Real Analysis - Homework Solutions: Chris Monico, May 2, 2013
Real Analysis - Homework Solutions: Chris Monico, May 2, 2013
Real Analysis - Homework Solutions: Chris Monico, May 2, 2013
1.1 (a) Rings (resp. -rings) are closed under finite (resp. countable) intersections.
(b) If R is a ring (resp. -ring) then R is an algebra (resp. -algebra) iff X R.
(c) If R is a (nonempty) -ring then {E X : E R or E c R} is a -algebra.
(d) If R is a -ring, then {E X : E F R for all F R} is a -algebra.
Solution:
(a) If R is a ring and E1 , E2 R, then since R is closed under differences, E1 (E1 E2 ) R. But
E1 (E1 E2 ) = E1 (E1 E2c ) = E1 (E1c E2 ) = E1 E2 .
It follows inductivelySthat R is closed under finite intersections. Suppose now that R isS a -ring and
fn = A En . Then E
fn R for all n 1 and so F = E1 E
f
{En } R. Let A = Ej and E
j=2 j is also
in R. But
\
\
\
\
\
c
fj = E1 (Ac Ej ) =
E
((E1 Ac ) (E1 Ej )) =
(E1 Ej ) =
Ej .
F = E1
j=2
j=2
j=2
j=2
j=1
(b) If R is a ring (resp. -ring) and X R, then for all E R we have E c = X E R so R is closed
under complements and therefore an algebra (resp. -algebra). On the other hand if R is an algebra (resp.
-algebra), then its nonempty so there exists E R and hence X = E E c R.
(c) Suppose R is a nonempty -ring and M = {E X : E R or E c R}. Clearly M is closed
under \
complements. Suppose that {En } M. By part (a), R is [
closed under countable intersections so
A=
Ejc R. Since R is closed under countable unions, B =
Ej R. Now
j1
Ejc R
j1
Ej R
[
j1
Ej
[
[
=
Ej
E
j
j1
Ejc R
j1
Ej R
c c
\
[
=
Ejc
E
j
j1
Ejc R
j1
Ej R
(A B)c .
S
Since A, B R and R is closed under differences, A B R, hence Ej = (A B)c M. Therefore M
is also closed under countable unions, hence its a -algebra.
(d) Suppose R is a -ring and A = {E X : E F R for all F R}. Suppose E A and let
F R. Then E F R and since R is closed under differences, E c F = F E = F (E S
F ) R, and
therefore
A
is
closed
under
complements.
If
{E
}
A
and
F
R
then
E
R
and
so
(
En ) F =
n
n
S
(En F ) R for all n, so A is closed under countable unions as well and therefore is a -algebra.
1.3 Let M be an infinite -algebra. (a) M contains an infinite sequence of (distinct) disjoint sets.
(b) card(M) c.
Solution:
Note: the word distinct here is not given as part of the problem, but part (a) becomes trivial without it,
and it is extremely helpful in solving part (b) anyway.
The elements of M are partially ordered by inclusion. By the Hausdorff Maximal Principle, there is
a maximal linearly ordered subset {E }A M. By way of contradiction, suppose that A is finite, say
A = {1, 2, . . . , n}. Since E1 E2 En , the -algebra M({E1 , . . . , En }) is clearly finite so there exists
a set E M M({E1 , . . . , En }). We must have E En , otherwise E1 E2 En En E would
contradict the maximality of {E }. Now since
n
n
[
[
E = E En = E E1
(Ej Ej1 ) = (E E1 )
(E (Ej Ej1 )),
j=2
j=2
M
[
Fk .
k1
bk =1
1.5 If M = M(E) then M is the union of all -algebras generated by F as F ranges over all countable
subsets of E. Hint: show that the latter object is a -algebra.
Solution:
[
Let M0 =
M(F). As per the hint, well first show that M0 is a -algebra. Suppose {En } M0 .
F E
countable
S
For each n 1 there is a countable subset Fn E such that En M(Fn ). Then F = Fn is a countable
subset of ESand for all k 1, Fk F implies M(Fn ) M(F), so Ek M(Fk ) M(F). Since M(F) is a
-algebra, Ek M(F). And since F is countable, M(F) M0 .
Now if E M0 then there exists a countable subset F E such that E M(F), in which case
E c M(F) M0 . Therefore, M0 is a -algebra.
For all F E, M(F) M(E), so M0 M(E). On the other hand if E E then E is contained in
M({E}) and {E} is countable, so E M0 and hence E M0 , so that M(E) M0 .
= (E N ) (since E N M )
E + N = E = 0 E (since E M).
((E F ) + (E F ) + (F E)) + (F E)
= (E F ) + (E F ).
1.10 Given a measure space (X, M, ) and E M, define E (A) = (A E) for all A M. Then
E is a measure.
Solution:
Certainly E = ( E) = () = 0. Suppose now that {An }n1 M is a disjoint collection. Then since
{E An }n1 M is a disjoint collection it follows that
[
[
[
X
X
E
An = E
An =
(E An ) =
(E An ) =
E An .
n1
n1
n1
n1
n1
1.14 If is a semifinite measure and E = then for every C > 0 there exists F E with
C < F < .
Solution:
Let C = {F E : F < } and = sup C. By way of contradiction, suppose
that < . For each
Sn
n 1 there is an Fn C such that Fn 1/n. Define Gn = k=1 Fk . Then Gn E and
3
S
Gn Fn 1/n for all n. By continuity of measure,
G
= limn Gn and since Gn
n
n1
for all n 1, we have
lim Gn lim( 1/n) = ,
S
S
so that ( Gn ) = , and hence Gn C. But
[
[
[
= E =
Gn + E
Gn = + E
Gn ,
S
S
so (E S
Gn ) = . Since is semifinite
there exists F 0 (E Gn ) with 0 < F 0 < . Then
S
(F 0 Gn ) = F 0 + > . Since F 0 Gn C, this contradicts the fact that = sup C, and therefore
= .
1.17 If is San outerPmeasure on X and {Aj } is a sequence of disjoint -measurable sets then
(E Aj ) = (E Aj ) for all E X.
Solution:
Let {Aj } be such a sequence and E X. By countable subadditivity,
[
[
X
E
Aj
=
(E Aj )
(E Aj ).
Set Bn =
Sn
j=1
= (E Bn An ) + (E Bn Acn )
= (E An ) + (E Bn1 ).
S
Pn
By induction, (EBn ) = j=1 (EAn ) for all n 1. By monotonicity of outer measure, E
j1 Aj
S
P
P
n
1.18 Let A P(X) be an algebra, 0 a premeasure on A and the induced outer measure.
(a) For all E X and > 0 there exists A A with E A and A E + .
(b) If E < then E is -measurable iff there exists B A with E B and (B E) = 0.
(c) If 0 is -finite, the restriction E < in (b) is superfluous.
Solution:
S
P
Let E X
S and > 0. By definition, there is a sequence {An } A so that E An and 0 An E +.
Let A = An . Then A A and
X
X
A
An =
0 An E + .
For (b), suppose first that E is -measurable with E < . ByTPart (a), for each n 1 there
exists Bn A such that E Bn and Bn E + 1/n. Set B = Bn . Then E B A and
B Bn E + 1/n for all n 1, so B E. But since E B it follows that B = E. Since
E is -measurable,
E = B
(B E) + (B E c )
E + (B E).
Since E < , it follows that (B E) = 0. Conversely, suppose that there exists B A with E B
and (B E) = 0. Then B is measurable since B A and B E is -measurable by the proof of
Caratheodorys Theorem. So (B E)c is -measurable as is
B (B E)c = B (B c E) = B E = E.
Suppose E X is -measurable and > 0. Then each E Xn is measurable with finite measure, and
by Part (a) there exists An,k A such that E Xn An,k and An,k < (E Xn ) + k21n . Since E Xn
is measurable,
An,k
(An,k (E Xn )) + (An,k (E Xn )c )
(E Xn ) + (An,k (E Xn )),
and since S
(E Xn ) < , we thus have that (An,k (E Xn )) = An,k (E Xn ) <
c
(Bk E) = (Bk E ) =
(An,k E )
n=1
1
k2n .
Thus,
X
1
1
= .
(An,k E)
(An,k (E Xn ))
n
k2
k
n=1
n=1
n=1
T
Now set B = k=1 Bk . Then E B A and (B E) (Bk E) k1 for all k 1, hence
(B E) = 0. The converse follows from our proof of the corresponding implication in Part (b) (which did
not make use of the E < assumption).
