Functional Analysis With Applications

Palle Jorgensen, and Feng Tian
arXiv:1408.1164v1 [math.FA] 6 Aug 2014
Functional Analysis with

Applications
2
A quote:
“ ‘The Journal of Functional Analysis’ is dedicated to the broadening of the horizons of

functional analysis. Accordingly, it encourages original research papers of high quality from
all branches of science, provided the core and flavor are of a functional analytic character
and the paper is in accordance with contemporary mathematical standards.”– From the cover
page of ‘The Journal of Functional Analysis’ (written in 1967.)
Dedicated to the memory of
William B. Arveson (22 November 1934 – 15
November 2011).
“What goes around has come around, and today

quantum information theory (QIT) has led us back into
a finite-dimensional context. Completely positive maps
on matrix algebras are the objects that are dual to
quantum channels; in fact, the study of quantum
channels reduces to the study of completely positive
maps of matrix algebras that preserve the unit. This is
an area that is still undergoing vigorous development
in efforts to understand entanglement, entropy, and
channel-capacity in QIT.” — William B. Arveson
(written around 2009.)
Foreword
Over the decades, Functional Analysis has been enriched and inspired on account of
demands from neighboring fields, within mathematics, harmonic analysis (wavelets
and signal processing), numerical analysis (finite element methods, discretiza-
tion. . . ), PDEs (diffusion equations, scattering theory), representation theory; iter-
ated function systems (fractals, Julia sets, chaotic dynamical systems), ergodic the-
ory, operator algebras, and many more. And neighboring areas, probability/statistics
(for example stochastic processes, Ito and Malliavin calculus), physics (representa-
tion of Lie groups, quantum field theory), and spectral theory for Schrödinger oper-
ators.
5
Preface
There are already many books in Functional Analysis, so why another?

The main reason is that we feel a need: in the teaching at the beginning grad-
uate level; more flexibility, more options for students and instructors in pursuing
new directions. And aiming for a book which will help students with primary inter-
ests elsewhere to acquire a facility with tools of a functional analytic flavor, say in
harmonic analysis, numerical analysis, stochastic processes, or in physics.
7
Acknowledgment
The first named author thanks his students in the Functional Analysis sequence 313-
314. Also he thanks postdocs and colleagues; among them: D. Alpay, W. Arveson,
O. Bratteli, P. Casazza, I. Cho, D. Dutkay, R. Kadison, W. Klink„ K. Kornelson, D.
Larson, P. Muhly, R. Niedzialomski, J. Packer, E. Pearse, R.S. Phillips, W. Polyzou,
D. Robinson, S. Sakai, I.E. Segal, K. Shuman, R. Werner.
9
Contents
Part I Introduction and Motivation
Part II Topics from Functional Analysis and operators in Hilbert space: a

selection
1 Elementary Facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.1 A Sample of Topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.2 Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.3 Transfinite Induction (Zorn and All That) . . . . . . . . . . . . . . . . . . . . . . 27
1.4 Basics of Hilbert Space Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.4.1 Orthonormal Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
1.4.2 Bounded Operators in Hilbert Space . . . . . . . . . . . . . . . . . . . . 35
1.5 Dirac’s Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
1.5.1 Three Norm-Completions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
1.5.2 Connection to Quantum Mechanics . . . . . . . . . . . . . . . . . . . . . 44
1.5.3 Probabilistic Interpretation of Parseval in Hilbert Space . . . . 48
1.6 The Lattice Structure of Projections . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2 Unbounded Operators in Hilbert Space . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

2.1 Domain, Graph, and Adjoints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.2 Characteristic Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.2.1 Commutants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.3 Normal Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
2.4 Polar Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
11
12 Contents
3 The Spectral Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

3.1 An Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.2 Multiplication Operator Version . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.2.1 Transformation of Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.2.2 Direct Integral Representation . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.2.3 Proof of Theorem 3.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.3 Projection-Valued Measure (PVM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.4 Convert Mϕ to a PVM (projection-valued measure) . . . . . . . . . . . . . . 90
3.5 The Spectral theorem for compact operators . . . . . . . . . . . . . . . . . . . . 95
3.5.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Part III Applications
4 GNS and Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

4.1 Definitions and Facts: An Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.2 The GNS Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.3 States, Dual and Pre-dual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
4.4 New Hilbert Spaces From “old” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
4.5 Abelian C∗ -algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.6 States and Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.6.1 Normal States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
4.6.2 A Dictionary of OP and QM . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4.7 Krein-Milman, Choquet, Decomposition of States . . . . . . . . . . . . . . . 132
4.7.1 Noncommutative Radon-Nikodym Derivative . . . . . . . . . . . . 136
4.7.2 Examples of Disintegration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
4.8 Examples of Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
4.9 Beginning of Multiplicity Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
5 Completely Positive Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

5.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
5.2 CP v.s. GNS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
5.3 Stinespring’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
5.4 Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
6 Brownian Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

6.1 The Path Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
6.2 Decomposition of Brownian motion . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
Contents 13
7 Applications to Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

7.0.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
7.0.2 Group - Algebra - Representations . . . . . . . . . . . . . . . . . . . . . . 175
7.1 Some examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
7.1.1 ax+b group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
7.2 Induced representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
7.2.1 Induced Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
7.3 Example - Heisenberg group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
7.3.1 ax + b group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
7.4 Co-adjoint Orbits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
7.4.1 review of some Lie theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
7.5 Gårding Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
7.6 Decomposition of representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
7.7 Summary of Induced Representations, the Example of d/dx . . . . . . . 208
7.7.1 Induced representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
7.8 Connections to Nelson’s Spectral Theory . . . . . . . . . . . . . . . . . . . . . . . 212
8 The Kadison-Singer Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219

8.1 Frames in Hilbert Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
Part IV Extension of Operators
9 Selfadjoint Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227

9.1 Extensions of Hermitian Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
9.2 Cayley Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
9.3 Boundary Triple . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
9.4 The Friedrichs Extension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
9.5 Rigged Hilbert Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
10 Unbounded Graph-Laplacians . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261

10.1 Basic Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
10.2 The Energy Hilbert Spaces HE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
10.3 The Graph-Laplacian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
10.4 The Friedrichs Extension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
10.5 A 1D Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
14 Contents
11 Reproducing Kernel Hilbert Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279

11.1 A Digression: Stochastic Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
11.2 Two Extension Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
11.3 The Reproducing Kernel Hilbert Space HF . . . . . . . . . . . . . . . . . . . . . 283
11.4 Type I v.s. Type II Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
11.5 The Case of e−|x| , |x| < 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
11.5.1 The selfadjoint Extensions Aθ ⊃ −iDF . . . . . . . . . . . . . . . . . . 295
11.5.2 The Spectra of the s.a. Extensions Aθ ⊃ −iDF . . . . . . . . . . . . 300
A An overview of Functional Analysis books (cast of characters) . . . . . . . 307
B Often cited above . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
Part I
Introduction and Motivation
17
Motivation: More traditional books on Functional Analysis tend to postpone ap-

plications till after all the big theorems from the theory have been covered. The
purpose of the present book is to give students a tour of a selection of applications.
We aim to do this by first offering a crash course in theory topics tailor made for
the purpose (part II). In order to stress the interplay between theory and applica-
tions (part III) we have emphasized the traffic in both directions. We believe that the
multitude of new applications makes Functional Analysis both a powerful, versatile,
and timeless tool in mathematics.
A glance at existing books in Functional Analysis and related areas (see list of
reviews in the Appendix) shows that related books so far already display a rich
variety, even if they may have the same title “Functional Analysis” or “Functional
Analysis with a subtitle, or a narrowing of the focus.”
Still we feel that the aims, and the contents of these other books go in a different
direction than ours. One thing they have in common is an emphasis on the Four Big
Theorems in Functional Analysis, The Hahn-Banach Theorem, The Open Mapping
Theorem, The Uniform Boundedness Principle, The Closed Range Theorem, and
duality principles.
By contrast, we do as follows; rather we select a list of topics that acquire a
degree of elegance when presented in a functional analytic setting. There are several
reasons for this different approach, the main ones are as follows: (i) the subject is
ever changing on account of demands from neighboring fields; (ii) students come
to our graduate functional analysis course with a diversity of backgrounds, and a
“one-size fits all” approach is not practical; (iii) well-motivated students can pick
up on their own what is needed from the Four Big Theorems; (iv) concentrating
on the Four Big Theorems leaves too little time for a variety of neighboring areas,
both within mathematics, and in neighboring sciences. (v) Also the more traditional
approach, beginning with the Four Big Theorems is already in many existing books
(see the Appendix).
A glance at the Table of Contents will reflect our aim: beginning with tools from
Hilbert space in chapters 1 & 2, but motivated by quantum physics; a preview of the
Spectral Theorem in chapter 3; some basic tools from the theory of operator algebras
in chapter 4, with an emphasis on the Gelfand-Naimark-Segal (GNS) construction;
and stressing the many links between properties of states, and the parallel properties
of the representations, and the operator algebras they generate.
18
In chapter 5, and motivated by physics and harmonic analysis, we discuss dilation

theory. This is the general phenomenon (pioneered by Stinespring and Arveson) of
studying problems in an initial Hilbert space by passing to an enlarged Hilbert space.
Chapter 6 (Brownian motion), while different from the others, still fits perfectly,
and inviting application of the tools already discussed in the first four chapter. The
applications we cover in chapter 7 are primarily to representations of groups and
algebras. Chapter 8 is an allocation of theorems from chapters 3-4 to the problem
named after Kadison and Singer, now abbreviated the KS-problem. It is 50 years
old, is motivated by Dirac’s formulation of quantum physics (observables, states,
and measurements); and it was solved only a year ago (as of present).
The last three chapters, 9 selfadjoint extensions, 10 graph-Laplacian, and 11 re-
producing kernel Hilbert spaces (RKHSs), are somewhat more technical, but they
are critical for a host of the questions inside the book. Some readers may be familiar
with this already. If not, a quick reading of chapters 2, 9, and 10 may be useful. Sim-
ilarly, in the appendix, to help students orient themselves, we give a birds-eye view
of, in the order of 20 books out there, all of which cover an approach to Functional
Analysis, and its many applications.
Note on cited books and papers. For readers looking for references on the foun-
dations, our suggestions are as follows: Operators in Hilbert space: [Arv76, Arv72].
Quantum mechanics: [OR07, Wei03, GG02]. Non-commutative functional analy-
sis and algebras of operators: [BJKR84, BR81, BR79]. Unitary representations of
groups: [Mac92, Mac85, Mac52].
19
Subject Example
´x
A analysis f (x) − f (0) = 0 f 0 (y) dy
B dynamical systems functions on fractals, Cantor set, etc.
C PDE Sobolev spaces
D numerical analysis discretization
E measures / probability theory probability space (Ω , F, P)
F quantum theory Hilbert spaces of quantum states
Table 0.1: Examples of Linear Spaces: Banach spaces, Banach algebras, Hilbert
spaces H , linear operators act in H .
Reader Guide.
Below we list chapter by chapter how the six areas in Table 0.1 are covered.
Ch 1: Areas A, E, F.
Ch 2: Areas A, C, F.
Ch 3: Areas B, C, E, F.
Ch 4: Areas A, B, E, F.
Ch 5: Areas A, E, F.
Ch 6: Area E.
Ch 7: Areas D, F.
Ch 8: Areas E, F.
Ch 9: Areas B, C, D.
Ch 10: Areas A, F.
Ch 11: Areas A, B, C, D, E.
In more detail, the six areas in Table 0.1 may be fleshed out as follows:
Examples of subjects within area A includes measure theory, transforms, con-

struction of bases, Fourier series, Fourier transforms, wavelets, and wavelet trans-
forms, as well as a host of operations in analysis.
20
Subjects from area B include solutions to ordinary differential equations (ODEs),

and the output of iteration schemes, such as the Newton iteration algorithm. Also
included are ergodic theory; and the study of fractals, including harmonic analysis
on fractals.
Item C encompasses the study of the three types of linear PDEs, elliptic, parabolic
and hyperbolic. Sample questions: weak solutions, a priori estimates, diffusion
equations, and scattering theory.
Area D encompasses discretization, algorithms (Newton etc), estimation of er-
ror terms, approximation (for example wavelet approximation, and the associated
algorithms.)
Area E encompasses probability theory, stochastic processes (including Brown-
ian motion), and path-space integration.
Finally part F includes the theory of unbounded operators in Hilbert space, the
three versions of the Spectral Theorem, as well as representations of Lie groups, and
of algebras generated by the commutation relations coming from physics.
Part II
Topics from Functional Analysis and
operators in Hilbert space: a selection
Chapter 1
Elementary Facts
“. . . the [quantum mechanical] observables are operators on a

Hilbert space. The algebra of operators on a Hilbert space is
noncommutative. It is this noncommutativity of operators on a
Hilbert space that provides a precise formulation of
[Heisenberg’s] uncertainty principle: There are operator
solutions to equations like pq - qp = 1. This equation has no
commutative counterpart. In fact, it has no solution in operators
p,q acting on a finite dimensional space. So if you’re interested
in the dynamics of quantum theory, you must work with
operators rather than functions and, more precisely, operators
on infinite dimensional spaces.”
— William B. Arveson (1934-2011. The quote is from 2009.)
I received an early copy of Heisenberg’s first work a little before

publication and I studied it for a while and within a week or two
I saw that the noncommutation was really the dominant
characteristic of Heisenberg’s new theory. It was really more
important than Heisenberg’s idea of building up the theory in
terms of quantities closely connected with experimental results.
So I was led to concentrate on the idea of noncommutation and
to see how the ordinary dynamics which people had been using
until then should be modified to include it. — P. A. M. Dirac
Problems worthy
of attack
prove their worth
by hitting back.
— Piet Hein
Below we outline some basic concepts, ideas, and examples which will be studied
inside the book itself. While they represent only a sample, and we favor the setting
of Hilbert space, the details below still tie in nicely with diverse tools and techniques
not directly related to Hilbert space.
23
24 1 Elementary Facts
1.1 A Sample of Topics
Classical functional analysis is roughly divided into two branches, each with a long
list of subbranches:
• study of function spaces (Banach space, Hilbert space)

• applications in physics and engineering
Within pure mathematics, it is manifested in
• representation theory of groups and algebras, among a long list of diverse

topics
We will consider three classes of algebraic objects of direct functional analytic

relevance: (i) generators and relations (for example the canonical commutation re-
lations for the operators P and Q); (ii) algebras, and (iii) groups.
In the case of (i), we illustrate the ideas with the canonical commutation relation
PQ − QP = iI (1.1)
The objective is to build a Hilbert space such that the symbols P and Q are rep-
resented by unbounded selfadjoint operators, defined on a common dense domain
in some Hilbert space, and with the operators satisfying (1.1) on this domain. In
class (ii), we consider both C∗ -algebras and von Neumann algebras (also called W ∗ -
algebras); and in case (iii), our focus is on unitary representations of the group G
under consideration. The group may be abelian or non-abelian, continuous or dis-
crete, locally compact or not. Our present focus will be the case when G is a Lie
group. In this case, we will study its representations with the use of the correspond-
ing Lie algebra.
• C∗ -algebras, von Neumann algebras
We will be considering C∗ -, and W ∗ -algebras axiomatically. In doing this we use

the theorem by S. Sakai to the effect that the W ∗ -algebras consist of the subset of
the C∗ -algebras that are the dual of a Banach space. If the W ∗ -algebra is given the
Banach space is called the pre-dual. Representations will be studied with the use
of states, and we stress the theorem of Gelfand, Naimark, and Segal (GNS) linking
states with cyclic representations.
• wavelets theory
1.2 Duality 25
A wavelet is a special basis for a suitable L2 -space which is given by generators

and relations, plus self-similarity. Our approach to wavelets will be a mix of func-
tional analysis and harmonic analysis, and we will stress a correspondence between
a family of representations of a particular C∗ -algebra, called the Cuntz-algebra, on
one side and wavelets on the other.
• harmonic analysis
Our approach to harmonic analysis will be general, – encompassing anyone of a

set of direct sum (or integral) decompositions. Further our presentation will rely on
representations.
• analytic number theory
Our notions from analytic number theory will be those that connect to groups,
and representations; such as the study of automorphic forms, and of properties of
generalized zeta-functions.
1.2 Duality
The “functional” in the name “Functional Analysis” derives from the abstract notion
of a linear functional: Let E be a vector space over a field F (we shall take F = R,
or F = C below.)
Definition 1.1. A linear function ϕ : E → R (or C) is said to be a linear functional,

i.e., we require
ϕ (u + λ v) = ϕ (u) + λ ϕ (v) , ∀λ ∈ R, ∀u, v ∈ E.
If E comes with a topology (for example from a norm, or from a system of semi-
norms), we say that ϕ is a continuous linear functional. Occasionally, continuity
will be implicit.
Definition 1.2. The set of all continuous linear functionals is denoted E ∗ , and it is
called the dual space. (In many examples there is a natural identification of E ∗ as
illustrated in the following examples (Table 1.1).
Definition 1.3. If E is a normed vector space, and if it is complete in the given norm,
we say that E is a Banach space.
Lemma 1.1. Let E be a normed space with dual E ∗ . For ϕ ∈ E ∗ , set
kϕk∗ := sup |ϕ (x)| .

kxk=1
Then (E ∗ , k·k∗ ) is a Banach space.
Proof. An exercise.
E E∗ how?
1
l p , 1 ≤ p < ∞ with lq, p + q1 = 1 x = (xi ) ∈ l p , y = (yi ) ∈ l q ,
l p -norm ϕy (x) = ∑i xi yi
´1
C (I), I = [0, 1] with signed Borel ϕµ ( f ) = 0 f (x) dµ (x), ∀ f ∈ C (I) .
max-norm measures µ on I, of
bounded variation
C∞ (R), system of E 0 all Schwartz ϕD ( f ) = D applied to f , f ∈ C∞ (R) .

seminorms distributions D on R
of compact support
Table 1.1: Examples of dual spaces.
Examples of Banach spaces are (i) l p , 1 ≤ p ≤ ∞, and (ii) L p (µ) where µ is a

positive measure on some given measure space; details below.
• l p : all p-summable sequences.

A sequence x = (xk )k∈N is in l p iff ∑k∈N |xk | p < ∞, and then
!1
p
p
kxk p := ∑ |xk | .
k∈N
• L p : all p-integrable function w.r.t. some fixed measure µ.

Let F : R → R be monotone, i.e., x ≤ y =⇒ F (x) ≤ F (y); then there is a
Borel measure µ on R (see [Rud87]) such that µ ((x, y]) = F (y) − F (x); and
´
ϕdµ will be the limit of the Stieltjes sums:
∑ ϕ (xi ) (F (xi+1 ) − F (xi )) , where x1 < x2 < · · · < xn .

i
1.3 Transfinite Induction (Zorn and All That) 27
´ p
We say that ϕ ∈ L p (µ) iff R |ϕ| dµ is well defined and finite; then
ˆ 1
p
p
kϕk p = |ϕ (x)| dµ (x) .
R
´ ´
(By Stieltjes integral, |ϕ| p dµ = |ϕ| p dF.)
Remark 1.1. For completeness of l p and of L p (µ), see [Rud87].
Remark 1.2. We gave the definition of L p (µ) in the case where µ = dF, but it ap-
plies more generally.
The following result is basic in the subject.
Theorem 1.1 (Banach). Let E be a Banach space, and let x ∈ E\ {0}, then there is
a ϕ ∈ E ∗ such that ϕ (x) = kxk.
Remark 1.3. In the examples (i) and (ii) above, i.e., l p or L p (µ), it is possible to
identify the needed elements in E ∗ . But the power of Theorem 1.1 is that it yields
existence for all Banach spaces, i.e., when E is given only by the axioms from
Definitions 1.1-1.2.
1.3 Transfinite Induction (Zorn and All That)
Let (X, ≤) be a partially ordered set. By partial ordering, we mean a binary relation
“≤” on the set X, such that (i) x ≤ x; (ii) x ≤ y and y ≤ x implies x = y; and (iii)
x ≤ y and y ≤ z implies x ≤ z.
A subset C ⊂ X is said to be a chain, or totally ordered, if x, y ∈ C implies that
either x ≤ y or y ≤ x. Zorn’s lemma says that if every chain has a majorant then there
exists a maximal element in X.
Theorem 1.2 (Zorn). Let (X, ≤) be a partially ordered set. If every chain C in X
has a majorant (or upper bound), then there exists an element m in X so that x ≥ m
implies x = m.
An illuminating example of a partially ordered set is the binary tree model (Figs
1.1-1.2). Another example is when X is a family of subsets of a given set, partially
ordered by inclusion.
m1 m2
m3
m5 m4
m6
Fig. 1.1: Finite Tree (natural order on the set of vertices). Examples of maximal
elements: m1 , m2 , . . .
All finite words in the alphabet {0, 1}, continued indefinitely.
Fig. 1.2: Infinite Tree (no maximal element!)
Zorn’s lemma lies at the foundation of set theory. It is in fact an axiom and is
equivalent to the axiom of choice, and to Hausdorff’s maximality principle.
Theorem 1.3 (Hausdorff Maximality Principle). Let (X, ≤) be a partially ordered
set, then there exists a maximal totally ordered subset L in X.
The axiom of choice is equivalent to the following statement on infinite product,
which itself is extensively used in functional analysis.
Theorem 1.4 (axiom of choice). Let Aα be a family of nonempty sets indexed by
α ∈ I. Then the infinite Cartesian product

Ω = ∏ Aα = ω : I → ∪α∈I Aα ω (α) ∈ Aα
α∈I
is nonempty.
The point of using the axiom of choice is that, if the index set is uncountable, there
is no way to verify whether (xα ) is in Ω , or not. It is just impossible to check for
each α that xα is contained in Aα .
1.4 Basics of Hilbert Space Theory 29
In case the set is countable, we simply apply the down to earth standard induction.
Note that the standard mathematical induction is equivalent to the Peano’s axiom:
Every nonempty subset of the set of natural number has a unique smallest element.
The power of transfinite induction is that it applies to uncountable sets as well.
In applications, the key of using the transfinite induction is to cook up in a clear
way a partially ordered set, so that the maximum element turns out to be the ob-
ject to be constructed. Examples include Hahn-Banach extension theorem, Krein-
Milman’s theorem on compact convex set, existence of orthonormal bases in Hilbert
space, Tychnoff’s theorem on infinite Cartesian product of compact spaces (follows
immediately from the axiom of choice.)
Theorem 1.5 (Tychonoff). Let Aα be a family of compact sets indexed by α ∈ I.

Then the infinite Cartesian product ∏α Aα is compact with respect to the product
topology.
We will apply transfinite induction (Zorn’s lemma) to show that every infinite di-
mensional Hilbert space has an orthonormal basis (ONB).
1.4 Basics of Hilbert Space Theory
Definition 1.4. Let X be a vector space over C. A norm on X is a mapping k·k : X →

C such that
• kcxk = |c| kxk, c ∈ C, x ∈ X;

• kxk ≥ 0; kxk = 0 implies x = 0, for all x ∈ X;
• kx + yk ≤ kxk + kyk, for all x, y ∈ X.
Definition 1.5. Let (X, k·k) be a normed space. X is called a Banach space if it is
complete with respect to the induced metric
d (x, y) := kx − yk , x, y ∈ X.
Definition 1.6. Let X be vector space over C. An inner product on X is a function

h·, ·i : X × X → C so that for all x, y ∈ H , and c ∈ C, we have
• hx, ·i : X → C is linear (linearity)

• hx, yi = hy, xi (conjugation)
• hx, xi ≥ 0; and hx, xi = 0 implies x = 0 (positivity)
Remark 1.4. The abstract formulation of Hilbert was invented by von Neumann in
1925. It fits precisely with the axioms of quantum mechanics (spectral lines, etc.)
A few years before von Neumann’s formulation, Heisenberg translated Max Born’s
quantum mechanics into mathematics.
Lemma 1.2 (Cauchy-Schwarz). 1 Let (X, h·, ·i) be an inner product space, then
|hx, yi|2 ≤ hx, xi hy, yi , ∀x, y ∈ X. (1.2)
Proof. By the positivity axiom in the definition of an inner product, we see that
* +
2 2 2
∑ ci c j xi , x j = ∑ ci xi , ∑ c j x j ≥ 0, ∀c1 , c2 ∈ C;
i, j=1 i=1 j=1
i.e., the matrix " #

hx1 , x1 i hx1 , x2 i
hx2 , x1 i hx2 , x2 i
is positive definite. Hence the above matrix has positive determinant, and (1.2) fol-
lows.
Corollary 1.1. Let (X, h·, ·i) be an inner product space, then
p
kxk := hx, xi, x ∈ X (1.3)
defines a norm.
Proof. It suffices to check the triangle inequality (Definition 1.6). For all x, y ∈ X,
we have
kx + yk2 = hx + y, x + yi
= kxk2 + kyk2 + 2ℜ {hx, yi}
≤ kxk2 + kyk2 + 2 kxk kyk (by (1.2))
= (kxk + kyk)2
1 Hermann Amandus Schwarz (1843 - 1921), German mathematician, contemporary of Weier-
strass, and known for his work in complex analysis. He is the one in many theorems in books on
analytic functions. There are other two “Schwartz” (with a “t”):
Laurent Schwartz (1915 - 2002), French mathematician, Fields Medal in 1950 for his work of
distribution theory.
Jack Schwartz (1930 - 2009), American mathematician, author of the famous book “Linear
Operators”.
and the corollary follows.
Definition 1.7. An inner product space (X, h·, ·i) is called a Hilbert space if X is
complete with respect to the metric
d (x, y) = kx − yk , x, y ∈ X;
where the RHS is given by (1.3).
Remark 1.5. An extremely useful method to build Hilbert spaces is the GNS con-
struction. For details, see chapter 4. The idea is to start with a positive definite
function ϕ : X × X → C, defined on an arbitrary set X. We say ϕ is positive definite,
if for all n ∈ N,
n
∑ ci c j ϕ (xi , x j ) ≥ 0 (1.4)
i, j=1
for all system of coefficients c1 , . . . , cn ∈ C, and all x1 , . . . , xn ∈ X.

Given ϕ, set
( )
H0 := ∑ cx δx : x ∈ X, cx ∈ C = spanC {δx : x ∈ X} ,
finite
and define a sesquilinear form on H0 by
∑ cx δx , ∑ dy δy ϕ
:= ∑ cx dy ϕ (x, y) .
Note that
2
∑ cx δx ϕ
:= ∑ cx δx , ∑ cx δx ϕ
= ∑ cx cy ϕ (x, y) ≥ 0
x,y
by assumption. (All summations are finite.)

However, h·, ·iϕ is in general not an inner product since the strict positivity axiom
may not be satisfied. Hence one has to pass to a quotient space by letting
n o
N= f ∈ H0 h f , f iϕ = 0 ,
and set H := completion of the quotient space H0 /N with respect to k·kϕ . (The
fact that N is really a subspace follows from (1.2).) H is a Hilbert space.
Corollary 1.2. Let X be a set, and let ϕ : X × X → C be a function. Then ϕ is

positive definite if and only if there is a Hilbert space H = Hϕ , and a function
Φ : X → H such that
ϕ (x, y) = hΦ (x) , Φ (y)iH (1.5)
for all (x, y) ∈ X × X, where h·, ·iH denotes the inner product in H .
Given a solution Φ satisfying (1.5), then we say that H is minimal if
H = span {Φ (x) : x ∈ X} . (1.6)
Given two minimal solutions, Φi : X → Hi , i = 1, 2 (both satisfying (1.5)); then there

is a unitary isomorphism U : H1 → H2 such that
UΦ1 (x) = Φ2 (x) U, ∀x ∈ X. (1.7)
Proof. These conclusions follow from Remark 1.5, and the definitions. (The miss-
ing details are left as an exercise to the student.)
Remark 1.6. It is possible to be more explicit about choice of the pair (Φ, H ) in
Corollary 1.2, where ϕ : X × X → C is a given positive definite function: We may
in fact choose H to be L2 (Ω , F, P) where P = Pϕ depends on ϕ, and (Ω , F, P) is
a probability space.
Example 1.1. In Remark 1.6, take X = [0, ∞) = R+ ∪ {0}, and set
ϕ (s,t) = s ∧ t = min (s,t) ,
see Fig 1.3. In this case, we may then take Ω = C (R) = all continuous functions on
R, and Φt (ω) := ω (t), t ∈ [0, ∞), ω ∈ C (R).
And in this case, further the measure Pϕ on the sigma-algebra in C (R) generated
by cylinder-sets is the Wiener-measure; and Φ on L2 (C (R) , cyl, P) is the standard
Brownian motion, i.e., Φ : [0, ∞) → L2 (C (R) , P) is a Gaussian process with
ˆ
EP (Φ (s) Φ (t)) = Φs (ω) Φt (ω) dP (ω) = s ∧ t.
C(R)
We refer to chapter 6 for full details on Brownian motion.
Remark 1.7. We see from the proof of Lemma 1.2 that the Cauchy-Schwarz inequal-
ity holds for all positive definite functions.
s jH×, t)
t
s
s 0 t
(a) ϕ (s,t) = s ∧ t (b) with t fixed
Fig. 1.3: Covariance function of Brownian motion.
1.4.1 Orthonormal Bases
Definition 1.8. Let H be a Hilbert space. A family of vectors {uα } in H is said to

be an orthonormal basis if
1. uα , uβ H = δαβ and
2. span {uα } = H . (Here “span” means “closure of the linear span.”)
We are ready to prove the existence of orthonormal basis for any Hilbert space.
The key idea is to cook up a partially ordered set satisfying all the requirements in
the transfinite induction, so that the maximum elements turns out to be an orthonor-
mal basis. Notice that all we have at hands are the abstract axioms of a Hilbert space,
and nothing else. Everything will be developed out of these axioms.
Theorem 1.6. Every Hilbert space H has an orthonormal basis.
To start out, we need the following lemmas.
Lemma 1.3. Let H be a Hilbert space and S ⊂ H . Then the following are equiv-
alent:
1. x ⊥ S implies x = 0
2. span{S} = H
Proof. Now we prove Theorem 1.6. If H is empty then the proof is done. Other-
wise, let u1 ∈ H . If ku1 k 6= 1, it can be normalized by u1 / ku1 k. Hence we may
assume ku1 k = 1. If span{u1 } = H the proof finishes again, otherwise there exists
/ span{u1 }. By Lemma 1.3, we may assume ku2 k = 1 and u1 ⊥ u2 . By induction,
u2 ∈
there exists a collection S of orthonormal vectors in H .
Let P (S) be the set of all orthonormal sets partially order by set inclusion. Let
C ⊂ P(S) be any chain and let M :=
S
E∈C E. M is clearly a majorant of C.
In fact, M is in the partially ordered system. For if x, y ∈ M, there exist Ex and Ey
in C so that x ∈ Ex and y ∈ Ey . Since C is a chain, we may assume Ex ≤ Ey . Hence
x, y ∈ Ey , and so x ⊥ y. This shows that M is in the partially ordered system and a
majorant.
By Zorn’s lemma, there exists a maximum element m ∈ P (S). It suffices to show
that the closed span of m is H . Suppose this is false, then by Lemma 1.3 there exists
x ∈ H so that x ⊥ span {m}.
Since m ∪ {x} ≥ m and m is maximal, it follows that x ∈ m, which implies x ⊥ x.
By the positivity axiom of the definition of Hilbert space, x = 0.
Corollary 1.3. Let H be a Hilbert space, then H is isomorphic to the l 2 space of

the index set of an ONB of H . Specifically, given an ONB {uα }α∈J in H , where J
is some index set, then
v= ∑ huα , vi uα , and (1.8)
α∈J
kvk2 = ∑ |huα , vi|2 , ∀v ∈ H . (1.9)

α∈J
((1.9) is called the Parseval identity.)
Remark 1.8. In summary, the correspondence H ←→ l 2 (index set of an ONB) is

functorial, and an isomorphism. Hence, there seems to be just one Hilbert space. But
this is misleading, because numerous interesting realizations of an abstract Hilbert
space come in when we make a choice of the ONB. The question as to which Hilbert
space to use is equivalent to a good choice of an ONB; in L2 (R), for example, a
wavelet ONB.
Example 1.2 (Wavelets). Suppose H is separable (i.e., having a countable ONB),

for instance let H = L2 (R). Then H ∼= l 2 (N) ∼
= l 2 (N × N), and it follows that
potentially we could choose a doubly indexed basis {ψ j,k : j, k ∈ N} for L2 (R). It
turns out that this is precisely the setting of wavelet basis! What’s even better is that
in the l 2 space, there are all kinds of diagonalized operators, which correspond to
selfadjoint (or normal) operators in L2 . Among these operators in L2 , we single out
the following two:
Uj
scaling: f (x) −−→ 2 j/2 f 2 j x

V k
translation: f (x) −−→ f (x − k)
for all j, k ∈ Z. However, U j and Vk are NOT diagonalized simultaneously though.

See below for details!
1.4.2 Bounded Operators in Hilbert Space
Definition 1.9. A bounded operator in a Hilbert space H is a linear mapping T :

H → H such that
kT k := sup {kT xkH : kxkH ≤ 1} < ∞.
We denote by B (H ) the algebra of all bounded operators in H .
Setting (ST ) (v) = S (T (v)), v ∈ H , S, T ∈ B (H ), we have kST k ≤ kSk kT k.
Lemma 1.4 (Riesz). There is a bijection H 3 h 7−→ lh between H and the space
of all bounded linear functionals on H , where
lh (x) := hh, xi , ∀x ∈ H , and

klh k := sup {|l (x)| : kxkH ≤ 1} < ∞.
Moreover, klh k = khk.
Corollary 1.4. For all T ∈ B (H ), there exists a unique operator T ∗ ∈ B (H ),

called the adjoint of T , such that
hx, Tyi = hT ∗ x, yi , ∀x, y ∈ H ;
and kT ∗ k = kT k.
Proof. Let T ∈ B (H ), then it follows from the Cauchy-Schwarz inequality that
|hx, Tyi| ≤ kT xk kyk ≤ kT k kxk kyk .
Hence the mapping y 7−→ hx, Tyi is a bounded linear functional on H . By Riesz’s
theorem, there exists a unique hx ∈ H , s.t. hx, Tyi = hhx , yi, for all y ∈ H . Set
T ∗ x := hx . One checks that T ∗ linear, bounded, and in fact kT ∗ k = kT k.
Exercise 1.1. Let T ∈ B (H ), then prove

kT ∗ T k = kT k2 . (1.10)
Definition 1.10. Let T ∈ B (H ). Then,
• T is normal if T T ∗ = T ∗ T
• T is selfadjoint if T = T ∗
• T is unitary is T ∗ T = T T ∗ = IH (= the identity operator)
• T is a (selfadjoint) projection if T = T ∗ = T 2
For T ∈ B (H ), we may write
1
R= (T + T ∗ )
2
1
S = (T − T ∗ )
2i
then both R and S are selfadjoint, and
T = R + iS.
This is similar the to decomposition of a complex number into its real and imaginary
parts. Notice also that T is normal if and only if R and S commute. Thus the study
of a family of normal operators is equivalent to the study of a family of commuting
selfadjoint operators.
Theorem 1.7. Let H be a Hilbert space. There is a one-to-one correspondence

between selfadjoint projections and closed subspaces of H . (See Fig 1.4.)
Proof. Let P be a selfadjoint projection in H , i.e., P2 = P = P∗ . Then
PH = {x ∈ H : Px = x}
is a closed subspace in H . Let P⊥ := I − P be the complement of P, so that

n o
P⊥ H = x ∈ H : P⊥ x = x = {x ∈ H : Px = 0} .
Since PP⊥ = P(1 − P) = P − P2 = P − P = 0, we have PH ⊥ P⊥ H .

Conversely, let W ( H be a closed subspace. Note the following “parallelogram
law” holds:
kx + yk2 + kx − yk2 = 2(kxk2 + kyk2 ), ∀x, y ∈ H . (1.11)

Let x ∈ H \W , and set

d := inf kx − wk .
w∈W
By definition, there exists a sequence {wn } in W so that kwn − xk → 0 as n → ∞.

Applying (1.11) to x − wn and x − wm , we get

k(x − wn ) + (x − wm )k2 + k(x − wn ) − (x − wm )k2 = 2 kx − wn k2 + kx − wm k2 ;
which simplifies to
2
wn + wm
4 x− + kwn − wm k2 = 2 kx − wn k2 + kx − wm k2 . (1.12)
2
Notice here all we require is 1

2 (wn + wm ) ∈ W , hence the argument carries over
if we simply assume W is a convex subset in H . We conclude from (1.12) that
kwn − wm k → 0, and so {wn } is a Cauchy sequence. Since H is complete, there is
a unique limit, Px := limn→∞ wn . See Fig 1.4.
We leave the rest to the reader. See, e.g., [Rud73], [Nel69, p.62].
H x
W
wn-1
wn
o Px =lim wn
Fig. 1.4: kx − Pxk = inf {kx − wk : w ∈ W }
The Gram-Schmidt Process
Every Hilbert space has an ONB, but it does not mean in practice it is easy to
select one that works well for a particular problem. The Gram-Schmidt orthogonal-
ization process was developed a little earlier than von Neumann’s formulation of
abstract Hilbert space. It is an important tool to get an orthonormal set out of a set
of linearly independent vectors.
Lemma 1.5 (Gram-Schmidt). Let {un } be a sequence of linearly independent vec-

tors in H then there exists a sequence {vn } of unit vectors so that hvn , vk i = δn,k
and span {uk } = span {vk }.
Proof. Given {un } as in the statement of the lemma, we set
u1
v1 = .
ku1 k
Suppose we have constructed the orthonormal set Fn := {v1 , . . . , vn }, and let PF be

the projection on Fn . For the induction step, we set
Proof.
x − PFn x
vn+1 := , n = 1, 2, . . . (1.13)
kx − PFn xk
Note the LHS in (1.13) a unit vector, and orthogonal to PFn H .
The formula for PFn is
n
PFn = ∑ |vk ih vk | .
k=1
Remark 1.9. If H is non separable, the standard induction does not work, and the
transfinite induction is needed.
Example 1.3. Let H = L2 [0, 1]. The polynomials {1, x, x2 , . . .} are linearly indepen-
dent in H , for if
n
∑ ck xk = 0
k=1
then as an analytic function, the left-hand-side must be identically zero. By Stone-

Weierstrass theorem, span{1, x, x2 , . . .} is dense in C([0, 1]) under the k·k∞ norm.
Since k·kL2 ≤ k·k∞ , it follows that span{1, x, x2 , . . .} is also dense in H .
By the Gram-Schmidt process, we get a sequence {Vn }∞
n=1 of finite dimensional
subspaces in H , where Vn has an orthonormal basis {h0 , . . . , hn−1 }, so that
Vn = span{1, x, . . . , xn−1 }
= span {h0 , h1 , . . . , hn−1 } .
Details: Set h0 = 1 = constant function, and
xn+1 − Pn xn+1
hn+1 := , n ∈ N.
kxn+1 − Pn xn+1 k
Then the set {hn : n ∈ N ∪ {0}} is an ONB in H .
Example 1.4. Let H = L2 [0, 1]. Consider the set of complex exponentials
i2πnx
e : n ∈ N ∪ {0} ,
or equivalently, one may also consider
{1, cos 2πnx, sin 2πnx : n ∈ N} .
This is already an ONB in H and leads to Fourier series.
In the next example we construct the Haar wavelet.

Definition 1.11. A function ψ ∈ L2 (R) is said to generate a wavelet if
ψ j,k (x) = 2 j/2 ψ 2 j x − k , j, k ∈ Z

(1.14)
is an ONB in L2 (R).
Note with the normalization in (1.14) we get
ˆ ˆ
2
ψ j,k (x) dx = |ψ (x)|2 dx, ∀ j, k ∈ Z.
R R
Example 1.5 (Haar wavelet). Let H = L2 (0, 1), and let ϕ0 be the characteristic
function of the unit interval [0, 1]. ϕ0 is called a scaling function. Define
ϕ1 := ϕ0 (2x) − ϕ0 (2x − 1), and

ψ j,k := 2 j/2 ϕ1 (2 j x − k), j, k ∈ Z.
Claim: ψ j,k : j, k ∈ Z is an orthonormal set (in fact, an ONB) in L2 (0, 1).

Proof. Fix k, if j1 6= j2 , then ψ j1 ,k and ψ j2 ,k are orthogonal since their supports are
nested. For fixed j, and k1 6= k2 , then ψ j,k1 and ψ j,k2 have disjoint supports, and so
they are also orthogonal. See Fig 1.5.
Exercise 1.2. Let Mt : L2 [0, 1] → L2 [0, 1] be the operator of multiplication by t.

Compute the matrix of Mt under the Haar wavelet basis in Example 1.5.
This problem is taken from Joel Anderson, who showed that A = A∗ implies that
A = D + K, where D is a diagonal operator and K is a compact perturbation. See
[And74] for details.
Ψj k
1 1
Ψj k
2 2
Fig. 1.5: Haar wavelet.
Exercise 1.3. Let z be a complex number, and P be a selfadjoint projection. Show

that U(z) = zP + (I − P) is unitary if and only if |z| = 1.
Hint: U(z)U(z)∗ = U(z)∗U(z) = |z|2 P + (I − P), so U (z) is unitary ⇐⇒ |z| = 1.
1.5 Dirac’s Notation
P.A.M Dirac was every efficient with notations, and he introduced the “bra-ket”
vectors [Dir35, Dir47].
Definition 1.12. Let H be a Hilbert space with inner product h·, ·i. We denote by
“bra” for vectors hx| and “ket” for vectors |yi, for x, y ∈ H .
With Dirac’s notation, our first observation is the following lemma.
Lemma 1.6. Let v ∈ H be a unit vector. The operator x 7→ hv, xi v can be written as
Pv = |v ih v| ,
i.e., a “ket-bra” vector. And Pv is a rank-one selfadjoint projection.
Proof. First, we see that
Pv2 = (|v ih v|) (|v ih v|) = |vi hv, vi hv| = |v ih v| = P.
Also, if x, y ∈ H then
D E
hx, Pv yi = hx, vi hv, yi = hx, viv, y = hhv, xi v, yi = hPv x, yi
1.5 Dirac’s Notation 41
so Pv = Pv∗ .
Corollary 1.5. Let F be a finite set of orthonormal vectors in H , then
PF := ∑ |vi ih vi |
vi ∈F
is the selfadjoint projection on F.
Proof. Indeed, we have
PF2 =

∑ (|vi ih vi |) v j ih v j = ∑ |vi ih vi | = PF
vi ,v j ∈F vi ∈F
PF∗ = ∑ (|vi ih vi |)∗ = ∑ |vi ih vi | = PF .

vi ∈F vi ∈F
Remark 1.10. More generally, any rank-one operator can be written in Dirac nota-
tion as
|u ih v| : H 3 x 7−→ hv, xi u ∈ H .
With the bra-ket notation, it is easy to verify that the set of rank-one operators forms
an algebra, which easily follows from the fact that
(|v1 ih v2 |) (|v3 ih v4 |) = hv2 , v3 i |v1 ih v4 | .
The moment that an orthonormal basis is selected, the algebra of operators on H

will be translated to the algebra of matrices (infinite). See Lemma 1.8.
Lemma 1.7. Let {uα }α∈J be an ONB in H , then we may write
IH = ∑ |uα ih uα | .
α∈J
Proof. This is equivalent to the decomposition
v= ∑ huα , vi uα , ∀v ∈ H .
α∈J
A selection of ONB makes a connection to the algebra operators acting on H

and infinite matrices. We check that using Dirac’s notation, the algebra of operators
really becomes the algebra of infinite matrices.
Lemma 1.8. MAB = MA MB , i.e., (MAB )i j = ∑k (MA )ik (MB )k j .
Proof. By AB we mean the operator in B (H ) given by (AB) (u) = A (B (u)), ∀u ∈

H.
Pick an ONB {ui } in H , A, B ∈ B(H ). We denote by MA = Ai j := ui , Au j
the matrix of A under the ONB. We compute ui , ABu j .
(MA MB )i j = ∑ Aik Bk j = ∑ hui , Auk i uk , Bu j

k k
= ∑ hA∗ ui , uk i uk , Bu j
k
= A∗ ui , Bu j
= ui , ABu j
= (MAB )i j
where we used that I = ∑ |ui ih ui |.
Exercise 1.4. Let H be a Hilbert space, and A a set which indexes a fixed ONB
{vα }α∈A . Now define T : H → l 2 (A), by
(T h) (α) := hvα , hi , ∀α ∈ A, h ∈ H .
Show that T is unitary, and onto l 2 (A), i.e.,
T T ∗ = Il 2 (A) , and
T ∗ T = IH .
1.5.1 Three Norm-Completions
Let H be a fixed Hilbert space, infinite-dimensional in the discussion below.

Let
F R (H ) = {all finite-rank operators H −→ H }

= span {|v ih w| : v, w ∈ H }
where |v ih w| denotes the Dirac ket-bra-operator.

Definition 1.13. On F R (H ) we introduce the following three norms: the uniform

norm (UN), the trace-norm (TN), and the Hilbert-Schmidt norm as follows:
• (UN) For all T ∈ F R (H ), set
kT kUN := sup {kT vk : kvk = 1} .
• (TN) Set √
kT kT N := trace T ∗T ;
• (HSN) Set
1
kT kHSN = (trace (T ∗ T )) 2 .
Theorem 1.8. The completion of F R (H ) with respect to k·kUN , k·kT N , and k·kHSN
are the compact operators, the trace-class operators, and the Hilbert-Schmidt
operators.
We will prove, as a consequence of the Spectral Theorem (Chapter 3), that the
k·kUN - completion agrees with the usual definition of the compact operators.
Definition 1.14. Let T ∈ B (H ), then we say that T is compact iff (Def.) T (H1 )
is weak∗ -compact in H , where H1 := v ∈ H kvk ≤ 1 .

A similar remark applies to the other two Banach spaces of operators. The
Hilbert-Schmidt operators forms a Hilbert space.
Exercise 1.5. If dim H = ∞, show that the identity operator IH is not compact.
Definition 1.15. The following is useful in working through the arguments above.
Lemma 1.9. Let v, w ∈ H , then
trace (|v ih w|) = hv, wiH .
Proof. Introduce an ONB {uα } in H , and compute:
trace (|v ih w|) = ∑ huα , |v ih w| uα iH

α
= ∑ huα , viH hw, uα iH
α
= hv, wiH
where we used Parseval in the last step of the computation.

Exercise 1.6. Let T ∈ F R (H ), and let the three norms be as in Definition 1.13.
Then show that
kT kUN ≤ kT kHSN ≤ kT kT N ; (1.15)
and conclude the following inclusions:
{Trace-class operators} ⊂ {Hilbert-Schmidt operators}

⊂ {compact operators} .
Remark 1.11. In the literature, the following notation is often used for the three
norms in Definition 1.13: 


kT kUN = kT k∞

kT kT N = kT k1 (1.16)


kT k
HSN = kT k2

Note that if T is a diagonal operator, T = ∑k xk |uk ih uk | in some ONB {uk }, then

the respective norms are kxk∞ , kxk1 , and kxk2 .
With the notation in (1.16), the inequalities (1.15) now take the form
kT k∞ ≤ kT k2 ≤ kT k1 , T ∈ F R (H ) .
1.5.2 Connection to Quantum Mechanics
Much of the motivation for the axiomatic approach to the theory of linear opera-
tors in Hilbert space dates back to the early days of quantum mechanics (Planck,
Heisenberg, and Schrödinger), but in the form suggested by J. von Neumann. (von
Neumann’s formulation is the one now adopted by most books on functional anal-
ysis.) Here we will be brief, as a systematic and historical discussion is far beyond
our present scope. Suffice it to mention here that what is known as the "matrix-
mechanics" of Heisenberg takes the form infinite by infinite matrices with entries
representing, in turn, transition probabilities, where "transition" refers to "jumps"
between energy levels. By contrast to matrices, in Schrödinger’s wave mechanics,
the Hilbert space represents wave solutions to Schrödinger’s equation. Now this
entails the study of one-parameter groups of unitary operators in H . In modern
language, with the two settings we get the dichotomy between the case when the
Hilbert space is an l 2 space (i.e., a l-sequence space), vs the case of Schrödinger

when H is an L2 -space of functions on phase-space. (See Appendix.)
In both cases, the observables are represented by families of selfadjoint operators
in the respective Hilbert spaces. For the purpose here, we pick the pair of selfadjoint
operators representing momentum (denoted P) and position (denoted Q). In one
degree of freedom, we only need a single pair. The canonical commutation relation
h
is PQ − QP = iI; or PQ − QP = ih̄I where h̄ = 2π is Planck’s constant.
A few years after the pioneering work of Heisenberg and Schrödinger, J. von
Neumann and M. Stone proves that the two approaches are unitarily equivalent,
hence they produce the same “measurements.” In modern lingo, the notion of mea-
surement take the form of projection valued measures, which in turn are the key
ingredient in the modern formulation of the spectral theorem for selfadjoint, or nor-
mal, linear operators in Hilbert space. Because of dictates from physics, the “inter-
esting” operators, such as P and Q are unbounded.
The first point we discuss about the pair of operators P and Q is non-commutativity.
As is typical in mathematical physics, non-commuting operators will satisfy condi-
tions on the resulting commutators. In the case of P and Q, the commutation re-
lation is called the canonical commutation relation; see below. For reference, see
[Dir47, Hei69, vN31, vN32d].
Quantum mechanics was born during the years from 1900 to 1913. It was created
to explain phenomena in black body radiation, hydrogen atom, where a discrete
pattern occurs in the frequencies of waves in the radiation. The radiation energy
E = ν h̄, with h̄ being the Plank’s constant. Classical mechanics runs into trouble.
During the years of 1925~1926, Heisenberg found a way to represent the energy
E as a matrix, so that the matrix entries v j , Evi represents the transition proba-
bility from energy i to energy j. A fundamental relation in quantum mechanics is
the commutation relation satisfied by the momentum operator P and the position
operator Q, where
PQ − QP = iI. (1.17)
Heisenberg represented the operators P, Q by infinite matrices, although his solution

to (1.17) is not real matrices, and not finite matrices.
(1.17)
Fact 1.1 Eq. (1.17) has no solutions for finite matrices.
Proof. The reason is for matrices, there is a trace operation where

trace(AB) = trace(BA).
This implies the trace on the left-hand-side is zero, while the trace on the RHS is
not.
This shows that there is no finite dimensional solution to the commutation rela-
tion above, and one is forced to work with infinite dimensional Hilbert space and
operators on it. Notice also that P, Q do not commute, and the above commutation re-
lation leads to the uncertainty principle (Hilbert, Max Born, von Neumann worked
out the mathematics), which says that the statistical variance 4P and 4Q satisfy
4P4Q ≥ h̄/2 . We will come back to this later.
We will show that non-commutativity always yields “uncertainty.”
However, Heisenberg [vN31, Hei69] found his “matrix” solutions by tri-diagonal
∞ × ∞ matrices, where
 
0 1
 √ 
1 0 2 
 √ √ 
1  2 0 3 
P= √   (1.18)
2 √ .. 

 3 0 .
.. ..
 
. .
and  
0 1
 √ 
 −1 0 2 
 √ √ 
1  − 2 0 3 
Q= √   (1.19)
i 2
 √ .. 
− 3 0 .


.. ..
 
. .
the complex i in front of Q is to make it selfadjoint.
Exercise 1.7. Using matrix multiplication for ∞ × ∞ matrices verify directly that the
two matrices P and Q satisfy PQ − QP = iI, where I is the identity matrix in l 2 (N0 ),
i.e., (I)i j = δi j . Hint: use the rules in Lemma 1.8.
Exercise 1.8. Raising and lowering operators (non-commutative complex variables.)

Set A∓ := P ± iQ; and show that the matrix representation for these operators is as
follows:
 
01 0 0 0 ···
. √ 
 .. 0 2 0 0 · · ·
 
√ . . √ 
A− = 2  .. .. 0 3 0 · · ·

 .. .. .. √
 

. . . 0
 4 · · ·
.. .. .. .. . . ..
. . . . . .
and  
0 0 ··· ··· ··· ···
 
1 0 0 · · · · · · · · ·
√ 0 √2 0 0
 
A+ = 2  · · · · · ·
.
.
. √ 
. 0 3 0 0 · · ·


.. .. √ .. ..

. . 0 4 . .
In other words, the raising operator A+ is a sub-banded matrix, while the lowering
operator A− is a supper-banded matrix. Both A+ and A− has 0s down the diagonal.
Further, show that
 
1 0 ··· ··· ··· ···
0 2 0 · · · · · · · · ·
 
. 
.
· · · · · ·

A− A+ = 2  . 0 3 0
 .. ..
 

. . 0 4
 0 · · · 
.. .. .. .. .. ..
. . . . . .

..
i.e., a diagonal matrix, with the numbers N down the diagonal inside . ; and
 
0 0 ··· ··· ··· ··· ···
 
0 1 0 · · · · · · · · · · · ·
 
 .. 
. 0 2 0 ··· · · · · · ·
 
A+ A− = 2 
 ...
;
 0 3 0 · · · · · ·

. .. 
.
. . 0 4 0 · · ·

.. .. .. .. .. ..
 
. . . . . .
so that
1
[A− , A+ ] = I.
2
Remark 1.12. In the canonical ONB (en )∞ 2

n=0 in l (N0 ), we have the following rep-
resentations of the two operators A± :
√ √
A+ en = 2 n + 1en+1 ; n = 0, 1, 2, . . . , and
√ √
A− en = 2 nen−1 ; n = 1, 2, . . . ,
A− e0 = 0.
The vector e0 is called the ground state, or vacuum vector.
Remark 1.13. The conclusion of the discussion above is that the Heisenberg com-
mutation relation (1.17) for pairs of selfadjoint operators has two realizations, one
in L2 (R), and the other in l 2 (N).
1 d
In the first one we have (P f ) (x) = i dx f , and(Q f ) (x) = x f (x), for f ∈ S ⊂
L2 (R), where S denotes the Schwartz test-function subspace in L2 (R).
The second realization is by ∞ × ∞ matrices, and it is given in detail above. In
section 7.3 we shall return to the first realization.
The Stone-von Neumann uniqueness theorem implies the two solutions are uni-
tarily equivalent; see chapter 7.
1.5.3 Probabilistic Interpretation of Parseval in Hilbert Space
Case 1. Let H be a complex Hilbert space, and let {uk }k∈N be an ONB, then Par-
seval’s formula reads:
hv, wiH = ∑ hv, uk i huk , wi , ∀v, w ∈ H . (1.20)

k∈N
Translating this into a statement about transition probabilities for quantum states,
v, w ∈ H , with kvkH = kwkH = 1, we get
Prob (v → w) = ∑ Prob (v → uk ) Prob (uk → w) . (1.21)

k∈N
See Fig 1.6. The states v and w are said to be uncorrelated iff (Def.) they are orthog-
onal.
Fix a state w ∈ H , then
kwk2 = ∑ |huk , wi|2 = 1.

k∈N
The numbers |hui , wi|2 represent a probability distribution over the index set, where
|huk , wi|2 is the probability that the quantum system is in the state |uk i.
Fig. 1.6: Transition of quantum-states
Case 2. If P : B (R) → Proj (H ) is a projection valued measure, we get the anal-

ogous assertions, but with integration, as opposed to summation. In this case eq.
(1.20) holds the following form:
ˆ
hv, wiH = hv, P (dλ ) wiH ; (1.22)
R
and for v = w, it reads: ˆ

2
kvkH = kP (dλ ) vk2H . (1.23)
R
Recall the other axioms of P (·) are:
1. P (A) = P (A)∗ = P (A)2 , ∀A ∈ B (R).

2. P (·) is countably additive on the Borel subsets of R, i.e.,
∑ P (A j ) = P (∪ j A j )
j
where A j ∈ B (R), Ai ∩ A j = 0,
/ i 6= j.
3. P (A ∩ B) = P (A) P (B), ∀A, B ∈ B (R).
1.6 The Lattice Structure of Projections
Notation: In this section we will denote projections P, Q, etc.

von Neumann invented the notion of abstract Hilbert space in 1928 as shown in
one of the earliest papers.2 His work was greatly motivated by quantum mechanics.
In order to express quantum mechanics logic operations, he created lattices of pro-
jections, so that everything we do in set theory with set operation has a counterpart
in the operations of projections. See Table 1.3.
SETS CHAR PROJECTIONS DEFINITIONS
A∩B χA χB P∧Q PH ∩ QH
A∪B χA∪B P∨Q span{PH ∪ QH }
A⊂B χA χB = χA P≤Q PH ⊂ QH
A1 ⊂ A2 ⊂ · · · χAi χAi+1 = χAi P1 ≤ P2 ≤ · · · Pi H ⊂ Pi+1 H
k=1 Pk H
S∞ S∞
k=1 Ak χ∪k A ∨∞
k=1 Pk span{ }
k=1 Pk H
T∞ T∞
k=1 Ak χ∩k Ak ∧∞
k=1 Pk
A×B (χA×X ) (χX×B ) P⊗Q P ⊗ Q ∈ pro j(H ⊗ K )

Table 1.3: Lattice of projections in Hilbert space.
For example, if P and Q are two projections in B (H ), then
PH ⊂ QH (1.24)
m
P = PQ (1.25)
m
P ≤ Q. (1.26)
2 Earlier authors, Schmidt and Hilbert, worked with infinite bases, and ∞ × ∞ matrices.
1.6 The Lattice Structure of Projections 51
This is similar to the following equivalence relation in set theory
A ⊂ B (containment of sets) (1.27)

m
A = A ∩ B. (1.28)
In general, product and sum of projections are not projections. But if PH ⊂ QH

then the product PQ is in fact a projection. Taking adjoint in (1.25) yields P∗ =
(PQ)∗ = Q∗ P∗ = QP. It follows that PQ = QP = P, i.e., containment of subspaces
implies the corresponding projections commute.
During the same time period as von Neumann developed his Hilbert space the-
ory, Lesbegue developed his integration theory which extends the classical Riemann
integral. The monotone sequence of sets A1 ⊂ A2 ⊂ · · · in Lebesgue’s integration
theory also has a counterpart in the theory of Hilbert space.
Lemma 1.10. Let P1 and P2 be orthogonal projections acting on H , then
P1 ≤ P2 ⇐⇒ kP1 xk ≤ kP2 xk , ∀x ∈ H (1.29)

m
P1 = P1 P2 = P2 P1 (1.30)
(see Table 1.3.)
Proof. Indeed, for all x ∈ H , we have
kP1 xk2 = hP1 x, P1 xi = hx, P1 xi = hx, P2 P1 xi ≤ kP1 P2 xk2 ≤ kP2 xk2 .
As a consequence, P1 ≤ P2 ≤ · · · implies that {kPk xk}∞

k=1 is a monotone increas-
ing sequence in R, and the sequence is bounded by kxk, since kPk xk ≤ kxk, for all
k ∈ N. Therefore the sequence {Pk }∞
k=1 converges to P (strongly), in symbols, we
write
∨Pk = lim Pk = P
k
and P really defines a selfadjoint projection. (We use “≤” to denote the lattice op-
eration on projection.) Note the convergence refers to the strong operator topology,
i.e., for all x ∈ H , there exists a vector, which we denote by Px, so that
lim kPk x − Pxk = 0.

k
The examples in section 1.4 using Gram-Schmidt process can now be formulated
in the lattice of projections.
Recall that for a linearly independent subset {uk } ⊂ H , the Gram-Schmidt pro-
cess yields an orthonormal set {vk } ⊂ H , with v1 := u1 / ku1 k, and
un − Pn un
vn+1 := , n = 1, 2, . . . ;
kun − Pn un k
where Pn is the orthogonal projection on the n-dimensional subspace Vn := span {v1 , . . . , vn }.

Note that
[
Vn ⊂ Vn+1 → Vn ∼ Pn ≤ Pn+1 → P
n
Pn⊥ ≥ Pn+1
⊥
→ P⊥ .
is dense in H , then P = I and P⊥ = 0. In lattice notations, we may

S
Assume n Vn
write
∨Pn = sup Pn = I
∧Pn⊥ = inf Pn⊥ = 0.
Lemma 1.11. Let P, Q ∈ Pro j (H ), then
P + Q ∈ Pro j(H ) ⇐⇒ PQ = QP = 0, i.e., P ⊥ Q.
Proof. Notice that

(P + Q)2 = P + Q + PQ + QP (1.31)
and so
(P + Q)2 = P + Q (1.32)
m
PQ + QP = 0. (1.33)
Suppose PQ = QP = 0 then (P + Q)2 = P + Q = (P + Q)∗ , i.e., P + Q ∈

Pro j (H ).
1.6 The Lattice Structure of Projections 53
Conversely, if P + Q ∈ Pro j(H ), then (P + Q)2 = P + Q =⇒ PQ + QP = 0 by

(1.33). Also, (PQ)∗ = Q∗ P∗ = QP, combining with (1.33) yields
(PQ)∗ = QP = −PQ. (1.34)
Then,
(PQ)2 = P (QP) Q = −P(PQ)Q = −PQ
(1.34)
which implies PQ (I + PQ) = 0. Hence,
PQ = 0 or PQ = I. (1.35)
But by (1.34), PQ is skew-adjoint, it follows that PQ = 0, and so QP = 0.

Remark 1.14. Eq. (1.31) is analogous to the following identity for characteristic
functions:
χA + χB = χA∪B − χA∩B
Therefore, (χA + χB )2 = χA + χB iff χA∩B = 0, i.e., iff A ∩ B = 0.

/
The set of projections in a Hilbert space H is partially ordered according to the
corresponding closed subspaces partially ordered by inclusion. Since containment
implies commuting, the chain of projections P1 ≤ P2 ≤ · · · is a family of commuting
selfadjoint operators. By the spectral theorem (chapter 3), {Pi } may be simultane-
ously diagonalized, so that Pi is unitarily equivalent to the operator of multiplication
by χEi on the Hilbert space L2 (X, µ), where X is compact and Hausdorff. Therefore
the lattice structure of projections in H is precisely the lattice structure of χE , or
equivalently, the lattice structure of measurable sets in X.
Lemma 1.12. Consider L2 (X, µ). The following are equivalent.
1. E ⊂ F;
2. χE χF = χF χE = χE ;
3. kχE f k ≤ kχF f k, for any f ∈ L2 ;
4. χE ≤ χF , in the sense that
h f , χE f i ≤ h f , χF f i , ∀ f ∈ L2 (X) .
Proof. The proof is trivial. Note that

ˆ
h f , χE f i = χE | f |2 dµ
ˆ ˆ
kχE f k2 = |χE f |2 dµ = χE | f |2 dµ
where we used that fact that

χE = χE = χE2 .
Lemma 1.13. Let P, Q ∈ Pro j (H ); then the following conditions are equivalent:
1. PQP ∈ Pro j (H );
2. PQ = QP.
Proof. First note that the operator A = PQ − QP is skew-symmetric, i.e., A∗ = −A,

and so its spectrum is contained in the imaginary line iR.
The implication 2⇒1 above is immediate so assume 1, i.e., that (PQP)2 = PQP.
And using this, one checks by a direct computation that A3 = 0. But with A∗ = −A,
and the spectral theorem, we therefore conclude that A = 0, in other words, 2 holds.

A summary of relevant numbers from the Reference List
For readers wishing to follow up sources, or to go in more depth with topics above,
we suggest: [Arv72, Ban93, BR79, Con90, DM85, DS88c, Lax02, RS75, RSN90,
Rud73, Rud87, AJS14, AJLM13, AJL13, AJ12].
Chapter 2
Unbounded Operators in Hilbert Space
We were [initially] entirely in Heisenberg’s footsteps. He had

the idea that one should take matrices, although he did not know
that his dynamical quantities were matrices.... And when one
had such a programme of formulating everything in matrix
language, it takes some effort to get rid of matrices. Though it
seemed quite natural for me to represent perturbation theory in
the algebraic way, this was not a particularly new way.
— Max Born
Quantum physics is one of the sources of problems in Functional Analysis, in par-

ticular the study of operators in Hilbert space. In the dictionary translating between
Quantum physics and operators in Hilbert space we already saw that “quantum ob-
servables” are “selfadjoint operators.”
As noted, even in a finite number of degrees of freedom, the relevant operators
such as momentum, position, and energy are unbounded. Most Functional Analysis
books stress the bounded case, and below we identify questions and theorems related
to key-issues for unbounded linear operators.
In this chapter, we review the basic theory of unbounded operators in Hilbert
space. For general notions, we refer to [DS88b, DS88c].
2.1 Domain, Graph, and Adjoints
Let H be a complex Hilbert space. An operator A is a linear mapping whose domain

dom (A) and range ran (A) are subspaces in H . The kernel ker (A) of A consists of
all a ∈ dom (A) such that Aa = 0. The operator A is uniquely determined by its graph
G (A) = {(a, Aa) : a ∈ dom (A)} (2.1)
i.e., a subspace in H ⊕ H with the inherited graph inner product
ha, biA = ha, bi + hAa, Abi , and (2.2)

kak2A = ha, aiA , ∀a, b ∈ dom (A) (2.3)
55
56 2 Unbounded Operators in Hilbert Space
In general, a subspace K ⊂ H ⊕H is the graph of an operator if and only if (0, a) ∈

K implies a = 0.
Given two operators A and B, we say B is an extension of A, denoted by B ⊃ A,
if G (B) ⊃ G (A) in H ⊕ H . The operator A is closable if G (A) is the graph of an
operator A, namely, the closure of A. We say A is closed if A = Ā.
Let A be a closed operator. A dense subspace K ⊂ H is called a core of A, if the
closure of the restriction A K
is equal to A.
 
1ac
Let G be the group of all 3 × 3 matrices 0 1 b with Lie algebra g consisting
 
001
 
0ac
of matrices 0 0 b, (a, b, c) ∈ R3 . Let U be a strongly continuous unitary repre-
 
000
sentation of G acting on a Hilbert space H . Then for every X ∈ g, t 7−→ U etX

defines a strongly continuous one-parameter group; and hence its infinitesimal gen-
erator, denoted dU (X) is a skew-adjoint operator with dense domain in H .
In the example below, we apply this to H = L2 (R), and the unitary representa-
tion U of G is defined as follows: For f ∈ H = L2 (R), set
(U (g) f ) (x) = ei(c+bx) f (x + a) , x ∈ R;
called the Schrödinger representation. Differentiate in the three directions in the Lie
algebra, we get 


(dU (X1 ) f ) (x) = f 0 (x) = dx
d
f,

(dU (X2 ) f ) (x) = ix f (x) , and (2.4)



(dU (X ) f ) (x) = i f (x) .
3
The first two operators in (2.4) are often written as follows:
dU (X1 ) = iP
where P is the momentum operator of a single quantum mechanical particle (wave

function); and
dU (X2 ) = iQ
where Q is the corresponding position operator (in a single degree of freedom.)

