Sei sulla pagina 1di 62

# lder Inequality

I. The Ho
Holder: kf gk1 kf kpkgkq for

1
p

1
q

= 1.

## What does it give us?

Holder: (Lp ) = Lq (Riesz Rep), also: relations between Lp spaces
I.1. How to prove H
older inequality.
p
q
(1) Prove Youngs Inequality: ab ap + bq .
(2) Then put A = kf kp, B = kgkq . Note: A, B 6= 0 or else trivial. Then
|g(x)|
let a = |f (x)|
A , b = B and apply Youngs:
ap bq
ab =

+
=
+
p
q
Z
Z
Z
1
|f (x)g(x)|d pA1 p |f |p d + qB1 q |g|q d
AB
R
R
but Ap = |f |p d and B q = |g|q d, so this is
|f (x)|p
pAp

|f (x)g(x)|
AB

1
kf kp kgkq kf gk1

1
p

1
q

|g(x)|q
qAq

=1

kf gk1 kf kpkgkq

## I.1.1. How to prove Youngs inequality.

There are many ways.
1. Use Math 9A. [Lapidus]
Wlog, let a, b < (otherwise, trivial).
p
Define f (x) = xp + 1q x on [0, ) and use the first derivative test:
f 0 (x) = xp1 1, so f 0 (x) = 0 xp1 = 1 x = 1.

## So f attains its min on [0, ) at x = 1. (f 00 0).

Note f (1) = 1p + 1q 1 = 0 (conj exp!).
So f (x) f (1) = 0 =

xp
p
xp
p

+ 1q x 0
+

1
q

ab1/(1p)
ab
ab

ap bq
p
ap
p
ap
p

+
p
1p

bq
q

1
q



p
1p

p
1p

p
1 1p
qb

= q

1
1 1p

=b

p
1p

## 2. Use Math 9B. [Cohn]

Consider the graph of t = sp1:

## Figure 1. The graph of t = sp1

t
t = s p-1

b
(1)
(2)
a

Since
1
p

1
q

= 1 =

1
p

=1

1
q

q1
q

= p = qq 1 = p 1 =

## this is also the graph ofis = tq1 .

Ra
p a
p
Now (1) = 0 sp1 = sp = ap ,
ib 0 q
R b q1
tq
and (2) = 0 t
= q = bq .

1
q1 ,

## Thus the area of the entire shaded region is (1) + (2) =

is clearly always larger than the box of area ab:

## I.1.2. A proof without Youngs inequality.

Use convexity [Rudin]:
((1 )x + y) (1 )(x) + (y).

ap
p

+ bq , which

t

t
t = s p-1

t = s p-1
b

extra

extra
ab

ab
a

## Since f Lp , g Lq , we have 0 < kf kp, kgkq < , wlog.

|g(x)|
Define F (x) = |fkf(x)|
kp and G(x) = kgkq so that
Z

F p d =

R
and Gq = 1 similarly.
Now define

R
|f |p d
|f (x)|p
d = R
=1
kf kpp
|f |p d

|f (x)|
s(x) = log
kf kp

p

|g(x)|
, t(x) = log
kgkq

so that
F (x) = es(x)/p and G(x) = et(x)/q .
Since ex is a convex function, put =
s(x) t(x)
+ q
e p

1
q

and get

s(x)
1
p
pe

= F (x)G(x)

F (x)p
p

t(x)
1
q
qe

G(x)q
q .

q

## Now integrate the left side and get

Z
kF Gk1 = |F G|d
Z
|f g|
=
d
kf kpkgkq
Z
1
=
|f g|d
kf kpkgkq
kf gk1
=
,
kf kpkgkq
and we integrate the right side to get

Z 
Z
Z
F (x)p G(x)q
+
d = 1p F p d + 1q Gq d
p
q
= 1p + 1q
=1
Thus,
kf gk1
1 = kf gk1 kf kpkgkq .
kf kpkgkq
Advantage of this method? No need for Youngs!

## I.1.3. Recap - 3 good ways to prove a functional inequality.

To prove a(x) b(x):
1. Use basic calculus on a difference function:
Define f (x) := a(x) b(x).
Use calculus to show f (x) 0 (by computing f 0 , etc)
2. Use geometry.
3. Exploit another inequality. E.g., for any convex function (x),
((1 )x + y) (1 )(x) + (y).

## I.1.4. What did we not do yet? case p = 1, .

|g(x)| 6ae
|f (x)| |g(x)| 6ae
|f (x)g(x)| 6ae
kf (x)g(x)k1 6ae

kgk
|f (x)| kgk
|f (x)| kgk
|f (x)| kgk

## p = is exactly the same.

I.2. How to use the H
older inequality. Assume (X, M, ) is a measure
space with X 1, f : X R is measure, and Lp = Lp(X, ).
1. For
if |f (x)| 1, then kf kp kf kq . If X = , then
R 1 p p Rq < ,
q
X |f | d = X |g| d = , so let X < .
Then
|f |p |f |q
Z
Z
p
|f | d
|f |q d.
X

(I.1)

R
If X |fR |q d = , it is trivial, so assume not.
Then X |f |q d < = f Lq , and (I.1) = f Lp.
Now p = q = kf kp = kf kq and we are done trivially, so let p < q. We
would like to use Holder with g(x) = 1 and some conjugate exponents

## Real Analysis Qual Seminar

, with 1 + 1 = 1.
Ansatz : Let = pq and =
1

q
qp ,
p
q

so
qp
q

q
q

= 1.

## Now use Holder with f = f p to get

kf pgk1 kf pk kgk .

(I.2)

## Now remembering that g = 1, we have

Z
p
p
kf gk1 = kf k1 =
|f |p = kf kpp, and
p

kf k =
kgk =

Z

Z

p q/p

|f |
1

p/q

1/

Z

Z

1
X

|f |

1/

p/q

## = kf kpq, and (*)

= (X)1/ .

So (I.2) becomes
kf kpp

kf kpq

qp
(X) q

kf kp kf kq (X)(qp)/pq
kf kp kf kq

q/p

## 2. For 1 p q < , |f (x)| 1 x X = kf kp kf kq .

p = q is trivial, so take p < q.
Then
p < q, |f | 1 = |f |p
Z
Z
|f |p |f |q
Z

|f |

1/p

Z

|f |

kf kp kf kq/p
q .

|f |q

1/p

## Real Analysis Qual Seminar

3. Show p q r and f Lp , f Lr = f Lq .
e
Let A = {|f | 1} and B = {|f | < 1} = A.

f Lp =

ZX

|f |p =

|f |p +

|f |p <

|f |p <
ZB
Z
Z
f Lr =
|f |r =
|f |r +
|f |r <
A
B
ZX
=
|f |r < .
=

R
R
On A, |f |q |f |r = RA |f |q RA |f |r .
On B, |f |q |f |p = B |f |q B |f |p .
So

|f |q =

ZA

<

|f |q +
|f |r +

ZB

|f |q
|f |p
by (I.3),(I.4)

q
shows that f L
R .
R
Moral: to show X f (x) X g(x), try splitting X.

(I.3)

(I.4)

## 4. Show there is a bounded linear operator : Lq (Lp ) given by

Z
f (g) = (f )(g) =
f gd, f Lq , g Lp
X
R
so that : g 7 f gd is the functional integration against f.
kk =

k(f )k
f Lq ,f 6=0 kf kq
sup

k(f )(g)k
f Lq ,f 6=0 gLp ,g6=0 kf kq kgkp
R
| X f g d|
= sup
sup
f Lq ,f 6=0 gLp ,g6=0 kf kq kgkp
=

sup

kf gk1
sup
sup
f Lq ,f 6=0 gLp ,g6=0 kf kq kgkp

kf kq kgkp
f Lq ,f 6=0 gLp ,g6=0 kf kq kgkp
sup

def of kk

sup

def of (f )(g)
|

f g|

sup

|f g|

Holder

so kk 1 and is bounded.
To see is linear, let f1 , f2, f Lq , g Lp, and R: we show two
things in (Lp ) are equal by showing that they act the same way on
any g Lp .
Z
(f1 + f2 )(g) = (f1 + f2) g d
Z
Z
= f1g d + f2g d
= (f1)(g) + (f2)(g) = ((f1) + (f2)) (g)

## shows (f1 + f2) = (f1) + (f2), and

Z
Z
(f )(g) = f g d = f g d = (f )(g)

shows (f ) = (f ).
Hence (by the linearity of the integral), is linear.