An
(by countable subadditivity)
X
=
An
(since = |M ).
It follows that E + E.
S
Suppose now that E = + E. Then for each k 1 there exists {An,k } M with E n=1 An,k and
[
X
1
An,k + E + .
An,k
k
n1
Then with A =
k1
S
n1
A
, we have A M , E A, and A + E + 1/k for all k, hence
n,k
n1
1.23 Let A be the collection of finite unions of sets of the form (a, b] Q, where a < b (if
b = , then it is understood that (a, b] = (a, )).
(a) A is an algebra on Q.
(b) The -algebra generated by A is P(Q).
(c) Define 0 on A by 0 = 0 and 0 A = for A 6= . Then 0 is a premeasure on A and there
is more than one measure on P(Q) whose restriction to A is 0 .
Solution:
As per the hint for (a), let E = {} {(a, b] Q : a < b }. Clearly E is closed under intersections
and for =
6 A E we have that A = (a, b] Q for some a < b. The complement of A is taken with respect
to Q (since the claim is that A is a -algebra on Q), and so Ac = (, a] (b, ] is a disjoint union of
members of E. Therefore E is an elementary family and so the collection of finite disjoint unions of members
of E is an algebra by Proposition 1.7. But every finite union of sets of the form (a, b] can be written as a
disjoint union of such sets, and hence this algebraTis precisely A.
For (b), suppose that r Q. Then {r} = n1 ((r 1/n, r] Q) S M(A). If X Q, then X is
countable, say X = {x1 , x2 , . . . }. But {xj } M(A) for all j 1, so X = j1 {xj } is in M(A). Therefore
M(A) = P(Q).
Its clear that 0 is a premeasure on A. The induced outer measure is
X
[
E = inf
0 En : {En } A, E
En
n1
, if E 6=
=
0,
otherwise.
It follows by Theorem 1.14 that = is a measure on P(Q) which extends 0 .
If is the counting measure on P(Q), then = 0 = 0 . And for a < b, E = (a, b] Q has infinite
cardinality, so E = = 0 E, and therefore |E = 0 |E . By finite additivity of , it follows that E =
= 0 E for all 6= E A, so that |A = 0 . But {1} = 1 < = {1}, so is a different measure on
P(Q) which extends 0 .
[
[
[
[
E=
Ej =
(Vj Njc ) =
Vj
Nj .
jZ
jZ
jZ
jZ
Therefore
jZ
Vj0 =
jZ
jZ
1.28 Let F be increasing and right-continuous, and let F be the associated measure. Then F {a} =
F (a) F (a), F [a, b) = F (b) F (a), F [a, b] = F (b) F (a), and F (a, b) = F (b) F (a).
Solution:
Let a, b R with a < b. Since F is increasing and F (a) < , the limit F (a) exists and limn F (a1/n) =
F (a). Since F (a 1, a] < , we have by continuity of measure that
\
F {a} = F
(a 1/n, a] = lim F (a 1/n, a] = lim (F (a) F (a 1/n)) = F (a) F (a).
n
n1
The remaining properties are easily shown using this; for example,
F [a, b) = F (a, b] + F {a} F (b) = F (b) F (a) + F (a) F (a) F (b) + F (b) = F (b) F (a).
rR
rR
and therefore mE = 0.
For (b), suppose that E [0, 1] is a subset with the property that every subset of E is measurable. Then
for each r R, the set E Nr is measurable. By the translation invariance and finite additivity of Lebesgue
measure, the set E1r N is therefore measurable, and hence must have measure zero by Part (a). Since
the collection {Nr }rR is disjoint we have that
!
[
X
X
mE = m
(E Nr ) =
m(E Nr ) =
m(E1r N ) = 0.
rR
rR
rR
1.30 If E L and mE > 0 then for every < 1 there is an open interval I such that m(E I) > mI.
Solution:
Its clear if 0, so assume that 0 < < 1. Suppose first that E L with
mE < . By way of
. Then there is
contradiction suppose that m(E I) mI for every open interval I. Let 0, (1)mE
S
P
a disjoint collection {Ik }k1 of open intervals such that E Ik and
mI)k mE + . It follows that
[
[
mE = m E
Ik
= m
(E Ik )
X
=
m(E Ik )
X
mIk
X
=
mIk (mE + ) < mE + (1 )mE = mE,
a contradiction.
Now if mE = then there is a k Z such that E 0 = E (k, k + 1] has positive measure. So for each
< 1 there is an open interval I such that m(E 0 I) > mI, and since m(E I) m(E 0 I), were done.
= E (x0 , x0 ]
E2
= E (x0 , x0 + ).
where B1/m (0) is the open ball of radius 1/m centered at zero. Since fn fN is measurable, En,N,m is
measurable for all n, N, m 1, and hence
\ [ \
En,N,m ,
E=
m1 N 1 nN
is measurable. Furthermore, x E iff m 1 N 1 such that |fn (x) fN (x)| < 1/m for all n N , so
E = {x X : lim fn (x) exists }.
so the -algebra generated by {(r, ] : r Q} contains {(a, ] : a R}, and hence it contains BR .
Therefore f is measurable by Proposition 2.1.
so f is measurable.
2.9 Let f : [0, 1] [0, 1] be the Cantor-Lebesgue function from 1.5 and let g(x) = f (x) + x.
a. g is a bijection from [0, 1] to [0, 2] and h = g 1 is continuous from [0, 2] to [0, 1].
b. If C is the Cantor set, m(g(C)) = 1.
c. By Exercise 1.29, g(C) contains a Lebesgue nonmeasurable set A. Let B = g 1 (A). Then B
is Lebesgue measurable but not Borel.
Solution:
(a) Since f is increasing and a(x) = x is strictly increasing, it follows that g is strictly increasing, hence
it is one-to-one. Since f is continuous, so is g, and so its clearly onto, hence bijective. Since g is strictly
increasing and continuous, for 0 a < b 1 we have
h1 ((a, b)) = g((a, b)) = (g(a), g(b)).
It follows that for every open O [0, 1], h1 (O) is open in [0, 2], so h is continuous.
S
1] so that C = [0, 1] k1 Ik and
P (b) There is a collection {I
Sk } of disjoint closed intervals
S in [0, S
mIk = 1. Then [0, 1] C = Ik , and so g([0, 1] C) = g ( Ik ) = g(Ik ). Since g is one-to-one, {g(Ik )}
is a disjoint collection so
X
2 = m(g([0, 1])) = m(g(C)) +
m(g(Ik )).
(0.1)
If Ik = [ak , bk ] then f is constant on Ik by construction, so g(x) = f (x) + x = f (ak ) + x for all x Ik , and
thus g(Ik ) = Ik + f (ak ). Therefore m(g(Ik )) = mIk and we have from (0.1) that
X
2 = m(g(C)) +
mIk = m(g(C)) + 1,
hence m(g(C)) = 1.
(c) Since B = g 1 (A) g 1 (g(C)) = C and mC = 0, it follows that B is Lebesgue measurable. Since h :
[0, 2] [0, 1] is continuous, it is measurable. If B [0, 1] were Borel then h1 (B) = g(B) = g(g 1 (A)) = A
would be measurable, so B cannot be Borel.
R
2.12 Prove Prop. 2.20: If f L+ and f < then {x : f (x) = } is a null set and {x : f (x) > 0}
is a -finite set.
Solution:
Let E = {x : f (x) = }. Then E is measurable and for all n 1, n = nE is a simple function with
0 n f , so
Z
Z
f
n = nE.
R
R
Thus, E n1 f for all n 1. Since 0 f < it follows that E = 0.
S
Now let F = {x : f (x) > 0} and Fn = {x : f (x) > 1/n}. Then F = n1 Fn . For each n,
R
R
n = (1/n)F is a simple function with 0 n f , so f n = n1 Fn , from which it follows that
n
R
Fn n f < for all n, so F indeed is a -finite set.
R
R
R
R
2.13 Suppose {fn } L+ , fn f pointwise, and
f = lim
fn < . Then E f = lim E fn for all
R
R
E M. However, this need not be true if f = lim fn = .
Solution:
Let A M and note that f A f so
Z
R
A
Z
f=
Z
(f A + f Ac ) =
10
Z
f+
f.
Ac
All terms above are finite; from this and another application of Fatous Lemma,
Z
Z
Z
Z
Z
lim sup
fn f
f=
f lim inf
fn .
