Notes for Math 806, Functional Analysis

Rakesh
University of Delaware

September 4, 2013
Acknowledgement

These notes are based on material in Real and Complex Analysis by Walter Rudin,Functional
Analysis by Walter Rudin, Introduction to Topology and Modern Analysis by G F Simmons,
Introductory Functional Analysis with Applications by Erwin Kreyszig, Functional Analysis
by Michael Reed and Barry Simon, and Functional Analysis by Yuli Edelman, Vitali Milman
and Antonis Tsolomitis.

These notes are based on the lectures I gave at the University of Delaware for Math 806,
Functional Analysis, during the Fall semester of 2012. I gratefully acknowledge the help of
the following students who typed a first draft of these notes: Kenneth Boyle, Fun-Choi Chan,
Michael dePersio, Irene De Teresa Trueba, Ryan Evans, Matt Hassell, Charles Kish, Shixu
Meng, Rafael Plaza, Tianyu Qiu, Christine Rakowski, David Rowan, Mike Stapf, Tonatiuh
Sanchez Vizuet and Tao Yuan.
Contents

1 Normed Linear Spaces 6

1.1 Basic Ideas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.2 Examples of Banach Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.2.1 The Banach space C(X) . . . . . . . . . . . . . . . . . . . . . . . . . 10

1.2.2 lp spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

1.2.3 Lp (R) spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2 Bounded Linear Maps, Finite Dimensionality 18

2.1 Bounded Linear Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.2 Finite Dimensional Normed Spaces . . . . . . . . . . . . . . . . . . . . . . . 24

3 Four Important Theorems 28

3.1 The Open Mapping Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3.2 The Uniform Boundedness Theorem . . . . . . . . . . . . . . . . . . . . . . . 30

3
 3.3 The Hahn-Banach Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3.4 The Banach-Alaoglu Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 35

3.5 Adjoint Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

4 Hilbert Spaces 39

4.1 Elementary properties of inner product spaces . . . . . . . . . . . . . . . . . 39

4.2 Projections onto closed subspaces . . . . . . . . . . . . . . . . . . . . . . . . 43

4.3 Functionals on a Hilbert space . . . . . . . . . . . . . . . . . . . . . . . . . . 46

4.4 Orthonormal Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

4.5 The Hilbert space adjoint of an operator . . . . . . . . . . . . . . . . . . . . 52

5 Spectral Theory 56

5.1 General Spectral Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

5.2 Compact Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

5.3 Spectral Theory for Compact Operators . . . . . . . . . . . . . . . . . . . . 63

5.4 Sturm-Liouville Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

5.4.1 An example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

5.4.2 The general case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

5.5 Spectral Theorem for Self-Adjoint Operators . . . . . . . . . . . . . . . . . . 75

5.5.1 An example of a self-adjoint operator . . . . . . . . . . . . . . . . . . 76

4
5.5.2 The Spectral Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 79

5
Chapter 1

Normed Linear Spaces

1.1 Basic Ideas

Definition 1.1 (Normed Linear Space). A vector space X over R (or C) is called a normed
linear space if there is a function k·k : X → R satisfying the following properties for all
x, y ∈ X and all α ∈ R (or C):

1. kxk ≥ 0 with equality iff x = 0.

2. kx + yk ≤ kxk + kyk.
3. kαk = |α| kxk.
Example. The following are normed linear spaces.

Pn
• Rn with the norm kxk = i=1 |xi | where x = (x1 , · · · , xn ).
• The set C[a, b] = {f : [a, b] → R : f is continuous on [a, b]}, with norm kf k = max |f (x)|.
x∈[a,b]

Observe that there are other choices for norms on each of these spaces.
Proposition 1.1. If X is a normed linear space then
| kxk − kyk | ≤ kx − yk , ∀x, y ∈ X
and the map from X to R, given by x → kxk, is continuous.

6
Proof. From the triangle inequality we have kxk ≤ kx − yk + kyk so kxk − kyk ≤ kx − yk,
and by a similar argument we have kyk − kxk ≤ ky − xk = kx − yk. Hence

| kxk − kyk | ≤ kx − yk .

This inequality also shows that the map x 7→ kxk is continuous, because if x is close to
y then kxk is close to kyk.

Proposition 1.2. If X is a normed linear space then X forms a metric space with distance
function d(x, y) = kx − yk.

Proof.

1. d(x, y) = kx − yk ≥ 0 and is zero iff x = y.

2. d(x, y) = kx − yk = ky − xk = d(y, x).

3. d(x, z) = kx − zk = k(x − y) + (y − z)k ≤ kx − yk + ky − zk = d(x, y) + d(y, z).

Remark. Since a normed space X is a metricP space, we can talk of sequences {xn } in X and
their limits. Hence we can also talk of series ∞n=1 xn in X and their sum as the limit of the
sequence of partial sums.

Definition 1.2 (Banach Space). A complete normed linear space is a Banach space.

Remark. Note that {xn } is a Cauchy sequence in X if for every  > 0 there exists an N such
that

kxn − xm k <  for all n ≥ N.

Definition 1.3. Suppose X is a vector space and A is a subspace of X. Define

X/A = {x + A : x ∈ X}

, where x + A = {x + a : a ∈ A}. On X/A we define an addition and a scalar multiplication.

(i) (x1 + A) + (x2 + A) := (x + x2 ) + A for all x1 , x2 ∈ X,

(ii) α(x + A) := αx + A for all x ∈ X, α ∈ F.

7
 Since there can be different x, x0 ∈ X with x + A = x0 + A we must check that our
definition is independent of the representative x of x + A. Suppose x1 + A = x01 + A and
x2 + A = x02 + A; we show that x1 + x2 + A = x01 + x02 + A..

Since x01 ∈ x01 + A = x1 + A, there exists and a1 ∈ A such that x1 + a1 = x01 ; similarly
there is an a2 ∈ A such that x2 + a2 = x02 . Thus x01 + x02 = x1 + x2 + a1 + a2 and hence
x01 + x02 + A = x1 + x2 + a1 + a2 + A = x1 + x2 + A; note here we used a + A = A for any
a ∈ A - check it.

Similarly, definition(ii) can be shown to be independent of the representative. Further,

one may quickly check that X/A is a vector space with these operations; here the zero
element is 0 + A and the inverse of x + A is −x + A. We also define a candidate for a norm
on X/A, we define
kx + Ak = inf kx + ak = inf kx − ak
a∈A a∈A
.

Theorem 1.1. If X is a normed space and A a closed subspace of X then X/A is a normed
space (with the definition earlier). Further if X is complete then X/A is a Banach space.

Proof. One easily checks that X/A is a vector space; we now check the properties for a norm.
If kx + Ak = 0 then there exists a sequence {an } in A such that kx − an k → 0; thus, an → x
in X. But A is closed so x ∈ A and hence x + A = A, the 0 element in X/A.

Next kα(x + A)k = αkx + Ak - check it. Finally, for arbitrary x1 , x2 ∈ A, we have

kx1 + A + x2 + Ak = kx1 + x2 + Ak = inf kx1 + x2 + ak

a∈A
= inf kx1 + x2 + a1 + a2 k
a1 ,a2 ∈A

≤ inf kx1 + a1 k + kx2 + a2 k

a1 ,a2 ∈A

= inf inf kx1 + a1 k + kx2 + a2 k

a1 ∈A a2 ∈A

= kx1 + Ak + kx2 + Ak.

It remains to show that if X is complete then X/A is complete. For x ∈ X we denote

x + A by x̄. Let {x̄n } be a Cauchy sequence in X/A; it is enough to construct a convergent
subsequence of {x̄n }. Since {x̄n } is a Cauchy sequence, for each k = 1, 2, · · · , there exists an
Nk such that
1
kx̄n − x̄m k < k , for all m, n ≥ Nk .
2

8
We can arrange to have N1 < N2 < N3 < · · · , so
1
xNk+1 − xNk + A = kxNk+1 − xNk k < , for all k = 1, 2, · · ·
2k
So, by definition, there exists an ak ∈ A such that
1
kxNk+1 − xNk + ak k ≤ , k = 1, 2, · · · .
2k
Define a0 := 0 and

yk := xNk + a0 + · · · + ak−1 , k = 1, 2, · · · ;

then ȳk = x̄Nk and

1
kyk+1 − yk k = kxNk+1 − xNk + ak k ≤ , k = 1, 2, · · · .
2k
Now {yk } is a Cauchy sequence in X because for m < n
n−1
X n−1
X
kyn − ym k = k yk+1 − yk k ≤ kyk+1 − yk k
k=m k=m
n−1
X 1 1
≤ k
≤ m−1 ;
k=m
2 2

thus given  > 0 if we choose N such that 2N1−1 <  then kyn − ym k <  if n ≥ m ≥ N . Since
X is complete, yk → x for some x ∈ X; then

kx̄Nk − x̄k = ky¯k − x̄k = kyk − xk = inf kyk − x + ak ≤ kyk − xk → 0.

a∈A

1.2 Examples of Banach Spaces

When applying Functional Analysis tools to a problem, say in PDEs, Probability Theory or
Linear Algebra, the problem must be placed in an appropriate Banach space. Many of these
Banach spaces are based on the ones described in this section.

9
1.2.1 The Banach space C(X)

Definition 1.4. If X is a compact metric space then define

C(X) = {f : X → R| f continuous}

and for any f ∈ C(X) define

kf k := maxx∈X |f (x)|

We note that C(X) is a vector space over R under the usual addition and scalar mul-
tiplication and one can check that C(X) is a normed space with this norm. Since uniform
convergence preserves continuity one can also show that C(X) is complete under this norm.
A similar Banach space can be defined if we work with complex valued functions instead of
real valued functions.

We now characterize the compact subsets of C(X) which will require the introduction of
a new term.

Definition 1.5. Suppose X is a compact metric space and E a subset of C(X). We say E
is equicontinuous if for each  > 0 there exists δ > 0 such that

|f (x) − f (x0 )| <  ∀x, x0 ∈ X with d(x, x0 ) < δ

and for all f ∈ E. So the same δ works for all f ∈ E.

Every continuous function on a compact set is uniformly continuous - equicontiuity re-

quires that the same δ work for all functions in the set. We now characterize the compact
subsets of C(X).

Theorem 1.2 (Arzela-Ascoli Theorem). If X is a compact metric space and E a subset of

C(X) then E is compact iff E is equicontinuous and a bounded subset in C(X).

Proof. If E is compact then clearly E is bounded - we now show that E is equicontinuous.

Given  > 0, we have an open covering of E given by

E ⊆ ∪f ∈C(X) B/3 (f )

and hence by the compactness of E, we can find f1 , · · · , fN in C(X) such that

E ⊆ ∪N
i=1 B/3 (fi ). (1.1)

10
Now, each fi is continuous on the compact set X so uniformly continuous. Hence there exist
δi > 0 so that for all i = 1, ·, N we have

|fi (x) − fi (x0 )| < ∀x, x0 ∈ X with d(x, x0 ) < δi
3
Let δ = min(δ1 , δ2 , ..., δN ). Given f ∈ E, by (1.1), there is an i such that kf − fi k < /3,
that is

|f (x) − fi (x)| < , ∀x ∈ X.
3
Hence, for x, x0 ∈ X with d(x, x0 ) < δ we have

|f (x) − f (x0 )| ≤ |f (x) − fi (x)| + |fi (x) − fi (x0 )| + |fi (x0 ) − f (x0 )|
  
≤ + +
3 3 3
=

proving E is equicontinuous.

We now outline the proof of the converse. We assume E is bounded and equicontinuous
and we wish to show that E is pre-compact. Suppose fn is a sequence in E - we have to show
that is has a subsequence which converges in the C(X) norm. Note that the boundedness
of the sequence in C(X) implies that there is a constant M such that

|fn (x)| ≤ M ∀x ∈ X, n ≥ 1.

If X contains only a finite number of points xi then each fn (xi ) is a bounded sequence
of complex numbers and hence has a convergent subsequence, so taking subsequences of
subsequences we would be successful in constructing a subsequence of fn which converges
uniformly on X. The problem is to handle the case when X has an infinite number of points
- the compactness of X (along with the equicontinuity) allows us to reduce the problem to
the finite sized X case. Details are available in the books for the course.

Definition 1.6. For any non-negative integer k define

C k [a, b] := {f : [a, b] → R : f has k continuous derivatives on [a, b]}

and kf k := maxx∈[a,b] kj=0 |f (j) (x)|.

P

Proposition 1.3. C k [a, b] is a Banach space.

Proof. Completeness is a consequence of a Theorem in Rudin about Uniform Convergence

and Differentiability - do it. Other parts are easy.

11
1.2.2 lp spaces

Definition 1.7. Suppose x = (xn )n≥1 is a sequence of complex numbers. For any p ≥ 1,
define
∞
! p1
X
kxkp = |xn |p , kxk∞ = sup |xn |,
n≥1
n=1

and for 1 ≤ p ≤ ∞ define

`p := {x = (xn )n≥1 : xn ∈ C, kxkp is finite }.

To prove that lp is a normed space one needs the following important inequalities.
Theorem 1.3.
1 1
(a) (Holder’s Inequality) If p > 1, q > 1 with + = 1 then
p q
∞
X
|xn yn | ≤ kxkp kykq , ∀x ∈ `p , y ∈ `q ,
n=1

with equality holding iff there exist real α, β (independent of n) such that α|xn |p = β|yn |q
for all n ≥ 1, with at least one of α, β non-zero.

(b) (Minkowski’s Inequality) If p > 1 then

kx + ykp ≤ kxkp + kykp , ∀x, y ∈ `p

with equality holding iff there are non-negative α, β such that αx = βy and at least one
of α, β non-zero.

Proof. (a) If x or y is zero then the result holds trivially, so we assume that x 6= 0 and
x y
y 6= 0. We now scale to unit vectors by defining X := , Y := and note that
kxkp kykq
kXkp = 1, kY kq = 1; we have to show ∞
P
n=1 |Xn Yn | ≤ 1. From Lemma 1.1 below, we
have
∞ ∞
X X |Xn |p |Yn |q 1 1 1 1
|Xn Yn | ≤ + = kXkpp + kY kqq = + = 1.
n=1 n=1
p q p q p q

Further, from Lemma 1.1, equality will occur iff |Xn |p = |Yn |q for all n ≥ 1, that is iff
kykq xn = kxkp yn for all n ≥ 1.

12
 (b) If x or y is zero then the result holds trivially. Next, noting that
∞
!1/p ∞
!1/p
X X
kx + ykp = |xn + yn |p ≤ (|xn | + |yn |)p ,
n=1 n=1

with equality holding iff xn , yn have the same sign for all n ≥ 1, it is enough to prove
the result when xn ≥ 0 and yn ≥ 0, so we assume that below, along with x 6= 0, y 6= 0.
Since p > 1, we have
∞
X ∞
X
(xn + yn )p = (xn + yn )(xn + yn )p−1
n=1 n=1
∞
X X
= xn (xn + yn )p−1 + xn (xn + yn )p−1 . (1.2)
n=1

Using Holder’s inequality and noting that q(p − 1) = p, we have

∞ ∞
!1/p ∞ !1/q ∞
!1/q
X X X X
xn (xn + yn )p−1 ≤ xpn (xn + yn )q(p−1) = kxn kp (xn + yn )p
n=1 n=1 n=1 n=1
∞ ∞
!1/p ∞
!1/q ∞
!1/q
X X X X
yn (xn + yn )p−1 ≤ ynp (xn + yn )q(p−1) = kyn kp (xn + yn )p .
n=1 n=1 n=1 n=1

Using these relations in (1.2) we have

∞ ∞
!1/q
X   X
(xn + yn )p ≤ kxkp + kykp (xn + yn )p .
n=1 n=1
1
Then noting that 1 − q
= p1 , we have

kx + ykp ≤ kxkp + kykp .

Equality will occur iff equality occurs where we used Holder’s inequality, that is iff there
are αi , βi , i = 1, 2 such that α1 xpn = β1 (xn + yn )q(p−1) and α2 ynp = β2 (xn + yn )q(p−1) ,
for all n ≥ 1, with one of α1 , β1 non-zero and one of α2 , β2 non-zero. Now both αi , βi
are non-zero because x 6= 0 and y 6= 0, hence we conclude that β2 α1 xpn = β1 α2 ynp for
all n ≥ 1 and clearly we can take all the constants to be non-negative.

1 1
Lemma 1.1. If p > 1, q > 1 with p
+ q
= 1 then
ap b q
ab ≤ + , ∀a, b ≥ 0
p q
with equality holding iff ap = bq .

13
Proof. The result holds trivially if a = 0 or b = 0, so assume that a > 0, b > 0. Hence we
can find real x, y such that ap = ex , bq = ey and we have to show that
x y x y 1 1
e p + q = e p e q ≤ ex + ey (1.3)
p q
with equality holding iff x = y. But ex is a strictly convex function for real x because the
second derivative of ex is strictly positive. Hence

etx+(1−t)y ≤ tex + (1 − t)ey , ∀x, y ∈ R, ∀t ∈ [0, 1]

with equality holding iff x = y. So (1.3) holds by taking t = 1/p and equality holds iff x = y.
So in the original relation, equality holds iff ap = ex = ey = bq .

We have done most of the work to prove that lp is a Banach space.

Theorem 1.4. For 1 ≤ p ≤ ∞, `p is a Banach space.

Proof. For 1 < p < ∞ Minkowski’s inequality together with the easily observed relation
kαxkp = |α| kxkp are enough to show that `p is a normed space (and in particular a vector
space). For p = 1, and p = ∞, one easily sees that Minkowski’s inequality is still valid
and kαxkp = |α| kxkp . So `p is a normed space even for p = 1, ∞. It remains to prove the
completeness of `p . We will prove completeness when 1 ≤ p < ∞; the proof for the p = ∞
is almost identical.