These two operators may be realized as follows:
2.1 Domain, Graph, and Adjoints 57
Remark 2.1. It would appear that the function fλ (x) = eiλ x , λ fixed is an eigenfunc-
1 d
tion for P = i dx . But then, for all λ ,
P fλ = λ fλ
holds pointwise, “spectrum” depends on the ambient Hilbert space H , in this case
H = L2 (R); and fλ ∈
/ L2 (R). Nonetheless, if we allow an interval for the λ vari-
able, e.g., a < λ < b, with a and b being finite, then
ˆ b
eibx − eiax
Fa,b (x) = eiλ x dλ =
a ix
is in L2 (R); and hence P has continuous spectrum. The functions Fa,b (·) are exam-
ples of wave-packets in quantum mechanics.
Example 2.1. d/dx and Mx in QM, acting on L2 (R) with dense domain the Schwartz
space.
An alternative way to get a dense domain, a way that works for all representations,
is to use Gårding space, or C∞ vectors. Let u ∈ H and define
ˆ
uϕ := ϕ(g)Ug udg
G
where ϕ ∈ Cc∞ , and U ∈ Rep(G, H ). Let ϕε be an approximation of identity. Then

for functions on G, ϕε ? ψ → ψ as ε → 0; and for C∞ vectors, uϕε → u, as ε → 0
in H i.e., in the k·kH - norm. The set {uϕ } is dense in H. It is called the Gårding
space, or C∞ vectors, or Schwartz space. Differentiate Ug and get a Lie algebra
representation
d
ρ(X) := t=0
U(etX ) = dU(X).
dt
Lemma 2.1. uϕε − u → 0, as ε → 0.
´
Proof. Since u − uϕε = ϕε (g)(u −Ug u)dg, we have
ˆ
uϕε − u = ϕε (g)(u −Ug u)dg
ˆ G
≤ ϕε (g) u −Ug u dg
G
where the integration on G is w.r.t. Haar measure, and where we used the fact that
´
G ϕε = 1. Notice that we always assume the representations are norm continuous
in the g variable, otherwise it is almost impossible to get anything interesting. i.e.,

assume U being strongly continuous. So for all δ > 0, there is a neighborhood O of
e ∈ G so that u −Ug u < δ for all g ∈ O. Choose εδ so that ϕε is supported in O
for all ε < εδ . Then the statement is proved.
Probability Kernels. There are about 7 popular kernels in probability theory. Look
1 1
at any book on probability theory. One of them is the Cauchy kernel π 1+x2 . Notice
that not only uϕ is dense in H , their derivatives are also dense in H .
Theorem 2.1. The following are equivalent.
1. A = Ā.
2. G (A) = G (A).
3. dom (A) is a Hilbert space with respect to the graph inner product h·, ·iA .
4. If (an , Aan ) is a sequence in G (A) and (an , Aan ) → (a, b), then (a, b) ∈
G (A). In particular, b = Aa.
Proof. All follow from definitions.
Let X be a vector space over C. Suppose there are two norms defined on X, such
that
k·k1 ≤ k·k2 . (2.5)
Let Xi be the completion of X with respect to k·ki , i = 1, 2. The ordering (2.5) implies
the identify map
ϕ : (X, k·k2 ) → (X, k·k1 )
is continuous, hence it has a unique continuous extension ϕe to X2 ; and (2.5) passes

to the closure X2 . If ϕe is injective, X2 is embedded into X1 as a dense subspace. In
that case, k·k1 and k·k2 are said to be topologically consistent.
Lemma 2.2. k·k1 and k·k2 are topologically equivalent if and only if
 
{xn } ⊂ X is Cauchy under k·k2 
 
(hence Cauchy under k·k1 ) =⇒ kxn k2 → 0.
 
kxn k1 → 0
 
Proof. Note ϕe is linear, and

( )
x ∈ X2 ∃ (xn ) ⊂ X, kxn − xk2 → 0, ϕe (x) = 0
ker ϕe = .
(note ϕe (x) = limn ϕ (xn ) = limn xn in X1 )
2.1 Domain, Graph, and Adjoints 59
The lemma follows from this.
Lemma 2.3. The graph norm of A is topologically equivalent to k·k + kA·k.
Proof. This follows from the estimate
1
(kxk + kAxk)2 ≤ kxk2 + kAxk2 ≤ (kxk + kAxk)2 , ∀x ∈ dom (A) .
2
Theorem 2.2. An operator A is closable if and only if k·k and k·kA are topologi-
cally consistent. (When they do, the completion of dom (A) with respect to k·kA is
identified as a subspace of H .)
Proof. First, assume A is closable. Let {xn } be a sequence in dom (A). Suppose {xn }
is a Cauchy sequence with respect to k·kA , and kxn k → 0. We need to show {xn }
converges to 0 under the A-norm, i.e., kxn kA → 0. Since {(xn , Axn )} ⊂ G (A), and
A is closable, it follows that (xn , Axn ) → (0, 0) ∈ G (A). Therefore, kAxn k → 0, and
(see lemma 2.3)
kxn kA = kxn k + kAxn k → 0.
Conversely, assume k·k and k·kA are topologically consistent. Let {xn } ⊂ dom (A),
such that
(xn , Axn ) → (0, b) in H ⊕ H . (2.6)
We proceed to show that b = 0, which implies that A is closable.

By (2.6), {xn } ⊂ dom (A) is a Cauchy sequence with respect to k·kA , and kxn k →
0. Since the two norms are topologically consistent, then kxn kA → 0 and so kAxn k →
0. We conclude that b = 0.
Corollary 2.1. An operator A with dense domain is closable if and only if its adjoint
A∗ has dense domain.
We will focus on unbounded operators. In the sequel, we will consider densely de-
fined Hermitian (symmetric) operators. Such operators are necessarily closable.
The following result is usually applied to operators whose inverses are bounded.
Proposition 2.1. Let A be a bounded operator with domain dom (A), and act in H .
Then dom (A) is closed in k·kA if and only if it is closed in k·k. (That is, for bounded
operators, k·k and k·kA are topologically equivalent. )
Proof. This is the result of the following estimate:
kxk ≤ kxkA = kxk + kAxk ≤ (1 + kAk) kxk , ∀x ∈ dom (A) .
Corollary 2.2. If A is a closed operator in H and A−1 is bounded, then ran (A) is
closed in both k·k and k·kA−1 .
Proof. Note the G (A) is closed iff G A−1 is closed; and ran (A) = dom A−1 .

Now, apply 2.1 to A−1 .
Let A be an operator in a Hilbert space H . The set G (A)⊥ consists of (−b∗ , b)

such that (−b∗ , b) ⊥ G (A) in H ⊕ H .
Proposition 2.2. The following are equivalent.
1. D (A) is dense in H .
2. (b, 0) ⊥ G (A) =⇒ b = 0.
3. If (b, −b∗ ) ⊥ G (A), the map b 7→ b∗ is well-defined.
Proof. Let a ∈ D (A), and b, b∗ ∈ H ; then
(−b∗ , b) ⊥ (a, Aa) in H ⊕ H ⇐⇒ hb∗ , ai = hb, Aai
and the desired results follow from this.
If any of the conditions is satisfied, A∗ : b 7→ b∗ defines an operator, called the

adjoint of A, such that
hb, Aai = hA∗ b, ai (2.7)
for all a ∈ D (A). G (A)⊥ is the inverted graph of A∗ . The adjoints are only defined
for operators with dense domains in H .
Example 2.2. A = d/dx on L2 [0, 1] with dense domain
D = f ∈ C1 f (0) = f (1) = 0 .

Integration by parts shows that A ⊂ −A∗ .
For unbounded operators, (AB) ∗ = B∗ A∗ does not hold in general. The situation is
better if one of them is bounded.
2.2 Characteristic Matrix 61
Theorem 2.3 ([Rud90, Theorem 13.2]). If S, T, ST are densely defined operators

then (ST )∗ ⊃ T ∗ S∗ . If, in addition, S is bounded then (ST )∗ = T ∗ S∗ .
The next theorem follows directly from the definition of the adjoint operators.
Theorem 2.4. If A is densely defined then H = R (A) ⊕ K (A∗ ).
Finally, we recall some definitions.
Definition 2.1. Let A be a linear operator acting in H . A is said to be
• selfadjoint if A = A∗ .
• essentially selfadjoint if A = A∗ .
• normal if A∗ A = AA∗ .
• regular if D (A) is dense in H , and closed in k·kA .
Definition 2.2. Let A be a linear operator on a Hilbert space H . The resolvent R (A)
is defined as
n o
R (A) = λ ∈ C : (λ − A)−1 exists (the resolvent set)
and the spectrum of A is the complement of R (A), and it is denoted by sp (A) or

σ (A).
Exercise 2.1. Let A be a linear operator in a Hilbert space H , and, for λi ∈ R (A),
i = 1, 2 consider two operators (λi − A)−1 . Show that
(λ1 − A)−1 − (λ2 − A)−1 = (λ2 − λ1 ) (λ1 − A)−1 (λ2 − A)−1 .
This formula is called the resolvent identity.
2.2 Characteristic Matrix
The method of characteristic matrix was developed by M.S. Stone’s [Sto51a]. It is

extremely useful in operator theory, but has long been overlooked in the literature.
We recall some of its applications in normal operators.
( If H! is a fixed Hilbert
) space, and A a given liner operator, then its graph G (A) =
u
: u ∈ dom (A) is a linear subspace in H ⊕ H , represented as column
Au
!
u
vectors , u, v ∈ H . In the case where A is assumed closed, we now compute
v
the projection onto G (A) = the H ⊕ H -closure of the graph.
Let A be an operator in a Hilbert space H . Let P = (Pi j ) be the projection from
H ⊕ H onto G (A). The 2 × 2 operator matrix (Pi j ) of bounded operators in H is
called the characteristic matrix of A.
Since P2 = P∗ = P, the following identities hold
Pi∗j = Pji (2.8)
∑ Pik Pk j = Pi j (2.9)
k
In particular, P11 and P22 are selfadjoint.

Theorem 2.5. Let P = (Pi j ) be the projection from H ⊕ H onto a closed subspace
K . The following are equivalent.
1. K is the graph of an operator.
2. " #" # " #
P11 P12 0 0
= =⇒ a = 0.
P21 P22 a a
3.
P12 a = 0, P22 a = a =⇒ a = 0.
If any of these conditions is satisfied, let A be the operator with G (A) = K , then
for all a, b ∈ H ,
" #" # " #
P11 P12 a P11 a + P12 b
= ∈ G (A) ;
P21 P22 b P21 a + P22 b
i.e.,
A : (P11 a + P12 b) 7→ P21 a + P22 b. (2.10)
In particular,
AP11 = P21 (2.11)

AP12 = P22 (2.12)
Proof. Let v := (a, b) ∈ H ⊕H . Then v ∈ K if and only if Pv = v; and the theorem

follows from this.
The next theorem describes the adjoint operators.
Theorem 2.6. Let A be an operator with characteristic matrix P = (Pi j ). The fol-
lowing are equivalent.
1. D (A) is dense in H .
2. " #
b
⊥ G (A) = 0 =⇒ b = 0.
0
" #
−b∗
3. If ∈ G (A)⊥ , the map A∗ : b 7→ b∗ is a well-defined operator.
b
4. " #" # " #
1 − P11 −P12 b b
= =⇒ b = 0.
−P21 1 − P22 0 0
5.
P11 b = 0, P21 b = 0 =⇒ b = 0.
If any of the above conditions is satisfied, then

" #" # " #
1 − P11 −P12 a (1 − P11 )a − P12 b
= ∈ G (A)⊥
−P21 1 − P22 b (1 − P22 )b − P21 a
that is,
A∗ : P21 a − (1 − P22 )b 7→ (1 − P11 )a − P12 b. (2.13)
In particular,
A∗ P21 = 1 − P11 (2.14)

A∗ (1 − P22 ) = P12 . (2.15)
Proof. (1) ⇔ (2) ⇔ (3) is a restatement of Proposition 2.2. Note the projection
⊥
from H ⊕ H on G (A)⊥ = G (A) is
" #
1 − P11 −P12
1−P =
−P21 1 − P22
and so " # " # " #

b b b
⊥ G (A) ⇐⇒ (1 − P) = .
0 0 0
Therefore, (2) ⇔ (4) ⇔ (5). (2.13)-(2.15) follows from the definition of A∗ .
Theorem 2.7. Let A be a regular operator (i.e., densely defined, closed) with char-
acteristic matrix P = (Pi j ).
1. The matrix entries Pi j are given by
P11 = (1 + A∗ A)−1 P12 = A∗ (1 + AA∗ )−1

(2.16)
P21 = A(1 + A∗ A)−1 P22 = AA∗ (1 + AA∗ )−1
2. 1 − P22 = (1 + AA∗ )−1 .

3. 1 + A∗ A, 1 + AA∗ are selfadjoint operators.
4. The following containments hold
A∗ (1 + AA∗ )−1 ⊃ (1 + A∗ A)−1 A∗ (2.17)

∗ −1 ∗ −1
A(1 + A A) ⊃ (1 + AA ) A (2.18)
Proof. By (2.11) and (2.14), we have

!
AP11 = P21
=⇒ A∗ AP11 = 1 − P11 , i.e., (1 + A∗ A)P11 = 1.
A∗ P21 = 1 − P11
−1
That is, 1 + A∗ A is a Hermitian extension of P11 . By (2.8), P11 is selfadjoint and so
−1 −1
is P11 . Therefore, 1 + A∗ A = P11 , or
P11 = (1 + A∗ A)−1 .
By (2.11),
P21 = AP11 = A(1 + A∗ A)−1 .
Similarly, by (2.12) and (2.15), we have

!
AP12 = P22
=⇒ AA∗ (1 − P22 ) = P22 , i.e.,
A∗ (1 − P22 ) = P12
(1 + AA∗ )(1 − P22 ) = 1.
This means 1 + AA∗ ⊃ (1 − P22 )−1 is a Hermitian extension of the selfadjoint oper-
ator (1 − P22 )−1 (note P22 is selfadjoint), hence 1 + AA∗ is selfadjoint, and
1 − P22 = (1 + AA∗ )−1 .

By (2.15),
P12 = A∗ (1 − P22 ) = A∗ (1 + AA∗ )−1 .
By (2.12),
P22 = AP12 = AA∗ (1 + AA∗ )−1
We have proved (1), (2) and (3).

Finally,
∗
P12 = P21 = (AP11 )∗ ⊃ P11 A∗
yields (2.17); and
∗
P21 = P12 = (A∗ (1 − P22 ))∗ ⊃ (1 − P22 )A
gives (2.18).
2.2.1 Commutants
Let A, B be operators in a Hilbert space H , and suppose B is bounded. B is said to

commute (strongly) with A if BA ⊂ AB.
Lemma 2.4. B commutes with A if and only if B commutes with Ā (assuming A

exists).
Proof. Suppose BA ⊂ AB, we check that BA ⊂ AB. The converse is trivial. For
(a, Aa) ∈ G (A), choose a sequence (an , Aan ) ∈ G (A) such that (an , Aan ) → (a, Aa).
By assumption, (Ban , ABan ) = (Ban , BAan ) ∈ G (A). Thus, (Ban , ABan ) → (Ba, BAa) ∈
G (A). That is, Ba ∈ D(A) and ABa = BAa.
Lemma 2.5. Let A be a closed operator with characteristic matrix P = (Pi j ). Let B
be a bounded operator, and " #
B0
QB := .
0B
1. B commutes with A ⇔ B leaves G (A) invariant ⇔ QB P = PQB P.

2. B commutes with Pi j ⇔ QB P = PQB ⇔ QB∗ P = PQB∗ ⇔ B∗ commutes
with Pi j .
3. If B, B∗ commute with A, then B, B∗ commute with Pi j .
Proof. Obvious.
A closed operator is said to be affiliated with a von Neumann algebra M if it

commutes with every unitary operator in M0 . By [KR97a, Thm 4.1.7], every oper-
ator in M0 can be written as a finite linear combination of unitary operators in M0 .
Thus, A is affiliated with M if and only if A commutes with every operator in M0 .
Remark 2.2. Let M be a von Neumann algebra. Let x ∈ M s.t. kxk ≤ 1 and x = x∗ .
√
Set y := x + i 1 − x2 . Then, y∗ y = yy∗ = x2 + 1 − x2 = 1, i.e., y is unitary. Also,
x = (y + y∗ ) /2.
Theorem 2.8. Let A be a closed operator with characteristic matrix P = (Pi j ). Let
M be a von Neumann algebra, and
" #
B0
QB := , B ∈ M0 .
0B
The following are equivalent:
1. A is affiliated with M.
2. PQB = QB P, for all B ∈ M0 .
3. Pi j ∈ M.
4. If D(A) is dense, then A∗ is affiliated with M.
Proof. Notice that M is selfadjoint. The equivalence of 1, 2, 3 is a direct conse-

quence of Lemma 2.5.
P⊥ := 1 − P is the projection onto the inverted graph of A∗ , should the latter
exists. PQB = QB P if and only if P⊥ QB = QB P⊥ . Thus, 1 is equivalent to 4.
2.3 Normal Operators
As a first application of Stone’s characteristic matrix, we give a new proof to fol-

lowing theorem concerning operators of the form A∗ A.
Theorem 2.9 (von Neumann). If A is a regular operator in a Hilbert space H ,

then
1. A∗ A is selfadjoint;
2. D(A∗ A) is a core of A, i.e.,
A D (A∗ A)
= A;
2.3 Normal Operators 67
3. In particular, D(A∗ A) is dense in H .

−1 −1
Proof. By Theorem 2.7, A∗ A = P11 − 1. Since P11 is selfadjoint, so is P11 . Thus,
AA∗ is selfadjoint.
Suppose (a, Aa) ∈ G (A) such that
(a, Aa) ⊥ G (A D (A∗ A)

); i.e.,
ha, bi + hAa, Abi = ha, (1 + A∗ A) bi = 0, ∀b ∈ D (A∗ A) .

−1
Since 1 + A∗ A = P11 , and P11 is a bounded operator, then
R(1 + A∗ A) = D(P11 ) = H .
It follows that a ⊥ H , and so a = 0.
Theorem 2.10 (von Neumann). Let A be a regular operator in a Hilbert space H .

Then A is normal if and only if D(A) = D(A∗ ) and kAak = kA∗ ak, for all a ∈ D(A).
Proof. Suppose A is normal. Then for all a ∈ D (A∗ A) (= D (AA∗ )), we have
kAak2 = hAa, Aai = ha, A∗ Aai = ha, AA∗ ai = hAa, A∗ ai = kA∗ ak2 ;
i.e., kAak = kA∗ ak, for all a ∈ D (A∗ A). It follows that

D A D (A∗ A)
= D A∗ D (AA∗ )
.
By Theorem 2.9, D (A) = D(A D (A∗ A) ) and D (A∗ ) = D(A∗ D (AA∗ )

). Therefore,
D(A) = D(A∗ ) and kAak = kA∗ ak, for all a ∈ D(A).
Conversely, the map Aa 7→ A∗ a, a ∈ D(A), extends uniquely to a partial isometry
V with initial space R(A) and final space R(A∗ ), such that A∗ = VA. By Theorem
2.3, A = A∗V ∗ . Then A∗ A = A∗ (V ∗V )A = (A∗V ∗ )(VA) = AA∗ . Thus, A is normal.

The following theorem is due to M.S. Stone.
Theorem 2.11 (M.S. Stone). Let A be a regular operator in a Hilbert space H . Let
P = (Pi j ) be the characteristic matrix of A. The following are equivalent.
1. A is normal.
2. Pi j are mutually commuting.
3. A is affiliated with an abelian von Neumann algebra.
Remark 2.3. For the equivalence of 1 and 2, we refer to the original paper of Stone.
The most interesting part is 1 ⇔ 3. The idea of characteristic matrix gives rise to an
elegant proof without reference to the spectral theorem.
Proof. Assuming 1 ⇔ 2, we prove that 1 ⇔ 3.

Suppose A is normal, i.e. Pi j are mutually commuting. Then A is affiliated with
the abelian von Neumann algebra {Pi j }00 . For if B ∈ {Pi j }0 , then B commutes Pi j ,
and so B commutes with A by Lemma 2.5.
Conversely, if A is affiliated with an abelian von Neumann algebra M, then Pi j ∈
M by Theorem 2.8. This shows that Pi j are mutually commuting, and A is normal.

2.4 Polar Decomposition
Let A be a regular operator in a Hilbert space H . By Theorem 2.9, A∗ A is a positive

√
selfadjoint operator and it has a unique positive square root |A| := A∗ A.
Theorem 2.12.
√
1. |A| := A∗ A is the unique positive selfadjoint operator T satisfying D(T ) =
D(A), and kTak = kAak for all a ∈ D(A).
2. ker (|A|) = ker (A), R(|A|) = R(A∗ ).
√
Proof. Suppose T = A∗ A, i.e. T ∗ T = A∗ A. Let D := D(T ∗ T ) = D(A∗ A). By The-
orem 2.9, D is a core of both T and A. Moreover, kTak = kAak, for all a ∈ D.
We conclude from this norm identity that D(T ) = D(A) and kTak = kAak, for all
a ∈ D(A).
Conversely, suppose T has the desired properties. For all a ∈ D(A) = D(T ), and
b ∈ D(A∗ A),
hT b, Tai = hAb, Aai = hA∗ Ab, ai
This implies that T b ∈ D(T ∗ ) = D(T ), T 2 b = A∗ Ab, for all b ∈ D(A∗ A). That is,
T 2 is a selfadjoint extension of A∗ A. Since A∗ A is selfadjoint, T 2 = A∗ A.
The second part follows from Theorem 2.4.
Consequently, the map |A|a 7→ Aa extends to a unique partial isometry V with

initial space R(A∗ ) and final space R(A), such that
2.4 Polar Decomposition 69
A = V |A| . (2.19)
Equation (2.19) is called the polar decomposition of A. It is clear that such decom-
position is unique.
Taking adjoints in (2.19) yields A∗ = |A|V ∗ , so that
AA∗ = VA∗ AV ∗ (2.20)
Restrict AA∗ to R(A), and restrict A∗ A restricted to R(A∗ ). Then the two restrictions
are unitarily equivalent. It follows that A∗ A, AA∗ have the same spectrum, aside from
possibly the point 0.
√
By (2.20), |A∗ | = V |A|V ∗ = VA∗ , where |A∗ | = AA∗ . Apply V ∗ on both sides
gives
A∗ = V ∗ |A∗ | . (2.21)
By uniqueness, (2.21) is the polar decomposition of A∗ .
Theorem 2.13. A is affiliated with a von Neumann algebra M if and only if |A| is
affiliated with M and V ∈ M.
Proof. Let U be a unitary operator in M0 . The operator UAU ∗ has polar decompo-
sition
UAU ∗ = (UVU ∗ )(U |A|U ∗ ).
By uniqueness, A = UAU ∗ if and only if V = UVU ∗ , |A| = U |A|U ∗ . Since U is

arbitrary, we conclude that V ∈ M, and A is affiliated with M.
we suggest: [BR81, DS88c, Jor08, Kat95, KR97b, Sto51a, Sto51b, Wei03, Yos95].
Chapter 3
The Spectral Theorem
As far as the laws of mathematics refer to reality, they are not

certain, and as far as they are certain, they do not refer to
reality. — Albert Einstein
A large part of mathematics which becomes useful developed

with absolutely no desire to be useful, and in a situation where
nobody could possibly know in what area it would become
useful; and there were no general indications that it ever
would be so. By and large it is uniformly true in mathematics
that there is a time lapse between a mathematical discovery and
the moment when it is useful; and that this lapse of time can be
anything from 30 to 100 years, in some cases even more; and
that the whole system seems to function without any direction,
without any reference to usefulness, and without any desire to
do things which are useful. – von Neumann
Most Functional Analysis books, when covering the Spectral Theorem, stress the
bounded case. Because of dictates from applications (especially quantum physics),
below we stress questions directly related to key-issues for unbounded linear op-
erators. These themes will be taken up again in chapters 9 and 10. In a number of
applications, some operator from physics may only be “formally selfadjoint” also
called Hermitian; and in such cases, one asks for selfadjoint extensions (if any),
chapter 9. Chapter 10 is a particular case in point, arising in the study of infinite
graphs.
3.1 An Overview
von Neumann’s spectral theorem states that an operator A acting in a Hilbert space
H is normal if and only if there exits a projection-valued measure on C so that
ˆ
A= zPA (dz) (3.1)
sp(A)
71
72 3 The Spectral Theorem
i.e., A is represented as an integral against the projection-valued measure PA over its

spectrum.
In quantum mechanics, an observable is represented by a selfadjoint operator.
Functions of observables are again observables. This is reflected in the spectral the-
orem as the functional calculus, where we may define
ˆ
ϕ(A) = ϕ(z)PA (dz) (3.2)
sp(A)
using the spectral representation of A.

When P is a selfadjoint projection, h f , P f iH = kP f k2H is a real number and it
represents the expected value of the observable P prepared in the state f , unit vector
in H . Hence, in view of (3.2), kPA (·) f k2H is a Borel probability measure on sp (A),
and ˆ
h f , ϕ (A) f iH = ϕ (z) kP (dz) f k2H (3.3)
sp(A)
is the expected value of the observable ϕ (A).
Remark 3.1. Let ϕ : R → R be measurable and let A = A∗ be given; then, for ev-
(A)
ery f ∈ H \ {0}, set dµ f (λ ) := kPA (dλ ) f k2 ∈ M+ (R) (the finite positive Borel
measures on R.) Then the transformation formula (3.3) takes the following equiva-
lent form:
(ϕ(A)) (A)
dµ f = dµ f ◦ ϕ −1 , i.e., (3.4)
(ϕ(A)) (A)
dµ f (4) = dµ f ◦ ϕ −1 (4) , ∀4 ∈ B (R) . (3.5)
Corollary 3.1. Let A = A∗ , and f ∈ H \ {0} be given, and let µ f and ϕ be as in

Remark 3.1, then ϕ (A)∗ = ϕ (A), and
f ∈ dom (ϕ (A)) ⇐⇒ ϕ ∈ L2 R, µ f ,

where “dom” is short for “domain.”
Remark 3.2. The standard diagonalization of Hermitian matrices in linear algebra

is a special case of the spectral theorem. Recall that any Hermitian matrix A can
be decomposed as A = ∑k λk Pk , where λk0 s are the eigenvalues of A and Pk0 s are the
selfadjoint projections onto the eigenspaces associated with λk0 s. The projection-
valued measure in this case can be written as P(E) = ∑λk ∈E Pk , for all E ∈ B (R);
i.e., the counting measure supported on λk0 s.
3.1 An Overview 73
Quantum mechanics is stated using an abstract Hilbert space as the state space. In
practice, one has the freedom to choose exactly which Hilbert space to use for a par-
ticular problem. Physical measurements remain unchanged when choosing different
realizations of a Hilbert space. The concept needed here is unitary equivalence.
Definition 3.1. Let A : H1 → H1 and B : H2 → H2 be operators. A is said to be

unitarily equivalent to B if there exists a unitary operator U : H1 → H2 such that
B = UAU ∗ .
Suppose U : H1 → H2 is a unitary operator, P : H1 → H1 is a selfadjoint projection.

Then UPU ∗ : H2 → H2 is a selfadjoint projection on H2 . In fact, (UPU ∗ )(UPU ∗ ) =
UPU ∗ where we used UU ∗ = U ∗U = I, since U is unitary. Let | f1 i be a state in H1
and | f2 i = |U f1 i be the corresponding state in H2 . Then
h f2 ,UPU ∗ f2 iH2 = hU ∗ f2 , PU ∗ f2 iH1 = h f1 , P f1 iH1
i.e., the observable P has the same expectation value. Since every selfadjoint oper-
ator is, by the spectral theorem, decomposed into selfadjoint projections, it follows
that the expectation value of any observable remains unchanged under unitary trans-
formations.
We will also consider family of selfadjoint operators. Heisenberg’s commutation
relation PQ − QP = −iI is an important example of two non-commuting selfadjoint
operators.
Example 3.1. The classical Fourier transform F : L2 (R) → L2 (R) is unitary, so

F ∗ F = FF ∗ = IL2 (R) , and in particular, the Parseval identity
kF f k2L2 (R) = k f k2L2 (R)
holds for all f ∈ L2 (R).
Example 3.2. Let Q and P be the position and momentum operators in quantum
mechanics. That is, Q = Mx = multiplication by x, and P = −id/dx both defined
on the Schwartz space S (R)–space of rapidly decreasing functions on R, which
is dense in the Hilbert space L2 (R). On S (R), the operators P and Q satisfy the
canonical commutation relation: PQ − QP = −iIL2 (R) .
Denote F the Fourier transform on L2 (R) as before. Specifically, setting
ˆ
1
(Fϕ) (x) = ϕb (x) = √ ϕ (ξ ) e−iξ x dξ , and
2π R
ˆ
1
(F ∗ ψ) (ξ ) = ψ ∨ (ξ ) = √ ψ (x) eiξ x dx, ξ ∈ R.
2π R
Note that F is an automorphism in S (R), continuous with respect to the standard

l.c. topology. Moreover,
1 d
(F ∗ QFϕ) (ξ ) = F ∗ (xϕb (x)) = ϕ (ξ ) , ∀ϕ ∈ S.
i dξ
Therefore,
P = F ∗ QF (3.6)
and so P and Q are unitarily equivalent.
A multiplication operator version of the spectral theorem is also available. It

works especially well in physics. It says that A is a normal operator in H if and
only if A is unitarily equivalent to the operator of multiplication by a measurable
function f on L2 (X, µ), where X is locally compact and Hausdorff. The two versions
are related via a measure transformation.
Example 3.3. Eq. (3.6) says that P is diagonalized by Fourier transform in the fol-
lowing sense.
Let ψ be any Borel function on R, and set Mψ = multiplication by ψ (x) in
L2 (R), with
dom Mψ = f f , ψ f ∈ L2 (R)

(3.7)
ˆ ∞
= f 1 + |ψ (x)|2 | f (x)|2 dx < ∞ ;
−∞
then we define, via eq. (3.6),
ψ (P) := F ∗ ψ (Q) F.
In particular, given any 4 ∈ B (R), let ψ = χ4 = characteristic function, then
E(4) = F ∗ Mχ4 F.
One checks directly that E (4)2 = E (4) = E (4)∗ , so E (4) is a selfadjoint pro-
jection. Indeed, E (·) is a convolution operator, where
ˆ b ∧
(E (4) f ) (x) = eiξ x fb(ξ ) dξ = f ∗ χ[a,b] (x) , ∀ f ∈ L2 (R) .
a
3.1 An Overview 75
Thus,
n o
E (4) L2 (R) = f ∈ L2 (R) supp fb ⊂ 4 ;
i.e., the space of “band-limited” functions, with the “pass-band” being 4.
Example 3.4. Below, it helps to denote the Fourier transformed space (or frequency
space) by L2 (R).
b Fix any f ∈ L2 (R), 4 ∈ B (R), then
µ f (4) := kE (4) f k2L2 (R) = h f , E (4) f iL2 (R)

D E
= F f , Mχ4 F 2 b
L (R)
ˆ 2
= fb(x) dx
4
which is a Borel measure on R, such that

ˆ ∞ 2
ˆ ∞
µ f (R) = fb(x) dx = | f (x)|2 dx = k f k2L2 (R)
−∞ −∞
by the Parseval identity.

Now, let ψ (x) = x be the identity function, and note that it is approximated point-
wisely by simple functions of the form ∑finite ci χ4i , where 4i ∈ B (R), and 4i ’s are
mutually disjoint, i.e., x = limn→∞ ∑ni=1 ci χ4i .
Fix f ∈ dom (Mx ), see (3.7), it follows from Lebesgue dominated convergence
theorem, that
ˆ ∞
h f , P f iL2 (R) = fb(x)x fb(x) dx
−∞
ˆ ∞ n
!
= lim fb(x) ∑ ci χ4i (x) fb(x) dx
n→∞ −∞
i=1
n
= lim
n→∞
∑ ci h f , E (4i ) f iL2 (R)
i=1
n
= lim
n→∞
∑ ci kE (4i ) f k2L2 (R)
i=1
ˆ ∞
= x kE (dx) f k2L2 (R) .
−∞
The last step above yields the projection-valued measure (PVM) version of the spec-
tral theorem for P, where we write
ˆ ∞
P= x dE (x) . (3.8)
−∞
1 d
Consequently, we get two versions of the spectral theorem for P = i dx S(R) :
1. Multiplication operator version, i.e., P ' Mx = multiplication by x in

L2 (R);
b and
2. PVM version, as in (3.8).
This example illustrates the main ideas of the spectral theorem of a single selfadjoint
operator in Hilbert space. We will develop the general theory in this chapter, and
construct both versions of the spectral decomposition.
Example 3.5. Applying the Gram-Schmidt process to all polynomials against the
2 /2
measure dµ = e−x dx, one gets orthogonal polynomials Pn in L2 (µ). These are
called the Hermite polynomials, and the associated Hermite functions are given by
n
−x2 /2 −x2 d 2 /2
hn := e Pn = e ex .
dx
The Hermite functions (after normalization) forms an ONB in L2 (R), which trans-
forms P and Q to Heisenberg’s infinite matrices in (1.18)-(1.19).
Example 3.6 (The harmonic oscillator Hamiltonian).
Let P, Q be as in the previous example. We consider the quantum Hamiltonian
1
H := (Q2 + P2 − 1).
2
It can be shown that

Hhn = nhn
or equivalently,
(P2 + Q2 )hn = (2n + 1)hn
n = 0, 1, 2, . . .. That is, H is diagonalized by the Hermite functions.
H is called the energy operator in quantum mechanics. This explains

mathematically why the energy levels are discrete (in quanta), being a multiple the
Plank’s constant h̄.
Example 3.7 (Purely discrete spectrum v.s. purely continuous spectrum). The two
operators P2 + Q2 and P2 − Q2 acting in L2 (R); see Fig 3.1.
3.2 Multiplication Operator Version 77
(a) Harmonic oscillator P2 + Q2 (b) Repulsive potential P2 − Q2 .

(bound-states).
Fig. 3.1: Illustration of forces: attractive v.s. repulsive.
In the following sections, we present some main ideas of the spectral theorem for
single normal operators acting in Hilbert space. Since every normal operator N can
be written as N = T1 + iT2 , where T1 and T2 are strongly commuting and selfadjoint,
the presentation will be focused on selfadjoint operators.
3.2 Multiplication Operator Version
This version of the spectral theory states that every selfadjoint operator A is unitarily
equivalent to the operator of multiplication by a measurable function on some L2 -
space.
Theorem 3.1. Let A be a linear operator acting in the Hilbert space H , then A = A∗
iff there exists a unitary operator U : L2 (X, µ) → H such that
Mϕ = U ∗ AU; (3.9)
where X is locally compact and Hausdorff, ϕ is a real-valued µ-measurable func-

tion, and

Mϕ f := ϕ f , ∀ f ∈ dom Mϕ , where (3.10)
dom Mϕ := h ∈ L2 (X, µ) : ϕh ∈ L2 (X, µ) .

(3.11)
Hence, the following diagram commutes.
HO
A /H
O
U U
Mϕ
L2 (X, µ) / L2 (X, µ)
If A ∈ B (H ), then ϕ ∈ L∞ (X, µ) and dom Mϕ = H .

Exercise 3.1. Prove that Mϕ in (3.10)-(3.11) is selfadjoint.
Exercise 3.2. Let Mt : L2 [0, 1] → L2 [0, 1], f (t) 7−→ t f (t). Show that Mt has no
eigenvalues in L2 [0, 1].
Before giving a proof of Theorem 3.1, we show below that one can go one step
further and get that A is unitarily equivalent to the operator of multiplication by
the independent variable on some L2 -space. This is done by a transformation of the
measure µ in (3.11).
3.2.1 Transformation of Measures
Definition 3.2. Let ϕ : X → Y be a measurable function, TX and TY be the respective

sigma-algebras. Fix a measure µ on TX , the measure
µϕ := µ ◦ ϕ −1 (3.12)
defined on TY is called the transformation of µ under ϕ. Note that
χE ◦ ϕ(x) = χϕ −1 (E) (x) (3.13)
for all E ∈ TY and x ∈ X.
Lemma 3.1. For all T -measurable function f ,

ˆ ˆ
f d µ ◦ ϕ −1 .

f ◦ ϕdµ = (3.14)
X Y
(This is a generalization of the substitution formula in calculus.)
Proof. For any simple function s = ∑ ci χEi , Ei ∈ TY , it follows from (3.13) that
ˆ ˆ
s ◦ ϕdµ = ∑ ci χEi ◦ ϕdµ
X
ˆX
= ∑ ci χϕ −1 (Ei ) dµ
X
= ∑ ci µ ϕ −1 (Ei )

ˆ
= s d µ ◦ ϕ −1 .

Y
Note all the summations in the above calculation are finite.

Since any measurable function f : X → Y is approximated pointwisely by simple
functions, eq. (3.14) follows.
Remark 3.3. If ϕ is nasty, even if µ is a nice measure (say the Lebesgue measure),
the transformation measure µ ◦ ϕ −1 in (3.14) can still be nasty, e.g., it could even be
singular.
To simplify the discussion, we consider bounded selfadjoint operators below.
Corollary 3.2. Let ϕ : X → X be any measurable function. Then the operator U f :=

f ◦ ϕ in L2 (X, µ) is isometric iff µ ◦ ϕ −1 = µ. Moreover, Mϕ U = UMt . In particular,
U is unitary iff ϕ is invertible.
Proof. Follows from 3.1 immediately.
Lemma 3.2. In Theorem 3.1, assume A is bounded selfadjoint, so that ϕ ∈ L∞ (X, µ),
and real-valued. Let µϕ := µ ◦ ϕ −1 (eq. 3.12), supported on the essential range of
ϕ. Then the operator W : L2 R, µϕ −→ L2 (X, µ), by

(W f ) (x) = f (ϕ (x)) , ∀ f ∈ L2 µϕ

(3.15)
is isometric, and
W Mt = Mϕ W, (3.16)
where Mt : L2 Y, µ f −→ L2 Y, µ f , given by

(Mt f ) (t) = t f (t) ; (3.17)
i.e., multiplication by the identify function.
Proof. For all f ∈ L2 Y, µϕ , we have

ˆ ˆ
k f k2L2 (Y,µϕ ) = 2
| f | dµϕ = | f ◦ ϕ|2 dµ = kW f k2L2 (X,µ)
Y X
so W is isometric. Moreover,
Mϕ W f = ϕ (x) f (ϕ (x))
W Mt f = W (tg (t)) = ϕ (x) f (ϕ (x))
hence (3.16)-(3.17) follows.
Corollary 3.3. Let ϕ be as in Lemma 3.2. Assume ϕ is invertible, we get that W in

(3.15)-(3.17) is unitary. Set F = UW : L2 R, µϕ −→ H , then F is unitary and

Mt = F ∗ AF ; (3.18)
i.e., the following diagram commutes.
H
A /H
? O O _
U U
Mϕ
F L2 (X, µ) / L2 (X, µ) F
O O
W W
L2 (R, µϕ )
Mt
/ L2 (R, µϕ )
Remark 3.4.
1. Eq. (3.18) is a vast extension of diagonalizing hermitian matrices in linear

algebra, or a generalization of Fourier transform.
2. Given a selfadjoint operator A in the Hilbert space H , what’s involved are
two algebras: the algebra of measurable functions on X, treated as mul-
tiplication operators, and the algebra of operators generated by A (with
identity). The two algebras are ∗-isomorphic. The spectral theorem allows
us to (i) represent the algebra generated by A by the algebra of functions
(in this direction, it helps to understand A); also (ii) represent the algebra
of functions by algebra of operators generated by A (in this direction, it
reveals properties of the function algebra and the underlying space X.)
3. Let A be the algebra of functions. By a representation π ∈ Rep(A, H ), we
mean π : A → B (H ) is a ∗ homomorphism, i.e., π (gh) = π (g) π (h), and
π (g) = π (g)∗ , for all g, h ∈ A. Given F as in (3.18), then
π(ψ) = FMψ F ∗ ∈ Rep (A, H ) , (3.19)

and we define operators
ψ(A) := π(ψ), ψ ∈ A.
This is called the spectral representation of A. In particular, the spectral

theorem of A implies the following substitution rule
∑ ck xk 7−→ ∑ ck Ak
is well-defined, and it extends to all bounded measurable functions.
π
4. To see that ψ −→ ψ(A) := FMψ F ∗ (eq. 3.19) is an algebra isomorphism,
one checks that
(ψ1 ψ2 ) (A) = FMψ1 ψ2 F ∗

= FMψ1 Mψ2 F ∗
= FMψ1 F ∗ FMψ2 F ∗

= ψ1 (A) ψ2 (A)
using the fact that

Mψ1 ψ2 = Mψ1 Mψ2
i.e., multiplication operators always commute.
Let ϕ be as in Lemma 3.2. For the more general case when ϕ is not necessarily
invertible, so W as in (3.16) may not be unitary, we may still diagonalize A, i.e., get
that A is unitarily equivalent to multiplication by the independent variable in some
L2 -space; but now the L2 -space is vector-valued, and what is involved is a direct
integral representation. This argument is briefly sketched in the next section.
3.2.2 Direct Integral Representation
Throughout, we assume all the Hilbert spaces are separable.

The multiplication operator version of the spectral theorem says that A = A∗ ⇐⇒
A ' Mϕ , where
Mϕ : L2 (X, µ) −→ L2 (X, µ) (3.20)

f 7−→ ϕ f . (3.21)
Note that ϕ is real-valued. Moreover, A is bounded iff Mϕ ∈ L∞ (X, µ). When the
Hilbert space H is separable, we may further assume that µ is finite, or take µ to
be a probability measure.
To further diagonalize Mϕ , in the case when ϕ is “nasty”, we will need the fol-
lowing tool from measure theory.
Definition 3.3. Let X be a locally compact and Hausdorff space, and µ a Borel
probability measure on X. Let ϕ : X → Y be a measurable function, and set ν :=
µ ◦ ϕ −1 . A disintegration of µ with respect to ϕ is a system of probability measures

µy : y ∈ Y on X, satisfying
1. µy is supported on the “fiber” ϕ −1 ({y}), i.e., µy X\ϕ −1 ({y}) = 0, ν-a.e;

2. For all Borel set E in X, the function y 7→ µy (E) is ν-measurable, and

ˆ
µ (E) = µy (E) dν (y) . (3.22)
Y
Now, back to the spectral theorem. Let Mϕ be as in (3.20)-(3.21), and let ν := µ ◦

ϕ −1 , i.e., a Borel probability measure on R. In fact, ν is supported on the essential
range of ϕ.
It is well-known that, in this case, there exists a unique (up to measure zero sets)
disintegration of µ with respect to ϕ. See, e.g., [Par82]. Therefore, we get the direct
integral decomposition
ˆ ⊕
2
L (µ) ' L2 (µy )dν (y) (unitarily equivalent) (3.23)

where µy : y ∈ essential range of ϕ ⊂ R is the system of probability measures as
in Definition 3.3.
The RHS in (3.23) is the Hilbert space consisting of measurable cross-sections
S 2
f :R→ L (µy ), where f (y) ∈ L2 (µy ), ∀y, and with the inner product given by
ˆ
h f , giL2 (ν) := h f (y) , g (y)iL2 (µy ) dν (y) .
Y
Theorem 3.2. Let Mϕ : L2 (X, µ) → L2 (X, µ) be the multiplication operator in

(3.20)-(3.21), ν := µ ◦ ϕ −1 as before. Then Mϕ is unitarily equivalent to multi-
´⊕ 2
plication by the independent variable on L (µy )dν (y).
For details, see, e.g., [Dix81, Seg50].
3.2.3 Proof of Theorem 3.1
We try to get the best generalization of diagonalizing Hermitian matrices in finite

dimensional linear algebra.
Nelson’s idea [Nel69] is to get from selfadjoint operators → cyclic representation
of function algebra → measure µ → L2 (µ).
Sketch proof of Theorem 3.1:
1. Start with a single selfadjoint operator A acting in an abstract Hilbert space

H . Assume A is bounded.
2. Fix u ∈ H . The set { f (A)u}, as f runs through some function algebra,
generates a subspace Hu ⊂ H . Hu is called a cyclic subspace, and u the
corresponding cyclic vector. The function algebra might be taken as the
algebra of polynomials, then later it is extended to a much bigger algebra
containing polynomials as a dense sub-algebra.
3. Break up H into a direct sum of mutually disjoint cyclic subspaces, H =
⊕ j H j , with the family of cyclic vectors u j ∈ H j .
4. Each H j leaves A invariant, and the restriction of A to each H j is unitarily
equivalent to Mx on L2 (sp (A) , µ j ), where sp (A) denotes the spectrum of
A.
5. Piecing together all the cyclic subspace: set
G G
X= sp (A) , µ= µj
j j
i.e., taking disjoint union as u j runs through all the cyclic vectors. When
H is separable, we get H = ⊕ j∈N H j , and we may set µ := ∑∞j=1 2− j µ j .
Details below.
Lemma 3.3. There exists a family of cyclic vector {uα } s.t. H = ⊕α Huα , orthog-
onal sum of cyclic subspaces.
Proof. An application of Zorn’s lemma. See Theorem 4.2.
Lemma 3.4. Set K := [− kAk , kAk]. For each cyclic vector u, there exists a Borel
measure µu s.t. supp (µu ) ⊂ K; and Huα ' L2 (K, µu ).
Proof. The map

f 7→ wu ( f ) := hu, f (A)uiH
is a positive, bounded linear functional on polynomials over K; the latter is dense in

C (K) by Stone-Weierstrass theorem. Hence wu extends uniquely to C(K). (wu is a
state of the C∗ -algebra C (K). See Chapter 4.) By Riesz, there exists a unique Borel
measure µu on K, such that
ˆ
wu ( f ) = hu, f (A) uiH = f dµu .
K
Therefore we get L2 (K, µu ), a Hilbert space containing polynomials as a dense sub-

space. Let
Hu = span{ f (A)u : f ∈ polynomials}
Define W : Hu −→ L2 (K, µu ), by
W f (A) u 7−→ f (3.24)
for polynomials f , which then extends to Hu by density.
Lemma 3.5. Let W be the operator in (3.24), then
1. W is an isometric isomorphism; and

2. WA = Mt W , i.e., W intertwines A and Mt . Hence W diagonalizes A.
Remark 3.5. WA = Mt W ⇐⇒ WAW ∗ = Mt . In finite dimension, it is less empha-

sized that the adjoint W ∗ equals the inverse W −1 . For finite dimensional case,
Mt = diag(λ1 , λ2 , . . . λn ) where the measure µ = ∑finite δλi , where δλi is the Dirac
measure at the eigenvalue λi of A.
Proof. For the first part, let f ∈ L2 (µu ), then

ˆ D E
k f k2L2 (K,µu ) = | f |2 dµu = u, | f |2 (A)u
R H
= u, f¯(A) f (A)u H
∗
= hu, f (A) f (A)uiH
= h f (A)u, f (A)uiH
= k f (A)uk2H .
Notice that strictly speaking, f (A∗ )∗ = f¯(A). Since A∗ = A, therefore f (A∗ )∗ =

π
f¯(A). Also, f −
→ f (A) is a ∗-representation of the algebra C (K); i.e., π is a homo-
morphism, and π( f¯) = f (A)∗ .
For the second part, let f (t) = a0 + a1t + · · · ant n be any polynomial, then
WA f (A) = WA(a0 + a1 A + a2 A2 + · · · + an An )
= W (a0 A + a1 A2 + a2 A3 + · · · + an An+1 )
= a0t + a1t 2 + a2t 3 + · · · + ant n+1
= t f (t)
= Mt W f (A)
thus WA = Mt W . The assertion then follows from standard approximation.
Finally we piece together all the cyclic subspaces.
Lemma 3.6. There exists a locally compact Hausdorff space X and a Borel measure
µ, a unitary operator F : H −→ L2 (X, µ), such that
A = F ∗ Mϕ F
where ϕ ∈ L∞ (µ).
Proof. Recall that we get a family of states w j , with the corresponding measures
µ j , and Hilbert spaces H j = L2 (µ j ). Note that all the L2 -spaces are on K = sp(A).
So it’s the same underlying set, but with possibly different measures.
To get a single measure space with µ, Nelson [Nel69] suggested taking the dis-
joint union
[
X := K × { j}
j
and µ := the disjoint union of µ 0j s. The existence of µ follows from Riesz. Then we
get
F
H = ⊕H j −−→ L2 (X, µ).
Remark 3.6. Note the representation of L∞ (X, µ) onto H = ∑⊕ H j is highly non

unique. There we enter into the multiplicity theory, which starts with breaking up
each H j into irreducible components.
Remark 3.7. A representation π ∈ Rep (A, B (H )) is said to be multiplicity free if

and only if π (A)0 is abelian. We say π has multiplicity equal to n if and only if
(π(A))0 ' Mn (C). This notation of multiplicity generalizes the one in finite dimen-
sional linear algebra. See section 4.9.
Exercise 3.3. Prove that π ∈ Rep(L∞ (µ), L2 (µ)) is multiplicity free. Conclude that
each cyclic representation is multiplicity free (i.e., it is maximal abelian.)
Hint: Suppose B ∈ B L2 (µ) commutes with all Mϕ , ϕ ∈ L2 (µ). Define g = B1,

where 1 is the constant function. Then, for all ψ ∈ L2 (µ), we have
Bψ = Bψ1 = BMψ 1 = Mψ B1 = Mψ g = ψg = gψ = Mg ψ
thus B = Mg .
Corollary 3.4. T ∈ B (H ) is unitary iff there is a unitary operator F : H →

L2 (X, dµ) such that T = FMz F, with |z| ∈ T1 = {z ∈ C : |z| = 1}.
Exercise 3.4 (unbounded s.a. operators).
Finish the proof of Theorem 3.1 for the case when A is unbounded and selfadjoint.
Hint: Suppose A = A∗ , unbounded. The Cayley transform
CA := (A − i) (A + i)−1
is then unitary. See section 9.2. Apply corollary 3.4 to CA , and convert the result
back to A. See, e.g., [Nel69, Rud73].
3.3 Projection-Valued Measure (PVM)
A projection valued measure (PVM) P satisfies the usual axioms of measures (here
Borel measures) but with the main difference:
1. P (4) is a projection for all 4 ∈ B, i.e., P (4) = P (4)∗ = P (4)2 .

2. We assume that P (41 ∩ 42 ) = P (41 ) P (42 ) for all 41 , 42 ∈ B; this
property is called “orthogonality”.
Remark 3.8. The notion of PVM extends the familiar notion of an ONB:
Let H be a separable Hilbert space, and suppose {uk }k∈N is a ONB in H ; then
for 4 ∈ B (R), set
3.3 Projection-Valued Measure (PVM) 87
P (4) := ∑ |uk ih uk | . (3.25)

k∈4
Exercise 3.5. Show that P is a PVM.

Note that, under the summation on the RHS in (3.25), we used Dirac’s notation
|uk ih uk | for the rank-one projection onto Cuk . Further note that the summation is
over k from 4; so it varies as 4 varies.
The PVM version of the spectral theorem says that A = A∗ iff

ˆ
A= xPA (dx)
where P is a projection-valued measure defined on the Borel σ -algebra of R.
Definition 3.4. Let B(C) be the Borel σ -algebra of C. H is a Hilbert space. P :

B(C) → Proj (H ) is a projection-valued measure (PVM), if
1. / = 0, P(C) = I, P (A) is a projection for all A ∈ B

P(0)
2. P(A ∩ B) = P(A)P(B)
3. P(∪k Ek ) = ∑ P(Ek ), Ek ∩ E j = φ if k 6= j. The convergence is in terms of
the strong operator topology. (By assumption, the sequence of projections
∑Nk=1 P(Ek ) is monotone increasing, hence it has a limit, limN→ ∑Nk=1 P(Ek ) =
P(∪Ek ).)
The standard Lebesgue integration extends to PVM.
hϕ, P(E)ϕi = hP(E)ϕ, P(E)ϕi = kP(E)k2 ≥ 0
since P is countably additive, the map E 7→ is also countably additive. Therefore,

each ϕ ∈ H induces a regular Borel measure µϕ on the Borel σ -algebra of R.
For a measurable function ψ,
ˆ ˆ
ψdµϕ = ψ(x) hϕ, P(dx)ϕi
ˆ
= ϕ, ψP(dx) ϕ
hence we may define ˆ

ψP(dx)
as the operator so that for all ϕ ∈ H ,

ˆ
ϕ, ψP(dx) ϕ .
Remark 3.9. P(E) = F χE F −1 defines a PVM. In fact all PVMs come from this way.
In this sense, the Mt version of the spectral theorem is better, since it implies the
PVM version. However, the PVM version facilitates some formulations in quantum
mechanics, so physicists usually prefer this version.
Remark 3.10. Suppose we start with the PVM version of the spectral theorem. How
to prove (ψ1 ψ2 )(A) = ψ1 (A)ψ2 (A)? i.e. how to check we do have an algebra iso-
morphism? Recall in the PVM version, ψ(A) is defined as the operator so that for
all ϕ ∈ H , we have ˆ
ψdµϕ = hϕ, ψ(A)ϕi .
As a standard approximation technique, once starts with simple or even step func-
tions. Once it is worked out for simple functions, the extension to any measurable
functions is straightforward. Hence let’s suppose (WLOG) that the functions are
simple.
Lemma 3.7. We have ψ1 (A) ψ2 (A) = (ψ1 ψ2 ) (A).
Proof. Let
ψ1 = ∑ ψ1 (ti )χEi
ψ2 = ∑ ψ2 (t j )χE j
then
ˆ ˆ
ψ1 P(dx) ψ2 P(dx) = ∑ ψ1 (ti )ψ2 (t j )P(Ei )P(E j )
i, j
= ∑ ψ1 (ti )ψ2 (ti )P(Ei )

i
ˆ
= ψ1 ψ2 P(dx)
where we used the fact that P(A)P(B) = 0 if A ∩ B = φ .
Remark 3.11. As we delve into Nelson’s lecture notes [Nel69], we notice that on
page 69, there is another unitary operator. By piecing these operators together is
precisely how we get the spectral theorem. This “piecing” is a vast generalization of
Fourier series.
3.3 Projection-Valued Measure (PVM) 89
Lemma 3.8. Pick ϕ ∈ H , get the measure µϕ where
µϕ (·) = kP (·) ϕk2
and we have the Hilbert space L2 (µϕ ). Take Hϕ := span{ψ(A)ϕ : ψ ∈ L∞ (µϕ )}.
Then the map
ψ(A)ϕ 7→ ψ
is an isometry, and it extends uniquely to a unitary operator from Hϕ to L2 (µϕ ).
Proof. We have
kψ(A)ϕk2 = hψ(A)ϕ, ψ(A)ϕi

= hϕ, ψ̄(A)ψ(A)ϕi
ˆ
= |ψ (λ )|2 kP (dλ ) ϕk2
R
D E
= ϕ, |ψ|2 (A)ϕ
ˆ
= |ψ|2 dµϕ .
Remark 3.12. Hϕ is called the cyclic space generated by ϕ. Before we can construct
Hϕ , we must make sense of ψ(A)ϕ.
Lemma 3.9 ([Nel69, p.67]). Let p = a0 + a1 x + · · · + an xn be a polynomial. Then

kp(A)uk ≤ max|p(t)|, where kuk = 1 i.e. u is a state.
Proof. M := span{u, Au, . . . , An u} is a finite dimensional subspace in H (automat-

ically closed). Let E be the orthogonal projection onto M. Then
p(A)u = E p(A)Eu = p(EAE)u.
Since EAE is a Hermitian matrix on M, we may apply the spectral theorem for finite
dimensional space and get
EAE = ∑ λk Pλk
where λk0 s are eigenvalues associated with the projections Pλk . It follows that

p(A)u = p(∑ λk Pλk )u = ∑ p(λk )Pλk u
and
2
kp(A)uk2 = ∑ |p(λk )|2 Pλk u
2
≤ max |p(t)| ∑ Pλk u
= max |p(t)|
since
2
∑ Pλk u = kuk2 = 1.
Notice that I = ∑ Pλk .
Remark 3.13. How to extend this? polynomials - continuous functions - measurable

functions. [−kAk, kAk] ⊂ R,
kEAEk ≤ kAk
is a uniform estimate for all truncations. Apply Stone-Weierstrass theorem to the

interval [−kAk, kAk] we get that any continuous function ψ is uniformly approxi-
mated by polynomials. i.e. ψ ∼ pn . Thus
kpn (A)u − pm (A)uk ≤ max|pn − pm |kuk = kpn − pm k∞ → 0
and pn (A)u is a Cauchy sequence, hence
lim pn (A)u =: ψ(A)u

n
where we may define the operator ψ(A) so that ψ(A)u is the limit of pn (A)u.
3.4 Convert Mϕ to a PVM (projection-valued measure)
Theorem 3.3. Let A : H → H be a selfadjoint operator. Then A is unitarily equiv-

alent to the operator Mt of multiplication by the independent variable on the Hilbert
space L2 (µ). There exists a unique projection-valued measure P so that
ˆ
A= tP (dt)
i.e. for all h, k ∈ H , ˆ

hk, AhiH = t hk, P (dt) hiH .
3.4 Convert Mϕ to a PVM (projection-valued measure) 91
Proof. The unique part follows from a standard argument. We will only prove the
existence of P.
Let F : L2 (µ) → H be the unitary operator so that A = FMt F ∗ . Define
P (E) := F χE F ∗
for all E in the Borel σ -algebra B of R. Then P (0)

/ = 0, P (R) = I; and for all
E1 , E2 ∈ B,
P (E1 ∩ E2 ) = F χE1 ∩E2 F −1

= F χE1 χE2 F −1
= F χE1 F −1 F χE2 F −1

= P (E1 ) P (E2 ) .
Suppose {Ek } is a sequence of mutually disjoint elements in B. Let h ∈ H and

write h = F b h ∈ L2 (µ). Then
h for some b
D E
hh, P (∪k Ek ) hiH = F b h, P(∪Ek )F bh
H
D E
−1
= b h, F P (∪Ek ) F b h 2
L (µ)
D E
= b h, χ∪Ek b
h 2
L (µ)
ˆ 2
= h dµ
b
∪Ek
ˆ 2
=∑ h dµ
b
k Ek
= ∑ hh, P(Ek )hiH .

k
Therefore, P is a projection-valued measure.

For any h, k ∈ H , write h = F bh and k = F b
k. Then
D E
hk, AhiH = F b k, AF bh
D EH
∗
= k, F AF h
b b
H
D E
= bk, Mt bh 2
L (µ)
ˆ
= tb k(t)bh(t)dµ(t)
ˆ
= t hk, P (dt) hiH .
´
Thus A = tP(dt).
Remark 3.14. In fact, A is in the closed (under norm or strong topology) span of
{P(E) : E ∈ B}. This is equivalent to say that t = F −1 AF is in the closed span of
the set of characteristic functions, the latter is again a standard approximation in
measure theory. It suffices to approximate tχ[0,∞] .
The wonderful idea of Lebesgue is not to partition the domain, as was the case
in Riemann integral over Rn , but instead the range. Therefore integration over an
arbitrary set is made possible. Important examples include analysis on groups.
Proposition 3.1. Let f : [0, ∞] → R, f (x) = x, i.e. f = xχ[0,∞] . Then there exists a
sequence of step functions s1 ≤ s2 ≤ · · · ≤ f (x) such that limn→∞ sn (x) = f (x).
Proof. For n ∈ N, define


i2−n x ∈ [i2−n , (i + 1)2−n )
sn (x) =
n x ∈ [n, ∞]
where 0 ≤ i ≤ n2−n −1. Equivalently, sn can be written using characteristic functions

as
n2n −1
sn = ∑ i2−n χ[i2−n ,(i+1)2−n ) + nχ[n,∞] .
i=0
Notice that on each interval [i2−n , (i + 1)2−n ),
sn (x) ≡ i2−n ≤ x
sn (x) + 2−n ≡ (i + 1)2−n > x
sn (x) ≤ sn+1 (x).
Therefore, for all n ∈ N and x ∈ [0, ∞],
x − 2−n < sn (x) ≤ x (3.26)
and sn (x) ≤ sn+1 (x).

It follows from (3.26) that
lim sn (x) = f (x)

n→∞
3.4 Convert Mϕ to a PVM (projection-valued measure) 93
for all x ∈ [0, ∞].
Corollary 3.5. Let f (x) = xχ[0,M] (x). Then there exists a sequence of step functions
sn such that 0 ≤ s1 ≤ s2 ≤ · · · ≤ f (x) and sn → f uniformly, as n → ∞.
Proof. Define sn as in proposition (3.1). Let n > M, then by construction
f (x) − 2−n < sn (x) ≤ f (x)
for all s ∈ [0, M]. Hence sn → f uniformly as n → ∞.
Proposition (3.1) and its corollary immediate imply the following.
Corollary 3.6. Let (X, S, µ) be a measure space. A function (real-valued or complex-

valued) is measurable if and only if it is the point-wise limit of a sequence of simple
function. A function is bounded measurable if and only if it is the uniform limit of
a sequence of simple functions. Let {sn } be an approximation sequence of simple
functions. Then sn can be chosen such that |sn (x)| ≤ | f (x)| for all n = 1, 2, 3 . . ..
Theorem 3.4. Let M f : L2 (X, S, µ) → L2 (X, S, µ) be the operator of multiplication

by f . Then,
1. if f ∈ L∞ , M f is a bounded operator, and M f is in the closed span of the

set of selfadjoint projections under norm topology.
2. if f is unbounded, M f is an unbounded operator. M f is in the closed span
of the set of selfadjoint projections under the strong operator topology.
Proof. If f ∈ L∞ , then there exists a sequence of simple functions sn so that sn → f

uniformly. Hence k f − sn k∞ → 0, as n → ∞.
Suppose f is unbounded. By proposition () and its corollaries, there exists a
sequence of simple functions sn such that |sn (x)| ≤ | f (x)| and sn → f point-wisely,
as n → ∞. Let h be any element in the domain of M f , i.e.
ˆ
|h| + | f h|2 dµ < ∞.

Then
lim |( f (x) − sn (x))h(x)|2 = 0
n→∞
and
|( f (x) − sn (x))h(x)|2 ≤ const|h(x)|2 .
Hence by the dominated convergence theorem,

ˆ
lim |( f (x) − sn (x))h(x)|2 dµ = 0
n→∞
or equivalently,
k( f − sn )hk2 → 0
as n → ∞. i.e. Msn converges to M f in the strong operator topology.
Quantum States.
Let H be a Hilbert space (corresponding to some quantum system), and let A be

a selfadjoint operator in H , possibly unbounded. Vectors f ∈ H represent quantum
states if k f k = 1.
The mean of A in the state f is
ˆ
h f,A f i = λ kPA (dλ ) f k2
R
where PA (·) denotes the spectral resolution of A.

The variance of A in the state f is
v f (A) = kA f k2 − (h f , A f i)2 (3.27)

ˆ ˆ 2
2 2 2
= λ kPA (dλ ) f k − λ kPA (dλ ) f k
ˆR R
= (λ − h f , A f i)2 kPA (dλ ) f k2 .

R
´
Exercise 3.6. Let H = L2 (R), and f ∈ L2 (R) given, with k f k2 = | f (x)|2 dx = 1.
Suppose f ∈ dom (P) ∩ dom (Q), where P and Q are the momentum and position
operators, respectively. Show that
1
v f (P) v f (Q) ≥ . (3.28)
4
Inequality (3.28) is the mathematically precise form of Heisenberg’s uncertainty

relation, often written in the form
1
σ f (P) σ f (Q) ≥
2
p p
where σ f (P) = v f (P), and σ f (Q) = v f (Q).
3.5 The Spectral theorem for compact operators 95
3.5 The Spectral theorem for compact operators
3.5.1 Preliminaries
The setting for the first of the Spectral Theorems we will consider is as following:
(Restricting assumptions will be relaxed in subsequent versions!)
Let H be a separable (typically infinite dimensional Hilbert space assumed
here!) and let A ∈ B (H ) \ {0} be compact and selfadjoint, i.e., A = A∗ , and A
is in the k·kUN -closure of F R (H ). See section 1.5.1.
Theorem 3.5. With A as above, there is an ONB {uk }k∈N , and a sequence {λk }k∈N
s.t. |λ1 | ≥ |λ2 | ≥ · · · ≥ |λk | ≥ |λk+1 | ≥ · · · , limk→∞ λk = 0, and
1. Auk = λk uk , k ∈ N,
2. A = ∑∞
k=1 λk |uk ih uk |,
3. spec (A) = {λk } ∪ {0},
dim v ∈ H Av = λk v < ∞, for all k ∈ N.

4.
The ONB {uk } contains an ONB (possibly infinite) subset, an ONB for the sub-
space
Ker (A) = v ∈ H Av = 0 .

We begin with some preliminaries in the preparation for the proof.
Theorem 3.6. Let f : X × X → C be a sesquilinear form on any set X, conjugate

linear in the first variable and linear in the second variable, then f satisfies the
polarization identity
1 3 k k
f (x, y) = ∑ i f i x + y, ik x + y .
4 k=0
Proof. Follows from a direct computation. Specifically, for all x, y ∈ X, we have
f (x + y, x + y) = f (x, x) + f (x, y) + f (y, x) + f (y, y)

f (ix + y, ix + y) = f (x, x) − i f (x, y) + i f (y, x) + f (y, y)
f (−x + y, −x + y) = f (x, x) − f (x, y) − f (y, x) + f (y, y)
f (−ix + y, −ix + y) = f (x, x) + i f (x, y) − i f (y, x) + f (y, y)
then
f (x + y, x + y) − f (−x + y, −x + y) = 2 ( f (x, y) + f (y, x))

i f (ix + y, ix + y) − i f (−ix + y, −ix + y) = 2 ( f (x, y) − f (y, x))
adding the above two equations gives
3
4 f (x, y) = ∑ ik f ik x + y, ik x + y
k=0
and this is the desired result.

Corollary 3.7. A bounded operator A in H is selfadjoint if and only if
hx, Axi ∈ R, ∀x ∈ H .
Proof. Suppose A is selfadjoint, i.e., hx, Ayi = hAx, yi, ∀x, y ∈ H . Setting x = y, we
get
hx, Axi = hAx, xi = hx, Axi =⇒ hx, Axi ∈ R, ∀x ∈ H .
Conversely, suppose hx, Axi ∈ R, ∀x ∈ H . Note that hx, Ayi and hAx, yi are both
sesquilinear forms defined on H , by theorem 3.6, they satisfy the corresponding
polarization identity. Thus,
4 hx, Ayi = hx + y, A (x + y)i + i hix + y, A (ix + y)i −

h−x + y, A (−x + y)i − i h−ix + y, A (−ix + y)i
and
4 hAx, yi = hA (x + y) , x + yi + i hA (ix + y) , ix + yi −
hA (−x + y) , −x + yi − i hA (−ix + y) , −ix + yi .
Since hx, Axi ∈ R, ∀x ∈ H , the RHSs of the above equations are equal, therefore,
hx, Ayi = hAx, yi, ∀x, y ∈ H , i.e., A is selfadjoint.
Theorem 3.7. Let H be a Hilbert space over C. Let f : H × H → C be a
sesquilinear form, and M := sup {| f (x, y)| : kxk = kyk = 1} < ∞. There exists a
unique bounded operator A in H , satisfying
f (x, y) = hx, Ayi , ∀x, y ∈ H ; and kAk = M.
Proof. Note that given x, y ∈ H , x, y nonzero, we have


x y
f , ≤ M ⇐⇒ | f (x, y)| ≤ M kxk kyk . (3.29)
kxk kyk
Thus, for each y ∈ H , the map x 7→ f (x, y) is a bounded linear functional on H .