Ac
Ac
Ac
Therefore, lim Ac fn exists and equals Ac f , and Rhence lim E fnR = E f for all E M.
To see that this need not be true when lim fRn = = R f , considerR (X, M) = (R, L)Rand fn =
n(0,1/n] + 1[1,). Then fn [1,) on R and lim fn = = f . But lim [0,1] fn = 1 6= 0 = [0,1] f .
R
+
+
2.14 If
R f L R let E = E f d for E M. Then is a measure on M and for every g L ,
g d = f g d. Hint: first suppose that g is simple.
Solution:
R
R
R
First note that = f d = f d = 0 d = 0. Suppose that {En } is a disjoint sequence in M. Then
Z
Z
[
En
=
f d = f S E d
S
n
En
Z X
XZ
(by Thm. 2.15)
=
f E d =
f E d
n
n
Z
X
X
=
f d =
En .
En
Therefore is a measure.
Suppose that E M. Then
Z
Z
E d =
Z
d = E =
Z
f d =
f E d.
R
R
It follows from linearity (for nonnegative functions and nonnegative constants) that d = f d for
every nonnegative simple function . Suppose now that g L+ . By Thm. 2.10, there is a sequence {n } of
increasing, nonnegative simple functions converging pointwise to g. By the Monotone Convergence Theorem
we have
Z
Z
Z
g d = lim n d = lim f n d.
But n n+1 and f 0 implies that f n f n+1 , so {f n } is an
R increasingR sequence of nonnegative
functions tending to f g, so again applying the MCT we find that lim f n d = f g d.
R
R
R
2.15 If {fn } L+ , fn decreases pointwise to f , and f1 < , then f = lim fn .
Solution:
R
R
Since
fRn is a decreasing
sequence converging
to f , we have Rf f1 and consequently
f f1 < . Now
R
R
R
R
R
R f + (f1 fR) = fR1 , and since f < , it follows that (f1 f ) = f1 f . It similarly follows that
(f1 fn ) = f1 fn . Since {f1 fn } is an increasing sequence of nonnegative functions tending to
f1 f , it follows from the MCT that
Z
Z
Z
Z
Z
Z
f1 f = (f1 f ) = lim (f1 fn ) = f1 lim fn .
11
R
+
2.16 If
R f LR and f < then for every > 0 there exists E M such that E < and
f .
f>
E
Solution:
R
R
R
Let > 0. By definition there is a simple function
with 0 f and >
f . Since < ,
Pn
the support of is finite; that is, if = j=1 aj E is the canonical representation, then aj = 0 for at
j
R
P
most one value of j, say a1 = 0. Then > =
aj Ej min{a2 , . . . , an }(E2 + . . . En ), and so
E = E2 En has finite measure. Then
Z
Z
Z
Z
f
f .
= >
E
2.17 Assume Fatous Lemma and deduce the MCT from it.
Solution:
Suppose that {fn } is an increasing sequence of nonnegative functions tending to f . By Fatous Lemma,
Z
Z
Z
f = (lim inf fn ) lim inf fn .
R
R
R
R
On the other hand, since fn f for all n 1, fn f for all n 1 and so lim sup fn f . Thus,
Z
Z
Z
lim sup fn f lim inf fn ,
so lim
f.
2.18 Fatous Lemma remains valid if the hypothesis that fn L+ is replaced by the hypothesis that
fn is measurable and fn g where g L+ L1 .
Solution:
Suppose {fn } is a measurable sequence, g L+ L1 and fn g for all n 1. Set fen = fn + g. Then
fen 0 so by Fatous Lemma
Z
Z
Z
Z
Z
Z
g + lim inf fn = lim inf fen lim inf fen = g + lim inf fn .
Since g L1 ,
fn .
R
R
R
R
2.20 If fn , gn , f, g L1 , fn f , gn g a.e., |fn | gn , and gn g then fn f .
Solution:
As in the proof of Theorem 2.24, begin by taking real and imaginary parts, so that it suffices to consider
the case where all functions are real-valued. Then |fn | gn implies gn + fn 0 and gn fn 0. Since
lim(fn + gn ) = f + g, applying Fatous Lemma in each of these cases 1 we have
Z
Z
Z
Z
Z
Z
g + f = lim(gn + fn ) lim inf (gn + fn ) = g + lim inf fn ,
and
Z
g
Since both
so lim
g and
Z
lim(gn fn ) lim inf
f=
Z
(gn fn ) =
Z
g lim sup
fn .
f.
2.22 Let be the counting measure on N. Interpret Fatous Lemma, MCT, and LDCT as statements
about infinite series.
Solution:
First note that every function f : N C is measurable in this case since each singleton {n} with n N
is measurable and every subset of N is a countable union of singletons. Furthermore, since {n} = 1 for all
n N,
Z
Z
X
|f (n)|.
d =
|f | d = sup
0|f |
1
n=1
Thus, f L iff
f (n) is absolutely convergent.
P
Fatous Lemma in this case translates to: If {ak,n } is a doubly-indexed nonnegative series with n=1 ak,n
absolutely convergent for every k 1, then
n=1
ak,n .
n=1
X
n=1
ak,n .
n=1
X
X
lim ak,n = lim
ak,n .
n=1
n=1
1 Note: Its not generally true that lim inf(a + b ) = lim inf a + lim inf b ; produce a counterexample to convince yourself
k
k
k
k
of this fact. However, we are using in this case the result that if lim ak = a then lim inf(ak + bk ) = a + lim inf bk . If youre
unfamiliar with this, you should prove it.
13
1
f (x rn )
dx = n
2n
2
1+rn
rn
dx
1
= n1 .
2
x rn
By Theorem 2.15,
Z
Z X
X Z f (x rn )
X
f (x rn )
1
|g(x)| dx =
dx =
dx =
= 2,
n1
2n
2n
2
n=1
> + g(x0 ).
2n x0 rn
Then
g(x0 )
1
f (x0 rn )
= n 0
> + g(x0 ),
n
2
2 x rn
Since 0 < < 1 was otherwise arbitrary, g is discontinuous at x0 . On the other hand, if g(x0 ) = and > 0
then there is an x0 (x0 , x0 + ) for which g(x0 ) is finite since g is finite a.e.. Therefore g is discontinuous
on R.
Let I R be a (nondegenerate) interval and rN Q a rational
point in the interior of I. Let (0, 1)
2
1
such that (rN , rN + ) I, and let > 2N . Then for all x rN , rN + 2N1
we have
g(x) =
X
f (x rN )
1
1
f (x rn )
= N
> N 1 = .
n
N
2
2
2
x rN
2
2N
n=1
2
Thus, for all sufficiently large , m g 1 ((, )) 2N1 > 0. This shows that g is unbounded on I and
remains so after any modification on a Lebesgue null set.
(c) If g is finite on E then so is g 2 . Therefore g 2 < a.e. Again, let I R be an interval and rN Q
be a point in the interior of I. Let (0, 1) so that (rN , rN + ) I. Then
Z
I
g2 =
2
Z X
f (x rn )
I
2n
Z
dx
I
f (x rN )
2N
2
rN +
dx
rN
14
dx
1
= N
N
4 (x rN )
4
rN +
rN
dx
= ,
x rN
Rx
2.26 If f L1 (m) and F (x) = f (t) dt then F is continuous on R.
Solution:
Let x0 R and let {xn } be a sequence converging to x0 . Set fn = f (,x ]. Then the sequence {fn }
n
converges pointwise to either f (,x0 ) or f (,x0 ]. In any case |fn | |f | and |f | L1 (m), so by the LDCT
Z
x0
F (x ) =
Z
f (t) dt = lim
xn
f = lim F (xn ).
fn = lim
Since {xn } was an arbitrary sequence converging to x0 , limxx0 F (x) = F (x0 ), so F is continuous at x0 , and
hence continuous on R.
2.27 Let P
fn (x) R= aenax benbx where 0 < a < b.
|fn |
Z
X
1
an
n=1
nax
ae
benbx dx
Z
X
nax
nbx
(ae
be
) dx
1
an
n=1
X
1
1
=
ne neb/a
n=1
1
1 X 1
= b/a
= .
e e
n
n=1
=
=
=
aenax benbx
n=1
a
a
n=1
X
n=1
enax b
1
eax
n
enbx
n=1
X
n=1
1
ebx
a
b
eax 1 ebx 1
a
eax 1
b
ebx 1
0. Therefore,
a
b
gn (x) =
[1/n,)(x) 0,
eax 1 ebx 1
15
1+
x n
sin(x/n) dx =
n
lim
1+
x n
sin(x/n) dx =
n
n(n1)
2n2
1
2
1
2n
1
4
for
0 dx = 0.
