222 BRUCE K.

DRIVER

12. Hilbert Spaces

12.1. Hilbert Spaces Basics.
Denition 12.1. Let H be a complex vector space. An inner product on H is a
function, h, i : H H C, such that
(1) hax +by, zi = ahx, zi +bhy, zi i.e. x hx, zi is linear.
(2) hx, yi = hy, xi.
(3) kxk
2
hx, xi 0 with equality kxk
2
= 0 i x = 0.
Notice that combining properties (1) and (2) that x hz, xi is anti-linear for
xed z H, i.e.
hz, ax +byi = ahz, xi +

bhz, yi.
We will often nd the following formula useful:
kx +yk
2
= hx +y, x +yi = kxk
2
+kyk
2
+hx, yi +hy, xi
= kxk
2
+kyk
2
+ 2Rehx, yi (12.1)
Theorem 12.2 (Schwarz Inequality). Let (H, h, i) be an inner product space, then
for all x, y H
|hx, yi| kxkkyk
and equality holds i x and y are linearly dependent.
Proof. If y = 0, the result holds trivially. So assume that y 6= 0. First o notice
that if x = y for some C, then hx, yi = kyk
2
and hence
|hx, yi| = || kyk
2
= kxkkyk.
Moreover, in this case :=
hx,yi
kyk
2
.
Now suppose that x H is arbitrary, let z x kyk
2
hx, yiy. (So z is the
orthogonal projection of x onto y, see Figure 28.) Then
Figure 28. The picture behind the proof.
0 kzk
2
=

x
hx, yi
kyk
2
y

2
= kxk
2
+
|hx, yi|
2
kyk
4
kyk
2
2Rehx,
hx, yi
kyk
2
yi
= kxk
2

|hx, yi|
2
kyk
2
from which it follows that 0 kyk
2
kxk
2
|hx, yi|
2
with equality i z = 0 or
equivalently i x = kyk
2
hx, yiy.
ANALYSIS TOOLS WITH APPLICATIONS 223
Corollary 12.3. Let (H, h, i) be an inner product space and kxk :=
p
hx, xi. Then
k k is a norm on H. Moreover h, i is continuous on H H, where H is viewed as
the normed space (H, kk).
Proof. The only non-trivial thing to verify that kk is a norm is the triangle
inequality:
kx +yk
2
= kxk
2
+kyk
2
+ 2Rehx, yi kxk
2
+kyk
2
+ 2kxk kyk
= (kxk +kyk)
2
where we have made use of Schwarzs inequality. Taking the square root of this
inequality shows kx +yk kxk +kyk. For the continuity assertion:
|hx, yi hx
0
, y
0
i| = |hx x
0
, yi +hx
0
, y y
0
i|
kykkx x
0
k +kx
0
kky y
0
k
kykkx x
0
k + (kxk +kx x
0
k) ky y
0
k
= kykkx x
0
k +kxkky y
0
k +kx x
0
kky y
0
k
from which it follows that h, i is continuous.
Denition 12.4. Let (H, h, i) be an inner product space, we say x, y H are
orthogonal and write x y i hx, yi = 0. More generally if A H is a set,
x H is orthogonal to A and write x A i hx, yi = 0 for all y A. Let
A

= {x H : x A} be the set of vectors orthogonal to A. We also say that a

set S H is orthogonal if x y for all x, y S such that x 6= y. If S further
satises, kxk = 1 for all x S, then S is said to be orthonormal.
Proposition 12.5. Let (H, h, i) be an inner product space then
(1) (Parallelogram Law)
(12.2) kx +yk
2
+kx yk
2
= 2kxk
2
+ 2kyk
2
for all x, y H.
(2) (Pythagorean Theorem) If S H is a nite orthonormal set, then
(12.3) k
X
xS
xk
2
=
X
xS
kxk
2
.
(3) If A H is a set, then A

is a closed linear subspace of H.

Remark 12.6. See Proposition 12.37 in the appendix below for the converse of
the parallelogram law.
Proof. I will assume that H is a complex Hilbert space, the real case being
easier. Items 1. and 2. are proved by the following elementary computations:
kx +yk
2
+kx yk
2
= kxk
2
+kyk
2
+ 2Rehx, yi +kxk
2
+kyk
2
2Rehx, yi
= 2kxk
2
+ 2kyk
2
,
and
k
X
xS
xk
2
= h
X
xS
x,
X
yS
yi =
X
x,yS
hx, yi
=
X
xS
hx, xi =
X
xS
kxk
2
.
224 BRUCE K. DRIVER

Item 3. is a consequence of the continuity of h, i and the fact that

A

=
xA
ker(h, xi)
where ker(h, xi) = {y H : hy, xi = 0} a closed subspace of H.
Denition 12.7. A Hilbert space is an inner product space (H, h, i) such that
the induced Hilbertian norm is complete.
Example 12.8. Let (X, M, ) be a measure space then H := L
2
(X, M, ) with
inner product
(f, g) =
Z
X
f gd
is a Hilbert space. In Exercise 12.6 you will show every Hilbert space H is equiv-
alent to a Hilbert space of this form.
Denition 12.9. A subset C of a vector space X is said to be convex if for all
x, y C the line segment [x, y] := {tx + (1 t)y : 0 t 1} joining x to y is
contained in C as well. (Notice that any vector subspace of X is convex.)
Theorem 12.10. Suppose that H is a Hilbert space and M H be a closed convex
subset of H. Then for any x H there exists a unique y M such that
kx yk = d(x, M) = inf
zM
kx zk.
Moreover, if M is a vector subspace of H, then the point y may also be characterized
as the unique point in M such that (x y) M.
Proof. By replacing M by M x := {mx : m M} we may assume x = 0.
Let := d(0, M) = inf
mM
kmk and y, z M, see Figure 29.
Figure 29. The geometry of convex sets.
By the parallelogram law and the convexity of M,
(12.4) 2kyk
2
+2kzk
2
= ky+zk
2
+kyzk
2
= 4k
y +z
2
||
2
+kyzk
2
4
2
+kyzk
2
.
Hence if kyk = kzk = , then 2
2
+ 2
2
4
2
+ ky zk
2
, so that ky zk
2
= 0.
Therefore, if a minimizer for d(0, )|
M
exists, it is unique.
ANALYSIS TOOLS WITH APPLICATIONS 225
Existence. Let y
n
M be chosen such that ky
n
k =
n
d(0, M). Taking
y = y
m
and z = y
n
in Eq. (12.4) shows 2
2
m
+ 2
2
n
4
2
+ky
n
y
m
k
2
. Passing to
the limit m, n in this equation implies,
2
2
+ 2
2
4
2
+ limsup
m,n
ky
n
y
m
k
2
.
Therefore {y
n
}

n=1
is Cauchy and hence convergent. Because M is closed, y :=
lim
n
y
n
M and because kk is continuous,
kyk = lim
n
ky
n
k = = d(0, M).
So y is the desired point in M which is closest to 0.
Now for the second assertion we further assume that M is a closed subspace of
H and x H. Let y M be the closest point in M to x. Then for w M, the
function
g(t) kx (y +tw)k
2
= kx yk
2
2tRehx y, wi +t
2
kwk
2
has a minimum at t = 0. Therefore 0 = g
0
(0) = 2Rehx y, wi. Since w M is
arbitrary, this implies that (x y) M. Finally suppose y M is any point such
that (x y) M. Then for z M, by Pythagoreans theorem,
kx zk
2
= kx y +y zk
2
= kx yk
2
+ky zk
2
kx yk
2
which shows d(x, M)
2
kx yk
2
. That is to say y is the point in M closest to x.
Denition 12.11. Suppose that A : H H is a bounded operator. The adjoint
of A, denote A

, is the unique operator A

: H H such that hAx, yi = hx, A

yi.
(The proof that A

exists and is unique will be given in Proposition 12.16 below.)