Suppose {xs }s≥1 is a Cauchy sequence in `p ; the n-th term of xs will be denoted by xsn ,
that is xs = (xs1 , xs2 , xs3 , · · · ). Given  > 0, there exists N such that kxs − xt kp <  if s, t ≥ N ,
that is !1/p
∞
p
X
xsn − xtn < if s, t ≥ N. (1.4)
n=1

Therefore, for each n, we have

xsn − xtn < , if s, t ≥ N.

Thus {xsn }s≥1 is a Cauchy sequence in R for each n and hence has a limit we call xn ; define
x = {xn }n≥1 .

Now pick any s ≥ N ; then from (1.4), for any integer M , we have

M
!1/p
X
|xsn − xtn |p < , for all t ≥ N.
n=1

14
So letting t → ∞, for any s ≥ N , for any M , we get

M
!1/p
X
|xsn − xn | p
≤ .
n=1

Now let M → ∞ and we get

∞
!1/p
X
|xsn − xn |p ≤ , ∀s ≥ N,
n=1

that is kxs − xkp ≤  for all s ≥ N . This shows x ∈ lp because

kxkp ≤ kx − xs kp + kxs k ≤  + kxs kp

is finite and also shows that xs → x since we found an N for each  > 0. So `p is complete.
Remark.

• Our definition of `p was based on sequences of complex numbers and this generated a
Banach space over C and we may also call it `p (C) to emphasize this . Similarly, we
can define an `p (R) to be based on sequences of real numbers and the same arguments
show that `p (R) is a Banach space over R.

• Fix a positive integer m and 1 ≤ p ≤ ∞; then Cm (and similarly Rm for lp (R))

may be considered a subspace of `p (C) if we identify x ∈ Cm with the sequence
(x1 , x2 , · · · , xm , 0, 0, · · · ) in `p (C). With this identification, it is easily seen that Cm is
a closed subspace of `p (C) and hence Cm with the `p norm is also a Banach space.

1.2.3 Lp (R) spaces

The `p spaces are useful for discrete problems; we now define a continuous version of `p .

Definition 1.8. Suppose p ≥ 1. Define

Z
p
L (R) := {f : R → R f measurable and |f |p is finite.}
R

and for f ∈ Lp (R) define

Z 1/p
p
kf kp = |f | .
R

15
 Note: Since kf kp = 0 only implies f = 0 a.e., if f, g ∈ Lp (R) with f = g a.e. we will
consider f and g to be equal as elements of Lp (R).

Since functions are defined only almost everywhere, the continuous analog of the l∞ norm
must be defined with care.

Definition 1.9. A measurable function f : R → R is said to be essentially bounded if there

exists M ≥ 0 such that |f (x)| ≤ M almost everywhere, that is the set {x ∈ R : |f (x)| > M }
has measure zero. We define

L∞ (R) := {f : R → R f measurable and essentially bounded}

and for f ∈ L∞ (R) define kf k∞ to be the infimum of all M which are an essential upper
bound for f .

Again, for f, g ∈ L∞ (R) with f = g a.e., we will consider f and g to be same element of
Lp (R). One may show (homework) that |f | ≤ kf k∞ almost everywhere, so considering the
identifications of functions which are equal a.e., we may assume that |f | ≤ kf k∞ .

Towards proving Lp (R) is a normed space we need a version of Minkowski’s inequality

for Lp (R).

Theorem 1.5.
1 1
(a) (Holder’s Inequality) If p > 1, q > 1 with + = 1 then
p q
Z
|f g| ≤ kf kp kgkq , ∀f ∈ lp , g ∈ lq ,
R

with equality holding iff there exist real α, β (independent of n) such that α|f |p = β|g|q
a.e., with at least one of α, β non-zero.

(b) (Minkowski’s Inequality) If p > 1 then

kf + gkp ≤ kf kp + kgkp , ∀f, g ∈ lp

with equality holding iff there are non-negative α, β such that αf = βg and at least one
of α, β non-zero.

Proof.

16
 (a) The proof is very similar to the one used for lp . The result follows easily if f = 0 or
g = 0 so we assume WLOG that kf kp 6= 0 and kgkq 6= 0. Define F := kffk , G := kgk g
;
p q
then kF kp = 1 and kGkq = 1 and we have to show that
Z
|F (x)| |G(x)|dx ≤ 1.
R

From Lemma 1.1 we have

|F (x)|p |G(x)|q
|F (x)| |G(x)| ≤ +
p q

with equality occurring iff |F (x)|p = |G(x)|q . Hence

Z Z Z
1 1 1 1 1 1
|F (x)||G(x)| ≤ p
|F (x)| + |G(x)|q = kF kpp + kGkqq = + = 1
R p R q R p q p q

with equality occurring iff |F (x)|p = |G(x)|q a.e. This proves (a).

(b) The proof of (b) is very similar to the lp case - please do it.

Theorem 1.6. For 1 ≤ p ≤ ∞, Lp (R) is a Banach space.

Proof. For 1 < p < ∞ Minkowski’s inequality together with the easily observed relation
kαxkp = |α| kxkp are enough to show that Lp (R) is a normed space (and in particular a
vector space). For p = 1, and p = ∞, one easily sees that Minkowski’s inequality is still
valid and kαxkp = |α| kxkp . So Lp (R) is a normed space even for p = 1, ∞. It remains to
prove the completeness of Lp (R) which we postpone.

17
Chapter 2

Bounded Linear Maps, Finite

Dimensionality

2.1 Bounded Linear Maps

Definition 2.1. A map T : X → Y from a normed space X to a normed space Y is said to

be bounded if
{kT xk : x ∈ X, kxk ≤ 1}
is a bounded subset of R.

The next proposition provides various conditions equivalent to the continuity of a linear
map.
Proposition 2.1. Suppose X and Y are normed spaces and T : X → Y is linear. The
following are equivalent:
(a) T is continuous on X;
(b) T is continuous at 0;
(c) T is bounded;
(d) There exists a C > 0 (independent of x) such that kT xk ≤ C kxk for all x ∈ X.

So saying that T is a continuous linear map is equivalent to saying T is a bounded linear

map.

18
Proof. (a) =⇒ (b) is clear. To see that (b) implies (a) we note that if xn → x in X, then
xn − x → 0 and hence the continuity of T at 0 implies T xn − T x = T (xn − x) → T 0 = 0,
which implies T xn → T x. To prove the rest of the Theorem we show (b) =⇒ (c) =⇒ (d)
=⇒ (b).

Suppose T is continuous at 0; we claim that {kT xk : x ∈ X, kxk ≤ 1} is bounded. If

not, then we can construct a sequence xn in X such that

kT xn k ≥ n but kxn k ≤ 1.

We construct a sequence converging to 0 by defining yn = xn /n; then kyn k = kxn k/n ≤ 1/n
so yn → 0 while
1 n
kT yn k = kT xn k ≥ = 1.
n n
This contradicts (b).

Next we show (c) =⇒ (d). Assume that {kT xk : x ∈ X, kxk ≤ 1} is bounded above,
say by C. (d) clearly holds for x = 0; if x ∈ X, x 6= 0 then define x0 = kxk
x
and note that
0
kx k = 1. Hence
kT xk = kT x0 k kxk ≤ C kxk.
Finally suppose (d) holds and xn is a sequence in X which converges to 0. Then

kT xn k ≤ Ckxn k → 0

and hence T is continuous at 0.

Definition 2.2. Suppose X and Y are normed spaces and T : X → Y is a bounded linear
map. Define
kT k := sup kT xk
x∈X, kxk≤1

Proposition 2.2. If X and Y are normed spaces and T : X → Y is a bounded linear map
then
(a) kT k = sup{kT xk : x ∈ X, kxk = 1},

(b) kT xk ≤ kT k · kxk for all x ∈ X.

(c) kT k = inf C, where C varies over all reals such that kT xk ≤ C kxk for all x ∈ X,

Proof. If x ∈ X with kxk ≤ 1, x 6= 0, define x0 := x/kxk; then kx0 k = 1 so

kT xk = kxk kT x0 k ≤ kT x0 k

19
which proves (a).

For any x ∈ X, x 6= 0, x/kxk is a unit vector, hence kT (x/kxk)k ≤ kT k be definition so

kT xk ≤ kT k kxk, which proves (b).

If kT xk ≤ Ckxk for all x ∈ X, then kT xk ≤ C for all x ∈ X with kxk = 1. Hence by

definition kT k ≤ inf C. By (b), kT k is one of the C 0 s so inf C ≤ kT k, which proves (c).
Example. Suppose k(x, y) is a continuous real valued function on [0, 1] × [0, 1]. Define the
map K : C[0, 1] → C[0, 1] by
Z 1
(Kf )(x) = k(x, y) f (y) dy, x ∈ [0, 1]
0

for all f ∈ C[0, 1]. One may verify (HW) that Kf is a continuous function on [0, 1] and K
is a linear map. We show that K is a bounded linear map and compute kKk.

Recall that kf k = supx∈[0,1] |f (x)|. Now, for any x ∈ [0, 1], we have
Z 1 Z 1 Z 1
| k(x, y) f (y) dy| ≤ |k(x, y)| |f (y)| dy ≤ kf k |k(x, y)| dy,
0 0 0
so Z 1 Z 1
kKf k = sup | k(x, y) f (y) dy| ≤ kf k sup |k(x, y)| dy.
x∈[0,1] 0 x∈[0,1] 0

Hence Z 1
kKk ≤ sup |k(x, y)| dy;
x∈[0,1] 0

however one can show (HW) that

Z 1
kKk = sup |k(x, y)| dy.
x∈[0,1] 0

Example. Similarly, for p > 1, q > 1 with 1/p + 1/q = 1, and k ∈ Lq (R × R), we may define
a map K : Lp (R) → Lq (R) with
Z
(Kf )(x) = k(x, y) f (y) dy, x ∈ R,
R

and one wishes to compute kKk.

Example. Suppose A is an m × n matrix with real entries. Then A generates a linear map
A : Rn → Rm mapping x to Ax where x is to be thought of as a column vector. If we use
the k · kp norm on Rn and the k · kp0 norm on Rm then one can show that A is a bounded
linear map. What is kAk ?

20
Definition 2.3. For normed spaces X and Y over the same field F, we define B(X, Y ) to be
the set of all bounded linear maps from X to Y and B(X) will denote the set of all bounded
linear operators from X to X.

If S, T ∈ B(X, Y ), we define S + T : X → Y by (S + T )(x) = S(x) + T (x). It is easily

verified that S + T is linear; further

k(S + T )xk = kS(x) + T (x)k ≤ kSxk + kT xk ≤ kSk kxk + kT k kxk = (kSk + kT k) kxk

so S + T ∈ B(X, Y ) and
kS + T k ≤ kSk + kT k .

If T ∈ B(X, Y ) and α ∈ F , then we define αT : X → Y by (αT )(x) = α(T x) for all

x ∈ X. We can verify that αT is linear; further k(αT )(x)k = kα(T x)k = |α| kT xk so

kαT k = sup kα(T x)k = |α| sup k(T x)k = |α| kT k

x∈X, kxk=1 x∈X, kxk=1

implying αT ∈ B(X, Y ) and

kαT k = |α| kT k.
Remark. Suppose X,Y,Z are normed spaces over the same field F and S ∈ B(X, Y ) and T ∈
B(Y, Z). Then the composition defines a linear map T S : X → Z by (T S)(x) = T (S(x)),
for each x ∈ X. We note that

k(T S)(x)k = kT (S(x))k ≤ kT k kSxk ≤ kT k kSk kxk , ∀x ∈ X

which proves that T S ∈ B(X, Z) and

kT Sk ≤ kT k kSk .

Theorem 2.1. If X, Y are normed spaces then B(X, Y ) is a normed space. Furthermore,
if Y is complete then B(X, Y ) is complete.

Proof. Our work above has showed that B(X, Y ) is a normed space; it remains to show that
B(X, Y ) is complete if Y is complete.

Suppose Y is complete and {Tn }n≥1 is a Cauchy sequence in B(X, Y ). Hence for each
integer k ≥ 1, there is an Nk such that
1
kTn − Tm k ≤ , ∀m, n ≥ Nk .
2k

21
We can arrange to have N1 < N2 < N3 < . . .; then
1
TNk+1 − TNk ≤ , ∀k ≥ 1.
2k
We will prove the sequence {TNk }k≥1 is convergent in B(X, Y ), which will be enough to show
that Tn is convergent in B(X, Y ).

Define Sk = TNk for k ≥ 1; then we have

1
kSk+1 − Sk k ≤ , ∀k ≥ 1. (2.1)
2k
Now
k−1
X
Sk = S1 + (S2 − S1 ) + (S3 − S2 ) + · · · + (Sk − Sk−1 ) = S1 + (Sj+1 − Sj ).
j=1

The candidate for the limit of Sk is the sum ofPthe series S1 + ∞

P
j=1 (Sj+1 − Sj ) in B(X, Y ),
assuming that the sum exists. Now the series ∞ j=1 kSj+1 − S j k is convergent because, from
∞
(2.1), it is bounded above by the convergent geometric series j=1 1/2j . However, this is not
P

enough1 to guarantee the convergence of ∞

P
j=1 (Sj+1 − Sj ) in B(X, Y ) because B(X, Y ) is
not known to be complete, yet. However, Y is complete, so we have to modify our argument
to a point-wise argument to prove the convergence of Sk .

We first find a candidate for the limit of Sk ; we define the map S : X → Y by

∞
X
Sx = S1 x + (Sj+1 − Sj )x, ∀x ∈ X.
j=1

Note that S is well defined because the series is convergent in Y , because Y is complete and,
using (2.1),
∞
X ∞
X
kS1 xk + k(Sj+1 − Sj )xk ≤ kS1 k kxk + kSj+1 − Sj k kxk
j=1 j=1
∞
X kxk
≤ kS1 k kxk + = (1 + kS1 k)kxk
j=1
2j

is finite. This also shows that kSxk ≤ (1 + kS1 k)kxk for all x ∈ X, and since the linearity
of S is easily verified, S is a bounded linear map.
1
P∞
All we can claim is that Sk , the sequence of partial sums of S1 + j=1 (Sj+1 − Sj ), is a Cauchy sequence.

22
 We now show that Sk → S in B(X, Y ). For any x ∈ X we have
∞
X k−1
X ∞
X
(S − Sk )x = Sx − Sk x = S1 x + (Sj+1 − Sj )x − S1 x − (Sj+1 − Sj )x = (Sj+1 − Sj )x.
j=1 j=1 j=k

Hence, for all x ∈ X, we have

X ∞
X
∞
kSx − Sk xk ≤ j=k (Sj+1 − Sj )x ≤ k(Sj+1 − Sj )xk
j=k
∞ +∞
X X 1
≤ kSj+1 − Sj k kxk ≤ kxk
j=k j=k
2j
kxk
≤ .
2k−1
Hence
1
kS − Sk k ≤
2k−1
proving Sk converges to S in B(X, Y ).

We now define an isomorphism of normed spaces.

Definition 2.4. Suppose X,Y are normed spaces and T : X → Y is a linear map. We say
T is an isomorphism if T is a bijection, T is continuous, and T −1 is continuous. We say T
is an isometry if kT xk = kxk for all x ∈ X. If T is an isometry and an isomorphism then
we say T is an isometric isomorphism.
Remark.
• If T is an isometry, then T is injective since T x = 0 ⇒ kT xk = 0 ⇒ kxk = 0 ⇒ x = 0.

• If T is an isometry and X is a Banach space then the range of T is a closed subspace

of Y (HW).

• If T is an isometry then kT k = 1.

• An isometry need not be surjective. Consider the map from T : lp → lp given by

(x1 , x2 , ...) 7→ (0, x1 , x2 , ...). T is a linear isometry, but is not surjective, since the first
term of T x is always zero.

• If T is an isometry and surjective then T is bijective. The isometry also implies that T
and T −1 are bounded since kT k = 1 and kT −1 k = 1; so T is an isometric isomorphism.
Remark. If T : X → Y is an isomorphism, then X and Y have the same linear structure and
topology but not the same metric or norm.

23
 Suppose T : X → Y is a linear bijection between normed spaces X and Y. Automatically
T −1 : Y → X exists and is linear (check it). Now T is continuous iff there is a c1 > 0 such
that
kT xk ≤ c1 kxk ∀x ∈ X.
Again, T −1 is continuous iff there is c2 > 0 such that
T −1 y ≤ c2 kyk ∀y ∈ Y ;
but T is surjective so every y ∈ Y may be written as y = T x for some x ∈ X, so the
continuity of T −1 is equivalent to the existence of a c2 such that
kxk ≤ c2 kT xk ∀x ∈ X
which is equivalent to the existence of a c3 > 0 such that
c3 kxk ≤ kT xk ∀x ∈ X.
Hence we have proved the following proposition.
Proposition 2.3. If X and Y are normed linear spaces and T : X → Y is a linear bijection,
then T −1 is continuous iff there is a c > 0 such that
ckxk ≤ kT xk, ∀x ∈ X.

2.2 Finite Dimensional Normed Spaces

The following result shows that any finite dimensional normed linear space is isomorphic to
Rn .
Theorem 2.2. If X is a finite dimensional normed space over R (or C) of dimension n,
then X is isomorphic to Rn (or Cn ) with the k · k2 norm.

Proof. Let {e1 , e2 , ..., en } be a basis for X and define a map T : Rn → X given by
T α = α1 e1 + · · · + αn en ∀α = (α1 , · · · , αn ) ∈ Rn .
Then one easily sees that T is a linear bijection - it remains to show that T and T −1 are
bounded. Now, noting that |αi | ≤ kαk2 , we have
kT αk = kα1 e1 + · · · αn en k ≤ |α1 | ke1 k + · · · + |αn | ken k
≤ kαk2 ke1 k + · · · + kαk2 ken k
= (ke1 k + · · · + ken k) kαk2 .

