Lecture notes on

OPERATOR ALGEBRAS AND THEIR

arXiv:2208.10151v1 [math-ph] 22 Aug 2022

APPLICATION IN PHYSICS

August 23, 2022

Lecturer Dr. A. Kegeles

TU Bergakademie Freiberg
Institute for Geophysics
Geomathematics and Geoinformatics Group
Preamble
Around the year 1927 physicist, lead by Heisenberg and Schrödinger, discovered a new
theory for subatomic particles, called quantum mechanics. This theory radically differed
from classical mechanics in its physical assumptions as well as in its mathematical formu-
lation. The formal developments of quantum mechanics rapidly made it one of the most
successful physical theories. Nevertheless, its physical foundations were very obscure and
unjustifiable, and its mathematical basis were often shaky. To provide a clear physical
justification and a rigorous mathematical foundation for the theory, several mathemati-
cians, lead by von Neumann, Weyl, and Jordan, set out to define an algebraic approach
for quantum mechanics. In this lecture we follow their path and see how the physi-
cal formulation of the standard Schrödinger quantum mechanics appears from general
physical arguments, paired with the brilliant observations of Heisenberg, and impeccable
mathematics of von Neumann and friends...

In the first part of the lecture, we will develop the basic theory of operator algebras.
Here we will not make any connections to physics and treat the subject in its pure form. In
particular, we discuss the function calculus and the Gelfand-Naimark-Segal construction
of representations for C ‹ -algebras. These tools will give us the necessary mathematical
background to deal with C ‹ -algebras when they appear in physics.
In the second part of the lecture we establish a connection between pure mathe-
matics and physics. We will see how general physical principles lead us to an algebraic
description of physical theories. Combining this description with the Heisenberg’s uncer-
tainty principle and the canonical commutation relations developed by Born and Jordan
we end up with the C ‹ -algebra that describes a quantum mechanical particle. The fi-
nal piece of the puzzle follows from the Stone-von Neumann theorem that justifies the
almost exclusive use of Schrödinger’s formulation for quantum mechanics.

2
Contents
1 Operator algebras: Banach and C ‹ -algebras 4
1.1 Basic properties of operator algebras . . . . . . . . . . . . . . . . . . . . 4
1.2 Function calculus for C ‹ -algebras . . . . . . . . . . . . . . . . . . . . . . 18
1.3 The GNS construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.3.1 Order structure on the C ‹ -algebra . . . . . . . . . . . . . . . . . . 24
1.3.2 States of the C ‹ -algebra . . . . . . . . . . . . . . . . . . . . . . . 27
1.3.3 Representations of a C ‹ -algebra . . . . . . . . . . . . . . . . . . . 31

2 Operator algebras in physics 38

2.1 Classical mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.1.1 Kinematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.1.2 Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.2 An algebraic construction for physical theories . . . . . . . . . . . . . . . 43

3 A quantum mechanical particle 53

3.1 Semi-historical note . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.2 The Weyl algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.3 Stone-von Neumann theorem . . . . . . . . . . . . . . . . . . . . . . . . 59
3.4 The Schrödinger representation of the Weyl algebra . . . . . . . . . . . . 66

3
1 Operator algebras: Banach and C ‹-algebras

1.1 Basic properties of operator algebras

Section remarks: This section is based on the volume 1 and 3 of the series by Kadison
and Ringrose [1, 2]. Some of the reminders and definitions that concern more standard
statements of functional analysis are mostly from [3]. Further reading: for the abstract
approach to C ‹ -algebras and von Neumann algebras the book by Sakai [4] is great but
is a little more advanced; another very satisfying read is the book by Pedersen [5]; the
four volumes of the aforementioned Kadison and Ringrose bible are anyway a treat for
the soul: the books cover most of the topics of old fashioned C ‹ -algebras and develop
the theory very slowly and self-consistently.

In a nutshell, an algebra is a vector space in which we can multiply vectors. More formally
we define the algebra as follows.

Definition 1.1 (Algebra). An algebra A over a field K “ pR, Cq is a vector space over K,
equipped with multiplication that
i ) distributes over addition (from left and right):

f ¨ pg ` hq “ f ¨ g ` f ¨ h, pf ` gq ¨ h “ f ¨ h ` g ¨ h; (1)

ii ) is associative:

pf ¨ gq ¨ h “ f ¨ pg ¨ hq; (2)

iii ) is compatible with scalar multiplication:

λpABq “ pλAqB “ ApλBq pλ P K, A, B P Aq. (3)

A unit of an algebra A is an element I P A such that I ¨ A “ A ¨ I “ A for all A P A.

It is clear that if such an element exists it must be unique. However, not every algebra
contains a unite; if it does we call the algebra unital. We will always use unital algebras
in the following, even if we do not mention it explicitly. In one of the exercises, we will
see that one can always adjoint a unite to an algebra to make it unital. For unital algebras
we can shorten the definition and state.

Definition 1.2. An algebra A over a field K is a vector space, whose underlying additive
group has a ring structure compatible with multiplication with scalars.

To recall the definitions of the vector spaces and rings consult the following remarks.

4
 Reminder: Vector space. A vector space over a field K is a set of points V equipped
with vector addition ` : V ˆ V Ñ V and scalar multiplication ¨ : F ˆ V Ñ V . These
operations have to satisfy the following requirements.
For f, g P V and α, β P K the scalar multiplication has to
i ) distribute over the vector addition: αpf ` gq “ αf ` αg,
ii ) distribute over the field addition: pα ` βqf “ αf ` βf ,
iii ) be compatible with the field multiplication: pαβqf “ αpβf q,
iv ) has the neutral element: 1f “ f .
The vector addition has to be commutative and V with ` : V Ñ V must be a
commutative group. That is for f, g, h P V and α, β P K the addition
i ) is commutative (aka. Abelian): f ` g “ g ` f ,
ii ) is associative: pf ` gq ` h “ f ` pg ` hq,
iii ) has a neutral element e P V : e ` f “ f
iv ) is such that for every f P V there is an inverse f ´1 P V : f ` f ´1 “ e.

Reminder: Ring. A ring is a commutative group pR, `q equipped with multiplica-

tion ¨ : R ˆ R Ñ R, such that with f, g, h P R the multiplication
i ) distributes from the left over addition: f ¨ pg ` hq “ f ¨ g ` f ¨ h
ii ) distributes from the right over addition: pf ` gq ¨ h “ f ¨ h ` g ¨ h
iii ) is associative: pf ¨ gq ¨ h “ f ¨ pg ¨ hq
iv ) has a neutral element I P R: I ¨ f “ f ¨ I “ f .

Example 1.3. To visualize the concept let us recall some algebras that we know and love.
i ) The simplest example of an algebra is the algebraic field K itself.
ii ) A more elaborate example is the space Mpn, Cq of n ˆ n matrices with complex
coefficients. This algebra is unital. The unit being the identity matrix.
iii ) Yet another example is the space R ˆ R with the Hadamard product defined by

p1, 0q ¨ p1, 0q “ p1, 0q, p0, 1q ¨ p0, 1q “ p0, 1q, p1, 0q ¨ p0, 1q “ p0, 0q, (4)

and required to be compatible with multiplication by scalars. Notice, however, that

this is not quite natural on the vector space R2 . This is because the definition of
the product depends on the choice of the basis. In particular, this product changes
under rotations, which is why it rarely appears in applications to physics.

5
 iv ) Finally, consider the space of all polynomials in a single variable. We will call this
space P l. It is a vector space, and equipped with the pointwise product it becomes
a unital algebra, whose unit is the constant 1.

Definition 1.4. A norm } ¨ } on an algebra A is said to be continuous if for A, B P A it

satisfies }A ¨ B} ď }A} }B}.
i ) An algebra with a continuous norm is called a normed algebra.
ii ) A normed algebra that is complete (in the metric induced by the norm) is called a
Banach algebra.

As we will see, Banach algebras can be quite small (that is contain few elements).
Nevertheless, they are very general, as they have only little structure compared to the
algebras we will primarily use. We define this additional structure as follows.

Definition 1.5. Let A be a Banach algebra. A mapping x ÞÑ x‹ of A into itself is called

an involution if it satisfies the following conditions:
i ) px‹ q‹ “ x,
ii ) px ` yq‹ “ x‹ ` y ‹ ,
iii ) pxyq‹ “ y ‹ x‹ ,
iv ) pλxq‹ “ λ̄x‹ , for λ P C and λ̄ being its complex conjugate.

Reminder: Adjoint. Let BpHq be the space of bounded operators on the Hilbert
space H, and let x¨, ¨y denote the inner product on H. Then, for A P BpHq we define
its adjoint as the operator A: P BpHq such that

xx, Ayy “ xA: x, yy px, y P Hq. (5)

It is not difficult to see that the adjoint on the space of bounded linear operators on
a Hilbert space satisfies all the requirements of Definition 1.5. Indeed, the adjoint is
the reason why we made these requirements in the first place. However, in the case of
bounded linear operators, we define an adjoint using the scalar product and then check
that it satisfies the properties in the Definition. For algebras, the logic is reversed. We
don’t have the scalar product at our disposal, and so we require the properties as the
defining features of the mapping. In this sense the involution, sometimes simply called
the star (operation), generalizes the concept of the adjoint for bounded linear operators.
With this additional structure we can now define the algebras, that will accompany
us for the rest of the lecture.

Definition 1.6. A Banach ‹-algebra B is a Banach algebra with an (isometric) involution,

that is with the property that }A‹ } “ }A} for every A P B.

6
Definition 1.7. A C ‹ -algebra A is a complex Banach ‹-algebra that has the C ‹ -property:

}A‹ A} “ }A}2 pA P Aq. (6)

In the definition of Banach ‹-algebras some authors do not require the involution to
be isometric. In our case, this difference will not play a big role, because we will only
use either Banach algebras (without the star), or C ‹ -algebras. For the latter, however,
the involution is always isometric: since if A is an element in the C ‹ -algebra A then the
C ‹ -property implies

}A}2 “ }A‹ A} ď }A‹ }}A}, (7)

and thus }A} ď }A‹ }. Upon replacing A Ñ A‹ we get the inverse inequality and thus
}A} “ }A‹ }. Notice, however, that the converse is not true: an isometric involution is
not enough to ensure the C ‹ -property. Thus, in general C ‹ -algebras have more structure
than Banach ‹-algebras.
The name giving for the elements of the Banach ‹-algebra follows the usual conven-
tion. We introduce the standard notation.

Common notation. Let A be a unital algebra with an involution. For A P A we call

the element A‹ the adjoint of A. We say that A is
i ) self-adjoint if A‹ “ A;
ii ) normal if AA‹ “ A‹ A;
iii ) unitary if AA‹ “ A‹ A “ I, with the unit I P A;
iv ) invertible if there exists a B P A with AB “ BA “ I. If such B exists it is
unique and we call it the two-sided inverse of A and denote it by A´1 ;
v ) self-inverse if A “ A´1 ;
It follows that the unit I is self-adjoint, self-inverse, and unitary since

pI ‹ Iq‹ “ pI ‹ q‹ “ I “ I ‹ I “ I ‹ . (8)

Further, we can state the following:

i ) The set of all self-adjoint elements of A is a real vector space.
ii ) The set of all unitary elements of A forms a multiplicative group.
iii ) Each element A P A can be uniquely expressed as H `ıK, where H “ 21 pA`A‹ q
and K “ 2ı pA‹ ´ Aq are self-adjoint elements of A, called the real and the
imaginary part of A, respectively. Moreover, A is normal if and only if H and
K commute.

Proposition 1.8. Let A be an element in a C ‹ -algebra. Then A has a (two-sided) inverse

if and only if A‹ is invertible and then pA‹ q´1 “ pA´1 q‹ .

7
Proof. Calculate

pA´1 q‹ A‹ “ pAA´1 q‹ “ 1, (9)

and

A‹ pA´1 q‹ “ pA´1 Aq‹ “ 1. (10)

It follows that pA´1 q‹ is a two-sided inverse of A‹ . By uniqueness of the two-sided inverse

we get pA´1 q‹ “ pA‹ q´1 .

Before we look at some examples of C ‹ -algebras, let us quickly refresh some basic
definitions in topology.

Reminder: Hausdorff space. A topological space X is called Hausdorff, if for every

pair of distinct points x, y P X there exists two open sets U, V Ă X with U X V “ H
and x P U and y P V .

Reminder: Compact set. A subset K of a topological space X is compact, if each

open cover of K, contains a finite sub-cover of K.

Example 1.9 (Commutative C ‹ -algebra). Let X be a compact Hausdorff space. Let

Cb pXq denote the space of (bounded) complex valued continuous functions on X. Define
a ring structure on Cb pXq by the pointwise product: for f, g P Cb pXq define pf gqpxq :“
f pxqgpxq, px P Xq. Next, topologize Cb pXq by the supremum norm: for f P Cb pXq
set }f } “ supxPX |f pxq|. Define an involution on Cb pXq by complex conjugation: for
f P Cb pXq define f ‹ “ f¯. Then Cb pXq is a commutative unital C ‹ -algebra and the unit
is a constant function I “ 1.

Proof. Clearly, Cb pXq is a vector space. Since a pointwise product of two continuous
functions is again a continuous function, the space Cb pXq is algebraically closed. Since
the uniform limit of continuous functions is continuous, the space Cb pXq is complete.
Hence, Cb pXq with the supremum norm and the pointwise product is a Banach algebra.
It is straight forward to check that complex conjugation defines an involution. In fact,
to a great extent complex conjugation inspired our definition of the involution. Hence,
Cb pXq is a Banach ‹-algebra. (Here we do not assume the involution to be continuous.
The continuity will follow, after we prove the C ‹ -property.) For the C ‹ -property we
observe for f P Cb pXq: |f pxqf¯pxq| “ |f pxq||f¯pxq| “ |f pxq||f pxq| “ |f pxq|2 , at any x P X.
Hence,
˘2
}f f¯} “ sup |f pxqf¯pxq| “ sup |f pxq|2 “ sup |f pxq| “ }f }2 . (11)
`
xPX xPX xPX

This completes the proof.

8
 Comment: Indeed, by the Gelfand transform (that we meet in a few lectures) we
will see that every unital commutative C ‹ -algebra is isometrically isomorphic to a
C ‹ -algebra of continuous functions on some compact Hausdorff space.

Now we consider a non-commutative example of a C ‹ -algebra.

Example 1.10. Consider our previous example: the algebra Mpn, Cq of n ˆ n matrices
with complex coefficients. Equipped with the operator norm and adjoint as an involution,
Mpn, Cq is a unital C ‹ -algebra.

Proof. This is a special case of the following example.

Example 1.11. Let H be a separable Hilbert space. Denote by BpHq the space of linear
bounded operators on H. Then with the operator product, the operator norm, and the
adjoint as involution, the space BpHq is a unital C ‹ -algebra.

Proof. We only show the C ‹ -property as the rest is trivial. Let A be in BpHq then

}A}2 “ sup }Ax} “ sup xAx, Axy “ sup xx, A: Axy “ }A: A}. (12)
xPH; }x}“1 xPH; }x}“1 xPH; }x}“1

The last equality holds because A: A is a self-adjoint operator.

Whenever we speak about C ‹ -algebras it is enlightening to consider, in the back of

your head, the space of linear bounded operators on a Hilbert space. However, you
should keep in mind that an abstract C ‹ -algebra is more general than BpHq for a fixed
Hilbert space H; it is more general because a C ‹ -algebra is not defined by the action
of operators on a Hilbert space. Instead, it is only a collection of elements that satisfy
some relations under multiplication. In particular, let A be an element of a BpHq. To
characterize it we often define its action on H. For example, if A maps a vector v in H
to a rotated vector ṽ, we say that A is a rotation. But how do we characterize elements
in an abstract algebra if we don’t define their action... we use their spectrum.

Definition 1.12. If A is an element of a Banach algebra A we say that λ P C is a spectral

value for A (relative to A) when A ´ λ1 does not have a two-sided inverse in A. The set
of spectral values of A is called the spectrum of A and is denoted spA pAq.

Example 1.13. Let A “ Cb pXq be the C ‹ -algebra of continuous real-valued functions on

a compact Hausdorff space X. For f P A what is its spectrum?
For λ R Imgpf q (the image of f ) consider f ´λ1. The function f ´λ1 is continuous and
defined on a compact domain. Hence, it attains all its minima. If f ´ λ1 never vanishes
then inf |f ´ λ1| ą 0 and the function g “ f ´λ
1
is continuous. Clearly, g is the inverse of
f ´ λ1. Hence spA pf q Ă Imgpf q .

9
 Take λ P Imgpf q, so for some x P X we have pf ´λ1qpxq “ 0. Since f ´λ1 is continuous
there is a sequence pxn q in X that converges to x and satisfies pf ´ λ1qpxn q ă 1{n. Let g
be the two-sided inverse such that pf ´ λ1qg “ gpf ´ λ1q “ 1. Then
1
(13)
` ˘
1 “ | pf ´ λ1qg pxn q| “ |gpxn q||pf ´ λ1qpxn q| ď |gpxq| ,
n
for every xn . Thus, |gpxn q| ě n, and g is not continuous at x. Consequently g R Cb pXq
and Imgpf q Ă spA pf q. Together with the opposite inclusion we get spA pf q “ Imgpf q.

Example 1.14. Let M be the C ‹ -algebra of complex n ˆ n matrices. For M P M consider

M ´ λ1. This element is invertible if and only if detpM ´ λ1q ‰ 0. Hence, if M is
diagonalizable its spectrum is the set of eigenvalues of M .

Warning. For a Banach algebra the spectrum of an element does indeed depend
on the algebra. Therefore, we use the subscript to specify the ambient algebra. The
following example shows that changing the algebra can change the spectrum.

Example 1.15. Let H be a separable Hilbert space with the basis tej u for j P Z. Let BpHq
be the Banach algebra of linear bounded operators on H. (Above we have shown that
BpHq is a C ‹ -algebra, so in particular it is a Banach algebra). Consider the left shift L in
BpHq defined on the basis of H by Lpej q “ ej`1 and extended by linearity and continuity
to the whole H. The adjoint of L is the operator L: pej q “ ej´1 . One can easily verify that
LL: “ L: L “ 1. Thus, L: is the two-sided inverse of L. In other words, 0 R spBpHq pLq.
Now consider a Banach algebra B generated by 1 and L. It is a subalgebra of BpHq.
Is it still true that 0 is not in the spectrum of L relative to B? No. To see this consider
a closed subspace S of H spanned by tej : j “ 1, 2, . . . u. Then L maps this subspace
into itself, and thus, each polynomial of L leaves this subspace invariant. It follows that
every element in B (the norm closure of such polynomials) maps S into itself. Assume
that L̃ is a two-sided inverse of L in B. Then L̃ is also in BpHq since B Ă BpHq. By the
uniqueness of the two-sided inverse we thus have L: “ L̃. However, L: does not preserve
S as it maps the element e1 to e0 R S. Thus, L: is not in the algebra B. We can thus
conclude that L does not have a two-sided inverse in B and 0 P spB pLq.

The above example shows that the spectrum of an element depends on the ambient
Banach algebra. One of the main results that we will obtain in the first part of the lecture
will be that for C ‹ -algebras this cannot happen.