A bounded operator A : H H is self - adjoint or Hermitian if A = A

.
Denition 12.12. Let H be a Hilbert space and M H be a closed subspace.
The orthogonal projection of H onto M is the function P
M
: H H such that for
x H, P
M
(x) is the unique element in M such that (x P
M
(x)) M.
Proposition 12.13. Let H be a Hilbert space and M H be a closed subspace.
The orthogonal projection P
M
satises:
(1) P
M
is linear (and hence we will write P
M
x rather than P
M
(x).
(2) P
2
M
= P
M
(P
M
is a projection).
(3) P

M
= P
M
, (P
M
is self-adjoint).
(4) Ran(P
M
) = M and ker(P
M
) = M

.
Proof.
(1) Let x
1
, x
2
H and F, then P
M
x
1
+P
M
x
2
M and
P
M
x
1
+P
M
x
2
(x
1
+x
2
) = [P
M
x
1
x
1
+(P
M
x
2
x
2
)] M

showing P
M
x
1
+P
M
x
2
= P
M
(x
1
+x
2
), i.e. P
M
is linear.
(2) Obviously Ran(P
M
) = M and P
M
x = x for all x M. Therefore P
2
M
=
P
M
.
226 BRUCE K. DRIVER

(3) Let x, y H, then since (x P

M
x) and (y P
M
y) are in M

,
hP
M
x, yi = hP
M
x, P
M
y +y P
M
yi
= hP
M
x, P
M
yi
= hP
M
x + (x P
M
), P
M
yi
= hx, P
M
yi.
(4) It is clear that Ran(P
M
) M. Moreover, if x M, then P
M
x = x implies
that Ran(P
M
) = M. Now x ker(P
M
) i P
M
x = 0 i x = x 0 M

.
Corollary 12.14. Suppose that M H is a proper closed subspace of a Hilbert
space H, then H = M M

.
Proof. Given x H, let y = P
M
x so that x y M

. Then x = y +(x y)
M +M

. If x M M

, then x x, i.e. kxk

2
= hx, xi = 0. So M M

= {0} .
Proposition 12.15 (Riesz Theorem). Let H

be the dual space of H (Notation

3.63). The map
(12.5) z H
j
h, zi H

is a conjugate linear isometric isomorphism.

Proof. The map j is conjugate linear by the axioms of the inner products.
Moreover, for x, z H,
|hx, zi| kxk kzk for all x H
with equality when x = z. This implies that kjzk
H
= kh, zik
H
= kzk . Therefore
j is isometric and this shows that j is injective. To nish the proof we must show
that j is surjective. So let f H

which we assume with out loss of generality is

non-zero. Then M = ker(f) a closed proper subspace of H. Since, by Corollary
12.14, H = M M

, f : H/M

= M

F is a linear isomorphism. This

shows that dim(M

) = 1 and hence H = M Fx
0
where x
0
M

\ {0} .
28
Choose z = x
0
M

such that f(x

0
) = hx
0
, zi. (So =

f(x
0
)/ kx
0
k
2
.) Then for
x = m+x
0
with m M and F,
f(x) = f(x
0
) = hx
0
, zi = hx
0
, zi = hm+x
0
, zi = hx, zi
which shows that f = jz.
Proposition 12.16 (Adjoints). Let H and K be Hilbert spaces and A : H K
be a bounded operator. Then there exists a unique bounded operator A

: K H
such that
(12.6) hAx, yi
K
= hx, A

yi
H
for all x H and y K.
Moreover (A+B)

= A

+

B

, A

:= (A

= A, kA

k = kAk and kA

Ak =
kAk
2
for all A, B L(H, K) and C.
28
Alternatively, choose x
0
M

\{0} such that f(x

0
) = 1. For x M

we have f(xx
0
) = 0
provided that := f(x). Therefore x x
0
M M

= {0} , i.e. x = x
0
. This again shows
that M

is spanned by x
0
.
ANALYSIS TOOLS WITH APPLICATIONS 227
Proof. For each y K, then map x hAx, yi
K
is in H

and therefore there

exists by Proposition 12.15 a unique vector z H such that
hAx, yi
K
= hx, zi
H
for all x H.
This shows there is a unique map A

: K H such that hAx, yi

K
= hx, A

(y)i
H
for all x H and y K. To nish the proof, we need only show A

is linear and
bounded. To see A

is linear, let y
1
, y
2
K and C, then for any x H,
hAx, y
1
+y
2
i
K
= hAx, y
1
i
K
+

hAx, y
2
i
K
= hx, A

(y
1
)i
K
+

hx, A

(y
2
)i
K
= hx, A

(y
1
) +A

(y
2
)i
K
and by the uniqueness of A

(y
1
+y
2
) we nd
A

(y
1
+y
2
) = A

(y
1
) +A

(y
2
).
This shows A

is linear and so we will now write A

y instead of A

(y). Since
hA

y, xi
H
= hx, A

yi
H
= hAx, yi
K
= hy, Axi
K
it follows that A

= A. he assertion that (A+B)

= A

+

B

is left to the
reader, see Exercise 12.1.
The following arguments prove the assertions about norms of A and A

:
kA

k = sup
kK:kkk=1
kA

kk = sup
kK:kkk=1
sup
hH:khk=1
|hA

k, hi|
= sup
hH:khk=1
sup
kK:kkk=1
|hk, Ahi| = sup
hH:khk=1
kAhk = kAk ,
kA

Ak kA

k kAk = kAk
2
and
kAk
2
= sup
hH:khk=1
|hAh, Ahi| = sup
hH:khk=1
|hh, A

Ahi|
sup
hH:khk=1
kA

Ahk = kA

Ak ,
wherein these arguments we have repeatedly made use of the Inequality.
Exercise 12.1. Let H, K, M be Hilbert space, A, B L(H, K), C L(K, M) and
C. Show (A+B)

= A

+

B

and (CA)

= A

L(M, H).
Exercise 12.2. Let H = C
n
and K = C
m
equipped with the usual inner products,
i.e. hz, wi
H
= z w for z, w H. Let A be an mn matrix thought of as a linear
operator from H to K. Show the matrix associated to A

: K H is the conjugate
transpose of A.
Exercise 12.3. Let K : L
2
() L
2
() be the operator dened in Exercise 9.12.
Show K

: L
2
() L
2
() is the operator given by
K

g(y) =
Z
X

k(x, y)g(x)d(x).
Denition 12.17. {u

}
A
H is an orthonormal set if u

for all 6=
and ku

k = 1.
228 BRUCE K. DRIVER

Proposition 12.18 (Bessels Inequality). Let {u

}
A
be an orthonormal set, then
(12.7)
X
A
|hx, u

i|
2
kxk
2
for all x H.
In particular the set { A : hx, u

i 6= 0} is at most countable for all x H.

Proof. Let A be any nite set. Then
0 kx
X

hx, u

iu

k
2
= kxk
2
2Re
X

hx, u

i hu

, xi +
X

|hx, u

i|
2
= kxk
2

|hx, u

i|
2
showing that
X

|hx, u

i|
2
kxk
2
.
Taking the supremum of this equation of A then proves Eq. (12.7).
Proposition 12.19. Suppose A H is an orthogonal set. Then s =
P
vA
v
exists in H i
P
vA
kvk
2
< . (In particular A must be at most a countable set.)
Moreover, if
P
vA
kvk
2
< , then
(1) ksk
2
=
P
vA
kvk
2
and
(2) hs, xi =
P
vA
hv, xi for all x H.
Similarly if {v
n
}

n=1
is an orthogonal set, then s =

P
n=1
v
n
exists in H i

P
n=1
kv
n
k
2
< . In particular if

P
n=1
v
n
exists, then it is independent of rearrange-
ments of {v
n
}

n=1
.
Proof. Suppose s =
P
vA
v exists. Then there exists A such that
X
v
kvk
2
=

X
v
v

2
1
for all A\ ,wherein the rst inequality we have used Pythagoreans theorem.
Taking the supremum over such shows that
P
vA\
kvk
2
1 and therefore
X
vA
kvk
2
1 +
X
v
kvk
2
< .
Conversely, suppose that
P
vA
kvk
2
< . Then for all > 0 there exists

A
such that if A\

,
(12.8)

X
v
v

2
=
X
v
kvk
2
<
2
.
Hence by Lemma 3.72,
P
vA
v exists.
For item 1, let

be as above and set s

:=
P
v
v. Then
|ksk ks

k| ks s

k <
and by Eq. (12.8),
0
X
vA
kvk
2
ks

k
2
=
X
v/

kvk
2

2
.
ANALYSIS TOOLS WITH APPLICATIONS 229
Letting 0 we deduce from the previous two equations that ks

k ksk and
ks

k
2

P
vA
kvk
2
as 0 and therefore ksk
2
=
P
vA
kvk
2
.
Item 2. is a special case of Lemma 3.72.
For the nal assertion, let s
N

N
P
n=1
v
n
and suppose that lim
N
s
N
= s exists
in H and in particular {s
N
}

N=1
is Cauchy. So for N > M.
N
X
n=M+1
kv
n
k
2
= ks
N
s
M
k
2
0 as M, N
which shows that

P
n=1
kv
n
k
2
is convergent, i.e.