## II. The Dual of Lp

Proposition II.1. Show that : Lq (Lp) by : f 7
isometry.

f gd is an

## Proof. So we must show k(f )k = kf k, f Lq . Let 1 < p, q < . Then

k(f )(g)k
kgkp
gLp ,g6=0
R
| X f g d|
= sup
kgkp
gLp ,g6=0

k(f )k =

sup

def of kk
def of (f )(g)

kf gk1
sup
gLp ,g6=0 kgkp

sup kf kq

f g|

|f g|

Holder

gLp ,g6=0

Hence k(f )k kf k. For k(f )k kf k, use the fact that k(f )k is defined
as a supremum: k(f )k is the smallest number such that
k(f )(g)k k(f )k kgk

## holds for all g (6= 0).

)(g)k
In other words, if we can find a g for which k(fkgk
kf k, then
n
o
)(g)k
kf k.
k(f )k = supgLp ,g6=0 k(fkgk

## Ansatz : let g = |f |q/p sgn f .

Then |g|p = |f |q = f g.1 Thus, f Lq = g Lp. Now
Z
Z
p
|g| = |f |q
Z

Z

|f |q

1/q Z

|g|p
|g|p

1/p

1/p

=
=

kf kq kgkp =
11 +
p

1
q

= 1 =

q
p

q
q

= q =

q
p

Z

Z

Z

|f |q
|f |q
|f |

1/p

1/q Z

1/q+1/p

|f |q
=

1/p

Z

|f |

1

= kf kqq

10

## Thus, (f )(g) = f g d = |f |q = kf kqq = kf kq kgkp

)(g)|
= |(f
kf kq .
kgkp
Now suppose p = 1, q = .
We have
R
| f gd|
kgk1 kf k
k(f )k = sup
sup
= kf k
kgk1
kgk1
R
R

as before. Now it remains to find a g L1 for which f g d
|g|d kf k .
We have f L , so note kf k < . Then fix > 0 and define
B = {f kf k },

and let A be any measurable subset of B such that 0 < A < .2 Define
g := A sgn f.3
R
Then |g |d = A and
Z
Z
f g d =
|f |d (kf k ) A
A
R
|f g|d
kf k
A

R
|f g |d
sup R
kf k .
|g |d
nR
o
|f g|d
R
Since this is true for any , let 0 and obtain sup
kf k.
|g|d

Now suppose p = ,R q = 1.
| f gd|
Again, k(f )k = sup kgk kf k1, as before. Now find g L such that
Z

Z
f g d kgk
|f |d .
Let > 0 and define g = sgn f be a constant function. Then
Z
Z
Z
f g d = |f |d = kgk |f |d.
2Note that if f is a constant function, B could be ! B > 0 by def of kf k .

11

## II.1. The Riesz Representation Theorem.

Theorem II.2. Let F be a bounded
linear functional on Lp, 1 p < .
R
Then g Lq such that F (f ) = f g d, f Lp , and kF k = kgkq .
There are two common proofs for this theorem. One uses step functions and
absolute continuity of functions; the other uses simple functions and absolute
continuity of measures. Both follow a similar strategy:
R
R
(1) Show F (A ) = g A d = A g d.
use absolute continuity of : [0, 1] R by gF (s) = F (s ); or
use absolute continuity of E = F (E ).
(2) Extend to a dense subspace of Lp
use step functions for gF (s)
use simple functions for E
(3) Establish kF k = kgkq
extend F to bounded
measurable functions, use Royden & Holder
R
define G(f ) = f gd on Lp and use density, continuity

(4) Extend to Lp
approx by step functions
use G
(5) Show uniqueness of g.

Method I:
uses step functions only applies for Lp ([a, b], )
requires reference to 3 thms of Royden
nice use of DCT, boundedness
absolute continuity of a function is a bit more concrete
Method II:
uses simple functions, so applies to Lp(X)
requires reference to 1 thm of Royden
smooth use of general topology
works for any -finite

12

## Important note: must add is -finite in order to do the case p = 1!

Method I (Royden)
Proof.
Rs

## 1. For s [0, 1], let s := [0,s] . Then F (s ) =

number, so define : [0, 1] R by
Z s
(s) = F (s ) =
g d.

g d is some real

## Claim: is absolutely continuous.

Fix > 0 and let {(ai, bi)}ni=1 be any finite collection of disjoint
subintervals of [0, 1] such that
X
(bi ai ) < .
P
Then
|(bi) (ai)| = F (f ) for
f=

Since

n
X
i=1

## (bi ai ) sgn ((bi) (ai)) .4

P
P R bi
PR p

1
d
=
(bi ai ) < ,5
=
|f |p =
(ai ,bi )
ai
X
|(bi) (ai)| = F (f ) kF k kf kp < kF k 1/p.

## Thus, total variation of over any finite collection of disjoint intervals

is less than , as long as the total length of these intervals is less than
=

p
kF kp ,

## which shows that is absolutely continuous.

Then has an antiderivative, by some theorem:
Z s
(s) =
g.

## 4Note that |(b ) (a )| = |F ( ) F (b )| = (F ( ) F ( )) sgn (( ) ( )) .

i
i
ai
i
ai
bi
bi
ai
5In the first equality, use ( ) =
and
|
sgn
h|
=
1
ae.
bi
ai
(ai ,bi )
6Royden, Lemma 5.14 on page 110.

13

Thus F (s ) =

R1
0

g s .

## 2. Since every step function on [0, 1] is a linear combination

X
=
ci si (-ae),
we get

F (f ) =
R
by the linearity of F, .

g
0

## 3. Now extend F to the bounded measurable functions f on [0, 1].

Let f be such a function, and find a bounded sequence {n} Step
which converges -ae to f . This is possible by Royden, Prop 3.22.
Then {|f n |p } is uniformly bounded (M such that |f n |p
M p , n, x) and tends to 0, -ae.
This bound allows us to use the DCT and get
Z
n
lim
|f n |p = 0
=
kf nkp 0.
n

## Then the boundedness of F and the fact that

|F (f ) F (n)| = |F (f n)| kF k kf nk
together imply that F (f ) =Rlimn F (
R n).
Also, |gn | |g| M = f g = lim gn by the DCT again.7
Putting this together,
Z
Z
F (f ) = lim F (n) = lim gn = f g, f (bdd,measurable).
Then by Proposition 5.12 on page 131 of Royden,

Z

f g d = |F (f )| kF k kf kpN kf kp

## and we have g Lq , kgkq N = kF k,.

Then by Prop II.1, kF k = kgkq .
7M is the uniform bound on the sequence { }.

14

## Real Analysis Qual Seminar

4. Extend to f Lp .
, Step such that kf kp < .
R
Then Step = bounded = F (f ) = g, by (3).
Hence,

Z
Z
Z

F (f ) f g = F (f ) F () + f f g

Z

Z

|F (f ) F ()| + f g f
Z

= |F (f )| + (g )f
kF k kf kp + kf kp kgkq

f g d.

## 5. If g1, g2 determine the same F in this way, then

Z
Z
Z
f g1 d f g2 d = f (g1 g2 ) d
gives the zero functional, and

kg1 g2 kq = 0

g1 =
g.
ae 2


Method II (Royden)
Again, given a bounded
linear functional F on Lp , we must find a g Lq
R
such that F (f ) = f g d.
Proof. First, consider a finite measure space (X, A, ).
Then f bounded = f Lp ().
1. Define on A by

E = F (E ).

15

## Then is a signed measure:

If E P
is the disjoint union of {En } A, let n = sgn F (En ) and define
f=
n En . Then F is bounded, so a lemma gives

X
X
|En| = F (f ) < and
En = F (E ) = E.
n=1

n=1

## So is a signed measure. Also,

= 0 = F (E ) = 0,
so
R . Then Radon-Nikodym applies and g measure such that
E = E g d.
Note: F bounded
= F (X ) = X < , so is finite.
R
Then X = X g d < = g L1().
2. Let be a simple function. Then by linearity of F and
Z
F () = g d.

, we have

## 3. (&4.) Define G(f ) =

which vanishes on S.

