streit for pointing out corrections to
some of the solutions.
1 ABSTRACT INTEGRATION
2
Deﬁne f : X → Rn by x → (u1 (x), . . . , un (x)). By Theorem 1.7(b), to prove that h is measurable, it is enough to prove that f is measurable. If R is any open rectangle in Rn which is the Cartesian product of n segments I1 , . . . , In , then f −1 (R) = u−1 (I1 ) ∩ · · · ∩ u−1 (In ), which is measurable n 1 since u1 , . . . , un is measurable. Finally, every open set of Rn is the countable union of such rectangles, so we are done. 3. Exercise. Prove that if f is a real function on a measurable space X such that {x  f (x) ≥ r} is measurable for every rational r, then f is measurable. Solution. Let U ⊆ R1 be an open set. First, U can be written as a union of countably many open balls with rational radii that are centered at rational points. So to prove that f −1 (U ) is measurable, it is enough to prove this when U is an open ball of this form, say with radius r and center c. Since the set of measurable sets is closed under complements and ﬁnite intersections, every set of the form {x  r1 > f (x) ≥ r2 } is measurable for rational r1 , r2 . Now note that {x  c + r > f (x) > c − r} can be written as the countable union −1 (U ) is measurable. n≥1 {x  c + r > f (x) ≥ c − r + 1/n}, so f 4. Exercise. Let {an } and {bn } be sequences in [−∞, ∞], and prove the following assertions: (a) lim sup(−an ) = − lim inf an .
n→∞ n→∞
(b) lim sup(an + bn ) ≤ lim sup an + lim sup bn
n→∞ n→∞ n→∞
provided none of the sums is of the form ∞ − ∞. (c) If an ≤ bn for all n, then lim inf an ≤ lim inf bn .
n→∞ n→∞
Show by an example that strict inequality can hold in (b). Solution. The supremum Ak of the set {−ak , −ak+1 , . . . } is the negative of the inﬁmum Ak of the set {ak , ak+1 , . . . }. Hence inf k {Ak } = − supk {Ak }, which implies (a). The relation sup{ak + bk , ak+1 + bk+1 , . . . } ≤ sup{ak , ak+1 , . . . } + sup{bk , bk+1 , . . . } is clear, so this implies (b). To see that the inequality in (b) can be strict, consider a1 = 1, ai = 0 for i > 1, and b1 = −1, bi = 0 for i > 1. Then lim sup(an +bn ) = 0, but lim sup an +lim sup bn = 1. Now suppose that an ≤ bn for all n. Then inf{ak , ak+1 , . . . } ≤ inf{bk , bk+1 , . . . } for all k, so (c) follows. 5. Exercise. (a) Suppose f : X → [−∞, ∞] and g : X → [−∞, ∞] are measurable. Prove that the sets {x  f (x) < g(x)}, are measurable. (b) Prove the set of points at which a sequence of measurable realvalued functions converges (to a ﬁnite limit) is measurable. {x  f (x) = g(x)}
Describe the corresponding measurable functions and their integrals. M is a σalgebra. then the integral of f is r∈A r. for every x ∈ X. . respectively. Let I be the set of n for which En is at most c countable. Then E = ∞ r=1 n=1 En. so E c ⊆ n∈J En . then the conclusion is a consequence of Lebesgue’s dominated convergence theorem (Theorem 1. the subset where both f and g take ﬁnite values. As for (b). ∞] is measurable for n = 1. the function f − fn is measurable (1. let fn be a sequence of measurable real functions. which is countable since J = ∅. let M be the collection of all sets E ⊂ X such that either E or E c is at most countable. fn (x) → f (x) as n → ∞.r ∞ which is deﬁned to be the preimage of f −fn of (−r. so µ(A) = µ(An ) = 0. X ∈ M. and f agrees with lim fn on E. Solution. which is also measurable. the set where f and g agree is (X \ h−1 ([−∞. 3. . and it is a measurable function. Thus. 0)) ∪ (Y+ \ Z+ ) ∪ (Z− \ Y− ). and f1 ∈ L1 (µ).e. Then these subsets are measurable: for example. Then f is measurable (Theorem 1. The ﬁrst set of (a) is h−1 ([−∞. In case Ac is at most countable. f1 ≥ f2 ≥ f3 ≥ · · · ≥ 0. If we let A ⊂ R1 denote the set of points such that f −1 (r) is not countable. If A is at most countable.22).r . For each n. Y+ is a countable intersection of the sets {x ∈ X  g(x) ≥ n} as n ranges over the positive integers. so is measurable. and put E = n≥1 En . The measurable functions on M consist of those functions f : X → R1 such that for each r ∈ R1 . so exactly one Ai is uncountable. µ(E) = 1 in the second. Now suppose En ∈ M for all n. and deﬁne µ(E) = 0 in the ﬁrst case. Suppose that Ai and Aj are both uncountable for i = j. Suppose fn : X → [0. so the set En. Let X be the complement of these sets. Then Ac ∪ Ac is countable and equal to X since Ai and Aj are disjoint. then so is each An . i j But this contradicts that X is uncountable. i. Let Y+ and Y− be the sets where g(x) = ∞ and −∞. 7. Otherwise. ∞])) ∪ (Y+ ∩ Z+ ) ∪ (Y− ∩ Z− ). If J = ∅. then E is a countable union of countable sets. If we ﬁrst assume that f1 (x) < ∞ for all x. 6. Since X c = ∅ is at most countable. r) is measurable. E c = n∈I En ∩ n∈J En . Also. so at least one Ai is uncountable. f −1 (r) is either at most countable. but En is at most countable. and so E c ∈ M. Hence µ is a measure on M. Solution. so that E = n∈I En ∪ n∈J En . 2. and deﬁne Z+ and Z− analogously for f . and let J be the set of n for which En is uncountable. or f −1 (R1 \ {r}) is at most countable. so is measurable. and c c c hence is countable. Now write a measurable set A as a disjoint union of measurable sets An . .14). . so the same is true for E c since (E c )c = E. So we can deﬁne the function h = f − g on X . Also. if E ∈ M. Prove that M is a σalgebra in X and that µ is a measure on M. Exercise.34) using g(x) = f1 (x) since f1 (x) ≥ . and let E be the set of x such that fn (x) converges as n → ∞. Prove that then n→∞ X lim fn dµ = X f dµ and show that this conclusion does not follow if the condition “f1 ∈ L1 (µ)” is omitted.1 ABSTRACT INTEGRATION 3 Solution. 0) ∪ (0. which means that µ(A) = µ(An ) = 1. then A is uncountable. then either E or E c is at most countable. Deﬁne f = lim sup fn . so E ∈ M. Exercise. Let X be an uncountable set..