Lemma 1.16. For a Banach ‹-algebra A, let A be in A and let spA pAq be its spectrum.
Then the spectrum of A‹ is
(´ ¯
spA pA‹ q “ λ̄ : λ P spA pAq “ spA pAq . (14)

10
Proof. If λ R spA pAq then pA ´ λIq´1 exists in A. Hence
˘´1 ˘‹
(15)
` `
pA ´ λIq‹ “ pA‹ ´ λ̄Iq´1 “ pA ´ λIq´1 ,
˘´1
and pA ´ λIq‹ exists. Thus, λ̄ R spA pA‹ q. Interchanging A and A‹ in the above
argument yields the desired conclusion.

Lemma 1.17. For a Banach algebra A, let A be an invertible element in A. Then the
spectrum of A´1 is
´ ¯
´1 ´1
spA pA q “ tλ : λ P spA pAqu “ spA pAq ´1
(16)

Proof. If A is invertible then 0 R spA pAq. Thus, if λ is in spA pAq then λ ‰ 0 and

A´1 ´ λ´1 1 “ pAλq´1 pλ ´ Aq. (17)

Hence, A´1 ´ λ´1 1 is invertible if and only if A ´ λ1 is invertible, meaning

spA pA´1 q “ tλ´1 : λ P spA pAqu. (18)

This shows the assertion.

In order to study the spectrum of an element it is good to have better understanding

of invertible elements. The next few lemmas provide us with this understanding.

Lemma 1.18 (Neumann series). If A is an element of a Banach algebra A and }A} ă 1

then
n
ÿ
Ak pwith A0 “ Iq, (19)
k“0

has a limit B P A as n goes to infinity. Moreover, it holds that

BpI ´ Aq “ pI ´ AqB “ I (20)

and B is the (two-sided) inverse of I ´ A.

Proof. For n ă m we can estimate

n
ÿ m
ÿ m
ÿ m
ÿ
} k
A ´ k
A }“} k
A }ď }A}k . (21)
k“0 k“0 k“n`1 k“n`1

Since }A} ă 1 the right-hand side of the above estimate goes to zero. Thus, nk“0 Ak is a
ř

Cauchy sequence. Since A is complete this sequence converges to an element B in A as

n goes to infinity. Then
n n n n
`ÿ ÿ ÿ `ÿ
(22)
k
˘ k k`1
˘
A pI ´ Aq “ A ´ A “ pI ´ Aq Ak “ I ´ An`1 .
k“0 k“0 k“0 k“0

11
 nÑ8
By continuity of the product we have }An`1 } ď }A}n`1 ÝÑ 0. Thus An`1 goes to zero
as n goes to infinity. Thence the right-hand side of (22) converges to I and we obtain
(again by continuity of multiplication)

BpI ´ Aq “ pI ´ AqB “ I. (23)

This completes the proof.

Proposition 1.19. If A is a Banach algebra, the set N Ă A of invertible elements in A is

an open subset of A and the operation

inv : A Ñ A´1 (24)

on N is continuous.

The following definition will be convenient for the proof of the proposition.

Definition 1.20. Let A be an element of a Banach algebra A. We define the left multi-
plication by A as the mapping LA : A Ñ A such that LA pBq “ AB. In the similar way we
define the right multiplication by A as the mapping RA : A Ñ A such that RA pBq “ BA.

Info: From the (joint) continuity of the product in a Banach algebra it follows directly
that LA and RA are continuous for every A in A.

Proof of proposition 1.19. Even though the proof is not complicated, it is still helpful to
understand the general idea of the argument.

Idea: To show that N is an open set we need to show that for every element in
N there exists an open set containing this element and that is contained in N . We
know that an open ball of radius 1 around the identity I is in N . The idea is to shift
this ball appropriately to prove the proposition. The multiplication operators that
we just defined provide the “shifting”. We now formalize the idea.

For A P N let LA be the left multiplication by A. Since A is invertible then so is LA

with the inverse LA´1 . As we mentioned above LA and LA´1 are both continuous (thus
LA is a homeomorphism).
From the Neumann series we know that B1 pIq — an open ball of radius 1 around I
— is a subset of N . Since LA is a homeomorphism (in particular open) it maps B1 pIq
to an open set U in A. The set U contains A since I P B1 pIq and LA pIq “ A; also U is
contained in N since if A and B are invertible then so is AB (with the inverse B ´1 A´1 ).
Thus, for every A P N there is an open set LA pB1 pIqq Ă N containing A and N is open.

12
 LA

A A

N N
A A

Figure 1: Idea of proof.

Now we prove that the mapping inv is continuous on N . Assume that B P A is such
that }I ´ B} ă 1, then by the Neumann series
8
ÿ ˘´1
(25)
`
pI ´ Bqk “ I ´ pI ´ Bq “ B ´1 .
k“0

Choosing }I ´ B} ă  ă 2´1 and the fact that I is self-inverse (i.e. II “ I) we can

estimate
8
ÿ
} invpBq ´ invpIq} “ }B ´1
´ I} ď }I ´ B}k (26)
k“1
8
´ÿ ¯ }I ´ B}
“ }I ´ B} k
}I ´ B} “ ď 2}I ´ B} ď 2. (27)
k“0
1 ´ }I ´ B}

Thus inv is continuous at the identity. To show that it is continuous at every element
observe that for any B P N the following identity holds

inv “ RB ´1 ˝ inv ˝LB ´1 . (28)

Thus for each fixed element A P N choosing the right-hand side of the above equality
to be RA´1 ˝ inv ˝LA´1 shows that the map inv is a combination of continuous mappings
L inv R
A´1 . Thus for each A P N the mapping inv is continuous.
A ´1A ´1
A Ñ IÑI Ñ

We will now show that every element in a Banach algebra has a spectrum. The
proof uses some known, but not entirely trivial statements from the complex and func-
tional analysis. Since, these topics are themselves very extensive, we only recall the basic
needed statements before continuing with the theorem.

13
 Reminder: Holomorphic function. A complex valued function f is called holo-
morphic at z0 P C if on some neighborhood of z0 the limit

f pzq ´ f pz0 q
lim (29)
zÑz0 z ´ z0
exists. A function that is holomorphic at every point of some open set O Ă C is said
to be holomorphic on O. If f pzq is holomorphic on an entire complex plain we call
it an entire function. By Liouville’s theorem, every bounded entire function must be
constant. In particular, if an entire function vanishes at infinite it must be zero.

Reminder: Hahn-Banach theorems. The name “Hahn-Banach theorem” refers to

several related results (consult e.g. [1, chap. 1 2]). There are two types of such the-
orems: separation and extension results. One version of the Hahn-Banach extension
theorem is the following.

Theorem. Let X be a vector space and X0 be a subspace of X. Let ρ0 : X0 Ñ C (or R)

be a linear functional, and p : X Ñ R be a semi-norm such that

|ρ0 pyq| ď ppyq py P X0 q. (30)

Then there exists a linear functional ρ : X Ñ C (or R) such that

|ρpxq| ď ppxq px P Xq, ρpyq “ ρ0 pyq py P X0 q. (31)

The corollaries of this very general theorem are often more useful than the theo-
rem itself. For example the next corollary assures that we can always extend a linear
functional.

Corollary. Let X0 be a subspace of a normed space X. Let ρ0 be a bounded linear

functional on X0 . There exists a bounded linear functional ρ on the whole X such that
}ρ} “ }ρ0 } and ρpxq “ ρ0 pxq for x P X0 .

Another useful corollary is the one we use in the proof of Theorem 1.21.

Corollary. If x is a non-zero vector in a normed space X, there is a bounded linear

functional ρ on X such that }ρ} “ 1 and ρpxq “ }x}.

Equipped with all that fundamental knowledge. We can now prove that the spectrum
of an element in a Banach algebra is never empty.

Theorem 1.21. If A is an element of a Banach algebra A then spA pAq is a non-empty closed
subset of the closed disc in C with center 0 and radius }A}.

14
Proof. Let λ P C be such that |λ| ą }A} then
´1 ¯
A ´ λI “ λ A´I . (32)
λ
Since }Aλ´1 } ă 1 the element λ1 A ´ I is invertible (by the Neumann series). Thus, A ´ λI
has the inverse λ1 p λ1 A ´ Iq´1 . It follows that spA pAq Ă D}A} p0q (a disc with center 0 and
radius }A}).
Now we need to show that spA pAq is not empty. Let λ R spA pAq. Then by the proposi-
tion 1.19, the element A´λI is invertible in a small neighborhood around λ. In particular
the set Cz spA pAq is open.

Reminder: Resolvent equation. Recall the first resolvent equation [3, thm. VI.5]

pA ´ λ1 Iq´1 ´ pA ´ λIq´1 “ pλ1 ´ λqpA ´ λ1 IqpA ´ λIq´1 . (33)

Let λ1 be close enough to λ so that A ´ λ1 I is invertible. Let ρ : A Ñ C be a continuous

linear functional on A. Then by the resolvent equation
´ ¯ ´ ¯
1 ´1 ´1
ρ pA ´ λ Iq ´ ρ pA ´ λIq ´ ¯ 1 ´ ¯
1 ´1 ´1 λ Ñλ ´2
“ ρ pA ´ λ Iq pA ´ λIq Ñ ρ pA ´ λIq .
λ1 ´ λ
The last limit follows by the continuity of ρ, the product, and the mapping inv on N . Thus
the function λ ÞÑ ρpA ´ λIq is holomorphic on Cz spA pAq. Also since p Aλ ´ Iq approaches
´I as |λ| grows, we have
´ ¯ 1 ´ 1 ¯
ρ pA ´ λIq ´1
“ ρ p A ´ Iq ´1
Ñ 0, for |λ| Ñ 8. (34)
λ λ
Now, if spA pAq were empty, the function λ ÞÑ ρpA ´ λIq would be an entire function that
vanishes at infinity. By the Liouville’s theorem we then had ρpA ´ λIq´1 “ 0 for all λ P C
and in particular for λ “ 0. But then ρpA´1 q “ 0 for every continuous linear functional ρ,
and by the Hahn-Banach theorem A´1 “ 0. This is a contradiction. Hence spA pAq ‰ H.
This completes the proof.

The above theorem states that the spectrum of an element in the Banach algebra is
bounded by its norm. However, the norm is just an upper bound not the smallest upper
bound. To study how different or similar is the norm and the smallest upper bound of
the spectrum, we give the latter quantity a proper name first.

Definition 1.22 (Spectral radius). The spectral radius rA pAq of an element A in a Banach
algebra A is

rA pAq “ supt|λ| : λ P spA pAqu (35)

15
 C

}A}

spA pAq
rA pAq

Figure 2: Spectrum of A relative to A bounded by the spectral radius and by }A||.

Info: The spectral radius is the radius of the smallest disc in C with center 0 contain-
ing spA pAq. Also from the preceding theorem, }A} is the radius of the largest disc in
C with center 0 that contains spA pAq. Hence we always have rA pAq ď }A}.

Surprisingly, the spectral radius — rather an algebraic property — can be stated in

analytical terms, and thus it relates to topology. For the proof consult e.g. [1, thm. 3.3.3]
Theorem 1.23 (Spectral radius formula). Let A be an element of a Banach algebra A.
The sequence p}An }1{n q converges to rA pAq such that
rA pAq “ lim }An }1{n . (36)
nÑ8

This relation between the spectrum and topology is as surprising as it is powerful.

We have already seen that the spectrum of an element depends on the ambient Banach
algebra. Nevertheless, from the spectral radius formula if follows that if B is a Banach
subalgebra of A then the spectral radii of A P B and A P A coincide, i.e. rA pAq “ rB pAq.
In a C ‹ -algebra the spectral radius formula provides even stronger results.
Proposition 1.24. Suppose that A is an element of a C ‹ -algebra A. Then
i ) if A is normal then rApAq “ }A};
ii ) if A is self-adjoint then spA pAq is a compact subset of the real line and it contains at
least one of the two real numbers ˘}A};
iii ) if A is unitary then }A} “ 1 and spA pAq is a compact subset of the unite circle ta P
C : |a| “ 1u.
Proof. i ) Let H be a self-adjoint element in A and n P N. Then by the C ‹ -property we
have
}H 2n } “ }pH n q‹ H n } “ }H n }2 . (37)

16
Hence, if q “ 2m for some m “ 1, 2, . . . then by induction we get

}H q } “ }H}q . (38)

Using the spectral radius formula we get

rA pHq “ lim }H q }1{q “ }H}. (39)

qÑ8

So the assertion holds for self-adjoint elements. Now, let A be normal and H “ A‹ A then

p1q p2q p3q p4q p5q

}A}2 “ }A‹ A} “ rA pA‹ Aq ď rA pA‹ qrA pAq “ rA pAq2 ď }A}2 . (40)

With the following justifications for the estimates: (1) is the C ‹ -property; (2) follows
from what we have just shown above; (3) follows from the fact that for normal elements
}pA‹ Aqn } ď }pA‹ qn }}An } and thus lim }pA‹ Aqn }1{n ď lim }pA‹ qn }1{n }An }1{n ; (4) follows
from the fact that spA pA‹ q “ spA pAq; (5) follows since }A} defines the largest radius of
a disc that contains the spectrum.
Since the left and the right hand side of the above estimate are equal all the "less or
equal" signs in the estimate have to be actual equalities. Hence rA pAq “ }A}.

ii ) By Theorem 1.21 the spectrum of A P A is a compact set of C. Thus the spectral

radius attains its maximum and there is a λ P spA pAq such that |λ| “ rA pAq. If A is
self-adjoint, then from the first part of the proposition we also know that rA pAq “ }A}
and thus there is an element λ P spA pAq with |λ| “ }A}. To prove the assertion we only
need to show that the spectrum of a self-adjoint element is a subset of a real line.
Assume that for some a, b P R the element a ` ıb is in spA pAq. For n P N consider the
element Bn “ A ` ınbI. Then, since a ` ıpn ` 1qb P spA pBn q, we have

|a ` ıpn ` 1qb|2 ď }Bn }2 ď }Bn‹ Bn } “ }A2 ` n2 b2 } ď }A}2 ` n2 b2 . (41)

On the other hand,

|a ` ıpn ` 1qb|2 “ pa ` ıpn ` 1qbqpa ´ ıpn ` 1qbq “ a2 ` n2 b2 ` 2nb2 ` b2 . (42)

So combining (41) and (42), we get a2 ` p2n ` 1qb2 ď }A}2 for every n P t0, 1, 2, . . . u.
Since A is, however, bounded b must be zero.

iii ) If A is unitary in A, then }A}2 “ }A‹ A} “ 1. Thus }A} “ 1. Therefore, if λ is in

spA pAq then |λ| ď }A} ď 1. On the other hand, since the inverse of A is its adjoint A‹
we have from Lemma 1.17 that λ´1 P spA pA´1 q “ spA pA‹ q and |λ´1 | ď }A‹ } “ }A} ď 1.
Thus, |λ| ď 1 and |λ´1 | ď 1 meaning that |λ| “ 1.

17
 Algebra vs. bounded operators on a Hilbert space. For bounded operators on
a Hilbert space we can use the inner product to prove that a self-adjoint operator
on a Hilbert space has a real spectrum. This makes the proof simpler (but also less
general). Consider a self-adjoint bounded linear operator A on a Hilbert space H.
Let λ P C be such that | Impλq| ą 0. Then for every x P H we have by the Cauchy-
Schwarz inequality,

}x}}pA ´ λ1qx} ě |xx, pA ´ λ1qxy| ě | Impλq|}x}2 ą 0. (43)

In particular, }pA ´ λ1qx} ě | Impλq|}x} for every x P H. Hence A ´ λ1 is injective

and has closed range. Also, since | Impλ̄q| “ | Impλq|, with the same argument we
get that pA ´ λ1q: is also injective, and A ´ λ1 has dense range in H. It follows that
A ´ λ1 is invertible and λ R spBpHq pAq.

1.2 Function calculus for C ‹ -algebras

Function calculus on C ‹ -algebras generalizes the idea of diagonalizing matrices. Diag-
onalizing a matrix allows us to define powers, square roots, exponential, logarithms of
matrices, and what not... For bounded operators on a Hilbert space the generalization
leads to the spectral calculus. The functional calculus in C ‹ -algebras essentially bears
the same meaning. We can even consider it a generalization of the spectral calculus, as it
is defined for abstract algebra elements. Generally speaking the function calculus allows
us to associate algebra elements to continuous functions, calculate with those functions,
and then associate the result back to algebra elements. This process, is very powerful.
It is hard to overestimate the use of the function calculus in the abstract theory and
applications...
In the following we will use the following notation. If p is a polynomial, say pptq “
k“0 ak t with ak P C, then we will write ppAq to denote the algebra element k“0 ak A .
řn k
řn k

Proposition 1.25 (Spectral calculus for polynomials). If A is an element in a Banach

algebra A and p is a polynomial in one variable, then
´ ¯
spA pppAqq “ tppλq : λ P spA pAqu “ ppspA pAqq . (44)

Proof. If λ P spA pAq, then A ´ λI has no two-sided inverse in A. If ppxq “ an xn ` ¨ ¨ ¨ ` a0

then

ppAq ´ ppλqI “ an pAn ´ λn Iq ` ¨ ¨ ¨ ` a1 pA ´ λIq. (45)

For k ě 2 we can rewrite Ak ´ λk I as

Ak ´ λk I “ pA ´ λIqpAk´1 ` λAk´2 ` ¨ ¨ ¨ ` λk´1 Iq

“ pAk´1 ` λAk´2 ` ¨ ¨ ¨ ` λk´1 IqpA ´ λIq.

18
Hence, we get
` ˘ ` ˘
ppAq ´ ppλqI “ pA ´ λIq an pAn´1 ` λAn´1 ` ¨ ¨ ¨ ` λn´1 Iq ` ¨ ¨ ¨ ` a1 “ ¨ ¨ ¨ pA ´ λIq,

and ppAq ´ ppλqI does not have a two-sided inverse (otherwise A ´ λI would have a
two-sided inverse). It follows that ppλq P spA pppAqq, and thus ppspA pAqq Ă spA pppAqq.
For the opposite inclusion consider that if γ is in spA pppAqq and λ1 , . . . , λn are the n
roots of ppxq ´ γ then

ppAq ´ γI “ cpA ´ λ1 Iq ¨ ¨ ¨ pA ´ λn Iq, (46)

and at least one of A ´ λj I is not invertible. Hence, λj P spA pAq for at least one of
λj ’s and by previous discussion γ “ ppλj q “ spA pppAqq (the first equality follows since
ppλj q ´ γ “ 0 by definition of roots). Hence, spA pppAqq Ă ppspA pAqq.

To unclutter the notation in the proofs we will use the following common notations.
We summarize them in the following paragraph.

Common notation. By P l we will denote the space of polynomials with complex

coefficients. If A is an element in a C ‹ -algebra A with the spectrum spA pAq then,

• by CpspA pAqq we will denote the C ‹ -algebra of continuous functions on spA pAq;

• by id : spA pAq Ñ spA pAq we will denote the identity mapping z ÞÑ z.

• when we write P l Ă CpspA pAqq we consider elements of P l as continuous func-

tions on spA pAq.

It is time to introduce the function calculus for self-adjoint operators. Almost the same
theorem, with minor modifications, holds also for normal elements in a C ‹ -algebra (see
the Remark at the end of the section). However, we will prove only the following result,
as it does not require any additional technicalities that are needed for normal elements.
Before the proof, we recall the Weierstrass approximation theorem for convenience.

Reminder: Weierstrass approximation theorem. Any continuous function on a

bounded interval can be uniformly approximated by polynomial functions.