## f g d for f Lp, so (G F ) is a BLT on Lp

(G F ) bounded (G F ) continuous,

so (G F ) = 0 on Lp . R
Hence, f Lp , F (f ) = f g d and kF k = kGk = kgkq .
5. If g1, g2 determine the same F in this way, then
Z
Z
Z
f g1 d f g2 d = f (g1 g2 ) d
gives the zero functional, and

kg1 g2 kq = 0

g1 =
g.
ae 2

## Now consider -finite. Choose {Xn } such that

[
X=
Xn , Xn Xn+1, Xn < , n.
n=1

16

## Then the previous

R case gives a gn for peach Xn such that gn vanishes outside
Xn and F (f ) = f gn d for all f L that vanish outside Xn .
By construction, the uniqueness of gn on Xn gives

gn+1 = gn ,
X
n

so for x X, define

## g(x) := gn (x), where x Xn .

Since gn differs from gm on a set of at most measure 0 (on any Xi where both
are defined), discrepancies may be safely ignored and g is well-defined.
Moreover, |gn | increases pointwise to |g|. By MCT,
Z
Z
q
|g| d = lim |gn |q d kF kq ,
so g Lq .

## For general f Lp , define

fn =
pw

Lp

f
0

on Xn
fn .
on X

Then fn f and fn f .
Then |fn g| |f g| L1, so DCT gives
Z
Z
DCT
f g d = lim fn g d
Z
DCT
= lim fn gn d
= lim F (fn )
= F (f ).


Side note: if is Lebesgue measure and is the point mass at 0 (on
d
be (if  )?
(R, B(R))), then what would d

17

## II.2. (L1) = L , but (L) 6= L1.

Let X be [0, 1] so that we can safely consider C(X) as a subspace of L (X).
Define : C(X) R by (f ) = f (0), so (C(X)) .
By HBT, (L ) such that (f ) = f (0) f C(X). To see that
cannot be given by integration against a function in L1, consider fn C(X)
defined by fn(x) = max{1 nx, 0}.
Figure 3. The functions fn .

1
fn

1
n

## Then (fn) = fn (0) = 1 n.

1
But fn (x)
R 0 x > 0, so fn g 0 g L .
If (fn) = fn g, then we would have
1 = (f ) = (lim fn )

= lim (fn)
Z
= lim fng
Z
= lim fng
Z
= 0

(L ) = continuous
hypothesis
DCT

=0

Slightly fancier version: use (f ) = f (p) and fn (x) = max{0, 1 n|x p|}.

18

## III. The Minkowski Inequality

(X, M, ), 1 p . Then kf + gkp kf kp + kgkp .
What does it give us? Lp is a normed vector space (4-ineq), Lp is Banach.

## III.1. How to prove Minkowski inequality.

III.1.1. Use H
older inequality.
case i) 1 < p < .
For f, g Lp , define q by
1
q

1
p

= 1

= q =
= p + q =
Now
|f + g|p1

q

1
p

1
q

= 1, so that p + q = pq:
p1
p

p
p1

p2 p
p1

p
p1

= |f + g|p1

p2
p1

=p

p/(p1)

p
p1

= pq

= |f + g|p .

## Since Lp is a vector space, we have f +g Lp and hence |f +g|p1 Lq .

We need to set up for Holder:
|f + g|p = |f + g|p1 |f + g|

## |f + g|p1|f | + |f + g|p1 |g|

Z
Z
Z
|f + g|p d |f + g|p1 |f | d + |f + g|p1 |g| d
Now Holder gives
Z
|f | |f + g|p1 d kf kpk|f + g|p1kq
Z

and

19

Thus
Z

## |f + g|p d (kf kp + kgkp) k|f + g|p1kq

= (kf kp + kgkp )

Z

|f + g|(p1)q d

1/q

R
If |f + g|p d = 0, then Minkowski is trivial, so assume not.
Then we may divide by
Z
1/q Z
1/q

p1 q
p
|f + g|
d
=
|f + g| d
to get

case ii) p = 1.
Then
Z

Z

|f + g|p d

11/q

kf kp + kgkp .

|f + g| |f | + |g|
Z
Z
|f + g|d |f |d + |g|d

4-ineq
integration

kf + gk1 kf k1 + kgk1

case iii) p = .
Then define the null sets
N1 := {|f (x)| > kf k},

## Then f, g L = N1 = N2 = 0, so (N1 N2) = 0 also. On the

complement of N1 N2 we have
|f (x) + g(x)| |f (x)| + |g(x)|.
Then taking suprema gives
kf + gk kf k + kgk .

20

## III.1.2. Use convexity.

Let = kf kp and = kgkp . Note: , 6= 0 or else trivial. Then define
f0 := 1 |f |,

|f | = f0 ,

## Note that this implies

Then we have

kf0kp = kg0kp = 1.

kf0kpp = kg0kpp = 1

Set
:=

|g| = g0 ,

g0 := 1 |g|

so 1 =

+
+

(III.1)

.
+


p

= ( + ) +
f0(x) +

4-ineq

+ g0 (x)

p

def of f0 , g0

## = ( + )p (f0 (x) + (1 )g0(x))p

def of
= ( + )p (f0 (x)p + (1 )g0(x)p)
cvxty of tp
Z
Z
|f (x) + g(x)|p d ( + )p (f0(x)p + (1 )g0 (x)p) d integrating

kf + gkpp ( + )p kf0kpp + (1 )kg0kpp
linearity
( + )p ( + (1 ))

by (III.1)

= ( + )p

kf + gkp ( + )

pth roots

21

## III.2. The Riesz-Fischer Theorem: Lp(X, M, ) is complete.

Split into the two cases 1 p < , p = . For each case:
(1) Invoke the Banach characterization lemma.
(2) Define
(P
fn (x) behaves
f (x) =
0
else

P
.
Use g(x) = k |fk (x)| for 1 p < , use Nk = {x .. |fk (x)| > kfk k }
for p = .
P
Lp
(3) Use Minkowski to show f Lp and nk=1 fk f.
case i) 1 p < .
1. By the lemma, it suffices to show that every series which converges
absolutely (in R) also converges in Lp , p [1, ).
P

p
2. Let
k kp < for some {fk }k=1 L .
k=1 kf
P
n
NTS: kf nk=1 fk kp 0 for some f Lp, since this is what
Pn
Lp
f

f means.
k=1 k
Define
g(x) =

X
k=1

|fk (x)|

n
X
k=1

|fk |

!p

(III.2)

n
X
k=1

|fk |

!p

n+1
X
k=1

|fk |

!p

(III.3)

22

Then we have
Z
kgkp =
lim
=

p !1/p
n
X

|fk | d

n

k=1
!p !1/p
Z
n
X
lim
|fk | d
n

= lim

k=1

n
X

k=1

n

X

= lim
|fk |
n

lim

X
k=1

k=1
n
X
k=1

|fk |

!p

kfk kp

!1/p

by (III.2)

by MCT,(III.3)
def of k kp
Minkowski

kfk kp ,

## which is finite, by hypothesis. Thus g Lp , so |g| 6ae .

Hence we may define
(P

## k=1 fk (x) |g(x)| <

f (x) =
0
|g(x)| =
so that f is measurable and

|f |p g p = f Lp .
P
P
p
Since limn |f (x) nk=1 fk (x)| = 0 and |f (x)
fk (x)|
k=1
P
n
g p are both true ae, the DCT gives kf nk=1 fk kp 0.

23

## Real Analysis Qual Seminar

case ii) p = .
1. By the lemma, it suffices to show that every series which converges
absolutely (in R) also converges in L .
P

2. Let
k k < for some {fk }k=1 L .
k=1 kf
P
n
NTS: kf nk=1 fk k 0 for some f L .
For each k, define
.
Nk := {x .. |fk (x)| > kfk k },

so that Nk = 0, k = (k Nk ) = 0. Then if x
/ k Nk ,
X
X
X
|fk (x)|
kfk k
=
fk (x) < ,
k

(P
/ k Nk
k fk (x) x
f (x) =
0
x k Nk

## so that f is -measurable and bounded, i.e., f L .

3. Since (k Nk ) = 0,

n

X

X

fk
fk
f

k=1

k=n+1

k=n+1

kfk k

by Mink

## Then taking limits,

n

X
X

fk lim
kfk k = 0.
lim f
n
n

k=n+1
k=1

Pn
n
Thus, kf k=1 fk k 0.

24

## Lemma III.1. (Banach Characterization Lemma).

Suppose that (X, k k) is a normed vector space. Then
X is Banach every absolutely convergent series in X is convergent.
Proof.
() Suppose every Cauchy
sequence converges.
P
Let {xk } be such that
kxk k < so {xk } is absolutely convergent.
Then let
n

X
X
sn :=
xk , ands := lim sn =
xk .
n

k=1

k=1

m

X

ksn sm k =
xk

k=n+1
m
X

k=n+1

kxk k

k=n+1

4-ineq
kxk k

## Hence, {sn } Cauchy implies that sn s = limn sn X.

() Suppose that every abs. convergent series is convergent.
Let {xn} be a Cauchy sequence.
NTS: xn x X.
Since {xn} is Cauchy, we can find a subsequence {xnk } which satisfies
kxnk xnk+1 k <

Define

v 1 = x n1 ,

1
2k .