0
|fn |
and 1 L1 ([0, 1], m). Since lim fn = {0}, it follows from the LDCT that
Z
lim
c. Let fn =
n sin(x/n)
x(1+x2 ) .
Z
fn dx =
{0} dx = 0.
x
x(1+x2 )
Z
fn dm =
1
1+x2 ,
dx
1
=
tan
x
= .
0
1 + x2
2
0,
1
n
1
= lim
/2,
dx
=
lim
tan
(nx)
tan
(an)
=
a
n
n 2
1 + (nx)2
,
if a > 0,
if a = 0,
if a < 0.
How does this accord with the various convergence theorems? Let
n
0, if x 6= 0,
f (x) = lim
=
, if x = 0.
n 1 + (nx)2
R
Since a f = 0 for all a, it is possible that the convergence theorems may apply for a > 0, but the hypotheses
of MCT and LDCT cannot be satisfied for a 0. Indeed, {fn } is not monotone on any interval containing
zero since {fn (0)} is an increasing sequence while for all x 6= 0, {fn (x)} is eventually decreasing.
16
If a > 0 then |fn | 1/x2 on [a, ), so LDCT may be applied in this case, but there is no function
g L1 ([a, )) that dominates {fn } on [a, ) if a 0.
Fatous Lemma, however, does apply and it supplies us with the conclusion that
Z
Z
f lim inf fn ,
0=
a
R
R
2.29 Show that 0 xn ex dx = n! by differentiating the equation 0 etx dx = 1/t. Similarly show
p
R 2n x2
R
2
that x e
dx = (2n)!
by differentiating the equation etx dx = /t.
4n n!
Solution:
R
We first show by induction that for all n 1 and t [1, 2], n!/tn+1 = 0 xn etx dx. Evaluation at t = 1
then produces the desired result.
f
tx
tx
exists for t [1, 2]. With g(x) = xex we have g L1 ([0, ))
Let
f (x, t) = e . Then t = xe
f
and t g for all (x, t) [0, ) [1, 2]. So, by Theorem 2.27,
1
=
2
t
t
Z
f (x, t) dx =
f
dx =
t
xetx dx.
n!
(n + 1)!
=
xn etx dx.
=
tn+2
t tn+1
t 0
x
e
dx
=
xn+1 etx dx.
tn+2
t
0
0
The second part of the problem is similar.
2
2
a. For a > 0, ex cos ax dx = ea /4 .
R1
P
b. For a > 1, 0 xa (1 x)1 log x dx = k=1 (a + k)2 .
R a1 x
P
(e 1)1 dx = (a)(a), where (a) = n=1 na .
c. For a > 1, 0 x
R ax 1
d. For a > 1, 0 e
x sin x dx = arctan(a1 ).
R ax
P
e. For a > 1, 0 e
J0 (x) dx = (a2 + 1)1/2 , where J0 (x) = n=0 (1)n x2n 4n (n!)2 is the
Bessel function of order 0.
Solution:
a. Using the power series expansion of cos ax we have
e
x2
cos ax =
ex
n=0
(ax)2n
(1)n .
(2n)!
Furthermore,
X
X
X
x2 (ax)2n
2
2
(ax)2n
(ax)k
n
x2
x2
e
(1)
=
e
e
= ex eax = eaxx .
(2n)!
(2n)!
k!
n=0
n=0
k=0
17
Since eaxx L1 (R) it follows from Theorem 2.25 and the result of Exercise 2.29 that
Z X
Z
2n
X
(1)n a2n 2n x2
x2 (ax)
n
x e
dx
e
(1) dx =
(2n)!
(2n)!
n=0
n=0
X
(1)n a2n (2n)!
=
(2n)!
4n n!
n=0
n
X
2
a2
1
=
= ea /4 .
4
n!
n=0
P
a
log x
b. Using the formula for the geometric series, we have for 0 < x < 1 that x 1x
= n=0 xn+a log x =
P n+a
n=0 x
log(1/x). Since each function xn+a log(1/x) is nonnegative on (0, 1), we have by Theorem 2.15
that
Z 1X
Z 1 a
x log x
dx =
xn+a log(1/x) dx
1x
0
0 n=0
Z 1
X
=
xn+a log(1/x) dx.
(0.2)
n=0
For j 1, let gj = xn+a log(1/x)[1/j,1]. Then {gj } is an increasing sequence of functions, so by MCT and
integration by parts,
Z 1
Z
Z
n+a
x
log(1/x) dx =
lim gj dm(x) = lim
gj dm(x)
(0,1] j
(0,1]
1
Z
=
1/j
=
=
xn+a log(1/x) dx
lim
lim
1
xn+a+1
xn+a+1 log x
+
.
n+a+1
(n + a + 1)2 1/j
1
(n + a + 1)2
Note: this type of argument shows that the usual technique for dealing with such improper integrals
works, provided the integrand is nonnegative for almost all values sufficiently close to the limit of integration;
henceforth, I will assume this rather than detail out the steps as Ive done here.
Combining with (0.2), we have
Z 1 a
X
X
x log x
1
1
dx =
=
.
2
1
x
(n
+
a
+
1)
(a
+
k)2
0
n=0
k=1
ex
1 ex
=
xa1 ekx ,
for x > 0.
k=1
=
=
Z
X
k=1
X
k=1
X
k=1
18
u a1
k
1
ka
eu
du
k
ua1 eu du
(a)
= (a)(a).
ka
X
(1)k x2k
k=0
Letting fk =
(1) x
(2k+1)!
Z
X
k=0
2k
(2k + 1)!
eax .
|fk | =
X
k=0
1
(2k + 1)!
2k ax
x e
dx
k=0
k=0
k=0
So by Theorem 2.25,
Z
!
fk
k=0
Z
X
fk
k=0
k=0
k=0
e. Let
fn =
t2k et dt
(2k + 1)
(2k + 1)!a2k+1
1
< ,
(2k + 1)a2k+1
for a > 1.
X
(1)k
x2k eax dx
(2k + 1)! 0
k=0
1
(2k + 1)!a2k+1
(1)k (2k + 1)
(2k + 1)!a2k+1
(1)k
= arctan(a1 ).
(2k + 1)a2k+1
(1)n x2n ax
e
.
4n (n!)2
Then
Z
X
n=0
1
|fn | =
n
4 (n!)2
n=0
2n ax
x e
dx =
=
=
If you are familiar with the well-known bound
inductively, or sloppily by:
2n
n
1
n
4 (n!)2 a2n+1
n=0
X
n=0
t2n et dt
(2n + 1)
4n (n!)2 a2n+1
(2n)!
.
n (n!)2 a2n+1
4
n=0
3
1
...
1.
4
2
P R
(The right-hand side here cannot be taken as 1/2 if n = 0). It thus follows that n=0 0 |fn | < . So by
Theorem 2.25 we have that
Z X X
Z
Z
X
(1)n
fn =
fn =
x2n eax dx
n (n!)2
4
0
0
0
n=0
n=0
(2n)!
(2n)(2n 1)(2n 2)(2n 3) . . . (2)(1)
=
=
n
2
4 (n!)
(2n)(2n)(2n 2)(2n 2) . . . (2)(2)
2n 1
2n
2n 3
2n 2
X
(1)n (2n + 1)
4n (n!)2 a2n+1
n=0
1
.
a2 + 1
(For the last equality, you may either look it up or derive it).
19
= {A B : A E1 , B E3 }
= {(E1 E2 ) E3 : Ej Mj }.
(1 2 )(E1 E2 )3 (E3 )
(1 2 3 )(E1 E2 E3 ).
It follows by countable additivity that ((1 2 ) 3 )(A) = (1 2 3 )(A) for every set A which is a
finite disjoint union of measurable cuboids (rectangular parallelpipeds). The collection A of all such sets A
is an algebra and the -algebra generated by A is 31 Mj ; since the measures (1 2 ) 3 and 1 2 3
are -finite and agree on A, they are equal by the uniqueness assertion in Theorem 1.14.
ZZ
D d d =
D(x, y) d(y) =
R
{x}
ZZ
Z
D(x, y) d(x) d(y) =
1 d(y) = {x} = 1, so
ZZ
D d d =
0 d(y) = 0.
Z
D(x, y) d(y) d(x) =
1 d(x) = 1.
