3.

Quantum mechanics as a statistical

theory
For this course, we explicitely do not assume any prior working knowledge in quantum
theory nor do we intend to provide such. This fact should not cause any depressions,
since knowledge of the physical content of quantum theory and its philosophical im-
plications is not essential for understanding quantum information theory (at least not
the part presented in this course.)
The first - comparatively informal - part of this lecture aims to give a hint how quan-
tum theories are connected to the real world - we will stress the statistical nature of
quantum theory, which allows to recognize a conceptual link to classical information
theory. Second, we give explicit mathematical definitions of some main entities of the
theory and some facts about their structure, which are essential for quantum informa-
tion theory.

3.1. Statistical Theories

Experience shows, that microscopic objects (e.g. atoms, photons,...) show a random
behaviour in experiments. Measurement outcomes fluctuate. As far as experiments
have demonstrated, the so-called statistical axiom is fulfilled. The statement of this
assumption reads as

• Statistical axiom If a measurement is performed on independent and equally

prepared systems in the same condition many times, the relative frequencies

number of occurencies of the measurement value i

f i :=
number of total measurements
stabilize (i.e. converge to probabilities)

The general scheme of a statistical experiment is usually divided into two

• preparation Setting the initial conditions.

• registration/measurement Setting rules what events have to be registered.

These considerations may seem rather abstract. In fact, this framework fits well to what
we do implicitely also in classical theory. A very simple example is how experiments
with a six-face are described. In this case, the preparation is set by fixing probabilities

15
3. Quantum mechanics as a statistical theory

for occurencies of the several faces (which we denote just by their eye numbers in the
following). So setting the preparation is fixing numbers p(1), . . . p(6) determines the ini-
tial conditions for all experiments with the dice. It is clear, that nearly most of the dies
produced fulfill p(1) = · · · = p(6) = 1/6 (fair preparation). To describe the registration
step in an experiment then is to give events to be registered when throwing the dice.
Usually it is suitable to register the number of eyes which appear on the upside of the
dice. Another possibility would be to just register if the eye number is even or odd.
Then the event that the eye number is odd would be given by the set E := {1, 3, 5, 7}. It
is intuitively clear how to derive the probability that the mentioned event occurs,

q(E) := q(i). (3.1)

i E

To put it short, for dice experiments the possible preparations is in one-to-one rela-
tionship with the probability distributions P (X ) on the set X := {1, . . . , 6}, while each
possible registration rule (or measurement) is given by a family {E}i I of subsets of X
which are mutually disjoint and union over all sets gives X . In the language we will
use a probability distribution on the set X will be called a state of the dice while a de-
composition of X is called a (dice) observable. It is clear, that in this easy example, it is
completely sufficient to register the “singleton” events, i.e. perform the measurement
described by the family of sets {1}, . . . , {6}. All other probabilities can be inferred from
it.

• include convexity, measurement -¿ affine map

The above discussion is definitely too narrow to really get the concepts. For the
reader being interested in more information on the conception and mathematics of
statistical theories we give the following list of references (ordered with increasing level
of formalization.)
1. A. Holevo, Quantum Systems, Channels, Information, de Gruyter, 2012, Chapter 2.

2. A. Holevo, Probabilistic and Statistical Aspcects of Quantum Theory, Edizione de

Scuola Normale Superiore Pisa, 2011, Chaper 1

3. S. Gudder, Stochastic Methods of Quantum Mechanics, Dover Publications, 2005,

Chapter 4

3.2. Quantum States and Measurements

In this section, we give the explicit mathematical objects, which represent states and
observables in case of quantum theory. In theoretical physics, the determination of
what is a “state” or an “observable” is usually subject to axioms, i.e. claims about the
relations of mathematical objects to real-world observations and experiments. Since
we are only interested in applying the corresponding mathematical theory to commu-
nication theoretic questions, we give definitions instead. Usually one of the axioms of

16
 3.2. Quantum States and Measurements

quantum mechanics is that to each sort of systems described by quantum theory can
be assigned a Hilbert space whose structure (i.e. dimension) derives from the degrees
of freedom of the systems. For this course, this terminology is a bit exaggerated, since
we only deal with systems whose assigned Hilbert space is finite-dimensional. Since
also in this special case the term Hilbert space is standard in the literature, we set the
following
Convention 8. When using the term “Hilbert space”, we mean a finite dimensional complex
euclidean space equipped with the standard scalar product.
Anyquantumstmurgggfmberntswh.se
Definition 9. The state of a quantum system with Hilbert space H is a density matrix on
H, i.e. a member of the set
describedbydensity matrix
S (H) := { S (H) : positive semidefinite, and tr( ) = 1}. (3.2)

Definition 10. An observable or a (quantum) measurement on a quantum system with

Hilbert space H with finite set X of measurement outcomes is a positive operator valued
measure (POVM), i.e. a family