By Riesz’s representation theorem, there exists a unique element Ay ∈ H , such that
f (x, y) = hx, Ayi .
We check y 7→ Ay is linear. For this, let c ∈ C, then
f (x, y1 + cy2 ) = hx, A (y1 + cy2 )i ; (3.30)
on the other hand, since f is linear in the second variable, and so
f (x, y1 + cy2 ) = f (x, y1 ) + c f (x, y2 ) (3.31)

= hx, Ay1 i + c hx, Ay2 i
= hx, Ay1 + cAy2 i .
Combining (3.30)-(3.30), we get
hx, A (y1 + cy2 )i = hx, Ay1 + cAy2 i , ∀x ∈ H ;
it follows that y 7→ Ay is linear.

By (3.29), we have
|hx, Ayi| ≤ M kxk kyk , ∀x, y ∈ H .
That is, kAyk ≤ M kyk, ∀y ∈ H , so that kAk ≤ M. But
| f (x, y)| = |hx, Ayi|

≤ kxk kAyk
≤ kAk kxk kyk , ∀x, y ∈ H ;
which then implies M ≤ A. Thus, we conclude that M = A.
Corollary 3.8. Any bounded operator A in H is uniquely determined by the corre-

sponding sesquilinear form hx, Ayi, x, y ∈ C.
Proof. Set f (x, y) = hx, Ayi, x, y ∈ H , in theorem 3.7. By the uniqueness part in the
theorem, A is recovered from f (·, ·).
Corollary 3.9. For any bounded operator A in H , the norm of A is given by
kAk = sup {|hx, Ayi| : kxk = kyk = 1} . (3.32)
Proof. Given A, f (x, y) := hx, Ayi is a sesquilinear form; and (3.32) follows imme-
diately from theorem 3.7.
Corollary 3.10. Let A be a bounded selfadjoint operator in H , then
kAk = sup {|hx, Axi| : kxk = 1} . (3.33)
Proof. Set M := sup {|hx, Axi| : kxk = 1}. Then

( )
|hx, Axi| ≤ kAxk kxk Cauchy-Schwarz
=⇒ M ≤ kAk .
≤ kAk kxk2
We proceed to verify that kAk ≤ M.

Using the polarization identity in theorem 3.7, we have
4 hx, Ayi = hA (x + y) , x + yi − hA (−x + y) , −x + yi

+ i hA (ix + y) , ix + yi − i hA (−ix + y) , −ix + yi .
Sine A is selfadjoint, the four inner products on RHS of the above equation are all
real-valued (3.7). Thus,
1
ℜ {hx, Ayi} = (hA (x + y) , x + yi − hA (−x + y) , −x + yi) .
4
Now, there exists a phase factor, eiθ depending on x, y, s.t.
|hx, Ayi| = eiθ hx, Ayi = ℜ {hx, Ayi} ;
where the second equality follows from the fact that eiθ hx, Ayi ∈ R. Then,
1
|hx, Ayi| = |hA (x + y) , x + yi − hA (−x + y) , −x + yi|
4
1
= |hA (x + y) , x + yi − hA (x − y) , x − yi|
4
1
≤ M kx + yk2 + kx − yk2
4
1
= kMk 2 kxk2 + 2 kyk2
4
= kMk , assuming kxk = kyk = 1;
therefore, by 3.9,
kAk = sup {|hx, Ayi| : kxk = kyk = 1} ≤ M.
we suggest: [Con90, FL28, Kat95, Kre55, Lax02, LP89, Nel69, RS75, Rud73,
Sto51a, Sto51b, Yos95, HJL+ 13, DHL09, HKLW07].
Part III
Applications
Chapter 4
GNS and Representations
If one finds a difficulty in a calculation which is otherwise quite

convincing, one should not push the difficulty away; one should
rather try to make it the centre of the whole thing.
— Werner Heisenberg
Since states in quantum physics are vectors (of norm one) in Hilbert space; two
question arise: “Where does the Hilbert space come from?” And “What are the
algebras of operators from which the selfadjoint observables must be selected?” In
a general framework, we offer an answer below, it goes by the name “the Gelfand-
Naimark-Segal (GNS) theorem, ” which offers a direct correspondence between
states and cyclic representations.
4.1 Definitions and Facts: An Overview
Let A be an algebra over C, with an involution A 3 a 7→ a∗ ∈ A, and the unit-element

1. Let A+ denote the set of positive elements in A; i.e.,
A+ = b∗ b b ∈ A .

Definition 4.1. We say A is a C∗ -algebra if it is complete in a norm k·k, which

satisfies:
(C∗ 1) kabk ≤ kak kbk, ∀a, b ∈ A;

(C∗ 2) k1k = 1;
(C∗ 3) kb∗ bk = kbk2 , ∀b ∈ A.1
Example 4.1. Let X be a compact Hausdorff space, the algebra C (X) of all continu-
ous function on X is a C∗ -algebra under the sup-norm.
1Kadison et al. in 1950’s reduced the axioms of C∗ -algebra from about 6 down to just one (C∗ 3)
on the C∗ -norm.
103
104 4 GNS and Representations
Example 4.2. B (H ): all bounded linear operators on a fixed Hilbert space H is a

C∗ -algebra.
Example 4.3. ON : the Cuntz-algebra, N > 1; it is the C∗ -completion of N generators

s1 , s2 , . . . , sN satisfying the following relations [Gli60, Gli61, BJO04]:
(ON 1) s∗i s j = δi j 1;
(ON 2) ∑Ni=1 si s∗i = 1.
Definition 4.2. We denote Rep (A, H ) the representations of A acting on some

Hilbert space H , i.e., π ∈ Rep (A, H ) iff π : A → B (H ) is a homomorphism
of ∗ -algebras, π(1) = IH = the identity operator in H ; in particular
hπ (b) u, viH = hu, π (b∗ ) viH , ∀b ∈ A, ∀u, v ∈ H . (4.1)
Let S (A) be the states ϕ : A → C on A; i.e., (axioms) ϕ ∈ A∗ = the dual of A,

ϕ (1) = 1, and
ϕ (b∗ b) ≥ 0, ∀b ∈ A. (4.2)
Theorem 4.1 (Gelfand-Naimark-Segal (GNS)). There is a bijection:
S (A) ←→ cyclic representations, up to unitary equivalence
as follows:
←− (easy direction): Given π ∈ Rep (A, H ), u0 ∈ H , ku0 k = 1, set
ϕ (a) = hu0 , π (a) u0 iH , ∀a ∈ A. (4.3)
−→ (non-trivial direction): Given ϕ ∈ S (A), there is a system (π, H , u0 ) such

that (4.3) holds. (Notation, we set π = πϕ to indicate the state ϕ.)
Proof. (−→) Given ϕ ∈ S (A), then on A × A consider the sesquilinear form
(a, b) 7−→ ϕ (a∗ b) (4.4)
∼
Hϕ = A/ b ∈ A ϕ (b∗ b) = 0

where ∼ refers to Hilbert completion in (4.4). Note
|ϕ (a∗ b)|2 ≤ ϕ (a∗ a) ϕ (b∗ b) ,∀a, b ∈ A.

4.1 Definitions and Facts: An Overview 105
Set Ω = class (1) in Hϕ , and
πϕ (a) (class (b)) = class (ab) , ∀a, b ∈ A.
Then it is easy to show that Hϕ , Ω , πϕ satisfies conclusion (4.3), i.e., ϕ (a) =

hΩ , π (a) Ω iHϕ , ∀a ∈ A.
Definition 4.3. Let ϕ ∈ S (A), we say it is a pure state iff ϕ ∈ extS (A) := the
extreme-points in S (A).
Facts 4.1
1. If A is a C∗ -algebra , then S (A) (⊂ A∗ ) is convex and weak *-compact.
2. Given ϕ ∈ S (A), and let πϕ ∈ Rep (A, H ) be the GNS-representation, see
(4.3); then
ϕ ∈ extS (A) , i.e., it is pure

m
πϕ is an irreducible representation
´
3. If ψ ∈ S (A), ∃ a measure Pψ on ext (S (A)) s.t. ψ = w dPψ (w), and then
ˆ ⊕
πψ = πw dPψ (w) .
Example 4.4 (Pure states, cases where the full list is known!). ϕ ∈ S (A):
A extS (A)
C (X) points x ∈ X, and ϕ = δx (Dirac mass); ϕ ( f ) = f (x), ∀ f ∈ C (X)
B (H ) v ∈ H , kvk = 1, ϕ = ϕv ; ϕv (A) = hv, Avi, ∀A ∈ B (H )
Partial list: u = (u1 , . . . , uN ) ∈ CN , ∑N1 |u j |2 = 1, ϕ = ϕu specified

ON by ϕ(si s∗j ) = ui u j , ∀i, j = 1, . . . , N; see (ON 1)-(ON 2).
Table 4.1: Examples of pure states.
Exercise 4.1. Using GNS, write down explicitly the irreducible representations of
the three C∗ -algebras in Table 4.1 corresponding to the listed pure states. (Hint: In
the case of C (X), the representations are one-dimensional, but in the other cases,
they are infinite-dimensional, i.e., dim Hπϕ = ∞.
Groups
Case 1. Groups contained in B (H ) where H is a fixed Hilbert space:
Definition 4.4. Set

B (H )−1 : all bounded linear operators in H with bounded inverse.
B (H )uni : all unitary operators u : H → H , i.e., u satisfies uu∗ = u∗ u = IH .
Definition 4.5. Fix a group G, and set:

Rep (G, H ): all homomorphisms, ρ ∈ G → B (H )−1
Repuni (G, H ): all homomorphisms, ρ : G → Buni (H ), i.e.,
ρ(g−1 ) = ρ (g) −1 = ρ (g)∗ , ∀g ∈ G
Repcont (G, H ): Elements ρ ∈ Rep (G, H ) s.t. ∀v ∈ H , G 3 g 7→ ρ (g) v is con-

tinuous from G into H ; called strongly continuous.
Remark 4.1. In the case of Repcont (G, H ) it is assumed that G is a continuous

group, i.e., is equipped with a topology s.t. the two operations:
(Multip.) G × G 3 (g1 , g2 ) 7−→ g1 g2 ∈ G

(Inv.) G 3 g 7−→ g−1 ∈ G
are both continuous.
Exercise 4.2. Let G be a locally compact group with µ = a left-invariant Haar mea-
sure. Set (ρL (g) f ) (x) := f g−1 x , g, x ∈ G, f ∈ L2 (G, µ). Then show that ρL is a

strongly continuous unitary representation of G acting in L2 (G, µ).
The Group Algebra

Let G be a group, and set C[G] := all linear combinations, i.e., finite sums
A = ∑ Ag g (4.5)
g
where Ag ∈ C, and making A := C[G] into a ∗-algebra with the following two oper-
ations on finite sums as in (4.5): A × A −→ A, given by
! ! !
∑ Ag g ∑ Bh h := ∑ ∑ Ah Bk g (4.6)
g∈G h∈G g hk=g
4.2 The GNS Construction 107
and !∗
∑ Ag g := ∑ Ag g−1 . (4.7)
g∈G g∈G
Fact 4.1 There is a bijection between Repuni (G, H ) and Rep (C[G], H ) as fol-
lows: If π ∈ Repuni (G, H ), set πe ∈ Rep (C[G], H ):
!
πe ∑ Ag g := ∑ Ag π (g) (4.8)
g∈G g∈G
where the element ∑g∈G Ag g in (4.8) is a generic element in C[G], see (4.5), i.e., is
a finite sum with Ag ∈ C, for all g ∈ G.
Exercise 4.3. Fill in the proof details of the assertion in Fact 4.1.
Example 4.5. Let G be a group, considered as a countable discrete group (the count-
ability is not important). Set H = l 2 (G), and
π (g) δh := δgh , ∀g, h ∈ G. (4.9)
Exercise 4.4. Show that π in (4.9) is in Rep G, l 2 (G) .

Definition 4.6. Let G, and π ∈ Rep G, l 2 (G) be as in (4.9), and let πe ∈ Rep C[G], l 2 (G)

be the corresponding representation of C[G]; see Fact 4.1. Set
∗
(G) := the norm closure of πe (C[G]) ⊂ B l 2 (G) ;

Cred
∗ (G) is called the radical C∗ -algebra of the group G.

then Cred
∗ (G) is a C∗ -algebra.
Exercise 4.5. Prove that Cred
∗ (F ) is simple, where F is the free group
Remark 4.2. It is known [Pow75] that Cred 2 2
on two generators. (“red” short for reduced; it is called the reduced C∗ -algebra on
the group.)
4.2 The GNS Construction
The GNS construction is a general principle for getting representations from given
data in applications, especially in quantum mechanics. It was developed indepen-
dently by I. Gelfand, M. Naimark, and I. Segal around 1960s.
Definition 4.7. Let A be a ∗-algebra with identity. A representation of A is a map

π : A → B(Hπ ), where Hπ is a Hilbert space, s.t. for all A, B ∈ A,
1. π(AB) = π(A)π(B)
2. π(A∗ ) = π(A)∗
The ∗ operation (involution) is given on A so that A∗∗ = A, (AB)∗ = B∗ A∗ ,

(λ A)∗ = λ A∗ , for all λ ∈ C.
Example 4.6. The multiplication version of the spectral theorem of a single self-
adjoint operator, say A acting on H , yields a representation of the algebra of
L∞ (sp (A)) (or C(sp (A))) as operators on H , where
L∞ (sp (A)) 3 f −→ f (A) ∈ B(H )

π
such that π( f g) = π( f )π(g) and π( f¯) = π( f )∗ .
The general question is given any ∗-algebra, where to get such a representation?
The answer is given by states. One gets representations from algebras vis states.
For abelian algebras, the states are Borel measures, so the measures come out as a
corollary of representations.
Definition 4.8. Let A be a ∗-algebra. A state on A is a linear functional ϕ : A → C

such that ϕ(1A ) = 1, and ϕ(A∗ A) ≥ 0, for all A ∈ A.
Example 4.7. Let A = C(X), i.e., C∗ -algebra of continuous functions on a compact

Hausdorff space X. Note that there is a natural involution f 7→ f ∗ := f by complex
conjugation. Let µϕ be a Borel probability measure on X, then
ˆ
C (X) 3 f 7→ ϕ( f ) = f dµϕ
X
is a state. In fact, in the abelian case, all states are Borel probability measures.
We are aiming at a proof of the GNS theorem (Theorem 4.1), and a way to get more
general representations of ∗-algebras. Indeed, any representation is built up by the
cyclic representations (Def. 4.9), and each cyclic representation is in turn given by
a GNS construction.
Definition 4.9. A representation π ∈ Rep(A, H ) is called cyclic, with a cyclic vec-

tor u ∈ H , if H = span π (A) u u ∈ A .

Theorem 4.2. Give any representation π ∈ Rep(A, H ), there exists an index set J,
closed subspaces H j ⊂ H such that
1. Hi ⊥ H j , ∀i 6= j;
2. ∑⊕j∈J H j = H ; and
3. there exists cyclic vectors v j ∈ H j such that the restriction of π to H j is
cyclic.
Remark 4.3. The proof of 4.2 is very similar to the construction of orthonormal
basis (ONB); but here we get a family of mutually orthogonal subspaces. Of course,
if H j ’s are all one-dimensional, then it is a decomposition into ONB. Note that not
every representation is irreducible, but every representation can be decomposed into
direct sum of cyclic representations.
Exercise 4.6. Prove Theorem 4.2. Hint: pick v1 ∈ H , and let
Hv1 := span {π (A) v1 : A ∈ A} ,
i.e., the cyclic subspace generated by v1 . If Hv1 6= H , then ∃v2 ∈ H \Hv1 , and the
cyclic subspace Hv2 , so that Hv1 and Hv2 are orthogonal. If Hv1 ⊕ Hv2 6= H , we
then build Hv3 and so on. Now use transfinite induction or Zorn’s lemma to show
the family of direct sum of mutually orthogonal cyclic subspaces is total. The final
step is exactly the same argument for the existence of an ONB of any Hilbert space.
Now we proceed to prove the GNS theorem (Theorem 4.1), which is restated below.
Theorem 4.3 (Gelfand-Naimark-Segal). There is a bijection between states ϕ and

cyclic representations π ∈ Rep (A, H , u), with kuk = 1; where
ϕ(A) = hu, π(A)ui , ∀A ∈ A. (4.10)
Moreover, fix a state ϕ, the corresponding cyclic representation is unique up to uni-

tary equivalence. Specifically, if (π1 , H1 , u1 ) and (π2 , H2 , u2 ) are two cyclic repre-
sentations, with cyclic vectors u1 , u2 , respectively, satisfying
ϕ(A) = hu1 , π1 (A)u1 i = hu2 , π2 (A)u2 i , ∀A ∈ A; (4.11)
then
W : π1 (A) u1 7−→ π2 (A) u2 , A ∈ A (4.12)
extends to a unitary operator from H1 to H2 , also denoted by W , and such that
π2 = W π1 . (4.13)
Remark 4.4. For the non-trivial direction, let ϕ be a given state on A, and we need
to construct a cyclic representation (π, Hϕ , uϕ ). Note that A is an algebra, and it
is also a complex vector space. Let us try to turn A into a Hilbert space and see
what conditions are needed. There is a homomorphism A → A which follows from
the associative law of A being an algebra, i.e., (AB)C = A(BC). To continue, A
should be equipped with an inner product. Using ϕ, we may set hA, Biϕ := ϕ (A∗ B),
∀A, B ∈ A. Then h·, ·iϕ is linear in the second variable, and conjugate linear in the
first variable. It also satisfies hA, Aiϕ = ϕ (A∗ A) ≥ 0. Therefore we take Hϕ :=
[A/{A : ϕ(A∗ A) = 0}]cl .
Proof. Given a cyclic representation π ∈ Rep (A, H , u), define ϕ as in (4.10).

Clearly ϕ is linear, and
ϕ (A∗ A) = hu, π(A∗ A)ui

= hu, π(A∗ )π (A) ui
= hπ(A)u, π (A) ui
= kπ (A) uk2 ≥ 0.
Thus ϕ is a state.
Conversely, fix a state ϕ on A. Set
( )
n
H0 := ∑ ci Ai ci ∈ C, n ∈ N
i=1
and define the inner product
∑ ci Ai , ∑ di Bi ϕ
:= ∑∑ ci d j ϕ (A∗i B j ) .
Note that, by definition,
2
∑ ci Ai ϕ
= ∑ ci Ai , ∑ ci Ai ϕ
= ∑∑ ci c j ϕ (A∗i A j ) ≥ 0. (4.14)
The RHS of (4.14) is positive since ϕ is a state. Recall that ϕ (A∗ A) ≥ 0, for all
A ∈ A, and this implies that for all n ∈ N, the matrix (ϕ (A∗i A j ))ni, j=1 is positive
definite, hence (4.14) holds.
Now, let Hϕ := completion of H0 under h·, ·iϕ modulo elements s s.t. kskϕ = 0.
See Lemma 4.1 below. Hϕ is the desired cyclic space, consisting of equivalence
classes [A], ∀A ∈ A. Next, let uϕ = [1A ] = equivalence class of the identity element,
and set
π (A) := [A] = [A1A ] = [A][1A ];
then one checks that π ∈ Rep A, Hϕ , and ϕ (A) = uϕ , π (A) uϕ

ϕ
, ∀A ∈ A.
For uniqueness, let (π1 , H1 , u1 ) and (π2 , H2 , u2 ) be as in the statement of the
theorem, and let W be as in (4.12). By (4.11), we have
ϕ (A∗ A) = kπ2 (A) u2 k2 = kπ1 (A) u1 k2
so that W is isometric. But since Hi = span {πi (A) ui : A ∈ A}, i = 1, 2, then W

extends by density to a unitary operator from H1 to H2 .
Finally, for all A, B ∈ A, we have
W π1 (A) (π1 (B) u1 ) = W π1 (AB) u1

= π2 (AB) u2
= π2 (A) (π2 (B) u2 )
therefore, by the density argument again, we conclude that π2 (A) = W π1 (A), ∀A ∈

A. This is the intertwining property in (4.13).
Lemma 4.1. {A ∈ A : ϕ(A∗ A) = 0} is a closed two-sided ideal in A.
Proof. This follows from the Schwarz inequality. Note that

" #
ϕ (A∗ A) ϕ (A∗ B)
ϕ (B∗ A) ϕ (B∗ B)
is a positive definite matrix, and so its determine is positive, i.e.,
|ϕ (A∗ B)|2 ≤ ϕ (A∗ A) ϕ (B∗ B) ; (4.15)
using the fact that ϕ (C∗ ) = ϕ (C)∗ , ∀C ∈ A. The lemma follows from the estimate
(4.15).
Example 4.8. Let A = C[0, 1]. Set ϕ : f 7→ f (0), so that ϕ ( f ∗ f ) = | f (0)|2 ≥ 0. Then,
ker ϕ = { f ∈ C [0, 1] f (0) = 0}
and C [0, 1] / ker ϕ is one dimensional. The reason is that if f ∈ C[0, 1] such that
f (0) 6= 0, then we have f (x) ∼ f (0) since f (x)− f (0) ∈ ker ϕ, where f (0) represents
the constant function f (0) over [0, 1]. This shows that ϕ is a pure state, since the
representation has to be irreducible.
Exercise 4.7. Fill in the details in the above proof of the GNS theorem.
Using GNS construction we get the following structure theorem for abstract C∗ -
algebras. As a result, all C∗ -algebras are sub-algebras of B (H ) for some Hilbert
space H .
Theorem 4.4 (Gelfand-Naimark). Every C∗ -algebra (abelian or non-abelian) is
isometrically isomorphic to a norm-closed sub-algebra of B(H ), for some Hilbert
space H .
Proof. Let A be any C∗ -algebra, no Hilbert space H is given from outside. Let
S(A) be the states on A, which is a compact convex subset of the dual space A∗ .
Here, compactness refers to the weak ∗-topology.
We use Hahn-Banach theorem to show that there are plenty of states. Specifically,
∀a ∈ A, ∃ϕ ∈ A∗ such that ϕ(a) > 0. It is done first on the 1-dimensional subspace
tA 7→ t ∈ R,
and then extends to A. (Note this is also a consequence of Krein–Milman, i.e.,

S(A) = cl(pure states). We will come back to this point later.)
For each state ϕ, one gets a cyclic representation (πϕ , Hϕ , uϕ ). Applying transfi-
nite induction, one concludes that π := ⊕πϕ is a representation on the Hilbert space
H := ⊕Hϕ . For details, see e.g., [Rud73].
Theorem 4.5. Let A be an abelian C∗ -algebra. Then there is a compact Hausdorff

∼ C(X).
space X, unique up to homeomorphism, such that A =
4.3 States, Dual and Pre-dual
Let V be a Banach space, i.e.,

4.3 States, Dual and Pre-dual 113
• V is a vector space over C;

• ∃ norm k·k
• V is complete with respect to k·k
The dual space V ∗ consists of linear functionals l : V → C satisfying
klk := sup |l (v)| < ∞.

kvk=1
These are the continuous linear functionals. The Hahn-Banach theorem implies that
for all v ∈ V , kvk 6= 0, there exists lv ∈ V ∗ such that l(v) = kvk. Recall the con-
struction is to define lv on one vector, then use transfinite induction to extend it to
V . Notice that V ∗ is always complete, even if V is an incomplete normed space. In
other words, V ∗ is always a Banach space.
Now V is embedded into V ∗∗ (as we always do this) via the mapping
V 3 v 7→ ψ (v) ∈ V ∗∗ , where
ψ (v) (l) := l (v) , ∀l ∈ V ∗ .
Example 4.9. Let X be a compact Hausdorff space. The algebra C(X) of all con-
tinuous functions on X with the sup norm, i.e., k f k∞ := supx∈X | f (x)|, is a Banach
space.
Example 4.10. The classical L p space: (l p )∗ = l q , (L p )∗ = Lq , for 1/p + 1/q = 1

and 1 ≤ p < ∞. If 1 < p < ∞, then (l p )∗∗ = l p , i.e., these spaces are reflexive. For
p = 1, however, we have (l 1 )∗ = l ∞ , but (l ∞ )∗ is much bigger than l 1 . Also note that
(l p )∗ 6= l q except for p = q = 2. And l p is a Hilbert space iff p = 2.
Let B be a Banach space and denote by B∗ its dual space. B∗ is a Banach space as
well, where the norm is defined by
k f kB∗ = sup | f (x)|.

kxk=1
Let B∗1 = { f ∈ B∗ : k f k ≤ 1} be the unit ball in B∗ .
Theorem 4.6 (Banach-Alaoglu). B∗1 is weak ∗ compact in B∗ .
Proof. This is proved by showing B∗1 is a closed subspace in ∏kxk=1 C, where the
latter is given its product topology, and is compact and Hausdorff.
As an application, we have
Corollary 4.1. Every bounded sequence in B∗ has a convergent subsequence in the

l.c. weak ∗-topology.
Corollary 4.2. Every bounded sequence in B(H ) contains a convergence subse-

quence in the l.c. weak ∗-topology.
We show in Theorem 4.7 that B (H ) = T1 (H )∗ , where T1 (H ) = trace-class op-

erators.
Now we turn to Hilbert space, say H :
• H is a vector space over C;

p
• it has an inner product h·, ·i, and the norm k·k := h·, ·i;
• H is complete with respect to k·k;
• H ∗ = H , i.e., H is reflexive;
• every Hilbert space has an orthonormal basis (by Zorn’s lemma)
The identification H = H ∗ is due to Riesz, and the corresponding map is given by
h 7→ hh, ·i ∈ H ∗
This can also be seen by noting that H is unitarily equivalent to l 2 (A), with some
index set A, and l 2 (A) is reflexive.
The set of all bounded operators B(H ) on H is a Banach space. We ask two
questions:
1. What is the dual B(H )∗ ?

2. Is B(H ) the dual space of some Banach space?
The first question is extremely difficult and we will discuss that later. For the present
section, we show that
B(H ) = T1 (H )∗
where we denote by T1 (H ) the trace-class operators in B(H ). (Details below! See

Theorem 4.7, and section 1.5.1.)
Let ρ : H → H be a compact selfadjoint operator. Assume ρ is positive, i.e.,
hx, ρxi ≥ 0 for all x ∈ H . By the spectral theorem of compact operators, we get the
following decomposition
ρ = ∑ λk Pk (4.16)
where λ1 ≥ λ2 ≥ · · · → 0, and Pk is the projection onto the finite dimensional

eigenspace of λk .
In general, we want to get rid of the assumption that ρ ≥ 0. This is done using
the polar decomposition, which we will consider in section 2.4 even for unbounded
operators. It is much easier for bounded operators: If A ∈ B (H ), A∗ A is positive,
√
selfadjoint, and so by the spectral theorem, we may take |A| := A∗ A. Then, one
checks that
D√ √ E
kAxk2 = hAx, Axi = hx, A∗ Axi = A∗ Ax, A∗ Ax = k|A| xk2 ,
thus
kAk = k|A|k (4.17)
and there is a partial isometry V : range (|A|) → range (A), and the following polar
decomposition holds:
A = V |A| (4.18)
We will come back to this point in section 2.4 when we consider unbounded opera-
tors.
Corollary 4.3. Let A ∈ T1 (H ), then A has the following decomposition
A = ∑ λn | fn ih en | (4.19)
n
where {en } and { fn } are ONBs in H .
Proof. Using the polar decomposition A = V |A|, we may first diagonalize |A| with
respect to some ONB {en } as
|A| = ∑ λn |en ih en | , then

n
A = V |A| = ∑ λn |Ven ih en | = ∑ λn | fn ih en |
n n
where fn := Ven .
With the above discussion, we may work, instead, with compact operators A :
H → H so that A is a trace class operator if |A| (positive, selfadjoint) satisfies
condition (4.20).
Definition 4.10. Let A be a compact operator with its polar decomposition A =

√
V |A|, where |A| := A∗ A. Let {λk }∞
k=1 be the eigenvalues of |A| as in (4.16). We
say A is a trace class operator, if
kAk1 := trace (|A|) = ∑ λn < ∞; (4.20)

n
in which case we may assume ∑ λn = 1.
We now continue the discussion from section 1.5.1 on spaces of operators.
Definition 4.11. Let A ∈ T1 (H ), and {en } an ONB in H . Set
trace (A) := ∑ hen , Aen i (4.21)

n
Note the RHS in (4.21) is independent of the choice of the ONB. For if { fn } is
another ONB in H , using the Parseval identity repeatedly, we have
∑ h fn , A fn i = ∑ ∑ h fn , em i hem , A fn i
n m
= ∑ ∑ h fn , em i hA∗ em , fn i
m n
= ∑ hA∗ em , em i
m
= ∑ hem , Aem i .
m
Corollary 4.4. Let A ∈ T1 (H ), then
|trace (A)| ≤ kAk1 .
Therefore, the RHS in (4.21) is absolutely convergent.
Proof. By Corollary 4.3, there exists ONBs {en } and { fn }, and A has a decomposi-
tion as in (4.19). Then,
|trace (A)| ≤ ∑ |hen , Aen i|

n
= ∑ λn |hen , fn i|
n
≤ ∑ λn ken k k fn k
n
= ∑ λn = kAk1 < ∞.
n
We have used the fact that trace (A) is independent of the choice of an ONBs.
Lemma 4.2. Let T1 (H ) be the trace class introduced above. Then,

1. T1 (H ) is a two-sided ideal in B(H ).

2. trace(AB) = trace(BA)
3. T1 (H ) is a Banach space with respect to the trace norm (4.20).
Exercise 4.8. Prove Lemma 4.2.
Lemma 4.3. Let ρ ∈ T1 (H ), then the map A 7→ trace(Aρ) is a state on B(H ).
Proof. By Lemma 4.2, Aρ ∈ T1 (H ) for all A ∈ B(H ). The map A 7→ trace(Aρ)

is in B(H )∗ means that the pairing (A, ρ) 7→ trace(Aρ) satisfies
|trace(Aρ)| ≤ kAk kρk1 .
By Corollary 4.4, it suffices to verify, instead, that
kAρk1 ≤ kAk kρk1 .
Indeed, if we choose an ONB {en } in H that diagonalizes |ρ|, i.e.,
|ρ| = ∑ λn |en ih en | , where ∑ λn < ∞, λn > 0, ∀n;

then
p
kAρk1 = trace ρ ∗ A∗ Aρ
p √
= trace ρ ∗ ρ A∗ A
= ∑ hen , |A| |ρ| en i
n
= ∑ λn hen , |A| en i
n
≤ kAk ∑ λk = kAk kρk1 .
k
Theorem 4.7. T1∗ (H ) = B(H ).
Proof. Let l ∈ T1∗ . How to get an operator A? It is supposed to be the case such that
l(ρ) = trace(ρA)
for all ρ ∈ T1 . How to pull an operator A out of the hat? The idea also goes back
to Dirac. It is in fact not difficult to find A. Since A is determined by its matrix, it
suffices to fine h f , A f i which are the entries in the matrix of A.
Proof. For any f1 , f2 ∈ H , the rank-one operator | f1 ih f2 | is in T1 , hence we know

what l does to it, i.e., we know the numbers l (| f1 ih f2 |). But since l (| f1 ih f2 |) is
linear in f1 , and conjugate linear in f2 , by the Riesz theorem for Hilbert space, there
exists a unique operator A such that
l(| f1 ih f2 |) = h f2 , A f1 i .
Now we check that l(ρ) = trace(ρA). By Corollary 4.3, any ρ ∈ T1 can be written
as ρ = ∑n λn | fn ih en |, where {en } and { fn } are some ONBs in H . Then,

trace (ρA) = trace | f
∑ n n n
λ ih e | A
n

= trace ∑ λn |A fn ih en |
n
= ∑ ∑ λn hum , A fn i hen , um i
m n

= ∑ λn ∑ hum , A fn i hen , um i
n m
= ∑ λn hen , A fn i (= l (ρ))
n
where {un } is an ONB in H , and the last step follows from Parseval’s identity.
Remark 4.5. If B is the dual of a Banach space, then we say that B has a predual.
For example l ∞ = (l 1 )∗ , hence l 1 is the predual of l ∞ .
Another example: Let H 1 be hardy space of analytic functions on the disk
[Rud87]. (H 1 )∗ = BMO, where BMO refers to bounded mean oscillation. It was
developed by Charles Fefferman in 1974 who won the fields medal for this theory.
See [Fef71]. (Getting hands on a specific dual space is often a big thing.)
Theorem 4.8 (C. Fefferman). H1∗ = BMO.
Definition 4.12. Let f be a locally integrable function on Rn , and let Q run through
all n-cubes ⊂ Rn . Set ˆ
1
fQ = f (y) dy.
|Q| Q
We say that f ∈ BMO iff (Def.)

4.4 New Hilbert Spaces From “old” 119
ˆ
1
sup | f (x) − fQ | dx < ∞. (4.22)
Q |Q| Q
In this case the LHS in (4.22) is the BMO-norm of f . Moreover, BMO is a Banach
space.
4.4 New Hilbert Spaces From “old”
Below we consider some cases of building new Hilbert spaces from given ones.
Only sample cases are fleshed out; and they will be needed in the sequel.
Case 1. GNS (section 4.2)
Case 2. Direct sum
(a) Let Hi , i = 1, 2 be two given Hilbert spaces, then the direct “orthogonal” sum
H = H1 ⊕ H2 is as follows:
H = {symbol pairs h1 ⊕ h2 , hi ∈ Hi , i = 1, 2} , and

kh1 ⊕ h2 k2H = kh1 k2H1 + kh2 k2H2 . (4.23)
(b) Given an indexed family of Hilbert spaces {Hα }α∈A where A is a set; then set
H := ⊕A Hα to be
2
⊕
∑ hα = ∑ khα k2Hα < ∞; (4.24)
α∈A H α∈A
i.e., finiteness of the sum in (4.24) is part of the definition.
Case 3. Hilbert-Schmidt operators (continuing the discussion in 1)
Let H be a fixed Hilbert space, and set
H S (H ) := T ∈ B (H ) T ∗ T is trace class

(4.25)
and set
kT k2H S := trace (T ∗ T ) ; (4.26)
similarly if S, T ∈ H S (H ), set
hS, T iH S := trace (S∗ T ) . (4.27)

Note that finiteness on the RHS in (4.26) is part of the definition.
Case 4. Tensor-product
Let H1 and H2 be two Hilbert spaces, and consider finite-rank operators (rank-1
in this case):
|h1 ih h2 | (Dirac ket-bra)

|h1 ih h2 | (u) = hh2 , uiH2 h1 , ∀u ∈ H2 , (4.28)
so T = |h1 ih h2 | : H2 −→ H1 with the identification
h1 ⊗ h2 ←→ |h1 ih h2 | . (4.29)
Set
kh1 ⊗ h2 k2 := trace (T ∗ T ) = kh1 k2H1 kh2 k2H2 . (4.30)
For the Hilbert space H1 ⊗ H2 we take the H S-completion of the finite rank opera-
tors spanned by the set in (4.28). The tensor product construction fits with composite
system in quantum mechanics.
Case 5. Contractive inclusion
Let H1 and H2 be two Hilbert spaces, and let T : H1 → H2 be a contractive

linear operator, i.e.,
IH1 − T ∗ T ≥ 0. (4.31)
On the subspace
R (T ) = T h1 h1 ∈ H1

(4.32)
(generally not closed in H2 ,) set
kT h1 knew := kh1 k , h1 ∈ H1 ; (4.33)
then with k·knew , R (T ) becomes a Hilbert space.
4.5 Abelian C∗ -algebras
Diagonalizing a commuting family of bounded selfadjoint operators may be for-

mulated in the setting of abelian C∗ -algebras. By the structure theorem of Gelfand
4.5 Abelian C∗ -algebras 121
and Naimark, every abelian C∗ -algebra containing the identity element is isomor-
phic to the algebra C (X) of continuous functions on some compact Hausdorff space
X, which is unique up to homeomorphism. The classification of all the represen-
tations abelian C∗ -algebras, therefore, amounts to that of C(X). This problem can
be understood using the idea of σ -measures (square densities). It also leads to the
multiplicity theory of selfadjoint operators. The best treatment on this subject can
be found in [Nel69].
Here we discuss Gelfand’s theory on abelian C∗ -algebras. Throughout, we as-
sume all the algebras contain unit element.
Definition 4.13. A is Banach algebra if it is a complex algebra and a Banach space

s.t. the norm satisfies kabk ≤ kak kbk, for all a, b ∈ A.
Let A be an abelian Banach. Consider the closed ideals in A (since A is normed, so

consider closed ideals) ordered by inclusion. By Zorn’s lemma, there exists maximal
ideals M, which are closed by maximality. Then A/M is 1-dimensional, i.e., A/M =
{tv} for some v ∈ A, and t ∈ R. Therefore the combined map
ϕ : A → A/M → C, a 7→ a/M 7→ ta
is a (complex) homomorphism. In particular, A 3 1A 7→ v := 1A /M ∈ A/M ' C,

and ϕ(1A ) = 1.
Conversely, the kernel of any homomorphism is a maximal ideal in A (since
the co-dimension = 1.) Therefore there is a bijection between maximal ideas and
homomorphisms.
Lemma 4.4. Let A be an abelian Banach algebra. If a ∈ A, and kak < 1, then 1A −a
is invertible.
Proof. It is easy to verify that (1 − a)−1 = 1 + a + a2 + · · · , and the RHS is norm

convergent.
Corollary 4.5. Any homomorphism ϕ : A → C is a contraction.
Proof. Let a ∈ A, a 6= 0. Suppose λ := ϕ (a) s.t. |λ | > kak. Then ka/λ k < 1 and
so 1A − a/λ is invertible by Lemma 4.4. Since ϕ is a homomorphism, it must map
invertible element to invertible element, hence ϕ (1A − a/λ ) 6= 0, i.e., ϕ (a) 6= λ ,
which is a contradiction.
Let X be the set of all maximal ideals, identified with all homomorphisms in A∗1 ,
where A∗1 is the unit ball in A∗ . Since A∗1 is compact (see Banach-Alaoglu, Theorem
4.6), and X is closed in it, therefore X is also compact. Here, compactness refers to
the weak ∗-topology.
Definition 4.14. The Gelfand transform F : A → C(X) is given by
F(a)(ϕ) = ϕ(a), a ∈ A, ϕ ∈ C (X) . (4.34)
Hence A/ ker F is homomorphic to a closed subalgebra of C(X). Note ker F =

{a ∈ A : ϕ (a) = 0, ∀ϕ ∈ X}. It is called the radical of A.
The theory is takes a more pleasant form when A is a C∗ -algebra. So there is an

involution, and the norm satisfies the C∗ axiom: kaa∗ k = kak2 , for all a ∈ A.
Theorem 4.9 (Gelfand). If A is an abelian C∗ -algebra then the Gelfand transform

(4.34) is an isometric ∗-isomorphism from A onto C (X), where X is the maximal
ideal space of A.
Example 4.11. Consider l 1 (Z), the convolution algebra:
(ab)n = ∑ ak bn−k
k
a∗n = a−n
kak = ∑ |an |
n
1A = δ0 (Dirac mass at 0)
To identity X in practice, we always start with a guess, and usually it turns out to be
correct. Since Fourier transform converts convolution to multiplication,
ϕz
l 2 (Z) 3 a −−−−−→ ∑ an zn
is a complex homomorphism. To see ϕz is multiplicative, we have
ϕz (ab) = ∑(ab)n zn
= ∑ ak bn−k zn
n,k
= ∑ ak zk ∑ bn−k zn−k
k n
! !
= ∑ ak zk ∑ bk zk .
k k
4.6 States and Representation 123
Thus {z : |z| = 1} is a subspace in the Gelfand space X. Note that we cannot use
|z| < 1 since we are dealing with two-sided l 1 sequence. (If the sequences were
truncated, so that an = 0 for n < 0 then we allow |z| < 1. )
ϕz is contractive: |ϕz (a)| = |∑ an zn | ≤ ∑n |an | = kak.
Exercise 4.9. Prove that every homomorphism of l 2 (Z) is obtained as ϕz for some
|z| = 1. Hence X = {z : |z| = 1}.
Example 4.12. l ∞ (Z), with kak = supn |an |. The Gelfand space in this case is X =
β Z, the Stone-Cech compactification of Z, which are the ultrafilters on Z. β Z is
much bigger then p-adic numbers. Pure states on diagonal operators correspond to
β Z. See Chapter 8 for details.
4.6 States and Representation
Let A be a ∗-algebra, a representation π : A → B(H ) generates a ∗-subalgebra

π(A) in B(H ). By taking norm closure, one gets a C∗ -algebra, i.e., a Banach ∗-
algebra with the axiom ka∗ ak = kak2 . On the other hand, by Gelfand and Naimark’s
theorem, all abstract C∗ -algebras are isometrically isomorphic to closed subalgebras
of B(H ), for some Hilbert space H (Theorem 4.4). The construction of H comes
down to states S(A) on A and the GNS construction. Therefore, the GNS construc-
tion gives rise to a bijection between states and representations.
Let A+ be the positive elements in A. s ∈ S(A), s : A → C and s(A+ ) ⊂ [0, ∞).
√
For C∗ -algebra, positive elements can be written f = ( f )2 by the spectral theorem.
In general, positive elements have the form a∗ a. There is a bijection between states
and GNS representations Rep(A, H ), where s(A) = hΩ , π(A)Ω i.
Example 4.13. A = C(X) where X is a compact Hausdorff space. sµ given by
´
sµ (a) = adµ is a state. The GNS construction gives H = L2 (µ), π( f ) is the
operator of multiplication by f on L2 (µ). {ϕ1 : ϕ ∈ C(X)} is dense in L2 , where 1
´ ´
is the cyclic vector. sµ ( f ) = hΩ , π( f )Ω i = 1 f 1dµ = f dµ, which is also seen
as the expectation of f in case µ is a probability measure.
We consider decomposition of representations or equivalently states, i.e., breaking
up representations corresponds to breaking up states.
The thing that we want to do with representations comes down to the smallest
ones, i.e., the irreducible representations. Irreducible representations correspond to
pure states which are extreme points in the states.
A representation π : A → B(H ) is irreducible, if whenever H breaks up into

two pieces H = H1 ⊕ H2 , where Hi is invariant under π(A), one of them is zero
(the other is H ). Equivalently, if π = π1 ⊕ π2 , where πi = π Hi
, then one of them is
zero. This is similar to the decomposition of natural numbers into product of primes.
For example, 6 = 2 × 3, but 2 and 3 are primes and they do not decompose further.
Hilbert spaces are defined up to unitary equivalence. A state ϕ may have equiv-
alent representations on different Hilbert spaces (but unitarily equivalent), however
ϕ does not see the distinction, and it can only detect equivalent classes of represen-
tations.
Example 4.14. Let A be a ∗-algebra. Given two states s1 and s2 , by the GNS con-
struction, we get cyclic vectors ξi , and representations πi : A → B(Hi ), so that
si (A) = hξi , πi (A)ξi i, i = 1, 2. Suppose there is a unitary operator W : H1 → H2 ,
such that for all A ∈ A,
π1 (A) = W ∗ π2 (A)W.
Then
hξ1 , π1 (A) ξ1 i1 = hξ1 ,W ∗ π2 (A)W ξ1 i1

= hW ξ1 , π2 (A)W ξ1 i2
= hξ2 , π2 (A) ξ2 i2 , ∀A ∈ A;
i.e., s2 (A) = s1 (A). Therefore the same state s = s1 = s2 has two distinct (unitarily
equivalent) representations.
Remark 4.6. A special case of states are measures when the algebra is abelian. Re-
call that all abelian C∗ -algebras with identity are C (X), where X is the correspond-
ing Gelfand space. Two representations are mutually singular π1 ⊥ π2 , if and only
if the two measures are mutually singular, µ1 ⊥ µ2 .
The theorem below is fundamental in representation theory. Recall that if M is a

subset of B (H ), the commutant M 0 consists of A ∈ B (H ) that commutes with
all elements in M.
Theorem 4.10 (Schur). Let π : A → B(H ) be a representation. The following are

equivalent.
1. π is irreducible.
2. The commutant (π(A))0 is one-dimensional, i.e., (π(A))0 = cIA , c ∈ C.
Proof. Suppose (π(A))0 has more than one dimension. Let X ∈ (π(A))0 , then by tak-
ing adjoint, X ∗ ∈ (π(A))0 . X +X ∗ is selfadjoint, and X +X ∗ 6= cI since by hypothesis
(π(A))0 has more than one dimension. Therefore X + X ∗ has a non trivial spectral
/ {0, I}. Let H1 = P(E)H and H2 = (I − P(E))H .
projection P(E), i.e., P(E) ∈
H1 and H2 are both nonzero proper subspaces of H . Since P(E) commutes with
π(A), for all A ∈ A, it follows that H1 and H2 are both invariant under π.
Conversely, suppose (π(A))0 is one-dimensional. If π is not irreducible, i.e., π =
π1 ⊕ π2 , then for
" # " #
IH1 0 0 0
PH1 = , PH2 = 1 − PH1 =
0 0 0 IH2
we have
PHi π(A) = π(A)PHi , i = 1, 2
for all A ∈ A. Hence (π(A))0 has more than one dimension.
Corollary 4.6. π is irreducible if and only if the only projections in (π(A))0 are 0
or I.
Thus to test invariant subspaces, one only needs to look at projections in the com-
mutant.
Corollary 4.7. If A is abelian, then π is irreducible if and only if H is one-

dimensional.
Proof. Obviously, if dim H = 1, π is irreducible. Conversely, by Theorem 4.10,

(π(A))0 = cI. Since π(A) is abelian, π(A) ⊂ π(A)0 . Thus for all A ∈ A, π(A) = cA I,
for some constant cA .
If instead of taking the norm closure, but using the strong operator topology, ones
gets a von Neumann algebra. von Neumann showed that the weak closure of A is
equal to A00
Corollary 4.8. π is irreducible ⇐⇒ (π(A))0 is 1-dimensional ⇐⇒ (π(A))00 =

B(H ).
Remark 4.7. In matrix notation, we write π = π1 ⊕ π2 as

" #
π1 (A) 0
π(A) = .
0 π2 (A)
If " #
X Y
∈ (π(A))0
UV
then
" #" # " #
X Y π1 (A) 0 Xπ1 (A) Y π2 (A)
=
UV 0 π2 (A) Uπ1 (A) V π2 (A)
" #" # " #
π1 (A) 0 X Y π1 (A)X π1 (A)Y
= .
0 π2 (A) UV π2 (A)U π2 (A)V
Hence
Xπ1 (A) = π1 (A)X

V π2 (A) = π2 (A)V
Uπ1 (A) = π2 (A)U
Y π2 (A) = π1 (A)Y.
Therefore,
X ∈ (π1 (A))0 , V ∈ (π2 (A))0 , and

U,Y ∈ int (π1 , π2 ) = intertwining operators of π1 , π2 .
This is illustrated by the diagram below.
π1 (A)
/ H1
C H1 [
Y U U Y
π2 (A)
H2 / H2
We say π1 and π2 are inequivalent if and only if int(π1 , π2 ) = 0. For π1 = π2 , π

has multiplicity 2. Multiplicity > 1 is equivalent to the commutant being non-abelian.
In the case π = π1 ⊕ π2 where π1 = π2 , (π(A))0 ' M2 (C).
Schur’s lemma addresses all representations. It says that a representation π : A →

B(H ) is irreducible if and only if (π(A))0 is 1-dimensional. When specialize to the
GNS representation of a given state s, this is also equivalent to saying that for all
positive linear functional t, t ≤ s ⇒ t = λ s for some λ ≥ 0. This latter equivalence
is obtained by using a more general result, which relates t and selfadjoint operators
in the commutant (π(A))0 .
We now turn to characterize the relation between state and its GNS representa-
tion, i.e., specialize to the GNS representation. Given a ∗ algebra A, the states S(A)
forms a compact convex subset in the unit ball of the dual A∗ .
Let A+ be the set of positive elements in A. Given s ∈ S(A), let t be a positive
linear functional. By t ≤ s, we mean t(A) ≤ s(A) for all A ∈ A+ . We look for relation
between t and the commutant (π(A))0 .
Lemma 4.5 (Schur-Sakai-Nicodym). Let t be a positive linear functional, and let s

be a state. There is a bijection between t such that 0 ≤ t ≤ s, and selfadjoint operator
A in the commutant with 0 ≤ A ≤ I. The relation is given by
t(·) = hΩ , π(·)AΩ i
Remark 4.8. This is an extension of the classical Radon-Nicodym derivative the-

orem to the non-commutative setting. We may write A = dt/ds. The notation
0 ≤ A ≤ I refers to the partial order of selfadjoint operators. It means that for all
ξ ∈ H , 0 ≤. See [Sak71, KR97b].
Proof. Easy direction, suppose A ∈ (π(A))0 and 0 ≤ A ≤ I. As in many applications,

the favorite functions one usually applies to selfadjoint operators is the square root
√ √ √
function ·. So let’s take A. Since A ∈ (π(A))0 , so is A. We need to show t(a) =
hΩ , π(a)AΩ i ≤ s(a), for all a ≥ 0 in A. Let a = b2 , then
t(a) = hΩ , π(a)AΩ i
= Ω , π(b2 )AΩ
= hΩ , π(b)∗ π(b)AΩ i
= hπ(b)Ω , Aπ(b)Ω i
≤ hπ(b)Ω , π(b)Ω i
= hΩ , π(a)Ω i
= s(a).
Conversely, suppose t ≤ s. Then for all a ≥ 0, t(a) ≤ s(a) = hΩ , π(a)Ω i. Again

write a = b2 . It follows that
t(b2 ) ≤ s(b2 ) = hΩ , π (a) Ω i = kπ (b) Ω k2 .

By Riesz’s theorem, there is a unique η, so that
t(a) = hπ(b)Ω , ηi .
Conversely, let a = b2 , then
t(b2 ) ≤ s(b2 ) = hΩ , π (a) Ω i = kπ (b) Ω k2
i.e., π(b)Ω 7→ t(b2 ) is a bounded quadratic form. Therefore, there exists a unique
A ≥ 0 such that
t(b2 ) = hπ(b)Ω , Aπ(b)Ω i .
It is easy to see that 0 ≤ A ≤ I. Also, A ∈ (π(A))0 , the commutant of π (A).
Corollary 4.9. Let s be a state. (π, Ω , H ) is the corresponding GNS construction.

The following are equivalent.
1. For all positive linear functional t, t ≤ s ⇒ t = λ s for some λ ≥ 0.

Proof. By Sakai-Nicodym derivative, t ≤ s if and only if there is a selfadjoint oper-

ator A ∈ (π(A))0 so that
t(·) = hΩ , π(·)AΩ i
Therefore t = λ s if and only if A = λ I.

Suppose t ≤ s ⇒ t = λ s for some λ ≥ 0. Then π must be irreducible, since
otherwise there exists A ∈ (π(A))0 with A 6= cI, hence A 3 a 7→ t(a) := hΩ , π(a)AΩ i
defines a positive linear functional, and t ≤ s, however t 6= λ s. Thus a contradiction
to the hypothesis.
Conversely, suppose π is irreducible. Then by Schur’s lemma, (π(A))0 is 1-
dimensional. i.e. for all A ∈ (π(A))0 , A = λ I for some λ . Therefore if t ≤ s, by
Sakai’s theorem, t(·) = hΩ , π(·)AΩ i. Thus t = λ s for some λ ≥ 0.
Definition 4.15. A state s is pure if it cannot be broken up into a convex combination

of two distinct states. i.e. for all states s1 and s2 , s = λ s1 + (1 − λ )s2 ⇒ s = s1 or s =
s2 .
The main theorem in this section is a corollary to Sakai’s theorem.
Corollary 4.10. Let s be a state. (π, Ω , H ) is the corresponding GNS construction.

The following are equivalent.
1. t ≤ s ⇒ t = λ s for some λ ≥ 0.
3. s is a pure state.
Proof. By Sakai-Nicodym derivative, t ≤ s if and only if there is a selfadjoint oper-

ator A ∈ (π(A))0 so that
t(a) = hΩ , π (a) AΩ i , ∀a ∈ A.
Therefore t = λ s if and only if A = λ I.

We show that (1) ⇔ (2) and (1) ⇒ (3) ⇒ (2).
(1) ⇔ (2) Suppose t ≤ s ⇒ t = λ s, then π must be irreducible, since otherwise
there exists A ∈ (π(A))0 with A 6= cI, hence t(·) := hΩ , π(·)AΩ i defines a positive
linear functional with t ≤ s, however t 6= λ s. Conversely, suppose π is irreducible.
If t ≤ s, then t(·) = hΩ , π(·)AΩ i with A ∈ (π(A))0 . By Schur’s lemma, (π(A))0 =
{0, λ I}. Therefore, A = λ I and t = λ s.
(1) ⇒ (3) Suppose t ≤ s ⇒ t = λ s for some λ ≥ 0. If s is not pure, then s =
cs1 + (1 − c)s2 where s1 , s2 are states and c ∈ (0, 1). By hypothesis, s1 ≤ s implies
that s1 = λ s. It follows that s = s1 = s2 .
(3) ⇒ (2) Suppose π is not irreducible, i.e. there is a non trivial projection P ∈
(π(A))0 . Let Ω = Ω1 ⊕ Ω2 where Ω1 = PΩ and Ω2 = (I − P)Ω . Then
s(a) = hΩ , π (a) Ω i
= hΩ1 ⊕ Ω2 , π (a) Ω1 ⊕ Ω2 i
= hΩ1 , π (a) Ω1 i + hΩ2 , π (a) Ω2 i

2 Ω1 Ω1 2 Ω2 Ω2
= kΩ1 k , π (a) + kΩ2 k , π (a)
kΩ1 k kΩ1 k kΩ2 k kΩ2 k

2 Ω 1 Ω 1 2
Ω
2 Ω2
= kΩ1 k , π (a) + 1 − kΩ1 k , π (a)
kΩ1 k kΩ1 k kΩ2 k kΩ2 k
= λ s1 (a) + (1 − λ ) s2 (a) .
Hence s is not a pure state.

4.6.1 Normal States
More general states in physics come from the mixture of particle states, which cor-
respond to composite system. These are called normal states in mathematics.
Let ρ ∈ T1 (H ) = trace class operator, s.t. ρ > 0 and tr(ρ) = 1. Define state
sρ (A) := tr(Aρ) , A ∈ B (H ). Since ρ is compact, by spectral theorem of compact
operators,
ρ = ∑ λk Pk
k
such that λ1 > λ2 > · · · → 0; ∑ λk = 1 and Pk = |ξk ih ξk |, i.e., the rank-1 projections.
(See section 4.3.) We have
• sρ (I) = tr(ρ) = 1; and

• for all A ∈ B (H ),
sρ (A) = tr(Aρ) = ∑ hun , Aρun i

n
= ∑ hA∗ un , ρun i
n
= ∑ ∑ λk hA∗ un , ξk i hξk , un i
n k

= ∑ λk ∑ hun , Aξk i hξk , un i
k n
= ∑ λk hξk , Aξk i ;
k
where {uk } is any ONB in H . Hence,
sρ = ∑ λk sξk = ∑ λk |ξk ih ξk |
k k
i.e., sρ is a convex combination of pure states sξk := |ξk ih ξk |.
Remark 4.9. Notice that tr(|ξ ihη|) = hη, ξ i. In fact, take any ONB {en } in H , then
tr (|ξ ih η|) = ∑ hen ξ i hη, en i = hη, ξ i

n
where the last step follows from Parseval identity. (If we drop the condition ρ ≥ 0
then we get the duality (T1 H )∗ = B(H ). See Theorem 4.7.)
4.6.2 A Dictionary of OP and QM
• states - unit vectors ξ ∈ H . These are all the pure states on B(H ).
• observable - selfadjoint operators A = A∗
• measurement - spectrum
The spectral theorem was developed by J. von Neumann and later improved by
Dirac and others. A selfadjoint operator A corresponds to a quantum observable,
and result of a quantum measurement can be represented by the spectrum of A.
• simple eigenvalue: A = λ |ξλ ihξλ |,
sξλ (A) = hξλ , Aξλ i
• compact operator: A = ∑λ λ |ξλ ihξλ |, such that {ξλ } is an ONB of H . If

ξ = ∑ cλ ξλ is a unit vector, then
sξ (A) = ∑ λ hξλ , Aξλ i

λ
where {|cλ |2 }λ is a probability distribution over the spectrum of A, and sξ is

the expectation value of A.
• more general, allowing continuous spectrum:
ˆ
A= λ E(dλ )
ˆ
Aξ = λ E(dλ )ξ .
We may write the unit vector ξ as
ˆ z }|
ξλ
{
ξ = E(dλ )ξ
so that ˆ
2
kξ k = kE(dλ )ξ k2 = 1
It is clear that kE(·)ξ k2 is a probability distribution on spectrum of A. sξ (A)

is again seen as the expectation value of A with respect to kE(·)ξ k2 , since
ˆ
sξ (A) = hξ , Aξ i = λ kE(dλ )ξ k2 .
4.7 Krein-Milman, Choquet, Decomposition of States
We study some examples of compact convex sets in locally convex topological

spaces.2 Typical examples include the set of positive semi-definite functions, tak-
ing values in C or B(H ).
Definition 4.16. A vector space is locally convex if it has a topology which makes
the vector space operators continuous, and if the neighborhoods {x + Nbh0 } have a
basis consisting of convex sets.
The context for Krein-Milman is locally convex topological spaces. It is in all func-
tional analysis books. Choquet’s theorem however comes later, and it’s not found
in most books. A good reference is the book by R. Phelps [Phe01]. The proof of
Choquet’s theorem is not specially illuminating. It uses standard integration theory.
Theorem 4.11 (Krein-Milman). Let K be a compact convex set in a locally convex

topological space. Then K is the closed convex hull of its extreme points E (X), i.e.,
K = conv(E(K)).
Proof. (sketch) If K % conv(E(K)), we get a linear functional w, such that w is zero

on conv(E(K)) and not zero on w ∈ K\conv(E(K)). Extend w by Hahn-Banach
theorem to a linear functional to the whole space, and get a contradiction.
Note: The dual of a normed vector space is always a Banach space, so the theo-
rem applies. The convex hull in an infinite dimensional space is not always
closed, so close it. A good reference to locally convex topological space is
the lovely book by F. Treves [Tre67].
A convex combination of points (ξi ) in K takes the form v = ∑ ci ξi , where ci > 0

and ∑ ci = 1. Closure refers to taking limit, so we allow all limits of such convex
combinations. Such a v is obviously in K, since K was assumed to be convex. The
point of the Krein-Milman’s theorem is the converse.
The decomposition of states into pure states was developed by Choquet et al; see
[Phe01]. The idea goes back to Krein and Choquet.
2 Almost all spaces one works with are locally convex.
4.7 Krein-Milman, Choquet, Decomposition of States 133
Theorem 4.12 (Choquet). K = S(A) is a compact convex set in a locally convex

topological space. Let E(K) be the set of extreme points on K. Then for all p ∈ K,
there exists a Borel probability measure µ p , supported on a Borel set bE(K) ⊃ E(K),
such that for all affine functions f , we have
ˆ
f (p) = f (ξ ) dµ p (ξ ). (4.35)
bE(X)
The expression in Choquet’s theorem is a generalization of convex combination. In

stead of summation, it is an integral against a measure. Since there are some bizarre
cases where the extreme points E(K) do not form a Borel set, the measure µ p is
actually supported on bE(K), such that µ p (bE(K) − E(K)) = 0.
This involves a theorem of Glimm [Gli60, Gli61]. Examples for this case include
the Cuntz algebra, free group with 2 generators, and the generators uvu−1 = u2
for wavelets. Other examples of such decompositions include Fourier transform,
Laplace transform, direct integrals [Sti59, Seg50].
Note: µ p in (4.35) may not be unique. If it is unique, K is called a simplex. The unit
disk has its boundary as extreme points. But representation of points in the
interior using points on the boundary is not unique. Therefore the unit disk
is not a simplex. A tetrahedron is (Fig. 4.1).
1 2
Fig. 4.1: A simplex. The four extreme points are marked.
Example 4.15. Let (X, M, µ) be a measure space, where X is compact and Haus-
dorff. The set of all probability measures P(X) is a convex set. To see this, let
µ1 , µ2 ∈ P(X) and 0 ≤ t ≤ 1, then tµ1 + (1 − t)µ2 is a measure on X, moreover
(tµ1 + (1 − t)µ2 )(X) = t + 1 − t = 1, hence tµ1 + (1 − t)µ2 ∈ P(X). Usually we
don’t want all probability measures, but a closed subset.
Example 4.16. We compute extreme points in the previous example. K = P(X) is
compact convex in C(X)∗ , which is identified as the set of all measures due to Riesz.
C(X)∗ is a Banach space hence is always convex. The importance of being the dual
of some Banach space is that the unit ball is always weak ∗ compact (Banach-
Alaoglu, Theorem 4.6). Note the weak ∗ topology is just the cylinder/product topol-
ogy. The unit ball B∗1 sits inside the infinite product space (compact, Hausdorff)
∏v∈B,kvk=1 D1 , where D1 = {z ∈ C : |z| = 1}. The weak ∗ topology on B∗1 is just the
restriction of the product topology onto B∗1 .
Claim: E(K) = {δx : x ∈ X}, where δx is the Dirac measure supported at x ∈ X. By
´
Riesz, to know the measure is to know the linear functional. f dδx = f (x). Hence
we get a family of measures indexed by X. If X = [0, 1], we get a continuous family
of measures. To see these really are extreme points, we do the GNS construction on
the algebra A = C(X), with the state µ ∈ P(X). The Hilbert space so constructed
is simply L2 (µ). It’s clear that L2 (δx ) is 1-dimensional, hence the representation is
irreducible. We conclude that δx is a pure state, for all x ∈ X.
There is a bijection between state ϕ and Borel measure µ := µϕ ,
ˆ
ϕ(a) = a dµϕ .
X
In C (X), 1A = 1 = constant function. We check that

ˆ
ϕ(1A ) = ϕ(1) = 1dµ = µ(X) = 1
√
since µ ∈ P (X) is a probability measure. Also, if f ≥ 0 then f = g2 , with g := f;
and ˆ
ϕ( f ) = g2 dµ ≥ 0.
Note: ν is an extreme point in P(X) if and only if
(ν ∈ [µ1 , µ2 ] = convex hull of {µ1 , µ2 }) =⇒ (ν = µ1 or ν = µ2 ) .
Example 4.17. Let A = B(H ), and S (A) = states of A. For each ξ ∈ H , the map
A 7→ wξ (A) := hξ , Aξ i is a state, called vector state.
Claim: E(S) = vector states.
To show this, suppose W is a subspace of H such that 0 W H , and suppose
W is invariant under the action of B(H ). Then ∃h ∈ H , h ⊥ W . Choose ξ ∈ W . The
wonderful rank-1 operator (due to Dirac) T : ξ 7→ h given by T := |h ih ξ |, shows that
h ∈ W (since TW ⊂ W by assumption.) Hence h ⊥ h and h = 0. Therefore W = H .
We say B(H ) acts transitively on H .
Note: In general, any C∗ -algebra is a closed subalgebra of B(H ) for some H

(Theorem 4.4). All the pure states on B(H ) are vector states.
Example 4.18. Let A be a ∗-algebra, S(A) be the set of states on A. w : A → C is

a state on A if w(1A ) = 1 and w(A) ≥ 0, whenever A ≥ 0. The set of completely
positive (CP) maps is a compact convex set. CP maps are generalizations of states
(chapter 5).
Exercise 4.10. Take the two state sample space Ω = ∏∞

1 {0, 1} with product topol-
ogy. Assign probability measure, so that we might favor one outcome than the
other. For example, let s = x1 + · · · xn , Pθ (Cx ) = θ s (1 − θ )n−1 , i.e. s heads, (n − s)
tails. Notice that Pθ is invariant under permutation of coordinates. x1 , x2 , . . . , xn 7→
xσ (1) xσ (2) . . . xσ (n) . Pθ is a member of the set of all such invariant measures (invariant
under permutation) Pinv (Ω ). Prove that
E(Pinv (Ω )) = [0, 1]
i.e., Pθ are all the possible extreme points.
Remark 4.10. Let σ : X → X be a measurable transformation. A (probability) mea-

sure µ is ergodic if
[E ∈ M, σ E = E] ⇒ µ(E) ∈ {0, 1}.
Intuitively, it says that the whole space X can’t be divided non-trivially into parts
where µ is invariant. The set X will be mixed up by the transformation σ .
Exercise 4.11 (Irrational rotation). Let θ > 0 be a fixed irrational number, and set
σθ (x) = θ x mod 1 (4.36)
i.e., multiplication by θ modulo 1. Show that σθ in (4.36) is ergodic in the measure

space R/Z ' [0, 1) with Lebesgue measure. See Fig 4.2.
1.0 1.0
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.2 0.4 0.6 0.8 1.0 0.2 0.4 0.6 0.8 1.0
√ √
(a) (x, 2x) mod 1, 1 ≤ x ≤ 5 (b) (x, 2x) mod 1, 0 ≤ x ≤ 15
Fig. 4.2: Irrational rotation.
4.7.1 Noncommutative Radon-Nikodym Derivative
Let w be a state on a C∗ -algebra A, and let K be an operator in A+ . Set

√ √
w( KA K)
wK (A) = .
w(K)
dw
Then wK is a state, and wK w, i.e., w(A) = 0 ⇒ wK (A) = 0. We say that K = dwK
is a noncommutative Radon-Nikodym derivative.
Check:
wK (1) = 1
√ √
∗ w( KA∗ A K)
wK (A A) =
w(K)
√ √
w((A K)∗ (A K))
= ≥0
w(K)
The converse holds too [Sak71] and is called the noncommutative Radon-Nikodym
theorem.
4.7.2 Examples of Disintegration
Example 4.19. L2 (I) with Lebesgue measure. Let


1 t ≥ x
Fx (t) =
0 t < x
Fx is a monotone increasing function on R, hence by Riesz, we get the corresponding

Riemann-Stieljes measure dFx .
ˆ ⊕
dµ = dFx (t)dx.
i.e. ˆ ˆ ˆ
f dµ = dFx ( f )dx = f (x)dx.
Equivalently, ˆ
dµ = δx dx
i.e. ˆ ˆ ˆ
f dµ = δx ( f )dx = f (x)dx.
µ is a state, δx = dFx (t) is a pure state, ∀x ∈ I. This is a decomposition of state into

direct integral of pure states. See [Sti59, Seg50].
Example 4.20. Ω = ∏t≥0 R̄, Ωx = {w ∈ Ω : w(0) = x}. Kolmogorov gives rise to Px

by conditioning P with respect to “starting at x”.
ˆ ⊕
P= Px dx
i.e. ˆ
P() = P(·|start at x)dx.
Example 4.21. Harmonic function on D

ˆ
h 7→ h(z) = fbdµz
∂D
Poisson integration.
4.8 Examples of Representations
We consider the Fourier algebra.
1. Discrete case: l 1 (Z) and the Gelfand transform
(a ∗ b)n = ∑ ak bn−k
k
∗
(a )n = a−n
1A = δ0
F
a −−−−→ F (z) := ∑ an zn
Gelfand n
We may specialize to z = eit , t ∈ R mod 2π. {F(z)} is an abelian algebra

of functions, with multiplication is given by
F(z)G(z) = ∑(a ∗ b)n zn

n
In fact, most abelian algebras can be thought of function algebras.
Homomorphism:
F
(l 1 , ∗) −
→ C(T1 )
(an ) 7→ F(z).
If we want to write F(z) as power series, then we need to drop an for n < 0.
Then F(z) extends to an analytic function over the unit disk. The sequence
space
{a0 , a1 , . . .}
was suggested by Hardy. Rudin has two nice chapters on H 2 , as a Hilbert

space, a RKHS. See [Rud87, ch16].
2. Continuous case: L1 (R)
ˆ ∞
( f ∗ g)(x) = f (s)g(x − s)ds
−∞
f ∗ (x) = f (−x)
4.8 Examples of Representations 139
The algebra L1 has no identity, but we may always insert one by adding
δ0 . So δ0 is the homomorphism f 7−→ f (0); and L1 (R) ∪ {δ0 } is again a
Banach ∗-algebra.
The Gelfand map is the classical Fourier transform, i.e.,

ˆ ∞
F
f −−−−→ fˆ (ξ ) = f (x) e−iξ x dx
Gelfand −∞
∗ g = fˆĝ.
where fd
Remark 4.11. C(T1 ) is called the C∗ -algebra completion of l 1 . L∞ (X, B, µ) = L1 (µ)∗

is also a C∗ -algebra. It is a W ∗-algebra, or von Neumann algebra. The W ∗ refers to
the fact that its topology comes from the weak ∗-topology. Recall that B(H ), for
any Hilbert space, is a von Neumann algebra.
Example 4.22. u f = eiθ f (θ ), v f = f (θ − ϕ), restrict to [0, 2π], i.e., 2π periodic

functions.
vuv−1 = eiϕ u
vu = eiϕ uv
u, v generate a noncommutative C∗ -algebra.
Example 4.23 (Quantum Mechanics). Consider the canonical commutation relation
[p, q] = −iI
where [x, y] := xy − yx denotes the commutator of x and y .

p, q generate an algebra, but they can not be represented by bounded operators.
But we may apply bounded functions to them and get a C∗ -algebra.
Exercise 4.12. Show that p, q and not be represented by bounded operators. Hint:
take the trace.
Example 4.24. Let H be an infinite dimensional Hilbert space. H is isomet-

rically isomorphic to a subspace of itself. For example, let {en } be an ONB.
H1 = span{e2n }, H2 = span{e2n+1 }. Let
V1 (en ) = e2n
V2 (en ) = e2n+1
then we get two isometries. Also,
V1V1∗ +V2V2∗ = I
Vi∗Vi = I
ViVi∗ = Pi
where Pi is a selfadjoint projection, i = 1, 2. This is the Cuntz algebra O2 . More

general ON , N > 2. Cuntz showed that this is a simple algebra in 1977. See, e.g.,
[Gli60, Gli61, BJO04].
4.9 Beginning of Multiplicity Theory
The main question here is how to break up a representation into smaller ones. The
smallest representations are the irreducible ones. The next would be multiplicity
free representation.
Let A be an algebra.
• commutative: e.g. function algebra

• non-commutative: matrix algebra, algebras generated by representation of
non-abelian groups
Smallest representation:
• irreducible - π ∈ Repirr (A, H ) where H is 1-dimensional. This is the start-

ing point of further analysis
• multiplicity free - π ∈ Rep(A, H ) assuming cyclic, since otherwise π =
⊕πcyc . π is multiplicity free if and only if A0 is abelian. In general A0 may or
may not be abelian.
Let C ⊂ B(H ) be a ∗-algebra, C0 = {X : H → H |XC = CX, ∀C ∈ C}. C0 is also

a ∗-algebra. It is obvious that C ⊂ C0 if and only if C is abelian.
Definition 4.17. Let π ∈ Rep(A, H ). Then multi(π) = n ⇐⇒ π(A)0 ' Mn (C).
Example 4.25. Let

4.9 Beginning of Multiplicity Theory 141
 
" 1 #
I2 0
A= 1 = .
 
