California State University, San Bernardino

CSUSB ScholarWorks

Theses Digitization Project John M. Pfau Library

2008

Operations in Hilbert space

Dennis Michael Gumaer

Follow this and additional works at: https://scholarworks.lib.csusb.edu/etd-project

Part of the Mathematics Commons

Recommended Citation
Gumaer, Dennis Michael, "Operations in Hilbert space" (2008). Theses Digitization Project. 4400.
https://scholarworks.lib.csusb.edu/etd-project/4400

This Thesis is brought to you for free and open access by the John M. Pfau Library at CSUSB ScholarWorks. It has
been accepted for inclusion in Theses Digitization Project by an authorized administrator of CSUSB ScholarWorks.
For more information, please contact [email protected].
 Operators in Hilbert Space

A Thesis

Presented to the

Faculty of

California State University,

San Bernardino

In Partial Fulfillment

of the Requirements for the Degree

Master of Arts

in

Mathematics

by

Dennis Michael Gumaer

September 2008
 Operators in Hilbert Space

A Thesis

Presented to the

Faculty of

California State University,

San Bernardino

by

Dennis Michael Gumaer

September 2008

Approved by:

uichiro Kakihara, Committee Chair Date

Dr. Charles Stanton, Committee Member

Dr. Belisari entura, Committee Member

_______________

Dr. Peter Williams, Chair, Joseph Chave^x-^

Department of Mathematics Graduate Coogffmator,
Department of Mathematics
 Abstract

This review of some of the major topics in elementary Hilbert space theory starts with
a study of basic operations on a Hilbert space. The theory of operators is developed by
providing details regarding several types of operators, in particular compact operators.
Compact operators on a Hilbert space are operators which map the closed unit
ball to a set whose closure is compact. This study of compact operators is the start of the
refinement of bounded linear operators to those which are also members of the Schatten
p-class operators - the final goal of this discussion.
 iv

Acknowledgements

First, I would like to extend my many thanks to my advisor and committee

chair, Dr. Yuichiro Kakihara. I appreciate the considerable time and effort you spent
commenting on the multiple versions of this text. Your explanations of the finer points
of the theory were instrumental to my understanding of this topic and my growth as a
mathematician.
I would like to thank the Department of Mathematics, and specifically Dr.
Joseph Chavez. The flexibility you demonstrated allowed the time so this thesis could be
properly completed. I would also thank Dr. Charles Stanton and Dr. Belisario Ventura
for the detailed reading and suggestions you both put forth.
I offer my unending gratitude to my family. To my parents for your love and
support; you instilled in me the belief I could do anything I set my mind to. To my
mother-in-law and father-in-law for the bets that kept me on track. I thank you all for
helping with the kids after a long day of work.
Most importantly, I credit my loving wife Danaca. Thank you for setting your
expectations high and helping me meet them. You’re the only one who spent as much
time on this thesis as I did. Every moment I spent writing you spent taking care of our
family. I dedicate this thesis to you.
 V

Table of Contents

Abstract iii

Acknowledgements iv

1 Introduction 1

2 The Basics 4
2.1 Norm and Inner Product...................................................................... 5
2.2 Orthogonality..................................................................................... 6
2.3 Applications........................................................................................ 7
3 Introduction to Operators 10

4 Real Versus Complex Hilbert Space 13

4.1 Properties of Real Hilbert Spaces......................................................... 13
4.2 Properties of Complex Hilbert Spaces ................................................. 14

5 Unitary Operators 17

6 Compact Operators 20

7 The Spectrum 25

8 Self-Adjoint Operators 28

9 The Polar Decomposition Theorem 31

9.1 The Schur Representation Theorem . .................................................... 31
9.2 Square Roots................................................... 32
9.3 Polar Decomposition........................................................................... 33

10 The Schmidt Representation 35

11 The Schatten Class Operators 38

12 Trace Class and Hilbert-Schmidt Operators 41

 vi

13 Conclusion 46

Appendix A Banach Space 47

Bibliography 50
 1

Chapter 1

Introduction

Deeper levels of abstraction arise when there is a general understanding of un­

derlying theory. Hilbert Space is one of those abstractions. When R2, R3, and C became
less and less mysterious, mathematicians looked to further dimensions. David Hilbert
worked with the extreme case and asked questions regarding spaces with arbitrary di­
mension, including infinite dimensions. There are some basics that would be desired in
this arbitrary dimensional space to help coincide with R2, R3, and C. A portion of the
thesis will deal with these desired properties and their consequences.
Hilbert Space H is a complete normed linear space with an inner product. First
off, Hilbert Space is a complete space. In any normed linear space X, if every Cauchy
sequence approaches a limit that is contained in X, then that space is said to be complete.
So sequences which have a limit can actually attain that limit within the given space.
Completeness is an essential component of the definition of Hilbert Space. It brings
structure to the space allowing further revelations, not to mention keeping properties of
Hilbert Space in line with those of R2, R3, and C. This property will be used throughout
the thesis. However, since it is part of the definition, it is usually not specifically pointed
out. This is not to belittle its importance.
The primary example of a Hilbert Space for the purpose of the thesis is a linear
space called The space is the space of all square summable sequences (an) of
complex numbers.
Ki |2 < Too >.
n=l j

Properties of this space will follow shortly.

 2

The norm is the arbitrary dimensional analogue of the absolute value. In higher
dimensions it is denoted ||-||. There are many ways for the norm to be defined. In the
case of the space the norm is defined as
/ o° \ I

||(a»)II = I 52 W2 ) ’ V sequences (an) 6 €2.

This norm defines a metric on the space allowing further structure as it narrows down
the spaces which can be called a Hilbert Space.
An inner product is the arbitrary dimensional analogue of the dot product. (The
convention used in my primary research resource, Hilbert Space: Compact Operators and
the Trace Theorem by J.R. Retherford[Ret93], is that inner products are denoted by (•, •).
This convention will be followed throughout the thesis.) Many different inner products
are possible. However each must satisfy a list of properties - one property of particular
importance will follow shortly. For our primary example ^2 of the thesis, the inner product
is defined over as
\ 00 _
V(an), (bn) G ^2) f (®n)j (^n) 1 = 5 >
' 71=1

with the bar denoting complex conjugation.

Hilbert Space is called a complete normed linear space with an inner product
because it has a special' connection between its norm and the given inner product. The
special connection can be shown to be

(rc,j;)2 = ||a;||, Vx G H.

(Note that £2 clearly meets this criteria.)

Initially, many of the problems discussed in the thesis will use the relationship
between norms and inner products. This work pervades all later work and is necessary
as complexity progresses. Several detailed examples will be used to allow the reader
sufficient background for this later work.
The remainder of the thesis will deal with operators on Hilbert Space. Most, if
not all, of the operators to be studied are called bounded operators. A linear transfor­
mation T is called a bounded operator if

3M G R+ such that Va: G ||Ta;|| < M ||a;||.

 3

Among bounded operators, there are many further designations: unitary, normal, self-
adjoint, compact, Hilbert-Schmidt, Trace and Schatten class operators. Each of these
operators have specific properties and uses and each will be touched on in the remainder
of thesis. Particular attention will be focused on the latter portion of the list.
As an example of what will be covered regarding these operators, a brief look
into trace operators is warranted. The linear transformation A from to Kn can be
described by an n x n matrix (%•). The trace of that n x n matrix is the sum of the
diagonal entries, or
n
Tr(A) = y an.
i=l

This trace is something that is useful for Hilbert Space as well. Yet going from an infinite
dimensional space to another infinite dimensional space would not admit a traditional
matrix to sum the diagonal terms. So, as with the abstraction of norms and inner
products, an abstraction of the trace is necessary to glean the desired information.
The means of the abstraction of operators with finite trace will make judicious
use of the Schmidt Decomposition Theorem. That is, for all x 6 H a compact operator
T can be written as oo
Tx = ^Jan(x,xn)yn
n=l

for some orthonormal sets (xn) and where an is an eigenvalue of [T,*T] 1/2. This
statement ties together a significant portion of the work that will be covered toward
the end of this thesis. The Schmidt Decomposition Theorem gives a form for the infi­
nite dimension by infinite dimension matrix that would be impossible to describe in the
traditional matrix form.
The formulation and properties of the operators listed above are the goal of the
text. Each operator will be defined and put into context. Then examples using each
operator will be explored. While practical applications abound, that is not the intent of
the thesis and none will be offered. The intent of the text is to familiarize the reader with
the theoretical background necessary for such applications.
 4

Chapter 2

The Basics

Planar geometry was the origin of ideas which mathematicians later extended
from R2 to arbitrary dimensional spaces. The familiar absolute value was generalized to
the norm denoted || • || which gives the length of the vector in the defined space. The dot
product used to find angles between vectors in R2 or in R3 became the more general inner
product denoted (•, ■) in arbitrary complex linear spaces. The concept of perpendicular
lines was extended via one of its planar properties. In higher dimensions orthogonal lines
are defined by the inner product.
In the Euclidean Space Rn, repeated use of the Pythagorean Theorem gives the
formula for the 2-norm:
n i
Vo: = (o?i, rr2, z3,..., xn) € Rn> ||a:||2 = ( ^2 ) *'
i=l

But the norm could be defined in a different way such as

n i
Vz= (x1,X2,x3,...,xn) 6 X, Ms = (52ki|3)3,
V i=l

or even
Vx = (x^x2,x3,... ,xn) e X, ll^lloo = marr(|a;i|,..., |zn|).