N
X
k=1

25

## Real Analysis Qual Seminar

so

kvk k <

1
2k

=1

k=1
k=1
P
shows Pk=1 kvk k converges.
Hence,
k=1 vk converges by hypothesis to some v X. Then

vk = v = lim

k=1

N
X
k=1

vk = lim xnN
N

n

26

## IV. Hilbert Space Review

Most material in this talk is from Reed & Simon.
Definition IV.1. A complex vector space is called an inner product space
(IPS) when
(i) hx, xi 0, and hx, xi = 0 iff x = 0,

## (ii) hx, y + zi = hx, yi + hx, zi,

(iii) hx, yi = hx, yi,
(iv) hx, yi = hy, xi.

An inner
p product space is a Hilbert space iff it is complete under the norm
kxk = hx, xi.
Definition IV.2. Two vectors x 6= y are orthogonal iff hx, yi = 0.
A collection {xi} is an orthonormal set iff
hxi, xii = 1 and hxi , xj i = 0

i 6= j.

## Proposition IV.3. (Pythagorean Theorem)

Let {xn}N
n=1 be an orthogonal set in an IPS. Then

2
N
N
X

X

kxnk2
xn =

n=1
n=1
P
P
P
P
Proof. k xn k2 = h xn , xn i = N
n,m=1 hxn , xm i .
ThenP
see that all the terms
PN with n2 6= m are 0 because of orthogonality, leaving
N
only n=1 hxn, xni = n=1 kxnk .

Proposition IV.4. (Bessels Inequality)
If {x }A is an orthonormal set in an IPS, then for any x,
X
|hx, xi|2 kxk2.
A

27

## Real Analysis Qual Seminar

P
Proof. It suffices to show that F |hx, x i|2 kxk2 for any finite F A:

2

X

0 x
hx, xix

F
*
+
X
X
= x
hx, x ix , x
hx, x ix
F

2
X

2
= kxk 2 Re x,
hx, x ix + hx, xix

F
F

2

X
X

2
= kxk 2 Re
hx, xi hx, xi + hx, x ix

F
F
X
X
= kxk2 2
|hx, x i|2 +
|hx, xi|2
*

= kxk2

X
F

(IV.1)

|hx, x i|2

## Where (IV.1) comes by the Pythagorean Thm.

.
Note that this theorem indicates { .. hx, xi 6= 0} is countable.

## Proposition IV.5. (Schwartz Inequality)

If x and y are vectors in an IPS, then
kxkkyk |hx, yi| .

y
Proof. The case y 0 is trivial, so suppose y 6= 0. The vector kyk
by itself
forms an orthonormal set, so applying Bessels inequality to any x gives

2

y
2
kxk hx, kyk i

|hx, yi|2
=
kyk2

28

## Real Analysis Qual Seminar

p
Proposition IV.6. kxk = hx, xi really is a norm.

## Proof. The first two properties of norm are clearly satisfied:

kxk = 0 x = 0,
To see the triangle inequality,

kxk 0,

kxk = ||kxk.

kx + yk2 = hx + y, x + yi = hx + y, xi + hx + y, yi
= hx, xi + hx, yi + hy, xi + hy, yi

linearity

z+z
2

## kxk2 + 2 |hx, yi| + kyk2

= Re z

Re z |z|

= (kxk + kyk)2


Proposition IV.7. (Parallelogram Identity)


kx + yk2 + kx yk2 = 2 kxk2 + kyk2 .

## kx yk2 = kxk2 2 Rehx, yi + kyk2.


Example.
`2 :=
with the inner product
D

{xn}
n=1

.. X
.
|xn |2 <
n=1

{xn }
n=1 , {yn }n=1

:=

X
n=1

xn y n .

29

## Real Analysis Qual Seminar

Example.
L2 :=
with the inner product

Example.
L2H

:=

## with the inner product

.
f : X C ..
hf, gi :=

.
f : X H ..

hf, gi :=

|f |2 d <

f g d.
X

kf (x)k2H d

<

hf (x), g(x)iH d.

IV.1. Bases.
Definition IV.8. An orthonormal basis of a Hilbert space H is a maximal
orthonormal set S (i.e., no other orthonormal set contains S as a proper
subset).
Theorem IV.9. Every Hilbert space has an orthonormal basis.
Proof. Let C be the collection of orthonormal subsets of H. Order C by
inclusion:
S1 S 2

S 1 S2 .

## Then (C, ) is a poset.

It is also nonempty since {x/kxk} is an orthonormal set, x H.
Now let {S }A be any linearly ordered subset of C.
Then A S is an orthonormal set which contains each S and is thus an
upper bound for {S }A .
Since every linearly ordered subset of C has an upper bound, apply Zorns
Lemma and conclude that C has a maximal element.
This maximal element is an orthonormal set not properly contained in any
other orthonormal set.


30

## Theorem IV.10. (Orthogonal Decomposition and Parsevals Relation)

Let S = {x }A be an orthonormal basis for a Hilbert space H. Then
y H:
X
X
2
y=
hx , yix , and kyk =
|hx , yi|2 .
A

## Proof. Proving Bessels inequality, we saw that

X
|hx , yi|2 kyk2 ,
A

## and that there are at most countably many nonzero summands.

Collect these s for which hx , yi 6= 0 to obtain a sequence {j }
j=1 .
PN
As a positive-term series, j=1 |hxj , yi|2 is monotone increasing.
It is also bounded above by kyk2 .
Thus, it converges to a finite limit as N . Define
n
X
yn :=
hxj , yixj .
j=1

2

n
n

X
X

2
|hxj , yi|2 ,
hxj , yixj =
kyn ym k =

j=m+1

j=m+1

## by the Pythagorean Thm. Letting n, m shows {yn } is Cauchy.

Since H is Hilbert, it is complete and {yn } must converge to some y 0 H.
Let x be any element of S. If ` = ` , then by the continuity of norms:
*
+
n
X
hy y 0 , x` i = lim y
hxj , yixj , x` = hy, x` i hy, x` i = 0
n

j=1

and if not,
hy y 0 , x i = lim

n
X
j=1

hxj , yixj , x

=0

31

## Real Analysis Qual Seminar

because
hy y 0 , x i = lim

n
X
j=1

= hy, x i lim

=0
=
=

j=1

hxj , yixj , x
n
X

j=1

hxj , yixj , x

hxj , yixj , x

hy, xj i xj , x

hy, x i = 0 for 6= `

j=1

X
j=1

hy, xj i 0

hxj , x i = 0 for 6= `

=0

## So y y 0 is orthogonal to every x in S. Since S is a orthonormal basis, this

means we must have y y 0 = 0. Thus
n
X
y = lim
hxj , yixj ,
n

j=1

## and we have shown the first part. Finally,

2

n

X

hxj , yixj
0 = lim y
n

j=1

2
* n
+ n

X
X

2

= lim kyk 2 Re y,
hxj , yixj + hxj , yixj
n

j=1
j=1
!
n
n
X
X

hx , yix 2
= lim kyk2 2 Re
hx , yi y, x +
j

= lim

j=1

j=1

n
n
X

2 X
2
x j , y +
hxj , yi 2 xj 2
kyk 2
j=1

j=1

32

= lim

kyk

= kyk2

X
A

n
X
j=1

|hxj , yi|2

|hx , yi|2

kyk2 =

X
A

|hx , yi|2 .


## Definition IV.11. The coefficients hx , yi are the Fourier coefficients of y

with respect to the basis {x }.
IV.2. The Riesz Representation Theorem Again.
Definition IV.12. Let M be a closed subspace of H. Then M is a Hilbert
space under the inner product it inherits as a subspace of H. Define the
orthogonal complement of M to be
.
M := {x H .. hx, yi = 0 y M}.
Theorem IV.13. (Projection Theorem)
If M is a closed subspace of H, then H = M M. That is, x H, x can
be uniquely expressed as x = y + z, where y M, z M. Moreover, y, z
are the unique elements of M and M whose distance to x is minimal.
If y H, then y (x) = hx, yi defines a functional on H.
By the linearity of inner prod, it is a linear functional.
By the Schwartz inequality8,
ky k = supkxk1 ky (x)k = supkxk1 |hx, yi| supkxk1 kxkkyk kyk

## Shows that this functional is bounded/ continuous.

8Recall, the Schwartz ineq is just the H
older ineq when p = 2.