P
n=1
kv
n
k
2
< .
Remark: We could use the last result to prove Item 1. Indeed, if
P
vA
kvk
2
<
, then A is countable and so we may writer A = {v
n
}

n=1
. Then s = lim
N
s
N
with s
N
as above. Since the norm kk is continuous on H, we have
ksk
2
= lim
N
ks
N
k
2
= lim
N

N
X
n=1
v
n

2
= lim
N
N
X
n=1
kv
n
k
2
=

X
n=1
kv
n
k
2
=
X
vA
kvk
2
.
Corollary 12.20. Suppose H is a Hilbert space, H is an orthonormal set and
M = span . Then
P
M
x =
X
u
hx, uiu, (12.9)
X
u
|hx, ui|
2
= kP
M
xk
2
and (12.10)
X
u
hx, uihu, yi = hP
M
x, yi (12.11)
for all x, y H.
Proof. By Bessels inequality,
P
u
|hx, ui|
2
kxk
2
for all x H and hence
by Proposition 12.18, Px :=
P
u
hx, uiu exists in H and for all x, y H,
(12.12) hPx, yi =
X
u
hhx, uiu, yi =
X
u
hx, uihu, yi.
Taking y in Eq. (12.12) gives hPx, yi = hx, yi, i.e. that hx Px, yi = 0
for all y . So (x Px) span and by continuity we also have (x Px)
M = span . Since Px is also in M, it follows from the denition of P
M
that
Px = P
M
x proving Eq. (12.9). Equations (12.10) and (12.11) now follow from
(12.12), Proposition 12.19 and the fact that hP
M
x, yi = hP
2
M
x, yi = hP
M
x, P
M
yi
for all x, y H.
12.2. Hilbert Space Basis.
Denition 12.21 (Basis). Let H be a Hilbert space. A basis of H is a maximal
orthonormal subset H.
Proposition 12.22. Every Hilbert space has an orthonormal basis.
230 BRUCE K. DRIVER

Proof. Let F be the collection of all orthonormal subsets of H ordered by

inclusion. If F is linearly ordered then is an upper bound. By Zorns
Lemma (see Theorem B.7) there exists a maximal element F.
An orthonormal set H is said to be complete if

= {0} . That is to say

if hx, ui = 0 for all u then x = 0.
Lemma 12.23. Let be an orthonormal subset of H then the following are equiv-
alent:
(1) is a basis,
(2) is complete and
(3) span = H.
Proof. If is not complete, then there exists a unit vector x

\ {0} .
The set {x} is an orthonormal set properly containing , so is not maximal.
Conversely, if is not maximal, there exists an orthonormal set
1
H such that
&
1
. Then if x
1
\ , we have hx, ui = 0 for all u showing is not
complete. This proves the equivalence of (1) and (2). If is not complete and
x

\ {0} , then span x

which is a proper subspace of H. Conversely

if span is a proper subspace of H,

= span

is a non-trivial subspace by
Corollary 12.14 and is not complete. This shows that (2) and (3) are equivalent.
Theorem 12.24. Let H be an orthonormal set. Then the following are
equivalent:
(1) is complete or equivalently a basis.
(2) x =
P
u
hx, uiu for all x H.
(3) hx, yi =
P
u
hx, ui hu, yi for all x, y H.
(4) kxk
2
=
P
u
|hx, ui|
2
for all x H.
Proof. Let M = span and P = P
M
.
(1) (2) By Corollary 12.20,
P
u
hx, uiu = P
M
x. Therefore
x
X
u
hx, uiu = x P
M
x M

= {0} .
(2) (3) is a consequence of Proposition 12.19.
(3) (4) is obvious, just take y = x.
(4) (1) If x

, then by 4), kxk = 0, i.e. x = 0. This shows that is

complete.
Proposition 12.25. A Hilbert space H is separable i H has a countable ortho-
normal basis H. Moreover, if H is separable, all orthonormal bases of H are
countable.
Proof. Let D H be a countable dense set D = {u
n
}

n=1
. By Gram-Schmidt
process there exists = {v
n
}

n=1
an orthonormal set such that span{v
n
: n =
1, 2 . . . , N} span{u
n
: n = 1, 2 . . . , N}. So if hx, v
n
i = 0 for all n then hx, u
n
i = 0
for all n. Since D H is dense we may choose {w
k
} D such that x = lim
k
w
k
and therefore hx, xi = lim
k
hx, w
k
i = 0. That is to say x = 0 and is complete.
ANALYSIS TOOLS WITH APPLICATIONS 231
Conversely if H is a countable orthonormal basis, then the countable set
D =
_
_
_
X
u
a
u
u : a
u
Q+iQ : #{u : a
u
6= 0} <
_
_
_
is dense in H.
Finally let = {u
n
}

n=1
be an orthonormal basis and
1
H be another ortho-
normal basis. Then the sets
B
n
= {v
1
: hv, u
n
i 6= 0}
are countable for each n N and hence B :=

S
n=1
B
n
is a countable subset of
1
.
Suppose there exists v
1
\ B, then hv, u
n
i = 0 for all n and since = {u
n
}

n=1
is an orthonormal basis, this implies v = 0 which is impossible since kvk = 1.
Therefore
1
\ B = and hence
1
= B is countable.
Denition 12.26. A linear map U : H K is an isometry if kUxk
K
= kxk
H
for all x H and U is unitary if U is also surjective.
Exercise 12.4. Let U : H K be a linear map, show the following are equivalent:
(1) U : H K is an isometry,
(2) hUx, Ux
0
i
K
= hx, x
0
i
H
for all x, x
0
H, (see Eq. (12.16) below)
(3) U

U = id
H
.
Exercise 12.5. Let U : H K be a linear map, show the following are equivalent:
(1) U : H K is unitary
(2) U

U = id
H
and UU

= id
K
.
(3) U is invertible and U
1
= U

.
Exercise 12.6. Let H be a Hilbert space. Use Theorem 12.24 to show there exists
a set X and a unitary map U : H
2
(X). Moreover, if H is separable and
dim(H) = , then X can be taken to be N so that H is unitarily equivalent to

2
=
2
(N).
Remark 12.27. Suppose that {u
n
}

n=1
is a total subset of H, i.e. span{u
n
} = H.
Let {v
n
}

n=1
be the vectors found by performing Gram-Schmidt on the set {u
n
}

n=1
.
Then {v
n
}

n=1
is an orthonormal basis for H.
Example 12.28. (1) Let H = L
2
([, ], dm) = L
2
((, ), dm) and
e
n
() =
1

2
e
in
for n Z. Simple computations show := {e
n
}
nZ
is an
orthonormal set. We now claim that is an orthonormal basis. To see this
recall that C
c
((, )) is dense in L
2
((, ), dm). Any f C
c
((, ))
may be extended to be a continuous 2 periodic function on R and hence
by Exercise 11.9), f may uniformly (and hence in L
2
) be approximated by
a trigonometric polynomial. Therefore is a total orthonormal set, i.e.
is an orthonormal basis.
(2) Let H = L
2
([1, 1], dm) and A := {1, x, x
2
, x
3
. . . }. Then A is total in
H by the Stone-Weierstrass theorem and a similar argument as in the rst
example or directly from Exercise 11.12. The result of doing Gram-Schmidt
on this set gives an orthonormal basis of H consisting of the Legendre
Polynomials.
232 BRUCE K. DRIVER

(3) Let H = L
2
(R, e

1
2
x
2
dx).Exercise 11.12 implies A := {1, x, x
2
, x
3
. . . } is
total in H and the result of doing Gram-Schmidt on A now gives an ortho-
normal basis for H consisting of Hermite Polynomials.
Remark 12.29 (An Interesting Phenomena). Let H = L
2
([1, 1], dm) and B :=
{1, x
3
, x
6
, x
9
, . . . }. Then again A is total in H by the same argument as in item 2.
Example 12.28. This is true even though B is a proper subset of A. Notice that A
is an algebraic basis for the polynomials on [1, 1] while B is not! The following
computations may help relieve some of the readers anxiety. Let f L
2
([1, 1], dm),
then, making the change of variables x = y
1/3
, shows that
(12.13)
Z
1
1
|f(x)|
2
dx =
Z
1
1

f(y
1/3
)

2
1
3
y
2/3
dy =
Z
1
1

f(y
1/3
)

2
d(y)
where d(y) =
1
3
y
2/3
dy. Since ([1, 1]) = m([1, 1]) = 2, is a nite mea-
sure on [1, 1] and hence by Exercise 11.12 A := {1, x, x
2
, x
3
. . . } is a total in
L
2
([1, 1], d). In particular for any > 0 there exists a polynomial p(y) such that
Z
1
1

f(y
1/3
) p(y)

2
d(y) <
2
.
However, by Eq. (12.13) we have

2
>
Z
1
1

f(y
1/3
) p(y)

2
d(y) =
Z
1
1

f(x) p(x
3
)

2
dx.
Alternatively, if f C([1, 1]), then g(y) = f(y
1/3
) is back in C([1, 1]). There-
fore for any > 0, there exists a polynomial p(y) such that
> kg pk
u
= sup{|g(y) p(y)| : y [1, 1]}
= sup

g(x
3
) p(x
3
)

: x [1, 1]

= sup

f(x) p(x
3
)

: x [1, 1]

.
This gives another proof the polynomials in x
3
are dense in C([1, 1]) and hence
in L
2
([1, 1]).
12.3. Weak Convergence. Suppose H is an innite dimensional Hilbert space
and {x
n
}

n=1
is an orthonormal subset of H. Then, by Eq. (12.1), kx
n
x
m
k
2
= 2
for all m 6= n and in particular, {x
n
}

n=1
has no convergent subsequences. From
this we conclude that C := {x H : kxk 1} , the closed unit ball in H, is not
compact. To overcome this problems it is sometimes useful to introduce a weaker
topology on X having the property that C is compact.
Denition 12.30. Let (X, kk) be a Banach space and X

be its continuous dual.