Theorem 1.26 (Function calculus for self-adjoint elements). Let A be a self-adjoint

element of a C ‹ -algebra A. There is a unique isometric ‹-isomorphism ϕ from CpspA pAqq
into A such that ϕpid q “ A. Moreover, for f P CpspA pAqq,
i ) ϕpf q is a normal element and it is self-adjoint if and only if f is real valued;
ii ) the set tϕpf q : f P CpspA pAqqu is a commutative C ‹ -algebra and it is the smallest
C ‹ -subalgebra of A that contains A;

19
 iii ) ϕpf q is the limit of a sequence of polynomials in I and A;
iv ) if B is in A that commutes with A, then it also commutes with ϕpf q.

Proof. Assume a polynomial p P P l of the form,

n
ÿ
pptq “ ak tk . (47)
k“0

By Proposition 1.24 ii ), the spectrum of A is a compact subset of the real line. Then we
can write p restricted to this subset using the identity mapping in CpspA pAqq such that
n
ÿ
p“ ak id k pwith id 0 “ 1q. (48)
k“0

Define ϕ0 : P l Ñ A by
n
ÿ n
ÿ
ϕ0 ppq “ ak Ak “ ppAq and ϕ0 ppq‹ “ āk Ak “ ppAq‹ . (49)
k“0 k“0

Then, ϕ0 is linear and ϕ0 pid k q ÞÑ Ak for k “ N. Also, ϕ0 ppq and ϕ0 ppq‹ commute, implying
that ϕ0 ppq is normal. Thus, Proposition 1.24 i ) applies and together with Proposition 1.25
we have

}ϕ0 ppq} “ rA pppAqq “ supt|λ| : λ P spA pppAqqu “ supt|ppλq| : λ P spA pAqu “ }p}, (50)

where the norm on the left-hand side is the C ‹ -norm on A and the norm on the right-
hand side is the supremum norm on CpspA pAqq. Thus, ϕ0 is an isometry. If p and q are
two polynomials in P l, identically equal on spA pAq, then

}ϕ0 ppq ´ ϕ0 pqq} “ }p ´ q} “ 0, (51)

and then ϕ0 ppq “ ϕ0 pqq. Thus, the mapping ϕ0 is well defined (unambiguous).
By the Weierstrass approximation theorem, P l Ă CpspA pAqq is everywhere dense in
CpspA pAqq. Since A is complete, ϕ0 extends to the mapping ϕ defined on CpspA pAqq. This
extension remains isometric; since if f P CpspA pAqq and tpn u is a sequence of polynomials
that approximates f then,

}ϕpf q} “ lim }ϕppn q} “ lim }pn } “ }f }. (52)

nÑ8 nÑ8

(This proves existence of the asserted mapping.)

Thence, since CpspA pAqq is complete its image, tϕpf q : f P CpspA pAqqu, is complete
as well. Therefore, it is a commutative C ‹ -subalgebra of A that contains I and A. We
will denote this subalgebra ApAq. Since polynomials are everywhere dense in CpspA pAqq
then every element of ApAq is a limit of a sequence of polynomials in A. A closed C ‹ -
subalgebra B of A that contains I and A necessarily contains all polynomials in A and

20
therefore it contains ApAq. Hence ApAq is the smallest C ‹ -subalgebra that contains I and
A. (This proves ii ) and iii ).)
By definition, ϕ0 is linear, multiplicative (i.e. ϕppqq “ ϕppqϕpqq for p, q P P l), preserves
the star (i.e. ϕpp̄q “ ϕppq‹ ), and carries the unite of P l onto that of A. These properties
extend by continuity to ϕ and thus, ϕ is a ‹-homomorphism. Since it is isometric it is
one-to-one and thus, a ‹-isomorphism. (This proves the isometry assertion.)
If φ is another such ‹-isomorphism that maps id to A then φ and ϕ coincide on the
dense space P l; thus they are equal. (This proves uniqueness of the asserted mapping.)
Since we can approximate every element in ApAq by polynomials in A, it follows at
once that ϕpf q is a normal element, and that it is self-adjoint if and only if f is real
valued. If B P A commutes with A then it commutes with all polynomials in A, and thus
it commutes with every element in ApAq. This concludes the proof.

The mapping ϕ is called the function calculus for the self-adjoint element A. To
simplify the notation, it is often common to write f pAq instead of ϕpf q. Henceforth, we
will use this convention.
We have already seen that a spectrum of a Banach algebra depends on the ambient
Banach algebra. For C ‹ -algebras this is not the case. Here, the spectrum of an element
is independent of the C ‹ -algebra in which we consider this element to be. Stated sloppy,
C ‹ -algebras are always large enough to contain an inverse of an element if it exists. With
the aid of the function calculus we can now prove this statement.

Proposition 1.27. Let A be a C ‹ -algebra, and B be a C ‹ -subalgebra of A. If B is in B,

then

spA pBq “ spB pBq. (53)

Proof. We have already discussed that spA pBq Ď spB pBq. Thus we need only to prove
the opposite. For this we will show that if A P B has an inverse in A then this inverse is
also in B.
We consider first a self-adjoint element A. Since 0 R spA pAq, the function f ptq “ t´1
is continuous on spA pAq. Thus, by the function calculus relative to the algebra A we get
an element f pAq in A. From Theorem 1.26 ii ) this element is contained in B. Since
tf ptq “ 1 for each t in spA pAq, it follows by multiplicativity of the function calculus that

I “ ϕp1q “ ϕpid ¨f q “ ϕpid qϕpf q “ Af pAq. (54)

Thus A´1 “ f pAq P B.

Now, consider a generic A P B that has an inverse A´1 in A. Then A‹ is also in
B and has an inverse pA´1 q‹ in A. Now, A‹ A is a self-adjoint element in B and it has
an inverse A´1 pA´1 q‹ in A. By the first part of the proof, A´1 pA´1 q‹ is in B. Thus,
A´1 “ pA´1 pA´1 q‹ qA‹ P B.

21
 The above proposition shows that the spectrum of an element in a C ‹ -algebra is inde-
pendent of the C ‹ -algebra this element is in. Thus, from now on, we can (and will) omit
the subscript referring to the ambient C ‹ -algebra and write sppAq instead of spA pAq.
In the proof of Proposition 1.27 we use the phrase “... by the function calculus relative
to the algebra A ...”. Implying, that the function calculus may depend on the original al-
gebra. Nevertheless, using Proposition 1.27 we can see that this cannot happen. Namely,
if B is a C ‹ -subalgebra of A, and A is a self-adjoint element in B, then let ϕ be the spec-
tral calculus sppAq Ñ A and ψ be the spectral calculus sppAq Ñ B. However, since the
spectrum of A relative to either algebra is the same, we can consider ψ as a spectral calcu-
lus into the larger algebra A. By uniqueness of the spectral calculus, ψ must then coincide
with ϕ. Thus, we do not need to worry about the ambient C ‹ -algebra when talking about
the spectral calculus. In particular, we can always choose a sufficiently small C ‹ -algebra,
ApAq “ tf pAq : f P CpsppAqqu such that the spectral calculus, CpsppAqq Ñ ApAq, is onto.
Then, we can state the following about the spectrum of f pAq.

Corollary 1.28. If A is a self-adjoint element of a C ‹ -algebra A, and f P CpsppAqq, then

sppf pAqq “ tf ptq : t P sppAqu. (55)

Proof. The function calculus f Ñ f pAq is a ‹-isomorphism from CpsppAqq onto the C ‹ -
subalgebra ApAq of A, and ‹-isomorphisms preserve the spectrum.

Another useful consequence of the function calculus is the definition of a square root.

Corollary 1.29. Let A be a self-adjoint element of a C ‹ -algebra A, with sppAq Ď R` .

There exists a unique self-adjoint element H in the commutative C ‹ -subalgebra ApAq of A
such that A “ HH. Moreover, sppHq Ď R` , and if B P A commutes with A then it also
commutes with H.
` ? ˘
Proof. Since sppAq is a subset of positive real numbers, the function f ptq “ t1{2 “ id
is real-valued, positive, and continuous on sppAq. Define H as f pAq. If ϕ denotes the
corresponding function calculus then

HH “ f pAqf pAq “ ϕpf qϕpf q “ ϕpf 2 q “ ϕpid q “ A. (56)

Since, f is real valued, the element H is self-adjoint by 1.26 i ). Since f is positive the
spectrum of H is positive by 1.28. If B P A commutes with A then it commutes with H
by 1.26 iv ).
To show uniqueness, suppose that K is any element in A with positive spectrum and
?
such that K 2 “ A. Let tpn u be a sequence of polynomials that converge to f “ id. Set
qn ptq “ pn pt2 q. Then we have,

lim qn ptq “ lim pn pt2 q “ f pt2 q “ t, (57)

nÑ8 nÑ8

22
uniformly for t P sppKq. Hence, by the function calculus for K we get

K “ lim qn pKq “ lim pn pK 2 q “ lim pn pAq “ f pAq “ H, (58)

nÑ8 nÑ8 nÑ8

which proves uniqueness.

There are more useful corollaries that we can derive from the function calculus. For
now, however, we close this chapter and proceed to the next ingredient that will be
necessary to understand the construction of quantum mechanics: the GNS construction.

Info: The function calculus can be extended to normal elements. The methods to
prove it, however, are a bit different. We provide the full theorem without proof.

Theorem 1.30 (Function calculus for normal elements). Let A be a normal element
of a C ‹ -algebra A. There is a unique isometric ‹-isomorphism ϕ from CpspA pAqq into
A such that ϕpid q “ A. Moreover, for f P CpspA pAqq,
i ) ϕpf q is a normal element of A;
ii ) the set tϕpf q : f P CpspA pAqqu is a commutative C ‹ -algebra and it is the smallest
C ‹ -subalgebra of A that contains A;
iii ) ϕpf q is the limit of a sequence of polynomials in I, A, and A‹ ;
iv ) moreover, the element A is

• self-adjoint iff sppAq Ď R; • unitary iff sppAq Ď C1 ;

• positive iff sppAq Ď R` ; • projection iff sppAq Ď t0, 1u.

23
1.3 The GNS construction
In this section we will see that any abstract C ‹ -algebra can be “represented” as an al-
gebra of bounded linear operators on some appropriate Hilbert space. The construction
used to show this was developed by Gelfand, Naimark, and Segal, hence the name GNS-
construction. This construction is vital for our purpose: to use C ‹ -algebras in applications
to quantum mechanics. To discuss the GNS construction we need additional tools — the
order structure on the C ‹ -algebra, the space of states, and a brief discussion of represen-
tations of a C ‹ -algebras. We begin with the order structure.

1.3.1 Order structure on the C ‹ -algebra

Definition 1.31. Let V be a vector space. A (positive) cone C of V is a subset of V with

the following properties:
i ) if A and ´A are in C , then A “ 0;
ii ) if a is a positive scalar and A is in C , then aA is in C ;
iii ) if A, B are in C then A ` B is also in C .

A cone on a vector space allows us to compare elements in that vector space. In

particular, we can define an element A P V to be “smaller then” B P V whenever B ´ A
is in the cone C . This relation defines a partial ordering on V . An element I P V is called
an order unit when, for a given A P V it holds that ´aI ď A ď aI for some positive
scalar a that may depend on A.

Reminder: partial order. A partial order on a set Ω is a relation over the set Ω such
that if a, b, c P Ω:
i ) a ď a, that is the relation is reflexive;
ii ) if a ď b and b ď a then a “ b, that is the relation is anti-symmetric;
iii ) if a ď b and b ď c then a ď c, that is the relation is transitive.

Example 1.32 (Cone on R2 ). Let V “ R2 a two dimensional Euclidean vector space.

Define the cone C on V as the set of all those elements that in the canonical basis of R2
have positive coefficients, that is

C “ tv P R2 : v “ pv1 , v2 q, and v1 ě 0, v2 ě 0u. (59)

It is easy to check that C is indeed a cone, and thus this provides a partial ordering for
the space R2 . However, this cone (and thus the partial ordering) depends on the basis
and is not preserved under rotations.

24
 aA
A`B
B

B´A
A
B A

(a) A cone in a vector space. (b) Partial order induced by the cone.

Figure 3: Cone structure and the partial order.

Since a (positive) cone defines a set of positive elements, it stands to reason that a
set of positive elements — defined in some meaningful way — will yield a cone. As is
common, we can borrow the intuition of positive elements from bounded linear operators
to define positive elements in a C ‹ -algebra.

Definition 1.33 (Positive element). A positive element of a C ‹ -algebra is a self-adjoint

element with a positive spectrum. By A` we denote the set of all positive elements of A.

Example 1.34. Consider a C ‹ -algebra CpXq on a compact Hausdorff space X. Then

f P CpXq is self-adjoint if and only if it is real-valued. Since, the spectrum of f is its
image, it follows that f is positive (in the C ‹ -algebra sense) if and only if f pxq ě 0
for each x P X. Thus the notion of positivity for elements in the C ‹ -algebra CpXq is
consistent with the notion of positivity for functions.

Example 1.35. If B is a bounded linear operator on some Hilbert space H then we call
it positive, if xx, Bxy ě 0 for every x P H. We have seen in our tutorial class, that B is
positive if and only if it is self-adjoint and its spectrum is positive. Hence, the notion of
positivity for elements in the C ‹ -algebra BpHq is consistent with the notion of positivity
for bounded linear operators.

Even though, positivity is a statement about “signs”, it has a very useful characteri-
zation in terms of the norm. We state this characterization in the following lemma.

Lemma 1.36. A self-adjoint element A in a C ‹ -algebra is positive if and only if

}A ´ aI} ď a, (60)

for some a P R with a ě }A}.

25
Proof. Fix a ě }A}, then sppAq Ď r´a, as, and

}A ´ aI} “ rpA ´ aIq “ supt|t ´ a| : t P sppAqu “ suptpa ´ tq : t P sppAqu. (61)

Where the first equality holds by 1.24 i ) and the second equality holds by 1.25. Hence
}A ´ aI} ď a if and only if sppAq Ď R` .

Now, we can show that the set of positive elements in a C ‹ -algebra defines a closed
positive cone.

Theorem 1.37. Let A be a C ‹ -algebra. Then,

i ) A` is closed in A;
ii ) for A P A` , if ´A is also in A` then A “ 0.
iii ) for A P A` and a P R` we have aA P A` ;
iv ) for A, B P A` we have A ` B P A` ;
v ) for A, B P A` and AB “ BA we have AB P A` ;

Proof. i ) From Lemma 1.36 we have

A` “ tA P A : A “ A‹ and ›A ´ }A}I › ď }A}u. (62)

› ›

By continuity of the norm, A` is closed.

ii ) If A and ´A are both in A` , then sppAq Ď R` X p´R` q “ t0u; so }A} “ rpAq “ 0.
iii ) For A P A` and a ě 0, the element aA is self-adjoint and

sppaAq “ taλ : λ P sppAqu Ď R` . (63)

iv ) For A, B P A` we have by Lemma 1.36 that

(64)
› › › ›
›A ´ }A}I › ď }A} ›B ´ }B}I › ď }B}.

Therefore,

(65)
› ›
›A ` B ´ p}A} ` }B}qI › ď }A} ` }B}.

With a “ }A} ` }B} ě }A ` B} it follows again from Lemma 1.36 that A ` B P A` .

v ) For A, B P A` , let H be the square root of A defined as in Corollary 1.29, and
K be the square root of B. By Corollary 1.29, if A commutes with B then also H com-
mutes with B, and by the same corollary, K commutes with H. Thus, AB “ HHKK “
HKHK “ pHKq2 . Since H and K are both self-adjoint and since they commute, HK
is self-adjoint, and thus its spectrum is a subset of a real line. By Corollary 1.28 we have
˘2
sppABq “ sp pHKq2 “ sppHKq Ď R` .
` ˘ `

26
 Condition, ii ), iii ), and iv ) state that A` is a (positive) cone. Hence, we can use it
to define partial ordering on the C ‹ -algebra A. Using this ordering, we can say that an
element A is positive if A ě 0. Also notice, that in one of our exercises we have shown
that }A}I ´ A is always positive, if A is self-adjoint. Hence, for a unital C ‹ -algebra the
unit of the algebra is also the order unit for this partial ordering,

´}A}I ď A ď }A}I. (66)

1.3.2 States of the C ‹ -algebra

So far, we have discussed the elements of the algebra. However, a great deal of structure
on the C ‹ -algebra comes from its dual space: the space of continuous linear functionals
on the algebra. We have encountered the power of linear functionals in the Theorem
1.21. Now we approach the topic of linear functionals more systematically. With the aid
of the positive cone on the C ‹ -algebra we can define a cone on its dual space.

Reminder. We briefly recall some basic definitions in topology.

Definition 1.38 (Convex set.). A set Y in a vector space V is called convex if x and
y in Y imply that the line tby1 ` p1 ´ bqy2 : 0 ď b ď 1u is also in Y .

Definition 1.39 (Locally convex space.). A locally convex space is a topological vector
space, in which the topology has a base consisting of convex sets. That means: Let
V be a topological vector space with topology τ . The basis of τ is the family β Ă τ
such that every open set O P τ is of the form O “ Yα Bα for some family tBα u Ă β.
If each set in β is convex, then V is a locally convex space.

Definition 1.40 (Extreme points.). Let V be a locally convex space and Y Ă V a

convex subset of V . Then a point y0 P Y is called an extreme point of Y if y0 “
p1 ´ aqy1 ` ay2 , with 0 ă a ă 1 and y1 , y2 P Y implies y1 “ y2 “ y0 .

Theorem 1.41 (Krein-Milman). If X is a non-empty compact convex set in a locally

convex space V , then X has an extreme point. Moreover, X is the closure of the convex
hull of all extreme points of X.

After this short reminder we are equipped with everything to dive into the discussion
of linear functional on a C ‹ -algebra. We begin with some general definitions.

Definition 1.42. Let V be a partially ordered vector space with order unit I. Then,
i ) a linear functional ρ on V is called positive if ρpvq ě 0 for every v ě 0 with v P V ;
ii ) a positive linear functional ρ on V is called a state if ρpIq “ 1;
iii ) the state space SpV q is the set of all states on V .

27
 If ρ is a linear functional on a C ‹ -algebra A, we can define another linear functional
ρ‹ , by the equation ρ‹ pAq “ ρpA‹ q for every A P A.

Definition 1.43. Let A be a C ‹ -algebra and ρ a linear functional on it. If ρpAq “ ρpA‹ q
for every A P A, we call ρ hermitian.

Recall that every element of a C ‹ -algebra uniquely decomposes into its real and imag-
inary part. Using this decomposition we can rephrase the condition for a state to be
hermitian only in terms of self-adjoint operators: let A be an element in a C ‹ -algebra A
with the real part H and imaginary part K, then for a linear functional ρ on A we have,

ρpAq “ ρpH ` ıKq “ ρpHq ` ıρpKq, ρpA‹ q “ ρpH ´ ıKq “ ρpHq ` ıρpKq. (67)

It follows that ρ is hermitian if and only if ρpHq “ ρpHq for every self-adjoint H P A.
Using this characterization of hermitian functionals we can deduce that a positive
linear functional is always hermitian: Let A be self-adjoint in A and ρ a positive linear
functional on A. Since }A}I ˘ A ě 0 (as we already have seen) it follows from positivity
of ρ that ρp}A}I ˘ Aq ě 0. In particular, ρp}A} ˘ Aq is real, and from
1´ ` ˘¯
(68)
˘ `
ρpAq “ ρ }A}I ` A ´ ρ }A}I ´ A ,
2
it follows that ρpAq itself is real.
Additionally, we can characterize positive linear functionals by their norm as follows.