24
Hence T is bounded, so T is continuous.

To show T −1 is bounded, we need to show there is a c > 0 such that c ≤ kT αk for all
α ∈ Rn with kαk = 1. Equivalently, we need to show that

inf kT αk > 0.
α∈Rn , kαk2 =1

We will use compactness of the unit ball in Rn to show this.

Consider the function F : Rn → R with F (α) = kT αk. F is the composition of the

map T : Rn → X and the function from X to R which maps x to kxk; these two maps are
continuous2 so F is continuous. Now the unit sphere on Rn is a compact subset of Rn hence
the infimum of the continuous map F (α) on the unit sphere is attained. Hence

inf kT αk = min kT αk = kT α∗ k
α∈Rn ,kαk2 =1 α∈Rn ,kαk2 =1

for some α∗ ∈ Rn , kα∗ k = 1. Now kT α∗ k > 0 because if kT α∗ k = 0 then T α∗ = 0 which,

using the injectivity of T , will imply α∗ = 0 which is not possible since kα∗ k = 1.

The above theorem has some useful consequences.

Proposition 2.4. Suppose X is a normed space.

(a) If X is finite dimensional then the closed unit ball {x ∈ X : kxk = 1} is compact and
X is complete.

(b) If Y is a finite dimensional subspace of X then Y is closed. In particular, if X is a

Banach space then every finite dimensional subspace of X is a Banach space.

Proof. The proposition follows from Theorem 2.2 - the proof of the proposition has been
assigned as a HW problem.

The converse of part (a) of Proposition 2.4 also holds. If X is a normed space whose
closed unit ball is compact, then X is finite dimensional. That is what we show next after
proving a lemma needed in the proof of our claim.

Lemma 2.1 (Riesz Lemma). Suppose Y is a closed subspace of a normed space X, Y 6= X.

Given  > 0, there is an x ∈ X with kxk = 1 and inf y∈Y kx − yk ≥ 1 − .
2
T was shown to be continuous above and the map x 7→ kxk is continuous because of the inequality
| kxk − kx0 k | ≤ kx − x0 k

25
Remark. If X had an inner product or if Y was finite dimensional then we could actually
find an x with inf y∈Y kx − yk = 1 while kxk = 1. In the inner product case we would just
take x to be orthogonal to Y and in the finite dimensional Y case we can use compactness.

Proof. Drawing a picture will help you understand the proof. Take any element x0 ∈ X
which is not in Y . Since Y is closed we have

d := inf kx0 − yk > 0.

y∈Y

Hence using the definition of the infimum, given any δ > 0, there is a y0 ∈ Y such that

kx0 − y0 k ≤ d + δ.

Of course kx0 − y0 k > 0 also because x0 is not in Y , now define

x0 − y 0
x1 = .
kx0 − y0 k

Then kx1 k = 1 and x1 is the candidate we sought because

x0 − y0 1
inf kx1 − yk = inf −y = inf k(x0 − y0 ) − y kx0 − y0 kk
y∈Y y∈Y kx0 − y0 k kx0 − y0 k y∈Y
1 1
= inf kx0 − y0 − y 0 k = inf kx0 − y 00 k
kx0 − y0 k y ∈Y
0 kx0 − y0 k y00 ∈Y
d
≥ = 1 − ,
d+δ
if we choose δ appropriately.

We are now ready to prove the converse of (a) of Proposition 2.4.

Theorem 2.3. If X is a normed space whose origin-centered closed unit ball is compact,
then X is finite dimensional.

Proof. We show that if X is infinite dimensional then the origin-centered closed ball is not
compact. Pick any e1 ∈ X with ke1 k = 1. Let E1 = Span(e1 ), then E1 is a closed subspace
of X because E1 is finite dimensional; further E1 6= X. So by Riesz’s lemma, taking  = 21 ,
there exists an e2 ∈ X, such that ke2 k = 1 and inf y∈E1 ke2 − yk ≥ 1/2. Now suppose we
have constructed {e1 , e2 , e3 , . . . , en } such that

• kei k = 1 for all i = 1, 2, · · · , n;

26
 1
• if Ek = Span(e1 , · · · , ek ) then inf y∈Ek kek+1 − yk ≥ 2
for all k = 1, · · · , n − 1.

Then En = Span(e1 , e2 , e3 , . . . en ) is a finite dimensional subspace of X, so En is closed;

further En is a strict subspace of X because X is infinite dimensional. So, by Riesz’s lemma,
there is an en+1 ∈ X with
1
ken+1 k = 1, and inf ken+1 − yk ≥ .
y∈En 2
In this fashion we have constructed a sequence {en }n≥1 in X with ken k = 1 and ken − em k ≥
1/2 for all n > m. So we have constructed a bounded sequence, in the closed unit ball of X,
which has no convergent subsequence. So the closed unit ball of X is not compact.

27
Chapter 3

Four Important Theorems

3.1 The Open Mapping Theorem

Theorem 3.1 (Open Mapping Theorem). If X and Y are Banach Spaces and T : X → Y
is a surjective bounded linear map then T is an open map, that is, T maps open sets to open
sets.

Proof. To be given later.

Remark. If T is a bijection then T −1 exists; further T −1 is continuous iff (T −1 )−1 (G) = T (G)
is open whenever G is open in X. Hence we have proved the following proposition.
Proposition 3.1. If X and Y are Banach spaces, and T : X → Y is a bounded linear
bijection, then T −1 is continuous, that is there exists a c > 0 such that

c kxk ≤ kT xk

Remark. So for Banach spaces X, Y and a bijective linear map T : X → Y , if there exists a
c1 > 0 such that
kT xk ≤ c1 kxk , ∀x ∈ X
then there exists a c2 > 0 such that

c2 kxk ≤ kT xk , ∀x ∈ X,

28
that is, if
sup kT xk is finite
x∈X, kxk=1

then
inf kT xk > 0.
x∈X, kxk=1

Definition 3.1. For normed spaces X and Y , X × Y = {(x, y) : x ∈ X, y ∈ Y } may be

given the structure of a normed space if we define

• (x, y) + (x0 , y 0 ) = (x + x0 , y + y 0 )
• α(x, y) = (αx, αy)
• k(x, y)k = kxkX + kykY

Further, X × Y is complete if X and Y are. The normed space X × Y is called the direct
sum of X and Y and denoted by X ⊕ Y .
Definition 3.2. Suppose X, Y are normed spaces and T : X → Y is linear. The graph of T
is the subset of X ⊕ Y defined by

Graph(T ) = {(x, T x) : x ∈ X} ⊆ X × Y.

Note that Graph(T ) is a subspace of X ⊕ Y and hence a normed space with the norm
acquired from X ⊕ Y .
Proposition 3.2 (Closed Graph Theorem). If X and Y are Banach spaces and T : X → Y
is linear, then T is bounded if and only if Graph(T ) is a closed subset of X ⊕ Y .
Remark. In general, to prove that T is continuous, one has to show that if xn → x then (a)
T xn is convergent and (b) T xn converges to T x. The Closed Graph Theorem says that to
prove the continuity of T it is enough to show that if (xn , T xn ) → (x, y), then y = T x, that
is, if xn → x and T xn → y then y = T x. So the Closed Graph Theorem makes it a little
easier to prove the boundedness of T .

Proof. If T : X → Y is bounded and (xn , T xn ) a sequence in Graph(T ) with (xn , T xn ) →

(x, y) ∈ X ⊕ Y , then xn → x and T xn → y. Now, since T is continuous, T xn → T x which
implies y = T x, and hence Graph(T ) is closed.

Conversely, suppose Graph(T ) is closed; then Graph(T ) is a closed subspace of the Banach
space X ⊕ Y and hence Graph(T ) is itself a Banach space. Now define S : Graph(T ) → X
with
S(x, T x) = x

29
and note that S is a linear bijection between Banach spaces. Further S is bounded because

kS(x, T x)k = kxk ≤ kxk + kT xk = k(x, T x)k

Hence by Proposition 3.1 S −1 is bounded, that is, there exists a c > 0 such that

c k(x, T x)k ≤ kS(x, T x)k , ∀x ∈ X,

equivalently, there is a c > 0 so that

c(kxk + kT xk) ≤ kx, k ∀x ∈ X.

Hence
kT xk ≤ (c−1 − 1) kxk , ∀x ∈ X
proving that T is continuous.
Remark. Note that the proof used the closed nature of Graph(T ) to claim that Graph(T )
was a Banach Space. If Graph(T ) was not complete we could not have used Proposition 3.1.

3.2 The Uniform Boundedness Theorem

Theorem 3.2 (Uniform Boundedness; Banach-Steinhaus). Suppose X is a Banach space,

Y a normed space, and Ti : X → Y , i ∈ I, a non-empty family of bounded linear operators.
If, for each x ∈ X with kxk = 1, there is a Cx such that

sup kT xk ≤ Cx ,
i∈I

then there exists a C independent of x such that

sup kTi xk ≤ C, ∀x ∈ X, kxk = 1;

i∈I

equivalently
sup kTi k ≤ C.
i∈I

Proof. To be given later.

Remark. Since the unit ball of a normed space is not compact for infinite dimensional spaces
the proof cannot be based on a compactness argument.

30
 One of the useful consequences of the Uniform Boundedness Theorem is a way to build
bounded linear maps by taking the point-wise limit of a sequence of bounded linear maps.
Proposition 3.3. Suppose X is a Banach space, Y is a normed space, and Tn : X → Y
a sequence of bounded linear maps such that lim Tn x exists for each x ∈ X. If we define
n→∞
T : X → Y by T x = lim Tn x then T is a bounded linear map.
n→∞

Note that this proposition does not guarantee that limn→∞ kTn − T k → 0.

Proof. One may quickly check that T is linear. Next, for each x ∈ X, the sequence {Tn x}n≥1
is convergent and hence bounded. Thus, for each x ∈ X with kxk = 1, there exists a Cx
such that
sup kTn (x)k ≤ Cx .
Hence, by the Banach-Steinhaus theorem, there exists a C independent of x such that
sup kTn k ≤ C
Hence, for any x ∈ X,
kT xk = lim kTn xk ≤ lim sup kTn k kxk ≤ C kxk ,
n→∞ n→∞

so that T is bounded and kT k ≤ C.

3.3 The Hahn-Banach Theorem

Definition 3.3. For any normed space X over C we define its dual as
X ∗ = {f : X → C : f linear and bounded} = B(X, C)
with a similar definition if X is a normed space over R. The elements of X ∗ are called
functionals sometimes. Note that by Theorem 2.1, X ∗ is a Banach space even if X is a
normed space.
Example. Pick an arbitrary a = (a1 , · · · , an ) ∈ Rn ; we define the function fa : Rn → R by
fa (x) = a1 x1 + · · · + an xn , ∀x = (x1 , · · · , xn ) ∈ Rn .
Then fa is bounded linear map under any norm on Rn . Further, if we define fi : Rn → R by
fi (x) = xi , ∀x ∈ Rn
then {f1 , . . . , fn } form a basis for (Rn )∗ and f = a1 f1 + · · · + an fn .

31
 Next we determine (lp )∗ for any p > 1; in fact we will show that if p > 1, q > 1 with
1/p + 1/q = 1 then (lp )∗ = lq via an isometric isomorphism. Given any a ∈ `q , a = (an )n≥1 ,
define fa : `p → C via
∞
X
fa (x) = x n an , ∀ x = (xn )n≥1 ∈ `p .
n=1

Note that infinite series in the definition of fa is convergent by Holder’s inequality because
∞
X
| xn an | ≤ kxkp kakq . (3.1)
n=1

Further, we can easily check that fa is a linear map, and from (3.1) we have
∞
X
|fa (x)| = | xn an | ≤ kxkp kakq ,
n=1

hence fa is a bounded map and kfa k ≤ kakq . We now show thatkfa k = kakq .

Given a ∈ `q with a 6= 0, we have to find a non-zero x ∈ `p so that |fa (x)| = kxkp kakq .
We can find such an x by using the equality condition from Holder’s inequality; we choose
to take a more direct approach. Define x = (xn )n≥1 where
(
ān |an |q/p−1 , if an 6= 0
xn = ;
0, an = 0

then, noting that |xn | = |an |q/p , we have

∞
!1/p ∞
!1/p
X X
kxkp = |xn |p = (|an |q/p )p = (kakq )q/p = kakqq−1
n=1 n=1

because 1/p + 1/q = 1 implies q/p = q − 1. Also, since an xn = |an |q/p+1 , we have
∞ ∞ ∞
X X q X
|fa (x)| = | an xn | = |an | p +1 = |an |q = (kakq )q = kxkp kakq
n=1 n=1 n=1

hence kfa kq = kakq .

So, we have constructed a map T : `q → (`p )∗ with

T (a) = fa , ∀ a ∈ `q .
T is linear - one easily checks that fa+b = fa + fb and fαa = αfa , and T is an isometry
because we have shown that kfa k = kakq . So T is injective and T would be an isometric
isomorphism between (`p )∗ and `q if we could show T is surjective. We leave the surjectivity
of T as a homework exercise. Hence we have proved part of the following theorem.

32
Proposition 3.4.

(a) If p, q are real numbers with p > 1, q > 1 and 1/p + 1/q = 1 then (`p )∗ is isometrically
isomorphic to `q .
(b) (`1 )∗ is isometrically isomorphic to `∞
(c) (`∞ )∗ is not isomorphic to `1

Proof.

(a) Done except for surjective of T (Homework exercise).

(b) Similar to (a).
(c) We will discuss this case later.

The Hahn-Banach Theorem below guarantees the existence of useful functionals on any
normed space X.
Theorem 3.3 (Hahn-Banach). If X is a normed space, M a subspace of X and f ∈ M ∗ ,
then f has an extension fe ∈ X ∗ with kfe k = kf k

Proof. Postpone.

The Hahn-Banach Theorem is useful in constructing functionals on normed spaces with

desired properties. To construct a functional on X which maps a specific finite set of linearly
independent elements of X in a certain way, let M be the span of these finite number of
points, define f ∈ M ∗ mapping the finite number of elements to the desired numbers and
then the Hahn-Banach theorem guarantees the existence of an extension fe ∈ X ∗ .
Corollary 3.1. If X is a normed space, x0 ∈ X, x0 6= 0 then there is an f ∈ X ∗ such that
f (x0 ) = kx0 k, kf k = 1.

Proof. Let M = span(x0 ) and define, f0 ∈ M ∗ by

f (αx0 ) = αkx0 k, ∀α ∈ C.

Then f0 is linear and |f0 (αx0 )| = |α|kx0 k = kαx0 k, therefore kf0 k = 1 and f0 ∈ M ∗ . Thus,
by the Hahn-Banach theorem, f0 has an extension f ∈ X ∗ with kf k = kf0 k = 1. Since f is
an extension of f0 we have that f (x0 ) = f0 (x0 ) = kx0 k.

33
Corollary 3.2. If X is a normed space and x1 , x2 ∈ X with x1 6= x2 , then there is f ∈ X ∗
with f (x1 ) 6= f (x2 ).

Proof. Since x1 6= x2 , we have x1 − x2 6= 0. So, by corollary 3.1, there exists an f ∈ X ∗ such

that kf k = 1, f (x1 − x2 ) = kx1 − x2 k =
6 0. Therefore, f (x1 ) 6= f (x2 ).
Remark (Significance of dual spaces). In some situations we do not know much about an
element x in a normed space X but instead we know f (x) for all f ∈ X ∗ . Corollary 3.2
assures us that knowing f (x) for all f ∈ X ∗ determines x uniquely.

In many situations we do not know an object but rather how it interacts with other
objects. If f ∈ C[a, b] then we can represent f in two ways:

• Give the value of f (x) for all x ∈ [a, b]

Rb
• Give a f (x)φ(x)dx for all φ ∈ C[a, b], that is give the functional Ff ∈ (C[a, b])∗ defined
by Z b
Ff (φ) = f (x)φ(x) dx, ∀ φ ∈ C[a, b].
a

The functional Ff completely determines f as seen from home work problems in Math 600.

This idea can be used to generalize the notion of a function. A distribution F on Rn is

a linear map F : Cc∞ (Rn ) → R which is continuous in a very specific (weak) topology on
Cc∞ (Rn ). Here Cc∞ (R) is the vector space of compactly supported smooth functions on Rn .
Every Lebesgue integrable function on Rn is a distribution via the functional
Z
φ 7→ f (x) φ(x) dx, ∀ φ ∈ Cc∞ (Rn ).
Rn

But we can also have distributions which are not functions, for example the Dirac delta
distribution, which is defined as the functional δ : Cc∞ (Rn ) → R with
δ(φ) = φ(0), ∀ φ ∈ Cc∞ (Rn ).
The Dirac delta function can be shown to be the limit of a sequence of smooth functions
whose limit is infinity at x = 0 and zero at all other points.

Another useful distribution is one which imitates being a function which is infinity on
some curve or surface and zero elsewhere. For example, we can define a distribution δ(x2 +
y 2 − 1) on R2 by the functional F : Cc∞ (R2 ) → R with
Z
1
F (φ) = φ(x, y) dSx,y ∀ φ ∈ Cc∞ (R2 ).
2 x2 +y2 =1

34
3.4 The Banach-Alaoglu Theorem

Definition 3.4. Suppose X is a normed space and {xn }n≥1 a sequence in X. We say xn
w
converges weakly to x ∈ X if f (xn ) → f (x) for all f ∈ X ∗ ; we write this as xn →x.