The weak topology,
w
, on X is the topology generated by X

. If {x
n
}

n=1
X
is a sequence we will write x
n
w
x as n to mean that x
n
x in the weak
topology.
Because
w
= (X

)
kk
:= ({kx k : x X} , it is harder for a function
f : X F to be continuous in the
w
topology than in the norm topology,
kk
.
In particular if : X F is a linear functional which is
w
continuous, then is

kk
continuous and hence X

.
Proposition 12.31. Let {x
n
}

n=1
X be a sequence, then x
n
w
x X as n
i (x) = lim
n
(x
n
) for all X

.
ANALYSIS TOOLS WITH APPLICATIONS 233
Proof. By denition of
w
, we have x
n
w
x X i for all X

and > 0
there exists an N N such that |(x) (x
n
)| < for all n N and .
This later condition is easily seen to be equivalent to (x) = lim
n
(x
n
) for all
X

.
The topological space (X,
w
) is still Hausdor, however to prove this one needs
to make use of the Hahn Banach Theorem 18.16 below. For the moment we will
concentrate on the special case where X = H is a Hilbert space in which case
H

= {
z
:= h, zi : z H} , see Propositions 12.15. If x, y H and z := y x 6= 0,
then
0 < := kzk
2
=
z
(z) =
z
(y)
z
(x).
Thus V
x
:= {w H : |
z
(x)
z
(w)| < /2} and V
y
:= {w H : |
z
(y)
z
(w)| < /2}
are disjoint sets from
w
which contain x and y respectively. This shows that (H,
w
)
is a Hausdor space. In particular, this shows that weak limits are unique if they
exist.
Remark 12.32. Suppose that H is an innite dimensional Hilbert space {x
n
}

n=1
is
an orthonormal subset of H. Then Bessels inequality (Proposition 12.18) implies
x
n
w
0 H as n . This points out the fact that if x
n
w
x H as n , it
is no longer necessarily true that kxk = lim
n
kx
n
k . However we do always have
kxk liminf
n
kx
n
k because,
kxk
2
= lim
n
hx
n
, xi liminf
n
[kx
n
k kxk] = kxk liminf
n
kx
n
k .
Proposition 12.33. Let H be a Hilbert space, H be an orthonormal basis for
H and {x
n
}

n=1
H be a bounded sequence, then the following are equivalent:
(1) x
n
w
x H as n .
(2) hx, yi = lim
n
hx
n
, yi for all y H.
(3) hx, yi = lim
n
hx
n
, yi for all y .
Moreover, if c
y
:= lim
n
hx
n
, yi exists for all y , then
P
y
|c
y
|
2
< and
x
n
w
x :=
P
y
c
y
y H as n .
Proof. 1. = 2. This is a consequence of Propositions 12.15 and 12.31. 2. =
3. is trivial.
3. = 1. Let M := sup
n
kx
n
k and H
0
denote the algebraic span of . Then for
y H and z H
0
,
|hx x
n
, yi| |hx x
n
, zi| +|hx x
n
, y zi| |hx x
n
, zi| + 2M ky zk .
Passing to the limit in this equation implies limsup
n
|hx x
n
, yi| 2M ky zk
which shows limsup
n
|hx x
n
, yi| = 0 since H
0
is dense in H.
To prove the last assertion, let . Then by Bessels inequality (Proposition
12.18),
X
y
|c
y
|
2
= lim
n
X
y
|hx
n
, yi|
2
liminf
n
kx
n
k
2
M
2
.
Since was arbitrary, we conclude that
P
y
|c
y
|
2
M < and hence we
may dene x :=
P
y
c
y
y. By construction we have
hx, yi = c
y
= lim
n
hx
n
, yi for all y
and hence x
n
w
x H as n by what we have just proved.
234 BRUCE K. DRIVER

Theorem 12.34. Suppose that {x

n
}

n=1
H is a bounded sequence. Then there
exists a subsequence y
k
:= x
n
k
of {x
n
}

n=1
and x X such that y
k
w
x as k .
Proof. This is a consequence of Proposition 12.33 and a Cantors diagonalization
argument which is left to the reader, see Exercise 12.14.
Theorem 12.35 (Alaoglus Theorem for Hilbert Spaces). Suppose that H is a
separable Hilbert space, C := {x H : kxk 1} is the closed unit ball in H and
{e
n
}

n=1
is an orthonormal basis for H. Then
(12.14) (x, y) :=

X
n=1
1
2
n
|hx y, e
n
i|
denes a metric on C which is compatible with the weak topology on C,
C
:=
(
w
)
C
= {V C : V
w
} . Moreover (C, ) is a compact metric space.
Proof. The routine check that is a metric is left to the reader. Let

be
the topology on C induced by . For any y H and n N, the map x H
hx y, e
n
i = hx, e
n
i hy, e
n
i is
w
continuous and since the sum in Eq. (12.14) is
uniformly convergent for x, y C, it follows that x (x, y) is
C
continuous.
This implies the open balls relative to are contained in
C
and therefore

C
. For the converse inclusion, let z H, x
z
(x) = hz, xi be an element of
H

, and for N N let z

N
:=
P
N
n=1
hz, e
n
ie
n
. Then
z
N
=
P
N
n=1
hz, e
n
i
e
n
is
continuous, being a nite linear combination of the
e
n
which are easily seen to be
continuous. Because z
N
z as N it follows that
sup
xC
|
z
(x)
z
N
(x)| = kz z
N
k 0 as N .
Therefore
z
|
C
is continuous as well and hence
C
= (
z
|
C
: z H)

.
The last assertion follows directly from Theorem 12.34 and the fact that sequen-
tial compactness is equivalent to compactness for metric spaces.
Theorem 12.36 (Weak and Strong Dierentiability). Suppose that f L
2
(R
n
)
and v R
n
\ {0} . Then the following are equivalent:
(1) There exists {t
n
}

n=1
R\ {0} such that lim
n
t
n
= 0 and
sup
n

f( +t
n
v) f()
t
n

2
< .
(2) There exists g L
2
(R
n
) such that hf,
v
i = hg, i for all C

c
(R
n
).
(3) There exists g L
2
(R
n
) and f
n
C

c
(R
n
) such that f
n
L
2
f and
v
f
n
L
2
g
as n .
(4) There exists g L
2
such that
f( +tv) f()
t
L
2
g as t 0.
(See Theorem 19.7 for the L
p
generalization of this theorem.)
Proof. 1. = 2. We may assume, using Theorem 12.34 and passing to a
subsequence if necessary, that
f(+t
n
v)f()
t
n
w
g for some g L
2
(R
n
). Now for
ANALYSIS TOOLS WITH APPLICATIONS 235
C

c
(R
n
),
hg, i = lim
n
h
f( +t
n
v) f()
t
n
, i = lim
n
hf,
( t
n
v) ()
t
n
i
= hf, lim
n
( t
n
v) ()
t
n
i = hf,
v
i,
wherein we have used the translation invariance of Lebesgue measure and the dom-
inated convergence theorem.
2. = 3. Let C

c
(R
n
, R) such that
R
R
n
(x)dx = 1 and let
m
(x) =
m
n
(mx), then by Proposition 11.24, h
m
:=
m
f C

(R
n
) for all m and

v
h
m
(x) =
v

m
f(x) =
Z
R
n

v

m
(x y)f(y)dy = hf,
v
[
m
(x )]i
= hg,
m
(x )i =
m
g(x).
By Theorem 11.21, h
m
f L
2
(R
n
) and
v
h
m
=
m
g g in L
2
(R
n
) as
m . This shows 3. holds except for the fact that h
m
need not have compact
support. To x this let C

c
(R
n
, [0, 1]) such that = 1 in a neighborhood of 0
and let

(x) = (x) and (

v
)