0 2
2
Then AC = CA if and only if
 
ab " #
B0
C =c d =
 
01
1
where B ∈ M2 (C).
Recall that A = A∗ if and only if

ˆ
A= tP(dt)
sp(A)
where P is a projection-valued measure (PVM). The simplest example of a PVM is

P(E) = χE . The spectral theorem states that all PVMs come this way.
Let A be a positive compact operator on H . A is positive means hx, Axi ≥ 0 for
all x ∈ H . The spectral theorem of A states that
A = ∑ λn Pn
where λn0 s are the eigenvalues of A, such that λ1 ≥ λ2 ≥ · · · λn → 0. and Pn0 s are
self-adjoin projections onto the corresponding finite dimensional eigenspace of λn .
P is a PVM supported on {1, 2, 3, . . .}, so that Pn = P({n}).
Question: what does A look like if it is represented as the operator of multiplica-
tion by the independent variable?
We may arrange the eigenvalues of A such that
s s s
z }|1 { z }|2 { z }|n {
λ1 = · · · = λ1 > λ2 = · · · = λ2 > · · · > λn = · · · = λn > · · · → 0.
We say that λi has multiplicity si . The dimension of the eigen space of λi is si , and
dimH = ∑ si .
Example 4.26. We represent A as the operator M f of multiplication by f on L2 (X, µ).

Let Ek = {xk,1 , . . . , xk,sk } ⊂ X, and let Hk = span{χ{xk, j } : j ∈ {1, 2, . . . , sk }}. Let
∞
f= ∑ λk χEk , (λ1 > λ2 > · · · > λn → 0)
k=1
Notice that χEk is a rank s1 projection. M f is compact if and only if it is of the given
form.
Example 4.27. Follow the previous example, we represent A as the operator Mt of

multiplication by the independent variable on some Hilbert space L2 (µ f ). For sim-
plicity, let λ > 0 and
f = λ χ{x1 ,x2 } = λ χ{x1 } + λ χ{x2 }
i.e. f is compact since it is λ times a rank-2 projection; f is positive since λ > 0.

The eigenspace of λ has two dimension,
M f χ{xi } = λ χ{xi } , i = 1, 2.
Define µ f (·) = µ ◦ f −1 (·), then
µ f = µ({x1 })δλ ⊕ µ({x2 })δλ ⊕ cont. sp δ0
and
L2 (µ f ) = L2 (µ({x1 })δλ ) ⊕ L2 (µ({x2 })δλ ) ⊕ L2 (cont. sp δ0 ).
Define U : L2 (µ) → L2 (µ f ) by
(Ug) = g ◦ f −1 .
U is unitary, and the following diagram commute.
Mf
L2 (X, µ) / L2 (X, µ)
U U

L2 (R, µ f )
Mt
/ L2 (R, µ f )
To check U preserves the L2 -norm,

ˆ
kUgk2 = kg ◦ f −1 ({x})k2 dµ f
4.9 Beginning of Multiplicity Theory 143
= kg ◦ f −1 ({λ })k2 + kg ◦ f −1 ({0})k2

ˆ
= |g(x1 )|2 µ({x1 }) + |g(x2 )|2 µ({x2 }) + |g(x)|2 dµ
X\{x1 ,x2 }
ˆ
= |g(x)|2 dµ
X
To see U diagonalizes M f ,
Mt Ug = λ g(x1 ) ⊕ λ g(x2 ) ⊕ 0g(t)χX\{x1 ,x2 }

= λ g(x1 ) ⊕ λ g(x2 ) ⊕ 0
UM f g = U(λ g(x)χ{x1 ,x2 } )
= λ g(x1 ) ⊕ λ g(x2 ) ⊕ 0
Thus
Mt U = UM f .
Remark 4.12. Notice that f should really be written as
f = λ χ{x1 ,x2 } = λ χ{x1 } + λ χ{x2 } + 0χX\{x1 ,x2 }
since 0 is also an eigenvalue of M f , and the corresponding eigenspace is the kernel

of M f .
Example 4.28. diagonalize M f on L2 (µ) where f = χ[0,1] and µ is the Lebesgue

measure on R.
Example 4.29. diagonalize M f on L2 (µ) where


2x x ∈ [0, 1/2]
f (x) =
2 − 2x x ∈ [1/2, 1]
and µ is the Lebesgue measure on [0, 1].
Remark 4.13. see direct integral and disintegration of measures.
In general, let A be a selfadjoint operator acting on H . Then there exists a second

Hilbert space K, a measure ν on R, and unitary transformation F : H → LK2 (R, ν)
such that
Mt F = FA
for measurable function ϕ : R → K,

ˆ
kϕkL2 (ν) = kϕ(t)k2K dν(t) < ∞.
K
Examples that do have multiplicities in finite dimensional linear algebra:
Example 4.30. 2-d, λ I, {λ I}0 = M2 (C) which is not abelian. Hence mult(λ ) = 2.
Example 4.31. 3-d,  

λ1 " #
λ1 I
=
 
 λ1
λ2
λ2
where λ1 6= λ2 . The commutant is

" #
B
b
where B ∈ M2 (C), and b ∈ C. Therefore the commutant is isomorphic to M2 (C),

and multiplicity is equal to 2.
Example 4.32. The example of Mϕ with repetition.
Mϕ ⊕ Mϕ : L2 (µ) ⊕ L2 (µ) → L2 (µ) ⊕ L2 (µ)

" #" # " #
Mϕ f1 ϕ f1
=
Mϕ f2 ϕ f2
the commutant is this case is isomorphic to M2 (C). If we introduces tensor product,

then representation space is also written asL2 (µ) ⊗V2 , the multiplication operator is
amplified to Mϕ ⊗ I, whose commutant is represented as I ⊗V2 . Hence it’s clear that
the commutant is isomorphic to M2 (C). To check
(ϕ ⊗ I)(I ⊗ B) = ϕ ⊗ B
(I ⊗ B)(ϕ ⊗ I) = ϕ ⊗ B.
we suggest: [Arv76, BR79, KR97a, Mac85, Seg50].
Chapter 5
Completely Positive Maps
From its shady beginnings devising gambling strategies and

counting corpses in medieval London, probability theory and
statistical inference now emerge as better foundations for
scientific models, especially those of the process of thinking and
as essential ingredients of theoretical mathematics, even the
foundations of mathematics itself. — David Mumford
In the last chapter we studied two question from the use of algebras of operators
in quantum physics: “Where does the Hilbert space come from?” And “What are
the algebras of operators from which the selfadjoint observables must be selected?”
An answer is given in “the Gelfand-Naimark-Segal (GNS) theorem;” a direct cor-
respondence between states and cyclic representations. But states are scalar valued
positive definite functions on ∗-algebras. For a host of applications, one must instead
consider operator valued “states.” For this a different notion of positivity is needed,
“complete positivity.”
The GNS construction gives a bijection between states and cyclic representa-
tions. An extension to the GNS construction is Stinespring’s completely positive
map. It appeared in an early paper by Stinespring in 1956 [Sti55]. Arveson in 1970’s
reformatted Stinespring’s result using tensor product [Arv72]. He realized that com-
pletely positive maps are the key in multivariable operator theory, and in noncom-
mutative dynamics.
5.1 Motivation
Let A be a ∗-algebra with identity. Recall that w : A → C is a state if w(1A ) = 1,

√
w(A∗ A) ≥ 0. If A was a C∗ -algebra, A ≥ 0 ⇔ sp(A) ≥ 0, hence we may take B = A
and A = B∗ B.
Given a state w, the GNS construction gives a Hilbert space K , a cyclic vector
Ω ∈ K , and a representation π : A → B(K ), such that
w(A) = hΩ , π(A)Ω i
145
146 5 Completely Positive Maps
K = span{π(A)Ω : A ∈ A}.
Moreover, the Hilbert space is unique up to unitary equivalence.

Stinespring modified a single axiom in the GNS construction. In stead of a state
w : A → C, he considered a positive map ϕ : A → B(H ), i.e., ϕ maps positive
elements in A to positive operators in B(H ). ϕ is a natural extension of w, since C
can be seen as a 1-dimensional Hilbert space, and w is a positive map w : A → B(C).
He soon realized that ϕ being a positive map is not enough to produce a Hilbert
space and a representation. It turns out that the condition to put on ϕ is complete
positivity:
Definition 5.1. Let A be a ∗-algebra. A map ϕ : A → B (H ) is completely positive,

if for all n ∈ N,
ϕ ⊗ IMn : A ⊗ Mn → B(H ⊗ Cn ) (5.1)
maps positive elements in A ⊗ Mn to positive operators in B(H ⊗ Cn ). ϕ is called

a completely positive map, or a CP map. (CP maps are developed primarily for
nonabelian algebras.)
The algebra Mn of n × n matrices can be seen as an n2 -dimensional Hilbert space

with an ONB given by the matrix units {ei j }ni, j=1 . It is also a ∗-algebra generated by
{ei j }ni, j=1 such that

e
il j=k
ei j ekl =
0 j 6= k
Members of A ⊗ Mn are of the form
∑ Ai j ⊗ ei j .
i, j
In other words, A ⊗ Mn consists of precisely the A-valued n × n matrices. Similarly,

members of H ⊗ Cn are the n-tuple column vectors with H -valued entries.
Let IMn : Mn → B(Cn ) be the identity representation of Mn onto B(Cn ). Then,
ϕ ⊗ IMn : A ⊗ Mn → B(H ) ⊗ B(Cn ) (= B(H ⊗ Cn )) (5.2)

!
ϕ ⊗ IMn ∑ Ai j ⊗ ei j = ∑ ϕ(Ai j ) ⊗ ei j . (5.3)
i, j i, j
Note the RHS in (5.3) is an n × n matrix with B(H )-valued entries.

5.1 Motivation 147
Remark 5.1. The algebra B(Cn ) of all bounded operators on Cn is generated by the
rank-one operators, i.e.,
IMn (ei j ) = ei ih e j . (5.4)
Hence the ei j on the LHS of (5.3) is seen as an element in the algebra Mn (C),
i.e., n × n complex matrices; while on the RHS of (5.3), ei j is treated as the rank one
operator ei ih e j ∈ B(Cn ). Using Dirac’s notation, when we look at ei j as operators,
we may write 
|e i j=k
i
ei, j (ek ) = |ei ih e j | |ek i =
0 j 6= k

|e ihe | j=k
i l
ei, j ekl = |ei ihe j | |ek ihel | =
0 j 6= k
This also shows that IMn is in fact an algebra isomorphism.

The CP condition in (5.1) is illustrated in the following diagram.

A → B(H ) : A 7→ ϕ(A)
⊗
M → M : x 7→ IMn (X) = X (identity representation of Mn )
n n
It is saying that if ∑i, j Ai j ⊗ ei j is a positive element in the algebra A ⊗ Mn , then

the n × n B(H )-valued matrix ∑i, j ϕ(Ai j ) ⊗ ei j is a positive operator acting on the
Hilbert space H ⊗ Cn .
Specifically, take any v = ∑nk=1 vk ⊗ ek in H ⊗ Cn , we must have
* +
∑ vl ⊗ el , (∑ ϕ(Ai j ) ⊗ ei j )(∑ vk ⊗ ek )
l i, j k
* +
= ∑ vl ⊗ el , ∑ ϕ(Ai j )vk ⊗ ei j (ek )
l i, j,k
* +
= ∑ vl ⊗ el , ∑ ϕ(Ai j )v j ⊗ ei
l i, j
= ∑ vl , ϕ (Ai j ) v j hel , ei i
i, j,l
= ∑ vi , ϕ (Ai j ) v j ≥ 0. (5.5)
i, j
Using matrix notation, the CP condition is formulated as:

For all n ∈ N, and all v ∈ H ⊗ Cn , i.e.,
 
v1
n .
v = ∑ vk ⊗ ek =  .
.
k=1
vn
we have
  
ϕ(A11 ) ϕ(A12 ) · · · ϕ(A1n ) v1
 ϕ(A21 ) ϕ(A22 ) · · · ϕ(A2n )   v2 
i
h  
v1 v2 · · · vn  . .. .. ..    .  ≥ 0. (5.6)
  
 .. . . .   .. 
ϕ(An1 ) ϕ(An2 ) · · · ϕ(Ann ) vn
5.2 CP v.s. GNS
The GNS construction can be reformulated as a special case of the Stinespring’s

theorem [Sti55].
Let A be a ∗-algebra, given a state ϕ : A → C, there exists a triple (K , Ω , π), all
depending on ϕ, such that
ϕ(A) = hΩ , π (A) Ω iK
where
Ω = π (1A ) ∈ K
K = span{π(A)Ω : A ∈ A}.
The 1-dimensional Hilbert space C is thought of being embedded into K (pos-

sibly infinite dimensional) via
V
C 3 t −→ tΩ ∈ CΩ (5.7)
where CΩ = the one-dimensional subspace in K generated by the unit cyclic vec-

tor cyclic Ω .
Lemma 5.1. The map V in (5.7) is an isometry, such that V ∗V = IC : C → C, and
VV ∗ : K → CΩ (5.8)
5.2 CP v.s. GNS 149
is the projection from K onto the 1-d subspace CΩ in K .

Moreover,
ϕ (A) = V ∗ π (A)V, ∀A ∈ A. (5.9)
Proof. Let t ∈ C, then kV tkK = ktΩ kK = |t|, and so V is an isometry.

For all ξ ∈ K , we have
hξ ,V tiK = hV ∗ ξ ,tiC = tV ∗ ξ .
By setting t = 1, we get
V ∗ ξ = hξ ,V 1iK = hξ , Ω iK = hΩ , ξ iK ⇐⇒ V ∗ = hΩ , ·iK .
Therefore,
V ∗V t = V ∗ (tΩ ) = hΩ ,tΩ iK = t, ∀t ∈ C ⇐⇒ V ∗V = IC
VV ∗ ξ = V (hΩ , ξ iK ) = hΩ , ξ iK Ω , ∀ξ ∈ K ⇐⇒ VV ∗ = |Ω ih Ω | .
It follows that
ϕ (A) = hΩ , π (A) Ω iK
= hV 1, π (A)V 1iK
= h1,V ∗ π (A)V 1iC
= V ∗ π (A)V, ∀A ∈ A
which is the assertion in (5.9).
In other words, Ω 7−→ π(A)Ω sends the unit vector Ω from the 1-dimensional
subspace CΩ to the vector π(A)Ω ∈ K , and hΩ , π (A) Ω iK cuts off the resulting
vector π(A)Ω and only preserves the component corresponding to the 1-d subspace
CΩ . Notice that the unit vector Ω is obtained from embedding the constant 1 ∈ C
via the map V , i.e., Ω = V 1. In matrix notation, if we identify C with its image CΩ
in K , then ϕ(A) is put into a matrix corner:
" #
ϕ(A) ∗
π(A) =
∗ ∗
so that when acting on vectors,

" #" #
h i ϕ(A) ∗ Ω
ϕ(A) = Ω 0 .
∗ ∗ 0
Equivalently
ϕ(A) = P1 π(A) : P1 K → C;
where P1 := VV ∗ = |Ω ih Ω | = rank-1 projection on CΩ .

Stinespring’s construction is a generalization of the above formulation: Let A be a
∗-algebra, given a CP map ϕ : A → B(H ), there exists a Hilbert space K = Kϕ ,

an isometry V : H → K , and a representation π = πϕ : A → K , such that

ϕ(A) = V ∗ π(A)V, ∀A ∈ A.
Notice that this construction starts with a possibly infinite dimensional Hilbert space
H (instead of the 1-dimensional Hilbert space C), the map V embeds H into a
bigger Hilbert space K . If H is identified with its image in K , then π(A) is put
into a matrix corner, " #
π(A) ∗
∗ ∗
so that when acting on vectors,
" #" #
h i π(A) ∗ Vξ
ϕ(A)ξ = V ξ 0 .
∗ ∗ 0
This can be formulated alternatively:

For every CP map ϕ : A → B(H ), there is a dilated Hilbert space K = Kϕ ⊃

H , a representation π = πϕ : A → B(K ), such that

ϕ(A) = PH π(A)
i.e., π(A) can be put into a matrix corner. K is chosen as minimal in the sense that
K = span{π(A)(V h) : A ∈ A, h ∈ H }.
H
V /K
ϕ(A) π(A)

H /K
V
5.3 Stinespring’s Theorem 151
Note: The containment H ⊂ K comes after the identification of H with its im-
age in K under the isometric embedding V . We write ϕ(A) = PH π(A), as
opposed to ϕ(A) = PH π(A)PH , since ϕ(A) only acts on the subspace H .
5.3 Stinespring’s Theorem
Theorem 5.1 (Stinespring [Sti55]). Let A be a ∗-algebra. The following are equiv-
alent:
1. ϕ : A → B(H ) is a completely positive map, and ϕ(1A ) = IH .

2. There exists a Hilbert space K , an isometry V : H → K , and a repre-
sentation π : A → B(K ) such that
ϕ(A) = V ∗ π(A)V, ∀A ∈ A. (5.10)
3. If the dilated Hilbert space K is taken to be minimum, then it is unique

up to unitary equivalence. Specifically, if there are two systems (Vi , Ki , πi ),
i = 1, 2, satisfying
ϕ(A) = Vi∗ πi (A)Vi (5.11)

Ki = span{πi (A)V h : A ∈ A, h ∈ H } (5.12)
then there exists a unitary operator W : K1 → K2 so that
W π1 = π2W (5.13)
Proof.
(Part 3, uniqueness) Let (Vi , Ki , πi ), i = 1, 2, be as in the statement satisfying
(5.11)-(5.12). Define
W π1 (A)V h = π2 (A)V h
then W is an isometry, since
kπi (A)V hk2K = hπi (A)V h, πi (A)V hiK

= hh,V ∗ πi (A∗ A)V hiH
= hh, ϕ(A∗ A)hiH .
Hence W extends uniquely to a unitary operator W : K1 → K2 . To see that W inter-

twines π1 , π2 , notice that a typical vector in Ki is πi (A)V h, and
W π1 (B)π1 (A)V h = W π1 (BA)V h

= π2 (BA)V h
= π2 (B)π2 (A)V h
= π2 (B)W π1 (A)V h.
Since such vectors are dense in the respective dilated space, we conclude that W π1 =
π2W , so (5.13) holds.
Note: kπ1 (B)π1 (A)V hk2 = hh,V ∗ π1 (A∗ B∗ BA)V hi. Fix A ∈ B(H ), the map B 7→
A∗ BA is an automorphism on B(H ).
(2 =⇒ 1)
Now suppose ϕ(A) = V ∗ π(A)V , and we verify it is completely positive.
Since positive elements in A ⊗ Mn are sums of the operator matrix
A∗1
 
 ∗ h
 A2  i
∑ A∗i A j ⊗ ei j = .
 . 
 A1 A2 · · · An
i, j  . 
A∗n
it suffices to show that

!
ϕ ⊗ IMn ∑ A∗i A j ⊗ ei j = ∑ ϕ (A∗i A j ) ⊗ ei j
i, j i, j
is a positive operator in B(H ⊗ Cn ), i.e., need to show that for all v ∈ H ⊗ Cn
ϕ (A∗1 A1 ) ϕ (A∗1 A2 ) · · · ϕ (A∗1 An )

  
v1
 ϕ (A∗2 A1 ) ϕ (A∗2 A2 ) · · · ϕ (A∗2 An )   v2 
i
h  
v1 v2 · · · vn  .. .. ..   .  ≥ 0. (5.14)
..

 . 
 . . . .  . 
ϕ (A∗n A1 ) ϕ (A∗n A2 ) · · · ϕ (A∗n An ) vn
This is true, since
RHS(5.14) = ∑ vi , ϕ (A∗i A j ) v j H
i, j
= ∑ vi ,V ∗ π (A∗i A j )V v j H
i, j
= ∑ π (Ai )V vi , π (A j )V v j K
i, j
2
= ∑ π (Ai )V vi ≥ 0.
i K
(1 =⇒ 2)
Given a completely positive map ϕ, we construct K = Kϕ , V = Vϕ and

π = πϕ . Recall that ϕ : A → B(H ) is a CP map means that for all n ∈ N,

ϕ ⊗ IMn : A ⊗ Mn → B(H ⊗ Mn )
is positive, and
ϕ ⊗ IMn (1A ⊗ IMn ) = IH ⊗ IMn .
The condition on the identity element can be stated using matrix notation as
    
ϕ 0 ··· 0 1A 0 · · · 0 IH 0 · · · 0
0 ϕ ··· 0  0 1A · · · 0 0 IH · · · 0
    
  
. .. . . ..   .. .. . . .. = .. .. . . .. .
.
. . . .
   
. .  . . .   . . . 
0 0 ··· ϕ 0 0 · · · 1A 0 0 · · · IH
Let K0 be the algebraic tensor product A ⊗ H , i.e.,

( )
K0 = span ∑ Ai ⊗ ξi : A ∈ A, ξ ∈ H .
finite
Define a sesquilinear form h·, ·iϕ : K0 × K0 → C, by

* +
n n
∑ Ai ⊗ ξi , ∑ B j ⊗ η j := ∑ ξi , ϕ (A∗i B j ) η j H
. (5.15)
i=1 j=1 ϕ i, j
By the CP condition (5.1), we have

* +
n n
∑ Ai ξi , ∑ A j ξ j = ∑ ξi , ϕ(A∗i A j )ξ j H
≥ 0.
i=1 j=1 ϕ i, j
n o
Let N := v ∈ K0 : hv, viϕ = 0 . Since the Schwarz inequality holds for any sesquilin-
ear form, it follows that
n o
N = v ∈ K0 : hs, viϕ = 0, ∀s ∈ K0 .
Thus N is a closed subspace in K0 . Let K = Kϕ be the Hilbert space by complet-

ing K0 /N with respect to

k·kK := h·, ·iϕ1/2 .
Let V : H → K0 , by
V ξ := 1A ⊗ ξ , ∀ξ ∈ H .
Then,
kV ξ k2ϕ = h1A ⊗ ξ , 1A ⊗ ξ iϕ
= hξ , ϕ(1∗A 1A )ξ iH
= hξ , ξ iH = kξ k2H
V
i.e., V is isometric, and so H −→ K0 is an isometric embedding.
Claim.
(i) V ∗V = IH ;
(ii) VV ∗ = projection from K0 on the subspace 1A ⊗ H .
Indeed, for any A ⊗ η ∈ K0 , we have
hA ⊗ η,V ξ iϕ = hA ⊗ η, 1A ⊗ ξ iϕ
= hη, ϕ(A∗ )ξ iH
= hϕ(A∗ )∗ η, ξ iH
which implies that

V ∗ (A ⊗ η) = ϕ(A∗ )∗ η.
It follows that
V ∗V ξ = V ∗ (1A ⊗ ξ ) = ϕ(1∗A )∗ ξ = ξ , ∀ξ ∈ H
i.e., V ∗V = IH . Moreover, for any A ⊗ η ∈ K0 ,
VV ∗ (A ⊗ η) = V (ϕ(A∗ )∗ η) = 1A ⊗ ϕ(A∗ )∗ η.
This proves the claim. It is clear that the properties of V pass to the dilated space
K = Kϕ = clϕ (K0 /N).

To finish the proof of the theorem, define π = πϕ as follows: Set
!
π(A) ∑Bj ⊗ηj := ∑ AB j ⊗ η j , ∀A ∈ A
j j
and extend it to K .
For all ξ , η ∈ H , then,
hξ ,V ∗ π(A)V ηiH = hV ξ , π(A)V ηiK

= h1A ⊗ ξ , π(A)1A ⊗ ηiK
= h1A ⊗ ξ , A ⊗ ηiK
= hξ , ϕ(1∗A A)ηiH
= hξ , ϕ(A)ηiH .
We conclude that ϕ(A) = V ∗ π(A)V , for all A ∈ A.
Application of Stinespring’s Theorem to Representations of ON
Corollary 5.1. Let N ∈ N, N > 1, and let Ai ∈ B (H ), 1 ≤ i ≤ N, be a system of

operators in a Hilbert space H such that
N
∑ A∗i Ai = IH ; (5.16)
i=1
then there is a second Hilbert space K , and an isometry V : H → K , and a

representation π ∈ Rep (ON , K ) such that
V ∗ π(si )V = A∗i , 1 ≤ i ≤ N, (5.17)
where {si }Ni=1 are generators for ON .
Proof. Given ON with generators {si }Ni=1 , then set
ϕ(si s∗j ) = A∗i A j
using (5.16), it is easy to see that ϕ is completely positive.

Now let (π, K ) be the pair obtained from Theorem 5.1 (Strinespring); then as a
block-matrix of operators, we have as follows
" #
∗ Ai ∗
π (si ) = (5.18)
0 ∗
relative to the splitting

K = V H ⊕ (K VH ), (5.19)
and so V ∗ π (si )∗ V = Ai , which is equivalent to (5.17).
5.4 Comments
In Stinespring’s theorem, the dilated space comes from a general principle (using
positive definite functions) when building Hilbert spaces out of the given data. We
illustrate this point with a few familiar examples.
Example 5.1. In linear algebra, there is a bijection between inner product structures
on Cn and positive-definite n × n matrices. Specifically, h·, ·i : Cn × Cn → C is an
inner product if and only if there exists a positive definite matrix A such that
hv, wiA = v∗ Aw
for all v, w ∈ Cn . We think of Cn as C-valued functions on {1, 2, . . . , n}, then h·, ·iA
is an inner product built on the function space.
This is then extended to infinite dimensional space.
Example 5.2. If F is a positive definite function on R, then on K0 = span{δx : x ∈

R}, F defines a sesquilinear form h·, ·iF : R × R → C, where
* +
∑ ci δxi , ∑ d j δx j := ∑ ci d j F(xi , x j ), and
i j F i, j
2 * +
∑ ci δxi := ∑ ci δxi , ∑ c j δx j = ∑ ci c j F (xi , x j ) ≥ 0.
i F i j F i, j
Let N = {v ∈ K0 : hv, vi = 0}, then N is a closed subspace in K0 . We get a Hilbert

space: K := clF (K0 /N) = the completion of K0 /N with respect to k·kF .
5.4 Comments 157
What if the index set is not {1, 2, . . . , n} or R, but a ∗-algebra?
Example 5.3. C (X), X compact Hausdorff. It is a C∗ -algebra, where k f k := supx | f (x)|.

By Riesz’s theorem, there is a bijection between positive states (linear functionals)
on C(X) and Borel probability measures on X.
Let B (X) be the Borel sigma-algebra on X, which is also an abelian algebra: The
associative multiplication is defined as AB := A ∩ B. The identity element is just X.
Let µ be a probability measure, then µ(A ∩ B) ≥ 0, for all A, B ∈ B (X), and
µ(X) = 1. Hence µ is a state. As before, we apply the GNS construction. Set
K0 = span{δA : A ∈ M} = span{χA : A ∈ M}
Note the index set here is B (X), and ∑i ci δAi = ∑i ci χAi , i.e., these are precisely the
simple functions . Define
* +
∑ ci χAi , ∑ d j χB j := ∑ ci d j µ(Ai ∩ B j )
i j i, j
which is positive definite, since

* +
∑ ci χAi , ∑ ci χAi = ∑ ci c j µ(Ai ∩ A j ) = ∑ |ci |2 µ (Ai ) ≥ 0.
i i i, j i
Here, N = {v ∈ K0 : hv, vi = 0} = µ-measure zero sets, and
H = clµ (K0 /N) = L2 (µ) .
Example 5.4 (GNS). Let A be a ∗-algebra. The set of C-valued functions on A is

A ⊗ C, i.e., functions of the form
( )
∑ Ai ⊗ ci = ∑ ci δAi (5.20)
i i
with finite summation over i. Note that C is naturally embedded into A ⊗ C as 1A ⊗

C (i.e., c 7→ cδ1A ), and the latter is a 1-dimensional subspace. In order to build a
Hilbert space out of (5.20), one needs a positive definite function. A state ϕ on A
does exactly the job. The sesquilinear form is given by
* +
∑ ci δAi , ∑ d j δB j := ∑ ci d j ϕ (A∗i B j )
i i ϕ i, j
so that
2
∑ ci δAi = ∑ ci c j ϕ (A∗i A j ) ≥ 0.
i ϕ i, j
Finally, let Kϕ = Hilbert completion of A ⊗ C/ ker ϕ. Define π(A)δB := δBA , so a

“shift” in the index variable, and extend to Kϕ .
In Stinespring’s construction, A ⊗ C is replaced by A ⊗ H , i.e., in stead of work-

ing with C-valued functions on A, one looks at H -valued functions. Hence we are
looking at functions of the form
( )
∑ Ai ⊗ ξi = ∑ ξi δAi
i i
with finite summation over i. H is embedded into A ⊗ H as 1A ⊗ H , by H 3

1A ⊗ ξ = ξ δ1A . 1A ⊗ H is in general infinite dimensional, or we say that the func-
tion ξ δ1A at 1A has infinite multiplicity. If H is separable, we are actually attaching
an l 2 sequence at every point A ∈ A.
How to build a Hilbert space out of these H -valued functions? The question
depends on the choice of a quadratic form. If ϕ : A → B(H ) is positive, i.e., ϕ
maps positive elements in A to positive operators on H , then
hA ⊗ ξ , B ⊗ ηiϕ := hξ , ϕ(A∗ B)ηiH
is indeed positive definite. But when extend linearly, one is in trouble. For
* +
∑ Ai ⊗ ξi , ∑ B j ⊗ η j
i j ϕ
=∑ ξi , ϕ(A∗i B j )η j H
i, j
ϕ(A∗1 B1 ) ϕ(A∗1 B2 ) · · · ϕ(A∗1 Bn )

  
ξ1
 ϕ(A∗2 B1 ) ϕ(A∗2 B1 ) · · · ϕ(A∗2 B1 )   ξ2 
i
h  
= ξ1 ξ2 · · · ξn 
 .. .. .. ..  . 
 . 
 . . . .  . 
ϕ(A∗n B1 ) ϕ(A∗n B2 ) · · · ϕ(A∗n Bn ) ξn
5.4 Comments 159
and it is not clear why the matrix (ϕ(A∗i B j ))ni, j=1 should be a positive operator acting
in H ⊗ Cn . But we could very well put this extra requirement into an axiom, so the
CP condition (5.1).
1. We only assume A is a ∗-algebra, not necessarily a C∗ -algebra. ϕ : A →

B(H ) is positive does not necessarily imply ϕ is completely positive. A
counterexample for A = M2 (C), and ϕ : A → B(C2 ) ' M2 (C) given by
taking transpose, i.e., A 7→ ϕ(A) = Atr . Then ϕ is positive, but ϕ ⊗ IM2 is
not.
2. The operator matrix (A∗i A j ), which is also written as ∑i, j A∗i A j ⊗ ei j is a
positive element in A ⊗ Mn . All positive elements in A ⊗ Mn are in such
form. This notation goes back again to Dirac, for the rank-1 projections
|v ih v| are positive, and all positive operators are sums of these rank-1 op-
erators.
3. Given a CP map ϕ : A → B(H ), we get a Hilbert space Kϕ , a represen-
tation π : A → B(Kϕ ) and an isometry V : H → Kϕ , such that
ϕ(A) = V ∗ π(A)V
for all A ∈ A. P = VV ∗ is a selfadjoint projection from Kϕ to the image of

H under the embedding. To see P is a projection, note that
P2 = VV ∗VV ∗ = V (V ∗V )V ∗ = VV ∗ .
Summary
Positive maps have been a recursive theme in functional analysis. A classical

example is A = Cc (X) with a positive linear functional Λ : A → C, mapping A into
a 1-d Hilbert space C.
In Stinespring’s formulation, ϕ : A → H is a CP map, then we may write
ϕ(A) = V ∗ π(A)V where π : A → K is a representation on a bigger Hilbert space
K containing H . The containment is in the sense that V : H ,→ K embeds H
into K . Notice that
V ϕ(A) = π(A)V =⇒ ϕ(A) = V ∗ π(A)V
but not the other way around. In Nelson’s notes [Nel69], we use the notation ϕ ⊂ π
for one representation being the subrepresentation of another. To imitate the situa-
tion in linear algebra, we may want to split an operator T acting on K into operators
action on H and its complement in K . Let P : K → H be the orthogonal projec-
tion. In matrix language, " #
PT P PT P⊥
.
P⊥ T P P⊥ T P⊥
A better looking would be
" # " #
PT P 0 ϕ1 0
=
0 P⊥ T P⊥ 0 ϕ2
hence
π = ϕ1 ⊕ ϕ2 .
Stinespring’s theorem is more general, where the off-diagonal entries may not be
zero.
Exercise 5.1. Let A = B(H ), ξ1 , . . . , ξn ∈ H . The map (ξ1 , . . . , ξn ) 7→ (Aξ1 , . . . , Aξn ) ∈

⊕n H is a representation of A if and only if
n times n times
z }| { z }| {
idH ⊕ · · · ⊕ idH ∈ Rep(A, H ⊕ · · · ⊕ H )
where in matrix notation, we have

  
idA (A) 0 ··· 0 ξ1
0 idA (A) · · · 0
  
   ξ2 

 .. .. .. ..  . 
 . 
 . . . .  . 
0 0 · · · idA (A) ξn
    
A 0 ··· 0 ξ1 Aξ1
0 A ··· 0   ξ2   Aξ2 
    
=
 .. .. .. =
..   ..   .. 
   
.
. . . .  .   . 
0 0 ··· A ξn Aξn
In this case, the identity representation idA : A → H has multiplicity n.
Exercise 5.2. Let Vi : H → H , and

5.4 Comments 161
 
V1
 V2 
 
 ..  : H → ⊕1 H .
n
V :=   (5.21)
 . 
Vn
h i
Let V ∗ : ⊕n1 H → H be the adjoint of V . Prove that V ∗ = V1∗ V2∗ · · · Vn∗ .
Proof. Let ξ ∈ H , then

 
V1 ξ
 V2 ξ 
 
 ..  ,
Vξ =  
 . 
Vn ξ
and
   
η1 V1 ξ
*   +
 η2   V2 ξ 
 .  ,  .  = ∑ hηi ,Vi ξ i = ∑ hVi∗ ηi , ξ i
 .   . 
 .   .  i i
ηn Vn ξ
 
η1
* +
h i
∗

 η2 
= ∗ ∗
V1 V2 · · · Vn  .  ,ξ .

 .. 
ηn
h i
This shows that V ∗ = V1∗ V2∗ · · · Vn∗ .
Exercise 5.3. Let V be as in (5.21). The following are equivalent:
1. V is an isometry, i.e., kV ξ k2 = kξ k2 , for all ξ ∈ H ;

2. ∑ Vi∗Vi = IH ;
3. V ∗V = IH .
Proof. Notice that

* +
2 2
kV ξ k = ∑ kVi ξ k = ∑ hξ ,Vi∗Vi ξ i = ξ,∑ Vi∗Vi ξ .
i i i
Hence kV ξ k2 = kξ k2 if and only if

* +
ξ,∑ Vi∗Vi ξ = hξ , ξ i
i
for all ξ ∈ H . Equivalently, ∑i Vi∗Vi = IH = V ∗V .
Corollary 5.2 (Krauss). Let dimH = n. Then all the CP maps are of the form
ϕ(A) = ∑ Vi∗ AVi .

i
(The essential part here is that for any CP mapping ϕ, we get a system {Vi }.)
This was discovered in the physics literature by Kraus. The original proof was
very intricate, but it is a corollary of Stinespring’s theorem. When dimH = n, let
{e1 , . . . en } be an ONB. Fix a CP map ϕ, and get (V, K , π). Set
Vi : ei 7→ Vei ∈ K , i = 1, . . . n;
then Vi is an isometry. So we get a system of isometries, and

  
A V1
i A   V2 
h  

ϕ(A) = V1∗ V2∗ ··· Vn∗ 
..
 . .
 . 
.

  . 
A Vn
Notice that ϕ(1) = 1 if and only if ∑i Vi∗Vi = 1.
Exercise 5.4 (tensor products). Prove the following.
1. ⊕n1 H ' H ⊗ Cn
2. ∑⊕∞
1 H ' H ⊗l
2
3. Given L2 (X, M, µ), then L2 (X, H ) ' H ⊗ L2 (µ); where L2 (X, H ) con-
sists of all measurable functions f : X → H such that
ˆ
k f (x)k2H dµ(x) < ∞
X
and ˆ
h f , gi = h f (x) , g (x)iH dµ (x) .
X
4. All the spaces above are Hilbert spaces.

5.4 Comments 163
Exercise 5.5 (Using tensor product in representations). Let (Xi , Mi , µi ), i = 1, 2,

be measure spaces. Let πi : L∞ (µi ) → L2 (µi ) be the representation such that πi ( f ) is
the operator of multiplication by f on L2 (µi ). Hence πi ∈ Rep(L∞ (Xi ), L2 (µi )), and
π1 ⊗ π2 ∈ Rep(L∞ (X1 × X2 ), L2 (µ1 × µ2 )),

π1 ⊗ π2 (ϕ̃) f˜ = ϕ̃ f˜
for all ϕ̃ ∈ L∞ (X1 × X2 ), and all f˜ ∈ L2 (µ1 × µ2 ).

Elementary tensors: Special form
ϕe (x1 , x2 ) = ϕ1 (x1 ) ϕ2 (x2 ) ,

fe(x1 , x2 ) = f1 (x1 ) f2 (x2 ) ,
(π1 ⊗ π2 ) (ϕ)
e f = π1 (ϕ1 ) f1 ⊗ π2 (ϕ2 ) f2 .
we suggest: [Arv98, BR81, Sti59, Tak79].
Chapter 6
Brownian Motion
The FBI has a database consisting of some 200 million

fingerprint records... As part of a modernization program, the
FBI is digitizing these records as 8-bit grayscale images, with a
spatial resolution of 500 dots per inch. This results in some 10
megabytes per card, making the current archive about 2,000
terabytes in size. — C.M. Brislawn
The concept of Brownian motion is not traditional included in Functional Analysis.

Below we offer a presentation which relies on almost all the big theorems from func-
tional analysis, and especially on L2 -Hilbert spaces, built from probability measures
on function spaces.
We have included a brief discussion of Brownian motion in order to illustrate
infinite Cartesian products (sect 1.0.1) and unitary one-parameter group {U (t)}t∈R
acting in Hilbert space. See [Jør14, Nel67, Nel59b, Hid80].
Definition 6.1. Let (Ω , F, P) be a probability space, i.e.,
• Ω = sample space
• F = sigma algebra of events
• P = probability measure defined on F.
A function X : Ω → R is called a random variable if it is a measurable function,

i.e., if for all intervals (a, b) ⊂ R the inverse image
X −1 ((a, b)) = ω ∈ Ω X (ω) ∈ (a, b)

(6.1)
is in F; and we write {a < X (ω) < b} ∈ F in short-hand notation.

We say that X is Gaussian (written N m, σ 2 ) if

ˆ b
1 2 2
P ({a < X (ω) < b}) = √ e−(x−m) /2σ dx. (6.2)
a σ 2π
(The function under the integral in (6.2) is called the Gaussian distribution.)
Definition 6.2. Events A, B ∈ F are said to be independent if
165
166 6 Brownian Motion
P (A ∩ B) = P (A) P (B) .
Random variables X and Y are said to be independent iff (Def.) X −1 (I) and
Y −1 (J) are independent for all intervals I and J.
Definition 6.3. A family {Xt }t∈R of random variables for (Ω , F, P) is said to be a

Brownian motion iff (Def.) for every n ∈ N,
1. the random variables Xt1 , Xt2 , . . . , Xtn are jointly Gaussian with
ˆ
E (Xt ) = Xt (ω) dP (ω) = 0, ∀t ∈ R;
Ω
2. if t1 < t2 < · · · < tn then Xti+1 − Xti and Xti − Xti−1 are independent;
3. for all s,t ∈ R,
E |Xt − Xs |2 = |t − s| .
Exercise 6.1. Using Definition 6.3, Corollary 1.2, Remark 1.6 and Example 1.1,
show that there is a unique strongly continuous unitary one-parameter group {U (t)}t∈R
acting in L2 (C (R) , Cyl, P), determined by
U (t) Xs = Xs+t , ∀s,t ∈ R. (6.3)
Hint: By 3 in Definition 6.3, we have

E |Xt2 − Xt1 |2 = E |Xt2 +s − Xt1 +s |2 , ∀t, s1 , s2 ∈ R. (6.4)
Hence, if U (t) is defined on the generator {Xs : s ∈ R} ⊂ L2 (P) as in (6.3), it fol-

lows by (6.4) that it preserves the L2 (P)-norm. Th remaining steps are left to the
reader.
Exercise 6.2. Discuss the infinitesimal generator of {U (t)}t∈R .
Exercise 6.3. Show that {U (t)}t∈R is induced by an ergodic action.
6.1 The Path Space
Theorem 6.1 (see e.g., [Nel67]). Set Ω = C (R) = (all continuous real valued func-
tion on R), F = the sigma algebra generated by cylinder-sets, i.e., determined by
6.1 The Path Space 167
finite systems t1 , . . . ,tn , and intervals J1 , . . . , Jn ;

Cyl (t1 , . . . ,tn , J1 , . . . , Jn ) = ω ∈ C (R) ω (ti ) ∈ Ji , i = 1, 2 . . . , n . (6.5)
The measure P is determined by its value on cylinder sets, and and an integral over
Gaussians; it is called the Wiener-measure. Set
Xt (ω) = ω (t) , ∀t ∈ R, ω ∈ Ω (= C (R)) .
If 0 < t1 < t2 < · · · < tn , and the cylinder set is as in 6.5, then
P (Cyl (t1 , . . . ,tn , J1 , . . . , Jn ))

ˆ ˆ
= · · · gt1 (x1 ) gt2 −t1 (x2 − x1 ) · · · gtn −tn−1 (xn − xn−1 ) dx1 · · · dxn
J1 Jn
where
1 −x2 /2t
gt (x) = √ e , ∀t > 0,
2πt
i.e., the N(0,t)-Gaussian. See Fig 6.1.
gt2 -t1 Ji+1
J2 gti -ti-1
gt1 Jn
gtn -tn-1
J1
Jn-1
Ji
t
t1 t2 ... ti ti+1 ... tn-1 tn
Fig. 6.1: Stochastic processes indexed by time: A cylinder set C is a special subset
of the space of all paths, i.e., functions of a time variable. A fixed cylinder set C is
specified by a finite set of sample point on the time-axis (horizontal), and a corre-
sponding set of “windows” (intervals on the vertical axis). When sample points and
intervals are given, we define the corresponding cylinder set C to be the set of all
paths that pass through the respective windows at the sampled times. In the figure
we illustrate sample points (say future relative to t = 0). Imagine the set C of all
outcomes with specification at the points t1 ,t2 , . . . etc.
I3 t
t1 t2 t3
I2
I1 '
I1
I2 '
I3 '
t
t1 t2 t3
(a) C optimistic (b) C0 pessimistic
I1 '' I2 ''
t
t1 t2 t3
I3 ''
(c) C00 mixed
Fig. 6.2: The cylinder sets C, C0 , and C00 .
Let Ω = ∏∞
k=1 {1, −1} be the infinite Cartesian product of {1, −1} with the prod-
uct topology. Ω is compact and Hausdorff by Tychnoff’s theorem.
For each k ∈ N, let Xk : Ω → {1, −1} be the kth coordinate projection, and assign
probability measures µk on Ω so that µk ◦ Xk−1 {1} = a and µk ◦ Xk−1 {−1} = 1 − a,
where a ∈ (0, 1). The collection of measures {µk } satisfies the consistency con-
dition, i.e., µk is the restriction of µk+1 onto the kth coordinate space. By Kolo-
mogorov’s extension theorem, there exists a unique probability measure P on Ω so
that the restriction of P to the kth coordinate is equal to µk .
It follows that {Xk } is a sequence of independent identically distributed (i.i.d.)
random variables in L2 (Ω , P) with E (Xk ) = 0 and Var[Xk2 ] = 1; and L2 (Ω , P) =
span{Xk }.
Remark 6.1. Let H be a separable Hilbert space with an orthonormal basis {uk }.
The map ϕ : uk 7→ Xk extends linearly to an isometric embedding of H into
L2 (Ω , P). Moreover, let F+ (H ) be the symmetric Fock space. F+ (H ) is the
closed span of the the algebraic tensors uk1 ⊗ · · · ⊗ ukn , thus ϕ extends to an iso-
morphism from F+ (H ) to L2 (Ω , P).
1 1

Exercise 6.4. Let Ω = ∏∞ 1 {−1, 1}, and let µ be the “fair-coin” measure 2 , 2 on
{±1} (e.g., “Head v.s. Tail”), let F be the cylinder sigma-algebra of subsets of Ω .
6.2 Decomposition of Brownian motion 169
Let P = ∏∞ 1 µ be the infinite-product measure on Ω . Set Zk (ω) = ωk , ω = (ωi ) ∈ Ω .

Finally, let ψ j j∈N be an ONB in L2 (0, 1), and set

∞ ˆ t
Xt (ω) := ∑ ψ j (s) ds Z j (ω) , t ∈ [0, 1] , ω ∈ Ω .
j=1 0
Show that the {Xt }t∈[0,1] is Brownian motion, where time “t” is restricted to [0, 1].
Hint: Let t1 ,t2 ∈ [0, 1], then
ˆ t1 ˆ t2 !
E (Xt1 Xt2 ) = E ∑ ψ j (s) ds Z j · ∑ ψk (s) ds Zk
j 0 k 0
ˆ t1 ˆ t2
= ∑∑ ψ j (s) ds ψk (s) dsE (Z j Zk )
j k 0 0 | {z }
=δ jk
= ∑ χ[0,t1 ] , ψ j L2
χ[0,t2 ] , ψ j L2
j
= χ[0,t1 ] , χ[0,t2 ] L2 (0,1)

= t1 ∧ t2 (:= min (t1 ,t2 )) .
6.2 Decomposition of Brownian motion
The integral kernel K : [0, 1] × [0, 1] → R
K(s,t) = s ∧ t
is a compact operator on L2 [0, 1], where

ˆ
K f (x) = (x ∧ y) f (y)dy.
K f is a solution to the differential equation
d2
− u= f
dx2
with zero boundary conditions.

K is also seen as the covariance functions of Brownian motion process. A
stochastic process is a family of measurable functions {Xt } defined on some sample
probability space (Ω , B, P), where the parameter t usually represents time. {Xt } is
a Brownian motion process if it is a mean zero Gaussian process such that
ˆ
E[Xs Xt ] = Xs Xt dP = s ∧ t.
Ω
It follows that the corresponding increment process {Xt − Xs } ∼ N(0,t − s). P is

called the Wiener measure.
Building (Ω , B, P) is a fancy version of Riesz’s representation theorem as in
Theorem 2.14 of Rudin’s book. It turns out that
Ω = ∏ R̄
t
which is a compact Hausdorff space.
Xt : Ω → R
is defined as
Xt (ω) = ω(t)
i.e. Xt is the continuous linear functional of evaluation at t on Ω .

For Brownian motion, the increment of the process 4Xt , in some statistical sense,
√
is proportional to 4t. i.e.
p
4Xt ∼ 4t.
It it this property that makes the set of differentiable functions have measure zero.
In this sense, the trajectory of Brownian motion is nowhere differentiable.
An very important application of the spectral theorem of compact operators in to
decompose the Brownian motion process.
∞
sin(nπt)
Bt (ω) = ∑ Zn (ω)
n=1 nπ
where ∞
sin (nπs) sin (nπt)
s∧t = ∑
n=1 (nπ)2
and Zn ∼ N(0, 1).
Exercise 6.5. Look up the Central Limit Theorem (CLT), and prove the following
approximation formula for Brownian motion:
6.2 Decomposition of Brownian motion 171
Let π be the “fair-coin-measure” on the two outcomes {±1}, i.e., winning or

loosing one unit, and let Ω = "N {±1}, P = "N π be the corresponding infinite prod-
uct measure. On Ω , set
Wk (ω) = ωk , ω = (ω1 , ω2 , . . .) ∈ Ω , k = 1, 2, . . . ; and
1 n
Sn (·) = √ ∑ Wk (·) . (6.6)
n k=1
Let Xt denote Brownian motion. Then show that
1 bntc
Xt (·) = lim √ ∑ Wk (·) (6.7)
n→∞ n
k=1
where bntc denotes the largest integer ≤ nt.

Hint: A good reference to the CLT is [CW14]. First approximate the CLT to the
sequence Sn in (6.6). Fig 6.3 illustrates the approximation formula in (6.7).
The Central Limit Theorem (CLT) states that the limit, n → ∞, of the sequence
Sn in (6.6) is a copy of N (0, 1)-random variable; i.e., limn→∞ Sn (·) = S (·) exists;
and ˆ b
1 1 2
√ e− 2 x dx,

P ω a ≤ S (ω) ≤ b =
a 2π
for all intervals (a, b) ⊂ R. Applying this to (6.7), we get existence of Xt as a limit,
and Xt ∼ N (0, 1), i.e.,
ˆ b
1 1 2
√ e− 2 x dx.

P ω a ≤ X (ω) ≤ b =
a 2π
We claim that
EP (Xs Xt ) = s ∧ t. (6.8)
Below we sketch the argument for the assertion in (6.8).

Fix s,t ∈ R+ , say s < t, and n ∈ N; then
! !!
bn sc bntc
1 1 1 bn sc bntc
E √ ∑ Wj √ ∑ Wk = ∑ ∑ E (W jWk )
n j=1 n k=1 n j=1 k=1
1 bn sc bntc
= ∑ ∑ δ j,k
n j=1 k=1
bn sc
= → s, as n → ∞.
n
Xt Xt
t t
0.5 1 0.5 1
(a) n = 10 (b) n = 50
Xt Xt
t t
0.5 1 0.5 1
(c) n = 100 (d) n = 500
Fig. 6.3: Monte-Carlo simulation of the standard Brownian motion process

{Xt : 0 ≤ t ≤ 1}, where E (Xt ) = 0, and E (Xs Xt ) = s ∧ t = min (s,t).
(n) j
For n = 10, 50, 100, 500, set X0 = 0 and X j/n = n−1/2 ∑k=1 Wk . Applying linear in-
terpolation between sample points { j/n : j = 0, . . . , n} yields the n-point approxi-
(n)
mation Xt , which converges in measure to Xt (standard BM restricted to the unit
interval), as n → ∞.
Hence by the CLT, the desired conclusion in 6.8 follows.

This is the key step in proving that the limit Xt in (6.7) is Brownian motion. The
remaining steps are routine left to the readers.
we suggest: [Hid80, Itô04, Itô06, Itô07, Nel67, Par09].
Chapter 7
Applications to Groups
Every axiomatic (abstract) theory admits, as is well known, an

unlimited number of concrete interpretations besides those from
which it was derived. Thus we find applications in fields of
science which have no relation to the concepts of random event
and of probability in the precise meaning of these words.
— A.N. Kolmogorov
We studied operators in Hilbert space of relevance to quantum physics. A source

of examples is relativistic physics. The symmetry group of Einstein’s theory is a
particular Lie group, called the Poincaré group. The study of its unitary representa-
tions is central to relativistic physics. But it turns out that there is a host of diverse
applications (including harmonic analysis) where other groups arise. Below we of-
fer a glimpse of the theory of unitary representations, and its many connections to
operators in Hilbert space.
7.0.1 Motivation
Every group G is also a ∗-semigroup, where the ∗-operation is given by
g∗ := g−1 .
G is not a complex ∗ algebra yet, in particular, multiplication by a complex num-

ber is not defined. As a general principal, G can be embedded into the ∗-algebra
AG = G ⊗ C, i.e. C-valued functions on G. Pointwise multiplication and scalar mul-
tiplication are defined naturally,
(g ⊗ cg )(h ⊗ ch ) = gh ⊗ cg ch
t(g ⊗ cg ) = g ⊗ tcg
with g, h ∈ G and cg , ch ,t ∈ C. The ∗-operation extends from G to AG as
173
174 7 Applications to Groups
(g ⊗ cg )∗ = g−1 ⊗ cg−1 .
Note: The ∗ operation so defined is the only way to make it a period-2, conjugate
linear, anti-automorphism. That is, the ∗-operation on AG is conjugate linear
and satisfies A∗∗ = A, and (AB)∗ = B∗ A∗ , for all A, B ∈ AG .
There is a bijection between representations of G and representations of AG . For

if π ∈ Rep(G, H ), it extends to π̃ = U ⊗ IdC ∈ Rep(AG , H ), where
π̃(g ⊗ cg ) = πg ⊗ cg .
Conversely, if ρ ∈ Rep(AG , H ) then
π =ρ G⊗1
is a representation of G ' G ⊗ 1.
Note: The notation g ⊗ cg is usually written as cg , which contains all the informa-
tion. Thus the pointwise multiplication takes the form
cg dh = lgh .
Equivalently,
lg = ∑ ch dh−1 g
h
which is the usual convolution. In more detail:

! !
∑ cg1 π (g1 ) ∑ dg2 π (g2 ) = ∑ lg π (g) .
g1 g2 g
The ∗-operation becomes

c∗g = cg−1 .
More generally, every locally compact group has a left (and right) Haar measure.
Thus the above construction has a “continuous” version. It turns out that AG is
the Banach ∗-algebra L1 (G). Again, there is a bijection between Rep(G, H ) and
Rep(L1 (G), H ). In particular, the “discrete” version is recovered if the measure µ
is discrete, in which case AG = l 1 (G).
7 Applications to Groups 175
7.0.2 Group - Algebra - Representations
Everything we say about algebras is also true for groups. In physics, we are inter-
ested in representation of symmetry groups, which preserve inner product or energy.
We always want unitary representations. Irreducible representation amounts to ele-
mentary particles which can not be broken up further. In practice, quite a lot work
goes into finding irreducible representations of symmetry groups. The idea is to go
from groups to algebras and then to representations.
G→A→π
• πG ∈ Rep(G, H ) 


π(g1 g2 ) = π(g1 )π(g2 )

π(eG ) = IH


π(g)∗

= π(g−1 )
• πA ∈ Rep(A, H ) 


π(A1 A2 ) = π(A1 )π(A)2

π(1A ) = IH


π(A)∗

= π(A∗ )
Case 7.1. G is discrete −→ A = G ⊗ l 1

! !
∑ a(g)g ∑ b(h)h = ∑ a(g)b(h)gh
g g g,h
= ∑ ∑ a(g0 h−1 )b(h)g0

g0 h
!∗ !
∑ c(g)g = ∑ c(g−1 )g
g g
where c∗ (g) = c(g−1 ). The multiplication of functions in A is a generalization of

convolutions.
G is locally compact −→ A = G ⊗ L1 (µ) ' L1 (G).
Existence of Haar measure: easy proof for compact groups, and extend to locally
compact cases. For non compact groups, the left / right Haar measures could be dif-
ferent. If they are always equal, the group is called unimodular. Many non compact
groups have no Haar measure.
Let B(G) be the Borel σ -aglebra of G, E ∈ B(G). Left Haar: λL (gE) = λL (E);
right Haar: λR (Eg) = λR (Eg). Theorem: the two measures are equivalent
λL λR λL
dλL
4G = :G→G
dλR
is called the modular function, 4G is a homomorphism, 4G (gh) = 4G (g)4G (h).
L1 (G)
ˆ
(ϕ1 ? ϕ2 )(g) = ϕ1 (gh−1 )ϕ2 (h)dλR (h)
ϕ ∗ (g) = ϕ(g−1 )4(g−1 )
Take C∗ completion in both cases!
Remark 7.1. (1) for the case l 1 (G), the counting measure in unimodular, hence 4(g)
does not appear; (2) In ϕ1 (gh−1 ), h−1 appears since operation on points is dual to
operation on functions. A change of variable shows
ˆ ˆ ˆ
ϕ(gh−1 )dλR (g) = ϕ(g0 )dλR (g0 h) = ϕ(g0 )dλR (g0 )
(3) L1 (G) is a Banach algebra. Fubini’s theorem shows that f ? g ∈ L1 (G), for all
f , g ∈ L1 (G).
There is a bijection between representations of groups and representations of alge-

bras. Given a unitary representation π ∈ Rep(G, H ), take a Haar measure dg in
L1 (G), then we get the group algebra representation
ˆ
πL1 (G) (ϕ) = ϕ(g)π(g)dg
G
πL1 (G) (ϕ1 ? ϕ2 ) = πL1 (G) (ϕ1 )πL1 (G) (ϕ2 )
πL1 (G) (ϕ)∗ = πL1 (G) (ϕ ∗ )
Conversely, given a representation of L1 (G), let (ϕi ) be a sequence in L1 such that

ϕi → δg . Then
7.1 Some examples 177
ˆ
ϕi (h)π(h)gdh → π(g)
i.e. the limit is a representation of G.
7.1 Some examples
7.1.1 ax+b group

" #
ab
a ∈ R+ , b ∈ R
01
" #" # " #
a0 b0 ab a0 a a0 b + b0
=
0 1 01 0 1
" #−1 " #
1
ab a − ab
=
01 0 1
0
The multiplication aa0 can be made into addition, by taking a = et , a0 = et so that
t0
et e = e t+t 0 . This is a transformation group
x 7→ ax + b
where composition gives
x 7→ ax + b 7→ a0 (ax + b) + b0 = aa0 x + (a0 b + b0 )
Left Haar measure:

ˆ ˆ ˆ
f (h−1 g)g−1 dg = f (g0 )(hg0 )−1 d(hg0 ) = f (g0 )g0−1 dg−1
" #" #
1
− ba da db dadb
dλL = g−1 dg = a =
0 1 0 0 a2
check:
0 a b−b0
" # " # " #" # " #
ab a0 b0 −1
1
a0 − ba0 ab a0 a0
g= ,h= ,h g= =
01 0 1 0 1 01 0 1
ˆ ˆ
a b − b0 dadb
f (h−1 g)dλL (g) = f( 0, ) 2
a a0 a
ˆ
d(a0 s)d(a0t + b0 )
= f (s,t)
(a0 s)(a0 s)
ˆ
dsdt
= f (s,t) 2
s
where with a change of variable
a
s= , da = a0 ds
a0
b − b0
t= , db = a0 dt
a0
dadb a02 dsdt dsdt
2
= 0 2
= 2
a (sa ) s
Right Haar:
ˆ ˆ
f (gh−1 )(dg)g−1 = f (g0 )d(g0 h)(g0 h)−1
ˆ
= f (g0 )(dg0 )(hh−1 )g0−1
ˆ
= f (g0 )(dg0 )g0−1
" #" #
1
−1 da db a − ba dadb
dλR = (dg)g = =
0 0 0 1 a
check:
0 0
" # " # " #" # " #
ab a0 b0 ab 1
a0 − ab0 a
a0 − ab
a0 + b
g= ,h= , gh−1 = =
01 0 1 01 0 1 0 1
ˆ ˆ
−1 a ab0 dadb
f (gh )dλR (g) = f ( 0 , − 0 + b)
a a a
ˆ 0
a dsdt
= f (s,t) 0
as
ˆ
dsdt
= f (s,t)
s
where with a change of variable

7.2 Induced representation 179
a
s= , da = a0 ds
a0
ab0
t =− + b, db = dt
a0
dadb a0 dsdt dsdt
= =
a a0 s s
7.2 Induced representation
Most of the Lie groups considered here fall in the following class:
Let V be a finite-dimensional vector space over R (or C). We will consider the
real case here, but the modifications needed for the complex case are straightfor-
ward.
Let q : V × V → R be a non-degenerate form, i.e., bilinear s.t. v → q (v, ·) ∈ V ∗
is 1-1. Let GL (V ) be the general linear group for V , i.e., all invertible linear maps
V → V . (If a basis in V is chosen, this will be a matrix-group.)
Lemma 7.1. Set

G (q) = g ∈ GL (V ) q (gu, gv) = q (u, v) , ∀u, v ∈ V . (7.1)
Then G (q) is a Lie group, and its Lie algebra consists of all linear mapping X : V →
V such that
q (Xu, v) + q (u, Xv) = 0, ∀u, v ∈ V. (7.2)
Proof. Fix a linear mapping X : V → V , and set

∞
tn
gX (t) = ∑ n! X n = exp (tX)
n=0
i.e., the matrix-exponential. Note that gX (t) satisfies (7.1) for all t ∈ R iff X satisfies
(7.2). To see this, differentiate: i.e., compute
d
q (exp (tX) , exp (tX) v)
dt
using that q is assumed bilinear.
We will address two questions: (1) How to induce a representation of a group

G from a representation of the a subgroup Γ ⊂ G? (2) Given a representation of
a group G, how to test whether it is induced from a representation of a subgroup

Γ ⊂ G? The main examples we have looked at so far are the Lie groups of
• ax + b
• Heisenberg
• SL2 (R)
• Lorentz
• Poincaré
Among these, the ax + b, Heisenberg and Poincaré groups are semi-direct prod-
uct groups. Their representations are induced from normal subgroups. It is ex-
tremely easy to find representations of abelian subgroups. Unitary representation
of abelian subgroups are one-dimensional, but the induced representation on an en-
larged Hilbert space is infinite dimensional. See the appendix for a quick review of
semi-direct product.
" #
ab
Example 7.1. The ax + b group (a > 0). G = {(a, b)}, where (a, b) = . The
01
multiplication rule is given by
(a, b)(a0 , b0 ) = (aa0 , ab0 + b)

1 b
(a, b)−1 = ( , − ).
a a
Γ = {(1, b)} is a one-dimensional abelian, normal subgroup of G.
• abelian: (1, b)(1, c) = (1, c + b)

• normal: (x, y)(1, b)(x, y)−1 = (1, xb), note that this is also Adg acting on the
normal subgroup Γ
• The other subgroup {(a, 0)} is isomorphic to the multiplicative group (R+ , ×).
Because we have
(a, 0)(a0 , 0) = (aa0 , 0)
by the group multiplication rule above.

