Chapter 1.
6, Page 37
Problem 2:
(a) Prove that x is in the Cantor set i x has a ternary expansion that
uses only 0s and 2s.
(b) The CantorLebesgue function is defined on the Cantor set by writing xs ternary expansion in 0s and 2s, switching 2s to 1s, and
reinterpreting as a binary expansion. Show that this is welldefined
and continuous, F (0) = 0, and F (1) = 1.
(c) Prove that F is surjective onto [0, 1].
(d) Show how to extend F to a continuous function on [0, 1].
Solution.
(a) The nth iteration of the Cantor set removes the open segment(s) consisting of all numbers with a 1 in the nth place of the ternary expansion.
Thus, the numbers remaining after n iterations will have only 0s and
2s in the first n places. So the numbers remaining at the end are precisely those with only 0s and 2s in all places. (Note: Some numbers
have a nonunique ternary representation, namely those that have a
representation that terminates. For these, we choose the infinitely repeating representation instead; if it consists of all 0s and 2s, it is in
the Cantor set. This works because we remove an open interval each
time, and numbers with terminating representations are the endpoints
of one of the intervals removed.)
(b) First, we show that this is welldefined. The only possible problem is
that some numbers have more than one ternary representation. However, such numbers can have only one representation that consists of
all 0s and 2s. This is because the only problems arise when one representation terminates and another doesnt. Now if a representation
terminates, it must end in a 2 if it contains all 0s and 2s. But then
the other representation ends with 12222... and therefore contains a
1.
Next we show F is continuous on the Cantor set; given > 0, choose
k such that 21k < . Then if we let = 31k , any numbers within will
agree in their first k places, which means that the first k places of their
images will also agree, so that their images are within 21k < of each
other.
The equalities F (0) = 0 and F (1) = 1 are obvious; for the latter,
1 = 0.2222 . . . so F (1) = 0.1111 = 1.
(c) Let x 2 [0, 1]. Choose any binary expansion of x, replace the 0s with
2s, and reinterpret as a ternary expansion. By part (a), this will
produce a member of the Cantor set whose image is x. (Note: Their
may be more than one preimage of x, e.g. F ( 13 ) = F ( 23 ) = 12 .)
(d) First, note that F is increasing on the Cantor set C. Now let
G(x) = sup{F (y) : y x, y 2 C}.
Note that G(x) = F (x) for x 2 C because F is increasing. G is
there
continuous at points not in C, because C is open, so if z 2 C,
is a neighborhood of z on which G is constant. To show that G is
continuous on C, let x 2 C and use the continuity of F (part b) to
1
choose > 0 such that G(x) G(z) < for z 2 C, x z < . Choose
z1 2 (x
, x), z2 2 (x, x + ) and let 0 < min(x z1 , z2 x). Then
for y x < 0 , if y 2 C we automatically have F (y) F (x) < . If
y2
/ C but y < x, G(x) > G(y) G(z1 ) > G(x) ; similarly, if y 2
/C
but y > x, G(x) < G(y) G(z2 ) < G(x) + .
Problem 3: Suppose that instead of removing the middle third of the segment at each step, we remove the middle , where 0 < < 1.
(a) Prove that the complement of C is the union of open intervals with
total length 1.
(b) Prove directly that m (C ) = 0.
Solution.
n
(a) At the nth
step
n (starting at n = 0), we remove 2 segments, each of
1
. The total length of these segments is
length 2
n
1
1
X
X
1
1
1
n
2
=
=
= 1.
n
2
(1
)
1
(1
)
n=0
n=0
n
(b) If Cn is the set remaining
nafter n iterations, then Cn is a union of 2
segments of length 1 2 . So
m(Cn ) = (1
)n .
Note that m(Cn ) ! 0. Since each Cn is a covering of C by almost
disjoint cubes, the infimum of the measures of such coverings is 0.
Problem 4: Construct a closed set C so that at the kth stage of the construction one removes 2k 1 centrally situated open intervals each of length
`k , with
`1 + 2`2 + + 2k 1 `k < 1.
P1
(a) If `j are chosen small enough, then k=1 2k 1 `k < 1. In this case,
> 0, and in fact,
show that m(C)
=1
m(C)
1
X
2k
`k .
k=1
then there exists a sequence xn such that xn 2
(b) Show that if x 2 C,
/ C,
yet xn ! x and xn 2 In , where In is a subinterval in the complement
of C with In  ! 0.
(c) Prove as a consequence that C is perfect, and contains no open interval.
(d) Show also that C is uncountable.
Solution.
(a) Let Ck denote the set remaining after k iterations of this process, with
C0 being the unit segment. Then
m([0, 1] \ Ck ) =
k
X
j=1
2j `j
since [0, 1] \ Ck is a union of disjoint segments with this total length.
Then
k
X
m(Ck ) = 1
2j `j .
j=1
so by Corollary 3.3,
Now Ck & C,
= lim m(Ck ) = 1
m(C)
n!1
1
X
2k `k .
k=1
(b) For k = 1, 2, . . . , let Jk be the interval of Ck which contains x. Let In
be the interval in C c which is concentric with Jn 1 . (Thus, at the nth
step of the iteration, the interval In is used to bisect the interval Jn 1 .)
Let xn be the center of In . Then xn 2 C c . Moreover, xn x Jn 1 
since Jn 1 contains both xn and x. Since the maximum length of the
intervals in Cn tends to 0, this implies xn ! x. Finally, xn 2 In C c ,
and In Jn 1 ) In  ! 0.
(c) Clearly C is closed since it is the intersection of the closed sets Cn . To
prove it contains no isolated points, we use the same construction from
This time, let xn be an endpoint of In ,
the previous part. Let x 2 C.
rather than the center. (We can actually take either endpoint, but for
specificity, well take the one nearer to x.) Because In is constructed
as an open interval, its endpoints lie in Cn . Moreover, successive
iterations will not delete these endpoints because the kth iteration
We also
only deletes points from the interior of Ck 1 . So xn 2 C.
have xn x Jn as before, so that xn ! x. Hence x is not an
isolated point. This proves that C is perfect.
(d) We will construct an injection from the set of infinite 01 sequences
To do this, we number the subintervals of Ck in order from
into C.
left to right. For example, C2 contains four intervals, which we denote
I00 , I01 , I10 , and I11 . Now, given a sequence a = a1 , a2 , . . . of 0s and
1s, let Ina denote the interval in Cn whose subscript matches the first
n terms of a. (For instance, if a = 0, 1, 0, 0, . . . then I4a = I0100 C4 .)
Finally, let
1
\
x2
Ina .
n=1
a
This intersection is nonempty because In+1
Ina , and the intersection
of nested closed intervals is nonempty. On the other hand, it contains
only one point, since the length of the intervals tends to 0. Thus, we
have constructed a unique point in C corresponding to the sequence a.
Since there is an injection from the uncountable set of 01 sequences
C is also uncountable.
into C,
Problem 5: Suppose E is a given set, and and On is the open set
1
On = {x : d(x, E) < }.
n
Show that
(a) If E is compact, then m(E) = lim m(On ).
n!1
(b) However, the conclusion in (a) may be false for E closed and unbounded, or for E open and bounded.
Proof. (a) First note that for any set E,
=
E
1
\
On
1
\
On .
n=1
since the closure of E consists of precisely those points whose distance
to E is 0. Now if E is compact, it equals its closure, so
E=
n=1
Note also that On+1 On , so that On & E. Now since E is bounded,
it is a subset of the sphere BN (0) for some N . Then O1 BN +1 (0),
so that m(O1 ) < 1. Thus, by part (ii) of Corollary 3.3,
m(E) = lim m(On ).
n!1
(b) Suppose E = Z R. Then m(On ) = 1 for all n, since On is a
collection of infinitely many intervals of length n2 . However, m(Z) = 0.
This shows that a closed unbounded set may not work. To construct a
bounded open counterexample, we need an open set whose boundary
has positive measure. To accomplish this, we use one of the Cantor > 0.
like sets C from Problem 4, with the `j chosen such that m(C)
Let E = [0, 1]\ C. Then E is clearly open and bounded. The boundary
since C contains no interval and hence has empty
of E is precisely C,
conversely,
interior. (This shows that the boundary of E contains C;
it cannot contain any points of E because E is open, so it is exactly
Hence E
= E [ @E = [0, 1]. Now
equal to C.)
1
1
1
< 1 =
On = x 2 R : d(x, E) <
= x 2 R : d(x, E)
,1 +
.
n
n
n
n
Clearly m(On ) ! 1, but m(E) = 1
< 1.
m(C)
Problem 6: Using translations and dilations, prove the following: Let B be
a ball in Rd of radius r. Then m(B) = vd rd , where vd = m(B1 ) and B1 is
the unit ball {x 2 Rd : x < 1}.
Solution. Let > 0. Choose a covering {Qj } of B1 with total volume
less than m(B1 ) + rd ; such a covering must exist because the m(B1 ) is
the infimum of the volumes of such cubical coverings. When we apply the
homothety x 7! rx to Rd , each Qj is mapped to a cube Q0j whose side
length is r times the side length of Qj . Now {Q0j } is a cubical covering of
Br with total volume less than rd m(B1 )+. This is true for any > 0, so we
must have m(Br ) rd m(B1 ). Conversely, if {Rj } is a cubical covering of
Br whose total volume is less than m(Br ) + , we can apply the homothety
x 7! 1r x to get a cubical covering {Rj0 } of B1 with total volume less than
5
1
(m(Br )
rd
+ ). This shows that m(B1 )
inequalities show that m(Br ) = rd m(B1 ).
1
(m(Br )
rd
+ ). Together, these
Problem 7: If = ( 1 , . . . , d ) is a dtuple of positive numbers with
and E Rd , we define E by
E = {( 1 x1 , . . . ,
d xd )
> 0,
: (x1 , . . . , xd ) 2 E}.
Prove that E is measurable whenever E is measurable, and
m( E) =
...
d m(E).
Solution. First we note that for an open set U , U is also open. We could
see this from the fact that x ! x is an invertible linear transformation, and
therefore a homeomorphism. More directly, if p 2 U , let Br (p) be a neighborhood of p which is contained in U ; then if we define = min( 1 , . . . , d ),
we will have B r ( p) U .
Next, we note that for any set S, m ( S) = 1 . . . d m (S). The proof of
this is almost exactly the same as Problem 6: the dilation x 7! x and
its inverse map rectangular coverings of S to rectangular coverings of S
and vice versa; but since the exterior measure of a rectangle is just its area
(Page 12, Example 4), the infimum of the volume of rectangular coverings
is the same as the infimum over cubical coverings. Hence a rectangular
covering within of the infimum for one set is mapped to a rectangular
covering within 1 ... n for the other.
As a more detailed version
Pof the preceding argument, suppose {Qj } is a
cubical covering of SP
with Qj  < m (S)+. Then { Qj } is a rectangular
covering of S with
 Qj  < 1 . . . d m (S) + 1 . . . P
d . Now for each rectangle Qj we can find a cubical covering {Q0jk } with k Q0jk  <  Qj + 2j .
P
Then \j,k Q0jk is a cubical covering of S with j,k Q0jk  < 1 . . . d m (S)+
(1 + 1 . . . d ). This implies that m ( S) 1 . . . d m (S). To get the reverse inequality we note that another type transformation goes the other
direction, i.e. S = 0 ( S) where 0 = (1/ 1 , . . . , 1/ d ).
Now let U E be an open set with m (U \ E) < 1 ...
. Then U
E is
d
an open set. Moreover, (E\U ) = E\ U , so m ( U \ E) = 1 . . . d m (U \
E) < . Hence E is also measurable. (Alternatively, we could prove that
E is measurable by appealing to Problem 8.)
Problem 8: Suppose L is a linear transformation of Rd . Show that if E is a
measurable subset of Rd , then so is L(E), by proceeding as follows:
(a) Note that if E is compact, so is L(E). Hence if E is an F set, so is
L(E).
(b) Because L automatically satisfies the inequality
L(x)
L(x0 ) M x
x0 
for some M , we can see that L maps any
p cube of side length ` into a
cube of side length cd M `, with cd = 2 d. Now if m(E)
P = 0, there is
a collection of cubes {Qj } such that E \j Qj , and j m(Qj ) < .
Thus m (L(E)) c0 , and hence m(L(E)) = 0. Finally, use Corollary 3.5. (Problem 4 of the next chapter shows that m(L(E)) =
 det Lm(E).)
Solution.
(a) Since linear transformations on finitedimensional spaces are always
continuous, they map compact sets to compact sets. Hence, if E is
compact, so is L(E). Moreover, because Rd is compact, any closed
set is the countable union of compact sets. So if
E=
1
\
Fn
1
\
Knj
n=1
where Fn is closed, then for each n we have
Fn =
j=1
where Knj is compact; then
\
E=
Knj
j,n
is a countable union of compact sets. Then
\
L(E) =
L(Knj )
j,n
is too, since L(Knj ) is compact. But compact sets are closed, so this
shows that L(E) is F .
(b) Let x be a corner of a cube Q of side p
length `. Then every point
0
x in the cube is a distance of at most d` away from x, since this
is
diagonally
opposite corner. Now x x0  <
p the distance to the
p
0
0
pd` ) L(x) L(x ) < dM `. Now if Q is the cube of side length
2 dMp` centered at x, the points on the exterior of the cube are all at
least dM ` away from x. L(Q) Q0 . Since a set of measure 0 has
a cubical covering with volume less thanp, its image under L has a
cubical covering with volume less than 2 dM . This implies that L
maps sets of measure 0 to sets of measure 0.
Finally, let E be any measurable set. By Corollary 3.5, E = C \ N
where C is an F set and N has measure 0. We have just shown
that L(C) is also F and L(N ) also has measure 0. Hence L(E) =
L(C) \ L(N ) is measurable.
Problem 9: Give an example of an open set O with the following property:
the boundary of the closure of O has positive Lebesgue measure.
Solution. We will use one of the Cantorlike sets from Problem 4; let C
be such a set with m(hatC) > 0. We will construct an open set whose
Let us number the intervals involved in the Cantor
closure has boundary C.
iteration as follows: If Cn is the set remaining after n iterations (with
C0 = [0, 1]), we number the 2n intervals in Cn in binary order, but with 2s
instead of 1s. For example, C2 = I00 \ I02 \ I20 \ I22 . The intervals in the
denoted by subscripted Js, are named according to the
complement of C,
intervals they bisected, by changing the last digit to a 1. For instance, in
C1 , the interval J1 is taken away to create the two intervals I0 and I2 . In
the next iteration, I0 is bisected by J01 to create I00 and I02 , while I2 is
bisected by J21 to create I20 and I21 , etc.
Having named the intervals, let G = J1 \ J001 \ J021 \ J201 \ J221 \ . . .
be the union of the intervals in C c which are removed during odd steps of
be the union of the other intervals,
the iteration, and G0 = [0, 1] \ (G \ C)
i.e. the ones removed during even steps of the iteration. I claim that the
Clearly this is a closed set (its complement in [0, 1]
closure of G is G \ C.
is the open set G0 ) containing G, so we need only show that every point in
C is a limit of points in G. To do this, we first note that with the intervals
numbered as above, an interval Iabc... whose subscript is k digits long has
length less than 21k . This is so because each iteration bisects all the existing
Is. In addition, an interval Jabc... with a kdigit subscript has length less
than 2k1 1 because it is a subinterval of an Iinterval with a (k 1)digit
Then x 2 \n Cn so for each n we can find an
subscript. Now let x 2 C.
(n)
interval I
containing x which has an ndigit subscript. Let J (n) be the
Jinterval with an ndigit subscript, whose first n 1 digits match those of
I (n) . Then I (n) and J (n) are consecutive intervals in Cn . Since they both
have length at most 2n1 1 , the distance between a point in one and a point
in the other is at most 2n1 2 . Thus, if we let yn be a sequence such that
yn 2 J (n) , then yn ! x. Now let yn0 be the subsequence taken for odd
n, so that yn0 G. Then we have constructed a sequence of points in G
which converge to x 2 C.
= G\ C.
It only remains to show that @(G\ C)
=
We have shown that G
C. Clearly @(G \ C) C since G is open and is therefore contained in the
Now let x 2 C.
By the same construction as above, we
interior of G \ C.
(n)
can choose a sequence yn 2 J
which converges to x. If we now take the
subsequence yn over even n, then yn 2 G0 and yn ! x. This proves that
Hence we have shown that G is an open set whose closure
x 2 @(G \ C).
which has positive measure.
has boundary C,
Problem 11: Let A be the subset of [0, 1] which consists of all numbers which
do not have the digit 4 appearing in their decimal expansion. Find m(A).
Proof. A has measure 0, for the same reason as the Cantor set. We can
construct A as an intersection of Cantorlike iterates. The first iterate
is the unit interval; the second has a subinterval of length 1/10 deleted,
with segments of lengths 3/10 and 6/10 remaining. (The deleted interval
corresponds to all numbers with a 4 in the first decimal place.) The next
has 9 subintervals of length 1/100 deleted, corresponding to numbers with
a non4 in the first decimal place and a 4 in the second. Continuing, we get
closed sets Cn of length (9/10)n , with A = \Cn . Clearly A is measurable
since each Cn is; since m(Cn ) ! 0, m(A) = 0.
Problem 13:
(a) Show that a closed set is G and an open set F .
(b) Give an example of an F which is not G .
(c) Give an example of a Borel set which is neither G nor F .
Proof.
(a) Let U be open. As is well known, U is the union of the open rational
balls that it contains. However, it is also the union of the closed
rational balls that it contains. To prove this, let x 2 U and r > 0 such
that Br (x) U . Choose a rational lattice point q with x q < 3r ,
and a rational d with 3r < d < 2r . Then Bd(q) Br (x) U and
x 2 Bd(q), so any x 2 U is contained in a closed rational ball within
U . Thus, U is a union of closed rational balls, of which there are only
countably many. For a closed set F , write the complement Rd \ F as
a union of rational balls Bn ; then F = \Bnc is a countable intersection
of the open sets Bnc , so F is G .
(b) The rational numbers are F since they are countable and single points
are closed. However, the Baire category theorem implies that they
are not G . (Suppose they are, and let Un be open dense sets with
Q = \Un . Define Vn = Un \ {rn }, where rn is the nth rational in some
enumeration. Note that the Vn are also open and dense, but their
intersection is the empty set, a contradiction.)
(c) Let A = (Q \ (0, 1)) [ ((R \ Q) \ [2, 3]) consist of the rationals in (0, 1)
together with the irrationals in [2, 3]. Suppose A is F , say A = [Fn
where Fn is closed. Then
(R \ Q) \ [2, 3] = A \ [2, 3] = ([Fn ) \ [2, 3] = [(Fn \ [2, 3])
is also F since the intersection of the two closed sets Fn and [2, 3] is
closed. But then
Q \ (2, 3) = \(Fnc \ (2, 3))
is G because Fnc \ (2, 3) is the intersection of two open sets, and
therefore open, for each n. But then if rn is an enumeration of the
rationals in (2, 3), (Fnc \(2, 3))\{rn } is also open, and is dense in (2, 3).
Hence \(Fnc \ (2, 3)) \ {rj } is dense in (2, 3) by the Baire Category
Theorem. But this set is empty, a contradiction. Hence A cannot be
F .
Similarly, suppose A is G , say A = \Gn where Gn is open. Then
Q \ (0, 1) = A \ (0, 1) = (\Gn ) \ (0, 1) = \(Gn \ (0, 1))
is also G since Gn \ (0, 1) is the intersection of two open sets and
therefore open. But then if {qn } is an erumeration of the rationals in
(0, 1), (Gn \ (0, 1)) \ {qn } is open and is dense in (0, 1), so
\ ((Gn \ (0, 1)) \ {qn })
must be dense in (0, 1). But this set is empty, a contradiction. Hence
A is not G .
Problem 16: BorelCantelli Lemma: Suppose {Ek } is a countable family
of measurable subsets of Rd and that
1
X
m(Ek ) < 1.
k=1
Let
E = {x 2 Rd : x 2 Ek for infinitely many k} = lim sup Ek .
Show that E is measurable.
Prove m(E) = 0.
Solution.
Let
Bn =
1
[
Ek
k=n
be the set of x which are in some Ek with k
many Ek i x 2 Bn for all n, so
1
1 [
1
\
\
E=
Bn =
Ek .
n=1
n. Then x is in infinitely
n=1 k=n
This is a countable intersection of a countable union of measurable
sets, and hence isPmeasurable.
Let > 0. Since
m(Ek ) converges, 9N such that
1
X
m(Ek ) < .
k=N
Then
m(BN ) = m
1
[
k=N
Ek
1
X
m(Ek ) <
k=N
by subadditivity. But m(\Bn ) m(BN ) by monotonicity, so m(E) <
for all . Hence m(E) = 0.
Problem 17: Let {fn } be a sequence of measurable functions on [0, 1] with
fn (x) < 1 for a.e. x. Show that there exists a sequence cn of positive
real numbers such that
fn (x)
! 0 a.e. x.
cn
Solution. We are given that for each n,
!
1
\
k
m
x : fn (x) >
=0
n
k=1
since this set is precisely the set where f (x) = 1. Since these sets are
nested, this implies
k
lim m
x : fn (x) >
= 0.
k!1
n
Hence, 9cn such that
n
cn o
1
m x : fn (x) >
< n.
n
2
Define
n
cn o
En = x : fn (x) >
.
n
10
Then m(En ) <
1
2n ,
so
m@
1 [
1
\
m=1 j=m
Ej A = 0
by the BorelCantelli lemma. But the complement of this set consists of
precisely those points that are in finitely many En , i.e. those points for
which fnc(x)
is eventually less than 21n . Hence we have found a set of measure
n
0 such that
fn (x)
cn
! 0 on the complement.
Problem 18: Prove the following assertion: Every measurable function is the
limit a.e. of a sequence of continuous functions.
Proof. Let f : R ! R be measurable. (The problem didnt specify whether
f can have 1 as a value, but Im assuming not.) Let Bn = [ n, n]. Then
by Lusins Theorem, there exists a closed (hence compact) subset En Bn
with m(Bn \En ) < 21n and f continuous on En . Then by Tietzes Extension
Theorem, we can extend f to a continuous function fn on all of R, where
fn = f on En . (Explicitly, such an extension could work as follows: Define
fn : R ! R by fn (x) = f (x) for x 2 En ; for x 2
/ En , since the complement
is open, x is in some open interval (a, b) Enc or in some unbounded open
interval ( 1, a) Enc or (b, 1) Enc . Let fn (x) = f (a) + xb aa f (b) in the
first case and fn (x) = f (a) in the other two cases.)
I claim that fn ! f almost everywhere. Suppose x is a point at which
fn 6! f . Then x 2 (Bnc ) [ (Bn \ En ) for infinitely many n since otherwise
fn (x) is eventually equal to f (x). Now a given x can be in only finitely many
Bnc , so it must be in infinitely many (Bn \ En ), i.e. x 2 lim sup(Bn \ En ).
But lim sup(Bn \ En ) has measure 0 by the BorelCantelli Lemma. Hence
the set of x at which fn (x) 6! f (x) is a subset of a set of measure 0, and
therefore has measure 0.
Problem 20: Show that there exist closed sets A and B with m(A) = m(B) =
0, but m(A + B) > 0:
(a) In R, let A = C, B = C/2. Note that A + B [0, 1].
(b) In R2 , observe that if A = I {0} and B = {0} I (where I = [0, 1]),
then A + B = I I.
Solution.
(a) As noted, let C be the Cantor set, A = C, and B = C/2. Then A
consists of all numbers which have a ternary expansion using only 0s
and 2s, as shown on a previous homework set. This implies that B
consists of all numbers which have a ternary expansion using only 0s
and 1s. Now any number x 2 0, 1] can be written as a + b where
a 2 A and b 2 B as follows: Pick any ternary expansion 0.x1 x2 . . . for
x. Define
(
2
(xn = 2)
an =
0
(else)
(
1
(xn = 1)
bn =
0
(else)
11
Then a = 0.a1 a2 2 A and b = 0.b1 b2 2 B, and a + b = x.
(b) Duly noted. (I feel like I should prove something, but Im not sure
what there is to prove here.)
Problem 21: Prove that there is a continuous function that maps a Lebesgue
measurable set to a nonmeasurable set.
Proof. As shown in the last homework, there is a continuous function f :
[0, 1] ! [0, 1] such that f (C) = [0, 1], where C is the Cantor set. Let
V [0, 1] be a Vitali set, which is nonmeasurable. Let E = f 1 (V ) \ C.
Then E is measurable since E is a subset of a set of measure 0. However,
f (E) = V which is not measurable.
Problem 22: Let [0,1] be the characteristic function of [0, 1]. Show that
there is no everywhere continuous function f on R such that
f (x) =
[0,1] (x)
almost everywhere.
Proof. Suppose that such an f exists. Then f (1) = [0,1] (1) = 1. By
continuity, 9 > 0 such that x 1 < ) f (x) 1 < 12 . In particular,
f (x) > 12 > 0 for x 2 (1, 1 + ). Thus
{x : f (x) 6=
[0,1] (x)}
(1, 1 + ) ) m({x : f (x) 6=
[0,1] (x))
> 0.
Problem 23: Suppose f (x, y) is a function on R2 that is separately continuous: for each fixed variable, f is continuous in the other variable. Prove
that f is measurable on R2 .
Solution. For all x 2 R and n 2 N, we define Dxn to be the largest nthorder
dyadic rational less than or equal to x, i.e. Dxn = 2kn where 2kn x < k+1
2n .
Let fn (x, y) = f (Dxn , y). I will show that fn is measurable and fn ! f
everywhere.
First we show that fn is measurable.
1
[
k
k+1
k
{(x, y) : fn (x, y) > a} =
(x, y) : n x <
,
f
,
y
>a
2
2n
2n
k= 1
1
[
k k+1
k
=
,
y:f
,y > a .
2n 2n
2n
k= 1
Now because f is continuous in y, {y : f ( 2kn , y) > a} is open, so the product
k
[ 2kn , k+1
2n ) {y : f ( 2n , y) > a} is measurable. Hence {fn > a} is a countable
union of measurable sets, and thus measurable.
Next, we show that fn ! f everywhere. Let > 0. For any x, y, because f
is continuous in x, there is some 1dimensional neighborhood (x , y) on
which f (x0 , y) f (x, y) < . Then for sufficiently large n, Dxn x < )
fn (x, y) f (x, y) = f (Dxn , y) f (x, y) < . Hence fn (x, y) is within of
f (x, y) for sufficiently large n. Since f is the pointwise limit of measurable
functions fn , it is measurable.
12
Problem 27: Suppose E1 and E2 are a pair of compact sets in Rd with
E1 E2 , and let a = m(E1 ) and b = m(E2 ). Prove that for any c with
a < c < b, there is a compact set E with E1 E E2 and m(E) = c.
Solution. Since E1 is measurable, there is an open set U E1 with m(U \
E1 ) < b c. Then E2 \ U c is compact (since its the intersection of a
compact set and a closed set) and has measure at least m(E1 ) m(U ) >
b (a + b c) = c a. If we can find a compact subset K E2 \ U c with
m(K) = c a, then K [ E1 will be a compact subset of E2 with measure
(c a) + a = c (since K and E1 are disjoint). Hence we have reduced the
problem to the following: Given a compact set F Rd with m(K) = ,
and given with 0 < < , find a compact subset F 0 F with m(F 0 ) = .
This can be solved as follows: Let f (y) = m(F \ By (0)). Then f (0) = 0
whereas f (y) = for sufficiently large y (because F is bounded). Moreover,
f is continuous: Given y and > 0, the continuity of m(By (0)) allows us
to find such that y 0 y < ) m(By0 (0)) m(By (0)) < . Then
f (y 0 ) f (y) = m(F \ (By0 (0) By (0))) < , because this is the measure
of a subset of the symmetric dierence (By0 (0) By (0)) which has measure
m(By0 (0)) m(By (0)) < . Hence f is continuous, so by the Intermediate
Value Theorem there is a value y0 such that f (y0 ) = . Then the compact
set F \ By0 (0) has measure as desired.
Problem 28: Let E R with m (E) > 0. Let 0 < < 1. Then there exists
an interval I R such that m (I \ E) m(I).
P
Proof. For any , we can find a cubical covering {Qj } of E with
Qj  <
m (E) + . Then, by expanding each cube to an open P
cube of size 2j more,
we can construct an open cubical covering {Ij } with
Ij  < m (E) + 2.
(We name these Ij because 1dimensional cubes are in fact intervals.) Then
[
[
X
E
Ij ) E = (E \ Ij ) ) m (E)
m (E \ Ij ).
j
Now we apply something like the Pigeonhole Principle: Suppose m (E \
Ij ) < m(Ij ) for all j. Then
X
X
m (E)
m (E \ Ij ) <
m(I)j < (m (E) + 2).
j
But if is chosen small enough, this is a contradiction; explicitly, we can
(E)
choose < (1 )m
. Thus there must be some interval Ij for which
2
m (E \ Ij ) m(Ij ).
Problem 29: Suppose E is a measurable subset of R with m(E) > 0. Prove
that the dierence set of E
{z 2 R : z = x
y for some x, y 2 E}
contains an open interval centered at the origin.
Solution. By Problem 28, there exists an interval I such that m(E \ I)
3
4 m(I). (We can replace m by m here because E is measurable.) Let
d = m(I) and 0 <  < d4 . Then the translated set I + intersects I in an
13
interval I 0 of length d > 34 d. Now m(E\I) 34 d so m((E+)\(I+))
3
4 d by the translation invariance of Lebesgue measure. Now
3
d m(E \I) = m(E \I 0 )+m(E \(I \I 0 )) m(E \I 0 )+m(I \I 0 ) = m(E \I 0 )+
4
so m(E\I 0 ) 3d
> d2 . Similarly, m((E+)\I 0 ) > d2 . Now if (E+) and
4
E were disjoint, this would imply m(I 0 ) m(E \ I 0 ) + m((E + d4 ) \ I 0 ) > d.
But m(I 0 ) = d , so E and E + must have nonempty intersection. Let
x 2 E \(E +). Then 9e1 , e2 2 E such that e1 = x = e2 + ) e1 e2 = .
Hence 2 E E for all 2 ( d4 , d4 ).
Problem 30: If E and F are measurable, and m(E) > 0, m(F ) > 0, prove
that
E + F = {x + y : x 2 E, y 2 F }
contains an interval.
Solution. We follow the preceding proof almost exactly. By the lemma,
there exist intervals I1 and I2 such that m(E \ I1 ) 34 m(I1 ) and m( F \
I2 ) 34 m(I2 ). WLOG assume m(I2 ) m(I1 ). Then 9t0 such that I2 +t0
I1 . Let d = m(I2 ). Then for 0 <  < d4 , I2 + t0 + intersects I1 in
an interval I 0 of length at least d > 3d
4 . By the same argument as
in problem 29, this implies that I1 and I2 + t0 + must have nonempty
intersection; let x be a point of this intersection. Then 9e 2 E, f 2 F such
that e = x = f + t0 + ) e + f = t0 + . Hence E + F contains the
interval (t0 d4 , t0 + d4 ).
Problem 34: Given two Cantor sets C1 , C2 [0, 1], there exists a continuous
bijection f : [0, 1] ! [0, 1] such that f (C1 ) = C2 .
Proof. Any Cantor set can be put in bijective correspondence with the set
of 01 sequences as follows: Given x 2 C, where C = C1 \ C2 \ . . . is a
Cantor set, define x1 = 0 if x is in the left of the two intervals in C1 (call
this left interval I0 ), and x1 = 1 if x is in the right interval I1 . Then define
x2 = 0 if x is in the left subinterval (either I00 or I10 ) in C2 , and x2 = 1 if x
is in the right subinterval. Continuing in this fashion, we obtain a bijection
from C to the set of 01 sequences. Note that this bijection is increasing in
the sense that if y > x for x, y 2 C, then yn > xn at the first point n in the
sequence at which xn and yn dier.