which contradicts f1 ∈ L1 (µ). we claim that the integrand is bounded from above by αf . X fn dµ = µ(E) if n is odd. Put fn = χE if n is odd. g(0) = 0. we can ignore E when integrating over X. But this is clearly true: if n ≤ x. so that fn → 0. fn (E) = 0 for n even. Then deﬁne fn (x) = ∞ for x ∈ [0. in which case. so the limit is 0. Exercise. the limit of the integral is also inﬁnite. then rewrite the integrand as α log(1 + (f /n)α )n . Then in general. 1 + (x/n)α nα n + xα So it is enough to show that nxα−1 ≤ nα +xα . n→∞ X 0 if 1 < α < ∞. fn (X \ E) = 1 for n even. We need to show that g(x) ≥ 0 for all x ≥ 0. This is an example where (lim inf fn ) dµ < lim inf X n→∞ n→∞ X fn dµ. Hence by Fatou’s lemma. Solution. If µ(E) > 0. and fn (X \ E) = 0 for n odd. then the integrand approaches ∞ since nα−1 → 0 as n → ∞. we take the derivative of g to get g (x) = α − n 1 αxα−1 nαxα−1 · =α− α . where 0 < c < ∞. and X fn dµ = µ(X \ E) if n is even. Prove that ∞ if 0 < α < 1. provided that µ(E) > 0 and µ(X \ E) > 0. so the integral is c. So the integral on the lefthand side is 0. and µ(E) is the length of E. . If α > 1. If 0 < α < 1. ∞) → R by g(x) = αx − n log(1 + (x/n)α ). then nxα−1 ≤ nα . 1/n]. First. so we may use Lebesgue’s dominated convergence theorem. fn = 1 − χE if n is even. let E = {x ∈ X  f1 (x) = ∞}. and α is a constant. When α ≥ 1. Then X fn dµ = ∞ for all n. µ(X \ E)) > 0. Suppose µ is a positive measure on X. ﬁrst note that lim inf fn (x) = 0 for all x ∈ X because fn (E) = 1 for n odd. and 0 elsewhere. 8. Hence the righthand side is min(µ(E). ∞] is measurable. and hence the integral is also 0. Otherwise. the numerator goes to f α and the denominator goes to ∞. and if n ≥ x. and we are back to the above discussion. Take X = R1 . α lim n log(1 + (f /n) ) dµ = c if α = 1. f : X → [0. What is the relevance of this example to Fatou’s lemma? Solution.1 ABSTRACT INTEGRATION 4 fn (x) ≥ 0 implies that f1 (x) ≥ fn (x). then X f1  dµ = ∞. 9. the integrand converges to log ef = f . then nxα−1 ≤ xα . nα−1 As n → ∞. To see this. we ignore the set where f (x) = ∞ since f is integrable and hence this set has measure 0. X f dµ = c. deﬁne a function g : [0. To see this. For α = 1. but X 0 dµ = 0. First. Exercise. On the other hand. Now suppose that f1 ∈ L1 (µ) no longer holds. So we conclude that µ(E) = 0.
By Theok=n rem 1. Then x ∈ ∞ Ek for all k. and hence prove the theorem without any reference to integration. Show that Proposition 1. and fn → f uniformly on X. and show that the hypothesis “µ(X) < ∞” cannot be omitted. Since fn → f uniformly. Then fn → 0 uniformly. and let C be the maximum of the largest value of f (x)±1 and C. .19(e). Since µ(Bn ) ≤ ∞ µ(Ek ). and we can integrate over F instead of E and get the same result since ∞ · 0 is deﬁned to be 0. Let F be the set where f is nonzero. {fn } is a sequence of bounded complex measurable functions on X. and C ∈ L1 (µ) because X C dµ = C µ(X) < ∞. there exists N such that n ≥ N implies that fn (x)−f (x) < 1 for all x ∈ X. Suppose µ(X) < ∞. Show that A= n=1 k=n ∞ ∞ Ek in Theorem 1. so the equality does not hold. and the bounding sum k=n approaches 0 as n → ∞. Solution. which / k=i means x ∈ B.34. So by Theorem 1. Eir } with i1 < · · · < ir . Solution. Deﬁne fn to be the constant function 1/n. then x is k=n contained in ﬁnitely many Ek . 13. let X = R1 with the usual measure. Then C ≥ fn (x) for all n. . Exercise. being the preimage of an open set. Denote the righthand side by B. we get that µ(B) = 0. . Then µ(B1 ) < ∞ by assumption. Exercise. then f1 + f2 is lower semicontinuous. fN −1 } is ﬁnite and consists of bounded sets. Then F is measurable. . say {Ei1 . Exercise. (b) If f1 and f2 are lower semicontinuous. Recall that A is the set of all x which lie in / inﬁnitely many Ek .41. but the X fn dµ = ∞ for all n. so we are done. n→∞ X lim fn dµ = X f dµ. then ∞f dµ = ∞ E E f dµ. . Prove that n→∞ X lim fn dµ = X f dµ.24(c) is also true for c = ∞.2 POSITIVE BOREL MEASURES 5 10. 11. . So x ∈ ∞ r +1 Ek . To see that µ(X) < ∞ is necessary. If x ∈ A. Pick x ∈ A. while X 0 dµ = 0. Then since {f1 . and consider the following four statements: (a) If f1 and f2 are upper semicontinuous. µ(Bn ) → µ(B) as n → ∞. we can take C to be the largest absolute value any of them obtains. so x ∈ B. / Now set Bn = ∞ Ek . . Then the integral over F is ∞ on both sides of the above equation. and B1 ⊃ B2 ⊃ · · · . We wish to prove that if f ≥ 0. Let {fn } be a sequence of real nonnegative functions on R1 . . . 2 Positive Borel Measures 1. Solution. Exercise. then f1 + f2 is upper semicontinuous. and hence A = B.
nor have we used that f1 and f2 are deﬁned on R1 . Then each fn is upper semicontinuous since the set of x such that fn (x) = 0 is open. Solution. fn is 0 at 0 and greater than 0 elsewhere. 6 Show that three of these are true and that one is false. t ∈ (x − δ. so i fi (x) > α. This proves that the LHS is i=1 m n contained in the RHS. deﬁne fn (x) = 1 on n . if i=1 fi (x) i=1 fi (x) > α. we claim that n {x ∈ R1  i fi (x) > α} = n≥1 {x ∈ R1  i=1 fi (x) > α}. Deﬁne f1 (x) = 0 on (−1. and hence that the set of points of continuity of an arbitrary complex function is a Gδ . If both f1 and f2 are instead lower semicontinuous. δ) = sup{f (s) − f (t)  s. 2. there must be some N such that N fi (x) > α. which i=1 implies that fn is lower semicontinuous. we can construct a counterexample as follows: let f0 (x) = −1 if x ≥ 1 and 0 otherwise. so is not a lower semicontinous function. Here we have we not used that these functions are deﬁned on R1 . Prove that ϕ is upper semicontinuous. − i+1 ] ∪ [ i+1 . 1 1 1 1 For n > 1. For any given α ∈ R. x + δ)}. then an analogous argument shows that f1 + f2 is also lower semicontinuous. Then fi (x) is lower semicontinuous for i i all i. Hence this set is open. Formulate and prove an analogous statement for general topological spaces in place of R1 . δ)  δ > 0}. ﬁrst pick x such that i fi (x) > α. However. then ∞ n=1 fn is upper semicontinuous. Exercise. that f is continuous at a point x if and only if ϕ(x) = 0. To see this. n−1 ∪ − n−1 . Conversely. (c) is a false statement. But if we drop the nonnegativity requirement. so f1 + f2 is upper semicontinuous. then (d) If each fn is lower semicontinuous. t ∈ U }. Let f be an arbitrary complex function on R1 . let fi (x) be the function 1 1 which is −1 on [− 1 . U x . Now by (a). − n and 0 elsewhere. The set {x ∈ R1  f1 (x) + f2 (x) < α} is the union of the sets {x ∈ R1  f1 (x) < β} ∩ {x ∈ R1  f2 (x) < α − β} where we range over all β ≤ α.2 POSITIVE BOREL MEASURES (c) If each fn is upper semicontinuous. Now suppose that we have a sequence {fn } of lower semicontinuous functions. However. and deﬁne ϕ(x. and for i > 0. ∞ n=1 fn is lower semicontinuous. What happens if the word “nonnegative” is omitted? Is the truth of the statements aﬀected if R1 is replaced by a general topological space? Solution. We have not used that the functions are nonnegative here. Since the partial sums n fi (x) converge i=1 to i fi (x). and deﬁne ϕ(x) = inf sup{f (s) − f (t)  s. each partial sum n fi (x) is lower semicontinuous. 1) and f1 (x) = 1 on the rest of R1 . First suppose that f1 and f2 are upper semicontinuous. 1 ] and 0 elsewhere. so is not upper semicontinuous. ϕ(x) = inf{ϕ(x. We formulate the general statement and prove that. then the same is true for whenever m ≥ n by the fact that the fi are nonnegative functions. and the sum f (x) = i fi (x) is 0 at x = 0 and negative elsewhere. so the set above is open. Let X be a topological space. let f : X → C be an arbitrary function.