33

## Theorem IV.14. (Riesz Representation Theorem for Hilbert Spaces)

If H , then !y H such that (x) = hx, yi x H. Also, kk = kyk.
Proof. If is the zero functional, then y = 0 and were done.
Otherwise, consider the nullspace
.
M := {x H .. (x) = 0}.

## M is a proper closed subspace of H and M 6= {0} by the Projection Thm.

Thus we can find z M with kzk = 1 and define
Then

u := (x)z (z)x.


## shows that u M and hence that u z. Thus,

0 = hz, ui = hz, (x)z (z)xi

linearity

## = (x) hx, (z)zi

kzk = 1

= (x)kzk2 (z)hz, xi

hz, zi = kzk2

## Thus, (x) = hx, yi where y = (z)z.

As for uniqueness, if hx, yi = hx, y 0 i for all x, take x = y y 0 and get
ky y 0 k2 = hy y 0 , y y 0 i = hy y 0 , yi hy y 0 , y 0 i = 0

y = y0.


## This shows that y 7 y is a conjugate linear isometry of H onto H.

Definition IV.15. Isomorphisms of Hilbert spaces are those transformations
U : H1 H2 which preserve the inner product:
hU x, U yiH2 = hx, yiH1 x, y H1 .

## Such operators are called unitary.

For U : H H, unitary operators are also characterized by U = U 1, where
T is the Hilbert space adjoint of T L(H) and is defined by
hT x, yi = hx, T yi.

34

## V. A Practical Guide to Integral Problems

This talk covers the relation between Riemann and Lebesgue integration,
when you can differentiate under an integral, and other practical applications
of Lebesgue theory to standard integral problems.

## V.1. Some related theorems.

Theorem V.1. Let f : [a, ) R be locally R-integrable. Then
Z
Z
Z
f L1 [a, ) R |f | dx < and R f dx = L
f d.
a

[a,)

Proof. () f L1 = f + , f L1 .
Define fn = f +[a,a+n) so that fn fn+1, and fn f + and fn L1.
Now for A := [a, ),
Z

f dx = lim

= lim

a+n

Za

f +dx

f +dx

R = L on bounded

[a,a+n)

= lim L fn d
n A
Z
= L lim fn d
n
ZA
= L f + d

def fn
MCT

Similarly,
Z

L
A

R
R
a f dx
f d =

Af

f f

d, so


d =

R
a

f f

dx =

f dx.

## () Define fn (x) = |f |[a,a+n] so that fn % |f | and fn is R-integrable.

(support is compact)

35

## Real Analysis Qual Seminar

Then

R
R [a+n]
fn dx
a
Z

L
A

exists and

|f | d = lim

= lim

R
R [a,a+n]
fn dx
a
fn d

[a,a+n] fn d,

so

MCT

a+n

fn dx

R = L on bounded

a+n

|f |dx
= lim
n a
Z
= R |f |dx

def of fn
def of impr int

<

hypothesis

shows that f L1 .

## (1) What is sufficient for F to be continuous? lim F (t) = F (t0), t0 .

tt0
Z
f
(2) What is sufficient for F to be differential? F 0 (t) =
t (x, t) d(x).
X

Royden:

## (1) (i) ft (x) = f (x, t) is a measurable function of x for each fixed t.

(ii) t, |f (x, t)| g(x) L1(X).
(iii) limtt0 f (x, t) = f (x, t0) for each x (i.e., f (x, t) is continuous in t
for each x).
The proof follows by applying DCT to f (x, tn), where tn t0 .
(2) (i) f
exists on X [a, b],
t
(ii)

f
t

## (iii) f is bounded on X [a, b],

(iv) For each fixed t, f is a measurable function of x.

36

## Real Analysis Qual Seminar

(3) Alternatively:
(i)

f
t

## exists on X [a, b],

f

(ii) t (x, t) g(x) L1(X) on X [a, b].

t need not be bounded

## Proof. (of the second version).

Pick any sequence {tn } [a, b] with tn t0 . Then define
hn (x) :=

## f (x, tn) f (x, t0)

.
tn t 0

Then f
(x, t0) = limn hn (x), so f
(x, t0) is measurable as a limit of meat
t
surable functions. It follows that f
t (x, t) is measurable. By the mean value
theorem, there is a t between tn and t0 for which
f (x, tn) f (x, t0) = (tn t0 ) f
t (x, t).
Then

f

|hn (x)| sup t (x, t) g(x),
t[a,b]

## since taking the supremum can only make it larger.

Invoke the dominated convergence theorem again and get
F (tn ) F (t0 )
F 0 (t0) = lim
= lim
tn t 0

hn (x) d(x) =

f
t (x, t)d(x).

lim g(x) = g(t)
xt

{xn } X, xn t.


37

## V.2. Solutions to the Nasty Integrals.

1. f : X [0, ] is measurable and
R be a constant. Show that

Proof.

## f d = c where 0 < c < . Let



 
Z
0 < < 1
f (x)
n log 1 +
d = c
lim
=1
n X

0 >1

case i) = 1.

By basic calculus, 1 +


f (x)
n

f (x)
n

n

n

## increases to ef (x) for each x, so

f (x) L1 (increasing).
gn (x) = log 1 +
R
R
Then limn X gn (x) = X f (x) = c by MCT.

## case ii) > 1.

R
Note that f (x) 0, X f d = c show f is finite -ae, i.e.:

M < and E M s.t. E = 0 and f E(x) M,

## (so E is where f is bounded). Since we can always find N such

f (x)
that n N = M
M
n < 1, n
n , -au. Were concerned
with n , so this means f (x)
n < 1 for our purposes. Hence,
> 1 = 1 > 0 implies

1
f (x)
0 n
< 1.
(V.1)
We need

n log 1 +

f (x)
n

 

so define
G(t) = n log 1 +

f (x),

 
t
n

t.

38

 1

0
nt
G (t) = n +t 1
 1

nt
n 1
 

t 1
= n
1

diff
drop the t
simp

<0

t = f (x),

f (x)
n

## so G is decreasing and G(t) < 0 for t > 0, i.e.,

 


f (x)
f (x).
n log 1 +
n

Set

1

< 1,

 
f (x)
gn (x) = n log 1 +
.
n
Since this is bounded by f and f L1 by hypothesis, DCT gives
Z
Z
lim
gn d =
lim gn d.
n

X n

## Now split the leading n and match the denominator:


  

 n
f
1
1
gn (x) = n n log 1 + n
=n
log 1 + nf
,
so that

lim 1 +

shows

f
n

 n

log 1 +
lim

= ef 6ae eM
f
n

n1


 n

f > 0 = log 1 +

= 0.

f
n

> 0,

## and ifRwe define RA := {f > 0}, then A > 0 because

Thus A f d = X f d, and


f
log 1 + n > 0 on A.

f d > 0.

39

## Real Analysis Qual Seminar

But then

log 1 +

f
n

 n

n1
because < 1 = 1 < 0.
Thus, lim gn = , so by Fatous Lemma,
Z
Z
Z
lim gn d lim gn =
=
lim
gn d = .
n

40

## Real Analysis Qual Seminar

ext
dx, for t > 0.
1 + x2
0
a) Show that F is well-defined as an improper Riemann integral and
as a Lebesgue integral.

2. Define F (t) =

xt

e
Riemann: 1+x
2 is continuous t, so it is R-integrable on any bounded
interval (a, b). So only remains to show the convergence of
Z a
ext
lim
1+x2 dx.
a

Z a
Z
xt
e
ba
=
dx
1+x2
0

b
0

ext
dx.
1+x2

## Thus, it suffices to consider a dominating function g(x):

ext
1+x2

Since

a
1

1
x2 dx

1
1+x2

1
x2

t > 0.

 a
a
= x1 1 = a1 + 1 1

Z a
ext
1+x2 dx 2 a,
0

## and thus it converges as a .

R ext
Lebesgue: R+ 1+x2 d exists because we can bound the integrand
as above.
b) Show F 00 (t) exists on (0, ).
xt

e
1
We have (t) = 1+x
2 L , so just need
order to use the theorem and get
Z xt
xe
0
F (t) =
dx.
2
1
+
x
0

0

(t)
t

xt

= xe
L1 in
1+x2

41

## Fix t > 0 and pick > 0 such that t > 0.

Observe: x ex for large enough x. Thus we pick M large enough
that x M = x ex , and split the integral:
Z

xext
dx =
1 + x2

M
0

xext
dx +
1 + x2

xext
dx.
1 + x2

Then we have
Z

M
0

xt
Z M
xe

xt

dx
xe dx,
1 + x2
0

## which as a continuous function over a compact space is clearly

finite, and hence the integrand is in L1(0, M ). Also,
Z

Z x xt
xt
e e
xe
dx

1 + x2 dx
1 + x2
M
Z (t)x
e

=
1 + x2 dx
M
Z (t)x
e
=
dx
1 + x2
0

Thus

xetx
1+x2

positive integrand

e(t)x
1+x2

L1 by (a).
 tx 

1
L . Now if t xe
=
1+x2

So t < 0 =

by choice of M

00

F (t) =

x2 etx
1+x2

L1 , well have

x2 etx
dx.
1 + x2

as before.