{Ex }x X ⇥ L(H), (3.3)

having the properties

1. 0 ⇤ Ex ⇤ H for all x X , and
⇥
2. x X Ex = H .

tr Ey (3.4)

for each y Y.
Exercise 11. Drop any condition in the Definition of a state of an observable (e.g. positive
semidefiniteness in Def. 9). Construct an example, where this modification implies the rule
in Eq. (3.4) does not lead to a probability distribution.
As the definitions suggest, in the statistical theory we will use, the density matrices
describe the preparation of the sytems, while POVS describe the registrations. In the
last paragraph we noticed, that convexity is an essential property of the set of prepara-
tion procedures. Mixing preparations should again lead to a preparation. The follow-
ing proposition ensures us, that §(H) is actually a convex set, and moreover provides
us with a characterisation of the exetremal elements of it.
Proposition 12. Let H be a Hilbert space. The following claims hold.
1. S (H) is a convex subset of L(H)

2. §(H) is an extremal element (called a pure state of S (H), if and only if is a rank-1
projector.

17
3. Quantum mechanics as a statistical theory

Proof. We first show convexity of S (H). Let ⇥1 , ⇥2 S (H) be any two density marices
on H, and ⇤ (0, 1). We define := ⇤⇥1 + (1 ⌅ ⇤)⇥2 . Positive semidefiniteness is easy to
check from the definitions, it holds for each x H

⇧x, x⌃ = ⇤ ⇧x, ⇥1 x⌃ + (1 ⌅ ⇤) ⇧x, ⇥2 x⌃ (3.5)

by linearity of the scalar product. It is easily observed, that the right hand side of Eq.
(3.5) is nonnegative. Also, we have

tr( ) = ⇤tr(⇥1 ) + (1 ⌅ ⇤)tr(⇥2 ) = ⇤ + (1 ⌅ ⇤) = 1. (3.6)

Consequently, is also a density matrix. For the second claim we first show the ⌥
implication. Assume, that is a rank one projection, i.e. it holds
2
= . (3.7)

We show, that any convex combination

= µ⌅1 + (1 ⌅ µ)⌅2 (3.8)

is necessarily trivial. It holds

2
⌅ = µ⌅1 + (1 ⌅ µ)⌅2 ⌅ µ2 ⌅12 ⌅ µ(1 ⌅ µ)(⌅1 ⌅2 + ⌅2 ⌅1 ) ⌅ (1 ⌅ µ)2 ⌅22 (3.9)
= µ(⌅1 ⌅ ⌅12 ) + (1 ⌅ µ)(⌅2 ⌅ ⌅222 ) + µ(1 ⌅ µ)(⇥2 )2 (3.10)
µ(1 ⌅ µ)2 (⌅1 ⌅ ⌅2 )2 (3.11)
0. (3.12)

The equalities above are by rearranging terms. The first inequality above (notice: the
inequality is a matrix inequality in the hermitian semiorder) is by the fact, that µ and
1 ⌅ µ are nonnegative and (⌅1 ⌅ ⌅12 ) as well as ⌅2 ⌅ ⌅22 are positive semidefinite matrices
(check this.) The second inequality follows, because (⌅1 ⌅ ⌅2 )2 is positive semidefinite.
Since ⌅ 2 = 0 ( is assumed to be a projections), it holds

µ(1 ⌅ µ)(⌅1 ⌅ ⌅2 )2 = 0. (3.13)

But this is only possible if µ {0, 1} or ⌅1 = ⌅2 , which is the case if and only if the
convex combination in Eq. (3.8) is trivial. For the remaining implication, let be an
extremal element of S (H). Consider a spectral decomposition
dim H
= ⇤i |⇧i ⌃ ⇧⇧i | (3.14)
i=1

of . Notice that this is a convex combination of . Since is assumed to be extremal

there is exactly one i0 with ⇤i0 = 1, while ⇤i = 0 for all i i0 . Consequently = |⇧i0 ⌃ ⇧⇧i0 |,
a rank one projection.

18
 3.3. Example: Qubit systems

Remark 13. (i) The extremal elements of S (H) are usually called pure states. Each unit
vector v H gives rise to a pure state = |v⌃ ⇧v|. the correspondence v ⌦ |v⌃ ⇧v| is
one-to-one up to global phases, i.e. for ⌃ , v and ei⌃ v give rise to the same pure
state.

(ii) The classical counterpart to S ( d ) is the set P ({1, . . . , d}) of probability distributions
on {1, . . . , d}. The extremal elements of this convex set are the dirac measures ⌥1 , . . . ⌥d
which are defined by
⇤
⌅
⇧1 if x = y
⌅
⌥x (y) := ⌅
⌅ (3.15)
⌃0 otherwise.