so ρ(x. Now suppose that f is continuous at x. then ρE (x) − ρE (y) < ε for all x. Then ρ(x. To show that f is continuous at x. examine the relevance of the function f (x) = to Urysohn’s lemma. where V1 and V2 are disjoint open sets. and put δ = ε/2. Now let A and B be disjoint nonempty closed subsets of X. so we have established that ρE is a uniformly continuous function on X. z) ≥ ρE (x). Then there exists an open set U x such that sup{f (s)−f (t)  s. it is enough to show that for every ε > 0. this means that ϕ(x) < ε. So E is open. In particular. Show that ρE is uniformly continuous function on X. We claim that if ρ(x. Conversely. 3. y) < δ. ρA (x) ρA (x) + ρB (x) . f (x) = 0 for x ∈ A. Pick ε > 0. Exercise. so that this is an analogous result to Urysohn’s lemma. 4. there is a neighborhood Uε x such that f (Uε ) ⊂ Bε (f (x)). we conclude that ρ(x. y)  y ∈ E}. z) ≥ ρ(x. Exercise. ϕ(t) < α. For any nonempty E ⊂ X. y ∈ X. and hence ϕ is upper semicontinuous. (b) If E ∈ MF . then E = N ∪ K1 ∪ K2 ∪ · · · . Pick a real number α. Then χB ≤ f ≤ χX\A . y). t ∈ U } < α. deﬁne ρE (x) = inf{ρ(x. We can ﬁnd z ∈ E such that ρ(y. with metric ρ. even if E1 and E2 are not in M. suppose that ϕ(x) = 0. then µ(E1 ∪ E2 ) = µ(E1 ) + µ(E2 ). y) + δ > ρE (x) − ρE (y). y) + δ > ρE (x) − ρE (y). We conclude that the set of points for which f is continuous is a Gδ since it is in the countable intersection of open sets n≥0 En where En = {x ∈ X  ϕ(x) < 1/n}. and f (x) ≤ 1 on X \(A∪B). If A and B are disjoint nonempty closed subsets of X. Then f (x) = 1 for x ∈ B. this means that for every t ∈ U . and consider the set E = {x ∈ X  ϕ(x) < α}. Let X be a metric space. there is an open set Uε x such that f (Uε ) ⊂ Bε (f (x)). Then ϕ is upper semicontinuous and f is continuous at x if and only if ϕ(x) = 0. Pick x ∈ E and ε > 0 such that ϕ(x)+ε < α. Solution. But the lefthand side is less than ε. +ρ(y. where Bε (f (x)) denotes the ball of radius ε around f (x). and consider f as deﬁned above. In particular. but this is clear from the deﬁnition. y) + ρE (y) + δ > ρ(x. By symmetry. so we conclude that ϕ(x) = 0. Then for every ε > 0.2 POSITIVE BOREL MEASURES 7 where U ranges over open sets containing x. Examine the proof of the Riesz theorem and prove the following two statements: (a) If E1 ⊂ V1 and E2 ⊂ V2 . z) < ρE (y) + δ by deﬁnition of ρE . where {Ki } is a disjoint countable collection of compact sets and µ(N ) = 0.
So v ≤ 0. However. Then set N = E \ n≥1 Kn . 1]. construct an open set E ⊂ [0. If v is lower semicontinuous and v ≤ χK . Solution. Since this is the only diﬀerence between sets in M and sets not in M. Inductively. we can ﬁnd a compact set Kn and open set Vn such that Kn ⊂ En−1 ⊂ Vn and µ(Vn \ Kn ) < 1/n. let U be a subset containing E1 ∪ E2 . so we have a contradiction. Now let v be a lower semicontinuous function with v ≤ χK . By Step V of the proof of the Riesz theorem. then K contains an interval (a. and we have established (a). Recall that the deﬁnition is µ(Ei ) = inf{µ(V )  Ei ⊂ V. 2. 1] such that 0 ≤ fn ≤ 1. Let E be Cantor’s familiar “middle thirds” set. 1] which is dense in [0. 9. Exercise. b) for a < b. such that m(E) = ε. so set E1 = E0 \ K1 . there is a compact set K1 and an open set V1 such that K1 ⊂ E0 ⊂ V1 and µ(V1 \ K1 ) < 1. Exercise. Note that in (Ex. even though E and R1 have the same cardinality. Now pick E ∈ MF . show that actually v ≤ 0. Hence µ(E1 ∪ E2 ) ≥ µ(E1 ) + µ(E2 ) by deﬁnition. Hence K is also totally disconnected. K is totally disconnected: if there were a connected component of K consisting of more than a point. we could have replaced 1 with an arbitrary number in (0. so µ(E1 ∪ E2 ) ≤ µ(E1 ) + µ(E2 ). and deﬁne En = En−1 \ Kn . Exercise. since it is the intersection n→∞ of closed sets. 1]. 1]. we see that n m(Kn ) = 1 − i=1 1 · 2i−1 = 1 − 22i n i=1 1 2i+1 =1− 1− 1 1 2n+2 −1 2 −1− 1 2 . such that 1 n→∞ 0 lim fn (x) dx = 0. Note that Step I of the proof of the Riesz theorem does not use the fact that Ei ∩ K ∈ MF for every compact set K. Solution. so must be empty. Finally. Deﬁne K0 = [0. we then get m(E) = 0. If 0 < ε < 1. so m(K) = lim m(Kn ) = 1/2. Then m(En ) = (2/3)n−1 . Hence χK cannot be approximated from below by lower semicontinuous functions. Since Kn has 2n connected components. Then µ(U ) = µ(U ∩ V1 ) + µ(U ∩ V2 ) ≥ µ(E1 ) + µ(E2 ). 7. Furthermore. The set of x where v(x) > 0 lies inside of K and is open. 1] to get the desired example. and inductively deﬁne En to be the the result of removing the open middle third of each connected component of En−1 . K is bounded and closed. V open}. and we have (b). Letting E = n≥0 En . Show that m(E) = 0.6). and set E0 = E. Then E0 \ K1 ∈ MF by Step VI. then take K = n≥0 Kn . so µ(N ) = 0. Construct a totally disconnected compact set K ⊂ R1 such that m(K) > 0. . Construct a sequence of continuous functions fn on [0. so K is compact. where the ﬁrst equality follows since V1 and V2 are disjoint. Then we just need to take the complement in [0. and inductively deﬁne Kn to be Kn−1 with an open interval of length 2−2n removed from the middle of each connected component. Exercise. 6. in the sense of the Vitali–Carath´odory e theorem. Conversely. 1) 2 (start with a smaller or larger set for K1 ). E contains uncountably many points because each decimal in base 3 with either no 1’s or exactly one 1 at the end is an element of E. because K has no interior. Solution. the proof follows just as before. 2−2n < b − a. 5.2 POSITIVE BOREL MEASURES 8 Solution. Then µ(N ) < 1/n for all n. First deﬁne E1 = [0. But for n suﬃciently large.