42

## c) (Extra credit) Show F (t) satisfies F 00 (t)+F (t) = 1t . Compute F (t).

We have F 00 (t) =
00

R x2 etx
1+x2

dx from (b), so

x2 etx + etx
dx
2
1
+
x
Z0
(1 + x2)etx
=
dx
2
1
+
x
Z0
=
etx dx
 0 1 tx 
= te
= 0 ( 1t ) = 1t .
0

F (t) + F (t) =

## Now solve the differential equation F 00 (t) + F (t) = 1t .

3. Let I be an open interval of R and suppose f : R R such that
x 7 ext f (x) is integrable for each fixed t I. Define F : I R by
Z
F (t) =
ext f (x) dx.
R

## Show that F is differentiable with derivative F 0 (t) =

each t I.

R xe

xt

f (x) dx at

Note: xetx f (x) may not be in L1! We would like to compute F 0 (t)
by

Z  tn x
e e t0 x
0
F (t0) = lim
f (x) dx,
(V.2)
n
tn t 0
where {tn } is a sequence in I with tn t0 .
To use DCT, we need to find g L1 such that
tx

e n e t0 x

tn t0 g(x) n.
Choose t0 I such that tn < t0 , n. This is possible, since otherwise
there would be a subsequence of {tn } converging to sup{t I}.
<
. I is open and tn t0 I.
By MVT,
tx

e n et0 x |xesx | |tn t0 | for some s [tn , t0].

43

## Real Analysis Qual Seminar

Since s < t0 ,

tx

tx

e n e t0 x
t0 x
e n et0 x t0 x

tn t0 xe = tn t0 f (x) xe f (x) ,

(V.3)

## and we have a bound which no longer depends on n.

0

To see that g(x) = xet x f (x) is integrable, split the integral: pick
> 0 such that t0 + I and choose M be such that
xM

x ex .

Now
Z M
Z M

t0 x

t0 x

xe f (x) dx M
e f (x) dx <
Z0
Z 0

t0 x

(t0 +)x

f (x) dx <
xe f (x) dx
e
M

t0 I, and
t0 + I.

R 0
R0
Thus we have 0 xet x f (x) dx < . For g(x) dx, pick > 0 such
that t0 I and let M be such that
xM
=
x ex ,
R
and proceed as for 0 g(x) dx.
Together, this gives g L1 . By (V.3), we can use the DCT in (V.2) to
obtain the result.

## 4.  Let f be a bounded measurable function on [0, ). Show that

Z
f (x)ext

F (t) =
dx, t > 0
x
0
is continuously differentiable on (0, ).
Let us denote the integrand by (x, t) := f (x)extx1/2.
We would like to find F 0 (t) by choosing any sequence {tn } such that

44

## Real Analysis Qual Seminar

tn t0 and computing

Z 
f (x)extn f (x)ext0
0

dx
F (t0) = lim
n 0
tn x
t0 x
Z
extn ext0 1/2
= lim
f (x)
x
dx
n 0
tn t 0
Since f is bounded, |f (x)| M . Then

xt

xt0
xtn
xt0
n

e

e
e

e
1/2
1/2

f (x)
x

M
x

tn t 0
.
tn t 0

## Since t ranges over (0, ) and tn t0 < , we can certainly pick a

strict lower bound of {tn }, i.e. a number such that inf{tn } > .
By MVT,
t x

e n et0 x xesx |tn t0 | for some s [tn, t0 ].
Since s > ,

t x
e n et0 x 1/2 sx 1/2

xe x
e x x1/2 L1 .
tn t 0 x

## To verify the integrability of the dominating function, note that

u = x1/2

dv = ex dx

2du = x1/2dx

v = ex

gives
Z

ex x1/2 = ex x1/2

= (0 0) + 4

= 4 2

= 2 .

+2

ex x1/2dx
2

eu du

put u =

45

Further,

R
0

ex x1/2dx = 2 =
Z

R
0

e x x1/2dx = 2

## extn ext0 1/2

x
dx
F (t0 ) = lim
f (x)
n 0
tn t 0
Z
extn ext0 1/2
=
lim f (x)
x
dx
tn t 0
0 n
 xt

Z
e ext0
=
f (x) lim
x1/2 dx
tt
t t0
0
Z0
 1/2
xt
=
f (x) t
e
x
dx
0
Z

=
f (x) xext x1/2 dx
0
Z
=
f (x)extx1/2 dx.
0

Thus

by DCT

G(t) =

f (x)extx1/2 dx.

## Since we are required to show that F is continuously differential, we

must show that G is continuous.
Notice that another u-substitution with u = (tx)1/2,
allows us to rewrite
G(t) =

2t

2
u du
t

= dx

f (x)xex dx.

## Thus, all we need to do is show the integral to be finite. Since f is

bounded by M , it suffices to show
Z

xex dx < .

46

Z
Z
x2
1
eu du
xe dx = 2
0
 0 u 
1
= 2 e 0
= 12 (0 (1))

1
2

## which is as finite as it gets.

Z

sin(x2t)
5.  Show F (t) =
dx is continuous on R.
2
1 + x
First, note that | sin x| 1 gives

sin(x2t)
1 L1 ,

1 + x2 1 + x2

## where the final inclusion is clear from

Z
dx
= [arctan x]
=
2
1 + x
Pick any {tn } R with tn t0. Then
Z
sin(x2tn )
lim F (tn ) = lim
dx
n
n 1 + x2
Z
sin(x2tn )
=
lim
dx
2
n
1
+
x

Z
sin(x2 limn tn )
=
dx
1 + x2

Z
sin(x2t0 )
=
dx
2
1 + x
= F (t0).


2 = .

def of F
DCT
contin of sin x, mult by x2
tn t0

Since this is true for all sequences tn t0 , we have limtt0 F (t) = F (t0 ).

47

Show that f 0.

R1
0

## By the Stone-Weierstra Theorem, there is a sequence of polynomials

unif
{Pk (x)} such that Pk (x) f (x). Then
Z
Z
k
Pk (x)f (x) dx [f (x)]2 dx
by uniformity.
But since any polynomial may be written
P (x) =

m
X
i=0

ai x i = a 0 + a 1 x + a 2 x 2 + + a m x m ,

## the linearity of the integral and the hypothesis

Z

P (x)f (x) dx = a0
0

dx + a1
0

1
0

R1
0

xn f (x) dx = 0 give

x dx + + am

1
0

= a0 0 + a 1 0 + a 2 0 + + a m 0
=0

So

Pk (x)f (x)dx = 0 k
Z
f 2dx = 0
f2 0

f2 0

f 0
7. Compute the limits
n
R
a) lim 0 1 + nx
sin
n

Note that

sin

x
n

x
n


1 and

dx

1+


x n
n

ex .

xm dx

48

## Real Analysis Qual Seminar


2

x n
Then 1 + n
n 2, so the DCT gives
1 + x2
Z 
Z
x

x
x n
x n
lim
1+
sin
dx =
lim 1 +
sin
dx
n 0
n
n
n
n
0 n
Z
=
ex sin(0) dx
0

= 0.

b) lim

n a

n(1 + n2x2 )1 dx

## We do a u-substitution with u = nx, du = n dx:

Z
Z
du
2 2 1
lim
n(1 + n x ) dx = lim
n a
n na 1 + u2
= lim [arctan u]
na
n

= lim

2

arctan na = 0

a=0
a>0.
a<0

## 8. a) Find the smallest constant c such that log(1 + et ) < c + t for

0 < t < .
First, observe that
1 + e t < ec et

1+et
et

< ec .

Note that
t

lim 1+e
et = 2 and
t0

Since

1+et
et

1+et
t
t e

lim

= 1.


d 1+et
d 1
=
+
1
= dtd et = et .
t
t
dt e
dt e

49

b) Does lim n1
n
f > 0?