[0,1]
R
Finally, D d( ) = ( )(D) which we will now show is , by arguing on the outer
S measure of
D. SupposeSthat {An Bn } is a countable
collection
of
measurable
rectangles
with
D
(An Bn ).
S
Then D = (D An Bn ), and so (An Bn ) [0, 1]. It follows that there is an N 1 for which
m(AN BN ) > 0. In particular, mAN > 0 and BN is infinite so that BN = . It follows that the outer
measure of D is infinite and since D is measurable, ( )(D) = .
2.49 Prove Theorem 2.39 by using Thm. 2.37 and Prop. 2.12 together with the following lemmas.
a. If E M N and ( )(E) = 0 then (Ex ) = (E y ) = 0 for almost all x, y.
b.
and f = 0 -a.e., then fx and f y are integrable for almost all x, y and
R If f is RL-measurable
y
fx d = f d = 0 for almost all x, y.
Solution:
a. One may prove the stronger result for E M N with ( )(E) = 0 trivially by invoking Theorem
2.36; I think this is what was intended, since its the first thing Im going to do in Part (b) anyway.
20
b. Let E = {(x, y) : f (x, y) 6= 0}. Then E = 0 and since is the completion of , there is a set
e M N such that E
e E and ( )(E)
e = 0. By Theorem 2.36,
E
Z
Z
e
e
e y d(y).
0 = ( )(E) = Ex d(x) = E
ex = 0 a.e. [] and E
e y = 0 a.e. []. Since Ex E
ex it follows by the completeness of
Therefore E
that Ex = 0 a.e. []. Similarly, from the completeness of it follows that E y = 0 a.e. []. Since
{y : fx (y) 6= 0} = Ex and Ex = 0 a.e. [], it follows that for almost all x, fx = 0 a.e.. By RProp. 2.11,
we therefore have that for almost all x, fx is measurable, |fx | = 0Ra.e. so fx is integrable, and fx d = 0.
Similarly, we conclude that for almost all y, f y is integrable, and |f y | d = 0.
Thm. 2.39: Assume the hypotheses of the Theorem and suppose that f is L-measurable and f 0. By
Prop. 2.12, there is an M N -measurable function fe so that fe = f a.e. []. We may further assume that
fe 0 (by possibly modifying fe on a -null set). Set g = f fe. Then g = 0 a.e. [], so by Part (b),
gx , g y are integrable, and in particular measurable, for almost all x, y. fex , fey are measurable by Prop. 2.34,
y
and so
for almost all x, y respectively. Furthermore,
byRTheorem 2.37, the functions
R fx , f are measurable
R
R
e
x 7 fx d and y 7 fey d are measurable. By Part (b), we have gx d = g y d = 0 for almost all x, y
resp., so the functions
Z
Z
Z
x 7
fx d = (fex + gx ) d = fex d,
for almost all x
Z
Z
Z
y 7
f y d = (fey + g y ) d = fey d,
for almost all y
are measurable for almost all x, y respectively. By Theorem 2.37,
ZZ
ZZ
ZZ
Z
f d d =
(fex + gx ) d d =
fex d d = fed( ),
RR
R
R
R
e
and similarly,
f d d = fed( ), so it remains only
R to show that R f d( ) = f d.
If E is M N measurable then ( )(E) = (E), so E d( ) = E d. By linearity, we have the
same result for M N-measurable simple functions. Let {n } be a sequence of simple functions increasing
to fe. Using the Monotone Convergence Theorem twice, we deduce that
Z
Z
Z
Z
fed( ) = lim n d( ) = lim n d = fed,
R
R
and since f = fe a.e. [], it follows that fed = f d.
Tonellis Theorem now follows from Fubinis as in the proof of Theorem 2.37.
21
3.9 Suppose {j } is
Pa sequence of positive measures. If j for all j then (
for all j then ( j ) .
Solution:
j ) . If j
(E Acj ) = 0.
dx
=
x
/2
Z
0
dx
= > 1 = ,
x
By the first part, there is a 2 > 0 responding to the -challenge for uniform integrability of {f1 , . . . , fN 1 },
so = min{1 , 2 } does the job for every fn .
3.16 Suppose , are -finite measures on (X, M) with and let = + . If f = d/ d, then
0 f < 1 -a.e. and d/ d = f /(1 f ).
Solution:
d
d
Since , we have by Prop 3.9 that d
= f d
.
22
Since
d
d
=1+
d
d
=1+f
d
d
and f,
d
d
d
= 1.
d
(0.3)
Extra: Suppose is P
a -finite measure on (X, M) and {En } a sequence of measurable sets. Define on
d
M by E = (E En ). Find the Radon-Nikodym derivative d
.
Solution:
Clearly is -finite and , so the Radon-Nikodym derivative in question exists. Furthermore, for every
E M,
X
E =
(E En )
Z X
XZ
XZ
E d,
E d =
=
EE d =
n
so that
d
d
E .
n
3.25 If E is a Borel set in Rn define DE (x) = limr0 m(E B(r, x))/mB(r, x), if the limit exists.
a. Show that DE (x) = 1 a.e. on E and DE (x) = 0 a.e. on E c .
b. Find examples of E and x s.t. DE (x) is a given number (0, 1) and an example for which
DE (x) does not exist.
Solution:
a. This is a direct application of Thm. 3.22: let E BRn and define A = m(A E) for all Borel sets A
d
(i.e., = m|E ). Then m and dm
= E. Since the family {B(r, x)}r>0 shrinks nicely to x, it follows by
Theorem 3.22 that DE (x) = limr0 B(r, x)/mB(r, x) = E(x) for m-almost all x Rn .
b. Let (0, 1) and E = {(t sin , t cos ) : t > 0, 0 2} R2 , and x = (0, 0). For each r > 0,
m(E B(r, x)) = 2r2 , so DE (x) = .
For an example in which the limit does not exist, consider the subset E of R given by
[ 1
1
1 1
1 1
E=
,
=
,
,
...,
22n+1 22n
8 4
32 16
n1
1
and x = 0. Consider the sequence rk = 2k . Then mB(rk , 0) = 2k1
. Now if k is even, say k = 2k 0 , then
X 1
X 1
1
1
1
4
m(E B(rk , 0)) =
m 2n+1 , 2n =
=
=
,
0
2n+1
k
2
2
2
24
3
3 2k1
0
2nk
nk
and so m(E B(rk , 0))/mB(rk , 0) = 1/3 when k is even. Now for odd k, say k = 2k 0 + 1 we have
X 1
X
1
1
1
4
1
m(E B(rk , 0)) =
m 2n+1 , 2n =
=
=
,
2n+1
k0 +1 3
2
2
2
2
4
3
2k
0
2nk
nk +1
and so m(E B(rk , 0))/mB(rk , 0) = 1/6 when k is odd. Therefore, limk m(E B(rk , 0))/mB(rk , 0) does
not exist and hence DE (0) is undefined.
23
3.26 If and are positive, mutually singular Borel measures on Rn and + is regular then so are
and .
Solution:
For compact K Rn , ( + )K < , and so K < , K < since and are positive.
Let A BRn so that A = Ac = 0. Then for all E BRn ,
E = (E A)
( + )(E A)
,
2j
Xj = j .
j
2
2
Oj < .
f (x) =
0 p
q x
p
q
1
,
q3
X
X
q=1 0pqx
X
X
1
bqxc
1
x 2
=
x
=
,
q3
q3
q2
6
q=1
q=1
a
b
X
0 p
q x
so f is discontinuous at x.
24
q3
X
0 p
q x
1
1
3,
q3
b
P
Suppose x is irrational and > 0. Choose N N so that q=N q12 < . For each b = 1, 2, . . . , N let nb N
x nb = b is minimal. Since x is irrational, each b is positive. Let = min{1, 1 , 2 , . . . , N }.
for which
b
Then x pq < q > N , so for 0 < y < ,
0 f (x) f (x ) =
X
x< p
q x
X
1
<
q3
X
X
X
1
1 + q
1
1
=
+
< + 2.
q3
q3
q3
q2
q=N qxq<pqx
q=N
q=N
q=N
Rb
3.33 If F is increasing on R then F (b) F (a) a F 0 (t) dt.
Solution:
Let < a < b < and assume WOLG that F is NBV by properly redefining it outside the interval [a, b].