These last two norms are obviously different from the usual norm in R2, but these norms
are not coming from the usual R2.
2.1 Norm and Inner Product

Definition 2.1. Let X be a complex linear space. A norm is a function from X into the
non-negative reals satisfying:

1. ||z|[ = 0 iff x ~ 0;

2. for each x 6 X and each scalar a 6 C, ||air|| = |a|||z||;

3. the triangle inequality ||a; + y\\ < ||e|| + ||y|| holds for all x, y G X.

(X, || • ||) or X with a norm || • || is called a normed space.

Definition 2.2. Let (xn) be a sequence in a normed space X. Then {xn} is called a
Cauchy sequence if

Vc > 0 H TV € 9 m,n > N => — a5m|| < e.

Definition 2.3. If every Cauchy sequence (xn) in a normed space X converges in X then
X is called a complete space.

Definition 2.4. Let X be a linear space over C. An inner product, (•, •), is a function
defined on X x X, the set of all pairs of elements in X, satisfying:

1- (#, z) > 0 with equality iff x = 0;

2. (x, y) = (y, x) where the bar denotes the complex conjugate;

3. VajG'C and Vx^x^y € X; (axi + 0xify) = a(a;i,y) + j0(a;2,y).

(X, (-,-)) or X with an inner product (•, •) is called an inner product space.

A complete normed linear space is called a Banach space, named after the Polish
mathematician Stefan Banach. Hilbert Space - the focus of this paper - is a specific type
of those spaces. Hilbert Space is a Banach Space with the added restriction that it has an
inner product derived from the norm of the space, more specifically Vac G H ||rc||2 = (x, x).
An important and recurring example of Hilbert Space is €2. A complex sequence
(an) is an element of ^2 provided the sum of the magnitude squared of each term is finite.
Mathematically
 6

When a space is said to be a Hilbert Space, it must be complete and a norm

must be given. The space I2 is well known to be complete. (A proof of a stronger result is
detailed in Proposition A.3.) For (an), (bn) e £2 the norm and the induced inner product
are given respectively by
/ 00 \ 2 00
II(pn)II = I y? Ian|2 I J (&n)) = an&n
*
\ n=l / n=l

2.2 Orthogonality

The idea of perpendicular lines is something that mathematicians would like

to have extended to other spaces even when there may not be a ninety degree angle to
visualize. It so happens that there is a way to transfer the usefulness of perpendicular
lines to any space with an inner product. The usual inner product from R2 (the dot
product), on two perpendicular lines is zero. It is this property of perpendicular lines
that will persist regardless of the inner product space. But the word perpendicular is
usually reserved for R2, C and other spaces where actual ninety degree angles exist. So
when dealing with spaces where there aren’t angles to measure, the word orthogonal is
used. Two different objects in an inner product space are called orthogonal when their
inner product is equal to zero.
In £2 there is an orthogonal basis with unit length, or orthonormal basis, denoted
(en). This sequence has a zero as each of its terms except for the nth term which is a one.
Mathematically,
e< = (0,0,1,...) for i = 1,2,...
'----------- v------------ '
ittl
one in the position

It should be noted that (e,, e-f) = and |[e»|| = 1.

One common equality in Hilbert Space theory is Parseval’s Equality, named
after the French mathematician Antoine Parseval. It is offered here, without proof, in a
form slightly modified from the general equality for our uses.
Theorem 2.5. (Parseval’s Equality). For a complete orthonormal set (yn) of Hilbert
Space £2 and Vx € £2 it holds that
00

n=l
 7

While this theorem holds for every complete orthonormal set, it is mostly com­
monly used with the canonical orthonormal basis et for £2-

2.3 Applications

After several definitions it is generally a good idea to give some properties and
applications using those definitions. These properties and applications follow.

Proposition 2.6. In an inner product space X, (re, u) = (a?, v) for all x implies that
u — v.

Proof. Observe that,

(x,u) = (x,v) => (x,u) — (x,v) = Q

=> (x,u — v) = 0.

This has to be true for all x, so in particular for x = u — v. This implies (u — v, u — v) = 0.

By Def. 2.4 part 1, u — v = 0 and u = v. □

In Hilbert Space a sequence can converge strongly or weakly. As the names im­
ply, a strongly convergent sequence meets stricter requirements than a weakly convergent
sequence. So a strongly convergent sequence is automatically weakly convergent. This
leads to the question, when does weak convergence imply strong convergence? But first,
the definitions of these terms are given.

Definition 2.7. For Hilbert Space H, a sequence (xn) C H is said to converge strongly
to the point x 6 H if ||a;n — m|| —> 0 as n oo. This type of convergence is also known
as norm convergence or convergence in the norm.

Definition 2.8. For Hilbert Space If, a sequence (a?n) C H is said to converge weakly to
x 6 H if, for all y 6 H, (xn, y) —> (z, y) as n —► oo.

Under an additional condition weak convergence implies strong convergence as

shown below.

Proposition 2.9. In Hilbert Space H let a sequence (xn) C H converge weakly to x e H.

If the sequence (||a;n||) converges to ||x|| then (xn) converges strongly to x.
 8

Proof. Start by using a property of norms

lim ||®n - as|| = lim \/(xn — xfxn — x)

n—*oo n—foo
= lim J(®n, Xn) ~ (®n, x) - (xn, x) + (x, x)
n — ’’
*eoO
ti—*O

= lim y||®n||2 - (xn,x) - (xn,x) + ||z||2

n—>oo ’

= o.

Thus as n approaches infinity ||®n — z|| = 0 as desired. □

Proposition 2.10. For x,y in Hilbert Space H,xly iffVa € C, ||rE + ay|| > ||n?||.

Proof. Assume xly. Then Va e C,

(x, x) = (x, x) + a(x, y) + a(y, x)

< (®, x) + (x, ay) + (ay, x) + aa(y, y)
= (x + ay,x + ay)

So [[sell2 < ||® + ay||2 and ||z|| < ||® + ay\\.

Assume Va 6 C, ||®|| < ||® + ay||. Then,

0 < ||® + ay||2 — ||®||2

= (x + ay,x + ay) - ||®||2
= a(y,®)+ a(®,y)+ aa||y||2>

leaving
0 < a(y, x) + a(x, y) + aa(y, y). (2.1)

Let a = i with n > 0. Substituting a into (2.1) above leaves part of the
inequality needed. So
111
0 < ~(y,x) + -(®,y) + -^v)
71 71 7Z
and
< ±(y,z) + ±(x,y). (2-2)
 9

Now let a = — £ with n > 0. Substituting a into (2.1) above solves for the remainder of
the required inequality.

o< + -^(^y) +
n n nr
111
(2-3)

Combining (2.2) and (2.3) above and multiplying through by n yields

~~(y,y) < (y,a) + (z,y) < -(y,y).

Tv 71*

Now take the limit as n approaches infinity, so that

1 1
Jim [- -(y,y) < (y^) + < -(y5y)]-
n—*oo n n
Hence, 0 < (y,rtr)+(z, y) < 0 so 0 = (y,x)+(x,y). Then letting (x, y) — p+iq for G R,
it also follows that (y, x) — So 0 = (y, a)4-(a;,y) => 0 = (/?+z7) + (/?-27) => 0 = 0.
Once it is shown that 7 is also zero the proof will be complete.

Return to (2.1) and solve when ot = and ct = simultaneously. This leads

to a similar inequality

— (y,y) < ifayj-ifax) < ~(y,y).

71 *
7c

Again take the limit as n approaches zero to get

0 < i(x,y) <0

and
2(^y)-^(y^) = 0-

Since (re, y) has been defined as 5 + 27 the equation turns into

i(/3 + iy) — i(j3 — 27) = 0=>2/? — 7 — — 7 = 0=>7 = 0.

□
 10

Chapter 3

Introduction to Operators

Hilbert Space has been carefully created to mimic Euclidean Space. As a result,
research in the area doesn’t need to focus on the space itself. Interest is usually concen­
trated on the functions - or more precisely operators - that act on the space and the
consequences of that action.
Functions are defined as a mapping from one set to another. In general terms,
a function is simply a set of rules which describes how elements in one set are mapped
to elements in another set. Restrictions are usually placed on these functions depending
on the focus of study. For the purpose of Hilbert Space, the functions are required to be
linear. In any linear space, function f is linear when for a?, y in the domain and scalars
a, b it holds that
f(ax + by) = af(x) + bf(y).
Linear functions from Hilbert Space H to Hilbert Space K are called linear operators.