it is the support of some continuous function. 9 i Solution. i+1 and n i+1 i+2 i−1 i deﬁne gn. Exercise. we need only construct it on each connected component. . Otherwise. and hence closed since X is Hausdorﬀ. and by compactness. . However.n−1 . 11. they are closed.i (x) = nx − (i − 1) on n . . take the indiscrete topology on a set X = {a. f2 . and hence its support is either all of R1 . Solution.r (x) such that if we deﬁne Fn = lim Fn. X}. Then think of gn. Let ϕ 2 be a homeomorphism of the interior of K to R1 .0 . Is it true that every compact subset of R1 is the support of a continuous function? If not. so assume that K is connected. and so its support must contain an open interval f −1 (0. {fn (x)} does not converge for any x ∈ [0. c} with open sets {∅. {a}. 1]. Then V c is closed and hence compact. However.0 .1 . if K is a compact set with nonempty interior. Since the Kα c are compact. Prove that there exist Borel functions g and h such that g(x) = h(x) a. and we can extend the function f ◦ ϕ to K by deﬁning it to be 0 at the end points of K. Conversely. which contradicts the deﬁnition of K. f contains 0 in its image. .i dx = 2/n if 0 < i < n − 1 and 1 otherwise the integral is equal to 3/2n. this description will not carry to general topological spaces. Let {f1 . . Prove that there is a compact set K ⊂ X (the carrier or support of µ) such that µ(K) = 1 but µ(H) < 1 for every proper compact subset H of K.i (x) = −nx + i + 2 on n .i (x) = 1 on n . . . Since µ(K c ) = 0. 13. Each Kα is compact. n−1 by gn. Restricting to each connected component of Supp f . namely because compact sets need not be closed. which implies µ(K) = 1 since µ is regular. . but cannot be the support of any continuous function. 1] because there are inﬁnitely many values of n for which fn (x) = 1 and inﬁnitely many values of n for which fn (x) = 0 for each x ∈ [0. then r→∞ . and 0 elsewhere. . To construct such a function. Let K be the intersection of all compact Kα such that µ(Kα ) = 1. Finally. . n . ik ) ∈ Zk . g2. Let V be an open set which contains K.i as functions on [0. so there exist compactly supported continuous functions Fn. . and g(x) ≤ f (x) ≤ h(x) for every x ∈ Rk . i1 +1]×· · ·×(ik . note that they are continuous. . then µ(H) = 1. gn. but it is not closed. b. n and gn. It is clear that the support of this function is K. Exercise. Then f  is bounded on En .i for i = 0. Exercise. α). Solution. . 14. Let f be a realvalued Lebesgue measurable function on Rk .0 . . . . For an example. we will deﬁne F on the box BI = (i1 . Pick I = (i1 . Then 0 gn. Let En be the set f −1 ((n. g2. .e. So all open sets containing K have measure 1. A point is a compact subset of R1 . . then µ(H) < 1. and hence can write V n 1 1 µ(V ) = 1. 1]. can you describe the class of all compact sets in R1 which are supports of continuous functions? Is your description valid in other topological spaces? Solution. . we c = (K c ∪ · · · ∪ K c ) ∩ V c . [m]. } be the 1 sequence {g1. . assume µ(X) = 1. Hence 0 fn (x) dx → 0 as n → ∞. if H is a compact set properly contained in V . Let µ be a regular Borel measure on a compact Hausdorﬀ space X. so K is closed. deﬁne n functions gn. this shows that µ(V c ) = 0. 1]. .2 POSITIVE BOREL MEASURES but such that the sequence {fn (x)} converges for no x ∈ [0. Then {a. so Kα ∩ V c forms an open cover of V c . Then we can deﬁne f : R1 → R1 by x → e−x . For a given n. ik +1]. Any nonzero continuous function f has some real number α in its image. . If not. gn. n + 1]) ∩ BI for all n ∈ Z.r . and hence compact since X is compact. }. b} is a compact set with nonempty interior. we see that each component needs to contain an open interval. .
Y lies in a proper linear subspace of Rk . It is easy to see that the measure of a proper linear subspace must be 0 because we can take arbtrarily thin open sets that contain the subspace. γ). y1 ). So by Lebesgue’s dominated convergence theorem. β) + ρ(β. that g = h a. . By similar considerations. and that the resulting metric space X is locally compact.e. β = (x2 . Let α = (x1 . . . let x1 . Deﬁne a function fn to be 1 − n ex/2 on [0.. Deﬁne the distance between two points (x1 . Show that this is indeed a metric. Now deﬁne a function ϕ : Rk → R1 by ϕ(x) = n if x ∈ Xn . The set where F and f diﬀer is then a countable union of sets of measure zero. x). / Then ϕ is a measurable function. Exercise. Exercise.e.e. We repeat for every I ∈ Zn and deﬁne a global function in this fashion. ∞ n→∞ 0 ∞ n lim fn dx = 0 e−x/2 dx = 2. . Then F is also Borel measurable. x Solution.20(e)? Solution. Then fn → e−x/2 as n → ∞. y) = 0 if and only if x = y. and the lefthand side is the ﬁrst limit to compute. y) = 0 for at least one y. It is easy to guess the limits of n 1− 0 x n n n ex/2 dx and 0 1+ x n n e−2x dx. 0 16.14. γ) ≤ 1 + y1 − y3  ≤ y1 − y2  + 1 + y2 − y3  ≤ ρ(α. so f = F a. It is also obvious that ρ(x. Let µ be the measure associated with this Λ by Theorem 2. Exercise. e−x/2 ≥ fn (x) for all n. Finally. In this case. and γ = (x3 . y. It is obvious that ρ(x. Prove that your guesses are correct. and that ρ(x. xn be those values of x for which f (x. Now deﬁne g = F + ϕ and h = F + ϕ + χX .24. and furthermore.14). If f ∈ Cc (X). 15. y) ≥ 0 for all x. y1 ) and (x2 . deﬁne F on I to be Fn on the set En . and e−x/2 ∈ L1 (R1 ). on En by the corollary to Theorem 2. as n → ∞. n + 1]) ∩ X. . and that g(x) ≤ f (x) ≤ h(x) for all x ∈ Rk . Why is m(Y ) = 0 in the proof of Theorem 2. Then ρ(α. 1 + y1 − y2  if x1 = x2 . show that µ(E) = ∞ although µ(K) = 0 for every compact K ⊂ E. Let X = {x  f (x) = F (x)}. y2 ) in the plane to be y1 − y2  if x1 = x2 . and ϕ(x) = 0 if x ∈ X. n] and 0 for x > n. 17. y) dy. We just need to verify the triangle inequality. Solution. y) = ρ(y. and deﬁne n ∞ Λf = j=1 −∞ f (xj . We can partition X into measurable sets Xn = (f − F )−1 ((n. If E is the xaxis.2 POSITIVE BOREL MEASURES 10 Fn = f a. y2 ). y3 ). the second integral is ∞ e−x dx = 1. Fn is Borel measurable (Corollary to Theorem 1. Since continuous functions are Borel measurable. Let ρ be the metric deﬁned. Then we see that g and h are Borel measurable.