• Notice that (R+ , ×) is not a normal subgroup, since (a, b)(x, 0)( 1a , − ba ) =
(ax, b)( 1a , − ba ) = (x − bx + b).
The multiplicative group (R+ , ×) acts on the additive group (R, +) by
ϕ : (R+ , ×) 7→ Aut((R, +))

ϕa (b) = ab
check:
(a, b)(a0 , b0 ) = (aa0 , b + ϕa (b0 ))

= (aa0 , b + ab0 )
(a, b)−1 = (a−1 , ϕa−1 (b−1 ))
1 b
= (a−1 , a−1 (−b)) = ( , − )
a a
(a, b)(1, x)(a, b−1 ) = (a, b + ϕa (x))(a, b−1 )
1 b
= (a, b + ax)( , − )
a a
= (1, b + ax − b)
= (1, ax)
= ϕa (x)
" # " #
10 01
Example 7.2. The Lie algebra of G is given by X = ,= . We check
00 00
that " #
tX et 0
e =
0 1
which is subgroup (R+ , ×); and

" #
sY 1s
e = I + sY + 0 + · · · + 0 =
01
which is subgroup (R, +). We also have [X,Y ] = Y .
Example 7.3. Form L2 (µL ) where µL is the left Haar measure. Then π : g →
π(g) f (x) = f (g−1 x) is a unitary representation. Specifically, if g = (a, b) then
x y−b
f (g−1 x) = f ( , ).
a a
Differentiate along the a direction we get
d x y−b ∂ ∂
X̃ f = a=1,b=0
f( , ) = (−x − y ) f (x, y)
da a a ∂x ∂y
d x y−b ∂
Ỹ f = a=1,b=0
f( , ) = − f (x, y)
db a a ∂y
therefore we have the vector field

∂ ∂
X̃ = −x −y
∂x ∂y
∂
Ỹ = −
∂y
or equivalently we get the Lie algebra representation dπ on L2 (µL ). Notice that
[X̃, Ỹ ] = X̃ Ỹ − Ỹ X̃
∂ ∂ ∂ ∂ ∂ ∂
= (−x − y )(− ) − (− )(−x − y )
∂x ∂y ∂y ∂y ∂x ∂y
∂2 ∂2 ∂2 ∂ ∂2
=x + y 2 − (x + +y 2)
∂ x∂ y ∂y ∂ x∂ y ∂ y ∂y
∂
=−
∂y
= Ỹ .
Notice that X̃ and Ỹ can be obtained by the exponential map as well.
d
X̃ f = t=0
f (e−tX x)
dt
d
= t=0
f ((e−t , 1)(x, y))
dt
d
= t=0
f (e−t x, e−t y + 1)
dt
∂ ∂
= (−x − y ) f (x, y)
∂x ∂y
d
Y˜f = t=0
f (e−tY x)
dt
d
= t=0
f ((1, −t)(x, y))
dt
d
= t=0
f (x, y − t)
dt
∂
=− f (x, y)
∂y
Example 7.4. We may parametrize the Lie algebra of the ax+b group using (x, y)
variables. Build the Hilbert space L2 (µL ). The unitary representation π(g) f (σ ) =
f (g−1 σ ) induces the follows representations of the Lie algebra
d
dπ(s) f (σ ) = s=0
f (e−sX σ ) = X̃ f (σ )
dx
d
dπ(t) f (σ ) = t=0
f (e−tY σ ) = Ỹ f (σ ).
dy
Hence in the parameter space (s,t) ∈ R2 we have two usual derivative operators
∂ /∂ s and ∂ /∂t, where on the manifold we have
∂ ∂ ∂
= −x − y
∂s ∂x ∂y
∂ ∂
=−
∂t ∂y
The usual positive Laplacian on R2 translates to

2 2
∂ ∂
−4 = +
∂s ∂t
= (X̃)2 + (Ỹ )2
∂ 2

∂ ∂ ∂ ∂
= −x − y −x − y + −
∂x ∂y ∂x ∂y ∂y
∂ 2 ∂ 2 ∂ 2 ∂ ∂
= x2 2 + 2xy + y2 + 1 2
+x +y ,
∂x ∂ x∂ y ∂y ∂x ∂y
2 2
where we used x ∂∂x = x2 ∂∂x + x ∂∂x . This is in fact an elliptic operator, since
the matrix " #
x2 xy
xy y2 + 1
has trace trace = x2 + y2 + 1 ≥ 1, and det = x2 ≥ 0. If instead we have “y2 ” then the
determinant is the constant zero.
Example 7.5. Notice the term “y2 + 1” is essential for 4 being elliptic. Also notice
that all the coefficients are analytic functions in the (x, y) variables.
Note: Notice the Γ is unimodular, hence it is just a copy of R. Its invariant measure
is the Lebesgue measure on R.
Example 7.6. Heisenberg group G = {a, b, c} where

 
1ac
(a, b, c) =  0 1 b 
 
001
The multiplication rule is given by

(a, b, c)(a0 , b0 , c0 ) = (a + a0 , b + b0 , c + ab0 + c0 )

(a, b, c)−1 = (−a, −b, −c + ab)
The subgroup Γ = {(0, b, c)} where

 
10c
(1, b, c) =  0 1 b 
 
001
is two dimensional, abelian and normal.
• abelian: (0, b, c)(0, b0 , c0 ) = (0, b + b0 , c + c0 )

• normal:
(a, b, c)(0, x, y)(a, b, c)−1 = (a, b, c)(0, x, y)(−a, −b, −c + ab)

= (a, b + x, c + y + ax)(−a, −b, −c + ab)
= (0, x, y + ax + ab − ab)
= (0, x, ax + y)
Note that this is also Adg acting on the Lie algebra of Γ .
The additive group (R, +) acts on Γ = {(0, b, c)} ' (R2 , +) by
ϕ : (R, +) → Aut(Γ )
" # " #" #
c 1a c
ϕ(a) =
b 01 b
" #
c + ab
=
b
check:
(a, (b, c))(a0 , (b0 , c0 )) = (a + a0 , (b, c) + ϕ(a)(b0 , c0 ))

= (a + a0 , (b, c) + (b0 , c0 + ab0 ))
= (a + a0 , b + b0 , c + c0 + ab)
(a, (b, c))−1 = (−a, ϕa−1 (−b, −c))
= (−a, (−b, −c + ab))
= (−a, −b, −c + ab)
(a, b, c)(0, b0 , c0 )(a, b, c)−1 = (a, b + b0 , c + c0 + ab0 )(−a, −b, −c + ab)

= (0, b0 , c0 + ab0 )
" #
c0
= ϕa 0
b
7.2.1 Induced Representation
This also goes under the name of “Mackey machine”. Its modern formulation is in
the context of completely positive map.
Let G be a locally compact group, and Γ ⊂ G be a closed subgroup. Recall the
modular functions come in when the translation was put on the wrong side, i.e.
ˆ ˆ
−1
f (gx)dx = 4(g ) f (x)dx
G G
or equivalently, ˆ ˆ
4(g) f (gx)dx = f (x)dx.
G G
δ is defined similarly for the measure dξ on Γ . Form the quotient M = Γ \G space,

and let π : G → Γ \G be the quotient map (the covering map). M carries a transitive
G action.
group right Haar measure modular function

G dg 4
Γ dξ δ
Note: M is called a fundamental domain or a homogeneous space. M is a group if

and only if Γ is a normal subgroup in G. In general, M may not be a group,
but it is still a very important manifold.
Note: µ is called an invariant measure on M, if µ(Eg) = µ(E), ∀g ∈ G.

µ is said to be quasi-invariant, if µ(E) = 0 ⇔ µ(Eg) = 0, ∀g.
In general there is no invariant measures on M, but only quasi-invariant mea-
sures.
G has an invariant measure if and only if G is unimodular (for example, G is
the Heisenberg group). Not all groups are unimodular. A typical example is
the ax + b group.
Define τ : Cc (G) → Cc (M) by

ˆ
(τϕ)(π(x)) = ϕ(ξ x)dξ .
Γ
Lemma 7.2. τ is surjective.
Note: Since ϕ has compact support, the integral is well-defined. τ is called condi-
tional expectation. It is simply the summation of ϕ over the orbit Γ x. This
is because if ξ runs over Γ , ξ x runs over Γ x. τϕ may also be interpreted as
taking average, only it does not divide out the total mass, but that only differs
by a constant.
Note: We may also say τϕ is a Γ -periodic extension, by looking at it as a

function defined on G. Then we check that
ˆ
(τϕ)(ξ1 x) = ϕ(ξ ξ1 x)dξ = τϕ(x)
Γ
because dξ is a right Haar measure. Thus τϕ, treated as function on

G, is Γ -periodic, in the sense that
τϕ(ξ x) = τϕ(x), ∀ξ ∈ Γ .
Example 7.7. G = R, Γ = Z with dξ being the counting measure on Z.

ˆ
(τϕ)(π(x)) = ϕ(ξ x)dξ = ∑ ϕ(z + x)
Γ z∈Z
As a consequence, τϕ is left translation invariant by integers, i.e. τϕ (as a function

on R) is Z-periodic,
(τϕ)(z0 + x) = ∑ ϕ(z0 + z + x)
z∈Z
= ∑ ϕ(z + x)
z∈Z
Since ϕ has compact support, ϕ(z + x) vanishes for all but a finite number of z.
Hence τϕ contains a finite summation, and so it is well-defined.
Let L : Γ → V be a unitary representation of Γ on a Hilbert space V . We now

construct a unitary representation U ind : G → H of G on an enlarged Hilbert space
H.
Let F∗ be the set of function f : G → V so that
f (ξ g) = ρ(ξ )1/2 Lξ f (g), ∀ξ ∈ Γ
where ρ = δ /4. For all f ∈ F∗ , let
(Rg f )(·) := f (·g)
be the right translation of f by g ∈ G. We claim that Rg f ∈ F∗ as well. That is, F∗ is

invariant under right translation by g ∈ G. To see this, let f ∈ F∗ then
(Rg f )(ξ x) = f (ξ xg) = ρ(ξ )1/2 Lξ f (xg) = ρ(ξ )1/2 Lξ (Rg f )(x)
so that Rg f ∈ F∗ .
Note: The factor ρ(ξ )1/2 comes in, since later we will defined an inner product on
F∗ so that k f (ξ ·)knew = k f (·)knew , ∀ξ ∈ Γ . Eventually, we will define the
induced representation by (Ugind f )(·) := (Rg f )(·), not on F∗ , but pass to a
quotient space.
Note: Let’s ignore the factor ρ(ξ )1/2 for a moment. Lξ is unitary implies
that for all f ∈ F∗ ,
k f (ξ ·)kV = kLξ f (·)kV = k f (·)kV , ∀ξ ∈ Γ .
Since Hilbert spaces exist up to unitary equivalence, Lξ f (g) and f (g)

really are the same function. As ξ running through Γ , ξ g running
through Γ g. Thus k f (ξ g)k is a constant on the orbit Γ g. It follows
that f (ξ g) is in fact a V -valued function defined on the quotient M =
Γ \G. We will later use these functions as multiplication operators.
Example 7.8. The Heisenberg group is unimodular, ρ = 1.
Example 7.9. For the ax + b group,
dadb
dλR =
a
dadb
dλL =
a2
dλL 1
4= =
dλR a
On the abelian normal subgroup Γ = {(1, b)}, where a = 1, 4(ξ ) = 1. Γ is uni-

modular, hence δ (ξ ) = 1. Therefore, ρ(ξ ) = δ (ξ )/4(ξ ) = 1, ∀ξ ∈ Γ .
For all f ∈ F∗ , the map µ f , f : Cc (M) → C given by

ˆ
µ f , f : τϕ 7→ k f (g)kV2 ϕ(g)dg
G
is a positive linear functional. By Riesz’s theorem, there exists a unique Radon mea-
sure µ f , f on M, such that
ˆ ˆ
k f (g)kV2 ϕ(g)dg = (τϕ)dµ f , f .
G M
Note: Recall that given a measure space (X, M, µ), let f : X → Y . Define a linear
functional Λ : Cc (Y ) → C by
ˆ
Λ ϕ := ϕ( f (x))dµ(x)
Λ is positive, hence by Riesz’s theorem, there exists a unique regular Borel

measure µ f on Y so that
ˆ ˆ
Λϕ = ϕdµ f = ϕ( f (x))dµ(x).
Y X
It follows that µ f = µ ◦ f −1 .
Note: Under current setting, we have a covering map π : G → Γ \G =: M,

and the right Haar measure µ on G. Thus we may define a measure
µ ◦ π −1 . However, given ϕ ∈ Cc (M), ϕ(π(x)) may not have compact
support, or equivalently, π −1 (E) is Γ periodic. For example, take G =
R, Γ = Z, M = Z\R. Then π −1 ([0, 1/2)) is Z-periodic, which has
infinite Lebesgue measure. What we really need is some map so that
the inverse of a subset of M is restricted to a single Γ period. This
is essentially what τ does. Taking τϕ ∈ Cc (M), get the inverse image
ϕ ∈ Cc (G). Even if ϕ is not restricted in a single Γ period, ϕ always
has compact support.
Hence we get a family of measures indexed by elements in F∗ . If choosing f , g ∈ F∗

then we get complex measures µ f ,g . (using polarization identity)
• Define k f k2 := µ f , f (M),
• Complete F∗ with respect to this norm to get an enlarged Hilbert space H .

• Define induced representation U ind on H as
Ugind f (x) = f (xg)
Ugind is unitary,
Ugind f = k f kH .
H
´
Note: µ f ,g (M) = M τϕdξ , where τϕ ≡ 1. What is ϕ then? It turns out that ϕ could
be constant 1 over a single Γ -period, or ϕ could spread out to a finite number
of Γ -periods. Therefore, in this case
ˆ
2
kfk = k f (g)kV2 ϕ(g)dg
ˆ G
= k f (g)kV2 ϕ(g)dg
1−period
ˆ
= k f (g)kV2 dg
1−period
ˆ
= k f (g)kV2 dg
M
Define P(ψ) f (x) := ψ(π(x)) f (x), for ψ ∈ Cc (M), f ∈ H , x ∈ G. P(ψ) ψ ∈ Cc (M)

is the abelian algebra of multiplication operators.
Lemma 7.3. We have

Ugind P(ψ)Ugind
−1 = P(ψ(·g)).
Proof. Check:
Ugind P(ψ) f (x) = Ugind ψ(π(x)) f (x)

= ψ(π(xg)) f (xg)
P(ψ(·g))Ugind f (x) = P(ψ(·g)) f (xg)
= ψ(π(xg)) f (xg).
Conversely, how to recognize induced representation? Answer:
Theorem 7.1 (Imprimitivity [Ørs79]). Let G be a locally compact group with a

closed subgroup Γ . Let M = Γ \G. Suppose the system (U, P) satisfies the covari-
ance relation,
Ug P(ψ)Ug−1 = P(ψ(·g)),
and P (·) is non-degenerate. Then, there exists a unitary representation L ∈ Rep(Γ ,V )

such that U ' indΓG (L).
7.3 Example - Heisenberg group
Theorem 7.2. The Schrödinger representations are induced.
Let G = {(a, b, c)} be the Heisenberg group, where

 
1ac
(a, b, c) =  0 1 b 
 
001
The multiplication rule is given by
(a, b, c)(a0 , b0 , c0 ) = (a + a0 , b + b0 , c + c0 + ab0 )

(a, b, c)−1 = (−a, −b, −c + ab)
The subgroup Γ = {(0, b, c)} where

 
10c
(1, b, c) =  0 1 b 
 
001
is two dimensional, abelian and normal.
• abelian: (0, b, c)(0, b0 , c0 ) = (0, b + b0 , c + c0 )

• normal:
(a, b, c)(0, x, y)(a, b, c)−1 = (a, b, c)(0, x, y)(−a, −b, −c + ab)

= (a, b + x, c + y + ax)(−a, −b, −c + ab)
= (0, x, y + ax + ab − ab)
= (0, x, ax + y)
i.e. Ad : G → GL(n), as
7.3 Example - Heisenberg group 191
Ad(g)(n) : = gng−1
(x, y) 7→ (ax + y)
the orbit is a 2-d transformation.
Fix h ∈ R\{0}. Recall the Schrödinger representation of G on L2 (R)
Ug f (x) = eih(c+bx) f (x + a)
We show that the Schrödinger representation is induced from a unitary representa-

tion L on the subgroup Γ .
1. Let L ∈ Rep(Γ ,V ) where Γ = {(0, b, c)}, V = C,
Lξ (b,c) = eihc .
The complex exponential comes in since we want a unitary representation.

The subgroup {(0, 0, c)} is the center of G. What is the induced represen-
tation? Is it unitarily equivalent to the Schrödinger representation?
2. Look for the family F∗ of functions f : G → C (V is the 1-d Hilbert space
C), such that
f (ξ (b, c)g) = Lξ f (g).
Since
f (ξ (b, c)g) = f ((0, b, c)(x, y, z)) = f (x, b + y, c + z)

Lξ (b,c) f (g) = eihc f (x, y, z)
f (x, y, z) satisfies
f (x, b + y, c + z) = eihc f (x, y, z).
i.e. we may translate the y, z variables by arbitrary amount, and the only
price to pay is multiplicative factor eihc . Therefore f is really a function
defined on the quotient
M = Γ \G ' R.
M = {(x, 0, 0)} is identified with R, and the invariant measure on the ho-
mogeneous space M is simply the Lebesgue measure. It is almost clear
at this point why the induced representation is unitarily equivalent to the

Schrödinger representation on L2 (R).
´
3. τϕ 7→ G k f (g)kV2 ϕ(g)dg induces a measure µ f , f on M. This can be seen
as follows.
ˆ ˆ
k f (g)kV2 ϕ(g)dg = | f (x, y, z)|2 ϕ(x, y, z)dxdydz
G G'R3
ˆ ˆ
2
= | f (x, y, z)| ϕ(x, y, z)dydz dx
M'R Γ 'R2
ˆ ˆ
2
= | f (x, y, z)| ϕ(x, y, z)dydz dx
R2
ˆR
= | f (x, y, z)|2 (τϕ)(π(g))dx
ˆR
= | f (x, y, z)|2 (τϕ)(x)dx
ˆR
= | f (x, 0, 0)|2 (τϕ)(x)dx
R
where
ˆ
(τϕ)(π(g)) = ϕ(ξ g)dξ
ˆΓ
= ϕ((0, b, c)(x, y, z))dbdc
2
ˆR
= ϕ(x, b + y, c + z)dbdc
2
ˆR
= ϕ(x, b, c)dbdc
2
ˆR
= ϕ(x, y, z)dydz
R2
= (τϕ)(x).
Hence Λ : Cc (M) → C given by

ˆ
Λ : τϕ 7→ k f (g)kV2 ϕ(g)dg
G
is a positive linear functional, therefore
Λ = µf,f
7.3 Example - Heisenberg group 193
i.e. ˆ ˆ
2
| f (x, y, z)| ϕ(x, y, z)dxdydz = (τϕ)(x)dµ f , f (x).
R3 R
4. Define
ˆ ˆ ˆ
k f k2ind := µ f , f (M) = 2
| f | dξ = 2
| f (x)| dx = | f (x, 0, 0)|2 dx
M R R
Ugind f (g0 ) := f (g0 g)
By definition, if g = g(a, b, c), g0 = g0 (x, y, z) then
Ugind f (g0 ) = f (g0 g)

= f ((x, y, z)(a, b, c))
= f (x + a, y + b, z + c + xb)
and U ind is a unitary representation by the definition of k f kind .

5. To see U ind is unitarily equivalent to the Schrödinger representation on
L2 (R), define
W : H ind → L2 (R)
(W f )(x) = f (x, 0, 0)
If put other numbers into f , as f (x, y, z), the result is the same, since
f ∈ H ind is really defined on the quotient M = Γ \G ' R.
W is unitary:
ˆ ˆ ˆ
kW f k2L2 = 2
|W f | dx = 2
| f (x, 0, 0)| dx = | f |2 dξ = k f k2ind
R R Γ \G
Intertwining: let Ug be the Schrödinger representation.
Ug (W f ) = eih(c+bx) f (x + a, 0, 0)
WUgind f = W ( f ((x, y, z)(a, b, c)))
= W ( f (x + a, y + b, z + c + xb))

= W eih(c+bx) f (x + a, y, z)
= eih(c+bx) f (x + a, 0, 0)
6. Since {U, L}0 ⊂ {L}0 , then the system {U, L} is reducible implies L is re-
ducible. Equivalent, {L} is irreducible implies {U, L} is irreducible. Con-
sequently, Ug is irreducible. Since Ug is the induced representation, if it is
reducible, L would also be reducible, but L is 1-dimensional.
Note: The Heisenberg group is a non abelian unimodular Lie group, so the Haar
measure on G is just the product measure dxdydz on R3 . Conditional expec-
tation becomes integrating out the variables correspond to the subgroup. For
example, given f (x, y, z) conditioning with respect to the subgroup (0, b, c)
amounts to integrating out the y, z variable and get a function f˜(x), where
¨
f˜(x) = f (x, y, z)dydz, i.e.,
Γ \G ' R.
7.3.1 ax + b group
" #
ab
a ∈ R+ , b ∈ R, g = (a, b) = .
01
Ug f (x) = eiax f (x + b)
could also write a = et , then
t
Ug f (x) = eie x f (x + b)
x f (x+b)
Ug(a,b) f (x) = eiae
d
[ , iex ] = iex
dx
[A, B] = B
or
x
U(et ,b) f = eite f (x + b)
" #
0b
01
7.4 Co-adjoint Orbits 195
1-d representation. Lb = eib . Induce indLG ' the Schrödinger representation.
7.4 Co-adjoint Orbits
It turns out that only a small family of representations are induced. The question is
how to detect whether a representation is induced. The whole theory is also under
the name of “Mackey machine” [Mac52, Mac88]. The notion of “machine” refers
to something that one can actually compute in practice. Two main examples are the
Heisenberg group and the ax + b group.
What is the mysteries parameter h that comes into the Schrödinger representa-
tion? It is a physical constant, but how to explain it in mathematical theory?
7.4.1 review of some Lie theory
Theorem 7.3 (Ado). Every Lie group is diffeomorphic to a matrix group.
The exponential function exp maps a neighborhood of 0 into a connected component

of G containing the identity element. For example, the Lie algebra of the Heisenberg
group is
 
0∗∗
0 0 ∗
 
000
All the Lie groups the we will ever encounter come from a quadratic form. Given a
quadratic form
ϕ : V ×V → C
there is an associated group that fixes ϕ, i.e. we consider elements g such that
ϕ(gx, gy) = ϕ(x, y)
and define G(ϕ) as the collection of these elements. G(ϕ) is clearly a group. Apply
the exponential map and the product rule,
d
t=0
ϕ(etX x, etX y) = 0 ⇐⇒ ϕ(Xx, y) + ϕ(x, Xy) = 0
dt
hence
X + X tr = 0
The determinant and trace are related so that
det(etX ) = et·trace(X)
thus det = 1 if and only if trace = 0. It is often stated in differential geometry that
the derivative of the determinant is equal to the trace.
Example 7.10. Rn , ϕ(x, y) = ∑ xi yi . The associated group is the orthogonal group
On .
There is a famous cute little trick to make On−1 into a subgroup of On . On−1 is not
normal in On . We may split the quadratic form into
n−1
∑ xi2 + 1
i=1
where 1 corresponds to the last coordinate in On . Then we may identity On−1 as a

subgroup of On " #
g0
g 7→
0I
where I is the identity operator.
Claim: On /On−1 ' Sn−1 . How to see this? Let u be the unit vector corresponding
to the last dimension, look for g that fixes u i.e. gu = u. Such g forms a subgroup of
On , and it is called isotropy group.
In = {g : gu = u} ' On−1
Notice that for all v ∈ Sn−1 , there exists g ∈ On such that gu = v. Hence
g 7→ gu
in onto Sn−1 . The kernel of this map is In ' On−1 , thus
On /On−1 ' Sn
Such spaces are called homogeneous spaces.

Example 7.11. visualize this with O3 and O2 .
7.4 Co-adjoint Orbits 197
Other examples of homogeneous spaces show up in number theory all the time. For
example, the Poincaré group G/discrete subgroup.
G, N ⊂ G normal subgroup. The map g·g−1 : G → G is an automorphism sending
identity to identity, hence if we differentiate it, we get a transformation in GL(g). i.e.
we get a family of maps Adg ∈ GL(g) indexed by elements in G. g 7→ Adg ∈ GL(g)
is a representation of G, hence if it is differentiated, we get a representation of g,
adg : g 7→ End(g) acting on the vector space g.
gng−1 ∈ N. ∀g, g · g−1 is a transformation from N to N, define Adg (n) = gng−1 .
Differentiate to get ad : n → n. n is a vector space, has a dual. Linear transformation
on vector space passes to the dual space.
ϕ ∗ (v∗ )(u) = v∗ (ϕ(u))

m
hΛ v , ui = hv∗ ,Λ ui .
∗ ∗
In order to get the transformation rules work out, have to pass to the adjoint or the
dual space.
Adg∗ : n∗ → n∗
the coadjoint representation of n.

Orbits of co-adjoint representation amounts precisely to equivalence classes of
irreducible representations.
Example 7.12. Heisenberg group G = {(a, b, c)} with

 
1ac
(a, b, c) =  0 1 b 
 
001
normal subgroup N = {(0, b, c)}

 
10c
(0, b, c) =  0 1 b 
 
001
with Lie algebra n = {(b, c)}

 
10c
(0, ξ , η) =  0 1 b 
 
001
Adg : n → n given by
gng−1 = (a, b, c)(0, y, x)(−a, −b, −c + ab)

= (a, b + y, c + x + ay)(−a, −b, −c + ab)
= (0, y, x + ay)
hence Adg : R2 → R2 " # " #

x x + ay
Adg : 7→ .
y y
The matrix of Adg is (before taking adjoint) is
" #
1a
Adg = .
01
The matrix for Adg∗ is

" #
10
Adg∗ = .
a1
We use [ξ , η]T for the dual n∗ ; and use [x, y]T for n. Then
" # " #
ξ ξ
Adg∗ : 7→
η aξ + η
What about the orbit? In the example of On /On−1 , the orbit is Sn−1 .
For ξ ∈ R\{0}, the orbit of Adg∗ is
" # " #
ξ ξ
7→
0 R
i.e. vertical lines with x-coordinate ξ . ξ = 0 amounts to fixed point, i.e. the orbit is
a fixed point.
The simplest orbit is when the orbit is a fixed point. i.e.
7.5 Gårding Space 199
" # " #
ξ ξ
Adg∗ : 7→ ∈ V∗
η η
where if we choose " # " #

ξ 0
=
η 1
it is a fixed point.
The other extreme is to take any ξ 6= 0, then
" # " #
ξ ξ
Adg∗ : 7→
0 R
i.e. get vertical lines indexed by the x-coordinate ξ . In this example, a cross section
is a subset of R2 that intersects each orbit at precisely one point. Every cross section
in this example is a Borel set in R2 .
We don’t always get measurable cross sections. An example is the construction
of non-measurable set as was given in Rudin’s book. Cross section is a Borel set
that intersects each coset at precisely one point.
Why does it give all the equivalent classes of irreducible representations? Since
we have a unitary representation Ln ∈ Rep(N,V ), Ln : V → V and by construction
of the induced representation Ug ∈ Rep(G, H ), N ⊂ G normal such that
Ug LnUg−1 = Lgng−1
i.e.
Lg ' Lgng−1
now pass to the Lie algebra and its dual
Ln → LA → LA∗ .
7.5 Gårding Space
Definition 7.1. Let U be a strongly continuous representation of a Lie group G, with

Lie algebra g, and let exp : g → G denote the exponential mapping from Lie theory.
Fro every ϕ ∈ Cc∞ (G), set
ˆ
U (ϕ) = ϕ (g) Ug dg
G
where dg is a left-invariant Haar measure on G; and set

n o
HGårding = U (ϕ) v ϕ ∈ Cc∞ (G) , v ∈ H .
Lemma 7.4. Fix X ∈ g, set
U (exp (tX)) v − v
dU (X) v = lim
t→0 t
then
HGårding ⊂
\
dom (dU (X))
X∈g
and
dU (X) U (ϕ) v = U Xϕ
e v,
for all ϕ ∈ Cc∞ (G), v ∈ H , where
e (g) = d

Xϕ t=0
ϕ (exp (−tX) g) , ∀g ∈ G.
dt
Proof. (Hint)
ˆ ˆ
ϕ (g) U (exp (tX)) U (g) dg = ϕ (exp (−tX) g) U (g) dg.
G G
We talked about how to detect whether a representation is induced. Given a group

G with a subgroup Γ let M := Γ \G. The map π : G → M is called a covering map,
which sends g to its equivalent class or the coset Γ g. M is given its projective topol-
ogy, so π is continuous. When G is compact, many things simplify. For example, if
G is compact, any irreducible representation is finite dimensional. But many groups
are not compact, only locally compact. For example, the groups ax + b, H3 , SLn .
Specialize to Lie groups. G and subgroup H have Lie algebras g and h respec-
tively.
g = {X : etX ∈ G, ∀t ∈ R}
Almost all Lie algebras we will encounter come from specifying a quadratic form
ϕ : G × G → C. ϕ is then uniquely determined by a Hermitian matrix A so that
ϕ(x, y) = xtr · Ay
Let G = G(ϕ) = {g : ϕ(gx, gy) = ϕ(x, y)}, then
d
t=0
ϕ(etX x, etX y) = 0
dt
and with an application of the product rule,
ϕ(Xx, y) + ϕ(x, Xy) = 0

(Xx)tr · Ay + xtr · AXy = 0
X tr A + AX = 0
hence
g = {X : X tr A + AX = 0}.
Let U ∈ Rep(G, H ), for X ∈ g, U(etX ) is a one parameter continuous group of

unitary operator, hence by Stone’s theorem, it must have the form
U(etX ) = eitHX (7.3)
for some selfadjoint operator HX (possibly unbounded). The RHS in (7.3) is given
by the Spectral Theorem. We often write
dU(X) := iHX
to indicate that dU(X) is the directional derivative along the direction X. Notice that
HX∗ = HX but
(iHX )∗ = −(iHX )
i.e. dU(X) is skew adjoint.

Example 7.13. G = {(a, b, c)} Heisenberg group. g = {X1 ∼ a, X2 ∼ b, X3 ∼ c}. Take
the Schrödinger representation Ug f (x) = eih(c+bx) f (x + a), f ∈ L2 (R).
• U(etX1 ) f (x) = f (x + t)
d d
t=0
U(etX1 ) f (x) = f (x)
dt dx
d
dU(X1 ) =
dx
• U(etX2 ) f (x) = eih(tx) f (x)
d
U(etX2 ) f (x) = ihx f (x)
t=0
dt
dU(X2 ) = ihx
• U(etX3 ) f (x) = eiht f (x)
d
t=0
U(etX3 ) f (x) = ih f (x)
dt
dU(X2 ) = ihI
Notice that dU(Xi ) are all skew adjoint.
d
[dU(X1 ), dU(X2 )] = [ , ihx]
dx
d
= ih[ , x]
dx
= ih
In case we want selfadjoint operators, replace dU(Xi ) by−idU(Xi ) and get
1 d
−idU(X1 ) =
i dx
−idU(X2 ) = hx
−idU(X1 ) = hI
1 d h
[ , hx] = .
i dx i
Below we answer the following question:
What is the space of functions that Ug acts on? L. Gårding /gor-ding/ (Swedish
mathematician) looked for one space that always works. It’s now called the Gårding
space.
Start with Cc (G), every ϕ ∈ Cc (G) can be approximated by the so called Gårding
functions, using the convolution argument. Define convolution as
ˆ
ϕ ? ψ(g) = ϕ(gh)ψ(h)dR h
ˆG
ϕ ? ψ(g) = ϕ(h)ψ(g−1 h)dL h
G
Take an approximation of identity ζ j (Fig. 7.1), so that

ϕ ? ζ j → ϕ, j → 0.
Ζj
e G
Fig. 7.1: Approximation of identity.
Define Gårding space as the span of the vectors in H , given by

ˆ
U(ϕ)v = ϕ(h)U(h)vdL h
where ϕ ∈ Cc (G), v ∈ H , or we say

ˆ
U(ϕ) := ϕ(h)U(h) dL h.
G
Since ϕ vanishes outside a compact set, and since U(h)v is continuous and bounded
in k·k, it follows that U(ϕ) is well-defined.
Every representation U of a Lie group G induces a representation (also denote
U) of the group algebra:
Lemma 7.5. U(ϕ1 ?ϕ2 ) = U(ϕ1 )U(ϕ2 ) (U is a representation of the group algebra)
Proof. Use Fubini,

ˆ ¨
ϕ1 ? ϕ2 (g)U(g)dg = ϕ1 (h)ϕ(h−1 g)U(g)dhdg
G
¨G×G
= ϕ1 (h)ϕ(g)U(hg)dhdg (dg is r-Haarg 7→ hg)
¨G×G
= ϕ1 (h)ϕ(g)U(h)U(g)dhdg
G×G
ˆ ˆ
= ϕ1 (h)U(h)dh ϕ2 (g)U(g)dg
G G
Choose ϕ to be an approximation of identity, then

ˆ
ϕ(g)U(g)vdg → U(e)v = v
G
i.e. any vector v ∈ H can be approximated by functions in the Gårding space. It

follows that
{U(ϕ)v}
is dense in H .
Lemma 7.6. U(ϕ) can be differentiated, in the sense that
dU(X)U(ϕ)v = U(X̃ϕ)v
where we use X̃ to denote the vector field.
Proof. need to prove
1
(U(etX ) − I)U(ϕ)v = U(X̃ϕ)v.

lim
t→0 t
Let vϕ := U(ϕ)v, need to look at in general U(g)vϕ .

ˆ
U(g)vϕ = U(g) ϕ(h)U(h)vdh
ˆ G
= ϕ(h)U(gh)dh
ˆG
= 4(g)ϕ(g−1 h)U(h)dh
G
set g = etX .
Note: If assuming unimodular, 4 does not show up. Otherwise, 4 is some cor-
rection term which is also differentiable. X̃ acts on ϕ as X̃ϕ. X̃ is called the
derivative of the translation operator etX .
Exercise 7.1. Show that Schwartz space S is the Gårding space for the Schrödinger
representation.
7.6 Decomposition of representation
We study some examples of duality.

7.6 Decomposition of representation 205
• G = T, Ĝ = Z
χn (z) = zn
χn (zw) = zn wn = χn (z)χn (w)
• G = R, Ĝ = R
χt (x) = eitx
• G = Z/nZ ' {0, 1, · · · , n − 1}. Ĝ = G.

This is another example where Ĝ = G.
Let ζ = ei2π/n be the primitive nth -root of unity. k ∈ Zn , l = {0, 1, . . . , n − 1}
2πkl
χl (k) = ei n
If G is a locally compact abelian group, Ĝ is the set of 1-dimensional representa-

tions.
Ĝ = {χ : g 7→ χ(g) ∈ T, χ(gh) = χ(g)χ(h), assumed continuous}.
Ĝ is also a group, with group operation defined by (χ1 χ2 )(g) := χ1 (g)χ2 (g). Ĝ is
called the group characters.
Theorem 7.4 (Pontryagin). If G is a locally compact abelian group, then G ' Ĝˆ
ˆ where “'” means “natural iso-
(isomorphism between G and the double dual Ĝ,)
morphism.”
Note: This result first appears in 1930s in the annals of math, when John von Neu-
mann was the editor of the journal at the time. The original paper was hand
written. von Neumann rewrote it, since then the theorem became very popu-
lar, see [Rud90].
There are many groups that are not abelian. We want to study the duality question
in general. Examples:
• compact group
• finite group (abelian, or not)
• H3 locally compact, nonabelian, unimodular
• ax+b locally compact, nonabelian, non-unimodular
If G is not abelian, Ĝ is not a group. We would like to decompose Ĝ into irreducible

representations. The big names in this development are Krein, Peter-Weyl, Weil,
Segal. See [AD86, ARR13, BR79, Emc00, JÓ00, KL14, KR97b, Rud73, Rud90,
Seg50, Sto51b].
Let G be a group (may not be abelian). The right regular representation is defined
as
Rg f (·) = f (·g), (translation on the right) .
Then Rg is a unitary operator acting on L2 (µR ), where µR is the right invariant Haar
measure.
Theorem 7.5 (Krein, Weil, Segal). Let G be locally compact unimodular (abelian
or not). Then the right regular representation decomposes into a direct integral of
irreducible representations
ˆ ⊕
Rg = ”irrep” dµ
Ĝ
where µ is called the Plancherel measure. See [Sti59, Seg50].
Example 7.14. G = T , Ĝ = Z. Irreducible representations {ein(·) }n ∼ Z
(Uy f )(x) = f (x + y)
= ∑ fˆ(n)χn (x + y)
n
= ∑ fˆ(n)ei2πn(x+y)
n
(Uy f )(0) = f (y) = ∑ fˆ(n)ei2πny

n
The Plancherel measure in this case is the counting measure.
Example 7.15. G = R, Ĝ = R. Irreducible representations {eit(·) }t∈R ∼ R.
(Uy f )(x) = f (x + y)
ˆ
= fˆ(t)χt (x + y)dt
ˆ R
= fˆ(t)eit(x+y) dt
R
ˆ
(Uy f )(0) = f (y) = fˆ(t)eity dt
R
7.6 Decomposition of representation 207
where the Plancherel measure is the Lebesgue measure on R.
As can be seen that Fourier series and Fourier integrals are special cases of the de-
composition of the right regular representation Rg of a unimodular locally compact
´⊕
group. =⇒ k f k = k fˆk. This is a result that was done 30 years earlier before
the non abelian case. Classical function theory studies other types of convergence,
pointwise, uniform, etc.
Example 7.16. G = H3 . G is unimodular, non abelian. Ĝ is not a group.

Irreducible representations: R\{0} Schrödinger representation, {0} 1-d trivial
representation
Decomposition: ˆ ⊕
h
Rg = Uirrep hdh
R\{0}
For all F ∈ L2 (G),

ˆ ⊕
(Ug f )(e) = U h f hdh, U h irrep
F(g) = (Rg F)(e)

ˆ ⊕
= eih(c+bx) f (x + a) hdh
ˆ
R\{0}
F̂(h) = (Ugh F)dg

G
Plancherel measure: hdh and the point measure δ0 at zero.
Example 7.17. G ax + b group, non abelian. Ĝ not a group. 3 irreducible represen-

tations: +, −, 0 but G is not unimodular.
The + representation is supported on R+ , the − representation on R− , and the 0
representation is the trivial one-dimensional representation.
The duality question may also be asked for discrete subgroups. This leads to re-
markable applications in automorphic functions, automorphic forms, p-adic num-
bers, compact Riemann surface, hyperbolic geometry, etc.
Example 7.18. Cyclic group of order n. G = Z/nZ ' {0, 1, · · · , n − 1}. Ĝ = G. This
is another example where the dual group is identical to the group itself. Let ζ =
ei2π/n be the primitive nth -root of unity. k ∈ Zn , l = {0, 1, . . . , n − 1}
2πkl
χl (k) = ei n
In this case, Segal’s theorem gives finite Fourier transform. U : l 2 (Z) → l 2 (Ẑ) where
1
U f (l) = √ ∑ ζ kl f (k)
N k
7.7 Summary of Induced Representations, the Example of d/dx
We study decomposition of group representations. Two cases: abelian and non

abelian. The non abelian case may be induced from the abelian ones.
non abelian
• semi product G = HN often N is normal.

• G simple. G does not have normal subgroups, i.e., the Lie algebra does not
have any ideals.
Exercise 7.2. (i) Find the normal subgroups in the Heisenberg group. (ii) Find the
normal subgroups in the ax + b group.
Example 7.19. SL2 (R) (non compact)

!
ab
, ad − bc = 1
cd
with Lie algebra

sl2 (R) = {X : tr(X) = 0}.
Note that sl2 is generated by

! ! !
01 0 −1 1 0
.
10 1 0 0 −1
! !
0 −1 cost − sint
In particular, generates the one-parameter group 'T
1 0 sint cost
whose dual group is Z, where
χn (g(t)) = g(t)n = eitn .

7.7 Summary of Induced Representations, the Example of d/dx 209
May use this to induce a representation of G. This is called principle series. Need to
do something else to get all irreducible representations.
A theorem by Iwasawa states that simple matrix group (Lie group) can be de-
composed into
G = KAN
where K is compact, A is abelian and N is nilpotent. For example, in the SL2 case,
! ! !
cost − sint es 0 1u
SL2 (R) = .
sint cost 0 e−s 01
The simple groups do not have normal subgroups. The representations are much
more difficult.
7.7.1 Induced representation
Suppose from now on that G has a normal abelian subgroup N C G, and G = H n N

The N ' Rd and N ∗ ' (Rd )∗ = Rd . In this case
χt (ν) = eitν
for ν ∈ N and t ∈ N̂ = N ∗ . Notice that χt is a 1-d irreducible representation on C.

Let Ht be the space of functions f : G → C so that
f (νg) = χt (ν) f (g).
On Ht , define inner product so that

ˆ ˆ
k f k2Ht := | f (g)|2 = k f (g)k2 dm
G G/N
where dm is the invariant measure on N\G ' H.

Define Ut = indNG (χt ) ∈ Rep(G, Ht ). Define Ut (g) f (x) = f (xg), for f ∈ Ht . No-
tice that the representation space of χt is C, 1-d Hilbert space; however, the represen-
tation space of Ut is Ht which is infinite dimensional. Ut is a family of irreducible
representations indexed by t ∈ N ' N̂ ' Rd .
Note: Another way to recognize induced representations is to see these functions

are defined on H, not really on G.
Define the unitary transformation W : Ht → L2 (H). Notice that H ' N\G
is a group, and it has an invariant Haar measure. By uniqueness on the Haar
measure, this has to be dm. It would be nice to cook up the same space L2 (H)
so that all induced representations indexed by t act on it. In other words, this
Hilbert space L2 (H) does not depend on t. Wt is defined as
W Ft (h) = Ft (h).
So what does the induced representation look like in L2 (H) then? Recall by
definition that
Ut (g) := W indχGt (g) W ∗
and the following diagram commutes.
indχGt
Ht / Ht
W W

L2 (H)
Ut
/ L2 (H)
Let f ∈ L2 (H).

Ut (g) f (h) = W indχGt (g) W ∗ f (h)

= indχGt (g)W ∗ f (h)
= (W ∗ f ) (hg) .
Since G = H n N, g is uniquely decomposed into g = gN gH . Hence hg =

hgN gH = gN g−1
N hgN gH = gN h̃gH and
Ut (g) f (h) = (W ∗ f )(hg)

= (W ∗ f )(gN h̃gH )
= χt (gN )(W ∗ f )(h̃gH )
= χt (gN )(W ∗ f )(g−1
N hgN gH )
This last formula is called the Mackey machine [Mac52, Mac88].

7.7 Summary of Induced Representations, the Example of d/dx 211
The Mackey machine does not cover many important symmetry groups in
physics. Actually most of these are simple groups. However it can still be
applied. For example, in special relativity theory, we have the Poincaré group
L n R4 where R4 is the normal subgroup. The baby version of this is when
L = SL2 (R). V. Bargman formulated this baby version. Wigner pioneered
the Mackey machine, long before Mackey was around.
Once we get unitary representations, differentiate it and get selfadjoint alge-
bra of operators (possibly unbounded). These are the observables in quantum
mechanics.
Example 7.20. Z ⊂ R, Ẑ = T . χt ∈ T , χt (n) = eitn . Let Ht be the space of functions

f : R → C so that
f (n + x) = χt (n) f (x) = eint f (x).
Define inner product on Ht so that

ˆ 1
2
k f kHt := | f (x)|2 dx.
0
Define indχRt (y) f (x) = f (x + y). Claim that Ht ' L2 [0, 1]. The unitary transforma-
tion is given by W : Ht → L2 [0, 1]
(W Ft )(x) = Ft (x).
Let’s see what indχRt (y) looks like on L2 [0, 1]. For any f ∈ L2 [0, 1],

W indχGt (y) W ∗ f (x) = indχGt (y)W ∗ f (x)
= (W ∗ f ) (x + y)
Since y ∈ R is uniquely decomposed as y = n + x0 for some x0 ∈ [0, 1), therefore

W indχGt (y) W ∗ f (x) = (W ∗ f )(x + y)
= (W ∗ f )(x + n + x0 )
= (W ∗ f )(n + (−n + x + n) + x0 )
= χt (n)(W ∗ f )((−n + x + n) + x0 )
= χt (n)(W ∗ f )(x + x0 )
= eitn (W ∗ f )(x + x0 )
Note: Are there any functions in Ht ? Yes, for example, f (x) = eitx . If f ∈ Ht , | f |
is 1-periodic. Therefore f is really a function defined on Z\R ' [0, 1]. Such
a function has the form
f (x) = (∑ cn ei2πnx )eitx = ∑ cn ei(2πn+t)x .
Any 1-periodic function g satisfies the boundary condition g(0) = g(1). f ∈

Ht has a modified boundary condition where f (1) = eit f (0).
7.8 Connections to Nelson’s Spectral Theory
In Nelson’s notes [Nel69], a normal representation has the form (counting multi-
plicity)
⊕
ρ = ∑ nπ Hn
, Hn ⊥ Hm
where
nπ = π ⊕ · · · ⊕ π (n times)
is a representation acting on the Hilbert space
⊕
∑ K = lZ2 n ⊗ K.
In matrix form, this is a diagonal matrix with π repeated on the diagonal n times. n
could be 1, 2, . . . , ∞. We apply this to group representations.
Locally compact group can be divided into the following types.
• abelian
• non-abelian: unimodular, non-unimodular
• non-abelian: Mackey machine, semidirect product e.g. H3 , ax + b; simple
group SL2 (R). Even it’s called simple, ironically its representation is much
more difficult than the semidirect product case.
We want to apply these to group representations.
Spectral theorem says that given a normal operator A, we may define f (A) for
quite a large class of functions, actually all measurable functions. One way to define
f (A) is to use the multiplication version of the spectral theorem, and let
f (A) = F f (Â)F −1 .
7.8 Connections to Nelson’s Spectral Theory 213
The other way is to use the projection-valued measure version of the spectral theo-
rem, write
ˆ
A= λ P(dλ )
ˆ
f (A) = f (λ )P(dλ ).
The effect is ρ is a representation of the abelian algebra of measurable functions

onto operators action on some Hilbert space.
ρ : f 7→ ρ( f ) = f (A)
ρ( f g) = ρ( f )ρ(g)
To imitate Fourier transform, let’s call fˆ := ρ( f ). Notice that fˆ is the multiplication

operator.
Example 7.21. G = (R, +), group algebra L1 (R). Define Fourier transform
ˆ
fˆ(t) = f (x)e−itx dx.
{eitx }t is a family of 1-dimensional irreducible representation of (R, +).
Example 7.22. Fix t, H = C, ρ(·) = eit(·) ∈ Rep(G, H ). From the group represen-
tation ρ, we get a group algebra representation ρ̃ ∈ Rep(L1 (R), H ) defined by
ˆ ˆ
ρ̃( f ) = f (x)ρ(x)dx = f (x)eitx dx
It follows that
fˆ(ρ) := ρ̃( f )
? g = fˆĝ
? g = fd
fd
i.e. Fourier transform of f ∈ L1 (R) is a representation of the group algebra L1 (R) on

to the 1-dimensional Hilbert space C. The range of Fourier transform in this case is
1-d abelian algebra of multiplication operators, multiplication by complex numbers.
Example 7.23. H = L2 (R), ρ ∈ Rep(G, H ) so that

ρ(y) f (x) := f (x + y)
i.e. ρ is the right regular representation. The representation space H in this

case is infinite dimensional. From ρ, we get a group algebra representation ρ̃ ∈
Rep(L1 (R), H ) where ˆ
ρ̃( f ) = f (y)ρ(y)dy.
Define
fˆ(ρ) := ρ̂( f )
then fˆ(ρ) is an operator acting on H .

ˆ
fˆ(ρ)g = ρ̃( f )g = f (y)ρ(y)g(·)dy
ˆ
= f (y)(Ry g)(·)dy
ˆ
= f (y)g(· + y)dy.
If we have used the left regular representation, instead of the right, then
ˆ
fˆ(ρ)g = ρ̃( f )g = f (y)ρ(y)g(·)dy
ˆ
= f (y)(Ly g)(·)dy
ˆ
= f (y)g(· − y)dy.
Hence fˆ(ρ) is the left or right convolution operator.
Back to the general case. Given a locally compact group G, form the group algebra
L1 (G), and define the left and right convolutions as
ˆ ˆ
(ϕ ? ψ)(x) = ϕ(g)ψ(g−1 x)dL g = ϕ(g)(Lg ψ)dL g
ˆ ˆ
(ϕ ? ψ)(x) = ϕ(xg)ψ(g)dR g = (Rg ϕ)ψ(g)dR g
Let ρ(g) ∈ Rep(G, H ), define ρ̃ ∈ Rep(L1 (G), H ) given by

ˆ
ρ̃(ψ) := ψ(g)ρ(g)dg
G
and write
ψ̂(ρ) := ρ̃(ψ).
ψ̂ is an analog of Fourier transform. If ρ is irreducible, the operators ψ̂ forms an

abelian algebra. In general, the range of this generalized Fourier transform gives rise
to a non abelian algebra of operators.
For example, if ρ(g) = Rg and H = L2 (G, dR ), then
ˆ ˆ
ρ̃(ψ) = ψ(g)ρ(g)dg = ψ(g)Rg dg
G G
and
ˆ ˆ
ρ̃(ψ)ϕ = ψ(g)ρ(g)ϕdg = ψ(g)(Rg ϕ)dg
ˆG G
= ψ(g)ϕ(xg)dg
G
= (ϕ ? ψ)(x)
Example 7.24. G = H3 ∼ R3 . Ĝ = {R\{0}} ∪ {0}. 0 ∈ Ĝ corresponds to the trivial

representation, i.e. g 7→ Id for all g ∈ G.
ρh : G → L2 (R)
ρh (g) f (x) = eih(c+bx) f (x + a) ' indHG (χh )
where H is the normal subgroup {b, c}. It is not so nice to work with indHG (χh )
directly, so instead, we work with the equivalent representations, i.e. Schrödinger
representation. See Folland’s book on abstract harmonic analysis.
ˆ
ψ̂(h) = ψ(g)ρh (g)dg
G
Notice that ψ̂(h) is an operator acting on L2 (R). Specifically,

ˆ
ψ̂(h) = ψ(g)ρh (g)dg
˚G
= ψ(a, b, c)eih(c+bx) f (x + a)dadbdc

¨ ˆ
= ψ(a, b, c)e dc f (x + a)eihbx dadb
ihc
¨
= ψ̂(a, b, h) f (x + a)eihbx dadb
ˆ ˆ
ihbx
= ψ̂(a, b, h)e db f (x + a)da
ˆ
= ψ̂(a, hx, h) f (x + a)da
= (ψ̂(·, h·, h) ? f ) (x)
Here the ψ̂ on the right hand side in the Fourier transform of ψ in the usual sense.
Therefore the operator ψ̂(h) is the one so that
L2 (R) 3 f 7→ (ψ̂(·, h·, h) ? f ) (x).
If ψ ∈ L1 (G), ψ̂ is not of trace class. But if ψ ∈ L1 ∩ L2 , then ψ̂ is of trace class.

ˆ ⊕ ˆ ˆ
tr (ψ̂ ∗ (h)ψ̂(h)) dµ = |ψ|2 dg = ψ̄ψdg
R\{0}
where µ is the Plancherel measure.

If the group G is non unimoduler, the direct integral is lost (not orthogonal).
These are related to coherent states from physics, which is about decomposing
Hilbert into non orthogonal pieces.
Important observables in QM come in pairs (dual pairs). For example, position

- momentum; energy - time etc. The Schwartz space S(R) has the property that
[ = S(R). We look at the analog of the Schwartz space. h 7→ ψ̂(h) should de-
S(R)
crease faster than any polynomials.
Take ψ ∈ L1 (G), Xi in the Lie algebra, form 4 = ∑ Xi2 . Require that
4n ψ ∈ L1 (G), ψ ∈ C∞ (G).
For 4n , see what happens in the transformed domain. Notice that
d
t=0
(RetX ψ) = X̃ψ
dt
where X 7→ X̃ represents the direction vector X as a vector field.

Let G be any Lie group. ϕ ∈ Cc∞ (G), ρ ∈ Rep(G, H ).
ˆ
dρ(X)v = (X̃ϕ)(g)ρ(g)vdg
where ˆ
v= ϕ(g)ρ(g)wdg = ρ(ϕ)w. generalized convolution
If ρ = R, the n ˆ
v= ϕ(g)R(g)
X̃(ϕ ? w) = (Xϕ) ? w.
Example 7.25. H3
∂
a→
∂a
∂
b 7→
∂b
∂
c 7→
∂c
get standard Laplace operator. {ρh (ϕ)w} ⊂ L2 (R) . ” = ” due to Dixmier. {ρh (ϕ)w}
is the Schwartz space.
2 2
d d
+ (ihx)2 + (ih)2 = − (hx)2 − h2
dx dx
Notice that 2
d
− + (hx)2 + h2
dx
is the Harmonic oscillator. Spectrum = hZ+ .
we suggest: [JÓ00, Mac52, Mac85, Mac92].
Chapter 8
The Kadison-Singer Problem
Le plus court chemin entre deux vérités dans le domaine réel

passe par le domaine complexe. — Jacques Hadamard
Born wanted a theory which would generalize these matrices or

grids of numbers into something with a continuity comparable
to that of the continuous part of the spectrum. The job was a
highly technical one, and he counted on me for aid.... I had the
generalization of matrices already at hand in the form of what is
known as operators. Born had a good many qualms about the
soundness of my method and kept wondering if Hilbert would
approve of my mathematics. Hilbert did, in fact, approve of it,
and operators have since remained an essential part of quantum
theory. — Norbert Wiener
The Kadison-Singer problem (KS) lies at the root of how questions from quantum
physics take shape in the language of functional analysis, and algebras of operators.
A brief sketch is included below, summarizing some recent advances (in fact the
KS-problem was recently solved.) It is of special interest as it is known that the
solution to KS at the same time answers a host of other questions; these with appli-
cations to engineering, especially to signal processing. The notion from functional
analysis here is “frame.” A frame of vectors in Hilbert space generalizes the notion
of orthonormal basis in Hilbert space.
The Kadison-Singer problem (KS) comes from functional analysis, but it was
resolved (only recently) with tools from areas of mathematics quite disparate from
functional analysis. More importantly, the solution to KS turned out to have impor-
tant implications for a host of applied fields from engineering.1
This reversal of the usual roles seem intriguing for a number of reasons: While
the applications considered so far, involve problems which in one way or the
other, derive from outside functional analysis itself, e.g., from physics, from signal-
processing, or from anyone of a number of areas of analysis, PDE, probability, statis-
tics, dynamics, ergodic theory, prediction theory etc.; the Kadison-Singer problem is
different. It comes directly from the foundational framework of functional analysis;
more specifically from the axiomatic formulation of C∗ -algebras. Then of course,
1 Atiyah and Singer shared the Abel prize of 2004.
219
220 8 The Kadison-Singer Problem
C∗ -algebras are a by-product of rigorous a formulation of quantum theory, as pro-

posed by P.A.M. Dirac.2
From quantum theory, we have such notions as state, observable, and measure-
ment. But within the framework of C∗ -algebras, each of these same terms, “state”,
“observable”, and “measurement” also has a purely mathematical definition, see
section 2.1 in chapter 2. Indeed C∗ -algebra theory was motivated in part by the de-
sire to make precise fundamental and conceptual questions in quantum theory, e.g.,
the uncertainty principle, measurement, determinacy, hidden variables, to mention
a few (see for example [Emc00]). The interplay between the two sides has been
extraordinarily fruitful since the birth of quantum mechanics in the 1920ties.
Now consider the following: (i) the Hilbert space H = l 2 (= l 2 (N)), all square
summable sequences, (ii) the C∗ -algebra B(l 2 ) of all bounded operators on l 2 , and
finally (iii) the sub-algebra A of B(l 2 ) consisting of all diagonal operators, so an
isomorphic copy of l ∞ . The Kadison-Singer problem (KS) is simply this: Does every
pure state of A have a unique pure-state extension to B(l 2 )? We remark that exis-
tence (of a pure-state extension) follows from the main theorems from functional
analysis of Krein and Krein-Milman, but the uniqueness is difficult. The difficulty
lies in the fact that it’s hard to find all states on l ∞ , i.e., the dual of l ∞ . The problem
was settled in the affirmatively only a year ago, after being open for 50 years.
Lemma 8.1. Pure states on B(H ) are unit vectors (in fact, the equivalent class of
unit vectors; equivalently, pure states sit inside the projective vector space. In Cn+1 ,
this is CPn .) Specifically, let u ∈ H , kuk = 1, then
B(H ) 3 A 7−→ wu (A) = hu, Aui
is a pure state. All pure states on B(H ) are of this form.
Remark 8.1. Since l ∞ is an abelian algebra Banach ∗-algebra, by Gelfand’s theorem,

l ∞ ' C(X) where X is a compact Hausdorff space. Indeed, X = β N, – the Stone-
Cech compactification of N. Points in β N are called ultrafilters. Pure states on l ∞
correspond to pure states on C(β N), i.e., Dirac measures on β N.
Let s be a pure state on l ∞ . Using Hahn-Banach theorem one may extend s, as a
linear functional, from l ∞ to s̃ on the Banach space B(H ). However, Hahn-Banach
2P.A.M. Dirac gave a lecture at Columbia University in the late 1950’s, in which he claimed
without proof that pure states on the algebra of diagonal operators (' l ∞ ) extends uniquely on
B (l 2 ). Two students Kadison and Singer sitting in the audience were skeptical about whether
Dirac knew what it meant to be an extension. They later formulated the conjecture in a joint paper
8 The Kadison-Singer Problem 221
theorem doesn’t guarantee the extension remains a pure state. Let E(s) be the set
of all states on B(H ) which extend s. E(s) is non-empty, compact and convex in
the weak-∗ topology. By Krein-Milman’s theorem, E(s) = closure(Extreme Points).
Any extreme point will then be a pure state extension of s; but which one to choose?
It’s the uniqueness part that is the famous KS problem.
Physics Mathematics
H ∗ = H, v ∈ H , kvk = 1 (state).
Spectral theorem: H ∼ PH (·)
projection-valued measure.
Measurement:
Prob (H ∈ (a, b)) =
kPH (a, b) vk2
Fig. 8.1: Observable, state, measurement. Left column: An idealized physics exper-
iment. Right: the mathematical counterpart, a selfadjoint operator H, its associated
projection-valued measure PH , and a norm-one vector v in Hilbert space.
Exercise 8.1. Extend + on N to a “+” on β N (the Stone–Čech compactification).

Hint:

(1) For subsets A ⊂ N, and n ∈ N, set A − n := k ∈ N k + n ∈ A .
(2) Let F and G be ultra-filters on N, and set

F + G := A ⊂ N {n ∈ N; A − n ∈ F} ∈ G . (8.1)
(3) Show that F + G is an ultra-filter.

(4) Show that the “addition” operation “+” in (8.1) is an operation on β N, i.e.,
β N × β N −→ β N which is associative, but not commutative, i.e., F + G 6= G + F
may happen.
(5) Fix F ∈ β N, and show that
β N 3 G 7−→ F + G ∈ β N
is continuous, where F + G is defined in (8.1).

Ultra-filters define pure states of l ∞ as follows: If (xn )n∈N ∈ l ∞ , and if F ∈ β N,

i.e., is an ultra-filter, then there is a well-defined limit
lim xn = ϕF (x) ;
F
and this defines ϕF as a state on l ∞ .

A related use of ultra-filters yield the famous Dixmier-trace. For this we need
ultra-filters ω on N with the following properties:
(i) xn ≥ 0 =⇒ limω xn ≥ 0.
(ii) If xn is convergent with limit x, then limω xn = x.
(iii) For n ∈ N, set

σN (x) = x1 · · · x1 , x2 · · · x2 , x3 · · · x3 , · · ·
| {z } | {z } | {z }
N times N times N times
then limω (xn ) = limω (σN (x)).

√
Let A be a compact operator, and assume the eigenvalues λk of |A| = A∗ A as
λ1 ≥ λ2 ≥ · · · , λk = λk (A)
and set
n
1
trDix,ω (A) = lim ∑ λk (A) . (8.2)
ω log (n + 1) k=1
Definition 8.1. We say that A has finite Dixmier trace if the limit in (8.2) is finite.
Exercise 8.2. Show that (8.2) is well-defined and that:

(i) A 7−→ trDix,ω (A) is linear, and positive.
(ii) trDix,ω (AB) = trDix,ω (BA) holds if B is bounded.
(iii) If ∑k λk (A) < ∞, then trDix,ω (A) = 0.
8.1 Frames in Hilbert Space
The proof of the KS-problem involves systems of vectors in Hilbert space called
frames. For details we refer to [Cas13].
Below a sketch. Let H be a separable Hilbert space, and let {uk }k∈N be an ONB,
then we have the following unique representation
8.1 Frames in Hilbert Space 223
w= ∑ huk , wiH uk (8.3)

k∈N
valid for all w ∈ H . Moreover,
kwk2H = ∑ |huk , wiH |2 , (8.4)

k∈N
the Parseval-formula.
Definition 8.2. A system {vk }k∈N in H is called a frame if there are constants A, B
s.t. 0 < A ≤ B < ∞, and
A kwk2H ≤ ∑ |hvk , wiH |2 ≤ B kwk2H (8.5)

k∈N
holds for all w ∈ H .
Note that (8.5) generalizes (8.4). Below we show that, for frames, there is also a
natural extension of (8.3).
Proposition 8.1. Let {vk }k∈N be a frame in H ; then there is a dual system v∗k

k∈N
⊂
H such that the following representation holds:
w= ∑ hv∗k , wiH vk (8.6)

k∈N
for all w ∈ H ; absolute convergence.
Proof. Define the following operator T : H → l 2 (N) by Tw = (hvk , wiH )k∈N , and
show that the adjoint T ∗ : l 2 (N) → H satisfies T ∗ ((xk )) = ∑k∈N xk vk . Hence
T ∗ Tw = ∑ hvk , wiH vk . (8.7)

k∈N
It follows that T ∗ T has a bounded inverse, in fact, A IH ≤ T ∗ T ≤ B IH in the

1
order of selfadjoint operators. As a result, (T ∗ T )−1 and (T ∗ T )− 2 are well-defined
bounded operators.
Substitute (T ∗ T )−1 into (8.7) yields:
D E
w= ∑ vk , (T ∗ T )−1 w vk
H
k∈N
D E
= ∑ (T ∗ T )−1 vk , w vk
H
k∈N
which is the desired (8.6) with v∗k := (T ∗ T )−1 vk .

−2 1
Corollary 8.1. Let {vk } be as in Proposition 8.1, and set v∗∗ ∗
k := (T T ) vk , k ∈ N;
then
w= ∑ hv∗∗ ∗∗
k , wiH vk
k∈N
holds for all w ∈ H ; absolute convergence.
Remark 8.2. We saw in Chapter 3 (sect. 3.3) that, if {vk }k∈N is an ONB in some
fixed Hilbert space H , then
P (4) = ∑ |vk ih vk | , 4 ∈ B (R) , (8.8)

k∈4
is a projection valued measure (PVM) on R.

Suppose now that some {vk }k∈N in the expression (8.8) is only assumed to be a
frame, see Definition 8.2.
Exercise 8.3. Write down the modified list of properties for P (·) in (8.8) which
generalize the axioms of Definition 3.4 for PVMs.
we suggest: [Arv76, BR81, Cas13, Cas14, AW13, MSS13].
The most current paper concerning the solution to KS appears to be [MSS13]
by Marcus, Spielman, and Strivastava. Paper [Cas14] by P. Casazza explains the
problem and its implications.
Part IV
Extension of Operators
Chapter 9
Selfadjoint Extensions
It will interest mathematical circles that the mathematical

instruments created by the higher algebra play an essential part
in the rational formulation of the new quantum mechanics. Thus
the general proofs of the conservation theorems in Heisenberg’s
theory carried out by Born and Jordan are based on the use of
the theory of matrices, which go back to Cayley and were
developed by Hermite. It is to be hoped that a new era of mutual
stimulation of mechanics and mathematics has commenced. To
the physicist it will seem first deplorable that in atomic problems
we have apparently met with such a limitation of our usual
means of visualisation. This regret will, however, have to give
way to thankfulness that mathematics, in this field too, presents
us with the tools to prepare the way for further progress.
— Niels Bohr
Because of dictates from applications (especially quantum physics), below we stress

questions directly related to key-issues for unbounded linear operators: Some oper-
ator from physics may only be “formally selfadjoint” also called Hermitian; and in
such cases, one ask for selfadjoint extensions (if any).
The axioms of quantum physics (see e.g., [BM13, OH13, KS02, CRKS79,
ARR13, Fan10, Maa10, Par09] for relevant recent papers), are based on Hilbert
space, and selfadjoint operators.
A quantum mechanical observable is a Hermitian (selfadjoint) linear operator
mapping a Hilbert space, the space of states, into itself. The values obtained in a
physical measurement are in general described by a probability distribution; and the
distribution represents a suitable “average” (or “expectation”) in a measurement of
values of some quantum observable in a state of some prepared system. The states
are (up to phase) unit vectors in the Hilbert space, and a measurement corresponds to
a probability distribution (derived from a projection-valued spectral measure). The
particular probability distribution used depends on both the state and the selfadjoint
operator. The associated spectral type may be continuous (such as position and mo-
mentum; both unbounded) or discrete (such as spin); this depends on the physical
quantity being measured.
227
228 9 Selfadjoint Extensions
Since the spectral theorem serves as the central tool in quantum measurements,
we must be precise about the distinction between linear operators with dense do-
main which are only Hermitian (formally selfadjoint) as opposed to selfadjoint.
This distinction is accounted for by von Neumann’s theory of deficiency indices
[AG93, DS88c, HdSS12]1 .
9.1 Extensions of Hermitian Operators
In order to apply spectral theorem, one must work with self adjoint operators in-
cluding the unbounded ones. Some examples first.
In quantum mechanics, to understand energy levels of atoms and radiation, the
energy level comes from discrete packages. The interactions are given by Column
Law where
c jk
H = −4r +
kr j − rk k
and Laplacian has dimension 3 × #(electrons).
In Schrödinger’s wave mechanics, one needs to solve for ψ(r,t) from the equa-
tion
1 ∂
Hψ = ψ.
i ∂t
If we apply spectral theorem, then ψ(t) = eitH ψ(r,t = 0). This shows that motion in
quantum mechanics is governed by unitary operators. The two parts in Schrödinger
equation are separately selfadjoint, but justification of the sum being selfadjoint
wasn’t made rigorous until 1957 when Kato wrote the book on “perturbation theory”
[Kat95]. It is a summary of the sum of selfadjoint operators.
In Heisenberg’s matrix mechanics, he suggested that one should look at two
states and the transition probability between them, such that
hψ1 , Aψ2 i = hψ1 (t) , Aψ2 (t)i , ∀t.
1 Starting with [vN32a, vN32d, vN32c], J. von Neumann and M. Stone did pioneering work in the
1930s on spectral theory for unbounded operators in Hilbert space; much of it in private correspon-
dence. The first named author has from conversations with M. Stone, that the notions “deficiency-
index,” and “deficiency space” are due to them; suggested by MS to vN as means of translating
more classical notions of “boundary values” into rigorous tools in abstract Hilbert space: closed
subspaces, projections, and dimension count.
9.1 Extensions of Hermitian Operators 229
If ψ(t) = eitH ψ, then it works. In Heisenberg’s picture, one looks at evolution of the
observables e−itH AeitH . In Schrödinger’s picture, one looks at evolution of states.
The two point of views are equivalent.
Everything so far is based on application of the spectral theorem, which requires
the operators being selfadjoint in the first place.
von Neumann’s index theory gives a complete classification of extensions of sin-
gle Hermitian unbounded operators with dense domain in a given Hilbert space. The
theory may be adapted to Hermitian representations of ∗-algebras [Nel59a].
Let A be a densely defined Hermitian operator on a Hilbert space H , i.e. A ⊂ A∗ .
If B is any Hermitian extension of A, then
A ⊂ B ⊂ B∗ ⊂ A∗ . (9.1)
Since the adjoint operator A∗ is closed, i.e., G (A∗ ) is closed in H ⊕ H , it fol-

lows that G (A) ⊂ G (A∗ ) is a well-defined operator graph, i.e., A is closable and
G (A) = G A . Thus, there is no loss of generality to assume that A is closed and

only consider its closed extensions.