Definition 3.1. In Hilbert Spaces H and K an operator T : H —> K is called a bounded

operator if
3M € R+ such that ¥x € H, ||Tz|| < M ||o:||.

The set of all linear bounded operators T on Hilbert Space H to K is denoted

L(H,K). That is,

L(H, K) — {T : H —* K | T is linear and bounded}.

Operators are treated as objects in their own right. As an object, it is necessary

to know what the norm of that object is.
 11

Definition 3.2. For T € L(H), the norm, ofT is defined by

||T|| =sup{||T2:|| | M<1}-

It is routine to verify that || • || is a norm on L(H). Also, it should be noted

that if T L(H) the above definition does not define a norm on a Hilbert Space. It is
boundedness that forces ||Tz|| to be finite for all x G H with M < 1.
Another common property of operators defined on a space is that of continuity.
The following definition focuses on Hilbert Space operators, but it could easily be modified
to work in a general Banach Space.

Definition 3.3. An operator T defined on Hilbert Space H is continuous when every

sequence (xn) which converges strongly to x implies that the sequence (Txn) converges
strongly to Tx.

Theorem 3.4. For Hilbert Spaces H, K and linear operator T from H to K the following
are equivalent:

1. T is bounded,

2. T is continuous at 0,

3. T is continuous over H.

Theorem 3.4 brings together two notable properties of operators, boundedness

and continuity. An operator being continuous is an important property. However, to
minimize redundancy, continuity is not usually listed as a property of an operator. The
phrase bounded operator is sufficient.
A few quick examples will help to solidify this theory.

Example 3.5. For the following examples, let the Hilbert Space H = £2 and let the
sequence (e^) be a countable orthonormal basis in £2.

1. Let Tei = ef+i for i > 1. This operator is known as a shift operator. Then ||7j| = 1.
oo oo
Proof Let x = which implies Tx = 2 (z> ei)e»+i-
2=1 2=1
 12

Parseval’s equation (Theorem 2.5) states that

oo

1=1

.or ||Tx|| = M. Thus ||r|| = 1. □

2. Let Tei = + ef+i. Then ||T|| = 2.

Proof. It is clear that ||ei|| = 1 and ||Tei|| = ||ef + e^i|| — \/2 Vi. For x 6 h,
00 oo
x = 52^’ ei)ei 911(1 Tx = 52ef)(ei + ei+x)-

i=l i=l

Hence

IM2 = (ru, ei)ei + e<) + (x, e^-i)]e»

i=2
oo

k®, ei)i2+52 i(x>ei)+ (x> ei-i)i2

i=2
oo
< [ (as, 6j) I2 + 2 52 [| (□:> e«) |2 + | (□;, e£-i)|2]
i=2

< 4|rf

Thus II^ZYeII < 2||ic|| and ||T|| < 2. Now it must be shown that ||T|| —> 2. Let
fn = ^(ei + e2 + ” -en). Then, ||/n|| = 1 and Tfn = ^(ei + 2e2H---- 2en4-en+i).
So ||T/n|| = ^jj^/l + 4(n-l) + l = y/4 - 2/n. Hence, ||T/n|| -> 2 as n -> oo.
Since the norm is the supremum of all possible values, ||T|| = 2. □

* of Hilbert Space H is the set of all linear bounded

Definition 3.6. The dual space H
operators, in this case linear functionals, T such that T maps elements of H to the scalar
field C.

* of H is identified with H itself as seen from the following.

The dual space H

Theorem 3.7. (Riesz Representation Theorem). Let H be a Hilbert Space.

A) Let y e H. Define fy(x) = (x,y). Then fy € 77

* and = \\y\\.

*.
B) Let f € H Then there is a unique y e H such that f(x) = (a:,y) Vie G H.
Moreover, ||/|| = ||y||.
 13

Chapter 4

Real Versus Complex Hilbert

Space

Recall that the inner product has, as part of its definition, a property (x,y) =
(y, x). Calculations concerning inner products of complex terms can quickly become more
involved than the same calculation with only real terms. However, as it will be shown,
the real case may be too simple and important information can be lost. A few examples
are in order.

4.1 Properties of Real Hilbert Spaces

In a real Hilbert Space 77, (ir,y) = (y,rr) can be written as (a?,y) = (y,x)- This
leads to simplifications that would be impossible if H is a complex Hilbert Space.

Proposition 4.1. In a real inner product space H, if x,y G H and ||x|| = ||y|| then
||aa? + by\\ = ||&xc 4- ay|| Va, b G R.

Proof. Observe that

||aa: 4- by\\2 = (ax + by, ax + by)

= (ax 4- by, ax) + (ax + by, by)
= (ax, ax) 4- (by, ax) 4- (ax, by) 4- (by, by)
= a2||z||2 + ba(y, x) 4- ab(x, y) 4- &2||y||2 (given ||x|| = ||y||.)
 14

= a2||z/1|2 + ab(y,x) + ba(x,y) + fc2M2

= (ay, ay) + (ay, bx) + (bx, ay) 4- (bx, bx)
(taking this back to a single inner product)
= (ay 4- bx, ay + bx)
= ||ay 4-6a?||2.

Finally, taking the square root of both sides leaves ||asc 4- by\\ = ||6a; + ay||. □

The definition of an isosceles triangle is well known. Plainly, it’s a triangle that
has two sides of equal length. This idea is generalized to a real Hilbert Space in the
following theorem.
Theorem 4.2. (The Isosceles Triangle Property) Let H be a real Hilbert Space. If
x,y,z 6 H, x + y 4- z = 0 and ||a?|| = ||y||, then ||a? — z|| = ||y — z||.
Proof It holds that

||a; — ^||2 = ||2a? 4- y||2 (since x + y 4- z = 0)

= (2x + y,2x + y)
= (2x, 2x) + (2a?, y) + (y, 2x) + (y, y)
= 4(®, x) + (2x, y) 4- (y, 2x) + (y, y)
= 4IMI2 + 2<X y) + 2fe> x) + flyII2 (since IMI = IIH)
= 4||y||2 + (x,2y) 4- (2y,x) 4- ||a:||2
= 4(jf, y) + (z, 2y) + (2y, x) + (x, x)
=' (2y, 2y) + (x, 2y) + (2y, x) 4- (x, x)
= (2y 4- x, 2y 4- a?)
= I|2y4-rc||2
= II?/ - ^ll2 (since x 4- y = ~z).

Then taking the square root of both sides gives the desired equality. □

4.2 Properties of Complex Hilbert Spaces

Having (a;, y) = (y, x) can obviously be beneficial in some cases. But sometimes
this property can allow terms to cancel even when they are necessary for the completion
 15

of a proof.

Proposition 4.3. For a real Hilbert Space H, x,y G H and ||z|| = ||y|| imply that
(x + y,x — y) = 0. If H is complex, this assertion is not true.

Proof. 1. For a real Hilbert Space H,

(x + y,x-y) = (x, x - y) + (y, x - y)

= (x, x) - (x, y) + (y, x) - (y, y)

= M2-(s, !/) + (*> 2/)- 1MI2

= IMI2 - M2 - y) + v) = 0-
2. For the complex case, a counterexample will suffice.
Let H be C and let x = + i^- and y = 1. Clearly ||ac]| = ||y||. But

/T2 + 2 , lX/2 ^-2 , .x/2

(x + y, x - y) = —2— + —2— +

Taking the inner product

(V2 + 2
= — yj'ii 7^ 0.
I 2
So for the x and y given above ||z|| = ||y|| but (x + y, x — y) 0. □
Proposition 4.4. In complex Hilbert Space H, T G L(H) and (Tx, x) = 0 Vac G H
implies T = 0. This is false if H is real.

Proof. 1. If H is complex:
(T(ac + y),x + y) = 0 by assumption. Now,

(T(ac + y),ac+ y) = (Tx + Ty, x + y)

= (Tx, x) + (Tx, y) + (Ty, x) + (Ty, y)
(given the ls£ and 4th terms are both zero.)
= (Tx,y) + (Ty,x').

So 0 = (T(x + y), x + y) = (Tx,y) + (Ty,x)

Similarly with (T(ix + y),ix + y) = 0 we have

(T(ix + y), ix + y) = (Tix + Ty, ix + y)

 16

= (Tix, ix) + (Tix, y) + (Ty, ix) -f (Tp, y)

(given the 1st and 4f/l terms are both zero.)
= (Tix,y) + (Ty,ix)
= i(Tx,y) + i(Ty,x)
= i[(Tx,y)-(Ty,x)].