. . Solution. Λf < ∞ because f is compactly supported. ∞) is upper semicontinuous. Now let α = sup{f (x)  x ∈ X}. i+2 . Then for all x ∈ R1 .n2 −1 . deﬁne n2 functions gn. . y) = 0 for at least one y must be ﬁnite because any collection of vertical lines is open. g2. which is a contradiction. In particular. xr }. Suppose that X is a metric space. Now let E be the xaxis. so the closure of such a neighborhood is the same as the closure when considering it as a subset of R1 .r (x) = n(i+2−n2 x) on i+1 . n−1 for >0 n = 2. ∞) for r = 0. p)  p ∈ X} and prove that . . this shows that µ(Uδ ) = ∞. prove that f attains its maximum at some point of X. In particular. x2 . If X is compact and f : X → (−∞. But each vertical line is closed since it is the complement of the other vertical lines. . y) where x ∈ {x1 . . . consider the sets X1 = [1. }. every compact set K contained in E must be a ﬁnite set of points {(x1 . δ(x))) for each x ∈ R1 where δ(x) > 0. −δ(x)). . so by compactness. xn } and −δ(x)/2 < y < n δ(x)/2. (x. and that f : X → [0. 1] 1 as n → ∞. we can ﬁnd a compactly supported continuous function f that is 1 for points of the form (x. To show that µ(E) = ∞. Since we have countably many sets. ∞] is lower semicontinuous. . be distinct values inside of Xi . . . . For n = 1. . We claim that f (x) = α for some x ∈ X. . and R1 is uncountable. . . ni2 . If not. 3. f2 .r (x) = n on ni2 . gn. and there is no ﬁnite subcover. it is enough to show that the open set Uδ = {(x. . ∞). . . Exercise. and such a compact set in X is a union of closed sets in ﬁnitely many vertical lines. there are ﬁnitely many that cover X. ∞) such that fn (x) → 0 for all x ∈ [0. . Exercise. so µ(E) = ∞. . . 1] → [0. .0 . Note that a set under ρ is open if and only if its intersection with each vertical line is open when considered as a copy of R1 . 2. . but supn fn is not in L1 . . gn. Then n2 n2 n2 n2 1 2 0 gn. In particular. f (p) < ∞ for at least one p ∈ X. . . y)  x ∈ {x1 . 3. i+1 . Solution. Letting the sequence {f1 . 22. . . Find continuous functions fn : [0. x ∈ X. . 20. To see this. .r (x) = n(n2 x−i+1) on i−1 . deﬁne gn (x) = inf{f (p) + nd(x. . we only need to take one such set. Hence f attains its maximum at some point of X. . there is some set Xi such that δ −1 (Xi ) is inﬁnite. with metric d. 1 we get 0 fn (x) dx → 0. Note that if f ∈ Cc (X). g2. 0). For each n. n2 − 1 by gn. and Xn = n . . 1 1 so µ(K) = 0 necessarily because for all n. (xr . y)  −δ(x) < y < δ(x)} has inﬁnite measure for all arbitrary 1 1 functions δ : R1 → R1 . . and supn fn = ∞. so is not L1 . and hence only a ﬁnite union of vertical lines can be compact. α − n )) cover X. The fact that Λ is a linear functional follows from the fact that one could sum over all x ∈ R1 and not change the value of Λ. Any open set V containing E must contain a segment of the form ((x. 21.0 . the set Un = {(x.2 POSITIVE BOREL MEASURES 11 so ρ is a metric. α)) are open for all α ∈ R1 . Every point in the plane has an open neighborhood in its vertical line whose closure is compact when thought of as a set in R1 . Also. .r : R1 → [0. . . Then for each n. .3 . then we can 1 1 ﬁnd some sequence αn such that 0 < α − αn < n . Exercise. . . then the values of x for which f (x. 0)}. The sets f −1 ((−∞. In particular. Then Λf ≥ 2i . and hence f is bounded. the sets f −1 ((−∞. − n < y < n } contains K and has measure µ(Un ) = r/n. 0 fn (x) dx → 0. . and gn. and the fact that the integral is a linear functional. let x1 . . gn. . So X is locally compact. . . However. δ(x) lies in some set.0 .r (x) dx ≤ n . } be {g1. as n → ∞.
This is enough to guarantee that −∞ f − g dx < 2ε by noting that −n ∞ Sn ≤ −∞ f − g dx + n f − g dx. A step function is. if f (x) = ∞. Pick x. then f ≥ gn for all n. One can partition R1 with sets Er for r ∈ Z where i=−n n n r ≤ f (x) < r + 1 for x ∈ Er . p) − f (p) ≥ gn (y) − gn (x). the expression −n ∞ Sn = −∞ f  dx + n f  dx → 0 as n → ∞. a ﬁnite linear combination of characteristic functions of bounded intervals in R1 . Exercise. From this it is clear that the limit of the integral above is 0. too. Hence for a given ε. (iii) gn (x) → f (x) as n → ∞. by deﬁnition. (i) Find the smallest constant c such that log(1 + et ) < c + t (ii) Does 1 n→∞ n lim 1 (0 < t < ∞). so in both cases we have gn → f as n → ∞. if f (x) = ∞. i+1 ] . Assume f ∈ L1 (R1 ). for n suﬃciently large. for all x ∈ X. Since f  dx < ∞. (ii) 0 ≤ g1 ≤ g2 ≤ · · · ≤ f . so f ≥ gn for all n in this case. y). Thus f is the pointwise limit of an increasing sequence of continuous functions. y ∈ X and p ∈ X with f (p) < ∞. Furthermore. y) ≥ nd(y. p) + f (p) − nd(x. It is clear that g1 ≥ 0. Otherwise. Solution. Without loss of generality. so gn (x) → ∞ as n → ∞. then d(x. so this shows (i). p) − nd(x. what is it? . p) = nd(y. For a given r. suppose that gn (y) ≥ gn (x). Otherwise. Let gn = n 2 f (i)χ[ i . So (ii) is established. Also. 12 Solution.2 POSITIVE BOREL MEASURES (i) gn (x) − gn (y) ≤ nd(x. p) ≥ gn−1 (x). p) > 0 for all p with f (p) < ∞. and prove that there is a sequence {gn } of step functions so that ∞ n→∞ −∞ 2 lim f (x) − gn (x) dx = 0. and also so that n ∞ −n f − g dx < ε. gn (x) = f (x) by taking p = x. so gn−1 (x) ≤ gn (x) for all x. We omit the precise details. f (p)+nd(x. p) ≥ f (p)+(n−1)d(x. log(1 + enf (x) ) dx 0 exist for every real f ∈ L1 ? If it exists. on each the subintervals of length we can bound f (x)−gn (x) on their intersection Er by an error directly proportional to r and inversely proportional to n. We have nd(x. 24. gn (x) = f (x) for all n. Again. 25. Exercise. we can choose n large enough so that Sn < ε.
Prove that the supremum of any collection of convex functions on (a. Pick λ ∈ [0. 1] be the set where f (x) ≥ 0. Then the lower limit f is deﬁned by f (x) = x if x ∈ [0. Exercise. y = 3/2. This implies log(1 + enf (x) ) dx ≤ X X (log 2 + nf (x)) dx = log 2 + X nf (x) dx. and hence f is convex. from which we divide by et to get e−t + 1 < ec . we have (1 − λ)f (x) + λf (y) ≥ (1 − λ)fα (x) + λfα (y) ≥ fα ((1 − λ)x + λy). and the integral approaches 0 on [0. Pick λ ∈ [0. and we have (1 − λ)fn (x) + λfn (y) ≥ fn ((1 − λ)x + λy) for all n. Since (1 − λ)fn (x) + λfn (y) → (1 − λ)f (x) + λf (y) and fn ((1 − λ)x + λy) → f ((1 − λ)x + λy) as n → ∞. So by deﬁnition of supremum. and assume that f is ﬁnite.3 Lp SPACES 13 Solution. 1] \ X as n → ∞. 1] and x. However. Now consider the sequence of functions on (0. Then f (x)/2 + f (y)/2 = 1/2. The lefthand side is decreasing with t. and this is not convex: pick x = 1/2. 1] and f (x) = 2 − x if x ∈ [1. and ﬁnally. b). So we conclude that the lower limit of a sequence of convex functions need not be convex. Since exp is an increasing function. Let X ⊂ [0. we apply log (which is an increasing function) to get log(e−t + 1) < c. b) (if it is ﬁnite) and that pointwise limits of sequences of convex functions are convex. log(1 + et ) < c + t becomes 1 + et < ec+t . y ∈ (a. b) is convex on (a. The proof is similar to the proof for pointwise convergent functions. but f (1) = 1. so the smallest c satisfying this inequality is limt→0 log(e−t + 1) = log 2. 3 Lp Spaces 1. so that f is convex. 1] and x. Then for all α. y ∈ (a. we conclude that (1 − λ)fn (x) + λfn (y) ≥ fn ((1 − λ)x + λy). there exist inﬁnitely many values of n (rather than all suﬃciently large n) for which (1 − λ)fn (x) + λfn (y) is within ε of (1 − λ)f (x) + λf (y). 2]. let f = supα fα . λ = 1/2. . and similarly for fn ((1 − λ)x + λy) and f ((1 − λ)x + λy). b). Now let {fn } be a sequence of convex functions that converges pointwise to f . the upper limit of convex functions will be convex. the integral becomes X f (x) dx since the lefthand side increases toward the second term as n → ∞. we conclude that (1 − λ)f (x) + λf (y) ≥ f ((1 − λ)x + λy). except that we use that for any ε > 0. What can you say about upper and lower limits of sequences of convex functions? Solution. so as n → ∞. b). Let {fα } be a collection of convex functions on (a. 2) deﬁned by fn (x) = x if n is even and fn (x) = 2 − x if n is odd.