Now we can create an increasing bijection f from C1 to C2 by mapping from
C1 to 01 sequences, and from there to C2 . This function will be continuous
on C1 because if x, y 2 C1 are close, their corresponding sequences {xn }
and {yn } will agree in their first N terms; then f (x) and f (y) will agree in
their first N terms as well, which means theyre in the same subinterval of
the N th iterate of C2 , which has length at most 21N . Hence f (x) and f (y)
can be made arbitrarily close if x and y are sufficiently close. Then since
C1 is compact, we can extend f to a continuous bijection on all of [0, 1] in a
piecewise linear fashion, because C1c is a disjoint union of open intervals on
which f can be made piecewise linear. This construction will also preserve
the bijectivity of f . Hence we have a continuous bijection f : [0, 1] ! [0, 1]
with f (C1 ) = C2 .
14
Problem 35: Give an example of a measurable functions f and a continuous
function
so that f
is nonmeasurable. Use the construction in the
hint to show that there exists a Lebesgue measurable set that is not a Borel
set.
Solution. We will make use of Problem 34. Let C1 and C2 be two Cantor
subsets of [0, 1] such that m(C1 ) > 0 and m(C2 ) = 0. Let N C1 such
that N is nonmeasurable. (Here we use the fact that every set E R
of positive measure has a nonmeasurable subset. This is easy to prove
by mimicking the Vitali construction but restricting it to E: Define the
equivalence relation on E by x y if x y 2 Q, and let N E contain one member from each equivalence class. Then N and its countably
many translates are all of E, so N cannot have either measure 0 or nonzero
measure.) Define f =
= N is nonmeasurable, since
(N ) . Then f
{x : N (x) > 12 } = N .
To show that there is a Lebesguemeasurable set which is not Borel measurable, we will use (without proof) the fact that every Borel set can be
represented by a finite number of union and intersection signs, followed by
some open sets. We also use the general fact that for any function and
1
1
any sets X ,
(\X ) = \ 1 X and
([X ) = [ 1 X . Together,
1
these imply that for a continuous function ,
of a Borel set is Borel,
because
[ \ \ [ \ \
1
1
Gn =
(Gn )
1
and
(Gn ) is open for continuous and open Gn . (Of course, the string
of cups and caps in this equation could just as well start with a cap.) Now
consider the set (N ) from our construction above. Since (N ) is a subset
of the set C2 which has measure 0, (N ) is measurable by the completeness
1
of Lebesgue measure. However, is a bijection so
( (N )) = N which
is not Borel, so (N ) cannot be Borel.
1. Chapter 1.7, Page 46
Problem 2: Any open set U can be written as the union of closed cubes, so
that U = [Qj with the following properties:
(i) The Qj have disjoint interiors.
(ii) d(Qj , U c ) side length of Qj . This means that there are positive
constants c and C so that c d(Qj , U c )/`(Qj ) C, where `(Qj )
denotes the side length of Qj .
Proof. Let U Rd be an open set. Let P0 = {Qj } be the partition of U
d(Qj ,U c )
into dyadic cUbes as described in the book (pp. 78). I claim that `(Q
j)
is bounded above for this partition (i.e. does
not
take
on
arbitrarily
large
p
valUes). This is so because if d(Qj , U c ) > d`(Qj ), then Qj is a subset of
a larger dyadic cube lying within U . (If we take the dyadic
p cube one size
larger than Qj and containing Qj , then Qj is at most d`Qj away from
any point in this larger dyadic cube.) But the constrUction of P0 is done
in such a way that every dyadic cube is maximal in U , i.e. p
is not contained
d(Qj ,U c )
in a larger dyadic cube lying within U . Hence `Qj d for all Qj .
15
Now we define an iterative refining procedUre on the partition P0 . Inp this
c
)
procedure, every cube Q in the partition Pn for which d(Q,U
< 4d is
`(Q)
replaced by its 2d dyadic subcubes. If R is one of these subcubes, then
d(R, U c )
`(R)
and
p
d(R, U c )
d + d`(R)
d(Q, U c ) p
=2
+ d.
`(R)
`(R)
`(Q)
ThUs, if (Q) =
Now let
d(Q, U c )
d(Q, U c )
=2
`(Q)
`(Q)
d(Q,U c )
`(Q) ,
then the subcubes have the property
p
2(Q) (R) 2(Q) + d.
P=
1 \
1
[
n=1 k=n
Pk
be the partition
consisting of those cubes that are eventUally in all the Pn .
p
p
c
)
I claim that 4d d(Q,U
2 d for any cube Q 2 P. Consider what
`(Q)
happens as oUr refinement process iterates. If a given cube has too small
a distancetoside ratio, its subcubes will have this ratio at least doubled
in the next iteration. Hence, after enough iterations its subcubes
will all
p
p
have their distancetoside ratio in the desired interval [ 4d , 2 d]. Once
they are in this interval, they are not subdivided any further. One the
other hand, none can ever achieve a ratio too large topbe in this interval,
since a cube is only subdivided if its ratio
p is less
p than pd/4, andpthe next
iteration can then make it at most 2( d/4) + d = 3 d/2 < 2 d. This
shows that P has all its cubes in the desired interval. Consider also that
P must have disjoint interiors and cover all of U : The only cubes with
overlapping interiors are those from distinct steps in our iteration scheme,
so taking the intersection to get P will weed out any such overlaps. Also,
for a given x 2 U , if we consider the sequence of cubes containing x in
our various partitions, this sequence will shrink for a finite number of steps
and then stay constant once the distancetoside ratio reaches a desirable
number. Hence x is contained in some cube that is eventually in all the Pn ,
so x is covered by P.
Problem 3: Find an example of a measurable subset C of [0, 1] such that
m(C) = 0, yet the dierence set of C contains a nontrivial interval centered
at the origin.
Solution. Let C be the Cantor middlethirds set. Note that the Cantor
dust described in the hint consists precisely of those point (x, y) for which
both x and y are in C. Note also that if a line of slope 1 passes through
any cube in any iteration of the Cantor dust, it must pass through one
of the subcubes of that cube in the next iteration. To see this, consider
WLOG the original cube. For a line y = x + a, if 13 a 1 then the line
passes through the upper left cube; if 13 a 13 then it passes through
the lower left cube; and if 1 a 13 it passes through the lower right
cube. Thus, if Cn is the nth iteration of the Cantor dust and L is a line
16
of the form y = x + a with 1 a 1, then L \ Cn is nonempty and
compact. (It is nonempty by the preceding remark, and compact because
it is the intersection of the compact sets Cn and L \ ([0, 1] [0, 1]).) Thus,
the infinite intersection of these nested compact sets is nonempty, i.e. L
intersects the Cantor dust at some point (x, y). Then since x, y 2 C, we
have x y = a. So a is in the dierence set of the Cantor set for any
a 2 [ 1, 1].
Problem 4: Complete the following outline to prove that a bounded function
on an interval [a, b] is Riemann integrable if and only if its set of discontinuities has measure zero. This argument is given in detail in the appendix
to Book I.
Let f be a bounded function on a compact interval J, and let I(c, r) denote the open interval centered at c of radius r > 0. Let osc(f, c, r) =
sup f (x) f (y) where the supremum is taken over all x, y 2 J \ I(c, r),
and define the oscillation of f at c by osc(f, c) = limr!0 osc(f, c, r). Clearly,
f is continuous at c 2 J if and only if osc(f, c) = 0.
Prove the following assertions:
(a) For every > 0, the set of points c 2 J such that osc(f, c)
is
compact.
(b) If the set of discontinuities of f has measure 0, then f is Riemann
integrable.
(c) Conversely, if f is Riemann integrable on J, then its set of discontinuities has measure 0.
Solution.
(a) Clearly this set is bounded, so we need only show that its closed.
Suppose c1 , c2 , . . . are a sequence of such points with cn ! c. We wish
to show that osc(f, c) as well. Let r > 0. Since cn ! c, there is
some cN with cN c < 2r . Then since osc(f, cN ) , there must be
x, y 2 J within 2r of cN such that f (x) f (y) . But then x and
y are within r of c, so osc(f, c, r) . Since this is true for all r, we
must have osc(f, c) .
(b) Let > 0. Let
A = {c 2 J : osc(f, c)
}.
Since A is a subset of the discontinuity set of f , it has measure 0.
Hence there it can be covered by an open set U with m(U ) < . Since
every open subset of R P
is a countable disjoint union of intervals, we
can write U = [In with In  < . Now because A is compact, there
is a finite subcover; by reordering the intervals we may write
A
N
[
In .
n=1
Now on the compact set J 0 = J \ [In , osc(f, c) < for all c. For
each x 2 J 0 , this means we can find rx such that osc(f, x, rx ) < .
Then J 0 is covered by the open intervals Ux = (x rx , x + rx ). Let
be the Lebesgue number of this covering, so that any subinterval of
J 0 with length at most must be contained in one of the Ux . Now
17
consider a partition of J with mesh size less than . The total length of
all subintervals which intersect [In is at most + 2N since enlarging
each In by will cover all such intervals. On each of these subintervals,
sup f inf f 2M where f  M on J. Hence the contribution these
intervals make to the dierence U (P, f ) L(P, f ) is at most 2M +4M .
The other subintervals are contained in J 0 and by construction of ,
each is contained within some Ux , so sup f inf f on each of them.
Hence the total contribution they make to U (P, f ) L(P, f ) is at most
m(J). Thus, we have
U (P, f )
L(P, f ) (2M + m(J) + 4 ).
By requiring to be less than some constant times , we have thus
shown that the dierence between upper and lower sums can be made
smaller than a constant times . Hence f is Riemann integrable.
(c) Suppose f is Riemann integrable, and let > 0. Let n 2 N. Then
there is a partition P of J with U (P, f ) L(P, f ) < n . Now if the
interior of any subinterval Ik of this partition intersects A1/n at some
1
1
x, then sup f inf f
n on Ik because osc(f, x, r)
n for all r, and
(x r, x + r) Ik for sufficiently small r. So the total length of the
subintervals whose interiors intersect A1/n is at most since otherwise
they would make a contribution of more than /n to U (P, f ) (P, f ).
Hence we have covered A1/n by a collection of intervals of total length
less than , which implies m(A1/n ) < . Now if A is the set of points
at which f is discontinuous, then
1
[
A=
A1/n .
n=1
Since An An+1 , the continuity of measure implies that
m(A) = lim m(A1/n ) .
n!1
Since m(A) for all , m(A) = 0.
Problem 6: The fact that the axiom of choice and the wellordering principle
are equivalent is a consequence of the following considerations.
One begins by defining a partial ordering on a set E to be a binary relation
on the set E that satisfies:
(i) x x for all x 2 E.
(ii) If x y and y x, then x = y.
(iii) If x y and y z, then x z.
If in addition x y or y x whenever x, y 2 E, then is a linear ordering
of E.
The axiom of choice and the wellordering principle are then logicall equivalent to the Hausdor maximal principle: Every nonempty partially ordered
set has a (nonempty) maximal linearly ordered subset. In other words, if
E is partially ordered by , then E contains a nonempty subset F which
is linearly ordered by and such that if F is contained in a set G also
linearly ordered by , then F = G.
An application of the Hausdor maximal principle to the collection of all
18
wellorderings of subsets of E implies the wellordering principle for E.
However, the proof that the axiom of choice implies the Hausdor maximal
principle is more complicated.
Solution. I dont like this problem for two reasons: (1) Its not very clearly
stated what exactly Im supposed to do. My best guess is that Im supposed
to use the axiom of choice to prove the Hausdor maximal principle, but
the problem never really comes out and says that. (2) What the crap. This
is supposed to be an analysis course, not a set theory course. We havent
defined any of the basic concepts of set theory, yet Im supposed to come
up with a complicated set theory proof. Since Ive never had a course in
set theory, I really dont feel like guessing my way to something that might
be a proof. So, heres the proof from the appendix to Rudins Real and
Functional Analysis:
For a collection of set F and a subcollection F, call a subchain of
F if is totally ordered by set inclusion.
Lemma 1. Suppose F is a nonempty collection of subsets of a set X such
that the union of every subchain of F belongs to F. Suppose g is a function
which associates to each A 2 F a set g(A) 2 F such that A g(A) and
g(A) \ A consists of at most one element. Then there exists an A 2 F for
which g(A) = A.
Proof. Let A0 2 F. Call a subcollection F 0 F a tower if A0 2 F 0 , the
union of every subchain of F 0 is in F 0 , and g(A) 2 F 0 for every A 2 F 0 .
Then there exists at least one tower because the collection of all A 2 F
such that A0 A is a tower. Let F0 be the intersection of all towers,
which is also a tower. Let be the collection of all C 2 F0 such that every
A 2 F0 satisfies either A C or C A. For each C 2 , let (C) be the
collection of A 2 F0 such that A C or g(C) A. We will prove that is
a tower. The first two properties are obvious. Now let C 2 , and suppose
A 2 (C). If A is a proper subset of C, then C cannot be a proper subset
of g(A) because then g(A) \ A would have at least two elements. Hence
g(A) C. If A = C, then g(A) = g(C). The third possibility for A is that
g(C) A. But since A g(A) this implies g(C) g(A). Thus, we have
shown that g(A) 2 (C) for any A 2 (C), so (C) is a tower. By the
minimality of F0 , we must have (C) = F0 for every C 2 . This means
that g(C) 2 for all C 2 , so is also a tower; by minimality again,
= F0 . This shows that F0 is totally ordered.
Now let A be the union of all sets in F0 . Then A 2 F0 by the second tower
property, and g(A) 2 F0 by the third. But A is the largest member of F0
and A g(A), so A = g(A).
Now let F be the collection of all totally ordered subsets of a partially
ordered set E. Since every singleelement subset of E is totally ordered, F
is not empty. Note that the union of any chain of totally ordered sets is
totally ordered. Now let f be a choice function for E. If A 2 F, let A be
the set of all x in the complement of A such that A [ {x} 2 F. If A 6= ;,
let
g(A) = A [ {f (A )}.
19
If A = ;, let g(A) = A.
By the lemma, A = ; for at least one A 2 F, and any such A is a maximal
element of F.
Problem 7: Consider the curve = {y = f (x)} in R2 , 0 x 1. Assume
that f is twice continuously dierentiable in 0 x 1. Then show that
m( + ) > 0 if and only if + contains an open set, if and only if f is
not linear.
Solution. We are asked to show the equivalence of the conditions (i) m( +
) > 0, (ii) + contains an open set, and (iii) f is not linear. We will
show that (ii) implies (i), which implies (iii), which implies (ii).
First, we should note that + is measurable. The problem doesnt ask for
this, but its worth pointing out. Consider G : [0, 1] [0, 1] ! R2 defined
by G(x, y) = (x + y, f (x) + f (y)). Then + is just the range of G. Since
dierentiable functions map measurable sets to measurable sets, + is
measurable. (We havent proved yet that dierentiable functions preserve
measurability, but I assume we will once we get further into dierentiation
theory.)
The easiest of our three implications is (ii) implies (i). Suppose +
contains an open set. Open sets have positive measure, so + is a
measurable set with a subset of positive measure, so it has positive measure.
Now suppose m( + ) > 0. We wish to show that f is not linear. Suppose
instead that f is linear, say f (x) = ax+b. Then for any x, x0 , (x+x0 , f (x)+
f (x0 )) = (x + x0 , a(x + x0 ) + 2b) so + is a subset of the line y = ax + 2b,
which has measure 0. Thus, if + has positive measure, f must not be
linear.
The third implication is the least trivial. Suppose f is not linear. Then
there are points x0 , y0 2 [0, 1] with f 0 (x0 ) 6= f 0 (y0 ). Then the Jacobian
DG =
1
1
= f 0 (y)
f 0 (x) f 0 (y)
f 0 (x)
is nonzero at the point (x, y) 2 [0, 1] [0, 1], where G(x, y) = (x + y, f (x) +
f (y)) as above. WLOG we may assume (x, y) 2 (0, 1) (0, 1) since a
nonlinear function on [0, 1] with continuous derivative cannot have constant derivative everywhere on (0, 1). Then the Inverse Function Theorem
guarantees that there is an open neighborhood of (x, y) on which G is a
dieomorphism; since dieomorphisms are homeomorphisms, this implies
that the image of G contains an open set.
Chapter 2.5, Page 89
Exercise 1: Given a collection of sets F1 , . . . , Fn , construct another collection
Sn
SN
F1 , . . . , FN , with N = 2n 1, so that k=1 Fk = j=1 Fj ; the collection
S
{Fj } is disjoint; and Fk = F Fk Fj for every k.
j
Solution. For j = 1, . . . , N , create Fj as follows: First write j as an ndigit
binary number j1 j2 . . . jn . Then for k = 1, . . . , n, let
(
Fk jk = 1
Gk =
Fkc jk = 0.
20
Finally, let
Fj =
n
\
Gk .
k=1
For example, 2 = 000 . . . 10 in binary, so
F2 = F1c \ F2c \ \ Fnc
\ Fn
\ Fnc .
Note that the Fj are pairwise disjoint because if j 6= j 0 , then they dier
in some binary digit, say j` 6= j`0 . Suppose WLOG that j` = 1 and j`0 = 0.
Then Fj F` whereas Fj0 F`c , so they are disjoint.
Also,
[
Fk =
Fj .
Fj Fk
To see this, note that the RHS is clearly a subset of the LHS since it is a
union of subsets. Conversely, suppose x 2 Fk . Define x1 , . . . , xn by xi = 1 if
x 2 Fi and 0 otherwise. Then if m has the binary digits m1 = x1 , . . . , mn =
xn , x 2 Fm
by definition of Fm
. Since Fm
Fk , the result follows.
This implies
n
N
[
[
Fi
Fj .
But Fj
i=1
j=1
Fi for each j, so the reverse inclusion holds as well.
Exercise 2: In analogy to Proposition 2.5, prove that if f is integrable on Rd
and > 0, then f ( x) converges to f (x) in the L1 norm as ! 1.
Solution. Let > 0. Since CC (Rd ) is dense in L1 (Rd ), we can choose
g 2 CC (Rd ) such that kf gk1 < 3 . We can also choose such that

1 < ) 1d < 32 . Now let K = supp(g), and choose M such that
K {x M }. Let B = {x M + 1}, which is also compact. Because
g is uniformly continuous on B, 9 0 > 0 such that 0 < 1 and for x, y 2 B,
x y < 0 ) g(x) g(y) < 6m(B)
. Now suppose we choose such that
0
1 < min( , M +1 ). Then
Z
kf ( x) f (x)k = f ( x) f (x)
Z
Z
f ( x) g( x) + f ( x)

g(x) +
g(x)
f (x)
by the triangle inequality. The third integral is just kf gk 3 . By the
dilation property of the integral, the first integral is 1d kf gk 32 3 = 2 .
Now for the second integral. I claim that the integrand is 0 outside B and
at most 6m(B)
inside. If x 2
/ B, then x > M + 1 so
0
0
 x =  x
1
x > 1
(M + 1) > M,
M +1
M +1
which implies that g( x) = g(x) = 0. Now suppose x 2 B. If g(x) 6= 0,
then x 2 K and
0
0
 x 1 +
x < 1 +
M <M +1
M +1
M +1
21
so x 2 B and  x x < 0 ) g( x) g(x) < 6m(B)
. If g( x) 6= 0,
then x 2 K B and again g( x) g(x) < 6m(B) because of uniform
continuity. Hence the integrand is 0 outside B and is at most 6m(B)
inside
B, so its integral is at most 6m(B) m(B) = 6 . Putting the pieces together,
we have kf ( x) f (x)k < .
Exercise 4: Suppose f is integrable on [0, b], and
Z b
f (t)
g(x) =
dt for 0 < x b.
t
x
Prove that g is integrable on [0, b] and
Z b
Z b
g(x)dx =
f (t)dt.
0
Solution. We may assume WLOG that f (t)
analyze f + and f separately. Now let
0, since otherwise we can
f (t)
{0<xtb} .
t
Then h 0 and h is clearly measurable since it is a quotient of measurable
functions times another measurable function. By Tonellis theorem,
Z 1
Z b
f (t)
h(x, t)dt =
{0<xb} dt
t
1
x
h(x, t) =
is a measurable function of x. Note that this is just equal to g(x) for
0 < x b and 0 elsewhere. Hence g is measurable (in general, g : (0, b] ! R
is measurable i g (0,b] : R ! R is). Moreover, Tonellis theorem also tells
us that
Z b
Z b Z t
f (t)
g(x) dx =
dt dx
t
0
x
Z0
=
h(x, t)
=
RR
b Z t
h(x, t) dx dt
f (t)
dx dt
t
f (t)
t
dt
t
f (t) dt.
Note that the fact that this integral is finite implies that g is integrable and
not just measurable.
Exercise 5: Suppose F is a closed set in R, whose complement has finite
measure, and let (x) denote the distance from x to F , that is,
(x) = d(x, F ) = inf{x
y : y 2 F }.
22
Consider
Z
I(x) =
(a) Prove that
condition
x
(y)
dy.
y2
is continuous by showing that it satisfies the Lipschitz
 (x)
(y) x
y.
(b) Show that I(x) = 1 for each x 2
/ F.
(c) Show that I(x) < 1 for a.e. x 2 F . This may be surprising in view of
the fact that the Lipschitz condition cancels only one power of x y
in the integrand of I.
Solution.
(a) Let > 0. Choose z 2 F such that y
(x) x
z x
y + y
z < x
z < (y) + . Then
y + (y) + ) (x)
(y) < x
y + .
Interchanging the roles of x and y, we also have
(y)
(x) < x
y + .
Hence  (x)
(y) < x y + for any > 0. This implies  (x)
(y) x y.
(b) Suppose x 2
/ F . Because F is closed, this implies (x) > 0, since
otherwise there would be a sequence of points in F converging to x.
Let = (x). By the Lipschitz condition from part (a), x y < 2 )
 (y)
 < 2 ) (y) 2 . Hence
I(x) =
Z
Z
1
1 x
x+ /2
x
(y)
dy
y2
2
/2
x+ /2
(y)
dy
y2
/2
/2
/2
x
y2
dy
1
dy = 1.
y2
(c) First, consider that for y 2
/ F,
Z
1
x
y2
dx 2
1
(y)
1
2
dx =
x2
(y)
23
since F {x : x y
(y)}. Now since I(x)
0, we have by
Tonellis theorem
Z
Z
(y)
I(x) dx =
F (x)
y2
F
RR x
Z Z
(y)
=
(x)
dx
dy
F
y2
R
R x
Z
Z
1
=
(y)
dx
dy
y2
c
F x
ZF
2
(y)
dy
(y)
Fc
= 2m(F c ) < 1.
R
Since F I(x) dx < 1, we must have I(x) < 1 for almost all x 2 F .
(This is actually not all that shocking, since I(x) is clearly less than 1
for an interior point. Of course, there are closed sets whose boundaries
have positive measure, but those are the nasty guys.)
Exercise 6: Integrability of f on R does not necessarily imply the convergence of f (x) to 0 as x ! 1.
(a) There exists a positive continuous function f on R such that f is
integrable on R, yet lim supx!1 f (x) = 1.
(b) However, if we assume that f is uniformly continuous on R and integrable, then limx!1 f (x) = 0.
Solution.
(a) Let
(
23n+4 d x, n, n +
f (x) =
0
c
1
22n+1
nxn+
else.
1
22n+1 , n
2Z
The graph of f consists of a series of triangular spikes withR height
1
n+2
n
2
P1 andn base 22n+1 . The nth such spike has area 2 , so f  =
= 2. But, because the spikes get arbitrarily high, lim sup f (x) =
n=0 2
1.
(b) Suppose f is uniformly continuous on R, and let > 0. Select > 0
such that x y < ) f (x) f (y) < 2 , and also require < 12 .
Since f (x) 6! 0, 9x1 > 0 such that f (x1 ) . Then f (y)
2 for
y 2 (x1 , x1 + ). Now since f (x) 6! 0, 9x2 > x1 +1 with f (x2 ) .
Then f  2 on (x2 , x2 + ). Continuing in this manner, we obtain
infinitely many intervals of length 2 on which f  2 . These intervals
are disjoint because of our requirements that xn+1 x > 1 and < 12 .
Hence, by Tchebyches inequality,
Z
f (x) dx
m({x : f (x)
}) = 1.
2
2
R
24
Exercise 8: If f is integrable on R, show that
Z x
F (x) =
f (t)dt
1
is uniformly continuous.
Solution. Let > 0. By the absolute
continuity of the integral (Prop 1.12b),
R
9 > 0 such that m(E) < ) E f  < . Then (assuming WLOG x > y),
Z x
Z x
x y < ) F (x) F (y) =
f (t)dt
f (t) dt <
y
because m([y, x]) = x
y < .
Exercise 10: Suppose f
0, and let E2k = {x : f (x) > 2k } and Fk = {x :
k
k+1
2 < f (x) 2
}. If f is finite almost everywhere, then
1
[
k= 1
Fk = {f (x) > 0}
and the sets Fk are disjoint.
Prove that f is integrable if and only if
1
X
k= 1
2k m(Fk ) < 1,
1
X
if and only if
k= 1
2k m(E2k ) < 1.
Use this result to verify the following assertions. Let
(
x a if x 1
f (x) =
0
otherwise
and
(
x
g(x) =
0
if x > 1
otherwise.
Then f is integrable on Rd if and only if a < d; also g is integrable on Rd
if and only if b > d.
Solution. Let
g(x) =
1
X
2k
Fk (x),
k= 1
h(x) =
1
X
2k+1
Fk (x).
k= 1
Then g(x) f (x) h(x) by definition of Fk . Then
Z
Z
1
X
f (x)dx < 1 ) g(x)dx =
2k m(Fk ) < 1
whereas
1
X
k= 1
2 m(Fk ) < 1 )
k
k= 1
f (x)dx <
h(x)dx =
1
X
k= 1
k+1
m(Fk ) = 2
1
X
k= 1
2k m(Fk ) < 1.
25
Now let
(x) =
1
X
2k
Ek (x).
k= 1
Then f (x) (x) 2f (x) because if 2k < f (x) 2k+1 , (x) =
1 + 1 + 2 + 4 + + 2k = 2k+1 . Hence
Z
Z
1
X
f (x)dx < 1 ,
(x)dx =
2k m(Ek ) < 1.
j= 1
2k =
k= 1
Now for the function f given,
(
{x 1}
Ek = {f (x) > 2 } =
{x 2 k/a }
k
so
Pk
(
2d
m(Ek ) =
2d 2
k0
k 1
k0
.
k 1
kd/a
So f is integrable i
1
X
k= 1
2k m(Ek ) =
0
X
2k 2d +
k= 1
1
X
2k 2d 2
kd/a
= 2d+1 + 2d
1
X
2(1
k= 1
k=1
d/a)k
< 1.
This infinite sum will converge i the constant 1 ad is negative, i.e. i
a < d.
For the function g given, let us redefine g(x) = 1 for x 1; clearly this
does not aect the integrablity of g. Now Ek is empty for k > 0, so we
need only consider negative values of k.
g(x) > 2k , x < 2
so Ek is a cube of volume 2d 2
0
X
k= 1
2k 2d 2
kd/b
kd/b
k/b
. Hence g is integrable i
= 2d
0
X
2(1
d/b)k
k= 1
d/b > 0 ) b > d.
R
Exercise 11: Prove that if f is integrable on Rd , and R E f (x)dx 0 for every
measurable E, then f (x) 0 a.e. x. As a result, if E f (x)dx = 0 for every
measurable E, then f (x) = 0 a.e.
converges. This will happen i 1
Proof. Suppose it is not true that f (x) 0 a.e., so m({f (x) < 0} is positive.
Now
1
1
[
X
1
{x : f (x) < 0} =
x : f (x) <
) m({x : f (x) < 0})
m
x : f (x) <
n
n=1
n=1
by countable additivity. Hence at least one of the sets
1
En = x : f (x) <
n
has positive measure. But then
Z
Z
1
1
f (x)dx
dx =
m(En ) < 0.
n
n
En
En
1
n
26
R
By contraposition, if E f (x)dx 0 for every measurable set E, then f (x)
0 a.e. R
R
Now
if E f (x)dx = 0 for every measurable E, then E f (x)dx
0 and
R
f
(x)dx
0,
which
means
f
0
a.e.
and
f
0
a.e.
Hence
f
=0
E
a.e.
Exercise 12: Show that there are f 2 L1 (Rd ) and a sequence {fn } with
fn 2 L1 (Rd ) such that
kf
but fn (x) ! f (x) for no x.
fn k1 ! 0,
Solution. To assist in constructing such a sequence, we first construct a
sequence of measurable sets En Rd with the property that m(En ) ! 0
but every x 2 Rd is in infinitely many En . We proceed as follows: Choose
integers N1 , N2 , . . . such that
1
1
1 + + +
> 10
2
N1
1
1
1
+
+ +
> 100
N1 + 1 N1 + 2
N2
1
1
1
+
+ +
> 1000
N2 + 1 N2 + 2
N3
etc. This is possible because of the divergence of the harmonic series.
For convience, we also define N0 = 0. Next, for each k = 0, 1, 2, . . . , let
BNk +1 be the cube of volume Nk1+1 centered at the origin. Then, for a
given k, define Bj for Nk + 1 < j Nk+1 to be the cube centered at the
origin with Bj  = Bj 1  + 1j . Finally, we define ENk +1 = BNk +1 , and for
Nk + 1 < j Nk+1 , Ej = Bj \ Bj 1 . I claim that the sets En have the
desired properties. First, note that m(En ) = n1 . This is obvious for Nk + 1;
for Nk +1 < j Nk+1 it is easy to see inductively that Bj  = Nk1+1 + + 1j
and since they are nested sets, Bj \ Bj 1  1j . Thus m(En ) ! 0. However,
for each k,
Nk+1
[
Ej
j=Nk +1
is a cube centered at the origin with a volume greater than 10k . For any
given x, these cubes will eventually contain x, i.e. there is some K such
that
Nk+1
[
k>K)x2
Ek .
j=Nk +1
Hence every x is in infinitely many Ej as desired.
Having constructed these sets, we simply let fn (x) = En (x) and f (x) = 0.
R
R
R
L1
Then fn (x)dx = n1 so kfn f k = fn f  = fn ! 0, i.e. fn ! f .
However, for any given x there are infinitely many n such that fn (x) = 1,
so fn (x) 6! f (x) for any x.
Exercise 13: Give an example of two measurable sets A and B such that
A + B is not measurable.
27
Solution. As suggested in the hint, let N R be a nonmeasurable set.
Let A = {0} [0, 1] and B = N {0}. Then A and B are both measurable
subsets of R2 because they are subsets of lines, which have measure 0. Now
A + B = N [0, 1]. By Proposition 3.4, if N [0, 1] were measurable, since
[0, 1] has positive measure, this would imply that N was measurable too, a
contradiction. Hence A + B is not measurable.
Exercise 15: Consider the function defined over R by
(
x 1/2 if 0 < x < 1,
f (x) =
0
otherwise.
For a fixed enumeration {rn } of the rationals Q, let
F (x) =
1
X
f (x
rn ).
n=1
Prove that F is integrable, hence the series defining F converges for almost
every x 2 R. However, observe that this series is unbounded on every
interval, and in fact, any function F that agrees with F a.e. is unbounded
in any interval.