Definition 3.5. Suppose X is a normed space and {fn }n≥1 a sequence in X ∗ . We say fn is
w∗
weak ∗ convergent to f ∈ X ∗ if fn (x) → f (x) for all x ∈ X; we write this as →.
w
Remark. If xn → x in the norm on X then xn →x because if f ∈ X ∗ then f is continuous and
w
hence f (xn ) → f (x). However, xn →x does not imply xn → x in the norm on X as seen from
the following example. In `2 , if e√n denotes the element with the n-th entry 1 and all other
entries zero, then ken − em k2 = 2 if n 6= m so no subsequence of {en }n≥1 is convergent in
w
`2 but en →0 2
P∞in ` because (`2 )∗ = `2 and for any a ∈ `2 we have fa (en ) = an → 0 because
the series n=1 |an |2 is convergent. Similar remarks apply to weak∗ convergence.

If X is an infinite dimensional normed space then the closed balls of X or X ∗ are not
compact. The Banach-Alaoglu theorem asserts that closed balls of X ∗ are compact in the
weak∗ topology.

Theorem 3.4 (Banach-Alaoglu). If X is a Banach space and {fn }n≥1 a bounded sequence
w∗
in X ∗ , then there exists a subsequence {fnk }k≥1 and an f ∈ X ∗ such that fnk →f , that is

lim fnk (x) = f (x), ∀ x ∈ X.

k→∞

Definition 3.6. A Banach space X is said to be embedded in a Banach space Y if there

is a bijective, isometric linear map from X to a subspace of Y . Because of completeness and
isometry, the range of such a linear map is a closed subspace of X, so we may regard X as
a closed subspace of Y .

There is a natural embedding of any normed space X into its double dual X ∗∗ = (X ∗ )∗ .
One of the advantages of this is that sequences in X can be thought of as sequences in X ∗∗ ,
which is the dual of X ∗ - a Banach space, so the Banach-Alaoglu theorem is applicable.
This gives us almost weak compactness of closed balls of X, but not quite so, because the
subsequential weak limit is in X ∗∗ and is not guaranteed to be in X.

We now describe this natural embedding of a normed space X into X ∗∗ . For any x ∈ X,
we define the map Fx : X ∗ → C by

Fx (f ) = f (x), ∀f ∈ X ∗ . (3.2)

35
Fx is linear because for all f1 , f2 ∈ X ∗ and α, β ∈ C we have

Fx (αf1 + βf2 ) = (αf1 + βf2 )(x) = αf1 (x) + βf2 (x)

= αFx (f1 ) + βFx (f2 )

Further, Fx is a bounded because for any f ∈ X ∗ we have

|Fx (f )| = |f (x)| ≤ kf k kxk

implying kFx k ≤ kxk. Actually kFx k = kxk; for x = 0 this is clear and for x 6= 0, by 3.1,
there is an f ∈ X ∗ with kf k = 1 and f (x) = kxk. Hence, for this f , we have

|Fx (f )| = |f (x)| = kxk = kxk kf k

proving that kFx k ≥ kxk. So we have proved the following theorem about a natural embed-
ding of X into X ∗∗ .
Theorem 3.5. If X is a normed space then the map F : X → X ∗∗ defined by x 7→ Fx is an
injective, isometric, linear map from X to X ∗∗ .

Next we give some useful properties of weak convergence.

w
Proposition 3.5. If X is a normed space and {xn }n≥1 a sequence in X such that xn →x for
some x ∈ X, then

(a) the weak limit is unique;

(b) (xn )∞
n=0 is a bounded sequence in X;

(c) kxk ≤ lim inf n→∞ kxn k .

w w
Proof. (a) If xn →x and xn →x0 , then ∀f ∈ X ∗ we have f (xn ) → f (x) and f (xn ) → f (x0 ) as
complex numbers, so f (x) = f (x0 ) for all f ∈ X ∗ , that is f (x − x0 ) = 0 for all f ∈ X ∗ ,
which by Corollary 3.1 implies x − x0 = 0.
(b) Define the map Fn : X ∗ → C by

Fn (f ) = f (xn ), ∀ f ∈ X ∗.

Since Fn = Fxn , Fn is a bounded linear map with kFn k = kxn k because of Theorem
w
3.5. Moreover, given f ∈ X ∗ , the weak convergence xn →x implies that the sequence of
complex numbers f (xn ) → f (x), so {f (xn )}n≥1 is bounded. Hence, for each f ∈ X ∗ ,
there is a real number Cf such that

sup |Fn (f )| ≤ Cf .
n≥1

36
 Hence by the Banach-Steinhaus theorem (note X ∗ is complete) there is a real number
C such that supn≥1 kFn k ≤ C, that is
sup kxn k ≤ C.
n≥1

(c) The result is obvious if x = 0 so we assume x 6= 0. By Corollary 3.1, there is an f ∈ X ∗

with |f (x)| = kxk and kf k = 1. By hypothesis, f (xn ) → f (x) = kxk, so
kxk = |f (x)| = lim |f (xn )| = lim inf |f (xn )| ≤ lim inf kf k kxn k = lim inf kxn k
n→∞ n→∞ n→∞ n→∞

as desired.

3.5 Adjoint Operators

Suppose X and Y are normed spaces and T : X → Y is a bounded linear map. Define the
adjoint map T ∗ : Y ∗ → X ∗ by
(T ∗ (g))(x) = g(T (x)), ∀g ∈ Y ∗ , ∀x ∈ X.
Note that for every g ∈ Y ∗ , T ∗ (g) defines a bounded linear bounded functional in X ∗ because
for x1 , x2 ∈ X and α1 , α2 ∈ C we have
T ∗ (g)(α1 x1 +α2 x2 ) = g(T (α1 x1 +α2 x2 )) = α1 g(T (x1 ))+α2 g(T (x2 )) = α1 T ∗ (g)(x1 )+α2 T ∗ (g)(x2 )
and
|T ∗ (g)(x)| = |g(T x)| ≤ kgk kT xk ≤ kgk kT k kxk , ∀g ∈ Y ∗ and x ∈ X.
Hence T ∗ g is bounded and
kT ∗ gk ≤ kgk kT k , ∀g ∈ Y ∗ . (3.3)
One may easily verify that T ∗ is a linear map so T ∗ ∈ B(Y ∗ , X ∗ ) and (3.3) shows that T ∗ is
a bounded map with kT ∗ k ≤ kT k. In fact, kT ∗ k = kT k as shown next.

I fT = 0 then it is clear, so we assume that T 6= 0. Pick an x ∈ X, x 6= 0, such that

T x 6= 0; then by corollary 3.1, there is a g ∈ Y ∗ such that g(T x) = kT xk and kgk = 1. Then
|T x| = |g(T x)| = |T ∗ (g)(x)| ≤ kT ∗ k kgk kxk = kT ∗ k kxk
for all x ∈ X with T x 6= 0. Hence
kT k ≤ kT ∗ k
proving our claim. So we have proved the following theorem.

37
Theorem 3.6. If X and Y are normed spaces then the map T → T ∗ is a linear isometry
from B(X, Y ) to B(Y ∗ , X ∗ ), that is, for all T1 , T2 ∈ B(X, Y ) and complex α1 , α2 we have

(α1 T1 + α2 T2 )∗ = α1 T1∗ + α2 T2∗

and
kT ∗ k = kT k , ∀T ∈ B(X, Y ).
Further, if X, Y, Z are normed spaces with S ∈ B(X, Y ) and T ∈ B(Y, Z) then

(T S)∗ = S ∗ T ∗ .

Proof. Above, we have proved all parts of the theorem except the last part. The last part
holds because for any h ∈ Z ∗ and any x ∈ X we have

(T S)∗ (h)(x) = h(T Sx) = h(T (Sx)) = T ∗ (h)(Sx) = (S ∗ (T ∗ (h)))(x)

so
(T S)∗ (h) = S ∗ T ∗ h, ∀h ∈ Z ∗ .

38
Chapter 4

Hilbert Spaces

4.1 Elementary properties of inner product spaces

Definition 4.1 (Inner Product Space). A Vector Space X over a field C (or R) is an inner
product space if there is a function h, i : X × X → C such that ∀x, y, z ∈ X, ∀α ∈ C the
following hold:

(a) hx, xi ≥ 0 with hx, xi = 0 iff x = 0.

(b) hx + y, zi = hx, zi + hy, zi.

(c) hαx, yi = αhx, yi.

(d) hx, yi = hy, xi.

For any x ∈ X we define p

kxk = hx, xi.

We observe that k·k is a norm on X.

Proposition 4.1. If X is an inner product space then ∀x, y, z ∈ X and for all α ∈ C

(a) hx, 0i = h0, xi = 0.

(b) hαx + βy, zi = αhx, zi + βhy, zi.

39
(c) hx, αyi = αhx, yi.

(d) kxk = 0 ⇐⇒ x = 0.

(e) kαxk = |α| kxk.

Proof. These properties are direct consequences of the definition of the inner product func-
tion.

Definition 4.2. If X is an inner product space and x, y are elements of X. We say x, y are
orthogonal (and denote it by x ⊥ y) if hx, yi = 0. We say x is parallel to y (and denote it
by x k y) if x = αy or y = αx for some α in C.
Remark. The zero vector is parallel to all other vectors - use α = 0.

Proposition 4.2. Suppose X is an inner product space and x, y ∈ X.

(a) (Pythagoras’s Theorem) If x ⊥ y then kx + yk2 = kxk2 + kyk2 .

(b) (Parallelogram Identity) kx + yk2 + kx − yk2 = 2(kxk2 + kyk2 ).

Proof.

(a) We have

kx + yk2 = hx + y, x + yi = kxk2 + hx, yi + hy, xi + kyk2 = kxk2 + kyk2 .

(b) The result follows from adding the following relations-

kx + yk2 = hx + y, x + yi = kxk2 + kyk2 + hx, yi + hy, xi

kx − yk2 = hx − y, x − yi = kxk2 + kyk2 − hx, yi − hy, xi.

Theorem 4.1. If X is an inner product space then

(a) (Cauchy-Schwarz Inequality) For any x, y ∈ X,

|hx, yi| ≤ kxk kyk ,

with equality holding iff x and y are parallel;

40
(b) (Triangle Inequality) For any x, y ∈ X,

kx + yk ≤ kxk + kyk

with equality occurring iff x is a non-negative multiple of y or vice versa.

Proof. If x = 0 or y = 0 then both results follows trivially so we assume that x and y are
non-zero.

hy, xi hy, xi
(a) Define z = y − 2 x; so y = x + z and one may easily check that z ⊥ x. So, by
kxk kxk2
the Pythagorean Theorem, we have
2
hy, xi |hy, xi|2 |hy, xi|2
kyk2 = x + kzk2 = kxk2
+ kzk2
= + kzk2
kxk2 kxk 4
kxk 2

|hy, xi|2
≥ ,
kxk2

hy, xi
which proves the inequality. Further, equality occurs iff kzk = 0, that is iff y = x,
kxk2
that is iff y k x.

(b) Using (a), we have

kx + yk2 = hx + y, x + yi = kxk2 + hx, yi + hy, xi + kyk2

= kxk2 + 2<hx, yi + kyk2
≤ kxk2 + 2|hx, yi| + kyk2
≤ kxk2 + 2 kxk kyk + kyk2
= (kxk + kyk)2 .

Further, equality occurs iff <hx, yi = |hx, yi| and |hx, yi| = kxk kyk. From (a), the second
relation holds iff x = αy for some α ∈ C and substituting this in the first relation we
can see that it holds iff α ≥ 0.

Proposition 4.3 (inner products are continuous). If X is an inner-product space and x0 ∈ X

then the maps x 7→ hx, x0 i and x 7→ hx0 , xi are continuous maps from X to C.

Remark. The map x 7→ hx, x0 i is linear while the map x 7→ hx0 , xi is ’conjugate’ linear.

41
Proof. Since
|hx, x0 i − hy, x0 i| = |hx − y, x0 i| ≤ kx − yk kx0 k
the map x 7→ hx, x0 i is continuous. The other map is continuous by a similar argument.
Definition 4.3 (Hilbert Space). An inner product space complete in the norm generated by
the inner-product is called a Hilbert Space.
Remark. An inner product space can always be thought of as a normed space with the norm
generated by the inner product. Similarly, a Hilbert space is also a Banach space with the
norm generated by the inner product.

We now list a few examples of Hilbert spaces.

Example. Consider the vector space Cn with the inner product
n
X
hx, yi = xk y k , x = (x1 , · · · , xn ), y = (y1 , · · · , yn ) in Cn .
k=1

Since kxk2 = hx, xi =

Pn
k=1 |xk |2 = kxk22 , Cn is a Hilbert space.
Example. On the Banach Space `2 , define the inner product hx, yi := ∞
P
n=1 xn yn for x =
`2 . Note that hx, yi is finite because of Holder’s inequality and since
(xn ), y = (yn ) inP
∞
kxk2 = hx, xi = 2 2 2
n=1 |xn | = kxk2 , the inner product generates the k·k2 norm, ` is a
Hilbert space.
R
Example. For f, g ∈ L2 (R) define hf, gi := R f g; then this is an inner product on L2 (R)
which generates the L2 norm. So L2 (R) is a Hilbert space.
Example. We can create a weighted version of L( R). Suppose w : R → R is a measurable
function such that w(x) > 0 almost everywhere. Define
Z
2
Lw (R) := {f : R → R|f measurable and w(x)|f (x)|2 dx is finite }
R

and on L2w (R) define

Z
hf, gi = w(x)f (x)g(x) dx, ∀f, g ∈ L2w (R).
R

If we define Z 1/2
2
kf kw,2 = w(x) |f (x)|
R

then one may prove a version of Holder’s inequality asserting that hf, gi ≤ kf kw,2 kgkw,2 so
the inner product is finite. Further, since w(x) > 0 almost everywhere, we have hf, f i ≥ 0
and zero iff f = 0 a.e. It is fairly easy to verify that the inner product axioms are satisfied
and with considerable effort one can show that L2w (R) is a Hilbert space.

42
4.2 Projections onto closed subspaces

Definition 4.4 (Convex Sets). A non-empty subset C of a vector space X is said to be

convex if tx + (1 − t)y ∈ C for every x, y ∈ C and every t ∈ [0, 1].

Remark. So C is convex iff for every x, y in C, the line segment joining x to y is also in C.

The next theorem is the crucial step in the construction of the projection of a Hilbert
space onto one of its closed subspaces.

Theorem 4.2. If X is a Hilbert Space and C is a closed, convex subset of X, then for every
x ∈ X there exists a unique c0 ∈ C such that

kx − c0 k = inf kx − ck .
c∈C

Proof. We first prove the uniqueness. Suppose there are c0 and c1 in C such that

kx − c0 k = kx − c1 k = δ := inf kx − ck .
c∈C

If c0 = (c0 + c1 )/2 then c0 ∈ C because C is convex and hence δ ≤ kx − c0 k. We now apply

the Parallelogram Identity to x − c0 and x − c1 . Noting that

x − c0 + x − c1 = 2(x − (c0 + c1 )/2) = 2(x − c0 )

(x − c0 ) − (x − c1 ) = c1 − c0

we obtain
4kx − c0 k2 + kc1 − c0 k2 = 2(kx − c0 k2 + kx − c1 k2 ) = 4δ 2
hence
kc1 − c0 |2 = 4δ 2 − 4kx − c0 k2 ≤ 4δ 2 − 4δ 2 = 0,
implying c0 = c1 .

We now prove the existence of a c0 . By definition, there exists a sequence {cn }n≥1 in C
such that
lim ||x − cn || = δ = inf ||x − c||.
n→∞ c∈C

Since X is a Hilbert space and C is closed, if we can show that {cn }n≥1 is a Cauchy sequence,
there will exist a c0 ∈ C such that cn → c0 . Hence

kx − c0 k = lim kx − cn k = δ.
n→∞

43
To prove that cn is a Cauchy sequence, we start with the Parallelogram identity applied to
x − cn and x − cn . Noting that

(x − cn ) + (x − cm ) = 2(x − (cn + cm )/2), (x − cn ) − (x − cm ) = cm − cn

we obtain
2
2 2 cn + cm
2(kx − cn k + kx − cm k ) = 4 x − + kcm − cn k2
2
which implies
2
2 2 2 cn + cm
kcm − cn k = 2(kx − cn k + kx − cm k ) − 4 x − . (4.1)
2

Since C is convex we have (cn + cm )/2 ∈ C so

2
cn + cm
4 x− ≥ 4δ 2 ,
2

hence (4.1) implies

kcm − cn k2 ≤ 2(kx − cn k2 + kx − cm k2 ) − 4δ 2 .

As m and n approach infinity the RHS approaches 2(δ 2 + δ 2 ) − 4δ 2 which is 0, so cn is a

Cauchy sequence.

Definition 4.5. If X is an inner product space and E a non-empty subset of X then we

define
E ⊥ = {x ∈ X : hx, ei = 0 for all ∈ E}.

Proposition 4.4. If X is an inner product space and E a non-empty subset of X then

(a) E ⊥ is a closed subspace of X

(b) E ⊂ (E ⊥ )⊥

(c) E ∩ E ⊥ ⊂ {0}

Proof. It is clear that E ⊥ is a subspace of X. If xn is sequence in E ⊥ and xn → x, then for

any e ∈ E we have h(ix, e) = limn→∞ hxn , ei = 0, hence x ∈ E ⊥ proving that E ⊥ is closed.
(b) and (c) may be shown easily.

We now show that if X is a Hilbert space and M a closed subspace of X then we can
define the projection of X onto M . This will be a very useful result later.

44
Theorem 4.3. Let X be a Hilbert space, M a closed subspace of X. Every x ∈ X may be
written uniquely as the sum of an element of M and an element of M ⊥ .