Once again 0 = (T(ix + y),ix + y) = (Tx, y) — (Ty, x).

This gives two equations and two unknowns. Namely,

(Tx, y) + (Ty, x) = 0
(Tx, y) - (Ty, x) = 0

So (Tx, y) = (Ty, x) = 0 Yx, y G H, And therefore T = 0.

2. If H is real:
/ 0 -1
Here is a counterexample. Let T =
U 0
Then
-1 W
0 J ^2

Set y = I X2 j. Does (Tx, x) = 0, Vz 6 R2 imply T = 0? Well, (Tx, x) = (y, x)

\ X1 /
and the inner product in R2 is the dot product. So (y, x) with x and y defined as
above is (—&2)(®1) + (^2)(zi) = —^2^1 T ^2^1 = 0. Thus B T e L(H) such that
(Tx,x) = 0 Yx E R2 but T f 0. □
 17

Chapter 5

Unitary Operators

The next notion to be covered is the extremely useful adjoint of operators. The
adjoint of an operator is actually a direct result of the Riesz Representation Theorem.

Definition 5.1. Let T G L(H) and let y G H. Define S G H

* such that

S(x) = (Tx,y).

The Riesz Representation Theorem (3.7) states Bz G H such that \/x S(x) ~
(x,z). So (Tx,y) = (x,z). Then there is an operator that maps y onto z. This operator
* and is called the adjoint of T. T
is T * is a bounded linear operator and ||T
*|| = ||T||.
A unitary operator is one which is defined by special properties involving its
adjoint. The definition of this operator follows.

Definition 5.2. Unitary Operators. Let T G L(H). If it is given that T-1 G L(H) and
* = T_1 then T is said to be unitary.
if T

Before a theorem regarding unitary operators, a brief lemma is in order. Let

0 be a Hermitian bilinear form on H x H. That is, </> is linear in the first term and
conjugate linear in the second. Let ip(x) = j>(x,x). Note that is called a quadradic
form associated with (/>.

Lemma 5.3. Given </> and as defined above, it follows that

+ vif
2 2 2 2
 18

Proof. Observe that

= l^x + y,x + y)
1
= + ^y) F^(y,x) + <Xy,y)],

*(M -

= ^>(x-y,x-y)

= i [<£ (a;, a?) — <f(x,y) - ^(y.x) -}- t/)(y,y)].

Combining the first two terms consolidates to

Similar calculations concerning the second pair of terms shows

Lastly, combining the two intermediate results leaves

as desired. □

It is important to note that the inner product is also a Hermitian form, so the
above lemma applies to the following theorem.

Theorem 5.4. T G L(H) is unitary iffT is an isometric isomorphism.

* — T^1. L(H) is defined to be

Proof. Given T is unitary, 3 T_1 6 L(H) such that T
the set of continuous operators so T and T-1 are continuous. The condition T
* = T"1
implies that TT~l = T^T and therefore T is one-to-one and onto. So if T is unitary
then T is an isomorphism. Then to show T is an isometry:

IIT^H2 = (Tx,Tx)
 19

*Tx,x)
= (T
= fa^Tx^) since T is unitary
= fax)
= IW2
So if T is unitary then T is an isometry and therefore T is an isometric isomorphism.

Given T is an isometric isomorphism, it follows that Va:, y G H

*Tx,y)
(T (Tx,Ty)
i1 f HTz + Tj/H2 - ||T® - Ty||2 + iHTo: + iTy||2 - i\\Tx - iTytf2
4
(by Lemma 5.3.)
* + y} II2 - ||T(x - y) [|2 + i||T(z + «/) II2 - 2||T(s - «/) I
J [||T( 2

j IIs + yll2 - Ik - yll2 + «lk + ®j/II2 - «lk - »y||2

So (T
*Ta?,y) = fay) Va:,y G H => T
* * = T”1 => T is unitary.
= I => T □
 20

Chapter 6

Compact Operators

In analysis, a significant amount of time is spent working on compactness of sets

and spaces. This is for several good reasons. In operator theory, it would be nice to have
operators that take bounded sets to compact sets. The following definition formalizes
this idea.

Definition 6.1. Given the unit ball A = {x G H : ||®|| < 1} a compact operator is
an operator T G L(H) such that T(A) is compact. The set of all compact operators is
denoted K(H).

When it is given that an operator is compact, it would be convenient to know

if related operators are also compact.

Proposition 6.2. Let H be a Hilbert space.

1. T G K(H) <=> T
* G K(H);

2. IfTe K(H),S G L(H) then both TS and ST are elements of K(H);

3. TeK(H) rTeK(H);

*
4. T G K(H) <=> TT G K(H);

5. T G K(H) and T-1 G L(H) imply that H is finite dimensional;

6. IfT G L(H) has finite rank, then T G K(H).

 21

In R2 and C there is a well known result that states every bounded sequence
has a convergent subsequence. For Hilbert space this isn’t the case. However, given a
few more restrictions a similar result can be obtained.
First, a quick definition. A Hilbert space is said to be separable if it contains a
countable, dense subset.

Theorem 6.3. Every bounded sequence (xn) in a separable Hilbert Space H has a weakly
convergent subsequence.

Proof. Space H is separable so G H - a countable dense set which is treated as a

sequence. Define
A= {(a?i,yi), (®2> yih Cc3, yi), • • •}•
A is a sequence of complex numbers, as consequently contains a convergent subsequence
Ai. That is, there is a subsequence (a?i,fc) of (xn) such that the subsequence Ai =
{(rEi,fc,yi)} converges. This particular subsequence of (xn) will referred to as the first
subsequence. So the first convergent sequence, which converges to ai, is

Al = {(a?i,x,yi), (a?i,2,yi), C^i^yi), •. •}, &i,k,yi) ai-

After choosing y2 from (y$) there is a second subsequence of (xn) such

that (x2]fc) C (x-if) and

A2 = {(£2,1^2), (%2t2,y2)> (®2,3,2/2), • ■ •}, ~* «2-

Continuing this process - the underlined parts would be explained later - gives
the following sequences:
A = {(^1,1,2/1), (a?i,2,yi)> (®i,3»2/1), •• •}, (®i,fc,yi)^«i

A2 = {(x2)l,y2), (x2,2,y2), (x2,3,y2)} ‘ (X2,k,y2) G2

A3 = {(3:3,1,2/3), (e3>2, 2ft), (a?3,3, y3),• • •}, (^,2/3) ^«3

A„i = y?n)j 2/m)) (^m,3> 2/m), • • ’ , 2/m), • • •}, (x ym) *

It also follows that (xn) □ (%i,k) 2 (%2,k) 2 • • • 2 (xm,m) 2 • • •

 22

A new sequence is created using the underlined terms in the above sequences:

B = {(^i,b 2/1) > (%2,2, 2/2), (^3,3> 2/3)> • • 2/m)i • •

And Vi, as m approaches infinity —* aj. This is a direct result of each'sequence

being a subsequence of the preceeding sequence.
As a matter of notation xm>m will be written as zm.
Let M be a bound for (xn). Let w 6 H and e > 0 be arbitrary. Then 3yi
such that ||w — j/i|| < 3^. Since ((zm,yi)) is a Cauchy sequence, 3AT > 1 such that
|(z£ — £j,2/i)l < > TV. Thus we have that

|(^,w) - (^,w)| = Ito,™) - (z^yi) + (^,y£) - + (zj,2/i) - (*

7, w)|

< I(^,w) - (Z£,2/01 + |(Z£,yi) - (Zj,yi)\ + |(zj,y$ - (

j,w)|
*

= l(z£,w- 2/f)| 4- l(z£ - Zj,yi)l + l(zj,yi - w)|

< IIM ■ Ih - 2/fII + l(^ - »2/i)I + Ill'll ■ 1,2/i - w||

Hence ((zm,w)) is convergent. So (xmtm) is weaky Cauchy and is therefore weakly con­
vergent. □

Riesz worked with operators which he called completely continuous. According

to Riesz, an operator T G L(H) is completely continuous if T maps weakly convergent
sequences to norm convergent sequences-these terms are defined in Definitions 2.7 and 2.8.
While the phrase completely continuous has fallen out of favor, the work that Riesz did
with this type of operator is worthwhile. It turns out that the operator called completely
continuous coincides perfectly with the compact operator defined in this chapter. While
the definitions of these operators are different, they are logically equivalent. This, of
course, must be shown. But first, a brief lemma.

Lemma 6.4. For weakly convergent (xn) G H, (xn) is bounded.

Proposition 6.5. Let H be a separable Hilbert space. T G K(H) iff T is completely

continuous in the Riesz sense.
 23

Proof. Let T G K(7?). If T were not completely continuous then there exists a weakly
convergent sequence (xn) that converges weakly to x such that

lim ||Txn — Tx|| 0 0.

n—*oo

Then there is an e > 0 and a subsequence (Txnf) of (Txn) such that

||Txnii — Tx|| > c, Vn > 1.