Theorem 1.44. A linear functional ρ on a (unital) C ‹ -algebra A is positive if and only if ρ

is bounded and }ρ} “ ρpIq.

Proof. For the “only if ” direction, assume that ρ is positive. Let A be in A, choose a P C
such that |a| “ 1 and aρpAq ě 0, and let H be the real part of aA. Since }H} ď }A} we
have H ď }H}I ď }A}I and thus

}A}ρpIq ´ ρpHq “ ρp}A}I ´ Hq ě 0. (69)

Since ρ is hermitian we have ρpāA‹ q “ ρpaAq “ ρpaAq and consequently,

1
|ρpAq| “ ρpaA ` āA‹ q “ ρpHq ď ρpIq}A}. (70)
2
Thus, ρ is bounded, with }ρ} ď ρpIq. Simultaneously, |ρpIq| ď }ρ}; implying }ρ} “ ρpIq.
For the “if ” direction, suppose that ρ is bounded and }ρ} “ ρpIq. After suitable rescal-
ing, it is enough to consider the case in which }ρ} “ ρpIq “ 1. With A P A` and some
a, b P R assume that ρpAq “ a ` ıb. We want to show that a ě 0 and b “ 0. For this we
observe that since sppAq Ă R` , there is a small positive number s such that,

sppI ´ sAq “ t1 ´ st : t P sppAqu Ď r0, 1s. (71)

28
So by the spectral radius formula we have }I ´ sA} “ rpI ´ sAq ď 1. From the continuity
of ρ we then have

1 ´ sa ď |1 ´ spa ` ıbq| “ |ρpI ´ sAq| ď }I ´ sA} ď 1, (72)

and a ě 0. To show that b “ 0 we proceed as in the proof of Proposition 1.24 ii ). For

n P N consider the element Bn “ A ` ınbI in A. Then,

}Bn }2 “ }Bn‹ Bn } “ }A2 ` n2 b2 I} ď }A}2 ` n2 b2 , (73)

and we get for every n “ 1, 2, . . .

a2 ` pn2 ` 2n ` 1qb2 “ |a ` ıb ` ınb|2 “ |ρpBn q|2 ď }A}2 ` n2 b2 , (74)

where the last inequality follows from the continuity of ρ and (73). Since A is bounded,
b has to vanish.

With this understanding we can summarize the similarities and differences between
positive elements in the algebra and its dual.
An element A in A is positive if An element ρ in A1 is positive if
i ) it is self-adjoint, i ) it is hermitian,
ii ) sppAq Ă R` , ii ) ρpAq ě 0 whenever A P A` ,
iii ) }A ´ aI} ď a. iii ) ρpIq “ }ρ} ă 8.
Before we proceed with the analysis of states, we recall some ideas regarding the
topologies on vector spaces.

Reminder: weak‹ -topology. Let V be a linear space and let F be a family of linear
functionals on V . Assume that F separates points of V in the sense that if x P V is
non-zero then there is a ρ P F such that ρpxq ‰ 0. We define the topology σpV, Fq
as the coarsest (weakest, one with the fewer open sets) topology on V relative to
which each element of F is a continuous mapping from V into C.
If V is a locally convex topological vector space, we denote its continuous (topo-
logical dual) as V 1 . If x is in V then the equation

x̂pρq “ ρpxq pρ P V 1 q, (75)

defines a linear functional x̂ on the dual V 1 . The set V̂ “ tx̂ : x P V } is a linear sub-
space of V 2 — the algebraic dual space of V 1 (the space of not necessarily continuous
functionals on V 1 ). The set V̂ separates points of V 1 since, if ρ P V 1 and x̂pρq “ 0 for
all x P V then ρpxq “ 0 for all x P V and thus ρ “ 0. The weak‹ -topology on V 1 is the
topology σpV 1 , V̂ q. We often denote it simply by σpV 1 , V q. Hence, the weak‹ -topology

29
 is the coarsest topology on V 1 for which each functional x̂ is continuous.

A bounded closed set in an infinite dimensional space is not always compact. How-
ever, in the weak‹ -topology this is indeed true (see the useful Lemma below). This makes
the weak‹ -topology an important tool in analysis.

Useful lemma. Let X be a normed space and X 1 be the topological dual of X. If S

is a bounded weak‹ -closed subset of X 1 , then S is weak‹ -compact.

From Theorem 1.44 we see that a state on A is a bounded linear functional on A,

with }ρ} “ 1. Hence, SpAq is contained in the unit sphere of the Banach dual space A1
and is therefore bounded; it is also convex, since for every ρ1 , ρ2 P SpAq and 0 ď a ď 1
the functional ρ “ aρ1 ` p1 ´ aqρ2 is again positive (since positive elements form a cone)
and ρpIq “ a ` p1 ´ aq “ 1. Additionally, SpAq is weak‹ -closed. To see this, observe that

SpAq “ tρ P A1 : ρpIq “ 1, ρpAq ě 0 when A P A` u. (76)

For A P A let  be the dual element of A1 such that ρpAq “ Âpρq for every ρ P A1 .
Since each  is by definition continuous in the weak‹ -topology, the set CA “ ´1 r0, 8q
` ˘

— the pre-image of r0, 8q under  — is closed, and so is the set C0 “ Iˆ´1 pt1uq. Then
SpAq “ XAPA` CA XC0 is an intersection of weak‹ -closed sets and thus is weak‹ -closed.
` ˘

In summary: SpAq is a bounded, weak‹ -closed, convex subset of A1 , and thus from the
“useful lemma” above, it is weak‹ -compact. By the Krein-Milman theorem SpAq is the
closed convex hull of its extreme points.

Definition 1.45. An extreme point of SpAq is called a pure state of A. The set of all pure
states of A will be PpAq. The weak‹ -closure of PpAq is called the pure state space of A.

Visually we can picture the state space as in Figure 4. In general, the set of pure states
is not weak‹ -closed, and PpAq can contain states which are not pure. For commutative
C ‹ -algebras, however, this does not happen. So if A is commutative then the space of all
pure states is already weak‹ -closed. This observation relates to a deep result attributed
to Gelfand — sometimes called the Gelfand representation of a C ‹ -algebra. We state the
result without proof.

Theorem 1.46 (Gelfand representation.). Let A be a commutative C ‹ -algebra, and PpAq

be the set of all pure states of A. Then PpAq is a compact Hausdorff space, relative to the
weak‹ -topology. For A P A let  be a complex-valued function on PpAq defined by Âpρq “
ρpAq. The mapping A Ñ Â is a ‹-isomorphism from A onto the C ‹ -algebra CpPpAqq.

30
 SpAq

mixed states

PpAq

Figure 4: Intuition for the space of states SpAq. The pure states lie on the black solid
boundary line. The line, however, is not closed; it may miss some points, which are
indicated in red. The weak‹ -closure of that line is the whole boundary and it is the space
of pure states PpAq. In particular, the space of pure states may contain elements which
are not pure states (the red points). The interior of the set is sometimes called the space
of mixed states.

Remark. By this construction we can associate a compact Hausdorff space to ev-

ery commutative C ‹ -algebra. This can be used to show that every Hausdorff space
admits a compactification — the Stone-Čeck compactification.

Remark. If A is a commutative C ‹ -algebra then by the Gelfand representation it

is ‹-isomorphic to the algebra CpPpAqq of continuous functions on the pure state
space PpAq of A. Thus, if ω is a pure state of A and f is in CpPpAqq then there is a
functional δω pf q “ f pωq. We call such functionals point evaluations. Therefore a pure
state of a commutative C ‹ -algebra A corresponds to a point evaluation on the algebra
CpPpAqq. Indeed it is a general result required to prove the Gelfand representation,
namely: if CpXq is a C ‹ -algebra of continuous functions on a compact Hausdorff
space X then a state ω P SpAq is pure if and only if it is a point evaluation, i.e. there
is an x P X such that ωpf q “ δx pf q “ f pxq for every f P CpXq.

1.3.3 Representations of a C ‹ -algebra

Up to now we discussed abstract C ‹ -algebras. Now, we will see that every C ‹ -algebra
can be represented as a (sub)algebra of bounded linear operators on some Hilbert space.
From the physicist point of view, it is this realization that allows us to formulate quantum
and classical mechanics in terms of algebras. After all, as a physicist we understand
operators in terms of how they act, not in terms of to what space they belong.

Definition 1.47. By a representation of C ‹ -algebra A on a Hilbert space H, we mean

a ‹-homomorphism ϕ from A onto BpHq. If in addition, ϕ is a ‹-isomorphism (that is
one-to-one) then it is called a faithful representation.

31
 The Gelfand-Naimark theorem assures that every C ‹ -algebra has a faithful represen-
tation on a Hilbert space. However, we only will discus the construction that provides a
(possibly not faithful) representation.
The following terminology is common when discussing representations. Let ϕ be a
representation of a C ‹ -algebra A on a Hilbert space H. Recall from our assignment 4 i)
that every ‹-homomorphism between C ‹ -algebras is continuous, and so ϕ is continuous,
i.e. }ϕpAq}H ď }A}A for every A P A. Also recall, that a ‹-isomorphism (an injective
‹-homomorphism) is isometric, and so if ϕ is faithful }ϕpAq}H “ }A}A . If you recall the
definition of ideals from assignment sheet 5, you realize that the set A P A : ϕpAq “ 0
(

is a closed two-sided ideal in A. We call this set the kernel of ϕ. If there is a vector x P H
for which the linear subspace

(77)
(
ϕpAqx “ ϕpAqx : A P A

is everywhere dense in H, then ϕ is said to be a cyclic representation, and x is said to be

the cyclic vector for ϕ.

Example 1.48. Recall, that we defines a ‹-homomorphism as a mapping that among

other things maps preserves the unit of the algebra. If for a second, we do not consider
this requirement then the map ϕ0 pAq “ 0 for every A P A is defines a trivial homomor-
phism and we can include to say that ϕ0 is a trivial representation.

Example 1.49. Let A be a subalgebra of bounded linear operators BpHq on the Hilbert
space H. Then the inclusion mapping ϕpAq “ A for every A P A is a representation. This
representation is faithful, since if A P A is not zero, then it is also non-zero in BpHq.
Let H be a Hilbert space K be a closed subspace of H, that is invariant under each
operator from A, i.e. the set Ax : A P A, x P K is a subset of K. Let A|K be the
(

restriction of the operator A to the subspace K. It can be checked that this restriction
is a ‹-homomorphism and thus A Ñ A|K : A Ñ BpKq is a representation of A on the
Hilbert space K. In the same way one can see that the orthogonal complement K K is
also invariant under A and thus the restriction A Ñ A|K K is a representation of A on K K .
If the inclusion representation A P A Ñ A P BpHq is cyclic then also the restriction
representations are cyclic as well. To see this let P be the orthogonal projection from H
onto K and x P H be the cyclic vector x P H for the inclusion representation. If P x “ 0
then x is in K K . This, violates the fact that the inclusion representation is cyclic since
K K is invariant under the action of A. Thus P x ‰ 0 and P x P K is a cyclic vector for the
representation A Ñ A|K , and pI ´P qx is a cyclic vector for the representation A Ñ A|K K .

For the following discussion it will be convenient to define an inner product as a

sesquilinear form which is not necessarily definite. In the following reminder we quickly
recall the properties of the inner product. We also refresh that the well known Cauchy-
Schwarz inequality does not depend on the definiteness of the inner product.

32
 Reminder: Inner product and the Cauchy-Schwarz inequality. In the following
we use the convention that an inner product on a complex vector space V may not
be definite. That is, we say that a mapping pv, wq Ñ v, w from V ˆ V to C is an
@ D

inner product on V when it is

i ) positive: v, v ě 0;
@ D

ii ) hermitian: v, w “ w, v ;
@ D @ D

iii ) sesquilinear: v, αw ` βz “ α A, B ` β A, C ;
@ D @ D @ D

with v, w, z P V and α, β P C. If in addition it is

iv ) definite: v, v “ 0 only when v “ 0,
@ D

then we call , a definite inner product.

@ D

@ D
Proposition 1.50. Let V be a complex vector space. Suppose that , is an inner
product on V . Then,
@ D @ D@ D
i ) the Cauchy-Schwarz inequality, | v, w |2 ď v, v w, w , holds for v, w P V ;
@ D
ii ) the set L “ tv P V : v, v “ 0u is a linear subspace of V ;
iii ) the equation

(78)
@ D @ D ` ˘
v ` L, w ` L q “ v, w v, w P V
@ D
defines a definite inner product , q on the quotient space V {L.

Proof. For the proof of the above Proposition consult e.g. [1, prop. 2.1.1]
(
Proposition 1.51. Let A be a C ‹ -algebra, ρ a state on A, and Lρ “ A P A : ρpA‹ Aq “ 0 .
The state ρ provides a definite inner product on the quotient space A{Lρ by the relation

(79)
@ D ` ˘
A ` Lρ , B ` Lρ “ ρpA‹ Bq A, B P A ,

Moreover, Lρ is a closed left ideal in A, and ρpA‹ Bq “ 0 whenever A P Lρ and B P A.

Proof. For A, B P A define A, B ρ “ ρpA‹ Bq. Then , ρ is positive since ρ is positive,
@ D @ D

hermitian since ρ is hermitian, and sesquilinear by definition. Thus, , is an inner

@ D

product on A with Lρ “ A P A : A, A ρ “ 0 . By Proposition 1.50, the set Lρ is a

@ D (

linear subspace of A, and A ` Lρ , B ` Lρ “ A, B ρ “ ρpA‹ Bq is a definite inner

@ D @ D

product on A{Lρ .
With A P Lρ and B P A we have by the Cauchy-Schwarz inequality

|ρpA‹ Bq|2 ď ρpA‹ AqρpB ‹ Bq “ 0, (80)

and so ρpA‹ Bq “ 0. It follows that

(81)
` ˘ ` ˘
ρ pBAq‹ BA “ ρ A‹ pB ‹ BAq “ 0,

33
and thus BA is in Lρ whenever A P Lρ and B P A. Hence, Lρ is a left ideal of A, which
is closed since ρ and the mapping A Ñ A‹ A are continuous.

We will call the space Lρ the left kernel of the state ρ. We can now summarize the
idea of the GNS-construction in a loosely way before proving it rigorously. The preceding
Proposition provides us with means to define a (definite) inner product on A. With the
norm induced by this inner product, the vector space A may not be closed, but after
closing it we obtain a Hilbert space. Moreover, since A is an algebra we can define an
action of A P A on the vector space A by setting B Ñ AB : A Ñ A. After proving that
this mapping is continuous we can extend it to the whole H. This will give us the desired
representation. As always, however, the devil lies in the complement of a dense set...

Theorem 1.52 (GNS representation). For every state ρ on a C ‹ -algebra A there is a cyclic
representation πρ of A on a Hilbert space Hρ and a unit cyclic vector xρ for πρ such that

(82)
@ D ` ˘
ρpAq “ xρ , πρ pAqxρ APA .

If ϕ is a cyclic representation of A on a Hilbert space H with the unit cyclic vector x for π
such that

(83)
@ D ` ˘
ρpAq “ x, ϕpAqx APA ,

then there is an isomorphism U : Hρ Ñ H from Hρ onto H such that

(84)
` ˘
x “ U xρ and ϕpAq “ U πρ pAqU ˚ APA .

Proof. Let Lρ be the left kernel of ρ. The quotient linear space A{Lρ is a pre-Hilbert space
relative to the definite inner product defined by

(85)
@ D ` ˘
A ` Lρ , B ` Lρ “ ρpA‹ Bq A, B P A .

Define Hρ as the completion of A{Lρ in the norm induced by the above scalar product.
To show that for A P A the mapping B ` Lρ Ñ AB ` Lρ defines an unambiguous
operator on Hρ , we need to show that it does not depend on the “representative” B of
the set B ` Lρ . For this assume B, C P A such that B ` Lρ “ C ` Lρ . Then B ´ C is
in Lρ and since Lρ is a left ideal of A we have for every A P A that AB ´ AC P Lρ and
consequently AB ` Lρ “ AC ` Lρ . Hence, the equation πpAqpB ` Lρ q “ AB ` Lρ defines
a linear operator πpAq acting on the pre-Hilbert space A{Lρ .
Next, we want to show that π is continuous. For this recall that

}A}2 I ´ A‹ A “ }A‹ A}I ´ A‹ A P A` , (86)

34
 ˘1{2
implies that the positive square root K “ }A}2 I ´ A‹ A exists, and B ‹ p}A}2 I ´
`

A‹ AqB “ pKBq‹ pKBq P A` . Therefore for each A, B P A we have

}A}2 }B ` Lρ }2 ´ }πpAqpB ` Lρ q}2 “ }A}2 }B ` Lρ }2 ´ }AB ` Lρ }2

@ D @ D
“ }A}2 B ` Lρ , B ` Lρ ´ AB ` Lρ , AB ` Lρ
“ }A}2 ρpB ‹ Bq ´ ρpB ‹ A‹ ABq
` ˘
“ ρ B ‹ p}A}2 I ´ A‹ AqB ě 0.

Thus πpAq is bounded, with }πpAq} ď }A}; and πpAq extends by continuity to a bounded
linear operator πρ pAq acting on Hρ . It remains to show that πρ is a ‹-homomorphism and
that it is cyclic.
With A, B, C P A and α, β P C we get:
linearity,

πρ pαA ` βBqpC ` Lρ q “ pαA ` βBqC ` Lρ “ αpAC ` Lρ q ` βpBC ` Lρ q

` ˘
“ απρ pAq ` βπρ pBq pC ` Lρ q;

multiplicativity,
` ˘ ` ˘ ` ˘
πρ pABq C ` Lρ “ ABC ` Lρ “ πρ pAq BC ` Lρ “ πρ pAqπρ pBq C ` Lρ ;

star preservation,
@ D @ D
B ` Lρ , πρ pAqpC ` Lρ q “ B ` Lρ , AC ` Lρ
` ˘ ` ˘
“ ρ B ‹ pACq “ ρ pA‹ Bq‹ C
@ D @ D
“ A‹ B ` Lρ , C ` Lρ “ πρ pA‹ qpB ` Lρ q, C ` Lρ .

Since A{Lρ is everywhere dense in Hρ these properties extend to the whole Hρ and thus
πρ is a representation of A on Hρ .
To see that πρ is cyclic let xρ be the vector I ` Lρ in A{Lρ , then

(87)
` ˘
πρ pAqxρ “ πρ pAqpI ` Lρ q “ A ` Lρ APA ,

and πρ pAqxρ is A{Lρ — an everywhere-dense subset of Hρ . Hence xρ is a cyclic vector for

πρ . Moreover, xρ , πρ pAqxρ “ I ` Lρ , A ` Lρ “ ρpAq for every A P A. In particular,
@ D @ D

}xρ }2 “ ρpIq “ 1. This proves the first part of the theorem.

For the second part, let ϕ be another representation as stated in the theorem. For
each A P A,
@ D @ D @ D
}ϕpAqx}2 “ ϕpAqx, ϕpAqx “ x, ϕpA‹ Aqx “ ρpA‹ Aq “ xρ , πρ pA‹ Aqxρ “ }πρ pAqxρ }2 .