Proof. Since M is closed and convex, given x ∈ X there exists a unique m0 ∈ M such that

kx − m0 k = inf kx − mk .
m∈M

Since x = m0 + (x − m0 ), if we can show that x − m0 ∈ M ⊥ we would have proved part of

the theorem. Given m ∈ M , define the function f : R → R with

f (t) := kx − m0 − tmk2 = kx − m0 k2 + t2 kmk2 − 2t<hx − m0 , mi.

By construction, the quadratic function f (t) has a minimum at t = 0, hence f 0 (0) = 0 which
gives us <hx − m0 , mi = 0 for all m ∈ M . Replacing m by im we also obtain

0 = <hx − m0 , imi = =hx − m0 , mi, ∀m ∈ M.

Hence
hx − m0 , mi = 0, ∀m ∈ M.

Uniqueness follows from the observation that if x = m1 + y1 = m2 + y2 , with mi ∈ M

and yi ∈ M ⊥ , then
y1 − y2 = m2 − m1 ∈ M ∩ M ⊥ = (0),
hence m1 = m2 , y1 = y2 .
Definition 4.6. Let X be Hilbert space and M a closed subspace. The projection of X onto
M is defined as the map P : X → X with

Px = m

where x = m + y is the unique representation of x as the sum of an element of M and an

element of M ⊥ .
Theorem 4.4. Let X be Hilbert space, M a closed subspace of X and P the projection of
X onto M .

(a) If x ∈ M then P x = x; if x ∈ M ⊥ then P x = 0.

(b) Range(P ) = M .

(c) kxk2 = kP xk2 + kx − P xk2 for all x ∈ X.

(d) For every x ∈ X, kx − P xk = inf m∈M kx − mk with the infimum attained only at m =
P x.

45
(e) P is linear and bounded.

(f ) P 2 = P .

Proof.

(a) If x ∈ M then x = x + 0 with x ∈ M and 0 ∈ M ⊥ so P x = x. If x ∈ M ⊥ then x = 0 + x

with 0 ∈ M and x ∈ M ⊥ so P x = 0.

(b) Cleary Range(P ) ⊂ M and P x = x for all x ∈ M so Range(P ) = M .

(c) x = P x + (x − P x) with P x ∈ M, x − P x ∈ M ⊥ so the result follows from Pythagoras’s

theorem.

(d) This is a consequence of the proof of Theorem 4.3.

(e) Let x, x0 ∈ X with x = m + y, x0 = m0 + y 0 where m, m0 ∈ M and y, y 0 ∈ M ⊥ . Then

x + x0 = (m + m0 ) + (y + y 0 ), m + m0 ∈ M, y + y 0 ∈ M ⊥

with m + m0 ∈ M and y + y 0 ∈ M ⊥ , hence P (x + x0 ) = m + m0 = P x + P x0 ; similarly

we may show that P (αx) = αP x.
Moreover, from (c) we have kP xk ≤ kxk hence ||P || ≤ 1. However, for any x ∈ M we
have P x = x, hence kP k = 1 (assuming M not empty).

(f) Since P x ∈ M for any x ∈ X, we obtain P 2 x = P (P x) = P x.

4.3 Functionals on a Hilbert space

Suppose X is a Hilbert space over1 C and y ∈ X. Define fy : X → C with

fy (x) = hx, yi, ∀x ∈ X.

Clearly fy is linear and also bounded since

|fy (x)| = |hx, yi| ≤ kxk kyk .

1
Similar results hold for Hilbert spaces over R

46
Hence kfy k ≤ kyk. Actually kfy k = kyk because |fy (y)| = kyk2 = kyk kyk. So every y in X
generates a functional fy ∈ X ∗ with kfy k = kyk. So we have a map from X to X ∗ mapping
y → fy . We may check that
fαy+βy0 = αfy + βfy0
and the map is injective because if hx, y1 i = hx, y2 i for all x ∈ X then hx, y1 − y2 i = 0 for all
x ∈ X and hence y1 = y2 as seen by taking x = y1 − y2 . The claim of the next theorem is
that this map is surjective, that is every element of X ∗ is generated in this fashion.

Theorem 4.5. If X is a Hilbert space and f ∈ X ∗ , then there is a unique y ∈ X such that
f (x) = hx, yi for all x ∈ X.

Proof. The uniqueness was shown above. If f = 0 then y = 0 will work so we assume that
f 6= 0. Let M be the null space of f , that is M = f −1 (0). Then M is a subspace of X and
M is closed because f is continuous. Since f 6= 0, we have M 6= X and hence M ⊥ 6= 0.
Since f (x) = 0 for all x ∈ M , clearly the candidate for y is some element of M ⊥ . One can
argue that M ⊥ is one dimensional though we will not prove it explicitly and not use this
claim. Since M ⊥ 6= (0), we can find a y0 ∈ M ⊥ so that f (y0 ) 6= 0; by scaling y0 if necessary
we may assume that f (y0 ) = 1. The candidate for y will be a scalar multiple of y0 , that is
y = αy0 , chosen so that f (y0 ) = hy0 , αy0 i, that is

α ky0 k2 = f (y0 ) = 1.

So take y = y0 / ky0 k2 ; note y ∈ M ⊥ . Then for any x ∈ X we have

f (x − f (x)y0 ) = f (x) − f (x)f (y0 ) = 0

implying x − f (x)y0 ∈ M . Hence

0 = hx − f (x)y0 , yi = hx, yi − f (x)hy0 , yi = hx, yi − f (x)hy0 , y0 / ky0 k2 i = hx, yi − f (x).

Hence f (x) = hx, yi for all x ∈ X.

4.4 Orthonormal Sets

Definition 4.7. If E is a non-empty subset of a vector space X then the span of E is the
set of all finite linear combinations of elements of E and is denoted by span(E). One may
check that span(E) is a subspace of X.

Definition 4.8. A subset {xi }i∈I of a vector space X is said to be linearly independent if
any finite subset of {xi }i∈I are linearly independent.

47
Definition 4.9. A subset {ui }i∈I of a Hilbert space X is said to be orthonormal if hui , uj i =
δij for all i, j ∈ I.

Proposition 4.5. Suppose X is an inner-product space and {u1 , u2 , ..., un } is an orthonormal

subset of X.
Pn
(a) If x = i=1 ci ui then ci = hx, ui i for all i = 1, 2, ..., n and
n
X n
X
kxk2 = |ci |2 = |hx, ui i|2 .
i=1 i=1

(b) {u1 , u2 , ..., un } is a linearly independent set of vectors.

Pn
Proof. If x = i=1 ci ui then

Xn n
X
hx, uj i = h ci ui , uj i = ci hui , uj i = cj .
i=1 i=1

Further, since {ui } are orthogonal, by the Pythagorean Theorem we have

n
X n
X
kxk2 = |ci |2 = |hx, ui i|.
i=1 i=1

Finally, if c1 u1 + c2 u2 , ... + cn un = 0 then, using the first part, cj = h0, uj i = 0 for all
j = 1, 2, ..., n, so {u1 , u2 , ..., un } is a lineary independent set of vectors.

Proposition 4.6. Suppose X is a Hilbert space and {u1 , · · · , un } an orthonormal subset of

X. Given x ∈ X we have
n n 2
2
X X
kxk − |hx, ui i|2 ≤ x − ci ui , ∀ci ∈ C
i=1 i=1

with equality occurring only when ci = hx, ui i for all i = 1, 2, · · · , n.

Proof. If M = Span(u1 , u2 , ..., un ) then M is a subspace of X and hence M is convex; also

M is closed because M is finite dimensional. So given x ∈ X, from Theorem 4.2, there exists
a unique m0 ∈ M such that

kx − m0 k ≤ kx − mk , ∀m ∈ M (4.2)

48
with equality occurring iff m = m0 . Further, from the proof of Theorem 4.3, we know that
x − m0 ⊥ M , so hx − m0 , ui i = 0 and hence hm0 , ui i = hx, ui i for i = 1, · · · , n. Since
m0 ∈ M = Span(u1 , · · · , un ) we conclude that
n
X
m0 = hx, ui iui .
i=1

Now x = m0 + (x − m0 ) with x − m0 ⊥ m0 so by Pythagoras’s theorem

kxk2 = km0 k2 + kx − m0 k2
which together with (4.2) implies that
n 2
2
X
kxk − hx, ui iui = kxk2 − km0 k2 = kx − m0 k2 ≤ kx − mk2 , i = 1, · · · , n
i=1

with equality occurring iff m = m0 .

Definition 4.10. If {αi }i∈I is a set of complex numbers then we define
( )
X X
|αi | = sup |αj | : J ⊆ I, J finite
i∈I j∈J
.
P
Remark. If i∈I |αi | isPfinite then we claim that only a countable number of αi are non-zero
as seen next. Let s = i∈I |αi | and define
1
In = {αi : |αi | >
}.
n
Clearly In has at most ns entries, so In is a finite set. Hence the number
S∞of non-zero αi is
countable because the set of indices corresponding to the non-zero αi is n=1 In .
Proposition 4.7. (Bessel’s Inequality) If X is a Hilbert space and {ui }i∈I is an orthonormal
subset of X then X
|hx, ui i|2 ≤ kxk2 , ∀x ∈ X.
i∈I

Remark. In particular, this means that hx, ui i =

6 0 for at most a countable number of ui .

Proof. If J is a finite subset of I and M is the span of {ui }i∈J then by Proposition 4.6 we
have X
kxk2 − |hx, ui i|2 = inf kx − mk2 ≥ 0.
m∈M
i∈J
Hence X
|hx, ui i|2 ≤ kxk2 , for every finite subset J of I,
i∈J
implying the proposition.

49
 Suppose X is a Hilbert space, {ui }i∈I is
Pan orthonormal subset of X and suppose x ∈ X;
then Bessel’s inequality guarantees that i∈I |hx, ui i|2 is finite. The next theorem asserts
the converse.

Theorem 4.6. (Riesz-Fischer Theorem) Suppose X is a Hilbert space and {u Pi }i∈I is2 an
orthonormal subset of X. If {αi }i∈I is a set of complex numbers such that i∈I |αi | is
finite, then there exists an x ∈ X with hx, ui i = αi for all i ∈ I.

Proof. Since i∈I |αi |2 is finite,Ponly a countable number of the αi are non-zero. We
P
P∞reindex
them as {αk }k≥1 and note that ∞ k=1 |α k |2
is given to be finite. We claim that x = k=1 αk uk
is the desired element of X. To show this we have to prove this series is convergent and
hxi , ui i = αi .

Define the sequence of partial sums xn = nk=1 αk uk ; then for n ≥ m, from Pythagoras’s
P
theorem
n 2 n
2
X X
kxn − xm k = αk uk = |αk |2 → 0,
k=m+1 k=m+1
P∞
as m, n approach infinity, because the series k=1 |αk |2 is convergent. Hence the sequence
of partial sums {xn }n≥1 is a Cauchy sequence in X and hence convergent to some element
x ∈ X. Further
n
X
hx, um i = lim hxn , um i = lim αk huk , um i = lim αm = αm .
n→∞ n→∞ n→∞
k=1

Also, if i is an index for which αi = 0 then i is not one of the indices defining x or xn , hence

hx, ui i = lim hxn , ui i = lim 0 = 0.

n→∞ n→∞

Definition 4.11. An orthonormal set {ui }i∈I , in a Hilbert space X, is said to be a maximal
orthonormal set in X if {ui }i∈I ∪ {x} is not orthonormal for any x ∈ X different from all
the ui , i ∈ I.

Theorem 4.7. Every Hilbert space has a maximal orthonormal set.

Proof. The proof is a fairly simple consequence of the Well-ordering Principle which is equiv-
alent to the Axiom of Choice.

Theorem 4.8. If X is a Hilbert space and {ui }i∈I is an orthonormal subset of X, then the
following are equivalent:

50
(a) {ui }i∈I is a maximal orthonormal set in X;

(b) If S is the span of {ui }i∈I then S is dense in X;

X
(c) x = hx, ui iui for all x ∈ X;
i∈I
X
(d) kxk2 = |hx, ui i|2 for all x ∈ X;
i∈I
X
(e) hx, yi = hx, ui ihy, ui i for all x, y ∈ X.
i∈I

Remark. If we combine the Riesz-Fischer theorem with (e) we conclude that, as a Hilbert
space, X is isomorphic to L2 (I) where
X
L2 (I) = {(αi )i∈I : αi ∈ C, |αi |2 finite}
i∈I

with the usual addition and scalar multiplication in L2 (I) and the inner product
X
hα, βi = αi βi , α, β ∈ L2 (I).
i∈I

Proof. We show that (a) implies (b), (b) implies (c), (c) implies (d), (d) implies (e) and (e)
implies (a).

(a) implies (b): We give a proof by contradiction. Let M denote the closure of the span
of all the ui and suppose M is a proper subset of X. Choose an x ∈ X which is not in
M ; then we may write x = P x + y where P is the projection of X onto M and of course
y = x−P x is in M ⊥ . Also note that y 6= 0 because x is not in M . Hence y/ kyk is a unit vec-
tor orthogonal to all the ui implying {ui }i∈I is not a maximal orthonormal set. Contradiction.

(b) implies (c): Since i∈I |hx, ui i|2 ≤ kxk2 , the sum is finite. So as shown in the proof of
P
Theorem 4.6, hx, ui i =
6 0 only for a countablePnumber of i ∈ I; further if these special i are
k
indexed by the positive integers then xk = n=1 hx, un iun is a Cauchy sequence in X and
hence converges to some x0 ∈ X. We show x0 = x.
If ui is one of un then

hx0 , un i = lim hxk , un i = lim hxk , un i = lim hx, un i = hx, un i.

k→∞ k→∞, k≥n k→∞

If ui is not one of the un (that is hx, ui i = 0) then

hx0 , ui i = lim hxk , ui i = lim hxk , ui i = lim 0 = 0 = hx, ui i.

k→∞ k→∞, k≥n k→∞

51
Hence hx−x0 , ui i = 0 for all i and the density of the span of the ui in X implies hx−x0 , yi = 0
for all y ∈ Y implying x = x0 . Note that this is true no matter which rearrangement of
P ∞
n=1 hx, un iun we use.

(c) implies (d): As above we use only those i for which hx, ui i =
6 0 and
Pk we assume that
the relevant i have been reindexed by the natural numbers. If xk = n=1 hx, un iun then
xk → x in X. Then using the Pythagorean theorem
k
X ∞
X X
2 2 2
kxk = lim kxk k = lim |hx, un i| = |hx, un i|2 = |hx, ui i|2 .
k→∞ k→∞
n=1 n=1 i∈I

(d) implies (e): This is essentially an argument showing that the inner product may be
recovered from the norm. We have
kx + yk2 = hx, xi + hy, yi + 2<hx, yi
which implies (using the fact that only a countable number of hx, ui i are non-zero)
2<hx, yi = kx + yk2 − kxk2 − kyk2
X∞ ∞
X ∞
X
= |hx + y, un i|2 − |hx, un i|2 − |hy, un i|2
n=1 n=1 n=1
∞
X
|hx + y, un i|2 − |hx, un i|2 − |hy, un i|2


=
n=1
X∞
= 2<(hx, un ihy, un i)
n=1
∞
X
= 2< hx, un ihy, un i, ∀x, y ∈ X.
n=1

If X is a real inner product space then we are done. If X is a complex inner product space
then y by iy in the above relation to also obtain the relationship between the imaginary parts.
P
(e) implies (a): If hx, ui i = 0 for all i ∈ I then by (e), hx, xi = i∈I hx, ui ihx, ui i = 0
implying x = 0. Hence {ui }i∈I is a maximal orthonormal set.

4.5 The Hilbert space adjoint of an operator

Suppose X is a Hilbert space and T : X → X is a bounded linear operator. For each y ∈ X

we may define the map f : X → C with
f (x) = hT x, yi, ∀x ∈ X.

52
Since T is linear and the inner product is linear in the first argument, f is linear. Further f
is bounded because
|f (x)| = |hT x, yi| ≤ kT xkkyk ≤ kT kkxkkyk
which implies kf k ≤ kT kkyk. So f ∈ X ∗ and hence by Theorem 4.5 there is a unique y 0 ∈ X
such that f (x) = hx, y 0 i for all x ∈ X. So to each y ∈ X we have associated a unique y 0 ∈ X
such that
hT x, yi = f (x) = hx, y 0 i ∀x ∈ X.
So we have a well defined map T ∗ : X → X with T ∗ y = y 0 , that is

hT x, yi = hx, T ∗ yi ∀x, y ∈ X.

T ∗ is called the Hilbert space adjoint of T .

Theorem 4.9. If X is a Hilbert space and T : X → X a bounded linear operator, then there
is a unique linear map T ∗ : X → X such that

hT x, yi = hx, T ∗ yi ∀x, y ∈ X.

Further T ∗ is linear, bounded and kT ∗ k = kT k.

Proof. The existence of such a map was shown above; we show the uniqueness. If we have
maps S1 , S2 : X → X such that

hT x, yi = hx, S1 yi = hx, S2 yi, ∀x, y ∈ X

then hx, S1 y − S2 yi = 0, ∀x ∈ X and hence S1 y − S2 y = 0 if we take x = S1 y − S2 y.