Without loss of generality. Pick λ ∈ [0. Then (1 − λ)ϕ(x) + λϕ(y) ≥ ϕ((1 − λ)x + λy). for n = 1. we know that fn → f a. Solution. the convexity of log ϕ implies the convexity of exp ◦ log ϕ = ϕ since exp is a nondecreasing and convex function. b). 2. For ϕ > 0. For ϕ > 0. 1] and x. as Solution. but not vice versa. Prove that f dµ · Ω g dµ ≥ 1. Hence f = g a. assume that ϕ(y) ≥ ϕ(x). but log x is not convex. 3. Let E be the set where lim fn = g and let F be the set where lim fn = f . Suppose fn ∈ Lp (µ). b). y ∈ (a. and suppose f and g are positive measurable functions on Ω such that f g ≥ 1. prove that ψ ◦ ϕ is convex on (a.e. y ∈ (a.e.3 Lp SPACES 14 2. What relation exists between f and g? p → 0 and fn → g a. we get f dµ · Ω Ω f g ≥ 1. ∞] is measurable. the converse is not true: the identity function x is convex. y ∈ (a. Suppose µ(Ω) = 1. prove that 1 + A2 ≤ Ω 1 + h2 dµ ≤ 1 + A. Exercise. √ Ω Solution.. Using the Cauchy–Schwarz inequality 2 2 g dµ ≥ Ω f g dµ ≥ Ω 1 dµ = µ(Ω)2 = 1. Exercise. (1 − λ)ψ(ϕ(x)) + λψ(ϕ(x)) ≥ ψ((1 − λ)ϕ(x) + λϕ(y)) ≥ ψ(ϕ((1 − λ)x + λy)). 12. The general case follows by continuity of ϕ: we can arbitrarily approximate ϕ((1 − λ)x + λy) by ϕ((1 − r)x + ry) where r is some rational number whose denominator is a power of 2. b). Assume that ϕ is a continuous real function on (a. 1] and x. If ϕ is convex on (a. Exercise. . . which has measure 0. If A= Ω h dµ. so ψ◦ϕ is convex on (a. Then f = g except possibly on E ∪ F . Pick λ ∈ [0. First note that f g ≥ 1 implies that √ √ on f and g. Suppose µ(Ω) = 1 and h : Ω → [0. Prove that ϕ is convex. and since ψ is convex and nondecreasing. and fn − f n → ∞. . 11. Exercise. 10. Solution. b) such that ϕ x+y 2 1 1 ≤ ϕ(x) + ϕ(y) 2 2 for all x. b) and if ψ is convex and nondecreasing on the range of ϕ. By repeated iterations of the above inequality. However. . if λ is a rational number whose denominator is a power of 2. then we can conclude that ϕ((1 − λ)x + λy) ≤ (1 − λ)ϕ(x) + λy. 3. b). . Since fn − f p → 0.e. Exercise. show that the convexity of log ϕ implies the convexity of ϕ. b). and we can choose r such that (1 − r)ϕ(x) + rϕ(y) ≤ (1 − λ)ϕ(x) + λϕ(y).
this has equality if and only if equality is obtained for f and f + g in H¨lder’s o inequality and is obtained for g and f + g. this is the inequality f + g ≤ f  + g. and prove your conjecture. √ < we go back to the above discussion and note that 1 + x2 √ x + 1 if x > 0. not both 0. this follows from Minkowski’s inequality. namely equation (3). Under what conditions on f and g does equality hold in the conclusions of Theorems 3.e. If 1 < p < ∞. either g(x) = 0 or f (x) = f ∞ . if 1 + A2 = Ω 1 + h2 dµ. Exercise. then f (x)g(x) = f ∞ g(x) holds if and only if for all x. and this obtains equality if and only if f and g are nonnegative functions. In the case p = ∞ or p = 1. If p = ∞. Prove that ∞ N =1 1 N N p an n=1 ≤ p p−1 p ∞ ap n n=1 if 1 < p < ∞. The function ϕ(x) = 1 + x2 is a convex function because its second derivative √ x2 +1 is always positive. 1] and h is continuous. The case for p = 1 is analogous. we are interested in the inequality f +g p ≤ f p + g p. Suppose {an } is a sequence of positive numbers. since ϕ is injective on [0. Hence the ﬁrst inequality follows from Jensen’s inequality. shows that ϕ(A) = ϕ(h(x)) a. For Theorem 3. square both sides to get 1 + x2 ≤ x2 + 2x + 1. The x4 +2x2 +1 √ second inequality is equivalent to Ω ( 1 + h2 − 1) dµ ≤ Ω h dµ since µ(Ω) = 1. and the ﬁrst inequality is equality if and only if h = A a. In the case that Ω = [0. Exercise. √ Solution. The ﬁrst claim is easy to establish. The intuition from this suggests that the second inequality is equality if and only if h = 0 a. and 0 the second inequality says that the longest path from (0. this is H¨lder’s inequality. Examining the proof of Minkowski’s inequality.3 Lp SPACES 15 If µ is Lebesgue measure on [0.e. Then A = f (1) − f (0). so assuming that o both quantities are ﬁnite. h = f .9. And 1 + A2 is the length of the hypotenuse of the right triangle whose legs are the path just described.8 and 3. The inequality in question for Theorem 3. and h = f is continuous. then trivially the ﬁrst inequality holds. then 1 1 + (f )2 dµ is the formula for the arc length of the graph of f . To see this.8 is fg 1 ≤ f p g q for p and q conjugate exponents.e. Solution. ∞). . From this. 13. and then going up the line x = 1 until y = f (1)..9? You may have to treat the cases p = 1 and p = ∞ separately..e. conjecture (for general Ω) under what conditions on h equality can hold in either of the above inequalities. such that αf p = βg q a. 15.. f (1)) is following along the √ line y = f (0) from x = 0 to x = 1. so the ﬁrst inequality says that the straight path is the shortest path.e. f (0)) to (1. This new √ inequality follows from the fact that 1 + x2 ≤ x + 1 for all nonnegative x. 1] and µ is the Lebesgue measure. If h is constant √ a. then an examination of the proof of Jensen’s inequality. Conversely.e. so h = A a. we know that equality holds if and only if there are constants α and β. the above inequalities have a simple geometric interpretation. For 1 < p < ∞.
where x 1 an + (x − x )an+1 .14(a)) to get F p ≤ p−1 f p . Call a metric space Y a completion of a metric space X if X is dense in Y and Y is complete. the sum on the lefthand side is biggest when they are nondecreasing because the earlier terms appear more often. Hence we deduce the general case from the nondecreasing case. and now we just used that t → ft (x) is continuous for all x to ﬁnd the nk used in the above proof. 3. If we drop the condition that µ(X) < ∞ and replace it with X is σﬁnite. by deﬁnition. but cannot converge uniformly on any unbounded set. j > nk . and if ε > 0. then / n p the above inequality trivially holds. there is a measurable set E ⊂ X. k). Hence µ(S(n. such that {fn } converges uniformly on E.n+1] . then this inequality implies the desired inequality by noting that F ( x ) ≤ F (x). we can make the same deﬁnitions. In the general case. State and prove a uniqueness theorem which justiﬁes this terminology. to the situation in which the sequence {fn } is replaced by a family {ft }. the assumptions are now that. k)) > µ(X) − ε/2. if {fn } is a sequence of complex measurable functions which converges pointwise at every point of X. Deﬁne S(n. d) and (Y . Solution. {fn } will converge uniformly on E because for every ε > 0. then Y and Y are isomorphic metric spaces. k) = {x ∈ X  fi (x) − fj (x) < 1/k}. k). we can ﬁnd nk for each k such that µ(S(nk . k)) → µ(X) as n → ∞. 3. Then the functions fn (x) = x/n converge pointwise to 0. Then f ∈ Lp if and only if n≥1 ap < ∞. and then every x ∈ E satisﬁes fi (x) − fj (x) < 1/k for all i. so µ(X \ E) < ε. with µ(X \ E) < ε. Set f = n≥1 an χ[n. take X = R1 with the Lebesgue measure. Prove Egoroﬀ’s theorem: If µ(X) < ∞. 16. there exists bijective maps f : Y → Y . where t ranges over the positive reals. Otherwise. (i) lim ft (x) = f (x) and t→∞ (ii) t → ft (x) is continuous. there is a k such that 1/k ≤ ε.j>n For each k. for all x ∈ X. Show that the theorem does not extend to σﬁnite spaces. So for ε > 0. and that n≥1 S(n. the conclusion does not necessarily hold. We claim that if (Y.15 reference was made to “the” completion of a metric space. i. Then take E = k≥1 S(nk . k) = X by our assumptions that the fn converge pointwise. More precisely. we can use (Ex. Among all permutations of the sequence {an }. Exercise. with essentially the same proof. For the extension to functions {ft } as t ranges over positive real numbers. Note that µ(E) ≥ µ(X) − ε/2. Show that the theorem does extend. k) ⊂ S(n + 1. note that S(n. For an example. In Sec. d ) are both completions of a metric space X.3 Lp SPACES 16 Solution. Solution. Also. Exercise. note that the righthand side stays the same if we rearrange the an to be nondecreasing since the an are positive and hence the sum is absolutely convergent. F (x) = x n=1 If we assume that an ≥ an+1 for all n. If f ∈ Lp . 21.