(x) := (
v
) (x). Then

v
(

h
m
) =
v

h
m
+

v
h
m
= (
v
)

h
m
+

v
h
m
so that

h
m
h
m
in L
2
and
v
(

h
m
)
v
h
m
in L
2
as 0. Let f
m
=

m
h
m
where
m
is chosen to be greater than zero but small enough so that
k
m
h
m
h
m
k
2
+k
v
(
m
h
m
)
v
h
m
k
2
< 1/m.
Then f
m
C

c
(R
n
), f
m
f and
v
f
m
g in L
2
as m .
3. = 4. By the fundamental theorem of calculus

tv
f
m
(x) f
m
(x)
t
=
f
m
(x +tv) f
m
(x)
t
=
1
t
Z
1
0
d
ds
f
m
(x +stv)ds =
Z
1
0
(
v
f
m
) (x +stv)ds. (12.15)
Let
G
t
(x) :=
Z
1
0

stv
g(x)ds =
Z
1
0
g(x +stv)ds
which is dened for almost every x and is in L
2
(R
n
) by Minkowskis inequality for
integrals, Theorem 9.27. Therefore

tv
f
m
(x) f
m
(x)
t
G
t
(x) =
Z
1
0
[(
v
f
m
) (x +stv) g(x +stv)] ds
and hence again by Minkowskis inequality for integrals,

tv
f
m
f
m
t
G
t

Z
1
0
k
stv
(
v
f
m
)
stv
gk
2
ds =
Z
1
0
k
v
f
m
gk
2
ds.
Letting m in this equation implies (
tv
f f) /t = G
t
a.e. Finally one more
application of Minkowskis inequality for integrals implies,

tv
f f
t
g

2
= kG
t
gk
2
=

Z
1
0
(
stv
g g) ds

Z
1
0
k
stv
g gk
2
ds.
236 BRUCE K. DRIVER

By the dominated convergence theorem and Proposition 11.13, the latter term tends
to 0 as t 0 and this proves 4. The proof is now complete since 4. =1. is trivial.
12.4. Supplement 1: Converse of the Parallelogram Law.
Proposition 12.37 (Parallelogram Law Converse). If (X, kk) is a normed space
such that Eq. (12.2) holds for all x, y X, then there exists a unique inner product
on h, i such that kxk :=
p
hx, xi for all x X. In this case we say that kk is a
Hilbertian norm.
Proof. If kk is going to come from an inner product h, i, it follows from Eq.
(12.1) that
2Rehx, yi = kx +yk
2
kxk
2
kyk
2
and
2Rehx, yi = kx yk
2
kxk
2
kyk
2
.
Subtracting these two equations gives the polarization identity,
4Rehx, yi = kx +yk
2
kx yk
2
.
Replacing y by iy in this equation then implies that
4Imhx, yi = kx +iyk
2
kx iyk
2
from which we nd
(12.16) hx, yi =
1
4
X
G
kx +yk
2
where G = {1, i} a cyclic subgroup of S
1
C. Hence if h, i is going to exists
we must dene it by Eq. (12.16).
Notice that
hx, xi =
1
4
X
G
kx +xk
2
= kxk
2
+ikx +ixk
2
ikx ixk
2
= kxk
2
+i

1 +i|
2

kxk
2
i

1 i|
2

kxk
2
= kxk
2
.
So to nish the proof of (4) we must show that hx, yi in Eq. (12.16) is an inner
product. Since
4hy, xi =
X
G
ky +xk
2
=
X
G
k (y +x) k
2
=
X
G
ky +
2
xk
2
= ky +xk
2
+k y +xk
2
+ikiy xk
2
ik iy xk
2
= kx +yk
2
+kx yk
2
+ikx iyk
2
ikx +iyk
2
= 4hx, yi
it suces to show x hx, yi is linear for all y H. (The rest of this proof may
safely be skipped by the reader.) For this we will need to derive an identity from
ANALYSIS TOOLS WITH APPLICATIONS 237
Eq. (12.2). To do this we make use of Eq. (12.2) three times to nd
kx +y +zk
2
= kx +y zk
2
+ 2kx +yk
2
+ 2kzk
2
= kx y zk
2
2kx zk
2
2kyk
2
+ 2kx +yk
2
+ 2kzk
2
= ky +z xk
2
2kx zk
2
2kyk
2
+ 2kx +yk
2
+ 2kzk
2
= ky +z +xk
2
+ 2ky +zk
2
+ 2kxk
2
2kx zk
2
2kyk
2
+ 2kx +yk
2
+ 2kzk
2
.
Solving this equation for kx +y +zk
2
gives
(12.17) kx +y +zk
2
= ky +zk
2
+kx +yk
2
kx zk
2
+kxk
2
+kzk
2
kyk
2
.
Using Eq. (12.17), for x, y, z H,
4 Rehx +z, yi = kx +z +yk
2
kx +z yk
2
= ky +zk
2
+kx +yk
2
kx zk
2
+kxk
2
+kzk
2
kyk
2

kz yk
2
+kx yk
2
kx zk
2
+kxk
2
+kzk
2
kyk
2

= kz +yk
2
kz yk
2
+kx +yk
2
kx yk
2
= 4 Rehx, yi + 4 Rehz, yi. (12.18)
Now suppose that G, then since || = 1,
4hx, yi =
1
4
X
G
kx +yk
2
=
1
4
X
G
kx +
1
yk
2
=
1
4
X
G
kx +yk
2
= 4hx, yi (12.19)
where in the third inequality, the substitution was made in the sum. So Eq.
(12.19) says hix, yi = ihix, yi and hx, yi = hx, yi. Therefore
Imhx, yi = Re (ihx, yi) = Rehix, yi
which combined with Eq. (12.18) shows
Imhx +z, yi = Rehix iz, yi = Rehix, yi + Rehiz, yi
= Imhx, yi + Imhz, yi
and therefore (again in combination with Eq. (12.18)),
hx +z, yi = hx, yi +hz, yi for all x, y H.
Because of this equation and Eq. (12.19) to nish the proof that x hx, yi is
linear, it suces to show hx, yi = hx, yi for all > 0. Now if = m N, then
hmx, yi = hx + (m1)x, yi = hx, yi +h(m1)x, yi
so that by induction hmx, yi = mhx, yi. Replacing x by x/m then shows that
hx, yi = mhm
1
x, yi so that hm
1
x, yi = m
1
hx, yi and so if m, n N, we nd
h
n
m
x, yi = nh
1
m
x, yi =
n
m
hx, yi
so that hx, yi = hx, yi for all > 0 and Q. By continuity, it now follows that
hx, yi = hx, yi for all > 0.
238 BRUCE K. DRIVER

12.5. Supplement 2. Non-complete inner product spaces. Part of Theorem

12.24 goes through when H is a not necessarily complete inner product space. We
have the following proposition.
Proposition 12.38. Let (H, h, i) be a not necessarily complete inner product space
and H be an orthonormal set. Then the following two conditions are equivalent:
(1) x =
P
u
hx, uiu for all x H.
(2) kxk
2
=
P
u
|hx, ui|
2
for all x H.
Moreover, either of these two conditions implies that H is a maximal ortho-
normal set. However H being a maximal orthonormal set is not sucient to
conditions for 1) and 2) hold!
Proof. As in the proof of Theorem 12.24, 1) implies 2). For 2) implies 1) let
and consider

x
X
u
hx, uiu

2
= kxk
2
2
X
u
|hx, ui|
2
+
X
u
|hx, ui|
2
= kxk
2

X
u
|hx, ui|
2
.
Since kxk
2
=
P
u
|hx, ui|
2
, it follows that for every > 0 there exists

such
that for all such that

x
X
u
hx, uiu

2
= kxk
2

X
u
|hx, ui|
2
<
showing that x =
P
u
hx, uiu.
Suppose x = (x
1
, x
2
, . . . , x
n
, . . . )

. If 2) is valid then kxk

2
= 0, i.e. x = 0. So
is maximal. Let us now construct a counter example to prove the last assertion.
Take H = Span{e
i
}

i=1

2
and let u
n
= e
1
(n+1)e
n+1
for n = 1, 2 . . . . Apply-
ing Gramn-Schmidt to { u
n
}

n=1
we construct an orthonormal set = {u
n
}

n=1
H.
I now claim that H is maximal. Indeed if x = (x
1
, x
2
, . . . , x
n
, . . . )

then
x u
n
for all n, i.e.
0 = (x, u
n
) = x
1
(n + 1)x
n+1
.
Therefore x
n+1
= (n + 1)
1
x
1
for all n. Since x Span{e
i
}

i=1
, x
N
= 0 for some
N suciently large and therefore x
1
= 0 which in turn implies that x
n
= 0 for all
n. So x = 0 and hence is maximal in H. On the other hand, is not maximal
in
2
. In fact the above argument shows that

in
2
is given by the span of v =
(1,
1
2
,
1
3
,
1
4
,
1
5
, . . . ). Let P be the orthogonal projection of
2
onto the Span() = v