R1
0

1+enf (x)
enf (x)
nf (x)

## enf (x) < 1 + enf (x) < 2e



nf (x)
nf (x) < log 1 + e
< nf (x) + c
(c = log 2)
Z 1
Z 1
Z 1


nf (x)
n
f (x) dx <
log 1 + e
dx < n
f (x) dx + c
0

Z 1
Z 1
Z 1


nf (x)
1
f (x) dx < n
log 1 + e
dx <
f (x) dx + nc
0

Z 1
Z 1
Z 1


nf (x)
1
f (x) dx lim n
log 1 + e
dx
f (x) dx.
n

## Since f is integrable by hypothesis, the Sandwich Theorem gives

Z 1
Z 1


lim n1
log 1 + enf (x) dx =
f (x) dx.
n

50

## VI. The Hahn-Banach Theorem and applications

[Folland] It is not obvious that there are any nonzero bounded functionals
on an arbitrary normed vector space. That such functionals exist in great
abundance is one of the fundamental theorems of functional analysis.
[Reed & Simon] In dealing with Banach spaces, one often needs to construct
linear functionals with certain properties. This is usually done in two steps:
first one defines the linear functional on a subspace of the Banach space where
it is easy to verify the desire properties; second, one appeals to (or proves) a
general theorem which says that any such functional can be extended to the
whole space while retaining the desired properties. One of the basic tools the
second step is the following theorem,
Theorem VI.1. Let X be a vector space and p : X R such that
(i) p(x) = p(x), 0, and

## If S is a subspace of X and there is a linear functional

f : S R such that f (s) p(s), s S, then f may be extended to
F : X R with F (x) p(x), x X, with F (s) = f (s) s S.
Proof. The idea of the proof is to first show that if x X but x
/ S, then
we can extend f to a functional having all the right properties on the space
spanned by x and S. We then use a Zorns Lemma / Hausdorff Maximality
argument to show that this process can be continued to extend f to the whole
space X.
(Sketch)
1. Consider the family
.
G := {g : D R .. g is linear; g(x) p(x), x D; g(s) = f (s), s S},
where D is any subspace of X which contains S. So G is roughly the
collection of all linear extensions of f which are bounded by p.
Now G is a poset under

= g1 .
g1 g 2

Dom(g1 )

## Real Analysis Qual Seminar

51

2. Use Hausdorff maximality Principle (or Zorn) to get a maximal linearly ordered subset {g } G which contains f . Define F on the union
of the domains of the {g } by F (x) = g (x) for x Dom(g ).
3. Show that this makes F into a well-defined linear functional which
extends f , and that F is maximal in that F G = F = G.
4. Show F is defined on all of X using the fact that F is maximal. Do
this by showing that a linear functional defined on a proper subspace
has a proper extension. (Hence F must be defined on all of X or it
wouldnt be maximal.)

Proposition VI.2. (Hausdorff Maximality Principle)
(A, ) is a poset = B A such that B is a maximal linearly ordered
subset. I.e., if C is linearly ordered, then B C A = C = B or C = A.
Of course, the HBT is also readily extendable to the complex case:
Theorem VI.3. Let X be a complex vector space and p a real-valued function defined on X satisfying
p(x + y) ||p(x) + ||p(y) x, y X, , C with || + || = 1.

## If S is a subspace of X and there is a complex linear functional

f : S R such that |f (s)| p(s), s S, then f may be extended to
F : X R with |F (x)| p(x), x X, with F (s) = f (s) s S.

52

## VI.1. Principle Applications of the HBT.

Most often, p(x) is taken to be the norm of the Banach space in question.

1. M is a closed
subspace of X and x X\M = f X such that
f (x) 6= 0, f M = 0. In fact, if = inf yM kx yk, f can be taken to
satisfy kf k = 1 and f (x) = .
Define f on M + Cx by f (y + x) = for y M, C. Then
but for m M ,

f (x) = f (0 + x) = 1 =

f (m) = f (m + 0) = 0 = 0.

For 6= 0, we have

|f (y + x)| = || || k1y + xk = ky + xk

## because = inf ky + xk k1y + xk (putting in 1 for y). Using

p(x) = kxk, apply the HBT to extend f from M + Cx to all of X.
2. If x 6= 0, x X, then f X such that kf k = 1 and f (x) = kxk.
M = {0} is trivially a closed subspace, so apply (1) with = kxk.
3. The bounded linear functionals on X separate points.
If x 6= y, then (2) shows f X such that f (x y) 6= 0. I.e.,
f (x) 6= f (y). This result indicates that X is BIG.
4. If x X, define x : X C by x(f ) = f (x), so x
X . Then
: x 7 x
is a linear isometry from X into X .
x(f + g) = (f + g)(x) = f (x) + g(x) = x
(f ) + x(g),
so x is linear. This verifies that x
X .
For (x) = x
, (ax + by) is defined by
by(f ) = f (ax + by) = af (x) + bf (y) = a
ax +
x(f ) + b
y (f ),

53

## Real Analysis Qual Seminar

by = a
so (ax+by) = ax +
x+b
y = a(x)+b(y) shows that : x 7 x
is linear. Finally,
|
x(f )| = |f (x)| kf k kxk
shows that
|
x(f )|
|f (x)|
kf k kxk
= sup
sup
= kxk.
kf k
kf k1 kf k
kf k1 kf k
kf k1

k
xk = sup

To get the reverse inequality, note that (2) provides a function f 0 for
which |
x(f0)| = |f0 (x)| = kxk and kf k = 1. Then
k
xk = sup |
x(f )| |
x(f0)| = kxk.
kf k=1

## We saw this for Hilbert spaces, but this example is applicable to

general Banach spaces and requires none of the Hilbert space machinery
(orthonormal basis, projection theorem, etc.) as the HBT takes care
of a lot.
VI.2. Corollaries to the HBT.
1. If X is a normed linear space, Y a subspace of X, and f Y , then
there exists F X extending f and satisfying kF kX = kf kY .
Proof. Apply HBT with p(x) = kf kY kxkX .

## 2. Let X be a Banach space. If X is separable, then X is separable.

Proof. Let {fn } be a dense set in X . Choose xn X, kxnk = 1 so
that
|fn (xn)| kfn k/2.
Let D be the set of all finite linear combinations of the kxnk with
rational coefficients. Since D is countable, we just need to show that
D is dense in X.
If D is not dense in X, then there is a y X\D and a linear functional
f X such that f (y) 6= 0 but f (x) = 0 for all x D, by application
(1).

54

## Let {fnk } be a subsequence of {fn} which converges to f . Then

kf fnk k |(f fnk )(xnk )|
= |(fnk )(xnk )|
kfnk k/2

## which implies kfnk k 0 as k . Thus f = 0. .<

Therefore D is dense and X is separable.

## The example of `1 and ` shows that the converse of this theorem

doesnt hold. In fact, this corollary offers a proof that `1 is not the
dual of ` , provided you can show ` is not separable.

55

## VII. The Baire Theorem and Consequences

= X, equivalently, iff U open in
Definition VII.1. D is dense in X iff D
X, U D 6= .

Definition VII.2. E is nowhere dense iff E is dense in X.
Definition VII.3. X is meager 10 iff X is a countable union of nowhere dense
sets.

Theorem VII.4. Let (X, d) be a complete metric space. (Note: can substitute LCH for complete, using f.i.p. def of compactness). Then
T
a) If {On }
n=1 are open dense subsets of X, then
n=1 On is dense in X.
b) No nonempty open subset of X is a countable union of nowhere dense
sets. In particular, X is not.

## Proof. The idea of the proof S

is straightforward: Suppose that X is a complete metric space and X =
n=1 An with each An nowhere dense. We will
construct a Cauchy sequence {xm } which stays away from each An so that
its limit point x (which is inSX by completeness) is in no An , thereby contradicting the statement X =
n=1 An .
Since A1 is nowhere dense, we can find x1
/ A1 and an open ball B1 about
x1 such that B1 A1 = , with the radius of B1 smaller than 1.
Since A2 is nowhere dense, we can find x2 B1 \A2. Let B2 be an open ball
about x2 such that B2 A2 = , with the radius of B2 smaller than 21 .
Proceeding inductively, we obtain a sequence {xn} where xn Bn1\An,
Bn Bn1, and Bn An = .
This sequence is Cauchy because n, m N implies that for xn, xm BN ,
N

## Let x = limn xn . Since xn BN for n N , we have

x BN BN 1.
S
Thus x
/ AN 1 for any N , which contradicts X =
n=1 An .
10Formerly called of first category.