Solution. First we compute the integral of f ; the improper Riemann integral is
Z 1
p 1
1
p dx = 2 x = 2,
x
0
0
but we only proved that the Lebesgue and Riemann integrals are equal for
the proper Riemann integral. Of course its true for improper integrals as
1
1
well; here, since (0, 1] = [1
n=1 ( n+1 , n ], we have by countable additivity that
Z
1 Z
X
f dx =
f dx
R
1
1
( n+1
,n
]
n=1
N Z
X
= lim
N !1
= lim
N !1
= lim
a!0
= 2.
n=1
Z 1
1
N +1
1
1
( n+1
,n
]
f dx
f dx
f dx
(since this limit exists)
By translation invariance, the integral of f (x rn ) is also 2. Now since f
is nonnegative everywhere, the partial sums are monotonely increasing, so
by the Monotone Convergence Theorem
Z
1 Z
1
X
X
n
F dx =
2 f (x rn )dx =
21 n = 2.
n=1
n=1
Since this integral is finite, F is integrable. This implies that F is finitevalued for almost all x 2 R.
Now let F be any function that agrees with F almost everywhere, and I
28
any interval on the real line. Let rN be some rational number contained in
I. Then for any M > 0, f (x rN ) > M on the interval (rN M12 , rN + M12 ),
which intersects I in an interval IM of positive measure. Since F agrees
with F almost everywhere, it must also be greater than M at almost all
points of this interval IM I. Hence F exceeds any finite value M on
I.
Exercise 17: Suppose f is defined on R2 as follows: f (x, y) = an if n x <
n + 1 and n y < n + 1, n
0; f (x, y) = an if n xP< n + 1 and
n + 1 y < n + 2, n 0; f (x, y) = 0 elsewhere. Here an = kn bk , with
P1
{bk } a positive sequence such that k=0 bk = s < 1.
R
y
(a) Verify that each
R Rslice f and fx is integrable. Also for all x, fx (y)dy =
0, and hence
( f (x, y)dy)dx = 0.
R
R y
(b) However, f y (x)dx = a0 if 0 y < 1, and
R y f (x)dx = an an 1
if n y < n + 1 with n 1. Hence y 7! f (x)dx is integrable on
(0, 1) and
Z Z
f (x, y)dx dy = s.
R
(c) Note that RR f (x, y)dxdy = 1.
Solution.
(a) Since f is constant on boxes and 0 elsewhere, the horizontal and vertical slices are constant on intervals and 0 elsewhere, and therefore
integrable. More precisely,
8
>
< abyc 1 byc 1 x < byc
y
f (x) = abyc
byc x < byc + 1
>
:
0
else
for y
1,
(
a0
f (x) =
0
y
for 0 y < 1, and
8
>
<abxc
fx (y) =
abxc
>
:
0
0x<1
else
bxc y < bxc + 1
bxc + 1 y < bxc + 2
else
where bxc is the greatest integer less than or equal to x. (For x < 0
the function fx (y) is identically
0, and for y < 0 the function f y (x)
R
is identically 0.) Clearly fx (y)dy = 0 for all x, since fx (y) is equal
to abxc on an interval of Rlength
1 and abxc on an interval of length 1
R
and 0 elsewhere. Hence
f (x, y)dydx = 0.
(b) Since all the integrals are of constants on intervals Rof length 1, it
immediately follows from the formulas in part (a) that f y (x)dx is a0
for 0 y < 1 and an an 1 = bn for n y < n + 1. Then
X
Z Z
1 Z n+1 Z
1
X
y
y
f (x)dx =
f (x)dx dy =
bn = s.
R
n=0
n=0
29
(c) Since f (x, y) is positive, we may use Tonellis theorem, so
Z
Z Z
f (x, y) =
fx (y)dy dx
RR
=
=
1 Z
X
n=0
1
X
n=0
n+1
fx (y)dx dx
2an = 1
since an > a0 so the terms in the sum are bounded away from 0.
Exercise 18: Let f be a measurable finitevalued function on [0, 1], and suppose that f (x) f (y) is integrable on [0, 1] [0, 1]. Show that f (x) is
integrable on [0, 1].
Solution. Let g(x, y) = f (x) f (y). By Fubinis Theorem, since g is
integrable on [0, 1] [0, 1], g y (x) is an integrable function of x for almost
all y 2 [0, 1]. Choose any such y. Then since f (x) f (y) f (x) f (y),
Z 1
Z 1
(f (x) f (y)) dx
f (x) f (y)dx < 1
0
so
f (x)dx f (y) +
f (x)
f (y)dx < 1.
Exercise 19: Suppose f is integrable on Rd . For each > 0, let E = {x :
f (x) > }. Prove that
Z
Z 1
f (x)dx =
m(E )d.
Rd
Solution. By Tonellis Theorem,
Z 1
Z 1 Z
m(E )d =
f (x)> dx d
0
0
Rd
Z
Z
1
=
f (x)> d dx
d
0
ZR
=
f (x)dx.
Rd
Exercise 22: Prove that if f 2 L1 (Rd ) and
Z
f() =
f (x)e 2ix dx,
Rd
then f() ! 0 as  ! 1. (This is the RiemannLebesgue lemma.)
30
Solution. By translation invariance,
Z
f() =
f (x)e 2ix dx
d
ZR
2i(x 2
2 )
=
f x
e
dx
2
2
Rd
2i
=
f x
e 2ix e 22 dx
2

Rd
=
f x
e 2ix dx
22
Rd
so, multiplying by 12 and adding the original expression,
Z
Z
1
2ix
f() =
f (x)e 2ix dx
f x
e
dx
2
22
Rd
Rd
Z
1
=
f (x) f x
e 2ix dx
2 Rd
22
and
Z
1
f () =
f (x) f x
e
2 Rd
22
Z
1
f (x) f x
e
2 Rd
22
Z
1
=
f (x) f x
dx
2 Rd
22
=
f (x) f x
.
2
22 1
As  ! 1,  2
2  ! 0, so kf (x)
of translation (Proposition 2.5).
f (x
22 )k1
2ix
2ix
dx
dx
! 0 by the L1 continuity
Exercise 23: As an application of the Fourier transform, show that there
does not exist a function I 2 L1 (Rd ) such that
f I =f
for all f 2 L1 (Rd ).
= f()
Solution. Suppose such an I exists. Then for every f 2 L1 , f()I()
for all . This implies that I()
= 1 for all . But this contradicts the
RiemannLebesgue Lemma (Problem 22). QED.
Note: To be totally complete, we should show that for any there is a
function g 2 L1 such that g() 6= 0. Otherwise, the equation fI = f
wouldnt necessarily imply I = 1 everywhere. But it is easy to show that
such a g exists for any ; for example, g could be equal to e2ix on some
compact set and 0 outside.
Exercise 24: Consider the convolution
Z
(f g)(x) =
f (x
Rd
y)g(y)dy.
(a) Show that f g is uniformly continuous when f is integrable and g
bounded.
31
(b) If in addition g is integrable, prove that (f g)(x) ! 0 as x ! 1.
Solution.
(a) Since g is bounded, 9M with g < M everywhere. Then
Z
Z
f g(x) f g(x0 ) =
f (x y)g(y)dy
f (x0 y)g(y)dy
Rd
Rd
Z
=
(f (x y) f (x0 y)) g(y)dy
Rd
Z
f (x y) f (x0 y) g(y)dy
Rd
Z
M
f (x y) f (x0 y)dy
d
ZR
=M
f (y) f (y + (x0 x))dy
Rd
= M kf (y)
f (y + (x
x0 ))k1 .
In the penultimate
step
R
R we have used translation invariance and the
fact that f (y)dy = f ( y)dy provided both integrals are taken over
all of Rd . Now by the L1 continuity of translation, 9 > 0 such that
x x0  < ) kf (y) f (y + (x x0 ))k1 < M
. This in turn implies
0
f g(x) f g(x ) < , so f g is uniformly continuous.
(b) Let > 0. Since CC (Rd ) is dense in L1 (Rd ), we may choose f such
that supp(f) = K is compact, f is continuous, and kf fk1 < 2M
,
where M is a bound for g as in part (a). Continuous functions with
compact support are bounded, so choose N such that f < N . Now
Z
f g(x) =
f (x y)g(y)dy
Rd
Z
=
f(x y) + (f (x y) f(x y)) g(y)dy
d
ZR
f(x y) + (f (x y) f(x y))g(y) dy
Rd
Z
Z
f(x y)g(y)dy +
f (x y) f(x y) g(y)dy.
Rd
Rd
Call the first integral I1 and the second I2 . Since g < M , I2
M kf fk1 <R 2 . Now since g is integrable, there must exist compact
F such that F c g < 2N
. Then if x is larger than the sum of the
diameters of K and N ,
Z
I2 =
f(x y)g(y)dy
Rd
Z
=
f(x y)g(y)dy
since f = 0 on K c
x y2K
Z
N
g(y)dy
x y2K
Z
N
g(y)dy <
2
c
F
32
since {y : x
I1 + I2 < .
y 2 K} F c . Thus, for sufficiently large x, f g(x)
Chapter 2.6, Page 95
R 2
Problem 1: If f is integrable on [0, 2], then 0 f (x)e inx dx ! 0 as n !
1. Show as a consequence that if E is a measurable subset of [0, 2], then
Z
m(E)
cos2 (nx + un )dx !
, as n ! 1
2
E
for any sequence {un }.
R 2
R 2
Solution. First, note that 0 f (x) cos(nx)dx ! 0 and 0 f (x) sin(nx) !
R 2
0 since these are the real and imaginary parts of 0 f (x)e inx dx. In
particular, if we let f (x) = E (x) for some measurable E [0, 2], then for
R 2
R 2
any > 0, 9N such that  0
E (x) sin(nx)dx and  0
E (x) cos(nx)dx
are both less than 2 provided n > N . Then for any sequence un ,
Z
Z
cos(2nx + 2un )dx =
cos(2nx) cos(2un ) sin(2nx) sin(2un )dx
E
= cos(2un )
E (x) cos(2nx)dx
sin(2un )
 cos(2un )
E (x) cos(2nx)dx +  sin(2un )
+1 =
2
2
R
for n > N . Hence E cos(2nx + un )dx ! 0 as n ! 1. Now
Z
Z
1
cos2 (nx + un ) =
(1 + cos(2(nx + un ))) dx
E
E 2
Z 2
m(E)
=
+
E (x) cos(2nx + 2un )dx
2
0
1
and we have shown that the second term tends to 0 as n ! 1.
E (x) sin(2nx)dx
E (x) sin(2nx)dx
Problem 2: Prove the CantorLebesgue theorem: if
1
X
n=0
An (x) =
1
X
(an cos nx + bn sin nx)
n=0
converges for x in a set of positive measure (or in particular for all x), then
an ! 0 and bn ! 0 as n ! 1.
p
Solution. We can rewrite An (x) = cn cos(nx + dn ) where cn = a2n + b2n
and
P dn is some phase angle (it can be arctan(bn /an ), for example). If
An (x) converges on some E with m(E) > 0, then An (x) ! 0 on E. By
33
Egorovs theorem, this implies An ! 0 uniformly on some E 0 e with
m(E 0 ) > 0. Then
u
on E 0
on E 0
cn cos(nx + dn ) ! 0
E0
) c2n cos2 (nx + dn ) ! 0
c2n cos2 (nx + dn )dx ! 0.
R0
0
)
But E cos2 (nx+dn )dx ! m(E
by the previous problem, so c2n ! 0, which
2
implies cn ! 0, which implies an ! 0 and bn ! 0.
Problem 3: A sequence {fk } of measurable functions on Rd is Cauchy in
measure if for every > 0,
m({x : fk (x)
f` (x) > }) ! 0
as k, ` ! 1.
We say that {fk } converges in measure to a (measurable) function f if
for every > 0,
m({x : fk (x)
f (x) > }) ! 0
as k ! 1.
This notion coincides with the convergence in probability of probability
theory.
Prove that if a sequence {fk } of integrable functions converges to f in L1 ,
then {fk } converges to f in measure. Is the converse true?
Solution. Suppose fn , f 2 L1 and fn ! f in L1 . By Chebyshevs Inequality,
kfn f k1
m ({x : fn (x) f (x) > })
and since the RHS tends to 0 as n ! 1, the LHS does as well, so fn !
f in measure. However, the converse does not hold. Consider the case
f (x) = 0, fn (x) = n [0, n1 ] . Then m({x : f (x) 6= fn (x)}) = n1 so for any ,
m({x : f (x) fn (x) > }) n1 ! 0. Hence fn ! f in measure. However,
kfn f k1 = 1 for all n, so fn 6! f in L1 .
Problem 4: We have already seen (in Exercise 8, Chapter 1) that if E is a
measurable set in Rd , and L is a linear transformation of Rd to Rd , then
L(E) is also measurable, and if E has measure 0, then so has L(E). The
quantitative statement is
m(L(E)) =  det(L)m(E).
As a special case, note that the Lebesgue measure is invariant under rotations. (For this special case see also Exercise 26 in the next chapter.)
The above identity can be proved using Fubinis theorem as follows.
(a) Consider first the case d = 2, and L a strictly upper triangular
transformation x0 = x + ay, y 0 = y. Then
L(E) (x, y)
E (L
(x, y)) =
E (x
ay, y).
34
Hence
m(L(E)) =
=
RR
RR
Z
Z
= m(E),
ay, y)dx dy
(x,
y)dx
dy
E
E (x
by the translationinvariance of the measure.
(b) Similarly m(L(E)) = m(E) if L is strictly lower triangular. In general,
one can write L = L1 L2 , where Lj are strictly (upper and lower)
triangular and
is diagonal. Thus m(L(E)) =  det Lm(E), if one
uses Exercise 7 in Chapter 1.
Solution.
(a) Im not quite sure what to comment on here, since the problem statement pretty much did all the work for me. I guess I should point out
that the use of Tonellis theorem to turn the double integral into an
iterated integral is justified because E is nonnegative; note that this
proves that the result holds even if m(E) = 1.
(b) The fact that lower triangular transformations work the same way is
obvious. Now supposing L = L1 L2 , we have
m(L(E)) = m(L1 ( (L2 (E)))) = m( (L2 (E)))
=  det( )m(L2 (E)) =  det( )m(E) =  det(L)m(E)
since det(L) = det(L1 ) det( ) det(L2 ) = det( ) and m( (E)) =
 det( )m(E) by Exercise 7 of Chapter 1. Thus, every linear transformation that has an LU decomposition works as we want it to. However, I think the problem is flawed because not every matrix has an
LU decomposition (in fact, not even every invertible matrix does.) In
particular, in the 2 2 case a matrix of the form L1 L2 will look like
either
1 c
d1 0
1 0
d1 + cd2 e cd2
=
0 1
0 d2
e 1
d2 e
d2
or
1 0
d1 0
1 c
d1
cd1
=
.
e 1
0 d2
0 1
ed1 ed1 c + d2
But a matrix of the form
0
b
a
0
with a, b 6= 0 cannot be put in either form, because in the first case we
would need d2 = 0 in order to make the lower right entry 0, and then
the upper right and lower left entries could not be nonzero; similarly,
in the second case, we would need d1 = 0 which would make a, b 6= 0
impossible. Hence, there are some matrices that cannot be factored
in the way this problem indicates; for the ones that can, though, we
know that they expand measures by a factor of  det .
35
Chapter 3.5, Page 145
R
Exercise 1: Suppose is an integrable function on Rd with Rd (x)dx = 1.
Set K (x) = d (x/ ), > 0.
(a) Prove that {K } >0 is a family of good kernels.
(b) Assume in addition that is bounded and supported in a bounded
set. Verify that {K } >0 is an approximation to the identity.
(c) Show that Theorem 2.3 holds for good kernels as well.
Solution.
(a) By
we have immediately
that
R the dilation Rproperties of the integral,
R
R
K
(x)dx
=
(x)dx
=
1
and
K
(x)dx
=

(x)dx
=
Rd
Rd
Rd
Rd
k k1 < 1. This proves the first two properties of good kernels. For
the last, we recall that for
2 L1 , for every > 0 there exists a
R
compact set F such that F c   < . Now compact subsets of Rd are
bounded, so K Br (0) for some radius Rr . Now if h > 0 is any fixed
number, for < rh this will imply that x>h K (x)dx < . Thus,
R
for any h > 0, x>h K (x)dx ! 0 as ! 0. Hence {K } is a family
of good kernels.
(b) Suppose   M everywhere and (x) = 0 for x
B. Let A =
M B d+1 . Then for any > 0,
x
1
K (x) = (d+1)
M (d+1)
A
A
( B)d+1
xd+1
for x B; for x > B we have K (x) = 0. Hence K (x) A /xd+1
for all x and , so {K } is an approximation to the identity.
(c) Suppose {K } is any family of good kernels. Then
Z
kf K
fk =
f K (x) f (x) dx
d
ZR Z
=
f (x y)K (y)dy f (x) dx
Rd
Rd
Z Z
=
(f (x y) f (x))K (y)dy dx
d
d
ZR Z R
f (x y) f (x)K (y)dydx
Rd Rd
Z Z
=
f (x y) f (x)K (y)dxdy
d
d
ZR RZ
Z
Z
=
f (x y) f (x)K (y)dxdy +
f (x y) f (x)K (y)dxdy
y
Rd
y>
Rd
for any > 0. Let us call the first integral I1 and the second I2 . Then
Z
Z
I1 =
K (y)
f (x y) f (x)dx dy
y
Rd
Z
=
K (y)kf (x y) f (x)k1 dy.
y
36
Now by the L1 continuity of translation, kf (x y) f (x)k1 ! 0 as
y ! 0. Thus, if is sufficiently small, this norm will be at most, say,
2A for all y 2 [ , ]. Then
Z
I1
K (y) dy
2A
2
y
R
since Rd K (y)dy A for all . Thus, by choosing sufficiently
small, we may make I1 as small as we like, independent of . On the
other hand,
Z
Z
I2 =
K (y)
f (x y) f (x)dxdy
y>
Rd
Z
Z
K (y)
(f (x y) + f (x))dxdy
y>
Rd
Z
=
K (y) 2kf k1 dy
y>
Z
= 2kf k1
K (y)dy ! 0
y>
as ! 0. Putting the two halves together, we see that choosing a
sufficiently small and then letting ! 0 makes kf K
f k1 ! 0.
Exercise 3: Suppose 0 is a point of (Lebesgue) density of the set E R.
Show that for each of the individual conditions below there is an infinite
sequence of points xn 2 E, with xn 6= 0, and xn ! 0 as n ! 1.
(a) The sequence also satisfies xn 2 E for all n.
(b) In addition, 2xn belongs to E for all n.
Generalize.
Solution.
(a) Since 0 is a point of density of E, there exists r0 > 0 such that m(E \
Br (0)) > 23 m(Br (0)) = 43 r for r r0 . By symmetry, m(( E) \
Br0 (0)) = m(E\Br0 (0)) 43 r0 . Now E and E both intersect Br0 (0),
which has measure 2r0 , in sets of measure at least 43 r0 . Therefore they
intersect each other in a set of measure at least 23 r0 . Since E \ ( E)
has positive measure, it is infinite, so it contains an infinite sequence
xn . This sequence satisfies xn 2 E and xn 2 E.
(b) Since 0 is a point of density of E, there exists r0 > 0 such that
m(E \ Br (0)) > 23 m(Br (0)) = 43 r for r r0 . Let E0 = E \ Br0 (0).
Then m( 12 E0 ) = 12 m(E0 ) 23 r0 as we showed in a previous homework
about the eect of dilation on Lebesgue measure. Now 12 E0 = ( 12 E) \
Br0 /2 (0) has measure at least 23 r0 , and we also know m(E\Br0 /2 (0))
r0
2
2
1
3 m(Br0 /2 ) = 3 r0 since 2 < r0 . So E and 2 E both intersect Br0 /2 ,
which has measure r0 , in sets of measure at least 23 r0 . Therefore they
intersect each other in a set of measure at least 13 r0 . Since E \ ( 12 E)
has positive measure, it must contain an infinite sequence xn . Then
xn 2 E and 2xn 2 E.
37
Clearly the above process generalizes to produce a sequence xn with xn 2 E
and cxn 2 E for any c 6= 0.
Exercise 5: Consider the function on R defined by
(
1
if x 1/2,
2
f (x) = x(log 1/x)
0
otherwise.
(a) Verify that f is integrable.
(b) Establish the inequality
c
for some c > 0 and all x 1/2,
f (x)
x(log 1/x)
to conclude that the maximal function f is not locally integrable.
Solution.
(a) We have
Z
Rd
f (x)dx = 2
1/2
1
1
dx = 2
x(log 1/x)2
log x1
1/2
=
0
2
.
log 2
1
2,
(b) For 0 < x
if B = (0, 2x) is the ball of radius x centered at x,
then
Z
Z 2
1
1
f (y)dy =
xf (y)dy
m(B) B
2x 0
Z 
1
1
x
dy
2x 0 y(log 1/y)2
1
1
=
.
2x log(1/x)
Since f (x) is the supremum of such integrals over all balls containing
1
x, it is at least equal to the integral over B, so f (x)
2x log(1/x) .
This function is not locally integrable, because if we integrate it in any
neighborhood around 0 we get
Z
1
dx =
2x log(1/x)
1
log log
x
= 1.
Exercise 6: In one dimension there is a version of the basic inequality (1) for
the maximal function in the form of an identity. We defined the onesided
maximal function
Z
1 x+h
f (y)dx.
f+
(x) = sup
h>0 h x
If E+ = {x 2 R : f+
(x) > }, then
Z
1
m(E+ ) =
f (y)dy.
E+
38
Solution. First, we note that
Z
1 x+h
x 2 E+ , 9h > 0 s.t.
f (y)dy >
h x
Z x+h
,
f (y)dy > h
x
x+h
,
,
x+h
f (y)dy
f (y)dy
(x + h) >
f (y)dy
(x + h) + x > 0
f (y)dy
x.
Rx
Thus, if we define F (x) = 0 f (y)dy x, the set E+ is precisely the
set {x : 9h > 0 s.t. F (x + h) > F (x)}. Note also that F is continuous
by the absolute continuity of the integral (we are assuming that f 2 L1 ,
naturally). By the Rising Sun Lemma,
E+ =
1
[
(aj , bj )
j=1
where the intervals (aj , bj ) are disjoint and
aj
F (aj ) = F (bj )
Z bj
f (y)dy aj =
f (y)dy
)
Then
Z
+
E
f (y)dy =
as desired.
bj
aj
1 Z
X
j=1
bj
f (y)dy = (bj
bj
aj
f (y)dy =
1
X
aj ).
(bj
aj ) = m(E+ )
j=1
Exercise 8: Suppose A is a Lebesgue measurable set in R with m(A) > 0.
Does there exist a sequence {sn } such that the complement of [1
n=1 (A+sn )
in R has measure zero?
Solution. Yes. Let x be any point of density of A, and let sn = qn x
where {qn } is an enumeration of the rationals. Let E = [(A + sn ). To
show that m(E c ) = 0, it is sufficient to show that m(E c \ [n, n + 1]) = 0
for all n 2 Z; since E is invariant under rational translations, it is sufficient
to show that E c \ [0, 1] has measure zero.
For m = 1, 2, . . . let Nm be an integer such that m(A \ B1/Nm (x)) (1
1
m )m(B1/Nm )(x). Such an Nm must exist because x is a point of density of
1
A. Then by the construction of E, m(A\B1/Nm (q)) (1 m
)m(B1/Nm )(q)
for any q 2 Q. Now the open balls
j
j
Um = B1/Nm
, j = 1, 2, . . . , 2Nm
2Nm
39
cover [0, 1], so
E c \ [0, 1]
and
m (E c \ [0, 1])
2N
m
X
j=1
2N
[m
j=1
j
m E c \ Um
Since m(E c \[0, 1])
4
m
j
E c \ Um
2N
m
X
j=1
1
1 2
4
j
m(Um
) = (2Nm )
= .
m
m Nm
m
for all m, m(E c \[0, 1]) = 0. Hence m(E c ) = 0.
Note: It is not sufficient to construct a set whose Lebesgue points are
dense in R. Consider an open dense set of measure (e.g. put an interval
of length 2k around the kth rational). Then every point is a Lebesgue point
since the set is open, yet its complement has positive measure.
Exercise 9: Let F be a closed subset in R, and (x) the distance function
from x to F , that is
(x) = inf{x
y : y 2 F }.
Clearly, (x + y) y whenever x 2 F . Prove the more refined estimate
(x + y) = o(y)
for a.e. x 2 F ,
that is, (x + y)/y ! 0 for a.e. x 2 F .
Solution. We note that is a function of bounded variation on any interval;
in fact, in general we have Vab ( ) b a because  (y)
(z) y z
for all y, z 2 R. Since has bounded variation, it is dierentiable almost
everywhere; in particular, it is dierentiable for a.e. x 2 F . But is a
nonnegative function that is 0 for all x 2 F , so it has a local minimum at
every point of F ; if it is dierentiable at x 2 F , its derivative is zero. By
the definition of derivative, this implies (x + y)/y ! 0.
Exercise 15: Suppose F is of bounded variation and continuous. Prove that
F = F1 F2 , where both F1 and F2 are monotonic and continuous.
Solution. Every boundedvariation function is a dierence of increasing
functions, so write F = G1 G2 where G1 and G2 are increasing. As
shown in Lemmas 3.1213, an increasing function is a continuous increasing
function plus a jump function. Hence G1 = F1 + J1 where F1 is continuous
and increasing, and J1 is a jump function; similarly, G2 = F2 + J2 . Then
F = (F1 F2 ) + (J1 J2 ). But J1 J2 is a jump function, and jump
functions are continuous only if theyre constant. Since F is continuous,
this implies that J1 J2 is constant; WLOG, J1 J2 = 0. (Otherwise we
could redefine F10 = F1 + (J1 J2 ) and F10 would also be continuous and
increasing.) Hence F = F1 F2 .
Exercise 17: Prove that if {K }>0 is a family of approximations to the
identity, then
sup(f K )(x) cf (x)
>0
for some constant c > 0 and all integrable f .
40
Exercise 18: Verify the agreement between the two definitions given for the
CantorLebesgue function in Exercise 2, Chapter 1 and Section 3.1 of this
chapter.
Solution. This is such a lame problem. Its so clear that theyre the same.
Probably the easiest way to see that is to think of the CantorLebesgue
function as the following process:
Given x, let y be the greatest member of the Cantor set such that
y x. (We know such a y exists because the Cantor set is closed.)
Write the ternary expansion of y.
Change all the 2s to 1s and reinterpret as a binary expansion. The
value obtained is F (x).
Its pretty clear that both the definitions of the CantorLebesgue function
given in the text do exactly this.
Exercise 19: Show that if f : R ! R is absolutely continuous, then
(a) f maps sets of measure zero to sets of measure zero.
(b) f maps measurable sets to measurable sets.
Solution.
(a) Suppose E R hasPmeasure zero. Let P> 0. By absolute continuity,
9 > 0 such that
bj aj  < )
f (bj ) f (aj ) < . Since
m(E) = 0, there is an open set U
E with m(U ) < . Every open
subset of R is a countable disjoint union of open intervals, so
U=
1
[
(aj , bj ) with
j=1
1
X
(bj
aj ) < .
j=1
For each j = 1, 2, . . . ,, let mj , Mj 2 [aj , bj ] be values of x with
f (mj ) = min f (x) and f (Mj ) = max f (x).
x2[a,b]
x2[a,b]
Such mj and Mj must exist because f is continuous and [aj , bj ] is
compact. Then
f (U )
But Mj
1
X
j=1
mj  bj
Mj
mj  <
1
[
[f (mj ), f (Mj )].
j=1
aj  so
)
1
X
j=1
f (Mj )
f (mj ) < .
Hence f (E) is a subset of a set of measure less that . This is true for
all , so f (E) has measure zero.
(b) Let E R be measurable. Then E = F [ N where F is F and
N has measure zero. Since closed subsets of R are compact, F is
compact. But then f (F ) is also compact since f is continuous.
Then f (E) = f (F ) [ f (N ) is a union of an F set and a set of measure
zero. Hence f (E) is measurable.
41
Exercise 20: This exercise deals with functions F that are absolutely continuous on [a, b] and are increasing. Let A = F (a) and B = F (b).
(a) There exists such an F that is in addition strictly increasing, but such
that F 0 (x) = 0 on a set of positive measure.
(b) The F in (a) can be chosen so that there is a measurable subset E
[A, B], m(E) = 0, so that F 1 (E) is not measurable.
(c) Prove, however, that for any increasing absolutely continuous F , and
E a measurable subset of [A, B], the set F 1 (E) \ {F 0 (x) > 0} is
measurable.
Solution.
(a) Let
F (x) =
C (x)dx
where C [a, b] is a Cantor set of positive measure and C (x) is
the distance from x to C. Note that C (x)
0 with equality i
x 2 C. Since C is continuous, this integral is welldefined, even in the
Riemann sense. Moreover, F is absolutely continuous by the absolute
continuity of integration of L1 functions. As shown in problem 9, F 0 (x)
exists and equals zero a.e. in C, hence on a set of positive measure.
However, F is strictly increasing: Suppose a x < y b. Since C
contains no interval, some point, and therefore some interval, between
x and y belongs to C C . The integral of C over this interval will be
positive, so F (y) > F (x).
(b) The same function from part (a) does the trick. Since F is increasing,
it maps disjoint open intervals to disjoint open intervals. Let U =
[a, b] \ C. Since U is open, we can write
1
[
U=
(aj , bj )
j=1
where the intervals (aj , bj ) are disjoint. Then
F (U ) =
1
[
(F (aj ), F (bj ))
j=1
and
m(F (U )) =
1
X
F ((bj )
F (aj )).
j=1
But
B
A = F (b)
F (a) =
(c)
(x)dx =
(x)dx =
1
X
(F (bj )
F (aj ))
j=1
R
since = 0 on C so C (x)dx = 0. Thus m(F (U )) = m(F ([a, b])), so
that m(F (C)) = 0. This implies that m(F (S)) = 0 for any subset S
C. But since C has positive measure, it has a nonmeasurable subset.
Then if E = F (S), m(E) = 0 so E is measurable, but F 1 (E) = S is
not measurable.
42
Exercise 22: Suppose that F and G are absolutely continuous on [a, b]. Show
that their product F G is also absolutely continuous. This has the following
consequences.
(a) Whenever F and G are absolutely continuous in [a, b],
Z b
Z b
0
F (x)G(x)dx =
F (x)G0 (x)dx + [F (x)G(x)]ba .
a
(b) Let F be absolutely continuous in [ , ] with F () = F ( ). Show
that if
Z
1
an =
F (x)e inx dx,
2
P
such that F (x) an einx , then
X
F 0 (x)
inan einx .
(c) What happens if F ( ) 6= F ()?
Proof. Since F and G are absolutely continuous, they are continuous and
therefore bounded on the compact interval [a, b]. Suppose F ,P
G M on
this interval.
Now
given
>
0,
we
can
choose
>
0
such
that
bj aj  <
P
P
) F (bj ) F (aj ) < M
and
G(bj ) G(aj ) < 2M
. Then
F (bj )G(bj )
F (aj )G(aj )
X1
=
(F (bj F (aj ))(G(bj ) + G(aj )) + (F (bj ) + F (aj ))(G(bj ) G(aj ))
2
X
1 X
F (bj ) F (aj )G(bj ) + G(aj ) +
F (bj ) + F (aj )G(bj ) G(aj )
2
X
1 X
(2M )F (bj ) F (aj ) +
(2M )G(bj ) G(aj )
2
1
2M
+ 2M
= .