Hence, if A, B P A are such that πρ pAqxρ “ πρ pBqxρ then inserting A ´ B instead

of A in the above relations, yields ϕpAqx “ ϕpBqx. Therefore, we can define a map

35
U0 : πρ pAqxρ Ñ ϕpAqx by the equation U0 πρ pAqxρ “ ϕpAqx for each A P A. It follows
that U0 is a norm-preserving linear operator from πρ pAqxρ onto ϕpAqx. Since the closure
of πρ pAqxρ is the entire Hρ and the closure of ϕpAqx is the entire H, it follows that U0
extends by continuity to an isomorphism U from Hρ onto H, and

U xρ “ U0 πρ pIqxρ “ ϕpIqx “ x. (88)

With A, B P A,

U πρ pAqπρ pBqxρ “ U πρ pABqxρ “ ϕpABqx “ ϕpAqϕpBqx “ ϕpAqU πρ pBqxρ . (89)

Since vectors of the form πρ pBqxρ with B P A form an everywhere-dense subset of Hρ , it

follows that U πρ pAq “ ϕpAqU , and thus ϕpAq “ U πρ pAqU ‹ .

The method used in the proof of the above Theorem is called the GNS (Gelfand-
Naimark-Segal) construction. To avoid bloated notation, we will often write pπ, H, xq to
refer to a cyclic representation π of a C ‹ -algebra A on the Hilbert space H with a cyclic
vector x for π. When its convenient we will also call the triple pπ, H, xq itself a cyclic
representation of A. If ρ is a state on A, and pπρ , Hρ , xρ q is the representation set out in
the Theorem 1.52, we will call it the GNS representation.
For a given C ‹ -algebra A, let π be a representation of A on the Hilbert space H
and ϕ be another representation of A on the Hilbert space K. We will call the two
representations (unitarily) equivalent if there exists an isomorphism U : H Ñ K such
that ϕpAq “ U πpAqU ‹ for each A P A. In particular, the GNS Theorem states that a
GNS representation πω of A and the cyclic state xω is uniquely determined (up to unitary
equivalence) by the condition ρp¨q “ xω , πω p¨qxω .
@ D

We already mentioned that if π is a representation of a C ‹ -algebra on a Hilbert space

H then H may have invariant subspaces K Ă H such that πpAqK Ď K. If this hap-
pens, we call π a reducible representation. If on the other hand, H has no such invariant
subspace, and the only invariant subspace is H itself, then we call π an irreducible rep-
resentation. By restricting a reducible representation to the invariant subspaces we can
get an irreducible representation from a reducible one. Conversely, by “piling up” (direct
sum) irreducible representations we can construct new representations, which will be
reducible. In this sense, irreducible representations are the building blocks of represen-
tations. Since C ‹ -algebra representations play a vital role in physics, it is important to
characterize them. From the above it follows that it is enough to characterize only the
irreducible representations. This is a more tractable undertaking.
Using the GNS construction one can draw a connection between irreducible repre-
sentations of a C ‹ -algebra and its pure states. However, we are not going to do this.
Instead we will now transition to the mathematical description of physical systems to
find out how C ‹ -algebras enter physics. We close this section with an example that will
be important for the next section.

36
Example 1.53. Consider a C ‹ -algebra A and a state ρ on A.
i ) The GNS representation for the state ρ is the ‹-homomorphism πρ : A Ñ BpHρ q,
with the Hilbert space Hρ and a cyclic (GNS) vector xρ ;
ii ) Let y be a unit vector in Hρ . Then we can use y instead of the cyclic vector for πρ . If
y is cyclic for πρ then pπρ , Hρ , yq is a new cyclic representation which is equivalent
to the GNS representation πρ . In fact, the ‹-homomorphism and the Hilbert spaces
are both the same, and thus U “ I is the desired isomorphism. (Notice that we do
not need that the cyclic vectors transform from one to the other).
iii ) For y from the previous example, set ωy pAq “ y, πω pAqy for every A P A. Then
@ D

ωy is a positive linear functional on A with ωy pIq “ 1; thus ωy is a state on A. The

GNS representation pπωy , Hωy , xωy q is thus unitarily equivalent to the representation
from the previous example.
From the above example we see that every cyclic state of the GNS Hilbert space Hω de-
fines a new state ωy which leads to a new GNS representation that is equivalent to the
original one. In fact, if the original state ω is pure, than its GNS representation is irre-
ducible and every vector y P Hω is cyclic. Thus, every vector from the GNS representation
of a pure state, defines an equivalent representation.

SpAq H SpAq

BpH1 q

BpHn q

Figure 5: The state space SpAq yields different GNS representations. Some are equiva-
lent some are not. In particular, each (cyclic) vector from the GNS Hilbert space defines
a state. The GNS representations of these states are all equivalent to the original repre-
sentation. Also other states (not created from vectors from the Hilbert space) can yield
equivalent representations.

37
2 Operator algebras in physics

Section remarks: This section is based on two books. The part on classical me-
chanics as well as the formulation of the Dirac-von Neumann axioms is adopted from
[6]. The axioms for a general physical system are from [7]. Nevertheless, axiom 4 is
shortened, because otherwise we would need too much additional notation. Also the
final axiom from [7] is missing entirely, because by now we have a classification of
Jordan-Banach algebras that make the axiom (more or less) obsolete. The operational
motivation (viewing observables as measurement devices) is again from [6]. Further
reading: the book by Strocchi [6] is very commendable for a simple introduction to the
physical side of the discussion. Despite being a book for mathematicians, it focuses on
the general picture to phrase quantum mechanics in the algebraic language. It is easy
to read and is not too heavy on the mathematical details. As a complementary read
the book by Emch [7] is mathematically much more precise and helps to understand
how careful and rigorous the algebraic formulation is constructed. For the algebraic
formulation of statistical physics Ruelle [8] is a great reference. Alternatively, the two
volumes by Bratteli and Robinson [9, 10] give a self-consistent introduction to the sub-
ject. The first volume covers the C ‹ -algebra theory, the second volume discusses the
algebraic formulation of quantum mechanics and non-relativistic quantum field the-
ory. Contrary to Strocchi and Emch, however, these books focus almost exclusively on
the mathematics. The deep physical ideas that historically lied underneath the subject
are not a matter of discussion in these two books.

2.1 Classical mechanics

We separate the description of a physical system into two parts: kinematics and dynamics.
Kinematics describes the possible configurations of the system; dynamics describes their
evolution in time.

2.1.1 Kinematics

Kinematics is concerned with describing configurations of a physical system — objects

that uniquely determine all experimental measurements. Such configurations are called
states. To describe the states we usually choose (independent) variables that determine
the state uniquely. Each such independent physical variable is sometimes termed the
(physical) degree of freedom.
One way to define a system and to identify its degrees of freedom is to use the La-
grangian formalism. Another, is to use the Hamiltonian description. In the latter, the
state is uniquely described in terms of canonical variables, which one can think of as

38
(generalized) positions and (generalized) momenta of a particle. The state of a mechan-
ical system is described by a set of canonical variables, tq, pu where q “ pq1 , . . . qn q P Rn
and p “ pp1 , . . . , pn q P Rn . The collection of all possible canonical variables for the system
is called the phase space Γ. For simplicity, we will assume in the following that the phase
space is compact, that is the range of possible positions and momenta of a particle are
bounded. Since the kinetic energy of a particle with the mass m is given by

p2
Ekin pq, pq “ , (90)
2m
this assumption means that the particle under consideration is confined to a bounded
region and has finite energy. Thus this assumption is reasonable.

Observables of the system are the physical quantities that result from the physical de-
grees of freedom. Thus in particular the position q and the momentum p are themselves
observables. But all its powers and sums are observables too, and so the observables
include the polynomials of the canonical variables q and p. Without loss of generality, we
can take the closure of this space in the uniform topology (supremum norm), to obtain
that all observables form the space of real-valued continuous function CpΓqR .
Equipped with a pointwise multiplication CpΓqR forms a real algebra. At the same
time, CpΓqR is the subspace of self-adjoint elements of the C ‹ -algebra CpΓq equipped
with the conjugation as the star operation and topologized with the supremum norm.
Extending CpΓqR to CpΓq is quite natural from the algebraic point of view. It is also not
harmful from the physics point of view since a C ‹ -algebra is uniquely determined by its
self-adjoint elements as we have seen. Thus, we see that the space of observables of a
classical system defines a C ‹ -algebra with identity I — the constant unit function on Γ.

Physical states define the physical system at hand. From our description, every state
is a point of the phase space tq, pu “ P P Γ, and it uniquely determines the value of
all the observables. Conversely, the value of all the observables uniquely determines the
state by the Urysohn’s and the Stone-Weierstrass theorem. This is sometimes called the
duality relation between the states and observables.
Identifying states of a classical system with points of a phase space is, however, an
idealization. It requires infinitely precise measurements to resolve single points of Γ. For
realistic measurements, we always have measurement errors and noise that corrupt the
measurement. The general procedure (and the one that is in fact used in physics) takes
into account these errors. To overcome the noise problem we typically repeat the exper-
iment several times and take the average of the measurement results. More precisely, let
us refer to the state of a classical system by ω. Then the value of an observable f P CpΓq
in the state ω is given by the following procedure: we take n replicas of ω and perform
measurements of f on each of the replicas. With this we obtain the measurement results

39
m1 pf q, m2 pf q, . . . , mn pf q; subsequently we compute the average
pωq pωq pωq

@ Dpωq 1 ` pωq
(91)
pωq ˘
f n “ m1 pf q ` m2 pf q ` ¨ ¨ ¨ ` mpωq
n pf q
n
It is one of the basic assumptions of experimental physics that for n Ñ 8 the above
quantity converges. This limit value is the expectation value of the observable f in the
state ω. It is this value that we assign to a measurement
@ Dpωq
ωpf q “ lim f n . (92)
nÑ8

Since the expectation values of all the observables is all there is that we can measure,
we can identify a state of a physical system with the set of expectation values of its
observables. Since ωpf q is operationally defined by the averaging process it follows that
such expectations are linear. That is, for f, g P CpΓq and α, β P C we have

ωpαf ` βgq “ αωpf q ` βωpgq. (93)

Also such expectations are positive

(94)
` ˘
ωpf ‹ f q ě 0 f P CpΓq ,

since each measurement mk pf ‹ f q yields a positive number for every k P N. Thus ω

pωq

defines a positive linear functional on the C ‹ -algebra CpΓq.

As we have seen such functionals induce an inner product on the C ‹ -algebra by

(95)
@ D ` ˘
f, g “ ωpf ‹ gq f, g P CpΓq ,

for which the Cauchy-Schwarz inequality holds

(96)
@ D @ D@ D ` ˘
| f, g |2 ď f, f g, g f, g P CpΓq .

In particular, this implies that for every f P CpΓq we have

|ωpf q| ď ωpIqωpf ‹ f q. (97)

If ωpIq “ 0 then ωpf q “ 0 for every f P CpΓq, meaning that ω is a trivial functional.
Conversely, if ω is non-trivial then ωpf q ‰ 0 for at leas one f P CpΓq and thus ωpIq ‰ 0.
In this case, we can normalize ω to satisfy ωpIq “ 1. Thus we can identify physical states
of a classical system with (algebraic) states on the algebra of observables CpΓq.

Reminder: Riesz-Markov representation theorem. Let X be a locally compact

Hausdorff space. For any positive linear functional ω on CpXq, there is a unique
Radon measure µ on X such that
ż
(98)
` ˘
ωpf q “ f pxq dµpxq f P CpXq .
X

40
 Proof. For the proof consult e.g. [3, thm. IV. 14].

From the Riesz-Markov theorem it follows that a state ω on CpΓq defines a Radon
measure (regular Borel measure) µω on Γ such that
ż
(99)
` ˘
ωpf q “ f pq, pq dµω pq, pq f P CpΓq .
Γ

Since 1 “ ωpIq “ Γ dµω “ µpΓq we see that µ defines a probability measure on Γ. Thus
ş

such an operational description of a physical system leads to a probabilistic description

of a state, and observables then get the meaning of random variables.
As a special case, when µδP is a Dirac measure such that, for O an open subset of Γ
$
&1 when P P O;
µδP pOq “ (100)
%0 when P R O

If P “ tq0 , p0 u we have
ż
δP pf q “ f pq, pq dµδP pq, pq “ f pP q “ f pq0 , p0 q. (101)
Γ

Thus δP is an evaluation state, and from the Gelfand representation Theorem 1.46 and
the remark that follows it, we saw that such evaluation states define pure state of the
C ‹ -algebra CpΓq. Thus the standard description of classical mechanics, where points of
the phase space define the states, corresponds to pure states in the algebraic formalism.
To quantify the precision of measurements for an observable f P CpΓq in a state ω,
we use the variance

(102)
` ˘
p∆ω f q2 “ ω pf ´ ωpf q2 q .

The smaller the variance, the sharper is the measurement — the more “reliable” is the
mean value ωpf q. Pure states such as δP (P P Γ) minimize the variance, as for them we
have δP pf 2 q “ δP pf q2 “ f pP q2 and thus ∆δP f “ 0. Thus such states are called dispersion-
free, as they represent measurements of a classical system without any uncertainty.
Since algebraic states capture both, the pure and the mixed states, they provide a uni-
fied framework to define sharp measurements as well as statistical fluctuations. For that
reason we can use the same formalism to describe statistical mechanics. In fact, in sta-
tistical mechanics the description of the system with pure states is unfeasible because we
would need to know the exact positions and momenta for 1023 particles. Therefore, the
only reasonable description of such systems is using probability distributions for states
and random variables to describe observables. This is what we usually encounter in sta-
tistical approach to thermodynamics: the partition function there defines a probability
distribution, and observables are random variables sampled from that distribution.

41
2.1.2 Dynamics

Dynamics of a physical system describes how the system evolves in time. When we use
canonical variables tq, pu to describe the system, then dynamics is defined by the time
evolution of these variables

q Ñ qptq “ qpt, q0 , p0 q p Ñ pptq “ ppt, q0 , p0 q. (103)

In general, the evolution of canonical variables is given by the solution of a system of

differential equation called the Hamiltonian equations of motion,

Bq BHpq, pq Bp BHpq, pq
“ “´ , (104)
Bt Bp Bt Bp
qpt0 q “ q0 ppt0 q “ p0 . (105)

where Hpq, pq is called the Hamiltonian of the system. If gradpHq is continuous, then the
solution to the Hamiltonian equations of motions exists; and if gradpHq is continuously
differentiable then the solution is unique. In this case, it can be extended to all times
t P R and the evolution tq, pu Ñ tqptq, pptqu is continuous in time t P R, and depends
continuously on the initial data tqpt0 q, ppt0 qu “ tq0 , p0 u.
The evolution of the canonical variables implies that the results of measurements
evolve in time. Since all observables are constructed from canonical variables, all ob-
servables evolve in time. In particular, this implies a one-parameter family of continuous
invertible mappings αt0 ,t : CpΓq Ñ CpΓq from the algebra of observables into itself so that

αt0 ,t : f pq0 , p0 q Ñ f pqptq, pptqq. (106)

If f, g are two polynomials of canonical variables then it is easy to see that for β, γ P C
and every t P R we have,

αt0 ,t pβf ` γgq “ βαt0 ,t pf q ` γαt0 ,t pgq “ βf pt, q0 , p0 q ` γgpt, q0 , p0 q, (107)

αt0 ,t pf ‹ q “ αt0 ,t pf q‹ “ f pt, q0 , p0 q‹ , (108)
αt0 ,t pf gq “ αt0 ,t pf qαt0 ,t pgq “ f pt, q0 , p0 qgpt, q0 , p0 q. (109)

By continuity, these properties extend to the whole algebra CpΓq, and thus αt0 ,t defines a
one-parameter family of ‹-automorphisms (‹-isomorphisms from the algebra into itself).
When the dynamics depends only on the difference t ´ t0 , i.e. it is time-translation invari-
ant, then the one-parameter family is actually a one-parameter group of automorphisms
tαt : t P Ru on CpΓq. By duality it defines a one-parameter group of transformations αt‹
on the state space as

(110)
` ˘
ωt pf q “ pαt‹ pωqqpf q “ ωpαt pf qq f P CpΓq .

42
 In summary we see, that observables of a classical system define the commutative C ‹ -
algebra of observables CpΓq, and physical states define (algebraic) state on that algebra;
dynamics of the canonical system is given by a one-parameter group of ‹-automorphisms
on CpΓq. In particular, all physically observable features of a physical system are captured
by the algebraic formulation. However, still in the standard approach we start with the
phase space that describes our system, and then derive the C ‹ -formulation.
The Gelfand representation Theorem 1.46, however, allows us to reverse the logic.
Instead of starting with the phase space, and deriving the algebra from it, we can start
with a commutative C ‹ -algebra A that captures an abstract characterization of observ-
ables of our system; define the physical states as the state space SpAq; and define the
dynamics as a ‹-automorphism on A. By the Gelfand representation Theorem the ab-
stract algebra A can be represented as an algebra of continuous functions CpXq on a
compact Hausdorff space X. Every dispersion-free state of A is associated with the point
in X and so the space X gets the meaning of the phase space. In this sense, the algebra
comes first, from which the phase space is then derived. This is the starting point for
the algebraic construction. Using this motivation we aim in the next section to define a
general algebraic framework for physical theories.

2.2 An algebraic construction for physical theories

The discovery of quantum mechanics soon proved to be pivotal for physics. This inspired
a rapid development in the field. This development was of physical nature, as more
and more quantum mechanical models described more and more experiments with ex-
traordinary precision. However, on the mathematical side the theory was shaky, and
the concepts were hardly justified on physical grounds. To formalize the theory, Dirac
and von Neumann provided 5 axioms to summarize a common mathematical structure
of quantum mechanics as developed by Heisenberg and Schrödinger — sometimes re-
ferred to as the Dirac-von Neumann axioms of quantum mechanics. Before we delve into
the algebraic framework we summarize a simplified version of these axioms that define
the standard approach to quantum mechanics.

States: The states of a physical system are described by rays (essentially unit vectors)
in a separable Hilbert space H.

Observables: The observables of a physical system are described by the set of bounded
self-adjoint operators in H.

Expectations: If a state ω is described by the vector ψ P H, for any observable A the

expectation ωpAq, defined by the average of the outcomes of replicated measurements of

43
A performed on the system in the state ω, is given by the Hilbert space matrix element

(111)
@ D
ωpAq “ ψ, Aψ

Dirac canonical quantization: Canonical coordinates and momenta satisfy the follow-
ing commutation relations

(112)
` ˘
rqi , qj s “ 0 “ rpi , pj s, rqi , pj s “ ı~δi,j , i, j “ 1, 2, . . . , n .

Schrödinger representation: The canonical commutation relations are represented

by the following operators in H “ L2 pRn , λq:

Bψ
(113)
` ˘
qi ψpxq “ xi ψpxq, pj ψpxq “ ´ı~ pxq i, j “ 1, 2, . . . , n .
Bxj

Even though the physical framework proved to be very successful, the physical basis
for many of its assumptions was unclear and unjustifiable. For that reason, a group
of mathematicians/physicists pioneered by Jordan, von Neumann, and Wigner, aimed
to provide a conceptually clear, and mathematically rigorous formalism for the above
postulates; to justify some of the tools in use of quantum mechanics; and to disregard
some of the assumptions not supported by mathematics or physical intuition. In this
section we discuss the summary of that approach that took a long time to be developed.
The idea of the algebraic approach is to use reasonable physical assumptions to define
an algebraic structure on the set of observables of a theory. Then to show that this
algebraic structure is rich enough to provide all the necessary tools used in quantum
mechanics as formulated in the above axioms. Following the literature, we formulate the
necessary assumptions on the set of observables as axioms.
@ D
Axiom 1. To each physical system Σ we can associate the triple pA, S, ; q formed by the set
@ D
A of all observables, the set S of all its states, and a mapping ; : pA, Sq Ñ R that assigns
@ D
to each pair pA, φq P pA, Sq a real number φ; A that we interpret as the expectation value
of the observable A in the state φ.