56

## VII.1. Applications of Baire.

1. Q is NOT a G .
Proof. Suppose it were. Then Q =
Note: Q
in R, n.

n=1 On ,

## where the On are open.

On = Q On , n, so Q dense in R = On dense

Let {qk }
k=1 be an enumeration of Q.
Consider the singleton set (not the sequence!) {qn }.
Then {qn } closed = {qn } open. Also, {qn } is dense in R. (It
contains all but 1 of the rationals, so, e.g. the sequence {qn m1 }
m=1

## {qn} and has qn as a limit point.)

Then On {qn } is open and dense. To see dense, note that
On {qn } = Q {qn }.

Then
\

\
\

(On {qn} ) = ( On )
{qn } = Q (R Q) = ,

## So is dense in R, by Baires Theorem.

S
However, Q is an F : Q =
n=1 {qn }.

<
.

## 2. A meager set with Lebesgue measure 1.

We start by constructing a nowhere dense set with positive measure.
Let C be the Cantor set formed by removing an open interval of

length 3n+1
from each of the remaining 2n pieces, where 0 < < 1.
T
n
Then C =
n=0 C is closed as an intersection of closed sets and
C = C . To show C is nowhere dense, it suffices to show that C
contains no nonempty open set.11 If O 6= is open, then it contains

11If a closed set F contains no open set, then it is nowhere dense: we use contrapositive. Suppose (F )
is not dense. Then x F which is not a limit point of Fe. If no sequence in Fe converges to x, every point
of Fe must be at least > 0 away from x. Thus F contains B (x). The conditions are actually equivalent.

57

_
3

C0
C1
C2
C3
C4

## an interval A of positive length > 0. Each of the 2n intervals of Cn

(nth step of construction) are of length ` < 2n, so for large enough
N , we have ` < 2N < . Thus A is longer than any component of
CN and hence cannot be contained in CN or C . So C is nowhere dense.
Now we use the measure lemma12 to compute the measure of C :
C = 1

X
n=0

2n
3n+1

=1

X
n=0


2 n
3

=1

1
12/3

= 1 .

## Thus, every singleton {C } is a trivially a countable union of nowhere

dense sets which have positive Lebesgue measure.
12The Measure Lemmae.
Proposition VII.5. Let (X, A, ) be a measure space.
a) If {Ak } is an increasing sequence (Ak Ak+1 ) of sets of A, then (Ak ) = lim Ak .
b) If {Ak } is a decreasing sequence (Ak+1 Ak ) of sets of A, and An < for some n, then (Ak ) =
lim Ak .
(An < is necessary, else let Ak = (k, ).)
Proposition VII.6. Let (X, A) be a measurable space and let be a finitely additive function : A [0, ]
such that () = 0. Then is a measure if either
a) lim Ak = (Ak ) for each increasing sequence {Ak } A; or
b) lim Ak = 0 holds for each decreasing sequence {Ak } A with

Ak = .

58

## Real Analysis Qual Seminar

However, to complete
S this in the manner suggested by Royden, we
now consider P = k=1 C1/k .
!

[
P =
C1/k = lim (1 k1 ) = 1.
k=1

## So P is a union of a countable infinite collection of nowhere dense sets,

and P has Lebesgue measure 1.
VII.2. Consequences of Baire.
Theorem VII.7 (Open Mapping Theorem). X, Y are Banach spaces, T
L(X, Y ). Then T surjective = T open.
Proof. Nasty. (technical and longish)

## Corollary VII.8. If X, Y are Banach and T L(X, Y ) is bijective, then T

is an isomorphism.
Proof.
T is an isomorphism T 1 L(X, Y )

T 1 is continuous
T is open

## T bijective = T surjective = T open.


Definition VII.9. The graph of T is the set
.
(T ) = {(x, y) X Y .. y = T x}.
Definition VII.10. T L(X, Y ) is closed iff (T ) is a closed subspace of
X Y.

59

## Theorem VII.11 (Closed Graph Theorem). If X, Y are Banach spaces, and

the linear map T : X Y is closed, then T is bounded.
Proof. Let 1 , 2 be the projections of

.
(T ) = {(x, y) X Y .. y = T x}

to X and Y respectively:

1 (x, T x) = x

and

2 (x, T x) = T x.

## Obviously, 1 L((T ), X) and 2 L((T ), Y ).

Since X, Y are complete, so is X Y . Hence T closed implies (T ) is a
closed subspace (by definition), and a closed subspace of a complete space is
complete, so (T ) is also complete.
Now 1 : (T ) X is a bijection, hence an isomorphism by the corollary
to OMT, i.e., 11 is bounded.
But then T = 211 is bounded.

Figure 5. T as a composition of projections
Y

T


-1
1

## CGT restated: Let X, Y be Banach spaces, and T : X Y be linear.

(1) Then T is bounded (T ) is closed.
(2) Then T is continuous (xn x, T xn y = y = T x).
Note. For X, Y be Banach spaces, and S : X Y unbounded ,
(a) (S) is not complete,
(b) T : X (S) is closed but not bounded, and
(c) T 1 : (S) X is bounded and surjective but not open.

60

## Theorem VII.12 (Uniform Boundedness Principle, aka Banach-Steinhaus

Theoreom). Suppose X, Y are normed vector spaces and and A L(x, y).
a) If supT A kT xk < for all x in some nonmeager set D X, then
supT A kT k < .

.

## Proof of (a). Let

\
..
.
En = {x X . sup kT xk n} =
{x X .. kT xk n}.
T A

T A

Thus the En are closed as intersections of preimages of closed sets under continuous maps. Since supT A kT xk N, x D by hypothesis, EN contains
a nontrivial closed ball
B(x0, r), r > 0.
But then
kxk < R = (x x0) EN
= kT xk kT (x x0)k + kT x0k 2N,
so kxk r = kT xk 2N T A, which implies that
kT k

2N
r

< .


## Proof of (b). X is a nonempty Banach space, so X is nonmeager by Baire.

(Baires Theorem says that every complete metric space is nonmeager.) Then
just apply (a).

Rephrase of the UBP:
Either M < such that kT k M, T A,
or else supT A kT xk = , x in some dense G X.
Geometrically, either there is a ball B Y (with radius M , center 0) such
that every T A maps the unit ball of X into B, or there exists an x X
(in fact, a whole dense G of them) such that no ball in Y contains T x, for
all T A simultaneously.

61

## VII.3. Related Problems.

1. Find a Banach space X, a normed linear space Y , and a continuous
linear bijection f : X Y such that f 1 is not continuous. (Note: Y
better not be Banach!)


Let X := L2 [0, 1], k k2 and Y := L2[0, 1], k k1 , and define f :
X Y by f (x) = x.
f is clearly bijective and linear. To see f is continuous, it suffices to
show f is continuous at 0:
Let X and fix > 0. Since X, Y are finite measure spaces,
p < q = kkp kkq .
Thus, kk1 kk2 shows kk2 = kf ()k1 = kk1 .
To see f 1 is not continuous, consider {n} Y defined by
1
n (x) = p
.
x + 1/n
R1
Now n(x) n+1(x) (x) = 1x , n, x. Since 0 1x dx = 2 =
L1 , the MCT gives
Z 1
dx
p
lim kk1 = lim
x + 1/n
0
Z 1
=
(lim n)dx
0
Z 1
dx

=
x
0
= 2 = kk1.

However,
knk2 =

Z

1
0

dx
x + 1/n

1/2

= (log[1 + n])1/2

## shows that lim f 1(n) does not converge.

R1
R1
Indeed, 0 2 dx = 0 dx
= . Hence, lim f 1(n) 6= f 1() shows
x
that f 1 is not continuous.

62

## 2. Let V be a Banach space.

a) If V is infinite-dimensional, construct an unbounded linear operator
f :V V.
b) If V is finite-dimensional, show that every linear operator on it is
bounded.
Since V is a vector space, it has some basis {e } . Then if {c }
is any
Psequence (or net) in R, !f : V V such that for every element
v = v e of V , we have
X
f (v) =
c v e .
(Just define f (e ) = c e .)

## a) For V infinite-dimensional, {c } unbounded implies f unbounded.

b) For V finite-dimensional, kf k k max{c } for some k. (Basically,
finite = max{c } exists.)

## See any book on Quantum Dynamics or Functional Analysis for

the examples of the position & momentum operators (which are unbounded), e.g., Reed & Simon.