.
Then

X
i=1
hx, u
n
iu
n
= Px = x
hx, vi
kvk
2
v,
so that

P
i=1
hx, u
n
iu
n
= x i x Span() = v

2
. For example if x =
(1, 0, 0, . . . ) H (or more generally for x = e
i
for any i), x / v

and hence

P
i=1
hx, u
n
iu
n
6= x.
ANALYSIS TOOLS WITH APPLICATIONS 239
12.6. Supplement 3: Conditional Expectation. In this section let (, F, P)
be a probability space, i.e. (, F, P) is a measure space and P() = 1. Let G F
be a sub sigma algebra of F and write f G
b
if f : C is bounded and f is
(G, B
C
) measurable. In this section we will write
Ef :=
Z

fdP.
Denition 12.39 (Conditional Expectation). Let E
G
: L
2
(, F, P) L
2
(, G, P)
denote orthogonal projection of L
2
(, F, P) onto the closed subspace L
2
(, G, P).
For f L
2
(, G, P), we say that E
G
f L
2
(, F, P) is the conditional expecta-
tion of f.
Theorem 12.40. Let (, F, P) and G F be as above and f, g L
2
(, F, P).
(1) If f 0, P a.e. then E
G
f 0, P a.e.
(2) If f g, P a.e. there E
G
f E
G
g, P a.e.
(3) |E
G
f| E
G
|f|, P a.e.
(4) kE
G
fk
L
1 kfk
L
1 for all f L
2
. So by the B.L.T. Theorem 4.1, E
G
extends
uniquely to a bounded linear map from L
1
(, F, P) to L
1
(, G, P) which we
will still denote by E
G
.
(5) If f L
1
(, F, P) then F = E
G
f L
1
(, G, P) i
E(Fh) = E(fh) for all h G
b
.
(6) If g G
b
and f L
1
(, F, P), then E
G
(gf) = g E
G
f, P a.e.
Proof. By the denition of orthogonal projection for h G
b
,
E(fh) = E(f E
G
h) = E(E
G
f h).
So if f, h 0 then 0 E(fh) E(E
G
f h) and since this holds for all h 0 in G
b
,
E
G
f 0, P a.e. This proves (1). Item (2) follows by applying item (1). to f g.
If f is real, f |f| and so by Item (2), E
G
f E
G
|f|, i.e. |E
G
f| E
G
|f|, P
a.e. For complex f, let h 0 be a bounded and G measurable function. Then
E [|E
G
f| h] = E
h
E
G
f sgn(E
G
f)h
i
= E
h
f sgn(E
G
f)h
i
E [|f| h] = E [E
G
|f| h] .
Since h is arbitrary, it follows that |E
G
f| E
G
|f| , P a.e. Integrating this
inequality implies
kE
G
fk
L
1 E |E
G
f| E[E
G
|f| 1] = E [|f|] = kfk
L
1.
Item (5). Suppose f L
1
(, F, P) and h G
b
. Let f
n
L
2
(, F, P) be a
sequence of functions such that f
n
f in L
1
(, F, P). Then
E(E
G
f h) = E( lim
n
E
G
f
n
h) = lim
n
E(E
G
f
n
h)
= lim
n
E(f
n
h) = E(f h). (12.20)
This equation uniquely determines E
G
, for if F L
1
(, G, P) also satises E(Fh) =
E(f h) for all h G
b
, then taking h = sgn(F E
G
f) in Eq. (12.20) gives
0 = E((F E
G
f) h) = E(|F E
G
f|).
240 BRUCE K. DRIVER

This shows F = E
G
f, P a.e. Item (6) is now an easy consequence of this charac-
terization, since if h G
b
,
E [(gE
G
f) h] = E [E
G
f hg] = E [f hg] = E [gf h] = E[E
G
(gf) h] .
Thus E
G
(gf) = g E
G
f, P a.e.
Proposition 12.41. If G
0
G
1
F. Then
(12.21) E
G
0
E
G
1
= E
G
1
E
G
0
= E
G
0
.
Proof. Equation (12.21) holds on L
2
(, F, P) by the basic properties of or-
thogonal projections. It then hold on L
1
(, F, P) by continuity and the density of
L
2
(, F, P) in L
1
(, F, P).
Example 12.42. Suppose that (X, M, ) and (Y, N, ) are two nite measure
spaces. Let = X Y, F = M N and P(dx, dy) = (x, y)(dx)(dy) where
L
1
(, F, ) is a positive function such that
R
XY
d ( ) = 1. Let

X
: X be the projection map,
X
(x, y) = x, and
G := (
X
) =
1
X
(M) = {AY : A M} .
Then f : R is G measurable i f = F
X
for some function F : X R
which is N measurable, see Lemma 6.62. For f L
1
(, F, P), we will now show
E
G
f = F
X
where
F(x) =
1
(x)
1
(0,)
( (x))
Z
Y
f(x, y)(x, y)(dy),
(x) :=
R
Y
(x, y)(dy). (By convention,
R
Y
f(x, y)(x, y)(dy) := 0 if
R
Y
|f(x, y)| (x, y)(dy) =
.)
By Tonellis theorem, the set
E := {x X : (x) = }

x X :
Z
Y
|f(x, y)| (x, y)(dy) =

is a null set. Since

E [|F
X
|] =
Z
X
d(x)
Z
Y
d(y) |F(x)| (x, y) =
Z
X
d(x) |F(x)| (x)
=
Z
X
d(x)

Z
Y
(dy)f(x, y)(x, y)

Z
X
d(x)
Z
Y
(dy) |f(x, y)| (x, y) < ,
F
X
L
1
(, G, P). Let h = H
X
be a bounded G measurable function, then
E [F
X
h] =
Z
X
d(x)
Z
Y
d(y)F(x)H(x)(x, y)
=
Z
X
d(x)F(x)H(x) (x)
=
Z
X
d(x)H(x)
Z
Y
(dy)f(x, y)(x, y)
= E [hf]
and hence E
G
f = F
X
as claimed.
ANALYSIS TOOLS WITH APPLICATIONS 241
This example shows that conditional expectation is a generalization of the notion
of performing integration over a partial subset of the variables in the integrand.
Whereas to compute the expectation, one should integrate over all of the variables.
See also Exercise 12.8 to gain more intuition about conditional expectations.
Theorem 12.43 (Jensens inequality). Let (, F, P) be a probability space and
: R R be a convex function. Assume f L
1
(, F, P; R) is a function such
that (for simplicity) (f) L
1
(, F, P; R), then (E
G
f) E
G
[(f)] , P a.e.
Proof. Let us rst assume that is C
1
and f is bounded. In this case
(12.22) (x) (x
0
)
0
(x
0
)(x x
0
) for all x
0
, x R.
Taking x
0
= E
G
f and x = f in this inequality implies
(f) (E
G
f)
0
(E
G
f)(f E
G
f)
and then applying E
G
to this inequality gives
E
G
[(f)] (E
G
f) = E
G
[(f) (E
G
f)]
0
(E
G
f)(E
G
f E
G
E
G
f) = 0
The same proof works for general , one need only use Proposition 9.7 to replace
Eq. (12.22) by
(x) (x
0
)
0

(x
0
)(x x
0
) for all x
0
, x R
where
0

(x
0
) is the left hand derivative of at x
0
.
If f is not bounded, apply what we have just proved to f
M
= f1
|f|M
, to nd
(12.23) E
G

(f
M
)

(E
G
f
M
).
Since E
G
: L
1
(, F, P; R) L
1
(, F, P; R) is a bounded operator and f
M
f
and (f
M
) (f) in L
1
(, F, P; R) as M , there exists {M
k
}

k=1
such that
M
k
and f
M
k
f and (f
M
k
) (f), P a.e. So passing to the limit in Eq.
(12.23) shows E
G
[(f)] (E
G
f), P a.e.
12.7. Exercises.
Exercise 12.7. Let (X, M, ) be a measure space and H := L
2
(X, M, ). Given
f L

() let M
f
: H H be the multiplication operator dened by M
f
g = fg.
Show M
2
f
= M
f
i there exists A M such that f = 1
A
a.e.
Exercise 12.8. Suppose (, F, P) is a probability space and A := {A
i
}

i=1
F
is a partition of . (Recall this means =
`

i=1
A
i
.) Let G be the algebra
generated by A. Show:
(1) B G i B =
i
A
i
for some N.
(2) g : R is G measurable i g =
P

i=1

i
1
Ai
for some
i
R.
(3) For f L
1
(, F, P), let E(f|A
i
) := E [1
A
i
f] /P(A
i
) if P(A
i
) 6= 0 and
E(f|A
i
) = 0 otherwise. Show
E
G
f =