Using the weak convergence of (xn) and Lemma 6.4, (xn) has a bound. Thus (Txnf)
has a convergent subsequence (Txn,2)- Define m such that (Txnf) converges to m. It is
given that (Txnf) converges weakly to Tx so m = Tx. So ||Txni2 — Tx|| approaches 0, a
contradiction. T is therefore completely continuous.
For the completely continuous operator T to also be compact, it must map the
unit ball {x^H: M < 1} into a set whose closure is compact.
Let A = {x e H : ||x|| < 1} and let T(A) = B. If it can be shown that every
sequence in B has a subsequence converging to some member of B, then B is relatively
compact. So T would then map the unit ball to a relatively compact set and T G K(H)
as desired.
Let {i/fc} C B. There exists a sequence {x^} C A such that Txk = Yk. Note
that {xfc} is bounded because ||x|| < 1. So by Theorem 6.3 3{zn) C {xjJ such that
is weakly convergent. Also (Tzn) is norm convergent since T is completely continuous.
So B is relatively compact. Therefore T is compact. □

Compact operators are one of the main focuses throughout the rest of the text
and are worthy of further investigation.
While this next proposition is instructive, it serves another purpose as well. It
offers insight into a topic that will be covered in the next chapter. Keep this proposition
in mind when reading Chapter 7.

Proposition 6.6. Let T G K(H). If X f 0 then ker(A7 — T) has finite dimension.

Proof. Assume ker(A7 — T) has infinite dimension. Then 3{xn} C ker(Al — T) such
that {xn} is an infinite orthonormal set. (Otherwise ker(A7 — T) would have a finite
orthonormal set which would serve as a finite basis for the space. For the moment, this
is assumed not to be the case.)
 24

Since T is compact and {xn} is weakly convergent to 0 there exists a convergent

subsequence of {acn} = {^Ta:n}. But \/Xi,Xj G {£n},z / j ll^i — 3j|| = x/2. So there
is no such convergent subsequence of {rrn}. A contradiction. Therefore ker(AZ — T) has
finite dimension. □

Theorem 6.7. An operator T G X(H) iff there are finite rank operators An such that

lim ||T — An|| = 0.

n—»oo

An outline of the proof follows.

Assume there is such an A„. Choose a bounded sequence (xn) and use the
compactness of An to find a subsequence (a?n,fc) of (%n) such that (An(xnm)) converges.
This sequence is Cauchy. This is used to show that (T(xnm)) is Cauchy and therefore T
is compact.
Then assuming T G let (u{) be a complete orthonormal set. The inner
product expansion of T and of the orthogonal projection Pn would be
co n
Tx — y^fTx,Uj)uj and Pnx— ^2(x,Ui)ui
i=l i=l

respectively. It can be shown that the required An is given by the composition An = PnT.
Thus

||T-An|| = ||T-P„T||
= sup||(T-P„TM
IM<1
co
sup 22 (Tx,Ui)ui
IM<1 i=n+l

Since ||a;|| < 1, the closure of Tx is compact. The above sum tends to zero for large
enough n when Tx is in a compact set. So limn_>oo ||T — An|| = 0. □
 25

Chapter 7

The Spectrum

It is a straightforward calculation to find the eigenvalues of a transformation

defined by a matrix. This procedure is taught in an elementary linear algebra course.
Unfortunately, this procedure fails when the transformation is an operator in Hilbert
Space. The obvious difficulty is that there is no matrix to calculate the eigenvalues from,
but the heart of the matter remains. If H is a Hilbert Space, is there a way to find out
when Tx — Ax is not invertible?
The following should seem familiar, as it is the same reasoning that is given
when explaining the algorithm for finding eigenvalues in basic linear algebra. Tx = Ax
implies

Tx — Ax = 0,
(T — AI)x = 0, I being the identity matrix.

This has a non-trivial solution when T—Al is not invertible. It is here that the explanation
for the purposes of this paper diverges from the explanation given in linear algebra. In
linear algebra the procedure concluded with writing down the matrix that corresponded
with T and subtracting Al. Then to find out where T—Al is not invertible, this procedure
dictates setting the determinant of the above difference to zero and solving for A.
In the Hilbert Space version, there is no matrix. Since no calculations can be
done, the generalization stops at (Tx — AZ) being not invertible. In the Hilbert Space
version, the result of this process is called the spectrum.

Definition 7.1. Let T G L(H). The spectrum of T, denoted cr(T), is defined to be

 26

exactly what it should be to match up with the equations above. Namely,

a(T) = {A G C | T — Al is not invertible}.

What of the elements of C that are not in cr(T)?

Definition 7.2. Let T G L(7Z). The resolvent of T, or p(T’), is the set of all complex
numbers not in cr(T). That is,
p(T) = C\<r(T).

The spectrum of arbitrary operators in Hilbert Space can be complicated. Later

work with or(T) will be restricted to operators that have more structure than most ar­
bitrary operators- compact operators being a notable example. This eases some of the
complication. This chapter, however, deals with the spectrum of arbitrary operators.
For an arbitrary operator T, cr(T) can be partitioned into three pairwise disjoint
sets:
cr(T) = <rp(T) U &r(T) U <tc(T).
The properties that distinguish these sets are as follows, for A G cr(T): A G ap(T)~ or the
point spectrum- when Al — T is not one to one; A G crr(T)~ or the residual spectrum-
when AZ — T is one to one but the range of AZ — T is not dense in ZZ; Ac ac(T)-or the
continuous spectrum- when AZ — T is one to one and has a dense range. It is important
to note that the elements of ap(T) are the eigenvalues of T.
An exercise in partitioning the spectrum will be instructive. But first, a theorem
to aid in that exercise. The proof of the theorem leads off topic and is omitted.
Theorem 7.3. The set <j(T) is closed and bounded for T G L(ZZ).

Example 7.4. Let T be an operator in Lfa) such that

oo
Ta; = 22(ar,en+i)en,
n=l

where (en) is the standard basis of C?. Show that if |A| < 1 then A G a(T). Partition the
spectrum.

Solution. First, find the actual eigenvalues of T. Let (AZ — T)(ac) = Ax — Tx = 0, i.e.,
A G <7p. Then
oo oo oo

5 ^n)en Cn-f-i)en 5 [^(®j ®n) (x, Cn+1)]0.

n=l n=l n=l
 27

Since {e^} is linearly independant, we have

X(x, ei) — (x, e2) = 0 => A(s, ei) = (x, e2)

X(x, e2) - (x, e3) = 0 => A2(a;, ei) = (x, e3)

X(x,en) - (a?,en+1) = 0 => Xn(x, er) = (x, en+1),

oo oo
An_1(a?, ei)en = ^(x, en)en = x.
n=l n=l

So oo

X=
n=l

This must be true for all x G l2. When (x, ei) 0, we have 0 < |A| < 1. This also implies
that if |A| > 1 then A G p(T).
Secondly, if A = 0 and Tx = 0 then 0(®, ei) — (a?, e2) = 0 and (a;, 62) = 0. This
leads to x = (sq, 0,0,...) but Tx = 0. So XI — T is not one to one and A = 0 is an
eigenvalue.
Since Co = {z G C | |z| < 1} C cr(T) and cr(T) is closed, we have that Ai G
Ci = {2 G C | |^| = 1} C ct(T). Then Ai ar(T) because A G o-r(T) implies that XI — T
has a dense range. Also Ai £ crp(T) because A G <JP(T) implies that XI — T is one to one.
Thus Ai 6 ac(T) since crp(T), ar(T) and ac(T) are disjoint and their union is cr(T).
This completes the partition. □
 28

Chapter 8

Self-Adjoint Operators

General operators with no restrictions are sometimes difficult to work with. So

much so, the very first topic regarding on operators in a Hilbert Space was the restriction
to bounded and linear operators. The introduction of the adjoint of operators allowed
work with inner products that had previously been impossible. Even further refinements
were given in Chapter 5. Many interesting results arose when the general theory of
operators was narrowed to unitary operators. Another such refinement of the adjoint of
operators will continue this specialization of general Hilbert Space theory.

Definition 8.1. A self-adjoint operator of a Hilbert Space H is an operator T G L(H)

such that T = T
*. In other words,

(Tx, y) = (x,Ty), Vx,y G H.

This definition has some major consequences. Given Hilbert Space H, for self-
adjoint operator T and V® G H, (T®, ®) = (®,T®). (It is important to note that, by
definition, T being self-adjoint implies that T G L(ZI).) But part of the definition of
inner product states that (T®,®) = (®,T®). From these equalities, it is obvious that
(Tx, ®) is real for all ® G H - a significant restriction.
One important property that results from working with strictly real terms is
defined next. This definition revolves around the fact that the real line is well ordered.
It follows that some sort of ordering can be imposed on self-adjoint operators.