The containment (9.1) suggests a detailed analysis in D (A∗ )\D (A). Since D (A)
is dense in H , the usual structural analysis in H (orthogonal decomposition, etc.)
is not applicable. However, this structure is brought out naturally when D (A∗ ) is
identified with the operator graph G (A∗ ) in H ⊕ H . That is, D (A∗ ) is a Hilbert
space under its graph norm. With this identification, D(A) becomes a closed sub-
space in D(A∗ ), and
D (A∗ ) = D (A) ⊕ (D (A∗ ) D (A)) . (9.2)
The question of extending A amounts to a further decomposition
D (A∗ ) D (A) = S ⊕ K (9.3)
in such a way that
e = A∗
A e) , where (9.4)
D (A
D A
e = D (A) ⊕ S

(9.5)
e ⊃ A.
defines a (closed) Hermitian operator A
e in (9.4)-(9.5) is Hermitian iff the closed subspace S ⊂ D (A∗ ) is

The extension A
symmetric, in the sense that
hA∗ y, xi − hy, A∗ xi = 0, ∀x, y ∈ S. (9.6)
Lemma 9.1. Let S be a closed subspace in D (A∗ ), where D (A∗ ) is a Hilbert space
under the A∗ -norm. The following are equivalent.
1. hA∗ y, xi = hy, A∗ xi, for all x, y ∈ S.
2. hx, A∗ xi ∈ R, for all x ∈ S.
Proof. If (1) holds, setting x = y, we get hx, A∗ xi = hA∗ x, xi = hx, A∗ xi, which im-
plies that hx, A∗ xi is real-valued.
Conversely, assume (2) is true. Since the mappings
(x, y) 7→ hy, A∗ xi
(x, y) 7→ hA∗ y, xi
are both sesquilinear forms on S × S (linear in the second variable, and conjugate
linear in the first variable), we apply the polarization identity:
1 3 kD E
hy, A∗ xi = ∑ i x + i k
y, A ∗
x + ik
y
4 k=0
1 3 D E
hA∗ y, xi = ∑ ik A∗ x + ik y , x + ik y
4 k=0
for all x, y ∈ D (A∗ ). Now, since A is Hermitian, the RHSs of the above equations
are equal; therefore, hy, A∗ xi = hA∗ y, xi, which is part (2).
Eqs (9.4)-(9.5) and lemma 9.1 set up a bijection between (closed) Hermitian
extensions of A and (closed) symmetric subspaces in D (A∗ ) D (A). Moreover, by
lemma 9.1, condition (9.6) is equivalent to
hx, A∗ xi ∈ R, ∀x ∈ D (A) . (9.7)
Let ϕ ∈ D (A∗ ), s.t. A∗ ϕ = λ ϕ, ℑ {λ } 6= 0; then hϕ, A∗ ϕi = λ kϕk2 ∈/ R. By

lemma 9.1 and (9.7), ϕ ∈ / D(A), where A is any possible Hermitian extension of
e e
A. This observation is in fact ruling out the “wrong” eigenvalues of A.
e Indeed, theo-
rem 9.1 below shows that A is selfadjoint if and only if ALL the “wrong” eigenvalues
of A∗ are excluded. But first we need the following lemma.
Lemma 9.2. Let A be a Hermitian operator in H , then
k(A − λ ) xk2 = k(A − a) xk2 + |b|2 kxk2 , ∀λ = a + ib ∈ C. (9.8)
In particular,
k(A − λ ) xk2 ≥ |ℑ {λ }|2 kxk2 , ∀λ ∈ C. (9.9)
Proof. Write λ = a + ib, a, b ∈ R; then
k(A − λ ) xk2
= h(A − a) x − ibx, (A − a) x − ibxi
= k(A − a) xk2 + |b|2 kxk2 − i (h(A − a) x, xi − hx, (A − a) xi)
= k(A − a) xk2 + |b|2 kxk2
≥ |b|2 kxk2 ;
where h(A − a) x, xi − hx, (A − a) xi = 0, since A − a is Hermitian.
Corollary 9.1. Let A be a closed Hermitian operator acting in H . Fix λ ∈ C with

ℑ {λ } 6= 0, then ran (A − λ ) is a closed subspace in H . Consequently, we get the
following decomposition

H = ran (A − λ ) ⊕ ker A∗ − λ . (9.10)
Proof. Set B = A − λ ; then B is closed, and so is B−1 , i.e., the operator graphs G (B)
and G B−1 are closed in H ⊕ H . Therefore, ran (B) = dom(B−1 ) is closed in

k·kB−1 -norm. But by (9.9), B−1 is bounded on ran (B), thus the two norms k·k and
k·kB−1 are equivalent on ran (B). It follows that ran (B) is also closed in k·k-norm,
i.e., it is a closed subspace in H . The decomposition (9.10) follows from this.
Theorem 9.1. Let A be a densely defined, closed, Hermitian operator in a Hilbert

space H ; then the following are equivalent:
n o n o
∃λ , ℑ {λ } 6= 0, ker (A∗ − λ ) = ker A∗ − λ = 0 ⇐⇒ A = A∗ .
=⇒ By 9.1, the hypothesis in the theorem implies that ran (A − λ ) =

Proof.
ran A − λ = H . Let y ∈ D (A∗ ), then
D E
hy, (A − λ ) xi = A∗ − λ y, x , ∀x ∈ D (A) . (9.11)
Since ran (A − z) = H , ∃y0 ∈ D (A) s.t.

A∗ − λ y = A − λ y0 .
Hence, RHS of (9.11) is

D E
A − λ y0 , x = hy0 , (A − λ ) xi . (9.12)
Combining (9.11)-(9.12), we then get
hy − y0 , (A − λ ) xi = 0, ∀x ∈ D (A) .
Again, since ran (A − λ ) = H , the last equation above shows that y − y0 ⊥ H . In

particular, y − y0 ⊥ y − y0 , i.e.,
ky − y0 k2 = hy − y0 , y − y0 i = 0.
Therefore, y = y0 , and so y ∈ D (A). This shows that A∗ ⊂ A.

The other containment A ⊂ A∗ holds since A is assumed to be Hermitian. Thus,
we conclude that A = A∗ .
To capture all the “wrong” eigenvalues, we consider a family of closed subspace

in H , ker (A∗ − λ ), where ℑ {λ } 6= 0.
Theorem 9.2. If A is a closed Hermitian operator in H , then
dim (ker (A∗ − λ ))
is a constant function on ℑ {λ } > 0, and ℑ {λ } < 0.
Proof. Fix λ with ℑ {λ } > 0. For ℑ {λ } < 0, the argument is similar. We proceed to
verify that if η ∈ C, close enough to λ , then dim (ker (A∗ − η)) = dim (ker (A∗ − λ )).
The desired result then follows immediately.
Since A is closed, we have the following decomposition (by 9.1),

H = ran A − λ ⊕ ker (A∗ − λ ) . (9.13)
Now, pick x ∈ ker (A∗ − η), and suppose x ⊥ ker (A∗ − λ ); assuming kxk = 1. By
(9.13), ∃x0 ∈ D (A) s.t.
x = A − λ x0 . (9.14)
Then,
0 = h(A∗ − η) x, x0 i = hx, (A − η) x0 i
D E
= x, A − λ x0 − η − λ x0

= kxk2 − η − λ hx, x0 i
≥ kxk2 − η − λ kxk2 kx0 k2 (Cauchy-Schwarz)

= 1 − |η − λ | kx0 k2 (9.15)
Applying lemma 9.2 to (9.14), we also have

2
1 = kxk2 = A − λ x0 ≥ |ℑ {λ }|2 kx0 k2 ;
substitute this into (9.15), we see that
0 ≥ 1 − |η − λ | kx0 k2 ≥ 1 − |η − λ | |ℑ {λ }|−2
which would be a contradiction if η was close to λ .

It follows that the projection from ker (A∗ − η) to ker (A∗ − λ ) is injective. For
otherwise, ∃x ∈ ker (A∗ − η), x 6= 0, and x ⊥ ker (A∗ − λ ). This is impossible as
shown above. Thus,
dim (ker (A∗ − η)) ≤ dim (ker (A∗ − λ )) .
Similarly, we get the reversed inequality, and so dim (ker (A∗ − η)) = dim (ker (A∗ − λ )).

A complete characterization of Hermitian extensions of a given Hermitian oper-

ator is due to von Neumann. Theorem 9.2 suggests the following definition:
Definition 9.1. Let A be a densely defined, closed, Hermitian operator in H . The

closed subspaces
D± (A) = ker (A∗ ∓ i) (9.16)

= {ξ ∈ D (A∗ ) : A∗ ξ = ±i ξ }
are called the deficiency spaces of A, and dimD± (A) are called the deficiency in-
dices.
Theorem 9.3 (von Neumann). Let A be a densely defined closed Hermitian opera-
tor acting in H . Then
D (A∗ ) = D (A) ⊕ D+ (A) ⊕ D− (A) ; (9.17)
where D(A∗ ) is identified with its graphG (A∗ ), thus a Hilbert space under the graph
inner product; and the decomposition in (9.17) refers to this Hilbert space.
Proof. By assumption, A is closed, i.e., D (A), identified with G (A), is a closed

subspace in D (A∗ ).
Note that D± (A) = ker (A∗ ∓ i) are closed subspaces in H . Moreover,
kxk2A∗ = kxk2 + kA∗ xk2 = 2 kxk2 , ∀x ∈ D± (A) ;
and so D± (A), when identified with the graph of A∗ , are also closed sub-
D± (A∗ )
spaces in D (A∗ ).
Next, we verify the three subspaces on RHS of (9.17) are mutually orthogonal.
For all x ∈ D (A), and all x+ ∈ D+ (A) = ker (A∗ − i), we have
hx+ , xiA∗ = hx+ , xi + hA∗ x+ , A∗ xi

= hx+ , xi − i hx+ , Axi
= −i (hx+ , i xi + hx+ , Axi)
= −i hx+ , (A + i) xi = 0
where the last step follows from x+ ⊥ ran (A + i) in H , see (9.10). Thus, D (A) ⊥
D+ (A) in D (A∗ ). Similarly, D (A) ⊥ D− (A) in D (A∗ ).
Moreover, if x+ ∈ D+ (A) and x− ∈ D− (A), then
hx+ , x− iA∗ = hx+ , x− i + hA∗ x+ , A∗ x− i

= hx+ , x− i + hi x+ , −i x− i
= hx+ , x− i − hx+ , x− i = 0.
Hence D+ (A) ⊥ D− (A) in D (A∗ ).

Finally, we show RHS of (9.17) yields the entire Hilbert space D (A∗ ). For this,
let x ∈ D (A∗ ), and suppose (9.17) holds, say, x = x0 + x+ + x− , where x ∈ D (A),
x± ∈ D± (A); then
(A∗ + i) x = (A∗ + i) (x0 + x+ + x− )

= (A + i) x0 + 2i x+ . (9.18)
But, by the decomposition H = ran (A + i) ⊕ ker (A∗ − i), eq. (9.10), there exist x0
and x+ satisfying (9.18). It remains to set x− := x − x0 − x+ , and to check x− ∈
D− (A). Indeed, by (9.18), we see that
A∗ x − Ax0 − i x+ = −i x + i x0 + i x+ ; i.e.,
A∗ (x − x0 − x+ ) = −i (x − x0 − x+ )
and so x− ∈ D− (A). Therefore, we get the desired orthogonal decomposition in

(9.17).
Another argument: Let y ∈ D (A∗ ) s.t. y ⊥ D± (A) in D (A∗ ). Then, y ⊥ D+ (A)
in D (A∗ ) =⇒
0 = hy, x+ i + hA∗ y, A∗ x+ i
= hy, x+ i + hA∗ y, i x+ i
= i (hi y, x+ i + hA∗ y, x+ i)
= i h(A∗ + i) y, x+ i , ∀x+ ∈ D+ (A) = ker (A∗ − i)
and so ∃ x1 ∈ D (A), and

(A∗ + i) y = (A + i) x1 . (9.19)
On the other hand, y ⊥ D− (A) in D (A∗ ) =⇒
0 = hy, x− i + hA∗ y, A∗ x− i
= hy, x− i + hA∗ y, −i x− i
= −i (h−i y, x− i + hA∗ y, x− i)
= i h(A∗ − i) y, x− i , ∀x− ∈ D− (A) = ker (A∗ + i) ;
hence ∃ x2 ∈ D (A), and

(A∗ − i) y = (A − i) x2 . (9.20)
Subtracting (9.19)-(9.20) then gives
x1 + x2
y= ∈ D (A) .
2
Remark 9.1. More generally, there is a family of decompositions
D (A∗ ) = D (A) + ker (A∗ − z) + ker (A∗ − z) , ∀z ∈ C, ℑ {z} 6= 0. (9.21)
However, in the general case, we lose orthogonality.
Proof. Give z ∈ C, ℑ {z} 6= 0, suppose x ∈ D (A∗ ) can be written as
x = x0 + x+ + x− ;
where x0 ∈ D (A), x+ ∈ ker (A∗ − z), and x− ∈ ker (A∗ − z). Then
A∗ x = Ax0 + zx+ + zx−

zx = zx0 + zx+ + zx−
and
(A∗ − z) x = (A − z) x0 + (z − z) x+ . (9.22)
Now, we start with (9.22). By the decomposition H = ran (A − z) ⊕ ker (A∗ − z),
there exist unique x0 and x+ such that (9.22) holds. This defines x0 and x+ . Then,
set
x− := x − x0 − x+ ;
and it remains to check x− ∈ ker (A∗ − z). Indeed, by (9.22), we have
A∗ x − Ax0 − zx+ = zx − zx0 − zx+ , i.e.,
A∗ (x − x0 − x+ ) = z (x − x0 − x+ )
thus, x− ∈ ker (A∗ − z).
Remark 9.2. In the general decomposition (9.21), if f = x+x+ +x− , g = y+y+ +y−
where f , g ∈ D (A), x+ , y+ ∈ ker (A∗ − z), and x− , y− ∈ ker (A∗ − z); then
hg, A∗ f i − hA∗ g, f i
= hy + y+ + y− , A∗ (x + x+ + x− )i − hA∗ (y + y+ + y− ) , x + x+ + x− i
= hy + y+ + y− , Ax + zx+ + zx− i − hAy + zy+ + zy− , x + x+ + x− i
= hy, Ax + zx+ + zx− i − hAy, x + x+ + x− i +

| {z }
0
hy+ + y− , Axi − hzy+ + zy− , xi +
| {z }
0
hy+ + y− , zx+ + zx− i − hzy+ + zy− , x+ + x− i
= hy+ , zx+ i − hzy+ , x+ i + hy− , zx− i − hzy− , x− i
+ hy+ , zx− i + hy− , zx+ i − hzy+ , x− i − hzy− , x+ i
= (z − z) hy+ , x+ i + (z − z) hy− , x− i +
z hy+ , x− i + z hy− , x+ i − z hy+ , x− i − z hy− , x+ i
| {z }
0
= (z − z) (hy+ , x+ i − hy− , x− i) .
tor in H .
1. The (closed) Hermitian extensions of A are indexed by partial isometries

with initial space in D+ (A) and final space in D− (A).
2. U ⊃ A is
Given a partial isometry U as above, the Hermitian extension Af
determined as follows:
U (x + (1 +U) x+ ) = Ax + i (1 −U) x+ , where

Af

D AfU = {x + x+ +Ux+ : x ∈ D (A) , x+ ∈ D+ (A)}
(9.23)
Proof. By the discussion in (9.6) and (9.7), and lemma 9.1, it remains to character-
ize the closed symmetric subspaces S in D+ (A) ⊕ D− (A) (⊂ D (A∗ )). For this, let
x = x+ + x− , x± ∈ D± (A), then
hx, A∗ xi = hx+ + x− , A (x+ + x− )i

= hx+ + x− , i (x+ − x− )i

= i kx+ k2 − kx− k2 − 2iℑ {hx+ , x− i}

= i kx+ k2 − kx− k2 + 2ℑ {hx+ , x− i} . (9.24)
Thus,
hx, A∗ xi ∈ R, ∀x ∈ S ⇐⇒ S = {(x+ , x− ) : kx+ k = kx− k , x± ∈ D± (A)}

i.e., S is identified with the graph of a partial isometry, say U, with initial space in
D+ (A) and final space in D− (A).
Corollary 9.2. Let A be a densely defined, closed, Hermitian operator on H , and

set d± = dim (D± (A)); then
1. A is maximally Hermitian if and only if one of the deficiency indices is 0;

2. A has a selfadjoint extension if and only if d+ = d− 6= 0;
3. A is selfadjoint if and only if d+ = d− = 0.
Proof. Immediate from theorem 9.3 and theorem 9.4.
Example 9.1. d+ = d− = 1. Let e± be corresponding eigenvalues. e+ 7→ ze− is the

unitary operator sending one to the other eigenvalue. It is clear that |z| = 1. Hence
the self adjoint extension is indexed by U1 (C).
Example 9.2. d+ = d− = 2, get a family of extensions indexed by U2 (C).
Remark 9.3. M. Stone and von Neumann are the two pioneers who worked at the
same period. They were born at about the same time. Stone died at 1970’s and von
Neumann died in the 1950’s.
There is a simple criterion to test whether a Hermitian operator has equal deficiency
indices.
Definition 9.2. An operator J : H → H is called a conjugation if
• J is conjugate linear, i.e., J (cx) = cx, for all x ∈ H , and all c ∈ C,

• J 2 = 1, and
• hJx, Jyi = hy, xi, for all x, y ∈ H .
tor in H . Set d± = dim (D± (A)). Suppose AJ = JA, where J is a conjugation, then
d+ = d− . In particular, A has selfadjoint extensions.
Proof. Note that, by definition, we have hJx, yi = Jx, J 2 y = hJy, xi, for all x, y ∈
H.
We proceed to show that J commutes with A∗ . For this, let x ∈ D (A), y ∈ D (A∗ ),
then
hJA∗ y, xi = hJx, A∗ yi = hAJx, yi = hJAx, yi = hJy, Axi . (9.25)
9.2 Cayley Transform 239
It follows that x 7→ hJy, Axi is bounded, and Jy ∈ D (A∗ ). Thus, JD (A∗ ) ⊂ D (A∗ ).
Since J 2 = 1, D (A∗ ) = J 2 D (A∗ ) ⊂ JD (A∗ ); therefore, JD (A∗ ) = D (A∗ ). More-
over, (9.25) shows that JA∗ = A∗ J.
Now if x ∈ D+ (A), then
A∗ Jx = JA∗ x = J (ix) = −iJx
i.e., JD+ (A) ⊂ D− (A). Similarly, JD− (A) ⊂ D+ (A).

Using J 2 = 1 again, D− (A) = J 2 D− (A) ⊂ JD+ (A); and we conclude that
JD+ (A) = D− (A).
Since the restriction of J to D+ (A) preserves orthonormal basis, we then get
dim (D+ (A)) = dim (D− (A)).

9.2 Cayley Transform
There is an equivalent characterization of Hermitian extensions, taking place en-

tirely in H and without the identification of D (A∗ ) ' G (A∗ ), where G (A∗ ) is seen
as a Hilbert space under its graph inner product. This is the result of the following
observation.
Lemma 9.3. Let A be a Hermitian operator acting in H ; then
k(A ± i) xk2 = kxk2 + kAxk2 , ∀x ∈ D (A) . (9.26)
Proof. See lemma 9.2. Or, a direct computation shows that
k(A + i) xk2 = h(A + i) x, (A + i) xi

= kxk2 + kAxk2 + i (hAx, xi − hx, Axi)
= kxk2 + kAxk2 ;
where hAx, xi − hx, Axi = 0 since A is Hermitian.
Theorem 9.6 (Cayley transform). Let A be a densely defined, closed, Hermitian

operator in H .
1. The following subspaces in H are isometrically isomorphic:

ran (A ± i) ' G (A) ' D (A) .
In particular, ran (A ± i) are closed subspace in H .

2. The map CA : ran (A + i) → ran (A − i) by
(A + i) x 7→ (A − i) x, ∀x ∈ D (A) (9.27)
is isometric. Equivalently,
CA x = (A − i) (A + i)−1 x (9.28)
for all x ∈ ran (A + i).

3. Moreover,
A = i (1 +CA ) (1 −CA )−1 . (9.29)
Proof. By (9.3), ran (A ± i) are isometric to the graph of A, and the latter is closed
(as a subset in H ⊕ H ) since A is closed (i.e., G (A) is closed). Thus, ran (A ± i)
are closed in H . Note this is also a result of 9.1.
The mapping (9.28) being isometric follows from (9.26).
By (9.27), we have
(1 −CA ) ((A + i) x) = (A + i) x − (A − i) x = 2ix

(1 +CA ) ((A + i) x) = (A + i) x + (A − i) x = 2Ax
for all x ∈ D (A). It follows that
(1 +CA ) (1 −CA )−1 (2ix) = (1 +CA ) ((A + i) x) = 2Ax; i.e.,
Ax = i (1 +CA ) (1 −CA )−1 x, ∀x ∈ D (A)
which is (9.29).
Theorem 9.7. Suppose A is densely defined, closed, and Hermitian in H . Then

the family of (closed) Hermitian extensions of A is indexed by partial isometries U
with initial space in D+ (A) and final space in D− (A). Given U, the corresponding
extension AeU ⊃ A is determined by
eU (x + (1 −U) x+ ) = x + i (1 +U) x+ , where

A
eU ) = {x + (1 −U) x+ : x ∈ D (A) , x+ ∈ D+ (A)}
dom(A
9.3 Boundary Triple 241
eU is selfadjoint if and only if U is unitary from D+ (A) onto D− (A).

Moreover, A
Proof. Since A is closed, we get the following decompositions (9.1)
H = ran (A + i) ⊕ ker (A∗ − i)

= ran (A − i) ⊕ ker (A∗ + i) .
By theorem 9.6, CA : ran (A + i) → ran (A − i) is isometric. Consequently, getting a

Hermitian extension of A amounts to choosing a partial isometry U with initial space
in ker (A∗ − i) (= D+ (A)) and final space in ker (A∗ + i) (= D− (A)), such that
CAe := CA ⊕U
U
eU ⊃ A.
is the Cayley transform of A
Given U as above, for all x ∈ D (A), x+ ∈ D+ (A), we have
CAe ((A + i) x ⊕ x+ ) = (A − i) x ⊕Ux+ .

U
Then,
(1 −CAe ) ((A + i) x ⊕ x+ ) = ((A + i) x + x+ ) − ((A − i) x +Ux+ )

U
= 2ix + (1 −U) x+
(1 +CAe ) ((A + i) x ⊕ x+ ) = ((A + i) x + x+ ) + ((A − i) x +Ux+ )
U
= 2Ax + (1 +U) x+ ;
and so

−1 1 1
i(1 +CAe )(1 −CAe ) x + (1 −U) x+ = Ax + (1 +U) x+ .
U U 2i 2
The theorem follows by setting x+ := 2iy+ .
9.3 Boundary Triple
In applications, especially differential equations, it is convenient to character-

ize selfadjoint extensions using boundary conditions. For recent applications, see
[JPT12c, JPT12b, JPT12a]. A slightly modified version can be found in [dO09].
Let A be a densely defined, closed, Hermitian operator acting in a Hilbert space

H . Assume A has deficiency indices (d, d), d > 0, and so A has non-trivial self-
adjoint extensions. By von Neumann’s theorem (theorem 9.3), for all x, y ∈ D (A∗ ),
we have the following decomposition,
x = x0 + x+ + x−
y = y0 + y+ + y−
where x0 , y0 ∈ D (A), x+ , y+ ∈ D+ (A), and x− , y− ∈ D− (A). Then,
hy, A∗ xi − hA∗ y, xi
= hy0 + y+ + y− , Ax0 + i (x+ − x− )i − hAy0 + i (y+ − y− ) , x0 + x+ + x− i
= hy0 , Ax0 i − hAy0 , x0 i + hy0 , i (x+ − x− )i − hAy0 , x+ + x− i +
| {z } | {z }
0 0
hy+ + y− , Ax0 i − hi (y+ − y− ) , x0 i +
| {z }
0
hy+ + y− , i (x+ − x− )i − hi (y+ − y− ) , x+ + x− i
= 2i {hy+ , x+ i − hy− , x− i} . (9.30)
Therefore, we see that

h i
x, y ∈ D A

e ⊃ A, Hermitian extension ⇐⇒ [RHS of (9.30) vanishes]
e ,A
For selfadjoint extensions, this is equivalent to choosing a partial isometry U from

D+ (A) onto D− (A), and setting
x− = Ux+ , y− = Uy+ ; so that
hy, A∗ xi − hA∗ y, xi = 2i {hy+ , x+ i − hUy+ ,Ux+ i}

= 2i {hy+ , x+ i − hy+ , x+ i} = 0.
The discussion above leads to the following definition:
Definition 9.3. Let A be a densely defined, closed, Hermitian operator in H . Sup-

pose A has deficiency indices (d, d), d > 0. A boundary space for A is a triple
(Hb , ρ1 , ρ2 ) consisting of a Hilbert space Hb and two linear maps ρ1 , ρ2 : D (A∗ ) →
Hb , such that
1. ρi (D (A∗ )) is dense in Hb , i = 1, 2; and

2. for all x, y ∈ D (A∗ ), ∃ c 6= 0, such that
hy, A∗ xi − hA∗ y, xi = c [hρ1 (y) , ρ1 (x)ib − hρ2 (y) , ρ2 (x)ib ] . (9.31)
Remark 9.4. In (9.30), we set
Hb = D+ (A)
ρ1 (x0 + x+ + x− ) = x+
ρ2 (x0 + x+ + x− ) = Ux+
for any x = x0 + x+ + x− in D (A∗ ). Then (Hb , ρ1 , ρ2 ) is a boundary space for A. In

this special case, ρ1 , ρ2 are surjective. It is clear that the choice of a boundary triple
is not unique. In applications, Hb is usually chosen to have the same dimension as
D± (A).
Consequently, theorem 9.4 can be restated as follows.
Theorem 9.8. Let A be a densely defined, closed, Hermitian operator in H . Sup-
pose A has deficiency indices (d, d), d > 0. Let (Hb , ρ1 , ρ2 ) be a boundary triple.
Then the selfadjoint extensions of A are indexed by unitary operators U : Hb → Hb ,
such that given U, the corresponding selfadjoint extension AfU ⊃ A is determined by
∗
U =A
Af , where
D (Af
U)

∗
D Af
U = {x ∈ D (A ) : Uρ1 (x) = ρ2 (x)} .
Certain variations of theorem 9.8 are convenient in the boundary value problems
(BVP) of differential equations. In [DM91, GG91], a boundary triple (Hb , β1 , β2 )
is defined to satisfy
hx, A∗ yi = hA∗ x, yi = c0 [hβ1 (x) , β2 (y)ib − hβ2 (x) , β1 (y)ib ] (9.32)
for all x, y ∈ D (A∗ ); and c0 is some nonzero constant. Also, see [JPT12c, JPT12b,
JPT12a].
The connection between (9.31) and (9.32) is via the bijection
β1 = ρ1 + ρ2 
( )  
ρ1 = β1 + iβ2 2
⇐⇒ . (9.33)
ρ2 = β1 − iβ2 β = ρ1 − ρ2 
2
2i
Lemma 9.4. Under the bijection (9.33), we have
hρ1 (x) , ρ1 (y)ib − hρ2 (x) , ρ2 (y)ib = 2i (hβ1 (x) , β2 (y)ib − hβ2 (x) , β1 (y)ib )
Proof. For convenience, we suppress the variables x, y. Then a direct computation

shows that,
hρ1 , ρ1 ib − hρ2 , ρ2 ib
= hβ1 + iβ2 , β1 + iβ2 ib − hβ1 − iβ2 , β1 − iβ2 ib
= i hβ1 , β2 ib − i hβ2 , β1 ib + i hβ1 , β2 ib − i hβ2 , β1 ib
= 2i (hβ1 , β2 ib − hβ2 , β1 ib )
which is the desired conclusion.
Theorem 9.9. Given a boundary triple (Hb , β1 , β2 ) satisfying (9.32), the family of
U ⊃ A is indexed by unitary operators U : Hb → Hb , such
selfadjoint extensions Af
that
∗
U =A
Af, where (9.34)
D (Af
U)

∗
D Af
U = {x ∈ D (A ) : (1 −U) β1 (x) = i (1 +U) β2 (x)} . (9.35)
Proof. By theorem 9.8, we need only pick a unitary operator U : Hb → Hb , such

that ρ2 = Uρ1 . In view of the bijection (9.33), this yields
β1 − iβ2 = U (β1 + iβ2 ) ⇐⇒ (1 −U) β1 = i (1 +U) β2
and the theorem follows.
d
Example 9.3. Let A = −i dx , and
D (A)
D (A) = f ∈ L2 (0, 1) : f 0 ∈ L2 (0, 1) , f (0) = f (1) = 0 .

Then A∗ = −i dx
d
, where
D (A∗ )
D (A∗ ) = f : f , f 0 ∈ L2 (0, 1) .

For all f , g ∈ D (A∗ ), using integration by parts, we get

1
hg, A∗ f i − hA∗ g, f i = −ig (x) f (x) = −i g (1) f (1) − g (0) f (0) .
0
Let Hb = C, i.e., one-dimensional, and set
ρ1 ( f ) = f (1) , ρ2 ( f ) = f (0) ; then
hg, A∗ f i − hA∗ g, f i = −i (hρ1 (g) , ρ1 ( f )ib − hρ2 (g) , ρ2 ( f )ib ) .
Therefore, (Hb , ρ1 , ρ2 ) is a boundary triple.

The family of selfadjoint extensions of A is given by the unitary operator
eiθ : Hb → Hb , s.t. ρ2 = eiθ ρ1 ;
i.e.,
fθ = −i d
A .
dx { f ∈D (A∗ ): f (0)=eiθ f (1)}
Example 9.4. Let A f = − f 00 , with D (A) = Cc∞ (0, ∞). Since A is Hermitian and A ≥
0, it follows that it has equal deficiency indices. Also, D (A∗ ) = f , f 00 ∈ L2 (0, ∞) ,

and A∗ f = − f 00 , ∀ f ∈ D (A∗ ).
For f , g ∈ D ∗ (A), we have
ˆ ∞ ˆ ∞
∗ 00
0 0 ∞

hg, A f i = − gf = − gf −g f 0 + g00 f
0 0
ˆ ∞
0 0

= g f (0) − g f (0) − g00 f
0
= g f 0 (0) − g0 f (0) + hA∗ g, f i

and so
hg, A∗ f i − hA∗ g, f i = g f 0 (0) − g0 f (0) .

Now, set Hb = C, i.e., one-dimensional, and
β1 (ϕ) = ϕ (0) , β2 (ϕ) = ϕ 0 (0) ; then
hg, A∗ f i − hA∗ g, f i = hβ1 (g) , β2 ( f )ib − hβ2 (g) , β1 ( f )ib .
This defines the boundary triple.

The selfadjoint extensions are parameterized by eiθ , where

1 − eiθ β1 ( f ) = i 1 + eiθ β2 ( f ) ; i.e.,
f (0) = z f 0 (0) , f ∈ D (A∗ )
where
1 + eiθ
z=i .
1 − eiθ
We take the convention that z = ∞ ⇐⇒ f 0 (0) = 0, i.e., the Neumann boundary
condition.
Example 9.5. A f = − f 00 , D (A) = Cc∞ (0, 1); then D (A∗ ) = f , f 00 ∈ L2 (0, 1) . In-

tegration by parts gives

ˆ 1
hg, A∗ f i = − g f 00
0
ˆ 1
1
= − g f 0 − g0 f 0 − g00 f

0
1
= − g f 0 − g0 f 0 + hA∗ g, f i .

Thus,
hg, A∗ f i − hA∗ g, f i = g f 0 (0) + g0 f (1) − g0 f (0) + g f 0 (1)

= hβ1 (g) , β2 ( f )ib − hβ2 (g) , β1 ( f )ib
where ! !
ϕ (0) ϕ 0 (0)
β1 (ϕ) = , β2 (ϕ) = .
ϕ 0 (1) ϕ (1)
The boundary space is Hb = C2 , i.e., 2-dimensional

The family of selfadjoint extensions is parameterized by U ∈ M (2, C). Given U,
the corresponding extension Af
U is determined by
∗
U =A
Af , where
D (Af
U)

∗
D Af
U = { f ∈ D (A ) : (1 −U) β1 ( f ) = i (1 +U) β2 ( f )} .
Remark 9.5. Another choice of the boundary map:
hg, A∗ f i − hA∗ g, f i = g f 0 (0) − g f 0 (1) − g0 f (0) − g0 f (1)

= hβ1 (g) , β2 ( f )ib − hβ2 (g) , β1 ( f )ib
where
9.4 The Friedrichs Extension 247
! !
ϕ (0) ϕ 0 (0)
β1 (ϕ) = , β2 (ϕ) = .
ϕ (1) −ϕ 0 (1)
The selfadjoint boundary condition leads to
! !
f (0) f 0 (0)
(1 −U) = i (1 +U) .
f (1) − f 0 (1)
For U = 1, we get the Neumann boundary condition:
f 0 (0) = f 0 (1) = 0.
For U = −1, we get the Dirichlet boundary condition:
f (0) = f (1) = 0.
9.4 The Friedrichs Extension
Let A be a densely defined Hermitian operator in a Hilbert space H . A is semi-

bounded if A ≥ c > −∞, in the sense that, hx, Axi ≥ c hx, xi, ∀x ∈ dom (A). Set
LA := inf {hx, Axi : x ∈ dom (A) , kxk = 1} (9.36)
and LA is called the lower bound of A.

In the following discussion, we first assume A ≥ 1 and eventually drop the con-
straint.
Let H1 = completion of dom (A) with respect to the inner product
hx, yi1 := hx, Ayi (9.37)
Theorem 9.10. Let A (≥ 1), H , and H1 be as above. Then
1. kxk ≤ kxk1 , ∀x ∈ dom (A).

2. k·k and k·k1 are topologically consistent, i.e., the identity map
ϕ : dom (A) → dom (A)
extends by continuity to
ϕe : H1 ,→ H
such that
kxk ≤ kxk1 , ∀x ∈ H1 . (9.38)
Therefore, H1 is identified as a dense subspace in H .

3. Moreover,
hy, xi1 = hy, Axi , ∀x ∈ dom (A) , ∀y ∈ H1 . (9.39)
4. Define
e := A∗
A , where
dom(A)
e
e∗ ) ∩ H1 .
e := dom(A
dom(A)
e∗ , and L e = LA .
e=A
Then A A
Proof. (1)-(2) The assumption A ≥ 1 implies that
kxk2 = hx, xi ≤ hx, Axi = kxk21 , ∀x ∈ dom (A) .
Hence ϕ is continuous and the norm ordering passes to the completions of dom (A)
with respect to k·k1 and k·k. Therefore (9.38) holds.
Next, we verify that ϕe is injective (i.e., kerϕe = 0.) Suppose (xn ) ⊂ dom (A) s.t.
k·k k·k
1
xn −−→ x ∈ H1 , and xn −→ 0. We must show that kxk1 = 0. But
kxk21 = lim hxm , xn i1 (limit exists by assumption)

m,n→∞
= lim hxm , Axn i

m,n→∞
= lim h0, Axn i = 0.

n→∞
In the computation, we used the fact that
|hxm − x, Axn i| ≤ kxm − xk kAxn k

≤ kxm − xk1 kAxn k → 0, as m → ∞.
(3) Let (yn ) ⊂ dom (A), and kyn − yk1 → 0. For all x ∈ dom (A), we have
hy, xi1 = lim hyn , xi1 = lim hyn , Axi = hy, Axi .
n→∞ n→∞
9.4 The Friedrichs Extension 249
Equivalently,
|hyn , Axi − hy, Axi| = |hyn − y, Axi|

≤ kyn − yk kAxk
≤ kyn − yk1 kAxk → 0, as n → ∞.
e ∃ (xn ) , (yn ) ⊂ dom (A) s.t. kxn − xk → 0 and kyn − yk →

(4) For all x, y ∈ dom(A), 1 1
0. Hence the following limit exists:

lim hxm , Ayn i = lim hxm , yn i1 .
m,n→∞ m,n→∞
Consequently,
lim lim hxm , Ayn i = lim lim hAxm , yn i

m→∞ n→∞ m→∞ n→∞
= lim hAxm , yi
m→∞
= lim hxm , A∗ yi
m→∞
= lim hxm , Ayi

e = hx, Ayi
e
m→∞
and
lim lim hxm , Ayn i = lim hx, Ayn i

n→∞ m→∞ n→∞
= lim hA∗ x, yn i
n→∞
= lim hAx,
e yn i = hAx,
e yi.
n→∞
Thus, A
e is Hermitian.
Fix y ∈ H . The map x 7−→ hy, xi, ∀x ∈ dom (A) ⊂ H1 , is linear and satisfies
|hy, xi| ≤ kyk kxk ≤ kyk kxk1 .
Hence it extends to a unique bounded linear functional on H1 , as dom (A) is dense

in H1 .
By Riesz’s theorem, there exists unique hy ∈ H1 s.t.
hy, xi = hy , x 1
, ∀x ∈ H1 . (9.40)
In particular,
hy, xi = hy , x 1
= hy , Ax , ∀x ∈ dom (A) .
Then, hy ∈ H1 ∩ dom(A∗ ) = dom(A), e = H . Note

e y = y. Therefore, ran(A)
e and Ah
we have established the identity
e xi = hy, xi , ∀y ∈ dom(A),
hAy, e ∀x ∈ dom (A) . (9.41)
1
e = H implies that A
Claim. ran(A) e is selfadjoint. In fact, for all x ∈ dom(A)
e and
e∗ ), we have
y ∈ dom(A
hy, Axi e∗ y, xi = hAh,

e = hA e xi = hh, Axi
e
e∗ y = Ah,
where A e for some h ∈ dom(A), e = H . Thus,
e using the assumption ran(A)
hy − h, Axi
e = 0, ∀x ∈ dom(A);
e
e = H . Therefore, y = h, and so y ∈ dom(A).

i.e., y − h ⊥ ran(A) e
Finally, we show LAe = LA . By the definition of lower bound, dom(A) ⊂ dom(A)

e
1 k·k
implies LAe ≤ LA . On the other hand, let (xn ) ⊂ dom (A) s.t. xn −−→ x ∈ dom(A),
e
then
hx, Axi
e = lim hx, Ax
e n i = lim hx, Axn i
n→∞ n→∞
= lim hx, xn i1 = hx, xi1 ≥ LA hx, xi

n→∞
which shows that LAe ≥ LA .
Remark 9.6. In the proof of theorem 9.10, we established an embedding ψ : H ,→

H ∗ by ψ : y 7→ hy with the defining equation (9.40). And we define Ah
1
e y = y, i.e.,
e−1 y. It follows that dom(A)
hy = A e =H .
e = ran (ψ), and ran(A)
Theorem 9.11. The Friedrichs extension of A is the unique selfadjoint operator sat-
isfying
e yi = hx, yi , ∀x ∈ dom(A),
hAx, e ∀y ∈ dom (A) .
1
See (9.41).
Proof. Suppose A ⊂ B,C ⊂ A∗ , and B,C selfadjoint, satisfying
hBx, yi = hx, yi1 , ∀x ∈ dom (B) , ∀y ∈ dom (A)

9.5 Rigged Hilbert Space 251
hCx, yi = hx, yi1 , ∀x ∈ dom (C) , ∀y ∈ dom (A) .
Then, for all x, y ∈ dom (A), we have
hBx, yi = hCx, yi = hx,Cyi (= hx, Ayi) .
Fix x ∈ dom (A), and the above identify passes to y ∈ dom (C). Therefore, y ∈ B∗ =
B, and By = Cy. This shows C ⊂ B. Since
C = C ∗ ⊃ B∗ = B
i.e., C ⊃ B, it then follows that B = C.
Remark 9.7. If A is only assumed to be semi-bounded, i.e., A ≥ c > −∞, then B :=

A − c + 1 ≥ 1, and we get the Friedrichs extension Be of B; and Be − 1 + c is the
Friedrichs extension of A.
9.5 Rigged Hilbert Space
In the construction of Friedrichs extensions of semi-bounded operators, we have im-

plicitly used the idea of rigged Hilbert spaces. We study this method systematically
and recover the Friedrichs extension as a special case.
Let H0 be a Hilbert space with inner product h·, ·i0 , and H1 be a dense subspace
in H0 , which by itself, is a Hilbert space with respect to h·, ·i1 . Further, assume the
ordering
kxk0 ≤ kxk1 , ∀x ∈ H1 . (9.42)
Hence, the identity map

id : H1 ,→ H0 (9.43)
is continuous with a dense image.

Let H−1 be the space of bounded conjugate linear functionals on H1 . By Riesz’s
theorem, H−1 is identified with H1 via the map
H−1 → H1 , f 7→ ξ f , s,t. (9.44)

f (x) = x, ξ f 1
, ∀x ∈ H1 . (9.45)
Then H−1 is a Hilbert space with respect to the inner product
h f , gi−1 = ξ f , ξg 1
, ∀ f , g ∈ H−1 . (9.46)
Remark 9.8. The map f 7→ ξ f in (9.44) is linear. For if c ∈ C, then x, ξc f 1 =

= x, cξ f 1 , for all x ∈ H1 ; i.e., ξc f − cξ f ⊥ H1 , and so ξc f =

c f (x) = c x, ξ f
1
cξ f .
Theorem 9.12. The mapping
H0 ,→ H−1 , x 7→ h·, xi0 , ∀x ∈ H0 (9.47)
is linear, injective, continuous, and having a dense image in H−1 .
Proof. Since cx 7→ h·, cxi = c h·, xi, the mapping in (9.47) is linear.
For all x ∈ H0 , we have 8.3
(9.42)
|hy, xi0 | ≤ kxk0 kyk0 ≤ kxk0 kyk1 , ∀y ∈ H1 ; (9.48)
hence h·, xi0 is a bounded conjugate linear functional on H1 , i.e., h·, xi0 ∈ H−1 .
Moreover, by (9.48),
kh·, xi0 k−1 ≤ kxk0 . (9.49)
If h·, xi0 ≡ 0 in H−1 , then hy, xi0 = 0, for all y ∈ H1 . Since H1 is dense in H0 ,
it follows that x = 0 in H0 . Thus, (9.47) is injective.
Now, if f ⊥ {h·, xi0 : x ∈ H1 } in H−1 , then
(9.46) (9.44)
D E
h f , h·, xi0 i−1 = ξ f , ξh·,xi0 = ξf ,x 0
= 0, ∀x ∈ H1 . (9.50)
1
Thus, ξ f 0
= 0, since H1 is dense in H0 . Since id : H1 ,→ H0 is injective, and
ξ f ∈ H1 , it follows that ξ f 1 = 0. This, in turn, implies k f k−1 = 0, and so f = 0
in H−1 . Consequently, the image of H1 (resp. H0 as it contains H1 ) under (9.47)
is dense in H−1 .
Combining (9.42) and theorem 9.12, we get the triple of Hilbert spaces
(9.43) (9.47)
H1 ,−−−→ H0 ,−−−→ H−1 . (9.51)
The following are immediate:

1. All mappings in (9.51) are injective, continuous (in fact, contractive), hav-
ing dense images.
2. The map x 7→ ξh·,xi0 is a contraction from H1 ⊂ H0 into H0 . This follows
from the estimate:
(9.42)
ξh·,xi0 ≤ ξh·,xi0
0 1
(9.46)
= kh·, xi0 k−1
(9.49) (9.42)
≤ kxk0 ≤ kxk1 , ∀x ∈ H1 . (9.52)
In particular, for x ∈ H1 , x 6= ξh·,xi0 in general.

3. The canonical bilinear form h·, ·i : H1 × H−1 → C is given by
f (x) = x, ξ f 1
, ∀x ∈ H1 , ∀ f ∈ H−1 . (9.53)
In particular, if f = h·, yi0 , y ∈ H0 , then

D E
x, ξ f 1
= x, ξh·,yi0
= hx, yi0 , ∀x ∈ H1 .
1
4. By theorem 9.12, H0 is dense in H−1 , and
hx, yi−1 := hh·, xi0 , h·, yi0 i−1

D E
= ξh·,xi0 , ξh·,yi0
D E D1 E
= x, ξh·,yi0 = ξh·,xi0 , y , ∀x, y ∈ H0 . (9.54)
0 0
Combined with the order relation kxk−1 ≤ kxk0 for all x ∈ H0 , we see that
D E
hx, yi−1 = x, ξh·,yi0
0
≤ kxk0 ξh·,yi0
−1
= kxk0 kyk−1 .
≤ kxk0 kyk0 , , ∀x, y ∈ H0 .
Thus h·, ·i−1 is a continuous extension of h·, ·i0 .
Theorem 9.13. Let H1 ,→ H0 ,→ H−1 be the triple in (9.51). Define B : H0 → H0

by
(9.47) (9.44) (9.43)

B : x 7−−−→ h·, xi0 7−−−→ ξh·,xi0 7−−−→ ξh·,xi0 . (9.55)
H0 H−1 H1 H0
Then,
1. For all x ∈ H1 , and all y ∈ H0 ,
hx, Byi1 = hx, yi0 . (9.56)
In particular,
hx, yi−1 = hBx, Byi1

= hx, Byi0 = hBx, yi0 , ∀x, y ∈ H0 ; (9.57)
D E
where hx, yi−1 := ξh·,xi0 , ξh·,yi0 , as defined in (9.54).
1
2. B is invertible.
3. ran (B) is dense in both H1 and H0 .
4. 0 ≤ B ≤ 1. In particular, B is a bounded selfadjoint operator on H0 .
Proof.
1. For y ∈ H0 , By = ξh·,yi0 ∈ H1 , where ξh·,yi0 is given in (9.44)-(9.45). Thus,

D E
hx, Byi1 = x, ξh·,yi0 = hx, yi0 , ∀x ∈ H1 .
1
(9.57) follows from this.

2. If kBxk0 = 0, then kBxk1 = 0, since Bx ∈ H1 and H1 ,→ H0 is injective.
But then
(9.57)
kBxk1 = kxk−1 = 0 =⇒ kxk0 = 0
since H0 ,→ H−1 is injective. This shows that B is injective.

3. Since H0 ,→ H−1 is dense, and H−1 ' H1 , it follows that ran (B) is dense
in H1 . Now if y ∈ H0 , and hy, Bxi0 = 0, for all x ∈ H0 , then
hBy, Bxi1 = 0, ∀x ∈ H0 ;
equivalently,
hy, xi−1 = 0, ∀x ∈ H0 .
Since H0 ,→ H−1 is dense, we have kyk−1 = 0. But y ∈ H0 and H0 ,→

H−1 is injective, it follows that kyk0 = 0, i.e., y = 0 in H0 . Therefore,
ran (B) is also dense in H0 .
4. For all x ∈ H0 , we have
(9.56)
hx, Bxi0 = hBx, Bxi1 ≥ 0 =⇒ B ≥ 0.
On the other hand,
(9.57) (9.49)
hx, Bxi0 = hBx, Bxi1 = hx, xi−1 ≤ hx, xi0 ;
and so B ≤ 1. Since B is positive and bounded, it is selfadjoint.
Another argument:
kBxk0 ≤ kBxk1 = kxk−1 ≤ kxk0 , ∀x ∈ H0 .
In view of applications, it is convenient to reformulate the previous theorem in

terms of B−1 .
Theorem 9.14. Let B, H1 , H0 , H−1 , be as in theorem 9.13, set A := B−1 , then
1. A = A∗ , A ≥ 1.
2. dom (A) is dense in H1 and H0 , and ran (A) = H0 .
3. For all y ∈ dom (A), x ∈ H1 ,
hx, yi1 = hx, Ayi0 . (9.58)
In particular,
hx, yi1 = hAx, Ayi−1

= hAx, yi0 = hx, Ayi0 , ∀x, y ∈ dom (A) . (9.59)
4. There is a unique selfadjoint operator in H0 satisfying (9.58).
Proof. Part (1)-(3) are immediate by theorem 9.13. For (4), suppose A, B are selfad-
joint in H0 s.t.
(i) dom (A), dom (B) are contained in H1 , dense in H0 ;
(ii)
hx, yi1 = hx, Ayi0 , ∀x ∈ H1 , y ∈ dom (A)

hx, yi1 = hx, Byi0 , ∀x ∈ H1 , y ∈ dom (B) .
Then, for all x ∈ dom (A) and y ∈ dom (B),
hx, Byi0 = hx, yi1 = hAx, yi0 .
Thus, x 7→ hAx, yi0 is a bounded linear functional on dom (A), and so y ∈ dom (A∗ ) =
dom (A) and A∗ y = Ay = By; i.e., A ⊃ B. Since A, B are selfadjoint, then
B = B∗ ⊂ A∗ = A.
Therefore A = B.
Theorem 9.15. Let H1 , H0 , A as in theorem 9.14. Then
1. H1 = dom A1/2 , and

D E
hx, yi1 = A1/2 x, A1/2 y , ∀x, y ∈ H1 . (9.60)
0
2. For all x, y ∈ H0 ,
D E
hx, yi−1 = A−1/2 x, A−1/2 y . (9.61)
0
Since H0 is dense in H−1 , then H−1 = completion of H0 under the

A−1/2 · 0 -norm.
3. For all x ∈ dom (A),

kAxk−1 = kxk1 = A1/2 x . (9.62)
0
Consequently, the map dom (A) 3 x 7→ Ax ∈ H0 extends by continuity to

a unitary operator from H1 = dom A1/2 onto H−1 , which is precisely

the inverse of (9.44)-(9.45).
Proof. (1) This is the result of the following observations:
a. dom (A) ⊂ dom A1/2 . With the assumption A ≥ 1, the containment is

clear. The assertion also holds in general: By spectral theorem,

ˆ
x ∈ dom (A) ⇐⇒ 1 + |λ |2 kP (dλ ) xk20 < ∞;
where P (·) is the projection-valued measure (PVM) of A, defined on the

set of all Borel sets in R, and dµx := kP (dλ ) xk20 is a finite positive Borel
measure on R. Thus,
ˆ
(1 + |λ |) kP (dλ ) xk20 < ∞
since L2 ⊂ L1 when the measure is finite. But this is equivalent to x ∈

dom A1/2 .

b. For any Hermitian operator T satisfying T ≥ c > 0, we have the estimate:
kT xk ≤ kxk + kT xk ≤ (1 + c) kT xk
Thus, the graph norm of T is equivalent to kT ·k.

dom (A) is dense in dom A1/2 . Note dom A1/2 is a Hilbert space

c.
with respect to the A1/2 -graph norm. By the discussion above, k·kA1/2 '
A1/2 · 0 .
Let y ∈ dom A1/2 , then

y ⊥ dom (A) in dom A1/2
m
D E
A1/2 y, A1/2 x = 0, ∀x ∈ dom A1/2
0
m

hy, Axi0 = 0, ∀x ∈ dom A1/2
m
y = 0 in H0
d. dom (A) is dense in H1 . See theorems 9.13-9.14.

e. A1/2 x 0 = kxk1 , ∀x ∈ dom (A). Indeed,
(9.58)
D E
hx, xi1 = hx, Axi0 = A1/2 x, A1/2 x , ∀x ∈ dom (A) .
0
Conclusion: (i) dom (A) is dense in H1 and dom A1/2 ; (ii) k·k1 and A1/2 · 0

agree on dom (A). Therefore the closures of dom (A) in H1 and dom A1/2 are

identical. This shows H1 = dom A1/2 . (9.60) is immediate.

(2) Given x, y ∈ H0 ,
(9.57) (9.60)
D E
hx, yi−1 = A−1 x, A−1 y 1
= A−1/2 x, A−1/2 y .
0
(3) Given x ∈ dom (A), we have
(9.57) (9.60)
kAxk−1 = A−1 (Ax) 1
= kxk1 = A1/2 x .
0
Application: Friedrichs extension.
Theorem 9.16 (Friedrichs). Let A be a densely defined Hermitian operator acting

in H0 , and assume A ≥ 1. There exists a selfadjoint extension S ⊃ A, such that
LS = LA , i.e., A and S have the same lower bound.
Proof. Given A, construct the triple H1 ,→ H0 ,→ H−1 as in theorem 9.10, so that

H1 = clh·,A·i0 (dom (A)), and
hy, xi1 = hy, Axi0 , ∀x ∈ dom (A) , ∀y ∈ H1 . (9.63)
By theorem 9.14, there is a densely defined selfadjoint operator S in H0 , s.t. (i)

H1 = dom S1/2 ⊃ dom (S); (ii)

D E
hy, xi1 = S1/2 y, S1/2 x , ∀x, y ∈ H1 . (9.64)
0
Combing (9.63)-(9.64), we get

D E
hy, Axi0 = S1/2 y, S1/2 x , ∀x ∈ dom (A) , ∀y ∈ dom S1/2 .
0
Therefore, S1/2 x ∈ dom S1/2 , i.e., x ∈ dom (S), and

hy, Axi0 = hy, Sxi0 , ∀x ∈ dom (A) , ∀y ∈ dom S1/2 .
Since H1 = dom S1/2 is dense in H0 , we conclude that Sx = Ax, for all x ∈

dom (A). Thus, S ⊃ A.

Clearly, S ⊃ A implies LA ≥ LS . On the other hand,
kxk21 = hx, Axi0 ≥ LA hx, xi0 = L kxk20 , ∀x ∈ dom (A)
and the inequality passes by continuity to all x ∈ H1 ; i.e.,
kxk21 ≥ LA kxk20 , ∀x ∈ H1 .
This is equivalent (by theorem 9.14) to

D E
S1/2 x, S1/2 x ≥ LA hx, xi0 , ∀x ∈ dom S1/2 (= H1 ) .
0
In particular,
hx, Sxi0 ≥ LA hx, xi0 , ∀x ∈ dom (S)
and so LS ≥ LA . Therefore, LA = LS .
we suggest: [AG93, Nel69, Dev72, DS88c, Kre46, Jor08, Rud73, Sto51b].
Chapter 10
Unbounded Graph-Laplacians
Mathematics is an experimental science, and definitions do not

come first, but later on. — Oliver Heaviside
It is nice to know that the computer understands the problem.

But I would like to understand it too. — Eugene Wigner
Below we study selfadjoint operators, and extensions in a particular case arising in

the study of infinite graphs; the operators here are infinite discrete Laplacians.
As an application of the previous chapter, we consider the Friedrichs extension
of discrete Laplacian in infinite networks [JP10, JP13a, JP13c, JT14a, JT14c].
By an electrical network we mean a graph G of vertices and edges satisfying
suitable conditions which allow for computation of voltage distribution from a net-
work of prescribed resistors assigned to the edges in G. The mathematical axioms
are prescribed in a way that facilitates the use of the laws of Kirchhoff and Ohm
in computing voltage distributions and resistance distances in G. It will be more
convenient to work with prescribed conductance functions c on G. Indeed with a
choice of conductance function c specified we define two crucial tools for our anal-
ysis, a graph Laplacian ∆ (= ∆c , ) a discrete version of more classical notions of
Laplacians, and an energy Hilbert space HE .
Because of statistical consideration, and our use of random walk models, we
focus our study on infinite electrical networks, i.e., the case when the graph G is
countable infinite. In this case, for realistic models the graph Laplacian ∆c will then
be an unbounded operator with dense domain in HE , Hermitian and semibounded.
Hence it has a unique Friedrichs extension.
Large networks arise in both pure and applied mathematics, e.g., in graph the-
ory (the mathematical theory of networks), and more recently, they have become
a current and fast developing research area; with applications including a host of
problems coming from for example internet search, and social networks. Hence, of
the recent applications, there is a change in outlook from finite to infinite.
More precisely, in traditional graph theoretical problems, the whole graph is
given exactly, and we are then looking for relationships between its parameters,
261
262 10 Unbounded Graph-Laplacians
variables and functions; or for efficient algorithms for computing them. By contrast,
for very large networks (like the Internet), variables are typically not known com-
pletely; – in most cases they may not even be well defined. In such applications,
data about them can only be collected by indirect means; hence random variables
and local sampling must be used as opposed to global processes.
Although such modern applications go far beyond the setting of large electrical
networks (even the case of infinite sets of vertices and edges), it is nonetheless true
that the framework of large electrical networks is helpful as a basis for the analysis
we develop below; and so our results will be phrased in the setting of large electrical
networks, even though the framework is much more general.
The applications of “large” or infinite graphs are extensive, counting just physics;
see for example [BCD06, RAKK05, KMRS05, BC05, TD03, VZ92].
10.1 Basic Setting
Let V be a countable discrete set, and let E ⊂ V ×V be a subset such that:
1. (x, y) ∈ E ⇐⇒ (y, x) ∈ E; x, y ∈ V ;
2. # {y ∈ V | (x, y) ∈ E} is finite, and > 0 for all x ∈ V ;
3. (x, x) ∈
/ E; and
4. ∃ o ∈ V s.t. for all y ∈ V ∃ x0 , x1 , . . . , xn ∈ V with x0 = o, xn = y, (xi−1 , xi ) ∈
E, ∀i = 1, . . . , n. (This property is called connectedness.)
5. If a conductance function c is given we require cxi−1 xi > 0. See Definition
10.1 below.
Definition 10.1. A function c : E → R+ ∪ {0} is called conductance function if
1. c (e) ≥ 0, ∀e ∈ E; and
2. Given x ∈ V , cxy > 0, cxy = cyx , for all (xy) ∈ E.
If x ∈ V , we set
c (x) := ∑ cxy . (10.1)
(xy)∈E
The summation in (10.1) is denoted x ∼ y; i.e., x ∼ y if (xy) ∈ E.
Definition 10.2. When c is a conductance function (see Def. 10.1) we set ∆ = ∆c

(the corresponding graph Laplacian
10.1 Basic Setting 263
(∆ u) (x) = ∑ cxy (u (x) − u (y)) = c (x) u (x) − ∑ cxy u (y) . (10.2)

y∼x y∼x
Given G = (V, E, c) as above, and let ∆ = ∆c be the corresponding graph Lapla-

cian. With a suitable ordering on V , we obtain the following banded ∞ × ∞ matrix-
representation for ∆ (eq. (10.3)). We refer to [GLS12] for a number of applications
of infinite banded matrices.
 
c (x1 ) −cx1 x2 0 ··· ··· ··· ··· 0 ···
 .. 
−cx2 x1 c (x2 ) −cx2 x3 0 ··· ··· ··· . · · ·
 
0
 
 0 −cx3 x2 c (x3 ) −cx3 x4 ··· ··· · · ·
 
 0 
 .. . .. . .. . .. .. .. .. 
 .
 0 . . . · · ·  (10.3)
 . .. .. .. .. .. ..
 ..

 . . . . . 0 . · · ·
 .
0

 .
 . ··· 0 −cxn xn−1 c (xn ) −cxn xn+1 0 · · ·

 . .. .. .. .. ..

.. . ··· ··· 0 . . . .
Remark 10.1 (Random Walk). If (V, E, c) is given as in Definition 10.2, then for
(x, y) ∈ E, set
cxy
pxy := (10.4)
c (x)

and note then pxy in (10.4) is a system of transition probabilities, i.e., ∑y pxy = 1,
∀x ∈ V , see Fig. 10.1 below.
y2
y3
p x y3
p x y2
x
p x y1
y1
Fig. 10.1: Transition probabilities pxy at a vertex x (in V ).

A Markov-random walk on V with transition probabilities (pxy ) is said to be

reversible iff ∃ a positive function ce on V such that
ce(x) pxy = ce(y) pyx , ∀ (xy) ∈ E. (10.5)
10.2 The Energy Hilbert Spaces HE
Let G = (V, E, c) be an infinite connected network introduced above. Set HE :=

completion of all compactly supported functions u : V → C with respect to
1
hu, viHE := ∑∑ cxy (u (x) − u (y)) (v (x) − v (y)) (10.6)
2 (x,y)∈E
1
kuk2HE : = ∑∑ cxy |u (x) − u (y)|2
2 (x,y)∈E
(10.7)
then HE is a Hilbert space [JP10, JT14c].
Lemma 10.1. For all x, y ∈ V , there is a unique real-valued dipole vector vxy ∈ HE
s.t.
vxy , u HE
= u (x) − u (y) , ∀u ∈ HE .
Proof. Apply Riesz’s theorem.
Definition 10.3. Let H be a Hilbert space with inner product denoted h·, ·i, or
h·, ·iH when there is more than one possibility to consider. Let J be a countable
index set, and let w j j∈J be an indexed family of non-zero vectors in H . We say

that w j j∈J is a frame for H iff (Def.) there are two finite positive constants b1

and b2 such that

2
b1 kuk2H ≤ ∑ w j, u H
≤ b2 kuk2H (10.8)
j∈J
holds for all u ∈ H . We say that it is a Parseval frame if b1 = b2 = 1.

For references to the theory and application of frames, see e.g., [HJL+ 13,
KLZ09, CM13, SD13, KOPT13, EO13].
Lemma 10.2. If w j j∈J is a Parseval frame in H , then the (analysis) operator

A = AH : H −→ l 2 (J),

Au = w j, u H j∈J
(10.9)
is well-defined and isometric. Its adjoint A∗ : l 2 (J) −→ H is given by

10.2 The Energy Hilbert Spaces HE 265

A∗ (γ j ) j∈J := ∑ γ jw j (10.10)
j∈J
and the following hold:
1. The sum on the RHS in (10.10) is norm-convergent;

2. A∗ : l 2 (J) −→ H is co-isometric; and for all u ∈ H , we have
u = A∗ Au = ∑ w j, u w j (10.11)
j∈J
where the RHS in (10.11) is norm-convergent.
Proof. The details are standard in the theory of frames; see the cited papers above.
Note that (10.8) for b1 = b2 = 1 simply states that A in (10.9) is isometric, and so
A∗ A = IH = the identity operator in H , and AA∗ = the projection onto the range
of A.
Theorem 10.1. Let G = (V, E, c) be an infinite network. Choose an orientation on

the edges, denoted by E (ori) . Then the system of vectors
n √ o
wxy := cxy vxy , (xy) ∈ E (ori) (10.12)
is a Parseval frame for the energy Hilbert space HE . For all u ∈ HE , we have the
following representation
u= ∑ cxy vxy , u vxy , and (10.13)

(xy)∈E (ori)
2
kuk2HE = ∑ cxy vxy , u (10.14)
(xy)∈E (ori)
Proof. See [JT14b, CH08].
Frames in HE consisting of our system (10.12) are not ONBs when resisters are
configured in non-linear systems of vertices, for example, resisters in parallel. See
Fig 10.2, and Example 10.1.
V = Band V = Z2
Fig. 10.2: non-linear system of vertices
Example 10.1. Let c01 , c02 , c12 be positive constants, and assign conductances on the
three edges (see Fig 10.3) in the triangle network.
c01 c02
c12
1 2

Fig. 10.3: The set vxy : (xy) ∈ E is not orthogonal.
√
In this case, wi j = ei j vi j , i < j, in the cyclic order is a Parseval frame but not
an ONB in HE [JT14b].
Note the corresponding Laplacian ∆ (= ∆c ) has the following matrix representa-
tion  
c (0) −c01 −c02
M := −c01 c (1) −c12  (10.15)
 
−c02 −c12 c (2)

n o
The dipoles vxy : (xy) ∈ E (ori) as 3-D vectors are the solutions to the equation
∆ vxy = δx − δy .
Hence,
h itr
Mv01 = 1 −1 0
h itr
Mv02 = 1 0 −1
10.3 The Graph-Laplacian 267
h itr
Mv12 = 0 1 −1
The Parseval frame from Lemma 10.2 is

√ √ tr
√ c01 c12 c01 c02
w01 = c01 v01 = ,− ,0
c01 c02 + c01 c12 + c02 c12 c01 c02 + c01 c12 + c02 c12
√ √ tr
√ c12 c02 c12 c01
w12 = c12 v12 = 0, ,−
c01 c02 + c01 c12 + c02 c12 c01 c02 + c01 c12 + c02 c12
√ √ tr
− c20 c12

√ c20 c01
w20 = c20 v20 = , 0, .
c01 c02 + c01 c12 + c02 c12 c01 c02 + c01 c12 + c02 c12
Remark 10.2. The dipole vxy is unique in HE as an equivalence class, not a func-
tion on V . Note ker M = harmonic functions = constant (see (10.15)), and so
vxy + const = vxy in HE . Thus, the above frame vectors have non-unique repre-
sentations as functions on V .
10.3 The Graph-Laplacian
Here we include some technical lemmas for graph Laplacian in the energy Hilbert
space HE .
Let G = (V, E, c) be as above; assume G is connected; i.e., there is a base point o
in V such that every x ∈ V is connected to o via a finite path of edges.
If x ∈ V , we set 
1 if y = x
δx (y) = (10.16)
0 if y 6= x
Definition 10.4. Let (V, E, c, o, ∆ ) be as above. Let V 0 := V \ {o}, and set
vx := vx,o , ∀x ∈ V 0 .
Further, let
D2 := span δx x ∈ V , and

(10.17)
n o
DE := ∑x∈V 0 ξx vx finite support ; (10.18)
where by “span” we mean of all finite linear combinations.

Lemma 10.3 below summarizes the key properties of ∆ as an operator, both in l 2 (V )

and in HE .
Lemma 10.3. The following hold:
1. h∆ u, vil 2 = hu, ∆ vil 2 , ∀u, v ∈ D2 ;

2. h∆ u, viHE = hu, ∆ viHE , ∀u, v ∈ DE ;
3. hu, ∆ uil 2 ≥ 0, ∀u ∈ D2 , and
4. hu, ∆ uiHE ≥ 0, ∀u ∈ DE .
Moreover, we have
5. hδx , uiHE = (∆ u) (x), ∀x ∈ V , ∀u ∈ HE .

6. ∆ vxy = δx − δy , ∀vxy ∈ HE . In particular, ∆ vx = δx − δo , x ∈ V 0 = V \ {o}.
7.
δx (·) = c (x) vx (·) − ∑ cxy vy (·) , ∀x ∈ V 0 .
y∼x
8. 


c (x) = ∑t∼x cxt if y = x

δx , δy HE
= −cxy if (xy) ∈ E


if (xy) ∈
/ E, x 6= y

0
Proof. See [JP10, JP11a, JT14b].
10.4 The Friedrichs Extension
Fix a conductance function c. In this section we turn to some technical lemmas we

will need for the Friedrichs extension of ∆ (= ∆c ).
It is known the graph-Laplacian ∆ is automatically essentially selfadjoint as a
densely defined operator in l 2 (V ), but not as a HE operator [Jor08, JP11b]. Since ∆
defined on DE is semibounded, it has the Friedrichs extension ∆Fri (in HE ).