X
i=1
E(f|A
i
)1
A
i
.
Exercise 12.9. Folland 5.60 on p. 177.
Exercise 12.10. Folland 5.61 on p. 178 about orthonormal basis on product
spaces.
Exercise 12.11. Folland 5.67 on p. 178 regarding the mean ergodic theorem.
242 BRUCE K. DRIVER

Exercise 12.12 (Haar Basis). In this problem, let L

2
denote L
2
([0, 1], m) with the
standard inner product,
(x) = 1
[0,1/2)
(x) 1
[1/2,1)
(x)
and for k, j N
0
:= N{0} with 0 j < 2
k
let

kj
(x) := 2
k/2
(2
k
x j).
The following pictures shows the graphs of
00
,
1,0
,
1,1
,
2,1
,
2,2
and
2,3
re-
spectively.
1 0.75 0.5 0.25 0
1
0.5
0
-0.5
-1
x
y
x
y
Plot of
0
, 0.
1 0.75 0.5 0.25 0
1
0.5
0
-0.5
-1
x
y
x
y
Plot of
1
0.
1 0.75 0.5 0.25 0
1
0.5
0
-0.5
-1
x
y
x
y
Plot of
1
1.
1 0.75 0.5 0.25 0
2
1
0
-1
-2
x
y
x
y
Plot of
2
0.
1 0.75 0.5 0.25 0
2
1
0
-1
-2
x
y
x
y
Plot of
2
1.
1 0.75 0.5 0.25 0
2
1
0
-1
-2
x
y
x
y
Plot of
2
2.
1 0.75 0.5 0.25 0
2
1
0
-1
-2
x
y
x
y
Plot of
2
3.
(1) Show := {1}

kj
: 0 k and 0 j < 2
k

is an orthonormal set, 1
denotes the constant function 1.
(2) For n N, let M
n
:= span

{1}

kj
: 0 k < n and 0 j < 2
k

.
Show
M
n
= span

{1
[j2
n
,(j+1)2
n
)
: and 0 j < 2
n

.
(3) Show

n=1
M
n
is a dense subspace of L
2
and therefore is an orthonormal
basis for L
2
. Hint: see Theorem 11.3.
ANALYSIS TOOLS WITH APPLICATIONS 243
(4) For f L
2
, let
H
n
f := hf, 1i1 +
n1
X
k=0
2
k
1
X
j=0
hf,
kj
i
kj
.
Show (compare with Exercise 12.8)
H
n
f =
2
n
1
X
j=0

2
n
Z
(j+1)2
n
j2
n
f(x)dx
!
1
[j2
n
,(j+1)2
n
)
and use this to show kf H
n
fk
u
0 as n for all f C([0, 1]).
Exercise 12.13. Let O(n) be the orthogonal groups consisting of n n real
orthogonal matrices O, i.e. O
tr
O = I. For O O(n) and f L
2
(R
n
) let
U
O
f(x) = f(O
1
x). Show
(1) U
O
f is well dened, namely if f = g a.e. then U
O
f = U
O
g a.e.
(2) U
O
: L
2
(R
n
) L
2
(R
n
) is unitary and satises U
O
1
U
O
2
= U
O
1
O
2
for all
O
1
, O
2
O(n). That is to say the map O O(n) U(L
2
(R
n
)) the
unitary operators on L
2
(R
n
) is a group homomorphism, i.e. a unitary
representation of O(n).
(3) For each f L
2
(R
n
), the map O O(n) U
O
f L
2
(R
n
) is continuous.
Take the topology on O(n) to be that inherited from the Euclidean topology
on the vector space of all nn matrices. Hint: see the proof of Proposition
11.13.
Exercise 12.14. Prove Theorem 12.34. Hint: Let H
0
:= span{x
n
: n N}
a separable Hilbert subspace of H. Let {
m
}

m=1
H
0
be an orthonormal basis
and use Cantors diagonalization argument to nd a subsequence y
k
:= x
n
k
such
that c
m
:= lim
k
hy
k
,
m
i exists for all m N. Finish the proof by appealing to
Proposition 12.33.
Exercise 12.15. Suppose that {x
n
}

n=1
H and x
n
w
x H as n . Show
x
n
x as n (i.e. lim
n
kx x
n
k = 0) i lim
n
kx
n
k = kxk .
Exercise 12.16. Show the vector space operations of X are continuous in the weak
topology. More explicitly show
(1) (x, y) X X x +y X is (
w

w
,
w
) continuous and
(2) (, x) F X x X is (
F

w
,
w
) continuous.
Exercise 12.17. Euclidean group representation and its innitesimal generators
including momentum and angular momentum operators.
Exercise 12.18. Spherical Harmonics.
Exercise 12.19. The gradient and the Laplacian in spherical coordinates.
Exercise 12.20. Legendre polynomials.
Exercise 12.21. In this problem you are asked to show there is no reasonable
notion of Lebesgue measure on an innite dimensional Hilbert space. To be more
precise, suppose H is an innite dimensional Hilbert space and m is a measure on
B
H
which is invariant under translations and satises, m(B
0
()) > 0 for all > 0.
Show m(V ) = for all open subsets of H.
244 BRUCE K. DRIVER

12.8. Fourier Series Exercises.

Notation 12.44. Let C
k
per
(R
d
) denote the 2 periodic functions in C
k
(R
d
),
C
k
per
(R
d
) :=

f C
k
(R
d
) : f(x + 2e
i
) = f(x) for all x R
d
and i = 1, 2, . . . , d

.
Also let h, i denote the inner product on the Hilbert space H := L
2
([, ]
d
) given
by
hf, gi :=

1
2

d
Z
[,]
d
f(x) g(x)dx.
Recall that

k
(x) := e
ikx
: k Z
d

is an orthonormal basis for H in particular

for f H,
(12.24) f =
X
kZ
d
hf,
k
i
k
where the convergence takes place in L
2
([, ]
d
). For f L
1
([, ]
d
), we will
write

f(k) for the Fourier coecient,
(12.25)

f(k) := hf,
k
i =

1
2

d
Z
[,]
d
f(x)e
ikx
dx.
Lemma 12.45. Let s > 0, then the following are equivalent,
(12.26)
X
kZ
d
1
(1 +|k|)
s
< ,
X
kZ
d
1
(1 +|k|
2
)
s/2
< and s > d.
Proof. Let Q := (0, 1]
d
and k Z
d
. For x = k +y (k +Q),
2 +|k| = 2 +|x y| 2 +|x| +|y| 3 +|x| and
2 +|k| = 2 +|x y| 2 +|x| |y| |x| + 1
and therefore for s > 0,
1
(3 +|x|)
s

1
(2 +|k|)
s

1
(1 +|x|)
s
.
Thus we have shown
1
(3 +|x|)
s

X
kZ
d
1
(2 +|k|)
s
1
Q+k
(x)
1
(1 +|x|)
s
for all x R
d
.
Integrating this equation then shows
Z
R
d
1
(3 +|x|)
s
dx
X
kZ
d
1
(2 +|k|)
s

Z
R
d
1
(1 +|x|)
s
dx
from which we conclude that
(12.27)
X
kZ
d
1
(2 +|k|)
s
< i s > d.
Because the functions 1+t, 2+t, and

1 +t
2
all behave like t as t , the sums
in Eq. (12.26) may be compared with the one in Eq. (12.27) to nish the proof.
Exercise 12.22 (Riemann Lebesgue Lemma for Fourier Series). Show for f
L
1
([, ]
d
) that

f c
0
(Z
d
), i.e.

f : Z
d
C and lim
k

f(k) = 0. Hint: If
f H, this follows form Bessels inequality. Now use a density argument.
ANALYSIS TOOLS WITH APPLICATIONS 245
Exercise 12.23. Suppose f L
1
([, ]
d
) is a function such that

f
1
(Z
d
) and
set
g(x) :=
X
kZ
d

f(k)e
ikx
(pointwise).
(1) Show g C
per
(R
d
).
(2) Show g(x) = f(x) for m a.e. x in [, ]
d
. Hint: Show g(k) =

f(k) and
then use approximation arguments to show
Z
[,]
d
f(x)h(x)dx =
Z
[,]
d
g(x)h(x)dx h C([, ]
d
).
(3) Conclude that f L
1
([, ]
d
) L

([, ]
d
) and in particular f
L
p
([, ]
d
) for all p [1, ].
Exercise 12.24. Suppose m N
0
, is a multi-index such that || 2m and
f C
2m
per
(R
d
)
29
.
(1) Using integration by parts, show
(ik)

f(k) = h

f,
k
i.
Note: This equality implies

f(k)