Definition 8.2. In Hilbert Space H, a self-adjoint operator T is said to be nonnegative,

 29

denoted T > 0, if
V® 6 H, (Tx, x) > 0.

Similarly, for self-adjoint operators A and B,

A > B iff A — B >0.

There are some basic properties of self-adjoint operators that will prove useful
in later work.

Proposition 8.3. In Hilbert Space H, for operator T 6 B(H) and for self-adjoint oper­
ators A,B,Ce L(H):

1. A > 0 and B>0^A + B>0;

2. A > B and B > A => A = B;

3. A> B and B > C => A> C;

*
4. T *
> 0 andTT > 0.

The first three statements in the above proposition are rather trivial. However,
the last deserves a second look as it will be used frequently in later work.

Proof In Hilbert Space H, let T G L(H). Then V® G H, (Tx,Tx) > 0 hence (T

*Tx,x) >
0. Then by Definition 8.2 T
* *,
> 0. For TT (T,T
the same process with *® ) will
suffice. □

This partial ordering of the self-adjoint operators of L(H) allows some interesting
results with the spectrum from the previous chapter.

Theorem 8.4. IfT is self-adjoint then a(T) has only real elements. Also, ifT>0, then
a(T) C [0,4-oo).

Theorem 8.5. For self-adjoint T and Hilbert Space H, if (Tx, x) = 0 V® G H then

T = 0.

Before the next proposition, a definition is required.

Definition 8.6. Normal Operators. If T

* *
= TT then operator T is said to be normal.
 30

Proposition 8.7. The operator T G L(H) is normal iff ||Tx|| = ||T

*x|| Yx G H.

Proof. Given T is normal. Then we have for x G H that

||Tx||2 = (Tx,Tx)
*Tx,x)
= (T
= (TT
*x,x)
= *x
(T),T
= IIT^II2.

so ||Tx|| = ||7”o:||.
*x||
Conversely when ||Tx|| = ||T Yx G H, it follows that

*Tx,x)
(T = (Tx,Tx)
= l|Tx||2
= l|T^||2
= *x
z),T
(T
= (TT
*x,x),

so (T
*Tx,x) = (TT
*x,x). Working with this equality,

T,x)
*x
(T x,x)-(TT = 0
*x
T,x)
(T x-TT = 0
*)((T
Tx,x)
-TT = 0.

Then let S = T
* *.
— TT S is shown to be self-adjoint since

Theorem 8.5 states that for self-adjoint 5,

YxtH, (Sx,x) = 0 => S = 0.

*
Then T = TT
* and T is normal as desired. □
 31

Chapter 9

The Polar Decomposition

Theorem

The Polar Decomposition Theorem defines a method for finding two operators
whose composition is equivalent to a given operator. This theorem is one of the last
pieces to the puzzle needed to prove the main classes of operators in this thesis.

9.1 The Schur Representation Theorem

There is a technique to systematically find eigenvalues of T when T is a compact,

self-adjoint operator. Theorem 9.1 will be the starting point for finding these eigenvalues.

Theorem 9.1. IfTfQ is both a compact and self-adjoint operator, then either [|T|| or
— ||T|| is an eigenvalue ofT.

Given this eigenvalue, an algorithm finding eigenvalues of an operator can be

described. This procedure is given on page 21 of R. Schatten’s book “Norm Ideals of
Completely Continuous Operators.” [Sch60] The proof provided is quite laborious and
only slightly instructive. An outline of the details will suffice.
Each dimension in space has an associated eigenvalue. Start with the eigenvalue
given by Theorem 9.1. The operator T is then restricted to the space orthogonal to the
eigenvector associated to that eigenvalue. This effectively reduces the dimension of the
space by one. The restricted operator is also compact and self-adjoint. If it is also non­
zero, then use of Theorem 9.1 will give a new eigenvalue and a new associated eigenvector.
 32

The restricted operator maybe zero, which is how this procedure is completed in finite
dimensional spaces. This algorithm is continued ad finitum if the original space is infinite
dimensional. Beginning with the eigenvalue from the first iteration,

|Ai| > ]A2| > |A3| > ...

terminating in the finite dimensional case. Also, the sequence (An) converges to zero by
Theorem 6.7.

Theorem 9.2. (Schur’s Theorem.) Let H be a Hilbert Space and let T G L(H) be a
compact and self-adjoint operator. For a sequence {An} of nonzero eigenvalues ofT, the
associated orthonormal sequence of eigenvectors (xn) and fx G H the following equality
holds: oo
Tx = An(o;, xn)xn, x G H.
n=l

This is the Schur Representation of the operator T. Also er(T) — {An} U {0}.

9.2 Square Roots

Positive self-adjoint operators have enough structure that they have many analo­
gies to real numbers. There is a partial ordering of these operators (as was discussed in
Proposition 8.3), similar to the ordering of the real numbers. Another analogy to real
numbers is the existence of a unique square root. Both existence and uniqueness must
be proven and proof is available in a variety of texts on Hilbert Space. The consequences
of this square root are numerous and useful in the remainder of this work.

Definition 9.3. Let T G £(/7) be positive and self-adjoint operator. If there is a self-
adjoint A G L(H) such that A2 = T then A is called a square root of T. If A is also
positive then A is called the (meaning unique) square root of T and denoted A = T1?2.

Proposition 9.4. For a positive self-adjoint operator T € K(H) there is a positive

operator A with A2 = T. A is the unique square root ofT.

Proof. The Schur Representation of positive self-adjoint T G K(H) is

oo
Tx = ^Xn(x,xn)xn.
n=l
 33

(Schur’s Theorem requires that T f 0. If T — 0 the theorem, and hence the above
representation, does not apply. However, in this case A = 0 fills all the requirements
to be the unique positive square root.) Theorem 8.4 states that [AJ are real numbers
(which are known to have unique square roots). Define fa — A^2.
There exists a positive self-adjoint operator A such that
oo
Ax = r € H.
n=l

Then A is self-adjoint and positive.

It follows that A is unique by the uniqueness of A^2. To show A2 = T a
few calculations are necessary. Start with the Schur Representation of A and apply the
operator A again. So, for x € H it holds that
oo / oo 1 X

A X = AAx ~ } fan I 5> ^m) j

m=l \ n=l /
oo
=
n=l
oo

n=l
= Tx.

Leaving the operator A2 — T. □

9.3 Polar Decomposition

There are two more details that need attention before the main topic of this
chapter. The polar decomposition, defined below, provides a connection between the
operators T, T
*, and its square root A. Note: Proposition 8.3 states T
* > 0. The
operator T
* then has a unique positive square root.

Theorem 9.5. (The Polar Decomposition Theorem.) Let T G L(H) and A = (71*? 1)1/2.
There is an operator U € L(H) such that

1. T=UA;

2. ||(7rc|| = ||rr|| for x G R(A);

 34

3. Ux = OforxeR(A)±.

Proof. Since A2 = T
* *,
— A for all x e H ||Ar||2 = (Ax, Ax) = (A
*Aa;,x) =
*Tx,x)
(T = ||2Yc||2, and ||Ax|| = ||Tx||. If we. let U : R(A) —> R(T) be given by
U(Ax) = Tx for x G H, then U is a well-defined bounded linear operator on R(A) and
11^ = ||T:r|| for x G Z?(A). Hence, U can be extended onto I?(A) and, moreover, we
can define U = 0 on Then, U G L(H) and it satisfies all the above statements.
□
This operator U has the added property of taking the set of eigenvectors of A
(which are orthonormal) to another set which is orthonormal as well.

Theorem 9.6. Let T G K(H) and let U be the operator defined by Theorem 9.5. Let
(xn) be the orthonormal eigenvectors of A — (PT)1/2. Then the set (Uxn) is also
orthonormal.

Proof. Given (xn) is orthonormal, the second part of Theorem 9.5 has ||t7a?n || = ||a;n|| = 1
for all n. Then for n m,

|\UXn Uxm || = ||t7(iEn 3?m)|| = IK^n ^m)!!2 = 2.

But

||Us?n — (Uxn Uxm,Uxn Uxm)

= (Uxm Uxn) (Uxn, Uxm) (Uxm, Uxn) 4” (Uxm, Uxm)
= 1 (fJxn, Uxm) (Uxm, Uxn) "f-1-

Combining the two equations, —(Uxn,Uxm) = (Uxn,Uxm). This implies that the real
part of (Uxn, Uxm) must be zero.