Lemma 10.4. Consider ∆ with dom (∆ ) := span vxy : x, y ∈ V , then
2
hϕ, ∆ ϕiHE = ∑ c2xy vxy , ϕ HE
.
(xy)∈E
Proof. Suppose ϕ = ∑ ϕxy vxy ∈ dom(∆ ). Note the edges are not oriented, and a
direct computation shows that
10.5 A 1D Example 269
2
hϕ, ∆ ϕiHE = 4 ∑ ϕxy .
x,y
Using the Parseval frames in Theorem 10.1, we have the following representation
1
ϕ= ∑ cxy vxy , ϕ v
HE xy
2
(xy)∈E | {z }
=:ϕxy

Note ϕ ∈ span vxy : x, y ∈ V , so the above equation contains a finite sum.
It follows that
2 2
hϕ, ∆ ϕiHE = 4 ∑ ϕxy = ∑ c2xy vxy , ϕ HE
(xy)∈E (xy)∈E
which is the assertion.

Theorem 10.2. Let G = (V, E, c) be an infinite network. If the deficiency indices of

∆ (= ∆c ) are (k, k), k > 0, where dom(∆ ) = span vxy , then the Friedrichs exten-
sion ∆Fri ⊃ ∆ is the restriction of ∆ ∗ to
n o
2
dom(∆Fri ) := u ∈ HE ∑(xy)∈E c2xy vxy , ϕ E
<∞ . (10.19)
Proof. Follows from Lemma 10.4, and the characterization of Friedrichs extensions
of semibounded Hermitian operators (chapter 9); see, e.g., [DS88c, AG93, RS75].

10.5 A 1D Example
Consider G = (V, E, c), where V = {0}∪Z+ . Observation: Every sequence a1 , a2 , . . .

in R+ defines a conductance cn−1,n := an , n ∈ Z+ , i.e.,
0o a1
/1o
a2
/2o
a3
/3 ··· no an+1
/ n+1 ···
The dipole vectors vxy (for x, y ∈ N) are given by


0

 if z ≤ x

vxy (z) = − ∑zk=x+1 a1 if x < z < y
 k

− ∑y
 1
if z ≥ y
k=x+1 ak
See Fig 10.4.
x z y
Fig. 10.4: The dipole vxy .
The corresponding graph Laplacian has the following matrix representation:

 
a1 −a1
−a1 a1 + a2 −a2
 

0
 
 

 −a2 a2 + a3 −a3 

 . 

 −a3 a3 + a4 . . 

 .. .. 
. . −an (10.20)
 
 
 

 −an an + an+1 −an+1 





0 −an+1
..
.
.. 
.


 .. .. 
 . .
 
.. ..
. .
That is,

(∆ u)0

 = a1 (u0 − u1 )

(∆ u)n = an (un − un−1 ) + an+1 (un − un+1 ) (10.21)


= (an + an+1 ) un − an un−1 − an+1 un+1 , ∀n ∈ Z+ .


Lemma 10.5. Let G = (V, c, E) be as above, where an := cn−1,n , n ∈ Z+ . Then u ∈

HE is the solution to ∆ u = −u (i.e., u is a defect vector of ∆ ) if and only if u satisfies
the following equation:
∞
∑ an hvn−1,n , uiHE (δn−1 (s) − δn (s) + vn−1,n (s)) = 0, ∀s ∈ Z+ ; (10.22)
n=1
where ∞ 2
kuk2HE = ∑ an hvn−1,n , uiHE < ∞. (10.23)
n=1
√
forms a Parseval frame in HE . In
∞
Proof. By Theorem 10.1, the set an vn−1,n n=1
fact, the dipole vectors are

0 s ≤ n−1
vn−1,n (s) = ; n = 1, 2, . . . (10.24)
− 1 s≥n
an
√
forms an ONB in HE ; and u ∈ HE has the representation
∞
and so an vn−1,n n=1
∞
u= ∑ an hvn−1,n , uiHE vn−1,n
n=1
see (10.11). Therefore, ∆ u = −u if and only if

∞ ∞
∑ an hvn−1,n , uiHE (δn−1 (s) − δn (s)) = − ∑ an hvn−1,n , uiHE vn−1,n (s)
n=1 n=1
for all s ∈ Z+ , which is the assertion.
Below we compute the deficiency space in an example with index values (1, 1).
Lemma 10.6. Let (V, E, c = {an }) be as above. Let Q > 1 and set an := Qn , n ∈ Z+ ;
then ∆ has deficiency indices (1, 1).
Proof. Suppose ∆ u = −u, u ∈ HE . Then,

1 1+Q 1
−u1 = Q (u1 − u0 ) + Q2 (u1 − u2 ) ⇐⇒ u2 = + u1 − u0
Q2 Q Q

1 1+Q 1
−u2 = Q2 (u2 − u1 ) + Q3 (u2 − u3 ) ⇐⇒ u3 = + u2 − u1
Q3 Q Q
and by induction,

1 1+Q 1
un+1 = + un − un−1 , n ∈ Z+
Qn+1 Q Q
i.e., u is determined by the following matrix equation:

" # " #" #

un+1 1
Qn+1
+ 1+Q 1
Q −Q un
=
un 1 0 un−1
The eigenvalues of the coefficient matrix are

 s 
1+Q 2 4 

1 1 1+Q 1
λ± = + ± + −
2 Qn+1 Q Qn+1 Q Q


1 1+Q

Q−1
 1
∼ ± = 1 as n → ∞.
2 Q Q 
Q

Equivalently, as n → ∞, we have

1+Q 1 1 1
un+1 ∼ un − un−1 = 1 + un − un−1
Q Q Q Q
and so
1
un+1 − un ∼ (un − un−1 ) .
Q
Therefore, for the tail-summation, we have:
(Q − 1)2
∑ Qn (un+1 − un )2 = const ∑ Qn+2
<∞
n n
which implies kukHE < ∞.
Next, we give a random walk interpretation of Lemma 10.6. See Remark 10.1,
and Fig 10.1.
Remark 10.3 (Harmonic functions in HE ). Note that in Example 10.5 (Lemma

10.6), the space of harmonic functions in HE is one-dimensional; in fact if Q > 1 is
fixed, then
u ∈ HE ∆ u = 0

1
is spanned by u = (un )∞
n=0 , un = Qn , n ∈ N; and of course k1/Qn k2HE < ∞.
Remark 10.4. For the domain of the Friedrichs extension ∆Fri , we have:
dom(∆Fri ) = f ∈ HE | ( f (x) − f (x + 1)) Qx ∈ l 2 (Z+ )

(10.25)
i.e.,
( )
∞
2
dom(∆Fri ) = f ∈ HE | ∑ | f (x) − f (x + 1)| 2x
Q <∞ .
x=0
Proof. By Theorem 10.1, we have the following representation, valid for all f ∈
HE :
D x
E x
f = ∑ f , Q 2 v(x,x+1) Q 2 v(x,x+1)
x HE
= ∑ ( f (x) − f (x + 1)) Qx v(x,x+1) ;

x
and
h f , ∆ f iHE = ∑ | f (x) − f (x + 1)|2 Q2x .
x
The desired conclusion (10.25) now follows from Theorem 10.2. Also see e.g.
[DS88c, AG93].
Definition 10.5. Let G = (V, E, c) be a connected graph. The set of transition prob-
abilities (pxy ) is said to be reversible if there exists c : V → R+ s.t.
c (x) pxy = c (y) pyx ; (10.26)
and then
cxy := c (x) pxy (10.27)
is a system of conductance. Conversely, for a system of conductance (cxy ) we set
c (x) := ∑ cxy , and (10.28)

y∼x
cxy
pxy := (10.29)
c (x)
and so (pxy ) is a set of transition probabilities. See Fig 10.5 below.

y y
cxy px y
x cxy ¢ x p x y¢
y¢ y¢
cxy , y ∼ x transition probabilities
Fig. 10.5: neighbors of x
Recall the graph Laplacian in (10.21) can be written as
(∆ u)n = c (n) (un − p− (n) un−1 − p+ (n) un+1 ) , ∀n ∈ Z+ ; (10.30)
where
c (n) := an + an+1 (10.31)
and
an an+1
p− (n) := , p+ (n) := (10.32)
c (n) c (n)
are the left/right transition probabilities, as shown in Fig 10.6.
p- p+ p- p+
0 1 2 n-1 n n+1
Fig. 10.6: The transition probabilities p+ , p− , in the case of constant transition prob-
abilities, i.e., p+ (n) = p+ , and p− (n) = p− for all n ∈ Z+ .
In the case an = Qn , Q > 1, as in Lemma 10.6, we have
c (n) := Qn + Qn+1 , and (10.33)
Qn+1 Q
p+ := p+ (n) = = (10.34)
Qn + Qn+1 1+Q
Qn 1
p− := p− (n) = n n+1
= (10.35)
Q +Q 1+Q
For all n ∈ Z+ ∪ {0}, set
(Pu)n := p− un−1 + p+ un+1 . (10.36)
Note (Pu)0 = u1 . By (10.30), we have
∆ = c (1 − P) . (10.37)
In particular, p+ > 21 , i.e., a random walker has probability > 1

2 of moving to the
right. It follows that
travel time (n, ∞) < ∞;
| {z }
= dist to ∞
and so ∆ is not essentially selfadjoint, i.e., indices (1, 1).

Lemma 10.7. Let (V, E, ∆ (= ∆c )) be as above, where the conductance c is given by
cn−1,n = Qn , n ∈ Z+ , Q > 1 (see Lemma 10.6). For all λ > 0, there exists fλ ∈ HE
satisfying ∆ fλ = λ fλ .

Proof. By (10.37), we have ∆ fλ = λ fλ ⇐⇒ P fλ = 1 − λc fλ , i.e.,

1 Q λ
fλ (n − 1) + fλ (n + 1) = 1 − n−1 f (n)
1+Q 1+Q Q (1 + Q) λ
and so
1+Q λ 1
fλ (n + 1) = − n fλ (n) − f (n − 1) . (10.38)
Q Q Q λ
This corresponds to the following matrix equation:
" # " #" #
1+Q
f (n + 1) Q − Qλn − Q1 f (n)
=
f (n) 1 0 f (n − 1)
" #" #
1+Q
Q − Q1 f (n)
∼ , as n → ∞.
1 0 f (n − 1)
The eigenvalues of the coefficient matrix are given by


1

1+Q Q−1
1

λ± ∼ ± = 1 as n → ∞.
2 Q Q 
Q

That is, as n → ∞,

1+Q 1
fλ (n + 1) ∼ fλ (n) − f (n − 1) ;
Q Q λ
i.e.,
1
fλ (n + 1) ∼ f (n) ; (10.39)
Q λ
and so the tail summation of k fλ k2HE is finite. (See the proof of Lemma 10.6.) We
conclude that fλ ∈ HE .
Corollary 10.1. Let (V, E, ∆ ) be as in the lemma. The Friedrichs extension ∆Fri has
continuous spectrum [0, ∞).
Proof. Fix λ ≥ 0. We prove that if ∆ fλ = λ fλ , f∈ H

∞E , then f λ ∈
/ dom(∆Fri ).
Note for λ = 0, f0 is harmonic, and so f0 = k Q1n for some constant k 6= 0.
n=0
See Remark 10.3. It follows from (10.25) that f0 ∈
/ dom(∆Fri ).
1
The argument for λ > 0 is similar. Since as n → ∞, fλ (n) ∼ Qn (eq. (10.39)), so
by (10.25) again, fλ ∈
/ dom(∆Fri ).
However, if λ0 < λ1 in [0, ∞) then
ˆ λ1
fλ (·) dλ ∈ dom(∆Fri ) (10.40)
λ0
and so every fλ , λ ∈ [0, ∞), is a generalized eigenfunction, i.e., the spectrum of ∆Fri
is purely continuous with Lebesgue measure, and multiplicity one.
The verification of (10.40) follows from eq. (10.38), i.e.,

1+Q λ 1
fλ (n + 1) = − n fλ (n) − f (n − 1) . (10.41)
Q Q Q λ
Set ˆ λ1
F[λ0 ,λ1 ] := fλ (·) dλ . (10.42)
λ0
Then by (10.41) and (10.42),

ˆ λ1
1+Q 1 1
F[λ0 ,λ1 ] (n + 1) = F (n) − n λ fλ (n) dλ − F (n − 1)
Q [λ0 ,λ1 ] Q λ0 Q [λ0 ,λ1 ]
´ λ1
and λ0 λ fλ dλ is computed using integration by parts.
we suggest: [Jor08, JP10, JP11a, JP11b, JP13a, JP13c, JT14b, RAKK05, Yos95].
Chapter 11
Reproducing Kernel Hilbert Space
The simplicities of natural laws arise through the complexities

of the language we use for their expression. — Eugene Wigner
... an apt comment on how science, and indeed the whole of

civilization, is a series of incremental advances, each building
on what went before. — Stephen Hawking
A special family of Hilbert spaces H are reproducing kernel Hilbert spaces (RKHSs).
We say that H is a RKHS if H is a Hilbert space of functions on some set X such
that for every x in X, the linear mapping f 7−→ f (x) is continuous in the norm of
H.
We begin with a general
Definition 11.1. Let S be a set. We say that H is a S-reproducing kernel Hilbert

space if:
RS1. H is a Hilbert space of functions on S; and

RS2. For all s ∈ S, the mapping H −→ C, by Es : h 7−→ h (s) is continuous
on H , i.e., by Riesz, there is a Ks ∈ H s.t. h (s) = hKs , hi, ∀h ∈ H .
Notation. The system of functions {Ks : s ∈ S} ⊂ H is called the associated re-

producing kernel.
Let H be a RKHS, see Definition 11.1, hence S is a set, and H is a Hilbert space
of functions on S such that RS2 holds. Fix s ∈ S, and note that Es : H −→ C is
then a bounded linear operator, where C is a 1-dimensional Hilbert space. Hence its
adjoint Es∗ : C −→ H ∗ ' H is well-defined.
Claim. The kernel Ks in Definition 11.1 is Es∗ (1) = Ks .
Proof. For λ ∈ C and h ∈ H , we have
hEs∗ (λ ) , hiH = λ Es (h) = λ h (s) = λ hKs , hiH = hλ Ks , hiH ;

(by (RS2))
and therefore Es∗ (λ ) = λ Ks as desired.
279
280 11 Reproducing Kernel Hilbert Space
In this chapter, we study two extension problems, and their interconnections.

The first class of extension problems concerns (i) positive definite (p.d.) continuous
functions on Lie groups G, and the second deals with (ii) Lie algebras of unbounded
skew-Hermitian operators in a certain family of reproducing kernel Hilbert spaces
(RKHS). The analysis is non-trivial even if G = Rn , and even if n = 1. If G = Rn , we
are concerned in (ii) with the study of systems of n skew-Hermitian operators {Si }
on a common dense domain in Hilbert space, and in deciding whether it is possible
to find a corresponding system of strongly commuting selfadjoint operators {Ti }
such that, for each value of i, the operator Ti extends Si .
From the postulates of quantum physics, we know that measurements of observ-
ables are computed from associated selfadjoint operators—observables. From the
corresponding spectral resolutions, we get probability measures, and of course un-
certainty. There are many philosophical issues (which we bypass here), and we do
not yet fully understand quantum reality. See for example, [Sla03, CJK+ 12].
The axioms are as follows: An observable is a Hermitian (selfadjoint) linear op-
erator mapping a Hilbert space, the space of states, into itself. The values obtained
in a physical measurement are, in general, described by a probability distribution;
and the distribution represents a suitable “average” (or “expectation”) in a measure-
ment of values of some quantum observable in a state of some prepared system.
The states are (up to phase) unit vectors in the Hilbert space, and a measurement
corresponds to a probability distribution (derived from a projection-valued spectral
measure). The spectral type may be continuous (such as position and momentum)
or discrete (such as spin). Information about the measures µ are computed with the
use of generating functions (on R), i.e., spectral (Bochner/Fourier) transforms of the
corresponding measure. Generating functions are positive definite continuous func-

tions F = Fµ on R. One then tries to recover µ from information about F. In this
chapter we explore the cases when information about F (x) is only available for x in
a bounded interval.
In probability theory, normalized continuous positive definite functions F, i.e.,
F(0) = 1, arise as generating functions for probability measures, and one passes
from information about one to the other; – from generating function to probability
measure is called “the inverse problem”, see e.g., [DM85]. Hence the study of par-
tially defined p.d. functions addresses the inverse question: ambiguity of measures
when only partial information for a possible generating function is available.
11.1 A Digression: Stochastic Processes 281
11.1 A Digression: Stochastic Processes
The interest in positive definite functions has at least three roots: (i) Fourier analy-
sis, and harmonic analysis more generally, including the non-commutative variant
where we study unitary representations of groups; (ii) optimization and approxi-
mation problems, involving for example spline approximations as envisioned by I.
Schöenberg; and (iii) the study of stochastic (random) processes.
A stochastic process is an indexed family of random variables based on a fixed
probability space; in our present analysis, the processes will be indexed by some
group G; for example G = R, or G = Z correspond to processes indexed by real
time, respectively discrete time. A main tool in the analysis of stochastic processes
is an associated covariance function, see (11.1).

A process Xg g ∈ G is called Gaussian if each random variable Xg is Gaus-
sian, i.e., its distribution is Gaussian. For Gaussian processes we only need two
moments. So if we normalize, setting the mean equal to 0, then the process is deter-
mined by the covariance function. In general the covariance function is a function
on G × G, or on a subset, but if the process is stationary, the covariance function will
in fact be a positive definite function defined on G, or a subset of G. We will be using
three stochastic processes in the Memoir, Brownian motion, Brownian Bridge, and
the Ornstein-Uhlenbeck process, all Gaussian, or Ito integrals.
We outline a brief sketch of these facts below.
Let G be a locally compact group, and let (Ω , F , P) be a probability space, F a
sigma-algebra, and P a probability measure defined on F . A stochastic L2 -process is
a system of random variables Xg g∈G , Xg ∈ L2 (Ω , F , P). The covariance function

cX of the process is the function G × G → C given by

cX (g1 , g2 ) = E X g1 Xg2 , ∀ (g1 , g2 ) ∈ G × G. (11.1)
´
To simplify will assume that the mean E (Xg ) = Ω Xg dP (ω) = 0 for all g ∈ G.
We say that (Xg ) is stationary iff
cX (hg1 , hg2 ) = cX (g1 , g2 ) , ∀h ∈ G. (11.2)
In this case cX is a function of g−1

1 g2 , i.e.,
E (Xg1 , Xg2 ) = cX g−1

1 g2 , ∀g1 , g2 ∈ G. (11.3)
(Just take h = g−1

1 in (11.2).)
We now recall the following theorem of Kolmogorov (see [PS75]). One direction
is easy, and the other is the deep part:
Definition 11.2. A function c defined on a subset of G is said to be positive definite

iff
g−1

∑ ∑ λi λ j c i gj ≥ 0
i j
for all finite summation, where λi ∈ C and g−1

i g j in the domain of c.
Theorem 11.1 (Kolmogorov). A function c : G → C is positive definite if and only

if there is a stationary Gaussian process (Ω , F , P, X) with mean zero, such that
c = cX .
Proof. To stress the idea, we include the easy part of the theorem, and we refer to
[PS75] for the non-trivial direction:
Let λ1 , λ2 , . . . , λn ∈ C, and {gi }Ni=1 ⊂ G, then for all finite summations, we have:
 
2
N
g−1

∑ ∑ λi λ j cX i gj = E
 ∑ λi Xgi  ≥ 0.
i j i=1
11.2 Two Extension Problems
While each of the two extension problems has received a considerable amount of
attention in the literature, our emphasis here will be the interplay between the two
problems: Our aim is a duality theory; and, in the case G = Rn , and G = Tn =
Rn /Zn , we will state our theorems in the language of Fourier duality of abelian
groups: With the time frequency duality formulation of Fourier duality for G = Rn
we have that both the time domain and the frequency domain constitute a copy
of Rn . We then arrive at a setup such that our extension questions (i) are in time
domain, and extensions from (ii) are in frequency domain. Moreover we show that
each of the extensions from (i) has a variant in (ii). Specializing to n = 1, we arrive
of a spectral theoretic characterization of all skew-Hermitian operators with dense
domain in a separable Hilbert space, having deficiency-indices (1, 1).
11.3 The Reproducing Kernel Hilbert Space HF 283
A systematic study of densely defined Hermitian operators with deficiency in-

dices (1, 1), and later (d, d), was initiated by M. Krein [Kre46], and is also part of
de Branges’ model theory; see [dB68, dBR66]. The direct connection between this
theme and the problem of extending continuous positive definite (p.d.) functions F
when they are only defined on a fixed open subset to Rn was one of our motivations.
One desires continuous p.d. extensions to Rn .
If F is given, we denote the set of such extensions Ext (F). If n = 1, Ext (F) is
always non-empty, but for n = 2, Rudin gave examples in [Rud70, Rud63] when
Ext (F) may be empty. Here we extend these results, and we also cover a number of
classes of positive definite functions on locally compact groups in general; so cases
when Rn is replaced with other groups, both Abelian and non-abelian.
The results in the framework of locally compact Abelian groups are more com-
plete than their counterparts for non-Abelian Lie groups, one reason is the availabil-
ity of Bochner’s duality theorem for locally compact Abelian groups; – not available
for non-Abelian Lie groups.
11.3 The Reproducing Kernel Hilbert Space HF
Reproducing kernel Hilbert spaces were pioneered by Aronszajn in the 1950s

[Aro50]; and subsequently they have been used in a host of applications; e.g.,
[SZ09, SZ07].
As for positive definite functions, their use and applications are extensive and
includes such areas as stochastic processes, see e.g., [JP13b, AJSV13, JP12, AJ12];
harmonic analysis (see [JÓ00]) , and the references there); potential theory [Fug74,
KL14]; operators in Hilbert space [Alp92, AD86]; and spectral theory [AH13,
Nus75, Dev72, Dev59]. We stress that the literature is vast, and the above list is
only a small sample.
Associated to a pair (Ω , F), where F is a prescribed continuous positive definite
function defined on Ω , we outline a reproducing kernel Hilbert space HF which
will serve as a key tool in our analysis. The particular RKHSs we need here will
have additional properties (as compared to a general framework); which allow us to
give explicit formulas for our solutions.
Definition 11.3. Let G be a Lie group. Fix Ω ⊂ G, non-empty, open and connected.
A continuous function
F : Ω −1 · Ω → C (11.4)
is positive definite (p.d.) if
xi−1 x j ≥ 0,

∑ ∑ ci c j F (11.5)
i j
for all finite systems {ci } ⊂ C, and points {xi } ⊂ Ω .

Equivalently, ˆ ˆ
ϕ (x)ϕ (y) F x−1 y dxdy ≥ 0,

(11.6)
Ω Ω
for all ϕ ∈ Cc (Ω ); where dx denotes a choice of left-invariant Haar measure on G.
For simplicity we focus on the case G = R, indicating the changes needed for gen-
eral Lie groups.
Definition 11.4. Fix 0 < a < ∞, set Ω := (0, a). Let F : Ω − Ω → C be a continuous
p.d. function. The reproducing kernel Hilbert space (RKHS), HF , is the completion
of
∑ c j F (· − x j ) : c j ∈ C (11.7)
finite
with respect to the inner product
hF (· − x) , F (· − y)iHF = F (x − y) , ∀x, y ∈ Ω , and
∑ ci F (· − xi ) , ∑ c j F (· − x j ) HF
= ∑ ∑ ci c j F (xi − x j ) , (11.8)
i j i j
Remark 11.1. Throughout, we use the convention that the inner product is conjugate
linear in the first variable, and linear in the second variable. When more than one
inner product is used, subscripts will make reference to the Hilbert space.
Notation. Inner product and norms will be denoted h·, ·i, and k·k respectively.
Often more than one inner product is involved, and subscripts are used for identifi-
cation.
Lemma 11.1. The reproducing kernel Hilbert space (RKHS), HF , is the Hilbert
completion of the functions
ˆ
Fϕ (x) = ϕ (y) F (x − y) dy, ∀ϕ ∈ Cc∞ (Ω ) , x ∈ Ω (11.9)
Ω
with respect to the inner product

ˆ ˆ
Fϕ , Fψ HF
= ϕ (x)ψ (y) F (x − y) dxdy, ∀ϕ, ψ ∈ Cc∞ (Ω ) . (11.10)
Ω Ω
In particular,
ˆ ˆ
2
Fϕ HF
= ϕ (x)ϕ (y) F (x − y) dxdy, ∀ϕ ∈ Cc∞ (Ω ) (11.11)
Ω Ω
and ˆ
Fϕ , Fψ HF
= ϕ (x)Fψ (x) dx, ∀φ , ψ ∈ Cc∞ (Ω ). (11.12)
Ω
Proof. Apply standard approximation, see lemma 11.2 below.
The remaining of this section is devoted to a number of technical lemmas which

will be used throughout the paper. Given a locally defined continuous positive def-
inite function F, the issues addressed below are: approximation (lemma 11.2), a
reproducing kernel Hilbert space (RKHS) HF built from F, an integral transform,
and a certain derivative operator D(F) , generally unbounded in the RKHS HF . We
will be concerned with boundary value problems for D(F) , and in order to produce
suitable orthonormal bases in HF , we be concerned with an explicit family of skew-
adjoint extensions of D(F) , as well as the associated spectra, see Corollaries 11.3 and
11.4.
Lemma 11.2. Let ϕ be a function s.t.
1. supp (ϕ) ⊂ (0, a);

2. ϕ ∈ Cc∞ (0, a), ϕ ≥ 0;
á
3. 0 ϕ (t) dt = 1.
Fix x ∈ (0, a), and set ϕn,x (t) := nϕ (n (t − x)). Then limn→∞ ϕn,x = δx , i.e., the
Dirac measure at x; and
Fϕn,x − F (· − x) HF
→ 0, as n → ∞. (11.13)
Hence Fϕ : ϕ ∈ Cc∞ (0, a) spans a dense subspace in HF . See figure 11.1.

j n,x
0 x a
Fig. 11.1: The approximate identity ϕn,x

Recall, the following facts about HF , which follow from the general theory
[Aro50] of RKHS:
• F(0) > 0, so we can always arrange F(0) = 1.

• F(−x) = F(x)
• HF consists of continuous functions ξ : Ω − Ω → C.
• The reproducing property:
hF (· − x) , ξ iHF = ξ (x) , ∀ξ ∈ HF , ∀x ∈ Ω ,
is a direct consequence of (11.8).
Remark 11.2. It follows from the reproducing property that if Fφn → ξ in HF , then
Fφn converges uniformly to ξ in Ω . In fact
Fφn (x) − ξ (x) = F (· − x) , Fφn − ξ HF
≤ kF (· − x)kHF Fφn − ξ HF
= F (0) Fφn − ξ HF
.
Lemma 11.3. Let F : (−a, a) → C be a continuous and p.d. function, and let HF
be the corresponding RKHS. Then:
á
1. the integral Fϕ := 0 ϕ (y) F (· − y) dy is convergent in HF for all ϕ ∈
Cc (0, a); and
2. for all ξ ∈ HF , we have:
ˆ a
Fϕ , ξ HF
= ϕ (x)ξ (x) dx. (11.14)
0
Proof. For simplicity, we assume the following normalization F (0) = 1; then for
all y1 , y2 ∈ (0, 1), we have
kF (· − y1 ) − F (· − y2 )k2HF = 2 (1 − ℜ {F (y1 − y2 )}) . (11.15)
Now, view the integral in (1) as a HF -vector valued integral. If ϕ ∈ Cc (0, a), this
á
integral 0 ϕ (y) F (· − y) dy is the HF -norm convergent. Since HF is a RKHS,
h·, ξ iHF is continuous on HF , and it passes under the integral in (1). Using
hF (y − ·) , ξ iHF = ξ (y) (11.16)

the desired conclusion (11.14) follows.
Corollary 11.1. Let F : (−a, a) → C be as above, and let HF be the corresponding

RKHS. For ϕ ∈ Cc1 (0, a), set
ˆ a
Fϕ (x) = (TF ϕ) (x) = ϕ (y) F (x − y) dy; (11.17)
0
then Fϕ ∈ C1 (0, a), and
d
Fϕ (x) = TF ϕ 0 (x) , ∀x ∈ (0, a) .

(11.18)
dx
á
Proof. Since Fϕ (x) = 0 ϕ (y) F (x − y) dy, x ∈ (0, a); the desired assertion (11.18)
follows directly from the arguments in the proof of lemma 11.3.
Theorem 11.2. Fix 0 < a < ∞. A continuous function ξ : (0, a) → C is in HF if and

only if there exists a finite constant A > 0, such that
∑ ∑ ci c j ξ (xi )ξ (x j ) ≤ A ∑ ∑ ci c j F (xi − x j ) (11.19)

i j i j
for all finite system {ci } ⊂ C and {xi } ⊂ (0, a). Equivalently, for all ϕ ∈ Cc∞ (Ω ),
ˆ a 2 ˆ aˆ a
ϕ (y) ξ (y) dy ≤ A ϕ (x)ϕ (y) F (x − y) dxdy (11.20)
0 0 0
We will use these two conditions (11.19)(⇔(11.20)) when considering for example
the von Neumann deficiency-subspaces for skew Hermitian operators with dense
domain in HF .
2
Proof. Note, if ξ ∈ HF , the LHS in (11.20) is Fϕ , ξ HF
, and so (11.20) holds,
since h·, ξ iHF is continuous on HF .
If ξ is continuous on [0, a], and if (11.20) holds, then
ˆ a
HF 3 Fϕ 7−→ ϕ (y) ξ (y) dy
0
is well-defined, continuous, linear; and extends to HF by density (see lemma 11.2).

Hence, by Riesz’s theorem, ∃! kξ ∈ HF s.t.
ˆ a
ϕ (y) ξ (y) dy = Fϕ , kξ HF
.
0
á
But using the reproducing property in HF , and Fϕ (x) = 0 ϕ (x) F (x − y) dy, we
get ˆ ˆ
a a
ϕ (x)ξ (x) dx = ϕ (x)kξ (x) dx, ∀ϕ ∈ Cc (0, a)
0 0
so ˆ a
ϕ (x) ξ (x) − kξ (x) dx = 0, ∀ϕ ∈ Cc (0, a) ;
0
it follows that ξ − kξ = 0 on (0, a) =⇒ ξ − kx = 0 on [0, a].

Definition 11.5 (The operator DF ). Let DF Fϕ = Fϕ 0 , for all ϕ ∈ Cc∞ (0, a), where
dϕ
ϕ0 = dt and Fϕ is as in (11.9).
Lemma 11.4. The operator DF defines a skew-Hermitian operator with dense do-
main in HF .
Proof. By lemma 11.2, dom (DF ) is dense in HF . If ψ ∈ Cc∞ (0, a) and |t| <
dist (supp (ψ) , endpoints), then
ˆ aˆ a
2 2
Fψ(·+t) HF
= Fψ HF
= ψ (x)ψ (y) F (x − y) dxdy (11.21)
0 0
see (11.11), so
d 2
F HF
=0
dt ψ(·+t)
which is equivalent to
DF Fψ , Fψ HF
+ Fψ , DF Fψ HF
= 0. (11.22)
It follows that DF is well-defined and skew-Hermitian in HF .
Lemma 11.5. Let F be a positive definite function on (−a, a), 0 < a < ∞ fixed. Let
DF be as in 11.5, so that DF ⊂ D∗F (lemma 11.4), where D∗F is the adjoint relative
to the HF inner product.
Then ξ ∈ HF (as a continuous function on [0, a]) is in dom (D∗F ) iff
ξ 0 ∈ HF where ξ 0 = distribution derivative, and (11.23)

D∗F ξ = −ξ 0
(11.24)
Proof. By theorem 11.2, a fixed ξ ∈ HF , i.e., x 7→ ξ (x) is a continuous function on

á 2 2
[0, a] s.t. ∃C, 0 ϕ (x) ξ (x) dx ≤ C Fϕ H .
F
ξ is in dom (D∗F ) ⇐⇒ ∃C = Cξ < ∞ s.t.
2
ˆ aˆ a
2
DF Fϕ , ξ HF
≤ C Fϕ HF
=C ϕ (x)ϕ (y) F (x − y) dxdy (11.25)
0 0
But LHS of (11.25) under |h·, ·i|2 is:
2
ˆ a
(11.14)
DF Fϕ , ξ HF
= Fϕ 0 , ξ HF
= ϕ 0 (x)ξ (x) dx, ∀ϕ ∈ Cc∞ (0, a) (11.26)
0
So (11.25) holds ⇐⇒
ˆ a 2
2
ϕ 0 (x)ξ (x) dx ≤ C Fϕ HF
, ∀ϕ ∈ Cc∞ (0, a)
0
i.e.,
ˆ a 2
2
ϕ (x)ξ 0 (x) dx ≤ C Fϕ HF
, ∀ϕ ∈ Cc∞ (0, a) , and
0
ξ 0 as a distribution is in HF , and
ˆ a
ϕ (x)ξ 0 (x) dx = Fϕ , ξ 0 HF
0
where we use the characterization of HF in (11.20), i.e., a function η : [0, a] → C

á
is in HF ⇐⇒ ∃C < ∞, 0 ϕ (x)η (x) dx ≤ C Fϕ H , ∀ϕ ∈ Cc∞ (0, a), and then
á F
0 ϕ (x)η (x) dx = Fϕ , η HF
, ∀ϕ ∈ Cc
∞ (0, a). See theorem 11.2.
∗
Corollary 11.2. h ∈ HF is in dom D2F iff h00 ∈ HF (h00 distribution derivative)

∗
and (D∗F )2 h = D2F h = h00 .

d 2
Proof. Application of (11.26) to DF Fϕ = Fϕ 0 , we have D2F Fϕ = Fϕ 00 = dx

Fϕ ,
∀ϕ ∈ Cc∞ (0, a), and
ˆ a
D2F Fϕ , h

HF
= Fϕ 00 , h H = ϕ 00 (x)h (x) dx
F
ˆ a 0
D ∗ E
= ϕ (x)h00 (x) dx = Fϕ , D2F h .
0 HF

Definition 11.6. [DS88c]Let D∗F be the adjoint of DF relative to HF inner product.
The deficiency spaces DEF ± consists of ξ± ∈ dom (D∗F ), such that D∗F ξ± = ±ξ± ,
i.e.,
n o
DEF ± = ξ± ∈ HF : Fψ 0 , ξ± HF
= Fψ , ±ξ± HF
, ∀ψ ∈ Cc∞ (Ω ) .
Corollary 11.3. If ξ ∈ DEF ± then ξ (x) = constant e∓x .
Proof. Immediate from lemma 11.5.
The role of deficiency indices for the canonical skew-Hermitian operator DF

(11.5) in the RKHS HF is as follows: using von Neumann’s conjugation trick
[DS88c], we see that the deficiency indices can be only (0, 0) or (1, 1).
We conclude that there exists proper skew-adjoint extensions A ⊃ DF in HF (in
case DF has indices (1, 1)). Then
DF ⊆ A = −A∗ ⊆ −D∗F (11.27)
(If the indices are (0, 0) then DF = −D∗F ; see [DS88c].)

Hence, set U (t) = etA : HF → HF , and get the strongly continuous unitary one-
parameter group
{U (t) : t ∈ R} , U (s + t) = U (s)U (t) , ∀s,t ∈ R;
and if
U (t) ξ − ξ
ξ ∈ dom (A) = ξ ∈ HF : s.t. lim exists
t→0 t
then
U (t) ξ − ξ
Aξ = s.t. lim . (11.28)
t→0 t
Now use Fx (·) = F (x − ·) defined in (0, a); and set
FA (t) := hF0 ,U (t) F0 iHF , ∀t ∈ R (11.29)
then using (11.13), we see that FA is a continuous positive definite extension of F

on (−a, a). This extension is in Ext1 (F).
Corollary 11.4. Assume λ ∈ R is in the point spectrum of A, i.e., ∃ξλ ∈ dom (A),
ξλ 6= 0, s.t. Aξλ = iλ ξλ holds in HF , then ξλ = const · eλ , i.e.,
ξλ (x) = const · eiλ x , ∀x ∈ [0, a] . (11.30)
Proof. Assume λ is in spec pt (A), and ξλ ∈ dom (A) satisfying
(Aξλ ) (x) = iλ ξλ (x) in HF , (11.31)

then since A ⊂ −D∗F , we get ξ ∈ dom (D∗F ) by lemma 11.5 and (11.27), and D∗F ξλ =
−ξλ0 where ξ 0 is the distribution derivative (see (11.24)); and by (11.27)
(11.31)
(Aξλ ) (x) = − (D∗F ξλ ) (x) = ξλ0 (x) = iλ ξλ (x) , ∀x ∈ (0, a) (11.32)
so ξλ is the distribution derivative solution to
ξλ0 (x) = iλ ξλ (x) (11.33)

m
ˆ a ˆ a
− ϕ 0 (x)ξλ (x) dx = iλ ϕ (x)ξλ (x) dx, ∀ϕ ∈ Cc∞ (0, a)
0 0
m

− DF Fϕ , ξλ HF
= iλ Fϕ , ξλ HF
, ∀ϕ ∈ Cc∞ (0, a) .
But by Schwartz, the distribution solutions to (11.33) are ξλ (x) = const · eλ (x) =
const · eiλ x .
In the considerations below, we shall be primarily concerned with the case when
a fixed continuous p.d. function F is defined on a finite interval (−a, a) ⊂ R. In this
case, by a Mercer operator, we mean an operator TF in L2 (0, a) where L2 (0, a) is
defined from Lebesgue measure on (0, a), given by
ˆ a
(TF ϕ) (x) := ϕ (y) F (x − y) dy, ∀ϕ ∈ L2 (0, a) , ∀x ∈ (0, a) . (11.34)
0
Lemma 11.6. Under the assumptions stated above, the Mercer operator TF is trace
class in L2 (0, a); and if F (0) = 1, then
trace (TF ) = a. (11.35)
Proof. This is an application of Mercer’s theorem [LP89, FR42, FM13] to the inte-
gral operator TF in (11.34). But we must check that F, on (−a, a), extends uniquely
by limit to a continuous p.d. function Fex on [−a, a], the closed interval. This is true,
and easy to verify, see e.g. [JPT14].
Corollary 11.5. Let F and (−a, a) be as in lemma 11.6. Then there is a sequence
(λn )n∈N , λn > 0, s.t. ∑n∈N λn = a, and a system of orthogonal functions {ξn } ⊂
L2 (0, a) ∩ HF such that
F (x − y) = ∑ λn ξn (x) ξn (y), and (11.36)

n∈N
ˆ a
ξn (x)ξm (x) dx = δn,m , n, m ∈ N. (11.37)
0
Proof. An application of Mercer’s theorem [LP89, FR42, FM13].
Corollary 11.6. For all ψ, ϕ ∈ Cc∞ (0, a), we have
Fψ , Fϕ HF
= Fψ , TF−1 Fϕ 2
. (11.38)
Consequently,
−1/2
khkHF = kTF hk2 , ∀h ∈ HF . (11.39)
Proof. Note
Fψ , TF−1 Fϕ 2
= Fψ , TF−1 TF ϕ 2 = Fψ , ϕ 2
ˆ a ˆ a
= ψ (x) F (y − x) dx ϕ (y) dy
ˆ0 a ˆ a0
= ψ (x)ϕ (y) F (x − y) dxdy = Fψ , Fϕ HF
.
0 0
Corollary 11.7. Let {ξn } be the ONB in L2 (0, a) as in 11.5; then

√
λn ξn is an
ONB in HF .
Proof. The functions ξn are in HF by theorem 11.2. We check directly (11.6) that
Dp p E p
λn ξn , λm ξm = λn λm ξn , T −1 ξm 2
HF
p
= λn λm λm−1 hξn , ξm i2 = δn,m .
11.4 Type I v.s. Type II Extensions
When a pair (Ω , F) is given, where F is a prescribed continuous positive definite

function defined on Ω , we consider the possible continuous positive definite exten-
sions to all of Rn . The reproducing kernel Hilbert space HF will play a key role in
11.4 Type I v.s. Type II Extensions 293
our analysis. In constructing various classes of continuous positive definite exten-

sions to Rn , we introduce operators in HF , and their dilation to operators, possibly
acting in an enlargement Hilbert space [JPT14, KL14]. Following techniques from
dilation theory we note that every dilation contains a minimal one. If a continuous
positive definite extensions to Rn has its minimal dilation Hilbert space equal to
HF , we say it is type 1, otherwise we say it is type 2.
Definition 11.7. Let G be a locally compact group, and let Ω be an open connected
subset of G. Let F : Ω −1 · Ω → C be a continuous positive definite function.
Consider a strongly continuous unitary representation U of G acting in some
Hilbert space K , containing the RKHS HF . We say that (U, K ) ∈ Ext (F) iff there
is a vector k0 ∈ K such that
F (g) = hk0 ,U (g) k0 iK , ∀g ∈ Ω −1 · Ω . (11.40)
1. The subset of Ext (F) consisting of (U, HF , k0 = Fe ) with
F (g) = hFe ,U (g) Fe iHF , ∀g ∈ Ω −1 · Ω (11.41)
is denoted Ext1 (F); and we set
Ext2 (F) := Ext (F) \Ext1 (F) ;
i.e., Ext2 (F), consists of the solutions to problem (11.40) for which K %
HF , i.e., unitary representations realized in an enlargement Hilbert space.
(We write Fe ∈ HF for the vector satisfying hFe , ξ iHF = ξ (e), ∀ξ ∈ HF ,
where e is the neutral (unit) element in G, i.e., e g = g, ∀g ∈ G.)
2. In the special case, where G = Rn , and Ω ⊂ Rn is open and connected, we
consider
F : Ω −Ω → C
continuous and positive definite. In this case,

n ˆ
Ext (F) = µ ∈ M+ (Rn ) µ b (x) = eiλ ·x dµ (λ ) (11.42)
n
oR
is a p.d. extensiont of F .
Remark 11.3. Note that (11.42) is consistent with (11.40): For if (U, K , k0 ) is a uni-
tary representation of G = Rn , such that (11.40) holds; then, by a theorem of Stone,
there is a projection-valued measure (PVM) PU (·), defined on the Borel subsets of

Rn s.t. ˆ
U (x) = eiλ ·x PU (dλ ) , x ∈ Rn . (11.43)
Rn
Setting
dµ (λ ) := kPU (dλ ) k0 k2K , (11.44)
it is then immediate that we have: µ ∈ M+ (Rn ), and that the finite measure µ
satisfies
b (x) = F (x) , ∀x ∈ Ω − Ω .
µ (11.45)
Set n = 1: Start with a local p.d. continuous function F, and let HF be the corre-
sponding RKHS. Let Ext(F) be the compact convex set of probability measures on
R defining extensions of F.
We now divide Ext(F) into two parts, say Ext1 (F) and Ext2 (F).
All continuous p.d. extensions of F come from strongly continuous unitary rep-
resentations. So in the case of 1D, from unitary one-parameter groups of course, say
U(t).
Let Ext1 (F) be the subset of Ext(F) corresponding to extensions when the uni-
tary representation U(t) acts in HF (internal extensions), and Ext2 (F) denote the
part of Ext(F) associated to unitary representations U(t) acting in a proper enlarge-
ment Hilbert space K (if any), i.e., acting in a Hilbert space K corresponding to a
proper dilation of HF .
11.5 The Case of e−|x| , |x| < 1
Our emphasis is von Neumann indices, and explicit formulas for partially defined
positive definite functions F, defined initially only on a symmetric interval (−a, a).
Among the cases of partially defined positive definite functions, the following ex-
ample F (x) = e−|x| , in the symmetric interval (−1, 1), will play a special role. The
present section is devoted to this example.
There are many reasons for this:
(i) It is of independent interest, and its type 1 extensions (see section §11.4)
can be written down explicitly.
(ii) Its applications include stochastic analysis [Itô06] as follows. Given a
random variable X in a process; if µ is its distribution, then there are two
11.5 The Case of e−|x| , |x| < 1 295
measures of concentration for µ, one called “degree of concentration,”

and the other “dispersion,” both computed directly from F (x) = e−|x|
applied to µ.
(iii) In addition, there are analogous relative notions for comparing different
samples in a fixed stochastic process. These notions are defined with the
use of example F (x) = e−|x| , and it will frequently be useful to localize
the x-variable in a compact interval.
(iv) Additional reasons for special attention to example F (x) = e−|x| , for
x ∈ (−1, 1) is its use in sampling theory, and analysis of de Branges
spaces [DM85], as well as its role as a Greens function for an important
boundary value problem.
(v) Related to this, the reproducing kernel Hilbert space HF associated to
this p.d. function F has a number of properties that also hold for wider
families of locally defined positive definite function of a single variable.
In particular, HF has Fourier bases: The RKHS HF has orthogonal
bases of complex exponentials eλ with aperiodic frequency distribu-
tions, i.e., frequency points {eλ } on the real line which do not lie on
any arithmetic progression, see Fig 11.3. For details on this last point,
see Corollaries 11.13, 11.14, 11.15, and 11.18.
11.5.1 The selfadjoint Extensions Aθ ⊃ −iDF
The notation “⊇” above refers to containment of operators, or rather of the respec-
tive graphs of the two operators; see [DS88c].
Lemma 11.7. Let F (x) = e−|x| , |x| < 1. Set Fx (y) := F (x − y), ∀x, y ∈ (0, 1); and
´1
Fϕ (x) = 0 ϕ (y) F (x − y) dy, ∀ϕ ∈ Cc∞ (0, 1). Define DF Fϕ = Fϕ 0 on the dense
subset
dom (DF ) = Fϕ : ϕ ∈ Cc∞ (0, 1) ⊂ HF .

(11.46)
Then the skew-Hermitian operator DF has deficiency indices (1, 1) in HF , where

the defect vectors are
ξ+ (x) = F0 (x) = e−x (11.47)

ξ− (x) = F1 (x) = ex−1 ; (11.48)
moreover,
kξ+ kHF = kξ+ kHF = 1. (11.49)
Proof. (Note if Ω is any bounded, open and connected domain in Rn , then a locally
defined continuous p.d. function, F : Ω −Ω :→ C, extends uniquely to the boundary
∂ Ω := Ω \Ω by continuity [JPT14].)
In our current settings, Ω = (0, 1), and Fx (y) := F (x − y), ∀x, y ∈ (0, 1). Thus,
Fx (y) extends to all x, y ∈ [0, 1]. In particular,
F0 (x) = e−x , F1 (x) = ex−1
are the two defect vectors, as shown in 11.3. Moreover, using the reproducing prop-
erty, we have
kF0 k2HF = hF0 , F0 iHF = F0 (0) = F (0) = 1

kF1 k2HF = hF1 , F1 iHF = F1 (1) = F (0) = 1
and (11.49) follows. For more details, see [JPT14, lemma 2.10.14].
Lemma 11.8. Let F be any continuous p.d. function on (−1, 1). Set
ˆ 1
h (x) = ϕ (y) F (x − y) dy, ∀ϕ ∈ Cc∞ (0, 1) ;
0
then
ˆ 1 ˆ 1
h (0) = ϕ (y) F (−y) dy, h (1) = ϕ (y) F (1 − y) dy (11.50)
0 0
ˆ 1 ˆ 1
h0 (0) = ϕ (y) F 0 (−y) dy, h0 (1) = ϕ (y) F 0 (1 − y) dy; (11.51)
0 0
where the derivatives F 0 in (11.50)-(11.51) are in the sense of distribution.
Proof. Note that

ˆ x ˆ 1
h (x) = ϕ (y) F (x − y) dy + ϕ (y) F (x − y) dy;
0 x
ˆ x ˆ 1
h0 (x) = ϕ (y) F 0 (x − y) dy + ϕ (y) F 0 (x − y) dy.
0 x
and so (11.50)-(11.51) follow.

11.5 The Case of e−|x| , |x| < 1 297
We now specialize to the function F (x) = e−|x| defined in (−1, 1).
Corollary 11.8. For F (x) = e−|x| , |x| < 1, set h = TF ϕ, i.e.,

ˆ 1
h := Fϕ = ϕ (y) F (· − y) dy, ∀ϕ ∈ Cc∞ (0, 1) ;
0
then
ˆ 1 ˆ 1
h (0) = ϕ (y) e−y dy, h (1) = ϕ (y) ey−1 dy (11.52)
0 0
ˆ 1 ˆ 1
h0 (0) = ϕ (y) e−y dy, h0 (1) = − ϕ (y) ey−1 dy (11.53)
0 0
In particular,
h (0) − h0 (0) = 0 (11.54)

h (1) + h0 (1) = 0. (11.55)
Proof. Immediately from lemma 11.8. Specifically,

ˆ x ˆ 1
h (x) = e−x ϕ (y) ey dy + ex ϕ (y) e−y dy
0 x
ˆ x ˆ 1
0 −x
h (x) = −e y
ϕ (y) e dy + ex
ϕ (y) e−y dy.
0 x
Setting x = 0 and x = 1 gives the desired conclusions.
Remark 11.4. The space

n o
h ∈ HF h (0) − h0 (0) = 0, h (1) + h0 (1) = 0
is dense in HF . This is because it contains Fϕ ϕ ∈ Cc∞ (0, 1) . Note

F0 + F00 = −δ0 , and

F1 − F10 = −δ1 ;
/ HF .
however, δ0 , δ1 ∈
By von Neumann’s theory [DS88c] and lemma 11.5, the family of selfadjoint ex-
tensions of the Hermitian operator −iDF is characterized by

Aθ h + c e−x + eiθ ex−1 = −i h0 + i c e−x − eiθ ex−1 , where
n o (11.56)
dom (Aθ ) := h + c e−x + eiθ ex−1 h ∈ dom (DF ) , c ∈ C .
Remark 11.5. In (11.56), h ∈ dom (DF ) (see (11.46)), and by 11.8, h satisfies the
boundary conditions (11.54)-(11.55). Also, by lemma 11.7, ξ+ = F0 = e−x , ξ− =
F1 = ex−1 , and kξ+ kHF = kξ− kHF = 1.
Proposition 11.1. Let Aθ be a selfadjoint extension of −iD as in (11.56). Then,
ψ (1) + ψ 0 (1) = eiθ ψ (0) − ψ 0 (0) , ∀ψ ∈ dom (Aθ ) .

(11.57)
Proof. Any ψ ∈ dom (Aθ ) has the decomposition

ψ (x) = h (x) + c e−x + eiθ ex−1
where h ∈ dom (DF ), and c ∈ C. An application of 11.4 gives

ψ (1) + ψ 0 (1) = h (1) + h0 (1) + c e−1 + eiθ + c −e−1 + eiθ = 2c eiθ
| {z }
=0 (by (11.55))

ψ (0) − ψ 0 (0) = h (0) − h0 (0) + c 1 + e−1 eiθ − c −1 + e−1 eiθ = 2c
| {z }
=0 (by (11.54))
which is the assertion in (11.57).
Corollary 11.9. Let Aθ be a selfadjoint extension of −iDF as in (11.56). Fix λ ∈ R,

then λ ∈ spec pt (Aθ ) ⇐⇒ eλ (x) := eiλ x ∈ dom (Aθ ), and λ is a solution to the
following equation:

−1 2λ
λ = θ + tan + 2nπ, n ∈ Z. (11.58)
λ2 −1
Proof. By assumption, eiλ x ∈ dom (Aθ ), so ∃hλ ∈ dom (DF ), and ∃cλ ∈ C s.t.

eiλ x = hλ (x) + cλ ex + eiθ ex−1 . (11.59)
Applying the boundary condition in 11.1, we have
eiλ + iλ eiλ = eiθ (1 − iλ ) ; i.e.,

11.5 The Case of e−|x| , |x| < 1 299
1 − iλ 1−iλ
eiλ = eiθ = eiθ ei arg( 1+iλ ) (11.60)
1 + iλ
where
1 − iλ 2λ
arg = tan−1
1 + iλ λ2 −1
and (11.58) follows. For a discrete set of solutions, see figure 11.2.
-6 Π -4 Π -2 Π 2Π 4Π 6Π
-Π
Fig. 11.2: Fix θ = 0.8, Λθ = {λn (θ )} = intersections of two curves.
Corollary 11.10. If Aθ ⊃ −iDF is a selfadjoint extension in HF , then
spect (Aθ ) = λ ∈ R eλ ∈ HF satisfying (11.57)

= λ ∈ R eλ ∈ HF , eλ = hλ + cλ ex + eiθ ex−1 ,

hλ ∈ dom (DF ) , cλ ∈ C .
Remark 11.6. The corollary holds for all continuous p.d. functions F : (−a, a) → C.
Corollary 11.11. All selfadjoint extensions Aθ ⊃ −iDF have purely atomic spec-
trum; i.e.,
Λθ := spect (Aθ ) = discrete subset in R. (11.61)
And for all λ ∈ Λθ ,
ker (Aθ − λ IHF ) = Ceλ , where eλ (x) = eiλ x (11.62)
i.e., all eigenvalues have multiplicity 1. (The set Λθ will be denoted {λn (θ )}n∈Z
following Fig. 11.2. )
Proof. This follows by solving eq. (11.58).

Corollary 11.12. Let A be a selfadjoint extension of −iDF as before. Suppose

λ1 , λ2 ∈ spec (A), λ1 6= λ2 , then eλi ∈ HF , i = 1, 2; and eλ1 , eλ2 HF
= 0.
Proof. Let λ1 , λ2 be as in the statement, then
(λ1 − λ2 ) eλ1 , eλ2 HF

= Aeλ1 , eλ2 HF
− eλ1 , Aeλ2 HF
= 0;
so since λ1 − λ2 6= 0, we get eλ1 , eλ2 HF

= 0.
For explicit computations regarding these points, see also Corollaries 11.16,
11.17, and 11.18 below.
11.5.2 The Spectra of the s.a. Extensions Aθ ⊃ −iDF
Let F (x) = e−|x| , |x| < 1. Define DF Fϕ = Fϕ 0 as before, where

ˆ 1
Fϕ (x) = ϕ (y) F (x − y) dy
0
ˆ 1
= ϕ (y) e−|x−y| dy, ∀ϕ ∈ Cc∞ (0, 1) .
0
And let HF be the RKHS of F.
Lemma 11.9. For all ϕ ∈ Cc∞ (0, 1), and all h, h00 ∈ HF , we have
= Fϕ , 21 h − h00 − 12 [W ]10

Fϕ , h HF 2
(11.63)
where " #
h Fϕ
W = det . (11.64)
h0 Fϕ 0
Setting l := Fϕ , we have
[W ]10 = −l (1) h (1) + h0 (1) − l (0) h (0) − h0 (0) .

(11.65)
Proof. Note
ˆ 1
Fϕ , h HF
= ϕ (x) h (x) dx (reproducing property)
0
D E
d 2
= 12 I − dx Fϕ , h
2
11.5 The Case of e−|x| , |x| < 1 301
= Fϕ , 12 h − h00 − 21 [W ]10 .

2
Set l := Fϕ ∈ HF , ϕ ∈ Cc∞ (0, 1). Recall the boundary condition in 11.8:
l (0) − l 0 (0) = l (1) + l 0 (1) = 0.
Then
[W ]10 = l 0 h − lh0 (1) − l 0 h − lh0 (0)

= −l (1) h (1) − l (1) h0 (1) − l (0) h (0) + l (0) h0 (0)

= −l (1) h (1) + h0 (1) − l (0) h (0) − h0 (0)

which is (11.65).
Corollary 11.13. eλ ∈ HF , ∀λ ∈ R.
Proof. By theorem 11.2, we need the following estimate: ∃C < ∞ s.t.

ˆ 1 2
2
ϕ (x) eλ (x) dx ≤ C Fϕ HF
. (11.66)
0
But
ˆ 1 D E
1 d 2

ϕ (x) eλ (x) dx = 2 I− dx Fϕ , eλ
2
0
00 1
= Fϕ , 12 eλ − eλ − [W ]10

2 2
= 21 1 + λ 2 Fϕ , eλ − 12 −l (1) (1 + iλ ) eiλ − l (0) (1 − iλ ) ;

2
see (11.63)-(11.65). Here, l := Fϕ .

It suffices to show
(i) ∃C1 < ∞ s.t.
2
|l (0)|2 and |l (1)|2 ≤ C1 Fϕ HF
.
(ii) ∃C2 < ∞ s.t.

2 2
Fϕ , eλ 2
≤ C2 Fϕ HF
.
For (i), note that
|l (0)| = hF0 , liHF ≤ kF0 kHF klkHF = kF0 kHF Fϕ HF

|l (1)| = hF1 , liHF ≤ kF1 kHF klkHF = kF1 kHF Fϕ HF
and we have
kF0 kHF = kF1 kHF = 1

2
klk2HF = Fϕ HF
= kTF ϕk22 ≤ λ12 kϕk22 < ∞
where λ1 is the top eigenvalue of the Mercer operator TF (lemma 11.6).

For (ii),
2
Fϕ , eλ 2
= |hTF ϕ, eλ i2 |2
D
1/2 1/2
E 2
= TF ϕ, TF eλ
2
1/2 2 1/2 2
≤ TF ϕ TF eλ (by Cauchy-Schwarz)
2 2
1/2 2
= hϕ, TF ϕi2 TF eλ
2
2 2 2
≤ Fϕ H keλ k2 = Fϕ H ;
F F
1/2 2
where we used the fact that TF eλ ≤ λ1 keλ k22 ≤ 1, since λ1 < 1 = the right
2
endpoint of the interval [0, 1] (see lemma 11.6), and keλ k2 = 1.
Therefore, the corollary follows.
Corollary 11.14. For all λ ∈ R, and all Fϕ , ϕ ∈ Cc∞ (0, 1), we have
1
1 + λ 2 Fϕ , eλ 2

Fϕ , eλ HF
= 2 (11.67)

+ 12 l (1) (1 + iλ ) eiλ + l (0) (1 − iλ ) .
Proof. By lemma 11.9,
= Fϕ , 12 eλ − e00λ − 12 [W ]10 .

Fϕ , eλ HF 2
where
1
eλ − e00λ = 1
1 + λ 2 eλ ; and

2 2
(11.65)
[W ]10 = −l (1) (1 + iλ ) eiλ − l (0) (1 − iλ ) , l := Fϕ .
Lemma 11.10. For all Fϕ , ϕ ∈ Cc∞ (0, 1), and all λ ∈ R,

11.5 The Case of e−|x| , |x| < 1 303
Fϕ , eλ HF
= hϕ, eλ i2 . (11.68)
In particular, set λ = 0, we get

ˆ 1 ˆ 1
1
Fϕ − Fϕ00 (x) dx

Fϕ , 1 HF
= ϕ (x) dx =
0 2 0
1
= Fϕ , 1 2
− Fϕ00 , 1 2
2
≤ C Fϕ H
Proof. Eq. (11.68) follows from basic fact of the Mercer operator. See Lemma 11.6
and its corollaries. It suffices to note the following estimate:
ˆ 1
Fϕ00 (x) dx = F 0 ϕ (1) − Fϕ0 (0)
0
ˆ 1 ˆ 1
= −e−1 ey ϕ (y) dy − e−y ϕ (y) dy
0 0
= −Fϕ (1) − Fϕ (0) ≤ 2 Fϕ H
.
Corollary 11.15. For all λ ∈ R,
λ2 +3
heλ , eλ iHF = . (11.69)
2
Proof. By 11.14, we see that
1
1 + λ 2 Fϕ , eλ 2

Fϕ , eλ HF
=
2
1
+ l (1) (1 + iλ ) eiλ + l (0) (1 − iλ ) ; l := Fϕ . (11.70)
2
Since Fϕ : ϕ ∈ Cc∞ (0, 1) is dense in HF , ∃Fϕn → eλ in HF , so that

Fϕn , eλ HF
→ heλ , eλ iHF
1 1 −iλ
= 1+λ2 + e (1 + iλ ) eiλ + (1 − iλ )
2 2
1 λ2 +3
= 1+λ2 +1 =

.
2 2
The approximation is justified since all the terms in the RHS of (11.70) satisfy
2
the estimate |· · · |2 ≤ C Fϕ H . See the proof of 11.13 for details.
F
Note lemma 11.9 is equivalent to the following:
Corollary 11.16. For all h ∈ HF , and all k ∈ dom TF−1 = Fϕ : ϕ ∈ Cc∞ (0, 1) ,

we have
1 1
hh, kiH = hh, ki0 + h0 , k0 0 + h (0)k (0) + h (1)k (1) (11.71)
2 2
and eq. (11.71) extends to all k ∈ HF , since dom TF−1 is dense in HF .