With a state of a physical system, we identify the method used for its preparation.
That is a (reproducible) way to prepare a configuration of the system. We assume that a
state is fully determined by all its expectation values, that is if φ, ϕ P S are such that

(114)
@ D @ D ` ˘
φ; A “ ϕ; A APA ,

then this must imply that φ “ ϕ. That is, we assume that there are no hidden properties
that we can never measure. We say the observables separate the states.
With an observable we identify a measurement device. Thus for any A P A we can
imagine a device with a pointer scale. The expectation value, is then the average over

44
n measurements of A in identically prepared states ω. Ideally, n shall go to infinity,
practically it has to be large enough to assure that our average is meaningful. If any
two devices produce the same expectation values for all states φ P S, then we identify
these devices. That is we assume that each measurement device produces a unique set
of expectation values, when evaluated in all states of S. We say, that S separates the
observables or simply that the set S is full with respect to A (see Figure 6 for intuition).

A S

A
B 1) switch on the stove
2) take 100g sugar
3) eat it
¨¨¨
C 4) hope for the best
..
. ..
.

(a) Observables: measurement devices. (b) States: methods of state preparation.

Figure 6: The set of observables and the set of states of a physical system. After preparing
a state according to a given method we can measure it with all observables and identify it
with a list of all expectation values. If any to methods lead to the same list of expectation
values we identify these methods. On the other hand, if any two observables provide two
identical lists of expectation values when measuring all states, we identify them either.

Next we realize that we can use the states to compare observables. For A, B P A we
say that A ď B if φ; A ď φ; B for all state φ P S. In particular, we call A ě 0 positive
@ D @ D

if φ; A ě 0 for every state φ P S. Especially, this implies that the states are positive
@ D

functionals on A.
Axiom 2. The relation ď defined above is a partial ordering on A; in particular A ď B and
B ď A imply A “ B.
With our interpretation of observables as measurement devices we see that the con-
dition φ; A ě 0 for every φ P S implies that every measurement A is positive. Since
@ D

the set of state is full with respect to the set of observables this meas that the scale of
the measurement device represented by A is positive. Thus we can identify the positive
element with measurement devices whose pointer scale ranges in positive real numbers.
Next we notice that for any two observables A, B P A the set t φ; A ` φ; B : φ P Su
@ D @ D

can be interpreted as a set of expectation values, thus a measurable quantity on Σ. We

strengthen this observation to the next axiom.

45
Axiom 3.
i ) For each observable A P A and any λ P R there exists an element pλAq P A such that

(115)
@ D @ D ` ˘
φ; λA “ λ φ; A φPS .

ii ) For any pair of observables A, B P A there exists an element pA ` Bq P A such that

(116)
@ D @ D @ D ` ˘
φ; A ` B “ φ; A ` φ; B φPS .

If A P A represents a measurement and λ is a real number then we can think of λA

as a new device, whose pointer scale is rescaled by λ.
This axiom, even though reasonable, is non-trivial when we interpret elements in A
as measurement devices. Using purely operational arguments, it is not clear that for
any two measurement devices represented by A and B there is always one device C
that produces the same expectation values as the sum of expectation values of A and B.
Strictly speaking, with this assumption we cannot interpret the set A solely as a set of
measurement devices, but we extend this set to all hypothetical measurements that result
from adding expectation values of measurements together. In particular, this defines a
linear structure on A making it into a vector space. Nevertheless, this assumption is still
quite reasonable. Being pragmatic, we can agree that the set A contains all theoretically
measurable devices. Hence also such devices that correspond to measurements of A ` B.
Recall from our analysis of a classical system, that dispersion-free states corresponded
to idealized measurements with infinite precision, i.e. to points of the phase space. Be-
cause of this, dispersion-free states uniquely defined all classical observables. It seems
reasonable to expect that this should hold in a general physical system. We use this
intuition to state the next axiom.

Axiom 4. For every observable A the set S contains the set of dispersion-free states SA for
A that uniquely determines the one-dimensional subspace of A generated by A; for any two
observables A, B the dispersion-free states satisfy SA X SB Ď SA`B and SI “ S.

In particular, this axiom implies that for two observables A and B if SA Ď SB and
φ; A “ φ; B for every φ P SA then A is equal to B. Thus dispersion-free states in a
@ D @ D

certain sense, uniquely determine the observables. This reflects our original motivation
to introduce this axiom. Nevertheless, this axiom requires an assumption of dispersion-
free states, which is an idealization of a physical process of measurement, as we dis-
cussed above. To argue that such states are reasonable, we can take the standpoint that
even though every realistic measurement is noisy, the noise can be arbitrary reduced.
Therefore there is no fundamental (nature given) threshold which we cannot surpass
by making a measurement more precise/less noisy. (In fact, this may be violated if one
takes General Relativity into account. But we are not going into this discussion here.)

46
 If A is an observable that represents a measurement device then we can interpret A2
as a measurement device whose scale is the square of the pointer scale of A. In particular,
A2 is therefore always positive. Similarly, we can define any power An for a non-negative
integer n, and set I “ A0 . We put the existence of such powers as an additional axiom.

Axiom 5. For any observable A and any non-negative integer n there exists at least one
element An P A such that
i ) SA Ď SAn ,
@ D @ Dn
ii ) φ; An “ φ; A for all φ P SA .

From the second point of this Axiom and Axiom 4 we see, that the element An is
unique in A and we can unambiguously call it the nth power of A.
With this axiom the vector space A obtains a power structure. Moreover, it follows
that A contains a unit I “ A0 for every A P A and an element O such that

and (117)
@ D @ D ` ˘
φ; I “ 1 φ; O “ 0 φPS .

From the second point of this axiom, it follows that SI “ SA0 “ S, and that the usual
rules apply: A0 “ I, A1 “ A, I n “ I and On “ O for n ą 1.
Furthermore, since φ; A2 ě 0 for all φ P SA and since SA fully determines A2 it
@ D

follows that A2 ě 0. In particular this implies that φ; A2 ě 0 for every state φ P S.

@ D

Since the set S is full with respect to A, we can conclude that A2 ` B 2 ` C 2 ` ¨ ¨ ¨ “ 0

implies A “ B “ C “ ¨ ¨ ¨ 0. This conditions is sometimes called reality, in analogy to the
fact that this implication holds for real numbers (as opposed to complex numbers).
Even though we have argued for a lot of structure on the space A, one important
property is still missing: the distributivity of powers over addition of any two elements
A, B. In general, for A, B P A we have that

pA ` Bqn “ pA ` BqpA ` Bq ‰ ApA ` Bq ` BpA ` Bq. (118)

Stated more precisely, we never defined the product between A and B that again defines
an observable AB. Notice, that in the usual approach to quantum mechanics, where we
identify observables with self-adjoint operators on a Hilbert space this is in particular
false: a product of self-adjoint elements is in general not self-adjoint.
On the other hand, for any observable A the product AA is defined by the unique
element A2 from the Axiom 5. We use this observation to equip A with a product.

Definition 2.1. With each pair of observables A, B we associate an element A ˝ B in A,

called the symmetric product of A and B, defined by
1`
(119)
˘
A˝B “ pA ` Bq2 ´ A2 ´ B 2 .
2

47
 This product depends on powers of the same observable (which we defined) and on
sums between observables (which we also defined). Moreover, this product is commuta-
tive, A ˝ B “ B ˝ A, and from Axiom 4 and 5 it satisfies

(120)
@ D @ D@ D ` ˘
φ : A ˝ B “ φ; A φ; B φ P SA X SB .

Thus, since dispersion-free states uniquely determine the observables it follows that An ˝
Am “ An`m . In particular, we have A ˝ I “ A. In other words, the power structure that
we had derived after the previous Axiom, can be generated by the symmetric product
on A. Therefore ˝ naturally extends the structure that we already have. This makes A
almost an algebra with identity, however, not quite yet.
The symmetric product, even though commutative, is neither associative, pA˝Bq˝C ‰
A˝pB ˝Cq, nor does it distribute over addition. The non-associativity is not a big loss, and
in general there is little algebraic reason to want to restore associativity. Distributivity, on
the other hand, is crucial for a definition of an algebra. At the same time, it seems hard
to find a physically compelling argument to make the product distributive. However,
one can show (even though we do not do it here) that distributivity follows from the
following mild looking assumption.

Axiom 6. For any pair of observables A, B the symmetric product is homogeneous in the
fist entry, i.e. for every real number λ it satisfies λpA ˝ Bq “ pλAq ˝ B.

Since the product is commutative, homogeneity in the first entry implies homogeneity
in both entries. With this assumption the symmetric product becomes distributive. In
fact, it also becomes weakly associative meaning that for A, B P A the product satisfies
A ˝ pB ˝ A2 q “ pA ˝ Bq ˝ A2 . A commutative algebra with a symmetric product that
satisfies the weak associativity is called a Jordan algebra; and a Jordan algebra that
satisfies the reality condition is called a real Jordan algebra. Therefore we can say that
with the first 6 Axioms the set of observables A becomes a real Jordan algebra. Notice,
that the Axioms 1-5 are very natural from the physicists perspective. The Axiom 6 is
the only questionable one. However, even though there is no direct evidence that the
product must be distributive, there is no direct physical reason why it should not be.
Mathematically, this assumption makes a significant difference.
At this point, our set A is an algebra, but it remains to be shown that this algebra
is indeed useful: that it provides a computational or conceptual advantage over the
standard approach to classical or quantum theory. In particular, it is important to classify
such algebras, and to understand their representations. To approach this question we
need to introduce topology on A.

Definition 2.2. We define a mapping } ¨ } : A Ñ R such that for each A P A,

(121)
@ D
}A} “ supt| φ; A | : φ P Su.

48
 Identifying observables with measurement devices, relation (121) assigns the largest
possible expectation value to each measurement. This mapping has almost all the prop-
erties of a norm: if }A} “ 0 then φ; A “ 0 for every φ P S, and since S separates the
@ D

observables we get A “ 0; by Axiom 3 each state is linear and thus, for A P A and λ P R
we have }λA} “ |λ|}A}; with B P A it is clear that }A ` B} ď }A} ` }B}. Moreover,
from the reality condition it also holds that }A2 } ď }A2 ` B 2 } for any A, B P A. The only
missing requirement such that (121) defines a norm, is that it is finite for every element
of A. We put this as an axiom.

Axiom 7. Each element A in A has finite norm and A is complete in the norm topology,
@ D
and we identify elements in S with positive continuous linear functional with φ; I “ 1.

The reason why we require the norm of observables to be bounded can be again un-
derstood with our identification of observables with measurement devices (even though,
by now the space A is significantly larger than it was in the beginning of our discussion).
Each practical measurement device has a finite (bounded) scale. Therefore, if A repre-
sents a measurement device then even the largest expectation value of A is going to be
bounded by the scale of that device. Measurements that contain infinity as a valid mea-
surement outcome are even theoretically not feasible. Therefore this axiom is reasonable
on physical grounds. Assuming that A is complete, on the other hand, is merely a practi-
cal requirement. Even though it does enlarge the space A, it does not significantly alter
the physical interpretation.
With this norm each state in S is by definition continuous. Thus a state φ P S is a
positive, linear, continuous functional on the algebra A with φpIq “ 1. Thus, S is the space
of states in the sense of our algebraic Definition 1.42 ii ). For that reason, henceforth we
will write SpAq instead of S and use our standard notation ωpAq instead of ω; A .
@ D

A non-trivial property of the introduced norm is stated in the following Lemma.

Lemma 2.3. For every observable A P A it holds

}A2 } “ }A}2 . (122)

Proof. By definition of the norm it follow that ωpAq ď }A} for every state ω P SpAq.
Thus by linearity of ω we have ωp}A}I ˘ Aq ě 0, so both }A}I ˘ A are positive. Thus
p}A}I ´ Aqp}A}I ` Aq is also positive (e.g. because it is uniquely determined by the
dispersion-free states/ two measurement devices both with positive pointer scale). Then
for every state ω P SpAq,

(123)
` ˘
}A}2 ´ ωpA2 q “ ω p}A}I ´ Aqp}A}I ` Aq ě 0,

which implies }A}2 ě }A2 }.

49
 On the other hand, the positivity of p}A}I ˘ Aq2 “ }A}2 ` A2 ˘ 2}A}A implies that
for any state ω we have

2}A}ωpAq ď }A}2 ` ωpA2 q ď }A}2 ` }A2 }, (124)

and therefore }A}2 ď }A2 }.

From the Proposition 1.50 we know that states induce an inner product on the vector
space on which they are defined; in this case on A by,

(125)
@ D ` ˘
ωpA ˝ Bq “ A, B A, B P A .

Moreover, we have seen that this product satisfies the Cauchy-Schwarz inequality. This
inequality, in turn, implies that the product is continuous since for A, B P A we have

|ωpA ˝ Bq|2 ď ωpA2 qωpB 2 q, (126)

for every ω P SpAq and every pair A, B P A. Taking the supremum on both sides of (126)
and using Lemma 2.3 implies }A ˝ B}2 ď }A2 }}B 2 } “ }A}2 }B}2 for every A, B P A.
A real Jordan algebra, which is a Banach space in the norm in which the product is
continuous is called a Jordan-Banach algebra. However, we have more structure. The
norm of our Jordan-Banach algebra also satisfies }A2 } “ }A}2 and the reality condition
}A2 } ď }A2 ` B 2 }. Such Jordan-Banach algebras are called JB-algebras.
In summary, A is a JB-algebra which is commutative, distributive, but not-necessarily
associative with a continuous product. From the mathematical point of view, however,
it is easier to treat an algebra which is associative but not-necessarily commutative. Be-
cause of this, it is convenient to ask the question when the JB-algebra can be embedded
into a larger algebra such that the symmetric product is given by the (associated but not
necessarily commutative) product AB such that
1` ˘ AB ` BA
A˝B “ pA ` Bq2 ´ A2 ´ B 2 “ . (127)
2 2
To achieve this, we can first complexify the algebra A by considering the set

tA ` ıB : A, B P Au, (128)

and then introduce the involution by ‹ : A ` ıB Ñ A ´ ıB. One can show that most of the
properties of the C ‹ -algebra are satisfied. Especially, one can extend the above norm such
that it satisfies the C ‹ -property }A‹ A} “ }A}2 and that }A‹ } “ }A}. However, the product
on this algebra may still be non-associative. This is the reason why the embedding may
sometimes fail.
A JB-algebra is called special if its embedding into a C ‹ -algebra exists; otherwise it
is called exceptional. Even though there is always a homomorphism from a JB-algebra
into a C ‹ -algebra [11, thm. 7.1.8], there may not exist an isometric isomorphism that

50
embeds a JB algebra into a C ‹ -algebra. A JB-algebra that is isometrically isomorphic to
a Jordan subalgebra of a C ‹ -algebra is called a JC-algebra. The following theorem [11,
thm. 7.2.3 and lem. 7.2.2] characterizes all exceptional JB algebras.

Theorem 2.4. Any exceptional JB-algebra A contains a unique purely exceptional ideal I
such that A{I is a JC-algebra. The ideal I is itself a Jordan algebra and each of its factor
representations is onto H3 pOq.

In the above Theorem H3 pOq denotes the algebra of 3 ˆ 3 self-adjoint matrices over
octonions. This algebra seems too poor for physical applications, and so an embedding
of a JB-algebra into a C ‹ -algebra is quite natural. We state this as a finial axiom.

Axiom 8. The algebra of observables A can be identified with the set of all self-adjoint
elements of a not necessarily commutative C ‹ -algebra A.

If such an embedding is possible then the JB-algebra (treated as a subset of self-

adjoint elements in A), generates the C ‹ -algebra A. In particular, completeness of A in
the JB-algebra norm already implies completeness of the ambient C ‹ -algebra, thus no
further closure assumption is needed.
We can summarize our above discussion as follows:
i ) A physical system is defined by its C ‹ -algebra of observables with identity.
ii ) The physical states are identified with the algebraic states on that C ‹ -algebra.
iii ) The dynamics of the system is a ‹-automorphism on the C ‹ -algebra of observables.

To conclude: We started with a set of observables A and states S; under mild physi-
cal assumptions we derived an algebraic structure on A. We further realized, that the
resulting algebra can be identified with a set of self-adjoint elements of a C ‹ -algebra A.
Now we have the entire algebraic machinery for C ‹ -algebras. In particular, by means
of the GNS construction we know that each state ω on A yields a cyclic representation
pπω , Hω , xω q. In particular, by the ‹-homomorphism property self-adjoint elements of A
are mapped to self-adjoint operators on the Hilbert space Hω . By the “Expectation” axiom
of Dirac-von Neumann and the unitarily equivalence clause of the GNS Theorem, we see
that the framework of quantum mechanics is unitarily equivalent to the so derived GNS
representation. Thus, we recovered the first three Dirac-von Neumann axioms.
The first three axioms of Dirac-von Neumann are of general nature. They do not refer
to a quantum system, but describe a physical system in general. The fourth and the fifth
axioms, on the other hand, refer specifically to a quantum system. It is for this reason
that with our general construction we recovered the general axioms. As we will see in
the next section, the fourth Dirac-von Neumann axiom will guide us to define an explicit
C ‹ -algebra for quantum mechanics. The fifth axiom drops entirely by the result of the
Stone-von Neumann theorem.

51
 Additionally we notice that in the algebraic formalism, the structure of quantum me-
chanics becomes derived, just like the structure of the phase space of classical mechanics
was derived from the commutative C ‹ -algebra. In particular, we see that the Hilbert
space of quantum mechanics is not a fixed construction of the framework, but a property
that arises from a choice of a physical state, i.e. from the method of preparation of a
physical system. Different methods of preparation may lead to different Hilbert spaces.
Indeed, in the standard approach to quantum mechanics, such transitions between states
that lead to different Hilbert spaces typically result in exploding physical quantities such
as the magnetization, susceptibility, conductivity, and the like. If this happens, we say
that a system undergoes a phase transition. Such an explosion of physical quantities
results from the fact, that the scalar product on the originally defined Hilbert space be-
comes inappropriate for the new physical system, since for the new system a new Hilbert
space is required. In the algebraic approach, such s transition corresponds to the change
between states ω1 and ω2 that lead to two not unitarily equivalent GNS representations
of the C ‹ -algebra. In this way, one avoids divergent quantities and can describe phase
transitions in a mathematically careful way.

52
3 A quantum mechanical particle

Section remarks: This section is based on several books and papers. The semi-
historical note is motivated by the discussion in the Heisenberg’s book [12] and his
papers [13, 14]. The uniqueness of the Weyl algebra is from [10]. The proof of the
Stone-von Neumann theorem and the Schrödinger representation of the Weyl algebra
is adopted from [6] and the original paper of von Neumann [15].