We now show that T ∗ is linear. For all x, y1 , y2 ∈ X and α, β ∈ C we have

hx, T ∗ (αy1 + βy2 )i = hT x, αy1 + βy2 i = ᾱhT x, y1 i + β̄hT x, y2 i

= ᾱhx, T ∗ y1 i + β̄hx, T ∗ y2 i = hx, αT ∗ y1 i + hx, βT ∗ y2 i

Now we compute kT ∗ k. For any x ∈ X we have

kT ∗ xk2 = hT ∗ x, T ∗ xi = hT T ∗ x, xi ≤ kT kkT ∗ xkkxk

therefore kT ∗ xk ≤ kT kkxk for all x ∈ X, implying kT ∗ k ≤ kT k. Similarly

kT xk2 = hT x, T xi = hx, T ∗ T xi ≤ kxkkT ∗ kkT xk

so kT xk ≤ kT ∗ kkxk implying kT k ≤ kT ∗ k.

53
 The operator T ∗ satisfies the following useful relations.

Theorem 4.10. Suppose X is a Hilbert space and S, T ∈ B(X) and α is a scalar; then

(a) (S + T )∗ = S ∗ + T ∗

(b) (αT )∗ = αT ∗

(c) (ST )∗ = T ∗ S ∗

(d) T ∗∗ = T

(e) kT ∗ k = kT k

(f ) kT ∗ T k = kT k2 = kT T ∗ k

Proof. (a) Check it.

(b) It holds because, for all x, y ∈ X, we have

hx, (αT )∗ yi = h(αT )x, yi = hα(T x), yi = αhT x, yi = αhx, T ∗ yi = hx, (ᾱT ∗ )yi.

(c) It holds because, for all x, y ∈ X, we have

hx, (ST ) ∗ yi = hST x, yi = hS(T x), yi = hT x, S ∗ yi = hx, T ∗ S ∗ y > i.

(d) Check it.

(e) Done earlier.

(f) We have
kT ∗ T k ≤ kT ∗k kT k = kT k kT k = kT k2 .
Next, for all x ∈ X, we have

kT xk2 = hT x, T xi = hT ∗ T x, xi ≤ kT ∗ T xk kxk ≤ kT ∗ T k kxk2

which implies kT k ≤ kT ∗ T k so kT k2 ≤ kT ∗ T k. To obtain the other relation replace

p

T by T ∗ in the first relaton.

Finally we state a useful consequence of the Banach-Alaoglu Theorem.

54
Theorem 4.11. If {xn }n≥1 is a bounded sequence in a Hilbert Space X, then there is a
subsequence {xnk }k≥1 and an x ∈ X so that hxnk , yi → hx, yi for each y ∈ Y .

Proof. We define the functionals fn ∈ X ∗ with fn (y) = hy, xn i. As shown elsewhere we

have kfn k = kxn k, so {fn }n≥1 is a bounded sequence in X ∗ . Hence, by the Banach-Alaoglu
theorem, there is an f ∈ X ∗ and a subsequence {fnk }k≥1 , such that fnk (y) → f (y) for all
y]inX. However, by the Riesz Representation Theorem, there is a unique x ∈ X so that
f (y) = hy, xi for all y in X. Hence, there is an x ∈ X and a subsequence {xnk }k≥1 such that
hxnk , yi → hx, yi for each y ∈ Y .

55
Chapter 5

Spectral Theory

5.1 General Spectral Theory

In Math 672 we studied the spectral theory of normal operators on Cn and of the self-
adjoint operators on Cn and Rn . Here we will study the spectral theory of operators on
infinite dimensional spaces.
Definition 5.1. Suppose X is a Banach space over F and T ∈ B(X). We say λ ∈ F is a
regular point for T if T − λI is invertible and we say λ is a singular point for T if T − λI
is not invertible. The set of regular points of T is denoted by ρ(T ) and the set of singular
points of T is denoted by σ(T ) - so ρ(T ) and σ(T ) are complements of each other. The set
σ(T ) is called the spectrum of T .

I use the mnemonic S for singular so σ(T ) contains the singular points and R for regular
so ρ(T ) contains the regular points. The set σ(T ) is usually divided further. From the Open
Mapping Theorem we know that T − λI will have a bounded inverse iff T − λI is bijective.
So λ ∈ σ(T ) iff three mutually exclusive scenarios occur as described below.
Definition 5.2. Suppose X is a Banach space and T ∈ B(X). The spectrum of T may be
subdivided into three disjoint sets which are the point spectrum

σp (T ) := {λ ∈ F| T − λI is not injective },

the continuous spectrum

σc (T ) := {λ ∈ F| T − λI is injective and Range(T − λI) is dense in X but not equal to X},

56
and the residual spectrum
σr (T ) := {λ ∈ F| T − λI is injective and Range(T − λI) is not dense in X}.
If λ is in σp (T ) then the null space of T − λI is non-zero and there is an x ∈ X, x 6= 0 so
that T x = λx. The λ ∈ σp (T ) are called the eigenvalues of T and the corresponding non-zero
x are called eigenvectors of T corresponding to the eigenvalue λ.
Remark. If X is a finite dimensional normed space and T ∈ B(X) then for any λ ∈ F
dim X = dim ker (T − λI) + dim range (T − λI).
So T − λI is bijective iff ker T − λI = (0). Hence every element in the spectrum of T is an
eigenvalue.

We state a useful lemma whose proof was a homework exercise.

Lemma 5.1. Suppose P X is a normed space and T ∈ B(X). If kT k < 1 then I − T is
invertible, (I − T )−1 = ∞ n −1
n=0 T and k(I − T ) k ≤ 1/(1 − kT k).

We now study an example which will help us become more comfortable with the termi-
nology introduced in this section.
Example. We compute σ(R) and σ(L) where R and L are the right and left shift operators
on `2 as defined below.
R : `2 → `2 L : `2 → `2
x = (xn )n≥1 7→ (0, x1 , x2 , x3 , . . . ) x = (xn )n≥1 7→ (x2 , x3 , x4 , . . . )

Since R is injective and the range of R is the closed subspace of `2 consisting of entries
of `2 whose first component is 0 we have 0 ∈ σr (R). Since kRxk = kxk for all x ∈ `2 we
have kRk = 1 and hence
σ(R) ⊆ {z ∈ C : |z| ≤ 1}.
For L we note that kLxk ≤ kxk and Lx = x if the first component of x is zero, so kLk = 1.
Hence
σ(L) ⊆ {z ∈ C : |z| ≤ 1}.
We now determine which λ ∈ C could be in the point spectrum of L. If Lx = λx then
(x2 , x3 , x4 , . . . ) = (λx1 , λx2 , λx3 , . . . )
implying xn+1 = λxn for each n ≥ 1. If we take x1 = 0 then x = 0 so we take x1 = 1.
HenceP x2 = λ, x3 = λ2 , and in fact xn = λn−1 for all n ≥ 1. Now x = (λn−1 )n≥1 is in
2 2(n−1)
` iff n=1 |λ| is finite which holds iff |λ| < 1. So λ ∈ σp (L) iff |λ| < 1 and the
corresponding eigenvector is (λn−1 )n≥1 . The determination of the complete σ(L) and σ(R)
is left as an exercise.

57
Proposition 5.1. If X is a Banach space and T ∈ B(X) then

(i) σ(T ) is contained in {z ∈ F : |z| ≤ |T |},

(ii) σ(T ) is compact.

Proof.
kT k
(i) Now T − λI = −λ I − Tλ , so if |λ| > kT k then k Tλ k = T


|λ|
< 1, so I − λ
is invertible,
and hence T − λI is invertible.

(ii) We show that the complement of σ(T ) is open. If λ0 is in the complement of σ(T ) then
T − λ0 I is invertible and let R0 be its (bounded) inverse. Then for any λ we have

T − λI = T − λ0 I − (λ − λ0 )I = (T − λ0 I)(I − (λ − λ0 )(T − λ0 I)−1 )

= (T − λ0 I)(I − (λ − λ0 )R0 ) (5.1)

Now if |λ − λ0 | < 1/ kR0 k then

k(λ − λ0 )R0 k = |λ − λ0 | kR0 k < 1

so (I − (λ − λ0 )R0 ) is invertible and hence, from (5.1), T − λI is invertible. So T − λI

is invertible if λ is in a certain disk around λ0 .

Next we state a lemma which is useful is relating the spectra of T and T ∗ . Note that the
next two results are stated only for Hilbert spaces.

Lemma 5.2. If X is a Hilbert space and T ∈ B(X) then

(i) Ker T = (Range T ∗ )⊥ ,

(ii) Ker T ∗ = (Range T )⊥ .

Remark. Note that the above lemma does not guarantee, say, that Range T = (Ker T ∗ )⊥ -
that would hold only if the range was closed which would be guaranteed, for example, if T
was S − λI with λ 6= 0 and S was compact (shown later).

Proof. Left to the reader.

58
 The following proposition relates the different parts of the spectra of T and T ∗ and in
particular the point spectrum of an operator to the continuous and residual spectra of the
adjoint - note that this stated for Hilbert spaces only. These relationships are useful in
determining the various parts of the spectrum of T and T ∗ simultaneously. One determines
the point spectra of T and T ∗ because it is usually easier to determine the point spectrum
and then use this to determine the other parts of the spectrum of T and T ∗ .

Proposition 5.2. If X is a Hilbert space and T ∈ B(X) then

(a) λ ∈ σ(T ) if and only if λ̄ ∈ σ(T ∗ ),

(b) λ ∈ σp (T ) implies λ̄ ∈ σp (T ∗ ) ∪ σr (T ∗ ),

(c) λ ∈ σr (T ) implies λ̄ ∈ σp (T ∗ ).

Proof.

(a) If λ ∈ ρ(T ) then T − λI is invertible, so there is an S ∈ B(X) such that S(T − λI) =
(T − λI)S = I. Taking adjoints we get (T ∗ − λ̄I)S ∗ = S ∗ (T ∗ − λ̄I) = I, so λ̄ ∈ ρ(T ∗ ).
Applying this to T ∗ gives the reverse relation. Therefore, λ ∈ ρ(T ) ⇐⇒ λ̄ ∈ ρ(T ∗ ).

(b) If λ ∈ σp (T ) then the kernel of T − λI 6= (0) and hence, by Lemma 5.2, (range (T ∗ −
λ̄))⊥ 6= (0). Since E ⊥⊥ = Ē for any subset E of a Hilbert space, we obtain range (T ∗ − λ̄) 6=
X, that is the range of T ∗ − λ̄ is not dense in X. Hence λ̄ ∈/ σc (T ∗ ), but λ̄ ∈ σ(T ∗ ) from
(a) so λ̄ ∈ σp (T ∗ ) ∪ σr (T ∗ ).

(c) If λ ∈ σr (T ) then ker T − λI = (0) and range (T − λI) 6= X. So, from Lemma 5.2, we
have
⊥
ker (T ∗ − λ̄I) = range (T − λI)⊥ = range (T − λI) 6= X ⊥ = {0}.

So λ̄ ∈ σp (T ∗ ).

5.2 Compact Operators

Definition 5.3 (Compact maps). If X and Y are normed spaces and T : X → Y is linear
then T is said to be compact if the closure of T (B) is compact for all bounded subsets B of
X.

59
 ¯ is
Because translation and scaling are continuous operations, T is compact iff T (B)
compact just when B is the zero-centered closed unit ball in X.
Proposition 5.3. Suppose X and Y are normed spaces and T : X → Y is linear.

(a) If T is compact then T is continuous.

(b) T is compact iff T xn has a convergent subsequence for every bounded sequence xn in X.

Proof.

(a) If B is the unit ball in X then T (B) is compact and hence bounded. So there is a C > 0
so that supkxk≤1, x∈X kT xk ≤ C and hence T is bounded and continuous.

(b) Suppose T is compact. If {xn } be a bounded sequence in X then {T xn }n≥1 is precompact

and hence it has a convergent subsequence in Y .
Now suppose the converse hypothesis holds and B is a bounded subset of X. To show
that T (B) is precompact, it is enough to show that every sequence in T (B) has a
convergent subsequence but this is guaranteed by the hypothesis and we are done.

Proposition 5.4. Suppose X, Y are normed spaces.

(a) If T : X → Y is linear and X is finite dimensional then T is compact.

(b) If T ∈ B(X, Y ) and Y is finite dimensional then T is compact.

(c) The compact linear maps form a subspace of B(X, Y ).

Note that (a) is not true if X is infinite dimensional because the identity operator on X
is not compact if X is infinite dimensional.

Proof.

(a) Since X is finite dimensional, the range of T will be finite dimensional and hence by
Theorem 2.2, the range will be isomorphic (as a normed space) to Cn (or Rn ). Now T
may be considered as a linear map from a finite dimensional space to a finite dimensional
Banach space and hence T is bounded. If B is a bounded subset of X then T (B) is
bounded and since it is a bounded subset of Cn (up to isomorphism) its closure will be
compact.

60
(b) Since Y is finite dimensional, Y is isomorphic to Cn . If B is a bounded subset of X then
T (B) is bounded because T is bounded. Hence T (B) is a bounded subset of Cn so its
closure is compact.

(c) If T ∈ B(X, Y ) is compact then using the sequential definition one can easily check that
αT is compact for any scalar α. Now suppose S, T ∈ B(X, Y ) are compact. If xn is
a bounded sequence in X then Sxn has a convergent subsequence Sxnk . Since xnk is
bounded, T xnk has a convergent subsequence T xnkl . Hence (S + T )xnkl is a convergent
subsequence of (S + T )xn .

We next give an example of a compact operator on the Banach space C[0, 1].
Example. Given the continuous function k : [0, 1] × [0, 1] → R define the map K : C[0, 1] →
C[0, 1] with Z 1
(Kf )(x) := K(x, y)f (y)dy, x ∈ [0, 1].
0
In the homework, we have shown that K is a bounded linear operator. We now claim that
K is compact. This can be proved by showing that the image of bounded subset of C[0, 1]
is an equicontinuous and bounded subset of C[0, 1] and then appeal to the Arzela-Ascoli
theorem. We leave this verification as a homework problem.

There is another way to prove the compactness of K using the Weierstrass approximation
theorem instead of the Arzela-Ascoli theorem. By Weierstrass’s theorem, we can approximate
k(x, y) by a sequence of polynomials kn (x, y) such that kk − kn k goes to zero. If K is the
operator corresponding to the kn then one can show that kKn − Kk approaches zero. Finally
because kn (x, y) is a polynomial one can see that the range of Kn is finite dimensional - for
example if kn (x, y) = x2 + xy + x + 2y then
Z 1 Z 1 Z 1
2
(Kn f )(x) = x f (y) dy + x yf (y) dy + 2 yf (y) dy
0 0 0

- and hence Kn is compact. So the compactness of K will follow from proposition 5.5.

Just as the Arzela-Ascoli theorem gives us the compact subsets of C(X), everyone’s
repertoire should also include the compact subsets of the other Banach spaces we have
studied - `p and Lp (R), for p ≥ 1.

We now state some additional useful properties of compact operators.

Proposition 5.5. Suppose X, Y, Z are Banach spaces.

61
(a) If S ∈ B(X, Y ) and T ∈ B(Y, Z), and one of S, T is compact then T S is compact.

(b) If Tn is a sequence of operator in B(X, Y ) converging to T ∈ B(X, Y ) in the operator

norm with each Tn compact, then T is compact.

So (b) combined with proposition 5.4 implies that the set of compact operators in B(X, Y )
is a closed subspace of B(X, Y ).

Proof.

(a) Suppose B is a bounded subset of X. If S is compact and T bounded then S(B) is

compact and hence T (S(B)) is compact because T is continuous. Hence T S(B) will be
precompact because it is a subset of T (S(B)).
If T is compact and S bounded then S(B) is bounded and hence T S(B) is precompact.

(b) Suppose xn is a bounded sequence in X, so ||xn || ≤ c for all n ≥ 1, for some c > 0.
Since T1 is compact, there is a subsequence {x1,k }k≥1 such that T1 (x1,k ) is convergent.
Since x1,k is also bounded and T2 is compact, there is a subsequence x2,k of x1,k such
that T2 (x2,k ) is convergent. In this fashion, we can construct sequences xn,k such that

• {xn+1,k }k≥1 is a subsequence of {xn,k }k≥1 ;

• {Tn (xn,k )}k≥1 is convergent for all n = 1, 2, ...

Clearly {xn,n }n≥1 is a subsequence of {xn }n≥1 ; we claim that the sequence {T xn,n }n≥1 is
convergent. Since Y is a Banach Space, it is complete, so to prove our claim it is enough
to show that {T (xn,n )}n≥1 is a Cauchy sequence.
Given  > 0, from the hypothesis, there is an N1 such that ||T − Tn || <  if n ≥ N . Also,
since {TN (xN,k )}k≥1 is convergent, there is an N2 such that

kTN (xN,j ) − TN (xN,k )k <  if j, k ≥ N2 .

Hence, for any m, n ≥ max(N1 , N2 ), we have

kT (xn,n ) − T (xm,m )k ≤ kT xn,n − TN xn,n k + kTN xn,n − TN xm,m k + kTN xm,m − T xm,m k
= k(T − TN )(xn,n )k + kTN xn,n − TN xm,m k + k(TN − T )(xm,m )k
≤ kT − TN k kxn,n k + kTN xn,n − TN xm,m k + kTN − T k kxm,m k
≤ c +  + c = (1 + 2c).

Above, we used the observation that {xn,n }n≥N is a subsequence of {xN,k }k≥1 .

62
Proposition 5.6. Suppose X and Y are Banach spaces and T ∈ B(X, Y ) then T is compact
iff its adjoint is compact.

Proof. If B is the closed unit ball of X then T (B) is compact. If {gn } be a bounded
sequence in Y ∗ and then the gn may be regarded as a sequence of continuous functions
on the compact set T (B) which are uniformly bounded and equicontinuous and hence we
can apply the Arzela-Ascoli theorem. Please see the details in Rudin’s book on Functional
Analysis.

5.3 Spectral Theory for Compact Operators

We now state some results about the spectrum of compact operators which will culminate
in the very useful Fredholm Alternative and the Spectral Theory of Self-adjoint Compact
operators on Hilbert spaces.