1
k

fk
H

1
k

fk
u
.
(2) Now let f =
P
d
i=1

2
f/x
2
i
, Working as in part 1) show
(12.28) h(1 )
m
f,
k
i = (1 +|k|
2
)
m

f(k).
Remark 12.46. Suppose that m is an even integer, is a multi-index and f
C
m+||
per
(R
d
), then
_
_
X
kZ
d
|k

f(k)

_
_
2
=
_
_
X
kZ
d
|h

f,
k
i| (1 +|k|
2
)
m/2
(1 +|k|
2
)
m/2
_
_
2
=
_
_
X
kZ
d

h(1 )
m/2

f,
k
i

(1 +|k|
2
)
m/2
_
_
2

X
kZ
d

h(1 )
m/2

f,
k
i

X
kZ
d
(1 +|k|
2
)
m
= C
m

(1 )
m/2

2
H
where C
m
:=
P
kZ
d
(1 +|k|
2
)
m
< i m > d/2. So the smoother f is the faster

f decays at innity. The next problem is the converse of this assertion and hence
smoothness of f corresponds to decay of

f at innity and visa-versa.
Exercise 12.25. Suppose s R and

c
k
C : k Z
d

are coecients such that

X
kZ
d
|c
k
|
2
(1 +|k|
2
)
s
< .
29
We view C
per
(R) as a subspace of H by identifying f C
per
(R) with f|
[,]
H.
246 BRUCE K. DRIVER

Show if s >
d
2
+m, the function f dened by
f(x) =
X
kZ
d
c
k
e
ikx
is in C
m
per
(R
d
). Hint: Work as in the above remark to show
X
kZ
d
|c
k
| |k

| < for all || m.

Exercise 12.26 (Poisson Summation Formula). Let F L
1
(R
d
),
E :=
_
_
_
x R
d
:
X
kZ
d
|F(x + 2k)| =
_
_
_
and set

F(k) := (2)
d/2
Z
R
d
F(x)e
ikx
dx.
Further assume

F
1
(Z
d
).
(1) Show m(E) = 0 and E + 2k = E for all k Z
d
. Hint: Compute
R
[,]
d
P
kZ
d |F(x + 2k)| dx.
(2) Let
f(x) :=
P
kZ
d
F(x + 2k) for x / E
0 if x E.
Show f L
1
([, ]
d
) and

f(k) = (2)
d/2

F(k).
(3) Using item 2) and the assumptions on F, show f L
1
([, ]
d
)
L

([, ]
d
) and
f(x) =
X
kZ
d

f(k)e
ikx
=
X
kZ
d
(2)
d/2

F(k)e
ikx
for m a.e. x,
i.e.
(12.29)
X
kZ
d
F(x + 2k) = (2)
d/2
X
kZ
d

F(k)e
ikx
for m a.e. x.
(4) Suppose we now assume that F C(R
d
) and F satises 1) |F(x)| C(1 +
|x|)
s
for some s > d and C < and 2)

F
1
(Z
d
), then show Eq. (12.29)
holds for all x R
d
and in particular
X
kZ
d
F(2k) = (2)
d/2
X
kZ
d

F(k).
For simplicity, in the remaining problems we will assume that d = 1.
Exercise 12.27 (Heat Equation 1.). Let (t, x) [0, ) R u(t, x) be a contin-
uous function such that u(t, ) C
per
(R) for all t 0, u := u
t
, u
x
, and u
xx
exists
and are continuous when t > 0. Further assume that u satises the heat equation
u =
1
2
u
xx
. Let u(t, k) := hu(t, ),
k
i for k Z. Show for t > 0 and k Z that
u(t, k) is dierentiable in t and
d
dt
u(t, k) = k
2
u(t, k)/2. Use this result to show
(12.30) u(t, x) =
X
kZ
e

t
2
k
2

f(k)e
ikx
ANALYSIS TOOLS WITH APPLICATIONS 247
where f(x) := u(0, x) and as above

f(k) = hf,
k
i =
1
2
Z

f(y)e
iky
dy.
Notice from Eq. (12.30) that (t, x) u(t, x) is C

for t > 0.
Exercise 12.28 (Heat Equation 2.). Let q
t
(x) :=
1
2
P
kZ
e

t
2
k
2
e
ikx
. Show that
Eq. (12.30) may be rewritten as
u(t, x) =
Z

q
t
(x y)f(y)dy
and
q
t
(x) =
X
kZ
p
t
(x +k2)
where p
t
(x) :=
1

2t
e

1
2t
x
2
. Also show u(t, x) may be written as
u(t, x) = p
t
f(x) :=
Z
R
d
p
t
(x y)f(y)dy.
Hint: To show q
t
(x) =
P
kZ
p
t
(x + k2), use the Poisson summation formula
along with the Gaussian integration formula
p
t
() =
1

2
Z
R
p
t
(x)e
ix
dx =
1

2
e

t
2

2
.
Exercise 12.29 (Wave Equation). Let u C
2
(RR) be such that u(t, ) C
per
(R)
for all t R. Further assume that u solves the wave equation, u
tt
= u
xx
. Let
f(x) := u(0, x) and g(x) = u(0, x). Show u(t, k) := hu(t, ),
k
i for k Z is twice
continuously dierentiable in t and
d
2
dt
2
u(t, k) = k
2
u(t, k). Use this result to show
(12.31) u(t, x) =
X
kZ

f(k) cos(kt) + g(k)

sinkt
k

e
ikx
with the sum converging absolutely. Also show that u(t, x) may be written as
(12.32) u(t, x) =
1
2
[f(x +t) +f(x t)] +
1
2
Z
t
t
g(x +)d.
Hint: To show Eq. (12.31) implies (12.32) use
cos kt =
e
ikt
+e
ikt
2
, and sinkt =
e
ikt
e
ikt
2i
and
e
ik(x+t)
e
ik(xt)
ik
=
Z
t
t
e
ik(x+)
d.
Exercise 12.30. (Worked Example.) Let D := {z C : |z| < 1} be the open
unit disk in C

= R
2
, where we write z = x + iy = re
i
in the usual way. Also let
=

2
x
2
+

2
y
2
and recall that may be computed in polar coordinates by the
formula,
u = r
1

r

r
1

r
u

+
1
r
2

u.
Suppose that u C(

D) C
2
(D) and u(z) = 0 for z D. Let g = u|
D
and
g(k) :=
1
2
Z

g(e
ik
)e
ik
d.
248 BRUCE K. DRIVER

(We are identifying S

1
= D :=

z

D : |z| = 1

with [, ]/ ( ) by the
map [, ] e
i
S
1
.) Let
(12.33) u(r, k) :=
1
2
Z

u(re
i
)e
ik
d
then:
(1) u(r, k) satises the ordinary dierential equation
r
1

r
(r
r
u(r, k)) =
1
r
2
k
2
u(r, k) for r (0, 1).
(2) Recall the general solution to
(12.34) r
r
(r
r
y(r)) = k
2
y(r)
may be found by trying solutions of the form y(r) = r

which then implies

2
= k
2
or = k. From this one sees that u(r, k) may be written as
u(r, k) = A
k
r
|k|
+ B
k
r
|k|
for some constants A
k
and B
k
when k 6= 0. If
k = 0, the solution to Eq. (12.34) is gotten by simple integration and the
result is u(r, 0) = A
0
+B
0
lnr. Since u(r, k) is bounded near the origin for
each k, it follows that B
k
= 0 for all k Z.
(3) So we have shown
A
k
r
|k|
= u(r, k) =
1
2
Z

u(re
i
)e
ik
d
and letting r 1 in this equation implies
A
k
=
1
2
Z

u(e
i
)e
ik
d = g(k).
Therefore,
(12.35) u(re
i
) =
X
kZ
g(k)r
|k|
e
ik
for r < 1 or equivalently,
u(z) =
X
kN0
g(k)z
k
+
X
kN
g(k) z
k
.
(4) Inserting the formula for g(k) into Eq. (12.35) gives
u(re
i
) =
1
2
Z

X
kZ
r
|k|
e
ik()
!
u(e
i
)d for all r < 1.
Now by simple geometric series considerations we nd, setting = ,
that
X
kZ
r
|k|
e
ik
=

X
k=0
r
k
e
ik
+

X
k=1
r
k
e
ik
=
1
1 re
i
+
re
i
1 re
i
=
1 re
i
+re
i

1 re
i

1 2r cos +r
2
=
1 r
2
1 2r cos +r
2
.
ANALYSIS TOOLS WITH APPLICATIONS 249
Putting this altogether we have shown
u(re
i
) =
1
2
Z

P
r
( )u(e
i
)d
where
P
r
() :=
1 r
2
1 2r cos +r
2
is the so called Poisson kernel.
Exercise 12.31. Show
P

k=1
k
2
=
2
/6, by taking f(x) = x on [, ] and
computing kfk
2
2
directly and then in terms of the Fourier Coecients

f of f.