After the last section we know that we need a C ‹ -algebra of observables to define a
physical system. For classical physics the algebra has to be commutative. In this case,
we can choose the algebra of continuous functions on the phase space. The question is,
what C ‹ -algebra do we choose to describe quantum mechanics. To answer this question
we use the canonical commutation relations and the Heisenberg uncertainty principle.
Before we proceed, we mention a semi-historical note that shall motivate our approach.

3.1 Semi-historical note

Heisenberg realized that the physical quantities like position, velocity, acceleration, and
the like, may not be defined for sub-atomic systems (systems of small size). He argued
that all physical properties of a particle are defined by measurements. A position of a
particle is nothing else than an outcome of a measurement of position. If no experiment
can be devised to measure a certain quantity then this quantity does not “exist”. The
concept of measurement is unavoidably connected to the property itself.
As an example he considered an electron. One way to measure the electrons position
is to look at it under the microscope. The image we see under the microscope appears
from the lightrays that scattered from the electron. The shorter the wave length of used
light, the sharper will be the image and the better will be the resolution of the electron’s
position. From the photoelectric effect, described by Einstein several years prior, we
know that momentum p of light particles — the photons — is indirectly proportional to
the wavelength λ such that
~
p“ , (129)
2πλ
where ~ is the Planck’s constant (up to factors of 2π). In other words, the sharper the
image, the higher the momentum of photons. As the light scatters from the electron, the
high momentum of photons will transfer onto the momentum of the electron — a process
described and measured by the Compton scattering. Moreover, this transfer of momen-
tum will happen discontinuously, so after measuring the position the momentum of the
electron becomes unknown. Heisenberg estimated the accuracy of the measurements to

53
be
~
∆x∆p ě , (130)
2π
where ∆x and ∆p denote the variance of position and momentum measurements. He
suggested that this uncertainty is a fundamental principle of nature; it is impossible
to measure position and momentum of an electron with better simultaneous precision.
Today we call it the Heisenberg’s uncertainty principle.
To describe this phenomenon quantitatively, he suggested to use the wave picture
of the electron: the electron is a wave packet, electron’s position is the localization of
the packet, and electrons velocity is the group velocity of the packet. The sharper the
wave packet is localized in space the more spread it is in momentum, thereby causing
the packet to disperse in time. Thus, a sharply peaked wave packet quickly spreads
and its group velocity is not clearly defined. The analysis of the problem requires care,
because the aforementioned wave packet describes the probability distribution for the
electron and not the electron itself. Moreover, the dispersion of the packet depends on
the implemented dynamics, as for non-linear processes the wave package may keep its
shape for an extended time period. Nevertheless, the above heuristic argument already
captures the main idea of the rigorous calculation.
Heisenberg’s idea was groundbreaking and is considered as the initial formulation of
quantum mechanics. Nevertheless, it was Born and Jordan who realized that Heisen-
berg defined observables as infinitely large matrices, i.e. linear operators on an infinitely
dimensional Hilbert space [16]. In particular, being operators, observables may not al-
ways commute. Using Heisenberg’s uncertainty principle Born and Jordan derived that
the position and the momentum of an quantum particle must satisfy what we nowadays
know as the canonical commutation relations or CCRs,

(131)
` ˘
rxi , xj s “ 0 “ rpi , pj s, rxi , pj s “ xi pj ´ pj xi “ ı~1, i, j “ 1, . . . , d ,

where d is the number of degrees of freedom for the electron (e.g. a free particle moving
in a d-dimensional space; no spin is considered); and xj (resp. pj ) denotes the position
(resp. momentum) observable in the jth direction. Often we define x “ px1 , . . . , xd q and
p “ pp1 , . . . , pd q as operators on d copies of the Hilbert space H, and phrase (131) as a
single CCR relation

rx, ps “ xp ´ px “ ı~1. (132)

In summery, the Heisenberg’s uncertainty principle implies that the position and the
momentum observables of an electron do not commute. Moreover, the CCRs (131) imply
that either of the operators, xi or pi (or both), must be unbounded for every i “ 1, . . . , d.

54
To see this, we write the CCRs (dropping the index for clarity) as xp “ ı~ ` px. Then,

x2 p ´ px2 “ xpxpq ´ pxx

“ xpı~ ` pxq ´ pxx
“ ı~x ` pxpqx ´ pxx
“ ı~x ` pı~ ` pxqx ´ pxx “ ı~ 2x.

By induction we then get

rxn , ps “ xn p ´ pxn “ ı~ nxn´1 . (133)

The above relation implies that the norms of the operators x and p satisfy

~ n}xn´1 } “ }rxn , ps} ď }pxn } ` }xn p} ď 2}p}}x}}xn´1 }. (134)

Now, }xn´1 } cannot be zero, otherwise 0 “ }xn´1 } “ }x}n´1 would imply x “ 0 which
violates the CCR. Thus, (134) reduces to
n
(135)
` ˘
~ ď }q}}p} nPN .
2
Because n can be arbitrary large, at least one of the operators has to have an unbounded
norm. This, in turn, implies that we have to be careful with the multiplication of opera-
tors xi and pi . In particular, this multiplication can be defined only on an empty set —
implying that it is not meaningful. Moreover, even if the multiplication can be defined on
the dense subset of the Hilbert space, it cannot be defined on the whole space. Because of
this, the whole concept of “commutation” for unbounded operators is non-trivial. Many
technical complications arise when we want to formulate CCRs using clean mathematics.

We now return to our algebraic approach. In the algebraic approach, the mathemat-
ics is simple because all elements have finite norm. We only need to define what algebra
is suitable to define observables of a particle. Ideally, we would like to use x and p to
motivate the algebra, but since these operators are unbounded, they cannot generate a
C ‹ -algebra. Despite this, we can use x and p in a heuristic way (not rigorous calcula-
tions), to get an intuition for the right C ‹ -algebra. One way to do this was suggested by
Weyl and it leads us to the definition of the Weyl algebra.

3.2 The Weyl algebra

To define the Weyl algebra we use the operators x and p as if they were bounded, or at
least as if there were no problems with unbounded operators. The resulting calculation
will serve as a “hint” towards an appropriate C ‹ -algebra of observables.

55
 Since the concept of “commutation” is complicated for unbounded operators, Weyl
suggested to look at the operators

(136)
` ˘
U pαq “ eıαx V pβq “ eıβp α, β P Rd ,

where we use the abbreviation αx “ di“1 αi xi and βp “ di“1 βi pi . The operators U pαq
ř ř

and V pβq are called the Weyl operators, and the vectors α and β label these operators.
Since x and p suppose to represent observables we assume them to be self-adjoint. In
this case, the Weyl operators are bounded and can be defined via the spectral calculus.
We want to use the Weyl operators to generate our algebra of observables. For this we
need to define a product between them. To this end we use the CCRs and the Baker-
Campbell-Hausdorff formula.

Reminder: the Baker-Campbell-Hausdorff formulat. A special case of the Baker-

Campbell-Hausdorff formula states that if the commutator rX, Y s of two operators
X and Y commutes with both, X and Y , then
1
eX eY “ eX`Y ` 2 rX,Y s . (137)

Proof. For the proof and a more general statement consult [17, p. 25 prop. 2].

Formally, the CCRs say that rx, ps is proportional to the identity; thus it commutes with
every operator, and we can use the Baker-Campbell-Hausdorff formula to define the
following replacement of the CCR for the Weyl operators,

U pαqV pβq “ V pβqU pαqe´ı~αβ , U pαqU pβq “ U pα ` βq, V pαqV pβq “ V pα ` βq, (138)

where αβ “ di“1 αi βi is the Euclidean scalar product. These relations are called the
ř

commutation relations in the Weyl form or simply the Weyl relations. Even though U pαq
and V pβq together with (138) is enough to proceed, it will be more convenient in the
following to work with a single operator instead of two. To this end, we use vectors
in Rd ˆ Rd , such that the first d-components represent the vector α, and the second d-
components represent the vector β. Then for v “ pα, βq and w “ pγ, δq in Rd ˆ Rd p“ R2d q
we define the following operation,
˜ ¸˜ ¸˜ ¸
α 0 1 γ
σpv, wq “ αδ ´ γβ “ . (139)
β ´1 0 δ

Especially, σpv, vq “ 0 for every v P R2d . The mapping σ : R2d ˆ R2d Ñ R is a particular
case of what is called a non-degenerate symplectic bilinear form. It has a profound meaning
in classical mechanics. Formulating our C ‹ -algebra using σ is thus preferable as it can
give us a general procedure how to define a quantum algebra from the classical case. In
other words: how to quantize a classical system.

56
 Indeed, it turns out that if we use v “ pα, βq P R2d and set
αβ αβ
W pvq “ e´ı~ 2 V pβqU pαq “ eı~ 2 U pαqV pβq. (140)

then the Weyl relations can be derived from the following relations
σpv,wq
W pvq‹ “ W p´vq, W pvqW pwq “ W pv ` wqe´ı~ 2 . (141)

In the following, we will not use U pαq and V pβq but instead only the W pvq’s labeled by
vectors in R2d . For this reason we will call W pvq the Weyl elements and refer to (141)
as to the Weyl relations. Notice, however, that we can recover the operators U pαq and
V pβq by using vectors of the form v “ pα, 0q P R2d or v “ p0, βq P R2 , respectively. Even
though, the above calculations were formal, we can now use (141) as a starting point to
define a clean C ‹ -algebra.
We start with the set of abstract elements tW pvq : v P R2d u, that transform under a ‹-
operation and satisfy the multiplication relations according to (141). Then we consider
an abstract algebra AW generated by this set. That is AW comprises elements of the form

a1 W pv1 q ` a2 W pv2 q ` ¨ ¨ ¨ ` an W pvn q (142)

for a1 , . . . , an P C and v1 , . . . , vn P R2d . Notice that due to the multiplication relations, all
monomials of Weyl operators can be rewritten as a single Weyl operator multiplied by a
complex number. To form a C ‹ -algebra, we need to equip AW with a norm that satisfies
the C ‹ -property. But what norm shall we choose? The following theorem provides an
answer to this question. For the proof see [10, thm. 5.2.8].

Theorem 3.1. Let V be a real linear space equipped with a non-degenerate symplectic
bilinear form σ. Let A1 and A2 be two C ‹ -algebras generated by nonzero elements Wi pvq,
v P V , satisfying (141). Then there exists a unique ‹-isomorphism α : A1 Ñ A2 such that

(143)
` ˘
α W1 pvq “ W2 pvq, pv P V q.

Since ‹-isomorphisms are isometric, there is essentially a unique C ‹ -norm which

makes AW a C ‹ -algebra. In other words, the Weyl relations (141) generate a unique
C ‹ -algebra, up to ‹-isomorphisms — the Weyl algebra W.
From the Weyl relations (141) we can state some basic, nonetheless important prop-
erties of the Weyl algebra and the Weyl operators.

Proposition 3.2. Let W be the Weyl algebra generated by elements tW pvq : v P R2d u that
satisfies the Weyl relations. Then
i ) W is unital and W p0q “ I is the unit of W;
ii ) for each v P R2d the element W pvq is unitary;
iii ) for each non-zero v P R2 the spectrum of W pvq is the entire unit circle of C.

57
Proof. For i ), observe from the Weyl relations that

(144)
` ˘
W pvqW p0q “ W pvq “ W p0qW pvq v P R2d .

Since each element in W is a limit of finite complex linear combinations of the Weyl
elements, for every A P W we have AW p0q “ W p0qA “ A, and thus W p0q “ I is the unit
of the Weyl algebra; in particular, W is unital.
For ii ), we use again the Weyl relations to see that

v P R2d . (145)
` ˘
W pvqW pvq‹ “ W pvqW p´vq “ I “ W p´vqW pvq “ W pvq‹ W pvq

Thence, W pvq is unitary.

For iii ), observe that since W pvq is unitary, by Proposition 1.24 iii ), its spectrum is a
subset of the unit circle in C . Suppose v ‰ 0. Let λ be in the spectrum of W pvq and α
be a real number in the interval r0, 2πq. Choose w P R2d such that σpv, wq “ α{~, which
is always possible since v ‰ 0 and σ is linear and non-degenerate. Then,
´ ¯
(146)
ıα
`
´ıα ıα
˘ ´ıα
` ˘ ‹
W pvq ´ e λ “ e e W pvq ´ λ “ e W pwq W pvq ´ λ W pwq .

The right hand side does not have a two-sided inverse, and consequently eıα λ is in the
spectrum of W pvq. It follows that the spectrum is invariant under rotation and since λ is
an element of the unit circle, the spectrum must be the entire unit circle in C.

Remark. The Theorem 3.1 holds even if V is infinite dimensional. In this case,
the Weyl algebra describes a system with infinitely many degrees of freedom. Such
a theory is called (non-relativistic) quantum field theory. Opposed to it, quantum
mechanics describes systems with a finite number of degrees of freedom and thus
the Weyl algebra is labeled by a finite dimensional V .

To conclude: Using the canonical commutation relations that follow from the Heisen-
berg uncertainty principle we motivated a C ‹ -algebra for the quantum particle. This
algebra is essentially unique and we call it the Weyl algebra W. To connect our algebraic
formalism to the usual (Schrödinger) quantum mechanics, we need to represent the al-
gebra of observables as bounded linear operators on a Hilbert space. As we have seen, we
can use the GNS construction for this. Nevertheless, we have also seen, that not all rep-
resentations are necessarily unitarily equivalent. And if the Weyl algebra has different
unitarily inequivalent representations, then these representations will yield physically
different descriptions of the quantum particle. Which one should we then choose? In
particular, if different inequivalent representations exist, what is their physical signifi-
cance? And then, how many different inequivalent representations of the Weyl algebra
can we construct. All these questions were answered by von Neumann. The answer is
known as the Stone-von Neumann theorem.

58
3.3 Stone-von Neumann theorem
We now want to investigate unitarily inequivalent representations of the Weyl algebra.
Recall that, unitarily inequivalent representations lead to different Hilbert spaces on
which we can represent the C ‹ -algebra as a (sub)algebra of bounded operators. If the C ‹ -
algebra suppose to define a physical system, then the Hilbert spaces (especially their inner
products) define the expectation values. Thus they are physically relevant. If several
inequivalent representations exist then we need to select the one which is best suited for
the description of the system.
Recall, that irreducible representations are the building blocks of all representations.
Thus, we only need to characterize irreducible representations of the Weyl algebra. We
can restrict the class of “interesting” representations even further by requiring a mild but
physically well motivated regularity condition. To formulate this condition we need to
refresh the definition of the strong operator topology.

Definition 3.3 (Strong operator topology). Let BpHq be the set of bounded linear
operators on the Hilbert space H. For every x P H define a map from BpHq into H by

A Ñ Ax. (147)

The strong operator topology is the weakest topology such that each of these maps is
continuous. In this topology, a sequence of operators pAn q converges to an operator A,
if the sequence of vectors pAn xq converges to the vector Ax for every x P H. It is a
consequence of the uniform boundedness principle that a strong limit of a sequence of
bounded linear operators is again a bounded operator.

Definition 3.4. An operator-valued function U ptq defined on R and acting on the Hilbert
space H is called a strongly continuous one-parameter unitary group if
i ) for each t, s P R, the operator U ptq is unitary, and U pt ` sq “ U ptqU psq;
ii ) and for x P H, the vector U ptqx converges to the vector U pt0 qx as t Ñ t0 .

We mention some well known results of functional analysis to point out the impor-
tance of the strong operator topology.

Stone’s theorem. We state the theorems and refer to e.g. [3, thm. VIII.7 and thm.
VIII.8] for a more complete statement and the proofs.

Theorem 3.5. Let A be a self-adjoint operator on a Hilbert space H. The operator-

valued function

(148)
` ˘
U ptq “ eıtA tPR ,

defines a strongly continuous one-parameter unitary group. Moreover, for x P dompAq,

59
 ` ˘ ` ˘
the limit lim 1t U ptqx ´ x exists and is equal to ıAx; conversely, if lim 1t U ptqx ´ x
tÑ0 tÑ0
exists, then x P dompAq.

The Stone’s theorem gives the converse of Theorem 3.5.

Theorem 3.6. If U ptq is a strongly continuous one-parameter unitary group on a

Hilbert space H, then U ptq “ eıtA for some self-adjoint operator A on H.

Stone’s theorem states that we can “differentiate” strongly continuous one-parameter

unitary groups with respect to the parameter. The derivative being the generator of the
group. The generator need not be bounded, and Theorem 3.5 provides a useful criteria
for x P H to be in the domain of the generator. Since we motivated the Weyl algebra
by considering operators eıαx and eıβp , we would expect that the generators of the Weyl
elements would correspond to the position and momentum operators. However, in order
for the generators to exist, we need to make sure that the Weyl operators are represented
as a strongly continuous unitary group. This is the reason for the following definition.

Definition 3.7. A representation π of the Weyl algebra W on a separable Hilbert space

H is called regular if π W pvq is strongly continuous in v P R2d .
` ˘

For v P R2d of the form vn “ p0, . . . , 0, t, 0, . . . , 0q P R2d (where all except the nth ele-
ment are zero), the Weyl operators define one-parameter unitary groups. In conjunction
with the Stone’s theorem, the regularity condition assures that the operators π W pvn q
` ˘

will have generators. Hence, we have a chance to define positions and momenta from
the Weyl algebra. Since this is desirable on physical grounds, we restrict our attention
only to regular representations. Thus we aim to characterize all regular irreducible rep-
resentations of the Weyl algebra.

Remark: The regularity condition says that if pvn q is a sequence in R2d that con-
verges to a vector v, then the sequence of operators pπ W pvn q q will strongly con-
` ˘

verge to an operator π W pvq . Notice, that this is not necessarily the case in the
` ˘

C ‹ -norm since
` σpvn ,vq ˘ σpvn ,vq
}W pvn q ´ W pvq}2 “ }W p´vq eı~ 2 W pvn ´ vq ´ I }2 “ }eı~ 2 W pvn ´ vq ´ I}2 .

Assume that vn ´ v ‰ 0 for all n P N, then by Proposition 3.2 iii ), the spectrum of
W pvn ´ vq is the entire unit circle in C and we have
σpvn ,vq
}eı~ 2 W pvn ´ vq ´ I} ě supt|λ ´ 1| : λ P C, |λ| “ 1u “ 2. (149)

Thus the sequence W pvn q is not even Cauchy in the C ‹ -norm.

` ˘

If U : R2d Ñ BpHq is a strongly continuous operator-valued function bounded such

that supt}U pvq} : v P R2d u “ }U } ď 8, and f is a continuous and integrable function

60
with respect to the measure µ on R2d , then the operator
ż
A“ f pvqU pvq dµpvq, (150)
R2d

can be defined as a limit of Riemann sums. We will use this construction in the following
proof, therefore we review it in some detail now.
Pn

Pm
C2

C1

Figure 7: Example of two partitions Pn , Pm of R2 . Cells of partition Pn are the large

squares, like the cell C1 . Cells of partition Pm are the smaller squares, like the cell C2 .
Partition Pm is finer than partition Pn and thus we can order them as Pm ě Pn .