Theorem 5.1. Suppose X is a Banach space, T ∈ B(X), and T is compact.

(i) If dim(X) = ∞ then 0 ∈ σ(T ).

(ii) If λ 6= 0 then the kernel of T − λI is finite dimensional.

(iii) If λ 6= 0 then the range of T − λI is closed.

Proof.

(i) If 0 is not in the spectrum of T then T − 0I = T is invertible. So there is an S ∈ B(X)

such that ST = I. Now T is compact so I = ST is compact so if B is the closed unit
ball B then ST (B) = I(B) = B = B is compact. Hence by Theorem 2.3 X is finite
dimensional.

(ii) If N = ker (T − λI) then N is a closed subspace of a Banach space X so N is also a

Banach space. So if B is the closed unit ball in N , then by the definition of N , we have
(T − λI)(B) = (0) so T (B) = λB. But B is bounded and T is compact so

λB = λB = T (B)

is compact, hence B is compact, so, by Theorem 2.3, N is finite dimensional.

63
(iii) We will prove this for any Hilbert space X; a variation of this works for Banach spaces.
To prove that the range of T − λI is closed one usually tries to show that there is an
estimate of the form
k(T − λI)xk ≥ c kxk , ∀x ∈ X
for some c > 0. However such an estimate cannot exist if λ is an eigenvalue because
then, for eigenvectors x, the LHS will be zero while the RHS will be non-zero. The way
to fix this problem is to drop the eigenvectors, as shown next.
If N = Ker(T − λI) then N is closed and hence X = N ⊕ N ⊥ . We define the operator
S : N ⊥ → X by
Sx = (T − λI)x, x ∈ N ⊥.
Note that the range of S is the same as the range of T − λI, and so it is enough to
show that Range(S) is closed. We do this by showing there is a c > 0 such that
kSxk ≥ c kxk , ∀x ∈ N ⊥ (5.2)
Note that the kernel of S is (0) so our earlier objection to existence of such an estimate
is no longer valid.
We first show why (5.2) implies that the Range(S) is closed. If y is a limit point of the
range of S then there is a sequence xn in N ⊥ such that Sxn converges to y. Hence Sxn
is a Cauchy sequence and hence
kSxn − Sxm k = kS(xn − xm )k ≥ c kxn − xm k
implies that xn is a Cauchy sequence in N ⊥ . Now N ⊥ is closed and a Banach space so
there is an x ∈ N ⊥ so that xn → x and hence Sxn → Sx. So y = Sx is in the range of
S and the range of S closed.
It remains to show that (5.2) holds or equivalently that
kSxk ≥ c > 0, ∀x ∈ N ⊥ , kxk = 1 (5.3)
Suppose not; then there is a sequence xn in N ⊥ such that kxn k = 1 and kSxn k → 0.
Since T is compact and {xn }n≥1 is a bounded sequence, {T xn }n≥1 has a convergent
subsequence {T xnk }k≥1 - say T xnk → y ∈ X. Now
Sxnk = (T − λI)xnk = T xnk − λxnk → 0.
Hence λxnk → y, so xnk → λ−1 y (note λ 6= 0) so
y = lim T xnk = λ−1 T y
k→∞

that (T − λI)y = 0, implying y ∈ N . Now λxnk → y, so y is the limit of a sequence in

N ⊥ so y ∈ N ⊥ ; also y ∈ N - hence y = 0. But λxnk → y so
kyk = lim kλxnk k = |λ| =
6 0
k→∞

This is a contradiction and so (5.3) holds.

64
Remark. If T ∈ B(X) is compact then, in the previous theorem, (i) asserts that T is not
invertible if dim(X) = ∞, (ii) asserts that if λ 6= 0 is an eigenvalue of T then the corre-
sponding eigenspace is finite dimensional, and (iii) asserts that the only possible entry in the
continuous spectrum of T is 0 because if λ 6= 0, T − λI is injective and the range of T − λI
is dense in X then the range equals X, so T − λI is a bijection and hence invertible and λ
would not be in the spectrum.

The following theorem is useful in the study of PDEs and Integral Equations. The
statement of the theorem includes results which were stated in some of the earlier theorems
so that all the relevant useful results are available in one place.

Theorem 5.2 (Spectral Theorem for Compact Operators). If X is a Hilbert Space, T ∈

B(X), and T is compact then the following hold:

(a) (Fredholm Alternative) If λ 6= 0 then ker (T − λI) and ker (T ∗ − λ̄I) are finite dimen-
sional, have the same dimension and

range (T − λI) = (ker (T ∗ − λ̄I))⊥

range (T ∗ − λ̄I) = (ker (T − λI))⊥ .

(b) λ ∈ σ(T ) if and only if λ̄ ∈ σ(T ∗ ).

(c) If λ ∈ σ(T ), λ 6= 0 then λ ∈ σp (T ).

(d) σ(T ) is compact, at most countable, and has at most one limit point - namely 0.

(e) If X is infinite dimensional then 0 ∈ σ(T ).

We discuss some of the implications of the results in Theorem 5.2.

Suppose T ∈ B(X) is compact, λ 6= 0, and we want to know whether the equation

(T − λI)x = y has a solution for every y ∈ X, that is whether range (T − λI) = X? From
(a), this is equivalent to the question whether (ker (T ∗ − λ̄I))⊥ = X or equivalently whether
ker (T ∗ − λ̄I) = {0}. From (a), the dimensions of the kernels of T − λI and T ∗ − λI are
equal, so T − λI is surjective iff ker (T − λI) = {0}. So, for T compact, so λ 6= 0, the
operator T − λI will be surjective iff T − λI is injective.

Since E ⊥ is always closed for any subset E of X, (a) also implies that the ranges of T −λI
and T ∗ − λ̄I are closed subsets of X.

65
 (c) is a consequence of (a) as seen next. Let T ∈ B(X), T compact, λ ∈ σ(T ) and λ 6= 0.
Suppose T − λI is injective; then by the equality of the dimensions of the kernels of T − λI
and T ∗ − λ̄I, we have ker (T ∗ − λ̄I) = {0} which, from (a), implies range (T − λI) = X, that
is (T − λI is surjective. But T − λI was assumed to be injective so by the Open Mapping
Theorem T − λI is invertible which contradicts the hypothesis that λ ∈ σ(T ).
Definition 5.4 (Self Adjoint). If X is a Hilbert space and T ∈ B(X) then T is said to be
self-adjoint if T ∗ = T.
Definition 5.5 (Separable). A Banach Space X is said to be separable if X has a countable
dense subset.
Lp (R), 1 ≤ p < ∞, and C[a, b] are examples of separable Banach spaces.

We now state the spectral theorem for compact self-adjoint operators which play an
important role in the proof of the existence of solutions of linear elliptic PDEs.
Theorem 5.3 (Spectral Theorem For Compact Self-adjoint Operators). If X is a separa-
ble, infinite dimensional Hilbert space and T ∈ B(X), T compact and self-adjoint then the
following hold:
(a) every singular value of T is an eigenvalue and real valued;
(b) there is a countable maximal orthonormal set {un }n≥1 of eigenvectors such that T un =
λn un and limn→∞ λn = 0.

5.4 Sturm-Liouville Theory

In this section all functions will be real valued, so L2 [0, 1] will consist only of real valued
functions. Further q will be a real valued function in L∞ [0, 1].

Our goal is to determine for which f will the boundary value problem
−u00 + q(x)u = f (x), 0 ≤ x ≤ 1, (5.4)
u(0) = 0 = u(1) (5.5)
has a solution and, if it does, how many solutions will it have. This problem is called a
Sturm-Liouville problem. For ODEs mostly initial value problems are studied, so for (5.4),
one usually prescribes u(0) and u0 (0) instead of the boundary condition (5.5), but boundary
value problems also arise naturally, particularly for partial differential equations.

In this section, we first study an example, and then study rigorously the boundary value
problem (5.4), (5.5), with the help of the spectral theorem for compact self-adjoint operators.

66
5.4.1 An example

We consider the boundary value problem

u00 + u = f (x), 0 ≤ x ≤ π, (5.6)

u(0) = 0 = u(π), (5.7)

and ask for which f will (5.6), (5.7) have a solution.

We first study the eigenvalues associated to this problem. So we ask for which λ will the
boundary value problem

u00 + u = λu, 0 ≤ x ≤ π, (5.8)

u(0) = 0 = u(π) (5.9)

have a non-zero solution. Since u00 +u = λu is equivalent to u00 +(1−λ)u = 0, using standard
methods one may argue the boundary value problem has a non-zero solution iff 1 − λ = n2
with n = 1, 2, · · · and the corresponding eigenfunctions are sin(nx). Summarizing, the
eigenvalues of the system are λn = 1 − n2 , n = 1, 2, · · · with the (normalized) eigenfunctions
1
un = √ sin(nx), that is, for n = 1, 2, · · ·
2π

u00n + un = λn un , 0 ≤ x ≤ π,
un (0) = 0 = un (π).

Most importantly we observe that the eigenfunctions {un }n≥1 form a maximal orthonormal
set for L2 [0, π]. This will be crucial.

So we return to our question: for which f ∈ L2 [0, π] will (5.6), (5.7) have a solution.
Since f ∈ L2 [0, π] it has an expansion
∞
X
f (x) = cn un (x),
n=1
Rπ
with cn = 0
f (x)un (x) dx, so we seek a solution of (5.6), (5.7) in the form
∞
X
u(x) = an un (x), (5.10)
n=1

that is, the cn are given and we ask whether we can find an such that u(x) satisfies (5.6),
(5.7).

67
 It is clear that u(x) given by the expansion (5.10) satisfies (5.7) because each of the un
does; we just have to satisfy (5.6). From (5.10)
∞
X ∞
X
00
u +u= an (u00n + un ) = an λ n u n
n=1 n=1

so (5.6) holds iff

an λn = cn , n = 1, 2, 3, · · · . (5.11)
So we get an = cn /λn for all n ≥ 1, except for those n where λn = 0, that is when n = 1.
When n = 1, (5.11) forces cP1 = 0 and a1 may be chosen arbitrarily. In addition we note that
since f ∈ L2 [0, π], we have ∞ 2
n=1 |cn | is finite, hence (noting λn ≥ 1 for n ≥ 2)

∞ ∞ ∞
X X |cn |2 X
|an |2 = 2
≤ |cn |2
n=2 n=2
|λn | n=2

is also finite so the u(x) given by (5.10) is in L2 [0, π].

Our conclusionR is the following: if f ∈ L2 [0, 2π], the boundary value problem (5.6), (5.7)
π
has a solution iff 0 f (x) sin x dx = 0, and in this case, the solutions are
∞
X cn
u(x) = un (x) + a1 sin x
λ
n=2 n
Rπ
where cn = √1 f (x) sin(nx) dx and a1 may be chosen arbitrarily.
2π 0

Just to reinforce the conclusion, consider the boundary value problem

u00 + u = sin x, 0 ≤ x ≤ π,
u(0) = 0 = u(π).

The general solution of the ODE is the general solution of the associated homogeneous
equation added to a particular solution (use variation of parameters) of the inhomogeneous
equation, that is
1
u(x) = a cos x + b sin x − x cos(x).
2
This u(x) must also satisfy the boundary condition u(0) = 0, u(π) = 0 which forces a = 0 and
−π cos π = 0 which is not true. So there is no solution, just as promised by our conclusion
because Z π
1
sin x sin x dx = 6= 0.
0 2π

68
5.4.2 The general case

Before we tackle the boundary value problem (5.4), (5.5) we first study formally the unique-
ness question - when does (5.4), (5.5) have at most one solution. Suppose u1 and u2 both
solve (5.4), (5.5); then v = u1 − u2 satisfies

−v 00 + q(x)v = 0, 0 ≤ x ≤ 1,
v(0) = 0 = v(1).

Then, using an integration by parts, we observe that

Z 1 Z 1 Z 1
00 0 1 0 2
0= v (−v + qv) dx = −vv |0 + (v ) dx + qv 2 dx
Z0 1 0 0

= (v 0 )2 + qv 2 dx.
0

So if q(x) ≥ 0 for all x ∈ [0, 1] then v 0 = 0 on [0, 1] so v is constant on [0, 1]; now v(0) = 0
so v = 0 on [0, 1] and hence u1 = u2 . So we see that if q(x) ≥ 0 for all x ∈ [0, 1] then the
boundary value problem (5.4), (5.5) has at most one solution. Our strategy for studying the
boundary value problem (5.4), (5.5) will be to first study the case where q(x) ≥ 0 for all
x ∈ [0, 1] and then relate the general q problem to this case.

Since q is bounded on [0, 1], we can find an α > 0 and a c > 0 so that

q(x) + α ≥ c > 0, ∀x ∈ [0, 1].

We will first study the following intermediate question; for which f ∈ L2 [0, 1] will

−u00 + (q + α)u = f, 0 ≤ x ≤ 1, (5.12)

u(0) = 0 = u(1) (5.13)

have a solution, and how many solutions does it have; note the α in the differential equation.

To answer this question, we first weaken what we mean by a solution. If u was a twice
differentiable solution of (5.12), (5.13), then, for any v ∈ C 1 [0, 1] with v(0) = 0 = v(1), we
would have
Z 1 Z 1 Z 1
00 0 1
fv = (−u + (q + α)u)v dx = −u v|0 + u0 v 0 + (q + α)uv dx
0
Z0 1 0

= u0 v 0 + (q + α)uv dx,
0

69
that is Z 1 Z 1
0 0
u v + (q + α)uv dx = f v dx (5.14)
0 0

for all v ∈ C 1 [0, 1] with v(0) = 0 = v(1). Conversely, if (5.14) holds for all v ∈ C 1 [0, 1] with
v(0) = v(1) = 0 and u is a C 2 function, then by taking v arbitrarily close to −u00 + (q + α)u
while satisfying v(0) = v(1) = 0, we can show that u must satisfy (5.12), (5.13).

Motivated by the above discussion, we define a Hilbert space where the solutions of (5.12),
(5.13) will reside. Define the vector space

X = u : [0, 1] → R : u ∈ L2 [0, 1], u0 ∈ L2 [0, 1], u(0) = u(1) = 0



with the inner product

Z 1
hu, viX = u0 v 0 + (q + α)uv dx, ∀u, v ∈ X.
0

We first observe that we do indeed have an inner product because

Z 1 Z 1
02 2
hu, uiX = u + (q + α)u dx ≥ c u2 ≥ ckuk2L2 [0,1] ≥ 0
0 0

and is zero iff u = 0. We also state without proof that X is complete and hence a Hilbert
space. We do not discuss what it means for a function in L2 [0, 1] to have a derivative or how
it can have a value at x = 0 and x = 1 when functions in L2 [0, 1] are only defined uniquely on
the complements of sets of measure zero. This is discussed in the theory of Sobolev spaces.
For use later we record that we have shown
1
kukL2 [0,1] ≤ √ kukX , ∀u ∈ X. (5.15)
c

Clearly there is a natural linear imbedding of i : X → L2 [0, 1] sending u → u and the

above calculation shows that i is bounded because
1
ki(u)kL2 [0,1] = kukL2 [0,1] ≤ √ kukX .
c

Further, one may show that X is dense in L2 [0, 1] because all infinitely differentiable functions
on R with support in [0, 1] are in X.

Finally, for the solution of the Sturm-Lioville problem, without proof, we make the very
important claim that the imbedding i is a compact operator. For i to be compact every
bounded sequence in X would need to have a subsequence which is convergent in L2 [0, 1].

70
To motivate why this holds, suppose that instead of using the L2 norm we use the max norm;
then a bounded sequence {un }n≥1 in X is one for which there is a real C so that
max |un (x)| + |u0n (x)| ≤ C, n = 1, 2, · · · .
x∈[0,1]

So the sequence is uniformly bounded in C[0, 1]; further because of the uniform bound on
u0n , by using the Mean Value Theorem on [0, 1], we can argue that the sequence {un }n≥1
is equicontinuous in C[0, 1]. Hence, by the Arzela-Ascoli Theorem, {un }n≥1 will have a
subsequence which is convergent in C[0, 1].

We now come to the main reason we defined the Hilbert Space X. We saw that if u was
a C 2 solution of (5.12), (5.13) then (5.14) holds for all v ∈ C 1 [0, 1] with v(0) = v(1) = 0.
Now (5.14) may also be written as
Z 1
hu, viX = fv
0

so we weaken the meaning of a solution of (5.12), (5.13).