2
2M
2M
This proves that F G is absolutely continuous on [a, b]. We now turn to the
consequences of this:
(a) Since F G is absolutely continuous, its dierentiable almost everywhere. By elementary calculus, (F G)0 = F 0 G + F G0 at any point
where all three derivatives exist, which
is almost everywhere. InteR
grating both sides and subtracting F G0 yields
Z b
Z b
Z b
F 0 (x)G(x)dx =
F (x)G0 (x)dx +
(F G)0 (x)dx.
a
Since F G is absolutely continuous, this implies
Z b
Z b
0
F (x)G(x)dx =
F (x)G0 (x)dx + [F (x)G(x)]ba .
a
(b) It would be nice if the problem would actually define this for us, but
Im assuming that the here means is represented by as opposed to
any kind of statement about whether the function actually converges
43
to its Fourier series or not. Then suppose bn are the Fourier coefficients
of F 0 , so by definition
Z
1
bn =
F 0 (x)e inx dx.
2
bn =
1
2
Using part (a), we have
F (x)( ine
inx
)dx+[F (x)e
inx b
]a
1
= in
2
F (x)e
inx
dx = inan .
(c) Then all bets are o. As one example, consider F (x) = x which is
clearly absolutely continuous on [ , ]. Then
Z
1
1 xe inx
e inx
2i
inx
an =
xe
dx =
+
= ( 1)n
2
2
2
in
n
n
R
for n 6= 0, and a0 =
xdx = 0. However, F 0 (x) = 1 which has
Fourier coefficients b0 = 1 and bn = 0 for n 6= 0.
Exercise 25: The following shows the necessity of allowing for general exceptional sets of measure zero in the dierentiation Theorems 1.4, 3.4, and
3.11. Let E be any set of measure zero in Rd . Show that:
(a) There exists a nonnegative integrable f in Rd , such that
Z
1
lim inf
f (y)dy = 1 for each x 2 E.
m(B)!0 m(B) B
x2B
(b) When d = 1 this may be restated as follows. There is an increasing
absolutely continuous function F such that
D+ F (x) = D F (x) = 1, for each x 2 E.
Solution.
(a) Since E has measure zero, thereP
exist open sets On with E On for
1
all n and m(On ) < 21n . Let f = n=1 On . Then f 2 L1 since
Z
1 Z
1
1
X
X
X
1
f=
=
m(O
)
= 1.
On
n
2n
d
Rd
n=1 R
n=1
n=1
Now let x 2 E. Since On is open, there exist open balls Bn On with
x 2 Bn . Then for any ball B 3 x,
Z
Z
1
1
1
X
X
X
1
m(Bn \ B)
f (y)dy =
m(On \B)
m(Bn \B) )
f (y)dy
.
m(B) B
m(B)
B
n=1
n=1
n=1
For any N , there exists > 0 such that m(B) < and x 2 B implies
B Bj for all j = 1, . . . , N . (This is true because B1 \ \ Bn is
an open set containing x and hence contains an open ball around x.)
Then
Z
1
X
1
m(Bn \ B)
f (y)dy
N
m(B) B
m(B)
n=1
44
for sufficiently small B. This proves that
Z
1
lim inf
f (y)dy = 1.
m(B)!0 m(B) B
x2B
(b) Let f be as in part (a), and
Z
F (x) =
f (y)dy.
Then F is absolutely continuous because it is the integral of an L1
function; it is increasing because f is nonnegative. Now
Z x+h
F (x + h) F (x)
D+ F (x) = lim inf
= lim inf
f (y)dy
h!0
h!0
h
x
h>0
h>0
and
D F (x) = lim inf
h!0
h<0
F (x + h)
h
F (x)
= lim inf
h!0
h<0
x+h
f (y)dy
The conclusion in (a) implies that both of these are infinite, since
one can consider integrals over the balls [x, x + h) and (x h, x].
(Technically, I suppose we should work with open balls, but one can
look at e.g. (x
, x + h) for sufficiently small .)
Exercise 30: A bounded function F is said to be of bounded variation on R if
F is of bounded variation on any finite subinterval [a, b] and supa,b TF (a, b) <
1. Prove
that such an F enjoys the following two properties:
R
(a) RR F (x + h) F (x)dx Ah, for some constant A and all h 2 R.
(b)  R F (x) 0 (x)dx A, where ranges over all C 1 functions of bounded
support with supx2R  (x) 1.
For the converse, and analogues in Rd , see Problem 6* below.
Solution.
(a) First, note that it is sufficient to treat the case where F is a bounded
increasing function. This is so because in general we can let F =
F1 F2 where F1 and F2 are bounded increasing functions; if they
both satisfy the given condition, with constants A1 and A2 , then
F (x + h)
R F (x) F1 (x + h) F1 (x) + F2 (x + h) F2 (x) for
all x, so F (x + h) F (x) A1 + A2 .
(In case someone asks why F must be a dierence of bounded increasing functions, we could redo the proof that was used on finite
intervals, using the positive and negative variations of f . This avoids
the problem of trying to extend from the bounded case and worrying
about whether such an extension is unique.)
Suppose now that F is bounded and increasing. Since F is increasing, F (x + h) F (x) = F (x + h) F (x) for h > 0. (By translation
invariance, it is sufficient
to treat theRcase of positive h, since
R
R
F (x h) F (x) = F (x) F (x h) = F (x + h) F (x).) Then
45
on any interval [a, a + h],
Z
F (x+h) F (x) F (a+2h) F (a) )
a+h
In particular,
Z (n+1)h
nh
F (x+h) F (x) h(F (a+2h) F (a)).
h(F ((n + 2)h)
F (nh))
for any n 2 Z. Then for any N ,
Nh
F (x + h)
Nh
F (x) h
N
X1
F ((n + 2)h)
F (nh)
n= N
h F ((N + 1)h) + F (N h)
2h(F (+1)
F ( 1))
since the sum telescopes. Since F (+1)
this proves the result.
(b) By some algebraic legwork,
F (x + h) (x + h)
F (x) (x) = F (x)
(x + h)
(x)
F (( N + 1)h)
F ( N h)
F ( 1) is a finite constant,
(x + h) F (x + h)
F (x) .
When we integrate both sides over R, the lefthand side integrates to
zero by translation invariance. Hence
Z
Z
F (x) (x + h)
(x) dx =
(x + h) F (x + h) F (x) dx.
R
By part (a), we have
F (x)
(x + h)
(x) dx =
Hence
(x + h) F (x + h)
F (x) dx
 (x + h) F (x + h)
F (x) dx
ZR
R
F (x + h)
F (x) dx
Ah.
Z
F (x)
(x + h)
h
(x)
dx A.
Now is supported on some compact set K, so 0 is a continuous
function which is 0 outside K. Hence it has a maximum M . By the
Mean Value Theorem, any dierence quotient of is at most M in
absolute value. Then if L is a bound for F on K, F (x) x+hh (x)
is dominated by M L K for all h. Hence we can use the Dominated
Convergence Theorem as h ! 0 to obtain
Z
F (x) 0 (x)dx A.
R
46
Exercise 31: Let F be the CantorLebesgue function described in Section
3.1. Consider the curve that is the graph of F , that is, the curve given by
x(t) = t and y(t) = F (t) with 0 t 1. Prove that the length L(
x) of the
segment 0 t x
of the curve is given by L(
x) = x
+ F (
x). Hence the
total length of the curve is 2.
Solution. It is true for any increasing function with F (0) = 0 that L(
x)
x
+ F (
x), because for any partition 0 = t0 < t1 < < tn = x
,
n q
X
(tj
tj
2
1)
+ (F (tj )
F (tj
2
1 ))
j=1
n
X
(tj tj
1 )+(F (tj )
F (tj
1 ))
=x
+F (
x).
j=1
We wish to show that this upper bound is in fact the least upper bound
when F is the CantorLebesgue function. Consider the iterates Fn (x) of
which this function is the limit. The interval [0, 1] can be divided into
2n+1 1 intervals on which Fn (x) alternately increases and stays constant;
suppose we label them I1 , C1 , I2 , C2 , . . . , C2n 1 , I2n . The intervals Cj have
varying lengths, since they correspond to intervals that are deleted from the
Cantor set at varying stages of the iteration; however, the Ij all have length
1
iteration of
3n since they correspond to the intervals remaining in the nth
2n
the Cantor set. Hence the sum of the lengths of the Ij is 3n , while the
2n
sum of the lengths of the Cj is 1
3n .
Now let x
2 [0, 1], and consider the partition Pn consisting of all points less
than or equal to x
which are an endpoint of one of the Cj or Ij . Thus we
have 0 = t0 < t1 < < tm = x
where Fn is increasing on [t0 , t1 ], constant
on [t1 , t2 ], increasing on [t2 , t3 ], etc. Note also that F (tj ) = Fn (tj ) since all
the tj are endpoints of the Ck intervals, which remain fixed in all successive
iterations. Then
m q
X
j=1
m
X
(tj
(tj
j=1
j odd
m
X
=F (
x) +
2
1)
+ (F (tj )
F (tj
2
1 ))
tj
2
1)
+ (F (tj )
F (tj
2
1 ))
m q
X
(tj
j=1
j even
(F (tj )
j=1
j odd
tj
X
k
=F (
x) + x
F (tj
Ck \ [0, x
]
X
X
k
=F (
x) + x
j=1
j even
F (
x) + x
1 ))
m
X
Ik \ [0, x
]
Ik 
n
2
.
3
(tj
tj
1)
tj
2
1)
+ (F (tj )
F (tj
2
1 ))
47
Letting n ! 1, this approaches x
+ F (
x), which proves L(
x) x
+ F (
x).
Since we already know L(
x) x
+ F (
x), we have L(
x) = x
+ F (
x) as
desired.
Exercise 32: Let f : R ! R. Prove that f satisfies the Lipschitz condition
f (x)
f (y) M x
y
for some M and all x, y 2 R, if and only if f satisfies the following two
properties:
(i) f is absolutely continuous.
(ii) f 0 (x) M for a.e. x.
Solution.
Suppose P
f is Lipschitz. Then for any > 0, if we let = M
, then
P
bj aj  < )
f (bj ) f (aj ) < . Hence f is absolutely continuous.
This implies f is dierentiable a.e.; if x is a point for which f 0 (x) exists, the
Lipschitz condition implies  f (x+h)h f (x)  M for all h. Taking the limit
as h ! 0 implies f 0 (x) M .
Conversely, suppose f is absolutely continuous and has bounded derivative a.e. Since absolutely continuous functions are the integrals of their
derivatives,
Z y
Z max(x,y)
Z max(x,y)
0
0
f (y) f (x) =
f (t)dt
f (t)dt
M dt = M x y
x
min(x,y)
min(x,y)
so f is Lipschitz.
Chapter 3.6, Page 152
Problem 4: A realvalued function defined on an interval (a, b) is convex
if the region lying above its graph {(x, y) 2 R2 : y > (x), a < x < b} is a
convex set. Equivalently, is convex if
(x1 + (1
)x2 ) (x1 ) + (1
) (x2 )
for every x1 , x2 2 (a, b) and 0 1. One can also observe as a consequence that we have the following inequality of the slopes:
(x + h)
h
(1)
(x)
(y)
y
(x)
x
(y)
(y
h
h)
whenever x < y, h > 0, and x + h < y. The following can then be proved.
(a) is continuous on (a, b).
(b) satisfies a Lipschitz condition of order 1 in any proper closed subinterval [a0 , b0 ] of (a, b). Hence is absolutely continuous in each subinterval.
(c) 0 exists at all but an at most denumerable number of points, and
0
= D+ is an increasing function with
Z y
0
(y)
(x) =
(t)dt.
x
(d) RConversely, if
is any increasing function on (a, b), then
x
(t)dt
is
a
convex
function in (a, b) for c 2 (a, b).
c
Solution.
(x) =
48
(a) Suppose to the contrary that is discontinuous at some x 2 (a, b).
This means there exists > 0 and an infinite sequence xn ! x with
xn 2 (a, b) and  (xn )
(x) > for all n. Since the sequence xn
is infinite, it must have infinitely many points in one of the following
categories:
(i) (xn ) > (x) +
(ii) (xn ) < (x)
We will treat each case separately and obtain a contradiction.
(i) Assume WLOG that the entire sequence xn is in this category
(otherwise, take a subsequence that is). We may also assume xn
converges to x monotonically, since otherwise we can again take
a subsequence. Let L() = (x) + (1 ) (x1 ) for 0 1.
This is a continuous function of , so 9 > 0 such that  < )
L() < (x) + . Now let n = xxn xx11 . Then 0 n 1 since
xn ! x monotonically. Note also that xn = n x + (1 n )x1 .
Now n ! 0 as n ! 1, so n < for n sufficiently large. But
this implies
(xn ) = (n x + (1
n )x1 ) L(n ) < (x) + ,
for sufficiently large n, a contradiction.
(ii) Again, we assume WLOG that xn ! x monotonically. Let y 2
(a, b) such that x is between y and xn for all n. Then n = xxnn xy
has the properties that 0 n 1 and n ! 1 as n ! 1. Now
x = n xn + (1 n )y, so
(x) n (xn ) + (1
n ) (y)
for all n. But n ! 1, and since (xn ) < (x) , this implies
(x) < (x) for sufficiently large n, a contradiction.
Hence is continuous.
(b) First, I prove an inequality of slopes that I like better than the one
given. I claim that for s < t < u with s, t, u 2 (a, b),
(2)
(3)
(t)
(s)
(u)
(s)
(u)
(t)
.
t s
u s
u t
This follows straightforwardly from the convexity condition:
u t
t s
t=
s+
u
u s
u s
u t
t s
) (t)
(s) +
(u)
u s
u s
) (u s) (t) (u t) (s) + (t s) (u)
) (u
s) (t)
(u
s) (s) (s
t) (s) + (t
s) (u)
(t)
(s)
(u)
(s)
.
t s
u s
Taking a dierent route from inequality (3) leads to
)
(u
t) (u)
)
(u
(u)
u
t) (s) (u
(s)
(u)
u
s) (u)
(t)
t
(u
s) (t)
49
as desired. Given this inequality of slopes, we can easily prove that
is Lipschitz on [a0 , b0 ]. Choose h > 0 such that [a0 h, b0 + h] (a, b).
Then for x, y 2 [a0 , b0 ], suppose WLOG that x < y; since a0 h < a0
x < y b0 < b0 + h, the slope inequality yields
(a0
h)
h
(a0 )
(y)
y
(a0 h)
(y)
a0 + h
y
(x)
x
(b0 + h)
b0 + h
(x)
x
(b0 + h)
h
The leftmost and rightmost terms above are constants, which we may
call m and M ; we thus have my x  (y)
(x) M y x,
whence is Lipschitz on [a0 , b0 ].
(c) Since is Lipschitz on any closed subinterval [x, y] (a, b), is absolutely
(x) =
R y 0 continuous on [x, y] by Exercise 32 above; hence (y)
(t)dt.
x
Now inequality (2) implies that (x+h)h (x) is an increasing function
of h at any x 2 (a, b). This implies that D+ = D+ and D = D ,
that D+ , D  < 1, and that D+
D . The inequality (1) tells
us that x < y ) D+ (x) D (y). This in turn implies that D+
and D are increasing. To show that D+ = D except at countably
many points, let {x } be those points in (a, b) for which this is not
true, and define j > 0 by j = D+ ( )(x ) D ( )(x ). Then on
any subinterval [a0 , b0 ], if {x1 , . . . , xn } [a0 , b0 ], we have
n+1
X
D+ ( )(xk )
k=1
n+1
X
D+ ( )(xk )
k=1
=D+ ( )(b0 )
D ( )(xk )
n+1
X
D ( )(xk ) +
D ( )(xk )
k=1
D ( )(a0 ),
D+ ( )(xk
1)
where we use the convention x0 = a0 and xn+1 = b0 . This implies that
X
x2[a0 ,b0 ]
D+ ( )(xk )
D ( )(xk )
is finite, because all finite subsums are bounded by the finite constant
D+ ( )(b0 ) D ( )(a0 ). So this sum can containly only countably many
nonzero terms, which means only countably many points in [a0 , b0 ] can
have D ( )(x) 6= D+ ( )(x). Since (a, b) is a countable union of closed
subintervals (e.g. \[a + n1 , b n1 ]), it can contain only countably many
points for which D+ 6= D . Everywhere else, the derivative exists.
(b0 )
50
(d) Because
x1 + (1
)x2 =
=
is increasing,
Z
x1 +(1 )x2
Zc x2
(t)dt
x2
(t)dt
Z
(t)dt
x2
(t)dt
x2
(t)dt
x2
(t)dt
(t)dt
x1 +(1 )x2
Z x2
x1 +(1 )x2
Z x2
x2
(1
x1
x1
(t)dt + (1
x2
(t)dt
x1 +(1 )x2
(t)dt
(1
(t)dt
x1 +(1 )x2
Z x2
)(x2
x1 ) (x1 + (1
x1 +(1 )x2
(t)dt
x1
(t)dt
Z x2
)
(t)dt
c
= (x1 ) + (1
So
(t)dt
x1 +(1 )x2
x2
) (x2 ).
is convex.
Chapter 4.7, Page 193
Exercise 4: Prove from the definition that `2 (Z) is complete and separable.
Solution. The proof that `2 is complete is exactly the same as the proof
(m)
that L2 is complete, from pp. 159160 of the textbook. Let {aj }1
m=1 be
2
a Cauchy sequence in ` (Z). For each k
1 we can choose nk such that
m, n
nk ) ka(m) a(n) k < 21k and nk < nk+1 . Then the subsequence
a(nk ) has the property that ka(nk+1 ) a(nk ) k 21k . Define sequences a =
{aj } and b = {bj } by
(n1 )
aj = aj
1
X
(nk+1 )
aj
(nk )
aj
k=1
and
(n1 )
bj = aj
+
and the partial sums
(a,K)
Sj
(n1 )
= aj
1
X
(nk+1 )
aj
(nk )
aj
k=1
K
X
aj
K
X
aj
(nk+1 )
aj
(nk+1 )
aj
(nk )
k=1
and
(b,K)
Sj
(n1 )
= aj
+
k=1
(nk )
)x2 )
51
Then
kS (b,K) k ka(n1 ) k +
K
X
1
2k
k=1
by the triangle inequality; letting K ! 1, kbk converges by the monotone
convergence theorem (for sums, but hey, sums are just integrals with discrete measures), so kak converges since it converges absolutely. Of course,
this implies that bj and hence aj converges for each j; since the partial
(n
)
sums are S (a,K) = a(nk+1 ) by construction, aj k+1 ! aj for all j. Now
given > 0, choose N such that ka(n) a(m) k < 2 for n, m > N , and let
nK > N such that ka(nK ) ak < 2 . Then
m > N ) ka(m)
ak ka(m)
a(nK ) k + ka(nK )
ak < .
Hence a
! a.
To prove that `2 is separable, consider the subset D consisting of all rational
sequences which are 0 except at finitely many values. This is countable
because
(m)
D = {rational sequences of finite length}
=
=
1
[
N =1
1
[
{sequences {an } with an 2 Q and an = 0 for n > N }
Q2N +1
N =1
is a countable union of countable sets. Now let {bn } 2 `2 be any square
summable sequence. Given > 0, there exists N such that
X
2
bn 2 <
2
n>N
since the infinite sum converges. Then for each j = N, . . . , N , we can
2
choose a rational number qj with qj bj 2 < 22+N
+j . If we also define
qj = 0 for j > N , then q 2 D and
kq
bk2 =
1
X
j= 1
qj
bj 2 =
j>N
bj
This shows that D is dense.
02 +
N
X
j= N
2N
qj
bj 2 <
2 X 2
+
< 2 .
2+s
2
2
s=0
Exercise 5: Establish the following relations between L2 (Rd ) and L1 (Rd ):
(a) Neither the inclusion L2 (Rd ) L1 (Rd ) nor the inclusion L1 (Rd )
L2 (Rd ) is valid.
(b) Note, however, that if f is supported on a set E of finite measure and
if f 2 L2 (Rd ), applying the CauchySchwarz inequality to f E gives
f 2 L1 (Rd ), and
kf k1 m(E)1/2 kf k2 .
(c) If f is bounded (f (x) M ), and f 2 L1 (Rd ), then f 2 L2 (Rd ) with
1/2
kf k2 M 1/2 kf k1 .
52
Solution.
(a) Let f (x) = x 1 x1d/2 and g(x) = x1 x1 d . Then Exercise 10 of
Chapter 2 shows that f and g 2 are integrable, but f 2 and g are not.
(b) Applying the CauchySchwarz inequality to the inner product of f E
and E ,
Z
1/2 Z
1/2
Z
kf k1 = kf E k1 = f E 
f 
= m(E)1/2 kf k2 .
E
(c) Since f  M ,
Z
Z
1/2
f 2 M f  ) kf k22 = f 2 M f  = M kf k1 ) kf k2 M 1/2 kf k1 .
Exercise 6: Prove that the following are dense subspaces of L2 (Rd ):
(a) The simple functions.
(b) The continuous functions of compact support.
Solution.
(a) It is sufficient to treat the case of nonnegative f , since every complex
L2 function is a linear combination of nonnegative L2 functions. We
know there exists a sequence of simple functions sn % f with 0 sn
fR . Then f sn p f p so by the Dominated Convergence Theorem,
f sn p ! 0. Hence sn ! f in Lp . Therefore the simple functions
are dense.
(b) Let s 2 Lp (Rd ) be a simple function. It is sufficient to find g 2
CC (Rd ) with kg skp < . If s = 0, s 2 CC (Rd ) and were done.
Otherwise, since s is simple, 0 < ksk1 < 1; since it is in Lp , it
must be supported on a set E of finite measure. Now Lusins
theorem
d
and
enables us to construct g 2 CC (R ) with m({g 6= s}) < 2ksk1
sup g sup s < 1. (To do this, construct g from Lusins theorem;
then if g
ksk1 on the closed set F , change g to ksk1 on F . See
Rudin, Real and Complex Analysis pp. 5556.)
Then g s 2ksk1
and is nonzero on a set of measure
Z
g
s 2 ksk1
p
2ksk1
2ksk1
= p ) kg
, so
skp < .
2
d
Exercise 7: Suppose { k }1
k=1 is an orthonormal basis for L (R ). Prove that
the collection { k,j }1k,j<1 with k,j (x, y) = k (x) j (y) is an orthonormal
basis of L2 (Rd Rd ).
R2d
Solution. First, note that k,j is indeed in L2 (Rd Rd , since
Z
Z
Z
2
2
2
2
2
 (k, j) =
 k (x)  j (y) dxdy =
 k (x) dx
 j (y) dy = 1
R2d
Rd
Rd
53
by Fubinis Theorem. Also,
Z
h j,k `,m i =
j,k `,m
2d
ZR
=
j (x) k (y) ` (x) m (y)dxdy
R2d
Z
Z
=
j (x) ` (x)dx
k (y)
=
Rd
` m
j k
Rd
m (y)dy
so { j,k } is an orthonormal set. To show that linear combinations of { j,k }
are dense in L2 (Rd ), one approach would be to blow this problem out
of the water with the StoneWeierstrass theorem. We know CC (Rd ) is
dense in L2 (Rd ), and the StoneWeierstrass theorem tells us that linear
combinations
of separable continuous functions (i.e. functions of the form
Pm
f
(x)g
(y))
are dense in CC (Rd ). It is easy to verify that functions of
i
i=1 i
this form can be approximated by { j,k }, so were done.
Alternatively, we can follow the approach given in the hint. Let f 2 L2 (R2d )
and suppose that hf, j,k i = 0 for all j and k. By Fubinis Theorem,
Z
0=
f (x, y) j,k (x, y)
R2d
Z Z
=
f (x, y) j,k (x, y)dy dx
Rd
Rd
Z Z
=
f (x, y) k (y)dy
j (x)dx
Rd
Rd
Hence, if we define
fk (x) =
we see that
Rd
Rd
f (x, y)
fk (x)
j (x)
k (y)dy,
=0
for all j. Because { j } is an orthonormal basis, this implies that fk (x) = 0
for all k. Because { k } is an orthonormal basis, this in turn implies that
f (x, y) = 0. Since f ? j,k ) f = 0, { j,k } is an orthonormal basis.
Exercise 9: Let H1 = L2 ([ , ]) be the Hilbert space of functions F (ei )
R
1
on the unit circle with inner product (F, G) = 2
F (ei )G(ei )d. Let
2
H2 be the space L (R). Using the mapping
i x
i+x
of R to the unit circle, show that:
(a) The correspondence U : F ! f , with
1
i x
f (x) = 1/2
F
i+x
(i + x)
x 7!
gives a unitary mapping of H1 to H2 .
54
(b) As a result,
1
1/2
i x
i+x
1
i+x
is an orthonormal basis of L (R).
2
1
n= 1
Solution.
(a) If we define = 2 tan 1 (x), then x = tan 2 , ii+xx = ei , 1 + x2 =
sec2 2 , and dx = 12 sec2 2 d. (Brings back memories of high school
calculus, dont it?) Then
2
Z
Z
1
i x
2
f (x) dx =
F
dx
2
i+x
R
R i + x
2
Z
1 1
i x
=
F
dx
2
i+x
R x +1
Z
1
1
i 2 1
2
=
F
(e
)
sec
d
2
2
2
sec (/2)
Z
1
=
F (ei )2 d
2
so kf kH2 = kF kH1 . So U is unitary.
(b) By the RieszFischer theorem, {ein } is an orthonormal basis for L2 (T ).
Because U is unitary,
n
1
i x
1
in
{U (e )} = p
i+x
i+x
is an orthonormal basis for L2 (R).
Exercise 10: Let S denote a subspace of a Hilbert space H. Prove that
(S ? )? is the smallest closed subspace of H that contains S.
Solution. Let
S =
V.
V H subspace
V closed
Then S is a closed subspace, because the intersection of closed sets is closed
and the intersection of subspaces is a subspace. It is obviously the smallest
closed subspace containing S. We want to show that S = (S ? )? . Clearly
S (S ? )? since the latter is a closed subspace containing S. To establish
? = S ? . Clearly (S)
? S?
the reverse inclusion, we first show that (S)
?
To show the opposite inclusion, let x 2 S . Then for any
since S S.
there is a sequence wn S with wn ! w. Because inner products
w 2 S,
This proves
are continuous, 0 = hwn , xi ! hw, xi so x ? w for any w 2 S.
? . Since theyre equal, we can use Proposition 4.2 to write
S ? (S)
H = S S ? .
Now let x 2 (S ? )? . Then we can write x = v +w where v 2 S and w 2 S ? .
Then
hx, wi = hv, wi + hw, wi = hw, wi.
55
But hx, wi = 0 because w 2 S ? and x 2 (S ? )? . Hence w = 0 and
x 2 S.
Exercise 11: Let P be the orthogonal projection associated with a closed
subspace S in a Hilbert space H, that is,
P (f ) = f if f 2 S and P (f ) = 0 if f 2 S ? .
(a) Show that P 2 = P and P = P .
(b) Conversely, if P is any bounded operator satisfying P 2 = P and P =
P , prove that P is the orthogonal projection for some closed subspace
of H.
(c) Using P , prove that if S is a closed subspace of a separable Hilbert
space, then S is also a separable Hilbert space.
Solution.
(a) Let x 2 H and write x = xS + xS ? . Then
P 2 (x) = P (P (x)) = P (xS ) = xS = P (x)
so P 2 = P . Moreover, if y = yS + yS ? is any other vector in H, then
hP x, yi = hxS , yS + yS ? i = hxS , yS i = hxS + xS ? , yS i = hx, P yi
so P = P .
(b) Let S = im(P ) which is a subspace of H. To show S is closed, suppose
xn 2 S and xn ! x. Then because P is bounded, its continuous,
so P xn ! P x. But P xn = xn , so xn ! P x which implies P x = x.
Hence x 2 S, so S is closed. Also, if w 2 S ? , then for all v 2 S,
hx, P wi = hP x, wi = hx, wi = 0
so P w 2 S ? . But P w 2 S, so P w = 0. Now using Proposition 4.2, if
y is any vector in H, then
y = y S + yS ? .
Then by linearity, P (y) = P (yS ) + P (yS ? ) = yS + 0 = yS . So P does
the same thing to y as orthogonal projection onto S, for any y 2 H.
(c) Let H be a separable Hilbert space, { n } a countable dense set, and S
a closed subspace. I claim that {PS n } is dense in S. For any x 2 S,
we can find a sequence k ! x. Then PS ( k x) = PS k x since
P x = x. Since projections do not increase length,
kPS
xk k
xk ) (PS
k)
! x.
Exercise 12: Let E be a measurable subset of R , and suppose S is the
subspace of L2 (Rd ) of functions that vanish for a.e. x 2
/ E. Show that the
orthogonal projection P on S is given by P (f ) = E f , where E is the
characteristic function of E.
d
Solution. Define a linear operator T : L2 (Rd ) ! L2 (Rd ) by T (f ) =
Then T 2 (f ) = 2E f = E f = T (f ). Moreover,
Z
Z
hT f, gi =
=
f E g = hf, T gi
Efg
Rd
Rd
Ef.
56
so T = T . We also note that T is bounded since T f (x) f (x) for all
x, so kT f k kf k. By problem 11c, T is a projection onto its image. But
im(T ) is precisely those functions which are 0 a.e. on E c . Hence T is the
desired projection.
Exercise 13: Suppose P1 and P2 are a pair of orthogonal projections on S1
and S2 , respectively. Then P1 P2 is an orthogonal projection if and only if
P1 and P2 commute, that is, P1 P2 = P2 P1 . In this case, P1 P2 projects onto
S1 \ S2 .
Solution. Suppose P1 P2 is an orthogonal projection. Then
P2 P1 = P2 P1 = (P1 P2 ) = P1 P2
so they commute. On the other hand, suppose they commute; then
(P1 P2 )2 = P12 P22 = P1 P2
and
(P1 P2 ) = P2 P1 = P2 P1 = P1 P2
so P1 P2 is an orthogonal projection. The image of P1 P2 is a subspace of
S1 because P1 P2 v = P1 (P2 v) 2 S1 for any v; similarly, its a subspace of S2
because P1 P2 v = P2 (P1 v) 2 S2 . Hence the image of P1 P2 is a subspace of
S1 \ S2 . But every vector in S1 \ S2 is fixed by both P1 and P2 , and hence
by P1 P2 . Thus, the image of P1 P2 is precisely S1 \ S2 .
Exercise 14: Suppose H and H0 are two completions of a preHilbert space
H0 . Show that there is a unitary mapping from H to H0 that is the identity
on H0 .
Solution. Define U : H ! H0 as follows: Given x 2 H, choose a sequence
fn ! x with fn 2 H0 . Define U (x) = lim fn in H0 . (This limit exists
because fn is Cauchy.) To show this is welldefined, suppose gn 2 H0 is
another sequence converging to x. Then fn gn ! 0 in both spaces (because
fn gn 2 H0 for all n), so they have the same limit in H0 as well. This
shows that U is welldefined. Clearly it is the identity on H0 . To show that
its unitary, we need only use the continuity of the norm: Suppose fn ! x
in H and fn ! U (x) in H0 . Then kxk = lim kfn k = kU (x)k. Hence U is
unitary.
Exercise 15: Let T be any linear transformation from H1 to H2 . If we
suppose that H1 is finitedimensional, then T is automatically bounded.
Solution. Let {e1 , . . . , en } be an orthonormal basis for H1 . (This is a basis
in the usual linear algebra sense, i.e. every vector is a finite linear combination of the basis vectors.)
(ei )k. Then if x 2 H1 is a unit
PnLet mi = kTP
n
vector, we can write x = i=1 ci ei with i=1 ci 2 = 1. By the Triangle
Inequality,
!
n
n
n
X
X
X
ci ei
=
ci T (ei )
ci kT (ei )k nM
kT (x)k = T
i=1
i=1
i=1
p
where MP= max(m
p i ). (In fact, this bound can be improved to M n,
because
ci  n by the CauchySchwarz inequality.)
57
Exercise 18: Let H denote a Hilbert space, and L(H) the vector space of all
bounded linear operators on H. Given T 2 L(H), we define the operator
norm
kT k = inf{B : kT vk Bkvk, for all v 2 H}.