A similar calculation for xn + ixm results in ||{Arn + Uixmll2 = 2 and thus

(Uxn, Uxm) = (Uxn, Uxm). So the imaginary part is also zero. Therefore (Uxn,Uxm) —
0 for all n 0 m and the set (Uxn) is orthonormal. □
 35

Chapter 10

The Schmidt Representation

There has now been enough background in operators to move onto the major
topics of this paper. Many of the previous topics will be refined and combined into a
much more focused underlying theory.
Schur’s Representation is helpful in calculations involving compact, self-adjoint
operators. A similar result would be useful for operators that don’t meet the strict
restriction of being self-adjoint. Combining topics from the previous chapters gives such
a result.
To ease notation, a brief lemma.

Lemma 10.1. Given T G K(H), let A = (T

T)
* X/2. The eigenvalues (rn{T) of A found
by the “procedure” - outlined on pages 31,32 - given by

an(T) — inf{||T — B|| : B is a finite rank operator with rank B < n}.

These crn(T) are also called the singular numbers of T. Some examples using
singular numbers will follow shortly.
Finally, the Schmidt Representation Theorem, due to the German mathemati­
cian Erhard Schmidt.

Theorem 10.2. (The Schmidt Representation Theorem.) Let T G K(H) and A =

*T)
(T X/2. Also let (crn(T')) be the eigenvalues of A and (xn) the associated orthonormal
eigenvectors. Then there is an orthonormal set (yn) such that
oo
Tx = ^2 °n(T)(x, xn)yn, xeH.
n=l
 36

Proof. For T G K(H'), A is compact and self-adjoint. A then has a Schur Representation
oo
Ax = ^<rn(T)(x,xn)xn
n=l

for the eigenvalues, an(Tf of A and associated eigenvectors (xn). Then Theorem 9.6 says
that yn = Uxn is an orthonormal set. Thus using the polar decomposition of the operator
T leaves oo oo
Tx = UAx = U^an(T)(x,xn)xn = ^(rn(T)(x,xn)yn
n=l n=l

as desired. □

Calculations using the Schmidt Representation deal with infinite sums and inner
products which are relatively easy to work with. It is the singular numbers crn(T) which
may take some work to gain intuition of their properties. Some examples will assist in
that intuition.

Example 10.3. Show ||T|| = <Ji(T) > ^(T) > ... for T G K(H).

Proof. The definition of an(T) states that ai(T) = inf{||T[|}. (Note that rank A < 1
implies that A = 0.) So crip1) = ||T||.
Since = inf{||T — A|| : rank A < n}, an+i(T) = inf{||T — A|| : rank A <
n 4- 1}, and

{A : A is of finite rank with rank A < n+1} C {A : A is of finite rank with rank A < n}

we see that crn(Tj > crn+i(T).

□
Next, an example demonstrating how to break up, or combine, the eigenvalues
of two operators.

Example 10.4. Show for S,T G am+n-i(S + T) < crn(S) 4- crm(T).

Proof. By definition, am^n-i(S + T) = inf{||S + T — A|| : rank A < m 4- n — 1}. Then

for A G L(H') with rank A < m + n — 1, BB, C G L(H) such that A = B + C, rank B <n,
rank C < m and B D range C = {0}. (The only element of both ranges is the element
zero.) So rank A = rank (B 4- C).
 37

Starting with one side

am+n-i(S + T) = inf{||S 4- T — A|| : rank A < m 4- n — 1}

= inf{||S 4-T — B — C|| : rank (B + C) < m + n - 1}
< inf{||S - B|| + \\T - CH : rank (B 4- C) < m 4- n - 1}
< inf{||S — B|| rank B < n — 1} 4- inf{||T — C|| rank C < m}
< infills' - B|| rank B < n} 4- inf{||T - CH rank C<m]
— &n(&) 4*

So <rm+n_i(S 4* T) < crn(S) 4- vm(T) as desired. □

Lastly, an example establishing a rule for the norm multiple operators and their
corresponding eigenvalues.

Example 10.5. Show ||B|| • ||T||crn(S) > an(RST) for operators R,S,T G K(H).

Proof. Starting with the left hand side

||B|| • ||T||an(S) = ||BK(S')||T||,

= ||B|| ■ inf{||S - A|| : rank A < n} • ||T||,
= inf{||B|||S - A||||T|| : rank A < n],
> inf{||B(S — A)T|| : rank A < n},
= inf {||RST - BAT|| : rank A < nl
> inf{||KST — B|| : rank B < n} = <jn(RST)

since rank A < n implies rank RAT < n. □

 38

Chapter 11

The Schatten Class Operators

This is where the work in the previous chapters comes together. This culmina­
tion begins with the work done by Robert Schatten. The Schatten p-class operators or
Sp operators are defined below.

Definition 11.1. The Schatten p-class operators on Hilbert space H are given by

SP(H) = {T e K(H)\(an(T)) G £p}, 1 < p < oo.

This definition states that a compact operator T is in SP(H) if the sequence of

eigenvalues of [T
*T] 1/2 are p-summable. Soon p will be fixed for certain situations, but
for the following examples p will be unspecified.

Example 11.2. For Hilbert space H, ifTe Sp(H) thenT can be factored into T = BDA
as follows:
H------- ---- ^H

A B

T
loo

where D((an)) = (crn(T)an) for all an e £<»•

Proof. Let x € H. Then x can be expanded and written as

oo
X = ^(x>xn)xn
n=l
 39

for the complete orthonormal basis xn created in Schur’s Theorem, Theorem 9.2. Let
oo
Ax = ((x,xn)) = (an). Then (an) G ^oo> because 52 W2 < oo- Let D((an)) =
n=l
(trn(T)an). But is (crn(T)an) G €p? Well
oo oo oo
£ |a„(T)a„f < £ U(an)IIEo ' K(T)|’ = 11 (*n)l|go ’ £ MT)IP < oo.
n=l n=l n=l

So (an(T)an) G €p.
oo
Now let B(an(T)an) = X &n(T)(an)yn, where T(xn) = yn for n > 1 and (yn)
n=l
is the orthonormal set found by using Theorem 9.6, on the orthonormal set (xn). Finally,
00

BDA(x) = BDA y^(g, xn)xn)

n=l
oo
= BD^(an)xn,
n=l
oo
— B'y (crn(T’)qra)xn,
n=l
oo
= ^an(T)(an)yn,
n=l
oo
= y2CTn(T)(x,xn)yn,
n=l
= Tx.

The last equality being the Schmidt Representation of Tx. □

As with any object defined so far, the Schatten class operators must have a norm
associated with it.

Definition 11.3. Given Hilbert space H and T G SP(H) the Schatten p-norm, written
sp(H) is
/ \ i/p

<n = (£^m • An=i 7

Theorem 11.4. Let T G K(H). If for all U,V G L(f2,H) ((TUe^Ven)) G £P) then
T G SP(H)

Proof Given T G K(H) so T has Schmidt representation

oo
Tx = ^Jcrn(T)(x,xn)yn
n=l
 40

for some orthonormal sets (xn), (yn)' Let U, V be operators which change the orthonormal
set (ej) as follows:
Uei — Xi and Vei = yi, i> 1.

The statement of the theorem says that for every, and as such this particular, U, V that
((TUen,Ven)) So ((Tx^yi)) G £p. On the other hand,

(Txi,yi) =

= (^i(^)yi^i)
(u.i)

Therefore (o/T)) G £p and T G SP(H) as desired. □

 41

Chapter 12

Trace Class and Hilbert-Schmidt

Operators

To complete this paper two final classes of operators will be discussed. Both
of these operators predate the Schatten p-class operators. It will be shown that the
trace class and Hilbert-Schmidt class operators are actually specific types of the Schatten
p-class operators.

Definition 12.1. An operator T G L(H) is in the trace class, denoted T G TC(ff), if

oo
Tx = ^2(x,yn)xn
n=l

where ||^n||||z/n|| < +oo. The norm of this operator is defined by

r(T) =inf|j2lWIWl}

with the infimum taken over all possible representations of T G TC(H).

Theorem 12.2. The trace class is identical with the Schatten 1-class. That is,

TC(ff) = Stiff)

and
si{T) = r(T) for T G TC(ff).
 42

Proof. If T G TC(H), then T G K(H) and T has a Schmidt representation. Let e > 0 be
given. Since T G TC(H) there is some representation'of T such that for all x G H there
is a (wn) and (zn) where
oo oo
Ta; = ^(x,wn)zn and 52 11'^1111^11 < (1 + c)t(T).
n=l n=l

oo
This gives a representation that admits £ llwn II Ikn || arbitrarily close to - but larger
n=l
than - r(T). For T to be in S/H), ^trn(jr) must be finite. This can be verified by
finding an upper bound for J3crn(T).
Let (xn), (yn) be orthonormal sets. Then, beginning with a calculation similar
to (11.1) in Example 11.4,
oo oo
= y~^(Ta?n, yn)
n=l n=l
00 ✓ oo

~ 5 f 5'2 kxnt wm)zm, yn

n=l ' m=l
oo oo

“ 5 5 (xni wm)(zm,yn)
n=lm=l
00 / oo \ x/2 ,f oo
< 521 52 52 l(Am, yn)|
m=l \ n=l / <n=l
oo
< 52 Ikmllll^mll
m=l
< (1 + e)r(T) < oo.
So TC(H) C 5i (S') and S1(T) < r(T).
Then let T G 5i(-£T). So T has a Schmidt representation,
oo

Ta; = 5 ' an(T)(x, Zn)xn

n=l

Letting vn(T)zn = yn,

OO

Tx =

71=1
00 oo
which means T is also in TC(H). And £ llz/nll||^n|| = 52 ffn(T). So r(T) < si(T).
Therefore Si(H) is identical to TC(H). □
 43

Definition 12.3. An operator T G L(H), is in the Hilbert-Schmidt class, denoted T G

HS(H), if there is a complete orthonormal set (yn) for H with

JSlI^ynll2 < +oo,

n=l

where the Hilbert-Schmidt norm of T operator is defined by,

oo 1/2
£ll^„||2)
n=l '

for any complete orthonormal set (yrt).