Definition 5.6. A function u ∈ X is a weak solution of (5.12), (5.13) if
Z 1
hu, viX = f v, ∀v ∈ X.
0

So our intermediate question is for which f ∈ L2 [0, 1] will (5.12), (5.13) have a weak
solution. We show that the intermediate problem (5.12), (5.13) has a unique weak solution
for all f ∈ L2 [0, 1].
Proposition 5.7. If q ∈ L∞ [0, 1] and f ∈ L2 [0, 1], then (5.12), (5.13) has a unique weak
solution in X, that is, there exists a unique u ∈ X such that
Z 1
hu, viX = f v, ∀v ∈ X.
0

Further kukX ≤ √1 kf kL2 [0,1] .

c

Proof. Given f ∈ L2 , define the functional F : X → R with

Z 1
F (v) = f v dx.
0

Clearly F is linear and F is bounded because

Z 1
1
|F (v)| = | f v dx| ≤ kf kL2 [0,1] kvkL2 [0,1] ≤ √ kf kL2 [0,1] kvkX .
0 c

71
In fact we have shown that (using (5.15)) that
1
kF k ≤ √ kf kL2 [0,1] .
c
So by the Riesz representation theorem there is a unique u ∈ X so that F (v) = hv, uiX for
all v ∈ X, that is Z 1
hu, viX = f v dx, ∀v ∈ X
0
Further, from the Riesz representation theorem
1
kukX = kF k ≤ √ kf kL2 [0,1] .
c

Proposition 5.7 permits us to define the solution map S : L2 [0, 1] → X where, for each
f ∈ L2 [0, 1], Sf = u is the weak solution of the intermediate boundary value problem (5.12),
(5.13), that is Z 1
hSf, viX = hu, viX = f v, ∀v ∈ X.
0
We note that S is linear because if u1 , u2 ∈ X are the weak solutions corresponding to
f1 , f2 ∈ L2 [0, 1] and β is any scalar then αu1 + u2 ∈ X is the weak solution corresponding
to βf1 + f2 because, for all v ∈ X, we have
Z 1 Z 1 Z 1
hβu1 + u2 , viX = βhu1 , viX + hu2 , viX = β f1 v dx + f2 v dx = (βf1 + f2 )v dx.
0 0 0

Further S is bounded because, by Proposition 5.7, we have

1
kSf kX = kukX ≤ √ kf kL2 [0,1] .
c

We define the bounded linear operator T : L2 [0, 1] → L2 [0, 1] with T = i ◦ S, that is

T f = u where u ∈ X is the weak solution of (5.12), (5.13) but now we think of u as residing
in L2 [0, 1] (which contains X). We note the important observation that T is compact because
i is compact. We also note that T is self adjoint which is shown next. If f1 , f2 ∈ L2 [0, 1] then
let u1 , u2 ∈ X be the weak solutions of (5.12), (5.13) with f replaced by f1 , f2 respectively.
Then, by definition, T fi = ui and
Z 1
hT f1 , viX = hu1 , viX = f1 v dx, ∀v ∈ X,
0
Z 1
hT f2 , viX = hu2 , viX = f2 v dx, ∀v ∈ X.
0

72
Hence using the above relations with v = u2 in the first relation and v = u1 in the second
relation we obtain
Z 1
hT f1 , f2 iL2 [0,1] = hu1 , f2 iL2 [0,1] = u1 f2 dx = hu2 , u1 iX
0
Z 1
hf1 , T f2 iL2 [0,1] = hf1 , u2 iL2 [0,1] = f1 u2 dx = hu1 , u2 iX ,
0

so T is self-adjoint because of the symmetry of the real inner product on X. Finally, T is

injective because if T f = 0 for some f ∈ L2 [0, 1] then
Z 1
0 = h0, viX = hT f, viX = f v dx, ∀v ∈ X.
0

This implies f = 0 because X is dense in L2 [0, 1]. Summarizing

• T : L2 [0, 1] → L2 [0, 1] is a compact self-adjoint operator whose range is contained in

X (note X is imbedded in L2 [0, 1]);

• for each f ∈ l2 [0, 1], u = T f ∈ X is the unique weak solution of the intermediate
problem (5.12), (5.13), that is
Z 1
hu, viX = hT f, viX = f v dx ∀v ∈ X;
0

• T is injective and hence 0 is not an eigenvalue for T .

Since T is compact and self-adjoint, from the spectral theorem for compact self-adjoint
operators, there is a maximal orthonormal set {un }n≥1 of functions in L2 [0, 1] where each
un is an eigenfunction of T and if T un = σn un then the σn are real and limn→∞ σn = 0. We
note that σn 6= 0 for all n because T is injective. Since the range of T is contained in X and
σn 6= 0, the un are in X for all n.

Having resolved the intermediate boundary value problem (5.12), (5.13), we now resolve
the original problem - for which f ∈ L2 [0, 1] will the boundary value problem (5.4), (5.5)
have a solution. Given f ∈ L2 [0, 1], u ∈ X is a weak solution of (5.4), (5.5) iff
Z 1 Z 1
fv = u0 v 0 + quv dx, ∀v ∈ X
0 0

that is iff Z 1 Z 1
(f + αu)v = u0 v 0 + (q + α)uv dx = hu, viX ∀v ∈ X
0 0

73
that is iff u = T (f + αu). So u ∈ X is a weak solution of (5.4), (5.5) iff
u − αT u = T f ; (5.16)
note that α 6= 0. We now attempt to find a solution of (5.16).

the expansion f = ∞
P∞
Since f ∈ L2 [0, 1] we have P 2
P
n=1 c n un with n=1 |cn | finite. Since
2 ∞
u ∈ X ⊆ L [0,P1], we have u = n=1 an un for some real an ; we just have to find the an and
∞
be sure that n=1 |an |2 is finite. Since T un = σn un we have
∞
X ∞
X
u − αT u = (1 − ασn )an un , Tf = σn cn un ,
n=1 n=1

hence (5.16) holds iff

(1 − ασn )an = σn cn . (5.17)
So if σn 6= 1/α then
σn
an =
cn (5.18)
1 − ασn
and if σn = 1/α then we must have cn = 0. Further if σn = 1/α then σn 6= 0 and, since T is
compact, the corresponding eigenspace will be finite dimensional, so if I = {n : σn = 1/α}
then I is a finite and possibly empty set. We also note that limn→∞ σn = 0 so there is a
σ > 0 so that
|1 − ασn | ≥ σ > 0 ∀n ∈
/ I.
So, for n ∈
/ I if we take an as given by (5.18) and take an to be arbitrary for n ∈ I then
∞
X X X X X σn2
|an |2 = |an |2 + |an |2 = |an |2 + |an |2
n=1 n∈I n∈I
(1 − ασn )2
n∈I
/ n∈I
/
X X
2 2
≤ |an | + C |an |
n∈I n∈I
/
P∞
which is finite, so u = n=1 an un will be in L2 [0, 1]. Of course, since this u ∈ L2 [0, 1] satisfies
(5.16), this u is actually in X because u = αT u + T f and the range of T is in X. Hence we
have proved the following proposition except for a few gaps.
Theorem 5.4 (Sturm-Liouville). Suppose q is a real valued function in L∞ [0, 1].R For any f ∈
1
L2 [0, 1], the boundary value problem (5.4), (5.5) has a weak solution u ∈ X iff 0 f un dx = 0
for all n ∈ I. In this case, all solutions of the boundary value problem (5.4), (5.5) are given
by
X∞
u= an u n
n=1
where an may be chosen arbitrarily for n ∈ I and
Z 1
σn
an = f un dx, ∀n ∈
/ I.
1 − ασn 0

74
 The solution of the problem required choosing an α and the construction of the opera-
tors S and T and the determination of the eigenvalues and eigenfunctions of T . We now
reinterpret the result in terms of the original boundary value problem without using α, S
and T .

If (σ, u) is an eigenpair for T then T u = σu, that is σu is the weak solution of the
intermediate boundary value problem (5.12), (5.13) with f = u, that is, assuming that u has
two derivatives,
−(σu)00 + (q + α)σu = u, 0 ≤ x ≤ 1,
u(0) = u(1) = 0.
This may be rewritten as
−u00 + qu = λu, 0 ≤ x ≤ 1, (5.19)
u(0) = u(1) = 0 (5.20)
where λ = (1 − ασ)/σ. In view of this, Theorem 5.4, regarding the solution of the boundary
value problem (5.4), (5.5), may be reinterpreted as given below.

• Find the eigenvalues and eigenfunctions of (5.19), (5.20). There is a countable maximal
orthonormal set of eigenfunctions {un }n≥1 in L2 [0, 1] so that if λn is the eigenvalue
corresponding to un then limn→∞ |λn | = ∞.
• If I = {n : λn = 0} then I is a finite and possibly empty set.
R1
• Given f ∈ L2 [0, 1], the boundary value problem (5.4), (5.5) has a solution iff 0
f un dx =
0 for all n ∈ I and in this case any solution will be given by
∞
X
u= an un
n=1

where an may be chosen arbitrarily for n ∈ I and

Z 1
1
an = f un dx, ∀n ∈
/ I.
λn 0

5.5 Spectral Theorem for Self-Adjoint Operators

Definition 5.7. If X and Y are Hilbert spaces then an operator U ∈ B(X, Y ) is said to be
unitary if U ∗ U = IX and U U ∗ = IY

75
 A unitary operator is an isomorphism of Hilbert spaces.

Proposition 5.8. If X and Y are Hilbert spaces and U ∈ B(X, Y ) then U is unitary iff U
is surjective and hU x, U x0 iY = hx, x0 iX for all x, x0 ∈ X.

Proof. If U is unitary then U is invertible and hence surjective. Further U ∗ U = IX implies

that
hx, x0 iX = hU ∗ U x, x0 iX = hU x, U x0 iY ∀x, x0 ∈ X.
Conversely, the above relation in reverse gives us U ∗ U = IX . This also implies U is injective
and since U is also surjective, U must be bijective. So, by the open mapping theorem, U is
invertible - let the inverse be T ∈ B(Y, X) and then T U = IX , U T = IY . Hence

U ∗ = U ∗ IY = U ∗ (U T ) = (U ∗ U )T = IX T = T

and hence U U ∗ = IY .

For a finite dimensional Hilbert space X, an operator T ∈ B(X) is self-adjoint iff X has
an orthonormal basis consisting only of eigenvectors of T . This is equivalent to having a
unitary operator U ∈ B(Fn , X) such that U ∗ T U is a diagonal matrix. This means that if
T is self-adjoint then there is an isomorphism of inner product spaces from X to Fn and
under this isomorphism T corresponds to a diagonal matrix. We show that, in some sense,
a similar result holds in infinite dimensions.

We first study an example which typifies self-adjoint operators on Hilbert spaces; then we
define a measure space, a concept needed in the spectral theorem for self-adjoint operators,
and lastly we state the spectral theorem for self-adjoint operators on a separable Hilbert
space.

5.5.1 An example of a self-adjoint operator

Below m(E) will denote the Lebesgue measure of a measurable subset E of R. Recall that
 Z b 
2 2
L (a, b) = f : (a, b) → C, |f | < +∞
a

is a Hilbert space with the inner product

Z b
hf, gi = f ḡ.
a

76
Further
L∞ (a, b) = {f : (a, b) → C, |f | bounded and measurable}
is a Banach space with the norm

kf k∞ = inf{M : |f | ≤ M a.e. on (a, b)}.

One may show that |f | ≤ kf k∞ almost everywhere on (a, b).

Definition 5.8 (Essential Range). For a measurable function φ : (a, b) → C, the essential
range of φ consists of all complex numbers t for which the set {x ∈ (a, b) : |f (x) − t| < }
has positive measure for all  > 0.

Suppose φ ∈ L∞ (a, b) and define the multiplication operator Tφ : L2 (a, b) → L2 (a, b) with

Tφ (f ) = φf, ∀f ∈ L2 (a, b).

Note that this is pointwise multiplication and not a composition of functions. Since
Z b Z b
2
2
kTφ f k = 2
|f φ| ≤ kφkL∞ |f |2 = kφk2∞ kf k22 ,
a a

and since Tφ is linear, Tφ is a bounded linear operator on L2 (a, b) with kTφ k ≤ kφk∞ ; actually
one may show that kTφ k = kφk∞ .

We now compute the adjoint of Tφ . For f, g ∈ L2 (a, b) we have

Z b Z b
hTφ f, gi = φf g = f φg = hf, Tφ gi;
a a

thus Tφ∗ = Tφ . Hence Tφ is a normal operator because

Tφ∗ Tφ = T|φ|2 = Tφ Tφ∗ ,

and Tφ is self-adjoint if and only if φ is real valued. We now compute the spectrum of Tφ .

Proposition 5.9. For any φ ∈ L∞ (a, b), Tφ is a normal operator and is self adjoint if φ is
real valued. The spectrum of Tφ is the essential range of φ and

• σp (Tφ ) consists of all λ ∈ C for which m({x ∈ (a, b) : φ(x) = λ}) > 0,

• σc (Tφ ) = ess range φ \ σp (Tφ ),

• σr (Tφ ) = ∅.

77
Proof. For any λ ∈ C we have

Tφ − λI = Tφ−λ ,

thus to prove the proposition it is enough to determine when 0 is in the point spectrum, the
continuous spectrum or the residual spectrum of Tφ .

If 0 is not in the essential range of φ then there is an  > 0 such that |φ| ≥  almost
everywhere and hence 1/φ ∈ L∞ (a, b). Therefore T1/φ Tφ = Tφ T1/φ = T1 = I so Tφ is
invertible and 0 is not in the spectrum of Tφ . If 0 is in the essential range of φ, then for each
positive integer n, the set

En = {x ∈ (a, b) : |φ(x)| < 1/n}

has positive measure (which will be finite). Therefore, for every n

Z Z
2 2 1 1
kTφ χEn k2 = |φ| ≤ 2 = 2 kχEn k22 ,
En n En n

which implies Tφ is not invertible because if Tφ was invertible then there would be a c > 0
such that
kTφ f k22 ≥ ckf k22 , ∀f ∈ L2 (a, b)
which would mean that 0 < c ≤ 1/n2 for all n, which is not possible. So we have shown that
0 ∈ σ(Tφ ) iff 0 ∈ ess range φ.

Now suppose 0 is in the essential range of φ. If the set

E = {x ∈ (a, b) : φ(x) = 0}

has positive measure (which will be finite) then χE ∈ L2 (a, b) and

Tφ χE = φχE = 0

so 0 is an eigenvalue. If m(E) = 0 then Tφ is injective because if Tφ f = 0 for some f ∈ L2 (a, b)

then f φ = 0 and hence f = 0 a.e. because φ is non-zero a.e. Further, if m(E) = 0 then the
range of Tφ is dense in L2 (a, b) because if g ∈ L2 (a, b) is orthogonal to the range of Tφ then
Z b
0 = hTφ f, gi = f φḡ, ∀f ∈ L2 (a, b)
a

and hence φḡ = 0 which implies g = 0 because φ 6= 0 almost everywhere. So we have shown
that 0 ∈ σc (Tφ ).

78
5.5.2 The Spectral Theorem

We need to define some new objects to state the spectral theorem for self-adjoint operators.
Consider a function f : M → C where M is an arbitrary set. To develop a theory for the
integral of f , we can attempt to imitate what we did for the Lebesgue integral. We would
need the integral of the real and imaginary parts of f which would require computing the
integral of a non-negative function. As in the Lebesgue measure case, given a non-negative
f we can construct an increasing sequence of non-negative functions sn : M → [0, ∞) so
that the range of each sn is a finite set (imitation of simple functions) and sn → f . Then
it would be enough to compute the integral of the sn . This would require something like a
Lebesgue measure of subsets of M with properties similar to those of the Lebesgue measure
- the measure of a set must be non-negative and the measure must be countably additive
for disjoint sets. As seen for the Lebesgue measure case, one may not be able to assign a
measure to all subsets of M and still preserve countable additivity - one may be able to do
this only for a special collection of subsets of M which may be considered as an imitation of
the Lebesgue measurable subsets. However, this special subcollection must have some basic
properties if we hope to develop a theory of integration - this collection must be closed under
complements and countable unions.

The discussion above leads to the following definitions.

Definition 5.9. Let M be an arbitrary non-empty set and M a collection of subsets of M .

M is called a σ-algebra if

• ∅ ∈ M,

• M ∈ M,

• if E ∈ M then E c ∈ M,

• if {En }n≥1 is a sequence in M then ∪n En ∈ M.

The subsets of M in M are called the measurable subsets of M .

We now need an analog of the Lebesgue measure.

Definition 5.10. If (M, M) is a σ-algebra, then a measure on M is a map µ : M → [0, ∞]

such that

• µ(∅) = 0,

79
 • if En , n ≥ 1, is a pairwise disjoint sequence of sets in M then
∞
X
µ (∪∞
n=1 En ) = µ(En ).
n=1

The triple (M, M, µ) is called a measure space; if µ(M ) is finite then µ is said to be a finite
measure.

As discussed in the first paragraph, with a measure space (M, M, µ) one can develop
a theory of integration of functions f : M → C analogous to the Lebesgue integral over
R. One would define measurable functions and integrable functions and the proof of the
main theorems for the Lebesgue integral would still hold - we would have the Monotone
Convergence Theorem, Fatou’s Lemma and the Dominated Convergence Theorem. One can
also define L2 (M ) and L∞ (M ).
Definition 5.11. If (M, M, µ) is a measure space then define
Z
2
L (M ) = {f : M → C : f measurable and |f |2 < +∞}
M

with an inner product Z

hf, gi = f ḡ.
M
Also, define
L∞ (M ) = {f : M → C : f bounded and measurable}
with
kf k∞ = inf{C : |f | ≤ C almost everywhere}.

One may show that L2 (M ) is a Hilbert space and L∞ (M ) is a Banach space; further
Holder’s and Minkowski’s inequality hold.

For any φ ∈ L∞ (M ) we may define the multiplication operator Tφ : L2 (M ) → L2 (M )

with
Tφ f = φf, ∀f ∈ L2 (M ).
Proposition 5.10. For any φ ∈ L∞ (M ), Tφ is a bounded linear, normal operator with
kTφ k = kφk∞ ; further Tφ is self-adjoint iff φ is real valued. The spectrum of Tφ is the
essential range of φ and

• σp (Tφ ) consists of all λ ∈ C for which µ({x ∈ M : φ(x) = λ}) > 0,

80
 • σc (Tφ ) = ess range φ \ σp (Tφ ),

• σr (Tφ ) = ∅.

We now state the spectral theorem for self-ajoint operators.

Theorem 5.5 (Spectral Theorem for Self-Adjoint Operators). Let X be a separable Hilbert
space and T a bounded, linear, self-adjoint operator on X. There is a finite measure space
(M, M, µ), a real-valued φ ∈ L∞ (M ), and a unitary operator U : X → L2 (M ) such that
U T U −1 = Tφ .

81