(a) Show that kT1 + T2 k kT1 k + kT2 k whenever T1 , T2 2 L(H).
(b) Prove that
d(T1 , T2 ) = kT1
T2 k
defines a metric on L(H).
(c) Show that L(H) is complete in the metric d.
Solution.
(a) This is easier if we use another expression for kT k, such as
kT k =
kT xk
.
xinH,x6=0 kxk
sup
If we call this supremum S, then we have kT xk Skxk for all x;
moreover, for any < S, there exists x with kT xk/kxk > ) kT xk >
kxk. Hence S is the infimum (in fact, the minimum) of such bounds,
so this definition really is equivalent. Then
kT1 + T2 k = sup
kT1 xk kT2 xk
kT1 xk
kT2 xk
+
sup
+ sup
= kT1 k + kT2 k.
kxk
kxk
kxk
kxk
(b) Since kT xk = k
T xk for all x, kT k = k
d(T1 , T2 ) = kT1
T2 k = kT2
T k. Then
T1 k = d(T2 , T1 ).
Clearly d(T1 , T1 ) = 0; conversely, if T1 6= T2 then there exists some x
with k(T1 T2 )xk > 0, so kT1 T2 k > 0. Finally, using part (a),
d(T1 , T3 ) = kT1 T3 k = k(T1 T2 )+(T2 T3 )k kT1 T2 k+kT2 T3 k = d(T1 , T2 )+d(T2 , T3 ).
Hence d is a metric.
(c) Let {Tn } be a Cauchy sequence in L(H). Define T (x) = lim Tn (x) for
all x 2 H. This limit exists because k(Tm Tn )xk kTm Tn kkxk
so {Tn x} is a Cauchy sequence in H. T is linear by the linearity of
limits. Finally, since Tn x ! T x, the continuity of the norm implies
kTn xk ! kT xk for all x. Hence kT k = lim kTn k which is finite because
(kTm k kTn k) kTm Tn k by the triangle inequality, so kTn k is a
Cauchy sequence of real numbers. So T is bounded.
Exercise 19: If T is a bounded linear operator on a Hilbert space, prove that
kT T k = kT T k = kT k2 = kT k2 .
58
Solution. We already know kT k2 = kT k2 from Proposition 5.4. Now
kT T k =
=
sup
kf k=kgk=1
sup
kf k=kgk=1
sup
kf k=kgk=1
hT T f, gi
hT f, T gi
kT f kkT gk
= sup kT f k sup kT gk
kf k=1
= kT k .
2
kgk=1
To show that equality is achieved, choose a sequence fn with kfn k = 1 and
kT fn k ! kT k. Then hT fn , T fn i ! kT k2 , so
kT T k =
sup
kf k=kgk=1
hT f, T gi
sup hT f, T f i
kf k=1
kT k2 .
Hence kT T k = kT 2 k. Finally, replacing T with T yields kT T k =
kT k2 = kT k2 and we are done.
Exercise 20: Suppose H is an infinitedimensional Hilbert space. We have
seen an example of a sequence {fn } in H with kfn k = 1 for all n, but
for which no subsequence of {fn } converges in H. However, show that for
any sequence {fn } in H with kfn k = 1 for all n, there exist f 2 H and a
subsequence {fnk } such that for all g 2 H, one has
lim (fnk , g) = (f, g).
n!1
One says that {fnk } converges weakly to f .
Solution. The proof is similar to the ArzelaAscoli theorem, and as with
that proof, the main hangup is notation. Let {en } be an orthonormal
basis for H. (Were assuming here that H is separable, of course, which the
book includes in its definition of Hilbert space.) Then we can define f by
defining its inner product with each en . Now hfn , e1 i is a sequence of real
numbers in [ 1, 1] (by the CauchySchwarz inequality, since these are both
unit vectors). Since [ 1, 1] is compact, there is a subsequence fnj which
converges to some limit `1 . Then hfnj , e2 i is a sequence of real numbers in
[ 1, 1], so some subsequence fnjr converges to a limit `2 . Continuing in this
fashion, we obtain sequences S0 , S1 , S2 , . . . where Sm+1 is a subsequence
n
(m)
(m)
of Sm , S0 = {fn }, and hvn , ej i ! `j for j = 1, . . . , m, where vn is the
(k)
nth term of Sm . Define a sequence S whose kth term is vk . Then for any
k
Sm , the tail of S is a subsequence of the tail of Sm . Hence hvk , ej i ! `j for
all j. Define f by hf, ej i = `j . Relabeling, vk = fnk is a subsequence with
hfnk , ej i ! hf, ej i for all basis elements ej , and thus hfnk , gi ! hf, gi for
all g 2 H.
Exercise 17: Fatous theorem can be generalized by allowing a point to approach the boundary in larger regions, as follows.
For each 0 < s < 1 and point z on the unit circle, consider the region
s (z) defined as the smallest closed convex set that contains z and the
closed disc Ds (0). In other words, s (z) consists of all lines joining z with
59
pointsn in Ds (0). Near the point z, the region s (z) looks like a triangle.
See Figure 2.
We say that a function F defined in the open unit disc has a nontangential limit at a point z on the circle, if for every 0 < s < 1, the
limit
F (w)
w!z
w2 s (z)
exists.
Prove that if F is holomorphic and bounded on the open unit disc, then
F has a nontangential limit for almost every point on the unit circle.
P1
n
Solution. Since F is holomorphic, we have F (z) = n=0 aP
n z for z < 1.
As shown on page 174 in the proof of Fatous theorem,
an 2 < 1 so
2
i
there is an L (T ) function F (e ) whose Fourier coefficients are an . Note
also that F (ei ) is bounded (almost everywhere) since, by Fatous theorem,
it is the a.e. radial limit of F (z), so F (z) M ) F (ei ) M .
We next prove a lemma about the Poisson kernel.
Lemma 2. For each s 2 (0, 1) there exists a constant ks such that
Pr (
for all (r, ) such that re
) ks Pr (
s.
Proof. By elementary arithmetic,
Pr (
)=
and
1
ei
r2
rei 2
1 r2
.
ei
r2
(This alternate formula can be found in any complex analysis book.) Our
task is thus reduced to proving
Pr (
ei
By the triangle inequality,
ei
r ei
)=
r ks ei
rei  + rei
rei .
1 = ei
Thus, our task is reduced to proving that
is bounded on s . But ei
is sufficient to prove that
2r sin 2
ei
rei 
rei 
rei  + 2r sin

2
r by the triangle inequality, so it
2r sin 2
1 r
is bounded on s . Now for each r, the maximum value of  (which will
maximize this quotient) is, as indicated in the diagram
below,
p
p one for which
2r sin 2 occurs in a triangle with 1 r and 1 s2
r2 s2 . Thus,
60
it is at most equal to the sum of them (actually quite a bit less). So we
finally need only to prove that
p
p
1 s2
r2 s2
1 r
is bounded as r ! 1. But this follows from the fact that it has negative
derivative, since
p
d p
r
d
1 s2
r2 s2 = p
< 1=
(1 r).
2
2
dr
dr
r
s
This completes the proof of the lemma. (I know, there are probably much
shorter proofs, but its late at night...)
61
Having established this lemma, the rest of the problem becomes trivial.
By the Poisson integral formula,
Z
1
F (rei ) =
Pr (
)F (ei d
2
Z
1
Pr (
)F (ei d
2
Z
1
ks
Pr ( )F (ei d
2
and by Math 245A (specifically, the fact that Pr is an approximate identity)
this last integral tends to zero. (Recall that we assumed F (1) = 0.)
Exercise 21: There are several senses in which a sequence of bounded operators {Tn } can converge to a bounded operator T (in a Hilbert space H).
First, there is convergence in the norm, that is, kTn T k ! 0 as n ! 1.
Next, there is a weaker convergence, which happens to be called strong
convergence, that requires that Tn f ! T f as n ! 1 for every f 2 H.
Finally, there is weak convergence (see also Exercise 20) that requires
(Tn f, g) ! (T f, g) for every pair of vectors f, g 2 H.
(a) Show by examples that weak convergence does not imply strong convergence, nor does strong convergence imply convergence in the norm.
(b) Show that for any bounded operator T there is a sequence {Tn } of
bounded operators of finite rank so that Tn ! T strongly as n ! 1.
Solution. (a) Let H = `2 (N). Let T be the zero operator and Tn = Rn
where R is the right shift operator; thus
Tn (a1 , a2 , . . . ) = (0, . . . , 0, a1 , a2 , . . . ).
Then for any fixed f = (a1 , a2 , . . . ) and g = (b1 , b2 , . . . ),
hTn f, gi =
1
X
k=1
ak bn+k = hf, Ln gi
where L is the leftshift operator defined by Lg = (b2 , b3 , . . . ). By the
CauchySchwarz inequality,
hTn f, gi kf kkLn gk.
P
But kLn gk ! 0 because bk 2 converges and the tails of a convergent
series tend to zero. Hence hTn f, gi ! 0 for all f, g 2 H. Thus,
Tn ! T weakly. However, Tn does not converge to T strongly; note
that kRf k = kf k for any f , so kTn f k = kRn f k = kf k by induction.
Hence Tn f 6! T f = 0.
To show that strong convergence does not imply convergence in norm,
let Tn = Ln in the same space. Then for any f , kTn f k = kLn f k ! 0
because the tails of a convergent series tend to zero. Hence Tn ! T
strongly. However, kTn k = 1 because a unit vector (0, . . . , 0, 1, 0, . . . )
with more than n initial zeros is mapped to another unit vector. Hence
kTn T k = kTn k 6! 0.
62
(b) (We assume H is separable.) Let {ei } be an orthonormal basis for H.
We can write
1
X
T ei =
cij ej
j=1
for each i = 1, 2, . . . . Define
Tn ei =
n
X
cij ej
j=1
and extend linearly from the basis to the rest of the space (actually,
extend linearly to finite linear combinations of the basis, and then take
limits to get the rest of the space...) Clearly each TnP
is of finite rank
1
since its range is spanned by e1 , . . . , en . Now let f = i=1 ai ei . Then
Tf =
1 X
1
X
ai cij ej
i=1 j=1
whereas
Tn f =
1 X
n
X
ai cij ej
i=1 j=1
which is just the nth partial sum (in j) and hence converges to T f .
(This is where we use the fact that T is a bounded operator, since
absolute convergence allows us to rearrange these sums.) Hence Tn f !
T f weakly for all f 2 H.
Exercise 22: An operator T is an isometry if kT f k = kf k for all f 2 H.
(a) Show that if T is an isometry, then (T f, T g) = (f, g) for every f, g 2 H.
Prove as a result that T T = I.
(b) If T is an isometry and T is surjective, then T is unitary and T T = I.
(c) Give an example of an isometry that is not unitary.
(d) Show that if T T is unitary then T is an isometry.
Solution. (a) By the polarization identity,
T gk2 + ikT f + iT gk2 ikT f iT gk2
4
kT (f + g)k2 kT (f g)k2 + ikT (f + ig)k2 ikT (f ig)k2
=
4
kf + gk2 kf gk2 + ikf + igk2 ikf igk2
=
4
= hf, gi.
hT f, T gi =
kT f + T gk2
kT f
This in turn implies
hf, T T gi = hf, Igi
for all f, g, so that T T = I.
(b) T preserves norms because its an isometry; its injective because normpreserving linear maps are always injective (since the kernel cannot
contain anything nonzero). Since its surjective as well, its a normpreserving linear bijection, which is by definition a unitary map. We
63
know T T = I from part (a). Since T is bijective, it has a linear 2sided inverse, and the equation T T = 1 shows that T is this inverse.
Hence T T = I.
(c) The rightshift operator on `2 (N) is isometric, since
k(0, a1 , a2 , . . . )k2 =
1
X
j=1
aj 2 = k(a1 , a2 , . . . )k2 .
However, this operator is not surjective, so its not unitary.
(d) First,
kT f k2 = hT f, T f i = hf, T T f i kf kkT T f k = kf k2
by the CauchySchwarz inequality; hence kT f k kf k. Then,
kf k2 = kT T f k2 = hT T f, T T f i = hT f, T T T f i kT f kkT (T T f )k kT f kkT T f k = kT f kkf k
where we have applied the previous inequality with T T f in place of f .
Dividing by kf k yields kf k kT f k. Putting the inequalities together,
kT f k = kf k. Hence T is an isometry.
Exercise 23: Suppose {Tk } is a collection of bounded operators on a Hilbert
space H, with kTk k 1 for all k. Suppose also that
Pn
Tk Tj = Tk Tj = 0
for all k 6= j.
Let SN = k= N Tk . Show that SN (f ) converges as N ! 1, for every
f 2 H. If T (f ) denotes the limite, prove that kT k 1.
Exercise 24: Let {ek }1
k=1 denote an orthonormal set in a Hilbert
P 2 space H.
If {ck }1
ck < 1, then
k=1 is a sequence of positive real numbers such that
the set
(1
)
X
A=
ak ek : ak  ck
k=1
is compact in H.
Solution. This is a standard diagonalization argument. Let be a sequence
of points in H. Consider the first components of the points in , i.e. their
components with respect to e1 . This is a sequence of complex numbers
in the compact ball {z : z c1 }, so some subsequence converges to a
complex number in this ball. If we take the corresponding subsequence of
, we obtain a subsequence S1 whose first components converge to some
number b1 with b1  c1 . Now consider the second components of the
points in S1 ; they form a sequence of complex numbers in the compact ball
{z : z c2 } and hence some subsequence converges to a complex number
b2 in this ball. Taking the corresponding subsequence S2 of S1 , we have a
sequence whose first and second components converge. Continuing, we may
inductively define sequences Sn for all n such that Sn+1 is a subsequence
of Sn , and the first n components of Sn converge. Finally, we define a
sequence S whose nth term is the nth term of Sn . This is a subsequence
of , and for any n it is eventually a subsequence of Sn (i.e. the tail of S
is a subsequence of the tail of Sn ). Hence it converges in every component.
64
But this implies convergence to a point in the Hilbert space (since the sizes
of the tails are uniformly bounded), so we are done.
Exercise 25: Suppose T is a bounded operator that is diagonal with respect
to a basis { k }, with T k = k k . Then T is compact if and only if k ! 0.
Solution. (From lecture) Suppose k ! 0. Let Tn be the nth truncation,
i.e. the operator that results when k is replaced with 0 for k > n. Then
T Tn is also a diagonal operator, with
kT
Tn k = sup
k>n
k
! 0.
Since T can be uniformly approximated by operators of finite rank, it is
compact. Conversely, suppose that k 6! 0, i.e. lim sup  n  > 0, so that
there is some subsequence nj with  nj  > 2 for some real number > 0.
Then T nj = nj nj and by orthonormality,
q
2 + 2 > p .
kT nj T nk k =
nj
nk
2
Since all the points of the sequence {T nj } are uniformly bounded away
from each other, it can have no convergent subsequence. These points all
lie in T (B), so T (B) is not compact.
Exercise 26: Suppose w is a measurable function on Rd with 0 < w(x) < 1
for a.e. x, and K is a measurable function on R2d that satisfies:
(i)
Z
K(x, y)w(y)dy Aw(x) for almost every x 2 Rd , and
Rd
(ii)
Rd
K(x, y)w(x)dx Aw(y) for almost every y 2 Rd .
Prove that the integral operator defined by
Z
T f (x) =
K(x, y)f (y)dy, x 2 Rd
Rd
2
d
is
R bounded on L (R ) with kTRk A. Note as a special case that if
K(x, y)dy A for all x, and K(x, y)dx A for all y, then kT k A.
Solution. First, note that
2 Z
p
2
p
p
p
1
K(x, y)f (y)dy =
K(x, y) w(y)
K(x, y)f (y) w(y)
Z
Z
2
1
K(x, y)w(y)dy
K(x, y)f (y) w(y) dy
(a.e.)
Aw(x)
K(x, y)f (y)2 w(y)
dy
65
by the CauchySchwarz inequality. Thus, for f 2 L2 ,
Z Z
2
kT f k2 =
K(x, y)f (y)dy dx
Z Z
K(x, y)f (y)dy dx
Z
Aw(x) K(x, y)f (y)2 w(y) 1 dydx
Z
Z
(T onelli)
2
1
=
A f (y) w(y)
K(x, y)w(x)dxdy
Z
A f (y)2 w(y) 1 Aw(y)dy
Z
= A2 f (y)2 dy
= A2 kf k2 .
Hence kT k A.
Exercise 27: Prove that the operator
Z
1 1 f (y)
T f (x) =
dy
0 x+y
is bounded on L2 (0, 1) with norm kT k 1.
Exercise 28: Suppose H = L2 (B), where B is the unit ball in Rd . Let
K(x, y) be a measurable function on B B that satisfies K(x, y) Ax
y d+ for some > 0, whenever x, y 2 B. Define
Z
T f (x) =
K(x, y)f (y)dy.
B
(a) Prove that T is a bounded operator on H.
(b) Prove that T is compact.
(c) Note that T is a HilbertSchmidt operator if and only if > d/2.
Solution.
(a) Let
C=
z2Rd :z2
dz
zd
which converges because the exponent is less than d. Then
Z
Z
Z
Ady
dz
K(x, y)dy
A
= AC
d
d
x
y
z
B
z2
so by problem 26 with w = 1, we have T bounded with kT k AC.
(b) As suggested, let
(
K(x, y) x y n1
Kn (x, y) =
0
else
and
Tn f (x) =
Kn (x, y)f (y)dy.
66
Then Tn is HilbertSchmidt (and therefore compact) since clearly Kn 2
L2 (BB) (Kn is, after all, bounded with compact support). Moreover,
Z
Z
Kn (x, y) K(x, y)dy
Ax y d+ = ACn
x y1/n
where we define
Cn =
z2Rd :z1/n
dz
.
zd
Since zd1 2 L1 (Rd ), the absolute continuity of the integral implies
Cn ! 0. By problem 26 again with w = 1, this implies kT Tn k ! 0.
Since Tn is compact, this implies that T is compact.
(c) This should actually say T is guaranteed to be HilbertSchmidt if and
only if... since K could be a lot less than the bound given. Anyhoo,
T necessarily HilbertSchmidt , Ax y d+ 2 L2 (B B)
Z Z
,
A2 x y 2d+2 < 1
B
2d + 2 >
,>
d
.
2
Exercise 29: Let T be a compact operator on a Hilbert space H and assume
6= 0.
(a) Show that the range of I T defined by
{g 2 H : g = ( I
T )f for some f 2 H}
is closed.
(b) Show by example that this may fail when = 0.
(c) Show that the range of I T is all of H if and only if the null space
of I T is trivial.
Exercise 30: Let H = L2 ([ , ]) with [ , ] identified as the unit circle.
Fix a bounded sequence { n }1
n= 1 of complex numbers, and define an
operator T f by
T f (x)
1
X
inx
n an e
whenever
n= 1
f (x)
1
X
an einx .
n= 1
Such an operator is called a Fourier multiplication operator, and the
sequence { n } is called the multiplier sequence.
(a) Show that T is a bounded operator on H and kT k = sup  n .
(b) Verify that T commutes with translations, that is, if we define h (x) =
f (x h) then
T
h = h T
for every h 2 R.
(c) Conversely, prove that if T is any bounded operator on H that commutes with translations, then T is a Fourier multiplier operator.
67
Exercise 34: Let K be a HilbertSchmidt kernel which is real and symmetric. Then, as we saw, the operator T whose kernel is K is compact and
symmetric. Let { k (x)} be the eigenvectors (with eigenvalues k ) that
diagonalize
T . Then
P
(a)
 k 2 < P
1.
(b) K(x, y)
k k (x) k (y) is the expansion of K in the basis {phik (x) k (y)}.
(c) Suppose T is a compact operator which
P is2 symmetric. Then T is of
HilbertSchmidt type if and only if
 n  < 1, where { n } are the
eigenvalues of T counted according to their multiplicities.
Exercise 35: Let H be a Hilbert space. Prove the following variants of the
spectral theorem.
(a) If T1 and T2 are two linear symmetric and compact operators on H
that commute, show that they can be diagonalized simultaneously. In
other words, there exists an orthonormal basis for H which consists of
eigenvectors for both T1 and T2 .
(b) A linear operator on H is normal if T T = T T . Prove that if T is
normal and compact, then T can be diagonalized.
(c) If U is unitary, and U = I T where T is compact, then U can be
diagonalized.
Solution.
(a) We can pretty much copy the proof verbatim with eigenvector replaced by common eigenvector. Let S be the closure of the subspace
of H spanned by all common eigenvectors of T1 and T2 . We want to
show S = H. Suppose not; then H = S S ? with S ? nonempty. If we
can show S ? contains a common eigenvector of T1 and T2 , we have a
contradiction. Note that T1 S S, which in turn implies T1 S ? S ?
since
g 2 S ? ) hT g, f i = hg, T f i = 0
for all f 2 S. Similarly, T2 S ? S ? . Now by the theorem for one
operator, T1 must have an eigenvector in S ? with some eigenvalue .
Let E be the eigenspace of (as a subspace of S ? ). Then for any
x2E ,
T1 (T2 x) = T2 (T1 x) = T2 ( x) = (T2 x)
so T2 x 2 E as well. Since T2 fixes E , it has at least one eigenvector in
E . This eigenvector is a common eigenvector of T1 and T2 , providing
us with our contradiction.
(b) This follows from part (a). Write
T =
T + T
T T
+i
.
2
2i
By a trivial calculation, both
over, since T is normal,
(T + T )(T
T ) = T 2 + T T
TT
T +T
2
and
T 2 = T 2
T
2i
are selfadjoint. More
T 2 = (T
T )(T + T )
so they commute as well. Hence, there exists an ONB of common
eigenvectors of T +T
and T 2iT . Any such common eigenvector is an
2
68
eigenvector of T , since
T + T
T T
x = x and
x=
2
2i
x ) T x = ( + i 0 )x.
(c)
Chapter 4.8, Page 202
Problem 1: Let H be an infinitedimensional Hilbert space. There exists a
linear functional ` defined on H that is not bounded.
Solution. It is a wellknown fact from linear algebra that every vector space
has a basis. This can be proved using Zorns lemma: linearly independent
sets are partially ordered by inclusion, and every chain has an upper bound
by union, so there exists a maximal linearly independent set, which is by
definition a basis. Applying this to our Hilbert space, we obtain an (algebraic) basis, i.e. one for which every vector is a finite linear combination
of basis elements. Let {en } be a countable subset of our algebraic basis.
Define `(en ) = nken k and `(f ) = 0 for f in our basis but f 6= en for any
n. We can then extend ` to the whole space in a welldefined manner, but
clearly ` is not bounded since `(en ) = nken k.
Problem 2: The following is an example of a nonseparable Hilbert space.
We consider the collection of exponentials {ei x } on R, where ranges over
the real numbers. Let H0 denote the space of finite linear combinations of
these exponentials. For f, g 2 H0 , we define the inner product as
Z T
1
(f, g) = lim
f (x)g(x)dx.
T !1 2T
T
(a) Show that this limit exists, and
(f, g) =
N
X
a k bk
k=1
PN
PN
if f (x) = k=1 a k ei k x and g(x) = k=1 b k ei k x .
(b) With this inner product H0 is a preHilbert space. Notice that kf k
supx f (x), if f 2 H0 , where kf k denotes the norm hf, f i1/2 . Let H
be the completion of H0 . Then H is not separable because ei x and
0
ei x are orthonormal if 6= 0 . A continuous function F defined on R
is called almost periodic if it is the uniform limit (on R) of elements
in H0 . Such functions can be identified with (certain) elements in the
completion H: We have H0 AP H, where AP denotes the almost
periodic functions.
(c) A continuous function F is in AP if for ever > 0 we can find a length
L = L such that any interval I R of length L contains an almost
period satisfying
supF (x + )
x
F (x) < .
69
(d) An equivalent characterization is that F is in AP if and only if every
sequence F (x + hn ) of translates of F contains a subsequence that
converges uniformly.
Problem 7: Show that the identity operator on L2 (Rd ) cannot be given as
an (absolutely) convergent integral operator. More precisely, if K(x, y) is a
measurable function onRRd Rd with the property that for each f 2 L2 (Rd ),
the integral T (f )(x) = Rd K(x, y)f (y)dy converges for almost every x, then
T (f ) 6= f for some f .
Solution. Suppose such a K exists. Let B1 and B2 be disjoint balls in Rd .
We will show that K = 0 a.e. in B1 B2 . Suppose not; then there is a
rectangle E1 E2 with E1 B1 and E2 B2 sets of positive measure,
on which K(x, y) > 0 or K(x, y) < 0; WLOG, K(x, y) > 0. (If K is allowed
complex values, we can change this condition to Re(K) > 0.) Let f = E2 .
Then for almost all x 2 E1 ,
Z
0 = f (x) = T f (x) =
K(x, y)dy.
E2
But the integral of a positive function over a set of positive measure is
nonzero, so we have a contradiction. Thus, K(x, y) = 0 a.e. in B1 B2 .
Now if we let = {(x, x) : x 2 Rd } R2d be the diagonal, we can cover
R2d \
with product sets B1 B2 . (One way to see this is topological:
products of balls form a basis for the product topology on Rd Rd , and
is closed, so its complement is open.) This implies K(x, y) = 0 a.e. on
R2d \ , and therefore a.e. on R2d . But then T f = 0 for all f , so any
nonzero f will have T f 6= f .
Problem 8: Suppose {Tk } is a collection of bounded operators on a Hilbert
space H. Assume that
kTk Tj k a2k
and kTk Tj k a2k
j,
P1
for positive constants {an } with the property that n= 1 an = A < 1.
PN
Then SN (f ) converges as N ! 1, for every f 2 H, with SN =
N Tk .
Moreover, T = limN !1 SN satisfies kT k A.
Solution. For any integers N and n,
n
kSN k2 = k(S S)2
=
N
X
N
X
j1 = N k1 = N
N
X
N
X
j1 = N k1 = N
j2n
j2n
N
X
1 = N k2n
N
X
1=
N k2n
Now since kTm k a0 for any m,
N
X
Tj1 Tk1 Tj2n
Tk2n
kTj1 Tk1 Tj2n
Tk2n
1= N
N
X
1=
k.
70
Problem 9: A discussion of a class of regular SturmLiouville operators follows. Other special examples are given in the problems below.
Suppose [a, b] is a bounded interval, and L is defined on functions f that
are twice continuously dierentiable in [a, b] (we write f 2 C 2 ([a, b]) by
d2 f
q(x)f (x).
dx2
Here the function q is continuous and realvalued on [a, b], and we assume
for simplicity that q is nonnegative. We say that
2 C 2 ([a, b]) is an
eigenfunction of L with eigenvalue if L( ) = , under the assumption
that satisfies the boundary conditions (a) = (b) = 0. Then one can
show:
(a) The eigenvalues are strictly negative, and the eigenspace corresponding to each eigenvalue is onedimensional.
(b) Eigenvectors corresponding to distinct eigenvalues are orthogonal in
L2 ([a, b]).
(c) Let K(x, y) be the Greens kernel defined as follows. Choose
(x)
to be a solution of L( ) = 0, with
(a) = 0 but 0 (a) 6= 0. Similarly, choose + (x) to be a solution of L( + ) = 0 with + (b) = 0 but
0
0
0
(x)
(x) + (x), be the Wronskian
+ (b) 6= 0. Let w = + (x)
of these solutions, and note that w is a nonzero constant.
Set
(
(x) + (y)
a x y b,
w
K(x, y) =
(y)
+ (x)
a y x b.
w
L(f )(x) =
Then the operator T defined by
Z b
T (f )(x) =
K(x, y)f (y)dy
a
is a HilbertSchmidt operator, and hence compact. It is also symmetric. Moreover, whenever f is continuous on [a, b], T f is of class
C 2 ([a, b]) and
L(T f ) = f.
(d) As a result, each eigenvector of T (with eigenvalue ) is an eigenvector of L (with eigenvalue = 1/ ). Hence Theorem 6.2 proves
the completeness of the orthonormal set arising from normalizing the
eigenvectors of L.
Solution.
(a) Let be an eigenfunction of L with eigenvalue . Then
00
= (q + ) )
00
= (q + )
Integrating by parts from a to b, we have
Z
Z
0 b
a
( 0 )2 = (q + ) 2 .
Since (a) = (b) = 0, this reduces to
Z
Z
0 2
( ) = (q + )
71
Now if is not almost everywhere zero, the LHS is strictly negative.
But the integrand on the RHS is everywhere nonnegative unless is
strictly negative. Hence the eigenvalues of L are all strictly negative.
Now suppose is an eigenvalue of L with eigenfunctions 1 and 2 .
Then
( + q)
1 2
=
)
)
)
00
1 2 =
0 0
1 2+
( 1 02 )0
0
1 2 =
00
2 1
00
1 2
=(
0
1 2
0 0
1 2
0
00
1 2
0
1 2)
+ C.
Plugging in a, we see that C = 0 because 1 (a) = 2 (a) = 0. Hence
0
0
1 2
1 2 = 0. Since the Wronskian of these solutions is zero, they
are linearly dependent. So the eigenspace can only be onedimensional.
(b) Suppose 1 and 2 are eigenfunctions with eigenvalues 1 and 2 respectively. Then
Z
Z
1
= ( 00 q )
Z
Z
00
=
q
Z
Z
0 0
= 0 ba
q
Z
Z
0 0
=
q
Z
Z
0 b
0 0
=
a
q
Z
Z
00
=
q
Z
= ( 00 q )
Z
= 2
.
R
If 1 6= 2 , this implies
1 2 = 0.
(c) The Wronskian is a constant because
w0 =
00
+
00
=q
= 0;
0
it is nonzero because plugging in at a yields
(a) + (a). We already
know 0 (a) 6= 0, and + (a) cannot be zero because then + would be
an eigenfunction with eigenvalue
R R 0, 2contradicting part (a). To show T
is HilbertSchmidt, consider
K . We will treat this on the region
R = {a y x b}; the other half is symmetric. Then
ZZ
ZZ
 + (x) (y)2
dxdy.
K(x, y)2 dxdy =
w2
R
R
Now w is a nonzero constant, as we saw above; + and
are both
continuous on a compact set and hence bounded. Thus, the integrand
is bounded, and the region of integration is compact, so the integral
72
is finite. So T is HilbertSchmidt. The symmetry of T is immediately evident from its definition. Now suppose f 2 C([a, b]). Then
T f 2 C 2 ([a, b]) because K 2 C 2 ([a, b]2 ) and the second partials of K
are bounded so that one can dierentiate T f under the integral sign.
Finally,
Z b
Z
Z
1 x
1 b
T f (x) =
K(x, y)f (y)dy =
(y)f (y)dy +
(x) + (y)f (y)dy
+ (x)
w a
w x
a
Z x
Z
(x) b
+ (x)
=
(y)f (y)dy +
+ (y)f (y)dy,
w
w
a
x
so
(T f ) (x) =
0
0
+ (x)
w
0
+ (x)
(y)f (y)dy +
+ (x)
(y)f (y)dy +
(x)
w
(x)f (x) +
Z
(x)
w
+ (y)f (y)dy
(x)
w
+ (x)f (x)
+ (y)f (y)dy
and
Z
(x) b
(T f ) (x) =
(y)f (y)dy +
(x)f (x) +
+ (y)f (y)dy
w
w
w
a
x
Z
Z
w q(x) + (x) x
q(x) (x) b
= f (x) +
(y)f (y)dy +
+ (y)f (y)dy
w
w
w
a
x
Z b
= f (x) + q(x)
K(x, y)f (y)dy
00
00
+ (x)
0
+ (x)
00
= f (x) + q(x)(T f )(x)
so L(T f ) = (T f )00 q(T f ) = f .