Integrals via Riemann sums. For a bounded set O Ă R2d we write Pn pOq “
tC1 , . . . Cn u to denote a partition of O into n (open) cells Ci Ă O. For two partitions
P1 pOq and P2 pOq we write P2 pOq ě P1 pOq and say that “P2 pOq is finer than P1 pOq” if
each cell C of P1 pOq partitions into a collection P2 pCq of cells in P2 pOq such that (see
Figure 7)
ď ÿ
C“ K µpCq “ µpKq. (151)
KPP2 pCq KPP2 pCq

Given a partition Pn pOq, choose one vector vC from each cell C and define an operator
ÿ
An “ f pvC qU pvC q µpCq. (152)
CPPn pOq

Observer, that since f U : R2d Ñ BpHq is strongly continuous, for x P H and  ą 0

there is a fine enough partition P0 pOq such that: for each cell C of P0

v, w P C implies }f pvqU pvq ´ f pwqU pwq} ď . (153)
µpOq

61
Then for any P1 pOq ě P0 pOq we have
› ÿ ÿ ›
}A0 x ´ A1 x} “ › f pvC qU pvC qµpCq ´ f pvK qU pvK qµpKq›
› ›
CPP0 pOq KPP1 pOq
› ÿ ÿ ` ˘ ›
f pvC qU pvC q ´ f pvK qU pvK q µpKq›
› ›
“›
CPP0 pOq KPP1 pCq
ÿ ÿ › ›
f f K ›µpKq
› ›
ď › Cpv qU pvC q ´ pvK qU pv q
CPP0 pOq KPP1 pCq

ď ,

where the second equality holds by (151) since P1 pOq is finer than P0 pOq. Thus, if
Pn pOq is an increasing sequence of partitions, i.e. Pn pOq ą Pm pOq when n ą m,
` ˘

then for every x P H there is a k P N for which Pk pOq is fine enough to satisfy (153)
and the sequence of vectors pAn xq is Cauchy; its limit defines a linear mapping AO
such that,
ż
(154)
` ˘
x ÞÑ lim An x “ f pvqU pvqx dµpvq “: AO x xPH .
nÑ8 O

Since the strong limit of bounded linear operators is bounded and thus AO defines
the integral of f U over O.
To see that the limit does not depend on the partition sequence, choose two
increasing sequences Pn pOq and P̃n pOq . Then there is a sequence P n pOq that is
` ˘ ` ˘ ` ˘

finer than both Pn pOq and P̃n pOq; for example define the cells of P n pOq as non-empty
intersections of each cell of Pn pOq with each cell of P̃n pOq. These three partition
sequences give rise to three sequences of operators pAn q for the partitions Pn pOq ,
` ˘

pBn q for the partitions P̃n pOq , and pCn q for the partitions P n pOq . From the above,
` ˘ ` ˘

these operator sequences converge to A, B, and C, respectively. Since P n pOq is finer

than other two we have that Cn Ñ A and Cn Ñ B, and by the uniqueness of limit
points we get A “ C “ B.
Now we need to “extend” O to the entire space R2d . For this let Bpnq denote an
open ball in R2d with radius n centered at the origin. Then for n ą m
›ż ż › ż
f pvqU pvq dµpvq ´ f pvqU pvq dµpvq› ď |f pvq|}U pvq} dµpvq
› ›
›
Bpnq Bpmq Bpnq{Bpmq
ż
ď }U } |f pvq| dµpvq.
Bpnq{Bpmq

Because f is integrable, the right hand side of the above estimate goes to zero as m
goes to infinity. Therefore the sequence of integrals over increasing balls is a Cauchy

62
 sequence of operators. The limit point of that sequence defines our desired integral
ż
A“ f pvqU pvq dµpvq. (155)
R2d

Notice that he matrix elements of A are the usual (Riemann) integrals of matrix
elements of U pvq since matrix elements of Riemann sums are the Riemann sums of
matrix elements. Explicitly, for x, y P H we have
ż
(156)
@ D @ D
x, Ay “ f pvq x, U pvqy dµpvq.
R2d

Since v Ñ x, U pvqy is a continuous bounded function, the integral is well defined.

@ D

Indeed we could use (156) as a definition for the operator A in terms of its matrix
elements. This construction of integral operators is often useful when we need to
define operators that are invariant under certain transformations. For this one has
to find the integral kernel f pvq such that the effect of the transformation can be
“averaged out”.

Now, we go back to the regular irreducible representations of the Weyl algebra. The
question is how many regular irreducible unitary inequivalent representations does a
Weyl algebra have. The Stone-von Neumann theorem provides the answer. To simplify
the notation, for the rest of this section we will assume ~ “ 1, which is always possible
to do by suitably redefining the units of meter, second, and kilogram. This is why ~ will
not appear in our commutation relations in the following.

Theorem 3.8 (Stone-von Neumann). All regular irreducible representations of the Weyl
algebra are unitarily equivalent.

Before starting the proof notice that each state on the Weyl algebra is uniquely defined
by its action on the Weyl elements. To see this, observe that each element A of the Weyl
algebra W is by definition a norm limit of linear combination of Weyl elements of the
form (142) such that
ÿ
A“ an W pvn q. (157)
nPN

Thence, if ω is a state on W then for each  ą 0 there is an N such that

N ˇ ´ N ¯ˇ ˇ ` ˘ ÿ N
ÿ ÿ ˘ˇˇ
(158)
`
 ě }A ´ an W pvn q} ě ˇω A ´ an W pvn q ˇ “ ˇω A ´ an ω W pvn q ˇ,
ˇ ˇ ˇ
n“0 n“0 n“0

and ωpAq is uniquely defined by the expectation values on the Weyl elements.

Proof of Theorem 3.8. The idea of proof is to show that if π is a regular irreducible rep-
resentation of W on the Hilbert space H then there exists a vector xF P H such that
2
(159)
@ ` ˘ D
xF , π W pvq xF “ e´|v| {4 .

63
 H1

BpH1 q xF |v|2
xF , π1 W pvq xF “ e´ 4
@ ` ˘ D
π1

H2
|v|2
yF , π2 W pvq yF “ e´ 4
@ ` ˘ D
π2 ` ˘ |v|2
W BpH2 q yF ωF W pvq “ e´ 4

H3
|v|2
π3 @
zF , π3 W pvq zF “ e´ 4
` ˘ D

BpH3 q zF

Figure 8: For every regular irreducible representation there is a vector xF (resp. yF , zF ),

that defines the same algebraic state ωF . By the GNS construction all the representations
are then equivalent, since they appear as the GNS representations of the state ωF .

Then such a vector defines an (algebraic) state that is uniquely specified by the relation
2
(160)
` ˘ @ ` ˘ D ` ˘
ωF W pvq “ xF , π W pvq xF “ e´|v| {4 v P R2d ,

and if xF is unique (up to normalization) then by the GNS construction, π is unitarily

equivalent to the GNS representation of the state ωF . In this case, we would have shown
that every regular irreducible representation of the Weyl algebra W is equivalent to the
GNS representation of the state ωF (Figure 8).
To construct such a vector xF we use the integral construction that we discussed
above. Let π be a regular irreducible representation of W on H. By Proposition 3.2 ii )
and since π is a ‹-homomorphism the operator π W pvq is unitary and thus bounded for
` ˘

all v P R2d such that supt}π W pvq } : v P R2d u ď 1. Since the function v Ñ e´}v} {4 is
` ˘ 2

continuous and integrable on R2d the conditions of our integral construction are satisfies
and we can define the operator
ż
1 |v|2
(161)
` ˘
P“ e´ 4 π W pvq dµpvq.
2π R2d

64
For x, y P H we have
ż
1 ´|v|2
(162)
@ D @ ` ˘ D
x, Py “ e 4 x, π W pvq y dµ
2π R2d
ż
1 ´|v|2
(163)
@ ` ˘ D
“ e 4 π W p´vq x, y dµ
2π R2d
ż
1 ´|v|2
(164)
@ ` ˘ D @ D
“ e 4 π W pvq x, y dµ “ Px, y ,
2π R2d
implying that P is self-adjoint. Moreover, P is not identically zero, otherwise for every
w P R2d and x, y P H we had
ż
1 |v|2 @
(165)
@ ` ˘ ` ˘ D ` ˘ D
0 “ π W pwq x, Pπ W pwq y “ e´ 4 x, π W pvq y eıσpv,wq dµpvq.
2π R2d
Writing v “ pα, βq and using the definition of σ we see that the right hand side of the
|v|2 @
above equation is the Fourier transform of the element e´ 4 x, π W pvq y . Thus (165)
` ˘ D

would imply that for every x, y P H the matrix element x, π W pvq y must vanish for
@ ` ˘ D

every v yielding that π W pvq “ 0 for every v P R2d , which is absurd.

` ˘

Next, we notice that by looking at the matrix elements Px, π W pzq Py for every
@ ` ˘ D

x, y P H we get the operator equality

ż
` ˘ 1 |v|2 |w|2 ` ˘
P π W pzq P “ 2 e´ 4 ´ 4 π W pz ` v ` wq e´ıpσpz,wq`σpv,z`wqq dµpwqdµpvq
4π R2d ˆR2d
Using the coordinate transformation v Ñ γ 21 pγ ´ z ` δq and w Ñ 21 pγ ´ z ´ δq, a lot of
patience, and Gaussian integrals, we eventually get (for a details see e.g. [15, p. 575])
|z|2
(166)
` ˘
P π W pzq P “ e´ 4 P.

Setting z “ 0 we get

P2 “ P. (167)

Since P is self-adjoint and non-zero it defines a non-trivial orthogonal projection on H.

In particular, there exists at least one vector x P H such that Px ‰ 0. Define
Px
xF “ . (168)
}Px}
Then xF “ PxF and for every v P R2d we get,
|v|2 @ |v|2
(169)
@ ` ˘ D @ ` ˘ D D
xF , π W pvq xF “ xF , Pπ W pvq PxF “ e´ 4 xF , PxF . “ e´ 4

To finish the proof, we need to show that such an xF is unique, that is that the range
of P is one-dimensional. Assume this is not the case, then there is a vector yF P H in the
range of P (i.e. PyF “ yF ), and orthogonal to xF . Then,
@ ` ˘ D @ ` ˘ D ´|v|2 @ D
yF , π W pvq xF “ yF , P π W pvq PxF “ e 4 yF , xF “ 0,

65
for every v P R2d , and thus by continuity yF , πpAqxF for every A P W. If yF ‰ 0, then
@ D

yF is in the orthogonal complement of the closure of the set of vectors tπpAqxF : A P Wu.
In particular, this set is a proper subset of H, invariant under the action of πpWq. This,
however, contradicts irreducibility of π. Therefore, yF “ 0.

Remark: It is a general fact that if π is an irreducible representation of an algebra A

on a Hilbert space H then every vector of H is cyclic. To see this assume that there
exists a non-cyclic vector z P H. In this case, the closure of tπpAqz : A P Wu Ă H
is a closed proper subspace of H that is invariant under the action of πpAq for each
A P W. This contradicts irreducibility.

The state ωF from the proof of the Stone-von Neumann Theorem is called the Fock
state. It is almost exclusively the only state that we use for explicit calculations in physics.
The GNS representation induced by the Fock state is called the Fock representation
and the GNS Hilbert space is (called) the Fock space.

In conclusion: To summarize the entire construction of quantum mechanics let us re-

view the steps:
i ) Using physical intuition we argued that a generic physical system can be described
by a C ‹ -algebra that defines observables of that system.
ii ) We realized that classical mechanics is described by commutative C ‹ -algebras, and
quantum mechanics is described by a non-commutative C ‹ -algebra.
iii ) To construct an explicit algebra for quantum mechanics we used the CCRs as a
guideline. From commutation of “position” and “momentum” we “guessed” the
product relations for the generating elements of the algebra (the Weyl relations).
Then theorem 3.1 showed us that there is essentially a unique C ‹ -algebra generated
by elements that satisfy the Weyl relations — the Weyl algebra.
iv ) The Stone-von Neumann theorem showed us that there is essentially only one reg-
ular irreducible representation of the Weyl algebra. So there is only one Hilbert
space (up to isometric isomorphism) that describes a quantum particle.
v ) Using the GNS construction we can explicitly construct the representation. At is
the point we are fully equipped to do explicit calculations.

3.4 The Schrödinger representation of the Weyl algebra

We finish this lecture by introducing the Schrödinger representation of quantum me-
chanics. It is the formulation of quantum mechanics that is commonly used in physics.
Opposed to the physics approach, however, in the algebraic approach we derived this

66
formulation from fundamental physical principles and rigorous mathematical analysis.
In particular, we know that the Schrödinger quantum mechanics is a representation of
an underlying algebra that describes a quantum particle. In this section, we only show
the general concepts, as we do not have time develop the details of the construction.
We begin with the Hilbert space L2 pRd q of square integrable functions with respect to
the Lebesgue measure. In the following, it will be convenient to get back to the original
Weyl operators U and V . So with v of the form v “ pα, 0q P R2d we define U pαq “ W pvq
and with w of the form w “ p0, βq P R2d we define V pβq “ W pwq. Now we define a
representation of the Weyl algebra. That is, we define the action of the Weyl operators
on the Hilbert space L2 pRd q. For ψ P L2 pRd q we set
(170)
` ˘ ` ˘
U pαqψ pxq “ eıαx ψpxq, V pβqψ pxq “ ψpx ` βq,
where the equality is meant to hold almost everywhere. After analyzing this action we
arrive at the conclusion that U pαq and V pβq define a strongly continuous irreducible
representation of the Weyl algebra.
Using the Stone theorem, we can derive the operators U pαq and V pβq with respect to
their parameters. The derivatives at the origin yield the generators of the Weyl operators
in the Schrödinger representation. We can explicitly calculate,
´ BU pαq ˇˇ ¯ ´ BV pβq ˇˇ ¯ ´ B ¯
´ı ˇ ψ pxq “ xψpxq ´ı ˇ ψ pxq “ ´ ı ψpxq, (171)
Bαi α“0 Bβi β“0 Bxi
for i P t1, . . . , du. For each i we call the generator of U pαi q the position operator qi and
group all these generators as q “ pq1 , . . . , qd q so that U pαq “ eıαq . Similarly, for each i we
call the generator of V pβi q the momentum operator pi and group them to p “ pp1 , . . . , pd q
so that V pβq “ eıβp . If we use our notation id pxq “ x to denote the identity function and
∇ to denote the Euclidean gradient on Rd then (171) defines the action of q and p on
L2 -functions as
qψ “ id ¨ψ pψ “ ´ı∇ψ. (172)
These operators map L2 -functions into d-copies of L2 , that is into the space of square
integrable vector fields L2 pRd ; Rd q. In particular, we see that neither q nor p are bounded.
Their domains of definition follow from the Stone’s theorem such that
dompqq “ tψ P L2 pRd q : xψ P L2 pRd ; Rd qu, (173)
domppq “ tψ P L2 pRd q : ∇ψ P L2 pRd ; Rd qu “ H 1 pRd q, (174)
where H 1 pRd q denotes the Sobolev space of square integrable functions, whose deriva-
tives are also square integrable. Especially, on these domains q and p are self-adjoint.
Their product can be defined on the common core domain of rapidly decreasing func-
tions — the Schwarz space SpRd q — and on that space we can easily calculate
(175)
` ˘
pqp ´ pqqf “ ´ıx∇f ` ı1f ` ıx∇f “ ı1f f P SpRd q .
So the CCRs are satisfied on the core of q and p.

67
 Remark: Notice, that we could as easily choose a different representation in which
U pαq would act by transition and V pβq would act by multiplication. In this case, we
would obtain the Fourier transform of the above choice. Notice, that this also implies
that the definition of position operator as a multiplication operator is ambiguous. It
really depends on the Hilbert space we choose. In fact, because of this symmetry
between representations, it is hardly possible to distinguish between position and
momentum operator at all. The only distinguishing feature will be the dynamics of
the system, so we can realize which is which, by looking at their time evolution.

Pure state on the algebra of observables correspond to unit-normalized functions in

L pRd q. For each unit vector ψ P H we can define the state ωψ such for A P BpHq we get
2

ż
ωψ pAq “ ψpxqpAψqpxq dλpxq. (176)
Rd

If A is a multiplication operator by some function F pxq then A generates a commutative

C ‹ -subalgebra of BpHq, and ωψ defines a state on this algebra. In particular, as in the
case of the classical mechanics we get the probabilistic interpretation of the state ωψ as
ż ż
ωψ pAq “ ψpxqpAψqpxq “ |ψpxq|2 F pxq dλpxq. (177)
Rd Rd

where |ψ|2 pxq defines the probability distribution finding the particle at position x. This
interpretation is possible for the algebra generated by A. In general, however, it is not
possible to find a state that has a probabilistic interpretation for all observables. This
distinguishes quantum mechanics from the classical case.
The next step is be to define the dynamics of the particle by defining the evolution
operator. This, however, is beyond the scope of this lecture.

68
References
[1] R. V. Kadison and J. R. Ringrose. Fundamentals of the theory of operator algebras
Volume I. Vol. 1. American Mathematical Soc., 1997.
[2] R. V. Kadison and J. R. Ringrose. Fundamentals of the theory of operator algebras
Volume III. Vol. 3. American Mathematical Soc., 1997.
[3] M. Reed. Methods of modern mathematical physics I: Functional analysis. London:
Academic Press, 1980.
[4] S. Sakai. C*-algebras and W*-algebras. Heidelberg, Berlin: Springer Science &
Business Media, 2012.
[5] G. K. Pedersen. Analysis now. Vol. 118. New York: Springer Science & Business
Media, 2012.
[6] F. Strocchi. An introduction to the mathematical structure of quantum mechanics:
a short course for mathematicians. Vol. 28. Singapore: World Scientific, 2008.
[7] G. G. Emch. Algebraic methods in statistical mechanics and quantum field theory. In-
terscience monographs and texts in physics and astronomy. Rochester, New York:
Wiley-Interscience, 2009.
[8] D. Ruelle. Statistical mechanics: Rigorous results. 2nd. New York: W. A. Benjamin
Inc., 1974.
[9] O. Bratteli and D. W. Robinson. Operator Algebras and Quantum Statistical Me-
chanics 1: C*-and W*-Algebras. Symmetry Groups. Decomposition of States. Vol. 1.
Berlin, Heidelberg: Springer Science & Business Media, 2012.
[10] O. Bratteli and D. W. Robinson. Operator Algebras and Quantum Statistical Me-
chanics 2: Quantum states and Models in Quantum Statistical Mechanics. Vol. 2.
Berlin, Heidelberg: Springer Science & Business Media, 2012.
[11] H. Hanche-Olsen and E. Størmer. Jordan operator algebras. Vol. 21. London: Pit-
man Advanced Publishing Program, 1984.
[12] W. Heisenberg. The physical principles of the quantum theory. New York: Dover
Publications, INC., 1949.
[13] W. Heisenberg. “Über quantentheoretische Umdeutung kinematischer und mech-
anischer Beziehungen.” In: Zeitschrift für Physik 33.1 (1925), pp. 879–893.
[14] W. Heisenberg. “Über den anschaulichen Inhalt der quantentheoretischen Kine-
matik und Mechanik”. In: Zeitschrift für Physik 43.3 (1927), pp. 172–198.
[15] J. v. Neumann. “Die eindeutigkeit der Schrödingerschen operatoren”. In: Mathe-
matische Annalen 104.1 (1931), pp. 570–578.

69
[16] M. Born and P. Jordan. “Zur quantenmechanik”. In: Zeitschrift für Physik 34.1
(1925), pp. 858–888.
[17] W. Rossmann. Lie groups: an introduction through linear groups. Vol. 5. Oxford:
Oxford University Press on Demand, 2006.

70