Let F be the Borel measure associated with F and
m, m,
F = + ,
its Lebesgue-Radon-Nikodym decomposition. Both and are regular measures, so by Theorem 1.16 there
exist right-continuous increasing functions G, H such that
R x = G , =R xH . Both G, H N BV , so by
Proposition 3.30, G0 = 0 a.e. and for x [a, b], H(x) = H 0 (t) dt = a H 0 (t) dt. But F = G + H =
G+H , so F = G + H a.e. and since G0 and H 0 exist a.e., it follows that F 0 = G0 + H 0 = H 0 a.e., so
Z b
Z b
0
F (b) F (a) = F (a, b] = G (a, b] + H (a, b] H (a, b] =
H (t) dt =
F 0 (t) dt.
a
3.37 Suppose F : R C. There is a constant M s.t. |F (x) F (y)| M |x y| for all x, y R iff F
is AC and |F 0 | M a.e..
Solution:
Suppose first that such a constant M > 0 exists. Let > 0 and set = /M . Suppose {(ai , bi )}N
i=1 is a
PN
disjoint collection of open intervals for which i=1 (bi ai ) < . Then
N
X
|F (bi ) F (ai )| M
N
X
(bi ai ) < M = ,
i=1
i=1
P
3.39 If {Fj } is a sequence of nonnegative
Fj (x) <
P 0 increasing functions on [a, b] so that F (x) =
0
for all x [a, b], then F (x) = Fj (x) for almost all x [a, b].
Solution:
As per the hint, it suffices to assume (by properly extending the Fj s and F to R) that the Fj s and F are
NBV. Let F and Fj be the associated Borel measures. These are positive -finite Borel measures, so they
admit Lebesgue-Radon-Nikodym decompositions:
dF = d + f dm,
P
where all of these measures are positive and m and j m. Since F and
Fj agree on h-intervals,
they are equal, and so
X
X
X
dF =
dFj =
dj +
fj dm .
P
P
Furthermore, since ( dj ) m and ( fj dm) m by Exercise
3.8, it follows
P
Pfrom the uniqueness
assertion
in
the
Lebesgue-Radon-Nikodym
Theorem
that
d
=
d
and
f
dm
=
fj dm; in particular,
j
P
f=
fj m-almost everywhere. Thus, with Er = (x, x + r] we have by Theorem 3.22 for m-almost all x,
F 0 (x) = lim
r0
F Er
F (x + r) F (x)
= lim
= f (x).
r0 mEr
r
fj (x) and another application of Theorem 3.22 yields for almost all x that
X
X
X
F Er
f (x) =
fj (x) =
lim j
=
Fj0 (x).
r0 mEr
(x, B) (x, A)
(x, B) + (x, A)
+ (x, A) =
< (x, B).
2
2
The same argument applied to V shows that V is open. By construction, we have that U V = and
A U , B V , so indeed the space is normal.
1
my
<
ry
2
1
my , xny
26
Rb
a
Let > 0. Since polynomial functions are dense in C([a, b], R), there is a polynomial p(x) such that
Rb
kf pku < . From the hypothesis, it follows by linearity that a p(x)f (x) dx = 0, so
Z
p(x)f (x) dx =
0=
f (x)2 dx,
and so
Z
Z
Z
Z b
b
b
b
2
|f (x)| dx.
|f (x) p(x)||f (x)| dx
f (x) dx = (f (x) p(x))f (x) dx
a
a
a
a
Rb
Rb
Since > 0 is arbitrary and f is fixed and a |f (x)| dx < , it follows that a f (x)2 dx = 0, so f = 0 a.e.
on [a, b]. But since f is continuous, this implies f 0 on [a, b].
5.1 If X is a normed vector space over K (= R or C) then addition and scalar multiplication are
continuous
from X X and K X to X. Moreover, the norm is continuous from X to [0, ); in
fact, kxk kyk kx yk.
Solution:
Let A : X X X be given by A(x1 , x2 ) = x1 + x2 . Then A is a linear map from the NVS X X to
the NVS X. Furthermore, for all (x1 , x2 ) X X,
kA(x1 , x2 )k = kx1 + x2 k kx1 k + kx2 k 2 max{kx1 k, kx2 k} = 2k(x1 , x2 )k,
so A is bounded, hence continuous, by Prop. 5.2.
Since K is a NVS over itself, K X is a NVS and M : K X X given by M (, x) = x is a linear
map. We will show that M is continuous at (0, 0), and result then follows from Prop. 5.2. Let > 0, and
= min{, 1}. Then for all (, x) K X with k(, x)k < we have max{||, kxk} < and 2 , so
kM (, x)k = kxk = ||kxk < 2 ,
so M is continuous at (0, 0).
Finally we show that k k is continuous. Let > 0, set = and suppose x, y X such that kx yk < .
Then kxk = kx y + yk kx yk + kyk, so kxk kyk kx yk. Also, kyk = ky x + xk ky xk + kxk,
so kyk kxk ky xk = kx yk, and therefore kxk kyk kx yk < , so k k is uniformly continuous
on X.
5.2 L(X, Y ) is a vector space and the function k k defined by (5.3) is a norm on it.
Solution:
The space Y X of all functions from X to Y is a vector space, so it suffices to show that L(X, Y ) is nonempty,
closed under addition and closed under scalar multiplication (that is, that L(X, Y ) is a subspace of Y X ).
Certainly the function Z(x) = 0 is in L(X, Y ) so its nonempty. If T1 , T2 L(X, Y ) then there are positive
constants C1 , C2 such that kT1 xk C1 kxk and kT2 xk C2 kxk for all x X. Then T1 + T2 is linear and for
all x X,
k(T1 + T2 )xk = kT1 x + T2 xk kT1 xk + kT2 xk (C1 + C2 )kxk,
so T1 + T2 is bounded, hence T1 + T2 L(X, Y ). If K then for all x X we have
k(T1 )xk = k(T1 x)k = ||kT1 xk (||C1 )kxk,
27
sup{||kT1 xk : kxk = 1}
= || sup{kT1 xk : kxk = 1}
= ||kT1 k,
so k k is a norm on L(X, Y ).
=
=
lim (1 Tn x1 + 2 Tn x2 )
1 lim Tn x1 + 2 lim Tn x2 ,
n
since addition and scalar multiplication on X are continuous by Problem 1. But the last expression equals
1 T x1 + 2 T x2 , so T is linear.
Since {Tn } is Cauchy in L(X, Y ) and kTm Tn k kTm k kTn k, it follows that {kTn k} is Cauchy in
R, hence convergent. From continuity of the norm on X we have that
kT xk = k lim Tn xk = lim kTn xk kxk lim kTn k,
n
=
=
lim kTn x Tm xk
Therefore, kTn T k for all n N , and so kTn T k 0. In particular, since kTn k kT k kTn T k,
we also have that kTn k kT k.
5.6 Suppose
that XPis a finite dimensional vector space. Let e1 , . . . , en be a basis for X and define
Pn
n
k 1 aj ej k1 = 1 |aj |.
a. k k1 is a norm on X. P
n
b. The map (a1 , . . . , an ) 7 1 aj ej is continuous from K n with the usual Euclidean topology to
X with the topology defined by k k1 .
c. {x X : kxk1 } is compact in the topology defined by k k1 .
d. All norms on X are equivalent.
Solution:
P
a. Its clear that for all x X, kxk1 0 and kxk1 = 0 iff x = 0. If x =
aj ej and K then
X
X
X
kxk1 =
aj ej
=
|aj | = ||
|aj | = kxk1 .
1
28
Finally, if x =
aj ej and y =
bj ej then
X
X
kx + yk1 =
|aj + bj |
(|aj | + |bj |) = kxk1 + kyk1 ,
so k k1 is a norm on X.
1/2
b. If we endow the vector space K n with the usual norm, k(a1 , . . . , an )k = |a1 |2 + + |an |2
, then
n
n
K n is a NVS and
this
norm
induces
the
Euclidean
toplogy
on
K
.
The
function
L
:
K
X
given
by
P
L(a1 , . . . , an ) = aj ej is linear, so by Prop. 5.2 is suffices to show that L is continuous at 0. Let > 0 and
set = /n. Let a K n with kak < . Then
> kak = |a1 |2 + + |an |2
1/2
kAk
for
all
n
1.
Since
kI
T
k
<
1,
it
follows
that
0 k(I T ) k
P
P
n
n
kI
T
k
<
,
so
the
series
(I
T
)
is
absolutely
convergent,
and
hence
convergent
by
Theorem
0
0
5.1 since L(X, X) is a Banach space by Prop 5.4.