(d) This is more just an observation in the problem statement than something for me to do.
Chapter 5.5, Page 253
Exercise 1: Suppose f 2 L2 (RRd ) and k 2 L1 (Rd ).
(a) Show that (f k)(x) = f (x y)k(y)dy converges for a.e. x.
(b) Prove that kf kk2 kf k2 kkk1 .
\
f() for a.e. .
(c) Establish (f
k)() = k()
(d) The operator T f = f k is a Fourier multiplication operator with
multiplier m() = k().
Solution.
(a) This will follow from part (b) because an L2 function must
R be finite almost everywhere (which will prove a.e. convergence of f (x
y)k(y)dy), and absolutely convergent integrals are convergent.
(b) Just so I have it for future reference, why dont I prove the Lp version
of this. Suppose f 2 Lp with 1 < p < 1 and k 2 L1 . Let q be the
(x)
w
+ (x)f (x)
73
conjugate exponent of p. Then
Z Z
p
kf kkpp =
f (x y)k(y)dy dx
p
Z Z
1/p
1/q
f (x y)k(y) k(y) dy dx
Z
p
kf (x y)k(y)1/p kp kk(y)1/q kq dx
=
Z Z
f (x
y) k(y)dy
p/q
= kkk1 kf p kk1
k(y)dy
p/q
dx
p/q
kkk1 kkk1 kf p k1
= kkkp1 kf kpp .
Here we have used Holders inequality on f k1/p 2 Lp and k1/q 2
Lq , as well as the bound for the L1 norm of a convolution of L1 functions. In the case p = q = 2, Holders inequality reduces to CauchySchwarz.
(c) If f 2 L1 \ L2 , we already know this from our theory of Fourier
transforms on L1 . Otherwise, since L1 \ L2 is dense in L2 , we may
L2
take a sequence fn 2 L1 \ L2 with fn ! f . Then
Z
Z
f[
k() f\
e 2ix (f (x y) fn (x y)) k(y)dydx
n k() =
Z Z
f (x y) fn (x y) k(y)dydx
kf
L2
fn k2 kkk1 ! 0
so
f[
k() = lim f\
n k() = lim fn ()k() = k() lim fn () = k()f ().
(d) This is just the definition of a Fourier multiplication operator applied
to part (c).
Exercise 3: Let F (z) be a bounded holomorphic function in the halfplane.
Show in two ways that limy!0 F (x + iy) exists for a.e. x.
(a) By using the fact that F (z)/(z
+ i) is in H 2 (R2+ ).
(b) By noting that G(z) = F i 11+zz is a bounded holomorphic function
in the unit disc, and using Exercise 17 in the previous chapter.
Solution.
(a) Since F (z) M for some M ,
F (x + iy)
M
p
x + i(y + 1)
x2 + 1
74
so
Hence
1
1
F (x + iy)
x + i(y + 1)
F (z)
z+i
dx
1
1
M2
dx = M 2 .
x2 + 1
2 H 2 (R2+ ). This implies that
F (x + iy)
y&0 x + i(1 + y)
lim
exists a.e., which in turn implies that lim F (x + iy) exists a.e.
(b) I assume that I can take for granted that z 7! i 11+zz is a conformal
mapping of the unit disc into the upper half plane, since we did this
w
on a previous homework. Then define G(w) = F (i 11+w
) which is a
bounded holomorphic function on D. It now suffices to show that w
approaches the unit circle nontangentially as y = Re(z) ! 0, where
w is now given by the inverse mapping
w=
x + (1 y)i
.
x + (1 + y)i
Then
x2 + (1 y)2
x2 + (1 + y)2
by straightforward arithmetic. Now if w were approaching the unit
2
circle in a tangential manner, we would have dw
dy y=0 = 0. However,
w2 =
dw2
4(y 2 x2 1)
= 2
dy
(x + (1 y)2
which is nonzero at y = 0.
Exercise 4: Consider F (z) = ei/z /(z + i) in the upper halfplane. Note that
F (x + iy) 2 L2 (R), for each y > 0 and y = 0. Observe also that F (z) ! 0
as z ! 1. However, F 2
/ H 2 (R2+ ). Why?
Solution. For any fixed y > 0,
ei/(x+iy)
x + i(1 + y)
e1/y
x2 + 1
which is integrable, so F (x + iy) 2 L2 (R). For y = 0,
ei/x
x+i
1
1
= 2
x + i2
x +1
which is again integrable. Also, as z ! 1, the numerator approaches 1 in
magnitude while the denominator becomes infinite. However, F 2
/ H 2 , as is
1/y
suggested by the fact that our bound includes an e
term, which blows up.
The problem, of course, is that F is not bounded in the upper half plane;
it has an essential singularity at 0, and Picards theorem tells us that it
takes on every complex value (except possibly 1) in every neighborhood of
the origin.
75
Exercise 6: Suppose is an open set in C = R2 , and let H be the subspace
of L2 () consisting of holomorphic functions on . Show that H is a closed
subspace of L2 (), and hence is a Hilbert space with inner product
Z
(f, g) =
f (z)
g (z)dxdy, where z = x + iy.
Solution. For any f 2 H, f 2 2 H as well, since the square of an analytic
function is analytic. Now by the mean value property, for any z 2 and
r d(z, c ),
ZZ
1
f (z)2 = 2
f ()2 dA,
r
 zr
whence
f (z)2
1
r2
ZZ
f ()2 dA
1
kf k2 .
r2
 zr
Now on any compact K , there is a minimum value r0 > 0 of d(z, c ) for
z 2 K. (This is because the distance between a compact set and a pclosed
set always attains a nonzero minimum.) Then we have f (z) r0 kf k
for all z 2 K and f 2 L2 (). So if {fn } is a Cauchy sequence in L2 (),
then kfm fn k ! 0, whence fm fn  ! 0 as well. Thus, {fn } converges
uniformly on any compact subset of . Now it is a theorem in complex
analysis that the uniform limit of analytic functions is analytic; this may be
proved, for example, by using the M L estimate to show that the integral of
the limit around any contour is zero, and then applying Moreras theorem.
(See e.g. Gamelin p. 136.) This theorem works on any domain, e.g. the
interior of any compact disc contained in . This allows us to prove that
the limit of {fn } is analytic at each point in , so its analytic on .
Exercise 7: Following up on the previous exercise, prove:
(a) If { n }1
n=0 is an orthonormal basis of H, then
1
X
n=0
(b) The sum
2
n (z)
B(z, w) =
c2
d(z, c )
1
X
for z 2 .
n (z) n (w)
n=0
converges absolutely for (z, w) 2 , and is independent of the
choice of the orthonormal basis { n } of H.
(c) To prove (b) it is useful to characterize the function B(z, w), called
the Bergman kernel, by the following property. Let T be the linear
transformation on L2 () defined by
Z
Tf =
B(z, w)f (w)dudv, w = u + iv.
Then T is the orthogonal projection of L2 () to H.
76
(d) P
Suppose that is the unit disc. Then f 2 H exactly when f (z) =
1
n
n=0 an z , with
1
X
n=0
an 2 (n + 1)
Also, the sequence { z
over, in this case
(n+1) 1
}n=0
1/2
B(z, w) =
< 1.
is an orthonormal basis of H. More1
(1
z w)
2
Solution.
(a) First we prove a lemma, whose relevance was so kindly pointed out by
Prof. Garnett:
Lemma 3. Let {bn }1
n=0 be a sequence of complex numbers. Then
v
u1
1
X
uX
t
bn 2 = P sup
an bn .
n=0
an 2 1 n=0
P
Proof. If bn 2 < 1 this follows from the CauchySchwarz inequality
applied to `2 (N), where equality is achieved when {an } is the unit
vector inPthe same direction (actually, the conjugate) as {bn }. Now
suppose
bn 2 = 1. Then for any N , define the truncated sequence
)
b(N ) = {b(N
n } by
(
b(N ) = bn n N
n
0 else.
Then b(N ) 2 `2 , so if a(N ) = {an }1
n=0 is the unit vector in the
conjugate direction of b(N ) , we have
(N )
1
X
)
a(N
n bn =
n=0
1
X
n=0
)(N )
a(N
= kb(N ) k.
n bn
Since this goes to infinity as N ! 1, we have
v
u1
X
uX
sup
an bn = 1 = t
bn 2 .
P
kan k2 1
n=0
Returning to the problem at hand, for any sequence an with
1,
1
X
g(z) =
an n (z)
an 2
n=0
is a unit vector in H. Applying problem 6, we have at any fixed z 2
that
p
p
1
X
an n (z) = g(z)
kgk =
.
c
d(z, )
d(z, c )
n=0
77
Applying the lemma with bn =
1
X
n=0
2
n (z)
1
X
sup
an
2 1
n (z),
an
we have
!2
n (z)
n=0
.
d(z, c )2
(b) The absolute convergence of this sum follows from part (a) and the
CauchySchwarz inequality: For fixed values of z and w, { n (z)} and
{ n (w)} are vectors in `2 , so by the CauchySchwarz inequality,
qX
qX
X
 n (z)2
 n (w)2 < 1.
n (z) n (w)
To prove that the sum is independent of the choice of basis, we use
part (c). Because integration against this sum is projection onto H,
and there is only one projection map, any two such sums must be
equal almost everywhere. Im not 100% sure how to go about showing they are in fact equal everywhere. Certainly B(z, w) is analytic
in z and analytic in w (with either variable fixed, its in H as a function of the other variable). However, I dont know anything about
functions of several complex variables; is a function thats analytic in
each variable separately necessarily analytic? Or, more to the point,
continuous? Assuming so, continuity plus equality almost everywhere
implies equality. Of course, one could say that since H is being viewed
as a subspace of L2 , a.e. equality is all we need for the functions to be
the same point in the Hilbert space.
(c) Since { n } is an ONB for the closed subspace H, we can extend it
to a basis for all of L2 by complementing it with another set { k } of
orthonormal vectors. For z 2 , define Bz (w) = B(z, w). Then
Z
T f (z) =
Bz (w)f (w)dw = hf, Bz i.
We can write f in our ONB as f (w) =
whence
T f (z) = hf, Bz i
* 1
X
=
an
=
n=0
an
1
X
j (z)h n ,
ji
j,n
an
bk
k=0
P1
n=0
an
n (w)+
! 01
X
j (z)
@
k ,
j=0
bk
j (z)h k ,
k,j
n (z).
P1
k (w)
k=0 bk
1+
A
ji
This is the formula for projection onto a closed subspaceT erases all
the components n
in the
complement.
q orthogonal
o
(d) The set {
n+1
h n, ni =
n}
= zn
n+1
z ndA =
D
2
is orthonormal since
n+2
r=0
=0
r2 nrdrd = 2(n + 1)
r2n+2
2n + 2
1
0
=1
78
and
h
n,
p
Z
Z 2
(m + 1)(n + 1) 1
=
rn+m e2i(n
r=0 =0
m)
rdrd = 0
for m 6= n. Now since every analytic function has
P a power series expansion, any analytic function can be written as
bn n . This proves
that { n } is a basis for H, and also gives us the condition for an
P
P an 2
analytic function to be in L2 :
bn 2 < 1 ,
n+1 < 1, since
q
bn = an n+1 .
To obtain an expression for B(z, w), we first note that for any complex
number with  < 1,
1
(1
)2
1
X
(n + 1) n .
n=0
P
This may be obtained by dierentiating the series 1 = n termwise,
or by squaring it and collecting like terms. Both are justified by the
uniform absolute convergence of this series on compact subdisks of the
unit disk. Then
r
1 r
1
X
n+1 n n+1 n
1X
B(z, w) =
z
w
=
(n + 1)(z w)
n=
.
(1
z w)
2
n=0
n=0
Exercise 8: Continuing with Exercise 6, suppose is the upper halfplane
R2+ . Then every f 2 H has the representation
p Z 1
f (z) = 4
f0 ()e2iz d, z 2 R2+ ,
where
R1
0
f0 ()2 d
< 1. Moreover, the mapping f0 ! f given by this
formula is a unitary mapping from L2 ((0, 1), d
) to H.
Solution. Following the proof of Theorem 2.1 on page 214, we define fy ()
to be the Fourier transform of the L2 function f (x+iy). (We know f (x+iy)
is an L2 function of x for almost all y since
Z Z
2
2
kf k2 =
f (x + iy) dx dy
R
so that f (x + iy)2 dx is an integrable function of y, and therefore finite
almost everywhere. Then we can show that fy ()e2y is independent of y
using exactly the same proof in the book. (Our proof of the boundedness
of f on closed halfplanes changes slightly: we now have
Z
1
1
f ( + z)2 dxdy 2 kf k22 .
f ()2 = 2
z<
Other than that the proof requires no modification.) Having established
this, we can then define f0 () to be the function that equals fy ()e2y for
79
almost all y. The Plancherel formula then gives us
Z 1
Z 1
f (x + iy)2 dx =
f0 ()2 e 4y d.
1
This tells us that f0 () = 0 for a.a. < 0 (since the integral in is infinite
for < 0), and also gives us the relation
Z 1Z 1
kf k22 =
f (x + iy)2 dxdy
1
1
Z 1Z 1
=
f0 ()2 e 4y ddy
1 0
Z 1
Z 1
Tonelli
=
f0 ()2
e 4y dyd
0
1
Z 1
1
2
=
f0 ()
d
4
0
This tells us that kf k = k p14 f0 kL2 ((0,1),d/) . We also have by Fourier
p R1
inversion that f (z) = 4 0 p14 f0 ()e2iz d. If we replace f0 by p14 f0 ,
we will have a unitary map f0 ! f , and
p Z 1
f (z) = 4
f0 ()e2iz d.
0
Exercise 9: Let H be the Hilbert transform. Verify that
(a) H = H, H 2 = I, and H is unitary.
(b) If h denotes the translation operator, h (f )(x) = f (x h), then H
commutes with h , h H = Hh .
(c) if a denotes the dilation operator, a (f )(x) = f (ax) with a > 0, then
H commutes with a , a H = H a .
Solution.
(a) Since the projection P and the identity I are both selfadjoint, 2P I
is selfadjoint, so H = i(2P I) is skewadjoint.
(b) Since H is a linear combination of I and P , it suffices to verify that
both of these commute with h . For I this is trivial. For P , we have
and
2ih
P\
(h f )() = ()d
f ()
h f () = ()e
2ih c
\
P f () = e2ih ()f().
h P (f )() = e
Since the Fourier transform on L2 is invertible, this implies h P f =
P h f , so P commutes with h .
(c) Again, it suffices to verify that P commutes with dilations.
\
c
a P (f )() = aP f (a) = a (a)f (a) = a ()f (a)
where (a) = () because a > 0. Similarly,
P[
a f () = () a f () = ()af (a).
Hence P commutes with dilations.
80
Exercise 15: Suppose f 2 L2 (Rd ). Prove that there exists g 2 L2 (Rd ) such
that
@
f (x) = g(x)
@x
in the weak sense, if and only if
(2i) f() = g() 2 L2 (Rd ).
Solution. (Help from Kenny Maples.) Let L =
Note in particular that
@
.
@x
Then L = ( 1)
@
.
@x
\
@
() = ( 1)
Ld
() = ( 1) (2i) () = (2i) ().
@x
Now suppose g = f()(2i) 2 L2 . Define g 2 L2 as the inverse Fourier
transform of g. Using Plancherels identity, for any 2 C01 we have
hg, i = h
g , i
Z
= g() ()d
Z
= f()(2i) ()d
Z
()d
= f()Ld
i
= hf, Ld
= hf, L i.
Hence g = Lf weakly.
Conversely, suppose there exists g 2 L2 such that g = Lf weakly. Using
Plancherel again,
Z
g() ()d = h
g , i
= hg, i
= hf, L i
i
= hf, Ld
Z
= f()(2i) ()d.
Since this is true for all
2 C01 , we must have g() = f()(2i) a.e.
2
2
Since g 2 L , g 2 L by Plancherel, so f()(2i) = g() 2 L2 .
Chapter 5.6, Page 259
Problem 6: This problem provides an example of the contrast between analysis on L1 (Rd ) and L2 (Rd ).
81
Recall that if f is locally integrable on Rd , the maximal function f is
defined by
Z
1
f (x) = sup
f (y)dx,
x2B m(B) B
where the supremum is taken over all balls containing the point x.
Complete the following outline to prove that there exists a constant C
so that
kf k2 Ckf k2 .
In other words, the map that takes f to f (although not linear) is bounded
on L2 (Rd ). This diers notably from the situation in L1 (Rd ), as we observed
in Chapter 3.
(a) For each > 0, prove that if f 2 L2 (Rd ), then
Z
2A
m({x : f (x) > })
f (x)dx.
f >/2
Here, A = 3d will do.
(b) Show that
Z
Z
f (x)2 dx = 2
Rd
m(E )d,
where E = {x : f (x) > }.
(c) Prove that kf k2 Ckf k2 .
Solution.
(a) Let G = {x : f (x) >
Z
Z
f (y)dy
G
2 }.
Then 1
on G , so
2
f 
2
2
f (y)2 dy kf k2 < 1.
Now let E = {x : f (x) > }. For any x 2 E , 9Bx with x 2 Bx and
Z
Z
Z
1
1
2
m(Bx ) <
f (y)dy <
m(Bx ) +
f (y)dy ) m(Bx ) <
f (y)dy.
Bx
2
G \Bx
G \Bx
Here we have broken up the integral into the integral over the portion
of Bx where f  2 and the region where f  > 2 and bounded each
portion. Now let K be any compact subset of E ; then K is covered
by finitely many balls Bx1 , . . . , BxN . By the Covering Lemma, there
exists a subcollection Bxn1 , . . . , BxnM such that
!
N
M
[
X
m
Bi 3d
m(Bxnj ).
i=1
j=1
Then
m(K) 3d
M
X
j=1
m(Bxnj )
Z
M Z
2 3d X
2 3d
f (y)dy
f (y)dy.
j=1 G \Bxn
G
j
By the regularity of Lebesgue measure,
m(E ) =
m(K)
KE cpct
) m(E )
2 3d
f (y)dy.
82
Rd
(b) Using Tonellis theorem,
Z Z 1
Z
f (x)2 dx =
f (x)2 >y dydx =
Rd
Substituting =
Z
2
m ({x : f (x) >
y}) dy.
y, dy = 2d, this equals
m({f (x) > })d.
(c)
Chapter 6.7, Page 312
Exercise 3: Consider the exterior Lebesgue measure m introduced in Chapter 1. Prove that a set E Rd is Caratheodory measurable if and only if
E is Lebesgue measurable in the sense of Chapter 1.
Exercise 4: Let r be a rotation of Rd . Using the fact that the mapping
x 7! r(x) preserves Lebesgue measure (see Problem 4 in Chapter 2 and
Exercise 26 in Chapter 3), show that it induces a measurepreserving mape
of the sphere S d 1 with its measure d .
where E
is the
Solution. Let E S d 1 . By definition, (E) = dm(E)
union of all radii with endpoints in E. Then if r is a rotation of Rd ,
by definition. But rE
= rE
since
(rE) = dm(rE)
, x =
x 2 rE
for some 1, 2 rE
, x = r(), 1, 2 E
, x = r() (since rotations are linear)
, x 2 r(E)
= dm(E)
= m(E), so r preserves measures on the
Thus, m(rE) = dm(rE)
sphere.
Exercise
R 5: Use2 the polar coordinate formula to prove the following:
(a) Rd e x dx = 1, when d = 2. Deduce from this that the same
identity
holds for all d.
R
1
r 2 d 1
(b)
e
r
dr (S d 1 ) = 1, and as a result, (S d 1 ) = 2 d/2 / (d/2).
0
d/2
(c) If B is the
R unit ball,
vd = m(B) = / (d/2 + 1), since this quantity
1 d 1
equals 0 r
dr (S d 1 ).
Solution.
(a) For d = 2, we have by polar coordinates
Z
Z Z 1
Z 1
2
2
e x dx =
e r rdrd = 2
e
Rd
Rd
S1
rdr =
Note that for general d,
Z
Z
Z
xd
(xd
++xd
)
1
d
e
dx =
e
dx = . . . e
Rd
r 2
xd
1
...e
xd
d
r 2 1
0
= 1.
dx1 . . . dxd =
R1
x2
d
dx
83
by Tonellis Theorem.R Since we have calculated that this equals 1 for
2
d = 2, it follows that R1 e x dx = 1, whence the integral over Rd is
1 for all d.
(b) Using integration by parts, for d 3,
Z 1
Z 1 r2
2
e r d 2 1
e
r 2 d 1
e
r
dr =
r
+
(d 2)rd 3 dr.
0
2
2
0
0
Since (S d 1 ) is just the reciprocal of this first integral (which follows
immediately from applying the polar coordinates formula to the result
in part (a)), it follows that (S d 1 ) = d22 (S d 3 ). We now prove
the formula (S d 1 ) = 2 d/2 / (d/2) by induction. The base cases are
p
(S 1 ) = 2 = 2 2/2 / (2/2) and (S 2 ) = 4 = 2 3/2 /( /2) =
p
2 3/2 / (3/2) since (3/2) = 1/2 (1/2) = 1/2 . Now if we let
ad = 2 d/2 / (d/2), then ad = d22 (d 2)/2 /((d 2)/2 ((d 2)/2) =
2
d 1
) satisfy the same recurrence and initial
d 2 ad 2 . Since ad and (S
conditions, they are equal for all d.
(c) By polar coordinates,
Z 1
1
2 d 2
d/2
d 1
m(B) = (S
)
rd 1 dr = 2 d/2 / (d/2) = d
=
.
d
(d/2 + 1)
2 2 (d/2)
0
Exercise 8: The fact that the Lebesgue measure is uniquely characterized by
its translation invariance can be made precise by the following assertion:
If is a Borel measure on Rd that is translationinvariant, and is finite
on compact sets, then is a multiple of Lebesgue measure m. Prove this
theorem by proceeding as follows.
(a) Suppose Qa denotes a translate of the cube {x : 0 < xj a, j =
1, . . . , d} of side length a. If we let (Q1 ) = c, then (Q1/n ) = cn d
for each integer n.
(b) As a result is absolutely continuous with respect to m, and there is
a locally integrable function f such that
Z
(E) =
f dx.
E
(c) By the dierentiation theorem (Corollary 1.7 in Chapter 3) it follows
that f (x) = c a.e., and hence = cm.
Solution.
(a) Because Q1 is a disjoint union of nd translates of the cube Q1/n ,
(Q1 ) = nd (Q1/n ) ) (Q1/n ) = n d (Q1 ) = cn d .
(b) Let E be Borel measurable with m(E) = 0. Then for any > 0 there
is an open set U with m(U \ E) < ) m(U ) < . We can write
U as a countable disjoint union of cubes Qj whose side lengths are
of the form 1/n, for example by decomposing U into dyadic rational
cubes as described
on pp.
(Qj ) = cm(Qj ) by part (a),
P
P 78. Then P
so (U ) =
(Qj ) =
cm(Qj ) = c m(Qj ) = cm(U ) < c. This
can be done for any , so (E) = 0. Thus, is absolutely continuous
84
wrt m, so there exists
R a locally integrable Borel measurable function
f such that (E) = E f dm.
(c) Let x be a Lebesgue point of f . Let Qn be a series of dyadic rational
cubes containing x. (Just to be clear, these are halfopen dyadic
rational cubes, i.e. ones of the form 2mnii xi < m2in+1
for i = 1, . . . , d.)
i
Then Qn shrinks regularly to x because the ratio of a cube to the
circumscribing ball is constant. For each cube we have
Z
1
1
1
f d =
(Qn ) =
cm(Qn ) = c,
m(Qn ) Qn
m(Qn )
m(Qn )
so f (x) = c at every Lebesgue point x of f (hence a.e.).
Exercise 9: Let C([a, b]) denote the vector space of continuous functions on
the closed and bounded interval [a, b]. Suppose we are given a Borel measure
on this interval, with ([a, b]) < 1. Then
Z b
f 7! `(f ) =
f (x)d(x)
a
is a linear functional on C([a, b]), with ` positive in the sense that `(f ) 0
if f 0.
Prove that, conversely, for any linear functional ` on C([a, b]) that is
positive in the above sense, there is a unique finite Borel measure so that
Rb
`(f ) = a f d for f 2 C([a, b]).
Solution. Define the notation f u to mean 0 f 1 and f = 1 on [a, u].
Let
F (u) = ` (f ).
f
Then F is increasing on [a, b], because for u0 > u, f u0 ) f u so F (u0 )
is the infimum of a smaller class of sets. To show F is right continuous, it
suffices to show that for every f
u and > 0 there exists a u0 > u and
0
0
0
f
u with `(f ) < `(f ) + . Let f
u and let C = `(f ). By continuity,
the function (1+ C )f is greater than 1 in some neighborhood u < x < x+ .
Let f 0 u + 2 and f 0 (y) = 0 for y
. (Such an f 0 can be constructed, for
0
example, as piecewise linear, say f = 1 on [a, u+ /2], f 0 (u+3 /4) = 0, and
f is linear between /2 and 3 /4.) Then f 0 (x) (1 + C )f (x) everywhere,
so `(f 0 ) (1 + C )`(f ) = C + . This proves that F is right continuous.
By Theorem 3.5, there exists a unique Borel measure on [a, b] such that
((a0 , b0 ]) = F (b0 ) F (a0 ) for all a a0 < b0 b.
Rb
Now we need to show `(f ) = a f d for all f 2 C([a, b]). Let L(f ) =
Rb
f d. Then it suffices to show `(f ) L(f ) since this will imply `(f ) =
a
`( f ) L( f ) = L(f ) ) `(f ) L(f ). Let > 0. Because continuous
functions can be uniformly approximated by step functions, we may choose
a step function f f with L(f ) < L(f ) + . Write
X
f =
ck (ak ,bk ] .
WLOG we may assume P
that the intervals (ak , bk ] are disjoint. Choose
a0k > ak and let f0 =
ck (a0k ,bk ] . Now by the definition of F there
exist continuous gk and hk with gk
b k , hk
a0k , and `(hk ) F (a0k ) <
85
`(gk ) F (bk ) < 2k . WLOG we may also assume f < ck (1
since otherwise we may take
(
max(f, ck (1 hk )) ak < x < a0k
0
ck (1 hk ) =
ck
a0k x < bk
hk ) on (ak , a0k )
and the function h0k defined by these relations will also be continuous, satisfy
h0k a0k , and have h0k < hk ) `(h0k ) < `(hk ). Now let
X
f =
ck (gk hk ).
Then we have f f by the above remarks concerning hk . Note also that
X
X
`(f ) =
ck (`(gk ) `(hk )) <
ck (F (bk ) F (a0k )) + = L(f0 ) + .
Since we also have the relations f f and f f0 , and both ` and L are
positive,
`(f ) < `(f ) < L(f 0 ) + < L(f ) + < L(f ) + 2.
This is true for all , so `(f ) L(f ).
Exercise 10: Suppose , 1 , 2 are signed measures on (X, M) and a (positive) measure on M. Using the symbols ? and defined in Section 4.2,
prove:
(a) If 1 ? and 2 ? , then 1 + 2 ? .
(b) If 1 and 2 , then 1 + 2 .
(c) 1 ? 2 implies 1  ? 2 .
(d) .
(e) If ? and , then = 0.
Solution.
(a) Let disjoint A1 and B1 be chosen such that 1 (E) = 1 (A1 \ E) and
(E) = (B1 \ E) for all measurable E. Similarly, choose A2 and
B2 disjoint with 2 (E) = 2 (A2 \ E) and (E) = (B2 \ E). Let
A = A1 [ A2 and B = B1 \ B2 . Note that A and B are disjoint
because A1 \ B A1 \ B1 = ; and similarly for A2 . Then for any
measurable E, (E) = (E \ B1 ) = (E \ B) + (E \ (B1 \ B2 ).
But (B1 \ B2 ) = ((B1 \ B2 ) \ B2 ) = 0, so (E) = (E \ B).
Similarly, 1 (E) = 1 (E \ A1 ) = 1 (E \ A) 1 (E \ (A \ A1 )), but
1 (A \ A1 ) = 1 ((A \ A1 ) \ A1 ) = 0, so 1 (E) = 1 (E \ A) and by the
same token, 2 (E) = 2 (E \ A), so (1 + 2 )(E) = (1 + 2 )(E \ A).
Thus, and 1 + 2 are supported on the disjoint sets A and B.
(b) (E) = 0 ) 1 (E) = 2 (E) = 0 ) (1 + 2 )(E) = 0.
(c) Choose disjoint A and B such that 1 (E) = 1 (E \ A) and 2 (E) =
2 (E \ B) for all measurable E. Then
X
X
1 (E) = sup
1 (Ej ) = sup
1 (Ej \ A) = 1 (E \ A)
j
and similarly 2 (E) = 2 (E \ B). Hence 1  and 2  are supported
on the disjoint setsP
A and B.
(d) (E) = 0 ) sup (Ej ) = 0 ) (Ej ) = 0 for all subsets Ej
E ) (E) = 0.
86
(e) Let disjoint A and B be chosen with (E) = (E \ A) and (E) =
(E \ B). Then for any measurable E, (E \ A) = ((E \ A) \ B) = 0
because A and B are disjoint. Then (E) = (E \ A) = 0 because
(E \ A) = 0 and .
Exercise 11: Suppose that F is an increasing normalized function on R, and
let F = FA + FC + FJ be the decomposition of F in Exercise 24 of Chapter
3; here FA is absolutely continuous, FC is continuous with FC0 = 0 a.e., and
FJ is a pure jump function. Let = A + C + J with , A , C , and
J the Borel measures associated to F, FA , FC , and FJ respectively. Verify
that:
(i) A is absolutely
continuous with respect to Lebesgue measure and
R
A (E) = E F 0 (x)dx for every Lebesgue measurable
R set E. R
(ii) RAs a result, if F is absolutely continuous, then f d = f dF =
f (x)F 0 (x)dx whenever f and f F 0 are integrable.
(iii) C + J and Lebesgue measure are mutually singular.
Solution.
(i) By definition
A (E) =
=
inf
E[(aj ,bj ]
inf
E[(aj ,bj ]
inf
E[(aj ,bj ]
FA (bj )
XZ
Z
bj
FA (aj )
F 0 (x)dx
aj
F 0 (x)dx
[(aj ,bj ]
F 0 (x)dx.
To prove the reverse inequality, let > 0 and use the absolute
continuR
ity of the integral to find a > 0 such that m(E) < ) E F 0 (x) < .
(In case the assumption that F 0 2 L1 is a problem, we can treat
the intersection of E with each interval [n, n + 1) separately.) Now
since E is Lebesgue measurable, there is an open set U
E such
be constructed by writing U as a disjoint
that m(U \ E) < . Let U
union of open intervals (aRj , bj ) and replacing
R each0 with (aj , bj ]. Then
\E) = m(U \E) and F 0 (x)dx =
\E
m(U
F (x)dx because U
U \E
U \E
is U \ E plus countably many points. Thus
Z
Z
Z
Z
F 0 (x)dx =
F 0 (x)dx +
F 0 (x)dx +
F 0 (x)dx.
\E
U
is one of the sets over which the infimum is taken in the definiBut U
R
tion of AR(E), so A (E) + E F 0 (x)dx. This is true for any , so
A (E) = E F R0 (x)dx. R
(ii) The equation E f d = E f F 0 (x)dx follows immediately from (a) in
the case where f is a characteristic function. By the linearity of the
integral, it holds for f a simple function as well. The result for nonnegative f follows from the Monotone Convergence Theorem: Choose
87
f d =
simple fn % f . Then
Z
Z
Z
Z
(lim fn )d = lim
fn d = lim
fn (x)F 0 (x)dx =
(lim fn )(x)F 0 (x)dx =
f (x)F 0 (x)dx.