3.3. Example: Qubit systems

To get into calculations with the mathematical objects defined above, we consider the
case of a quantum system with two degrees of freedom, i.e. the underlying Hilbert
space is two-dimensional. Despite the fact, that such systems are often considered in
physics1 , they can be regarded as quantum counterparts of classical bit systems bit
systems with alphabet |X | = 2. The set S (C 2 ) of qubit states has a convenient pictoral
representation in 3 , which we derive next. The matrices
⌥ ⌥
1 0 0 1
⇥0 = ⇥1 =
0 1 1 0
⌥ ⌥
0 ⌅i 1 0
⇥2 = ⇥3 =
i 0 0 ⌅1

are called Pauli matrices and form an orthogonal basis in L( 2 ), it holds

⇧⇥i , ⇥j ⌃HS = 2⌥ij . (3.16)

↵
If we normalize each of the matrices with a factor 1/ 2 we obtain an orthonormal basis.
We can write each matrix A L(H) as a linear combination
3
1
A= ri ⇥i . (3.17)
2
i=0

with ri = ⇧A, ⇥i ⌃HS . We aim to derive conditions on the numbers r0 , . . . , r3 being equiva-
lent to A S (H). We have
1. Each ri has to be real, because A is Hermitian.

2. r0 = 1 holds because of the property tr(A) = 1.

1 Real-world examples of such systems are e.g. the spin of an electron system or the polarization of light.

19
3. Quantum mechanics as a statistical theory

3. Since A 0 holds,
1 2 2 2 2
(r ⌅ r ⌅ r ⌅ r ) = det A 0. (3.18)
4 0 1 2 3
Since r0 = 1, we obtain the condition r ⇤ 1 for the vector (r1 , r2 , r3 )T 3.

On the other hand, if A is represented as in (3.17) with r0 = and (r1 , r2 , r3 )T an element

of B1 (0), the euclidean ball around 0 with radius one, then A S (H) is implied. Indeed,
1 = r0 = trA, and det A = 1⌅r12 ⌅r22 ⌅r32 0. Consequently, A is a density matrix. Since the
basis coefficients r0 , r1 , r2 , r3 of an element of L(H) are unique, the map which connects
each density matrix with its bloch vector (r1 , r2 , r3 )T is a one-to-one. By linearity of
the Hilbert-Schmidt scalar product, it is clear that this map is also affine. Therefore, we
have introduced an affine bijection of S (H) onto the radius one euclidean ball in 3 . By
this fact, it is clear, that the set of extremals of S (H) correspond to the set of extremals
of B1 (0), i.e. the unit sphere around 0 in 3 .
Remark 14. In quantum optics, it is common, to specify the polarization preparation of a
leaser beam by giving the corresponding bloch vector r = (r1 , r2 , r3 )T . Examples are (according
to www.wikipedia.de)
⌦⌦ 1 ⌦⌦ ⌅1
⌦⌦ ⌦⌦
⌦⌦ 0
⌦↵ (linear horizontal) ⌦⌦ 0
⌦↵ (linear vertical)
0 0
⌦⌦ 0 ⌦⌦ 0
⌦⌦ ⌦⌦
⌦⌦ 1
⌦↵ (linear 45 ) ⌦⌦ ⌅1
⌦↵ (linear -45 )
0 0
⌦⌦ 0 ⌦⌦ 0
⌦⌦ ⌦⌦
⌦⌦ 0
⌦↵ (right circular) ⌦⌦ 0
⌦↵ (left circular)
1 ⌅1
⌦⌦ 0
⌦⌦
⌦⌦ 0
⌦↵ (unpolarized).
0
By calculating their block vectors, one can verify, that the pure states Pi := |ei ⌃ ⇧ei |,
i {0, 1} are located at the north and south poles of the Bloch ball. The set of states
which lie on the straight line connecting the pure states P0 and P1 are parameterized by
probability distributions on {0, 1},
{P(p) := p(0)P0 + p(1)P1 } .
In fact, each straight line connecting two antipodes on the Bloch sphere can be regard-
ing as a version of the classical bit states. Proceeding with this discussion, we observe
an essential di⇥erence between the set of bit and the set of qubit states. ???
Exercise 15. Calculate the Bloch vectors of the states P± := |f ± ⌃ ⇧f ± | where
f ± := ↵1 (e0 ± e1 ).
2
Where are their images located on the Bloch ball?

20
 3.4. Exercises

3.4. Exercises
Exercise 16. Prove, that the set of POVMs on a Hilbert space all indexed by the same alpha-
bet form a convex set.

Exercise 17. Consider the set of probability distributions P (X ) on the alphabet X := {1, 2, 3}
find a useful geometric picture in 3 .

Exercise 18. Where are the pure qubit states located in the Bloch ball, where the maximally
mixed state
⌥ 1
2 0
2 := (3.19)
0 12

Exercise 19. Consider the orthonormal basis B := {f + , f ⌅ } defined by

f + := ↵1 (e0 + e1 ) and f ⌅ := ↵1 (e0 ⌅ e1 ). (3.20)

2 2

Calculate the Bloch vectors r1 , r2 , r3 , and draw them into the Bloch ball picture.

21