PN
For N 0 let SN = n=0 (I T )n . Then
T SN = (I (I T ))SN = SN
N
X
(I T )n+1 = I (I T )N +1 .
n=0
Since T is continuous,
T ( lim SN ) = lim T SN = I lim (I T )N +1 .
N
29
But
lim (I T )N +1
= lim k(I T )N +1 k lim kI T kN +1 = 0,
N
N +1
so limN (I T )
= 0. Therefore
I = T ( lim SN ) = T
N
(I T )n ,
(0.4)
n=0
P
P
and so, if T is bijective, then n=0 (I T )n = T 1 . We have already that n=0 (I T )n L(X, X), so it
remains only to show that T is bijective.
Suppose BWOC that x ker T and x 6= 0. Then
kxk = k(I T )xk kI T kkxk < kxk,
a contradiction, so ker T = {0} and T is therefore one-to-one. It follows immediately from (0.4) that T is
surjective.
b. kST 1 Ik = k(S T )T 1 k kS T kkT 1 k < 1, so by part (a), ST 1 is invertible, say A =
(ST 1 )1 . Then S(T 1 A) = (ST 1 )A = I, and since T 1 A L(X, X), S is invertible.
kx2 k
30
5.19(a) Let X be an infinite dimensional vector space. There is a sequence {xj } in X such that kxj k = 1
for all j and kxj xk k 1/2 for j 6= k.
Solution:
Let x1 be an arbitrary element of X with kx1 k = 1. We inductively construct xn as follows: suppose that
x1 , . . . , xn1 have been constructed which satisfy the conclusions. Let M = Span {x1 , . . . , xn1 }. By Ex.
18b, M is closed. Since X is infinite dimensional, M is a proper subspace. So, by Ex. 12b with = 1/2,
there is an element xn X such that kxn k = 1 and kx + Mk 1/2. In particular, since xi M for
1 i < n, it follows that
kxn xi k inf{kx + yk : y M} = kxn + Mk 1/2,
sup{kT gk : kgk = 1}
kT g0 k = kg0 T k
This proves the claim and it follows that kT k sup{kT xk : kxk = 1} = kT k, hence kT k = kT k.
b. Let x X. Then T x
b=x
b T . For all g Y we have that
(b
x T )(g) = x
b(T g) = (T g)(x) = g(T x).
b and Y with Yb .
Therefore T = x
b T = Tcx, so T |X = T , after identification of X with X
c. Suppose first that T (X) is dense in Y . Suppose g Y and T g = 0. Then for all x X,
0 = (T g)(x) = g(T (x)). Since g is continuous and vanishes on the dense subset T (X) of Y , it is identically
zero, hence T is injective.
Conversely, suppose T is injective. BWOC, suppose that T (X) is not dense in Y . Then M = T (X)
is a closed proper subspace of Y by Ex. 5.5. Let y0 Y M, By Thm 5.8a, there exists g Y such
that g|M = 0 and g(y0 ) 6= 0. Then for all x X, (T g)(x) = g(T (x)) = 0, so T g = 0 X . But since
g 6= 0 Y , it follows that T is not injective, a contradiction.
d. If the range of T is dense in X , then by part (c), T is injective. Under the (bijective) identification
in part (b), T |X = T , so T is injective.
b = X so T = T after identification of X with
Suppose X is reflexive and T is injective. Then X
b
b
X = X and Y with Y . Since T is injective, by part (c), the range of T is dense in X .
5.27 There exist meager subsets of R whose complements have Lebesgue measure zero.
Solution:
For each n 1, there is an open subset On R such that Q On and mOn < 1/n. Set Fn = Onc . Then
31
\
\
mX c = m
Fnc = m
On = 0.
n1
n1
Fn . Then X is a
n1
= kT (x x0 ) + T x0 (y y0 ) y0 k
kT x0 y0 k kT (x x0 )k ky y0 k
= r kT (x x0 )k ky y0 k
r r/3 r/3 = r/3.
Therefore B(, (x0 , y0 )) (T )c and so (T )c is open, hence (T ) is closed.
f (a)f (a) =
aA
|f (a)|2 .
aA
Suppose {fn }n1 is a Cauchy sequence in `2 (A). Then for each > 0 there is an N N such that for all
m, n N
X
2 > kfm fn k2 =
|fm (a) fn (a)|2 ,
aA
and in particular, for each a0 A we have > |fm (a0 ) fn (a0 )| for all m, n N . Therefore {fn (a0 )} is a
Cauchy sequence in C for each a0 A, so we define
f (a) = lim n fn (a) for all a A.
Claim: f `2 (A). Since {fn } is Cauchy and kfm k kfn k kfm fn k for all m, n 1, it follows that
kfn k is Cauchy in R, hence convergent. By continuity of the norm, kf k = limn kfn k < , so f `2 (A).
Furthermore,
lim kf fn k = k lim (f fn )k = 0,
n
so fn f in ` (A).
32
1
kx + yk2 kx yk2 + ikx + iyk2 ikx iyk2 .
4
b. If H 0 is another Hilbert space, a linear map from H to H 0 is unitary iff it is isometric and
surjective.
Solution:
For (a), fix x, y H and let denote the RHS of the given equation. Then
1
(hx + y, x + yi hx y, x yi + ihx + iyi ihx iy, x iyi)
4
1
i
(2hx, yi + 2hy, xi) + (2hx, iyi + 2hiy, xi)
4
4
i
1
hx, yi + hx, yi +
hx, iyi + hx, iyi
2
2
1
i
(2 Re hx, yi) + (2 Re hx, iyi)
2
2
Re hx, yi + i Re (ihx, yi)
=
=
=
=
For (b), suppose that U : H H 0 is unitary. Then U is surjective by definition and for all x H,
kU xk2 = hU x, U xi = hx, xi = kxk2 , so U is an isometry.
Conversely, suppose U : H H 0 is a surjective isometry. Then U is one-to-one since U x = 0 kU xk =
0 kxk = 0 x = 0. So U is bijective and therefore has a linear inverse. For each y H 0 there is a unique
x H such that U x = y, so that kyk = kU xk = kxk, and so kU 1 yk = kxk = kyk, so U 1 is a bounded
linear map. Finally, to see that U is unitary it remains only to show that it preserves the inner product. For
all x, y H we have by (a) that
1
kU x + U yk2 kU x U yk2 + ikU x + iU yk2 ikU x iU yk2
4
1
kU (x + y)k2 kU (x y)k2 + ikU (x + iy)k2 ikU (x iy)k2
=
4
1
=
kx + yk2 kx yk2 + ikx + iyk2 ikx iyk2
4
= hx, yi.
hU x, U yi =
since kgk |g| is finite on X. Furthermore, since kgk |g| 0 a.e., the integrand is nonegative a.e., so
|f |(kgk |g|) = 0 a.e. on E, and hence kgk = |g| a.e. on E since f 6= 0 on E.
33
b. The only norm property which is not immediately obvious is the triangle inequality, so suppose
f, g L . Then |f | kf k a.e. and |g| kgk a.e. imply |f |+|g| kf k +kgk a.e., but |f +g| |f |+|g|,
so |f + g| kf k + kgk a.e., hence kf + gk kf k + kgk .
c. Suppose kfn f k 0. Then for all k N there is an Nk so that kfn f k 1/k for all n Nk .
So, for
Ek M such that Ekc = 0 and |fn f | 1/k on Ek for all n Nk . Set
Teach k N therec is an S
that |fm fn | < 1/k on Ek for all m, n Nk and Ekc = 0. Set E = k=1 Ek . Then E c = 0 and for all
x E, {fn (x)} is Cauchy, so let f (x) = limn fn (x)E(x). Then fn f uniformly on E, so by part (c),
kfn f k 0. In particular, there is an N so that kfN f k < 1 and hence kf k kfN k + 1 < , so
f L and L is therefore complete.
e. Let f L and > 0. It suffices to show that there is a simple function L with kf k < .
Suppose first that f is real-valued. Let n N such that 1/n < and let M = kf k . For each integer
j [M n, M n] let
j+1
j
f (x)
,
Ej = x X :
n
n
PmN
and set = j=M n (j/n)E . Then L is simple and for all x with |f (x)| M there is a unique j
j
such that x Ej and so |f (x) (x)| 1/n < . Thus, |f | < a.e., so kf k < .
For the general case, let 1 , 2 be simple functions for which k Re (f )1 k < /2 and k Im (f )2 k <
/2. Then = 1 + i2 is a simple function in L and kf k = k Re (f ) 1 + i( Im (f ) 2 )k
k Re (f ) 1 k + ki( Im (f ) 2 )k < .