Finally, linearity allows us to extend to functions
in L1 ().
R
R Note that
1
0
1
f 2 L () i f F 2 L (dx) because both f d and fRF 0 dx are
defined
as suprema of integrals of simple functions, and cd =
R
gF 0 dx for g simple so the suprema are taken over the same sets.
The condition that f be integrable is, so far as I can tell, superfluous,
unless it means integrable, in which case its redundant.
(iii) By Exercise 10a, it is sufficient to prove that P
C and J are both
1
singular wrt Lebesgue measure. Write FJ (x) = k=1 ck [xk ,1) . Let
A = {xk } which is countable and therefore has Lebesgue measure zero,
so that Lebesgue measure is supported on Ac . Now Ac is open, so it
is covered by countably many intervals (aj , bj ]. (We can write Ac as a
countable disjoint union of open intervals, and any open interval is a
countable union of halfclosed intervals.) Thus, any subset E Ac can
be covered by countably many intervals (aj , bj ] with FJ (aj ) = FJ (bj ),
so J (E) = 0. Thus, J ? m.
The proof that C ? m is a bit trickier. (Help from Paul Smith on
this part.) We will use the following lemma, taken from page 35 of
Folland:
Lemma 4. If F is the Borel measure corresponding to the increasing,
rightcontinuous function F , then for any measurable set E,
X
(E) =
inf
((aj , bj )).
E[(aj ,bj )
In words, this lemma says that it is equivalent to use coverings of
open intervals instead of halfopen intervals. This is nice because it
enables us to use theorems about open covers. The straightforward
but unenlightening proof is in Folland for anyone who cares.
Let A = {x 2 R : FC0 (x) = 0}c . Then m(A) = 0 by hypothesis, so
Lebesgue measure is supported on Ac . We wish to show that C (Ac ) =
0, so that C is supported on A. It is sufficient to show that C (Ac \
[0, 1]) = 0 since Ac = [Ac \[n, n+1] and replacing FC (x) by FC (x n)
shifts Ac \ [n, n + 1] to Ac \ [0, 1]. Thus, let B = Ac \ [0, 1]. Let > 0.
Now for any x 2 B, FC0 (x) = 0, so 9hx > 0 such that FC (y) FC (x) <
y x for y 2 [x h, x + h]. By the Dreiecksungleichung, this implies
FC (x + h) FC (x h) < 2h. Thus, if we let Ix = (x h, x + h), we
have (Ix ) ((x h, x+h]) = FC (x+h) FC (x h) < 2h = m(Ix ).
The intervals Ix are an open cover of B, so we can take a countable
subcover In . (Here we use the fact that every subset of R is a Lindelof
space. The proofbyjargon of this fact is that every subspace of a
separable metric space is separable, and a metric space is separable
i it is Lindel
of. The direct proof is that every open subset of R is a
countable union of rational intervals, so for each rational interval we
can take a member of our open cover (if there is any) which contains
that interval, and the resulting countable subcollectionP
will still cover
our set.) Next, we shrink the intervals In to make
m(In ) < 3.
88
We do this inductively: For a given In , if In Ij for some j < n, we
discard In entirely; similarly, if Ij In for some j < n then we discard
Ij . Once this is done, In \ ([j<n Ij ) is an open subset of In , hence a
disjoint union of open intervals; but none of these intervals can have
both its endpoints within In because this would imply Ij In for some
j < n. Hence if In = (a, b), then In \ ([j<n Ij ) = (a, ) [ ( , b) for
1
some a < b. By replacing a with max(a,
2n+1 ) and b with
1
0
min(b, + 2n+1 ), we form a new interval In In with the property
that m(In \ ([j<n Ij )) < 21n . However, [nj=1 Ij0 = [nj=1 Ij because we
only delete parts of an interval that are covered by other intervals.
We would also like In0 to still have the property that (In0 ) < m(In0 ).
Unfortunately, this is only guaranteed as long as In0 contains the central
point xn from the interval In . So far I have not been able to close this
hole in the proof.
Overlooking this problem, we have found an open cover In0
B with
the properties that (In0 ) < m(In0 ) for each n, and m(In0 \([j<n Ij0 )) <
1
1 3
0
2n . We may additionally assume that In [ 2 , 2 ] for all n since
we
1]. This will then imply that
P are only interested in covering [0,
m(In ) < 3 because if we write In0 = An [ Bn where An = In \
([j<n Ij0 ) and Bn = In0 \ ([j<n Ij0 ), then the An are disjoint and have
union equal to [In0 , and
X
X
X
X 1
m(In0 ) =
m(An ) + m(Bn ) = m ([An ) +
m(Bn ) 2 +
=3
2n
where m ([An ) < 2 because [An [ 12 , 32 ]. This then implies
X
X
(B) ([In0 )
(In0 )
m(In0 ) < 3.
This is true for any , so (B) = 0.
Exercise 14: Suppose (Xj , Mj , j ), 1 j k, is a finite collection of measure spaces. Show that parallel with the case k = 2 considered in Section 3
one can construct a product measure 1 k on X = X1 Xk . In
fact, for any set E X such that E = E1 Ek , with Ej 2 Mj for all
Qk
j, define 0 (E) = j=1 j (Ej ). Verify that 0 extends to a premeasure on
the algebra A of finite disjoint unions of such sets, and then apply Theorem
1.5.
Solution. First a hand should at least be waved at the fact that A is an
algebra. It is closed under complements because
(E1 Ek )c = (E1c X2 Xk )
[ (E1 E2c X3 Xk )
[ [ (E1 E2 Ek
Ekc ).
This is a stopping time argument: we divide the complement into k
sets based on which is the first of the Ek that a point fails to be in. The
intersection of two unions of disjoint measurable rectangles is another union
89
of disjoint measurable rectangles, so we only need to check unions. This
follows from
(E1 Ek ) [ (F1 Fk ) = (E1 Ek )
[ (F1 \ E1 F2 Fk )
[ (F1 \ E1 F2 \ E2 F3 Fk )
[ (F1 \ E1 F2 \ E2 F3 \ E3 F4 Fk )
[ [ (F1 \ E1 Fk
\ Ek
Fk \ Ek ).
Finally, to show that the extension from rectangles to a premeasure on
the algebra A generated by them is welldefined, let E1 Ek be a
measurable rectangle, and suppose E1 Ek = [j E1j Ekj where
the union is disjoint. This immediately implies
1
X
E1 (x1 ) . . . Ek (xk ) =
E j (x1 ) . . . E j (xj )
j=1
for all (x1 , . . . , xk ) 2 X1 Xk . We can integrate both sides with respect
to x1 , using the monotone convergence theorem to move the integral inside
the sum on the RHS, to obtain
1
X
1 (E1 ) E2 (x2 ) . . . Ek (xk ) =
1 (E1j ) E j (x2 ) . . . E j (xk ).
j=1
We then integrate each side wrt x2 , etc. After doing this k times we obtain
1
1
X
X
j
j
1 (E1 ) . . . k (Ek ) =
1 (E1 ) . . . k (Ek ) ) 0 (E1 Ek ) =
0 (E1j Ekj )
j=1
j=1
as desired.
Since 0 is a premeasure on A, it extends to a measure on the algebra
generated by A by Theorem 1.5.
Exercise 15: The product theory extends to infinitely many factors, under
the requisite assumptions. We consider measure spaces (Xj , Mj , j ) with
j (Xj ) = 1 for all but finitely many j. Define a cylinder set E as
{x = (xj ), xj 2 Ej , Ej 2 Mj , Ej = Xj for all but finitely many j}.
Q1
For such a set define 0 (E) = j=1 j (Ej ). If A is the algebra generated
by the cylinder sets, 0 extends to a premeasure on A, and we can apply
Theorem 1.5 again.
Solution. First, note that finite disjoint unions of cylinder sets form an algebra, which is therefore the algebra A. To see this, we can just apply
Exercise 14 because of the condition that finitely many indices inQthe cylinder have Ej 6= Xj . For example, to see how unions work, let
Q Ej be a
cylinder set (where Ej = Xj for all but finitely many j) and Fj another
cylinder set. Then there are finitely many j for which either Ej or Fj is not
Xj ; we may apply the decomposition from Exercise 14 to these components
while
Q leaving
Q the others untouched, and hence obtain a decomposition of
( Ej ) [ ( Fj ) into finitely many disjoint cylinder sets. Similar comments
apply to intersections and complements.
90
To verify that 0 extends to a premeasure on A, let
set, and suppose
1
1 Y
1
Y
[
Ej =
Ejk
j=1
Ej be a cylinder
k=1 j=1
where the union is disjoint and all but finitely many Ejk are equal to Xj for
any fixed k. The characteristicfunction version of this statement is
1
1 Y
1
Y
X
Ej (xj ) =
Ejk (xj ).
j=1
k=1 j=1
Integrating both sides with respect to x1 and using the monotone convergence theorem to move the integral inside the sum on the right,
1
1
1
Y
X
Y
k
1 (E1 )
(x
)
=
(E
)
Ej
j
1
Ejk (xj ).
1
j=2
k=1
j=2
Repeating the process ` times, we have
1
1
1
Y
X
Y
1 (E1 ) . . . ` (E` )
1 (E1k ) . . . ` (E`k )
Ej (xj ) =
j=`+1
k=1
Ejk (xj ).
j=`+1
Q
As ` ! 1, the LHS approaches 0 ( Ej ); in fact, it will equal it after
a finite number of steps because all but finitely many Ej equal Xj and
j (Xj ) = 1 for all but finitely many Xj . For the RHS, we apply monotone
convergence again (this time in `) to see that it approaches
0
1
1 Y
1
1
1
X
X
Y
j (Ejk ) =
0 @
Ejk A
k=1 j=1
k=1
j=1
as desired. Hence 0 extends to a premeasure on A, and therefore to a
measure on the sigmaalgebra generated by A by Theorem 1.5.
Exercise 16: Consider the ddimensional torus Td = Rd /Zd . Identify Td
as T1 T1 and let be the product measure on Td given by =
1 d , where j is Lebesgue measure on Xj identified with the circle
T. That is, if we represent each point in Xj uniquely as kj with 0 < xj 1,
then the measure j is the induced Lebesgue measure on R restricted to
(0, 1].
(a) Check that the completion is Lebesgue measure induced on the cube
Q = {x : 0 < xj 1, j = 1, . . . , d}.
(b) For each function f on Q let f be its extension to Rd which is periodic,
that is, f(x + z) = f(x) for every z 2 Zd . Then f is measurable on Td
if and only if f is measurable on Rd , and f is continuous on Td if and
only if f is continuous on Rd .
d
(c) Suppose f and
R g are integrable on T . Show that the integral defining
(f g)(x) = Td f (x y)g(y)dy is finite for a.e. x, that f g is integrable
over Td , and that f g = g f .
(d) For any integrable function f on Td , write
X
f
an e2inx
n2Zd
91
R
to mean that an = Td f (x)e 2inx dx. Prove that if g is also integrable,
P
and g n2Zd bn e2inx , then
X
f g
an bn e2inx .
n2Zd
2inx
(e) Verify that {e
} is an orthonormal basis for L2 (Td ). As a result
P
kf kL2 (Td ) = n2Zd an 2 .
(f) Let f be any continuous periodic function on Td . Then f can be uniformly approximated by finite linear combinations of the exponentials
{e2inx }n2Zd .
Solution.
(a) This follows from the translation invariance of . To show that the
product of translation invariant measures is translation invariant, let
E = E1 Ed be a measurable rectangle in Td , and x = (x1 , . . . , xd ) 2
Td . Then
(E+x) = (E1 +x1 ) (Ed +xd ) = 1 (E1 +x1 ) . . . d (Ed +xd ) = 1 (E1 ) . . . d (Ed ) = (E)
so is translation invariant on measurable rectangles. This implies
that the outer measure generated by coverings of measurable rectangles is also translation invariant. But is just the restriction of
to the sigmaalgebra of Caratheodorymeasurable sets, so it is translation invariant as well. This implies that is a multiple of Lebesgue
measure; since (Td ) = m(Q) = 1, we must have m = (modulo the
correspondence between Q and Td ).
(b) This is blindingly obvious, but
f mble (resp. cts) , f 1 (U ) mble (resp. open) for open U
, f 1 (U ) mble (resp. open) in Rd
, f mble (resp. cts).
Here we use the fact that f 1 (U ) is a lattice consisting of translates
of f 1 (U ) by Zd ; such a set is open or measurable i f 1 (U ) is.
(c) This is a simple application of Tonellis Theorem, exactly analogous
to the case in L1 (Rd ):
Z
Z Z
f g(x)dx =
f (x y)g(y)dy dx
d
d
Td
ZT Z T
f (x y)g(y)dydx
Td Td
Z Z
=
f (x y)g(y)dxdy
Td
Td
= kf kL1 (Td ) kgkL1 (Td )
so f g is integrable on Td . This in turn implies that it is finite a.e.
Finally, the change of variables u = x y shows that
Z
Z
f g(x) = f (x y)g(y)dy = f (u)g(x u)du = g f (x).
92
(d) Once again, there is absolutely nothing dierent from the onevariable
case. Since f g is integrable by our above remarks, and e 2inx  = 1,
f g(x)e 2inx is also integrable, so by Fubinis theorem
Z
Z
Z
f g(x)e 2inx dx =
e 2inx
f (x y)g(y)dydx
Td
Td
Td
Z
Z
=
g(y)
f (x y)e 2inx dxdy
Td
Td
Z
Z
2iny
=
g(y)e
f (x y)e 2in(x y) dxdy
Td
Td
Z
=
g(y)e 2iny an dy
Td
= an bn .
Td
(e) The orthonormality of this system is evident, since
Z
d Z
d
Y
Y
e 2inx e2imx dx =
e2i(m n)x dx =
e2i(mj nj )xj dxj =
Td
j=1
nj
mj
n
m
j=1
where is the Kronecker delta function. To show completeness, we
use the fact that an orthonormal system {en } in a Hilbert space is
complete i hf, en i = 0 for all n ) f = 0. Suppose
Z
Z
Z
Z
2inx
2inx
2in1 x1
2in2 xn
0 = hf, e
i=
f (x)e
dx =
e
e
...
e 2ind xd f (x1 , . . . , xd )dxd . . . dx1
Td
Td
for all n1 , . . . , nd , where the use of Fubinis theorem is justified by the
integrability of f and the fact that e 2ink xk  = 1. Let
Z
Z
Z
F1 (x1 ) =
e 2in2 x2
e 2in3 x3 . . . e 2ind xd f (x1 , . . . , xd )dxd . . . dx2 .
T
T
T
R
Then T e 2in1 x1 F1 (x1 )dx1 = 0 for all n1 , so F1 (x1 ) = 0 a.e. by the
completeness of exponentials in the 1dimensional case. Now let
Z
Z
F2 (x1 , x2 ) =
e 2in3 x3 . . . e 2ind xd f (x1 , . . . , xd )dxd . . . dx1 .
T
For any fixed
R value of x1 , Fn (x1 , x2 ) is a function of x2 with the property that T e 2in2 xn F2 (x1 , x2 )dx2 = F1 (x1 ) = 0 a.e., so we must
have F2 (x1 , x2 ) = 0 for a.e. x2 . Continuing inductively, we see that
f (x1 , . . . , xd ) = 0 for a.e. x1 , . . . , xd .
(f) This problem is begging for the StoneWeierstrass theorem, but since
we havent covered that in class, RIll reluctantly do the convolution
stu. Let g (x) = 1d [0, ]d . Then g (x)dx = 1, so
Z
Z
f (x) f g (x) = f (x) g (y)dy
f (x y)g (y) dy
Z
f (x) f (x y)g (y)dy.
By the uniform continuity of f , given > 0 there exists
that x y < 0 ) f (x) f (y) < . For < 0 ,
Z
f (x) f g (x) g (y)dy = .
> 0 such
93
Thus, f g (x) ! f (x) uniformly as ! 0. However,
P if2 an are the
Fourier
coefficients
of
f
and
b
those
of
g
,
then
an  < 1 and
n
P
bn 2 < 1 because f and g are in L2 so their Fourier
P transforms
are as well. Then by the CauchySchwarz inequality,
an bn  < 1.
1
This implies that f\
g 2 L1 and
since
f
g
2
L
,
Fourier
inversion
P
holds and we have f g (x) = an bn e2inx a.e. (in fact, everywhere,
since both sides are continuous). Now we can choose such that
f f g  < 2 everywhere. For this , since the tails of the convergent
P
P
series an bn go to zero, we can choose some truncation nN an bn 
P
such that n>N an bn  < 2 . Then for any x,
f (x)
nN
an bn e2inx f (x)
f g (x) + f g (x)
= f (x)
f g (x) +
f (x)
f g (x) +
e2inx
nN
an bn e2inx
n>N 
n>N 
an bn 
+ = .
2 2
Thus, f can be uniformly approximated by trigonometric polynomials.
Exercise 17: By reducing to the case d = 1, show that each rotation x 7!
x + of the torus Td = Rd /Zd is measure preserving, for any 2 Rd .
m(
Solution. We first suppose that E Td is a measurable rectangle E =
E1 Ed where Ek T for k = 1, . . . , d. Then
(E)) = m(E ) = m((E1 1 ) (Ed d )) = m(E1 ) . . . m(Ed ) = m(E).
Hence is measurepreserving on measurable rectangles. But since the
measure of any set is computed in terms of its coverings by measurable
rectangles, is measurepreserving on all measurable sets. (We actually use
here the fact that 1 is measurepreserving as well; if {Rk } is a covering of
E by measurable rectangles, then { 1 (Rk )} is a covering of 1 (E) with
the same measure; conversely, if {Rk0 } is a covering of 1 (E) by measurable
rectangles, then { (Rk0 )} is a covering of E with the same measure.)
Exercise 18: Suppose is a measurepreserving transformation on a measure
space (X, ) with (X) = 1. Recall that a measurable set E is invariant
if 1 (E) and E dier by a set of measure zero. A sharper notion is to
require that 1 (E) equal E. Prove that if E is any invariant set, there is
a set E 0 so that E 0 = 1 (E), and E and E 0 dier by a set of measure zero.
Solution. Let
E0 =
1 [
1
\
n=1 k=n
(E).
94
Then
E0 \ E =
1 [
1
\
n=1 k=n
and
E \ E0 =
1 \
1
[
n=1 k=n
(E) \ E
E\
(E).
But E \ k (E) and k (E) \ E both have measure zero (this follows from
m(E E 0 ) = 0 by an easy induction), and countable unions and intersections of null sets are null, so m(E E 0 ) = 0. Moreover,
(E 0 ) =
1 [
1
\
(E) =
n=1 k=n
1 [
1
\
(E) = E 0
n=2 k=n
because the sets inside the intersection are nested so we get the same set
whether we start at n = 1 or n = 2.
Exercise 19: Let be a measurepreserving transformation on (X, ) with
(X) = 1. Then is ergodic if and only if whenever is absolutely continuous with respect to and is invariant (that is, ( 1 (E)) = (E) for
all measurable sets E), then = c, with c a constant.
Solution. We use the fact that is ergodic i the only
R functions
R with f =
f a.e. are constant a.e., as well as the fact that E f d = 1 (E) f d
for measurepreserving maps . Let be ergodic and let be an
invariant measure. By the RadonNikodym theorem, d = hd for some
function h 2 L1 (). Then for any measurable E,
Z
Z
(E) =
hd =
h d.
E
By the invariance of , this equals
Z
( 1 (E)) =
1 (E)
1 (E)
hd.
Since this is true for any measurable E, h = h a.e. But since is
ergodic, this implies h is constant a.e., so = c for some constant c.
Conversely, suppose every invariant absolutely continuous measure is a
constant times . Let f 2 L1 be any function with the property that
f = f a.e. Define an absolutely continuous measure by d = f d.
Then the above calculation (run in reverse) shows that is invariant, so
= c. But this implies that f d = d = cd so f is constant a.e. Hence
is ergodic.
Exercise 20: Suppose is a measurepreserving transformation on (X, ). If
(
(E) \ F ) ! (E)(F )
as n ! 1 for all measurable sets E and F , then (T n f, g) ! (f, 1)(1, g)
whenever f, g 2 L2 (X) with (T f )(x) = f ( (x)). Thus is mixing.
95
Solution. Suppose ( n (E) \ F ) ! (E)(F ) for measurable E and F .
This means that (T n f, g) ! (f, 1)(1, g) if f and g are characteristic functions: say f = E and g = F , then
Z
Z
n
n
(T n f, g) =
(
(x))
(x)dx
=
(E)\F ) ! (E)(F ) = (f, g).
E
F
n (E) F = m(
By linearity in f and conjugatelinearity in g, this implies (T n f, g) !
(f, 1)(1, g) if f and g are measurable simple functions. Now let f, g 2 L2
and let fm ! f and gm ! g be sequences of measurable simple functions.
For each m we have
(T n f, g) = (T n f, g
gm ) + (T n (f
fm ), g) + (T n fm , gm ).
Since T is an isometry, (T n f, g gm ) kT n f kkg gm k = kf kkg gm k ! 0
as m ! 1. Similarly, (T n (f fm ), g) ! 0 uniformly in n as m ! 1.
n
m
Finally, (T n fm , gm ) ! (fm , 1)(1, gm ) ! (f, 1)(1, g). (To be more precise
about this business of taking limits in two dierent variables, choose fm
and gm with kfm f k, kgm gk < . Then since T is an isometry,
(T n f, g) = (T n fm , gm ) + h(n)
where kh(n)k < 2 for all n. Letting n ! 1, we see that (T n f, g) is eventually within 2 of (fm , 1)(1, gm ), which in turn is within C of (f, 1)(1, g)
for some constant C. This is true for all , so (T n f, g) ! (f, 1)(g, 1).)
Exercise 21: Let Td be the torus, and : x 7! x + the mapping arising
in Exercise 17. Then is ergodic if and only if = (1 , . . . , d ) with
1 , . . . , d , and 1 are linearly independent over the rationals. To do this
show that:
(a)
Z
m 1
1 X
k
f ( (x)) !
f (x)dx
m
Td
k=0
as n ! 1, for each x 2 Td , whenever f is continuous and periodic
and satisfies the hypothesis.
(b) Prove as a result that in this case is uniquely ergodic.
Solution.
(a) Suppose first that 1 , . . . , d and 1 are dependent over Q, say
p
a1 1 + + ad d = ,
q
where i 2 Q. Let
1
},
2
where {z} = z bzc denotes the fractional part of z. Then m(E) = 12
but E = 1 (E), so is not ergodic.
On the other hand, suppose 1 , . . . , d and 1 are independent over Q.
Let f (x) = e2inx be any complex exponential. If n = 0, then
Z
m 1
1 X
k
f (( (x)) = 1 =
f (x)dx.
m
Td
E = {(x1 , . . . , xd ) : 0 < {q(a1 x1 + + ad xd )} <
k=0
96
If n 6= 0, then
m 1
m
X1
1 X
1
e2inx 1 e2imn
f (x)dx = e2inx
e2ikn =
.
m
m
m
1 e2in
k=0
k=0
Since 1
 2, this goes to zero as m ! 1, so
Z
m 1
1 X
k
f (( (x)) !
f (x)dx.
m
Td
2imn
k=0
Am f (x)
Finally, since complex exponentials are uniformly dense in the continuous periodic functions by exercise 16f, the above limit holds for any
continuous periodic function: Let f be a continuous periodic function and P a finite linear combination of complex exponentials with
f
P  < everywhere. Choose n sufficiently large that Am P
R
Pm 1
1
k
P dx < for all m > n, where Am g(x) = m
k=0 g( (x)). Then
for m > n,
Z
Z
Z
f (x)dx Am f (x) Am P (x) + Am P (x)
P (x)dx +
(P (x)
< + + = 3.
(b) Unique ergodicity follows by the same logic as the 1variable case.
Let be any invariant measure;Rthen part (a) plus the Mean Ergodic
Theorem shows that P (f ) = f dx, where P is the projection in
L2 () onto the subspace of invariant functions. This implies that the
image of P is just the constant functions. R But we know that the
2
L
of f onto the constants is f d, so we must have
R () projection
R
f dx = f d for continuous f . Since characteristic functions of
open rectangles can be L2 approximated by continuous functions, this
implies that m(R) = (R) for any open rectangle R. But this implies
that m and agree on the Borel sets. Hence m is uniquely ergodic for
this .
Exercise 26: There is an L2 version of the maximal ergodic theorem. Suppose is a measurepreserving transformation on (X, ). Here we do not
assume that (X) < 1. Then
f (x) = sup
m 1
1 X
f ( k (x))
m
k=0
satisfies
kf kL2 (X) ckf kL2 (X) ,
whenever f 2 L2 (X).
The proof is the same as outlined in Problem 6, Chapter 5 for the maximal
function on Rd . With this, extend the pointwise ergodic theorem to the
case where (X) = 1, as follows:
Pm 1
1
k
(a) Show that limm!1 m
k=0 f ( (x)) converges for a.e. x to P (f )(x)
2
for every f 2 L (X), because this holds for a dense subspace of L2 (X).
(b) Prove that the conclusion holds for every f 2 L1 (X), because it holds
for the dense subspace L1 (X) \ L2 (X).
f (x))dx
97
Solution.
(a) We use the subspaces S = {f 2 L2 : f = f } and S1 = {g T g :
g 2 L2 } from the proof of the mean ergodic theorem. As shown there,
L2 (X) = S S1 . Given f 2 L2 , let > 0 and write f = f0 + f1 + f2
where f0 2 S, f1 + f2 2 S1 , f1 2 S1 , kf2 k < . Since f1 2 S1 ,
f1 = g T g for some g 2 L2 . Let h = f0 + f1 . Then Am f0 = f0 = P f0
for all m, and
m 1
1 X k
Am f1 =
T (g
m
T g) =
k=0
1
(g
m
T m g).
Clearly
! 0 for all x as m ! 1. Moreover, as shown on page
1
301, m
T g(x) ! 0 for almost all x; one can see this from the fact
that, by the monotone convergence theorem,
Z X
Z
1
1
1
1
X
X
X
1
1
1
1
m
2
m
2
m
2
T
(g)(x)
=
T
(g)(x)
=
kT
gk
=
kgk
< 1.
2
2
2
m X
m
m2
X m=1 m
m=1
m=1
m=1
P 1 m
2
Since
m2 T (g)(x) is integrable, it is finite almost everywhere,
which means the terms in the series tend to zero for almost all x. The
upshot is that Am f1 (x) ! P f1 (x) for a.a. x, so that Am h(x) ! P h(x)
a.e. Finally, let
n
o
E = x 2 X : lim sup Am f (x) P f (x) > .
1
m g(x)
m
m!1
Since Am f P f = Am h P h+Am (f h) P (f h), E N [F [G ,
where N = {x : Am h(x) 6! P h(x)},
n
o
F = x 2 X : (f h) (x) >
,
2
and
n
o
G = x 2 X : P (f g)(x) >
.
2
Now (N ) = 0 as we have already established, and by the L2 maximal
2 2
k(f h) k2
theorem, (F ) < (/2)2 2 < 4c2 ; similarly, since kP yk kyk for
kf
hk2
y 2 L2 , (G ) < (/2)22 < 4
2 . Since is arbitrary, (F ) = (G ) =
0 for all > 0. Thus, (E ) = 0 for all > 0, so Am f (x) ! P f (x)
a.e.
(b) Let f 2 L1 . For any > 0, choose g 2 L2 \ L1 with kf gk1 < .
Then
Am f (x)
P f (x) = Am g(x)
P g(x) + Am (f
g)(x)
P (f
g)(x).
Let
E = {x 2 X : lim sup Am f (x)
P f (x) > 2} .
Then E N [ F [ G where N = {x : Am g(x)
and
F = {x 2 X : lim sup Am (f
g)(x) > } ,
G = {x 2 X : lim sup P (f
g)(x) > } .
P g(x) 6! 0},
98
Now (N ) = 0 by part (a), and by inequality (24) on page 297,
(F )
Also kP (f
g)k kf
(G )
A
kf
gk1 <
A
.
gk < , so by Chebyshevs inequality,
1
kf
gk1 <
1
.
Since is arbitrary, this implies (F ) = (G ) = (E ) = 0 for all
> 0. Hence Am f (x) ! P f (x) for a.a. x.
Exercise 27: We saw that if kfn kL2 1, then fnn(x) ! 0 as n ! 1 for a.e.
x. However, show that the analogue where one replaces the L2 norm by the
L1 norm fails, by constructing a sequence {fn }, fn 2 L1 (X), kfn kL1 1,
but with lim sup fnn(x) = 1 for a.e. x.
Solution. This is yet another example of why Stein & Shakarchi sucks. The
problem doesnt say anything about conditions on X. In fact, the hint seems
to assume that X = [0, 1]. I will assume that X is finite. I will also assume
that the measure has the property that for any measurable set E and any
real number with 0 (E), there is a subset S E with (S) = .
(This property holds, for example, in the case of Lebesgue measure, or any
measure which is absolutely continuous with respect to Lebesgue measure.
In fact, we dont need quite this stringent a requirementit isnt necessary
that every subset of X have this nice property, but only that we can find a
nested sequence of subsets of Xn whose measures we can control this way,
where X = [Xn and (Xn ) < 1.)
Given these assumptions, let X = [Xn where (Xn ) < 1. We construct
a sequence En of measurable subsets with the properties that (En )
1
n log n and that every x 2 X is in infinitely many En . To do this, we
will construct countably many finite sequences and then string them all
together. The first sequence E2 , . . . , EN1 will have the property that X1 =
1
[N
j=2 Ej , and (Ej ) j log j . To do this, let E2 be any subset of X1 with
1
1
measure 2 log
2 , unless (X1 ) 2 log 2 , in which case E2 = X1 . Let E3 be
1
1
any subset of X1 \ E2 with measure 3 log
3 , unless (X1 \ E2 ) 3 log 3 , in
which case E3 = X1 \E2 . Let E4 be a subset of X1 \(E2 [E3 ) with measure
1
1
at most 4 log
4 log 4 , or X1 \ (E2 [ E3 ) if this has measure P
4 . This process
1
will terminate in finitely many steps because
diverges
and (X1 )
n log n
is finite.
We then construct a second finite sequence of sets EN1 +1 , . . . , EN2 whose
union is X1 [ X2 , a third finite sequence whose union is X1 [ X2 [ X3 , etc.
Let En be the concatenation of all these finite sequences. Then every point
1
in X is in infinitely many En , and (En ) n log
n.
Now let fn = n log n En . Then kfn k1 = n log n(En ) 1. However,
fn (x)
= log n En (x) and since x is in infinitely many En , lim sup fnn(x) = 1
n
for all x.
99
Exercise 28: We know by the BorelCantelli lemma that P
if {En } is a collec1
tion of measurable sets in a measure space (X, ) and n=1 (En ) < 1,
then E = lim sup{En } has measure zero.
In the opposite direction, if is a mixing measurepreserving
transforP1
mation on X with (X) = 1, then whenever n=1 (En ) = 1, there are
integers m = mn so that if En0 = mn (En ), then lim sup(En0 ) = X except
for a set of measure 0.
P
Solution. Let Fn = Enc . Since
(En ) = 1, a theorem on infinite products (e.g. Corollary Q
5.6 on page 166 of Conway, Functions of One Complex
Variable) says that (Fn ) = 0. Let Fn0 = F1 . Because is mixing, there
exists m2 such that
1
( m2 (F2 ) \ F1 ) < (F1 )(F2 ) + 2 .