xn ) = d (f (x). and taking n → ∞. then let {xn } be a sequence converging to y. xm ) < ε/2. xni ) + d(xni . From our deﬁnition it is obvious that f and g are inverses of one another. . . We claim that x is the limit of {xn }. Then deﬁne f (y) to be the limit of this sequence in Y . We can show / / that the metric is preserved in this case by taking a double limit instead of a single limit in the previous argument. This makes sense since X is dense in Y and since Y is complete. 3. . . un = vn / vn . note that both f and g are continuous since they preserve limits. . The case x ∈ X and y ∈ X is handled / by the symmetric property of metrics. 2. and identify these two images.e. f (z)). and such that f and g are inverses of one another. xn ) ≤ d(x. Since 0 = (x. then f (y) = y. there exists i such that d(x. Put u1 = x1 / x1 . We can deﬁne g : Y → Y in an analogous manner. . . we conclude that z = 0. Exercise. . choose a sequence {xn } where xn ∈ X that converges to y. Does it follow that X is complete? Solution. Also. In this section. .. z). prove that M = (M ⊥ )⊥ . and T denotes the unit circle. Suppose X is a metric space in which every Cauchy sequence has a convergent subsequence. Conversely. By symmetry. The inclusion M ⊆ (M ⊥ )⊥ is obvious. y ∈ X. Show that the following construction yields an orthonormal set {un } such that {x1 . So suppose that X can be embedded in both Y and Y . .11(a)). suppose that x ∈ X and y ∈ X. xni ) < ε. f (y)) by our identiﬁcation of the image of X in both Y and Y . z)+(z. We deﬁne f as follows. ui )ui . Hence we have shown that Y and Y are isomorphic. uN } have the same span for all N . un−1 deﬁne n−1 vn = xn − i=1 (xn . y ∈ Y . } be a linearly independent set of vectors in H. H denotes a Hilbert space. . 2. so by continuity of the function z → d (f (x). Otherwise. y) = d (f (x). and there exists I such that i ≥ I implies that d(x. If x ∈ X and y ∈ X. . y ∈ Y . if x ∈ M ⊥ . Pick x. xn ) < ε. Now suppose M is not necessarily closed. M ⊆ (M ⊥ )⊥ . y) = d (f (x). then x ∈ M because y → (y. Having u1 . We can write x = y +z where y ∈ M and z ∈ M ⊥ (Theorem 4. then d(x. We just need to check that f and g preserve the metrics. f (y)) for all x. z) = (z. we see that d(x. i. so x ∈ M . If y ∈ X. Is there a similar true statement for subspaces M which are not necessarily closed? Solution. 1. On the other hand. . If M is a closed subspace of H. so M ⊆ (M ⊥ )⊥ . f (y)). Also. . it is enough to do so for f . 22. Finally. so we are done. f (xn )) for all n. m ≥ N implies that d(xn .4 ELEMENTARY HILBERT SPACE THEORY 17 and g : Y → Y which preserve the metric. Let {xn  n = 1. z) = (y. Then for n suﬃciently large. Hence M = (M ⊥ )⊥ . We then / have d(x. Then there exists N such that n. Let {xn } be a Cauchy sequence in X with a convergent subsequence {xni } with limit x. . y) = d (f (x). 4 Elementary Hilbert Space Theory Notation. If x. and ⊥ (M ⊥ )⊥ is a closed set. xN } and {u1 . pick x ∈ (M ⊥ )⊥ . Pick ε > 0. d(x. Then one can conclude that M ⊥ = ((M ⊥ )⊥ )⊥ because M ⊥ is a closed subspace of H. Exercise. . x) is a continuous map. one can say that M = (M ⊥ )⊥ . That it is welldeﬁned is a consequence of the uniqueness of limits. Indeed. Hence X is complete. Exercise.
uN } from {x1 . . Suppose {an } is a sequence of positive numbers such that bn ≥ 0 and b2 < ∞. that un = 1 for all n. . If M = {x  Lx = 0}. since it consists of ﬁnite linear combinations of an at most countable set. Then we will show that (un . In the case that M = H. so H is separable. uN } for each N by deﬁnition of the ui . Exercise. and hence can only converge to some multiple of y. It is obvious from the deﬁnition. then H contains a maximal orthonormal system which is at most countable by (Ex.4 ELEMENTARY HILBERT SPACE THEORY 18 Note that this leads to a proof of the existence of a maximal orthonormal set in separable Hilbert spaces which makes no appeal to the Hausdorﬀ maximality principle. . . there is nothing to show. . um ) = 0 for i < n. um ) = 0 for m < n by induction on n. and must be countable. . 4. Then one can ﬁnd an inﬁnite number of disjoint sets n 1 E1 . Exercise. . . from which we can extract an at most countable basis for it. If H is separable. n k k Sk n≥1 n∈E k≥1 k≥1 k≥1 k but b2 = n n≥1 k≥1 n∈Ek c2 a2 = k n k≥1 1 π2 = . Prove that a2 < ∞. Exercise. and hence a2 < ∞. Since they un and vn diﬀer only by scalars. y = 0. . so pick n > 1. 7. ui )ui . Suppose that a2 = ∞. we may assume that this basis is orthonormal. . 4. 5. Then S 1 √k ≥ an bn = ck a2 = = ∞.18. so dim N = 1. and N ⊥ = M . n . Then N = (N ⊥ )⊥ by (Ex. y) for all x ∈ H. Indeed. um ) = 0. . such that Sk = n∈Ek a2 > 1. and we are done. .12. Conversely. Then the subspace spanned by this basis is dense by Theorem 4. it is enough to show that (vn . Then by Theorem 4. 4. By Theorem 4. . . xN } also by the deﬁnition of the ui . Solution. . xN } lies inside of the span of {u1 . xn − i=1 (xn . there is a unique y ∈ H such that Lx = (x. Solution. E2 . prove that M ⊥ is a vector space of dimension 1 (unless M = H).2). Now put ck = k√S . . k2 6 which contradicts our hypothesis on {an }. Now we need to show orthogonality.1). um = (xn . then there is a countable dense subset for H. . where L is a continuous linear functional on H. Since u1 = x1 / x1 . If n = 1. um ) = since (ui . and deﬁne bn = ck an when n k n ∈ Ek . um ) − (xn . .18. Solution. We get n−1 (vn . and we do showing that (un . the N is closed because any sequence in this subspace consists of multiples of y. suppose that H contains a maximal orthonormal system which is at most countable. where N is the subspace spanned by y. um ) = 0 for all m < n by induction on m. If H is separable. n n an bn < ∞ whenever Solution. . um ) = 0 It is clear that {x1 . we know that it is a maximal orthonormal basis for H. Doing the above transformation. Show that H is separable if and only if H contains a maximal orthonormal system which is at most countable. We claim that M ⊥ = N . we can inductively build up the set {u1 .