Example 11.1. Take h = k = eλ , λ ∈ R, then (11.71) gives
1 1 λ2 +3
heλ , eλ iH = 1 + λ 2 + (1 + 1) =
2 2 2
as in (11.69).
Corollary 11.17. Let Aθ ⊃ −iD be any selfadjoint extension in HF . If λ , µ ∈

spect (Aθ ), s.t. λ 6= µ, then eλ , eµ HF
= 0.
Proof. It follows from (11.71) that

i(µ−λ )
2 eλ , e µ H
= eλ , eµ 0
+ λ µ e λ , e µ 0
+ 1 + e

= (1 + λ µ) eλ , eµ 0 + 1 + ei(µ−λ )
ei(µ−λ ) − 1
= (1 + λ µ) + 1 + ei(µ−λ ) (11.72)
i (µ − λ )
By 11.9, eq. (11.60), we have
1 − iλ iθ 1 − iµ iθ
eiλ = e , eiµ = e
1 + iλ 1 + iµ
and so
(1 − iµ) (1 + iλ )
ei(µ−λ ) = .
(1 + iµ) (1 − iλ )
Substitute this into (11.72) yields
−2 (1 + λ µ) 2 (1 + λ µ)
2 eλ , e µ H
= + = 0.
(1 + iµ) (1 − iλ ) (1 + iµ) (1 − iλ )

11.5 The Case of e−|x| , |x| < 1 305
Corollary 11.18. Let F (x) = e−|x| , |x| < 1. Let DF Fϕ = Fϕ 0 , ∀ϕ ∈ Cc∞ (0, 1), and

Aθ ⊃ −iDF be a selfadjoint extension in HF . Set eλ (x) = eiλ x , and
Λθ := spect (Aθ ) (= discrete subset in R by Cor. 11.11) (11.73)
Then
2
Feθ (x) = ∑ e (x) , ∀x ∈ R (11.74)
λ ∈Λθ
λ2 +3 λ
is a continuous p.d. extension of F to the real line. Note that both sides in eq. (11.74)
depend on the choice of θ .
The type 1 extensions are indexed by θ ∈ [0, 2π) where Λθ is given in (11.73), see
also (11.58) in 11.9.
Corollary 11.19 (Sampling property of the set Λθ ). Let F (x) = e−|x| in |x| < 1,
HF , θ , and Λθ be as above. Let TF be the corresponding Mercer operator. Then for
all ϕ ∈ L2 (0, 1), we have
ϕb (λ ) iλ x
(TF ϕ) (x) = 2 ∑ e , for all x ∈ (0, 1) .
λ ∈Λθ
λ2 +3
Proof. This is immediate from 11.18.
Remark 11.7. Note that the system {eλ | λ ∈ Λθ } is orthogonal in HF , but not in
L2 (0, 1).
Proof. We saw that Aθ has pure atomic spectrum. By (11.15), the set
(r )
2
e : λ ∈ Λθ
λ +3 λ
2
is an ONB in HF . Hence, for F = F0 = e−|x| , we have the corresponding p.d. exten-

sion:
1
Fθ (x) = ∑ heλ , FiHF eλ (x)
λ ∈Λθ keλ k2HF
2
= ∑ e (x) , ∀x ∈ [0, 1] . (11.75)
λ ∈Λθ
λ2 +3 λ
where heλ , FiHF = eλ (0) = 1 by the reproducing property. But the RHS of (11.75)
extends to R. See figure 11.3.
1.0
0.8
0.6
0.4
0.2
-4 -2 2 4
Fig. 11.3: θ = 0. A type 1 continuous p.d. extension of F (x) = e−|x| [−1,1]

in HF .
(−1, 1), and let Ho

Corollary 11.20. Let F (x) = e−|x| innq F be the RKHS. Let θ ∈
θ is an ONB in HF .
2
[0, 2π), and let Λθ be as above; then e
λ 2 +3 λ
| λ ∈ Λ
we suggest: [AD86, JPT14, Nus75, Rud63].
Appendix A
An overview of Functional Analysis books (cast
of characters)
If people do not believe that mathematics is simple, it is only

because they do not realize how complicated life is.
— John von Neumann
Below we offer a list of related Functional Analysis books; they cover a host of
diverse areas of functional analysis and applications, some different from what we
select here: Our comments are telegraphic-review form:
Akhiezer and Glazman, “Theory of linear operators in Hilbert space” [AG93]

– a classic book set covering the detailed structure of unbounded operators and
their applications; and now in a lovely Dover edition.
Arveson, “An invitation to C∗ -algebras” [Arv76]

– an introduction to C∗ -algebras and their representations on Hilbert spaces. –
covers the most basic ideas, as simply and concretely as we could. – Hilbert spaces
are separable and C∗ -algebras are GCR. Representations are given a concrete para-
metric description, including the irreducible representations of any C∗ -algebra, even
if not GCR. For someone interested in Borel structures, see Chapter 3. Chapter 1 is
a bare-bones introduction to C∗ -algebras.
Bachman and Narici, “Functional analysis” [BN00]

The book by Bachman and Narici’s is a systematic introduction to the fundamen-
tals of functional analysis. It is easier to follow than say Rudin’s Functional Analysis
book, but it doesn’t go as far either. Rather it helps readers reinforcing topics from
real analysis and other masters level courses. It serves to bridge the gap between
more difficult treatments of functional analysis. It is excellent for self-study. As a
Dover book, it is cheap, fits a small student budget. (Dover reprints classics in a
cheap paper back format.)
307
308 A An overview of Functional Analysis books (cast of characters)
Bratteli and Robinson, “Operator algebras and quantum statistical mechanics”

[BR79, BR81]
This is a widely cited two volume book-set, covering the theory of operator al-
gebras, and its applications to quantum statistical mechanics. It is one of the more
authoritative treatments of the many exciting applications of functional analysis to
physics. Both books are self-contained; with complete proofs; – a useful text for
students with a prior exposure to basic functional analysis. One of the main themes
in v1 is decomposition theory, and the use of Choquet simplices. Example: the set
of KMS-states for a C∗ -algebraic dynamical system typically forms a Choquet sim-
plex. An introductory chapter covers algebraic techniques and their use in statistical
physics; this is followed up in v2. Indeed, a host of applications are covered in v2.
The new edition has a more comprehensive discussion of dissipative operators and
analytic elements; – and it includes a positive resolution of the question of whether
maximal orthogonal probability measure on the state space of algebra is automati-
cally maximal among all the probability measures on the space.
Conway, “A course in functional analysis” [Con90]

– a comprehensive introduction to functional analysis. The style is very formal
and rigorous (as it have to be) and you need to have a good background in mea-
sure theory and general topology (as it is mentioned in the introduction), chapters
1-4 & 6 of Folland’s real analysis book would be more than enough. – an excellent
book, – is designed to be used in grad courses. Recommend for anyone who really
want/need to learn true functional analysis for graduate work. Through its eleven
chapters, J. Conway masterfully wrote a beautiful exposition of this core subject.
Dunford and Schwartz, “Linear operators” [DS88b, DS88c, DS88a]

This classic three-volume book set, the first Functional analysis, and the sec-
ond the theory of linear operators. And for the theory of unbounded operators it is
unsurpassed. – written by two notable mathematicians, it constitutes a comprehen-
sive survey of the general theory of linear operations, and their diverse applications.
Dunford and Schwartz are influenced by von Neumann, and they emphasize the sig-
nificance of the relationships between the abstract theory and its applications. The
two first volumes are for the students. – treatment is relatively self-contained. Now
a paperback edition of the original work, unabridged, in three volumes.
Kolmogorov and Fomin, “Introductory real analysis” [KF75]

A An overview of Functional Analysis books (cast of characters) 309
This book is two books bound as one; and in the lovely format from Dover. Part
1: metric spaces, and normed linear spaces. Part 2: Lebesgue integration and ba-
sic functional analysis. Numerous examples are sprinkled through the text. To get
the most out of this book, it helps if you have already seen many of the results
presented elsewhere. History: The book came from original notes from Andrei Kol-
mogorov’s lectures given at Moscow’s Lomonosov University in the 1940’s, and it
still stands as timely introduction to real and functional analysis. Strengths: step by
step presentation of all the key concepts needed in the subject; proceeding all the
way from set theory to Fredholm integral equations. Offers a wonderful and refresh-
ing insight. Contents (sample): Elements of Set Theory; Metric and Topological
Spaces; Normed and Topological Linear Spaces; Linear Functionals and Linear Op-
erators; Elements of Differential Calculus in Linear Spaces; Measure, Measurable
Functions, Integral; Indefinite Lebesgue Integral, Differentiation Theory; Spaces of
Summable Functions; Trigonometric Series, Fourier Transformation; Linear Inte-
gral Equations.
Kadison, “Fundamentals of the theory of operator algebras” [KR97a, KR97b]

Here we cite the first two volumes in a 4-volume book-set. It begins with the
fundamentals in functional analysis, and it aims at a systematic presentations of
the main areas in the theory of operator algebras, both C∗ -algebras, von Neumann
algebras, and their applications, so including subfactors, Tomita-Takesaki theory,
spectral theory, decomposition theory, and applications to ergodic theory, represen-
tations of groups, and to mathematical physics.
Lax, “Functional analysis” [Lax02]

The subject of functional analysis, while fundamental and central in the land-
scape of mathematics, really started with seminal theorems due to Banach, Hilbert,
von Neumann, Herglotz, Hausdorff, Friedrichs, Steinhouse,...and many other of, the
perhaps less well known, founding fathers, in Central Europe (at the time), in the pe-
riod between the two World Wars. It gained from there because of its many success
stories, – in proving new theorems, in unifying old ones, in offering a framework
for quantum theory, for dynamical systems, and for partial differential equations.
The Journal of Functional Analysis, starting in the 1960ties, broadened the subject,
reaching almost all branches of science, and finding functional analytic flavor in the-
ories surprisingly far from the original roots of the subject. Peter Lax has himself,
– alone and with others, shaped some of greatest successes of the period, right up
to the present. That is in the book!! And it offers an upbeat outlook for the future.
It has been tested in the class room, – it is really user-friendly. At the end of each
chapter P Lax offers personal recollections; – little known stories of how several of
the pioneers in the subject have been victims, – in the 30ties and the 40ties, of Nazi
atrocities. The writing is crisp and engaged, – the exercises are great; – just right for
students to learn from. This is the book to teach from.
MacCluer, “Elementary functional analysis” [Mac09]

I received extremely positive student-feedback on MacCluer’s very nice book. It
covers elementary functional analysis, is great for self-study, and easy to follow. It
conveys the author’s enthusiasm for her subject. It includes apposite quotes, anec-
dotes, and historical asides, all making for a wonderful personal touch and drawing
the reader into dialogue in a palpable way. Contents: six chapters, each introduced
by a well-chosen quote, often hinting in a very useful manner at the material that
is to follow. I particularly like MacCluer’s choice of Dunford and Schwartz to start
off her third chapter: “In linear spaces with a suitable topology one encounters three
far-reaching principles concerning continuous linear transformations. . . ” We find
out quickly that these “Big Three” (as the chapter is titled) are uniform bounded-
ness, the open mapping theorem, and Hahn-Banach. MacCluer quickly goes on to
cover these three gems in a most effective and elegant manner, as well as a num-
ber of their corollaries or, in her words, “close cousins,” such as the closed graph
theorem and Banach-Steinhaus. The book takes the reader from Hilbert space pre-
liminaries to Banach- and C∗ -algebras and, to the spectral theorem.
Nelson, “Topics in Dynamics I: Flows” [Nel69]

This is a book in the Princeton Math Lecture Notes series, appearing first in 1972,
but since Prof Nelson kindly made it available on his website. In our opinion, it is the
best account of general multiplicity for normal operators, bounded and unbounded,
and for abelian *-algebras. In addition it contains a number of applications of func-
tional analysis to geometry and to physics.
Riesz et al., “Functional analysis” [RSN90]

A pioneering book in F.A., first published in the early 50s, and now in a Dover
edition, very readable. The book starts with an example of a continuous function
which is not differentiable and then proves Lebesgue’s theorem which tells you
when a function does have a derivative. The 2nd part of the book is about integral
equations which again starts with some examples of problems from the 19th century
mathematicians. The presentation of Fredholm’s method is a gem.
Rudin, “Functional analysis” [Rud73]

“Modern analysis” used to be a popular name for the subject of this lovely book.
It is as important as ever, but perhaps less “modern”. The subject of functional anal-
ysis, while fundamental and central in the landscape of mathematics, really started
with seminal theorems due to Banach, Hilbert, von Neumann, Herglotz, Hausdorff,
Friedrichs, Steinhouse,...and many other of, the perhaps less well known, found-
ing fathers, in Central Europe (at the time), in the period between the two World
Wars. In the beginning it generated awe in its ability to provide elegant proofs of
classical theorems that otherwise were thought to be both technical and difficult.
The beautiful idea that makes it all clear as daylight: Wiener’s theorem on abso-
lutely convergent (AC) Fourier series of 1/ f if you can divide, and if f has the AC
Fourier series, is a case in point. The new subject gained from there because of its
many success stories, – in proving new theorems, in unifying old ones, in offering
a framework for quantum theory, for dynamical systems, and for partial differential
equations. And offering a language that facilitated interdisciplinary work in science!
The Journal of Functional Analysis, starting in the 1960ties, broadened the subject,
reaching almost all branches of science, and finding functional analytic flavor in
theories surprisingly far from the original roots of the subject. The topics in Rudin’s
book are inspired by harmonic analysis. The later part offers one of the most elegant
compact treatment of the theory of operators in Hilbert space, I can think of. Its
approach to unbounded operators is lovely.
Sakai, “C∗ -algebras and W ∗ -algebras” [Sak71]

The presentation is succinct, theorem, proof, ... qed; but this lovely book had a
profound influence on the subject. It’s scope cover nearly all major results in the
subject up until that time. In order to accomplish this goal (without expanding into
multiple volumes), the author omits examples, motivation,. . . . It is for students who
already have an interest in operator theory. As a student, myself (PJ), I learned a lot
from this wonderful book.
Shilov, “Elementary functional analysis” [Shi96]

Elementary Functional Analysis by Georgi E. Shilov is suitable for a beginning
course in functional analysis and some of its applications, e.g., to Fourier series, to
harmonic analysis, to partial differential equations (PDEs), to Sobolev spaces, and it

is a good supplement and complement to two other popular books in the subject, one
by Rudin, and another by Edwards. Rudin’s book is entitled “Functional Analysis”
(not in the Dover series) and it is my favorite. Rudin’s book is of newer vintage, and
it goes more in depth, and includes new material on unbounded operators in Hilbert
space. Edwards’ book “Functional Analysis: Theory and Applications;” is in the
Dover series, and it is twice as thick as Shilov’s book. Topics covered in Shilov:
Function spaces, L p -spaces, Hilbert spaces, and linear operators; the standard Ba-
nach, and Hahn-Banach theorems. It includes many exercises and examples. Well
motivated with applications. Book Comparison: Shilov book is gentler on students,
and it is probably easier to get started with: It stresses motivation a bit more, the ex-
ercises are easier, and finally Shilov includes a few applications; fashionable these
days. And of course, the books in the Dover series are cheap in comparison.
Stein et al., “Functional analysis” [SS11]

This book is the fourth book in a series: Elias Stein’s and Rami Shakarchi’s
Princeton lectures in analysis. Elias Stein is a world authority on harmonic anal-
ysis. The book is of more recent vintage than the others from our present list. The
book on functional analysis is actually quite different from other texts in functional
analysis. For instance Rudin’s textbook on functional analysis has quite a different
emphasis from Stein’s. Stein devotes a whole chapter to applications of the Baire
category theory while Rudin devotes a page. Stein does this because it provides
some insights into establishing the existence of a continuous but nowhere differen-
tiable function as well as the existence of a continuous function with Fourier series
diverging a point. Thus what he is doing is providing a much more holistic and
integrated approach to the subject than occurs in other approaches. It is closest in
philosophical approach to Frigyes Riesz-Bela Sz.-Nagy). A special touch in Stein:
Inclusion of Brownian motion, and of process with independent increments, a la
Doob’s. Stein’s approach to the construction of Brownian motion is different and
closer to the approaches taken in books on financial math. Stein et al develop Brow-
nian motion in the context of solving Dirichlet’s problem.
Stone, “Linear Transformations in Hilbert Space and Their Applications to Anal-

ysis” [Sto51b]
Stone’s book is a classic, came out in 1932, and was the unique source on spec-
tral multiplicity, and a host of applications of the theory of unbounded operators to
analysis, to approximation theory, and to special functions. The last two chapters
illustrate the theory with a systematic study of (infinite × infinite) Jacobi matrices;
i.e., tri-diagonal infinite matrices; assumed formally selfadjoint (i.e., Hermitian).
Sample results: A dichotomy: Their von Neumann indices must be (0, 0) or (1, 1).
Some of the first known criteria for when they are one or the other are given; plus a
number of applications to classical analysis.
Takesaki, “Theory of operator algebras” [Tak79]

– written by one of the most prominent researchers of the area, provides an in-
troduction to this rapidly developing theory. ... These books are recommended to
every graduate student interested in this exciting branch of mathematics. Further-
more, they should be on the bookshelf of every researcher of the area.
Trèves, “Topological vector spaces, distributions and kernels” [Trè06]

Covers topological vector spaces and their applications, and it is a pioneering
book. It is antidote for those who mistakenly believe that functional analysis is about
Banach and Hilbert spaces. It’s also about Frechet spaces, LF spaces, Schwartz dis-
tributions (generalized functions), nuclear spaces, tensor products, and the Schwartz
Kernel Theorem (proved by Grothendieck). Treves’s book provides the perfect back-
ground for advanced work in linear differential, pseudodifferential, or Fourier inte-
gral operators.
Yosida “Functional analysis” [Yos95]

Yosida’s book is based on lectures given decades ago at the University of Tokyo.
It is intended as a textbook to be studied by students on their own or to be used
in a course on Functional Analysis, i.e., the general theory of linear operators in
function spaces together with salient features of its application to diverse fields
of modern and classical analysis. Necessary prerequisites for the reading of this
book are summarized, with or without proof, in Chapter 0 under titles: Set The-
ory, Topological Spaces, Measure Spaces and Linear Spaces. Then, starting with
the chapter on Semi-norms, a general theory of Banach and Hilbert spaces is pre-
sented in connection with the theory of generalized functions of S.L. SOBOLEV
and L. SCHWARTZ. While the book is primarily addressed to graduate students, it
is hoped it might prove useful to research mathematicians, both pure and applied.
The reader may pass, e.g., from Chapter IX (Analytical Theory of Semi-groups) di-
rectly to Chapter XIII (Ergodic Theory and Diffusion Theory) and to Chapter XIV
(Integration of the Equation of Evolution). Such materials as “Weak Topologies and

Duality in Locally Convex Spaces” and “Nuclear Spaces” are presented in the form
of the appendices to Chapter V and Chapter X, respectively. These might be skipped
for the first reading by those who are interested rather in the application of linear
operators.
Some relevant books: Classics, and in the Dover series:
Banach, “Theory of Linear Operations” [Ban93]

Georgi, “Weak Interactions and Modern Particle Theory” [Geo09]
Prenter, “Splines and Variational Methods” [Pre89]
In chapters 4 and 8 above we have cited pioneers in quantum physics, the founda-
tions of quantum mechanics. The most central here are Heisenberg (matrix mechan-
ics), Schrödinger (wave mechanics, the Schrödinger equation), and Dirac (Dirac’s
equation is a relativistic wave equation, describes all spin-½ massive particles free
form, as well as electromagnetic interactions). We further sketched von Neumann’s
discovery of the equivalence of the answers given by Heisenberg and Schrödinger,
and the Stone-von Neumann uniqueness theorem. The relevant papers and books
are as follows: [Hei69, Sch32a, vN31, HN28, Dir35, Dir47].
Functional Analysis
O
u *
Linear Mathematical
Operators (A) Physics (C)
O c ; O

Harmonic Representation
Analysis (B) i 5 Theory (D)
) # { u
Probability Theory / Statistics (E)
Fig. A.1
(A)
 bounded differential operators, ODE/PDE
x
unbounded generators of diffusion
y
 geometry
x
spectral theory Schrödinger operators
y
spectral representation wave operators
single operators scattering operators
system of operators
operator commutation relations
Table A.1
(B)
analysis / synthesis
Fourier analysis, wavelet analysis
commutative
x

non-commutative Applications:
y
signal processing
physics
statistics
analysis on fractals
Table A.2
(C)
quantum physics
x

classical mechanics
y
quantum information
statistical physics states and decomposition
x

quantum field theory equilibrium: Gibbs, KMS, ...
y
relativistic
x

non-relativistic
y
Table A.3
(D)
 groups (abelian, non-abelian) locally compact, non-locally compact

x
algebras


generators and relations Lie groups ←→ Lie algebras
y
induced representations
decomposition of representations
groups over R, C, or other local fields
Table A.4
(E)
discrete
continuous
Gaussian
Brownian motion
stochastic processes non-Gaussian
Lévy
solutions of diffusion equations with the use of
functional integrals (i.e., probability measure on
infinite-dimensional spaces such as C (R) or
Schwartz space S)
Table A.5
Appendix B
Often cited above
mathematical ideas originate in empirics. But, once they are
conceived, the subject begins to live a peculiar life of its own
and is . . . governed by almost entirely aesthetical motivations.
In other words, at a great distance from its empirical source, or
after much “abstract” inbreeding, a mathematical subject is in
danger of degeneration. Whenever this stage is reached the only
remedy seems to me to be the rejuvenating return to the source:
the reinjection of more or less directly empirical ideas.
– von Neumann
Inside the book, the following authors are cited frequently, W. Arveson, S. Banach,
P. Dirac, K. Friedrichs, I. Gelfand, W. Heisenberg, D. Hilbert, K. Ito, M. Krein,
E. Nelson, R. Phillips, E. Schrödinger, H.A. Schwarz, L. Schwartz, J. Schwartz, I.
Segal, M. Stone, J. von Neumann, N. Wiener. Below a short bio.
W. Arveson. [Arv72, Arv98]. Cited in connection with C∗ -algebras and their states
and representations.
William Arveson (1934 – 2011); known for his work on completely positive maps,
and their extensions; powerful generalizations of the ideas of Banach, Krein, and
Stinespring . An early results in this area is an extension theorem for completely
positive maps with values in the algebra of all bounded operators. This theorem
led to injectivity of von-Neumann algebras in general, and work by Alain Connes
relating injectivity to hyperfiniteness. In a series of papers in the 60’s and 70’s,
Arveson introduced non-commutative analogues of several concepts from classical
harmonic analysis including the Shilov and. Choquet boundaries.
319
320 B Often cited above
S. Banach. [Ban93]. The Banach of “Banach space.” Banach called them

“B-spaces” in his book. They were also formalized by Norbert Wiener (who
traveled in Europe in the 1920ties.) But the name “Banach space” stuck.
Stefan Banach (1892 – 1945), one of the founders of modern functional analysis
and one of the original members of the Lwów School of Mathematics, in Poland be-
tween the two World Wars. His 1932 book, Théorie des opérations linéaires (The-
ory of Linear Operations), is the first monograph on the general theory of functional
analysis.
P. Dirac. [Dir35, Dir47]. Cited in connection with the “Dirac equation” and
especially our notation for vectors and operators in Hilbert space, as well as the
axioms of observables, states and measurements.
Paul Adrien Maurice Dirac (1902 – 1984), an English theoretical physicist; fun-
damental contributions to the early development of both quantum mechanics and
quantum electrodynamics. He was the Lucasian Professor of Mathematics at the
University of Cambridge. Notable discoveries, the Dirac equation, which describes
the behavior of fermions and predicted the existence of antimatter. Dirac shared the
Nobel Prize in Physics for 1933 with Erwin Schrödinger, “for the discovery of new
productive forms of atomic theory.”
K. Friedrichs. [Fri80, FL28]. Is the Friedrichs of the Friedrichs extension;

referring to the following Theorem: Every semibounded operator S with dense
domain in Hilbert space has a selfadjoint extension having the same lower bound
as S. There are other semibounded and selfadjoint extensions of S; – they were
found later by M. Krein.
Kurt Otto Friedrichs (1901 – 1982), a noted German-American mathematician; a

co-founder of the Courant Institute at New York University and recipient of the
National Medal of Science.
B Often cited above 321
I. Gelfand. [GJ60, GS60, GG59]. Is the “G” in GNS (Gelfand-Naimark-Segal), the

correspondence between states and cyclic representations.
Israel Moiseevich Gelfand, also written Israïl Moyseyovich Gel’fand, or Izrail M.

Gelfand, (1913 – 2009) , a Russian-American mathematician; major contributions
to many branches of mathematics: representation theory and functional analysis.
The recipient of numerous awards and honors, including the Order of Lenin and
the Wolf Prize, – a lifelong academic, serving decades as a professor at Moscow
State University and, after immigrating to the United States shortly before his 76th
birthday, at Rutgers University.
W. Heisenberg. [Hei69]. Is the Heisenberg of the Heisenberg uncertainty principle

for the operators P (momentum) and Q (position), and of matrix mechanics, as the
first mathematical formulation of quantum observables. In Heisenberg’s picture,
the dynamics, the observables are studied as function of time; by contrast to
Schrödinger’s model which have the states (wave-functions) functions of time, and
satisfying a PDE wave equation, now called the Schrödinger equation. In the late
1920ties, the two pictures, that of Heisenberg and of Schrödinger were thought to
be irreconcilable. Work of von Neumann in 1932 demonstrated that they in fact are
equivalent.
Werner Karl Heisenberg (1901 – 1976); one of the key creators of quantum me-
chanics. A 1925 paper was a breakthrough. In the subsequent series of papers with
Max Born and Pascual Jordan, this matrix formulation of quantum mechanics took a
mathematical rigorous formulation. In 1927 he published his uncertainty principle.
Heisenberg was awarded the Nobel Prize in Physics for 1932 “for the creation of
quantum mechanics.” He made important contributions to the theories of the hydro-
dynamics of turbulent flows, the atomic nucleus.
D. Hilbert. [Hil24, Hil22, Hil02]. Cited in connection with the early formulations
of the theory of operators in (what is now called) Hilbert space. The name Hilbert
space was suggested by von Neumann who studied with Hilbert in the early
1930ties, before he moved to the USA. (The early papers by von Neumann are in
German.)
David Hilbert (1862 – 1943) is recognized as one of the most influential and univer-
sal mathematicians of the 19th and early 20th centuries. Discovered and developed
invariant theory and axiomatization of geometry. In his his 1900 presentation of a
collection of research problems, he set the course for much of the mathematical
research of the 20th century.
K. Ito. [Itô07, Itô04]. Cited in connection with Brownian motion, Ito-calculus, and
stochastic processes. Making connection to functional analysis via the theory of
semigroups of operators (Hille and Phillips.)
M. Krein. [Kre46, Kre55]. Is the Krein of Krein-Millman on convex weak

∗-compact sets.
E. Nelson. [Nel69, Nel59a]. Cited in connection with spectral representation, and

Brownian motion.
R. Phillips. [LP68]. Cited in connection with the foundations of functional

analysis, especially the theory of semigroups of bounded operators acting on
Banach space.
E. Schrödinger. [Sch99, Sch40, Sch32b]. Is the Schrödinger of the Schrödinger

equation; the PDE which governs the dynamics of quantum states (as
wave-functions).
Erwin Rudolf Josef Alexander Schrödinger (1887 – 1961), a Nobel Prize in physics.
– quantum theory forming the basis of wave mechanics: he formulated the wave
equation (stationary and time-dependent Schrödinger equation) , and he Schrödinger
proposed an original interpretation of the physical meaning of the wave function;
formalized the notion of entanglement. He was critical the conventional Copen-
hagen interpretation of quantum mechanics (using e.g. the paradox of Schrödinger’s
cat).
H.A. Schwarz. [Sch70]. Is the Schwarz of the Cauchy-Schwarz inequality. H.A.

Schwarz is German and is a contemporary of K. Weierstrass.
Karl Hermann Amandus Schwarz (1843 –1921), a German mathematician, known

for his work in complex analysis. At Göttingen, he pioneered of function theory,
differential geometry and the calculus of variations.
L. Schwartz. [Sch95, Sch58, Sch57]. Is the Schwartz (French) of the theory of

distributions (dating the 1950ties), also now named “generalized functions” in the
books by Gelfand et al. Parts of this theory were developed independently on the
two sides of the Iron-Curtain;– in the time of the Cold War.
J. Schwartz. [DS88b, DS88c, DS88a]. Is the Schwartz of the book set “linear
operators” by Dunford and Schwartz. Vol II [DS88c] is one of the best presentation
of the theory of unbounded operators.
I. Segal. [Seg50]. Cited in connection with the foundations of functional analysis,

and pioneering research in mathematical physics. Is the “S” in GNS
(Gelfand-Naimark-Segal).
M. Stone. [Sto51a, Sto51b]. Is the “S” in the Stone-Weierstrass theorem; and in

the Stone-von Neumann uniqueness theorem; the latter to the effect that any two
representations of Heisenberg’s commutation relations in the same (finite!) number
of degrees of freedom are unitarily equivalent.
J. von Neumann. [vN31, vN32b]. Cited in connection with the Stone-von

Neumann uniqueness theorem, the deficiency indices which determine parameters
for possible selfadjoint extensions of given Hermitian (formally selfadjoint,
symmetric) with dense domain in Hilbert space.
John von Neumann (1903 – 1957), Hungarian-American; inventor and polymath.

He made major contributions to: foundations of mathematics, functional analysis,
ergodic theory, numerical analysis, physics (quantum mechanics, hydrodynamics,
and economics (game theory), computing (von Neumann architecture, linear pro-
gramming, self-replicating machines, stochastic computing (Monte-Carlo1 )), – was
1 “Monte-Carlo” means “simulation” with computer generated random number.

a pioneer of the application of operator theory to quantum mechanics, a principal

member of the Manhattan Project and the Institute for Advanced Study in Princeton.
– A key figure in the development of game theory, cellular automata, and the digital
computer.
N. Wiener. [Wie53, WS53]. Cited in connection with Brownian motion, Wiener

measure, and stochastic processes. And more directly, the “Wiener” of
Paley-Wiener spaces; –at the crossroads of harmonic analysis and functional
analysis. Also the Wiener of filters in signal processing; high-pass/low-pass etc.
Der skal et par dumheder

med i en bog ....
for at også de dumme
skal syns, den er klog.
— Piet Hein.
Translation:
Your book should include a few stupidities
mixing them in, – this is art.
so that also the stupid will think it is smart.
References
AD86. Daniel Alpay and Harry Dym, On applications of reproducing kernel spaces to the
Schur algorithm and rational J unitary factorization, I. Schur methods in operator
theory and signal processing, Oper. Theory Adv. Appl., vol. 18, Birkhäuser, Basel,
1986, pp. 89–159. MR 902603 (89g:46051)
AG93. N. I. Akhiezer and I. M. Glazman, Theory of linear operators in Hilbert space, Dover
Publications Inc., New York, 1993, Translated from the Russian and with a preface by
Merlynd Nestell, Reprint of the 1961 and 1963 translations, Two volumes bound as
one. MR 1255973 (94i:47001)
AH13. Giles Auchmuty and Qi Han, Spectral representations of solutions of linear elliptic
equations on exterior regions, J. Math. Anal. Appl. 398 (2013), no. 1, 1–10. MR
2984310
AJ12. Daniel Alpay and Palle E. T. Jorgensen, Stochastic processes induced by singular op-
erators, Numer. Funct. Anal. Optim. 33 (2012), no. 7-9, 708–735. MR 2966130
AJL13. Daniel Alpay, Palle Jorgensen, and Izchak Lewkowicz, Extending wavelet filters: infi-
nite dimensions, the nonrational case, and indefinite inner product spaces, Excursions
in harmonic analysis. Volume 2, Appl. Numer. Harmon. Anal., Birkhäuser/Springer,
New York, 2013, pp. 69–111. MR 3050315
AJLM13. Daniel Alpay, Palle Jorgensen, Izchak Lewkowicz, and Itzik Marziano, Representation
formulas for Hardy space functions through the Cuntz relations and new interpola-
tion problems, Multiscale signal analysis and modeling, Springer, New York, 2013,
pp. 161–182. MR 3024468
AJS14. Daniel Alpay, Palle Jorgensen, and Guy Salomon, On free stochastic processes and
their derivatives, Stochastic Process. Appl. 124 (2014), no. 10, 3392–3411. MR
3231624
AJSV13. Daniel Alpay, Palle Jorgensen, Ron Seager, and Dan Volok, On discrete analytic func-
tions: products, rational functions and reproducing kernels, J. Appl. Math. Comput.
41 (2013), no. 1-2, 393–426. MR 3017129
Alp92. Daniel Alpay, On linear combinations of positive functions, associated reproducing
kernel spaces and a non-Hermitian Schur algorithm, Arch. Math. (Basel) 58 (1992),
no. 2, 174–182. MR 1143167 (92m:46039)
325
And74. Joel Anderson, On compact perturbations of operators, Canad. J. Math. 26 (1974),

247–250. MR 0333774 (48 #12098)
Aro50. N. Aronszajn, Theory of reproducing kernels, Trans. Amer. Math. Soc. 68 (1950), 337–
404. MR 0051437 (14,479c)
ARR13. Luigi Accardi, Habib Rebei, and Anis Riahi, The quantum decomposition of ran-
dom variables without moments, Infin. Dimens. Anal. Quantum Probab. Relat. Top.
16 (2013), no. 2, 1350012, 28. MR 3078823
Arv72. William Arveson, Lattices of invariant subspaces, Bull. Amer. Math. Soc. 78 (1972),
515–519. MR 0298451 (45 #7503)
Arv76. , An invitation to C∗ -algebras, Springer-Verlag, New York-Heidelberg, 1976,
Graduate Texts in Mathematics, No. 39. MR 0512360 (58 #23621)
Arv98. , Subalgebras of C∗ -algebras. III. Multivariable operator theory, Acta Math.
181 (1998), no. 2, 159–228. MR 1668582 (2000e:47013)
AW13. Charles Akemann and Nik Weaver, A lyapunov type theorem from kadison-singer,
arXiv:1308.5276 (2013).
Ban93. Stefan Banach, Théorie des opérations linéaires, Éditions Jacques Gabay, Sceaux,
1993, Reprint of the 1932 original. MR 1357166 (97d:01035)
BC05. R. Burioni and D. Cassi, Random walks on graphs: ideas, techniques and results, J.
Phys. A 38 (2005), no. 8, R45–R78. MR 2119174 (2006b:82059)
BCD06. Lali Barrière, Francesc Comellas, and Cristina Dalfó, Fractality and the small-world
effect in Sierpinski graphs, J. Phys. A 39 (2006), no. 38, 11739–11753. MR 2275879
(2008i:28008)
BJKR84. O. Bratteli, P. E. T. Jorgensen, A. Kishimoto, and D. W. Robinson, A C∗ -algebraic
Schoenberg theorem, Ann. Inst. Fourier (Grenoble) 34 (1984), no. 3, 155–187. MR
762697 (86b:46105)
BJO04. Ola Bratteli, Palle E. T. Jorgensen, and Vasyl0 Ostrovs0 kyı̆, Representation theory and
numerical AF-invariants, Mem. Amer. Math. Soc. 168 (2004), no. 797, xviii+178. MR
2030387 (2005i:46069)
BM13. John P. Boyd and Philip W. McCauley, Quartic Gaussian and inverse-quartic Gaussian
radial basis functions: the importance of a nonnegative Fourier transform, Comput.
Math. Appl. 65 (2013), no. 1, 75–88. MR 3003386
BN00. George Bachman and Lawrence Narici, Functional analysis, Dover Publications, Inc.,
Mineola, NY, 2000, Reprint of the 1966 original. MR 1819613 (2001k:46001)
BR79. Ola Bratteli and Derek W. Robinson, Operator algebras and quantum statistical me-
chanics. Vol. 1, Springer-Verlag, New York-Heidelberg, 1979, C∗ - and W ∗ -algebras,
algebras, symmetry groups, decomposition of states, Texts and Monographs in Physics.
MR 545651 (81a:46070)
BR81. O. Bratteli and D.W. Robinson, Operator algebras and quantum statistical mechanics:
Equilibrium states models in quantum statistical mechanics, Texts and Monographs in
Physics, Springer-Verlag, 1981.
BR81. Ola Bratteli and Derek W. Robinson, Equilibrium states of a Bose gas with repulsive
interactions, J. Austral. Math. Soc. Ser. B 22 (1980/81), no. 2, 129–147. MR 593999
(82b:82022)
Cas13. Peter G. Casazza, The Kadison-Singer and Paulsen problems in finite frame theory,
Finite frames, Appl. Numer. Harmon. Anal., Birkhäuser/Springer, New York, 2013,
pp. 381–413. MR 2964016
Cas14. Peter G. Casazza, Consequences of the marcus/spielman/stivastava solution to the
kadison-singer problem, arXiv:1407.4768 (2014).
CH08. D. A. Croydon and B. M. Hambly, Local limit theorems for sequences of simple
random walks on graphs, Potential Anal. 29 (2008), no. 4, 351–389. MR 2453564
(2010c:60218)
CJK+ 12. Jianxin Chen, Zhengfeng Ji, David Kribs, Zhaohui Wei, and Bei Zeng, Ground-state
spaces of frustration-free Hamiltonians, J. Math. Phys. 53 (2012), no. 10, 102201, 15.
MR 3050570
CM13. B. Currey and A. Mayeli, The Orthonormal Dilation Property for Abstract Parseval
Wavelet Frames, Canad. Math. Bull. 56 (2013), no. 4, 729–736. MR 3121682
Con90. John B. Conway, A course in functional analysis, second ed., Graduate Texts in Math-
ematics, vol. 96, Springer-Verlag, New York, 1990. MR 1070713 (91e:46001)
CRKS79. P. Cotta-Ramusino, W. Krüger, and R. Schrader, Quantum scattering by external met-
rics and Yang-Mills potentials, Ann. Inst. H. Poincaré Sect. A (N.S.) 31 (1979), no. 1,
43–71. MR 557051 (81h:81129)
CW14. K. L. Chung and R. J. Williams, Introduction to stochastic integration, second ed.,
Modern Birkhäuser Classics, Birkhäuser/Springer, New York, 2014. MR 3136102
dB68. Louis de Branges, Hilbert spaces of entire functions, Prentice-Hall Inc., Englewood
Cliffs, N.J., 1968. MR 0229011 (37 #4590)
dBR66. Louis de Branges and James Rovnyak, Canonical models in quantum scattering theory,
Perturbation Theory and its Applications in Quantum Mechanics (Proc. Adv. Sem.
Math. Res. Center, U.S. Army, Theoret. Chem. Inst., Univ. of Wisconsin, Madison,
Wis., 1965), Wiley, New York, 1966, pp. 295–392. MR 0244795 (39 #6109)
Dev59. Allen Devinatz, On the extensions of positive definite functions, Acta Math. 102 (1959),
109–134. MR 0109992 (22 #875)
Dev72. , The deficiency index of a certain class of ordinary self-adjoint differential
operators, Advances in Math. 8 (1972), 434–473. MR 0298102 (45 #7154)
DHL09. Dorin Dutkay, Deguang Han, and David Larson, A duality principle for groups, J.
Funct. Anal. 257 (2009), no. 4, 1133–1143. MR 2535465 (2010i:22006)
Dir35. P. A. M. Dirac, The electron wave equation in de-Sitter space, Ann. of Math. (2) 36
(1935), no. 3, 657–669. MR 1503243
Dir47. , The Principles of Quantum Mechanics, Oxford, at the Clarendon Press, 1947,
3d ed. MR 0023198 (9,319d)
Dix81. Jacques Dixmier, Von neumann algebras, North Holland, 1 1981.
DM85. H. Dym and H.P. McKean, Fourier series and integrals, Probability and Mathematical
Statistics, ACADEMIC PressINC, 1985.
DM91. V. A. Derkach and M. M. Malamud, Generalized resolvents and the boundary value
problems for Hermitian operators with gaps, J. Funct. Anal. 95 (1991), no. 1, 1–95.
MR 1087947 (93d:47046)
dO09. César R. de Oliveira, Intermediate spectral theory and quantum dynamics, Progress in
Mathematical Physics, vol. 54, Birkhäuser Verlag, Basel, 2009. MR 2723496
DS88a. N. Dunford and J.T. Schwartz, Linear operators, spectral operators, Linear Operators,
Wiley, 1988.
DS88b. Nelson Dunford and Jacob T. Schwartz, Linear operators. Part I, Wiley Classics Li-
brary, John Wiley & Sons Inc., New York, 1988, General theory, With the assistance
of William G. Bade and Robert G. Bartle, Reprint of the 1958 original, A Wiley-
Interscience Publication. MR 1009162 (90g:47001a)
DS88c. , Linear operators. Part II, Wiley Classics Library, John Wiley & Sons Inc.,
New York, 1988, Spectral theory. Selfadjoint operators in Hilbert space, With the as-
sistance of William G. Bade and Robert G. Bartle, Reprint of the 1963 original, A
Wiley-Interscience Publication. MR 1009163 (90g:47001b)
Emc00. G.G. Emch, Mathematical and conceptual foundations of 20th-century physics, North-
Holland Mathematics Studies, Elsevier Science, 2000.
EO13. Martin Ehler and Kasso A. Okoudjou, Probabilistic frames: an overview, Finite
frames, Appl. Numer. Harmon. Anal., Birkhäuser/Springer, New York, 2013, pp. 415–
436. MR 2964017
Fan10. Mark Fannes, An introduction to quantum probability, Theoretical foundations of
quantum information processing and communication, Lecture Notes in Phys., vol. 787,
Springer, Berlin, 2010, pp. 1–38. MR 2762151 (2012d:81174)
Fef71. Charles Fefferman, Characterizations of bounded mean oscillation, Bull. Amer. Math.
Soc. 77 (1971), 587–588. MR 0280994 (43 #6713)
FL28. Kurt Friedrichs and Hans Lewy, Über die Eindeutigkeit und das Abhängigkeitsgebiet
der Lösungen beim Anfangswertproblem linearer hyperbolischer Differentialgleichun-
gen, Math. Ann. 98 (1928), no. 1, 192–204. MR 1512400
FM13. J. C. Ferreira and V. A. Menegatto, Positive definiteness, reproducing kernel Hilbert
spaces and beyond, Ann. Funct. Anal. 4 (2013), no. 1, 64–88. MR 3004212
FR42. W. H. J. Fuchs and W. W. Rogosinski, A note on Mercer’s theorem, J. London Math.
Soc. 17 (1942), 204–210. MR 0008270 (4,272g)
Fri80. Kurt Otto Friedrichs, Spectral theory of operators in Hilbert space, Applied Mathe-
matical Sciences, vol. 9, Springer-Verlag, New York-Berlin, 1980, Corrected reprint.
MR 635783 (82j:47001)
Fug74. Bent Fuglede, Boundary minimum principles in potential theory, Math. Ann. 210
(1974), 213–226. MR 0357827 (50 #10293b)
Geo09. H. Georgi, Weak interactions and modern particle theory, Dover Books on Physics
Series, Dover Publications, Incorporated, 2009.
GG59. I. M. Gel0 fand and M. I. Graev, Geometry of homogeneous spaces, representations of
groups in homogeneous spaces and related questions of integral geometry. I, Trudy
Moskov. Mat. Obšč. 8 (1959), 321–390; addendum 9 (1959), 562. MR 0126719 (23
#A4013)
GG91. V. I. Gorbachuk and M. L. Gorbachuk, Boundary value problems for operator differ-
ential equations, Mathematics and its Applications (Soviet Series), vol. 48, Kluwer
Academic Publishers Group, Dordrecht, 1991, Translated and revised from the 1984
Russian original. MR 1154792 (92m:34133)
GG02. M. Gadella and F. Gómez, A unified mathematical formalism for the Dirac formula-
tion of quantum mechanics, Found. Phys. 32 (2002), no. 6, 815–869. MR 1917012
(2003j:81070)
GJ60. I. M. Gel0 fand and A. M. Jaglom, Integration in functional spaces and its applications
in quantum physics, J. Mathematical Phys. 1 (1960), 48–69. MR 0112604 (22 #3455)
Gli60. James G. Glimm, On a certain class of operator algebras, Trans. Amer. Math. Soc. 95
(1960), 318–340. MR 0112057 (22 #2915)
Gli61. James Glimm, Type I C∗ -algebras, Ann. of Math. (2) 73 (1961), 572–612. MR 0124756
(23 #A2066)
GLS12. Martin J. Gander, Sébastien Loisel, and Daniel B. Szyld, An optimal block iterative
method and preconditioner for banded matrices with applications to PDEs on irregular
domains, SIAM J. Matrix Anal. Appl. 33 (2012), no. 2, 653–680. MR 2970224
GS60. I. M. Gel0 fand and Do-šin Sya, On positive definite distributions, Uspehi Mat. Nauk
15 (1960), no. 1 (91), 185–190. MR 0111991 (22 #2849)
HdSS12. Seppo Hassi, Hendrik S. V. de Snoo, and Franciszek Hugon Szafraniec (eds.), Operator
methods for boundary value problems, London Mathematical Society Lecture Note
Series, vol. 404, Cambridge University Press, Cambridge, 2012. MR 3075434
Hei69. W Heisenberg, Über quantentheoretische umdeutung kinematischer und mechanischer
beziehungen.(1925) in: G, Ludwig, Wellenmechanik, Einführung und Originaltexte,
Akademie-Verlag, Berlin (1969), 195.
Hid80. Takeyuki Hida, Brownian motion, Applications of Mathematics, vol. 11, Springer-
Verlag, New York, 1980, Translated from the Japanese by the author and T. P. Speed.
MR 562914 (81a:60089)
Hil02. David Hilbert, Mathematical problems, Bull. Amer. Math. Soc. 8 (1902), no. 10, 437–
479. MR 1557926
Hil22. , Die logischen Grundlagen der Mathematik, Math. Ann. 88 (1922), no. 1-2,
151–165. MR 1512123
Hil24. , Die Grundlagen der Physik, Math. Ann. 92 (1924), no. 1-2, 1–32. MR
1512197
HJL+ 13. Deguang Han, Wu Jing, David Larson, Pengtong Li, and Ram N. Mohapatra, Dilation
of dual frame pairs in Hilbert C∗ -modules, Results Math. 63 (2013), no. 1-2, 241–250.
MR 3009685
HKLW07. Deguang Han, Keri Kornelson, David Larson, and Eric Weber, Frames for undergrad-
uates, Student Mathematical Library, vol. 40, American Mathematical Society, Provi-
dence, RI, 2007. MR 2367342 (2010e:42044)
HN28. D. Hilbert and J. v. Neumann, Über die Grundlagen der Quantenmechanik, Math. Ann.
98 (1928), no. 1, 1–30. MR 1512390
Itô04. Kiyosi Itô, Stochastic processes, Springer-Verlag, Berlin, 2004, Lectures given at
Aarhus University, Reprint of the 1969 original, Edited and with a foreword by Ole
E. Barndorff-Nielsen and Ken-iti Sato. MR 2053326 (2005e:60002)
Itô06. Kiyoshi Itô, Essentials of stochastic processes, Translations of Mathematical Mono-
graphs, vol. 231, American Mathematical Society, Providence, RI, 2006, Translated
from the 1957 Japanese original by Yuji Ito. MR 2239081 (2007i:60001)
Itô07. Kiyosi Itô, Memoirs of my research on stochastic analysis, Stochastic analysis and
applications, Abel Symp., vol. 2, Springer, Berlin, 2007, pp. 1–5. MR 2397781
JÓ00. Palle E. T. Jorgensen and Gestur Ólafsson, Unitary representations and Osterwalder-
Schrader duality, The mathematical legacy of Harish-Chandra (Baltimore, MD, 1998),
Proc. Sympos. Pure Math., vol. 68, Amer. Math. Soc., Providence, RI, 2000, pp. 333–
401. MR 1767902 (2001f:22036)
Jor08. Palle E. T. Jorgensen, Essential self-adjointness of the graph-Laplacian, J. Math. Phys.
49 (2008), no. 7, 073510, 33. MR 2432048 (2009k:47099)
Jør14. Palle E. T. Jørgensen, A universal envelope for Gaussian processes and their kernels,
J. Appl. Math. Comput. 44 (2014), no. 1-2, 1–38. MR 3147727
JP10. Palle E. T. Jorgensen and Erin Peter James Pearse, A Hilbert space approach to effec-
tive resistance metric, Complex Anal. Oper. Theory 4 (2010), no. 4, 975–1013. MR
2735315 (2011j:05338)
JP11a. Palle E. T. Jorgensen and Erin P. J. Pearse, Resistance boundaries of infinite networks,
Random walks, boundaries and spectra, Progr. Probab., vol. 64, Birkhäuser/Springer
Basel AG, Basel, 2011, pp. 111–142. MR 3051696
JP11b. , Spectral reciprocity and matrix representations of unbounded operators, J.
Funct. Anal. 261 (2011), no. 3, 749–776. MR 2799579
JP12. P. E. T. Jorgensen and A. M. Paolucci, q-frames and Bessel functions, Numer. Funct.
Anal. Optim. 33 (2012), no. 7-9, 1063–1069. MR 2966144
JP13a. Palle E. T. Jorgensen and Erin P. J. Pearse, A discrete Gauss-Green identity for un-
bounded Laplace operators, and the transience of random walks, Israel J. Math. 196
(2013), no. 1, 113–160. MR 3096586
JP13b. , A discrete Gauss-Green identity for unbounded Laplace operators, and the
transience of random walks, Israel J. Math. 196 (2013), no. 1, 113–160. MR 3096586
JP13c. , Multiplication operators on the energy space, J. Operator Theory 69 (2013),
no. 1, 135–159. MR 3029492
JPT12a. Palle Jorgensen, Steen Pedersen, and Feng Tian, Restrictions and extensions of semi-
bounded operators, Complex Analysis and Operator Theory (2012) (English).
JPT12b. , Translation representations and scattering by two intervals, J. Math. Phys. 53
(2012), no. 5, 053505, 49. MR 2964262
JPT12c. Palle E.T. Jorgensen, Steen Pedersen, and Feng Tian, Momentum operators in two in-
tervals: Spectra and phase transition, Complex Analysis and Operator Theory (2012)
(English).
JPT14. Palle Jorgensen, Steen Pedersen, and Feng Tian, Harmonic analysis of a class of re-
producing kernel Hilbert spaces arising from groups, arXiv:1401.4782 (2014).
JT14a. Palle Jorgensen and Feng Tian, Frames and factorization of graph laplacians,
arXiv:1404.1424 (2014).
JT14b. , Frames and factorization of graph laplacians, arXiv:1404.1424 (2014).
JT14c. , Infinite networks and variation of conductance functions in discrete lapla-
cians, arXiv:1404.4686 (2014).
Kat95. Tosio Kato, Perturbation theory for linear operators, Classics in Mathematics,
Springer-Verlag, Berlin, 1995, Reprint of the 1980 edition. MR 1335452 (96a:47025)
KF75. A. N. Kolmogorov and S. V. Fomı̄n, Introductory real analysis, Dover Publications,

Inc., New York, 1975, Translated from the second Russian edition and edited by
Richard A. Silverman, Corrected reprinting. MR 0377445 (51 #13617)
KL14. Mark G. Krein and Heinz Langer, Continuation of Hermitian Positive Definite Func-
tions and Related Questions, Integral Equations Operator Theory 78 (2014), no. 1,
1–69. MR 3147401
KLZ09. Victor Kaftal, David R. Larson, and Shuang Zhang, Operator-valued frames, Trans.
Amer. Math. Soc. 361 (2009), no. 12, 6349–6385. MR 2538596 (2010h:42060)
KMRS05. Norio Konno, Naoki Masuda, Rahul Roy, and Anish Sarkar, Rigorous results on the
threshold network model, J. Phys. A 38 (2005), no. 28, 6277–6291. MR 2166622
(2006j:82054)
KOPT13. Gitta Kutyniok, Kasso A. Okoudjou, Friedrich Philipp, and Elizabeth K. Tuley, Scal-
able frames, Linear Algebra Appl. 438 (2013), no. 5, 2225–2238. MR 3005286
KR97a. Richard V. Kadison and John R. Ringrose, Fundamentals of the theory of operator
algebras. Vol. I, Graduate Studies in Mathematics, vol. 15, American Mathematical
Society, Providence, RI, 1997, Elementary theory, Reprint of the 1983 original. MR
MR1468229 (98f:46001a)
KR97b. , Fundamentals of the theory of operator algebras. Vol. II, Graduate Studies
in Mathematics, vol. 16, American Mathematical Society, Providence, RI, 1997, Ad-
vanced theory, Corrected reprint of the 1986 original. MR MR1468230 (98f:46001b)
Kre46. M. Krein, Concerning the resolvents of an Hermitian operator with the deficiency-
index (m, m), C. R. (Doklady) Acad. Sci. URSS (N.S.) 52 (1946), 651–654. MR
0018341 (8,277a)
Kre55. M. G. Krein, On some cases of the effective determination of the density of a nonuni-
form string by its spectral function, 2 Pine St., West Concord, Mass., 1955, Translated
by Morris D. Friedman. MR 0075403 (17,740f)
KS02. Vadim Kostrykin and Robert Schrader, Statistical ensembles and density of states,
Mathematical results in quantum mechanics (Taxco, 2001), Contemp. Math., vol. 307,
Amer. Math. Soc., Providence, RI, 2002, pp. 177–208. MR 1946030 (2004g:82007)
Lax02. Peter D. Lax, Functional analysis, Pure and Applied Mathematics (New York), Wiley-
Interscience [John Wiley & Sons], New York, 2002. MR 1892228 (2003a:47001)
LP68. P. D. Lax and R. S. Phillips, Scattering theory, Proc. Internat. Congr. Math. (Moscow,
1966), Izdat. “Mir”, Moscow, 1968, pp. 542–545. MR 0237960 (38 #6237)
LP89. Peter D. Lax and Ralph S. Phillips, Scattering theory, second ed., Pure and Applied
Mathematics, vol. 26, Academic Press Inc., Boston, MA, 1989, With appendices by
Cathleen S. Morawetz and Georg Schmidt. MR 1037774 (90k:35005)
Maa10. H. Maassen, Quantum probability and quantum information theory, Quantum informa-
tion, computation and cryptography, Lecture Notes in Phys., vol. 808, Springer, Berlin,
2010, pp. 65–108. MR 2768446
Mac52. George W. Mackey, Induced representations of locally compact groups. I, Ann. of
Math. (2) 55 (1952), 101–139. MR 0044536 (13,434a)
Mac85. , Quantum mechanics from the point of view of the theory of group representa-
tions, Applications of group theory in physics and mathematical physics (Chicago,
1982), Lectures in Appl. Math., vol. 21, Amer. Math. Soc., Providence, RI, 1985,
pp. 219–253. MR 789292 (86j:81051)
Mac88. , Induced representations and the applications of harmonic analysis, Harmonic
analysis (Luxembourg, 1987), Lecture Notes in Math., vol. 1359, Springer, Berlin,
1988, pp. 16–51. MR 974302 (90c:22021)
Mac92. , The scope and history of commutative and noncommutative harmonic analy-
sis, History of Mathematics, vol. 5, American Mathematical Society, Providence, RI;
London Mathematical Society, London, 1992. MR 1171011 (93g:22006)
Mac09. Barbara D. MacCluer, Elementary functional analysis, Graduate Texts in Mathematics,
vol. 253, Springer, New York, 2009. MR 2462971 (2010b:46001)
MSS13. Adam Marcus, Daniel A Spielman, and Nikhil Srivastava, Interlacing families ii:
Mixed characteristic polynomials and the kadison-singer problem, arXiv:1306.3969
(2013).
Nel59a. Edward Nelson, Analytic vectors, Ann. of Math. (2) 70 (1959), 572–615. MR 0107176
(21 #5901)
Nel59b. , Regular probability measures on function space, Ann. of Math. (2) 69 (1959),
630–643. MR 0105743 (21 #4479)
Nel67. , Dynamical theories of Brownian motion, Princeton University Press, Prince-
ton, N.J., 1967. MR 0214150 (35 #5001)
Nel69. , Topics in dynamics. I: Flows, Mathematical Notes, Princeton University
Press, Princeton, N.J., 1969. MR 0282379 (43 #8091)
Nus75. A. Edward Nussbaum, Extension of positive definite functions and representation of
functions in terms of spherical functions in symmetric spaces of noncompact type of
rank 1, Math. Ann. 215 (1975), 97–116. MR 0385473 (52 #6334)
OH13. Anatol Odzijewicz and Maciej Horowski, Positive kernels and quantization, J. Geom.
Phys. 63 (2013), 80–98. MR 2996399
OR07. V. S. Olkhovsky and E. Recami, Time as a quantum observable, Internat. J. Modern
Phys. A 22 (2007), no. 28, 5063–5087. MR 2371443 (2009b:81008)
Ørs79. Bent Ørsted, Induced representations and a new proof of the imprimitivity theorem, J.
Funct. Anal. 31 (1979), no. 3, 355–359. MR 531137 (80d:22007)
Par82. K.R. Parthasarathy, Probability measures on metric spaces, AMS Chelsea Publishing
Series, Acad. Press, 1982.
Par09. K. R. Parthasarathy, An invitation to quantum information theory, Perspectives in math-
ematical sciences. I, Stat. Sci. Interdiscip. Res., vol. 7, World Sci. Publ., Hackensack,
NJ, 2009, pp. 225–245. MR 2581746 (2011d:81061)
Phe01. Robert R. Phelps, Lectures on choquet’s theorem (lecture notes in mathematics), 2nd
ed., Springer, 5 2001.
Pow75. Robert T. Powers, Simplicity of the C∗ -algebra associated with the free group on two
generators, Duke Math. J. 42 (1975), 151–156. MR 0374334 (51 #10534)
Pre89. P. M. Prenter, Splines and variational methods, Wiley Classics Library, John Wiley
& Sons, Inc., New York, 1989, Reprint of the 1975 original, A Wiley-Interscience
Publication. MR 1013116 (90j:65001)
PS75. K. R. Parthasarathy and K. Schmidt, Stable positive definite functions, Trans. Amer.
Math. Soc. 203 (1975), 161–174. MR 0370681 (51 #6907)
RAKK05. G. J. Rodgers, K. Austin, B. Kahng, and D. Kim, Eigenvalue spectra of complex net-
works, J. Phys. A 38 (2005), no. 43, 9431–9437. MR 2187996 (2006j:05186)
RS75. M. Reed and B. Simon, Methods of modern mathematical physics: Fourier analysis,
self-adjointness, Fourier analysis, Self-adjointness, no. v. 2, Academic Press, 1975.
RSN90. Frigyes Riesz and Béla Sz.-Nagy, Functional analysis, Dover Books on Advanced
Mathematics, Dover Publications, Inc., New York, 1990, Translated from the sec-
ond French edition by Leo F. Boron, Reprint of the 1955 original. MR 1068530
(91g:00002)
Rud63. Walter Rudin, The extension problem for positive-definite functions, Illinois J. Math. 7
(1963), 532–539. MR 0151796 (27 #1779)
Rud70. , An extension theorem for positive-definite functions, Duke Math. J. 37 (1970),
49–53. MR 0254514 (40 #7722)
Rud73. , Functional analysis, McGraw-Hill Book Co., New York, 1973, McGraw-Hill
Series in Higher Mathematics. MR MR0365062 (51 #1315)
Rud87. , Real and complex analysis, third ed., McGraw-Hill Book Co., New York,
1987. MR 924157 (88k:00002)
Rud90. , Fourier analysis on groups, Wiley Classics Library, John Wiley & Sons Inc.,
New York, 1990, Reprint of the 1962 original, A Wiley-Interscience Publication. MR
1038803 (91b:43002)
Sak71. Shôichirô Sakai, C∗ -algebras and W ∗ -algebras, Springer-Verlag, New York, 1971,
Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 60. MR MR0442701 (56
#1082)
Sch32a. E. Schrödinger, Sur la théorie relativiste de l’électron et l’interprétation de la mé-
canique quantique, Ann. Inst. H. Poincaré 2 (1932), no. 4, 269–310. MR 1508000
Sch32b. , Sur la théorie relativiste de l’électron et l’interprétation de la mécanique
quantique, Ann. Inst. H. Poincaré 2 (1932), no. 4, 269–310. MR 1508000
Sch40. , A method of determining quantum-mechanical eigenvalues and eigenfunc-
tions, Proc. Roy. Irish Acad. Sect. A. 46 (1940), 9–16. MR 0001666 (1,277d)
Sch57. Laurent Schwartz, Théorie des distributions à valeurs vectorielles. I, Ann. Inst. Fourier,
Grenoble 7 (1957), 1–141. MR 0107812 (21 #6534)
Sch58. , La fonction aléatoire du mouvement brownien, Séminaire Bourbaki; 10e an-
née: 1957/1958. Textes des conférences; Exposés 152 à 168; 2e éd. corrigée, Exposé
161, vol. 23, Secrétariat mathématique, Paris, 1958. MR 0107312 (21 #6037)
Sch70. H. A. Schwarz, Gesammelte mathematische abhandlungen (ams chelsea publishing),
2nd ed., American Mathematical Society, 1 1970.
Sch95. Laurent Schwartz, Les travaux de L. Gårding sur les équations aux dérivées partielles
elliptiques, Séminaire Bourbaki, Vol. 2, Soc. Math. France, Paris, 1995, pp. Exp. No.
67, 175–182. MR 1609224
Sch99. E. Schrödinger, About Heisenberg uncertainty relation (original annotation by A. An-
gelow and M.-C. Batoni), Bulgar. J. Phys. 26 (1999), no. 5-6, 193–203 (2000), Transla-
tion of Proc. Prussian Acad. Sci. Phys. Math. Sect. 19 (1930), 296–303. MR 1782215
(2001i:81002)
SD13. F. A. Shah and Lokenath Debnath, Tight wavelet frames on local fields, Analysis
(Berlin) 33 (2013), no. 3, 293–307. MR 3118429
Seg50. I. E. Segal, An extension of Plancherel’s formula to separable unimodular groups, Ann.

of Math. (2) 52 (1950), 272–292. MR 0036765 (12,157f)
Shi96. Georgi E. Shilov, Elementary functional analysis, Dover Publications, Inc., New York,
1996, Revised English edition translated from the Russian and edited by Richard A.
Silverman, Corrected reprint of the 1974 English translation. MR 1375236 (98a:46002)
Sla03. D. A. Slavnov, Quantum measurements and Kolmogorov’s probability theory, Teoret.
Mat. Fiz. 136 (2003), no. 3, 436–443. MR 2025366 (2004m:81046)
SS11. Elias M. Stein and Rami Shakarchi, Functional analysis, Princeton Lectures in Anal-
ysis, vol. 4, Princeton University Press, Princeton, NJ, 2011, Introduction to further
topics in analysis. MR 2827930 (2012g:46001)
Sti55. W. Forrest Stinespring, Positive functions on C∗ -algebras, Proc. Amer. Math. Soc. 6
(1955), 211–216. MR 0069403 (16,1033b)
Sti59. , Integrability of Fourier transforms for unimodular Lie groups, Duke Math. J.
26 (1959), 123–131. MR 0104161 (21 #2921)
Sto51a. M. H. Stone, On unbounded operators in Hilbert space, J. Indian Math. Soc. (N.S.) 15
(1951), 155–192 (1952). MR 0052042 (14,565d)
Sto51b. M.H. Stone, Linear transformation in hilbert space: and their applications to analysis,
Colloquium publications, American Mathematical Society, 1951.
SZ07. Steve Smale and Ding-Xuan Zhou, Learning theory estimates via integral operators
and their approximations, Constr. Approx. 26 (2007), no. 2, 153–172. MR 2327597
(2009b:68184)
SZ09. , Geometry on probability spaces, Constr. Approx. 30 (2009), no. 3, 311–323.
MR 2558684 (2011c:60006)
Tak79. Masamichi Takesaki, Theory of operator algebras. I, Springer-Verlag, New York-
Heidelberg, 1979. MR 548728 (81e:46038)
TD03. Christophe Texier and Pascal Degiovanni, Charge and current distribution in graphs,
J. Phys. A 36 (2003), no. 50, 12425–12452. MR 2025876 (2004i:81264)
Tre67. Francois Treves, Topological vector spaces, distributions and kernels, first edition ed.,
Academic Press, 7 1967.
Trè06. François Trèves, Topological vector spaces, distributions and kernels, Dover Publi-
cations, Inc., Mineola, NY, 2006, Unabridged republication of the 1967 original. MR
2296978 (2007k:46002)
vN31. J. v. Neumann, Die Eindeutigkeit der Schrödingerschen Operatoren, Math. Ann. 104
(1931), no. 1, 570–578. MR 1512685
vN32a. J. von Neumann, Über adjungierte Funktionaloperatoren, Ann. of Math. (2) 33 (1932),
no. 2, 294–310. MR 1503053
vN32b. , Über adjungierte Funktionaloperatoren, Ann. of Math. (2) 33 (1932), no. 2,
294–310. MR 1503053
vN32c. , Über einen Satz von Herrn M. H. Stone, Ann. of Math. (2) 33 (1932), no. 3,
567–573. MR 1503076
vN32d. , Zur Operatorenmethode in der klassischen Mechanik, Ann. of Math. (2) 33
(1932), no. 3, 587–642. MR 1503078
VZ92. H. Vogt and A. Zippelius, Invariant recognition in Potts glass neural networks, J. Phys.
A 25 (1992), no. 8, 2209–2226. MR 1162879 (93c:82044)
Wei03. Joachim Weidmann, Lineare Operatoren in Hilberträumen. Teil II, Mathematische

Leitfäden. [Mathematical Textbooks], B. G. Teubner, Stuttgart, 2003, Anwendungen.
[Applications]. MR 2382320 (2008k:47002)
Wie53. Norbert Wiener, Ex-prodigy. My childhood and youth, Simon and Schuster, New York,
1953. MR 0057817 (15,277f)
WS53. Norbert Wiener and Armand Siegel, A new form for the statistical postulate of quantum
mechanics, Physical Rev. (2) 91 (1953), 1551–1560. MR 0057763 (15,273f)
Yos95. Kōsaku Yosida, Functional analysis, Classics in Mathematics, Springer-Verlag, Berlin,
1995, Reprint of the sixth (1980) edition. MR 1336382 (96a:46001)
Index
algebras free, 133

C∗ -algebra, 24, 103, 112, 121, 219 Heisenberg, 183, 190
Banach algebra, 121 Lie, 5, 195, 200, 283
Cuntz algebra, 104, 133, 140 locally compact, 174, 185, 190, 207, 283,
Cuntz-algebra, 25 293
group algebra, 106, 175, 213 non-abelian, 24
Lie algebra, 181, 199, 200 Poincaré, 197
von Neumann algebra (W ∗ -algebra), 24, 66,
125, 139 Heisenberg, W.K.
commutation relation, 24, 45, 73, 139
Banach space, 19, 25, 114 uncertainty principle, 46, 220
double-dual, 113 Hilbert space
dual, 24 L2 , 113, 134
pre-dual, 24 l 2 , 35, 113
reflexive, 113 direct sum, 119
Brownian motion, 169 energy, 264
reproducing kernel, 280
compact, 29, 105, 112, 115, 121, 132, 168,
rigged, 251
175, 220
tensor product, 120
convex, 29, 105, 112, 132
index
Dirac, P.A.M
deficiency, 228, 269, 282
bra-ket, 40
Ito, K.
ket-bra, 40
integral, 281
dual, 112, 127, 197, 205, 220
Friedrichs extension, 247, 258, 261, 268, 272 measure

Borel, 108, 188
groups counting, 176
abelian, 24, 205, 283 Dirac, 134, 220
compact, 175 ergodic, 135
337
338 Index
Haar, 174, 175, 210, 284 composite system, 120

Lebesgue, 137, 143, 276 energy operator, 76
Plancherel, 206 measurement, 45, 220
probability, 108, 133, 168, 281 momentum operator, 45, 56, 73
projection-valued, 49, 87, 221, 294 observable, 45, 72, 211, 220
quasi-invariant, 185 position operator, 45, 56, 73
Wiener, 170 uncertainty principle, 94, 220
Nelson, E, 83, 85, 160, 212 representation

cyclic, 83, 109
operators
GNS, 103, 123, 148
bounded, 92
Heisenberg, 190
compact, 96
induced, 180
contractive, 120
irreducible, 123
domain of, 56
Lie algebra, 182
extension of, 228
multiplicity free, 140
formally selfadjoint, 228
of ax + b, 177
graph of, 56
of algebra, 108
Hermitian, 72, 228
Hilbert-Schmidt, 120 of group, 174
isometric, 154 of Lie group, 174
Laplacian, 261 spectral, 72, 81
multiplication, 77
Schrödinger, E. R., 45, 56, 195
normal, 66
Stinespring, W. F., 145, 151
selfadjoint, 36, 45, 77, 131, 227
stochastic process, 170, 281
trace class, 116
Stone, M. H., 39, 45, 61, 123, 238
unbounded, 227
unitary, 44, 106
transform
product Cayley, 239
infinite, 29 Fourier, 74, 80, 122, 213
semi-direct, 212
tensor, 120, 145 von Neumann, J., 24, 30, 45, 66, 71
projection, 40, 49
wavelet, 25, 39
quantum mechanics Wiener, N., 32, 166

Functional Analysis With Applications

Uploaded by

Functional Analysis With Applications

Uploaded by

Palle Jorgensen, and Feng Tian

arXiv:1408.1164v1 [math.FA] 6 Aug 2014

Functional Analysis with

“ ‘The Journal of Functional Analysis’ is dedicated to the broadening of the horizons of

“What goes around has come around, and today

There are already many books in Functional Analysis, so why another?

Part I Introduction and Motivation

Part II Topics from Functional Analysis and operators in Hilbert space: a

2 Unbounded Operators in Hilbert Space . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

3 The Spectral Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

Part III Applications

4 GNS and Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

5 Completely Positive Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

6 Brownian Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

7 Applications to Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

8 The Kadison-Singer Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219

Part IV Extension of Operators

9 Selfadjoint Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227

10 Unbounded Graph-Laplacians . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261

11 Reproducing Kernel Hilbert Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279

A An overview of Functional Analysis books (cast of characters) . . . . . . . 307

B Often cited above . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319

Motivation: More traditional books on Functional Analysis tend to postpone ap-

In chapter 5, and motivated by physics and harmonic analysis, we discuss dilation

B dynamical systems functions on fractals, Cantor set, etc.

C PDE Sobolev spaces

D numerical analysis discretization

E measures / probability theory probability space (Ω , F, P)

F quantum theory Hilbert spaces of quantum states

Examples of subjects within area A includes measure theory, transforms, con-

Subjects from area B include solutions to ordinary differential equations (ODEs),

“. . . the [quantum mechanical] observables are operators on a

I received an early copy of Heisenberg’s first work a little before

1.1 A Sample of Topics

• study of function spaces (Banach space, Hilbert space)

Within pure mathematics, it is manifested in

• representation theory of groups and algebras, among a long list of diverse

We will consider three classes of algebraic objects of direct functional analytic

• C∗ -algebras, von Neumann algebras

We will be considering C∗ -, and W ∗ -algebras axiomatically. In doing this we use

A wavelet is a special basis for a suitable L2 -space which is given by generators

Our approach to harmonic analysis will be general, – encompassing anyone of a

• analytic number theory

Definition 1.1. A linear function ϕ : E → R (or C) is said to be a linear functional,

ϕ (u + λ v) = ϕ (u) + λ ϕ (v) , ∀λ ∈ R, ∀u, v ∈ E.

Lemma 1.1. Let E be a normed space with dual E ∗ . For ϕ ∈ E ∗ , set

kϕk∗ := sup |ϕ (x)| .

Then (E ∗ , k·k∗ ) is a Banach space.

C∞ (R), system of E 0 all Schwartz ϕD ( f ) = D applied to f , f ∈ C∞ (R) .

Table 1.1: Examples of dual spaces.

Examples of Banach spaces are (i) l p , 1 ≤ p ≤ ∞, and (ii) L p (µ) where µ is a

• l p : all p-summable sequences.

• L p : all p-integrable function w.r.t. some fixed measure µ.

∑ ϕ (xi ) (F (xi+1 ) − F (xi )) , where x1 < x2 < · · · < xn .

Remark 1.1. For completeness of l p and of L p (µ), see [Rud87].

The following result is basic in the subject.

1.3 Transfinite Induction (Zorn and All That)

All finite words in the alphabet {0, 1}, continued indefinitely.

Fig. 1.2: Infinite Tree (no maximal element!)

Theorem 1.5 (Tychonoff). Let Aα be a family of compact sets indexed by α ∈ I.

1.4 Basics of Hilbert Space Theory

Definition 1.4. Let X be a vector space over C. A norm on X is a mapping k·k : X →

• kcxk = |c| kxk, c ∈ C, x ∈ X;

Definition 1.6. Let X be vector space over C. An inner product on X is a function

• hx, ·i : X → C is linear (linearity)

|hx, yi|2 ≤ hx, xi hy, yi , ∀x, y ∈ X. (1.2)

i.e., the matrix " #

and the corollary follows.

where the RHS is given by (1.3).

for all system of coefficients c1 , . . . , cn ∈ C, and all x1 , . . . , xn ∈ X.

and define a sesquilinear form on H0 by

by assumption. (All summations are finite.)

Corollary 1.2. Let X be a set, and let ϕ : X × X → C be a function. Then ϕ is

H = span {Φ (x) : x ∈ X} . (1.6)

Given two minimal solutions, Φi : X → Hi , i = 1, 2 (both satisfying (1.5)); then there

UΦ1 (x) = Φ2 (x) U, ∀x ∈ X. (1.7)