2
n
and let F20 = m (F2 ). Next, choose m3 such that
1
( m3 (F3 ) \ (F10 \ F20 )) < (F10 \ F20 )(F3 ) + 3
2
and
1
( m3 (F3 ) \ F20 ) < (F20 )(F3 ) + 3
2
and let F30 = m3 (F3 ). Note that this implies
1
1
1
1
0
0
0
(F1 \F2 \F3 ) < (F3 ) (F1 )(F2 ) + 2 + 4 < (F1 )(F2 )(F3 )+ 2 (F3 )+ 3
2
2
2
2
and
(F20 \ F30 ) < (F2 )(F3 ) +
1
.
23
Continuing, we choose mk such that
0
1
0
1
k\1
k\1
1
@ mk (Fk ) \
Fj0 A < (Fk ) @
Fj0 A + k
2
j=`
for all ` = 1, . . . , k
k
Y
j=`
j=`
1; by induction this is less than
(Fj ) +
k
k
X
1 Y
(Fi )
2j i=j+1
j=`+1
where the latter product is taken to be 1 if empty (i.e. when j = k). I
will
Q1 prove that this tends to 0 as k ! 1, for any fixed `. As noted before,
of the product
j=` (Fj ) = 0 (any tail of the sum diverges, so any tail
Q1
tends to zero) so the first term tends to zero. Similarly, j=i (Fj ) = 0 for
any fixed i. Thus, if we split the sum into a sum from ` + 1 to M and a
sum from M to k, the finitely many terms from ` + 1 to M can all be made
arbitrarily small because they each approach zero as k ! 1. On the other
Qk
hand, each term 21j i=j+1 (Fi ) is at most 21j , so their sum is less than
P1
1
j=M 2j which can be made arbitrarily small by an appropriate choice of
M . This proves that
0
1
1
\
@ Fj0 A = 0
j=`
100
for all `. Then
1 \
1
[
`=1 j=`
Fj0 A = 0.
But the complement of this set is just lim sup Ej0 where Ej0 =
Thus, lim sup Ej0 is almost all of X.
mj
(Ej ).
Chapter 6.8, Page 319
Problem 1: Suppose is a C 1 bijection of an open set O in Rd with another
open set O0 in Rd .
(a) If E is a measurable
subset of O, then (E) is also measurable.
R
0
(b) m(
(E))
=

det
(x)dx, where 0 is the Jacobian of .
E
R
R
(c) O0 f (y)dy = O f ( (x)) det 0 (x)dx whenever f is integrable on O0 .
Solution.
(a) If K is compact, then (K) is also compact by continuity. Now if A
is F , then A is compact (since every closed set in Rn is a countable
union of compact sets), so (A) is also compact and hence F . Thus,
maps F sets to F sets. Since measurable sets are precisely those
that dier from F sets by a set of measure zero, it suffices to show that
maps sets of measure zero to sets of measure zero. (The following
argument is adopted from Rudin p. 153.) Let E O have measure
zero. For each integer n, define
Fn = {x 2 E : 
(x) < n}.
(x)
Then for any x 2 Fn , the definition of derivative implies that  (y)
<
y x
n for x y < for some > 0. Define Fn,p Fn to be the set of
x for which = p1 works. Now m(Fn,p ) = 0 because its a subset of
E. I claim that we can cover Fn,p by balls of radius less than p1 with
centers in Fn,p and total measure at most . To do this, we can first
cover Fn,p by an open set of arbitrarily small measure. This open set
can be decomposed into cubes of arbitrarily small diameter, as shown
in chapter 1. If the diameter is sufficiently small (less than a constant
times p1 ), we can cover whichever of these cubes intersect Fn,p with
a ball centered at a point of Fn,p and radius less than p1 ; the other
cubes we discard. The total measure of the resulting balls is at most
a constant times the measure of the open set. (This constant comes
from finding the maximal possible ratio of volumes of the ball covering
one of these smalldiameter cubes to the cube, which comes when the
center p
of the ball is at one corner of the cube and the radius of the
ball is 2 times the cubes diameter.) This proves the claim. Now if
we cover Fn,p by such cubes Bj , centered at xj and with radius rj < p1
then for x 2 Bj we have T (x) T (xj ) nx xj  nrj . This implies
that
[
X
X
T (Fn,p )
Bnrj (xj ) ) m(T (Fn,p )
m(Bnrj (xj )) = nd
m(Brj (xj )) < nd .
j
101
Since was arbitrary, this shows that m(T (Fn,p )) = 0. Since E =
[n,p Fn,p , this implies m( (E)) = 0.
(b) I will prove this in the case where E is a closed rectangle contained in
O, and then show that this implies the general case.
Thus, assume E R U for a compact rectangle R. By absolute
continuity of each of the n components of 0 , given any > 0 there
exists > 0 such that x y < , for x, y 2 R, implies ( 0 (x)
0
(y))j  < for j = 1, . . . , n. Divide R into cubes Qk of diameter less
than . Let ak be the center of Qk . Then for x 2 Qk , the mean value
theorem implies (x)j
(ak )j = 0 (c)(x ak ) for some c on the line
0
segment connecting ak to k. Then c 2 Qk so  0 (c)
(ak ) < .
0
Thus,  (x)j
(ak )j
(ak )j (x ak )j  < x ak j . This then
0
implies k (x)
(ak )
(ak )(x ak )k < kx ak k since a vector
that is larger in each component has larger norm. This local statement
translates into the global statement
(ak ) + (1
(ak )(Qk
ak ) (Qk ) (ak ) + (1 + )
(ak )(Qk
ak )
which may be verified by checking it for each x 2 Qk , since x must
lie in the space between the cubes (ak ) + (1 ) 0 (ak )(Qk ak )
and (ak ) + (1 + ) 0 (ak )(Qk ak ). Now is a bijection so (Qk )
are almost disjoint, which means their images are also almost disjoint
since maps sets of measure
zero to sets of measure zero by part (a).
P
Thus, m( ([Qk )) = m( (Qk )) and
X
X
X
m ( (ak ) + (1 ) 0 (ak )(Qk ak ))
m( (Qk ))
m ( (ak ) + (1 + ) 0 (ak )(Qk
Now by Exercise 4 of Chapter 2,
(1
m ( (ak ) + (1 )
since
(ak )(Qk
ak )) = (1 ) det
(ak )m(Qk )
is linear. Thus,
X
X
)
 det 0 (ak )m(Qk )
m( (Qk )) (1 + )
 det 0 (ak )m(Qk ).
R
P
But
 det 0 (ak )m(Qk ) is a Riemann sum for R  det 0 (x)dx, and
Riemann integration works because 0 is continuous on the compact
set R. Thus,
Z
Z
(1 )
 det( 0 (x))dx m( (R)) (1 + )
 det( 0 (x))dx
0
for all , so m( (R)) = R  det( 0 (x))dx.
Now let E be any measurable subset of O. Let Un O be open
sets containing E with m(Un \ E) < n1 . Then \Un = E [ N where
m(N ) = 0. Then
(E [ N ) = (\Un ) = \ (Un ).
(In general it is only true that (\Un ) \ (Un ), but here we have
equality because is bijective.) Since m( (N )) = 0,
m( (E)) m( (E [ N )) m( (E)) + m( (N )) ) m( (E [ N )) = m( (E)).
ak )) .
102
Now
(Un ) are nested sets of finite measure, so
Z
Z
Z
m(\ (Un )) = lim m( (Un )) = lim
 det 0  =
 0 =
\Un
Un
E[N
=
.
R
0
Thus, we have (finally) that
R m( (E)) =R E  .
(c) We proved in part (b) that O0 g(y)dy = O g( (x)) det 0 (x)dx holds
if g is a characteristic function E for measurablePE O. By linearity,
this extends to a measurable simple function
cj Ej . Now for f
nonnegative, take a sequence fn % f of simple functions; then by the
Monotone Convergence Theorem,
Z
Z
f (y)dy =
lim fn (y)dy
O0
O0
Z
= lim
fn (y)dy
0
ZO
= lim
fn ( (x)) det 0 (x)dx
O
Z
=
lim fn ( (x)) det 0 (x)dx
O
Z
=
f ( (x)) det 0 (x)dx.
O
Finally, we can extend to complex integrable f by linearity, since f is
a linear combination of four nonnegative integrable functions.
Problem 2: Show as a consequence of the previous problem: the measure
d = dxdy
in the upper halfplane is preserved by any fractional linear
y2
a b
transformation z 7! az+b
,
where
belongs to SL2 (R).
cz+d
c d
Solution. We first note that such a transformation does in fact map the
upper half plane to itself: if z = x + iy, then
az + b
a(x + iy) + b
ac(x2 + y 2 ) + (ad + bc)x + bd
(ad bc)y
=
=
+i
2
2
cz + d
c(x + iy) + d
(cx + d) + (cy)
(cx + d)2 + (cy)2
and since y > 0 and ad bc = 1 this has positive imaginary part. Now if
0 0
we write the map z 7! az+b
cz+d in terms of its components as (x, y) 7! (x , y ),
0
0
then using the ugly formulas for x and y from the above expression, we can
compute the even uglier partial derivatives and the Jacobian. However, we
can shortcut that by using the fact that the Jacobian is always the square
norm of the complex derivative. In case this needs proof, suppose z 7! f (z)
is a complex dierentiable function. Then if f 0 (z0 ) = + i, the linear
map z 7! f 0 (z0 )(z z0 ) can be rewritten as
(x+iy) 7! (+ i)((x x0 )+(y y0 )) = ((x x0 )
or
7!
y
(y y0 ))+i( (x x0 )+(y y0 )),
x0
y0
103
which has Jacobian 2 +
f 0 (z) =
If we let
then
= f 0 (z0 )2 . Now in our case,
(ad bc)z
1
=
.
2
(cz + d)
(cz + d)2
denote the mapping z 7! z 0 (and equivalently (x, y) 7! (x0 , y 0 )),
Z
1
dx0 dy 0
0 2
(E) (y )
Z
1
=
 det 0 dxdy
0 2
E (y )
Z
((cx + d)2 + (cy)2 )2
=
(c(x + iy) + d)2 2 dxdy
y2
E
Z
((cx + d)2 + (cy)2 )2
=
((cx + d)2 + (cy)2 )2 dxdy
y2
E
Z
1
=
dxdy
2
y
E
= (E).
( (E)) =
Problem 3: Let S be a hypersurface in Rd = Rd
S = {(x, y) 2 R
d 1
R, given by
R : y = F (x)},
with F a C 1 function defined on an open set Rd 1 . For each subset E
for the corresponding subset of S given by E
= {(x, F (x) :
we write E
x 2 E}. We note that the Borel sets of S can be defined in terms of the
metric on S (which is the restriction of the Euclidean metric on Rd ). Thus
is a Borel subset of S.
if E is a Borel set in , then E
(a) Let be the Borel measure on S given by
Z p
(E) =
1 + rF 2 dx.
E
= {(x, y) 2 Rd , d((x, y), B)
< }. Show that
If B is a ball in , let B
= lim
(B)
!0 2
),
m((B)
where m denotes the ddimensional Lebesgue measure. This result is
analogous to Theorem 4.4 in Chapter 3.
(b) One may apply (a) to the case when S is the (upper) half of the
unit sphere in Rd , given by y = F (x), F (x) = (1 x2 )1/2 , x < 1,
x 2 Rd 1 . Show that in this case d = d , the measure on the sphere
arising in the polar coordinate formula in Section 3.2.
(c) The above conclusion allows one to write an explicit formula for d in
terms of spherical coordinates. Take, for example, the case d = 3, and
write y = cos , x = (x1 , x2 ) = (sin cos , sin sin ) with 0 < 2 ,
0 < 2. Then according to (a) and (b) the element of area d
equals (1 x2 ) 1/2 dx. Use the change of variable theorem in Problem
1 to deduce that in this case d = sin dd . This may be generalized
104
to d dimensions, d
appendix in Book I.
2, to obtain the formulas in Section 2.4 of the
Solution.
(a) We proceed in the steps outlined in Prof. Garnetts hint:
Since B is compact and c is closed, the distance between them
is greater than zero, so we can choose < d(B, c ). Let V =
{x : d(X, B) < } . For each x 2 V define I (x) = {y 2 R :
} and h(x, ) = m(I (x)).
(x, y) 2 B
V R since for (x, y) 2 Rd 1 R, d((x, y), B)
Note that B
d(x, B). By Tonellis theorem,
Z
Z Z
Z
)=
m(B
(y)
=
(y)dydx
=
h(x, )dx.
I (x)
I (x)
R
(x,y)2V R
For x 2 B, let M = rF  evaluated at x. Let ~v be the unit
vector in the direction of rF . Then rF ~v = M at x. By the
continuity of rF , for any > 0 we may choose 0 > 0 such that
rF  < M + and rF ~v > M at all pointspwithin 0 of x.
Suppose < 0 . Then for y 2 R, if F (x) y
1 + (M )2 ,
(M )
consider F (x + t~v ) for 0 t 1+(M )2 . By the construction
of , F (x) + (M )t < F (x + t~v ) < F (x) + (M + )t. Hence,
y(M )
p (M ) 2 < ,
because 1+(M
)2 <
1+(M
x+
y(M )
~v
1 + (M )2
> F (x) +
y(M )2
.
1 + (M )2
By the intermediate value theorem, there is some t0 <
2
y(M )
1+(M )2
y(M )
at which F (x + t0~v ) = F (x) + 1+(M
v . Then
)2 . Let x0 = x + t0~
(x0 , F (x0 )) 2 B; moreover, the distance squared from (x, y) to
(x0 , F (x0 )) is
t20 + y
t2 + (y
y(M )2
1 + (M )2
F (x0 ))2
= t20 +
y2
1 + (M
)2
2
2
y(M )
y2
+
1 + (M )2
1 + (M )2
y2
=
<
1 + (M )2
p
so y 2 Ix, . On the other hand, if y > F (x) +
1 + (M + )2 ,
suppose (x, y) 2 Ix, , so there is some x0 with dist((x, y), (x0 , F (x0 ))) <
. Clearly this implies x0 x < ; suppose x0 x = t. Then because rF  < M + between x and x0 , F (x0 ) < F (x) + (M + )t.
Then the distance squared from (x, y) to (x0 , F (x0 )) is
t2 + (y
(M + )t) = (1 + (M + )2 )t2
2y(M + )t + y 2 .
105
This is a quadratic polynomial in t; the minimum occurs when
the derivative is zero, i.e. when
2t(1 + (M + )2 ) = 2y(M + ) ) t =
y(M + )
.
1 + (M + )2
At this point, the value of the quadratic is
2
y(M + )
y(M + )
y2
2
2
(1 + (M + ) )
2y
+
y
=
> ,
1 + (M + )2
1 + (M + )2
1 + (M + )2
within of (x, y), a contraso in fact there is no point in B
diction. Thus, y 2
/ Ip
.
By
symmetry,
x,
p the same results hold
2
for p
y < F (x), so [ p1 + (M ) ,
1 + (M )2 ] Ix,
2
2
[
1 + (M + ) ,
1 + (M + ) ]. Thus,
p
for
h(x, ) p
< 1 + (M + )2
2
< 0 . This proves that
1 + (M
)2
h(x, ) p
= 1 + rF 2 .
!0
2
lim
Because rF is continuous on the compact set B, it attains a
maximum M on B. For any x 2 B, if y F (x) > (M +1) then
y2
/ Ix, , because any point (x0 , F (x0 )) within of (x, y) would
have to have x0 x < , and then the bound on rF implies
)
F (x0 ) F (x) < M ) y F (x0 ) > . Thus, h(x,
< M +1
2
for all .
Since were interested in the limit of small , we can restrict
our attention toR below some
R cuto value, say a, where a <
d(B, c ). Then V h(x, ) = Va h(x, ) because h(x, ) = 0 for
x 2 Va \ V . This enables us to take all the integrals over the
same region. Moreover, since Va has finite measure, the constant
M + 1 is integrable over Va . So by the dominated convergence
theorem,
Z
1
h(x, )
dx
lim m(B ) = lim
!0 2
!0 V
2
a
Z
h(x, )
=
lim
dx
2
ZVap
=
1 + rF 2 dx
B
= (B).
(The region of integration becomes B in the limit because h(x, )
is eventually 0 for x outside B.)
(b) For two points , 2 S d 1 , let a(, ) denote the angular distance
between and , i.e. the angle between the radius vectors to and .
Similarly, for a point 2 S d 1 and a set E S d 1 , define a(, E) =
inf 2E a(, ). Now given a ball B Rd 1 and corresponding ball
106
Sd
B
, define B 0 = {p 2 S d
[1
B
< arcsin( )}. I claim that
: a(p, B)
B 0 [1
, 1 + ] (B)
, 1 + ].
Here the product is in spherical coordinates, of course. The first inclu [1 , 1 + ], then (, 1) 2 B
and
sion is obvious because if (, r) 2 B
is a distance 1 r away. For the second inclusion, let (, r) be
The distance
any point in Rd . f 2
/ B 0 , let (, 1) be any point in B.
from (, 1) to the line through the origin and (, r) is sin(a(, )) by
elementary trigonometry. By hypothesis this is greater than , so the
distance from (, 1) to (, r) is greater than . On the other hand,
if r 2
/ [1
, 1 + ], then no point in S d 1 is within of (, r). This
proves the second inclusion. Now by the spherical coordinates formulas
derived in section 3,
Z 1+
(1 + )d (1
)d
[1
rd 1 dr =
m(B
, 1 + ]) = (B)
(B),
d
1
so
[1
m(B
, 1 + ])
1 (1 + )d (1
)d
=
(B).
2
d
2
d
Since (1+ ) 2 (1 ) is a dierence quotient for the function f (x) = xd
d d
[1
at x = 1, it approaches dx
x x=1 = d as ! 0, so 21 m(B
,1+
]) ! (B). Similarly,
1
1 (1 + )d (1
)d
m(B 0 [1
, 1 + ]) =
(B 0 ).
2
d
2
As ! 0, this approaches lim !0 (B 0 ) (provided the latter exists,
lim (B 0 ) =
of course). Now since the B 0 are nested and \B 0 = B,
= (B)
by the continuity of measures. (It hardly bears pointing
(B)
out here that (B 0 ) are all finite.) By the Sandwich Theorem,
1
) = lim
(B
!0 2
) = (B 0 ).
m(B
Thus, = on balls, but since these generate the Borel sets, = .
(c) By the change of variable theorem,
dx =
@x1
@
@x2
@
@x1
@
@x2
@
dd =
cos cos
cos sin
Note also that
@F @F
rF =
,
=
@x1 @x2
so
1 + rF 2 = 1 +
x21
x21
x22
x22
x21
sin sin
sin cos
x1
1
x21
x22
x22
dd = sin cos dd .
,p
1
1
x21
x22
x2
x21
=
x22
1
1
=
.
2
2
cos
sin
Then
d =
p
1 + rF 2 dx =
1
sin cos dd = sin dd .
cos
107
In n dimensions, one may use for spherical coordinates the n angles
1 , . . . , n with
y = cos 1
x1 = sin 1 cos 2
x2 = sin 1 sin 2 cos 3
....
..
xn
= sin 1 . . . sin n
cos n
xn = sin 1 . . . sin n .
Then the Jacobian
@(x1 ,...,xn )
@(1 ,...,n )
c1 c2
c1 s2 c3
c1 s2 s3 c4
..
.
s1 s2
s1 c2 c3
s1 c2 s3 c4
..
.
0
s1 s2 s3
s1 s2 c3 c4
..
.
c1 s2 . . . sn
s1 c2 s3 . . . sn
s1 s2 c3 s4 . . . sn
is
0
0
s1 s2 s3 s4
..
.
...
...
...
..
.
s1 s2 s3 c4 s5 . . . sn
...
0
0
0
..
.
s1 . . . sn
1 cn
where ci = cos i and si = sin i . Call this Jn (1 , . . . , n ). Then we
can compute these inductively: a cofactor expansion along the first
row yields
Jn (1 , . . . , n ) = c1 c2 sn1
Jn
1 (2 , 3 , . . . , n )
+ s1 s2 sn2
Jn
1 (1 , 3 , . . . , n ).
Using as our initial case J2 (1 , 2 ) = c1 s1 as computed above, one has
by an easy induction that
Jn (1 , . . . , n ) = c1 sn1
Since 1 + rF  =
2
d = sinn
1
cos2 1
1 sinn
1 n 2 n 3
s2 s3
. . . sn
1.
as before, this yields the area element
2
2 . . . sin n
1 d1
. . . dn .
Note: If the point of this exercise was to calculate the area element on
the unit sphere, it seems that a more direct way is to use the change of
variables formula to compute the volume element in spherical coordinates
(r2 sin drdd in three dimensions), and since the measure of a set E on
the unit sphere is defined to be the measure of the corresponding conical
segment, we can compute it as
Z
Z
(E) = 3
r2 sin drdd =
sin dd ,
[0,1]E 0
E0
where E is the region in the
plane corresponding to the spherical
region E. This implies d = sin dd , with similar formulas in higher
dimensions.
Problem 6: Consider an automorphism A of Td = Rd /Zd , that is, A is a
linear isomorphism of Rd that preserves the lattice Zd . Note that A can
be written as a d d matrix whose entries are integers, with det A = 1.
Define the mapping : Td ! Td by (x) = A(x).
(a) Observe that is a measurepreserving isomorphism of Td .
0
108
(b) Show that is ergodic (in fact, mixing) if and only if A has no eigenvalues of the form e2ip/q , where p and q are integers.
(c) Note that is never uniquely ergodic. (Hint.)
Solution.
(a) Duly noted.
OK, I guess Im supposed to prove it. :) Since A is a linear isomor be the
phism, A 1 is as well. Let E Td be measurable, and let E
d
corresponding subset of the unit cube in R . Then
=  det A 1 m(E)
= m(E)
= (E)
(A 1 E) = m(A 1 (E))
where is the measure on Td induced by the Lebesgue measure m on
Rd .
(b) Suppose that A has an eigenvector of the form e2ip/q ; then AT does
as well, since it has the same characteristic polynomial. Then (AT )q
has 1 as an eigenvector. Since AT I has all rational entries, there is a
(nonzero) eigenvector in Qd , and hence, by scaling, in Zd . Let n 2 Zd
with (AT )q n = n. Consider the function f (x) = e2inx for x 2 Td .
T k
Then since n (Ax) = (AT n) x, T k f (x) = e2i((A ) n)x . The averages
of this function are
m 1
1 X 2i((AT )k n)
Am f (x) =
e
.
m
k=0
But T
f = T f for all k, so Ajq f (x) = Aq f (x) for any integer j.
Since Aq f (x) is not zero (it is a linear combination of exponentials
with distinct periods, since we may assume WLOG that
R q is as small
as possible), the averages do not converge a.e. to Td f (x)dx = 0.
Hence cannot be ergodic.
On the other hand, suppose A has no eigenvector e2ip/q . Let f (x) =
e2inx and g(x) = e2imx for any m, n 2 Zd . Then
Z
T k
hT k f, gi =
e2i((A ) n m)x dx.
k+q
Td
If m = n = 0 the integrand is 1 and the integral is 1 = hf, gi for all k.
If m and n are not both zero, then (AT )k n m) is eventually nonzero;
if not, there would be values k1 and k2 at which it were zero, but then
(AT )k1 n = (AT )k2 n so (AT )k1 n is an eigenvector of (AT )k2 k1 with
eigenvalue 1. (Since A is invertible, this eigenvector is nonzero). This
then implies that A has an eigenvector which is a (k2 k1 )th root of
unity, a contradiction. Thus, (AT )k n m is eventually nonzero, so
hT k f, gi is eventually equal to 0 = hf, gi. Hence T is mixing.
As a side note, the fact that (1 is an eigenvalue of Ak ) implies (some kth
root of 1 is an eigenvalue of A) is not completely trivial. The converse
is trivial, of course, but this direction is not if A is not diagonalizable
(at least not for any reason that Ive found). It follows, however, from
the Jordan canonical form. If Jm ( ) is a Jordan block of size m with
on the diagonal, then Jm ( )k is triangular with k on the diagonal,
so it has k as a kfold eigenvalue. This implies that the algebraic
multiplicity of k in Ak is the same as the algebraic multiplicity of
109
in A; since A and Ak have the same dimension, Ak can have no other
eigenvalues besides kth powers of eigenvalues of A (since the sum of
the multiplicities of the ki is already equal to the dimension of Ak ).
(c) If is the Dirac measure, i.e. (E) = 1 if 0 2 E and 0 otherwise, then
( 1 (E)) = (E) because the linearity of guarantees 0 2 1 (E) )
0 2 E, and the invertibility of guarantees 0 2 E ) 0 2 1 (E).
Problem 8: Let X = [0, 1), (x) = h1/xi, x 6= 0, (0) = 0. Here hxi denotes
dx
the fractional part of x. With the measure d = log1 2 1+x
, we have of course
(X) = 1.
Show that is a measurepreserving transformation.
Solution. Let (a, b) [0, 1). Then
1
1
[
[
1
1
1
1
h i 2 (a, b) , 2
(n + a, n + b) , x 2
,
x
x n=1
n+b n+a
n=1
so
1
X
1
1
,
n+b n+a
n=1
Z 1/(n+a)
1
X
1
dx
=
log
2
1
+x
1/(n+b)
n=1
!
1
1
1 + n+a
1 X
=
log
1
log 2 n=1
1 + n+b
((a, b))) =
1
Y
1
(n + a + 1)(n + b)
=
log
.
log 2
(n + a)(n + b + 1)
n=1
But
1
Y
1+b
(n + a + 1)(n + b)
=
(n + a)(n + b + 1)
1+a
n=1
because the product telescopes; all terms cancel except 1 + b on the top
and 1 + a on the bottom. Hence
1
1+b
1
( ((a, b)) =
log
= ((a, b)).
log 2
1+a
Since is measurepreserving on intervals and these generate the Borel sets,
it is measurepreserving.
Note: By following the hint and telescoping a sum rather than a product,
it is possible to prove is measurepreserving for all Borel sets directly
rather than proving it for intervals and then passing to all Borel sets. Let
E [0, 1) be Borel, let E + k denote the translates of E, and 1/(E + k) =
{1/(x + k) : x 2 E}. Since
1
X 1
1
=
1+x
k+x
k=1
X
1
1
=
,
k+1+x
(k + x)(k + 1 + x)
k=1
110
it follows that
(E) =
1
log 2
1
log 2
=
=
1
log 2
1
log 2
1
dx
1
+
x
E
Z X
1
E k=1
1
XZ
k=1 E
1 Z
X
k=1
1
dx
(x + k)(x + k + 1)
1
dx
(x + k)(x + k + 1)
E+k
1
dx
x(x + 1)
where we have used the monotone convergence theorem to interchange the
sum and the integral. Now if we make the change of variable x = y1 , then
it turns out that
dx
x(x+1)
1
1 X
log 2
k=1
dy
1+y ,
1/(E+k)
so this equals
1
X
1
dy =
y+1
k=1
=
= (
1
[
k=1
1
1
E+k
1
E+k
(E)).
Chapter 7.5, Page 380
Exercise 2: Suppose E1 and E2 are two compact subsets of Rd such that
E1 \ E2 contains at most one point. Show directly from the definition of
the exterior measure that if 0 < d, and E = E1 \ E2 , then
m (E) = m (E1 ) + m (E2 ).
Solution. If E\ E2 = ; then d(E1 , E2 ) > 0 because both are compact, so
for < d(E1 , E2 ), every cover of E1 [ E2 is a disjoint union of a cover
of E1 and a cover of E2 . This implies
H (E) = H (E1 ) + H (E2 )
and taking the limit as ! 0 yields m (E) = m (E1 ) + m (E2 ).
Now suppose E1 \ E2 = {z}. Let F1 = E1 \ B (z) and F2 = E2 \ B (z).
For any cover {Fj } of E, let {Ai } be the collection of those Fj which
intersect F1 , and {Bk } the collection of those that intersect F2 . Note that
. Then {Ai } [ {B (z)}
these collections are disjoint because d(F1 , F2 )
is a cover for E1 , and {Bk } [ {B (z)} is a cover for E2 . Thus
H (E1 ) + H (E2 ) H (E) + 2
Taking limits as ! 0, we have m (E) m (E1 ) + m (E2 ). The reverse
inequality always holds, of course, because m is an outer measure.
Exercise 3: Prove that if f : [0, 1] ! R satisfies a Lipschitz condition of
exponent > 1, then f is a constant.
111
Solution. Suppose f (x) f (y) M x y with
and let h = y x. Then for any integer n,
f (x)
f (y) =
n
X1
j+1
x+
h
n
j+1
x+
h
n
j=0
n
X1
j=0
M
n
j=0
n
X1
= M h n1
> 1. Let 0 x < y 1
j
x+ h
n
j
x+ h
n
This is true for all n, and the bound approaches 0 as n ! 1, so f (x) =
f (y).
Exercise 4: Suppose f : [0, 1] ! [0, 1] [0, 1] is surjective and satisfies a
Lipschitz condition
Prove that
f (x)
f (y) Cx
y .
1/2 directly, without using Theorem 2.2.
Solution. Suppose > 12 . By constructing a lattice of spacing n1 , we can
find n2 points in [0, 1]2 with the property that any two are at least n1
apart. Call these points yk . For each yk , let xk 2 [0, 1] be any point in the
1/
yj yk 
1
preimage. Then for any k 6= j, xj xk 
. But for
C
(Cn)1/
n sufficiently large, (Cn)1/ < n2 , so the n2 points xk must all be farther
than n12 from each other. This is manifestly impossible; by the Pigeonhole
2
Principle, some interval [ nj2 , j+1
1 must contain two of
n2 ] for j = 0, . . . , n
the points.
Exercise 5: Let f (x) = xk be defined on R, where k is a positive integer,
and let E be a Borel subset of R.
(a) Show that if m (E) = 0 for some , then m (f (E)) = 0.
(b) Prove that dim(E) = dim f (E).
Solution.
(a) Since
f (E) = f
1
[
n= 1
E \ [n, n + 1]
1
[
n= 1
f (E \ [n, n + 1]),
it is sufficient to show m (f (E \ [n, n + 1])) = 0. But f is Lipschitz
on [n, n + 1] because
f (x)
f (y) = xk
y k  = x
yxk
+ xk
y + + xy k
+ yk
and the second term is continuous on the compact set [n, n + 1], hence
bounded. By Lemma 2.2, this implies m (f (E \ [n, n + 1])) = 0.
112
(b) Let > dim E. Then m (E) = 0 ) m (f (E)) = 0 so dim f (E).
Hence dim f (E) dim E. To show the reverse, it suffices to show that
m (f (E)) = 0 implies m (E) = 0, since then we can apply exactly
the same logic with E and f (E) interchanged. Let
(
( x)1/k x < 0 and k even
g(x) =
x1/k
else.
Note that g(f (x)) = x, so E g(f (E)) [ g(f (E)). Thus, it will
suffice to prove that g is Lipschitz. Since
1
1
1
1
[
[
[
[
1
1
1
1
R=
[ n 1, n] [
[n, n + 1] [
,
[
,
[ {0},
n
n
+
1
n
+
1
n
n=1
n=1
n=1
n=1
and g is Lipschitz on each of these compact sets (it is C 1 on all but the
last), the result follows. (If we let {Kn } denote all the sets in the above
decomposition, then g is Lipschitz on Kn , so m (g(f (E)\Kn )) = 0 by
Lemma 2.2, and by countable additivity m (g(f (E))) = 0. Similarly
m ( g(f (E)) = 0, so m (E) = 0.)
Exercise 6: Let {Ek } be a sequence of Borel sets in Rd .
dim Ek for some and all k, then
[
dim Ek .
Show that if
Solution. Suppose to the contrary that dim [Ek > . Choose 0 with
< 0 < dim [Ek . Then m0 ([Ek ) = 1 because 0 < dim [Ek . But
0
m
P0 (Ek ) = 0 for each k because > dim Ek , which implies m0 ([Ek )
0
m (Ek ) = 0 by countable subadditivity. This is a contradiction, so
dim [Ek .