on E. Let n1 < n2 < n3 < · · · be positive integers. Using (Ex. and deﬁne xn = n+1 un for all positive integers n. For n = m. (γ may be called a “curve with orthogonal increments. Solution. If x0 ∈ H and M is a closed linear subspace of H. Prove that m(E) = 0. Find a nonempty closed set E in L2 (T ) that contains no element of smallest norm. we can write x0 = P x0 + Qx0 where P x0 ∈ M and Qx0 ∈ M ⊥ . y)  y ∈ M ⊥ . Pick y ∈ M ⊥ such that y = 1. Solution. which is a contradiction for ε < 1/2 if i = j. y = 1}. we have 1 √ 2 and let E2 be 0= E1 f (x) = m(E1 ) √ . and for suﬃciently large i. Then (x0 . so we have proved our claim. Finally. y) = Qx0 . y ) = (Qx0 .11. y ) ≤ Qx0 . We claim that y = Qx0 / Qx0 maximizes the quantity (x0 . suppose that A1 ≤ A2 . so we can ﬁnd an isomorphism from H1 to H. . From the relation 2 sin2 α = 1 − cos 2α. prove that min{ x − x0  x ∈ M } = max{(x0 . x − xni < ε. m(E2 ) = 0.e. Similarly. ] . then for ε > 0. } contains no element of smallest norm. Exercise.9) again. Then H is a Hilbert space that has a maximal orthonormal basis of cardinality A1 . 10. and furthermore. Solution.e. By Theorem 4. which means that X is closed. and let H be the closure of the subspace generated by this subset. on E. Hence {xni } is a constant sequence. 4. we see that k→∞ 2f (x)2 = 1 − lim cos 2nk x. x2 . Let un (t) = eint . Exercise. where the last inequality is the Schwarz inequality. Solution. and let E be the set of all x ∈ [0. Exercise. If H1 and H2 are two Hilbert spaces. prove that one of them is isomorphic to a subspace of the other. Exercise. Without loss of generality. The integral of the righthand side is 0 by (Ex. and hence f (x) = ± √2 a. Let f (x) = lim sin nk x on E. Qx0 = P x0 − x0 . . 4. so 2f (x)2 = 1 k→∞ 1 a. 16.9). Exercise. But (x0 . y) where y ∈ M ⊥ and y = 1. 11. 2π] at which {sin nk x} converges. 17. But xnj − xni = x − xni − (x − xnj ) ≤ x − xni + x − xnj < 2ε. Then either A1 ≤ A2 or A2 ≤ A1 .”) [ I think this is to be interpreted as there exists some Hilbert space H such that this is true. xn − xm = xn + xm = n+1 m+1 + ≥2 n m because the xn form an orthogonal set. because this is impossible for H = R1 . Then we can ﬁnd a subset of β2 with cardinality A1 . . 2 so we conclude that m(E1 ) = 0. 1] into H such that γ(b) − γ(a) is orthogonal to γ(d) − γ(c) whenever 0 ≤ a ≤ b ≤ c ≤ d ≤ 1. Show that there is a continuous onetoone mapping γ of [0. Then the set n X = {x1 . If x is a limit point of the sequence {xni }. Let A1 and A2 be the cardinalities of maximal orthonormal bases β1 and β2 of both H1 and H2 . and so m(E) = m(E1 ) + m(E2 ) = 0. y ) = Qx0 (y. Let E1 be the set where f (x) = 1 the set where f (x) = − √2 . that P x0 − x0 = min{ x − x0  x ∈ M }. so we have established the desired equality.4 ELEMENTARY HILBERT SPACE THEORY 19 8.
4 ELEMENTARY HILBERT SPACE THEORY 20 Solution. g) is welldeﬁned. Fix a positive integer N .b] . show that (f. so the orthogonality relation is obvious (similarly if c = d). t ∈ R1 . (f1 + f2 . Since f (x)f (x) = f (x)2 . us ) = lim For r = s. then one could ﬁnd a countable basis for H. and that (f. Hence X is a unitary space. Hence H is a nonseparable Hilbert space. γ(b) − γ(a) = χ(a. but since f is a ﬁnite linear combination of exponential functions. it follows that (f. g) + (f2 . f ) ≥ 0. Now we verify that this is an inner product structure on X. Show also that (x. f ) = 0 implies that f = 0 a. y) = N N x + ωny 2ωn n=1 that hold in every inner product space if N > 3. Let X be the complex vector space consisting of all ﬁnite linear combinations of these functions us . g) = α(f. 2A(s − r) dt = 1. and deﬁne γ(a) = χ[0. −A so this shows that (f. In this case. and (αf. we get (ur . it follows that {us  s ∈ R1 } is a maximal orthonormal set in H by Theorem 4.a] . s ∈ R1 . on R1 .d] dt = 0 since the integrand is 0. ur ) = lim 1 A→∞ 2A A 1 A→∞ 2A A cos((s − r)t) + i sin((s − r)t) dt = lim −A A→∞ 2i sin((s − r)A) = 0. This is clearly continuous and injective.b] χ(c. put ω = e2πi/N . g) for α ∈ C. If a = b. for r = s. this implies f = 0. If f ∈ X and g ∈ X. as well as showing that {us  s ∈ R1 } is an orthonormal set. and furthermore. it is clear that (f. f . 19. If H were to have a countable dense subset. and by linearity of integrals. g) = (f1 . g) = lim 1 A→∞ 2A A f (t)g(t) dt −A exists. Exercise. it is enough to do so when f = ur and g = us for r. To show that (f. (ur . χ(c. Since X is dense in H. Show also that {us  s ∈ R1 } is a maximal orthonormal set in H.18. 18. By deﬁnition. So in the case that a < b ≤ c < d. y) = 1 2π π x + eiθ y 2 eiθ dθ. Now let H be the completion of X. g) is welldeﬁned. prove the orthogonality relations 1 N N ω nk = n=1 1 0 if k = 0 if 1 ≤ k ≤ N − 1 and use them to derive the identities 1 (x.d] ) = 0 χ(a.e. g). we 1 have (χ(a.b] whenever a < b. which contradicts the fact that we have an uncountable orthonormal set. Exercise. Deﬁne us (t) = eist for all s ∈ R1 . Show that this inner product makes X into a unitary space whose completion is a nonseparable Hilbert space H. then γ(b) − γ(a) = 0. Solution. g) = g. Take H = L2 ([0. −π . 1]).
which is 0 for N > 1. I thank Jorun Bomert. then the sum n=1 N ω nk is the sum of all primitive N th roots of unity. To show the second identity. then the sum is k times the sum over all primitive N th roots of unity. expand the righthand side: 1 N N n=1 1 x+ω y ω = N n 2 n N (x + ω n y. k which is also 0. if n=1 k is not relatively prime. we get the desired result. x + ω n y)ω n n=1 N N N N = 1 N (x. y)ω 3n n=1 = (x. x)ω n + n=1 (x. The last identity is obtained by taking the limit as N → ∞ of the sum on the righthand side π 1 x + eiθ y 2 eiθ dθ. The identity N N ω 0 = 1 is obvious. The result is the Riemann integral 2π −π since the lefthand side of the second identity is independent of N . Acknowledgements. of the second identity.4 ELEMENTARY HILBERT SPACE THEORY 21 1 Solution. If k is relatively prime to N . x)ω 2n + n=1 (y. and otherwise. and Brian Streit for pointing out corrections to a some of the solutions. Caio Guimar˜es Souza. . y)ω −n + n=1 (y. y) using the ﬁrst identity. Of course.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.