It should be noted that hs(T) is independent of the choice of the complete

orthonormal set of (yn).
The Hilbert-Schmidt operators were studied at length before Schatten began
work with the Schatten p-classes. It would be nice to bring the solutions found using
Hilbert-Schimdt operators into the framework of the Schatten p-class operators. The
following theorem will classify this connection.

Theorem 12.4. The Hilbert-Schmidt class, HS(H), is identical with the Schatten 2-
class, S2(H). Their norms also coincide.

Proof. Let T G HS(H) and let (xn) be a complete orthonormal set in H. Then for e > 0
there is an TV such that
00 \ V2

E
n=Ar+l
iitm2
/
<e-
N
Define the orthogonal projection Px ~ f^(x,xi)xt. Thus for x G H,
i=l

oo

E
n=7V+l
(z, Xi)Txi

°° \ I 00
< E ( E
n=7V+l / \n=7V+l

< ||a:||e.
This implies that \\T — TP\\ < e hence T G K(H). So T has a Schmidt representation
oo
Tx = ^ok(T)(x,Uk)vk
k=l
 44

for orthonormal sets (uk) and (vt). Using the representation of T,

OO 00

£||T®n||2 = y(Txn,Txn)
n=l n=l
OO OO

= E£^CD2|(
*».^)| 2
n=l fc=l
OO OO

= E£^(r)2i(^,^)i2
fc=l n=l
00
= j>
*
(T) 2-
fc=l

By definition T G 82(H) and hs(fT) = S2(T).

Assuming T G 82(H), the above proof works in reverse to show T G HS(T). So
these two classes are identical. □

Definition 12.5. For a Banach space X, a sequence (yn) C X is said to be weakly square
summable provided
00
J2l/(?/n)|2 < +00

n=l

* and square summable if

for all / G X

52 iwi2 < +o°-

n=l

Definition 12.6. Let X and Y be Banach spaces. An operator T G L(X, Y) is called

absolutely 2-summing if T maps weakly square summable sequences in X into square
summable sequences in Y.
The set of absolutely 2~summing operators is denoted ^)- Any operator
T in this set is called a []2 operator.

Once again, the new operator should be put into context with respect to Sp.

Theorem 12.7. The Schatten 2-class operators are identical to the absolutely 2-summing
operators.

Proof. As a reminder, S2 = {T 6 K(H)\(an(T)) G £2} and

 45

The proof is by double inclusion. First, it will be shown that J]2 c *^2- Note:
Theorem 12.4 states SifH) is equivalent to
Let T 6 f^C^) 311(1 let fen) C H be a complete orthonormal set. Then (yn)
is a weakly square summable sequence. Then T G EM^) imPlies that (Tyn) is square
summable, ie. ||Tyn||2 < oo« Thus T G HS(H) and therefore T e S2(#)- So J]2 C 52-
Next choose T G S2 and let (yn) be weakly square summable. Theorem 11.4
states that T G /C(7T) implies VC7, V G £(^2>77)> ((7Ven, Ven)) G ^2. This is true for all
U, V\ so in particular, the U where Uen = yn and V where V = TU. Then ((Tyn, Tyn)) €
^2 and (||Tyn ||2) G €2, which implies
oo
^IITshH^+oo.
n=l

Thus T G n2- Therefore J]2(#) = ^2 (H). □

 46

Chapter 13

Conclusion

In an effort to extend information from to an infinite dimensional space many

abstractions have been explored. Among those abstractions were the familiar notions of
length, dot product, and eigenvalues which were generalized to norm, inner product, and
the spectrum respectively. An appropriate framework for considering these abstractions
was given along with meaningful examples.
The push for an appropriate analogue of the trace for an operator in a Hilbert
space has led to the classification of compact operators, self-adjoint operators, the square
root of operators and finally the trace class operators. In fact, the trace class operator
was found to be just one type of a wide range of interesting operators, the Schatten p-class
operators.
 47

Appendix A

Banach Space

A complete normed linear space is called a Banach space. This space was touched
on very briefly in Section 2.1. While the focus of this paper is on Hilbert space, there Eire
undeniable links between these two types of spaces. One such link is studied when dealing
with the Schatten p-class operators, so a short primer on Banach spaces is warranted
[Roy88].
The Banach spaces of interest are called spaces. The familiar space ^2 is one
of these spaces. (A Hilbert space is after all a Banach space with an inner product.) The
spaces are defined
f
tp = * (ttn) 22l®nlP < +°°
n=l
Again, a norm is required and it is defined to be
/ 00 \ Vp
ll(®n)llP= ( '
This space is also defined for p = 00. In that case,

||(an)||oo = sup [an|.

n>l

There are some fundamental theorems in Banach space theory, such as Holder’s
Inequality and Minkowski’s Inequality, which are given here in a form conducive for work
in Hilbert space. These inequalities are focused on Banach space which is a secondary
goal. Consequently, only brief outlines of the proofs are given. Complete proofs are
readily available.
 48

Theorem A.l. (Holder's Inequality). Let cq,..., an, . ,bn C C and let p, q > 0 be
such that A + 1 = 1. For p> 1 it holds that

n n n P
1.

52^1 -
3=1 .3=1 Z>l’
Ew” .3=1
This inequality is also obtained for p = 1;
n

52
i=i
- .3=1 : 1 < i < n}

Proof. The proof of Holder’s Inequality comes easily using Young’s Inequality, ab <
~ + y, and a clever substitution for a and b. □

Minkowski’s Inequality provides the triangle inequality for £p spaces.

Theorem A.2. (Minkowski's Inequality). Let cq,..., an, tq,... ,bn cC and let p, q> 0
be such that | ~ = 1. For p>l,

n
1
P n n1
P

< £l«dp + £im5

.3=1
.3=1 .3=1
Rewriting using the definition of the norm,

a + 6||p < ||a||p + ||b||p.

Proof. The proof follows from Holder’s Inequality by factoring out a term of |a + 6| from
the sum notation. Note p = p/q + 1. □

In Section 2.1 was not shown to be complete. Instead, it will be shown here
that lp is complete.

Proposition A.3. The space lp is complete for 1 < p < 00.

Proof. First, as a matter of notation, build a sequence of sequences in the following

manner. Let G Cp and {aWjgj. be a Cauchy sequence Given the
sequence is Cauchy says Ve > 0, BA such that i,j > N => ||aW — a^||p < e and by
definition,
< 6.
 49

For any k we have

p oo p
ak(i) - a
a akU} <€P.
k=o
For this to be true for all e, -a, must approach zero as i,j —> oo. Thus (a^) is
Cauchy in R. So limf_ooafc> = at exists. Let a = All that is left to show is
— a||p —> 0 and a G Z?p.
For i,j > N,
/ 00
E
vfe=l
O*’ - ak

and taking the limit as j goes to infinity

f 00
lim ( V

Taking the limit

and ||n^ — a||p < e. Hence aft —> a as ? —> oo.

Finally, pick an i such that, using Minkowski’s inequality, llallp < ||a - oft ||p +
||a^||p. Since |]a^ — a||p < e, ||a||p < e + ||a^ ||p which is finite as it is in €p. Therefore
||a||p < oo and a G £p. □

While all the work in this section holds for p = oo, the above results are sufficient
for present purposes.
 50

Bibliography

[Ret93] J. R. Retherford. Hilbert Space: Compact Operators and the Trace Theorem,
volume 27 of London Mathematical Society Student Texts. Cambridge University
Press, Cambridge, 1993.

[Roy88] H. L. Royden. Real Analysis. Macmillan Publishing Company, New York, third
edition, 1988.

[Sch60] Robert Schatten. Norm Ideals of Completely Continuous Operators. Ergebnisse

der Mathematik und ihrer Grenzgebiete. N. F., Heft 27. Springer-Ver lag, Berlin,
1960.