6.5 Suppose 0 < p < q < . Then Lp 6 Lq iff X contains sets of arbitrarily small positive measure.
Lq 6 Lp iff X contains sets of arbitrarily large finite measure. What about the case q = ?
Solution:
Suppose X contains sets of arbitrarily small positive measure. Let F1 X be a measurable set with
0 < F1 < S1/2. Inductively, let Fn+1 be a measurable set with 0 < Fn+1 < (1/4)Fn . For n 1 set
\
k=i+1
!
Fkc
Fj
Fkc
k=j+1
!
Fkc
Fj = (since )i + 1 j.
k=i+1
P
Furthermore En Fn k=n+1 Fk Fn (1/2)Fn > 0 and En Fn < 2n . Therefore {En } is
a disjoint sequence
0 < En < 2n .
P with
1
Set = n=1 (nEn )1/q E . Since the En s are disjoint, this sum converges everywhere on X. Finally,
n
we have
Z
X
X
En
1
q
q
kkq = || =
=
= ,
nE
n
n
n=1
n=1
so 6 Lq . But
kkpp
Z
=
X
X
X
En
(En )1p/q
1p/q
|| =
=
<
(En )
<
2n(1p/q) < ,
(nEn )p/q
np/q
n=1
n=1
n=1
n=1
p
34
Conversely, suppose Lp 6 Lq and let f Lp with f 6 Lq . For each positive integer n let En = {x X :
|f (x)| > n}. By Prop. 6.10, if f L then f Lq , so we must have kf k = and therefore En > 0 for
all n 1. On the other hand, for each n 1,
Z
Z
Z
p
p
p
kf kp = |f |
|f |
np = np En ,
En
En
kf kp
p
np
so limn En limn
= 0, so En 0 as n .
For the second question, suppose first that X contains sets of arbitrarily
Pn large finite measure. Let
F1 X with 1 SF1 < . Inductively, let Fn+1 X with > Fn+1 > k=1 Fk + P
1. Set E1 = F1 and
n
n
En+1P= Fn+1 k=1 Fk for n 1. The sequence {En } is disjoint and En Fn+1 k=1 Fk > 1. Let
1
f = n=1 (nEn )1/p E . Then
n
Z
X
En
kf kpp = |f |p =
= ,
nE
n
n=1
so f 6 Lp . But
kf kqq =
|f |q =
X
X
(En )1q/p
1
En
=
<
< ,
q/p
q/p
q/p
(nE
)
n
n
n
n=1
n=1
n=1
so f Lq and hence Lq 6 Lp .
Conversely,
suppose Lq 6 Lp and let f Lq Lp . Let En = {x X : |f (x)| > 1/n}. Since
R
q
q
n
kf kq = |f | > E
all n 1, it follows that En nq kf kqq < for all n. Furthermore, the sets En are
nq for S
increasing, so with E = n1 En we have E = limn En . But E = {x X : f (x) 6= 0}, so if E <
then by Prop. 6.12 applied to the measure space (E, M|E , |E ), wed have kf kp < , a contradiction. Thus
limn En = . (Alternatively, apply Holder as in the proof of Theorem 6.12 to the function f E.)
What about the case q = ? The first assertion continues to hold with q = . Suppose X contains
n
sets of arbitrarily
small positive measure. As P
above, let {EP
.
n } be a disjoint sequence with 0 < En < 2
P
np
p
p
p
p
n En < n=1 2n < , so f L and hence L 6 L .
Set f =
nE . Then f 6 L but kf kp =
n
Conversely, if Lp 6 L then let f Lp L . Set En = {x X : |f (x)| n}. Since f 6 L we have
R
kf kp
En > 0 for all n and kf kpp En |f |p np En so En npp 0 as n , so X contains sets of
arbitrarily small positive measure.
For the second statement, the only if fails with q = . Let X = {x0 }, M = {, X} and = 0,
X = . The constant function f = 1 is in L but not Lp , so L 6 Lp bu X does not contain sets
of arbitrarily large finite measure. The converse, however, does still hold with q = . If X contains sets
of arbitrarily
large positive measure, let {En } be a disjoint sequence
withS1 En < for all n. Let
R
P
f = n=1 E . Then |f | 1 on X so f L , but kf kpp = |f |p = ( En ) = , so f 6 Lp , hence
n
L 6 Lp .
6.6 Suppose 0 < p0 < p1 . Find examples of functions f on (0, ) (with Lebesgue measure) such
that f Lp iff (a) p0 < p < p1 , (b) p0 p p1 , (c) p = p0 .
Solution:
Assume first that p1 < . For (a), set f = x1/p1 (0,1) + x1/p0 [1,). Then
Z
|f | =
0
dx
xp/p1
Z
+
1
dx
.
xp/p0
Both integrals are nonnegative, so kf kp < iff both integrals are finite. The first is finite iff p < p1 and the
second is finite iff p > p0 , hence f Lp iff p0 < p < p1 .
For (b), set
1
1
f = 1/p
.
x 1 | log x|2/p1 (0,1/2) x1/p0 | log x|2/p0 [2,)
R
It is a calculus exercise to verify that |f |p < iff p0 p p1 . For (c), take p1 = p0 in the function
defined f defined in (b).
1
If p1 = , for (a) take f = | log(x1)|
[1,). For (b) take f = x1/p0 | log
.
x1/p0
x|2/p0 [2,)
35
6.7 If f Lp L for some p < so that f Lq for all q > p, then kf k = limq kf kq .
Solution:
Its clear if kf k = 0, so assume kf k > 0. Let 0 < < kf k , and set A = {x X : |f (x)| > kf k }.
Then f Lp implies that A < , and by construction, A > 0. For all q p,
1/q Z
1/q
Z
q
q
|f |
(kf k )(A)1/q ,
|f |
kf kq =
A
so lim inf q kf kq (kf k ) lim inf q (A)1/q = kf k . Since this holds for all > 0, it follows
that lim inf q kf kq kf k .
For the reverse inequality, we appeal to Prop. 6.10 with r = to conclude that for all q > p, kf kq
p/q
1p/q
p/q
1p/q
kf kp kf k . Therefore, lim supq kf kq lim supq kf kp kf k
= kf k .
6.12 If p 6= 2 the Lp norm does not arise from an inner product on Lp except in the trivial cases when
dim Lp 1.
Solution:
Let 1 p , and suppose that (X, M, ) is a measure space and dim Lp > 1. If every E M had
E {0, } then the dimension of Lp would be zero, so there must exist a set E1 M with 0 < E1 < .
Furthermore, if every E M with 0 < E < satisfied (E4E1 ) = 0, then every function in Lp would
equal a constant multiple of E a.e., which would give dim Lp = 1. Thus, there is a set E2 M with
1
0 < E2 < and E1 E2 = .
Suppose now that p < . Set f1 = E /(E1 )1/p and f2 = E /(E2 )1/p . Then kf1 kp = kf2 kp = 1.
1
2
Suppose that the norm on Lp does arise from an inner product. Then the parallelogram law holds, so that
kf1 + f2 k2p + kf1 f2 k2p = 2(kf1 k2p + kf2 k2p ) = 4.
Computing the LHS we have
kf1 +
f2 k2p
+ kf1
f2 k2p
p 2/p Z
p 2/p
Z
E
E
E1
E1
2
2
=
+
(E )1/p + (E )1/p
(E )1/p (E )1/p
1
2
1
2
Z
E 2/p Z E
E 2/p
E1
2
1
2
=
+
+
(since E1 E2 = )
E1
E2
E1
E2
=
1
for 1 i n such that kf kp < /3. Let M = max{|ai aj | : 1 i < j
p L 1, so by
R n}. Then
Theorem 2.26 there is a step function with finite measure support such that | | < 3 M p1 . It
follows that
1/p
1/p
k k1 k k11/p
k k1 M 11/p < /3.
k kp k k1 k k11/p
< /3.
kri (c
i ,di )
to do it).
Pn
0
0
0
i=1 ri (c0 ,d0 ) with the intervals (ci , di ) disjoint. For each 1 i n let ri , ci , di Q
i i
0
ri (c0 ,d0 )kp < /(3n) (its clear that they exist, but the argument is straightforward if
i i P
Pn
n
Then s = i=1 ri (c ,d ) S and ks kp i=1 kri (c ,d ) ri0 (c0 ,d0 )kp < /3 and so
i i
i i
Suppose =
such that
you want
37